text
stringlengths 87
880k
| pmid
stringlengths 1
8
| accession_id
stringlengths 9
10
| license
stringclasses 2
values | last_updated
stringlengths 19
19
| retracted
stringclasses 2
values | citation
stringlengths 22
94
| decoded_as
stringclasses 2
values | journal
stringlengths 3
48
| year
int32 1.95k
2.02k
| doi
stringlengths 3
61
| oa_subset
stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|
==== Front
World J Surg OncolWorld Journal of Surgical Oncology1477-7819BioMed Central London 1477-7819-3-701623232510.1186/1477-7819-3-70Case ReportPost-ERCP pancreatogastric fistula associated with an intraductal papillary-mucinous neoplasm of the pancreas – a case report and literature review Koizumi Masaru [email protected] Naohiro [email protected] Koji [email protected] Munetoshi [email protected] Katsumi [email protected] Yoshikazu [email protected] Hideo [email protected] Department of Surgery, Jichi Medical School, 3311-1, Yakushiji Minamikawachi Tochigi, 329-0498, Japan2005 19 10 2005 3 70 70 3 6 2005 19 10 2005 Copyright © 2005 Koizumi et al; licensee BioMed Central Ltd.2005Koizumi et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Fistula formation has been reported in intraductal papillary-mucinous neoplasms (IPMNs) with or without invasion of the adjacent organs. The presence or absence of invasion is mostly determined by postoperative histological examination rather than by preoperative work-up.
Case presentation
A 72 year-old Japanese woman showed remarkable dilatation of the main pancreatic duct (MPD) in the distal region of the pancreas. Subsequent ERCP also showed MPD dilatation, after which the patient suffered moderate pancreatitis. A subsequent gastroscopy revealed a small ulceration that had not been observed in a gastroscopy performed 3 months prior. Mucinous discharge from the ulceration suggested it might be the orifice of a fistula connected to the MPD. En bloc resection including the distal region of the pancreas, spleen, stomach and part of the transverse colon was performed under the pre- and intraoperative diagnosis of an invasive malignant IPMN. However, histopathology revealed the lesion to be of "borderline malignancy" without apparent invasion of the stomach. Light microscopy showed inflammatory cellular infiltrates (mainly neutrophils) around the pancreatogastric fistula, but there was no evidence of neoplastic epithelia lining the fistulous tract.
Conclusion
This case highlights that a pancreatogastric fistula can develop after acute inflammation of the pancreas in the absence of cancer invasion. Further information regarding IPMN-associated fistulae is necessary to clarify the pathogenesis, diagnosis, appropriate surgical intervention and prognosis for this disorder.
==== Body
Background
Intraductal papillary-mucinous neoplasms (IPMNs) of the pancreas have unique clinico-pathological characteristics and show a wide spectrum of histological types, ranging from adenomatous hyperplasia to invasive cancer. While non-invasive IPMNs show slow growth and good prognosis after resection, the outcome can become poor if they transform into invasive ductal cancers. There are three types of IPMN: main duct,, branch, and combined. The main duct and combined types are associated with higher rates of invasive cancer than the branch type [1].
A feature of IPMNs is occasional fistula formation with surrounding organs such as the duodenum, bile ducts, stomach and even the peritoneal or pleural cavity [2-10]. While such fistula formation is generally thought to be associated with invasive IPMNs, some non-invasive IPMNs also form fistulas. Thus, IPMN fistula formation raises intriguing issues concerning pathogenesis, diagnosis and patient prognosis.
We report here a rare non-invasive "borderline-malignant" IPMN that formed a pancreatogastric fistula. Fistula formation was triggered by acute pancreatitis following endoscopic retrograde cholangiopancreatography (ERCP).
Case presentation
A 72 year-old Japanese woman underwent abdominal ultrasonography (US) as part of a routine physical checkup in April 2001. This procedure incidentally detected marked dilatation of the main pancreatic duct (MPD) in the distal region of the pancreas. Magnetic resonance cholangiopancreatography (MRCP) and abdominal computed tomography (CT) showed MPD dilatation and a multi-lobular cystic lesion. ERCP confirmed these findings, and no extraluminal leakage of the contrast medium was observed (Figure 1). Cytology on pancreatic fluid indicated non-neoplastic epithelia. Endoscopic ultrasonography (EUS) identified 8 mm papillary projections in the cystic lesion, which was a key indication for surgery. A routine gastroscopy performed simultaneously with EUS showed no significant findings.
Figure 1 ERCP on presentation showing a dilated MPD and cystic lesions in the distal region of the pancreas.
Following the ERCP, the patient complained of abdominal pain and showed hyperamylasemia and leukocytosis. Distinct swelling of the pancreas and thickening of the gastric wall were subsequently observed by abdominal CT (Figure 2), which indicated a mild post-ERCP pancreatitis. Gastroscopy performed for the assessment of gastric wall thickness revealed a small ulceration in the body of the stomach (Figure 3), which was accompanied by a mucinous discharge suggesting the presence of a fistula opening into the MPD. These findings, along with abdominal angiography data indicating encasement of the splenic artery suggested a malignant invasion into the stomach by the primary lesion. All other laboratory data, including tumor marker expression (e.g., CEA and CA19-9), were within normal limits.
Figure 2 Abdominal CT just prior to surgery showing distinct swelling of the pancreas and the gastric wall, causing the border between the two organs to be unclear.
Figure 3 Gastroscopy just prior to surgery showing a small ulcer producing a mucinous discharge into the body of the stomach.
A laparotomy was performed 1 month after the ERCP. Firm fibrous tissue between the pancreas and the mesenterium of the transverse colon suggested the pancreatic neoplasm was of an invasive nature. An en bloc resection of the distal pancreas, spleen, stomach and part of the transverse colon was performed. Postoperative histopathology showed the distal pancreas had an IPMN whose lining epithelia showed only "borderline-malignancy" and no apparent invasion of adjacent tissues or organs. Light microscopy examination revealed inflammatory cellular infiltration consisting mainly of neutrophils around a pancreatogastric fistula which had no neoplastic epithelia (Figure 4). The patient was discharged after an uneventful postoperative recovery and showed no signs of recurrence 3 years after surgery.
Figure 4 A cross-section of the pancreatogastric fistula (Hematoxylin and Eosin stain × 40). The fistula was significantly infiltrated with neutrophils and showed no evidence of neoplastic infiltration.
Discussion
Fistula formation associated with an IPMN was first reported by Ohhashi et al in 1980 as a pancreatobiliary fistula [2]. Since then there have been 41 cases reported between 1980 and 2004 (Table 1) [3-9]. According to these reports, the organs most frequently affected by fistula formation were the duodenum (24 cases, 59%), common bile duct (21 cases, 51%) and stomach (7 cases, 17%). Apparent cancer invasion was the main cause of the fistulae in 16 cases (37%), while similar to the present case, 17 cases (41%) showed fistulae into neighboring organs in the absence of tumor invasion (Table 2). Five of the six previous IPMN cases associated with pancreatogastric fistulae were linked to apparent cancer invasion.
Table 1 Published cases of developing fistulae from IPMNs (1980–2004)
No. Author Year Affected Organs Cancer Invasion Pathological Diagnosis
1 Ohhashi [2] 1980 C - cystadenocarcinoma
2 Koike [3] 1980 C, D unknown pap
3 Sukisaki [3] 1984 S, D + pap (+ muc)
4 Usui [3] 1985 D - pap
5 Suyama [3] 1986 C + pap
6 Suyama [3] 1986 D + pap
7 Yamao [3] 1986 D unknown pap
8 Yamao [3] 1986 C unknown pap
9 Yamao [3] 1986 C unknown pap
10 Miyagawa [3] 1988 C, D + pap (+ muc)
11 Kuga [3] 1988 D - pap
12 Tohyama [3] 1989 C - pap
13 Nakajima [3] 1989 C - pap
14 Ohnuma [3] 1990 S + pap (+ muc)
15 Sakimoto [3] 1991 C + pap (+ muc)
16 Mayumi [3] 1991 C, D + muc (+ pap)
17 Hayashi [3] 1991 C unknown cystadenocarcinoma
18 Oda [3] 1993 C, D - pap
19 Saitoh [3] 1993 C - pap
20 Kobayashi [4] 1993 S + muc (+ pap)
21 Kobayashi [4] 1993 S, D + muc (+ pap)
22 Kobayashi [4] 1993 C, D unknown pap
23 Kobayashi [4] 1993 D unknown pap
24 Kobayashi [4] 1993 D unknown muc (+ pap)
25 Mori [3] 1994 C + pap
26 Ieda [3] 1995 D + tub
27 Yago [5] 1995 S - pap
28 Takekuma [3] 1996 D - pap
29 Nakamura [3] 1996 C, D + pap
30 Murata [3] 1996 C, D - pap
31 Hirota [3] 1996 D - pap
32 Tadokoro [6] 1996 D - unknown
33 Nakatsuka [3] 1997 S, D + muc
34 Matsubayashi [7] 1998 D + pap
35 Kawaharada [3] 1999 D + pap
36 Kawaharada [3] 1999 D + pap
37 Shiroko [8] 2000 C, D - unknown
38 Kurihara [9] 2000 C - pap
39 Kurihara [9] 2000 C - pap
40 Fujisawa [3] 2001 C - pap
41 Koizumi(present case) 2004 S - low grade malignancy
Affected Organ; C: Common bile duct, D: Duodenum, S: Stomach
Pathological Diagnosis; pap: papillary carcinoma, muc: mucinous carcinoma
tub: tubular adenocarcinoma
Table 2 Organs developing IPMN-associated fistulae (Revised from Table 1)
Cancer invasion
Affected organs Case + - unknown
C 13 3 7 3
D 13 5 5 3
C and D 8 3 3 2
S 4 2 2 (this case) 0
S and D 3 3 0 0
Total 41 16 17 8
Affected Organ; C: Common bile duct, D: Duodenum, S: Stomach
To our knowledge, the present report is only the second describing a case of IPMN associated with pancreatogastric fistulae without invasive cancer. Yago et al reported a rare case of non-invasive intraductal papillary adenocarcinoma of the pancreas that was associated with development of a pancreatogastric fistula in the absence of cancer invasion [5]. In that case, neoplastic epithelia were observed only on the surface of the fistula and gastric mucosa, while the structure of the gastric wall beside the fistula was maintained without invasive cancer. Those authors speculated that the pancreatogastric fistula developed due to high pressure in the MPD. Baek et al and Watanabe et al reported the same phenomenon in mucinous cystic tumors (MCTs) [11,12]. The mechanism of such fistula formation without cancer spread could be explained by a combination of high pressure in the MPD and inflammatory stimulation [3,4,13].
In the present case, ERCP played a key role in fistula formation. To our knowledge, this is the first report to show an absence of a pancreatogastric fistula immediately prior to ERCP, and then the presence of a fistula following post-ERCP pancreatitis. Post-ERCP pancreatitis, which is reported to occur in approximately 1.8–10% of patients, should be carefully investigated in IPMN cases [14,15].
We emphasize here that development of an IPMN-associated fistula does not necessarily represent underlying invasive cancer. Fistula formation can result in an incorrect preoperative diagnosis of invasive cancer and lead to the undertaking of unnecessary surgical procedures. Although the lesion in the present case was a non-invasive tumor, surgery was appropriate given the indications of invasive cancer. While non-invasive IPMNs have a good prognosis, there remains the possibility of dissemination by tumor penetration. When treating IPMNs associated with fistula, the extent of resection should depend on the extent of cancer invasion. To avoid dissemination, the fistula should be removed independent to cancer invasion. The extent of resection of affected organs should be individually determined based on the results of preoperative images and precise intraoperative histological examinations.
Conclusion
The present case highlights that a pancreatogastric fistula can develop in the absence of invasive cancer. Further data regarding IPMN-associated fistulae are necessary in order to clarify the pathogenesis, diagnosis, appropriate surgical intervention and prognosis for this disorder.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
MK, NS, KY, MT, KK, YY and HN were involved in performing surgery, undertook the literature search and drafted the manuscript for submission.
HN supervised preparation of the manuscript and edited the final version for publication. All authors read and approved the manuscript.
Acknowledgements
Patient consent was obtained for publication of case records, MRCP, abdominal CT, ERCP, US, gastroscopy and histopathological images.
==== Refs
Sugiyama M Atomi Y Intraductal papillary mucinous tumors of the pancreas: imaging studies and treatment strategies Ann Surg 1998 228 685 691 9833807 10.1097/00000658-199811000-00008
Ohhashi K Tajiri H Gondo M Yokoyama Y Maruyama M Takekoshi T Matsuura Y Kasumi F Takagi K Kato Y A case of cystadenocarcinoma of the pancreas forming bilio-pancratic fistula Progress of Digestive Endoscopy 1980 17 261 264
Fujisawa T Osuga T Mita M Sakamoto N Sakaguchi K Onishi Y Toyoda M Maeda M Kawase Y Nishigami T Intraductal papillary adenocarcinoma of the pancreas accompanied by biliopancreatic fistula; report of a case Ganno Rinsyo (Jpn J Cancer Clin) 2001 47 144 150
Kobayashi G Fujita N Noda Y Kimura K Watanabe H Chonan A Matsunaga A Yuki T Ando M Sato Y Tominaga G Mochizuki F Yamazaki T Study of cases of mucin producing tumors of the pancreas showing penetration of other organs Nippon Shokakibyo Gakkai Zasshi (Jpn J Gastroenterology) 1993 90 3081 3089
Yago A Fujita N Noda Y Kobayashi G Kimura K Mochizuki F Yamazaki T Ikeda T A case of non-invasive intraductal papillary adenocarcinoma of the pancreas showing penetration of the stomach Fukubu Gazo Shindan 1995 15 496 500
Tadokoro H Toki F Yoshida K Nishino T Kojima S Shiratori K Watanabe S Kozu T Takeuchi T Hayashi N Imaizumi T Suzuki S Two cases of pancreaticoduodenal fistula Gastroenterological Endoscopy 1996 38 1208 13
Matsubayashi H Kokuno S Itoi T Mizumura Y Niki S Takeda K Onoda K Ogiwara M Ohno H Horibe T Miwa K Shinohara Y Magami Y Niido T Seki T Saitoh T A case of intraductal papillary adenocarcinoma of the pancreas with pancreatico-duodenal fistula Progress of Digestive Endoscopy 1998 52 180 181
Shiroko J Takai S Yoshikawa T Ohnishi T Nakai M Asano H Tanabashi S Kametani M Okamoto K A case of Mucin-producing pancreatic cancer penetrating to the common bile duct and the duodenum Ganno Rinsyo (Jpn J Cancer Clin) 2000 46 967 971
Kurihara K Nagai H Kasahara K Kanazawa K Kanai N Biliopancreatic fistula associated with intraductal papillary-mucinous pancreatic cancer: institutional experience and review of the literature Hepato-Gastroenterology 2000 47 1164 1167 11020905
Inoue M Ikeda Y Kikui M Ogawa T Yasumitsu T Mucin-producing tumor of the pancreas associated with pyothorax: report of a case Surg Today 2001 31 538 541 11428610 10.1007/s005950170118
Baek Y Midorikawa T Nagasaki H Kikuchi H Kitamura N Takeuchi S Koh Y Yagi H Yoshizawa Y Kumada N A case report of pancreatic mucinous cystadenocarcinoma with penetration to the stomach Nippon Shokakibyo Gakkai Zasshi (Jpn J Gastroenterology) 1999 96 685 690
Watanabe N Hasegawa H Tsuneya Y Kumazawa M Baba M Kusano I Watanabe M Yamao Y A case of mucinous cystadenocarcinoma with a fistula between cyst and stomach Nippon Shokakibyo Gakkai Zasshi (Jpn J Gastroenterology) 2003 100 349 353
Fukushima N Mukai K Mechanism of invasion through the duct wall by intraductal papillary adenocarcinoma of the pancreas Tan to Sui (J Bil Panc) 1999 20 17 20
Pannu HK Fishman EK Complications of endoscopic retrograde cholangiopancreatography: Spectrum of abdominalities demonstrated with CT Radio Graphics 2001 21 1441 1453
Masci E Toti G Mariani A Curioni S Lomazzi A Dinelli M Minoli G Crosta C Comin U Fertitta A Prada A Passoni GR Testoni PA Complications of diagnostic and therapeutic ERCP: A prospective multicenter study Am J Gastroenterol 2001 96 417 423 11232684 10.1111/j.1572-0241.2001.03594.x
|
16232325
|
PMC1266404
|
CC BY
|
2021-01-04 16:39:04
|
no
|
World J Surg Oncol. 2005 Oct 19; 3:70
|
utf-8
|
World J Surg Oncol
| 2,005 |
10.1186/1477-7819-3-70
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35291613734010.1186/cc3529ResearchPulse high-volume haemofiltration for treatment of severe sepsis: effects on hemodynamics and survival Ratanarat Ranistha [email protected] Alessandra 2Piccinni Pasquale 3Dan Maurizio 3Salvatori Gabriella 4Ricci Zaccaria 4Ronco Claudio [email protected] Fellow, Department of Nephrology, Dialysis and Transplantation, St Bortolo Hospital, Vicenza, Italy, and Instructor, Department of Medicine, Faculty of Medicine Siriraj Hospital, Mahidol University, Bangkok, Thailand2 Nephrologist and Consultant in Nephrology, Department of Nephrology, Dialysis and Transplantation, St Bortolo Hospital, Vicenza, Italy3 Head of Department, Department of Anesthesia and Intensive Care, St Bortolo Hospital, Vicenza, Italy4 Fellow, Department of Nephrology, Dialysis and Transplantation, St Bortolo Hospital, Vicenza, Italy5 Professor and Head of Department, Department of Nephrology, Dialysis and Transplantation, St Bortolo Hospital, Vicenza, Italy2005 28 4 2005 9 4 R294 R302 16 2 2005 9 3 2005 17 3 2005 5 4 2005 Copyright © 2005 Ratanarat et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Severe sepsis is the leading cause of mortality in critically ill patients. Abnormal concentrations of inflammatory mediators appear to be involved in the pathogenesis of sepsis. Based on the humoral theory of sepsis, a potential therapeutic approach involves high-volume haemofiltration (HVHF), which has exhibited beneficial effects in severe sepsis, improving haemodynamics and unselectively removing proinflammatory and anti-inflammatory mediators. However, concerns have been expressed about the feasibility and costs of continuous HVHF. Here we evaluate a new modality, namely pulse HVHF (PHVHF; 24-hour schedule: HVHF 85 ml/kg per hour for 6–8 hours followed by continuous venovenous haemofiltration 35 ml/kg per hour for 16–18 hours).
Method
Fifteen critically ill patients (seven male; mean Acute Physiology and Chronic Health Evaluation [APACHE] II score 31.2, mean Simplified Acute Physiology Score [SAPS] II 62, and mean Sequential Organ Failure Assessment 14.2) with severe sepsis underwent daily PHVHF. We measured changes in haemodynamic variables and evaluated the dose of noradrenaline required to maintain mean arterial pressure above 70 mmHg during and after pulse therapy at 6 and 12 hours. PHVHF was performed with 250 ml/min blood flow rate. The bicarbonate-based replacement fluid was used at a 1:1 ratio in simultaneous pre-dilution and post-dilution.
Results
No treatment was prematurely discontinued. Haemodynamics were improved by PHVHF, allowing a significant reduction in noradrenaline dose during and at the end of the PHVHF session; this reduction was maintained at 6 and 12 hours after pulse treatment (P = 0.001). There was also an improvement in systolic blood pressure (P = 0.04). There were no changes in temperature, cardiac index, oxygenation, arterial pH or urine output during the period of observation. The mean daily Kt/V was 1.92. Predicted mortality rates were 72% (based on APACHE II score) and 68% (based on SAPS II score), and the observed 28-day mortality was 47%.
Conclusion
PHVHF is a feasible modality and improves haemodynamics both during and after therapy. It may be a beneficial adjuvant treatment for severe sepsis/septic shock in terms of patient survival, and it represents a compromise between continuous renal replacement therapy and HVHF.
See related commentary
==== Body
Introduction
Severe sepsis represents the leading cause of mortality and morbidity in critically ill patients worldwide. The sepsis syndrome is associated with an overwhelming, systemic overflow of proinflammatory and anti-inflammatory mediators, which leads to generalized endothelial damage, multiple organ failure and altered cellular immunological responsiveness. Although our understanding of the complex pathophysiological alterations that occur in severe sepsis and septic shock has increased greatly as a result of recent clinical and preclinical studies, mortality associated with the disorder remains unacceptably high, ranging from 30% to 50% [1-4].
The cornerstone of therapy continues to be early recognition, prompt initiation of effective antibiotic therapy, source control, and goal-directed haemodynamic, ventilatory and metabolic support as necessary. To date, attempts to improve survival with innovative, predominantly anti-inflammatory therapeutic strategies have been disappointing, with the exception of physiological doses of corticosteroid replacement therapy [5,6] and activated protein C (drotrecogin alfa [activated]) [7] in selected patients.
'Renal dose' haemofiltration rate of 2000 ml/hour has successfully been used to treat acute renal failure for years [8]. This dose suffices for renal replacement therapy and can remove inflammatory mediators; however, it does not alter plasma levels of these mediators, suggesting that its ability to clear inflammatory mediators is suboptimal [9]. This was reflected in one study [10] by failure to demonstrate an improvement in organ dysfunction and survival. Hence, the indication for its use in septic patients was abandoned, beyond its function to provide renal support in the presence of renal dysfunction [11]. However, the theory that underpins increasing plasma water exchange or higher dose haemofiltration seems reasonable.
Ronco and coworkers [12] demonstrated survival benefits by increasing the haemofiltration dose (35 ml/kg per hour) beyond the conventional renal dose (20 ml/kg per hour), but no further benefit was achieved, even at higher doses (45 ml/kg per hour), in the overall studied population. Nevertheless, there was an improvement in survival at the highest haemofiltration doses in that study for the subset of patients with sepsis. Additionally, benefits have been demonstrated in several animal models of sepsis. Improvements in cardiac function and haemodynamics were replicated in these animal studies using ultrafiltration (UF) rates up to 120 ml/kg per hour [13-16]. Septic dose haemofiltration, or high-volume haemofiltration (HVHF), was thus conceived and applied in human sepsis. Findings of improvements in haemodynamics with decreased vasopressor requirements [17-19] and trends toward improved survival [19,20] are evidence that HVHF may be efficacious. Because HVHF technique requires high blood flows, tight UF control and large amounts of expensive sterile fluids, we proposed a new technique, namely 'pulse HVHF' (PHVHF) [21,22]. PHVHF is application of HVHF for short periods (up to 6–8 hours/day), providing intense plasma water exchange, followed by conventional continuous venovenous haemofiltration (CVVH).
We hypothesized that daily 'PHVHF' may have beneficial effects in severe sepsis by unselectively removing of proinflammatory and anti-inflammatory mediators, and hence improving patient outcomes. The present study evaluates the feasibility of PHVHF and the effect of this treatment on haemodynamics, oxygenation and 28-day all-cause mortality.
Materials and methods
This is a prospective interventional study conducted in the intensive care unit (ICU) of St. Bortolo Hospital, Vicenza, Italy. Fifteen patients with severe sepsis receiving continuous renal replacement therapy (CRRT) were enrolled in the study. Patients were included in the study if they had severe sepsis or septic shock, as defined using the criteria reported by Bone and coworkers [23], and if they fulfilled one of the previously reported criteria for initiating renal replacement therapy in critically ill patients [24]. Exclusion criteria were age less than 18 years, death imminent within 24 hours, and very high weight (>140 kg). All patients were treated using the same, recently developed management guideline for severe sepsis and septic shock [25]. All except one patient were receiving mechanical ventilation because of respiratory failure. Broad spectrum antibiotics were given to all patients and were altered according to blood culture and sensitivity findings.
Eight out of 15 patients received activated protein C (drotrecogin alfa [activated]). The drug was not used in seven patients: one had underlying ruptured abdominal aortic aneurysm; the second was admitted because of multiple fractures and severe head trauma; the third had an Acute Physiology and Chronic Health Evaluation (APACHE) II score less than 25 at admission; and the remaining four had severe thrombocytopenia (<15,000/mm3) and/or impaired coagulation (international normalized ratio >3.0). The use of activated protein C (drotrecogin alfa [activated]) in approximately 50% of the patients included might therefore have contributed to any improved outcome identified. Clinical data are summarized in Table 1.
The APACHE II score, Simplified Acute Physiology Score (SAPS) II, and Sequential Organ Failure Assessment score were calculated from physiological measurements obtained during the first 24 hours of ICU admission. Expected mortality rates for APACHE II and SAPS II scores were computed using the logistic regression calculations suggested in the original reports [26,27]. The study protocol was approved by the hospital ethics committee.
Description of pulse high-volume haemofiltration technique
PHVHF was performed using a multifiltrate CRRT machine (Fresenious Medical Care, Bad Hamburg, Germany). This recently designed machine provides high-precision scales (equipped with software for online continuous testing and high capacity) and powerful heating systems for maintaining the large volumes of infusion solution at sufficiently high temperature.
Vascular access was obtained with 14-Fr central venous haemodialysis catheter. Blood flow rates of 250–300 ml/min, as permitted by the access, were used to achieve a filtration fraction of 20–25% and to prevent premature clotting of extracorporeal circuit.
PHVHF was performed using a UF rate of 85 ml/kg per hour for 6 hours/day followed by standard continuous venovenous haemofiltration (CVVH; UF rate 35 ml/kg per hour) for 18 hours, resulting in a cumulative dose of approximately 48 ml/kg per hour. Treatments were given on a daily basis, and were terminated if the patient died or if the physician considered the septic process to have ended and the patient's clinical parameters improved.
Commercially available bicarbonate-buffered replacement fluid containing sodium 142 mmol/l, potassium 2 mmol/l, chloride 113.5 mmol/l, bicarbonate 32 mmol/l and calcium 1.75 mmol/l (Bi-intensive; B-Braun, Bologna, Italy) was used at a ratio of 1:1 in simultaneous pre-dilution and post-dilution. Additional potassium and phosphate were administered intravenously to prevent hypokalaemia and hypophosphataemia. A highly biocompatible synthetic membrane with surface area of 1.8–2 m2 was also utilized. Anticoagulation was initiated with 1000–2000 IU bolus injection of heparin followed by an infusion of 250–500 IU/hour. Net fluid removal was set according to the patient's condition and clinical need.
Measurements
Haemodynamic monitoring was done using a thermodilution pulmonary artery catheter with continuous cardiac output monitoring (Vigilance; Edwards Lifesciences, Irvine, CA, USA). A radial or a femoral arterial catheter was used to measure blood pressure and obtain arterial blood for blood gas analysis. Systolic blood pressure, mean arterial pressure (MAP), body temperature, heart rate, cardiac index and noradrenaline (norepinephrine) dose required to maintain MAP above 70 mmHg were measured immediately before PHVHF, mid-PHVHF, immediately after PHVHF, and 6 hours and 12 hours after completion of the PHVHF session. The bedside nurse was instructed to maintain MAP above 70 mmHg by adjusting the dose of noradrenaline infused. pH, partial oxygen tension and bicarbonate were measured using a clinical blood gas analyzer (Rapidpoint 400; Bayer HealthCare, Newbury, UK) at similar time intervals.
Blood samples were also collected at immediately before initiation of treatment, immediately on discontinuation of PHVHF and 12 hours after the session had ended, in order to measure blood urea nitrogen, creatinine and electrolytes. Observed mortality was recorded during the day on which patients received PHVHF and at 28 days.
Data analysis
One-sample Kolmogorov–Smirnov test was utilized to assess whether the distribution of haemodynamic and metabolic variables were normal. Normally distributed data are presented as means ± standard deviation, and differences of serially measured variables were analyzed using analysis of variance for repeated measurements with Bonferroni correction. For non-normally distributed variables, results are reported as medians with 25th to 75th percentile range, and Friedman's two-way analysis of varience with post hoc Wilcoxon signed rank test was used to identify whether changes had occurred over time. Comparison between expected mortality (based on APACHE II and SAP II scores) and observed mortality was done using the standardized ratio and 95% confidence interval calculated by dividing the observed by expected mortality [28]. P < 0.05 was considered statistically significant.
Results
Patient outcomes
Of the 15 patients enrolled, 50 PHVHF treatments were performed on a daily basis. The mean number of treatments per patient was 3.4 (1–9). No treatment was prematurely discontinued because of extracorporeal circuit clotting or high pressure problems. Demographic data are presented in Table 1. The observed patient hospital mortality was 46.7%, as compared with a rate of 72% predicted by APACHE II and 68% predicted by SAPS II severity scores. Hospital mortality ratios (95% confidence interval) [28] were 0.65 (0.48–0.87) and 0.69 (0.51–0.92), as compared with the expected mortality calculated from APACHE II and SAPS II scores, respectively.
With respect to causes of death, one patient died from acute myocardial infarction with cardiogenic shock during day 7 of ICU admission. The second patient, with acute endocarditis, underwent PHVHF for 2 days and all vasopressors (noradrenaline, adrenaline and dopamine) were discontinued on day 3.
Unfortunately, the patient had cardiogenic shock from a ruptured aortic valve on day 7 and died on day 9 after admission. The third patient died because her underlying disease was multiple myeloma grade IIIb, which did not respond to chemotherapy, and the physician decided to withhold the treatment, in accordance with hospital policy, on day 9 after admission. Only the remaining four patients died from refractory septic shock.
Table 2 summarizes baseline demographic and physiological parameters, stratifying patients by whether they were alive at 28 days. Before initiation of PHVHF there were no significant differences between survivors and nonsurvivors at 28 days with respect to age, body weight, MAP, cardiac index, oxygenation, severity scores (APACHE II, SAPS II and Sequential Organ Failure Assessment) and number of organ failures. Interestingly, the mean number of PHVHF treatments per patient was significantly higher in the group of survivors (4.8 ± 2.7) than in the nonsurvivor group (1.9 ± 0.7; P = 0.02).
Haemodynamic outcomes
All patients except three received noradrenaline at the start of PHVHF treatment, with a median dose of 48 μg/min (Table 3). In fact, dopamine is generally the first-choice vasoactive/inotropic agent in our unit; however, once the dopamine infusion has exceeded 10 μg/kg per min or low systemic vascular resistance is identified by pulmonary artery catheter, our policy is to initiate noradrenaline and taper dopamine. As a result, noradrenaline was the sole vasoactive agent in one patient only. The remaining three patients were receiving dopamine with or without dobutamine at the initiation of PHVHF therapy. The median number of concurrently administered vasopressors per patient before PHVHF was 2, and this did not change after PHVHF. No patients developed threatening hypotension during pulse therapy, and none needed de novo institution of vasopressors during this treatment.
Haemodynamic changes are shown in Table 3. MAP before PHVHF was 82 ± 18 mmHg, after PHVHF it was 87 ± 18 mmHg, and 12 hours after PHVHF it was 87 ± 22 mmHg (P = 0.2). However, systolic blood pressure increased significantly over time (pre-PHVHF 124 ± 26 mmHg, mid-PHVHF 127 ± 22 mmHg, post-PHVHF 133 ± 25 mmHg, 6 hours after PHVHF 133 ± 24 mmHg, and 12 hours after PHVHF 133 ± 26 mmHg; P = 0.04). As expected, MAP and cardiac index did not change significantly over time during PHVHF and after treatment, and MAP was maintained at the target levels in accordance with the study protocol (Table 3). The dose of noradrenaline required for maintenance of target MAP decreased significantly by the mid-point of the PHVHF session, and this decrease was maintained at 6 and 12 hours after treatment (P = 0.001; Table 3 and Fig. 1).
By setting the temperature of the replacement fluid at around 38.5–39°C, body temperature was constant during pulse treatment (Table 3). Positive fluid balance on the day before PHVHF (1374 ± 2618 ml/day) was not different from that during the day on which patients underwent PHVHF (1514 ± 2548 ml/day; P = 0.9). Oxygenation (arterial oxygen tension/fractional inspired oxygen ratio) did not change over time.
Solute control and renal outcomes
Seven out of eight survivors underwent CVVH after the termination of daily PHVHF treatments because of renal failure. In one survivor renal function recovered by the time of cessation of daily PHVHF. All except two kidney transplant recipients (in whom the graft was lost because of septic shock) could be withdrawn from renal replacement therapy and had complete renal recovery.
Four nonsurvivors at 28 days with refractory septic shock died while they were still receiving daily PHVHF. As mentioned above, three nonsurvivors died for reasons other than septic shock and were treated with CVVH following improvement in their haemodynamic parameters and cessation of PHVHF.
Solutes and acid base status before and after PHVHF are presented in Table 4. Daily Kt/V was 1.92 ± 0.29. As expected, serum blood urea nitrogen and creatinine levels diminished greatly after pulse treatment (P < 0.0001; Table 4). Daily urine output on the day before treatment (median 310 ml, range 75–1916 ml) did not differ from that on the day of initiation of PHVHF treatment (median 268 ml, range 77–1905 ml).
Discussion
The sepsis syndrome is associated with an overwhelming, systemic overflow of proinflammatory and anti-inflammatory mediators, which leads to generalized endothelial damage, multiple organ failure and altered cellular immunological responsiveness. The complex inflammatory network involved is redundant, synergistic and acts like a cascade. It includes mediators with autocrine and paracrine actions, as well as cellular and intracellular components. A large number of proinflammatory mediators, including tumour necrosis factor-α, interleukin-1, interleukin-6, platelet-activating factor and nitric oxide, play important roles in the cascade, but attempts to improve survival in human trials using innovative, predominantly anti-inflammatory therapeutic strategies have been disappointing [29]. Almost paralleling the surge in proinflammatory mediators, there is a rise in anti-inflammatory substances, resulting in induction of a state of immunoparalysis or monocyte hyporesponsiveness [30]. Both proinflammatory and anti-inflammatory factors become upregulated and interact with each other, leading to various rises in mediator levels that change over time. Neither therapies directed at single mediators nor single-dose interventions therefore seem appropriate, in part because of a discrepancy between the biological timing of the syndrome and the clinical timing of symptoms.
CRRT has made extracorporeal treatment possible in septic patients even when they are haemodynamically unstable; such treatment is given to balance hypercatabolism and fluid overload. In addition, 'high volume' and convective modalities have the advantage of removing higher molecular weight substances, which include many inflammatory mediators. Multiple animal studies [13-16] have shown a beneficial effect of HVHF on survival in endotoxaemic models. Recent studies in humans [17-19] have demonstrated that HVHF improves haemodyamics, with decreased vasopressor requirements.
A daily PHVHF regimen was utilized as the intervention in the present study for the following reasons. First, the very high UF volume requires very close surveillance, which is difficult to maintain over 24 hours. Second, solute kinetics may render high volumes useless after a few hours because of saturation of membrane adsorption [17,31]. Third, standard CVVH (UF rate 35 ml/kg per hour) may help to maintain the effect of pulse therapy and prevent post-treatment rebound from sudden changes. Instead of using a fixed dose (i.e. UF rate 6 l/hour), we applied a dose of 85 ml/kg per hour during pulse treatment because body size is the main predictor of patient outcome [12,18]. Additionally, 'continuous' removal of soluble mediators may be the most logical and best approach to a complex and lengthy process such as sepsis; we therefore performed PHVHF on a daily basis and terminated treatment when haemodynamic variables improved. We hypothesized that beneficial haemodynamic effects would be achieved during PHVHF, and that these effects would be perpetuated after cessation of the pulse treatment by standard CVVH. We also hypothesized that they would be accompanied by improvement in oxygenation and reduction in mortality. The present pilot study in patients with severe sepsis/septic shock was conducted to test our hypotheses.
Overall, PHVHF was well tolerated by critically ill patients and appeared to offer many of the benefits conferred by continuous HVHF [19] while avoiding its drawbacks. Six to eight hours PHVHF during the daytime was widely accepted by the ICU nursing staff because it reduced the labour intensity of the protocol during the night shift. No treatment was prematurely discontinued because of extracorporeal circuit clotting or high pressure problems. It appears that PHVHF is a feasible modality and can safely be performed on a daily and prolonged basis. The greatest duration of treatment in any patient our study was 9 days. The most clinically relevant finding that emerged from our investigation is that adjuvant PHVHF in septic shock patients is associated with improvement in haemodynamic parameters (Fig. 1), permitting a significant reduction in vasopressor requirements as soon as halfway through and at the end of the PHVHF session, and this was maintained at 6 hours and 12 hours after treatment. The haemodynamic benefits of short-term HVHF regimens were recently demonstrated by Cole and coworkers [17] (UF rate 6 l/hour, duration 8 hours) and Honore and colleagues [18] (UF rate 8.75 l/hour, duration 4 hours). We proved that this beneficial hemodynamic effect can be maintained after HVHF by continuing with standard dose CVVH (UF rate 35 ml/kg per hour). Indeed, in practice our regimen could be adjusted on the basis of the individual patient's clinical response.
Several mechanisms are potentially responsible for the reduced need for pressor therapy with PHVHF. For mediator-independent factors, we were unable to demonstrate any differences in body temperature and arterial pH before and after PHVHF, including 12 hours after treatment. It is clear that cooling-induced vasoconstriction and correction of severe acidosis cannot account for this positive haemodynamic effect. The daily fluid balance on the day before initiation of PHVHF and that on the day of intervention were similar. Based on these findings, we argue that PHVHF permits continuous removal of soluble vasodilatory mediators or molecules identified in sepsis by either convection or adsorption, resulting in reduction in vasopressor requirements.
Unlike recent studies conducted by Honore and coworkers [18] and Joannes-Boyau and colleagues [19], we was unable to demonstrate any benefit of HVHF on cardiac index. The possible explanation for this is that we recruited patient at an earlier time point in septic shock (i.e. during hyperdynamic state); the mean cardiac index of our patients was 3.4 l/min per m2, whereas those in the other two studies were less (2.0 l/min per m2 [18] and 2.9 l/min per m2 [19]). The aim of haemodynamic support in our sepsis patients was to maintain condiac index at 2.5 l/min per m2 or above because the studies that attempted to maintain a supraphysiologic cardiac index of above 4.0 to 4.5 l/min per m2 have not shown consistent benefit [32,33]. Interestingly, this indicates that the improved haemodynamics and decreased vasopressor requirement conferred by daily PHVHF are clinically significant even during hyperdynamic septic shock.
Although it is beyond the scope of this report to provide a full comparison of mortality rates between standard sepsis treatment and such treatment combined with PHVHF, it appears that PHVHF may have beneficial immunomodulatory effects with prolonged daily use, especially with respect to patient outcome. The 28-day all-cause mortality was 47%, as compared with 72% as predicted by APACHE II and 68% as predicted by SAPS II severity scores. This is consistent with the findings of another study [19], in which 96 hours of continuous HVHF was given to patients with septic shock (46% observed and 70% predicted mortality rate). In fact, of the seven deaths at 28 days in our study, only four were attributable to refractory septic shock. How long would it take for a clinically relevant benefit to manifest? Tailoring our daily PHVHF regimen according to clinical response should permit sufficient duration of HVHF. In addition, because absolute or relative contraindications were met in seven patients, only the remaining eight patients in the present study received activated protein C (drotrecogin alfa [activated]) – a drug that has shown the benefit in terms of 28-day mortality in recent trials [7]. However, we can state that PHVHF is feasible and, as a treatment for severe sepsis/septic shock, can affect physiological end-points. In terms of mortality, the only way to demonstrate the effect of PHVHF in this population is to conduct a prospective, randomized, controlled study on a larger scale. Nevertheless, we can hypothesize that the use of activated protein C (drotrecogin alfa [activated]) in 50% of the population might have contributed to the improved outcome. If so, then the combination of activated protein C (drotrecogin alfa [activated]) and PHVHF might be particularly useful.
The present study is limited by the fact that the population was highly heterogeneous, relatively small and reflective of patients seen in a single center. We did not measure mediator levels in plasma and in the UF over time, which might have helped to explain the mechanisms of mediator removal. However, the nonselective, simultaneous removal of different mediators demonstrated by a reduction of the circulating cytokines or an increase their levels in the UF may not necessarily implicate as the gold standard of blood purification for sepsis patients. A more effective strategy would be to attempt to influence the functional responses of cells that are implicated in the pathogenesis of sepsis. Such approaches are under evaluation, and findings reported in a preliminary paper [21] are encouraging. Also, we did not evaluate removal of sedative drugs with vasodilatory effect, such as midazolam and sufentanil. For ethical reasons, we could not conduct the trial in sepsis patients who did not have acute renal failure. In the context of acute renal failure in sepsis, it is clear that metabolic compounds partly accumulate as a consequence of the loss of renal function. Uraemic toxins rapidly accumulate in tissues and plasma, and they may be responsible for the immune dysregulation associated with sepsis.
Conclusion
In summary, PHVHF appears to be feasible and is a promising technique for the treatment of severe sepsis. We demonstrated a clinically and statistically significant beneficial effect of this therapy on vasopressor requirements during treatment and after therapy. It may be a beneficial adjuvant treatment for severe sepsis/septic shock in terms of patient survival, and it represents a compromise between CRRT and HVHF. Further confirmation is required in large, properly designed clinical trials to establish the benefit of PHVHF.
Key messages
• PHVHF represents a feasible compromise between CRRT and HVHF, in which HVHF is applied for short periods of up to 6–8 hours/day and followed by standard dose CVVH.
• PHVHF, when applied in patients with septic shock/severe sepsis, can achieve beneficial effects on vasopressor requirements.
• PHVHF applied on the daily basis and tailored according to clinical response may represent a beneficial adjuvant treatment for severe sepsis/septic shock in terms of patient survival.
Abbreviations
APACHE = Acute Physiology and Chronic Health Evaluation; CRRT = continuous renal replacement therapy; CVVH = continuous venovenous haemofiltration; HVHF = high-volume haemofiltration; ICU = intensive care unit; MAP = mean arterial pressure; PHVHF = pulse high-volume haemofiltration; SAPS = Simplified Acute Physiology Score; UF = ultrafiltration;
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
RR conducted the study, collected data, performed statistic analysis and drafted the manuscript. AB and ZR conducted the intervention in the study. PP and MD carried out the haemodynamic measurements. GS helped to collect the data. CR conceived the study, participated in its design and helped to draft the manuscript. All authors read and approved the final manuscript.
Acknowledgements
We appreciate the critical care nephrology nurses (CCNN) group of St. Bortolo Hospital for their cooperation and support. We also thank S Udompunthurak for help in the statistic analysis.
Figures and Tables
Figure 1 Haemodynamic variables. Variables were recorded during the pulse high-volume haemofiltration (PHVHF) session, and 6 hours and 12 hours after completion of the session. Noradrenaline (norepinephrine [NE]) requirement decreased significantly during treatment, and this reduction persisted at 6 hours and 12 hours after treatment (P = 0.0001). *P < 0.05 and aP < 0.01 for difference between pre-PHVHF and other measures over time. Sytolic blood pressure (SBP) increased significantly during treatment, and this was maintained 6 hours and 12 hours after treatment (P = 0.04). All reported values are means.
Table 1 Clinical features of patients with septic shock/severe sepsis treated with pulse high volume hemofiltration
Age (years)/sex/body weight (kg) Number of treatments Diagnosis Microbiology Number of organ failures APACHE II scorea SAPS II scorea SOFA score 28-day survival
66/M/77 1 CHF, septic shock Negative 4 35 (83%) 79 (92%) 14 D
62/M/70 2 Lobar pneumonia Negative 4 27 (61%) 53 (53%) 11 D
77/M/70 2 Ruptured abdomonal aortic aneurysm, pancreatitis Nonfermentative Gram-negative bacilli 4 32 (76%) 53 (53%) 14 D
37/M/87 5 Necrotizing fasciitis Negative 5 29 (67%) 58 (64%) 17 A
69/F/68 3 Kidney transplant, disseminated candidiasis, septicaemia (uncertain source) Candida glabrata, coagulase-negative Staphylococcus 4 34 (81%) 86 (95%) 13 A
54/M/80 2 Bronchopneumonia Negative 3 23 (46%) 46 (37%) 12 A
54/F/45 2 Myelodysplasia, acute endocarditis Staphylococcus aureus, Escherichia coli 5 29 (67%) 55 (58%) 17 D
58/F/65 3 Obstructive uropathy, pyelonephritis Escherichia coli 4 28 (64%) 46 (37%) 15 A
64/M/80 1 Exfoliative dermatitis, erysipilas Haemolytic Streptococcus group A 4 39 (90%) 82 (94%) 16 D
74/F/90 2 Nosocomial pneumonia, catheter-related sepsis Pseudomonas aeruginosa, coagulase-negative Staphylococcus 4 33 (79%) 61 (70%) 14 A
43/F/63 6 Kidney transplant, disseminated candidiasis, UTI Escherichia coli, Candida albicans 3 26 (57%) 32 (42%) 11 A
33/M/85 3 Multiple trauma, infected wound Coagulase-negative Staphylococcus 5 31 (73%) 70 (84%) 13 D
69/F/82 2 Multiple myeloma, peritonitis Nonfermentative Gram-negative bacilli 4 33 (79%) 74 (88%) 14 D
44/F/83 8 Kidney transplant, septicaemia (uncertain source) Pseudomonas aeruginosa, Enterococcus faecalis 4 36 (85%) 68 (81%) 16 A
59/F/63 9 Rheumatoid arthritis, pneumonia Streptococcal pneumonia 5 33 (79%) 67 (80%) 16 A
aShown in parentheses is the predicted chance of hospital mortality. A, alive; APACHE, Acute Physiology and Chronic Health Evaluation score; CHF, congestive heart failure; D, died; SAPS, Simplified Acute Physiology Score; SOFA, Sequential Organ Failure Assessment; UTI, urinary tract infection.
Table 2 Baseline demograpic and physiological variables stratified by outcome (28-day survival)
Variables Survivor (n = 8) Nonsurvivor (n = 7) P
Age (years) 55 ± 13 61 ± 14 NS
Body weight (kg) 75 ± 11 73 ± 13 NS
SBP (mmHg) 98 ± 20 120 ± 32 NS
MAP (mmHg) 68 ± 12 72 ± 13 NS
CI (l/min per m2) 4.1 ± 1.1 2.7 ± 1.0 NS
PaO2/FiO2 ratio 216 ± 99 172 ± 49 NS
APACHE II score 30.3 ± 4.5 32.2 ± 3.9 NS
SAPS II score 58.0 ± 16.6 66.6 ± 12.7 NS
SOFA score 14.3 ± 2.1 14.1 ± 2.0 NS
Number of organ failures 4.0 ± 0.8 4.3 ± 0.5 NS
Number of PHVHF treatments 4.8 ± 2.7 1.9 ± 0.7 0.02
Values are expressed as mean ± standard deviation. APACHE, Acute Physiology and Chronic Health Evaluation score; CI, cardiac index; MAP, mean arterial pressure; PaO2/FiO2, arterial oxygen tension/fractional inspired oxygen; PHVHF, pulse high-volume haemofiltration; SAPS, Simplified Acute Physiology Score; SBP, systolic blood pressure; SOFA, Sequential Organ Failure Assessment.
Table 3 Effects of pulse high-volume haemofiltration on haemodynamic variables
Variables Pre-PHVHF Mid-PHVHF End-PHVHF 6 hours after PHVHF 12 hours after PHVHF P
Noradrenaline Dose (μg/min) 48 (0–114) 40 (0–97)* 40 (0–93) 40 (0–69)* 33 (0–67)** 0.001
SBP (mmHg) 124.32 ± 25.63 126.64 ± 22.10 133.00 ± 24.55 133.06 ± 23.88 133.16 ± 26.15 0.04
MAP (mmHg) 82.16 ± 18.31 85.02 ± 18.82 86.88 ± 17.56 87.76 ± 20.65 87.26 ± 22.05 NS
CI (l/min per m2) 3.4 ± 1.1 3.4 ± 1.2 3.5 ± 1.0 3.5 ± 1.1 3.5 ± 1.2 NS
HR (beats/min) 97.28 ± 25.53 99.62 ± 22.94 100.06 ± 21.79 99.94 ± 20.71 95.62 ± 20.66 0.04
Temperature (°C) 36.7 ± 1.0 36.8 ± 0.8 36.8 ± 0.8 36.9 ± 0.8 36.7 ± 0.9 NS
PaO2/FiO2 230.9 ± 109.1 232.8 ± 104.4 243.0 ± 105.6 230.2 ± 109.9 234.6 ± 106.4 NS
Normally distributed values are reported as mean ± standard deviation, and the statistical test used was analysis of variance for repeated measurements. Non-normally distributed values are reported as median (25th to 75th percentile), and P value was determined using Friedman's two-way analysis of varience with post-hoc Wilcoxon signed rank test. *P < 0.05, **P < 0.01 versus baseline. HR, heart rate; CI, cardiac index; MAP, mean arterial pressure; PaO2/FiO2, arterial oxygen tension/fractional inspired oxygen; PHVHF, pulse high-volume haemofiltration; SBP, systolic blood pressure.
Table 4 Effects of pulse high-volume haemofiltration on metabolic variables
Variables Pre-PHVHF End-PHVHF 12 hours after PHVHF P
Blood urea nitrogen (mg/dl) 102.5 (80.5–150.5) 86.0 (68.5–109.0)* 94.0 (69.0–138.0)* <0.0001
Creatinine (mg/dl) 2.5 (1.4–3.5) 1.8 (1.2–2.9)* 1.9 (1.2–2.8)* <0.0001
pH 7.38 (7.34–7.45) 7.40 (7.35–7.46) 7.39 (7.33–7.45) NS
HCO3- (mmol/l) 23.9 (21.3–25.9) 24.0 (22.4–25.4) 24.0 (22.1–25.7) NS
Reported values are median (25th to 75th percentiles); P values determined using Friedman's two-way analysis of varience with post-hoc Wilcoxon signed rank test. *P < 0.0001 versus baseline. PHVHF, pulse high volume haemofiltration.
==== Refs
Laupland KB Davies HD Church DL Louie TJ Dool JS Zygun DA Doig CJ Bloodstream infection-associated sepsis and septic shock in critically ill adults: a population-based study Infection 2004 32 59 64 15057568 10.1007/s15010-004-3064-6
Dellinger RP Cardiovascular management of septic shock Crit Care Med 2003 31 946 955 12627010 10.1097/01.CCM.0000057403.73299.A6
Friedman G Silva E Vincent JL Has the mortality of septic shock changed with time Crit Care Med 1998 26 2078 2086 9875924 10.1097/00003246-199812000-00045
Angus DC Linde-Zwirble WT Lidicker J Clermont G Carcillo J Pinsky MR Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of care Crit Care Med 2001 29 1303 1310 11445675 10.1097/00003246-200107000-00002
Annane D Sebille V Charpentier C Bollaert PE Francois B Korach JM Capellier G Cohen Y Azoulay E Troche G Effect of treatment with low doses of hydrocortisone and fludrocortisone on mortality in patients with septic shock JAMA 2002 288 862 871 12186604 10.1001/jama.288.7.862
Cooper MS Stewart PM Corticosteroid insufficiency in acutely ill patients N Engl J Med 2003 348 727 734 12594318 10.1056/NEJMra020529
Bernard GR Vincent JL Laterre PF LaRosa SP Dhainaut JF Lopez-Rodriguez A Steingrub JS Garber GE Helterbrand JD Ely EW Efficacy and safety of recombinant human activated protein C for severe sepsis N Engl J Med 2001 344 699 709 11236773 10.1056/NEJM200103083441001
Venkataraman R Kellum JA Palevsky P Dosing patterns for continuous renal replacement therapy at a large academic medical center in the United States J Crit Care 2002 17 246 250 12501152 10.1053/jcrc.2002.36757
Silvester W Mediator removal with CRRT: complement and cytokines Am J Kidney Dis 1997 S38 S43 9372978
Cole L Bellomo R Hart G Journois D Davenport P Tipping P Ronco C A Phase II randomized controlled trial of continuous hemofiltration in sepsis Crit Care Med 2002 30 100 106 11902250 10.1097/00003246-200201000-00016
Kellum JA Mehta RL Angus DC Palevsky P Ronco C ADQI Workgroup The first international consensus conference on continuous renal replacement therapy Kidney Int 2002 62 1855 1863 12371989 10.1046/j.1523-1755.2002.00613.x
Ronco C Bellomo R Homel P Brendolan A Dan M Piccinni P La Greca G Effects of different doses in continuous venovenous haemofiltration on outcomes of acute renal failure: a prospective randomised trial Lancet 2000 356 26 30 10892761 10.1016/S0140-6736(00)02430-2
Grootendorst AF van Bommel EF van Leengoed LA Nabuurs M Bouman CS Groeneveld AB High volume hemofiltration improves hemodynamics and survival of pigs exposed to gut ischemia and reperfusion Shock 1994 2 72 78 7735987
Grootendorst AF van Bommel EF van der Hoven B van Leengoed LA van Osta AL High volume hemofiltration improves right ventricular function in endotoxin-induced shock in the pig Intensive Care Med 1992 18 235 240 1430589 10.1007/BF01709839
Rogiers P Zhang H Smail N Pauwels D Vincent JL Continuous venovenous hemofiltration improves cardiac performance by mechanisms other than tumor necrosis factor-alpha attenuation during endotoxic shock Crit Care Med 1999 27 1848 1855 10507609 10.1097/00003246-199909000-00024
Bellomo R Kellum JA Gandhi CR Pinsky MR Ondulik B The effect of intensive plasma water exchange by hemofiltration on hemodynamics and soluble mediators in canine endotoxemia Am J Respir Crit Care Med 2000 161 1429 1436 10806135
Cole L Bellomo R Journois D Davenport P Baldwin I Tipping P High-volume haemofiltration in human septic shock Intensive Care Med 2001 27 978 986 11497156 10.1007/s001340100963
Honore PM Jamez J Wauthier M Lee PA Dugernier T Pirenne B Hanique G Matson JR Prospective evaluation of short-term, high-volume isovolemic hemofiltration on the hemodynamic course and outcome in patients with intractable circulatory failure resulting from septic shock Crit Care Med 2000 28 3581 3587 11098957 10.1097/00003246-200011000-00001
Joannes-Boyau O Rapaport S Bazin R Fleureau C Janvier G Impact of high volume hemofiltration on hemodynamic disturbance and outcome during septic shock ASAIO J 2004 50 102 109 14763500 10.1097/01.MAT.0000104846.27116.EA
Oudemans-van Straaten HM Bosman RJ van der Spoel JI Zandstra DF Outcome of critically ill patients treated with intermittent high-volume haemofiltration: a prospective cohort analysis Intensive Care Med 1999 25 814 821 10447538 10.1007/s001340050957
Brendolan A D'Intini V Ricci Z Bonello M Ratanarat R Salvatori G Bordoni V De Cal M Andrikos E Ronco C Pulse high volume hemofiltration Int J Artif Organs 2004 27 398 403 15202817
Tetta C Bellomo R Kellum J Ricci Z Pohlmeiere R Passlick-Deetjen J Ronco C High volume hemofiltration in critically ill patients: why, when and how? Contrib Nephrol 2004 144 362 375 15264423
Bone RC Sibbald WJ Sprung CL The ACCP-SCCM consensus conference on sepsis and organ failure Chest 1992 101 1481 1483 1600757
Bellomo R Ronco C Indications and criteria for initiating renal replacement therapy in the intensive care unit Kidney Int 1998 S106 S109
Dellinger RP Carlet JM Masur H Gerlach H Calandra T Cohen J Gea-Banacloche J Keh D Marshall JC Parker MM Surviving Sepsis Campaign guidelines for management of severe sepsis and septic shock Intensive Care Med 2004 30 536 555 14997291 10.1007/s00134-004-2398-y
Knaus WA Draper EA Wagner DP Zimmerman JE APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Le Gall JR Lemeshow S Saulnier F A new Simplified Acute Physiology Score (SAPS II) based on a European/North American multicenter study JAM 1993 270 2957 2963 10.1001/jama.270.24.2957
Morris JA Gardner MJ Calculating confidence intervals for relative risks (odds ratios) and standardised ratios and rates Br Med J (Clin Res Ed) 1988 296 1313 1316 3133061
Wheeler AP Bernard GR Treating patients with severe sepsis N Engl J Med 1999 340 207 214 9895401 10.1056/NEJM199901213400307
Adib-Conquy M Adrie C Moine P Asehnoune K Fitting C Pinsky MR Dhainaut JF Cavaillon JM NF-kappaB expression in mononuclear cells of patients with sepsis resembles that observed in lipopolysaccharide tolerance Am J Respir Crit Care Med 2000 162 1877 1883 11069829
Goldfarb S Golper TA Proinflammatory cytokines and hemofiltration membranes J Am Soc Nephrol 1994 5 228 232 7994003
Gattinoni L Brazzi L Pelosi P Latini R Tognoni G Pesenti A Fumagalli R A trial of goal-oriented hemodynamic therapy in critically ill patients: SvO2 Collaborative Group N Engl J Med 1995 333 1025 1032 7675044 10.1056/NEJM199510193331601
Hayes MA Timmins AC Yau EH Palazzo M Hinds CJ Watson D Elevation of systemic oxygen delivery in the treatment of critically ill patients N Engl J Med 1994 330 1717 1722 7993413 10.1056/NEJM199406163302404
|
16137340
|
PMC1269433
|
CC BY
|
2021-01-04 16:04:53
|
no
|
Crit Care. 2005 Apr 28; 9(4):R294-R302
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3529
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35311613734310.1186/cc3531ResearchInspiratory oscillatory flow with a portable ventilator: a bench study Frank Guenther E [email protected] Helmut [email protected] Robert D [email protected] Director, Department of Anaesthesiology and Intensive Care, General Hospital Barmherzige Brüder Eisenstadt, Austria2 Director, Department of Anaesthesiology and Intensive Care, General Hospital Wiener Neustadt, Austria3 Director, Ludwig Boltzmann Institute for Economics of Medicine in Anesthesia and Intensive Care, Vienna, Austria2005 17 5 2005 9 4 R315 R322 7 2 2005 1 3 2005 24 3 2005 6 4 2005 Copyright © 2005 Frank et al, licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
We observed an oscillatory flow while ventilating critically ill patients with the Dräger Oxylog 3000™ transport ventilator during interhospital transfer. The phenomenon occurred in paediatric patients or in adult patients with severe airway obstruction ventilated in the pressure-regulated or pressure-controlled mode. As this had not been described previously, we conducted a bench study to investigate the phenomenon.
Methods
An Oxylog 3000™ intensive care unit ventilator and a Dräger Medical Evita-4 NeoFlow™ intensive care unit ventilator were connected to a Dräger Medical LS800™ lung simulator. Data were registered by a Datex-S5™ Monitor with a D-fend™ flow and pressure sensor, and were analysed with a laptop using S5-Collect™ software. Clinical conditions were simulated using various ventilatory modes, using various ventilator settings, using different filters and endotracheal tubes, and by changing the resistance and compliance. Data were recorded for 258 combinations of patient factors and respirator settings to detect thresholds for the occurrence of the phenomenon and methods to overcome it.
Results
Under conditions with high resistance in pressure-regulated ventilation with the Oxylog 3000™, an oscillatory flow during inspiration produced rapid changes of the airway pressure. The phenomenon resulted in a jerky inspiration with high peak airway pressures, higher than those set on the ventilator. Reducing the inspiratory flow velocity was effective to terminate the phenomenon, but resulted in reduced tidal volumes.
Conclusion
Oscillatory flow with potentially harmful effects may occur during ventilation with the Dräger Oxylog 3000™, especially in conditions with high resistance such as small airways in children (endotracheal tube internal diameter <6 mm) or severe obstructive lung diseases or airway diseases in adult patients.
==== Body
Introduction
Transport ventilators, until recently, were simple flow interrupters with constant flow, allowing only a few parameters to be changed and with no, or only very limited, alarm and monitoring functions. These devices are still in use by emergency services and for mechanical ventilation of critically ill patients during intrahospital and interhospital transport [1]. The development of transport ventilators in recent years has introduced flow and pressure monitoring and has enabled the setting of positive end expiratory pressure (PEEP), inspiration to expiration ratio, and pressure limits. This made it possible to use these devices not only in emergency medicine, but also for transport of critically ill patients with severe lung injury. Nevertheless, the continuation of sophisticated mechanical ventilation during transport of critically ill patients with acute lung failure often still required the use of an intensive care ventilator. The higher weight, the higher power consumption, and the additional need for compressed air, as well as the larger dimensions, make transport with conventional intensive care ventilators more complicated and trouble-prone [2-6].
The Oxylog 3000™ transport ventilator (Dräger Medical, Best, The Netherlands) combines the properties of a modern intensive care ventilator with the advantages of a compact transport ventilator, such as low weight, small dimensions, and low power consumption. The main innovation of the Oxylog 3000™ is the possibility to use pressure-controlled, pressure-limited ventilation and pressure support.
We therefore used the Oxylog 3000™ routinely since 2002 for interhospital transfer, but have detected an undesired oscillatory flow during inspiration in paediatric patients and in adult patients with airway obstruction. The phenomenon occurred during pressure-regulated or pressure-limited ventilation and was characterised by four to eight rapid changes in flow velocity. The peak airway pressure exceeded the previously set pressure values and the phenomenon was accompanied by a reduction in minute ventilation. The phenomenon clinically impressed with a staccato-like breathing sound, similar to jet ventilation, and it was sometimes possible to detect a jerky thorax excursion during inspiration, even if the patient received neuromuscular blocking agents. Following these experiences, we conducted a bench study simulating different ventilator settings and respiratory conditions with the Oxylog 3000™ in comparison with a standard intensive care respirator – the Evita-4 NeoFlow™ (Dräger Medical).
Materials and methods
The Oxylog 3000™ allows the setting of all common modes of ventilation used in critically ill patients, including intermittent positive pressure ventilation (IPPV), biphasic intermittent positive airway pressure (BIPAP), which can be used as pressure controlled ventilation, assisted spontaneous breathing (ASB), which is equivalent to pressure support ventilation, continuous positive airway pressure, synchronised intermittent mandatory ventilation, and noninvasive ventilation with leakage compensation. Flow is generated and regulated by means of four magnetic valves. Only oxygen is required as the gas supply since ambient air is added for adjustment of the FIO2 from 40% to 100% by means of a Venturi valve. Further adjustable parameters are the I:E ratio, the tidal volume (VT), the respiratory rate, the pressure limit, the PEEP, the ramp of inspiratory flow in BIPAP and ASB (slow, standard, fast), and the flow trigger. A flow sensor is positioned close to the patient. The pressure curve, the flow curve and the following parameters can be shown on the display of the Oxylog 3000™: peak Paw, mean Paw, plateau Paw, PEEP, expiratory VT, respiratory rate and expiratory minute volume.
The setting of the bench study is demonstrated in Fig. 1. The ventilators, the lung simulator, and the test laboratory were provided by Dräger Medical™ (Vienna, Austria). Prior to performing the tests, all apparatus were checked for faults and correct function. Reusable tubing was used for both ventilators. A spirometry sensor, a heat and moisture exchange filter (DAR Tyco™ Healthcare, Mansfield, MA, USA), and an endotracheal tube were connected between the ventilator and the lung simulator in an airtight manner by means of the inflated cuff, which was checked for leakage prior to the measurements. The spirometry was performed with a Datex-S5™ monitor (Datex-Ohmeda™, Helsinki, Finland) with D-fend™ sensors in different sizes (paediatric, adult). This Datex-S5™ monitor is routinely used in anaesthesia and intensive care medicine, and uses a double line sensor inserted between the heat and moisture exchange filter and the tubing. The Datex-S5™ monitor was connected to a laptop using specific software (S5-Collect™, Datex-Ohmeda™, Helsinki, Finland) to store and analyse the measured data.
The stepwise changed parameters and respirator settings of the 258 tests are summarised in Table 1. The special combinations of patient factors and respirator settings were chosen to detect thresholds for the occurrence of the phenomenon. All measurements in the IPPV mode were taken with a pressure limit of 20 cmH2O.
The following parameters were recorded in all tests. From the display of the ventilators, the respiratory rate, the VT, the peak Paw, the FIO2, warnings and alarms were read and recorded manually. The S5-Collect™ software stored the data measured by the spirometry module, including the peak Paw, the mean Paw, the plateau Paw, the PEEP, the intrinsic PEEP, the inspiratory VT, the expiratory VT, the compliance, the resistance, the duration of inspiration and the I:E ratio.
For each measurement the Paw curve and the flow curve of six respiratory cycles and the trend data of all measured parameters were stored in a separate file. The curves were quantitatively analysed with Microsoft Excel™ to evaluate the duration of the oscillations as a percentage of the inspiration time, the frequency of the oscillations, the amplitudes of the pressure oscillation (Paw-ampl), the amplitude of the flow oscillation, and the maximal inspiratory flow (Figs 2 and 3). If the peak Paw was higher than the set inspiratory pressure in the pressure-regulated modes or higher than the set pressure limit in IPPV, the difference between these values was calculated and stored as the airway pressure overshoot (Fig. 2). The magnitude of the Paw-ampl and the amount of the airway pressure overshoot were used to describe the severity of the phenomenon.
Selected tests were performed with both the Oxylog 3000™ and the Evita-4 NeoFlow™ to validate the measurements and to compare ventilation with the two ventilators under exactly the same conditions. One hundred and ninety-eight tests were performed with the Oxylog 3000™ and 60 comparative measurements were taken with the Evita-4 NeoFlow™, producing a total of 258 tests. Differences in the peak Paw, the mean Paw, and the expiratory VT between the two ventilators were calculated.
The maximal inspiratory flow velocity seemed to have an important influence on the occurrence and severity of the phenomenon, and we therefore calculated the ratio of the maximal flows between the Oxylog 3000™ and the Evita-4 NeoFlow™.
Statistics
As indicated by Kolmogorov-Smirnov tests, the data showed deviations from a normal distribution, thus precluding the computation of parametric descriptive and inference statistics. Results are thus presented as the median with the interquartile range, minimum and maximum, and the Spearman rank correlations were computed. The Mann-Whitney U test was used to examine the differences between the Oxylog 3000™ and the Evita-4 NeoFlow™. P ≤ 0.05 was considered significant.
Results
No oscillatory flow was detected in any test using the Evita-4 NeoFlow™ respirator. Overall with the Oxylog 3000™, an oscillatory flow was detected in 90% of all respective measurements. The phenomenon was seen in the pressure-regulated modes BIPAP, ASB, continuous positive airway pressure, and in pressure-limited IPPV. No significant difference in the expiratory VT was detected when comparing the Oxylog 3000™ measurements with oscillatory inspiratory flow with the corresponding Evita-4 NeoFlow™ tests. Nevertheless, the oscillations resulted in significant higher peak and mean Paw on comparing the two ventilators (Table 2).
The duration and the shape of the pressure oscillations depended on the mode, on the ramp, and on whether a leakage was simulated. In general the curve oscillated around the normally seen Paw curve, which could be observed in comparison with the Paw curve of the Evita-4 NeoFlow™ measurements.
It was possible to measure the frequency of the oscillations in 153 tests with a median frequency of 5 Hz(interquartile range, 1.25 Hz; minimum, 2.78 Hz; maximum, 12.5 Hz). There was a trend to lower frequencies of the oscillations when the phenomenon was more severe. In the tests with a measured frequency of 4.17 Hz the median Paw-ampl was 23.8 cmH2O, with a frequency of 5 Hz the median was 15.8 cmH2O, and with a frequency of 12.5 Hz the median Paw-ampl was 6.9 cmH2O.
Concerning the severity of the phenomenon, the median values of Paw-ampl and the amount of the airway pressure overshoot are summarised in Table 3.
The severity of the phenomenon increased when the resistance on the LS800™ lung simulator (Dräger Medical, Best, The Netherlands) was increased (Fig. 4). The steepness of the ramp, set on the Oxylog 3000™, correlated positively with the severity of the oscillations (Fig. 5).
Changing the compliance on the LS800™ lung simulator did not have any influence on the occurrence and severity of the phenomenon. This was also true for the respiratory rate, the PEEP, the time of inspiration, and the I:E ratio. An intrinsic PEEP was detected in 126 of the Oxylog 3000™ measurements but did not show any correlation to the phenomenon. No influence of the set FIO2 on the occurrence of the phenomenon was seen. The phenomenon also occurred with 100% oxygen when the Venturi valve was not active.
We investigated the differences between the maximal inspiratory flow velocities generated by the two ventilators. The flow generated by the Oxylog 3000™ was usually higher than that with the Evita-4 NeoFlow™. The ratio of the maximal flows between the Oxylog 3000™ and the Evita-4 NeoFlow™ correlated well to the severity of the phenomenon, expressed as Paw-ampl (Fig. 6).
The oscillations almost exclusively occurred during inspiration. In general, the flow pattern of the expirations was no different compared with the corresponding Evita-4 NeoFlow™ measurements. An oscillatory flow during expiration was only seen in some of the measurements with simulated leakage, when the ventilator had to generate a flow directed to the test lung during expiration to maintain the PEEP. The oscillations during expiration had a Paw-ampl of 5 cmH2O and a frequency of 4.17 Hz. In comparison with equivalent measurements without leakage there seemed to be an attenuating effect of the leakage on the severity of the inspiratory oscillations.
Discussion
While no oscillatory flow could be detected with the Evita-4 NeoFlow™, the phenomenon was found in a high percentage of tests with the Oxylog 3000™. We have to point out, however, that this high percentage is due to the setting of our tests, chosen to induce and investigate the phenomenon. Fifty per cent of the tests were taken with small endotracheal tubes, and the phenomenon was surprisingly seen in all tests with endotracheal tubes with internal diameter ≤ 6 mm in pressure-regulated modes, irrespective of the test lung conditions. Only five tests with endotracheal tubes of internal diameter ≤ 6 mm showed no oscillatory flow, and all of them were taken in the IPPV mode without reaching the pressure limit (constant flow).
Two parameters had an impact on the occurrence and severity of the phenomenon: the resistance, and the peak velocity of the inspiratory flow. The latter is influenced by the ramp set on the Oxylog 3000™ and by the interaction of test lung conditions and ventilator settings. The ratio of the maximal inspiratory flow measured with the Oxylog 3000™ to the maximal flow measured with the Evita-4 NeoFlow™ is another way to describe an inappropriate high flow at the beginning of the inspiration, and the value correlated well to the severity of the phenomenon.
The following hypothesis was made to explain why an oscillatory flow occurs under conditions with high resistance and high initial flow. The maximal inspiratory flow, reached during pressure-regulated ventilation, mainly depends on the airway resistance. The initial flow generated by the Oxylog 3000™ in the BIPAP and ASB modes depends on the ramp, and in the mode of pressure-limited IPPV it depends on the set VT. The flow, initially generated by the Oxylog 3000™, sometimes is much higher than the flow that can traverse the resistance set on the test lung. After initiation of the inspiration with an inappropriate high flow, the inspiratory pressure or the pressure limit (set on the Oxylog 3000™) is reached very rapidly and the flow is downregulated or stopped by the Oxylog 3000™. This does not occur rapidly enough and the Paw exceeds the target value. The pressure drops after the reduction or interruption of the flow and the flow is generated too late and too high again. Thus the pressure oscillates around the desired level. The oscillations are a result of rapid changes between exceedingly high and low flow velocities. The feedback mechanism between measured Paw and flow generation does not work rapidly enough or sensitively enough to smoothly adjust the inspiratory flow to an appropriate level.
We might explain the expiratory oscillation, seen in some measurements with a simulated leakage, by the fact that during expiration the leakage has to be compensated by a flow delivered by the Oxylog 3000™ to maintain the PEEP.
The results of the bench study predict a high probability for the phenomenon to occure in paediatric patients with narrow airways. This is exactly what we have seen in clinical practice. The phenomenon occurred frequently in paediatric patients and it was not possible to use BIPAP in patients with an endotracheal tube < 6 mm ID. The mode had to be changed to IPPV, but an inspiratory oscillatory flow still occurred in the IPPV when the pressure limit was active. We adjusted the VT carefully to avoid an oscillatory flow, on the one hand, and to avoid low minute ventilation, on the other.
You would not expect the Paw to be higher than that set on the ventilator in a pressure-regulated or pressure-limited mode, but exactly this happens when an oscillatory flow occurs. Unfortunately we have not measured the pressures in the test lung, but the following points led to our conclusion that the phenomenon is potentially dangerous and harmful. There is an airway pressure overshoot, the mean airway pressure is increased and the peak airway pressure may reach values above 50 cmH2O. The oscillations led to a jerky inspiration, showing that the pressure spikes really do reach the lung. Finally the pressure limit of the Oxylog 3000™ does not protect against the pressure overshoot.
The phenomenon of oscillatory inspiratory flow may impose as a malfunction of the device but it actually reflects a kind of limitation, of which the user should be aware and know how to deal with. We informed Dräger Medical in The Netherlands about our experiences and the results of the bench study. In the meantime, Dräger Medical started their own measurements and confirmed the validity of the problem. An adaptation of the operator's manual of the Oxylog 3000™ seems necessary, especially because the Oxylog 3000™ is licensed for ventilation with a very low VT (≥ 50 ml).
Limitations
Spirometry was not obtained by a pneumotachograph, but with a spirometry module normally used for clinical purpose. This may especially affect the measurements of the compliance and the resistance, particularly in the tests with an oscillatory flow. Nevertheless, the flow curves and pressure curves obtained with this device were of good quality, and the results of the VT and other parameters were in accordance with the values displayed by the ventilators.
Conclusion
Under conditions with high resistance an oscillatory inspiratory flow may occur during ventilation with the Oxylog 3000™ in the BIPAP, ASB, continuous positive airway pressure, and pressure-limited IPPV modes. The phenomenon results in elevated airway pressures and jerky inspiration. The unexpected high airway pressures may be potentially harmful, and therefore ventilation should be checked for the phenomenon in paediatric patients with narrow endotracheal tubes and in adult patients with severe obstructive airway or lung disease. If oscillations are present, the ventilator setting has to be adjusted by reducing the steepness of the ramp in BIPAP and ASB or by reducing the VT in pressure-limited IPPV.
Key messages
• An oscillatory flow during inspiration may occur in pressure-regulated modes with the Oxylog 3000™, especially when airway resistance is high.
• The oscillatory flow results in a jerky inspiration and in elevated airway pressures.
• Peak airway pressures are markedly elevated above the set upper pressure limit and may cause lung injury.
Abbreviations
ASB = assisted spontaneous breathing; BIPAP = biphasic intermittent positive airway pressure; FIO2 = fraction of inspired oxygen; I:E = inspiration to expiration time ratio; IPPV= intermittent positive pressure ventilation; Paw = airway pressure; Paw-ampl = amplitudes of the pressure oscillation; PEEP = positive end expiratory pressure; VT = tidal volume.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
GEF discovered the phenomenon under clinical conditions, designed the study, conducted the bench study, and analysed the results. HT assisted in designing the study and participated in interpreting the results. RDF performed statistical analysis and drafted the manuscript. All authors read and approved the final manuscript.
Acknowledgements
The authors thank Claus Lamm, Ph.D, certified statistician, for his support in statistical analysis. They also thank Dräger Medical™ Austria for providing the test laboratory, the tested ventilators, and the lung simulator, and Sanitas™ Austria for supplying the Datex-S5™ monitor with a spirometry module and the S5-Collect™ software.
This article won the first scientific award of the ÖAMTC Christophorus Helicopter Emergency Medical Service, 2004, and was in part presented at the XIII Innsbrucker Notfallsymposium, Innsbruck, 5–6 November 2004.
Figures and Tables
Figure 1 Scheme for the experimental set-up. HME, heat and moisture exchange filter.
Figure 2 Airway pressure oscillation (Paw). Typical oscillatory Paw curve of the Oxylog 3000™ (solid line) in comparison with the Evita-4 NeoFlow™ (broken line). Settings: endotracheal tube internal diameter, 5 mm; biphasic intermittent positive airway pressure, 20/5 cmH2O; respiratory rate, 20/min; inspiration to expiration time ratio, 1:1; ramp, fast. The maximal amplitude of the pressure oscillation is the airway pressure amplitude (Paw-amplitude). Paw-overshoot, airway pressure overshoot.
Figure 3 Flow curve oscillation. Typical oscillatory flow curve of the Oxylog 3000™ (solid line) in comparison with the Evita-4 NeoFlow™ (broken line), with the same settings as Fig. 1. The minimal flow was calculated by dividing the expiratory tidal volume, measured by the Oxylog 3000™, through the time of inspiration.
Figure 4 Influence of the test lung – resistance. Stepwise increase of the resistance on the LS800™ lung simulator resulted in an increase of the amplitude of the airway pressure oscillations (Paw-amplitude) as well as in an increase in the airway pressure overshoot (Paw-overshoot), defined as peak airway pressure minus the upper pressure limit.
Figure 5 Influence of the steepness of the ramp on the phenomenon. A stepwise increase of the ramp, set on the Oxylog 3000™, resulted in an increase of the amplitude of the airway pressure oscillations (Paw-amplitude) as well as in an increase in the airway pressure overshoot (Paw-overshoot), defined as the peak airway pressure minus the upper pressure limit.
Figure 6 Peak inspiratory flow. Correlation between the airway pressure amplitude (Paw-amplitude) and the ratio of the peak inspiratory flow with the Oxylog 3000™ to the peak inspiratory flow with the Evita-4 NeoFlow™ (flow-ratio ox/ev). Only measurements in the biphasic intermittent positive airway pressure or the intermittent positive pressure ventilation modes without leakage and without single spikes in the flow curve were included.
Table 1 Course of the measurements including the settings of the lung simulator and the ventilators
Compliance, LS800™ (l/cmH2O) Resistance, LS800™ (cmH2O/l/s) Leak, LS800™ Endotracheal tube ID (mm) HME type D-fend™ type Mode FIO2 Respiratory rate (/min) Tinsp (s) I:E ratio PEEP (cmH2O) Pinsp (cmH2O) Ramp, Oxylog 3000™ VT (ml) Number of measurements, Oxylog 3000™ Comparable measurements, Evita-4 NeoFlow™
0.010 2, 4, 8, 16, 32, 64, 128 No 4 Baby Paed BIPAP 0.4 20 1.5 1/1 5 20 Slow std fast 21 15
0.007 2, 32, 64, 128 No 4 Baby Paed BIPAP 0.4 30 1 1/1 5 20 Slow std fast 12 4
0.015 2, 32, 64, 128 No 4 baby Paed BIPAP 0.4 20 1.5 1/1 5 20 Std 4 4
0.020 2, 32, 64, 128 No 5 Baby Paed BIPAP 0.4 20 1.5 1/1 5 20 Slow std fast 12 4
0.020 32 No 5 Baby Paed IPPV 0.4 20 1.5 1/1 5 180 – 500 8 0
0.020 2, 8, 32, 64, 128 No 7 Adult Adult BIPAP 0.8 20 1.5 1/1 15 35 Slow std fast 15 5
0.075 2, 8, 16, 32, 64, 128 No 7 Adult Adult BIPAP 0.8 16 1.9 1/1 10 28 Slow std fast 18 6
0.020 2, 8, 16, 32 No 7 Adult Adult ASB 0.8 6 20 Slow std fast 12 0
0.020 2, 8, 16, 32 No 7 Adult Adult CPAP 0.5 6 6 4 0
0.020 2, 8, 16, 32 No 7 Adult Adult ASB 0.5 6 20 Slow std fast 12 4
0.030 2, 8, 16, 32, 64, 128 No 7 Adult Adult BIPAP 0.5 12 1.6 1/2 8 26 Slow std fast 18 6
0.020 2, 8, 16, 32, 64, 128 No 7 Adult Adult BIPAP 0.5 12 1.6 1/2 8 26 Slow std fast 18 6
0.020 2 no 5 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 Yes 5 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 no 5.5 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 Yes 5.5 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 No 6 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 Yes 6 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Slow std fast 3 0
0.020 2 Yes 5 Paed Paed BIPAP 0.5 20 1.2 1/1.5 5 20 Std 1 0
0.020 2 Yes 5 Paed Paed BIPAP 0.4, 0.6, 0.8, 1 20 1.2 1/1.5 5 20 Fast 4 0
0.020 2 Yes 5 Paed Paed BIPAP 0.4, 0.6, 0.8, 1 20 1.2 1/1.5 5 20 Std 4 0
0.020 2 Yes 5 Paed Paed IPPV 0.5 20 1.2 1/1.5 5 240 – 600 8 0
0.020 2 No 5 Paed Paed IPPV 0.5 20 1.2 1/1.5 5 180 – 600 9 0
Related test-series with changes in one or maximal two parameters are grouped in one row. Numbers are values of the set parameters and do not reflect measured results. ID, internal diameter; Tinsp = inspiration time; I:E = inspiration to expiration time ratio; PEEP = positive end expiratory pressure; Pinsp, = set inspiratory airway pressure; VT = tidal volume; Paed, paediatric; std, standard; BIPAP = biphasic intermittent positive airway pressure; IPPV = intermittent positive pressure ventilation; ASB = assisted spontaneous breathing; CPAP, continuous positive airway pressure;
Table 2 Comparison between measurements showing oscillatory inspiratory flow andcorresponding Evita-4 NeoFlow™ measurements.
Parameter n Median P value
Oxylog 3000™ Evita-4 NeoFlow™
Peak airway pressure (cmH2O) 40 32.65 20.8 <0.001
Mean airway pressure (cmH2O) 40 13.05 12.1 0.007
Expiratory tidal volume (ml) 40 133.05 136.35 0.840
Peak inspiratory flow (l/s) 40 16.05 11.85 0.023
Only tests with biphasic intermittent positive airway pressure and without leakage were included.
Table 3 Severity of the phenomenon
Parameter Median Interquartile range Minimum Maximum n
Pressure amplitude (cmH2O) 12.15 17.75 0.10 37.9 153
Flow amplitude (l/s) 13.35 27.33 0.10 63.8 153
Airway pressure overshoot (cmH2O) 6.6 12.53 0.10 27.6 128
==== Refs
Warren J Fromm RE Orr RA Rotello LC Horst HM Guidelines for the inter- and intrahospital transport of critically ill patients Crit Care Med 2004 32 256 262 14707589 10.1097/01.CCM.0000104917.39204.0A
Zanetta G Robert D Guérin C Evaluation of ventilators used during transport of ICU patients – a bench study Intensive Care Med 2002 28 443 451 11967599 10.1007/s00134-002-1242-5
Waydhas C Intrahospital transport of critically ill patients Crit Care 1999 3 R83 R89 11094486 10.1186/cc362
Stevenson VW Haas CF Wahl WL Intrahospital transport of the adult mechanically ventilated patient Respir Care Clin North Am 2002 8 1 35 10.1016/S1078-5337(02)00014-X
Uusaro A Parviainen I Takala J Ruokonen E Safe long-distance interhospital ground transfer of critically ill patients with acute severe unstable respiratory and circulatory failure Intensive Care Med 2002 28 1122 1125 12185435 10.1007/s00134-002-1348-9
Reynolds HN Habashi NM Cottingham CA Frawley PM McCunn M Interhospital transport of the adult mechanically ventilated patient Respir Care Clin North Am 2002 8 37 50 10.1016/S1078-5337(02)00015-1
|
16137343
|
PMC1269434
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 17; 9(4):R315-R322
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3531
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35321613734410.1186/cc3532ResearchTime course of endothelial damage in septic shock: prediction of outcome Hein Ortrud Vargas [email protected] Klaudia 1Tessmann Jan-Peer 1van Dossow Vera 1Krimphove Michael 1Spies Claudia [email protected] Department of Anesthesiology and Intensive Care, University Hospital Charité, Campus Mitte, Berlin, Germany2005 13 5 2005 9 4 R323 R330 7 11 2004 9 1 2005 29 3 2005 7 4 2005 Copyright © 2005 Vargas Hein et al, licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
Endothelial damage accounts greatly for the high mortality in septic shock. Higher expression of mediators (IL-6, IL-8, soluble intercellular adhesion molecule 1 [sICAM-1], soluble endothelial-linked adhesion molecule 1 [sELAM-1]) have been described for non-survivors in comparison with survivors. We investigated the predictive value of the mediators IL-6, IL-8, sELAM-1 and sICAM-1 and their time course in intensive care unit patients who developed septic shock with respect to outcome.
Materials and methods
We measured serum levels of IL-6, IL-8, sELAM-1 and sICAM-1 in 40 intensive care unit patients who developed septic shock. Measurements were performed until death or until resolution of septic shock. Clinical and laboratory data were also recorded.
Results
After 48 hours the levels of sELAM-1 and sICAM-1 increased in non-survivors and decreased in survivors. sELAM-1 was predictive for outcome on the third day (P = 0.02) and the fourth day (P = 0.02) after diagnosis of septic shock. This difference in the time course between survivors and non-survivors occurred 7 days before death of the patients (median, 10 days). sICAM-1 levels increased significantly in non-survivors over the study period (P < 0.001). sELAM-1 (P = 0.04), IL-6 (P = 0.04) and IL-8 (P = 0.008) were significantly higher in non-survivors over the whole study period. The age and norepinephrine dose >0.5 μg/kg/min were significantly different between the groups.
Conclusion
sELAM-1 showed a markedly opposing course after 48 hours of septic shock. This adhesion molecule may be a useful early predictor of disease severity in the course of septic shock after early initial treatment of the patients, and might suggest considering endothelial-restoring therapy.
==== Body
Introduction
Endothelial damage accounts for much of the pathology of sepsis, resulting in capillary leak, hypotension, microvascular thrombosis with consecutive tissue hypoxia and, finally, multiple organ failure (MOF) and lethal outcome [1-3]. Endothelial damage is worsened in septic shock [4]. The mortality of septic shock is higher than the mortality in sepsis (35–60% versus 20–40%) [4,5]. The release of cytokines (IL-6, IL-8) and adhesion molecules (soluble endothelial-linked adhesion molecule 1 [sELAM-1], soluble intercellular adhesion molecule 1 [sICAM-1]) has been shown to correlate well with endothelial damage in an experimental setting – especially for sELAM-I, which is specific for endothelial tissue [2,6,7]. Although the release of these mediators is not only sepsis related, the levels are significantly higher in sepsis and in septic shock than after trauma, postoperatively or after myocardial infarction [8-12]. In addition, these mediators have higher levels in non-survivors than in survivors, and the baseline levels have been correlated with outcome [2,3,8,10-15].
The time of admission to the study and the onset of therapy are of major relevance for outcome, however, as shown by Rivers and colleagues in the early goal-directed therapy study in severe sepsis and septic shock patients [16]. As early clinical intervention improves outcome and as there are increasing levels of cytokines in non-survivors, in comparison with a decrease in survivors, differences in the mediator time course between survivors and non-survivors after early onset of therapy could be predictive for the outcome and for trend-setting for further therapy measures [10,11,15,17-19].
We investigated the predictive value of the mediators IL-6, IL-8, sELAM-1 and sICAM-1 and their time course, as primary outcome measures, in intensive care unit (ICU) patients who developed septic shock with respect to outcome. In addition, IL-8 as an early chemoattractant cytokine and IL-6 as an inflammatory tissue damage marker were investigated. Clinical data, such as age, the use of hemodynamically active substances and myocardial ischemia, were investigated as secondary outcome measures.
Materials and methods
Patients
After ethical committee approval and written informed consent from the legal representatives, 42 patients suffering from septic shock were enrolled in this observational study. Two patients had to be excluded after enrollment because of imminent surgery, so 40 patients completed the study. All patients fulfilled the clinical and laboratory criteria of septic shock as outlined in the 1992 Consensus Conference [20]. Exclusion criteria were age <18 years, pregnancy, patients who have had surgery within 48 hours before inclusion and patients who have had cardiac surgery and neurosurgery. Patients with an acute history of severe cardiac insufficiency (New York Heart Association class III-IV) [21] and coronary artery disease before the development of septic shock were also excluded [22].
Monitoring and management
The study was initiated in the first 24 hours after septic shock had been diagnosed. All patients were already admitted to the ICU and were under ICU standard therapy and monitoring [23]. All patients received analgesia, sedation and mechanical ventilation. The patients were screened twice a day. The study ended in the case of death or in resolution of septic shock.
A fiber optic pulmonary artery flotation catheter (Baxter Swan-Ganz® Intelicath™ continuous cardiac output thermodilution catheter 139H, 7.5 Fr; Baxter/Edwards Critical-Care, Irvine, CA, USA) and a radial artery catheter were inserted as part of the routine for continuous cardiovascular monitoring in septic shock. Hemodynamic measurements were recorded at study entry and every 8 hours during the study. Fluids were given to achieve an optimal left atrial pressure. After adequate fluid resuscitation, norepinephrine (maximum 4.0 μg/kg/min) was titrated to maintain a mean arterial pressure >70 mmHg. Catecholamine therapy in the case of low-output failure was performed primarily with dobutamine (maximum 20 μg/kg/min) or dopamine (maximum 10 μg/kg/min) at the discretion of the physician on duty. Enoximone (maximum 10 μg/kg/min) was added if low-output failure persisted, and then epinephrine infusion (maximum 2.0 μg/kg/min) was initiated if low-output failure remained. The target value was a cardiac index >3.0 l/min/m2. The amount of different positive inotropic substances was expressed as the number used in each group.
A 12-lead Holter electrocardiogram (ECG) was recorded every 8 hours to determine possible myocardial ischemia, defined by Spies and colleagues [22]. The oxygenation index was calculated as the quotient of partial arterial oxygen pressure and the inspired oxygen fraction (mmHg).
Group assignment
It was decided a priori to assign patients to the survivors group when they were discharged from the ICU to a regular ward. Those patients who died due to septic shock were assigned to the non-survivors group. Patients who died from a cause other than septic shock and consecutive MOF during their ICU stay were excluded from the study.
Laboratory data
Blood gas analysis was performed every 8 hours to determine the levels of hematrocrit and hemoglobin, and the arterial partial oxygen pressure (ABL 500; Radiometer, Copenhagen, Denmark).
Creatin kinase and the creatin kinase-myocardial bands were determined every 8 hours (BM/Hitachi 717 analyser; Boehringer Mannhein, Inc., Mannheim, Germany). The creatin kinase/creatin kinase-myocardial band fraction was calculated and a result >6% was recorded positive for myocardial ischemia [22]. Blood samples for the determination of IL-6 concentrations (Enzymeimmunoassay [Milenia®]; DPC Biermann GmbH, Bad Nauheim, Germany), of IL-8 concentrations (Enzymeimmunoassay [Milenia®]; DPC Biermann GmbH), of sICAM-1 concentrations (enzyme immunoassay kit BBE 1b; R&D Systems, Minneapolis, MN, USA), of sELAM-1 concentrations (enzyme immunoassay kit BBE 2b; R&D Systems) and of troponin T concentrations (enzyme-linked immunosorbent assay Enzymun-Test™ batch ELISA ES 300 analyser; Boehringer Mannheim Inc.) were withdrawn every 8 hours and were centrifuged, and the plasma was stored at -80°C until analysis.
Statistical analysis
Data are expressed as the median and range. Intergroup statistical analysis for determined time intervals was performed using the Mann–Whitney U test for continuous variables and using the Pearson chi-square test for dichotomous variables. Intragroup statistical analysis for the determined time intervals was performed with the Wilcoxon matched-pairs signed-rank sum test. For intergroup and intragroup analysis over the whole study period, the two-factorial non-parametric (analysis of variance)-type rank variance analysis for longitudinal data and small sample sizes using the SAS System software (SAS Institute Inc., Cary, NC, USA) was used. Variables that were significantly different between groups were analysed as predictors for outcome (group variable, survivor/non-survivor), determining the area under the receiver operating characteristics curve (AUC). The AUC, the P value and the 95% confidence intervals are stated. P < 0.05 was considered statistically significant.
Results
Forty patients were included in the study and 16 (40%) patients were discharged from the ICU to a normal ward. Twenty-four (60%) patients died due to septic shock. Patients in the non-survivor group were significantly older and stayed a significantly shorter time in the ICU than the survivors (Table 1). Survivors had a significantly higher rate of pneumonia as the sepsis focus whereas non-survivors had a significantly higher rate of peritonitis as the focus (Table 1). The Acute Physiology and Chronic Health Evaluation III baseline score and the Acute Physiology and Chronic Health Evaluation III maximum score did not significantly differ between the groups (Table 1). All patients required norepinephrine therapy but significantly more non-survivors than survivors required norephinephrine infusion >0.5 μg/kg/min (Table 2). The number of positive inotropic agents necessary and the markers for myocardial ischemia (monitored by ECG), for creatin kinase/creatin kinase-myocardial band fraction >6% and for troponin T were not significantly different between survivors and non-survivors (Table 2).
Intergroup analysis of variance between survivors and non-survivors showed significantly higher levels for IL-6 (P = 0.04), for IL-8 (P = 0.008) and for sELAM-1 (P = 0.04) in the non-survivors group. sICAM-1 (P = 0.25) was not significantly higher in levels in the non-survivors group. The intragroup analysis for IL-6 showed a significant decline between the first value and the last value (before discharge from the study or death) for survivors (P = 0.002) and non-survivors (P = 0.04) (Fig. 1). The intragroup analysis for IL-8 between the first value and the last value (before discharge from the study or death) was not significantly different in both groups (survivors, P = 0.17; non-survivors, P = 0.78) (Fig. 2).
After a comparable course in the first 2 days, non-survivors showed an increase in median values of sELAM-1 and sICAM-1 whereas survivors' adhesion molecule levels decreased markedly (Figs 3 and 4). This increase was significant for sICAM-1 in the non-survivor group when comparing the first value with last value before discharge from the study or death of the patients (P < 0.001) (Fig. 4). The marked decline of median values for sELAM-1 in the survivor group was significant in the comparison of the first time point with the last time point before discharge from the study or death (P = 0.04) (Fig. 3). When comparing survivors and non-survivors at single time points, sELAM-1 was significantly higher in non-survivors from the third day onwards (P = 0.02) (Fig. 3).
The AUC values for baseline, the third day and the fourth day measurements of IL-6, IL-8, sELAM-1 and sICAM-1 are presented in Table 3. IL-8 was most predictive for outcome at baseline, and sELAM-1 most predictive on the third and fourth days (Table 3). The AUC for age (AUC, 0.761; P = 0.01; 95% confidence interval, 0.624–0.898) and that for median norepinephrine dosage (AUC, 0.766; P = 0.001; 95% confidence interval, 0.636–0.896) were also significantly predictive for outcome.
Discussion
The most important finding in this study was the different time courses of the markers of endothelial damage (sELAM-1 and sICAM-1) after the second day in survivors and non-survivors of septic shock. After a comparable course at different levels in the first 2 days, non-survivors had an increase in adhesion molecule concentrations whereas survivors' adhesion molecule levels decreased markedly. SELAM-1 was predictive for outcome on the third and fourth days after the diagnosis of septic shock. This difference in time courses between survivors and non-survivors was evident on the third day and, therefore, far before death of the patients (median, 10 days).
Endothelial damage accounts for much of the pathology of septic shock, resulting finally in MOF and lethal outcome [1-3]. sELAM-1 is specific for endothelial tissue [2,7]. The latter marker and sICAM-1 have been shown to be significantly elevated at baseline and inconsistent in levels over the whole study period in sepsis, in comparison with trauma patients or critically ill patients without sepsis [2,3,8-12]. The levels of adhesion molecules in septic shock patients have been described as markedly elevated at baseline in comparison with septic patients without shock [10,12,24]. In addition, sELAM-1 and sICAM-1 have been shown to be markedly elevated at baseline in non-survivors in comparison with survivors, as shown in the present study [2,8,10-12,24].
In the present study, non-survivors (in comparison with survivors) showed elevated adhesion molecule levels over the whole study period. After a comparable time course at different levels over the first 48 hours, the endothelial mediator levels increased in non-survivors and decreased in survivors. sELAM-1 was predictive for outcome at the third and fourth days. Kayal and colleagues investigated patients with severe sepsis (56%) and with septic shock (44%) on admission to the ICU or during ICU hospitalisation. Seventy-two percent of the septic shock patients had a putative sepsis onset >6 hours before inclusion in the study, and 82% of the septic shock patients died after a median time of 3 days [10]. Fifty percent of the severe sepsis patients had a putative sepsis onset >6 hours before inclusion into the study, 14% of which died after 6 days in the ICU [10]. Kayal and colleagues observed an increase in sICAM-1 and sELAM-1 levels for 3–4 days after study inclusion in non-survivors, sELAM-1 then returning to levels similar to those observed in survivors whereas sICAM-1 continued to increase in non-survivors [10]. Those authors concluded that baseline sICAM-1 and sELAM-1, as markers of endothelial cell activation, predicted disease severity – and sICAM-1 more then sELAM-1 reflected the intensity of inflammation and tissue damage in late sepsis [10].
Boldt and colleagues investigated septic patients already admitted to the ICU at the onset of sepsis, 40% of which died [11]. The authors also demonstrated that sELAM-1 decreased over time in septic patients while sICAM-1 increased further [11]. Cowley and colleagues investigated adhesion molecule levels of patients admitted to the ward or the ICU within 12 hours after the onset of systemic inflammatory response syndrome, with or without signs of organ dysfunction or hypoperfusion – 60% of them died [18]. This study group observed increased levels of sELAM-1 over the study period in patients with sustained organ dysfunction and in non-survivors, whereas sELAM-1 levels decreased in patients whose organ dysfunction resolved [18]. Sessler and colleagues measured sICAM-1 levels of septic patients (64% in septic shock, from which 75% died) within 12 hours after admission to the ICU for sepsis, of which 48% died [12]. The authors were able to show that baseline sICAM-1 levels correlate independently with outcome [12]. Cummings and colleagues investigated sELAM-1 levels within 24 hours of admission to the ICU of 119 critically ill patients (7% had no systemic inflammatory response syndrome, 37% had non-infectious systemic inflammatory response syndrome, 56% were septic, 34% were in shock) [24]. The authors found a modest correlation between day 1 sELAM-1 levels and organ dysfunction as well as survival [24].
The inclusion time of patients into the study could be crucial for the course and interpretation of mediator levels in relation to outcome [17]. If admission and therapy is delayed, mediator levels might already be high at admission [17]. The clinical signs of septic shock become evident when the inflammatory insult is already ongoing and initialising therapy might be delayed, leading to a worse outcome [16]. The early goal-directed therapy performed by Rivers and colleagues in septic shock patients provided a significant outcome benefit [16]. Our patients, who were already under standardised ICU therapy before septic shock began, died 7 days (median) after possible outcome prediction by enhanced endothelial damage markers in non-survivors. The monitoring of sELAM-1 and sICAM-1 over the time course of septic shock could probably indicate when the patients' course is leading to lethal outcome and could help physicians to intervene and monitor further therapy before the patients die. Such therapies aim at recruiting the endothelium; for example, the application of activated protein C.
IL-6 has been described to have pro-inflammatory and anti-inflammatory properties in different animal and human septic and non-septic models [2,15,25,26]. IL-6 is widely accepted as a marker for disease severity in septic shock but elevations are not sepsis specific [13,15,27-29]. However, as has been demonstrated for adhesion molecules, IL-6 levels in septic shock patients were significantly higher and stayed higher in non-survivors than in survivors, as shown in the present study [13-15,17,27,28]. The predictive value of IL-6 on admission has been described for septic patients and septic shock patients [14,15,19]. Baseline values in our study were not predictive for outcome, perhaps because of the early entry time into the study as described earlier.
IL-6 tended to correlate with outcome on the third and fourth days after onset of septic shock. Pinsky and colleagues described the persistence of high levels of IL-6, and not the peaks of IL-6, as being predictive for outcome [17]. IL-6 continuously dropped in survivors whereas it showed a variable course in non-survivors. This variability has been described in patients suffering from sepsis and from septic shock [13,15]. In both groups, however, IL-6 levels decreased significantly from admission until the end of this study, in contrast to other cytokines such as tumor necrosis factor alpha or to other adhesion molecules, as shown in other studies and our own [14,15]. Presterl and colleagues observed a steady decrease in IL-6 over a 7-day period in survivors and observed persistent high levels in non-survivors [13]. This course could be related to an initial pro-inflammatory characteristic and a later anti-inflammatory characteristic of IL-6 when compared with the explicit pro-inflammatory cytokine tumor necrosis factor alpha [14,15,25,26].
IL-8 was significantly higher in non-survivors than in survivors, and it was predictive for lethal outcome at baseline. IL-8, a chemoattractant, is an early pro-inflammatory component released in sepsis by endothelial cells and other cells [7]. High levels of IL-8 have been described in sepsis, in shock and in MOF with poor outcome, consistent with our study [29-31]. These results, however, are conflicting in the literature [29,31,32]. Especially for early detection of nosocomial pneumonias and newborn infections, IL-8 has been shown to be an adequate marker and predictor [33-36]. The predictive value of this parameter at baseline, as shown in the present study, might be a hint that patients were in the phase of early septic progression.
The rate of pneumonias and peritonitis as the septic focus was significantly different between survivors and non-survivors. After revision of the literature, no data could be found regarding possible differences in expression of endothelial damage markers and outcome looking at different infection sites.
All our patients required norepinephrine therapy. Significantly more non-survivors needed norepinephrine at a dose >0.5 μg/kg/min than survivors, probably due to profound volume-refractory vasodilation. Norepinephrine follows dopamine as the first-choice vasopressor in septic shock and has been applied in dosages as high as 5 μg/kg/min [4,37]. The use of positive inotropic therapy to achieve supramaximal hemodynamic values for oxygen delivery, for mixed venous oxygen saturation and for cardiac index has been reported to worsen the outcome of patients in septic shock [38-40]. In the present study the use of positive inotropic therapy did not differ between survivors and non-survivors. Although myocardial dysfunction has been extensively described in sepsis, the main pathophysiology developing in septic shock is the peripheral vasodilation with consecutive hypotension [4,37,41,42]. As myocardial dysfunction/ischemia may be contributing factors influencing study results and the outcome, patients with an acute history of severe cardiac insufficiency and coronary artery disease before the development of septic shock were excluded from the study. The laboratory parameters for myocardial ischemia and the ECGs performed did not show differences in signs of myocardial ischemia between survivors and non-survivors. The high incidence of myocardial ischemic signs observed in the ECGs has to be interpreted carefully. Other studies have described the low specificity of ECG in comparison with troponin T for the diagnosis of myocardial ischemia [8].
Patients in the non-survivor group in this study were significantly older than the survivors. Age was also a significant predictor of lethal outcome in the AUC analysis. The patients' age has been described as a risk factor of fatal outcome in patients with sepsis, explained by a possibly diminished physiologic reserve and a poor immune status [1,5,19,43]. Boldt and colleagues were able to show higher levels of sELAM-1 and sICAM-1 in patients older than 70 years in comparison with patients younger than 50 years, indicating an association with more extensive endothelial damage [43]. In the present study, sELAM-1 was significantly higher in patients older than 65 years (P = 0.01). When excluding non-survivors, however, sELAM-1 was no longer significantly higher in patients older than 65 years (P = 0.60).
A major limitation of the present study is the low number of patients. This fact could be the cause for the large range in standard deviation of the markers measured. A far greater number of patients will be needed to verify the results presented.
Conclusion
The endothelial marker sELAM-I showed a markedly opposing and predictive course after 48 hours of septic shock. Our data suggest that the adhesion molecule sELAM-1 might be useful in assessing disease severity in the course of septic shock after early initiation of treatment. This might provide a valuable means of monitoring and a means of guidance of therapy with substances known to reduce endothelial damage (such as, for example, activated protein C).
Key messages
• sELAM-1 showed early prediction of outcome in septic shock patients
Abbreviations
AUC = area under the receiver operating characteristics curve; ECG = electrocardiogram; ICU = intensive care unit; IL = interleukin; MOF = multiple organ failure; sELAM-1 = soluble endothelial-linked adhesion molecule 1; sICAM-1 = soluble intercellular adhesion molecule 1.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
OVH and CS completed the proposal writing and experimental design. OVH, J-PT and KM participated in the research coordination, data analysis, presentation and conduction of all experimental aspects of the study. OVH, VvD, MK and CS prepared the manuscript.
Figures and Tables
Figure 1 IL-6 for survivors and non-survivors over time.
Figure 2 IL-8 for survivors and non-survivors over time.
Figure 3 Soluble endothelial-linked adhesion molecule 1 (sELAM-1) for survivors and non-survivors over time. * Significant difference (P < 0.05) for sELAM-1 between survivors and non-survivors.
Figure 4 Soluble intercellular adhesion molecule 1 (sICAM-1) for survivors and non-survivors over time.
Table 1 Baseline and outcome data
Survivors (n = 16, 40%) Non-survivors (n = 24, 60%) P*
Age (years) 59 (28–82) 65 (33–86) 0.03
Sex (male/female) 11/5 11/13 0.15
Intensive care unit stay (days) 27 (11–48) 8 (2–57) <0.01
Sepsis focus (n)
Pneumonia 9 4 0.02
Peritonitis 2 13 0.04
Wound infection 3 4 >0.99
Abscess 2 3 >0.99
Hemoglobin (g/dl) 10 (7–14) 11 (7–13) 0.61
Oxygenation index (mmHg) 245 (114–421) 199 (89–384) 0.44
APACHE III baseline score 55 (23–88) 61 (12–100) 0.91
APACHE III maximum score 75 (52–108) 86 (52–117) 0.47
MODS baseline 6 (2–11) 7 (2–12) 0.36
MODS max 9 (5–15) 9 (4–14) 0.73
Data presented as median (range). APACHE, Acute Physiology and Chronic Health Evaluation; MODS, multiple organ dysfunction syndrome.
*P value for intergroup baseline and outcome data: Mann–Whitney U test, and Pearson chi-square and Fisher exact tests, respectively.
Table 2 Clinical and laboratory data
Survivors (n = 16, 40%) Non-survivors (n = 24, 60%) P*
Norepinephrine (n) 16 (100%) 24 (100%)
Norepinephrine >0.5 μg/kg/min mean values (n) 8 22 <0.01
Number of + inotropic medications (dobutamine or dopamine, enoximone and epinephrine) (n) 0.79
0 3 6
1 9 11
2 4 6
3 0 1
Myocardial ischemia signs in electrocardiogram (n) 12 15 0.41
Troponin T >0.2 (ng/ml) 5 7 0.89
CK/CK-MB fraction >6% 1 0 0.22
Data presented as median (range). CK/CK-MB, creatin kinase/creatin kinase-myoglobin band.
*P value for intergroup data analysis: Pearson chi-square and Fisher exact tests.
Table 3 Predictive parameters determined by the area under the receiver operating characteristics curve (AUC)
Time point AUC 95% confidence interval P
Baseline
IL-8 0.777 0.619–0.935 0.004
IL-6 0.648 0.462–0.834 0.14
sELAM-1 0.600 0.400–0.800 0.30
sICAM-1 0.548 0.360–0.735 0.622
Third day
IL-8 0.133 0.402–0.923 0.25
IL-6 0.727 0.501–0.953 0.087
sELAM-1 0.808 0.599–1.017 0.02
sICAM-1 0.677 0.427–0.927 0.18
Fourth day
IL-8 0.775 0.529–1.021 0.05
IL-6 0.737 0.504–0.971 0.09
sELAM-1 0.847 0.631–1.064 0.02
sICAM-1 0.694 0.433–0.956 0.18
sELAM-1, soluble endothelial-linked adhesion molecule 1; sICAM-1, soluble intercellular adhesion molecule 1.
==== Refs
Peters K Unger RE Brunner J Kirkpatrick CJ Molecular basis of endothelial dysfunction in sepsis Cardiovasc Res 2003 60 49 57 14522406 10.1016/S0008-6363(03)00397-3
Reinhart K Bayer O Brunkhorst F Meisner M Markers of endothelial damage in organ dysfunction and sepsis Crit Care Med 2002 30 S302 S312 12004252 10.1097/00003246-200205001-00021
Endo S Inada K Kasai T Takakuwa T Yamada Y Koike S Wakabayashi G Niimi M Taniguchi S Yoshida M Levels of soluble adhesion molecules and cytokines in patients with septic multiple organ failure J Inflamm 1995 46 212 219 8878795
Ruokonen E Parviainen I Uusaro A Treatment of impaired perfusion in septic shock Ann Med 2002 34 590 597 12553499 10.1080/078538902321117814
Gogos CA Lekkou A Papageorgiou O Siagris D Skoutelis A Bassaris HP Clinical prognostic markers in patients with severe sepsis: a prospective analysis of 139 consecutive cases J Infect 2003 47 300 306 14556754 10.1016/S0163-4453(03)00101-4
Ridings PC Windsor AC Jutila MA Blocher CR Fisher BJ Sholley MM Sugerman HJ Fowler AA III A dual-binding antibody to E- and L-selectin attenuates sepsis-induced lung injury Am J Respir Crit Care Med 1995 152 247 253 7541277
Carlos TM Harlan JM Leukocyte–endothelial adhesion molecules Blood 1994 84 2068 2101 7522621
Spies C Haude V Fitzner R Schroder K Overbeck M Runkel N Schaffartzik W Serum cardiac troponin T as a prognostic marker in early sepsis Chest 1998 113 1055 1063 9554647
Moss M Gillespie MK Ackerson L Moore FA Moore EE Parsons PE Endothelial cell activity varies in patients at risk for the adult respiratory distress syndrome Crit Care Med 1996 24 1782 1786 8917025 10.1097/00003246-199611000-00004
Kayal S Jais JP Aguini N Chaudiere J Labrousse J Elevated circulating E-selectin, intercellular adhesion molecule 1, and von Willebrand factor in patients with severe infection Am J Respir Crit Care Med 1998 157 776 784 9517590
Boldt J Muller M Kuhn D Linke LC Hempelmann G Circulating adhesion molecules in the critically ill: a comparison between trauma and sepsis patients Intensive Care Med 1996 22 122 128 8857119
Sessler CN Windsor AC Schwartz M Watson L Fisher BJ Sugerman HJ Fowler AA III Circulating ICAM-1 is increased in septic shock Am J Respir Crit Care Med 1995 151 1420 1427 7735595
Presterl E Staudinger T Pettermann M Lassnigg A Burgmann H Winkler S Frass M Graninger W Cytokine profile and correlation to the APACHE III and MPM II scores in patients with sepsis Am J Respir Crit Care Med 1997 156 825 832 9310000
Calandra T Gerain J Heumann D Baumgartner JD Glauser MP High circulating levels of interleukin-6 in patients with septic shock: evolution during sepsis, prognostic value, and interplay with other cytokines. The Swiss-Dutch J5 Immunoglobulin Study Group Am J Med 1991 91 23 29 1907101 10.1016/0002-9343(91)90069-A
Martin C Boisson C Haccoun M Thomachot L Mege JL Patterns of cytokine evolution (tumor necrosis factor-alpha and interleukin-6) after septic shock, hemorrhagic shock, and severe trauma Crit Care Med 1997 25 1813 1819 9366763 10.1097/00003246-199711000-00018
Rivers E Nguyen B Havstad S Ressler J Muzzin A Knoblich B Peterson E Tomlanovich M Early goal-directed therapy in the treatment of severe sepsis and septic shock N Engl J Med 2001 345 1368 1377 11794169 10.1056/NEJMoa010307
Pinsky MR Vincent JL Deviere J Alegre M Kahn RJ Dupont E Serum cytokine levels in human septic shock. Relation to multiple-system organ failure and mortality Chest 1993 103 565 575 8432155
Cowley HC Heney D Gearing AJ Hemingway I Webster NR Increased circulating adhesion molecule concentrations in patients with the systemic inflammatory response syndrome: a prospective cohort study Crit Care Med 1994 22 651 657 7511496
Calandra T Baumgartner JD Grau GE Wu MM Lambert PH Schellekens J Verhoef J Glauser MP Prognostic values of tumor necrosis factor/cachectin, interleukin-1, interferon-alpha, and interferon-gamma in the serum of patients with septic shock. Swiss-Dutch J5 Immunoglobulin Study Group J Infect Dis 1990 161 982 987 2109023
Bone RC Balk RA Cerra FB Dellinger RP Fein AM Knaus WA Schein RM Sibbald WJ Definitions for sepsis and organ failure and guidelines for the use of innovative therapies in sepsis. The ACCP/SCCM Consensus Conference Committee. American College of Chest Physicians/Society of Critical Care Medicine Chest 1992 101 1644 1655 1303622
Bennett JA Riegel B Bittner V Nichols J Validity and reliability of the NYHA classes for measuring research outcomes in patients with cardiac disease Heart Lung 2002 31 262 270 12122390 10.1067/mhl.2002.124554
Spies CD Kern H Schroder T Sander M Sepold H Lang P Stangl K Behrens S Sinha P Schaffartzik W Myocardial ischemia and cytokine response are associated with subsequent onset of infections after noncardiac surgery Anesth Analg 2002 95 9 18 table 12088935 10.1097/00000539-200207000-00002
Kox WJ Spies C Ckeck-up Anästhesiologie Standards -Anästhesie, -Intensivmedizin, -Schmerztherapie, -Notfallmedizin 2003 Berlin: Springer Verlag
Cummings CJ Sessler CN Beall LD Fisher BJ Best AM Fowler AA III Soluble E-selectin levels in sepsis and critical illness. Correlation with infection and hemodynamic dysfunction Am J Respir Crit Care Med 1997 156 431 437 9279220
Tilg H Dinarello CA Mier JW IL-6 and APPs: anti-inflammatory and immunosuppressive mediators Immunol Today 1997 18 428 432 9293158 10.1016/S0167-5699(97)01103-1
Tilg H Trehu E Atkins MB Dinarello CA Mier JW Interleukin-6 (IL-6) as an anti-inflammatory cytokine: induction of circulating IL-1 receptor antagonist and soluble tumor necrosis factor receptor p55 Blood 1994 83 113 118 8274730
Geppert A Steiner A Zorn G Delle-Karth G Koreny M Haumer M Siostrzonek P Huber K Heinz G Multiple organ failure in patients with cardiogenic shock is associated with high plasma levels of interleukin-6 Crit Care Med 2002 30 1987 1994 12352031 10.1097/00003246-200209000-00007
Damas P Ledoux D Nys M Vrindts Y De Groote D Franchimont P Lamy M Cytokine serum level during severe sepsis in human IL-6 as a marker of severity Ann Surg 1992 215 356 362 1558416
Cavaillon JM Adib-Conquy M Fitting C Adrie C Payen D Cytokine cascade in sepsis Scand J Infect Dis 2003 35 535 544 14620132 10.1080/00365540310015935
Marty C Misset B Tamion F Fitting C Carlet J Cavaillon JM Circulating interleukin-8 concentrations in patients with multiple organ failure of septic and nonseptic origin Crit Care Med 1994 22 673 679 8143477
Hack CE Hart M van Schijndel RJ Eerenberg AJ Nuijens JH Thijs LG Aarden LA Interleukin-8 in sepsis: relation to shock and inflammatory mediators Infect Immun 1992 60 2835 2842 1612748
Endo S Inada K Ceska M Takakuwa T Yamada Y Nakae H Kasai T Yamashita H Taki K Yoshida M Plasma interleukin 8 and polymorphonuclear leukocyte elastase concentrations in patients with septic shock J Inflamm 1995 45 136 142 7583359
Muehlstedt SG Richardson CJ West MA Lyte M Rodriguez JL Cytokines and the pathogenesis of nosocomial pneumonia Surgery 2001 130 602 609 11602890 10.1067/msy.2001.117105
Franz AR Steinbach G Kron M Pohlandt F Interleukin-8: a valuable tool to restrict antibiotic therapy in newborn infants Acta Paediatr 2001 90 1025 1032 11683191 10.1080/080352501316978110
Dembinski J Behrendt D Heep A Dorn C Reinsberg J Bartmann P Cell-associated interleukin-8 in cord blood of term and preterm infants Clin Diagn Lab Immunol 2002 9 320 323 11874870 10.1128/CDLI.9.2.320-323.2002
Lin KJ Lin J Hanasawa K Tani T Kodama M Interleukin-8 as a predictor of the severity of bacteremia and infectious disease Shock 2000 14 95 100 10947149
Martin C Viviand X Leone M Thirion X Effect of norepinephrine on the outcome of septic shock Crit Care Med 2000 28 2758 2765 10966247 10.1097/00003246-200008000-00012
Hayes MA Timmins AC Yau EH Palazzo M Watson D Hinds CJ Oxygen transport patterns in patients with sepsis syndrome or septic shock: influence of treatment and relationship to outcome Crit Care Med 1997 25 926 936 9201043 10.1097/00003246-199706000-00007
Hinds C Watson D Manipulating hemodynamics and oxygen transport in critically ill patients N Engl J Med 1995 333 1074 1075 7675052 10.1056/NEJM199510193331609
Gattinoni L Brazzi L Pelosi P Latini R Tognoni G Pesenti A Fumagalli R A trial of goal-oriented hemodynamic therapy in critically ill patients. SvO2 Collaborative Group N Engl J Med 1995 333 1025 1032 7675044 10.1056/NEJM199510193331601
Goncalves JA JrHydo LJ Barie PS Factors influencing outcome of prolonged norepinephrine therapy for shock in critical surgical illness Shock 1998 10 231 236 9788653
Krishnagopalan S Kumar A Parrillo JE Kumar A Myocardial dysfunction in the patient with sepsis Curr Opin Crit Care 2002 8 376 388 12357104 10.1097/00075198-200210000-00003
Boldt J Muller M Heesen M Papsdorf M Hempelmann G Does age influence circulating adhesion molecules in the critically ill? Crit Care Med 1997 25 95 100 8989183 10.1097/00003246-199701000-00019
|
16137344
|
PMC1269435
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 13; 9(4):R323-R330
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3532
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35361613734210.1186/cc3536ResearchFactors that predict outcome of intensive care treatment in very elderly patients: a review de Rooij Sophia E [email protected] Ameen [email protected] Marcel [email protected] Jonge Evert [email protected] Head, Department of Geriatrics, Academic Medical Center, University of Amsterdam, Amsterdam2 Adjunct Head, Department of Medical Informatics, Academic Medical Center, University of Amsterdam, Amsterdam3 Professor and Head, Department of Internal Medicine, Cardiology and Pulmonary Disease, Academic Medical Center, University of Amsterdam, Amsterdam4 Adjunct Head Department of Intensive Care, Academic Medical Center, University of Amsterdam, Amsterdam2005 17 5 2005 9 4 R307 R314 13 1 2005 11 3 2005 6 4 2005 8 4 2005 Copyright © 2005 de Rooij et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Advanced age is thought to be associated with increased mortality in critically ill patients. This report reviews available data on factors that determine outcome, on the value of prognostic models, and on preferences regarding life-sustaining treatments in (very) elderly intensive care unit (ICU) patients.
Methods
We searched the Medline database (January 1966 to January 2005) for English language articles. Selected articles were cross-checked for other relevant publications.
Results
Mortality rates are higher in elderly ICU patients than in younger patients. However, it is not age per se but associated factors, such as severity of illness and premorbid functional status, that appear to be responsible for the poorer prognosis. Patients' preferences regarding life-sustaining treatments are importantly influenced by the likelihood of a beneficial outcome. Commonly used prognostic models have not been calibrated for use in the very elderly. Furthermore, they do not address long-term survival and functional outcome.
Conclusion
We advocate the development of new prognostic models, validated in elderly ICU patients, that predict not only survival but also functional and cognitive status after discharge. Such a model may support informed decision making with respect to patients' preferences.
See related commentary
==== Body
Introduction
Projections by the US Census Bureau [1] suggest that the population aged 85 years and older is likely to grow from about 4 million in 2000 to 19 million by 2050. This 'greying' of the population has also been identified in European countries and in Japan. Ageing of the population increases the proportion of people with chronic conditions, with corresponding expectations of eventual decline in function. Advanced age is associated with increased mortality in intensive care unit (ICU) patients [2]. Furthermore, the life expectancy of all elderly patients, remains limited, even after successful ICU treatment. In the UK life expectancy at age 80 years increased from 5.8 years in 1981 to 7.2 years in 2002 for males, and from 7.5 to 8.7 years for females [3]. Thus, the costs per year of life gained, both economical and emotional, are relatively high for elderly patients. Indeed, life-sustaining treatment is more often withdrawn or withheld in older patients. However, few data are available to help identify patients who will benefit from ICU treatment from those who will not.
In this review we focus on the most important factors that may influence outcomes in very elderly critically ill patients, on models that predict short-term and long-term outcome, and on the available data on patients' preferences regarding life-sustaining treatment and how these preferences are influenced by the likelihood of a beneficial outcome.
Materials and methods
A Medline search (January 1966 to January 2005) was performed using the terms 'frail elderly', 'geriatric', 'very elderly' and 'octogenarians'; and 'critical illness', 'critical care', 'intensive care' and 'intensive care units'; in combination with the terms 'prognosis', 'predictor', or 'outcome'.
Based on title and abstract, we selected English language articles containing clinical data on the outcomes of ICU treatment in very elderly patients. The reference lists of all reports were cross-checked for other potentially relevant articles. In the reports identified in this search, we examined factors that influence outcome in elderly patients such as age, diagnosis, comorbidity, functional status (including cognitive functioning) before hospital admittance, delirium, malnutrition, dehydration, acute renal failure, length of stay, and complications such as nosocomial infections and pressure ulcers. It was envisaged that the studies would be too heterogeneous to combine in a formal meta-analysis, and therefore a narrative synthesis, mainly focusing on prospective studies or very large retrospective studies, was undertaken.
In accordance with published criteria [4], we consider patients aged 80 years and older to be 'very elderly'. However, as several published studies used different criteria for defining a patient as elderly, we also consider data based on studies in other patient groups (e.g. those older than 70 years). Where data specific to elderly patients are not available, we briefly review best knowledge based on studies in patients of all ages.
Results and discussion
Factors influencing outcome in elderly patients
Age
When discussing the influence of age on ICU outcome, it is important to appreciate that all published studies, either prospective or retrospective, were performed in selected populations of elderly patients after admission to an ICU. Because intensive treatments, including intensive care, are often withheld in elderly patients [4,5], patients with severe comorbidity may be under-represented in these studies. This could result in an over-optimistic view on the effects of age on ICU outcome in the selected patient groups. On the other hand, high mortality rates in the studies may partly be accounted for by decisions to withhold life-sustaining treatments because of advanced age.
For this overview, we consider those patients aged 80 years or older to be 'very elderly' patients, in accordance with the definitions proposed by the SUPPORT (Study to Understand Prognosis and Preferences for Outcomes and Risks of Treatment) investigators [4].
We found 12 prospective cohort studies or retrospective studies based on large databases that addressed the influence of age on outcome in ICU patients (Table 1) [6-17]. In 1995 Cohen and Lambrinos [8] presented the results of a study of the impact age has on outcome of mechanical ventilation in a 41,848-patient, state-wide database. They found that in-hospital mortality in patients receiving mechanical ventilation aged 85 years or older was 70%, as compared with 32% in patients aged 29 years or younger. Only 14% of patients aged 85 years or older went home without home health care, as compared with 47% in patients aged 29 years or younger. Another large retrospective cohort study [9], conducted in data from consecutive ICU admissions to 38 ICUs, showed increased risk for hospital death with more advanced age. Relative to patients younger than 35 years, the adjusted odds of death in patients aged 80–84 years and ≥90 years were 3.9 and 4.7, respectively. These findings were adjusted for severity of illness, Acute Physiology Score, admission source, diagnosis and comorbidity. These conclusions are in accordance with the findings of the SUPPORT study [4]. In that study the risk of death was shown to increase by 1.0% for year of age in patients aged 18–70 years, and by 2.0% for patients aged 70 years or older.
Figure 1 shows the effect of age on in-hospital mortality in 54,021 patients admitted to various ICUs participating in the Dutch National Intensive Care Evaluation (NICE) registry [18]. The in-hospital mortality rate in patients aged 85 years or older was fourfold higher than in patients younger than 65 years.
Although advanced age clearly increases the risk for not surviving an ICU stay, this does not mean that all critically ill elderly patients have a poor prognosis. Studies in specific subgroups of elderly patients have shown that mortality may be as low as 4.3% or 22.1% for patients older than 85 years admitted to a surgical ICU [19,20], 15–25% in neurosurgical ICU patients, and 39–48% for medical ICU patients [21].
Despite potential bias in all studies, many suggest that older patients are more likely to die or experience adverse outcomes of their ICU treatment. However, several studies, using multivariate analysis, showed that age was not an independent predictor of mortality [6,16,21-23]. It appears that it is not advanced age per se but other factors associated with advanced age that determine prognosis in elderly patients.
Diagnosis
The conclusion that very elderly ICU patients are at substantially increased risk for dying may not hold true for all subgroups of patients. It was found that the effects of age on prognosis very much depend on other factors such as diagnosis. In patients aged 80–84 years hospital mortality was 85% for those with infection as their reason for admission, as compared with 58% for those with diagnoses of gastrointestinal disorder [8]. In another study [24], whereas in-hospital mortality in elderly patients on mechanical ventilation due to pneumonia was 62%, it was 40% in ventilated trauma patients. Outcome after brain injury in geriatric trauma patients is notoriously poor, with mortality and functional disability rates twice those in younger patients [25]. In a general population (all ages), it was shown that 13.6% of the predictive power of the Acute Physiology and Chronic Health Evaluation (APACHE) III model was due to admitting diagnosis [26]. Our data from the Dutch NICE database [18] show that, between 1997 and 2002, in-hospital mortality in ICU patients aged 80 years or older was 16.5% in those who had undergone cardiac surgery but 46% in other patients.
We can conclude that the reason for admission to an ICU has a major influence on prognosis.
Comorbidity
Comorbidity, defined as the total burden of illness unrelated to a patient's principal diagnosis, contributes to clinical outcomes (e.g. mortality, surgical results, complication rates, functional status and length of stay) as well as to economic outcomes (e.g. resource utilization, discharge destination and intensity of treatments) [27-29]. Most information on the influence of comorbidity on outcome after ICU admission comes from studies in patients of all ages. In 1987, Charlson and coworkers [29] developed a weighted index of comorbidity that takes into account the number as well as the seriousness of comorbid diseases. This index was shown to predict the 1-year mortality of hospitalized medical patients.
Some studies investigated the relationship between comorbidity and mortality in critically ill patients of all ages. Among the severity of illness models that predict mortality in critically ill patients, comorbidity is included in APACHE II and III [30,31] but not in the Simplified Acute Physiology Score (SAPS) II [32] or Mortality Probability Model (MPM) II [33]. It was shown that the APACHE II model was a very good predictor of mortality in critically ill patients, but that the chronic health points components of APACHE II did not have discriminating ability [34]. Furthermore, it was also shown that the Charlson index had some predictive value in critically ill patients but with an area under the receiver operating characteristic (ROC) curve of only 0.67, indicating limited discriminating ability. In a retrospective cohort study conducted in more than 17,000 ICU patients [35], comorbidity was found to account for only 8.4% of the predictive ability of APACHE II, as compared with 67.7% for laboratory values and 17.7% for diagnosis [35].
Comorbidity is commonly present in elderly patients. However, we could not find any study of the possible influence of comorbidity on outcome conducted specifically in (very) elderly critically ill patients.
Functional status before hospital admittance
Functional status, including physical, cognitive and social functioning, has been shown to be an important predictor of the hospital outcomes of older patients. Not surprisingly, impaired functioning in daily life is more likely to be prevalent in older patients and was found to form an independent predictor of mortality [36-38].
Functional status is generally not assessed by physiologically based models such as SAPS II and APACHE II and III. In ICU patients of all ages, an association between functional status and mortality was found by some investigators [39] but not by others [22,40]. Few clinical studies described the value of premorbid functional status in predicting ICU outcomes in the very elderly.
In 1991, Mayer-Oakes and coworkers [41] found in older ICU patients that those who died were significantly more likely to be totally dependent on help for activities of daily living than were those who survived. It was recently reported that long-term survival after admission to a medical ICU is dependent on functional status before admission [5]. In a more recent study [16], the prognosis of elderly patients hospitalized in a medical ICU depended not only on APACHE II scores but also on the loss of functional independence and on the presence of moderate to severe cognitive impairment before ICU admission. Mortality was 30% in patients who had an Activities of Daily Living score of 1–6 (dependent), as compared with 7.8% in patients with a score of 0 (independent). Likewise, mortality was 55.9% in patients with severe cognitive impairment versus 8.2% in those without cognitive impairment. Also, in older patients with severe pneumonia requiring mechanical ventilation, the Activities of Daily Living score before admission was shown to be an important predictor of discharge outcome [42].
Another recent study [22] showed that, in a population of very old patients, mortality after ICU discharge occurred predominantly during the first 3 months.
Although various instruments for measurement of impaired functioning were employed in the reviewed studies, both age and prior limitation of activity were associated with risk for dying during the ICU stay.
In a recent prospective cohort study conducted in 817 adult patients receiving prolonged mechanical ventilation, long-term ICU outcome, defined as mortality after 1 year of follow up, was also found to be associated with advanced age and poor functional status before hospitalization [22,43].
Other factors related to intensive care outcome in very elderly patients
Risk adjustment indices, which are mainly based on demographic data, and the existing prognostic models may underestimate the effects on prognosis of complicating conditions that are frequently present in older patients and that are under-reported in administrative databases. Examples of these are malnutrition and delirium.
Low body mass index has been shown to be an independent predictor of in-hospital mortality [36,44,45]. Malnutrition was common in older hospitalized patients with medical illness, and was also associated with delayed functional recovery and higher rates of nursing home use. These adverse outcomes were not accounted for by greater severity of acute illness, comorbidity, or functional dependence in malnourished patients on hospital admission [36]. This relation between nutrition, in some studies expressed as a low body mass index, and mortality was also confirmed in ICU patients aged 65 years and older [16]. Delirium, an often overlooked complication in older ICU patients, is an independent predictor of reintubation, prolonged hospital stay and mortality [46-48].
Other factors that may have an effect on prognosis are complications, such as adverse drug events [49], nosocomial infections [50] and pressure ulcers [16]. However, no studies were found concerning the impacts of these complications on outcome specifically in very elderly critically ill patients treated in ICUs.
Patient preferences
Patients do not necessarily prefer life-extending treatment over care focused on relieving pain and discomfort. The willingness to receive life-sustaining treatment depends on the burden of treatment, the outcome and the likelihood of the outcome. In a population of patients with limited life expectancy and aged 60 years or older, 74% stated that they would not choose treatment if the burden of treatment were high and the anticipated outcome survival with severe functional impairment [51]. Under the same conditions, 88% of patients opted not to undergo treatment if cognitive impairment was the expected outcome. The number of participants who stated that they would choose treatment declined as the likelihood of an adverse outcome increased. In another study conducted in patients aged 65 years and older [52], patients' willingness to receive cardiopulmonary resuscitation if they suffered a cardiac arrest decreased from 41% to 22% after learning the probability of survival (10–17%). Only 6% of patients aged 86 years or older opted for cardiopulmonary resuscitation under these conditions. Substantial differences in the willingness to receive life-sustaining treatment exist that may depend on ethnicity, religion, the role of family and other variables [53].
Unfortunately, physicians are often unaware of the treatment preferences of their patients. In a study conducted in 4556 patients [4], physicians did not know the preference of their patient in 25% of cases. Furthermore, their assessments of patients preferences were correct for 45% and incorrect for the remaining 30% of patients. Physicians were more likely to believe incorrectly that patients did not want life-extending care when patients were older (79% of the time for patients older than 80 years, as compared with 36% for patients younger than 50 years).
Prognostic models in intensive care
Patients and their representatives base their decisions regarding what treatments they wish to undergo to a large extent on the likelihood of a favourable outcome. This underscores the importance of reliable information on what outcome can be expected. In order to help physicians to estimate the likelihood of survival of their patients, several severity-of-illness based mortality prediction models were developed for use in multidiagnostic patient groups. They were developed using logistic regression and incorporate information about physiological derangement, admitting diagnosis, age and sometimes comorbid disease. In the general ICU population, these prognostic models, such as SAPS II [32], MPM II [33] and APACHE II and III [30,31], predict the probability of survival of critically ill patients reasonably well.
The information derived from these models can be used to evaluate ICU performance and to improve medical decision making, and perhaps it can also provide patients and their relatives with better information about the ICU stay and its possible outcomes. Unfortunately, when using prognostic models for individual decision making, the risk cannot be ruled out that these models will become self-fulfilling prophecies. If treatment is withdrawn in patients with a high risk for dying, then all high-risk patients indeed will die.
A potential limitation of these models is the fact that they are exclusively based on data obtained during the first 24 hours after ICU admission and that they do not take into account complications that may develop during treatment. It has been shown that the accuracy of prognostic models based on data from the first 24 hours after ICU admission is maintained at an acceptable level only in patients who stay in the ICU for a short period of time [54]. After this period has elapsed discriminative power decreases, probably resulting from excess risk for death associated with acquired infections or other iatrogenic complications during the ICU stay. Different models have been developed that use scores calculated on a daily basis in a general ICU patient population, showing good discriminating value [55,56]. Other potential limitations of prognostic models include the influence of organizational factors on patient outcomes [57,58], between country differences in performance of models [59] and mistakes in data collection [60].
The commonly used prognostic models have not been calibrated for use in the elderly. In a prospective cohort study conducted in patients on mechanical ventilation for pneumonia [61], the predictive values for mortality of the APACHE II, SAPS II and MPM II models were found to be significantly lower for patients aged 75 years or older as compared with younger patients.
Using the technique of recursive partitioning, El Solh and coworkers [42] developed a classification tree to predict hospital mortality in elderly ICU patients with pneumonia. This model exhibited good accuracy, with an area under the ROC of 0.93 versus 0.71 for the APACHE II model. However, that study is limited by the limited number of studied patients (n = 104) and the lack of a different population in which to validate the model.
Another model specifically developed to predict mortality and functional outcome in very elderly ICU patients used demographic and physiologic data as well as attributes of ICU treatment and ICU illnesses, such as the use of mechanical ventilation and the development of sepsis [21]. Although the model was developed in a relatively small number of patients (n = 243), it exhibited good discriminating performance for short-term outcome (predicting death and discharge to home or to a nursing facility).
Conclusion
The ICU population is ageing, and it may be concluded that very elderly patients admitted to ICUs represent a distinct and important subgroup of patients. In general, very elderly patients have poorer outcomes than do younger patients, but prognosis is more dependent on severity of illness and functional status before admission than on high age itself. A number of prognostic models have been developed that predict survival in critically ill patients, but these models are not calibrated for use in very old patients. Furthermore, they do not take into account some known risk factors, such as comorbid conditions, and functional and cognitive status before ICU admission. Finally, they do not give a prognosis regarding (long-term) functional status after hospital discharge. We suggest that a model should be developed for predicting outcome of ICU treatment in very old patients, taking into account all discussed prognostic factors. Such a model could more precisely predict the (long-term) discharge outcome of these patients and support informed decision making, in accordance with the preferences of the patients and their relatives.
Key messages
• ICU mortality is higher in elderly patients.
• High age alone is not responsible for the poorer outcome, but premorbid functional status and severity of illness also contribute.
• Present prognostic models are not suited for elderly individuals
• All (premorbid) prognostic factors should be taken into account in a prognostic model to support informed decision making.
Abbreviations
APACHE = Acute Physiology and Chronic Health Evaluation; ICU = intensive care unit; MPM = Mortality Probability Model; ROC = receiver operating characteristic; SAPS = Simplified Acute Physiology Score.
Competing interests
The author(s) declare that they have no competing interests
Authors' contributions
EdJ acquired and interpreted data, and participated in preparing the manuscript. SEdR interpreted data and participated in preparing the manuscript. AA-H analyzed and interpreted data. ML interpreted data. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 In-hospital mortality by age group in the Dutch National Intensive Care Evaluation database (n = 54021) [18]. Numbers indicate patients per age group.
Table 1 Studies concerning intensive care outcome and age
Study Sample Study type Age Main findings
Chelluri et al. (1993) [10] 97 ICU patients Prospective chart investigation 65–74 years (n = 43) and ≥75 years (n = 54) Age itself was not an adequate predictor of long-term survival and quality of life, but severity of illness was
Dardaine et al. (1995) [7] 110 ICU patients on mechanical ventilation Prospective cohort study > 70 years ICU mortality was 31% and 6-month mortality was 52%; outcome predictors were shock on admission and previous health status
Cohen and Lambrinos (1995) [8] 14,848 ICU patients on mechanical ventilation Retrospective cohort study >18 years In-hospital mortality, in patients receiving mechanical ventilation aged ≥85 years, was 70% versus 32% in patients aged ≤29 years
Dewar et al. (1999) [9] 37,573 patients on prolonged mechanical ventilation Retrospective database analysis > 18 years Inverse relation between age and survival; older survivors were often discharged to residential health care facilities
Ely et al. (1999) [12] 300 ICU patients Prospective cohort study <75 years versus >75 years No difference in duration of artificial ventilation
Montuclard et al. (2000) [13] 75 ICU patients Prospective cohort study > 70 years ICU mortality was 60% in elderly patients receiving ICU treatment
Ely et al. (2002) [14] 902 Patients with acute lung injury or ARDS Prospective cohort study <70 years (n = 729) and >70 years (n = 173) Patients aged 70 years and older were twice as likely to die than were younger patients, and had greater difficulty achieving liberation from the ventilator
Rosenthal et al. (2002) [15] 156,136 Consecutive admissions to medical, surgical, neurological, and mixed medical/surgical ICUs Retrospective cohort study 18–100 years The adjusted odds of death increased with each 5-year age increment
Djaiani and Ridley (1997) [17] 474 ICU patients Prospective cohort study >70 years The 1-year survival of patients aged <85 years was 56%, which was significantly better than that of patients aged >85 years (27%)
Bo et al. (2003) [16] 659 Medical ICU patients Prospective cohort study ≥ 65 years Independent predictors of mortality were functional dependence and cognitive impairment before admission, high APACHE II score and low body mass index
Tang et al. (2003) [11] 365 ICU patients on mechanical ventilation Prospective cohort study ≥ 65 years (n = 206) and <65 years (n = 159) Severity of acute illness and chronic co-morbidities, but not age, were predictors of medical ICU and hospital mortality in elderly ventilated patients
Chelluri et al. (2004) [43] 817 ICU patients on mechanical ventilation Prospective cohort study Mean age 65 years Long-term mortality rate was associated with old age and poor pre-hospitalization functional status
Esteban et al. (2004) [62] 5183 ICU patients on mechanical ventilation International prospective cohort study >70 years (n = 1612) Patients older than 70 years had higher in-hospital mortality (55%) but similar duration of mechanical ventilation and length of stay
Boumendil et al. (2004) [5] 233 ICU patients aged 80 years and older Prospective cohort study >80 years Long-term survival after ICU stay was mainly related to the underlying condition and preadmission functional status
Vosylius et al. (2004) [63] 2067 ICU patients Prospective observational cohort study >75 years (n = 477) Mortality in elderly patients was higher than in younger patients; most important risk factors were severity of illness, impaired level of conciousness and infection.
APACHE, Acute Physiology and Chronic Health Evaluation; ARDS, acute resppiratory distress syndrome; ICU, intensive care unit.
==== Refs
US Census Bureau Population Projections of the United States by Age, Sex, Race, Hispanic Origin and Nativity: 1999–2100 2000 Washington: US Census Bureau
Wood KA Ely EW What does it mean to be critically ill and elderly? Curr Opin Crit Care 2003 9 316 320 12883288 10.1097/00075198-200308000-00011
Gastrell J Annual update: mortality statistics 2001: general Health Stat Q 2004 21 67 69 15615153
Hamel MB Teno JM Goldman L Lynn J Davis RB Galanos AN Desbiens N Connors AF JrWenger N Phillips RS Patient age and decisions to withhold life-sustaining treatments from seriously ill, hospitalized adults. SUPPORT Investigators. Study to Understand Prognoses and Preferences for Outcomes and Risks of Treatment Ann Intern Med 1999 130 116 125 10068357
Boumendil A Maury E Reinhard I Luquel L Offenstadt G Guidet B Prognosis of patients aged 80 years and over admitted in medical intensive care unit Intensive Care Med 2004 30 647 654 14985964 10.1007/s00134-003-2150-z
Chelluri L Pinsky MR Grenvik AN Outcome of intensive care of the "oldest-old"; critically ill patients Crit Care Med 1992 20 757 761 1597028
Dardaine V Constans T Lasfargues G Perrotin D Ginies G Outcome of elderly patients requiring ventilatory support in intensive care Aging 1995 7 221 227 8541375
Cohen IL Lambrinos J Investigating the impact of age on outcome of mechanical ventilation using a population of 41,848 patients from a statewide database Chest 1995 107 1673 1680 7781366
Dewar DM Kurek CJ Lambrinos J Cohen IL Zhong Y Patterns in costs and outcomes for patients with prolonged mechanical ventilation undergoing tracheostomy: an analysis of discharges under diagnosis-related group 483 in New York State from 1992 to 1996 Crit Care Med 1999 27 2640 2647 10628603 10.1097/00003246-199912000-00006
Chelluri L Pinsky MR Donahoe MP Grenvik A Long-term outcome of critically ill elderly patients requiring intensive care JAMA 1993 269 3119 3123 8505814 10.1001/jama.269.24.3119
Tang EY Hsu LF Lam KN Pang WS Critically ill elderly who require mechanical ventilation: the effects of age on survival outcomes and resource utilisation in the medical intensive care unit of a general hospital Ann Acad Med Singapore 2003 32 691 696 14626803
Ely EW Evans GW Haponik EF Mechanical ventilation in a cohort of elderly patients admitted to an intensive care unit Ann Intern Med 1999 131 96 104 10419447
Montuclard L Garrouste-Org Timsit JF Misset B De Jonghe B Carlet J Outcome, functional autonomy, and quality of life of elderly patients with a long-term intensive care unit stay Crit Care Med 2000 28 3389 3395 11057791 10.1097/00003246-200010000-00002
Ely EW Wheeler AP Thompson BT Ancukiewicz M Steinberg KP Bernard GR Recovery rate and prognosis in older persons who develop acute lung injury and the acute respiratory distress syndrome Ann Intern Med 2002 136 25 36 11777361
Rosenthal GE Kaboli PJ Barnett MJ Sirio CA Age and the risk of in-hospital death: insights from a multihospital study of intensive care patients J Am Geriatr Soc 2002 50 1205 1212 12133014 10.1046/j.1532-5415.2002.50306.x
Bo M Massaia M Raspo S Bosco F Cena P Molaschi M Fabris F Predictive factors of in-hospital mortality in older patients admitted to a medical intensive care unit J Am Geriatr Soc 2003 51 529 533 12657074 10.1046/j.1532-5415.2003.51163.x
Djaiani G Ridley S Outcome of intensive care in the elderly Anaesthesia 1997 52 1130 1136 9485964 10.1111/j.1365-2044.1997.237-az0369.x
de Jonge E Bosman RJ van der Voort PH Korsten HH Scheffer GJ de Keizer NF Intensive care medicine in the Netherlands, 1997–2001. I. Patient population and treatment outcome [in Dutch] Ned Tijdschr Geneeskd 2003 147 1013 1017 12811973
Margulies DR Lekawa ME Bjerke HS Hiatt JR Shabot MM Surgical intensive care in the nonagenarian. No basis for age discrimination Arch Surg 1993 128 753 756 8317956
Van Den NN Vogelaers D Afschrift M Colardyn F Intensive care for very elderly patients: outcome and risk factors for in-hospital mortality Age Ageing 1999 28 253 256 10475859 10.1093/ageing/28.3.253
Nierman DM Schechter CB Cannon LM Meier DE Outcome prediction model for very elderly critically ill patients Crit Care Med 2001 29 1853 1859 11588439 10.1097/00003246-200110000-00001
Somme D Maillet JM Gisselbrecht M Novara A Ract C Fagon JY Critically ill old and the oldest-old patients in intensive care: short- and long-term outcomes Intensive Care Med 2003 29 2137 2143 14614546 10.1007/s00134-003-1929-2
Rockwood K Noseworthy TW Gibney RT Konopad E Shustack A Stollery D Johnston R Grace M One-year outcome of elderly and young patients admitted to intensive care units Crit Care Med 1993 21 687 691 8482089
Meinders AJ van der Hoeven JG Meinders AE The outcome of prolonged mechanical ventilation in elderly patients: are the efforts worthwhile? Age Ageing 1996 25 353 356 8921138
Jacobs DG Plaisier BR Barie PS Hammond JS Holevar MR Sinclair KE Scalea TM Wahl W EAST Practice Management Guidelines Work Group Practice management guidelines for geriatric trauma: the EAST Practice Management Guidelines Work Group J Trauma 2003 54 391 416 12579072
Knaus WA Wagner DP Zimmerman JE Draper EA Variations in mortality and length of stay in intensive care units Ann Intern Med 1993 118 753 761 8470850
Kaplan MH Feinstein AR The importance of classifying initial co-morbidity in evaluatin the outcome of diabetes mellitus J Chronic Dis 1974 27 387 404 4436428 10.1016/0021-9681(74)90017-4
Greenfield S Aronow HU Elashoff RM Watanabe D Flaws in mortality data. The hazards of ignoring comorbid disease JAMA 1988 260 2253 2255 3172404 10.1001/jama.260.15.2253
Charlson ME Pompei P Ales KL MacKenzie CR A new method of classifying prognostic comorbidity in longitudinal studies: development and validation J Chronic Dis 1987 40 373 383 3558716 10.1016/0021-9681(87)90171-8
Knaus WA Draper EA Wagner DP Zimmerman JE APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Knaus WA Wagner DP Draper EA Zimmerman JE Bergner M Bastos PG Sirio CA Murphy DJ Lotring T Damiano A The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalized adults Chest 1991 100 1619 1636 1959406
Le Gall JR Loirat P Alperovitch A Glaser P Granthil C Mathieu D Mercier P Thomas R Villers D A simplified acute physiology score for ICU patients Crit Care Med 1984 12 975 977 6499483
Lemeshow S Teres D Avrunin JS Gage RW Refining intensive care unit outcome prediction by using changing probabilities of mortality Crit Care Med 1988 16 470 477 3359785
Poses RM McClish DK Smith WR Bekes C Scott WE Prediction of survival of critically ill patients by admission comorbidity J Clin Epidemiol 1996 49 743 747 8691223 10.1016/0895-4356(96)00021-2
Johnston JA Wagner DP Timmons S Welsh D Tsevat J Render ML Impact of different measures of comorbid disease on predicted mortality of intensive care unit patients Med Care 2002 40 929 940 12395026 10.1097/00005650-200210000-00010
Covinsky KE Justice AC Rosenthal GE Palmer RM Landefeld CS Measuring prognosis and case mix in hospitalized elders. The importance of functional status J Gen Intern Med 1997 12 203 208 9127223 10.1046/j.1525-1497.1997.012004203.x
Reuben DB Rubenstein LV Hirsch SH Hays RD Value of functional status as a predictor of mortality: results of a prospective study Am J Med 1992 93 663 669 1466363 10.1016/0002-9343(92)90200-U
Inouye SK Peduzzi PN Robison JT Hughes JS Horwitz RI Concato J Importance of functional measures in predicting mortality among older hospitalized patients JAMA 1998 279 1187 1193 9555758 10.1001/jama.279.15.1187
Goldstein RL Campion EW Thibault GE Mulley AG Skinner E Functional outcomes following medical intensive care Crit Care Med 1986 14 783 788 3091318
Lemeshow S Teres D Pastides H Avrunin JS Steingrub JS A method for predicting survival and mortality of ICU patients using objectively derived weights Crit Care Med 1985 13 519 525 4006490
Mayer-Oakes SA Oye RK Leake B Predictors of mortality in older patients following medical intensive care: the importance of functional status J Am Geriatr Soc 1991 39 862 868 1885860
El Solh AA Sikka P Ramadan F Outcome of older patients with severe pneumonia predicted by recursive partitioning J Am Geriatr Soc 2001 49 1614 1621 11843993
Chelluri L Im KA Belle SH Schulz R Rotondi AJ Donahoe MP Sirio CA Mendelsohn AB Pinsky MR Long-term mortality and quality of life after prolonged mechanical ventilation Crit Care Med 2004 32 61 69 14707560 10.1097/01.CCM.0000098029.65347.F9
Landi F Onder G Gambassi G Pedone C Carbonin P Bernabei R Body mass index and mortality among hospitalized patients Arch Intern Med 2000 160 2641 2644 10999978 10.1001/archinte.160.17.2641
Galanos AN Pieper CF Kussin PS Winchell MT Fulkerson WJ Harrell FE JrTeno JM Layde P Connors AF JrPhillips RS Relationship of body mass index to subsequent mortality among seriously ill hospitalized patients. SUPPORT Investigators. The Study to Understand Prognoses and Preferences for Outcome and Risks of Treatments Crit Care Med 1997 25 1962 1968 9403743 10.1097/00003246-199712000-00010
Ely EW Optimizing outcomes for older patients treated in the intensive care unit Intensive Care Med 2003 29 2112 2115 12879233 10.1007/s00134-003-1845-5
Ely EW Stephens RK Jackson JC Thomason JW Truman B Gordon S Dittus RS Bernard GR Current opinions regarding the importance, diagnosis, and management of delirium in the intensive care unit: a survey of 912 healthcare professionals Crit Care Med 2004 32 106 112 14707567 10.1097/01.CCM.0000098033.94737.84
Ely EW Gautam S Margolin R Francis J May L Speroff T Truman B Dittus R Bernard R Inouye SK The impact of delirium in the intensive care unit on hospital length of stay Intensive Care Med 2001 27 1892 1900 11797025 10.1007/s00134-001-1132-2
Kane SL Weber RJ Dasta JF The impact of critical care pharmacists on enhancing patient outcomes Intensive Care Med 2003 29 691 698 12665997
Bochicchio GV Joshi M Knorr KM Scalea TM Impact of nosocomial infections in trauma: does age make a difference? J Trauma 2001 50 612 617 11303154
Fried TR Bradley EH Towle VR Allore H Understanding the treatment preferences of seriously ill patients N Engl J Med 2002 346 1061 1066 11932474 10.1056/NEJMsa012528
Murphy DJ Burrows D Santilli S Kemp AW Tenner S Kreling B Teno J The influence of the probability of survival on patients' preferences regarding cardiopulmonary resuscitation N Engl J Med 1994 330 545 549 8302322 10.1056/NEJM199402243300807
Clarfield AM Gordon M Markwell H Alibhai SM Ethical issues in end-of-life geriatric care: the approach of three monotheistic religions-Judaism, Catholicism, and Islam J Am Geriatr Soc 2003 51 1149 1154 12890081 10.1046/j.1532-5415.2003.51364.x
Lemeshow S Klar J Teres D Avrunin JS Gehlbach SH Rapoport J Rue M Mortality probability models for patients in the intensive care unit for 48 or 72 hours: a prospective, multicenter study Crit Care Med 1994 22 1351 1358 8062556
Timsit JF Fosse JP Troche G De Lassence A Alberti C Garrouste-Orgeas M Azoulay E Chevret S Moine P Cohen Y Accuracy of a composite score using daily SAPS II and LOD scores for predicting hospital mortality in ICU patients hospitalized for more than 72 h Intensive Care Med 2001 27 1012 1021 11497133 10.1007/s001340000840
Ferreira FL Bota DP Bross A Melot C Vincent JL Serial evaluation of the SOFA score to predict outcome in critically ill patients JAMA 2001 286 1754 1758 11594901 10.1001/jama.286.14.1754
Rosenberg AL Hofer TP Strachan C Watts CM Hayward RA Accepting critically ill transfer patients: adverse effect on a referral center's outcome and benchmark measures Ann Intern Med 2003 138 882 890 12779298
Morales IJ Peters SG Afessa B Hospital mortality rate and length of stay in patients admitted at night to the intensive care unit Crit Care Med 2003 31 858 863 12626997 10.1097/01.CCM.0000055378.31408.26
Livingston BM MacKirdy FN Howie JC Jones R Norrie JD Assessment of the performance of five intensive care scoring models within a large Scottish database Crit Care Med 2000 28 1820 1827 10890627 10.1097/00003246-200006000-00023
Polderman KH Thijs LG Girbes AR Interobserver variability in the use of APACHE II scores Lancet 1999 353 380 9950452 10.1016/S0140-6736(05)74953-9
Sikka P Jaafar WM Bozkanat E El Solh AA A comparison of severity of illness scoring systems for elderly patients with severe pneumonia Intensive Care Med 2000 26 1803 1810 11271088 10.1007/s001340000719
Esteban A Anzueto A Frutos-Vivar F Alia I Ely EW Brochard L Stewart TE Apezteguia C Tobin MJ Nightingale P Outcome of older patients receiving mechanical ventilation Intensive Care Med 2004 30 639 646 14991097 10.1007/s00134-004-2160-5
Vosylius S Sipylaite J Ivaskevicius J Determinants of outcome in elderly patients admitted to the intensive care unit Age Ageing 2005 34 157 162 15713860 10.1093/ageing/afi037
|
16137342
|
PMC1269437
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 17; 9(4):R307-R314
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3536
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35371613734110.1186/cc3537ResearchCircadian pattern of activation of the medical emergency team in a teaching hospital Jones Daryl 1Bates Samantha 2Warrillow Stephen 3Opdam Helen 3Goldsmith Donna 2Gutteridge Geoff 2Bellomo Rinaldo [email protected] Intensive Care Registrar. Department of Intensive Care and Department of Surgery (Melbourne University), Austin Hospital, Melbourne, Australia2 Research Nurse. Department of Intensive Care and Department of Surgery (Melbourne University), Austin Hospital, Melbourne, Australia3 Intensive Care Consultant. Department of Intensive Care and Department of Surgery (Melbourne University), Austin Hospital, Melbourne, Australia2005 28 4 2005 9 4 R303 R306 11 2 2005 16 3 2005 28 3 2005 8 4 2005 Copyright © 2005 Jones et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Hospital medical emergency teams (METs) have been implemented to reduce cardiac arrests and hospital mortality. The timing and system factors associated with their activation are poorly understood. We sought to determine the circadian pattern of MET activation and to relate it to nursing and medical activities.
Method
We conducted a retrospective observational study of the time of activation for 2568 incidents of MET attendance. Each attendance was allocated to one of 48 half-hour intervals over the 24-hour daily cycle. Activation was related nursing and medical activities.
Results
During the study period there were 120,000 consecutive overnight medical and surgical admissions. The hourly rate of MET calls was greater during the day (47% of calls in the 10 hours between 08:00 and 18:00), but 53% of the 2568 calls occurred between 18:00 and 08:00 hours. MET calls increased in the half-hour after routine nursing observation, and in the half-hour before each nursing handover. MET service utilization was 1.25 (95% confidence interval [CI] = 1.11–1.52) times more likely in the three 1-hour periods spanning routine nursing handover (P = 0.001). The greatest level of half-hourly utilization was seen between 20:00 and 20:30 (odds ratio [OR] = 1.76, 95% CI = 1.25–2.48; P = 0.001), before the evening nursing handover. Additional peaks were seen following routine nursing observations between 14:00 and 14:30 (OR = 1.53, 95% CI = 1.07–2.17; P = 0.022) and after the commencement of the daily medical shift (09:00–09:30; OR = 1.43, 95% CI = 1.00–2.04; P = 0.049).
Conclusion
Peak levels of MET service activation occur around the time of routine observations and nursing handover. Our findings raise questions about the appropriate frequency and methods of observation in at-risk hospital patients, reinforce the need for adequately trained medical staff to be available 24 hours per day, and provide useful information for allocation of resources and personnel for a MET service.
==== Body
Introduction
The medical emergency team (MET) concept is an evolving hospital system change that aims to reduce morbidity and mortality in acutely ill ward patients [1-3]. The MET is most often comprised of intensive care-based staff who are mobilized by ward-based doctors and nurses to review critically ill patients on the ward. The success of the MET system relies on the assumption that early intervention in the course of clinical deterioration improves patient outcome [4]. It would be important to gain insight into the possible processes that lead to MET calls and to understand their circadian variation in order to plan appropriate staff allocation.
We recently reported that the implementation of a MET system in our hospital resulted in a 65% relative risk reduction for in-hospital cardiac arrest over a 4-month period [4]. Analysis of the pattern of activation of the MET service in that study revealed a trend toward increased activation during the evening (P = 0.12). Lee and coworkers [5] reported that 36% of 522 MET calls registered over a 1-year period occurred between the hours 20:00 and 08:00. No information, however, exists on the possible relationship between routine nursing or medical activity and MET calls.
Available evidence suggests that between 69% and 82% of MET calls are initiated by a nurse [5,6]. The criteria for MET activation at our institution are based on derangements in vital signs that are typically measured or assessed at times of routine nursing observation and handover. Thus, we hypothesized that activation of the MET service at our institution would cluster around these times. To test this hypothesis we analyzed the frequency of MET activation at half-hourly intervals over a 24-hour period and related this to aspects of nursing and medical daily routine.
Materials and methods
The hospital
Austin Health is a university-affiliated teaching hospital with three hospital campuses situated in Melbourne, Australia. The Austin Hospital is the acute care hospital in which the MET service operates. It has 400 beds and receives approximately 60,000 day and overnight admissions per year.
Hospital emergency response teams
The acute care hospital has two levels of medical emergency responses and teams. A traditional cardiac arrest ('code blue') team is comprised of a cardiology fellow and coronary care nurse, as well as an intensive care fellow and nurse, and the receiving medical unit fellow. All wards are equipped with resuscitation trolleys containing resuscitation drugs and defibrillators.
In September 2000 a MET system was introduced into the acute campus following an extensive preparation and education process [4]. The team consists of an intensive care fellow and nurse, as well as the receiving medical unit fellow. It can be activated by any member of the hospital staff according to preset criteria for physiological instability. All code blue and MET calls are communicated by the switchboard operators through the hospital loudspeakers and paging system, and a detailed log of all calls is maintained.
Criteria for medical emergency team activation
Calling criteria for our MET service are based on acute changes in heart rate (<40 or >130 beats/min), systolic blood pressure (<90 mmHg), respiratory rate (<8 or >30 breaths/min), conscious state, urine output (<50 ml over 4 hours), and oxygen saturation derived from pulse oximetry (<90%, despite oxygen administration). In addition, the calling criteria contain a 'staff member is worried' category to allow staff to summon senior assistance to manage any possible emergency situation.
Outcome measures
Information on the activation of all MET calls is maintained on a hospital switchboard logbook that includes the date and time of the call, as well as the ward where the MET review occurred. The details of 2568 MET calls were manually entered into an MS Excel™ spreadsheet by two investigators who worked together and cross-checked the entries to minimize errors.
Each call was allocated to one of 48 half-hourly intervals over a 24-hour period (24:00–00:30, 00:31–01:00, 01:01–01:30, 01:31–02:00, etc.). A graph was then constructed from the 2568 episodes of MET service review to illustrate the frequency of activation at various times over the 24-hour period.
Episodes of activation were related to the periods of routine nursing handover (07:00, 13:00 and 21:00), routine nursing observations (02:00, 06:00, 10:00, 14:00, 18:00 and 22:00), and commencement and completion of the daily medical shift (08:00–18:00).
Statistical analysis
The frequency of MET service activation during peak periods was compared with the average activation over the 24-hour period. In the case of nursing handover, the 1-hour period spanning handover (the half-hour before and the half-hour after, repeated three times per day for a total of 3 hours) was compared with the average activation over the 24-hour period. Statistical significance was determined by analysis with Fisher's exact test using MS Windows Statview (Abacus Concepts, Berkeley, CA, USA). P < 0.05 was considered statistically significant.
Results
During the study period (August 2000 to September 2004) there were 120,000 consecutive overnight medical and surgical admissions to the Austin Hospital and 2568 activations of the MET service. Activation of the MET service was not uniform over the 24-hour period (Fig. 1).
Over the study period, 53% of the 2568 calls occurred in the 14 hours between 18:00 and 08:00 (58% of the day). On an hourly basis, MET call utilization was more common during the hours covered by the parent unit doctors (47% of MET calls during 42% of the day). In the 5 years that the MET system has operated, there has been a trend for an increasing proportion of calls to occur after hours (18:00–08:00; Fig. 2). Thus, in 2004, 374 out of 669 (55.9%) MET calls occurred after hours, compared with 69 out of 139 (49.6%) during the year 2000 (odds ratio [OR] = 1.13, 95% CI = 0.82–1.54; P = 0.19).
On average there were 106 calls (2568/24) for each hour period, or 53 calls (2568/48) per half-hour period. Increased activity of the MET service was typically seen in the half-hour following routine observations, and in the half-hour before routine nursing handover (Fig. 1). A total of 401 calls were made in the three 1-hour periods spanning nursing handover. During these periods, activation of the MET service was 1.25 times more likely (95% CI = 1.11–1.52) when compared with the average activation over the 24-hour period (P = 0.001).
The highest level of MET service activation for any given half-hour period was seen between 20:00 and 20:30, when use of the MET service was 1.8 (95% CI = 1.25–2.48) times greater than average half-hourly utilization (P = 0.001). Additional peaks of activity were seen between 14:00 and 14:30 (OR = 1.53, 95% CI = 1.07–2.17; P = 0.022) and between 09:00 and 09:30 (OR = 1.43, 95% CI = 1.00–2.04; P = 0.049). All other peaks of activity failed to achieve statistical significance.
Discussion
We report, for the first time, a detailed analysis of the level of utilization of a MET service over a 24-hour period and found a significant increase in the number of MET calls around periods of nursing handover and routine nursing observation. In addition, although MET calls occurred more frequently during the hours 08:00–18:00 (47% of calls during 42% of the day), a substantial proportion of MET calls occur after normal working hours (53% of calls during 58% of the day), with the peak time of activity occurring between 20:00 and 20:30. These findings have important implications for the frequency and method of patient monitoring, as well as for allocation of critical care resources and MET personnel, and require detailed discussion.
In a previous study at our institution [6] there was a trend toward more frequent activation of the MET service in the evening. In a study of 522 MET calls over a 1-year period, Lee and coworkers [5] demonstrated that 36% of MET calls were registered during the nightshift (20:00–8:00). Although the rate of MET calls did not vary during periods of reduced staffing, the investigators emphasized the importance of providing appropriately trained medical staff on a 24-hour basis.
In the present study, 53% of all calls occurred 'out of hours' (18:00–08:00) when wards are not staffed by parent unit doctors. In addition, there was a trend toward increased frequency of activation of the MET service during these hours in the 5 years following the introduction of the MET system. When directly compared with the study conducted by Lee and coworkers [5], 46.2% of the 2568 MET calls registered in the present study occurred between 20:00 and 08:00 hours. Our findings suggest a greater utilization of the MET service in the hours not covered by the parent unit medical staff than has previously been reported.
The frequent use of the MET service after 18:00 has important implications for allocation of resources to the MET service out of hours, and further reinforces previously reported opinion [5] that appropriately trained medical staff should be available on a 24-hour basis to assess and treat acutely ill hospital patients.
Utilization of a MET system has been associated with a reduction in all-cause hospital mortality in our institution [4,6]. Thus, our observation that MET service activation clusters around times of nursing handover and routine nursing observations raises questions about the appropriate frequency and methods of observations in 'at-risk' hospital patients. A more frequent or automated (e.g. telemetry) observation system for such at-risk patients may result in further reductions in mortality and morbidity. It is unlikely that patients would develop acute illness more frequently at specific times that happen to coincide with nursing observations or handover. It is more likely that the patient was discovered to be unwell only during a 'scheduled visit' by his/her care givers. In the case of medical staff, this would correspond to the morning medical ward round. In the case of nursing staff, we have clearly demonstrated increased levels of MET activity during periods when nurses are more likely to be tending to the patient. It is likely, therefore, that a substantial proportion of these patients would have been ill for some time before the call was made, and were only identified during routine observations or at the time of nursing handover. It is also possible that the diurnal variation of identifying 'patients in crisis' observed in the present study would not be seen in an environment with more automated and/or continuous monitoring.
The present study has a number of limitations. First, it is an observational study and does not demonstrate the effect of MET service utilization on patient outcome. However, we know from previous studies [4,6] that the introduction of the MET service was associated with significant beneficial effects on morbidity and mortality. Second, the pattern of fluctuation of the MET service at our institution is likely to be based on the calling criteria that we have implemented. The study may not apply to other hospitals where alternative calling criteria are employed. However, we deliberately employed simple calling criteria to increase the ease of utilisation of the MET system at our institution. Furthermore, the timing and frequency of patient observations reported in the study would be typical of most hospitals. Finally, information on episodes of MET review was obtained from the hospital switchboard log and did not provide information on the member of staff who activated the system. It would be interesting to known whether there was variation in the nature of the member (doctor versus nurse) and seniority of staff at various times of the day. We are currently collecting information on this aspect of MET operation.
Conclusion
In our institution, peak levels of MET service utilization occur around the time of routine nursing observations and nursing handover, and the majority of calls occur after hours. Our findings raise questions about the appropriate frequency and technology of observations in hospital ward patients. They also provide useful information to guide appropriate resource allocation for the provision of the MET service.
Key messages
• More than half of MET calls occur after hours.
• The peak time of MET activation is at 20:00, just before nursing handover.
• Other peak activities occur around nursing handover times or medical ward round times.
• These findings suggest that critical illness detection in hospital is episodic.
• More systematic approaches to hospital patient monitoring may be desirable in order to provide more timely intervention.
Abbreviations
CI = confidence interval; MET = medical emergency team; OR = odds ratio.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
DJ conceived the study, constructed the data base, and was the principle author of the manuscript. SB, DG, and SW assisted with construction of the data base. HO, GG, and RB contributed with the study design and authorship of the manuscript. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 Medical emergency team (MET) calls over 24 hours. Shown is a graph illustrating the number of MET calls made per half-hour over a 24-hour period for 2568 episodes of MET review in relation to aspects of daily nursing and medical routine. Arrows demonstrate periods of nursing handover (red, up-pointing arrows), the beginning and end of the daily medical shift (green, down-pointing arrows), and periods of routine nursing observations (pink, shorter, up-pointing arrows). The dotted line represents the average number of MET calls made per half-hour interval. *P < 0.05.
Figure 2 Medical emergency team (MET) calls during periods 08:00–18:00 and 18:00–08:00 comparison. Shown is a comparison of the percentage of MET calls made during the periods 08:00–18:00 and 18:00–08:00 for the years 2000–2004.
==== Refs
Buist MD Jarmolowski E Burton PR Bernard SA Waxman BP Anderson J Recognising clinical instability in hospital patients before cardiac arrests or unplanned admissions to intensive care Med J Aust 1999 171 22 25 10451667
Franklin C Mathew J Developing strategies to prevent in-hospital cardiac arrest: analyzing responses of physicians and nurses in the hours before the event Crit Care Med 1994 22 244 247 8306682
Schein RM Hazday N Pena M Ruben BH Sprung CL Clinical antecedents to in-hospital cardiopulmonary arrests Chest 1990 98 1388 1392 2245680
Bellomo R Goldsmith D Uchino S Buckmaster J Hart GK Opdam H Silvester W Doolan L Gutteridge G A prospective before-and-after trial of a medical emergency team Med J Aust 2003 179 283 287 12964909
Lee A Bishop G Hillman KM Daffurn K The medical emergency team Anaesth Intensive Care 1995 23 183 186 7793590
Bellomo R Goldsmith D Uchino S Buckmaster J Hart G Opdam H Silvester W Doolan L Gutteridge G Prospective controlled trial of effect of medical emergency team postoperative morbidity and mortality rates Crit Care Med 2004 32 916 921 15071378 10.1097/01.CCM.0000119428.02968.9E
|
16137341
|
PMC1269438
|
CC BY
|
2021-01-04 16:04:53
|
no
|
Crit Care. 2005 Apr 28; 9(4):R303-R306
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3537
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35381613734510.1186/cc3538ResearchDrotrecogin alfa (activated) in patients with severe sepsis presenting with purpura fulminans, meningitis, or meningococcal disease: a retrospective analysis of patients enrolled in recent clinical studies Vincent Jean-Louis [email protected] Simon [email protected] Demetrios J [email protected] RT Noel [email protected] S Betty [email protected] Virginia L [email protected] Joan E [email protected] Carol L [email protected] Samiha [email protected] Stephen M [email protected] Jonathan M [email protected] Head, Department of Intensive Care, University of Brussels (Erasme Hospital), Brussels, Belgium2 Consultant in Paediatric Intensive Care, Department of Paediatrics, Imperial College London (St. Mary's Hospital), London, UK3 Assistant Professor, Department of Public Health Sciences, Division of Critical Care Medicine, University of Alberta (Royal Alexandra Hospital), Edmonton, Alberta, Canada4 Professor, Division of Critical Care Medicine, University of Alberta (University of Alberta Hospital), Edmonton, Alberta, Canada5 Research Fellow, Lilly Research Laboratories, Indianapolis, IN, USA6 Associate Consultant, Project Management, Lilly Research Laboratories, Indianapolis, IN, USA7 Clinical Development Associate, Lilly Research Laboratories, Indianapolis, IN, USA8 Associate Global Medical Information Consultant, Lilly Research Laboratories, Indianapolis, IN, USA9 Statistician, Lilly Research Laboratories, Indianapolis, IN, USA10 Scientific Communications Associate, Lilly Research Laboratories, Indianapolis, IN, USA11 Medical Advisor, Lilly Research Centre, Erl Wood Manor, Windlesham, Surrey, UK2005 17 5 2005 9 4 R331 R343 28 1 2005 24 2 2005 4 4 2005 8 4 2005 Copyright © 2005 Vincent et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
We report data from adult and pediatric patients with severe sepsis from studies evaluating drotrecogin alfa (activated) (DrotAA) and presenting with purpura fulminans (PF), meningitis (MEN), or meningococcal disease (MD) (PF/MEN/MD). Such conditions may be associated with an increased bleeding risk but occur in a relatively small proportion of patients presenting with severe sepsis; pooling data across clinical trials provides an opportunity for improving the characterization of outcomes.
Methods
A retrospective analysis of placebo-controlled, open-label, and compassionate-use trials was conducted. Adult patients received infusions of either DrotAA or placebo. All pediatric patients (<18 years old) received DrotAA. 189 adult and 121 pediatric patients presented with PF/MEN/MD.
Results
Fewer adult patients with PF/MEN/MD met cardiovascular (68.3% versus 78.8%) or respiratory (57.8% versus 80.5%) organ dysfunction entry criteria than those without. DrotAA-treated adult patients with PF/MEN/MD (n = 163) had an observed 28-day mortality rate of 19.0%, a 28-day serious bleeding event (SBE) rate of 6.1%, and an intracranial hemorrhage (ICH) rate of 4.3%. Six of the seven ICHs occurred in patients with MEN (three of whom were more than 65 years old with a history of hypertension). DrotAA-treated adult patients without PF/MEN/MD (n = 3,088) had an observed 28-day mortality rate of 25.5%, a 28-day SBE rate of 5.8%, and an ICH rate of 1.0%. In contrast, a greater number of pediatric patients with PF/MEN/MD met the cardiovascular organ dysfunction entry criterion (93.5% versus 82.5%) than those without. DrotAA-treated PF/MEN/MD pediatric patients (n = 119) had a 14-day mortality rate of 10.1%, an SBE rate of 5.9%, and an ICH rate of 2.5%. DrotAA-treated pediatric patients without PF/MEN/MD (n = 142) had a 14-day mortality rate of 14.1%, an SBE rate of 9.2%, and an ICH rate of 3.5%.
Conclusion
DrotAA-treated adult patients with severe sepsis presenting with PF/MEN/MD had a similar SBE rate, a lower observed 28-day mortality rate, and a higher observed rate of ICH than DrotAA-treated patients without PF/MEN/MD. DrotAA-treated pediatric patients with severe sepsis with PF/MEN/MD may differ from adults, because all three outcome rates (SBE, mortality, and ICH) were lower in pediatric patients with PF/MEN/MD.
==== Body
Introduction
Despite the development of novel anti-infective therapies and improved patient management, severe sepsis remains a serious healthcare concern with an unacceptable mortality rate and an increasing incidence rate that has resulted in a significant economic and societal burden [1-3]. Although there has been a theoretical basis for blocking the excessive inflammatory response evoked during sepsis, so far such approaches have not led to the licensing of new compounds for the treatment of severe sepsis [4]. The contribution of coagulopathy to the pathophysiology of sepsis has become more widely understood [4-7] and has increased the interest in compounds that modulate the coagulation cascade such as antithrombin, tissue factor pathway inhibitor, and activated protein C [8-10]. Although several of these agents have been evaluated in large clinical trials, only recombinant human activated protein C (drotrecogin alfa (activated) (DrotAA; Xigris®); Eli Lilly and Company, Indianapolis, IN, USA] has been found to reduce 28-day all-cause mortality. DrotAA has been approved for treatment of adult patients with severe sepsis in more than 50 countries: in the USA, it is indicated for the reduction of mortality in adult patients with severe sepsis (sepsis associated with acute organ dysfunction) who have a high risk of death (for example, an Acute Physiology and Chronic Health Evaluation II (APACHE II) score of 25 or more); in the European Union, it is indicated (when added to best standard care) for the treatment of adult patients with severe sepsis and multiple organ failure.
Like endogenous activated protein C, DrotAA is a regulator of coagulation, fibrinolysis, and inflammation [11]. Consistent with its anticoagulant and profibrinolytic activity is its association with an increased incidence of serious bleeding events (SBEs), particularly in patients predisposed to bleeding [9,12]. Although the bleeding risk is modest, questions have arisen about treatment with DrotAA in patients predisposed to bleeding such as those with disseminated intravascular coagulation. In this relatively prevalent (about 30%) subpopulation of sepsis patients [13], retrospective analysis of data derived from a single trial recently demonstrated a favorable benefit-risk profile for DrotAA [14].
To examine additional safety information in smaller subgroups of patients, it is often helpful to pool experience across studies. Purpura fulminans (PF), with its attendant consumptive coagulopathy, and meningitis (MEN), with its attendant risk of intracranial hemorrhage (ICH), are two conditions seen in septic patients that, although not rare, are much less prevalent than disseminated intravascular coagulation [15-17]. Because both coagulopathy and MEN are sequelae of Neisseria meningitidis infection, patients with meningococcal disease (MD) may represent an additional population predisposed to bleeding complications [18,19]. Owing in part to the low incidence of PF, MEN, and MD (3% or less) in sepsis studies, limited data are available characterizing SBEs in septic patients with these conditions [15,19-21]. Uncertainty about the true SBE rates in the sepsis population confounds the interpretation of safety data from the few case reports describing the use of DrotAA in patients with PF, MEN, or MD [22-28].
The recent completion of several clinical studies evaluating DrotAA as an adjunctive treatment in severe sepsis affords an opportunity to improve our understanding of patients presenting with clinical signs and symptoms of PF, MEN, or MD. Here we report the baseline characteristics, mortality outcomes, and observed incidence rates of serious adverse events (especially SBEs and ICHs) in patients with and without PF, MEN, or MD.
Materials and methods
Data collection
Data were extracted from four clinical studies investigating DrotAA in adult and pediatric patients with severe sepsis. A database of 4,360 patients (4,096 adult, 264 pediatric) was assembled and, using retrospectively defined criteria, 310 patients (189 adult, 121 pediatric) with signs and symptoms of PF, MEN, or MD were identified, most of whom received DrotAA (165 adult, 121 pediatric). The studies pooled included, first, one multicenter, placebo-controlled, randomized, double-blind, phase 3 trial ('PROWESS', 1,690 adult patients enrolled; 850 DrotAA-treated, 840 placebo); second, one multicenter, open-label phase 3b study ('ENHANCE', 2,378 adult and 188 pediatric patients enrolled); third, one phase 2b open-label pediatric trial (EVAO, 83 patients enrolled); and fourth, one open-label compassionate-use study (EVAS, 28 adult and 14 pediatric patients enrolled) [9,29,30]. Pediatric patients (n = 21) enrolled in the dose-escalation phase of EVAO were not included in the present investigation [29]. Study investigators adhered to good clinical practices and ethical principles as stated in the Declaration of Helsinki of 1975, revised in 1983.
Trial inclusion and exclusion criteria
PROWESS and ENHANCE, as detailed previously, used similar inclusion criteria: proven or suspected infection; three or more signs of systemic inflammatory response syndrome (SIRS) (two or more signs of SIRS for pediatric patients); and evidence of one or more sepsis-induced organ dysfunctions (cardiovascular, respiratory, renal, hematologic, or metabolic acidosis) [9,30]. In comparison with PROWESS, the ENHANCE study design resulted in a longer time between the identification of acute organ dysfunction and initiation of the study drug. The EVAO study enrolled pediatric patients with severe sepsis and used the following inclusion criteria: proven or suspected infection; two or more signs of SIRS within 24 hours of study entry; and evidence of one or more sepsis-induced organ dysfunctions (cardiovascular, respiratory, renal, or hematologic) [29]. The original protocol for EVAO allowed enrollment on the basis of either cardiovascular or respiratory organ dysfunction but was subsequently amended to include renal and hematologic dysfunction in addition. The single inclusion criterion for the EVAS study was a clinical diagnosis of PF. Exclusion criteria were largely similar between trials: body weight more than 135 kg (and less than 3 kg for pediatric patients); platelet count less than 30,000/mm3 (the EVAS study did not exclude patients on the basis of platelet count); congenital or acquired conditions that increase the risk of serious bleeding; moribund state and presumed imminent death (within 24 hours for PROWESS, ENHANCE, and EVAO trials; within 6 hours for EVAS); and recent pharmacologic intervention that might induce a hypocoagulable state [9,29,30].
Patient selection and definitions
PF, MEN, and MD were not prospectively defined subgroups in the four trials, with the exception that a diagnosis of PF was required for enrollment in the compassionate-use study. A two-step identification process was developed for this retrospective analysis. Both study case report forms and investigator reports of serious adverse events were interrogated for medical and microbiological terms associated with PF, MEN, or MD. Data from patients identified in step one were then reviewed in detail and, on the basis of predefined selection criteria, patients were assigned to one or more of the following groups: PF, MEN, and MD. Because a prospective diagnosis of PF was required for enrollment in EVAS, all these patients were included in the PF group.
In a similar manner to previous retrospective analyses [31-33], the diagnosis of MEN was based on the following criteria: cerebrospinal fluid (CSF) findings consistent with MEN (positive CSF culture, leukocytosis, or positive CSF Gram stain); a clinical picture consistent with MEN (meningismus, headache, stiff neck, photophobia) together with the positive culture of a MEN-associated microorganism; or clinical diagnosis of MEN listed in the case comments. The diagnosis of PF was based on clinical diagnosis or purpuric rash, necrosis of digits, or gangrene recorded in the case comments or serious adverse event reports. A diagnosis of MD was based on clinical diagnosis in case comments or the identification of N. meningitidis in CSF or blood (positive culture, positive Gram stain, or other techniques such as polymerase chain reaction).
Bleeding events reported as serious adverse events (namely SBEs) included fatal or life-threatening events (patient at risk of death at the time of event occurrence), ICHs, or events associated with the following transfusion requirements: at least 3 units of packed red blood cells (RBCs) per day for two consecutive days (adult patients and pediatric patients 12 years to less than 18 years old); at least 20 ml of packed RBCs per kilogram per 24 hours (pediatric patients less than 1 year old); at least 10 ml of packed RBCs per kilogram per 24 hours (pediatric patients 1 year to less than 12 years old).
Drug administration
Adult and pediatric patients were to receive an intended 96-hour continuous infusion of DrotAA (24 μg kg-1 h-1); EVAS patients could have received up to an 168-hour continuous infusion of DrotAA (24 μg kg-1 h-1). In PROWESS, placebo patients received either 0.1% albumin or saline. No pediatric patients received placebo.
Statistical analysis
Data were extracted from validated clinical trial databases. All calculations were derived with SAS version 8.2 (SAS Institute, Inc., Cary, NC, USA). Continuous data were summarized by means of measures of central tendency and dispersion. Categorical data were summarized with incidence rates and counts. All analyses were exploratory and descriptive; no adjusted statistical analyses of event rates were performed across clinical trials, patient groups, or treatment groups. Twenty-eight-day mortality rates were calculated for adult patients. Pediatric mortality rate calculations were limited to 14-day endpoints because of differences in study design. Mortality and SBE rates are presented with 95% confidence intervals (CIs) generated with the exact CI method. Unadjusted odds ratios with 95% CIs were generated for the effect of diagnostic group membership (with or without PF, MEN, or MD) on mortality.
Results
Adult patients
One hundred eighty-nine (4.6%) of the total 4,096 adult patients with severe sepsis were identified as having PF, MEN, or MD. Because patients could be classified as having multiple diagnoses, there was substantial overlap between patient groups (Fig. 1). Most of the 189 patients were derived from either the ENHANCE (DrotAA, n = 111) or PROWESS (DrotAA, n = 26; PLC, n = 24) trials, and the remaining patients were enrolled in the EVAS compassionate-use study (DrotAA, n = 28).
Baseline characteristics of adult patients with severe sepsis presenting with PF, MEN, or MD are shown in Table 1. Patients with PF, MEN, or MD were younger, with less sepsis-associated organ dysfunction and fewer underlying comorbidities but with more thrombocytopenia. Less time elapsed from first organ dysfunction to the start of DrotAA treatment in patients with PF, MEN, or MD (mean 18.3 hours) than in those without (mean 22.6 hours). PF patients had the shortest mean time to treatment (mean 13.5 hours) and the lowest median baseline protein C level (30% of normal adult pooled plasma level). Although ENHANCE potentially allowed a longer window than PROWESS from first organ dysfunction to the start of treatment, the median time-to-DroAA treatment for patients with PF, MEN, or MD from ENHANCE was 15.7 hours; for those treated with DroAA from PROWESS it was 18.6 hours.
Predominant etiologic pathogens for patients with PF (67 of 77 (87%) had a culture result available) were N. meningitidis (50 of 67; 75%) and Streptococcus pneumoniae (11 of 67; 16%). Similarly, for patients with MEN the most common pathogens (111 of 128 (87%) had a culture result available) were N. meningitidis (51 of 111; 46%) and S. pneumoniae (37 of 111; 33%). For the 24 placebo-treated patients with PF, MEN, or MD, the baseline mean APACHE II score was 26.0 (SD 8.3) and the baseline median number of organ dysfunctions was two.
Table 2 summarizes 28-day all-cause mortality and safety data for adults with PF, MEN, or MD treated with DrotAA. The observed mortality rates for patients with and without PF, MEN, or MD were 19.0% and 25.5%, respectively. The unadjusted odds ratio for patients with versus those without PF, MEN, or MD was 0.69 (95% CI 0.44 to 1.03). Although not shown in Table 2, the mortality rate for placebo-treated patients with PF, MEN, or MD (all from the PROWESS clinical trial) was 25.0% (6 of 24).
During the DrotAA infusion period (defined as the duration of DrotAA infusion plus one full calendar day), rates of total SBEs were similar between patients with and without PF, MEN, or MD (3.7% versus 3.2%), including both fatal (0.6% versus 0.4%) and life-threatening (1.2% versus 1.4%) events. SBE rates during the 28-day study period were also similar between patients with and without PF, MEN, or MD (6.1% versus 5.8%).
ICH rates seemed to differ between the two main diagnostic groupings. Among the DrotAA-treated patients with PF, MEN, or MD, two-thirds (4 of 6) of the SBEs observed during the infusion period were ICHs (ICH rate 2.5%; 4 of 163), whereas 13 of 100 SBEs were ICHs (ICH rate 0.4%; 13 of 3,088) in patients without PF, MEN, or MD. The ICH rate for the 28-day study period was 4.3% for patients with PF, MEN, or MD and 1.0% for patients without PF, MEN, or MD. Among the 24 placebo-treated patients with PF, MEN, or MD from PROWESS, only one SBE (an ICH in a patient with PF and pneumococcal sepsis) was reported.
Because DrotAA has been approved for the treatment of adults with severe sepsis with two or more organ dysfunctions (for example in the European Union) or at high risk of death in the USA (for example an APACHE II score of 25 or more), mortality as well as SBEs for DrotAA-treated patients are also presented by baseline disease severity in Table 3. Baseline APACHE II and organ dysfunction data were available for only 137 of the total 163 DrotAA-treated adults with PF, MEN, or MD; it was not collected for the 26 DrotAA-treated adults with PF, MEN, OR MD from the compassionate-use open-label trial EVAS. DrotAA-treated adults with PF, MEN, or MD with either a baseline APACHE II score of 25 or more or with at least two baseline organ dysfunctions still had lower 28-day mortality rates than those in the high-severity subgroups without PF, MEN, or MD. Observed serious bleeding rates (infusion as well as 28-day) in the stratified groups were similar to all-event rates.
Table 4 (each column represents data for one patient) summarizes disease categories, baseline disease severity scores, and organ failure assessment scores for the 10 PF, MEN, or MD patients experiencing an SBE during the 28-day study period. All four ICHs during infusion and six of seven ICHs during the 28-day study period occurred in patients with MEN. Nearly half (three of seven) were observed in patients over 65 years old with a history of hypertension. Two of the four ICHs observed during the infusion period were associated with platelet counts less than 80,000/mm3 on the day before the event.
Pediatric patients
Of the 264 pediatric patients with severe sepsis, 121 (45.8%) were identified as having PF, MEN, or MD. As shown in Fig. 2, substantial overlap between these patient groups was observed. About 67% (81 of 121) of the patients originated from the pediatric arm of the ENHANCE trial, whereas the remaining 33% were enrolled in either the EVAO (n = 26) or EVAS (n = 14) studies.
Table 5 shows the baseline characteristics of pediatric patients with PF, MEN, or MD. Patients with PF, MEN, or MD were more likely to require vasopressor support but were less likely to receive ventilator support than those without PF, MEN, or MD. As in adults, DrotAA treatment began sooner after the first organ dysfunction for patients with PF, MEN, or MD (mean 13.0 hours) than in those without (mean 22.3 hours). Among pediatric patients with MEN (42 of 50 (84%) had a culture result available), the most common etiologic pathogens were N. meningitidis (31 of 42; 74%) and Group B streptococci (7 of 42; 17%). N. meningitidis (68 of 71; 96%) predominated in PF patients (71 of 87 (82%) had a culture result available).
Table 6 summarizes 14-day all-cause mortality and safety data for pediatric patients with severe sepsis with PF, MEN, or MD treated with DrotAA. Patients with PF, MEN, or MD had a lower observed mortality rate than patients without (10.1% versus 14.1%). The unadjusted odds ratio for patients with versus those without PF, MEN, or MD was 0.68 (95% CI 0.29 to 1.60).
During DrotAA infusion, the observed SBE rate was lower for patients with PF, MEN, or MD than for those without PF, MEN, or MD (1.7% versus 7.0%). Furthermore, there were no instances of ICH in patients with PF, MEN, or MD during DrotAA infusion, whereas patients without PF, MEN, or MD had an observed ICH rate of 1.4%. SBE rates for the entire study period were more equivalent between patient groups (with PF, MEN, or MD, 5.9%; without, 9.2%). The reported ICH rate during the study period was also similar between patients with and without PF, MEN, or MD (2.5% versus 3.5%).
Table 7 provides detailed information for pediatric patients experiencing an SBE during the study period. All seven SBEs occurred in patients with PF or MD, and five patients had signs and symptoms consistent with both diagnoses. All patients experiencing an SBE had baseline platelet counts of less than 75,000/mm3 (four patients had platelet counts of 30,000/mm3 or less). Four of the six patients for whom data were available also had a baseline activated partial thromboplastin time of more than 100 s.
Discussion
Because most adult and all pediatric patients with PF, MEN, or MD in this database were from open-label studies, the ability to make comparisons with a placebo group is limited. In view of the clinical overlap between PF, MEN, and MD, we considered these diagnoses collectively as well as individually. This approach is further supported by the likelihood that, given the retrospective nature of this study, it might not have been possible to complete a full clinical classification of all patients.
Incidence rates of PF, MEN, and MD in patients with severe sepsis are not widely available for comparison. In this analysis fewer than 5% (189 of 4,096) of adult patients with severe sepsis were identified as having PF, MEN, or MD, a finding similar to epidemiological analyses reporting a MEN incidence rate of 3.0% [21]. PF, MEN, and MD were much more prevalent among pediatric patients with severe sepsis with 46% (121 of 264) being diagnosed with or having signs or symptoms of PF, MEN, or MD. PF has been reported in 10 to 20% of patients with MD [34]. In our sample about 54% (50 of 92) of adult MD patients and 79% (71 of 90) of pediatric MD patients also had signs and symptoms of PF, although incidence rates might have been inflated by patient and site selection methods in these clinical trials.
There were important differences in demographic and clinical characteristics within and between diagnostic groupings. Collectively, adults with PF, MEN, or MD were younger with fewer underlying comorbidities than those without PF, MEN, or MD. Considered separately, adults with MEN were slightly older and more frequently had pre-existing hypertension or diabetes than adults with PF or MD. However, adult patients with PF and MD had evidence of greater baseline coagulopathy. For example, protein C deficiency was most severe in the adult PF group, followed by the MD group. Pediatric protein C levels were more consistent between the three diagnostic groupings. However, because protein C levels in children are highly dependent on age [35], baseline imbalances in age between the comparator groups potentially confound the interpretation of protein C deficiency. Protein C levels in the adult patients were more in line with the general perception that patients with PF and MD have worse coagulopathy than MEN patients.
The time from the first organ dysfunction to the start of DrotAA treatment differed between those with and without PF, MEN, or MD. A time-to-treatment difference was even more striking when individual diagnoses were examined, because adult patients with PF and MD began DrotAA treatment sooner than all other adult subgroups. A reduced time to DrotAA treatment probably reflects the more marked and unambiguous clinical presentation of PF and MD.
Mortality rates for MEN in the literature vary widely by pathogen and patient age [20,36,37]. For adults and adolescents, reported mortality rates for bacterial MEN range from 11 to 37% [16,17,31-33,38-42]. For children, mortality rates for MEN tend to be closer to 10% but have been reported to be 21% for those also presenting with shock [37,41,43]. The case fatality rate for MD has been reported to be between 8% and 14%, although can be as high as 20% in those less than 1 year of age [3,36,44]. PF has a much wider reported mortality range of 37 to 60% [45-47]. In the present analysis, mortality rates for patients with PF, MEN, or MD were 19.0% for adults and 10.1% for pediatric patients. However, it is difficult to directly compare clinical trial data, potentially confounded by entry and exclusion criteria, with data from epidemiological reports that may comprise a broader spectrum of patients.
There was insufficient evidence (for example small numbers of patients) to suggest that any mortality rate differences were statistically significant; however, the mortality rate for patients with PF, MEN, or MD certainly does not seem higher for patients without such diagnoses or complications. This trend holds also for patients assessed to have a higher risk of death at baseline, by either APACHE II scores of 25 or more or with at least two organ dysfunctions. Importantly, DrotAA treatment did not seem to increase mortality in adult patients with PF, MEN, or MD, because mortality rates for DrotAA-treated and placebo-treated patients with PF, MEN, or MD were 19.0% and 25.0%, respectively. These findings are consistent with previously published reports showing a mortality reduction associated with DrotAA treatment [9,30]. However, patients with PF, MEN, or MD were younger, presented with fewer underlying comorbidities, and began receiving DrotAA sooner after first organ dysfunction than those without PF, MEN, or MD. Because any of these baseline parameters could influence patient outcome, the data presented here must be interpreted with caution.
DrotAA is a recombinant form of an endogenous regulator of coagulation and, consistent with its antithrombotic properties, is associated with an increased risk of SBEs. In trials evaluating DrotAA in adults, SBE rates range from 3.5 to 5.5% for DrotAA-treated patients, compared with 2.0% for placebo controls [9,30]. As a reference, reported SBE rates ranged from 1 to 6% in the placebo arm of other recently completed clinical trials in severe sepsis [8,10,48]; however, SBE definitions may vary between trials, limiting inter-trial comparisons. In this study, DrotAA-treated adults with and without PF, MEN, or MD generally had similar SBE rates (including fatal or life-threatening bleeding) both during the infusion and 28-day study periods. Patients with PF had a higher SBE rate when considered separately, a finding consistent with the greater baseline coagulopathy observed in this group.
Because of its associated morbidity and mortality, ICH is among the most serious of SBEs. In a 6-year retrospective study of intensive care unit patients developing ICH (n = 2,198), Oppenheim and colleagues [49] found a spontaneous ICH rate of 0.4% in the critically ill; patients with sepsis accounted for five of nine patients (56%) developing ICH in their report. Other conditions or comorbidities associated with ICH included thrombocytopenia and impaired renal and hepatic function [49]. Central nervous system bleeding events (including ICHs), seen in the placebo arms of large clinical trials in severe sepsis, tend to be about 0.3% or 0.4% [8,10]. By comparison, the ICH rate for DrotAA-treated patients was 0.2% versus 0.1% for placebo controls in the PROWESS clinical trial [9]. Sharshar and colleagues [50] suggest that the rate of ICH might be much higher in patients with septic shock, as post-mortem examination revealed evidence of cerebral hemorrhage in 6 of 23 septic shock patients (26%). However, direct comparison of cerebral hemorrhage incidence between survivors and non-survivors of septic shock was not conducted, and ICH rates may differ between those who do and do not survive septic shock.
Adults with PF, MEN, or MD had a higher ICH rate than those without PF, MEN, or MD, both during DrotAA infusion (2.5% versus 0.4%) and during the 28-day study period (4.3% versus 1.0%). Considered separately, patients with MEN had the highest ICH rates (3.8% during infusion and 5.7% during the study period). Factors other than MEN that may increase the risk of ICH were also present in patients developing ICH. Nearly half (three of seven) of the patients with ICH were more than 65 years old and had pre-existing hypertension. Moreover, thrombocytopenia was evident in two of four patients with ICH during infusion, and two patients had either hepatic or both hepatic and renal organ dysfunction at the time of the ICH event. Using a database composed of a similar sample of patients from the current study, Bernard and colleagues [12] found that almost half of ICH events during DrotAA infusion occurred in patients with MEN or thrombocytopenia. However, in the analysis by Bernard and colleagues, patients with PF, MEN, or MD were not studied as a collective subgroup, nor were comparisons of mortality outcome and safety made with those in patients without PF, MEN, or MD.
The ICH rates reported here are consistent with previous reports of acute bacterial MEN in non-DrotAA-treated patients. In previous reports, ICH incidence ranged from 1 to 9% [16,17,51]. Despite the apparent increased incidence of ICH in adult PF, MEN, or MD patients, the rates of fatal or life-threatening SBEs did not differ markedly between those with and without PF, MEN, or MD. The observed ICH rates for adults with PF, MEN, or MD receiving either placebo or DrotAA during the study period was similar (4.2% versus 4.3%), although the small placebo sample limits conclusions derived from such a comparison. The data suggest that adults with MEN are at increased risk of ICH. However, the quantity of any additional potential risk resulting from DrotAA treatment is not clear from this analysis.
In contrast to the findings in adults, pediatric patients with PF, MEN, or MD had lower SBE and ICH rates than those without, both during the DrotAA infusion and overall study period. Whereas most SBEs occurred during the infusion period (6 of 10) for adult patients with PF, MEN, or MD, for pediatric patients most SBEs occurred during the post-infusion period. A possible explanation of why pediatric Drot-AA treated patients with PF, MEN, or MD had lower ICH rates than their adult counterparts is that they did not have two of the four risk factors (age more than 65 years, pre-existing hypertension, thrombocytopenia, MEN) that seemed to be associated with increased ICH rates in adult DrotAA-treated patients with PF, MEN, or MD.
In addition, the lack of observed ICHs during DrotAA infusion and an ICH rate of 2.5% during the study period for pediatric patients with PF, MEN, or MD are particularly interesting in view of a recent study of recombinant tissue plasminogen activator treatment in children with meningococcal PF (reported ICH rate 8%; 5 of 62) [52]. However, it is difficult to compare open-label clinical trials and observational case studies directly, because patients enrolled in clinical trials might not represent the same spectrum of disease severity observed in observational studies. For example, the mortality rate for pediatric PF, MEN, or MD patients described in our study was 10.1%, compared with 47% for PF patients in the tissue plasminogen activator study [52].
Differences in both mortality and SBE outcomes between adult and pediatric patients with severe sepsis are intriguing and probably reflect differences in microbial etiology, physiology (for example physiologic reserve), associated underlying disease, and treatment strategies. For example, mortality rates for adult and pediatric patients with MEN in this database were 17.9% and 8.3%, respectively. S. pneumoniae, associated with higher mortality in MEN patients, was reported in 33% of adult versus 2% of pediatric patients with MEN.
This study has several limitations. Most adult and all pediatric patients with PF, MEN, or MD in this database were enrolled in open-label studies. Correspondingly, few patients received placebo, making comparisons between DrotAA and placebo groups difficult. Another limitation was that patient subgroups (with or without PF, MEN, and MD) were not defined prospectively. Because these patients were identified through retrospective case review and there was not a prospective requirement to collect all of the clinical data needed to make these diagnoses (except for PF in EVAS), it is possible that some patients having PF, MEN, or MD might have been missed or classified as having only one rather than multiple diagnoses. In addition, patient data were combined from clinical trials with similar but non-identical entry criteria. As a result of differences in study design, not all information of interest was collected for each patient. The combination of small sample size and incomplete data sets precludes robust statistical assessment of the impact of DrotAA treatment on either mortality or SBE and ICH incidence in patients with PF, MEN, or MD. As a result, this study does not definitively address reported differences between patient groups defined by either disease category or treatment.
A general limitation extends from comparing data from clinical trials to results obtained from epidemiological studies or case reviews. All patients described here met specific enrollment criteria. Although comparisons with epidemiological data might be of some utility, any inferences should be interpreted with this caveat in mind.
Despite the limitations, this study provides novel information. So far, data on DrotAA use in the management of PF, MEN, or MD have been limited to case reports [22-28]. Four reports noted a positive outcome in 9 of 10 patients and no serious bleeding complications with meningococcal PF [22,24,27,28]. One report documented a subarachnoid hemorrhage and a fatal outcome for a 67-year-old patient with pneumococcal MEN and septicemia [23]. Two other case reports had positive outcomes: one in an adult with pneumococcal sepsis and PF, the other in an adolescent with congenital protein C deficiency and PF. Findings from our study are in line with current case report data and complement it. One advantage of this study is the large cohort of patients hospitalized with a uniform diagnosis of severe sepsis; thus, the number of patients considered here exceeds that regularly examined in observational studies or case series reviews. Another advantage is the use of prospectively defined primary endpoints and serious adverse events to investigate clinical experience and outcome.
Given our study's limitations, no recommendation can be made about the use of DrotAA in patients with severe sepsis presenting with PF, MEN, or MD, despite this group's observed lower mortality than in those patients without PF, MEN, or MD. Information from this study might be of use to clinicians considering DrotAA treatment in PF, MEN, or MD patients with severe sepsis: safety information from the largest cohort of such patients is made available. No obvious connection between severity of illness as indicated by either an APACHE II score of at least 25 or two or more organ dysfunctions and the occurrence of SBEs is suggested.
Conclusion
Patients with severe sepsis with signs and symptoms of PF, MEN, or MD are generally perceived to be at higher risk of bleeding complications. In this retrospective study, neither adult nor pediatric patients receiving DrotAA and exhibiting signs and symptoms of PF, MEN, or MD had increased serious bleeding rates (including life-threatening or fatal events) compared with patients without PF, MEN, or MD. Adult, but not pediatric, patients with MEN seemed at increased risk for developing ICH. These findings should be borne in mind when considering DrotAA in the management of patients with severe sepsis with evidence of PF, MEN, or MD. DrotAA is approved only for adult patients with severe sepsis at high risk of death (United States Package Insert) or with multiple organ failure (European Union Summary of Product Characteristics). A large ongoing placebo-controlled study evaluating DrotAA treatment in pediatric patients with severe sepsis will permit a more robust analysis of the benefit-risk profile of DrotAA in pediatric patients.
Key messages
• Patients with severe sepsis with signs and symptoms of purpura fulminans (PF), meningitis (MEN), or meningococcal disease (MD) are generally perceived to be at higher risk of bleeding complications.
• In this retrospective study, neither adult nor pediatric patients receiving DrotAA and exhibiting signs and symptoms of PF, MEN, or MD had increased serious bleeding rates (including life-threatening or fatal events) compared with patients without PF, MEN, or MD
• Adult, but not pediatric, patients with MEN seemed to be at an increased risk for developing intracranial hemorrhage
• These findings should be borne in mind when considering DrotAA in the management of adult patients with severe sepsis with evidence of PF, MEN, or MD
• DrotAA is not currently approved for treatment of pediatric patients with severe sepsis, and the results of an ongoing placebo controlled study are awaited
Abbreviations
APACHE = Acute Physiology and Chronic Health Evaluation; CI = confidence interval; CSF = cerebrospinal fluid; DrotAA = drotrecogin alfa (activated); ICH = intracranial hemorrhage; MD = meningococcal disease; MEN = meningitis; PF = purpura fulminans; RBC = red blood cell; SBE = serious bleeding event; SIRS = systemic inflammatory response syndrome.
Competing interests
SBY, VLW, JEB, CLM, SS, SMS, and JMJ are employees of Eli Lilly and Company. JLV and SN are consultants for Eli Lilly and Company. DJK and RTNG have declared that they have no competing interests.
Authors' contributions
All authors were involved in discussions regarding the design and objectives of the study. SS provided statistical expertise and constructed the database. JEB, VLW, SBY, SS, and JMJ reviewed case report forms for adverse event, microbiological, and diagnostic details. SBY, CLM, and SMS performed the literature review. Each author either drafted sections of the manuscript or provided critical revision of important intellectual content. All authors read and approved the final manuscript.
Acknowledgements
Michael J Mihm and Justin H Northrup provided writing assistance and technical help.
Figures and Tables
Figure 1 Venn diagram of adult patient distribution by disease category.
Figure 2 Venn diagram of pediatric patient distribution by disease category.
Table 1 Baseline characteristics of adult severe sepsis patients with purpura fulminans, meningitis, or meningococcal disease
Baseline characteristics1 No PF, MEN, or MD n = 3,907 (816 PLC) PF, MEN, or MD n = 189 (24 PLC) PF n = 77 (5 PLC) MEN n = 128 (21 PLC) MD n = 92 (10 PLC)
Demographics and disease severity
Age (years), mean ± SD 60.3 ± 16.5 44.4 ± 19.6 34.6 ± 14.4 48.2 ± 20.4 34.9 ± 16.1
Male (%) 57.8 53.4 58.4 53.1 50.0
Caucasian (%) 86.8 91.0 92.2 91.4 89.1
Organ dysfunctions, median; q1-q3 3.0; 2.0–3.0 2.0; 1.0–4.0 3.0; 2.0–4.0 2.0; 1.0–3.0 2.0; 2.0–4.0
First organ dysfunction to infusion (h), mean ± SD 22.6 ± 13.0 18.3 ± 13.2 13.5 ± 11.8 18.5 ± 13.4 14.9 ± 11.3
APACHE II score, mean ± SD 23.2 ± 7.6 22.6 ± 8.1 22.3 ± 7.8 22.0 ± 7.9 21.9 ± 7.8
GCS score, mean ± SD 12.2 ± 3.8 10.3 ± 4.1 11.8 ± 3.8 9.8 ± 4.0 11.2 ± 4.1
Underlying comorbidities
Congestive cardiomyopathy (%) 5.7 0 0 0 0
COPD (%) 19.4 3.1 0 4.0 1.3
Diabetes (%) 21.0 9.9 2.0 11.3 5.1
Hypertension (%) 36.7 18.0 6.1 20.2 7.7
Liver disease (%) 2.9 2.5 4.1 2.4 2.6
Cancer (%) 16.3 7.5 2.0 8.1 5.1
Myocardial infarction (%) 11.6 1.9 0 2.4 0
Pancreatitis (%) 3.7 1.2 0 1.6 1.3
Recent trauma (%) 3.9 1.9 4.1 1.6 1.3
Recent surgery (%) 35.5 1.9 4.1 1.6 1.3
Coagulation biomarkers
Protein C level (%), median; q1-q3 46; 30–64 47; 30–67 30; 20–45 54; 38–74 40; 27–53
Platelet count, median; q1-q3 172; 105–249 91; 45–142 59; 35–95 119; 65–159 70; 34–121
APTT (s), median; q1-q3 41; 34–50 44; 35–55 48; 41–63 40; 34–49 48; 38–62
PT (s), median; q1-q3 18; 15–21 17; 15–22 17; 14–23 17; 15–21 19; 16–23
Cardiovascular and respiratory measures
Vasopressor (%) 69.6 60.1 76.6 47.2 72.5
Ventilator (%) 79.3 79.8 80.5 76.4 76.9
Cardiovascular study entry criteria (%) 78.8 68.3 77.6 64.5 79.5
Respiratory study entry criteria (%) 80.5 57.8 59.2 54.0 47.4
1Patients with missing data were excluded from this analysis. APACHE II, Acute Physiology and Chronic Health Evaluation II; APTT, activated partial thromboplastin time; COPD, chronic obstructive pulmonary disease; GCS, Glasgow Coma Scale; MD, meningococcal disease; MEN, meningitis; PF, purpura fulminans; PLC, placebo; PT, prothrombin time.
Table 2 Serious bleeding and mortality rates in adult severe sepsis patients treated with drotrecogin alfa (activated)
Period and type of event1 No PF, MEN, or MD (n = 3,088) PF, MEN, or MD (n = 163) PF (n = 70) MEN (n = 106) MD (n = 80)
SBEs during infusion
All events, % (n); 95% CI 3.2 (100); 2.6–3.9 3.7 (6); 1.4–7.8 4.3 (3); 1.0–12.0 3.8 (4); 1.0–9.4 3.8 (3); 0.8–10.6
Fatal, % (n) 0.4 (12) 0.6 (1) 0 0.9 (1) 0
Life-threatening, % (n) 1.4 (43) 1.2 (2) 1.4 (1) 0.9 (1) 0
ICH, % (n) 0.4 (13) 2.5 (4) 1.4 (1) 3.8 (4) 2.5 (2)
SBEs over 28 days
All events, % (n); 95% CI 5.8 (178); 5.0–6.6 6.1 (10); 3.0–11.0 8.6 (6); 3.2–17.7 5.7 (6); 2.1–11.9 3.8 (3); 0.8–10.6
Fatal, % (n) 0.8 (24) 0.6 (1) 0 0.9 (1) 0
Life-threatening, % (n) 2.6 (81) 2.5 (4) 4.3 (3) 1.9 (2) 0
ICH, % (n) 1.0 (32) 4.3 (7) 4.3 (3) 5.7 (6) 2.5 (2)
28-day mortality
Mortality, % (n); 95% CI 25.5 (788); 24.0–27.1 19.0 (31); 13.3–26.0 21.4 (15); 12.5–32.9 17.9 (19); 11.2–26.6 8.8 (7); 3.6–17.2
1 Patients lost to follow-up (No PF, MEN, or MD = 3; PF, MEN, or MD = 2) were excluded from this analysis. DrotAA, drotrecogin alfa (activated); ICH, intracranial hemorrhage; MD, meningococcal disease; MEN, meningitis; PF, purpura fulminans; SBE, serious bleeding event.
Table 3 Serious bleeding and mortality rates in DrotAA-treated adults by baseline disease severity
Period and type of event No PF, MEN, or MD N = 3,088 PF, MEN, or MD N = 1371
28-day mortality
APACHE II
≥ 25 35.3 (433/1,227) 22.5 (11/49)
<25 19.1 (355/1,861) 12.5 (11/88)
Number of organ failures
≥ 2 27.3 (693/2,538) 17.8 (18/101)
<2 17.3 (95/550) 11.1 (4/36)
SBEs during infusion
APACHE II
≥ 25 3.7 (45/1,227) 4.1 (2/49)
<25 3.0 (55/1,861) 3.4 (3/88)
Number of organ failures
≥ 2 3.4 (85/2,538) 4.0 (4/101)
<2 2.7 (15/550) 2.8 (1/36)
SBEs, 28-day
APACHE II
≥ 25 6.2 (76/1,227) 6.1 (3/49)
<25 5.5 (102/1,861) 4.6 (4/88)
Number of organ failures
≥ 2 6.0 (151/2,538) 5.0 (5/101)
<2 4.9 (27/550) 5.6 (2/36)
1Baseline APACHE II and baseline organ dysfunction data available for 137 patients (not collected for the 26 DrotAA-treated adults from the compassionate-use open-label EVAS trial. APACHE, Acute Physiology and Chronic Health Evaluation; DrotAA, drotrecogin alfa (activated); MD, meningococcal disease; MEN, meningitis; PF, purpura fulminans; SBE, serious bleeding event.
Table 4 Characteristics of DrotAA-treated adults with PF, MEN, or MD and experienced a serious bleeding event
Characteristic During infusion After infusion
Relative day of event onset1 1 1 2 3 5 6 6 7 11 18
Bleeding event
ICH No No Yes Yes Yes Yes Yes Yes Yes No
Fatal No No No No Yes No No No Yes No
DrotAA-related2 Yes No No No Yes Yes No Yes No No
Disease category
PF Yes Yes No No No Yes Yes No Yes Yes
MEN No No Yes Yes Yes Yes No Yes Yes No
MD Yes No Yes No No Yes No No No Yes
Baseline characteristics
Age (years) 19 20 47 67 73 24 41 77 51 40
Protein C level (%) 11 NA 11 68 55 79 - 52 - 24
Platelet count 38 58 154 102 93 30 136 56 11 51
APTT (s) 198 86.7 65.0 31.1 - 63.0 - 32.0 - 47
PT (s) 58.9 16.6 - 14.1 - - 16.3 - 2.6 -
Organ failure 5 NA 2 1 3 3 - 1 - 4
APACHE II score 30 NA 20 23 24 26 - 20 - 25
Hypertension No No No Yes Yes No No Yes No No
SOFA3
Platelet count 38 NA 70 120 181 34 NA NA NA NA
Hematology SOFA score 3 NA 2 1 0 3 NA NA NA NA
Hepatic SOFA score 1 NA 1 0 - 2 NA NA NA NA
Renal SOFA score 2 NA 2 0 0 0 NA NA NA NA
1Day 0 is defined as the calendar day on which DrotAA treatment began; 2this denotes whether or not the investigator considered the bleeding event to be related to DrotAA treatment; 3values reported are those obtained 1 day before the relative onset day of the ICH. Data available only during first 6 days of enrollment in the PROWESS and ENHANCE trials. APACHE II, acute physiology and chronic health evaluation II; APTT, activated partial thromboplastin time; BL, baseline; DrotAA, drotrecogin alfa (activated); ICH, intracranial hemorrhage; MD, meningococcal disease; MEN, meningitis; NA, not available; PC, protein C activity; PF, purpura fulminans; PT, prothrombin time; SOFA, Sequential Organ Failure Assessment. A dash indicates missing data.
Table 5 Baseline characteristics of pediatric severe sepsis patients with purpura fulminans, meningitis, or meningococcal disease
Baseline characteristics1 No PF, MEN, or MD N = 143 PF, MEN, or MD N = 121 PF N = 87 MEN N = 50 MD N = 90
Demographics and disease severity
Age (years), mean ± SD 6.8 ± 6.3 5.6 ± 5.8 6.1 ± 5.9 5.7 ± 6.2 5.9 ± 5.8
Male (%) 46.2 53.7 50.6 52.0 55.9
Caucasian (%) 63.6 86.8 90.8 80.0 92.2
Organ dysfunctions, median; q1-q3 2.0; 1.0–3.0 2.0; 1.0–3.0 2.0; 1.0–3.0 2.0; 1.0–3.0 2.0; 1.0–3.0
First organ dysfunction to infusion (h), mean ± SD 22.3 ± 13.3 13.0 ± 10.0 13.2 ± 9.5 12.3 ± 9.1 13.0 ± 10.2
Coagulation biomarkers
Protein C level (%), median; q1-q3 38; 24–59 27; 19–38 29; 19–38 24; 16–41 26; 18–35
Platelet count, median; q1-q3 115; 58–205 88; 51–142 85; 42–122 107; 67–178 87; 53–132
APTT (s), median; q1-q3 46; 35–59 57; 41–79 57; 40–82 49; 40–70 57; 40–80
PT (s), median; q1-q3 16; 14–21 20; 15–26 20; 16–26 17; 15–23 20; 16–25
Cardiovascular and respiratory measures
Vasopressor (%) 70.8 87.4 90.0 81.6 87.5
Ventilator (%) 90.6 80.0 82.9 76.3 79.2
Cardiovascular study entry criteria (%) 82.5 93.5 97.3 90.9 96.2
Respiratory study entry criteria (%) 64.3 34.6 37.0 34.1 30.8
1Patients with missing data were excluded from this analysis. APTT, activated partial thromboplastin time; MD, meningococcal disease; MEN, meningitis; PF, purpura fulminans; PT, prothrombin time.
Table 6 Serious bleeding and mortality rates in pediatric severe sepsis patients treated with drotrecogin alfa (activated)
Period and type of event1 No PF, MEN, or MD N = 142 PF, MEN, or MD N = 119 PF N = 85 MEN N = 48 MD N = 88
Serious bleeding events during infusion
All events, % (n); 95% CI 7.0 (10); 3.4–12.6 1.7 (2); 0.2–6.0 2.4 (2); 0.3–8.2 2.1 (1); 0.05–11.1 2.3 (2); 0.3–8.0
Fatal, % (n) 0 0 0 0 0
Life-threatening, % (n) 2.1 (3) 1.7 (2) 2.4 (2) 0 2.3 (2)
ICH, % (n) 1.4 (2) 0 0 0 0
Serious bleeding events over 28 days2
All events, % (n); 95% CI 9.2 (13); 5.0–15.2 5.9 (7); 2.4–11.7 7.1 (6); 2.6–14.7 4.2 (2); 0.1–14.3 6.8 (6); 2.5–14.3
Fatal, % (n) 0.7 (1) 1.7 (2) 1.2 (1) 2.1 (1) 2.3 (2)
Life-threatening, % (n) 2.1 (3) 3.4 (4) 4.7 (4) 0 3.4 (3)
ICH, % (n) 3.5 (5) 2.5 (3) 2.4 (2) 2.1 (1) 2.3 (2)
14-day mortality
Mortality, % (n); 95% CI 14.1 (20); 8.8–20.9 10.1 (12); 5.3–17.0 9.4 (8); 4.2–17.7 8.3 (4); 2.3–20.0 10.2 (9); 4.8–18.5
1Patients lost to follow-up (no PF, MEN, or MD = 1; PF, MEN, or MD = 2) were excluded from this analysis; 2duration of follow-up for the open-label and compassionate-use studies was 28 days, and follow-up for the phase 2b open-label study was 14 days. DrotAA, drotrecogin alfa (activated); ICH, intracranial hemorrhage; MD, meningococcal disease; MEN, meningitis; PF, purpura fulminans.
Table 7 Characteristics of DrotAA-treated pediatric patients with PF, MEN, or MD and experienced a serious bleeding event
Characteristic During infusion After infusion
Day of event1 1 1 7 8 10 UD UD
Bleeding event
ICH No No No Yes No Yes Yes
Fatal No No No Yes No Yes No
DrotAA related2 No No No Yes No No No
Disease category
PF Yes Yes Yes Yes Yes No Yes
MEN No Yes No No No Yes No
MD Yes Yes Yes Yes Yes Yes No
Baseline characteristics
Protein C level (%) 36 - 16 41 - 34 -
Platelet count 26 23 33 30 68 71 14
APTT (s) 68 165 121 220 73 111 -
PT (s) 23 38 22 - - 17 16
Organ failure 3 4 3 4 1 - -
1Day 0 is defined as the calendar day on which DrotAA treatment began; 2This denotes whether or not the investigator considered the bleeding event to be related to treatment with DrotAA. APTT, activated partial thromboplastin time; BL, baseline; DrotAA, drotrecogin alfa (activated); ICH, intracranial hemorrhage; PT, prothrombin time; UD, unknown day after infusion. A dash indicates missing data.
==== Refs
Angus DC Linde-Zwirble WT Lidicker J Clermont G Carcillo J Pinsky MR Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of care Crit Care Med 2001 29 1303 1310 11445675 10.1097/00003246-200107000-00002
Martin GS Mannino DM Eaton S Moss M The epidemiology of sepsis in the United States from 1979 through 2000 N Engl J Med 2003 348 1546 1554 12700374 10.1056/NEJMoa022139
Watson RS Carcillo JA Linde-Zwirble WT Clermont G Lidicker J Angus DC The epidemiology of severe sepsis in children in the United States Am J Respir Crit Care Med 2003 167 695 701 12433670 10.1164/rccm.200207-682OC
Marshall JC Such stuff as dreams are made on: mediator-directed therapy in sepsis Nat Rev Drug Discov 2003 2 391 405 12750742 10.1038/nrd1084
Dellinger RP Inflammation and coagulation: implications for the septic patient Clin Infect Dis 2003 36 1265 10.1086/374835
Freeman BD Buchman TG Coagulation inhibitors in the treatment of sepsis Expert Opin Investig Drugs 2002 11 69 74 11772322 10.1517/13543784.11.1.69
Polderman KH Girbes AR Drug intervention trials in sepsis: divergent results Lancet 2004 363 1721 1723 15158636 10.1016/S0140-6736(04)16259-4
Abraham E Reinhart K Opal S Demeyer I Doig C Rodriguez AL Beale R Svoboda P Laterre PF Simon S Efficacy and safety of tifacogin (recombinant tissue factor pathway inhibitor) in severe sepsis: a randomized controlled trial JAMA 2003 290 238 247 12851279 10.1001/jama.290.2.238
Bernard GR Vincent JL Laterre PF LaRosa SP Dhainaut JF Lopez-Rodriguez A Steingrub JS Garber GE Helterbrand JD Ely EW Efficacy and safety of recombinant human activated protein C for severe sepsis N Engl J Med 2001 344 699 709 11236773 10.1056/NEJM200103083441001
Warren BL Eid A Singer P Pillay SS Carl P Novak I Chalupa P Atherstone A Penzes I Kubler A Caring for the critically ill patient. High-dose antithrombin III in severe sepsis: a randomized controlled trial JAMA 2001 286 1869 1878 11597289 10.1001/jama.286.15.1869
Esmon CT The protein C pathway Chest 2003 124 Suppl 3 26 32 10.1378/chest.124.3_suppl.26S
Bernard GR Macias WL Joyce DE Williams MD Bailey J Vincent JL Safety assessment of drotrecogin alfa (activated) in the treatment of adult patients with severe sepsis Crit Care 2003 7 163 10.1186/cc2167
Levi M Ten Cate H Disseminated intravascular coagulation N Engl J Med 1999 341 586 592 10451465 10.1056/NEJM199908193410807
Dhainaut JF Yan SB Joyce DE Pettila V Basson B Brandt JT Sundin DP Levi M Treatment effects of drotrecogin alfa (activated) in patients with severe sepsis with or without overt disseminated intravascular coagulation J Thromb Haemost 2004 2 1924 1933 15550023 10.1111/j.1538-7836.2004.00955.x
Darmstadt GL Acute infectious purpura fulminans: pathogenesis and medical management Pediatr Dermatol 1998 15 169 183 9655311 10.1046/j.1525-1470.1998.1998015169.x
Kastenbauer S Pfister HW Pneumococcal meningitis in adults: spectrum of complications and prognostic factors in a series of 87 cases Brain 2003 126 1015 1025 12690042 10.1093/brain/awg113
Pfister HW Borasio GD Dirnagl U Bauer M Einhaupl KM Cerebrovascular complications of bacterial meningitis in adults Neurology 1992 42 1497 1504 1641143
Pathan N Faust SN Levin M Pathophysiology of meningococcal meningitis and septicaemia Arch Dis Child 2003 88 601 607 12818907 10.1136/adc.88.7.601
Rosenstein NE Perkins BA Stephens DS Popovic T Hughes JM Meningococcal disease N Engl J Med 2001 344 1378 1388 11333996 10.1056/NEJM200105033441807
Schuchat A Robinson K Wenger JD Harrison LH Farley M Reingold AL Lefkowitz L Perkins BA Bacterial meningitis in the United States in 1995. Active Surveillance Team N Engl J Med 1997 337 970 976 9395430 10.1056/NEJM199710023371404
Brun-Buisson C The epidemiology of the systemic inflammatory response Intensive Care Med 2000 26 Suppl 1 64 74 10.1007/s001340051121
Bachli EB Vavricka SR Walter RB Leschinger MI Maggiorini M Drotrecogin alfa (activated) for the treatment of meningococcal purpura fulminans Intensive Care Med 2003 29 337 12594601
King D Higgins D Subarachnoid haemorrhage following activated protein C for bacterial meningitis Anaesthesia 2003 58 913 914 12911371 10.1046/j.1365-2044.2003.03362_4.x
Wcisel G Joyce D Gudmundsdottir A Shasby DM Human recombinant activated protein C in meningococcal sepsis Chest 2002 121 292 295 11796469 10.1378/chest.121.1.292
Manco-Johnson MJ Knapp-Clevenger R Activated protein Cconcentrate reverses purpura fulminans in severe genetic protein C deficiency J Pediatr Hematol Oncol 2004 26 25 27 14707707 10.1097/00043426-200401000-00008
Cone LA Waterbor B Sofonio MV Purpura fulminans due to Streptococcus pneumoniae sepsis following gastric bypass Obes Surg 2004 14 690 694 15186640 10.1381/096089204323093507
Thomas GL Wigmore T Clark P Activated protein C for the treatment of fulminant meningococcal septicaemia Anaesth Intens Care 2004 32 284 287
Martinon-Torres F Iglesias Meleiro JM Fernandez Sanmartin M Rodriquez Nunez A Martinon Sanchez JM Recombinant human activated protein C in the treatment of children with meningococcal purpura fulminans An Pediatr (Barc) 2004 61 690 694
Barton P Kalil AC Nadel S Goldstein B Okhuysen-Cawley R Brilli RJ Takano JS Martin LD Quint P Yeh TS Safety, pharmacokinetics, and pharmacodynamics of drotrecogin alfa (activated) in children with severe sepsis Pediatrics 2004 113 7 17 14702440 10.1542/peds.113.1.7
Bernard GR Margolis BD Shanies HM Ely EW Wheeler AP Levy H Wong K Wright TJ Extended evaluation of recombinant human activated protein C United States Trial (ENHANCE US): a single-arm, phase 3B, multicenter study of drotrecogin alfa (activated) in severe sepsis Chest 2004 125 2206 2216 15189943 10.1378/chest.125.6.2206
Durand ML Calderwood SB Weber DJ Miller SI Southwick FS Caviness VS JrSwartz MN Acute bacterial meningitis in adults. A review of 493 episodes N Engl J Med 1993 328 21 28 8416268 10.1056/NEJM199301073280104
McMillan DA Lin CY Aronin SI Quagliarello VJ Community-acquired bacterial meningitis in adults: categorization of causes and timing of death Clin Infect Dis 2001 33 969 975 11528567 10.1086/322612
Sigurdardottir B Bjornsson OM Jonsdottir KE Erlendsdottir H Gudmundsson S Acute bacterial meningitis in adults. A 20-year overview Arch Intern Med 1997 157 425 430 9046894 10.1001/archinte.157.4.425
Faust SN Levin M Harrison OB Goldin RD Lockhart MS Kondaveeti S Laszik Z Esmon CT Heyderman RS Dysfunction of endothelial protein C activation in severe meningococcal sepsis N Engl J Med 2001 345 408 416 11496851 10.1056/NEJM200108093450603
Kuhle S Male C Mitchell L Developmental hemostasis: pro- and anticoagulant systems during childhood Semin Thromb Hemost 2003 29 329 338 14517745 10.1055/s-2003-42584
Cohen J Cristofaro P Carlet J Opal S New method of classifying infections in critically ill patients Crit Care Med 2004 32 1510 1526 15241096 10.1097/01.CCM.0000129973.13104.2D
Saez-Llorens X McCracken GH Jr Bacterial meningitis in children Lancet 2003 361 2139 2148 12826449 10.1016/S0140-6736(03)13693-8
Auburtin M Porcher R Bruneel F Scanvic A Trouillet JL Bedos JP Regnier B Wolff M Pneumococcal meningitis in the intensive care unit: prognostic factors of clinical outcome in a series of 80 cases Am J Respir Crit Care Med 2002 165 713 717 11874820
de Gans J van de BD European Dexamethasone in Adulthood Bacterial Meningitis Study Investigators Dexamethasone in adults with bacterial meningitis N Engl J Med 2002 347 1549 1556 12432041 10.1056/NEJMoa021334
Flores-Cordero JM Amaya-Villar R Rincon-Ferrari MD Leal-Noval SR Garnacho-Montero J Llanos-Rodriguez AC Murillo-Cabezas F Acute community-acquired bacterial meningitis in adults admitted to the intensive care unit: clinical manifestations, management and prognostic factors Intensive Care Med 2003 29 1967 1973 12904848 10.1007/s00134-003-1935-4
Stanek RJ Mufson MA A 20-year epidemiological study of pneumococcal meningitis Clin Infect Dis 1999 28 1265 1272 10451164
Zimmerli W Acute bacterial meningitis: time for a better outcome Intensive Care Med 2003 29 1868 1870 14669755 10.1007/s00134-003-1934-5
Odetola FO Bratton SL Characteristics and immediate outcome of childhood meningitis treated in the pediatric intensive care unit Intensive Care Med 2004 32 92 97
Booy R Habibi P Nadel S de Munter C Britto J Morrison A Levin M Meningococcal Research Group Reduction in case fatality rate from meningococcal disease associated with improved healthcare delivery Arch Dis Child 2001 85 386 390 11668100 10.1136/adc.85.5.386
J5 Study Group Treatment of severe infectious purpura in children with human plasma from donors immunized with Escherichia coli J5: a prospective double-blind study J Infect Dis 1992 165 695 701 1552198
Fourrier F Leclerc F Aidan K Sadik A Jourdain M Tournoys A Noizet O Combined antithrombin and protein C supplementation in meningococcal purpura fulminans: a pharmacokinetic study Intensive Care Med 2003 29 1081 1087 12761614 10.1007/s00134-003-1784-1
Powars D Larsen R Johnson J Hulbert T Sun T Patch MJ Francis R Chan L Epidemic meningococcemia and purpura fulminans with induced protein C deficiency Clin Infect Dis 1993 17 254 261 8399877
Schuster DP Metzler M Opal S Lowry S Balk R Abraham E Levy H Slotman G Coyne E Souza S Recombinant platelet-activating factor acetylhydrolase to prevent acute respiratory distress syndrome and mortality in severe sepsis: Phase IIb, multicenter, randomized, placebo-controlled, clinical trial Crit Care Med 2003 31 1612 1619 12794395 10.1097/01.CCM.0000063267.79824.DB
Oppenheim-Eden A Glantz L Eidelman LA Sprung CL Spontaneous intracerebral hemorrhage in critically ill patients: incidence over six years and associated factors Intensive Care Med 1999 25 63 67 10051080 10.1007/s001340050788
Sharshar T Annane D de la Grandmaison GL Brouland JP Hopkinson NS Francoise G The neuropathology of septic shock Brain Pathol 2004 14 21 33 14997934
Gironell A Domingo P Mancebo J Coll P Marti-Vilalta JL Hemorrhagic stroke as a complication of bacterial meningitis in adults: report of three cases and review Clin Infect Dis 1995 21 1488 1491 8749641
Zenz W Zoehrer B Levin M Fanconi S Hatzis TD Knight G Mullner M Faust SN Use of recombinant tissue plasminogen activator in children with meningococcal purpura fulminans: a retrospective study Crit Care Med 2004 32 1777 1780 15286558 10.1097/01.CCM.0000133667.86429.5D
|
16137345
|
PMC1269439
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 17; 9(4):R331-R343
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3538
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35401613734910.1186/cc3540ResearchRenal blood flow in sepsis Langenberg Christoph 1Bellomo Rinaldo [email protected] Clive 3Wan Li 1Egi Moritoki 1Morgera Stanislao 41 Research fellow, Department of Intensive Care and Department of Medicine, Austin Hospital, and University of Melbourne, Heidelberg, Melbourne, Australia2 Director of Intensive Care Research, Department of Intensive Care and Department of Medicine, Austin Hospital, and University of Melbourne, Heidelberg, Melbourne, Australia3 Senior Researcher, Howard Florey Institute, University of Melbourne, Parkville, Melbourne, Australia4 Consultant Nephrologist, Department of Nephrology, Charité Campus Mitte, Berlin, Germany2005 24 5 2005 9 4 R363 R374 20 1 2005 14 3 2005 1 4 2005 14 4 2005 Copyright © 2005 Langenberg et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
To assess changes in renal blood flow (RBF) in human and experimental sepsis, and to identify determinants of RBF.
Method
Using specific search terms we systematically interrogated two electronic reference libraries to identify experimental and human studies of sepsis and septic acute renal failure in which RBF was measured. In the retrieved studies, we assessed the influence of various factors on RBF during sepsis using statistical methods.
Results
We found no human studies in which RBF was measured with suitably accurate direct methods. Where it was measured in humans with sepsis, however, RBF was increased compared with normal. Of the 159 animal studies identified, 99 reported decreased RBF and 60 reported unchanged or increased RBF. The size of animal, technique of measurement, duration of measurement, method of induction of sepsis, and fluid administration had no effect on RBF. In contrast, on univariate analysis, state of consciousness of animals (P = 0.005), recovery after surgery (P < 0.001), haemodynamic pattern (hypodynamic or hyperdynamic state; P < 0.001) and cardiac output (P < 0.001) influenced RBF. However, multivariate analysis showed that only cardiac output remained an independent determinant of RBF (P < 0.001).
Conclusion
The impact of sepsis on RBF in humans is unknown. In experimental sepsis, RBF was reported to be decreased in two-thirds of studies (62 %) and unchanged or increased in one-third (38%). On univariate analysis, several factors not directly related to sepsis appear to influence RBF. However, multivariate analysis suggests that cardiac output has a dominant effect on RBF during sepsis, such that, in the presence of a decreased cardiac output, RBF is typically decreased, whereas in the presence of a preserved or increased cardiac output RBF is typically maintained or increased.
See related commentary
==== Body
Introduction
Acute renal failure (ARF) affects 5–7% of all hospitalized patients [1-3]. Sepsis and, in particular, septic shock are important risk factors for ARF in wards and remain the most important triggers for ARF in the intensive care unit (ICU) [4-8]. Among septic patients, the incidence of ARF is up to 51% [9] and that of severe ARF (i.e. ARF leading to the application of acute renal replacement therapy) is 5% [7,10]. The mortality rate associated with severe ARF in the ICU setting remains high [2-5,11].
A possible explanation for the high incidence and poor outcome of septic ARF relates to the lack of specific therapies. This, in turn, relates to our poor understanding of its pathogenesis. Nonetheless, a decrease in renal blood flow (RBF), causing renal ischaemia, has been proposed as central to the pathogenesis of septic ARF [12-14]. However, the bulk of knowledge about RBF in sepsis is derived from animal studies using a variety of different models and techniques. This creates uncertainty regarding the applicability of these studies to humans. Furthermore, the findings of studies in which experimental sepsis was induced and RBF measured have not been systematically assessed. Accordingly, we obtained all electronically identifiable publications reporting RBF in sepsis and analyzed the data according to changes in RBF. We also studied the possible influences of several technical and model-related variables on RBF.
Materials and methods
We conducted a systematic interrogation of the literature using a standardized approach as described by Doig and Simpson [15] and Piper and coworkers [16]. We used two electronic reference libraries (Medline and PubMed), and searched for relevant articles using the following search terms: 'renal blood flow', 'kidney blood flow', 'renal blood supply', 'kidney blood supply', 'organ blood flow', 'organ blood supply', 'sepsis', 'septic shock', 'septicemia', 'caecal puncture ligation', 'cecum puncture ligation', 'lipopolysaccharide' and 'endotoxin'. We selected all animal studies published in the English language literature. Using the reference lists from each article, we identified and obtained other possible studies that might have reported information on RBF in septic ARF and that had not been identified by our electronic search strategy.
We assessed all human articles in detail. Because of the heterogeneity animal studies and the methods they employed, we also assessed all animal articles systematically for information on variables that might have influenced RBF in sepsis. The variables of interest were as follows: size of animal; technique of measurement for RBF (direct measurement via flow probe or microsphere technique or other technique); consciousness of animals during the study; recovery period between preparation surgery and the experiment; timing of RBF measurement in relation to septic insult; method used to induce sepsis (lipopolysaccharide [LPS], live bacteria, or caecal ligation–perforation technique); fluid administration during the experiment; cardiac output (CO); and haemodynamic patterns (hypodynamic and hyperdynamic sepsis).
Information obtained on RBF from these groups was compared. Comparisons were performed using the ?2 test or Fisher exact test where appropriate. Variables were also entered into a multivariate logistic regression analysis (MVLRA) model with RBF as the dependent variable. P < 0.05 was considered statistically significant.
Results
Human studies
We found only three studies conducted in septic ICU patients in which RBF was measured [17-19]. The findings of these studies suggest an increase in RBF during sepsis (Table 1). In only one patient was renal plasma flow (RPF) determined in the setting of oliguric ARF [19]. Such RPF was markedly increased at 2000 ml/min (normal 650 ml/min).
Animal models
We found 159 [20-178] animal studies that measured RBF in sepsis. Of these, 99 (62%) reported a decrease, whereas the remaining 60 (38%) studies reported no change or an increase in RBF (Table 2, Fig. 1).
Animal size
Experimental studies were conducted in a large variety of animals. We divided experimental animals into small (rats, mice, rabbits and piglets) and large (dogs, pigs and sheep). We identified 65 (41%) studies that were conducted in small animals and 94 (59%) that were conducted in large animals (Table 2). Of studies conducted in small animals, 46 found decreased and 19 (29%) unchanged or increased RBF. In large animals, 53 (56%) studies reported a decreased and 41 (44%) an unchanged or increased RBF (P = 0.066; Fig. 2).
Technique for measuring renal blood flow
The techniques used for the measurement of RBF varied widely. Therefore, we compared studies using direct measurement of RBF via ultrasonic or electromagnetic flow probes ('direct' techniques) with measurement by microsphere technique or para-aminohippurate (PAH) clearance or other techniques such as measurement of blood velocity via video microscopy ('indirect' techniques). Of 80 studies using flow probes, 49 (61%) showed a decreased and 31 (39%) an unchanged or increased RBF (Table 2). Of 79 studies using other methods, 50 (63%) reported a decreased and 29 (37%) reported an unchanged or increased RBF (P = 0.791; Table 2, Fig. 2).
Consciousness of animals
The use of awake or unconscious animals might also have influenced RBF. For this reason, we compared studies using conscious with those using unconscious animals. Of 127 experiments conducted in unconscious animals (Table 2), 86 (68%) reported a decreased and 41 (32%) an unchanged or increased RBF. Of 32 studies conducted in conscious animals (Table 2), 13 (41 %) reported a decreased and 19 (59%) reported no change or an increase in RBF (P = 0.005; Fig. 1).
Recovery period between surgical preparation and actual experiment
Before conducting the experiments, a surgical procedure is typically needed to prepare the animals. We compared studies starting the experiment immediately after surgery with studies with a recovery period after anaesthesia. Of 33 studies with a recovery period (Table 2), 11 (33%) showed a decreased and 22 (67%) showed an unchanged or increased RBF. Of 126 studies without a recovery period (Table 2), 88 (70%) reported a decreased and 38 (30%) reported no change or an increase in RBF (P < 0.001; Fig. 1).
Time from septic insult
The duration of RBF observation after the septic insult varied widely. We divided the studies into those with a 'short' (<2 hours; early period after induction of sepsis) or 'long' (>2 hours; late period after the induction of sepsis) observation time. Among 47 experiments with short periods of observation after the induction of sepsis (Table 2), 32 (68%) showed a decreased and 15 (32%) showed an unchanged or increased RBF. Among the 112 experiments with long periods of observation after the induction of sepsis (Table 2), 67 (60%) showed a decreased and 45 (40%) showed an unchanged or increased RBF (P = 0.327; Fig. 2).
Methods of inducing sepsis
Many different methods of induction of sepsis were used. We compared LPS-induced sepsis with sepsis induced by injection of live bacteria or caecal ligation–perforation. Of 100 articles that used LPS (Table 2), 67 (67%) showed a decreased and 33 (33%) showed an unchanged or increased RBF. Among the other 59 studies (Table 2), 32 (54%) reported a reduced and 27 (46%) reported an unchanged or increased RBF (P = 0.109; Fig. 2).
Fluid administration
We compared studies according to whether there was fluid administration during the experiments. Thirty-four articles did not mention fluid administration. Among the 20 studies with no fluid administration (Table 2), 16 (80%) reported a decreased and 4 (20%) reported an unchanged or increased RBF. Of the 106 studies in which fluid was given (Table 2), 63 (59%) showed a decrease and 43 (41%) showed no change or an increase in RBF (P = 0.081; Fig. 2).
Haemodynamic patterns
Most septic patients exhibit a hyperdynamic state with elevated CO and decreased blood pressure, when CO is measured. Therefore, we compared studies in which animals had a hyperdynamic state (low peripheral vascular resistance [PVR]) of sepsis with studies in which this state was not present (normal or high PVR). There were 84 studies in which the hypodynamic versus hyperdynamic pattern could be assessed. Of 42 studies that fulfilled criteria for hypodynamic sepsis (Table 2), 38 (90%) showed a reduced and 4 (10%) showed no change or an increase in RBF. Of the 42 studies conducted in hyperdynamic sepsis (Table 2), 14 (33%) reported a decreased and 28 (67%) reported an unchanged or increased RBF (P < 0.001; Fig. 1).
Cardiac output
We compared those studies with increased or unchanged CO with studies with decreased CO. Some studies gave no indication of CO. Of the 51 studies with decreased CO (Table 2), 46 (90%) reported a decreased and 5 (10%) reported an unchanged or increased RBF. Among the 67 studies with an unchanged or increased CO (Table 2), 27 (40%) showed a reduced and 45 (60%) showed an unchanged or increased RBF (P < 0.001; Fig. 1).
Using MVLRA, we created a model to test for independent determinants of a RBF and found that only CO remained in the model (P < 0.001) as a significant predictor for RBF (Table 3).
Discussion
We interrogated two electronic databases to assess the changes that occur in RBF during human and experimental sepsis in order to examine what might be the determinants of sepsis-associated changes in RBF. Variables that might influence RBF were used to categorize the heterogeneous data we found.
We found only a few human studies reporting RBF in a septic setting and found that the techniques used to measure RBF had poor accuracy and reproducibility. Only in a single patient with septic oliguric ARF was RBF measured. Nonetheless, within these serious limitations, we found that an increase in RBF was typically seen during sepsis.
We found that most animal studies reported a decrease in RBF in sepsis. However, we found that, in one-third of studies, RBF was either maintained or increased. We also found contradictory and inconsistent experimental findings with regard to RBF, which appeared to be affected by factors other than the induction of sepsis itself, including the consciousness of the animal, the recovery time after surgery and the haemodynamic pattern (hypodynamic or hyperdynamic state). More importantly, using MVLRA, we found that all of the above factors could be reduced to the dominant effect of CO on RBF. Thus, a low CO predicted a decreased RBF and a preserved or high CO predicted an unchanged or increased RBF. These findings are complex and require detailed discussion.
Human studies
Currently, only invasive techniques for measuring RBF have a high degree of accuracy. They require renal vein sampling. Because of the risks associated with such invasive measurement of RBF, only a few such studies have been conducted in humans with sepsis. Noninvasive methods of measurement such as the PAH clearance method are also possible but they assume a constant PAH extraction ratio of 0.91, such that RPF can be calculated with measurement of PAH concentrations in blood and urine. Unfortunately, the 'constant' PAH extraction ratio is not at all constant, is markedly unstable and is influenced by many factors, all of which apply in sepsis and ARF [18,19]. Therefore, in order to achieve improved accuracy, this method must be made invasive by inserting a renal vein catheter in order to calculate the true PAH extraction ratio. The RPF measured by this method is called the true RPF. Finally, a third method uses a thermodilution renal vein catheter. RPF and RBF determined by the thermodilution method were reported to correlate with corrected PAH clearances (r = 0.79) [17].
However, a recently reported study [179] demonstrated that both methods have a low reproducibility and a within group error of up to 40%. Therefore, these methods are not sufficiently accurate to detect potentially important changes in RBF. Nonetheless, within the boundaries of the technology, true RPF measurements from human studies (Table 1) consistently suggest that renal blood flow is increased during human sepsis. In only one study [19] was RBF estimated in a septic patient with ARF. The RPF was found to be 2000 ml/min in this patient, which contrasts with the normal RPF in humans of 600–700 ml/min [180].
Animal models
Animal size
In small animals, RBFs values are very small (7.39 ml/min [40]). The changes estimated in different settings are even smaller (1.4 ml/min [40]). On the other hand, absolute blood flows in large animals are up to 250 times greater (330 ml/min [55]). We hypothesized that measurement accuracy might therefore change with animal size and lead to different observations. We found a strong trend in this direction, which just failed to achieve statistical significance.
Technique of measurement if renal blood flow
Using the flow probe technique, it is possible to measure the RBF continuously. Microsphere techniques are also accurate and can distinguish between cortical and medullar RBF, but using the latter technique it is only possible to take several 'snapshot views' of blood flow during the experiment. We hypothesized that the technique of measurement might have influenced findings. However, there was no significant difference between techniques.
Consciousness of animals
Most studies were conducted in unconscious animals. Within this group, RBF was significantly more likely to be decreased than in conscious animals. This effect might partly be explained by anaesthesia rather than sepsis itself. Our observations highlight this as an important area of concern in drawing conclusions about the effect of sepsis per se on RBF.
Time from septic insult
A recently published animal study [55] described the time-dependent development of hyperdynamic sepsis after live Escherichia coli injection. In that study the CO decreased immediately after injection, recovered and then increased by 2 hours until a hyperdynamic state was reached. Therefore, we divided the studies in experiments with less or greater than 2 hours of observation time after the septic insult in order to determine whether there were differences between early and late septic states. We hypothesized that studies with longer periods of observation after the insult (late sepsis) might show a different RBF. However, there was no difference between the two groups.
Recovery period
Surgical preparation was performed in many of the reviewed studies just before the experiments were started. The negative effect on RBF of immediately beginning the experiments after surgery might be explained by the prolonged anaesthesia time and the negative effect of anaesthesia. We found that lack of an adequate recovery period after surgical preparation increased the likelihood of RBF being decreased.
Method of inducing sepsis
Many different techniques are used to induce sepsis such as LPS injection, live bacteria injection and caecal ligation–perforation. Previous reports [181,182] described a strong hypodynamic effect of injecting a bolus of LPS. Therefore, we hypothesized that studies using LPS might show decreased RBF. We found a trend in this direction that approached statistical significance.
Fluid administration
Most of the studies administered fluid during the experiments to counteract the hypotensive of effect of sepsis [14]. These fluids might maintain CO, central venous pressure and blood pressure, and thus affect RBF. As might be expected, we found a strong trend toward a higher RBF when fluid resuscitation was given, but this failed to achieve statistical significance.
Haemodynamic patterns
In septic patients, CO, blood pressure and PVR can be assessed. Most of these patients have an increased CO, a low blood pressure and a decreased PVR [14,183-189]. To assess what might happen to RBF in a haemodynamic situation simulating human sepsis, we compared studies with animals that had developed hyperdynamic sepsis (increased CO and decreased blood pressure) with those studies with hypodynamic sepsis (normal or increased PVR). Animals with hyperdynamic sepsis were more likely to exhibit preserved or increased RBF.
Cardiac output
In a recently published article [190] using a crossover animal model, CO was found to be the most important variable influencing organ blood flows. Thus, we compared studies showing an unchanged or increased CO with studies showing a decreased CO. We found a clear association between decreased CO and decreased RBF and between a preserved or increased CO and a preserved or increased RBF. Multivariate logistic analysis confirmed the role of CO as the most powerful independent predictor of RBF in sepsis (Table 2).
Limitations
We only interrogated two English language electronic reference libraries and might have missed original contributions reported in other languages. However, we believe that it is unlikely that enough such studies would exist to change our conclusions materially.
In order to make comparisons, we categorized experiments according to pre-set criteria (small versus large animals, methods of induction of sepsis, high versus low CO, etc.) that we hypothesized, on grounds of biological plausibility, were likely to affect experimental findings. We acknowledge that such criteria are by definition arbitrary and the subject of individual judgement. Furthermore, other criteria that we did not consider could be tested. Nonetheless, we found that many of these criteria appeared to have some effect in reality. We also found that such effects appeared to be mostly related to their association with the CO state, which overwhelmingly was the only independent predictor in MVLRA for the outcome of RBF. We consider it unlikely that the choice of other criteria for comparison would materially affect our conclusions.
The observation time in the reviewed articles varied widely as well. We compared articles with a shorter period after the insult (2 hours) versus studies with a longer period of observation. We acknowledge that this division is artificial and might not truly reflect what happened, because some groups waited until the animal reached defined criteria before starting their observation time and others begun the observation immediately after the septic insult, making this variable extremely heterogeneous. Nonetheless, once again, given the overwhelming effect of CO on RBF, we consider that refinements to this criterion are unlikely to influence our conclusions.
Our observations suggest that the widely held paradigm that RBF decreases in sepsis [12-14] and that such a decrease is responsible for the development of ARF is indeed sustained by the majority of studies. However, the reality beyond such a simplistic observation is much more complex. The animal studies are extraordinarily heterogeneous in their design and monitoring of RBF. Furthermore, the support that the bulk of the data offer to the concept of decreased RBF in sepsis is conditional upon a particular model of sepsis being present (hypodynamic sepsis without an increase in CO). If the CO is increased and PVR is decreased, then the most common finding is actually one of increased or preserved RBF. In the light of this review, we suggest that measurement of CO is a vital component of all future experimental studies measuring RBF in sepsis.
We note that, in human sepsis, systemic vasodilatation with a high CO is the dominant clinical finding. Such vasodilatation might also affect the afferent and efferent arterioles of the kidney. If the efferent arteriole dilated proportionately more than the afferent arteriole, then there would be a decrease in glomerular filtration pressure. This change in filtration pressure would decrease glomerular filtration rate and lead to oliguria and loss of small solute clearance. Accordingly, loss of glomerular filtration rate can occur with either vasoconstriction or vasodilatation.
Our findings have important implications for clinicians and for future strategies directed at preserving renal function in sepsis. They highlight the absence of human data. They show the heterogeneity and model dependence of the animal data. They also emphasize the limitations of the indirect data upon which clinical strategies are based. Much research remains to be done if we are to establish what happens to renal blood flow in human sepsis, and techniques are needed that permit such measurements to be taken noninvasively.
Conclusion
We interrogated the two major English language electronic reference libraries to examine changes in RBF in sepsis and septic ARF. We found that inadequate data exist to allow any conclusions to be drawn on the typical RBF or changes in RBF in humans. We also found that experimental data are extraordinarily heterogeneous in nature but show the dominant effect of CO on RBF, such that a low CO predicts a decreased RBF and an increased or preserved CO predicts an increased or preserved RBF. Given that CO is typically increased when measured in human sepsis in the ICU, the widely held paradigm that decreased RBF is pivotal to the pathogenesis of septic ARF might require reassessment.
Key messages
• It is unknown whether RBF is increased, decreased, or unchanged in human sepsis.
• Techniques to measure RBF in humans are invasive and of limited accuracy.
• Data on RBF from animals are heterogenous and do not allow firm conclusions to be drawn.
• RBF findings in experimental sepsis depend on the model used.
• CO is the most important independent predictor of RBF in sepsis: if CO is increased, then RBF is typically increased; and if CO is decreased, the RBF is typically decreased.
Abbreviations
ARF = acute renal failure; CO = cardiac output; ICU = intensive care unit; LPS = lipopolysaccharide; MVLRA = multivariate logistic regression analysis; PAH = para-aminohippurate; PVR = peripheral vascular resistance; RBF = renal blood flow; RPF = renal plasma flow.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
CL conducted the searches and reviewed all necessary material, wrote the initial draft of the manuscript and performed statistical analysis. RB designed the study, critically reviewed the material and supervised the writing of the manuscript, CM co-designed the study and assisted with the completion of the manuscript. LW assisted with data assessment. ME assisted with data assessment and statistical analysis. SM assisted with study design and assessment, and completion of the manuscript.
Figures and Tables
Figure 1 Effect of variables on renal blood flow: statistically significant findings. All of the differences between the shaded areas are statistically significant (P < 0.05). CO, cardiac output; inc, increased; RBF, renal blood flow; unc, unchanged.
Figure 2 Effect of variables on renal blood flow: nonsignificant findings. None of the differences between and shaded areas are statistically significant. lps, lipopolysaccharide.
Table 1 Details of human studies conducted in septic patients measuring renal blood flow
Reference Measurement of PAH-RPF/true RPF (n/n) PAH-RPF (ml/min) True RPF (ml/min)
[17] 6 (0) - 690
[18] 40 (11) 475 1116
[19] 22 (6) 474 1238
PAH-RPF, renal plasma flow calculated using para-aminohippurate clearance with no renal vein sampling; true RPF, true renal plasma flow (flow calculated with renal vein sampling for PAH).
Table 2 References for studies reporting various findings pertaining to RBF in experimental sepsis
Finding/study characteristic Number of studies (%) References
Decrease in RBF 99 (62%) 20, 21, 23, 24, 26-29, 37-45, 49-54, 58-64, 68-70, 73, 74, 76, 78, 80, 83-86, 88, 90-95, 98-101, 103-107, 109, 110, 112, 113, 118-121, 123, 124, 126, 128-131, 134, 135, 140, 143-145, 149, 150, 152-157, 159, 160, 163, 165, 168, 169, 171-175 and 178
No change or a decrease in RBF 60 (38%) 22, 25, 30-36, 46-48, 55-57, 65-67, 71, 72, 75, 77, 79, 81, 82, 87, 89, 96, 97, 102, 108, 111, 114-117, 122, 125, 127, 132, 133, 136-139, 141, 142, 146-148, 151, 158, 161, 162, 164, 166, 167, 170, 176 and 177
Conducted in small animals (rats, mice, rabbits and piglets) 65 (41%) 20, 24, 27, 28, 38, 40, 43-45, 49, 50, 59, 61, 62, 65-67, 71-74, 77, 78, 82, 87, 88, 90, 92, 93, 99, 100, 102-105, 109, 110, 112, 115, 116, 119-121, 126, 128, 133, 138, 139, 141, 145, 149, 150, 152, 155-157, 159, 160 and 164-170
Conducted in largane animals (dogs, pigs and sheep) 94 (59%) 21-23, 25, 26, 29-37, 39, 41, 42, 46-48, 51-58, 60, 63, 64, 68-70, 75, 76, 79-81, 83-86, 89, 91, 94-98, 101, 106-108, 111, 113, 114, 117, 118, 122-125, 127, 129-132, 134-137, 140, 142-144, 146-148, 151, 153, 154, 158, 161-163 and 171-178
Measurement of RBF using flow probes 80 (50%) 21, 23-26, 28, 29, 32, 33, 36, 42, 43, 46, 53-58, 60, 63, 65, 66, 68-70, 75, 79, 81-85, 89, 91, 92, 95, 98, 101, 104-108, 110, 111, 113, 115, 116, 118, 122, 124-127, 129, 131, 132, 134-136, 142-144, 148, 153, 158-163 and 171-178
Measurement of RBF using other methods 79 (50%) 20, 22, 27, 30, 31, 34, 35, 37-41, 44, 45, 47-52, 59, 61, 62, 64, 67, 71-74, 76-78, 80, 86-88, 90, 93, 94, 96, 97, 99, 100, 102, 103, 109, 112, 114, 117, 119-121, 123, 128, 130, 133, 137-141, 145-147, 149-152, 154-157 and 164-170
Conducted in unconscious animals 127 (80%) 21-29, 37-44, 46, 53, 54, 57, 60-63, 65-75, 77-89, 91-101, 103-107, 109-111, 113, 115, 116, 118-121, 123-136, 138-160, 163, 164 and 166-178
Conducted in conscious animals 32 (20%) 20, 30-36, 45, 47-52, 55, 56, 58, 59, 64, 76, 90, 102, 108, 112, 114, 117, 122, 137, 161, 162 and 165
Conducted following a recovery period (after surgical preparation) 33 (21%) 30-36, 47-52, 55-59, 64, 68, 70, 76, 102, 108, 112, 114, 117, 122, 137, 161, 162, 166 and 170
Conducted with no recovery period 126 (79%) 20-29, 37-46, 53, 54, 60-63, 65-67, 69, 71-75, 77-101, 103-107, 109-111, 113, 115, 116, 118-121, 123-136, 138-160, 163-165, 167-169 and 171-178
Short period of observation following induction of sepsis (<2 hours) 47 (29%) 22, 26, 27, 40, 41, 47, 49, 50, 57, 59-61, 67, 70, 79, 80, 82, 86, 89, 92, 99, 100, 103, 105, 106, 109, 111, 117, 120, 121, 123, 124, 129, 130, 145-147, 149-151, 153, 154, 156, 158, 163, 164 and 167
Long period of observation following induction of sepsis (>2 hours) 112 (71%) 20, 21, 23-25, 28-39, 42-46, 48, 51-56, 58, 62-66, 68, 69, 71-78, 81, 83-85, 87, 88, 90, 91, 93-98, 101, 102, 104, 107, 108, 110, 112-116, 118, 119, 122, 125-128, 131-144, 148, 152, 155, 157, 159-162, 165, 166 and 168-178
Use of LPS to induce sepsis 100 (63%) 21, 23-26, 28, 29, 37, 39, 40, 42-46, 50, 54, 58-61, 63, 65, 66, 68-72, 76, 79, 80, 82, 86-97, 101, 103-106, 109-111, 114-118, 120-127, 129-136, 141-144, 147-150, 153-158, 160-164, 171, 172 and 174-178
Use of injection of live bacteria or caecal ligation–perforation to induce sepsis 59 (37%) 20, 22, 27, 30-36, 38, 41, 47-49, 51-53, 55-57, 62, 64, 67, 73-75, 77, 78, 81, 83-85, 98-100, 102, 107, 108, 112, 113, 119, 128, 137-140, 145, 146, 151, 152, 159, 165-170 and 173
Fluid administered during the experimenta 20 (13%) 22, 27, 61, 68, 69, 72, 77, 78, 83, 85, 91, 113, 118, 121, 130, 135, 136, 144, 145 and 150
Fluid not administered during the experimenta 106 (67%) 21, 23-26, 28-32, 34-41, 43-46, 48-52, 54-59, 62-66, 71, 73-76, 79, 80, 82, 84, 87, 90, 92-101, 103-105, 107, 108, 111, 112, 114-116, 119, 122-129, 131, 137-140, 143, 146-148, 151-153, 155, 157-159, 161, 162, 165-167, 169, 170, 173-176 and 178
Conducted in hypodynamic sepsisb 42 (26%) 37, 39, 42-44, 53, 54, 58, 61, 63, 68, 69, 80, 84, 86, 89, 98, 101, 103, 107, 113, 118, 120, 121, 127, 129, 132, 140, 144, 149, 151, 154-157, 165, 172-174 and 178
Conducted in hyperdynamic sepsisb 42 (26%) 20, 26, 30-36, 41, 46-48, 51, 55-57, 76-79, 81, 83, 96, 97, 100, 102, 105, 111, 117, 122, 123, 125, 131, 150, 153, 158, 161, 162 and 175-177
Decreased COc 51 (32%) 21, 25, 29, 37-39, 42-44, 53, 54, 58, 59, 61, 63, 68, 69, 80, 84, 86, 88, 89, 98, 101, 103, 107, 112, 113, 118, 120, 121, 127-130, 132, 140, 144, 149, 151, 154-157, 165, 168, 169, 172-174 and 178
Unchanged or decreased COc 67 (42%) 20, 26, 27, 30-36, 40, 41, 46-52, 55-57, 64, 73, 74, 76-79, 81, 83, 90, 96, 97, 99, 100, 102, 105, 108, 111, 114, 117, 119, 122, 123, 125, 131, 133, 137-139, 141, 145, 148, 150, 152, 153, 158, 161, 162, 164, 166, 167, 170 and 175-177]
aSome studies did not mention fluid administration. bIt was not possible to assess in some studies whether a septic hyperdynamic versus hypodynamic state was present. cSome studies gave no indication of CO. CO, cardiac output; LPS, lipopolysaccharide; RBF, renal blood flow.
Table 3 Multivariate logistic regression analysis of possible predictors of renal blood flow in experimental sepsis
Variable Regression coefficient 95% confidence interval P
Cardiac output 3.658 5.916–254.468 <0.001
Blood pressure -0.796 0.076–2.669 0.380
Recovery period 2.767 0.340–745.908 0.159
Consciousness -2.650 0.001–4.318 0.207
Fluid administration 2.066 0.543–114.722 0.130
Animal size 1.043 0.362–22.230 0.321
Technique measurement 0.608 0.390–8.666 0.442
Duration 1.496 0.849–23.482 0.077
Method of insult 0.501 0.374–7.284 0.508
==== Refs
Hou SH Bushinsky DA Wish JB Cohen JJ Harrington JT Hospital-acquired renal insufficiency: a prospective study Am J Med 1983 74 243 248 6824004 10.1016/0002-9343(83)90618-6
Nash K Hafeez A Hou S Hospital-acquired renal insufficiency Am J Kidney Dis 2002 39 930 936 11979336 10.1053/ajkd.2002.32766
Thadhani R Pascual M Bonventre JV Acute renal failure N Engl J Med 1996 334 1448 1460 8618585 10.1056/NEJM199605303342207
Brivet FG Kleinknecht DJ Loirat P Landais PJ Acute renal failure in intensive care units: causes, outcome, and prognostic factors of hospital mortality; a prospective, multicenter study. French Study Group on Acute Renal Failure Crit Care Med 1996 24 192 198 8605788 10.1097/00003246-199602000-00003
Jorres A Acute renal failure. Extracorporeal treatment strategies Minerva Med 2002 93 329 324 12410165
Liano F Junco E Pascual J Madero R Verde E The spectrum of acute renal failure in the intensive care unit compared with that seen in other settings. The Madrid Acute Renal Failure Study Group Kidney Int Suppl 1998 66 S16 S24 9580541
Silvester W Bellomo R Cole L Epidemiology, management, and outcome of severe acute renal failure of critical illness in Australia Crit Care Med 2001 29 1910 1915 11588450 10.1097/00003246-200110000-00010
Uchino S Doig GS Bellomo R Morimatsu H Morgera S Schetz M Tan I Bouman C Nacedo E Gibney N Diuretics and mortality in acute renal failure Crit Care Med 2004 32 1669 1677 15286542 10.1097/01.CCM.0000132892.51063.2F
Rangel-Frausto MS Pittet D Costigan M Hwang T Davis CS Wenzel RP The natural history of the systemic inflammatory response syndrome (SIRS). A prospective study JAMA 1995 273 117 123 7799491 10.1001/jama.273.2.117
Cole L Bellomo R Baldwin I Hayhoe M Ronco C The impact of lactate-buffered high-volume hemofiltration on acid-base balance Intensive Care Med 2003 29 1113 1120 12783161 10.1007/s00134-003-1812-1
Chew SL Lins RL Daelemans R De Broe ME Outcome in acute renal failure Nephrol Dial Transplant 1993 8 101 107 8384329
Badr KF Sepsis-associated renal vasoconstriction: potential targets for future therapy Am J Kidney Dis 1992 20 207 213 1519601
De Vriese AS Bourgeois M Pharmacologic treatment of acute renal failure in sepsis Curr Opin Crit Care 2003 9 474 480 14639066 10.1097/00075198-200312000-00003
Schrier RW Wang W Acute renal failure and sepsis N Engl J Med 2004 351 159 169 15247356 10.1056/NEJMra032401
Doig GS Simpson F Efficient literature searching: a core skill for the practice of evidence-based medicine Intensive Care Med 2003 29 2119 2127 12955188 10.1007/s00134-003-1942-5
Piper RD Cook DJ Bone RC Sibbald WJ Introducing Critical Appraisal to studies of animal models investigating novel therapies in sepsis Crit Care Med 1996 24 2059 2070 8968277 10.1097/00003246-199612000-00021
Brenner M Schaer GL Mallory DL Suffredini AF Parrillo JE Detection of renal blood flow abnormalities in septic and critically ill patients using a newly designed indwelling thermodilution renal vein catheter Chest 1990 98 170 179 2361386
Lucas CE Rector FE Werner M Rosenberg IK Altered renal homeostasis with acute sepsis. Clinical significance Arch Surg 1973 106 444 449 4696717
Rector F Goyal S Rosenberg IK Lucas CE Sepsis: a mechanism for vasodilatation in the kidney Ann Surg 1973 178 222 226 4723431
Alden KJ Motew SJ Sharma AC Ferguson JL Effect of aminoguanidine on plasma nitric oxide by-products and blood flow during chronic peritoneal sepsis Shock 1998 9 289 295 9565258
Aranow JS Wang H Zhuang J Fink MP Effect of human hemoglobin on systemic and regional hemodynamics in a porcine model of endotoxemic shock Crit Care Med 1996 24 807 814 8706458 10.1097/00003246-199605000-00014
Auguste LJ Stone AM Wise L The effects of Escherichia coli bacteremia on in vitro perfused kidneys Ann Surg 1980 192 65 68 6996623
Beck RR Abel FL Papadakis E Influence of ibuprofen on renal function in acutely endotoxemic dogs Circ Shock 1989 28 37 47 2731320
Begany DP Carcillo JA Herzer WA Mi Z Jackson EK Inhibition of type IV phosphodiesterase by Ro 20-1724 attenuates endotoxin-induced acute renal failure J Pharmacol Exp Ther 1996 278 37 41 8764333
Bellomo R Kellum JA Pinsky MR Transvisceral lactate fluxes during early endotoxemia Chest 1996 110 198 204 8681628
Bellomo R Kellum JA Wisniewski SR Pinsky MR Effects of norepinephrine on the renal vasculature in normal and endotoxemic dogs Am J Respir Crit Care Med 1999 159 1186 1192 10194164
Bloom IT Bentley FR Garrison RN Escherichia coli bacteremia exacerbates cyclosporine-induced renal vasoconstriction J Surg Res 1993 54 510 516 8361177 10.1006/jsre.1993.1079
Boffa JJ Just A Coffman TM Arendshorst WJ Thromboxane receptor mediates renal vasoconstriction and contributes to acute renal failure in endotoxemic mice J Am Soc Nephrol 2004 15 2358 2365 15339984 10.1097/01.ASN.0000136300.72480.86
Bond RF Peripheral vascular adrenergic depression during hypotension induced by E coli endotoxin Adv Shock Res 1983 9 157 169 6880967
Bone HG Fischer SR Schenarts PJ McGuire R Traber LD Traber DL Continuous infusion of pyridoxalated hemoglobin polyoxyethylene conjugate in hyperdynamic septic sheep Shock 1998 10 69 76 9688094
Bone HG Schenarts PJ Fischer SR McGuire R Traber LD Traber DL Pyridoxalated hemoglobin polyoxyethylene conjugate reverses hyperdynamic circulation in septic sheep J Appl Physiol 1998 84 1991 1999 9609794
Booke M Hinder F McGuire R Traber LD Traber DL Nitric oxide synthase inhibition versus norepinephrine for the treatment of hyperdynamic sepsis in sheep Crit Care Med 1996 24 835 844 8706462 10.1097/00003246-199605000-00018
Booke M Armstrong C Hinder F Conroy B Traber LD Traber DL The effects of propofol on hemodynamics and renal blood flow in healthy and in septic sheep, and combined with fentanyl in septic sheep Anesth Analg 1996 82 738 743 8615490 10.1097/00000539-199604000-00011
Booke M Hinder F McGuire R Traber LD Traber DL Nitric oxide synthase inhibition versus norepinephrine in ovine sepsis: effects on regional blood flow Shock 1996 5 362 370 9156793
Booke M Hinder F McGuire R Traber LD Traber DL Selective inhibition of inducible nitric oxide synthase: effects on hemodynamics and regional blood flow in healthy and septic sheep Crit Care Med 1999 27 162 167 9934911 10.1097/00003246-199901000-00045
Booke M Hinder F McGuire R Traber LD Traber DL Noradrenaline and nomega-monomethyl-L-arginine (L-NMMA): effects on haemodynamics and regional blood flow in healthy and septic sheep Clin Sci (Lond) 2000 98 193 200 10657275
Breslow MJ Miller CF Parker SD Walman AT Traystman RJ Effect of vasopressors on organ blood flow during endotoxin shock in pigs Am J Physiol 1987 252 H291 300 3544876
Bressack MA Morton NS Hortop J Group B streptococcal sepsis in the piglet: effects of fluid therapy on venous return, organ edema, and organ blood flow Circ Res 1987 61 659 669 3311447
Bronsveld W van Lambalgen AA van den Bos GC Thijs LG Koopman PA Regional blood flow and metabolism in canine endotoxin shock before, during, and after infusion of glucose-insulin-potassium (GIK) Circ Shock 1986 18 31 42 3510757
Burnier M Waeber B Aubert JF Nussberger J Brunner HR Effects of nonhypotensive endotoxemia in conscious rats: role of prostaglandins Am J Physiol 1988 254 H509 H516 3126671
Carroll GC Snyder JV Hyperdynamic severe intravascular sepsis depends on fluid administration in cynomolgus monkey Am J Physiol 1982 243 R131 R141 7046476
Cavanagh D Rao PS Sutton DM Bhagat BD Bachmann F Pathophysiology of endotoxin shock in the primate Am J Obstet Gynecol 1970 108 705 722 4990505
Cheng X Pang CC Pressor and vasoconstrictor effects of methylene blue in endotoxaemic rats Naunyn Schmiedebergs Arch Pharmacol 1998 357 648 653 9686941
Chin A Radhakrishnan J Fornell L John E Effects of tezosentan, a dual endothelin receptor antagonist, on the cardiovascular and renal systems of neonatal piglets during endotoxic shock J Pediatr Surg 2002 37 482 487 11877672 10.1053/jpsu.2002.30871
Chou DE Cai H Jayadevappa D Porush JG Regional expression of inducible nitric oxide synthase in the kidney stimulated by lipopolysaccharide in the rat Exp Physiol 2002 87 153 162 11856960 10.1113/eph8702289
Cohen RI Hassell AM Marzouk K Marini C Liu SF Scharf SM Renal effects of nitric oxide in endotoxemia Am J Respir Crit Care Med 2001 164 1890 1895 11734442
Cronenwett JL Lindenauer SM Distribution of intrarenal blood flow during bacterial sepsis J Surg Res 1978 24 132 141 633878 10.1016/0022-4804(78)90165-8
Cronenwett JL Lindenauer SM Hemodynamic effects of cecal ligation sepsis in dogs J Surg Res 1982 33 324 331 7120985 10.1016/0022-4804(82)90045-2
Cryer HM Unger LS Garrison RN Harris PD Prostaglandins maintain renal microvascular blood flow during hyperdynamic bacteremia Circ Shock 1988 26 71 88 3056633
Cryer HG Bloom IT Unger LS Garrison RN Factors affecting renal microvascular blood flow in rat hyperdynamic bacteremia Am J Physiol 1993 264 H1988 H1997 8322929
Cumming AD Kline R Linton AL Association between renal and sympathetic responses to nonhypotensive systemic sepsis Crit Care Med 1988 16 1132 1137 3168506
Cumming AD Driedger AA McDonald JW Lindsay RM Solez K Linton AL Vasoactive hormones in the renal response to systemic sepsis Am J Kidney Dis 1988 11 23 32 3276170
Dedichen H Schenk WG Jr Hemodynamics of endotoxin shock in the dog Arch Surg 1967 95 1013 1016 6058789
Dedichen H Hemodynamic changes in experimental endotoxin shock Acta Chir Scand 1972 138 215 225 5036818
Di Giantomasso D May CN Bellomo R Vital organ blood flow during hyperdynamic sepsis Chest 2003 124 1053 1059 12970037 10.1378/chest.124.3.1053
Di Giantomasso D May CN Bellomo R Norepinephrine and vital organ blood flow during experimental hyperdynamic sepsis Intensive Care Med 2003 29 1774 1781 12698246 10.1007/s00134-003-1736-9
Di Giantomasso D Morimatsu H May CN Bellomo R Intrarenal blood flow distribution in hyperdynamic septic shock: effect of norepinephrine Crit Care Med 2003 31 2509 2513 14530759 10.1097/01.CCM.0000084842.66153.5A
Doursout MF Kilbourn RG Hartley CJ Chelly JE Effects of N-methyl-L-arginine on cardiac and regional blood flow in a dog endotoxin shock model J Crit Care 2000 15 22 29 10757195 10.1053/jcrc.2000.0150022
Dziki AJ Lynch WH Ramsey CB Law WR Beta-adrenergic-dependent and -independent actions of naloxone on perfusion during endotoxin shock Circ Shock 1993 39 29 38 8386986
Emerson TE JrWagner TD Gill CC In situ kidney pressure-flow-resistance relationship during endotoxin shock in dogs Proc Soc Exp Biol Med 1966 122 366 368 5342688
Etemadi AR Tempel GE Farah BA Wise WC Halushka PV Cook JA Beneficial effects of a leukotriene antagonist on endotoxin-induced acute hemodynamic alterations Circ Shock 1987 22 55 63 3038364
Fantini GA Shiono S Bal BS Shires GT Adrenergic mechanisms contribute to alterations in regional perfusion during normotensive E. coli bacteremia J Trauma 1989 29 1252 1257 2671401
Fink MP Nelson R Roethel R Low-dose dopamine preserves renal blood flow in endotoxin shocked dogs treated with ibuprofen J Surg Res 1985 38 582 591 3839278 10.1016/0022-4804(85)90079-4
Fox GA Lam CJ Darragh WB Neal AM Inman KJ Rutledge FS Sibbald WJ Circulatory sequelae of administering CPAP in hyperdynamic sepsis are time dependent J Appl Physiol 1996 81 976 984 8872670
Gardiner SM Kemp PA March JE Bennett T Influence of aminoguanidine and the endothelin antagonist, SB 209670, on the regional haemodynamic effects of endotoxaemia in conscious rats Br J Pharmacol 1996 118 1822 1828 8842449
Gardiner SM Kemp PA March JE Bennett T Regional haemodynamic responses to infusion of lipopolysaccharide in conscious rats: effects of pre- or post-treatment with glibenclamide Br J Pharmacol 1999 128 1772 1778 10588933 10.1038/sj.bjp.0702985
Garrison RN Wilson MA Matheson PJ Spain DA Nitric oxide mediates redistribution of intrarenal blood flow during bacteremia J Trauma 1995 39 90 96 discussion 96-97. 7636915
Gullichsen E Nelimarkka O Halkola L Niinikoski J Renal oxygenation in endotoxin shock in dogs Crit Care Med 1989 17 547 550 2721213
Gullichsen E Renal perfusion and metabolism in experimental endotoxin shock Acta Chir Scand Suppl 1991 560 7 31 1828126
Guntheroth WG Hougen T Kaplan EL Absence of pooling with endotoxin shock in the canine kidney J Appl Physiol 1972 32 512 515 4554677
Hallemeesch MM Cobben DC Dejong CH Soeters PB Deutz NE Renal amino acid metabolism during endotoxemia in the rat J Surg Res 2000 92 193 200 10896821 10.1006/jsre.2000.5867
Hallemeesch MM Soeters PB Deutz NE Renal arginine and protein synthesis are increased during early endotoxemia in mice Am J Physiol Renal Physiol 2002 282 F316 F323 11788446
Haybron DM Townsend MC Hampton WW Schirmer JM Fry DE Effective renal blood flow and renal energy charge in murine peritonitis Am Surg 1986 52 625 629 3789539
Haybron DM Townsend MC Hampton WW Schirmer WJ Schirmer JM Fry DE Alterations in renal perfusion and renal energy charge in murine peritonitis Arch Surg 1987 122 328 331 3827574
Haywood GA Tighe D Moss R al-Saady N Foshola TO Riley SP Pearson I Webb A McKenna WJ Bennett ED Goal directed therapy with dobutamine in a porcine model of septic shock: effects on systemic and renal oxygen transport Postgrad Med J 1991 S36 S39 discussion S40-S41 1924077
Hazelzet JA Stubenitsky R Petrov AB van Wieringen GW van der Voort E Hess J Hop WC Thijs LG Duncker DJ Poolman JT Cardiovascular aspects of experimental meningococcal sepsis in young and older awake piglets: age-related differences Shock 1999 12 145 154 10446896
Heemskerk AE Huisman E van Lambalgen AA Appelmelk BJ van den Bos GC Thijs LG Tangelder GJ Gram-negative shock in rats depends on the presence of capsulated bacteria and is modified by laparotomy Shock 1996 6 418 425 8961392
Heemskerk AE Huisman E van Lambalgen AA van den Bos GC Hennekes M Thijs LG Tangelder GJ Renal function and oxygen consumption during bacteraemia and endotoxaemia in rats Nephrol Dial Transplant 1997 12 1586 1594 9269634 10.1093/ndt/12.8.1586
Henderson JL Statman R Cunningham JN Cheng W Damiani P Siconolfi A Horovitz JH The effects of nitric oxide inhibition on regional hemodynamics during hyperdynamic endotoxemia Arch Surg 1994 129 1271 1274 discussion 1275 7527209
Henrich WL Hamasaki Y Said SI Campbell WB Cronin RE Dissociation of systemic and renal effects in endotoxemia. Prostaglandin inhibition uncovers an important role of renal nerves J Clin Invest 1982 69 691 699 7037854
Hermreck AS Thal AP Mechanisms for the high circulatory requirements in sepsis and septic shock Ann Surg 1969 170 677 695 4898969
Heyman SN Darmon D Goldfarb M Bitz H Shina A Rosen S Brezis M Endotoxin-induced renal failure. I. A role for altered renal microcirculation Exp Nephrol 2000 8 266 274 10940726 10.1159/000020678
Hiltebrand LB Krejci V Banic A Erni D Wheatley AM Sigurdsson GH Dynamic study of the distribution of microcirculatory blood flow in multiple splanchnic organs in septic shock Crit Care Med 2000 28 3233 3241 11008987 10.1097/00003246-200009000-00019
Hiltebrand LB Krejci V Sigurdsson GH Effects of dopamine, dobutamine, and dopexamine on microcirculatory blood flow in the gastrointestinal tract during sepsis and anesthesia Anesthesiology 2004 100 1188 1197 15114217 10.1097/00000542-200405000-00022
Hinshaw LB Solomon LA Holmes DD Greenfield LJ Comparison of canine responses to Escherichia coli organisms and endotoxin Surg Gynecol Obstet 1968 127 981 988 4878684
Hussain SN Roussos C Distribution of respiratory muscle and organ blood flow during endotoxic shock in dogs J Appl Physiol 1985 59 1802 1808 4077788
Jepson MM Cox M Bates PC Rothwell NJ Stock MJ Cady EB Millward DJ Regional blood flow and skeletal muscle energy status in endotoxemic rats Am J Physiol 1987 252 E581 E587 3555111
Keeler R Barrientos A Lee K Circulatory effects of acute or chronic endotoxemia in rats Can J Physiol Pharmacol 1981 59 204 208 7013954
Kellum JA Bellomo R Kramer DJ Pinsky MR Hepatic anion flux during acute endotoxemia J Appl Physiol 1995 78 2212 2217 7665420
Kikeri D Pennell JP Hwang KH Jacob AI Richman AV Bourgoignie JJ Endotoxemic acute renal failure in awake rats Am J Physiol 1986 250 F1098 F1106 3521325
Kirkebo A Tyssebotn I Renal blood flow distribution during E. coli endotoxin shock in dog Acta Physiol Scand 1980 108 367 372 6998257
Knight RJ Bowmer CJ Yates MS Effect of the selective A1 adenosine antagonist 8-cyclopentyl-1,3-dipropylxanthine on acute renal dysfunction induced by Escherichia coli endotoxin in rats J Pharm Pharmacol 1993 45 979 984 7908041
Knotek M Rogachev B Wang W Ecder T Melnikov V Gengaro PE Esson M Edelstein CL Dinarello CA Schrier RW Endotoxemic renal failure in mice: role of tumor necrosis factor independent of inducible nitric oxide synthase Kidney Int 2001 59 2243 2249 11380827
Knuth OE Wagenknecht LV Madsen PO The effect of various treatments on renal function during endotoxin shock. An experimental study in dogs Invest Urol 1972 9 304 309 5058769
Koyama S Participation of central alpha-receptors on hemodynamic response to E. coli endotoxin Am J Physiol 1984 247 R655 R662 6093560
Kreimeier U Hammersen F Ruiz-Morales M Yang Z Messmer K Redistribution of intraorgan blood flow in acute, hyperdynamic porcine endotoxemia Eur Surg Res 1991 23 85 99 1936082
Kreimeier U Brueckner UB Gerspach S Veitinger K Messmer K A porcine model of hyperdynamic endotoxemia: pattern of respiratory, macrocirculatory, and regional blood flow changes J Invest Surg 1993 6 143 156 8512888
Krejci V Hiltebrand LB Erni D Sigurdsson GH Endothelin receptor antagonist bosentan improves microcirculatory blood flow in splanchnic organs in septic shock Crit Care Med 2003 31 203 210 12545016 10.1097/00003246-200301000-00031
Krysztopik RJ Bentley FR Spain DA Wilson MA Garrison RN Free radical scavenging by lazaroids improves renal blood flow during sepsis Surgery 1996 120 657 662 8862374
Krysztopik RJ Bentley FR Wilson MA Spain DA Garrison RN Vasomotor response to pentoxifylline mediates improved renal blood flow to bacteremia J Surg Res 1996 63 17 22 8661165 10.1006/jsre.1996.0215
Laesser M Oi Y Ewert S Fandriks L Aneman A The angiotensin II receptor blocker candesartan improves survival and mesenteric perfusion in an acute porcine endotoxin model Acta Anaesthesiol Scand 2004 48 198 204 14995942 10.1111/j.0001-5172.2004.00283.x
Lang CH Bagby GJ Ferguson JL Spitzer JJ Cardiac output and redistribution of organ blood flow in hypermetabolic sepsis Am J Physiol 1984 246 R331 R337 6703086
Law WR Ferguson JL Naloxone alters organ perfusion during endotoxin shock in conscious rats Am J Physiol 1988 255 H1106 H1113 3189572
Levy B Mansart A Bollaert PE Franck P Mallie JP Effects of epinephrine and norepinephrine on hemodynamics, oxidative metabolism, and organ energetics in endotoxemic rats Intensive Care Med 2003 29 292 300 12594589
Levy B Vallee C Lauzier F Plante GE Mansart A Mallie JP Lesur O Comparative effects of vasopressin, norepinephrine, and L-canavanine, a selective inhibitor of inducible nitric oxide synthase, in endotoxic shock Am J Physiol Heart Circ Physiol 2004 287 H209 H215 14988074 10.1152/ajpheart.00946.2003
Linder MM Hartel W Alken P Muschaweck R Renal tissue oxygen tension during the early phase of canine endotoxin shock Surg Gynecol Obstet 1974 138 171 173 4589904
Lindgren S Almqvist P Arvidsson D Montgomery A Andersson KE Haglund U Lack of beneficial effects of milrinone in severe septic shock Circ Shock 1990 31 365 375 2204496
Lingnau W McGuire R Booke M Traber LD Traber DL Effects of alpha-trinositol on systemic inflammation and renal function in ovine bacterial sepsis Shock 1997 8 179 185 9377164
Lugon JR Boim MA Ramos OL Ajzen H Schor N Renal function and glomerular hemodynamics in male endotoxemic rats Kidney Int 1989 36 570 575 2681930
Maitra SR Homan CS Pan W Geller ER Henry MC Thode HC Jr Renal gluconeogenesis and blood flow during endotoxic shock Acad Emerg Med 1996 3 1006 1010 8922005
Malay MB Ashton JL Dahl K Savage EB Burchell SA Ashton RC JrSciacca RR Oliver JA Landry DW Heterogeneity of the vasoconstrictor effect of vasopressin in septic shock Crit Care Med 2004 32 1327 1331 15187515 10.1097/01.CCM.0000128578.37822.F1
Martin CM Sibbald WJ Modulation of hemodynamics and organ blood flow by nitric oxide synthase inhibition is not altered in normotensive, septic rats Am J Respir Crit Care Med 1994 150 1539 1544 7524982
Mellins RB Levine OR Wigger HJ Leidy G Curnen EC Experimental menigococcemia: model of overwhelming infection in unanesthetized monkeys J Appl Physiol 1972 32 309 314 5010040
Meyer J Hinder F Stothert J JrTraber LD Herndon DN Flynn JT Traber DL Increased organ blood flow in chronic endotoxemia is reversed by nitric oxide synthase inhibition J Appl Physiol 1994 76 2785 2793 7523359
Millar CG Thiemermann C Intrarenal haemodynamics and renal dysfunction in endotoxaemia: effects of nitric oxide synthase inhibition Br J Pharmacol 1997 121 1824 1830 9283724
Millar CG Thiemermann C Carboxy-PTIO, a scavenger of nitric oxide, selectively inhibits the increase in medullary perfusion and improves renal function in endotoxemia Shock 2002 18 64 68 12095136 10.1097/00024382-200207000-00012
Miller RL Forsyth RP Hoffbrand BI Melmon KL Cardiovascular effects of hemorrhage during endotoxemia in unanesthetized monkeys Am J Physiol 1973 224 1087 1091 4633704
Mitaka C Hirata Y Yokoyama K Nagura T Tsunoda Y Amaha K Improvement of renal dysfunction in dogs with endotoxemia by a nonselective endothelin receptor antagonist Crit Care Med 1999 27 146 153 9934909 10.1097/00003246-199901000-00043
Morisaki H Sibbald W Martin C Doig G Inman K Hyperdynamic sepsis depresses circulatory compensation to normovolemic anemia in conscious rats J Appl Physiol 1996 80 656 664 8929612
Mulder MF van Lambalgen AA van Kraats AA Scheffer PG Bouman AA van den Bos GC Thijs LG Systemic and regional hemodynamic changes during endotoxin or platelet activating factor (PAF)-induced shock in rats Circ Shock 1993 41 221 229 8143350
Mulder MF van Lambalgen AA van den Bos GC Thijs LG The fall of cardiac output in endotoxemic rats cannot explain all changes in organ blood flow: a comparison between endotoxin and low venous return shock Shock 1996 5 135 140 8705391
Murphey ED Traber DL Cardiopulmonary and splanchnic blood flow during 48 hours of a continuous infusion of endotoxin in conscious pigs: a model of hyperdynamic shock Shock 2000 13 224 229 10718380
Nishijima MK Breslow MJ Miller CF Traystman RJ Effect of naloxone and ibuprofen on organ blood flow during endotoxic shock in pig Am J Physiol 1988 255 H177 H184 3394818
Nishiyama A Miura K Miyatake A Fujisawa Y Yue W Fukui T Kimura S Abe Y Renal interstitial concentration of adenosine during endotoxin shock Eur J Pharmacol 1999 385 209 216 10607878 10.1016/S0014-2999(99)00716-5
Offner PJ Robertson FM Pruitt BA Jr Effects of nitric oxide synthase inhibition on regional blood flow in a porcine model of endotoxic shock J Trauma 1995 39 338 343 7545764
O'Hair DP Adams MB Tunberg TC Osborn JL Relationships among endotoxemia, arterial pressure, and renal function in dogs Circ Shock 1989 27 199 210 2650915
Oldner A Konrad D Weitzberg E Rudehill A Rossi P Wanecek M Effects of levosimendan, a novel inotropic calcium-sensitizing drug, in experimental septic shock Crit Care Med 2001 29 2185 2193 11700420 10.1097/00003246-200111000-00022
Ottosson J Dawidson I Brandberg A Idvall J Sandor Z Cardiac output and organ blood flow in experimental septic shock: effect of treatment with antibiotics, corticosteroids, and fluid infusion Circ Shock 1991 35 14 24 1742857
Oyama T Toyooka K Sato Y Kondo S Kudo T Effect of endotoxic shock on renal and hormonal functions Can Anaesth Soc J 1978 25 380 391 29700
Passmore JC Neiberger RE Eden SW Measurement of intrarenal anatomic distribution of krypton-85 in endotoxic shock in dogs Am J Physiol 1977 232 H54 H58 835721
Pastor CM Vascular hyporesponsiveness of the renal circulation during endotoxemia in anesthetized pigs Crit Care Med 1999 27 2735 2740 10628619 10.1097/00003246-199912000-00022
Preiser JC Sun Q Hadj-Sadok D Vincent JL Differential effects of a selective inhibitor of soluble guanylyl cyclase on global and regional hemodynamics during canine endotoxic shock Shock 2003 20 465 468 14560112 10.1097/01.shk.0000092267.01859.e9
Prins HA Houdijk AP Wiezer MJ Teerlink T van Lambalgen AA Thijs LG van Leeuwen PA The effect of mild endotoxemia during low arginine plasma levels on organ blood flow in rats Crit Care Med 2000 28 1991 1997 10890653 10.1097/00003246-200006000-00051
Rao PS Bhagat B Effect of dopamine on renal blood flow of baboon in endotoxin shock Pflugers Arch 1978 374 105 106 98752 10.1007/BF00585704
Rao PS Cavanagh D Marsden KA Knuppel RA Spaziani E Prostaglandin D2 in canine endotoxic shock. Hemodynamic, hematologic, biochemical, and blood gas analyses Am J Obstet Gynecol 1984 148 964 972 6424477
Rao PS Cavanagh DM Fiorica JV Spaziani E Endotoxin-induced alterations in renal function with particular reference to tubular enzyme activity Circ Shock 1990 31 333 342 2357774
Raper RF Sibbald WJ Hobson J Rutledge FS Effect of PGE1 on altered distribution of regional blood flows in hyperdynamic sepsis Chest 1991 100 1703 1711 1959417
Ravikant T Lucas CE Renal blood flow distribution in septic hyperdynamic pigs J Surg Res 1977 22 294 298 839776 10.1016/0022-4804(77)90146-9
Sam AD IISharma AC Rice AN Ferguson JL Law WR Adenosine and nitric oxide regulate regional vascular resistance via interdependent and independent mechanisms during sepsis Crit Care Med 2000 28 1931 1939 10890644 10.1097/00003246-200006000-00041
Schaer GL Fink MP Chernow B Ahmed S Parrillo JE Renal hemodynamics and prostaglandin E2 excretion in a nonhuman primate model of septic shock Crit Care Med 1990 18 52 59 2403507
Schirmer WJ Schirmer JM Naff GB Fry DE Systemic complement activation produces hemodynamic changes characteristic of sepsis Arch Surg 1988 123 316 321 3341912
Selmyer JP Reynolds DG Swan KG Renal blood flow during endotoxin shock in the subhuman primate Surg Gynecol Obstet 1973 137 2 6 4197452
Shanbour LL Lindeman RD Archer LT Tung SH Hinshaw LB Improvement of renal hemodynamics in endotoxin shock with dopamine, phenoxybenzamine and dextran J Pharmacol Exp Ther 1971 176 383 388 5568782
Somani P Saini RK A comparison of the cardiovascular, renal, and coronary effects of dopamine and monensin in endotoxic shock Circ Shock 1981 8 451 464 7023738
Spain DA Wilson MA Garrison RN Nitric oxide synthase inhibition exacerbates sepsis-induced renal hypoperfusion Surgery 1994 116 322 330 discussion 330-321 7519364
Stone AM Stein T LaFortune J Wise L Effect of steroids on the renovascular changes of sepsis J Surg Res 1979 26 565 569 439888 10.1016/0022-4804(79)90051-9
Stone AM Stein T LaFortune J Wise L Changes in intrarenal blood flow during sepsis Surg Gynecol Obstet 1979 148 731 734 432786
Tanigawa K Bellomo R Kellum JA Kim YM Zar H Lancaster JR JrPinsky MR Ondulick B Nitric oxide metabolism in canine sepsis: relation to regional blood flow J Crit Care 1999 14 186 190 10622753 10.1016/S0883-9441(99)90033-3
Tempel GE Cook JA Wise WC Halushka PV The improvement in endotoxin-induced redistribution of organ blood flow by inhibition of thromboxane and prostaglandin synthesis Adv Shock Res 1982 7 209 218 6753535
Tempel GE Cook JA Wise WC Halushka PV Corral D Improvement in organ blood flow by inhibition of thromboxane synthetase during experimental endotoxic shock in the rat J Cardiovasc Pharmacol 1986 8 514 519 2425166
Tindel MS Stone AM Stein TA Wise L Effect of steroids on the cardiac output and renal changes of bacteremia Am Surg 1985 51 716 720 3907448
Townsend MC Hampton WW Haybron DM Schirmer WJ Fry DE Effective organ blood flow and bioenergy status in murine peritonitis Surgery 1986 100 205 213 3738752
Treggiari MM Romand JA Burgener D Suter PM Aneman A Effect of increasing norepinephrine dosage on regional blood flow in a porcine model of endotoxin shock Crit Care Med 2002 30 1334 1339 12072691 10.1097/00003246-200206000-00032
van Lambalgen AA Runge HC van den Bos GC Thijs LG Regional lactate production in early canine endotoxin shock Am J Physiol 1988 254 E45 E51 3337225
van Lambalgen AA van Kraats AA van den Bos GC Stel HV Straub J Donker AJ Thijs LG Renal function and metabolism during endotoxemia in rats: role of hypoperfusion Circ Shock 1991 35 164 173 1777955
van Lambalgen AA van Kraats AA van den Bos GC Teerlink T Stel HV Donker AJ Thijs LG Development of renal failure in endotoxemic rats: can it be explained by early changes in renal energy metabolism? Nephron 1993 65 88 94 8413798
van Lambalgen AA van Kraats AA Mulder MF van den Bos GC Teerlink T Thijs LG Organ blood flow and distribution of cardiac output in dopexamine- or dobutamine-treated endotoxemic rats J Crit Care 1993 8 117 127 8102078 10.1016/0883-9441(93)90016-E
Wanecek M Rudehill A Hemsen A Lundberg JM Weitzberg E The endothelin receptor antagonist, bosentan, in combination with the cyclooxygenase inhibitor, diclofenac, counteracts pulmonary hypertension in porcine endotoxin shock Crit Care Med 1997 25 848 857 9187606 10.1097/00003246-199705000-00022
Wang P Zhou M Rana MW Ba ZF Chaudry IH Differential alterations in microvascular perfusion in various organs during early and late sepsis Am J Physiol 1992 263 G38 G43 1636714
Wang W Jittikanont S Falk SA Li P Feng L Gengaro PE Poole BD Bowler RP Day BJ Crapo JD Interaction among nitric oxide, reactive oxygen species, and antioxidants during endotoxemia-related acute renal failure Am J Physiol Renal Physiol 2003 284 F532 F537 12556364
Weber A Schwieger IM Poinsot O Klohn M Gaumann DM Morel DR Sequential changes in renal oxygen consumption and sodium transport during hyperdynamic sepsis in sheep Am J Physiol 1992 262 F965 F971 1621820
Weber A Schwieger IM Poinsot O Morel DR Time course of systemic and renal plasma prostanoid concentrations and renal function in ovine hyperdynamic sepsis Clin Sci (Lond) 1994 86 599 610 8033513
White FN Gold EM Vaughn DL Renin-aldosterone system in endotoxin shock in the dog Am J Physiol 1967 212 1195 1198 4290363
Xu D Qi L Guillory D Cruz N Berg R Deitch EA Mechanisms of endotoxin-induced intestinal injury in a hyperdynamic model of sepsis J Trauma 1993 34 676 682 discussion 682-683. 8497002
Yang S Zhou M Koo DJ Chaudry IH Wang P Pentoxifylline prevents the transition from the hyperdynamic to hypodynamic response during sepsis Am J Physiol 1999 277 H1036 H1044 10484426
Yang S Cioffi WG Bland KI Chaudry IH Wang P Differential alterations in systemic and regional oxygen delivery and consumption during the early and late stages of sepsis J Trauma 1999 47 706 712 10528605
Yang S Koo DJ Chaudry IH Wang P The important role of the gut in initiating the hyperdynamic response during early sepsis J Surg Res 2000 89 31 37 10720450 10.1006/jsre.1999.5807
Yang S Zhou M Fowler DE Wang P Mechanisms of the beneficial effect of adrenomedullin and adrenomedullin-binding protein-1 in sepsis: down-regulation of proinflammatory cytokines Crit Care Med 2002 30 2729 2735 12483065 10.1097/00003246-200212000-00018
Yang S Zhou M Chaudry IH Wang P Novel approach to prevent the transition from the hyperdynamic phase to the hypodynamic phase of sepsis: role of adrenomedullin and adrenomedullin binding protein-1 Ann Surg 2002 236 625 633 12409669 10.1097/00000658-200211000-00013
Yang S Chung CS Ayala A Chaudry IH Wang P Differential alterations in cardiovascular responses during the progression of polymicrobial sepsis in the mouse Shock 2002 17 55 60 11795670 10.1097/00024382-200201000-00010
Yao K Ina Y Nagashima K Ohno T Karasawa A Effect of the selective adenosine A1-receptor antagonist KW-3902 on lipopolysaccharide-induced reductions in urine volume and renal blood flow in anesthetized dogs Jpn J Pharmacol 2000 84 310 315 11138732 10.1254/jjp.84.310
Young JS Passmore JC Hemodynamic and renal advantages of dual cyclooxygenase and leukotriene blockade during canine endotoxic shock Circ Shock 1990 32 243 255 2175681
Zellner JL Cook JA Reines HD Smith EF 3rdKessler LD Halushka PV Effect of a LTD4 receptor antagonist in porcine septic shock Eicosanoids 1991 4 169 175 1663380
Zhang H Spapen H Manikis P Rogiers P Metz G Buurman WA Vincent JL Tirilazad mesylate (U-74006F) inhibits effects of endotoxin in dogs Am J Physiol 1995 268 H1847 H1855 7771536
Zhang H Rogiers P Friedman G Preiser JC Spapen H Buurman WA Vincent JL Effects of nitric oxide donor SIN-1 on oxygen availability and regional blood flow during endotoxic shock Arch Surg 1996 131 767 774 8678780
Zhang H Rogiers P Smail N Cabral A Preiser JC Peny MO Vincent JL Effects of nitric oxide on blood flow distribution and O2 extraction capabilities during endotoxic shock J Appl Physiol 1997 83 1164 1173 9338425
Zhang H De Jongh R De Backer D Cherkaoui S Vray B Vincent JL Effects of alpha – and beta -adrenergic stimulation on hepatosplanchnic perfusion and oxygen extraction in endotoxic shock Crit Care Med 2001 29 581 588 11373424 10.1097/00003246-200103000-00020
Zhang H De Jongh R Cherkaoui S Shahram M Vray B Vincent JL Effects of nucleoside transport inhibition on hepatosplanchnic perfusion, oxygen extraction capabilities, and TNF release during acute endotoxic shock Shock 2001 15 378 385 11336198
Sward K Valsson F Sellgren J Ricksten SE Bedside estimation of absolute renal blood flow and glomerular filtration rate in the intensive care unit. A validation of two independent methods Intensive Care Med 2004 30 1776 1782 15375650
Guyton AC Textbook of Medical Physiology 1986 7 WB Saunders Company 394 396 [AU: please indicate the title of the chapter you are citing, and the place of the publisher.]
Gilbert RP Mechanisms of the hemodynamic effects of endotoxin Physiol Rev 1960 40 245 279 13850016
Wichterman KA Baue AE Chaudry IH Sepsis and septic shock–a review of laboratory models and a proposal J Surg Res 1980 29 189 201 6997619 10.1016/0022-4804(80)90037-2
Parker MM Shelhamer JH Natanson C Alling DW Parrillo JE Serial cardiovascular variables in survivors and nonsurvivors of human septic shock: heart rate as an early predictor of prognosis Crit Care Med 1987 15 923 929 3652707
Thijs A Thijs LG Pathogenesis of renal failure in sepsis Kidney Int Suppl 1998 66 S34 S37 9573570
Winslow EJ Loeb HS Rahimtoola SH Kamath S Gunnar RM Hemodynamic studies and results of therapy in 50 patients with bacteremic shock Am J Med 1973 54 421 432 4696004 10.1016/0002-9343(73)90038-7
Villazon SA Sierra UA Lopez SF Rolando MA Hemodynamic patterns in shock and critically ill patients Crit Care Med 1975 3 215 221 1201657
Mathiak G Szewczyk D Abdullah F Ovadia P Feuerstein G Rabinovici R An improved clinically relevant sepsis model in the conscious rat Crit Care Med 2000 28 1947 1952 10890646 10.1097/00003246-200006000-00043
Koo DJ Zhou M Chaudry IH Wang P The role of adrenomedullin in producing differential hemodynamic responses during sepsis J Surg Res 2001 95 207 218 11162047 10.1006/jsre.2000.6013
Wang P Chaudry IH Mechanism of hepatocellular dysfunction during hyperdynamic sepsis Am J Physiol 1996 270 R927 R938 8928923
Di Giantomasso D May CN Bellomo R Norepinephrine and vital organ blood flow Intensive Care Med 2002 28 1804 1809 12447527 10.1007/s00134-002-1444-x
|
16137349
|
PMC1269441
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 24; 9(4):R363-R374
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3540
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc35441613734710.1186/cc3544ResearchDNase and atelectasis in non-cystic fibrosis pediatric patients Hendriks Tom 1de Hoog Matthijs 2Lequin Maarten H 3Devos Annick S 3Merkus Peter JFM [email protected] Pediatrician, Catharina Hospital, Eindhoven, The Netherlands2 Pediatric Intensivist, Division of Intensive Care, Department of Pediatrics, Erasmus University and Erasmus Medical Centre/Sophia Children's Hospital, Rotterdam, The Netherlands3 Pediatric Radiologist, Division of Radiology, Department of Pediatrics, Erasmus University and Erasmus Medical Centre/Sophia Children's Hospital, Rotterdam, The Netherlands4 Pediatric Pulmonologist, Division of Respiratory Medicine, Department of Pediatrics, Erasmus University and Erasmus Medical Centre/Sophia Children's Hospital, Rotterdam, The Netherlands2005 20 5 2005 9 4 R351 R356 26 11 2004 18 1 2005 14 4 2005 20 4 2005 Copyright © 2005 Hendriks et al, licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
No evidence based treatment is available for atelectasis. We aimed to evaluate the clinical and radiologic changes in pediatric patients who received DNase for persistent atelectasis that could not be attributed to cardiovascular causes, and who were unresponsive to treatment with inhaled bronchodilators and physiotherapy.
Methods
All non-cystic fibrosis pediatric patients who received nebulised or endotracheally instilled DNase for atelectasis between 1998 and 2002, with and without mechanical ventilation, were analysed in a retrospective descriptive study. The endpoints were the blood pCO2, the heart rate, the respiratory rate, the FiO2 and the chest X-ray scores before and after treatment.
Results
In 25 of 30 patients (median [range] age, 1.6 [0.1–11] years) who met inclusion criteria, paired data of at least three endpoints were available. All clinical parameters improved significantly within 2 hours (P < 0.01), except for the heart rate (P = 0.06). Chest X-ray scores improved significantly within 24 hours after DNase treatment (P < 0.001). Individual improvement was observed in 17 patients and no clinical change was observed in five patients. Temporary deterioration (n = 3) was associated with increased airway obstruction and desaturations. No other complications were observed.
Conclusion
After treatment with DNase for atelectasis of presumably infectious origin in non-cystic fibrosis pediatric patients, rapid clinical improvement was observed within 2 hours and radiologic improvement was documented within 24 hours in the large majority of children, and increased airway obstruction and ventilation–perfusion mismatch occurred in three children, possibly due to rapid mobilisation of mucus. DNase may be an effective treatment for infectious atelectasis in non-cystic fibrosis pediatric patients.
See related commentary
==== Body
Introduction
Atelectasis is a problem in many children with respiratory infections or who require ventilation. At least 8% of children on mechanical ventilation develop pulmonary atelectasis, with a concomitant increase in the morbidity and the length of stay [1]. There is no 'golden standard' for treatment of atelectasis in children. Efficacy of treatment modalities such as inhaled bronchodilators, steroids, physiotherapy and nebulised sodium chloride (NaCl 0.9%) has not been demonstrated [2].
Atelectasis is commonly caused by sputum blocking the airways. Mucus in patients with cystic fibrosis (CF) [3], in patients with bronchiectasis [4] and in patients with respiratory syncytial virus (RSV) bronchiolitis [5] contains significant amounts of extracellular DNA from degenerating leucocytes and epithelial debris. DNA increases the viscosity and adhesiveness of lung secretions [6]. Recombinant human DNase (rhDNase) has proven to be an effective treatment in opening airways in CF [7-9]. In infections complicated by atelectasis, bronchial secretions and mucus plugs also have a high concentration of DNA [5], such that DNase could also be an effective treatment in this situation.
Until now only case reports on the efficacy of DNase treatment in atelectasis have been published, suggesting efficacy [5,10-14] No randomised study has been published. The present study analysed the resolution of atelectasis following treatment with DNase in a large series of hospitalised children who were refractory to conventional treatment.
Materials and methods
Study subjects
This retrospective study included all patients who received rhDNase as treatment for atelectasis in patients with suspected or proven lower airway infection at Sophia Children's hospital, Rotterdam between 1998 and 2002. Patients were identified through the computerised pharmacy registration. Patients were included in the analysis when they had pulmonary atelectasis of at least one lobe, and when rhDNase was administered for that reason. CF patients were excluded and all other patients were evaluated. Thirty patients and 32 episodes of atelectasis were identified, and this included patients described previously [14]. When more than one atelectasis was treated, only the first episode was included in the analysis.
Table 1 presents the demographic data of the study group. The median age of the study group was 1.6 years. Sixteen patients were younger than 1 year of age; two of these patients were born prematurely at 26 weeks gestation, but were 6 months and 7 months corrected postnatal age at the time of admission. Sixteen patients were intubated in the days before receiving rhDNase (median, 3 days; range, 1–16 days). Twenty-five patients were treated in a pediatric intensive care unit, and five patients were treated in a medium care unit. Underlying illnesses or predisposing factors for severe lower airways infections were: airway malacia (seven patients), severe psychomotor retardation (five patients), congenital heart disease (five patients: three children with ventricular septal defect, one child with tetralogy of fallot, and one child with hypoplastic right ventricle, pulmonary atresia and blalock taussig shunt), tracheostomy (four patients), bronchopulmonary dysplasia (three patients), epilepsy (three patients), neuromuscular disease (two patients) and bronchiectasis (one patient). Because rhDNase was administered as a part of patient care and not as a medical trial, formal approval from an institutional review board or a medical ethics committee was not required in our hospital, and was therefore not requested.
Methods
RhDNase was only considered and administered when patients did not demonstrate clinical improvement following empirical treatment for atelectasis and still demonstrated significant elevated work of breathing, could not be weaned easily off the ventilator or improved too slowly or not at all. RhDNase (Pulmozyme®; Roche, Basel, Switzerland) was administered either as a 2.5 mg dose nebulised twice daily with a jet nebuliser, using a tight-fitting mask and high-flow oxygen, in children breathing spontaneously, or 10% of this dose was diluted to 5 ml with NaCl 0.9% and given slowly as droplets during 30 min into the endotracheal tube or the canula twice daily. This treatment was continued until the atelectasis had improved sufficiently, preferably based on the chest X-ray (CXR) of the next day. This dose was chosen as it was estimated that pulmonary deposition of a regular 2.5 mg dosage would be a maximal 10%. When rhDNase was instilled endotracheally, it was attempted to position the head as favorably as possible for the DNase to reach the affected lobe(s).
RhDNase was administered twice daily until patients improved clinically because it was assumed that deposition in the peripheral airways of these children would be significantly diminished due to airway obstruction. All ventilated patients were sedated according to protocol, but were not paralysed, and the ventilators were standard not in the controlled ventilatory mode but in the pressure-regulated volume control mode. Blood gas control values were only obtained when the patients were stable, at least 30 min after manipulation or endotracheal suctioning.
Following rhDNase administration, the ventilator settings were not altered until the results of the blood gas analyses were available (2 hours later), except in the case of clinical deterioration. The nursing staff was only instructed to taper off FiO2 as much as possible. The clinical parameters analysed were the heart rate (HR), the respiratory rate (RR), the capillary or arterial pCO2, and the FiO2 before and within 2 hours following administration of rhDNase. Evidence for atelectasis on CXR was quantified using a CXR score before and within 24 hours of treatment with rhDNase. In children on mechanical ventilation, the peak inspiratory pressure, the mean airway pressure, days on the ventilator before receiving rhDNase and the time to extubation were recorded. All parameters and side effects were collected from patient files. In pediatric intensive care patients, dates were also obtained from the computerised data management system.
As cardiorespiratory endpoints we considered the change in clinical parameters (RR, HR, FiO2, PCO2) before and within 2 hours after the first dose of rhDNase, and considered the CXR score before and within 24 hours after the first dose of rhDNase. Parameters were compared using the Wilcoxon signed rank test. Interobserver comparisons of CXR scores were made using Cohen's kappa. Overall individual improvement in patients was defined as the improvement of two or more endpoints.
Analysis
Clinical parameters
Parameters were compared before and within 2 hours after treatment with rhDNase.
Individual improvement of the FiO2, the RR and the HR was defined as >10% decrease, and deterioration was defined as >10% increase. Individual improvement of pCO2 was defined as a decrease >1 kPa, and deterioration was defined as an increase >1 kPa. When patients were on mechanical ventilation, the peak inspiratory pressure and the mean airway pressure improvement was defined as >3 cmH2O decrease of pressure and their deterioration was defined as >3 cmH2O increase of pressure.
Radiology
Anteroposterior CXRs before and within 24 hours after treatment with rhDNase were coded, blinded and interpreted randomly by two independent pediatric radiologists. Since a validated scoring system for atelectasis is lacking, a scoring system based on available literature [15-17] and personal experience of our radiologists was defined as follows.
Each X-ray was scored for atelectasis, hyperinflation and mediastinal shift. The presence or absence of hyperinflation was marked as 1 point or 0 points, respectively. The presence or absence of a mediastinal shift was scored as 1 or 0. Atelectasis was scored for each lobe. A partial atelectasis of one pulmonary lobe was scored as 1 point, and complete atelectasis of one lobe was marked as 2 points. The distinction between infiltrate and atelectasis was left up to the pediatric radiologist, and was judged similarly to that in routine clinical care. These results were summed for each CXR. The CXR score before rhDNase treatment was compared with the CXR score within 24 hours after treatment.
Results
Treatment
Conventional treatment of atelectasis before the use of rhDNase consisted of nebulised bronchodilators in 25 patients, nebulised NaCl 0.9% in 16 patients, systemic or inhaled glucocorticoids in 18 patients and physiotherapy in all patients.
RhDNase was nebulised in 18 patients and was given as a droplet in 12 patients on mechanical ventilation. All patients received antibiotics after obtaining the appropriate cultures because bacterial infections could not be ruled out and these children had elevated inflammatory markers and were seriously ill.
RhDNase administration
In 25 out of 30 patients who met the inclusion criteria, paired data of at least three cardiorespiratory endpoints were available. Individual values before and after rhDNase treatment are shown in Fig. 1. Group results are presented in Table 2. All clinical variables, except the HR, show a significant improvement within 2 hours following rhDNase treatment (P < 0.01, Fig. 1).
Anteroposterior CXRs before and within 24 hours after treatment were obtained in 22 of 30 patients. On average, the median CXR score improved from 4.0 to 2.0 (P < 0.001). No paired CXRs were available in eight patients. In seven of these eight patients, the hospital records documented that no CXR was made after rhDNase treatment because the clinical improvement was very obvious and another CXR was unnecessary. One patient died before a post-treatment CXR was made when all treatment, including mechanical ventilation, was discontinued because of the very poor prognosis of her mitochondrial encephalomyopathy. Three patients showed complete resolution of all atelectasis within 24 hours. Individual CXR improvement was observed in 17 patients, no clear change was seen in two patients and deterioration was observed in three patients. Agreement between the CXR scores by the two observers expressed as Cohen's kappa was 0.61 (P < 0.001).
Individual improvement of at least two endpoints was seen in 17 patients, but no clinical change was observed in five patients. Three patients on mechanical ventilation (two ex-premature infants now 6 months and 7 months corrected postgestational age with RSV bronchiolitis, and one 6-month-old full-term child with congenital airway narrowing with an adenovirus respiratory infection) showed an immediate deterioration after administration of rhDNase. This was possibly due to excessive mobilisation of mucus leading to temporary increased airway obstruction. Oxygenation was extremely difficult for 2 hours after administration of rhDNase in two of these patients. Oxygen saturations remained between 85% and 90%, despite increased ventilator settings. No other side effects were observed.
Twelve of the 16 patients on mechanical ventilation were extubated within 6 days after the onset of rhDNase treatment; six of these were extubated the day after administration of rhDNase. Four patients remained on mechanical ventilation for longer than 10 days.
Discussion
This retrospective series, being the largest study so far, suggests that rhDNase may be an effective drug in the treatment of refractory atelectasis in non-CF patients. Quick resolution of atelectasis is important to reduce the number of days on mechanical ventilation with associated morbidity and to prevent the need for therapeutic bronchoscopy. Several case reports [10-14] and one randomised study on RSV bronchiolitis [5] suggest a clinical and radiological improvement after treatment with rhDNase in patients with atelectasis. In the present study rhDNase was administered when an infection was proven or suspected, and when patients had severe respiratory problems due to atelectasis.
Individual improvement of at least two endpoints was observed in 17 patients, and mobilisation of sputum occurred so rapidly in three children that this resulted in temporary worsening of the ventilation perfusion mismatch due to increased airway obstruction. Clinical improvement was observed in most of the children within 2 hours, and the mean values of the RR, pCO2 and FiO2 all improved significantly after rhDNase treatment. One could argue that the improvement of these respiratory parameters is statistically significant, but is not clinically relevant. However, the extubation rate and CXRs do suggest a clinically relevant effect. A substantial number of ventilated patients (six out of 16) with refractory atelectasis could be extubated within 24 hours following the first dose of rhDNase, and CXRs improved in 17 out of 22 cases within 24 hours. As in the study by Nasr and colleagues [5], CXRs were scored by two independent radiologists, thereby preventing observer bias. Agreement between the two pediatric radiologists can be considered satisfactory. Furthermore, the degrees of improvement of CXRs scored by each radiologist were also similar. Nasr and colleagues studied CXRs in infants with RSV bronchiolitis before and after treatment with rhDNase [5]. It is difficult to compare their results with those of the present study as they interpreted the change in CXR on admission and at discharge, they administered rhDNase only once daily and they used a different CXR score, reflecting CXR features of bronchiolitis. They found a small but significant improvement of the CXR score after rhDNase, while their control group showed a significant deterioration.
The deterioration with increased airway obstruction and ventilation-perfusion mismatch in three children on mechanical ventilation was interpreted as a result of rapid mobilisation of mucus in these three patients. This deterioration was observed in three out of 12 infants, but not in the seven children who were younger than these three patients. In a subanalysis no relationship was found between this deterioration and viral infection. We speculate that this deterioration may be due to the mode of administration. If the drug is instilled, the effective lung dose may be far greater than when it is inhaled; in all three patients, rhDNase was instilled endotracheally. Instillation may be an attractive option in that nebulised rhDNase administration in patients on mechanical ventilation results in significant deposition of the drug in the ventilator tubing, but it may also imply a risk. To prevent deterioration following instillation, however, lower starting doses of instilled rhDNase may be warranted. The dose we used for instillation was higher than reported by Boeuf and colleagues [11] and was lower than that administered bronchoscopically by Durward and colleagues [12]. Incidentally, in the case reports mentioned earlier, no clinical deterioration was observed when rhDNase was nebulised or was instilled bronchoscopically. In the randomised clinical trial on RSV bronchiolitis [5], a beneficial effect and no adverse events were observed. This was possibly also explained by using nebulised rhDNase rather than instilled rhDNase. In the present study, none of the known adverse effects such as pharyngitis, airway irritation, laryngitis, conjunctivitis or rash were observed, nor any rebound effects within 24 hours following administration.
There are several limitations to this study. First, it is a retrospective and open study potentially suffering from selection effects, and lacking a control group. Second, there are no validated scoring systems for atelectasis on CXR. Third, the sputum DNA content was not known, and rhDNase was administered irrespectively. Hence, the present study does not provide ultimate proof that improvement should be attributed to rhDNase treatment.
However, an association between drug treatment and a beneficial clinical response is more likely to be causal when the response follows immediately or quickly after administration of the drug, when the response is consistent, when the response is marked and when the response is plausible with respect to the pathophysiology behind the disorder. In addition, the quick alterations following rhDNase administration are also consistent with in vitro experiments that demonstrated a quick effect on sputum characteristics [1,18]. We therefore think that there is a causal relationship between DNase administration and the clinical outcome, but further randomised control trials are required to confirm this.
In addition, because rhDNase is expensive, a cost-benefit calculation in such a trial is warranted. Effective treatment of atelectasis is likely to reduce the stay in the hospital. Daily treatment costs of €60 should ideally reduce the length of hospital stay and outweigh the costs of hospital stay (which would be €1000/day in The Netherlands).
Conclusion
We conclude that our observations suggest efficacy of the drug in at least 17/25 (68%) of the patients, and show complete resolution of all atelectasis in three patients within 1 day. RhDNase may hence have a place in the treatment of children with atelectasis. However, randomised controlled studies are needed to prove this, and also to assess whether it is cost-beneficial and can shorten the hospital stay of children with atelectasis.
Key messages
• Rapid clinical and radiological improvement was observed in the large majority of children following rhDNase treatment for atelectasis.
• Increased airway obstruction and ventilation-perfusion mismatch may occur when rhDNase is instilled endotracheally, possibly due to rapid mobilisation of mucus.
• DNase may be an effective and cost-beneficial treatment for atelectasis in non-CF pediatric patients.
Abbreviations
CF = cystic fibrosis; CXR = chest X-ray; FiO2 = Fraction of inspired Oxygen ; HR = heart rate; pCO2 = Pressure of CO2; rhDNase = recombinant human DNase; RR = respiratory rate; RSV = respiratory syncytial virus
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
TH performed data analysis and contributed to the manuscript. MdH provided clinical care and contributed to the manuscript. MHL performed radiology data analyses and designed the study. ASD performed radiology data analyses and designed the study. PJFM performed data analysis and contributed to the manuscript.
Figures and Tables
Figure 1 Change of cardiorespiratory parameters before and after treatment. Individual changes of the heart rate, breathing frequency, blood pCO2 and FiO2 before and within 2 hours, and the chest X-ray score before and within 24 hours after administration of DNase in 30 children treated for atelectasis. Although changes were statistically significant, a significant overlap was present before and after recombinant human DNase treatment (see also Table 1).
Table 1 Demographic data of the study group
Sex (male/female) 19/11
Median (range) age 1.6 years (14 days–12 years)
Median (range) duration of disease before atelectasis 12 days (2–365 days)
On ventilator/no ventilator 16/14
Intensive care/medium care 25/5
Table 2 Clinical variables before and after administration of recombinant human DNase
Variable Before After P
Heart rate (beats/min) 150 (121–164) 130 (115–145) 0.06
Respiratory rate 41 (26–60) 36 (25–40) 0.01
PCO2 (kPa) 6.55 (5.7–8.6) 6.00 (5.3–7.5) 0.005
FiO2 (%) 50 (40–100) 40 (21–75) <0.001
Chest X-ray score 4 (2–6) 2 (0–4) <0.001
Data presented as the median (interquartile range). Comparisons were made using the Wilcoxon signed rank test.
==== Refs
Rivera R Tibballs J Complications of endotracheal intubation and mechanical ventilation in infants and children Crit Care Med 1992 20 193 199 1737455
Peroni DG Boner AL Atelectasis: mechanisms, diagnosis and management Paediatr Respir Rev 2000 1 274 278 12531090 10.1053/prrv.2000.0059
Shak S Capon DJ Hellmiss R Marsters SA Baker CL Recombinant human DNase I reduces the viscosity of cystic fibrosis sputum Proc Natl Acad Sci USA 1990 87 9188 9192 2251263
Picot R Das I Reid L Pus, deoxyribonucleic acid, and sputum viscosity Thorax 1978 33 235 242 26989
Nasr SZ Strouse PJ Soskolne E Maxvold NJ Garver KA Rubin BK Moler FW Efficacy of recombinant human deoxyribonuclease I in the hospital management of respiratory syncytial virus bronchiolitis Chest 2001 120 203 208 11451839 10.1378/chest.120.1.203
Puchelle E Zahm JM de Bentzmann S Grosskopf C Shak S Mougel D Polu JM Effects of rhDNase on purulent airway secretions in chronic bronchitis Eur Respir J 1996 9 765 769 8726943 10.1183/09031936.96.09040769
Fuchs HJ Borowitz DS Christiansen DH Morris EM Nash ML Ramsey BW Rosenstein BJ Smith AL Wohl ME Effect of aerosolized recombinant human DNase on exacerbations of respiratory symptoms and on pulmonary function in patients with cystic fibrosis. The Pulmozyme Study Group N Engl J Med 1994 331 637 642 7503821 10.1056/NEJM199409083311003
Harms HK Matouk E Tournier G von der Hardt H Weller PH Romano L Heijerman HG FitzGerald MX Richard D Strandvik B Multicenter, open-label study of recombinant human DNase in cystic fibrosis patients with moderate lung disease. DNase International Study Group Pediatr Pulmonol 1998 26 155 161 9773909 10.1002/(SICI)1099-0496(199809)26:3<155::AID-PPUL1>3.0.CO;2-K
Quan JM Tiddens HA Sy JP McKenzie SG Montgomery MD Robinson PJ Wohl ME Konstan MW Pulmozyme Early Intervention Trial Study Group A two-year randomized, placebo-controlled trial of dornase alfa in young patients with cystic fibrosis with mild lung function abnormalities J Pediatr 2001 139 813 820 11743506 10.1067/mpd.2001.118570
Voelker KG Chetty KG Mahutte CK Resolution of recurrent atelectasis in spinal cord injury patients with administration of recombinant human DNase Intensive Care Med 1996 22 582 584 8814475 10.1007/s001340050134
Boeuf B Prouix F Morneau S Marton D Lacroix J Safety of endotracheal rh DNase (Pulmozyme) for treatment of pulmonary atelectasis in mechanically ventilated children Pediatr Pulmonol 1998 26 147 9727769 10.1002/(SICI)1099-0496(199808)26:2<147::AID-PPUL14>3.0.CO;2-1
Durward A Forte V Shemie SD Resolution of mucus plugging and atelectasis after intratracheal rhDNase therapy in a mechanically ventilated child with refractory status asthmaticus Crit Care Med 2000 28 560 562 10708200 10.1097/00003246-200002000-00045
El Hassan NO Chess PR Huysman MW Merkus PJ de Jongste JC Rescue use of DNase in critical lung atelectasis and mucus retention in premature neonates Pediatrics 2001 108 468 470 11483817 10.1542/peds.108.2.468
Merkus PJ de Hoog M van Gent R de Jongste JC DNase treatment for atelectasis in infants with severe respiratory syncytial virus bronchiolitis Eur Respir J 2001 18 734 737 11716180
Friis B Eiken M Hornsleth A Jensen A Chest X-ray appearances in pneumonia and bronchiolitis. Correlation to virological diagnosis and secretory bacterial findings Acta Paediatr Scand 1990 79 219 225 2321485
Babcook CJ Norman GR Coblentz CL Effect of clinical history on the interpretation of chest radiographs in childhood bronchiolitis Invest Radiol 1993 28 214 217 8486486
Nasr SZ Kuhns LR Brown RW Hurwitz ME Sanders GM Strouse PJ Use of computerized tomography and chest x-rays in evaluating efficacy of aerosolized recombinant human DNase in cystic fibrosis patients younger than age 5 years: a preliminary study Pediatr Pulmonol 2001 31 377 382 11340684 10.1002/ppul.1061
Dawson M Wirtz D Hanes J Enhanced viscoelasticity of human cystic fibrotic sputum correlates with increasing microheterogeneity in particle transport J Biol Chem 2003 50 50393 50401 10.1074/jbc.M309026200
|
16137347
|
PMC1269442
|
CC BY
|
2021-01-04 16:04:55
|
no
|
Crit Care. 2005 May 20; 9(4):R351-R356
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3544
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37141613734810.1186/cc3714ResearchA quantitative analysis of the acidosis of cardiac arrest: a prospective observational study Makino Jun [email protected] Shigehiko [email protected] Hiroshi [email protected] Rinaldo [email protected] Staff specialist in emergency, Tertiary Emergency Medical Center, Tokyo Metropolitan Bokuto Hospital, Tokyo, Japan2 Staff specialist in intensive care, Department of Emergency and Critical Care Medicine, Saitama Medical Center, Saitama Medical School, Saitama, Japan3 Staff specialist in intensive care, Department of Anesthesiology and Resuscitology, Okayama University Medical School, Okayama, Japan4 Director of intensive care research, Department of Intensive Care and Department of Medicine, Austin & Repatriation Medical Centre, Melbourne, Australia2005 23 5 2005 9 4 R357 R362 11 1 2005 22 2 2005 27 3 2005 25 4 2005 Copyright © 2005 Makino et al. licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Metabolic acidosis is common in patients with cardiac arrest and is conventionally considered to be essentially due to hyperlactatemia. However, hyperlactatemia alone fails to explain the cause of metabolic acidosis. Recently, the Stewart–Figge methodology has been found to be useful in explaining and quantifying acid–base changes in various clinical situations. This novel quantitative methodology might also provide useful insight into the factors responsible for the acidosis of cardiac arrest. We proposed that hyperlactatemia is not the sole cause of cardiac arrest acidosis and that other factors participate significantly in its development.
Methods
One hundred and five patients with out-of-hospital cardiac arrest and 28 patients with minor injuries (comparison group) who were admitted to the Emergency Department of a tertiary hospital in Tokyo were prospectively included in this study. Serum sodium, potassium, ionized calcium, magnesium, chloride, lactate, albumin, phosphate and blood gases were measured as soon as feasible upon arrival to the emergency department and were later analyzed using the Stewart–Figge methodology.
Results
Patients with cardiac arrest had a severe metabolic acidosis (standard base excess -19.1 versus -1.5; P < 0.0001) compared with the control patients. They were also hyperkalemic, hypochloremic, hyperlactatemic and hyperphosphatemic. Anion gap and strong ion gap were also higher in cardiac arrest patients. With the comparison group as a reference, lactate was found to be the strongest determinant of acidosis (-11.8 meq/l), followed by strong ion gap (-7.3 meq/l) and phosphate (-2.9 meq/l). This metabolic acidosis was attenuated by the alkalinizing effect of hypochloremia (+4.6 meq/l), hyperkalemia (+3.6 meq/l) and hypoalbuminemia (+3.5 meq/l).
Conclusion
The cause of metabolic acidosis in patients with out-of-hospital cardiac arrest is complex and is not due to hyperlactatemia alone. Furthermore, compensating changes occur spontaneously, attenuating its severity.
See related commentary
==== Body
Introduction
Metabolic acidosis is common in patients with cardiac arrest and is conventionally considered to be due essentially to hyperlactatemia [1-6]. However, hyperlactatemia alone fails to explain the cause of metabolic acidosis in some patients [3]. Traditional measures using the anion gap, standard bicarbonate and standard base excess might help to understand this acidosis [7,8]. However, they give little information about the mechanisms involved and the quantitative contribution of each variable [9-12], especially in the presence of major changes in serum electrolytes and albumin concentration.
Recently, the Stewart–Figge methodology [13,14] has been found to be useful in explaining and quantifying acid–base changes in clinical situations in which conventional analysis was deficient [15-18]. This novel quantitative methodology might also provide useful insight into the factors responsible for the acidosis of cardiac arrest.
We proposed that hyperlactatemia is not the sole cause of cardiac arrest acidosis and that other factors participate significantly in its development. We tested this hypothesis by conducting a prospective study of patients admitted to the Emergency Department of a tertiary hospital in Tokyo, Japan, and by applying quantitative principles to the assessment of their acid–base disorders.
Materials and methods
This study took place in an Emergency Department of a tertiary hospital in Tokyo, Japan. We prospectively examined out-of-hospital patients with cardiac arrest admitted to the department from May 2003 to October 2003. Because of the anonymous and non-interventional manner of the study, informed consent was not obtained.
Cardiac arrest was defined as the absence of both spontaneous respiration and palpable pulse. Cardiac arrest was described as witnessed arrest if the collapse of a patient was witnessed by a bystander or the emergency ambulance service. All patients were resuscitated in accordance with the Guidelines 2000 for Cardiopulmonary Resuscitation and Emergency Cardiovascular Care [19]. Data evaluated included age, sex, initial electrocardiographic rhythm at the scene, and possible cause. To compare the acid–base characteristics of these patients, we used a comparison group consisting of patients with minor injuries who were discharged within 2 days after admission. We used this group of patients as a comparison group because we routinely measured all variables required for the analysis such as lactate and phosphate in these patients.
Arterial samples were collected in heparinized plastic syringes and analyzed with a blood-gas analyzer (ABL 725; Radiometer, Copenhagen, Denmark) at the time of admission. The analyzer measured samples at 37°C. We collected the following data from the analyzer output: pH, partial pressure of carbon dioxide, bicarbonate, standard base excess, lactate and ionized calcium. Blood samples were also analyzed at the hospital central laboratory for the measurement of multiple biochemical variables including sodium, potassium, total magnesium, chloride, albumin and phosphate (Hitachi 7350 and 7360; Hitachi Industry, Tokyo, Japan). No sodium bicarbonate was administered before blood sampling.
Conceptual framework for the interpretation of quantitative acid–base analysis
Quantitative physicochemical analysis of the results was performed with Stewart's quantitative biophysical methods [13] as modified by Figge and colleagues [14] to take into account the effects of plasma proteins. This method involves first calculating the apparent strong ion difference (SIDa):
SIDa = [Na+] + [K+] + [Mg2+] + [Ca2+] - [Cl- ] - [lactate- ]
(all concentrations in meq/l).
However, this equation does not take into account the role of weak acids (CO2, albumin and phosphate) in the balance of electrical charges in plasma water. This is expressed through the calculation of the effective strong ion difference (SIDe). The formula for SIDe as determined by Figge and colleagues [14] is as follows:
SIDe = 1000 × 2.46 × 10-11× PCO2/(10-pH) + [albumin] × (0.123 × pH - 0.631) + [phosphate] × (0.309 × pH - 0.469)
In this equation, PCO2 (the partial pressure of CO2) is measured in mmHg, albumin in g/l, and phosphate in mmol/l. This formula accounts quantitatively for the contribution of weak acids to the electrical charge equilibrium in plasma.
Once weak acids are taken into account quantitatively, SIDa - SIDe should equal 0 (electrical charge neutrality) unless there are unmeasured charges to explain this 'ion gap'. Such charges are described by the strong ion gap (SIG):
SIG = SIDa- SIDe
A positive value for SIG must represent unmeasured anions (such as sulfate, oxo acids, citrate, pyruvate, acetate and gluconate) that must be included to account for measured pH. The traditional anion gap was also calculated as anion gap = [Na+] + [K+] - [Cl-] - [HCO3-], with a reference range of 12–20 mmol/l [20].
Data are expressed as means ± s.d., or as percentage. Student's t-test was used to compare the study group and the comparison group (StatView; Abacus Concepts, Berkeley, CA, USA). P < 0.05 was considered statistically significant.
Results
One hundred and five patients with out-of-hospital cardiac arrest were included in this study. The demographics of these patients are presented in Table 1. They had a mean age of 62.2 years and included 75 (71%) males and 30 (29%) females. Most of the patients had an initial rhythm of asystole (54%) or pulseless electrical activity (38%), and the number of witnessed arrests was 10 (10%). The main cause of collapse was cardiogenic (57%), followed by trauma (12%) and hanging (9%); 19% of the patients had a return of spontaneous circulation.
These 105 patients were compared with 28 patients with minor injuries as a comparison group (mean age 40.2 years; 19 males and 8 females). The mean interval from arrival at the emergency room to blood sampling was similar between the two groups (cardiac arrest, 12.9 ± 10.3 min; minor injury, 12.3 ± 5.5 min; P = 0.78).
The acid–base variables in cardiac arrest and minor injuries are shown in Table 2. Except for sodium and SIDa, all variables were significantly different between the two groups. In brief, patients with cardiac arrest were acidemic (pH 6.90 versus 7.39; P < 0.0001), secondary to metabolic acidosis (standard base excess -19.1 versus -1.5 meq/l; P < 0.0001) compared with the comparison group. Patients with cardiac arrest were also hyperkalemic, hypochloremic, hyperlactatemic and hyperphosphatemic. The anion gap and SIG were also higher in patients with cardiac arrest.
Figure 1 shows the acid–base impact of each variable in patients with cardiac arrest compared with the comparison group. Lactate was the strongest determinant of acidemia, accounting for -11.8 meq/l of acidifying effect. However, SIG contributed -7.3 meq/l of acidifying effect and phosphate -2.9 meq/l. This acidemia was attenuated by the alkalinizing effect of several variables. A decrease in chloride had the strongest alkalinizing effect (+4.6 meq/l), followed by an increase in potassium (+3.6 meq/l) and a decrease in albumin (+3.5 meq/l).
Discussion
It has been known for decades that patients with cardiac arrest invariably develop a severe metabolic acidosis [1-6,21-23]. This acidosis has been thought secondary to hyperlactatemia [2]. However, the correlation between standard base excess and lactate has been reported to be poor, suggesting that other factors might participate in the pathogenesis of cardiac arrest acidosis [3]. The problem is that no previous studies quantitatively analyzed the cause of metabolic acidosis in these patients. We therefore sought to define and quantify acid–base status in these patients by applying the quantitative principles of acid–base analysis described by Stewart, Figge and colleagues [13,14]. Using this methodology, we found that the causes of acidosis were much more complex than previously thought. Although lactate was the biggest contributor to metabolic acidosis and the development of acidemia in these patients, it accounted for only about 50% of it, whereas SIG and phosphate combined contributed an almost equal percentage (about 33% and 13%, respectively). However, this acidosis was associated with strong compensating responses, which attenuated its severity. These responses included hypochloremia, hyperkalemia, hypoalbuminemia and, to a smaller extent, hypermagnesemia and hypercalcemia.
A key finding was that patients with cardiac arrest had a disproportionately higher SIG than the comparison group (12.4 versus 5.1 meq/l; P < 0.0001). Although the source of this disproportionate increase in unmeasured anions was not specifically addressed, possible candidates include sulphate, urate, oxo acids, amino acids and other organic acids. Kaplan and colleagues reported increased SIG in patients with major vascular injury, another type of global tissue hypoperfusion [24]. It therefore seems that unmeasured anions are likely to be generated during global tissue hypoxic states.
Hyperphosphatemia in patients with cardiac arrest has been underemphasized as a contributor of acidosis (2.95 mmol/l in our study patients). Although causes of this abnormality remain unclear, transcellular shift, cellular injury and phosphate release might be responsible [25,26]. Hyperphosphatemia is also related to other types of metabolic acidosis [27,28]. However, because phosphate is not included in calculations of the anion gap, its impact on acid–base status is often poorly appreciated. The Stewart–Figge methodology can reveal its importance. For example, we previously reported that, in patients with acute renal failure, hyperphosphatemia accounted for about 20% of the difference in the acid–base status of patients compared with controls [29].
These acidifying effects were partly attenuated by a concomitant metabolic alkalosis, due mainly to hypochloremia, hyperkalemia and hypoalbuminemia. Their alkalinizing effects in cardiac patients were 4.6, 3.6 and 3.5 meq/l, respectively. The alkalinizing effects of hypercalcemia and hypermagnesemia accounted for less than 1 meq/l of alkalinizing effect. These abnormalities in four serum electrolytes and albumin can be explained by transcellular electrolyte shifts and, perhaps, previous co-morbidities. The existence of metabolic alkalosis in patients with cardiac arrest is not intuitive, unless each element is quantified and compared with a control.
There are several limitations in this study. First, only 10% of patients had their arrest witnessed and most of patients had an initial rhythm of asystole or pulseless electrical activity. Furthermore, drug injection by the emergency ambulance service is not approved in Japan, which inevitably delays the start of advanced life support. Our results might therefore not be applicable to patients in other institutions or countries. However, published clinical studies examining acidemia during cardiac arrest in a quantitative manner are lacking, particularly in terms of patients who have no received advanced life support interventions. Our study presents the first quantitative acid–base analysis for this disorder. Furthermore, the lack of drug or fluid administration in the field provides a unique opportunity to study these disorders with minimal iatrogenic modifications.
Second, we used patients with minor injuries as a comparison group. Although they were well enough to be discharged from the hospital within 2 days, mild hyperlactatemia (2.5 mmol/l) was present in these patients. However, all other variables, including pH and bicarbonate, were in the normal ranges. Considering the large difference in lactate between the two groups (11.8 mol/l), this group of patients, although imperfect, seems adequate for comparison.
Third, fluid resuscitation had started just before or at the time of blood sampling in some patients. This might have affected acid–base status in both groups. Unfortunately, we did not collect precise information on such fluid administration (timing, amount or number of patients so treated) because of the logistic difficulty of collecting such detailed information while attempts were being made to save the life of the patients. This is a significant limitation of our study because some of the fluid given might have affected our findings. However, blood sampling was conducted 12 min on average after admission to the emergency room in both groups, and only small amounts of fluid resuscitation (acetate Ringer in both groups) would have been given to these patients before sampling. Although SIG might have been somewhat affected by acetate in the fluids, the difference in the amount of acetate Ringer given before blood sampling is unlikely to have been large enough to fully explain the significant difference in SIG between the two groups.
The last limitation was that, although we attempted to collect arterial samples, it was often difficult to distinguish which sample, arterial or venous, was actually collected from patients with cardiac arrest. Some of the differences we found might therefore have been due to the higher incidence of venous sampling in the cardiac arrest group. These difficulties are inherent in research in the emergency setting.
Conclusion
Using the Stewart–Figge methodology, we studied the acid–base status of out-of-hospital cardiac arrest patients and found its pathogenesis to be complex. We found that lactate accounts for only 50% of the metabolic acidosis and consequent acidemia seen in such patients and that an increase in unmeasured anions and phosphate accounts for the rest. We also found that their acidifying effect was partly attenuated by the alkalinizing effect of hypochloremia, hyperkalemia, hypoalbuminemia, hypermagnesemia and hypercalcemia. The clinical and prognostic significance of these changes requires further investigation.
Key messages
• Lactate accounts for only 50% of metabolic acidosis in cardiac arrest, and SIG and phosphate combined contributed almost an equal percentage.
• This acidosis is attenuated by hypochloremia, hyperkalemia and hypoalbuminemia.
• acid–base status in patients with cardiac arrest is more complex than previously thought.
Abbreviations
PCO2, partial pressure of CO2; SIDa = apparent strong ion difference; SIDe = effective strong ion difference; SIG = strong ion gap.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
JM and SU conceived the study, designed the trial and supervised the conduct of the trial and data collection. HM provided statistical advice on analyzed data, and RB chaired the data oversight committee. JM drafted the manuscript, and all authors contributed substantially to its revision. JM takes responsibility for the paper as a whole. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 The impact of each variable on the acid–base status of patients with out-of-hospital cardiac arrest. Each value is presented as the difference between the mean for the comparison group and the study group. A negative value suggests an acidifying effect, and a positive value an alkalinizing effect. Alb, albumin; Ca, calcium; Cl, chloride; K, potassium; Lac, lactate; Mg, magnesium; Na, sodium; Phos, phosphate; SIG, strong ion gap.
Table 1 Demographics of patients with cardiac arrest
Parameter Value
Age (years) 62.2 ± 15.5
Male (%) 75 (71%)
Arrest witnessed (%) 10 (10%)
Initial rhythm (%)
Asystole 57 (54%)
Pulseless electrical activity 40 (38%)
Ventricular fibrillation 8 (8%)
Cause of arrest (%)
Cardiogenic 60 (57%)
Trauma 13 (12%)
Hanging 9 (9%)
Respiratory 4 (4%)
Neurological 4 (4%)
Other 13 (12%)
ROSC (%) 20 (19%)
ROSC, return of spontaneous circulation.
Table 2 acid–base variables in patients with cardiac arrest and with minor injuries
Variable Cardiac arrest Minor injury P
pH 6.90 ± 0.21 7.39 ± 0.08 0.0001
PCO2 (Torr) 78.3 ± 42.8 39.2 ± 9.1 0.0001
Bicarbonate (mmol/l) 13.8 ± 5.4 22.8 ± 3.5 0.0001
Standard base excess (mmol/l) -19.1 ± 6.2 -1.5 ± 3.6 0.0001
Sodium (mmol/l) 140.4 ± 5.9 139.6 ± 2.8 0.49
Potassium (mmol/l) 7.3 ± 2.6 3.6 ± 0.4 0.0001
Ionized calcium (mmol/l) 1.31 ± 0.12 1.17 ± 0.08 0.0001
Total magnesium (mmol/l) 1.17 ± 0.27 0.86 ± 0.14 0.0001
Chloride (mmol/l) 98.6 ± 5.9 103.0 ± 4.6 0.0002
Lactate (mmol/l) 14.3 ± 5.8 2.5 ± 1.8 0.0001
Albumin (g/dl) 3.4 ± 0.7 3.9 ± 0.5 0.0002
Phosphate (mmol/l) 2.95 ± 1.07 1.06 ± 0.36 0.0001
Anion gap (meq/l) 20.1 ± 7.4 11.0 ± 3.5 0.0001
SIDa (meq/l) 38.9 ± 4.6 41.0 ± 2.9 0.22
SIDe (meq/l) 26.5 ± 6.1 35.9 ± 4.0 0.0001
Strong ion gap (meq/l) 12.4 ± 6.0 5.1 ± 3.6 0.0001
PCO2, partial pressure of CO2; SIDa, apparent strong ion difference; SIDe, effective strong ion difference.
==== Refs
Tuchschmidt JA Mecher CE Predictors of outcome from critical illness. Shock and cardiopulmonary resuscitation Crit Care Clin 1994 10 170 195
Capparelli EV Chow MS Kluger J Fieldman A Differences in systemic and myocardial blood acid–base status during cardiopulmonary resuscitation Crit Care Med 1989 17 442 446 2707015
Prause G Ratzenhofer-Comenda B Pierer G Smolle-Juttner F Glanzer H Smolle J Comparison of lactate or BE during out-of-hospital cardiac arrest to determine metabolic acidosis Resuscitation 2001 51 297 300 11738782 10.1016/S0300-9572(01)00424-5
Cairns CB Niemann JT Pelikan PC Sharma J Ionized hypocalcemia during prolonged cardiac arrest and closed-chest CPR in a canine model Ann Emerg Med 1991 20 1178 1182 1952301
Leavy JA Weil MH Rackow EC Lactate washout following circulatory arrest J Am Med Assoc 1988 260 662 664 10.1001/jama.260.5.662
Sato S Kimura T Okubo N Naganuma T Tanaka M End-tidal CO2 and plasma lactate level: a comparison of their use as parameters for evaluating successful CPR Resuscitation 1993 26 133 139 8290808 10.1016/0300-9572(93)90173-N
Astrup PJK Jorgensen K Andersen OS Engel K The acid–base metabolism – a new approach Lancet 1960 1 1035 1039 13794904 10.1016/S0140-6736(60)90930-2
Siggaard-Andersen O Fogh-Andersen N Base excess or buffer base (strong ion difference) as measure of a non-respiratory acid–base disturbance Acta Anaesthesiol Scand Suppl 1995 107 123 128 8599264
Figge J Rossing TH Fencl V The role of serum proteins in acid–base equilibria J Lab Clin Med 1991 117 453 467 2045713
McAuliffe JJ Lind LJ Leith DE Fencl V Hypoproteinemic alkalosis Am J Med 1986 81 86 90 3089010 10.1016/0002-9343(86)90187-7
Rossing TH Maffeo N Fencl V acid–base effects of altering plasma protein concentration in human blood in vitro J Appl Physiol 1986 61 2260 2265 3100499
Gilfix BM Bique M Magder S A physical chemical approach to the analysis of acid–base balance in the clinical setting J Crit Care 1993 8 187 197 8305955 10.1016/0883-9441(93)90001-2
Stewart PA Modern quantitative acid–base chemistry Can J Physiol Pharmacol 1983 61 1444 1461 6423247
Figge J Mydosh T Fencl V Serum proteins and acid–base equilibria: a follow-up J Lab Clin Med 1992 120 713 719 1431499
Liskaser FJ Bellomo R Hayhoe M Story D Poustie S Smith B Letis A Bennett M Role of pump prime in the etiology and pathogenesis of cardiopulmonary bypass-associated acidosis Anesthesiology 2000 93 1170 1173 11046201 10.1097/00000542-200011000-00006
Story D Poustie S Bellomo R Quantitative physical chemistry analysis of acid–base disorders in critically ill patients Anaesthesia 2001 56 530 533 11412158 10.1046/j.1365-2044.2001.01983.x
Wilkes P Hypoproteinemia strong-ion difference, and acid–base status in critically ill patients J Appl Physiol 1998 84 1740 1748 9572825
Fencl V Jabor A Kazda A Figge J Diagnosis of metabolic acid–base disturbances in critically ill patients Am J Respir Crit Care Med 2000 162 2246 2251 11112147
Anon Guidelines 2000 for Cardiopulmonary Resuscitation and Emergency Cardiovascular Care Circulation 2000 102 Suppl I I-1 I-370 10964901
Shapiro BA Peruzzi WT Shoemaker WC, Ayres SM, Grenik A, Holbrook P Interpretation of blood gasses Textbook of Critical Care 1995 3 Philadelphia: WB Saunders Company 274 290
Stewart JS Stewart WK Gillies HG Cardiac arrest and acidosis Lancet 1962 ii 964 967 10.1016/S0140-6736(62)90729-8
Edmonds-Seal J acid–base studies after cardiac arrest. A report on 64 cases Acta Anaesthesiol Scand 1966 235 241
Chazan JA Stenson R Kurland GS The acidosis of cardiac arrest N Engl J Med 1968 278 360 364 5638354
Kaplan LJ Kellum JA Initial pH, base deficit, lactate, anion gap, strong ion difference, and strong ion gap predict outcome from major vascular injury Crit Care Med 2004 32 1120 1124 15190960 10.1097/01.CCM.0000125517.28517.74
Oster JR Alpert HC Vaamonde CA Effect of acid–base status on plasma phosphorus response to lactate Can J Physiol Pharmacol 1984 62 939 942 6435844
Barsotti G Lazzeri M Cristofano C Cerri M Lupetti S Giovannetti S The role of metabolic acidosis in causing uremic hyperphosphatemia Miner Electrolyte Metab 1986 12 103 106 3007964
Wang F Butler T Rabbani GH Jones PK The acidosis of cholera. Contributions of hyperproteinemia, lactic acidemia, and hyperphosphatemia to an increased serum anion gap N Engl J Med 1986 315 1591 1595 3785323
Kirschbaum B The acidosis of exogenous phosphate intoxication Arch Intern Med 1998 158 405 408 9487238 10.1001/archinte.158.4.405
Rocktaeschel J Morimatsu H Uchino S Goldsmith D Poustie S Story D Gutteridge G Bellomo R acid–base status of critically ill patients with acute renal failure: analysis based on Stewart–Figge methodology Crit Care 2003 7 60 66 10.1186/cc2333
|
16137348
|
PMC1269443
|
CC BY
|
2021-01-04 16:04:55
|
no
|
Crit Care. 2005 May 23; 9(4):R357-R362
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3714
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37181613735310.1186/cc3718ResearchStudy protocol: The DOse REsponse Multicentre International collaborative initiative (DO-RE-MI) Kindgen-Milles Detlef [email protected] Didier [email protected] Roberto [email protected] Sergio 4Maynar Javier [email protected] Anibal [email protected] Irene [email protected] Alessandra 8Formica Marco [email protected] Sergio [email protected] Mariella 11Marchesi Mariano [email protected] Filippo [email protected] Gianpaola [email protected] Elena 15Silengo Daniela 16Ronco Claudio [email protected] Scientific Committee member; Leading Consultant, Anesthesiology Clinic, University of Düsseldorf, Germany2 Scientific Committee member; Director, Anesthesiology and Intensive Care Service, Hospital European Georges-Pompidou, Paris, France3 Scientific Committee member; Associate Professor, Department of Anesthesiology and Intensive Care, Medicine and Surgery Faculty, University of Milan, Italy4 Scientific Committee member; Director, Department of Anesthesiology and Intensive Care, Ospedale Niguarda, Milan, Italy5 Scientific Committee member; Vice-Head, Anesthesiology and Intensive Care Unit, Hospital Santiago Apostol, Vitoria, Spain6 Scientific Committee member; Vice-Head, Anesthesiology and Intensive Care Unit, Hospital Geral Sant Antonio, Porto, Portugal7 Steering Committee member; Epidemiology Consultant, Department of Nephrology, Hospital San Bortolo, Vicenza, Italy8 Steering Committee member; Vice-Head, Department of Nephrology, Hospital San Bortolo, Vicenza, Italy9 Steering Committee member; Director, Department of Nephrology, Hospital Santa Croce e Carle, Cuneo, Italy10 Steering Committee member; Director, Intensive Care Unit, Hospital G.Bosco, Torino, Italy11 Steering Committee member; Vice-Head, Intensive Care Unit, Hospital G.Bosco, Torino, Italy12 Steering Committee member; Vice-Head, Department of Anesthesiology and Intensive Care, Hospital Riuniti di Bergamo, Bergamo, Italy13 Steering Committee member; Vice-Head, Nephrology and Dialysis Unit, CTO Hospital, Turin, Italy14 Steering Committee member; Vice-Head, Department of Anesthesiology and Intensive Care,, Hospital Niguarda, Milan, Italy15 Steering Committee member; Vice-Head, Department of Anesthesiology and Intensive Care, Hospital Riuniti di Bergamo, Bergamo, Italy16 Steering Committee member; Vice-Head, Intensive Care Unit, Hospital G.Bosco, Torino, Italy17 Scientific Committee member; Director, Department of Nephrology, St. Bortolo Hospital, Vicenza, Italy2005 14 6 2005 9 4 R396 R406 12 4 2005 26 4 2005 Copyright © 2005 Ronco et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Current practices for renal replacement therapy in intensive care units (ICUs) remain poorly defined. The DOse REsponse Multicentre International collaborative initiative (DO-RE-MI) will address the issue of how the different modes of renal replacement therapy are currently chosen and performed. Here, we describe the study protocol, which was approved by the Scientific and Steering Committees.
Methods
DO-RE-MI is an observational, multicentre study conducted in ICUs. The primary end-point will be the delivered dose of dialysis, which will be compared with ICU mortality, 28-day mortality, hospital mortality, ICU length of stay and number of days of mechanical ventilation. The secondary end-point will be the haemodynamic response to renal replacement therapy, expressed as percentage reduction in noradrenaline (norepinephrine) requirement. Based on the the sample analysis calculation, at least 162 patients must be recruited. Anonymized patient data will be entered online in electronic case report forms and uploaded to an internet website. Each participating centre will have 2 months to become acquainted with the electronic case report forms. After this period official recruitment will begin. Patient data belong to the respective centre, which may use the database for its own needs. However, all centres have agreed to participate in a joint effort to achieve the sample size needed for statistical analysis.
Conclusion
The study will hopefully help to collect useful information on the current practice of renal replacement therapy in ICUs. It will also provide a centre-based collection of data that will be useful for monitoring all aspects of extracorporeal support, such as incidence, frequency, and duration.
==== Body
Introduction
The systemic inflammatory response syndrome is characterized by widespread endothelial damage caused by persistent inflammation from both infectious and noninfectious stimuli. The host employs hormonal and immunological mechanisms to counter the systemic inflammatory response syndrome. Hypoperfusion and shock result when homeostatic mechanisms are no longer able to keep the system in balance, leading to organ dysfunction [1].
Septic shock can be defined as sepsis with hypotension, despite adequate fluid resuscitation, along with evidence of perfusion abnormalities. It is the leading cause of acute renal failure (ARF) and mortality in intensive care patients. The pathogenesis usually involves a nidus of infection, which progresses to a bloodstream infection, followed by activation of mediators and eventual shock or multiorgan failure [2]. Both septic shock and severe bacterial infections are associated with increased levels of plasma cytokines such as tumour necrosis factor-α, IL-1, IL-6, IL-8 and IL-10, IL-1 receptor antagonist, and soluble tumour necrosis factor receptors types I and II. These mediators are produced in response to constituents of both Gram-negative and Gram-positive bacteria. Lipopolysaccharides of Gram-negative bacteria, and peptidoglycans, lipoteichoic acid and exotoxins of Gram-positive bacteria are largely responsible for the initial inflammatory cascade[3,4].
Various continuous and intermittent modalities of renal replacement therapy (RRT) are currently used. There has been slow acceptance of continuous RRT (CRRT) in intensive care units (ICUs) for the management of ARF, but this therapy is not new. In 1977 Kramer and coworkers [5] developed this technique following their accidental accessing of the femoral artery rather than the vein, creating an arterovenous circuit that yielded a very primitive but innovative approach. Problems with low blood flow and coagulation meant that this idea remained dormant for some time. It was not until the application of blood pumps and the substitution of arteriovenous with venovenous circuitry that the current practice of CRRT was born. In recent years remarkable advances in CRRT technology have been made, driven by nephrologists dedicated to improving efficiency and function. Today, however, intensivists are the most familiar with these techniques. Nevertheless, in some countries such as the USA, CRRT is still infrequently employed [6]. Other modalities include intermittent haemodialysis (IHD), slow extended daily dialysis [7], or daily haemodialysis [8]. Some of the reasons for the considerable variability worldwide in extracorporeal treatment of ARF include local practice (e.g. whether management is by nephrologists or intensivists), the centre's experience with the various techniques, organization and health resources. Various methods of extracorporeal treatment, whether intermittent or continuous, are currently being employed and no guidelines exist. This variability was highlighted in a recently completed observational study (the Beginning and Ending of Supportive Therapy for the Kidney [BEST Kidney] trial), which collected data on ARF management in 1743 patients in 54 ICU from 23 countries worldwide.
The practice of CRRT has apparently not changed, even following the prospective studies conducted by Ronco and coworkers [9]. Despite the positive findings of that prospective trial, the practice of a higher intensity CRRT has not been widely adopted into routine ICU practice. The most outstanding examples are Australia and New Zealand, where almost 100% of treatments are CRRT. A survey of several units active in the Australian and New Zealand Intensive Care Society Clinical Trials Group (Bellomo R, unpublished data, 2002) found that very few units had adopted the intensive CRRT regimen proposed by Ronco and coworkers [9]. Data from such Australian units shows instead that the vast majority (>90%) prescribe a 'fixed' standard CRRT dose of 2 l/hour, which is not adjusted for body weight. Thus, a 100 kg man would receive 20 ml/kg per hour – the dose shown to have the worst outcome in the study by Ronco and coworkers [9]. In another recent study that involved several Australian units (the BEST Kidney study), the median body weight for Australian patients was 80 kg, thus indicating that the vast majority receive a CRRT intensity of approximately 25 ml/kg per hour of effluent. Finally, although in the study conducted by Ronco and colleagues [9] the technique of CRRT was uniform in the form of continuous venovenous haemofiltration (CVVH) with postfilter fluid replacement, current practice includes a variety of techniques in addition to CVVH, such as continuous venovenous haemodialysis (CVVHD) and continuous venovenous haemodiafiltration (CVVHDF). Scarce information exists on the practice of CRRT in Europe, particularly regarding the actually delivered dose of therapy in critically ill patients with ARF (i.e. in those who could potentially derive more benefit from high volume convective therapy).
In a recent preliminary collaborative study [10] we reported that there was no significant difference between prescribed and delivered ultrafiltration rate (both in ml/min and in l/hour), which was related to the reduced down-time associated with the technique. However, of greater relevance is that the dose of dialysis was over 40 ml/kg per hour.
If we are to understand how dialysis doses are actually delivered in routine clinical practice in ICUs, an observational clinical study is needed to confirm how, to what extent and with what clinical indication the different modalities of RRT are administered. With this in mind we have initiated the DOse REsponse Multicentre International collaborative initiative (DO-RE-MI) trial. The primary end-point of DO-RE-MI is mortality (ICU mortality, 28-day mortality and hospital mortality), and the secondary end-point is the haemodynamic response to RRT, expressed as percentage reduction in noradrenaline (norepinephrine) requirement to maintain blood pressure.
Materials and methods
Figure 1 presents a study flowchart. Only incident patients with an indication for RRT will be recruited. The study is intended to describe current practices of RRT in all patients admitted to ICUs who are in need of RRT, with or without ARF. All data listed herein will be entered in electronic case report forms (CRFs) that are available via the internet [11]. The following rules will be applied without exception:
First, all patient data will be entered anonymously. To this aim, each centre will have a code, and patients will be consecutively assigned a unique number. Under no circumstances will there be any written or oral transmission of data that may make it possible to identify any patient. Failure to adhere to this will be followed by cancellation of the data from the website by the webmaster.
Second, data for each patient will be entered in a separate CRF. These data may be copied from paper CRFs in order to make the reporting of data from bed to computer station easier. All fields may be amended at any time until the patient's CRF is completed and closed. At this point, one may access the patients' CRF but it will be no longer be possible to amend the CRF. In the case of overt inconsistency, corrections must be detailed in writing (e-mail) by the person responsible for data quality for the centre or by the center itself. In all cases, no corrections will be permitted in the absence of an express written request. The person responsible for data quality will have access to the centre's CRF in printed form only. A registry will collect correspondence between the person responsible for data quality and the centre.
Third, completion of some fields in the CRF is mandatory. Failure to complete them will prevent progression to the following CRF and closure of the opened CRF. Failure to complete a CRF electronically will result in the patient being excluded from the study.
Finally, Each centre will be able to open CRFs for its own patients but never CRFs for patients from other centres.
Case report form compilation
A guide to CRF compilation is presented in Table 1.
Case report form: Admission (step 1)
This CRF will automatically provide the patient's consecutive number. The user must enter the following data:
• sex,
• date of birth,
• weight,
• height,
• date/time of hospital admission,
• premorbid plasma creatinine levels,
• date/time of ICU admission,
• diagnosis at admission,
• Simplified Acute Physiology Score (SAPS) II (the index will be automatically calculated once each requested field is completed),
• Sequential Organ Failure Assessment (SOFA; the index will be automatically calculated once each requested field is completed).
Case report form: Criteria to initiate RRT (step 2)
This CRF will automatically provide the patient's consecutive number. The user must enter the date and time when the following clinical events (indexed numerically) occurred:
• 1. Oliguria (urine output <200 ml/12 hours),
• 2. Anuria (urine output <50 ml/12 hours),
• 3. High urea/creatinine,
• 4. Hyperkalaemia (>6.5 mmol/l or rapidly rising potassium),
• 5. Metabolic acidosis,
• 6. Fluid overload,
• 7. Hyperthermia (>41°C),
• 8. Immunomodulation,
• 9. What RIFLE (Risk Injury Failure Loss of function End stage renal disease) criteria [12] are applicable?
• 10. Others (to specify)
The user will also be asked to prioritize the criteria (from 1 to 3) when two or more specified. In addition, the modality chosen must be specified (defined as following and indexed numerically)
• 1. CVVH (as defined as ≤ 35 ml/kg per hour ultrafiltration rate in postdilution or <40 ml/kg per hour in predilution),
• 2. CVVHDF (defined as use of dialysate + replacement [define]),
• 3. High volume haemofiltration (defined as >35 ml/kg per hour in postdilution or >45 ml/kg per hour in predilution),
• 4. Pulse high volume haemofiltration (from 85 ml/kg per hour to 100 ml/kg per hour for 6–8 hours, followed by CVVH at 35 ml/kg per hour),
• 5. Coupled plasma filtration adsorption (CPFA) plus CVVH ,
• 6. IHD ('intermittent' includes conventional haemodialysis and slow extended daily dialysis, thereby encompassing all treatments in which sessions are separated from one another for 10 hours or more).
The CRF will then permit the user to specify any other relevent criteria, including the following:
• Staff problems,
• Technical problems,
• Product (e.g. fluids, lines, filters, machine) availability problems,
• Logistics,
• Others (to be specified).
Case report form: Modality-specific assessment time (step 3)
On the basis of the modality chosen in the 'Criteria to initiate RRT' CRF (see above), a specific CRF will be opened. The CRF for IHD will require data entry at baseline, at 4 hours and at treatment end. The CRF will automatically indicate the different visits (i.e. 0.0, 4.0, and treatment end). The CRF for IHD will request the following information:
• Decision taken (date/time),
• Start (date/time),
• Prescribed duration (only at assessment time 0),
• Delivered duration,
• Prescribed blood flow rate (only at assessment time 0),
• Delivered blood flow rate,
• Total weight loss (kg/session),
• Type of haemodialyzer (specify only commercial name),
• Surface of haemodialyzer (m2),
• Type of buffer (code number for lactate or bicarbonate),
• Anticoagulation (code number for heparin, citrate, prostacyclin, saline flushes, no anticoagulation),
• Arterial site of vascular access (code number for radial, femoral, pedidial, axillary access),
• Venous site of vascular access (code number for subclavian catheter, femoral catheter, jiugular, axillary catheter),
• Type of vascular access (code number for double lumen catheter, single lumen catheter),
• Vascular access gauge,
• Treatment interrupted (date/time),
• Resumption of treatment (specify date/time),
• End (day/time),
• Change in modality.
The CRF for CVVH will require data entry at 0 hours, at 1 hour, 3 hours, 6 hours, 12 hours, 24 hours, and every 24 hours thereafter and at treatment end. The CRF will automatically indicate the different assessment times (i.e. 0.0, 1.0, 3.0, 6.0, 12.0, 24.0, and so forth). Assessment times at 1.0, 3.0, 6.0 and 12.0 are optional, while assessment time at 24.0 and for multiples of 24 are mandatory. The following information will be requested:
• Decision taken (date/time),
• Start (date/time),
• Prescribed duration (only in assessment time 0),
• Delivered duration,
• Prescribed blood flow rate (only in assessment time 0),
• Delivered blood flow rate,
• Prescribed effluent (ml/hour; only at assessment time 0),
• Total effluent (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Prescribed reposition rate (ml/hours; only at assessment time 0),
• Total reposition (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Total volume removed from the patient (ml/24 hours),
• Type of haemodialyzer (as above),
• Surface (m2),
• Type of buffer,
• Anticoagulation (as above),
• Arterial site of vascular access (as above),
• Venous site of vascular access (as above),
• Type of vascular access (as above),
• Vascular access gauge,
• Treatment interrupted (date/time),
• Resumption of treatment (specify date/time),
• End (day/time),
• Change in modality.
The CRF for CVVHD (Note: I would suggest to indicate the modality in bold for clarity's sake) will require data entry at 0 hours, at 1 hour, 3 hours, 6 hours, 12 hours, 24 hours, and every 24 hours thereafter and at treatment end. The CRF will automatically indicate the different assessment times (i.e. 0.0, 1.0, 3.0, 6.0, 12.0, 24.0, and so forth). Assessment times at 1.0, 3.0, 6.0 and 12.0 are optional, while assessment times at 24.0 and for multiples of 24 are mandatory. The following information will be requested:
• Decision taken (date/time),
• Start (date/time),
• Prescribed duration (only in assessment time 0),
• Delivered duration (only at assessment time 24 or last assessment time before treatment interruption/end),
• Prescribed blood flow rate (only at assessment time 0),
• Dialysate (ml/24 hours),
• Effluent (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Total volume removed from patient (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Type of haemodialyzer (as above),
• Surface (m2),
• Type of buffer,
• Anticoagulation (as above),
• Arterial site of vascular access (as above),
• Venous site of vascular access (as above),
• Type of vascular access (as above),
• Vascular access gauge,
• Treatment interrupted (date/time),
• Resumption of treatment (specify date/time),
• End (day/time),
• Change in modality.
The CRF for CVVHDF will require data entry at 0 hours, at 1 hour, 3 hours, 6 hours, 12 hours, 24 hours, and every 24 hours thereafter and at treatment end. Assessment times at 1.0, 3.0, 6.0 and 12.0 are optional, while assessment times at 24.0 and for multiples of 24 are mandatory. The CRF will automatically indicate the different assessment times (i.e. 0.0, 1.0, 3.0, 6.0, 12.0, 24.0, and so forth). The following information will be requested:
• Decision taken (date/time),
• Start (date/time),
• Prescribed duration (hours; only at assessment time 0),
• Delivered duration (hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Prescribed blood flow rate (ml/min; only at assessment time 0),
• Prescribed effluent (ml/hour; only at assessment time 0),
• Delivered effluent (ml/ 24 hour; only at assessment time 24 or last assessment time before treatment interruption/end),
• Prescribed reposition rate (ml/hour),
• Delivered reposition rate (ml/24 hours; (only at assessment time 24 or last assessment time before treatment interruption/end),
• Dialysate (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Total volume removed from the patient (ml/24 hours; only at assessment time 24 or last assessment time before treatment interruption/end),
• Type of haemodialyzer (as above),
• Surface (m2),
• Type of buffer,
• Anticoagulation (as above),
• Arterial site of vascular access (as above),
• Venous site of vascular access (as above),
• Type of vascular access (as above),
• Vascular access gauge,
• Treatment interrupted (date/time),
• Resumption of treatment (specify date/time),
• End (day/time),
• Change of modality.
Independently of modality chosen, all 'Modality-specific assessment time' CRFs include the following additional fields:
• SOFA (full set of data; only at assemement time 24 and for multiples of 24),
• Creatinine,
• Urea,
• Na,
• K,
• White blood cells (103/μl),
• Platelets (103/μl),
• Hb,
• pH,
• PaO2,
• PCO2,
• Bicarbonate,
• FiO2,
• Body temperature,
• Urine volume,
• Fluid balance (only at assessment time 24),
• Bicarbonate,
• Fractional inspired oxygen,
• Urine volume (ml/24 hours),
• Fluid balance (ml/24 hours),
• Systolic blood pressure (mmHg),
• Diastolic pressure (mmHg),
• Mixed venous oxygen saturation,
• Heart rate,
• Cardiac output,
• Cardiac index
• Pulmonary artery pressure,
• Systemic vascular resistance index,
• Intravascular blood volume index,
• Extravascular lung water index,
• Stroke volume variation,
• Vasopressor administration (milligrams of vasopressors/previous 24 hours): adrenaline (μg/kg per min), noradrenaline (μg/kg per min), dobutamine (μg/kg per min), dopamine (μg/kg per min), vasopressin (units/previous 24 hours), terlipressin (mg/previous 24 hours),
• Vasodilator administration,
• Other treatments: steroids (mg/24 hours; specify what type), recombinant human activated protein C, antithrombin III, protein C,
• Coagulation: activated partial thromboplastin time (diff versus control), activated clotting time (diff versus control), INR (%),
• Factors complicating RRT: logistics, organization, vascular access, anticoagulation, circuit patency, haemodialyzer performance.
Guidelines given in the CRF
'Treatment interruption' is defined as when a treatment is stopped and resumed within 18 hours. In the case of treatment interruption the CRF will be continued and the treatment that follows will be considered in the context of the preceding one. The only exception is when, after RRT interruption, the modality is changed (see below under 'Case report form: Change modality (step 3)'; Fig. 2).
'Treatment end' is defined as when a given RRT is stopped because of clinical or other factors for more than 12 hours or when clinical or other factors have changed since the start of RRT. Should the patient be started on another RRT, then the latter shall be considered a new one.
In the case that the modality is changed, a new CRF will need to be filled in (see 'Criteria to change RRT'). This will be followed by a new CRF 'Modality-specific assessment time' (also see Table 1).
Case report form: 'Change to modality'
Each centre will be asked to define the clinical/practical reasons for changing a modality. The change to modality may be necessary after treatment is interrupted. In this case, the following treatment will be considered a new treatment. This CRF aims to provide information on why the modality was chosen. It is similar to the CRF: Criteria to initiate RRT.
Case report form: Outcome (step 4)
At discharge of the patient, the following information should be provided:
• SAPS II (all sets of data; previous 24 hours before discharge from ICU),
• SOFA (all set of data; previous 24 hours before discharge from ICU),
• IHD needed in ward (yes/no),
• Creatinine at discharge (μmol/l; mg/%),
• Urea at discharge (μmol/l; mg/%),
• In-ICU mortality (yes/no),
• Ventilation days (number of days),
• 28-day mortality (yes/no),
• Discharged from ICU (date),
• Discharged from hospital (date),
• Hospital survival (yes/no),
• Date of last RRT session.
Calculation of dialysis dose
The dialysis dose will be calculated differently according to the type of modality. In the case of CVVH, solute transport is achieved by pure convection. The solute flux across the membrane is proportional to the ultrafiltration rate (Qf) and the ratio between the concentration of the solute in the ultrafiltrate and in plasma water (sieving coefficient S). For solutes freely crossing the membrane, S values are equal or close to 1. Because clearance is calculated from the product Qf × S, when S is proximal to 1, as for urea, clearance is assumed to be equal to Qf, provided that replacement solution is given in postdilution mode. For diffusive techniques (IHD or CVVHD), the clearances will be calculated on the basis of the delivered operational parameters on an experimentally constructed relationship (blood flow versus clearance) at three different dialysate flow rates (in CVVHD at 1 and 2 l/hour) for each given haemodialyzer. In mixed convective/diffusive techniques (e.g. CVVHDF), this relationship will constructed at two dialysate flows (1 and 2 L/hour) and at three ultrafiltration rates.
Statistical analysis
Primary end-point
The power for a test of the null hypothesis (logistic regression, one continuous predictor) was calculated as follows. The one goal of the proposed study was to test the null hypothesis (i.e. that there is no relationship between clearance and event rate). Under the null condition, the event rate (0.51) is the same at all values of clearance or, equivalently, the odds ratio is 1.0, the log odds ratio (beta) is 0.0 and the relative risk is 1.0.
Power is computed to reject the null hypothesis under the following alternate hypothesis. For clearance values of 29.8 and 35.0, the expected event rates are 0.51 and 0.25. This corresponds to an odds ratio of 0.32, beta (log odds ratio) of -0.22, and a relative risk of 0.49. This effect was selected as the smallest effect that would be important to detect, in the sense that any smaller effect would not be of clinical or substantive significance. It is also assumed that this effect size is reasonable, in the sense that an effect of this magnitude could be anticipated in this field of research. In these computations, we assume that the mean clearance value will be 29.8 with a standard deviation of 10.0, and that the event rate at this mean will be 0.51 (Figure 3).
The sample size will be of a total of 110 patients.
Alpha and tails. The criterion for significance (alpha) has been set at 0.01. The test is two-tailed, which means that an effect in either direction will be interpreted (Figure 4).
Power. For this distribution (clearance mean of 29.8, standard deviation of 10.0), baseline (mean event rate of 0.51), effect size (log odds ratio of -0.22), sample size (n = 110) and alpha (0.01, two-tailed), the power is 1.00. This means that close to 100% of studies would be expected to yield a significant effect, rejecting the null hypothesis that the odds ratio is 1.0 [13-16]. The software used will be Epi Info (Utilities StatCalc Epi Info™ version 3.3, release date: 5 October 2004; Division of Public Health Surveillance and Informatics, Centers for Disease Control and Prevention, Atlanta, GA, USA) [17] and Power And Precision™ (version 2.0; release date: 20 December 2000) [18].
Secondary end-point
Based on data from one participating center (Milan Niguarda), approximately 20% of all RRT-treated patients have high noradrenaline requirements. A sample size of 27 patients will have 80% power to detect a difference in means of 0.295 (e.g. a mean of 2.5 μg/kg per min, assuming a standard deviation for differences of 0.600, using a paired t-test with a 0.05 one-sided significance level). Therefore, a minimun of 135 patients should be enrolled. Assuming a 20% dropout rate, the minimum number of patients to be recruited is 162.
Study limitations
This will be an observational study. Based on the conventional meaning [19], an observational study cannot modify actual practice or therapy. In this study, the decision as to whether RRT should be commenced is at the discretion of the attending physician.
Discussion
The practice of CRRT has been subject to much debate. Only a few prospective randomized studies have been performed and published on the relationship between CRRT and outcome, and so conclusions are difficult to draw [20,21]. As emphasized in a recent editorial [22], in the field of artificial organs, prospective observational studies, despite their inherent limitations, have been performed because they are more affordable but are also capable of providing useful information from practical and medical standpoints.
Guerin and coworkers [23] studied 587 patients requiring haemodialysis and followed them until hospital discharge. Among the 587 patients, 354 received CRRT and 233 intermittent RRT as first choice. CRRT patients had a greater number of organ dysfunctions on admission and at the time of ARF, as well as higher SAPS II. Mortality was 79% in the CRRT group and 59% in the intermittent RRT group. Logistic regression analysis showed decreased patient survival to be associated with SAPS II on admission, oliguria, admission from hospital or emergency room, number of days between admission and ARF, cardiac dysfunction at time of ARF, and ischaemic ARF. No underlying disease or nonfatal disease, and absence of hepatic dysfunction were associated with an increase in patient survival. The type of RRT was not significantly associated with outcome. Those authors concluded that RRT mode was not of prognostic value.
The largest observational study ever performed (the BEST Kidney) was recently completed and reported in part [24]. A total of 1743 consecutive patients, who either were treated with RRT (CRRT or IHD) or fulfilled predefined criteria for ARF, were studied. Importantly, the findings indicated a marked difference in mortality rates across the different ICUs, suggesting that the practice of RRT may yet exert an influence on mortality [Bellomo R, unpublished observation]. Increasing the dose to 35 ml/kg per hour would be associated with a significantly greater survival in all ARF patients. However, higher dialysis doses (45 ml/kg per hour) had no statistically significant impact in the ARF patients studied. However, in a subgroup analysis including only those patients with sepsis, there was a trend suggesting that this might be the case.
Despite the numerous publications that suggest a benefit from delivering higher dialysis doses (for review [25]), the real impact in critically ill patients is unclear. An observational clinical survey to evaluate what modality, for what reasons and what outcomes are important is needed if we are to understand how dialysis is delivered and at what dose in routine ICU practice; what the benefits, if any, are in terms of haemodynamics; and, finally, what are the benefits in terms of patient outcome as the primary end-point.
Current treatments for multiorgan dysfunction with ARF include many forms of CRRT that differ with respect to following factors: dose of dialysis, the extent of convection and diffusion, flow rates (blood, dialysate and replacement fluids) and anticoagulation protocols (heparin, citrate, flushes of saline). Ancillary to these factors are the choices of predilution or postdilution, of haemodialyser (surface, membrane) and of vascular access. It is still unknown whether and to what extent the prescribed dose comforms with evidence-based literature and, more importantly, how the delivered dose diverges from the prescribed one.
The present study, as indicated in the present protocol, should help to resolve at least some aspects of this still largely undefined area of critical care.
Conclusion
The present study should provide insight into how RRT is currently practiced in ICUs and should hopefully provide answers to as yet undefined questions, such as the following: what are the criteria for beginning and ending treatment?; what is the currently delivered dose of dialysis?; how is fluid control taken care of?; what schedules are mostly used?; how is technology used (or not used)?; and, finally, what are the reasons for down-time in RRT? The ultimate goal will be to define how the dialysis dose actually delivered may impact on the outcome primary end-points of ICU patients.
Key messages
• Choice of RRT in renal and nonrenal indications
• Delivered dose of dialysis and its impact on outcome measures (primary end-point)
• Hemodynamic response to RRT (secondary end-point)
• Causes for down-time in CRRT
Abbreviations
ARF = acute renal failure; CRF = case report form; CRRT = continuous renal replacement therapy; CVVH = continuous venovenous haemofiltration; CVVHD = continuous venovenous haemodialysis; CVVHDF = continuous venovenous haemodiafiltration; ICU = intensive care unit; IHD = intermittent haemodialysis; IL = interleukin; RRT = renal replacement therapy; SAPS = Simplified Acute Physiology Score; SOFA = Sequential Organ Failure Assessment.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
The Scientific Committee comprised Kindgen-Milles D (Duesseldorf, Germany), Journois D (Paris, France); Fumagalli R (Bergamo, Italy), Ronco C (Vicenza, Italy), Vesconi S (Milan, Italy), Maynar J (Vitoria, Spain) and Marinho A (Porto, Portugal), who reviewed the different versions of the study protocol prepared by the Steering Committee and gave the final approval of the version to be published. The Steering Committee comprised the following individuals: Livigni S, Maio M (Torino, Italy), Marchesi M (Bergamo, Italy), Monti GP (Milano, Italy) and Silengo D (Torino, Italy), who made substantial contributions to conception and design and to establishing the CRF; Bolgan I (Vicenza, Italy) defined the way in which data will be analyzed and interpreted; Brendolan A (Vicenza, Italy), Formica M.(Cuneo, Italy), and Mariano F (Torino, Italy) helped in the definition of RRT modalities and reviewed the final case report forms.
Figures and Tables
Figure 1 Flowchart of the DO-RE-MI observational study. All incident patients admitted to the intensive care unit (ICU) and requiring renal replacement therapy (RRT) will be followed up during RRT. At discharge, primary and secondary end-points will be recorded. All data will be entered in electronic case report form (CRF) and stored in a website [11]. The rectangles indicate the type of information that will be available from this study. ARF, acute renal failure; DO-RE-MI, DOse REsponse Multicentre International collaborative initiative; SAPS, Simplified Acute Physiology Score.
Figure 2 Examples of how the different case report forms will be applied. Four different cases are summarized, encompassing treatment interruption or end in relation to the compilation of case report forms (CRFs). Case 1 is the easiest case. The patient is admitted to the intensive care unit (ICU), is treated with renal replacement therapy (RRT), ends treatment and is discharged. The patient has a single CRF. In case 2 the patient is admitted and is treated with RRT, but this treatment is stopped for longer than 18 hours (this is defined as treatment end). However, the patient is later started on RRT again. A new CRF (even if the modality is the same) will need to be completed. In this case, the patient has two or more CRFs (as in the case of more than one treatment stoppages for longer than 18 hours). In case 3 the patient is admitted and is started on RRT, which is stopped for less than 18 hours (defined as interruption). The patient is then restarted and the compilation is continued on the same CRF. Case 4 is similar to case 3, with the important difference being related to the change in modality following treatment interruption. In this case, each change of modality will require a new CRF.
Figure 3 Mortality rate as a function of dialysis dose (expressed as urea clearance ml/min).
Figure 4 Power as a function of sample size.
Table 1 Guide to case report form compilation
Step Details
Step 1 Complete CRF: Admission
Press 'save' and open CRF: Criteria to initiate RRT
Step 2 Complete CRF: Criteria to initiate RRT:
• Indicate one or more criteria to initiate RRT and their priority score (from 1 [low] to 3 [high])
• In patients with ARF, choose which RIFLE criteria are applicable
• Indicate what you expect to happen using the technique you have chosen
Press 'save'; the next CRF for the chosen modality will automatically open
Step 3 Complete CRF: Change to modality:
• Fill in all mandatory fields using the measure/legend
• Be advised that there is one CRF for each hour of observation. This depends on the chosen RRT modality (for IHD: 0.0 hours, 4.0 hours and treatment end; for CVVH, CVVHD, CVVHDF, HVHF, CPFA: 0.0 hours, at 1.0 hour, 3.0 hours, 6.0 hours, 12.0 hours, 24.0 hours, and every 24 hours thereafter and at treatment end)
• In the case of treatment interruption or end, specify date/time (for definition of treatment interruption or treatment end, see under 'Guidelines given in the CRF', in the text)
• In the case of change of treatment modality after treatment interruption, fill in CRF: Criteria to modality. Then go back to the start of step 3. Once in CRF: Change to modality, do not forget to select the new modality chosen
Press 'save' and terminate CRF: Change to modality
Step 4 At discharge, please complete CRF: Outcome
ARF, acute renal failure; CPFA, coupled plasma filtration adsorption; CRF, case report form; CVVH, continuous venovenous haemofiltration; CVVHD, continuous venovenous haemodialysis; CVVHDF, continuous venovenous haemodiafiltration; HVHF, high-volume haemofiltration; RIFLE, Risk Injury Loss of fucntion End stage renal disease; RRT, renal replacement therapy
==== Refs
Cohen J The immunopathogenesis of sepsis Nature 2002 420 885 891 12490963 10.1038/nature01326
Cavaillon JM Munoz C Fitting C Misset B Carlet J Circulating cytokines: The tip of the iceberg? Circ Shoc 1992 38 145 152
Pinsky MR Sepsis: a pro- and anti-inflammatory disequilibrium syndrome Contrib Nephrol 2001 132 354 366 11395903
Cavaillon JM Adib-Conquy M Cloez-Tayarani I Fitting C Immunodepression in sepsis and SIRS assessed by ex vivo cytokine production is not a generalized phenomenon: a review J Endotoxin Res 2001 7 85 93 11521088 10.1179/096805101101532576
Kramer P Wigger W Rieger J Matthaei D Scheler F Arteriovenous haemofiltration: a new and simple method for treatment of over-hydrated patients resistant to diuretics Klin Wochenschr 1977 55 1121 1122 592681 10.1007/BF01477940
Mehta RL Letteri JM Current status of renal replacement therapy for acute renal failure. A survey of US nephrologists. The National Kidney Foundation Council on Dialysis Am J Nephrol 1999 19 377 382 10393374 10.1159/000013481
Marshall MR Golper TA Shaver MJ Alam MG Chatoth DK Sustained low-efficiency dialysis for critically ill patients requiring renal replacement therapy Kidney Int 2001 60 777 785 Erratum in Kidney Int 2001 60:1629. 11473662 10.1046/j.1523-1755.2001.060002777.x
Schiffl H Lang SM Fischer R Daily hemodialysis and the outcome of acute renal failure N Engl J Med 2002 346 305 310 11821506 10.1056/NEJMoa010877
Ronco C Bellomo R Homel P Brendolan A Dan M Piccinni P La Greca G Effects of different doses in continuous veno-venous haemofiltration on outcomes of acute renal failure: a prospective randomized trial Lancet 2000 356 26 30 10892761 10.1016/S0140-6736(00)02430-2
Brendolan A D'Intini V Ricci Z Bonello M Ratanarat R Salavtori G Bordoni V DeCal M Andrikos E Ronco C Pulse high volume haemofiltration Int J Artif Organs 2004 27 398 403 15202817
Acute Vision
Bellomo R Acute renal failure – definition, outcome measures, animal models, fluid therapy and information technology needs: the Second International Consensus Conference of the Acute Dialysis Quality Initiative (ADQI) Group Crit Care 2004 8 R204 R212 15312219 10.1186/cc2872
Fleiss JL Statistical Methods for Rates and Proportions 1981 2 Wiley & Sons, New York
Fabbris L L'indagine Campionaria(Statistical Survey) 1996 NIS, Nuova Italia Scientifica; Rome
Borenstein M Planning for precision in survival studies J Clin Epidemiol 1994 47 1277 1285 7722564 10.1016/0895-4356(94)90133-3
Rothstein H Borenstein M Cohen J Pollack S Statistical power analysis for multiple regression / correlation: A computer program Educational & Psychological Measurement 1995 50 819 830
EpiInfo (utilities StatCalc), Version 33 Atlanta, USA
Power And Precision, Relase 20 New Jersey, USA
Rosenbaum PR Observational Studies 2002 2 Berlin: Springer
Metha R McDonald B Gabbai FB Pahl M Pascual MT Farkas A Kaplan RM Collaborative Group for Treatment of ARF in the ICU A randomized clinical trial of continuous versus intermittent dialysis for acute renal failure Kidney Int 2001 60 1154 1163 11532112 10.1046/j.1523-1755.2001.0600031154.x
Tonelli M Manns B Feller-Kopman D Acute renal failure in the intensive care unit: a systematic review of the impact of dialytic modality on mortality and renal recovery Am J Kidney Dis 2002 100 158 160
Ronco C Evidence-based medicine: can we afford it? Int J Artif Organs 2004 27 819 820 15560674
Guerin C Girard R Selli JM Ayzac L Intermittent versus continuous renal replacement therapy for acute renal failure in intensive care units: results from a multicenter prospective epidemiological survey Int Care Med 2002 28 1411 1418 10.1007/s00134-002-1433-0
Uchino S Doig GS Bellomo R Morimatsu H Morgera S Schetz M Tan I Bouman C Nacedo E Gibney N Beginning and Ending Supportive Therapy for the Kidney (B.E.S.T. Kidney) Investigators. Diuretics and mortality in acute renal failure Crit Care Med 2004 32 1669 1677 15286542 10.1097/01.CCM.0000132892.51063.2F
Tetta C Bellomo R Kellum J Ricci Z Pohlmeier R Passlick-Deetjen J High volume hemofiltration critically ill patients: why, when and how? Contrib Nephrol 2004 144 362 375 15264423
|
16137353
|
PMC1269446
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 14; 9(4):R396-R406
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3718
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37221613734610.1186/cc3722ResearchDecrease in serum procalcitonin levels over time during treatment of acute bacterial meningitis Viallon Alain [email protected]'h Pantéa 1Guyomarc'h Stéphane 1Tardy Bernard 1Robert Florianne 1Marjollet Olivier 1Caricajo Anne 2Lambert Claude 3Zéni Fabrice 1Bertrand Jean-Claude 11 Emergency and Intensive Care Units, Bellevue Hospital, Saint-Etienne, France2 Microbiology Laboratory, Bellevue Hospital, Saint-Etienne, France3 Immunology Laboratory, Bellevue Hospital, Saint-Etienne, France2005 20 5 2005 9 4 R344 R350 23 2 2005 9 3 2005 25 4 2005 27 4 2005 Copyright © 2005 Viallon et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
The aim of this study was to describe the change in serum procalcitonin levels during treatment for community-acquired acute bacterial meningitis.
Methods
Out of 50 consecutive patients presenting with bacterial meningitis and infection at no other site, and who had received no prior antibiotic treatment, 48 had a serum procalcitonin level above 0.5 ng/ml on admission and were enrolled in the study.
Results
The mean age of the patients was 55 years, and mean Glasgow Coma Scale score on admission was 13. The time from symptom onset to admission was less than 24 hours in 40% of the patients, 24–48 hours in 20%, and more than 48 hours in 40%. The median (interquartile) interval between admission and initial antibiotic treatment was 160 min (60–280 min). Bacterial infection was documented in 45 patients. Causative agents included Streptococcus pneumoniae (n = 21), Neisseria meningitidis (n = 9), Listeria monocytogenes (n = 6), other streptococci (n = 5), Haemophilus influenzae (n = 2) and other bacteria (n = 2). The initial antibiotic treatment was effective in all patients. A lumbar puncture performed 48–72 hours after admission in 34 patients showed sterilization of cerebrospinal fluid. Median (interquartile) serum procalcitonin levels on admission and at day 2 were 4.5 (2.8–10.8) mg/ml and 2 (0.9–5.0) mg/ml, respectively (P < 0.0001). The corresponding values for C-reactive protein were 120 (21–241) mg/ml and 156 (121–240) mg/ml, respectively. Five patients (10%) died from noninfectious causes during their hospitalization.
Conclusions
Serum procalcitonin levels decrease rapidly with appropriate antibiotic treatment, diminishing the value of lumbar puncture performed 48–72 hours after admission to assess treatment efficacy.
==== Body
Introduction
Community-acquired acute bacterial meningitis (ABM) in adults remains a serious disease, with mortality rates of 10–25% [1,2]. In the context of emergency presentation, the management decisions to be made once the diagnosis has been established concern the initial antibiotic treatment [2], adjuvant therapies [3,4] and treatment of organ failure [5].
Antibiotic treatment must be started rapidly [6] and must be appropriate, particularly when risk factors are present [7,8], although the timing of antibiotic therapy initiation does not appear to be an independent prognostic factor [9-11]. The choice of antibiotic treatment, addressed in numerous national [12] and international [2,13-15] recommendations, is based on aetiological indices, risk factors, the results of direct examinations and knowledge of bacterial ecology.
The efficacy of this initial antibiotic therapy is assessed on the basis of clinical course of the disease and analysis of cerebrospinal fluid (CSF) samples obtained 48–72 hours after the start of treatment, when available, although cytochemical CSF parameters appear to be little modified by appropriate antibiotic treatment [16]. A marker that can demonstrate efficacy at an earlier stage would be extremely useful.
In 1993 Assicot and coworkers [17] demonstrated the value of serum procalcitonin (PCT) as a marker of infectious states of bacterial orgin in neonates and infants, as well as the rapid decrease in its concentration with appropriate antibiotic treatment. The aim of the present study was to describe the variation in serum PCT levels over time during the treatment of ABM.
Materials and methods
Patients
This was a prospective study and included patients admitted to the adult emergency department with community-acquired bacterial meningitis between January 1997 and October 2003. The demographic and clinical characteristics of the patients were recorded on admission.
Bacterial meningitis was diagnosed if pathogenic bacteria were detected in the CSF. In the absence of documented evidence of bacterial infection, this diagnosis was made if the polymorphonuclear leucocyte count in the CSF exceeded 250/mm3 and the CSF/serum glucose ratio was below 0.4, with a compatible clinical state, necessitating antibiotic treatment for 7 days or longer. Patients presenting with a further site of infection in addition to meningitis on admission, having received prior antibiotic treatment for more than 2 consecutive days or showing a serum PCT level of 0.5 mg/ml or less, were excluded from the study.
Laboratory tests
Blood samples for C-reactive protein (CRP), PCT, fibrinogen, lactate and creatinine assays, and complete blood count were taken on admission, then once daily during the first week. Lumbar puncture (for total and polymorphonuclear leucocyte count and assay of proteins, lactate and glucose) and bacteriological sampling (blood cultures) were performed before starting the initial antibiotic treatment. These tests could be repeated between 48 and 72 hours later at the discretion of the clinician.
The interval between admission and administration of the first dose of antibiotic was recorded. Bacterial sensitivity to antibiotics was routinely tested by determining the minimum inhibitory concentrations (MICs) of penicillin, amoxicillin, cefotaxime and ceftriaxone. With regard to penicillin, bacteria were considered to be sensitive if the MIC was 0.1 mg/l or less, of intermediate resistance if the MIC was above 0.1 mg/l but no greater than 1 mg/l, and highly resistant if the MIC was above 1 mg/l. For amoxicillin, cefotaxime and ceftriaxone, bacteria were considered to be sensitive if the MIC was 0.5 mg/l or less, of intermediate resistance if the MIC was above 0.5 mg/l but no greater than 2 mg/l, and highly resistant if the MIC was greater than 2 mg/l.
Serum PCT levels were determined using an immunoluminometric assay (Brahms Diagnostica, Berlin, Germany) with a limit of detection of 0.07 mg/ml.
Treatment and course of illness
The efficacy of initial antibiotic treatment was assessed on the basis of in vitro bacterial sensitivity to antibiotics, bacteriological analysis of CSF samples drawn 48–72 hours after treatment initiation, and clinical course. The nature and duration of antibiotic treatment, and any modifications to this, were recorded. Mortality and sequelae were assessed at 30 days.
Statistical analysis
Results are expressed as mean ± standard deviation or as median (interquartile range). The box plots are presented with the interquartile range. The values at day 0 (D0; admission) and day 2 (D2) were compared using Wilcoxon's nonparametric test for repeated measurements for quantitative parameters and the χ2 test for qualitative parameters, with the threshold of significance set at P < 0.05.
Results
During the study period (82 months), 59 patients presenting with ABM were admitted to the emergency department. Eleven patients were excluded for the following reasons: antibiotic treatment before admission (n = 6), presence of another site of infection (pneumopathy, n = 2; spontaneous bacterial peritonitis, n = 1), and serum PCT concentration 0.5 mg/ml or less on admission (one ABM due to pneumococci, one ABM due to unidentified bacteria).
The clinical characteristics of the 48 patients are summarized in Table 1. The mean interval between admission and lumbar puncture was 90 ± 40 min and the mean time elapsing from admission to injection of the first dose of antibiotic was 120 ± 70 min.
The microbiological results obtained are shown in Table 2. Among the 21 pneumococcal infections documented, 13 isolates (62%) were sensitive to penicillin, six (29%) were of intermediate resistance and two (10%) were resistant. Among the six strains with intermediate resistance to penicillin, four were sensitive to amoxicillin or to ceftriaxone, and two exhibited intermediate resistance. The two strains resistant to penicillin exhibited intermediate resistance to amoxicillin or to ceftriaxone. The initial antibiotic treatment comprised amoxicillin (150–200 mg/kg per day), ceftriaxone (60–80 mg/kg per day), or a combination of these. All of the pneumococcal strains with reduced sensitivity to penicillin were at least exposed to ceftriaxone during the initial treatment, with analysis of CSF samples drawn between 48 and 72 hours after the start of treatment showing sterilization in all cases. Antibiotic treatment was simplified eight times out of 21 on the basis of the results of microbiological analysis of CSF samples.
All of the other bacteria identified were sensitive to amoxicillin, with the initial antibiotic therapy being appropriate in all cases. The treatment was simplified between 24 and 72 hours after the start of treatment in six patients out of 24 on the basis of the results of microbiological analysis of the CSF.
For the three patients with a CSF culture not showing any evidence of bacteria, antibiotic treatment with amoxicillin and ceftriaxone was started on admission and continued for 15–20 days.
The changes in serum and CSF cytochemical parameters are shown in Tables 3 and 4 and in Fig. 1. With regard to the CSF, only the lactate concentration differed significantly between D0 and D2. Sterilization of the CSF was noted in the 34 patients who underwent a second lumbar puncture. In the 14 patients who did not undergo a repeat lumbar puncture, the duration of antibiotic treatment was 12–16 days, resulting in cure in all cases, and the duration of hospital stay was between 13 and 18 days. With respect to serum parameters, the decrease in PCT level was the only significant difference observed between D0 and D2.
Among the 48 patients, five patients (9%) died between 12 and 28 days after their admission to hospital. Only one of these patients was younger than 75 years. All of these patients underwent a second lumbar puncture during treatment, with analysis of the resulting sample showing sterilization of the CSF in all cases, which was confirmed by a third lumbar puncture in three of the five patients. A serum PCT concentration below 0.5 ng/ml was observed in all patients between 6 and 9 days after admission. The cause of death was multiple organ failure (n = 1), cerebral thrombophlebitis (n = 1) and cerebral oedema (n = 3). Four patients had neurological sequelae at 30 days.
Discussion
In the present study a significant and early decrease in serum PCT concentration was associated with cure of meningitis. In contrast, analysis of CSF showed a significant decrease only in lactate concentration between 48 and 72 hours after the first lumbar puncture.
The value of repeat lumbar puncture at 48 hours remains debatable, and second-line antibiotic treatment is based essentially on the MIC of various antibiotics for the bacteria identified or on the clinical course [2,6,12,13,18]. Apart from the microbiological data, the CSF parameters traditionally described during ABM appear to be little modified by appropriate antibiotic therapy within 48 hours. Blazer and coworkers [16], studied the effect of antibiotic treatment on the CSF parameters of 68 children presenting with ABM. None of the cytochemical parameters studied (proteins, glucose, total and polymorphonuclear leucocytes) exhibited a significant decrease between the first lumbar puncture and a second lumbar puncture performed 44–68 hours after the start of antibiotic therapy, whereas two bacteria were still detectable in the repeat CSF samples drawn. Similar findings were reported by Bland and coworkers [19] concerning the changes in these cytochemical parameters after 24–72 hours of treatment for ABM in 15 children.
Different results were obtained in an animal study [20]. In five sheep, treatment for an experimentally induced meningitis due to Escherichia coli resulted in a rapid decrease in polymorphonuclear leucocyte count in the CSF, which was associated with an increase in glucose concentration and a decrease in protein concentration. However, in that study antibiotic treatment was administered intrathecally. In the present study there was no significant decrease in the polymorphonuclear leucocyte count or protein concentration in the CSF after 48–72 hours of appropriate antibiotic treatment. The glucose concentration measured in the CSF remained stable, but there was a significant decrease in lactate concentration.
Although numerous articles have demonstrated the value of assaying lactate during the course of ABM [19,21-27], few data exist concerning the changes in this parameter during the treatment of this disease. In 21 patients with ABM, Gontroni and coworkers [23] showed a rapid decrease in lactate concentration in the CSF during the first 24 hours of treatment. Gould [22], Bland [19] and Genton [21] and their groups obtained similar results concerning the change in lactate concentration in CSF after 24–72 hours of treatment in 6, 15 and 25 patients with ABM, respectively. In the study reported by Bland and coworkers [19], the mean lactate concentration in the CSF was 75.1 ± 6.6 mg/100 ml at the time of the first lumbar puncture and 49.5 ± 5.7 mg/100 ml after 24–72 hours of treatment.
With regard to the changes in serum parameters, the present study revealed a rapid decrease in PCT concentration within the first 24 hours of treatment, which was accompanied by an increase in CRP, with the level of CRP diminishing only after 2–3 days.
In 1993, Assicot and coworkers [17] demonstrated that serum PCT concentration was a marker of infectious states of bacterial origin in children, exhibiting a rapid decrease following antibiotic treatment. Although several studies have demonstrated the value of serum PCT concentration in the differential diagnosis of ABM and viral or aseptic meningitis [28-31], few data are available concerning the change in serum PCT during treatment for ABM. Schwartz and colleagues [31] reported a reduction in median serum PCT concentration from 1.75 mg/ml at baseline to 1.05 mg/ml after 48 hours of treatment in 11 patients with ABM. In three of these patients, the PCT concentration remained unchanged, or increased, in conjunction with an unfavourable clinical course. In the study reported by Gendrel and coworkers [28], conducted in eight children receiving treatment for ABM, the serum PCT concentration diminished within 24 hours of treatment in all but two cases.
Although appropriate antibiotic therapy appears to be correlated with a rapid decrease in PCT levels, the absence of patients receiving an inappropriate treatment in our series did not allow us to determine the change in PCT levels under these circumstances. What are the arguments in support of a relationship between decrease in PCT levels and appropriate antibiotic treatment? Smith and coworkers [32] investigated the value of PCT in 43 patients presenting with melioidiosis of various grades of severity. Among the 16 patients with a severe infection, 13 exhibited a decrease in PCT levels from the first day of treatment. In the three other patients an increase in PCT levels was observed in relation to infectious complications (pulmonary abscess, septic arthritis, splenic abscess). In two patients the Pseudomonas pseudomallei infection detected was resistant to the initial antibiotic therapy.
Although the change in serum levels of CRP has been shown to be of value for tracking the course of a bacterial infection during treatment [33,34], the characteristics of this protein are such that its concentration reaches a maximum only after 24–48 hours [35]; this is in contrast to PCT, which attains a peak serum concentration more rapidly. After injection of endotoxin, the peak serum concentration of PCT is reached within approximately 8 hours [36].
Certain limitations of the present study should be mentioned. This was a descriptive study of the variation in serum PCT concentrations over time in patients who had received appropriate antibiotic treatment from the moment they were admitted to hospital. We currently have no data on changes in serum PCT levels occurring in patients who did not receive suitable treatment. Two patients with bacterial meningitis were not included in the study on the grounds that they presented with a serum procalcitonin level below 0.5 ng/ml on admission. At present there is no clear explanation for this finding. Several studies have reported low levels of serum PCT during ABM [30,31,37]. For the most part, this occurred in patients presenting with bacterial meningitis caused by intracellular bacteria or nosocomial infections [31,37].
Conclusion
The change in serum PCT level during treatment for community-acquired ABM appears to be a valuable parameter for evaluating the efficacy of antibiotic therapy. This hypothesis needs confirmation, particularly in patients presenting with bacterial meningitis that is not microbiologically documented.
Key messages
• After appropriate antibiotic treatment, serum PCT level decrease within the first 24 hours.
• After appropriate antibiotic teatment, serum CRP level decrease between days 2 and 3.
• The value of repeat lumbar puncture at 48 hours remains debatable.
• We have no data on changes in serum PCT levels in patients who do not receive an appropriate antibiotic.
• Some patients presenting with ABM have a low serum PCT level.
Abbreviations
ABM = acute bacterial meningitis; CRP = C-reactive protein; CSF = cerebrospinal fluid; MIC = minimum inhibitory concentration; PCT = procalcitonin.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
AV conceived of the study, and participated in its design and coordination and drafted the manuscript. PG participated in the inclusion and treatment of patients and drafted the manuscript. SG performed the statistical analysis. BT participated in the inclusion and treatment of patients. FR participated in the inclusion and treatment of patients. OM participated in the inclusion and treatment of patients. AC carried out the the microbiology. CL carried out the immunoassays. FZ participated in the design of the study and drafted the manuscript. JCB helped to draft the manuscript. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 Evolution of CRP and PCT levels over 72 hours. Change in serum levels of C-reactive protein (CRP) and procalcitonin (PCT) from admission (day 0) to 72 hours after the start of treatment.
Table 1 Patient characteristics on admission
Characteristics Number of patients (n = 48)
Demographic characteristics
Age (years; mean ± SD) 55 ± 21
Patients >75 years (n [%]) 8 (17)
Male (n [%]) 21 (44)
Duration of symptoms (hours; n [%])
<12 hours 4 (8)
12–23 hours 15 (31)
24–48 hours 10 (21)
>48 hours 19 (40)
Clinial characteristics (n [%])
Fever (>38°) 43 (90)
Headache 29 (60)
Nuchal rigidity 35 (73)
Seizures 4 (8)
Purpura 5 (10)
Focal neurological deficit 5 (10)
Glasgow Coma Scale score (mean ± SD) 13 ± 2
Simplified Acute Physiology Score II (mean ± SD) 18 ± 10
SD, standard deviation.
Table 2 Bacteriology of CSF on admission
Organism CSF Gram stain (n = 48) Culture (n = 48)
Streptococcus pneumoniae (n [%]) 12 (25) 21 (44)
Neisseria meningitidis (n [%]) 6 (12) 9 (19)
Other streptococci (n [%]) 0 5 (10)
Listeria monocytogenes (n [%]) 0 6 (13)
Haemophilus influenzae (n [%]) 0 2 (4)
Escherichia coli (n [%]) 0 1 (2)
Staphylococcus aureus (n [%]) 0 1 (2)
Total (n [%]) 18 (38) 45 (94)
CSF, cerebrospinal fluid.
Table 3 Cytochemical parameters of CSF and CSF/serum ratio on admission and after 2–3 days of treatment
Parameter Day 0 (admission; n = 48) Day 2 (2–3 days of treatment; n = 34)
Leucocyte count (cells/mm3) 757 (366–2730) 580 (309–2025)
Polymorphonuclear leucocyte count (cells/mm3) 605 (258–2482) 417 (254–1762)
Protein level (g/l) 4.2 (2–6.2) 3.9 (1.8–5)
Glucose CSF level (mmol/l) 2.4 (0.8–3.6) 2.5 (1.2–3)
CSF/serum glucose ratio 0.31 (0.1–0.48) 0.35 (0.17–0.5)
Lactate CSF level (mmol/l) 8.74 (5.5–13) 5* (3–9)
CSF/serum lactate ratio 3.22 (2.4–4.7) 2.73 (1.5–3.3)
Values are expressed as median (interquartile range). CSF, cerebrospinal fluid. *P < 0.001.
Table 4 Change in serum biological parameters from admission to day 4 of treatment
Parameter Day 0 Day 1 Day 2 Day 3 Day 4
Leucocyte (109/l) 14.2 (10–19) 14 (11.8–18) 13.5 (11–17) 10.5 (8.5–12) 10.2 (8–13)
Fibrinogen (g/l) 4.8 (4–6.4) 6.2 (5–7.5) 6.2 (5.6–8.2) 6.4 (5.8–8.4) 6.3 (5–8)
C-reactive protein (mg/l) 120 (48–241) 221 (141–299) 156 (121–240) 93 (67–170) 82 (43–130)
Procalcitonin ng/ml) 4.5 (2.8–10.8) 3.8 (1.5–10.7) 2* (0.9–5) 1.4 (0.4–3) 0.7 (0.4–1.3)
Values are expressed as median (interquartile range). *P < 0.0001 versus day 0.
==== Refs
Durand ML Caldewood SB Weber DJ Miller SI Southwick FS Caviness VS Swartz MN Acute bacterial meningitis in adults. A review of 493 episodes N Engl J Med 1993 328 21 28 8416268 10.1056/NEJM199301073280104
Quagliarello VJ Scheld WM Treatment of bacterial meningitis N Eng J Med 1997 336 708 716 10.1056/NEJM199703063361007
De Gans J van de Beek D Dexamethasone in adults with bacterial meningitis N Engl J Med 2002 347 1549 1556 12432041 10.1056/NEJMoa021334
Lindvall P Ahlm C Ericsson M Gothefors L Naredi S Koskinen LO Reducing intracranial pressure may increase survival among patients with bacterial meningitis Clin Infect Dis 2004 38 384 390 14727209 10.1086/380970
Rivers E Nguyen B Havstad S Ressler J Muzzin A Knoblich B Peterson E Tomlanovich M Early goal-directed therapy in the treatment of severe sepsis and septic shock N Engl J Med 2001 345 1368 1377 11794169 10.1056/NEJMoa010307
Tunkel AR Scheld WM Acute bacterial meningitis Lancet 1995 346 1675 1680 8551828 10.1016/S0140-6736(95)92844-8
Aronin SI Peduzzi P Quagliarello VJ Community-acquired bacterial meningitis: risk stratification for adverse clinical outcome and effect of antibiotic timing Ann Intern Med 1998 129 862 869 9867727
Meyer CN Samuelsson IS Galle M Bangsborg JM Adult bacterial meningitis: aetiology, penicillin susceptibility, risk factors, prognostic factors and guidelines for empirical antibiotic treatment Clin Microbiol Infect 2004 10 709 717 15301673 10.1111/j.1469-0961.2004.00925.x
Kipli T Anttila M Kallio MJ Peltola H Lenght of prediagnostic history related to the course and sequelae of childhood bacterial meningits Pediatr Infect Dis J 1993 12 184 188 8451093
Kallio MJ Kilpi T Anttila M Peltola H The effect of a recent previous visit to a physician on outcome after childhood bacterial meningitis JAMA 1994 272 787 791 8078143 10.1001/jama.272.10.787
Lebel MH McCracken GH Jr Delayed cerebrospinal fluid sterilization and adverse outcome of bacterial meningitis in infants and children Pediatrics 1989 83 161 167 2913547
Anonymous Community-acquired purulent meningitis. Short text of the 9th consensus conference on anti-infectious therapy [in French] Presse Med 1998 27 1145 1150
Begg N Cartwright KAV Cohen J Kaczmarski EB Innes JA Leen CL Nathwani D Singer M Southgate L Todd WT Consensus statement on diagnosis, investigation, treatment and prevention of acute bacterial meningitis in immunocompetent adults J Infect 1999 39 1 15 10468122 10.1016/S0163-4453(99)90095-6
Van de Beek D de Gans J Spanjaard L Vermeulen M Dankert J Antibiotic guidelines and antibiotic use in adult bactérial meningitis in The Nertherlands J Antimicrob Chemother 2002 49 661 666 11909840 10.1093/jac/49.4.661
Moller K Skinhoj P Guidelines for managing acute bacterial meningitis BMJ 2003 320 1290 1292 10.1136/bmj.320.7245.1290
Blazer S Berant M Alon U Bacterial meningitis. Effect of antibiotic treatment on cerebrospinal fluid Am J Clin Pathol 1983 80 386 387 6881104
Assicot M Gendrel D Carsin H Raymond J Guilbaud J Bohuon C High serum procalcitonin concentrations in patients with sepsis and infection Lancet 1993 341 515 518 8094770 10.1016/0140-6736(93)90277-N
Heyderman RS Lambert HP O'Sullivan I Stuart JM Taylor BL Wall RA Early management of suspected bacterial meningitis and meningococcal septicaemia in adults J Infect 2003 46 75 77 12634067 10.1053/jinf.2002.1110
Bland RD Lister RC Ries JP Cerebrospinal fluid lactic acid level and pH in meningitis Am J Dis Child 1974 128 151 156 4850943
Nazifi S Rezakhani A Badran M Evaluation of hematological, serum biochemical and cerebrospinal fluid parameters in experimental bacterial meningitis in the calf J Vet Med A 1997 44 55 63
Genton B Berger JP Cerebrospinal fluid lactate in 78 cases of adult meningitis Intensive Care Med 1990 16 196 200 2191022
Gould IM Irwin WJ Wadhwani RR The use of cerebrospinal fluid lactate determination in the diagnosis of meningitis Scand J Infect Dis 1980 12 185 188 7433918
Gontroni G Rodriguez WJ Deane CA Hicks JM Ross S Cerebrospinal fluid lactate determination: a new parameter for the diagnosis of acute and partially treated meningitis Chemotherapy 1976 1 175 182
Berg B Gärdsell P Skansberg P Cerebrospinal fluid lactate in the diagnosis of meningitis. Diagnostic value compared to standard biochemical methods Scand J Infect Dis 1982 14 111 115 7100821
Curtis GDW Slack MPE Tompkins DS Cerebrospinal fluid lactate and the diagnosis of meningitis J Infect 1981 3 159 165 7185956
Spranger M Schwab S Krempien S Maiwald M Bruno K Hacke W Excess glutamate levels in the cerebrospinal fluid predict clinical outcome of bacterial meningitis Arch Neurol 1996 53 992 996 8859061
Knight JA Dudek SM Haymond RE Early (chemical) diagnosis of bacterial meningitis: cerebrospinal fluid glucose, lactate, and lactate dehydrogenase compared Clin Chem 1981 27 1431 1434 7273404
Gendrel D Raymond J Assicot M Moulin F Iniguez JL Lebon P Bohuon C Measurement of procalcitonin levels in children with bacterial or viral meningitis Clin Inf Dis 1997 24 1240 1242
Viallon A Zeni F Lambert C Pozzetto B Tardy B Venet C Bertrand JC High sensitivity and specificity of serum procalcitonin levels in adults with bacterial meningitis Clin Infect Dis 1999 28 1313 1316 10451174
Jereb M Muzlovic I Hojker S Strle F Predictive value of serum and cerebrospinal fluid procalcitonin levels for the diagnosis of bacterial meningitis Infection 2001 29 209 212 11545482 10.1007/s15010-001-1165-z
Schwarz S Bertram M Schwab S Andrassy K Hache W Serum procalcitonin levels in bacterial and abacterial meningitis Crit Care Med 2000 28 1828 1832 10890628 10.1097/00003246-200006000-00024
Smith MD Suputtamongkol Y Chaowagul W Assicot M Bohuon C Petitjean S White NJ Elevated serum procalcitonin levels in patients with melioidosis Clin Infect Dis 1995 20 641 645 7756489
Povoa P C-reactive protein: a valuable marker of sepsis Intensive Care Med 2002 28 235 243 11904651 10.1007/s00134-002-1209-6
Cox ML Rudd AG Gallimore R Hodkinson HM Pepys MB Real-time measurement of serum C-reactive protein in the management of infection in the elderly Age Ageing 1986 15 257 266 3776747
Mary P Veinberg F Couderc R Acute meningitis, acute phase proteins and procalcitonin Ann Biol Clin 2003 61 127 137
Dandona P Nix D Wilson MF Aljada A Love J Assicot M Bohuon C Procalcitonin increase after endotoxin injection in normal subjects J Clin Endocrinol Metab 1994 79 1605 1608 7989463 10.1210/jc.79.6.1605
Hoffmann O Reuter U Masuhr F Holtkamp M Kassim N Weber JR Low sensitivity of serum procalcitonin in bacterial meningitis in adults Scand J Infect Dis 2001 33 215 218 11303813 10.1080/00365540151060905
|
16137346
|
PMC1269448
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 May 20; 9(4):R344-R350
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3722
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37241613735110.1186/cc3724ResearchIntrapulmonary percussive ventilation in acute exacerbations of COPD patients with mild respiratory acidosis: a randomized controlled trial [ISRCTN17802078] Vargas Frédéric [email protected] Hoang Nam 1Boyer Alexandre 1Salmi Louis Rachid 3Gbikpi-Benissan Georges 1Guenard Hervé 2Gruson Didier 1Hilbert Gilles 121 Département de Réanimation Médicale, Hôpital Pellegrin-Tripode, Bordeaux, France2 Laboratoire de Physiologie EA 518, Université Victor Segalen Bordeaux 2, Bordeaux, France3 Institut Fédératif de Recherche de Santé Publique (IFR-99), Université Victor Segalen Bordeaux 2, Bordeaux, France2005 1 6 2005 9 4 R382 R389 1 3 2005 15 3 2005 5 4 2005 2 5 2005 Copyright © 2005 Vargas et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
We hypothesized that the use of intrapulmonary percussive ventilation (IPV), a technique designed to improve mucus clearance, could prove effective in avoiding further deterioration in patients with acute exacerbations of chronic obstructive pulmonary disease (COPD) with mild respiratory acidosis.
Methods
The study was performed in a medical intensive care unit of a university hospital. Thirty-three patients with exacerbations of COPD with a respiratory frequency ≥ 25/min, a PaCO2 > 45 Torr and 7.35 ≤ pH ≤ 7.38 were included in the study. Patients were randomly assigned to receive either standard treatment (control group) or standard treatment plus IPV (IPV group). The IPV group underwent two daily sessions of 30 minutes performed by a chest physiotherapist through a full face mask. The therapy was considered successful when both worsening of the exacerbation and a decrease in pH to under 7.35, which would have required non-invasive ventilation, were avoided.
Results
Thirty minutes of IPV led to a significant decrease in respiratory rate, an increase in PaO2 and a decrease in PaCO2 (p < 0.05). Exacerbation worsened in 6 out of 17 patients in the control group versus 0 out of 16 in the IPV group (p < 0.05). The hospital stay was significantly shorter in the IPV group than in the control group (6.8 ± 1.0 vs. 7.9 ± 1.3 days, p < 0.05).
Conclusion
IPV is a safe technique and may prevent further deterioration in patients with acute exacerbations of COPD with mild respiratory acidosis.
==== Body
Introduction
Acute exacerbations of chronic obstructive pulmonary disease (COPD) are a frequent cause of admission to hospital and the intensive care unit (ICU) [1]. Despite a well conducted medical treatment, worsening can occur in patients with acute exacerbations of COPD and lead to a decompensation phase. Acute respiratory failure can lead to the requirement of mechanical ventilation; in these cases, non-invasive ventilation (NIV) must be considered in order to avoid invasive mechanical ventilation and its related complications [2-4]. Studies by Plant et al. [4] offer very strong arguments for the delivery of NIV as soon as the patient develops an increase in PaCO2 and respiratory acidosis. Numerous patients are hospitalized with mild respiratory acidosis. Airway inflammation, bronchospasm and the increase in sputum volume are constant in these patients and are responsible for an increase in airway resistance and air trapping [5]. This air trapping and increased airway resistance result in hyperinflation and intrinsic positive expiratory pressure (PEEPi), which are common features during acute exacerbations of COPD patients and are responsible for increasing the work required to breathe and respiratory muscle failure. Methods of treatment directed against the onset of decompensation are attractive in theory, although the benefits of mucus clearance therapies have been regularly challenged [6-10]. Nevertheless, in COPD patients, there is a pathophysiological rationale for the use of a mucus clearance therapy. Hypersecretion of mucus, changes in mucus viscoelasticity and surface adhesion, and impaired ciliary function lead to entrapped mucus. Retained airway secretions can form mucus plugs and bronchial casts that cannot be expelled by coughing. Airway plugging causes impaired ventilation, resulting in lower ventilation-to-perfusion ratios. Increased airway resistance to airflow and air trapping result in hyperinflation of the chest and inspiratory loading of the respiratory muscles, leading to fatigue [11-14]. Two studies have shown that chest physiotherapy based on a mucus clearance strategy could represent a useful therapeutic option in exacerbations of patients with COPD [15,16].
Intrapulmonary percussive ventilation (IPV), intended for the therapeutic mobilization of bronchial secretions, has been primarily used in patients with cystic fibrosis in a stable state [17-22]. In addition, small pilot studies have shown the IPV device to be useful for increasing sputum production in patients with COPD [23,24]. IPV could offer a treatment directed against the onset of decompensation, specifically the increase in airway mucus that is responsible for increasing airway resistance.
We hypothesized that the use of IPV could be effective in avoiding further deterioration in patients admitted with acute exacerbations of COPD. In this prospective, randomized, controlled study, we compared the efficacy of standard medical treatment with supplemental oxygen and no ventilatory support to standard medical treatment with supplemental oxygen plus IPV in patients with acute exacerbations of COPD.
Materials and methods
Patients
Adult patients hospitalized in our ICU due to an acute exacerbation of COPD were prospectively studied. The experimental protocol was approved by the institutional review board of the hospital, and all patients or their next of kin provided written informed consent.
Inclusion and exclusion criteria
Patients were eligible for the study if they were admitted as an emergency with an exacerbation of COPD (on the basis of the clinical history, physical examination, and chest radiograph) [25], and a respiratory rate (RR) ≥ 25/min, PaCO2 > 45 Torr after the patient had been breathing room air for at least 10 minutes, and 7.35 ≤ pH ≤ 7.38 without metabolic acidosis. Exclusion criteria were: the requirement for emergency intubation for cardiopulmonary resuscitation, respiratory arrest, or in the case of rapid deterioration in neurological status (Glasgow coma scale [26] of ≤ 8); hemodynamic instability defined as a systolic blood pressure of less than 80 mmHg or evidence on electrocardiography of ischemia or clinically significant ventricular arrhythmias; failure of more than two additional organs; or tracheotomy, pneumothorax, facial deformity, or a recent history of oral, oesophageal or gastric surgery. Patients were randomly assigned to receive standard treatment or standard treatment plus IPV through a face mask. Random assignments were made with sealed envelopes.
Monitoring
Arterial oxygen saturation was monitored continuously with a bedside pulse oximeter (Oxisensor, Nellcor, Hayward, CA, USA); heart rate and RR were displayed on the screen of the monitor.
Standard treatment
Patients assigned to standard treatment received oxygen with nasal cannulae to maintain a target oxygen saturation (recorded by pulse oximetry) of 88% to 92%. In all patients, the heart rate and RR were monitored continuously. The head of the bed was kept elevated at a 45-degree angle. The standard drug protocol consisted of nebulised salbutamol (5 mg every 4 h) or terbutaline, nebulised ipratropium bromide (500 μg every 6 h), subcutaneous heparin, corticosteroids (methyl prednisolone 2 mg/kg of body weight intravenously per day for three days; then decreasing doses of oral methyl prednisolone for 15 days), and an antibiotic [27]. Medication included the correction of electrolyte abnormalities.
Intrapulmonary percussive ventilation
The IPV device was developed by Forrest M. Bird in 1979 (Fig. 1). IPV is a ventilatory technique that delivers small bursts of high flow respiratory gas into the lung at high rates. This causes airway pressures to oscillate between 5 and 35 cmH2O and the airway walls vibrate in synchrony with these oscillations. A unique sliding venturi called a phasitron (Fig. 2) is powered by compressed gas at 25 to 40 pounds per square inch and generates these oscillations in the range of 80 to 650 cycles per minute [28]. During inspiration the high frequency gas pulse expands the lungs and vibrates and enlarges the airways. This technique may be associated with nebulization [29] and has the potential to improve secretion clearance [30]. During the percussive bursts of air into the lungs, a continued pressure is maintained, while a high velocity percussive inflow opens airways and enhances intra-bronchial secretion mobilization.
Patients assigned to the IPV group received the same medication as the patients in the standard-treatment group with the addition of two sessions of IPV per day. No patient in either group received externally applied treatments designed to clear mucus. IPV sessions were performed by the specialized and trained respiratory therapist and delivered to the patient through a full face mask (La Cigogne, Pessac, France). The mask was adjusted and connected to the intrapulmonary percussive ventilator (IPV1 device, Percussionaire Corp., Sandpoint, ID, USA). After the mask had been secured, the percussions were delivered into the lungs of the patient. The frequency of the percussion was initially set at 250/minute and the peak pressure was initially set at 20 cmH2O. Frequency and peak pressure were adjusted for each patient to improve comfort and to be certain that the entire thorax was being percussed; this was done on the basis of visualization of external thoracic movements and perception of thrill on the patient's thorax. The inspiration-to-expiration ratio was adjusted to 1/2.5. During IPV sessions, the nebulizer delivered only NaCl 0.9%. Oxygen was fed into the mask to maintain oxygen saturation between 88% and 92%. The duration of each IPV session was 30 minutes. Between periods of IPV, patients breathed oxygen spontaneously while arterial oxygen saturation was continuously monitored. IPV sessions were stopped when the patients reached a RR of < 25/min and a pH > 7.38 in spontaneous breathing without worsening for 24 h.
Success of therapy
Therapy was considered to be successful when it enabled the avoidance of both a worsening of the exacerbation and a decrease in pH to under 7.35 (which would have required NIV), and allowed the patient to be discharged from the ICU.
Criteria for non-invasive ventilation
NIV was used as previously described [31,32] when patients were tachypnoeic with a RR of more than 25/min and a respiratory acidosis defined by a PaCO2 > 45 Torr and a pH lower than 7.35 without metabolic acidosis, at any time during the study.
Follow-up
The RR and arterial-blood gas levels were recorded at base line and at the end of the first IPV session. On subsequent days, these data were obtained once daily during the morning. IPV session comfort was assessed using a five-point verbal-rating scale (comfortable, mildly uncomfortable, moderately uncomfortable, very uncomfortable, and intolerable). To assess the patient's severity of illness on ICU admission Simplified Acute Physiologic Score (SAPS II) was recorded for each patient [33]. Pulmonary function data were obtained from previous spirometric tests in 13 patients in the IPV group and in 14 patients in the control group (the last test was retained). For the other patients (three in the IPV group and three in the control group), reliable pulmonary function data were obtained within two months of inclusion in the study.
Hospital stay
Patients were discharged from the ICU when their pH was higher than 7.38. Patients were discharged from the hospital when their clinical status and gas exchange were comparable to the stable state.
Statistical analysis
The primary outcome variable was the avoidance of a worsening of the acute exacerbation leading to decompensation, defined by a pH < 7.35, and so the need for NIV at any time during the study. The secondary end point was the length of the hospital stay. Results are given as means ± standard deviation (SD). The group means were compared with the t-test. Repeated-measures analysis of variance was used to compare the partial pressure of arterial oxygen, PaCO2, RR, and bicarbonate values measured at base line and at the end of the first IPV session. A p-value of less than 0.05 was considered to indicate statistical significance. Analyses were done using SPSS statistical software, version 10 (SPSS, Chicago, IL, USA).
Results
Patient characteristics
A total of 81 patients who underwent episodes of acute exacerbation of COPD were admitted to our ICU over one 18 month period. Of these, 48 patients were excluded from the study: 40 patients required NIV because of a pH < 7.35 despite a well conducted medical treatment; and eight patients required immediate endotracheal intubation. In the end, 33 patients were included, of which 17 were randomly assigned to standard treatment, and 16 to IPV. The two groups had similar characteristics on admission (Table 1). The patients in both groups had respiratory disease of a similar severity; the functional steady-state characteristics were thus similar in the two groups. The same was true for the severity of exacerbation; there were also no significant differences in the SAPS II and the arterial blood gas levels between the standard treatment group and the IPV group. Chest X-ray findings were similar between the two groups. Four patients in the standard group and three patients in the IPV group had been previously non-invasively ventilated for a similar episode. The same medication was administered to the patients of both groups.
Clinical outcome
As shown in Table 2, 6 out of 17 patients (35.3%) in the control group progressed to the point of requiring NIV, compared with 0 out of 16 in the IPV group (p < 0·05). The mean interval between entry into the study and decompensation was 48 ± 12 h for the six patients of the control group. NIV was successful every time and none of these six patients needed invasive mechanical ventilation. No patient included in the study died.
Physiological outcomes
The values for the physiological variables in the IPV group on inclusion and at the end of the first IPV session are given in Table 3. IPV led to an improvement in PaO2, PaCO2 and RR (p < 0.05). No statistical difference was observed concerning the pH.
IPV group
Patients assigned to the IPV group were given this method of treatment for a mean duration of 3 ± 1 days. The duration of IPV was half an hour twice daily. The mean frequency of the percussions was 250 ± 50 per minute. The mean peak pressure was 20 ± 5 cmH2O. IPV sessions were well tolerated. The median comfort score of IPV sessions was 2 (mildly uncomfortable). As judged by the physiotherapist in charge of the patient, mucus clearance was greatly improved by the application of the IPV.
Hospital stay
The hospital stay was significantly longer in the group receiving standard treatment than in the group receiving IPV (7.9 ± 1.3 versus 6.8 ± 1 days, p < 0.05).
Discussion
In this randomized trial, the use of IPV helped avoid further deterioration in patients admitted with acute exacerbation of COPD and mild acidosis. When compared with the patients who received standard treatment, the patients who received IPV had a lower incidence of NIV use and a shorter duration of hospital stay.
In acute exacerbations of COPD patients, NIV has profoundly changed the management and outcome of these patients [2]. The results of several prospective randomized controlled studies on NIV favor an early use of ventilatory methods as soon as the patient develops an increase in PaCO2 and respiratory acidosis [2,4]. To our knowledge, however, the potential benefits of NIV in COPD patients with mild respiratory acidosis have not been studied. We hypothesized that the use of IPV could be effective in avoiding further deterioration in this situation.
To date, few studies have been published on the use of IPV in adult patients with pulmonary disease. IPV has been used primarily, however, for the treatment of atelectasis and retained secretions in patients in a stable state, as occurs in a wide variety of conditions, including cystic fibrosis and neuromuscular disease [17-22]. Ravez et al. [23] studied the use of IPV in a small group of adults with chronic bronchitis. They found that total lung clearance of radioaerosol was enhanced with IPV therapy, but it was unclear how much IPV stimulated cough contributed to the observed benefit [23]. In addition, small pilot studies with the IPV device have shown it to be useful for the relief of lobar atelectasis and for increased sputum production in patients with COPD [24].
In this study, the most important question is: how does IPV improve the clinical status of patients with COPD?
IPV is a mucus clearance device
Inflammatory cells are abundant in the sputum of patients with chronic mucus retention. These cells are able to release mediators that can alter the secretion and clearance of mucus. The end result is airway plugging, which causes bronchial obstruction resulting in atelectasis, impaired lung mechanics and gas exchange. There is a physiopathological rationale for the use of mucus clearance therapies because even small decreases in airway resistance may be important to achieve recompensation [11-14]. The benefits of mucus clearance strategies using physical and respiratory therapies, however, have been regularly challenged [27]. Three randomized, controlled trials of chest physiotherapy [6-8] and one observational study [9], showed that mechanical percussion of the chest as applied by physical or respiratory therapists was ineffective and perhaps even detrimental in the treatment of patients with acute exacerbations of COPD. None of these randomized trials reported any improvement in ventilatory function with respect to either forced expiratory volume in one second (FEV1) or functional vital capacity [6-8]. Furthermore, one trial described a significantly lower FEV1 in patients who received chest percussion therapy compared with controls [7]. No other adverse effects were reported. Does the lack of evidence, however, mean that there is a lack of benefit? Two studies have shown that chest physiotherapy based on a mucus clearance strategy could represent a useful therapeutic option in patients with exacerbations of COPD [15,16]. Bellone et al. [15] demonstrated that chest physiotherapy using a positive expiratory pressure mask in patients with mild acidosis (mean pH = 7.33) requiring NIV with pressure support could produce benefits in sputum clearance and could reduce the amount of time that the patient requires NIV [15]. Wolkove et al. [16] reported significant improvement in lung function after inhaled bronchodilator therapy, and the prior use of a mucus clearance device, compared to a sham mucus clearance device, improved the subsequent bronchodilator response in patients with stable COPD [16]. The mucus clearance device used (flutter device), promotes the clearance of sputum through the generation of low frequency pressure waves [16]. Both the IPV and the flutter device appeared equally effective in removing obstructing secretions from airways [34,35]. In our study, mucus clearance obtained by the application of IPV was judged greatly improved by the physiotherapist in charge of the patient; however, we did not measure the quantity of expectoration in both groups. Thus we can not conclude that IPV in addition to standard treatment increased the elimination of mucus to a significant increment.
IPV theoretically increases mean airway pressure
PEEPi has been identified in patients with exacerbations of COPD because of severe airway obstruction [36-40]. In order to initiate inspiratory airflow, the respiratory muscles must generate a negative pressure equal in magnitude to PEEPi. The presence of PEEPi also implies dynamic hyperinflation, with consequent worsening of thoracic wall geometry and muscle length-tension relationships. This further increases the workload of muscles as their efficiency and mechanical advantage are reduced. The application of positive end-expiratory pressure (PEEP) at the airway opening should decrease the pressure gradient between the mouth and alveoli at the end of expiration and so reduce the inspiratory threshold load [38-40]. During the percussive sessions, IPV maintains an intrapulmonary pressure, which serves to stabilize airway patency. Improvement may occur via the beneficial effects of this intrapulmonary pressure, including the reduction of PEEPi and the amount of work required to breathe, which may allow respiratory muscles to regain efficiency.
High frequency oscillatory ventilation like effects
Considering the effect of high frequency oscillatory (HFO) ventilation on gas exchange and breathing pattern, one can hypothesize similar effects with IPV. Indeed, any high frequency ventilation is a positive pressure ventilation, which would increase the airway pressure (Paw), induce a 'PEEP effect' and thus improve oxygenation [41]. Two mechanisms explain gas transport with respect to the clearance of CO2 during HFO,: convection and molecular diffusion. HFO maximizes CO2 removal primarily through facilitated diffusion [41]. The theoretical increase of mean Paw observed with IPV, however, is less important than the increase of Paw observed with HFO. Similarly, the frequency in HFO, generally set below 5 Hz, is more important than in IPV.
IPV and lung volume
Any high frequency ventilation induces a PEEP effect that can increase lung volume. But according to the 'waterfall theory', if PEEPi is the result of expiratory flow limitation, application of extrinsic PEEP should decrease the pressure gradient between the mouth and alveoli at the end of expiration. This should be achieved without further hyperinflation. Several studies in patients during acute exacerbations of COPD have demonstrated this effect [36,38,39]. O'Donoghue et al. [40], however, found that only high levels of continuous positive airway pressure (CPAP) reduce PEEPi and indices of muscle effort in patients with severe but stable COPD, but only at the expense of a substantial increase in lung volume.
IPV sessions were well tolerated by the patients. Except for one episode of transient haemoptysis reported in a patient with cystic fibrosis [18], no serious adverse effects of IPV have been reported in previous studies [17-19]. All types of interfaces could be used to perform IPV sessions. On the basis of our previous experience with NIV in patients with acute exacerbations of COPD [31,32], we used a full face mask to perform IPV. This interface was well tolerated as most patients found the interface comfortable or only mildly uncomfortable.
The hospital stay was significantly shorter in the group receiving IPV than in the group receiving standard treatment. This result suggests that IPV may be a cost-saving measure. In addition, IPV sessions could be performed in general respiratory wards in COPD patients with mild respiratory acidosis and so could minimize the transfer of these patients to ICU. Further studies are needed to test this hypothesis.
Our study has several limitations. It is impossible to eliminate bias when a study cannot be blinded, so we have to be very careful concerning the shorter duration of the hospital stay in the IPV group. The study included only selected COPD patients with acute exacerbations who were treated in a single ICU. In the study, the control group didn't receive normal saline by nebulizer. Hypertonic saline has the potential of a mucotropic agent by shielding the excess of fixed negative charges that develop on mucins in airway disease [42]. Despite this, pharmacological mucus clearance strategies have not been demonstrated to shorten the course of treatment for patients with acute exacerbations of COPD, although there is a possibility that these agents improve symptoms [27]. Miro et al. [43], in a descriptive study, showed the benefit of CPAP sessions in seven COPD patients with acute hypercapnic respiratory failure in an attempt to avoid endothacheal intubation and mechanical ventilation. Goldberg et al. [44] conclude that the non-invasive application of CPAP to spontaneously breathing patients with severe COPD in acute respiratory failure decreases inspiratory effort and dyspnea while improving breathing pattern. End-expiratory lung volume remained stable at the lower levels of CPAP, with only modest increases at the higher levels [44]. These patients were at a more severe stage of decompensation. To our knowledge, there are no data concerning the utility of CPAP in acute exacerbations of COPD with mild respiratory acidosis. In a recent study, however, O'Donoghue et al. [40] found that only high levels of CPAP reduce PEEPi and indices of muscle effort in patients with severe but stable COPD, but only at the expense of a substantial increase in lung volume. It would be interesting, therefore, to perform another study with a control group of patients using CPAP to correct for lung volume change during intervention.
We have not evaluated the long term effect of IPV. Treatment with IPV was stopped when patients reached a RR of < 25/min and pH > 7.38 in spontaneous breathing without worsening for 24 h. No COPD patient worsened after the withdrawal of IPV. Also, we hadn't planned to evaluate IPV in patients recovering from their acute exacerbation. Another limitation of the trial was the small number of patients included. Further studies are also needed to determine the mechanisms of improvement with IPV in patients with acute exacerbations of COPD.
Conclusion
This randomized controlled study shows that chest physiotherapy by IPV may prevent the deterioration of acute exacerbations of COPD with mild respiratory acidosis. The method we employed was well tolerated. In addition, the technique performed at an early stage could be a cost-saving measure. Further studies are needed to determine the long term effects of IPV and to clarify the impact of IPV in the management of patients with acute exacerbations of COPD.
Key messages
• IPV led to a significant decrease in respiratory rate, an increase in PaO2 and a decrease in PaCO2.
• Clinical deterioration occurred in 6 out of 17 patients in the control group versus 0 out of 16 in the IPV group.
• The hospital stay was significantly shorter in the IPV group than in the control group.
• IPV is a safe technique and may prevent deterioration in cases of acute exacerbations of COPD with mild respiratory acidosis.
Abbreviations
COPD = chronic obstructive pulmonary disease; CPAP = continuous positive airway pressure; FEV1 = forced expiratory volume in one second; HFO = high frequency oscillatory ventilation; ICU = intensive care unit; IPV = intrapulmonary percussive ventilation; NIV = non-invasive ventilation; PEEPi = intrinsic positive expiratory pressure; PEEP = positive end-expiratory pressure; RR = respiratory rate; SAPS II = simplified acute physiologic score; SD = standard deviation.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
FV, NHB, AB, GGB, HG, DG, GH conceived of the study, and participated in its design and coordination and helped to draft the manuscript. LRS participated in the design of the study and performed the statistical analysis.
Acknowledgements
We are grateful to the junior doctors, nursing staff, and physiotherapists (Philippe Wibart, Isabelle Dauguet) at Pellegrin-Tripode University Hospital. We would also like to thank Tara Embleton for stylistic editing of the manuscript.
Figures and Tables
Figure 1 The intrapulmonary percussive ventilation (IPV) device (Percussionaire Corp., Sandpoint, ID, USA) and the full face mask used in the study.
Figure 2 Schematic of the phasitron. The sliding venturi body moves back and forth to open and close the phasitron exhalation gate by an orificed diaphragm. Burst of air and aerosol are delivered to the patient during forward movement of the phasitron.
Table 1 Characteristics of COPD patients assigned to receive intrapulmonary percussive ventilation or standard treatment at inclusion
IPV (n = 16) Standard treatment (n = 17) p-value
Age (years) 69.2 ± 6.0 70.2 ± 5.0 NS
SAPS II 25.4 ± 6.0 25.4 ± 4.0 NS
Hb (g/dl) 13.7 ± 1.3 14.0 ± 1.4 NS
HR (beats/min) 117 ± 16 115 ± 19 NS
RR (breaths/min) 36 ± 2 36 ± 3 NS
SBP (mmHg) 141 ± 15 143 ± 16 NS
pH 7.37 ± 0.01 7.37 ± 0.01 NS
PaCO2 (Torr) 57.6 ± 4.5 58.0 ± 3.0 NS
PaO2 (Torr)a
56.9 ± 3.0 56.7 ± 3.0 NS
Bicarbonate (mmol/l) 33 ± 3 33 ± 3 NS
FEV1 (%)b 39 ± 7 38 ± 8 NS
aPatients received oxygen with nasal cannulae to maintain a target oxygen saturation (recorded by pulse oximetry) of 88% to 92%. Mean oxygen flow was 2.0 ± 0.5 l/min and 2.0 ± 0.5 l/min in the IPV group and the control group, respectively (p = NS).
bFEV1 was obtained from previous spirometric tests in 13 patients in the IPV group and in 14 patients in the control group (the last test was retained). For the other patients (three in the IPV group and three in the control group), reliable pulmonary function data were obtained within two months of inclusion. FEV1, forced expiratory volume in one second; Hb, hemoglobin; HR, heart rate; IPV, intrapulmonary percussive ventilation; RR, respiratory rate; SAPS II, simplified acute physiologic score; SBP, systolic arterial blood pressure; NS, not statistically significant.
Table 2 Clinical outcome of COPD patients assigned to receive intrapulmonary percussive ventilation or standard treatment
IPV (n = 16) Standard treatment (n = 17) p-value
Worsening of exacerbation with pH < 7.35 (NIV required) (%) 0 (0) 6 (35.3) <0.05
Hospital stay (days) 6.8 ± 1.0 7.9 ± 1.3 <0.05
Hospital death 0 0 -
Table 3 Values of arterial blood gas and respiratory rate
IPV (n = 16)
Inclusion Post IPV session p-value
Arterial pH 7.37 7.38 NS
PaCO2 (Torr) 57.6 ± 4.5 53.5 ± 2.3 <0.05
PaO2 (Torr)a
56.9 ± 3.0 61.0 ± 0.8 <0.05
RR (breaths/min) 36 ± 2 31 ± 2 <0.05
Measurements were taken at inclusion and at the end of the first intrapulmonary percussive ventilation session (post IPV session; in the thirtieth minute). aOxygen flow was not changed between both periods and was 2.0 ± 0.5 l/min. IPV, intrapulmonary percussive ventilation; RR, respiratory rate; NS, not statistically significant.
==== Refs
Brochard L Non-invasive ventilation for acute exacerbations of COPD. A new standard of care Thorax 2000 55 817 818 10992531 10.1136/thorax.55.10.817
American Thoracic Society. International Consensus Conferences in Intensive Care Medicine: Noninvasive positive pressure ventilation in acute respiratory failure Am J Respir Crit Care Med 2001 163 283 291 11208659
Brochard L Mancebo J Wysocki M Lofaso F Conti G Rauss A Simonneau G Benito S Gasparetto A Lemaire F Noninvasive ventilation for acute exacerbations of chronic obstructive pulmonary disease N Engl J Med 1995 333 817 822 7651472 10.1056/NEJM199509283331301
Plant PK Owen JL Elliott MW Early use of non-invasive ventilation for acute exacerbations of chronic obstructive pulmonary disease on general respiratory wards: a multicentre randomised controlled trial Lancet 2000 355 1931 1935 10859037 10.1016/S0140-6736(00)02323-0
Anthonisen NR Manfreda J Warren CP Hershfield ES Harding GK Nelson NA Antibiotic therapy in exacerbations of chronic obstructive pulmonary disease Ann Inter Med 1987 106 196 204
Petersen ES Esmann V Honcke P Munkner C A controlled study of the effect of treatment on chronic bronchitis: an evaluation using pulmonary function tests Acta Med Scan 1967 182 293 305
Newton DA Bevans HG Physiotherapy and intermittent positive-pressure ventilation of chronic bronchitis Br Med J 1978 2 1525 1528 365289
Wollmer P Ursing K Midgren B Eriksson L Inefficiency of chest percussion in the physical therapy of chronic bronchitis Eur J Respir Dis 1985 66 233 239 4018176
Campbell AH O'Connell JM Wilson F The effect of chest physiotherapy upon the FEV1 in chronic bronchitis Med J Aust 1975 1 33 35 1128356
Snow V Lascher S Mottur-Pilson C The evidence base for management of acute exacerbations of COPD: Clinical practice guideline, Part 1 Chest 2001 119 1185 1189 11296188 10.1378/chest.119.4.1185
King M Rubin BK Derenne JP, Whitelaw WA, Similowski Mucus physiology and pathophysioloy Acute respiratory failure in chronic obstructive pulmonary disease 1996 New York: Dekker 391 405
Menkes H Britt J Rationale for physical therapy Am Rev Respir Dis 1980 122 127 131 7458040
Rochester DF Goldberg SK Techniques of respiratory physical therapy Am Rev Respir Dis 1980 122 133 146 7458041
Rossman CM Waldes R Sampson D Newhouse MT Effect of chest physiotherapy on the removal of mucus in patients with cystic fibrosis Am Rev Respir Dis 1982 126 131 135 7091898
Bellone A Spagnolatti L Massobrio M Bellei E Vinciguerra R Barbieri A Iori E Bendinelli S Nava S Short-term effects of expiration under positive pressure in patients with acute exacerbation of chronic obstructive pulmonary disease and mild acidosis requiring non-invasive positive pressure ventilation Intensive Care Med 2002 28 581 585 12029406 10.1007/s00134-002-1210-0
Wolkove N Kamel H Rotaple M Baltzan MA Jr Use of a mucus clearance device enhances the bronchodilator response in patients with stable COPD Chest 2002 121 702 707 11888949 10.1378/chest.121.3.702
Natale JE Pfeifle J Hommick DN Comparison of intrapulmonary percussive ventilation and chest physiotherapy: A pilot study in patients with cystic fibrosis Chest 1994 105 1789 1793 8205878
Homnick DN White F de Castro C Comparison of effects of an intrapulmonary percussive ventilator to standard aerosol and chest physiotherapy in treatment of cystic fibrosis Pediatr Pulmonol 1995 20 50 55 7478782
Birnkrant D Pope J Lewarsky J Stegmaier J Besunder J Persistent pulmonary consolidation treated with intrapulmonary percussive ventilation: A preliminary report Pediatr Pulmonol 1996 21 246 249 9121855 10.1002/(SICI)1099-0496(199604)21:4<246::AID-PPUL8>3.0.CO;2-M
Toussaint M De Win H Steens M Soudon P Effect of intrapulmonary percussive ventilation on mucus clearance in duchenne muscular dystrophy patients: a preliminary report Respir Care 2003 48 940 947 14525630
Varekojis SM Douce FH Flucke RL Filbrun DA Tice JS McCoy KS Castile RG A comparison of the therapeutic effectiveness of and preference for postural drainage and percussion, intrapulmonary percussive ventilation, and high-frequency chest wall compression in hospitalized cystic fibrosis patients Respir Care 2003 48 24 28 12556258
Deakins K Chatburn RL A comparison of intrapulmonary percussive ventilation and conventional chest physiotherapy for the treatment of atelectasis in the pediatric patient Respir Care 2002 47 1162 1167 12354335
Ravez P Richez M Godart G Vanthiel J Gauchet P Robience YJ Effect of intermittent high-frequency intrapulmonary percussive breathing on mucus transport Eur J Respir Dis 1986 69 suppl 146 285 289 3817062
McInturff SL Shaw LI Hodgkin JE Rumble L Bird FM Intrapulmonary percussive ventilation (IPV) in the treatment of COPD Respir Care 1985 30 885
American Thoracic Society Standards for the diagnosis and care of patients with chronic obstructive pulmonary disease and asthma Am J Respir Crit Care Med 1995 152 877 921
Teasdale G Jennett B Assessment of coma and impaired consciousness: a practical scale Lancet 1974 2 81 84 4136544 10.1016/S0140-6736(74)91639-0
McCrory DC Brown C Gelfand SE Bach PB Management of acute exacerbations of COPD: a summary and appraisal of published evidence Chest 2001 119 1190 1209 11296189 10.1378/chest.119.4.1190
White G Equipment theory for respiratory care 1996 Albany, NY: Delmar Thomson learning 229 232
Reychler G Keyeux A Cremers C Veriter C Rodenstein DO Liistro G Comparison of lung deposition in two types of nebulization: intrapulmonary percussive ventilation vs jet nebulization Chest 2004 125 502 508 14769731 10.1378/chest.125.2.502
Langenderfer B Alternatives to percussion and postural drainage, a review of mucus clearance therapies: Percussion and postural drainage, autogenic drainage, positive expiratory pressure, flutter valve, intrapulmonary percussive positive, and high-frequency chest compression with the thairapy vest J Cardiopulm Rehabil 1998 18 283 289 9702607 10.1097/00008483-199807000-00005
Hilbert G Gruson D Portel L Gbikpi-Benissan G Cardinaud JP Noninvasive pressure support ventilation in COPD patients with postextubation hypercapnic respiratory insufficiency Eur Respir J 1998 11 1349 1353 9657578 10.1183/09031936.98.11061349
Hilbert G Vargas F Valentino R Gruson D Gbikpi-Benissan G Cardinaud JP Guenard H Noninvasive ventilation in acute exacerbations of chronic obstructive pulmonary disease in patients with and without home noninvasive ventilation Crit Care Med 2002 30 1453 1458 12130961 10.1097/00003246-200207000-00009
Le Gall JR Lemeshow S Saulnier F A new simplified acute physiology score (SAPS II) based on a European/North American multicenter study J Am Med Assoc 1993 270 2957 2963 10.1001/jama.270.24.2957
Newhouse PA White F Marks JH Homnick DN The intrapulmonary percussive ventilator and flutter device compared to standard chest physiotherapy in patients with cystic fibrosis Clin pediatr 1998 37 427 432
Pryor JA Physiotherapy for airway clearance in adults Eur Respir J 1999 14 1418 1424 10624775 10.1183/09031936.99.14614189
Smith TC Marini JJ Impact of PEEP on lung mechanics and work of breathing in severe airflow obstruction J Appl Physiol 1988 65 1488 1499 3053583
Tobin MJ Lodato RF PEEP, auto-PEEP, and waterfalls Chest 1989 96 449 451 2670461
Petrof BJ Legare M Goldberg P Milic-Emili J Gottfried SB Continuous positive airway pressure reduces work of breathing and dyspnea during weaning from mechanical ventilation in severe chronic obstructive pulmonary disease Am Rev Respir Dis 1990 141 281 289 2405757
Appendini L Patessio A Zanaboni S Carone M Gukov B Donner CF Rossi A Physiologic effects of positive end-expiratory pressure and mask pressure support during exacerbations of chronic obstructive pulmonary disease Am J Respir Crit Care Med 1994 149 1069 1076 8173743
O'Donoghue FJ Catcheside PG Jordan AS Bersten AD McEvoy RD Effect of CPAP on intrinsic PEEP, inspiratory effort, and lung volume in severe stable COPD Thorax 2002 57 533 539 12037230 10.1136/thorax.57.6.533
Ritacca FV Stewart TE Clinical review: High-frequency oscillatory ventilation in adults – a review of the literature and practical applications Critical Care 2003 7 385 390 12974971 10.1186/cc2182
Robinson M Regnis JA Bailey DL King M Bautovich GJ Bye PT Effect of hypertonic saline, amiloride and cough on mucociliary clearance in patients with cystic fibrosis Am J Respir Crit Care Med 1996 153 1503 1509 8630593
Miro AM Shivaram U Hertig I Continuous positive airway pressure in COPD patients in acute hypercapnic respiratory failure Chest 1993 103 266 268 8417894
Goldberg P Reissmann H Maltais F Ranieri M Gottfried SB Efficacy of noninvasive CPAP in COPD with acute respiratory failure Eur Respir J 1995 8 1894 1900 8620959 10.1183/09031936.95.08111894
|
16137351
|
PMC1269449
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 1; 9(4):R382-R389
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3724
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37271613735510.1186/cc3727ResearchClinical investigation: thyroid function test abnormalities in cardiac arrest associated with acute coronary syndrome Iltumur Kenan [email protected] Gonul [email protected]ıturk Zuhal [email protected] <[email protected] Tuncay [email protected] Nizamettin [email protected] Assistant Professor, Dicle University Medical Faculty Department of Cardiology, Diyarbakir, Turkey2 Assistant Professor, Dicle University Medical Faculty Department of Anesthesia and Reanimation, Diyarbakir, Turkey3 Resident, Dicle University Medical Faculty Department of Cardiology, Diyarbakir, Turkey4 Professor, Dicle University Medical Faculty Department of Cardiology, Diyarbakir, Turkey2005 9 6 2005 9 4 R416 R424 23 11 2004 9 2 2005 25 4 2005 3 5 2005 Copyright © 2005 Iltumur et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
It is known that thyroid homeostasis is altered during the acute phase of cardiac arrest. However, it is not clear under what conditions, how and for how long these alterations occur. In the present study we examined thyroid function tests (TFTs) in the acute phase of cardiac arrest caused by acute coronary syndrome (ACS) and at the end of the first 2 months after the event.
Method
Fifty patients with cardiac arrest induced by ACS and 31 patients with acute myocardial infarction (AMI) who did not require cardioversion or cardiopulmonary resuscitation were enrolled in the study, as were 40 healthy volunteers. The patients were divided into three groups based on duration of cardiac arrest (<5 min, 5–10 min and >10 min). Blood samples were collected for thyroid-stimulating hormone (TSH), tri-iodothyronine (T3), free T3, thyroxine (T4), free T4, troponin-I and creatine kinase-MB measurements. The blood samples for TFTs were taken at 72 hours and at 2 months after the acute event in the cardiac arrest and AMI groups, but only once in the control group.
Results
The T3 and free T3 levels at 72 hours in the cardiac arrest group were significantly lower than in both the AMI and control groups (P < 0.0001). On the other hand, there were no significant differences between T4, free T4 and TSH levels between the three groups (P > 0.05). At the 2-month evaluation, a dramatic improvement was observed in T3 and free T3 levels in the cardiac arrest group (P < 0.0001). In those patients whose cardiac arrest duration was in excess of 10 min, levels of T3, free T3, T4 and TSH were significantly lower than those in patients whose cardiac arrest duration was under 5 min (P < 0.001, P < 0.001, P < 0.005 and P < 0.05, respectively).
Conclusion
TFTs are significantly altered in cardiac arrest induced by ACS. Changes in TFTs are even more pronounced in patients with longer periods of resuscitation. The changes in the surviving patients were characterized by euthyroid sick syndrome, and this improved by 2 months in those patients who did not progress into a vegetative state.
See related commentary
==== Body
Introduction
The most common reason for cardiac arrest in adults is coronary heart disease [1]. In particular, sudden and unexpected cardiac arrest may occur after an acute myocardial infarction (AMI) [2,3]. Prompt intervention (such as cardioversion and cardiopulmonary resuscitation [CPR]) can successfully resuscitate cardiac arrest patients [4,5]. Cardiac output rarely reaches 25% of its normal level during CPR in cardiac arrest, which renders cerebral blood flow inadequate. Cerebral blood flow is less than 30% at this stage [6], which results in varying degrees of hypoxic encephalopathy [7].
The hypophysis and hypothalamus are intracerebral organs, and if blood flow is inadequate then the function of these organs may be critically impaired. It is known that the hypothalamus-pituitary-thyroid axis is affected in patients with brain death. Although the underlying mechanism has not been elucidated, it is generally considered an endocrine abnormality characterized by 'euthyroid sick syndrome' (ESS) [8]. It is also known that certain nonthyroid critical conditions, including heart disease, may also lead to ESS [9-19]. The ESS (or the 'low T3 syndrome') occurs as a result of impairment in normal feedback response due to low tri-iodothyronine (T3) levels and disruption in conversion of the precursor hormone thyroxine (T4) to T3. Furthermore, the inactive metabolite reverse T3 accumulates in ESS [13,19].
Thyroid hormones have a major impact on the cardiovascular system [20-22]. Low T3 concentrations are known to be major independent indicators of mortality in patients hospitalized for cardiac causes [23]. Previous studies [24-27] reported critical impairments in thyroid homeostasis during the acute stage of cardiac arrest. However, it is not certain how, for how long and in which patient population this critical condition occurs. In addition, to our knowledge, thyroid functions have not yet been systematically assessed in patients with cardiac arrest caused by acute coronary syndrome (ACS). In the present study, conducted in patients who were resuscitated following cardiac arrest caused by ACS, we evaluated alterations that occur in thyroid hormone metabolism during the acute stage of cardiac arrest and at the end of the first 2 months after the event.
Materials and methods
A total of 50 patients with cardiac arrest caused by ACS (35 males and 15 females) who had been resuscitated (by cardioversion or CPR) and hospitalized in the intensive care unit (ICU) within the first 72 hours, and 31 AMI patients who did not require cardioversion or CPR (25 males and 6 females) were enrolled in the study, as were 40 healthy volunteers (28 males and 12 females). All patients or, in the case of unconsciousness, their closest relative signed a written informed consent form. The protocol was approved by the local ethics committee.
Patients were excluded if they were known to have thyroid function test (TFT) abnormalities that could not be related to AMI or cardiac arrest. We also excluded those patients who had previously suffered acute coronary events, who had previously undergone percutaneous transluminal coronary angioplasty or bypass surgery, who had a history of heart failure, and who received medication that could alter thyroid function, such as amiodarone and phenytoin (excluding β-blockers, heparin and dopamine), or who had comorbid conditions (malignancy, hepatic, or renal failure).
Cardiac arrest group
Of the 50 patients (35 males and 15 females; mean age 59 ± 8 years) in the cardiac arrest group, 28 patients were resuscitated using CPR, whereas the remaining 22 patients only underwent cardioversion. In cardiac arrest patients, three subgroups were defined based on the duration of intervention in order to investigate whether this had any impact on TFTs: cardiac arrest group 1, <5 min (n = 24; mostly consisting of patients who underwent cardioversion); cardiac arrest group 2, 5–10 min (n = 14); and cardiac arrest group 3, >10 min (n = 12). Postischaemic anoxic encephalopathy (cerebral postresuscitation syndrome or disease) grading was done according to the classification reported by Maiese and Caronna [7]. The possible outcomes they distinguished are as follows: dead, decerebrate, persistent vegetative state, severe focal neurological deficit, amnesic syndrome and neurologically intact (but often with psychological changes).
Patients with cardiac arrest were followed up in the ICU until their cardiac function became stable. The patients received standard therapies, depending on the aetiology of cardiac arrest (ACS with or without ST-segment elevation). A total of 23 patients did not receive thrombolytic thera and the remaining 27 patients underwent thrombolytic therapy with streptokinase. The patients with severe arrhythmia were administered lidocaine, an antiarrhythmic agent. Furthermore, four patients received dopamine because of low blood pressure. All patients received therapy required to achieve a normal metabolic condition and acid–base balance.
Acute myocardial infarction group
The AMI group included 31 (25 males and 6 females; mean age 57 ± 9 years) consecutive AMI patients admitted to the ICU within the first 12 hours after the event and who did not require cardioversion or CPR. Myocardial infarction was defined using the European Society of Cardiology/American College of Cardiology guidelines [28]. All patients received standard medical therapy, consisting of aspirin, heparin, intravenous nitrates and β-blockers, where it was not contraindicated. Furthermore, all patients with AMI were treated with streptokinase (1.5 million IU in 60 min). Continuous electrocardiogram telemetry monitoring was done in all patients during their stay in the coronary care unit.
Control group
The control group included 40 volunteers (28 males and 12 females; mean age 58 ± 6 years) without angina pectoris and with the same age distribution and similar male/female ratios as the cardiac arrest and AMI groups. History, physical examination, electrocardiography, chest radiography and routine chemical analysis identified no evidence of coronary heart disease in these individuals.
Laboratory measurements
Fasting blood samples were collected for thyroid hormone profile from cardiac arrest and AMI groups after an average period of 72 hours following the initial event. Blood samples were also taken during the first 12 hours in the AMI group. Furthermore, blood samples were collected again for follow-up assessment from surviving patients in both groups at the end of the second month. Fasting blood samples from the control individuals were collected once. Blood samples drawn from brachial vein were centrifuged, and measurements of T3, free T3, T4, free T4 and thyroid-stimulating hormone (TSH) were taken. Serum T3, free T3, T4, free T4 and TSH serum levels were assessed using a Roshe-170E modular analytics device (Roshe Diagnostics GmbH, Mannheim for USA, US Distributor: Roshe Diagnostics, Indianapolis, IN), employing the electrochemiluminescence method. The reference intervals for our laboratory are as follows: T3, 0.85–2.02 ng/ml; T4, 5.13–14.06 μg/dl; free T3, 0.18–0.46 ng/ml; free T4, 0.93–1.71 ng/dl; and TSH, 0.27–4.2 μIU/ml. Standard procedures were used to determine serum levels of creatine kinase-MB and troponin I.
Echocardiographic examination was performed with a HP SONOS 4500 (Agilent Technologies Andover, Canada), using a 3.5 or 2.5 MHz transducer. Echocardiographical images were obtained from parasternal and apical views. Parasternal long axis, short axis, and apical four chamber views were assessed according to the criteria recommended by the American Echocardiography Society (29). The left ventricular ejection fraction (LVEF) was assessed echocardiographically, using the Simpson biplane formula [29].
Patients remained in the ICU until they were stable in terms of their ischaemic heart disease. Those with complications other than ischaemic heart disease (severe neurological deficit, or persistent vegetative or decerebrate state) were monitored in neurology departments. Coronary angiography was performed if indicated in those patients whose condition became stable. Iopromid (Ultravist; 370 mg iodine/ml Schering Alman, Istanbul Turkey)) was used as the contrast medium in coronary angiography.
Statistics
All values were expressed as mean ± standard deviation. The data were analyzed by analysis of variance for repeated measurements, followed by post hoc analysis for pairwise comparisons, and were corrected by Tukey test or paired t-test when indicated. P < 0.05 was considered statistically significant.
Results
Although patients in the cardiac arrest group were older than the AMI patients and control individuals, the difference was not statistically significant (P > 0.05). Most of the patients were men. The patients in cardiac arrest group were classified according to the Maiese and Caronna classification as follows: 21 were neurologically intact, 13 were amnesic, four had severe neurological deficit, two were in a persistent vegetative state, eight were decerebrate and two were dead. Of the cardiac arrest patients, 23 had anterior myocardial infarction, nine had inferior myocardial infarction, 14 had inferior myocardial infarction with right ventricular involvement, and four had non-Q-wave myocardial infarction. The AMI group included 14 patients with anterior myocardial infarction, 10 with inferior myocardial infarction, and seven with inferior myocardial infarction with right ventricular involvement.
Of the cardiac arrest patients, the duration of intervention was under 5 min for 24 patients (22 underwent cardioversion), 5–10 min for 14 patients, and longer than 10 min for 12 patients. Although 22 of the cardiac arrest patients died within the first 2 months, only one patient died in the AMI group. Of the cardiac arrest patients who died, 11 had an intervention lasting longer than 10 min, eight had an intervention lasting 5–10 min, and three had an intervention lasting less than 5 min. It was observed that, although troponin and CK-MB levels were higher, LVEF was lower in the cardiac arrest group compared with those parameters for the AMI group (P < 0.0001, P < 0.05 and P < 0.05, respectively). The characteristics of the patients and control inidividuals are summarized in Table 1.
Coronary angiography was performed in a total of 37 patients. Of these patients, 15 were in the cardiac arrest group and 22 were in the AMI group. The mean volume of contrast medium used in coronary angiography was 110 ± 19 ml. In the statistical analysis applied, at the end of the second month the TFT results for patients undergoing angiography were similar to those in patients not undergoing angiography (angiography versus no angiography: T3, 1.16 ± 0.25 versus 1.12 ± 0.22 ng/ml; free T3, 0.29 ± 0.06 versus 0.28 ± 0.09 ng/ml; T4, 8.45 ± 2 versus 7.84 ± 1.99 μg/dl; free T4, 1.31 ± 0.19 versus 1.29 ± 0.26 ng/dl; TSH, 1.35 ± 0.73 versus 1.19 ± 0.61 μIU/ml; P > 0.05 for all comparisons).
The T3 and free T3 levels on day 3 in the cardiac arrest group were significantly lower than those in the AMI group and control group (P < 0.0001). In contrast, T4, free T4 and TSH levels did not differ significantly between groups (P > 0.05; Table 2). The cardiac arrest group had lower T3 (0.9 ± 0.31 versus 1.13 ± 0.24 ng/ml) and free T3 (0.22 ± 0.12 versus 0.29 ± 0.07 ng/ml) levels than did the AMI group on day 3, even when subgroups were analyzed and only the surviving patients were considered (for both, P < 0.01). However, at the 2-month follow-up visits, T3 and free T3 levels were found to have improved dramatically in the cardiac arrest group (P < 0.0001; Fig. 1 and Table 3).
When the subgroup of patients who underwent cardioversion alone was compared with the subgroup of patients who underwent CPR alone, it was observed that T3 and free T3 levels were lower in the CPR subgroup (P < 0.006 and P < 0.02, respectively). No significant difference was observed between the other thyroid hormones and TSH (P > 0.05). It was also noted that, although troponin-I and CK-MB values were high, LVEF was low in the CPR subgroup (P < 0.03, P < 0.02 and P < 0.05, respectively; Table 4). At the 2-month follow-up visit, T3 and free T3 levels were similar between the CPR-alone and cardioversion-alone subgroups (T3, 1.12 ± 0.18 versus 1.17 ± 0.28 ng/ml; free T3, 0.28 ± 0.94 versus 0.29 ± 0.9 ng/ml; P > 0.05).
When the duration of cardiac arrest was considered, it was observed that T3 (0.6 ± 0.15 versus 0.93 ± 0.31 ng/ml) and free T3 (0.11 ± 0.03 versus 0.24 ± 0.11 ng/ml) levels were lower in patients with interventions of more than 10 min than in those with interventions of less than 5 min (P < 0.001). Similarly, TSH (8.9 ± 6.1 versus 13.9 ± 5.8 μIU/ml; P < 0.05) and T4 (6 ± 1.2 versus 8.5 ± 2.4 μg/dl; P < 0.005) levels were lower in those who had interventions of more than 10 min.
Although the day 1 values for thyroid hormones and TSH were lower in the AMI group than in the control group, the difference was not significant (P > 0.05). However, day 3 levels of T3 and free T3 were significantly lower in the AMI group than in the control group (P < 0.01). In contrast, serum levels of T4, free T4 and TSH did not differ significantly between these groups (P > 0.05). Thyroid hormones and TSH were lower on day 3 than on day 1 for the AMI group. However, only free T3 levels were significantly lower on day 3 when the day 1 and day 3 values were compared (P < 0.05; Table 5). T3 and free T3 values of the patients who died within the first 2 months in the cardiac arrest group were markedly lower than those in survivors (P = 0.02 and P = 0.03, respectively). T4, free T4 and TSH levels were low in patients who died, but this finding was not statistically significant (P > 0,05). It was also observed that the troponin and CK-MB values in those who died were higher than in survivors, but the LVEF value was lower (P < 0.001; Table 6).
When the 2-month TFTs for the cardiac arrest and AMI groups were compared with those in the control group, it was found that the level of free T3 (control 0.32 ± 0.02 ng/ml, cardiac arrest 0.29 ± 0.09 ng/ml, AMI 0.29 ± 0.05 ng/ml; P > 0.05) and TSH (control 1.2 ± 0.5 μIU/ml, cardiac arrest 1.25 ± 0.48 μIU/ml, AMI 1.27 ± 0.82 μIU/ml; P > 0.05) were similar in all three groups. In contrast, the level of T3 was lower both in cardiac arrest and AMI groups than in the control group. However, T3 in all groups was within the normal reference range (control 1.32 ± 0.28 ng/ml, cardiac arrest 1.15 ± 0.24 ng/ml, AMI 1.18 ± 0.23 ng/ml; P < 0.05).
The 2-month follow-up visit revealed that depressed T3 and free T3 levels in two patients, who were in vegetative state, had persisted. Furthermore, one of those patients was observed to have lower T4 and free T4 levels, but the TSH level did not change significantly.
Discussion
To the best of our knowledge, no other published study has demonstrated major alterations in standard thyroid homeostasis during the acute stage of cardiac arrest, which then normalized by the second month in patients who survived cardiac arrest induced by ACS. In severe illnesses of nonthyroid origin [10,11], including cardiac diseases [12], downregulation of the thyroid hormone system can occur. This condition, which has been called the ESS or the 'low T3 syndrome', is characterized by a change in thyroid homeostasis. This condition occurs as a result of impairment in the normal feedback response due to low T3 levels and disruption in conversion of precursor hormone T4 to T3. The significantly lower T3 and free T3 levels in the cardiac arrest group than in the uncomplicated AMI group noted here reflects the critical changes in thyroid homeostasis that occur in cardiac arrest
The hypothalamohypophysial–thyroid axis must function properly to ensure normal thyroid homeostasis. We had postulated that this axis would be disrupted in patients with cardiac arrest caused by impairment in the circulation to the hypophysis and hypothalamus, which would lead to significant TFT abnormalities. In fact, the study revealed that while T3 and free T3 levels were significantly lower in the cardiac arrest group, TSH was lower as well, albeit it not significantly so. In general, TSH rises in response to lower T3 levels. However, in cardiac arrest patients this was not found to be the case, which confirmed the occurrence of ESS in these cardiac arrest patients. Although hormonal changes were more prominent in the cardiac arrest group than in the AMI group, the changes in the two groups paralleled each other. The fact that the changes in thyroid function were observed to return to normal at the 2-month follow-up visit was another indication of the presence of ESS in cardiac arrest. It is known that thyroid functions normalize in ESS patients following improvement in the pathology causing ESS [9,13]. However, it must be noted that some of the patients, who had undergone CPR for a lengthy period, died within the first 2 months. This might have contributed to the difference in results. Normally, secondary hypothyroidism is expected in severe ischaemia of the hypophysis [30]. However, a possible explanation for our findings, characterized by ESS, are as follows: even during critical hypotension, brain perfusion continues via autoregulation of cerebral blood flow, and this prevents more severe complications in intracerebral organs.
Various vasoactive substances have been described that contribute to the physiological regulation of cerebral perfusion, either by vasoconstriction or by vasodilatation [31]. In particular, during severe hypotension, nitric oxide mediated autoregulation has been suggested to play an important role in maintaining brainstem perfusion, which is needed to preserve the integrity of vital brainstem functions [32]. Although cerebral blood flow is inadequate, brain perfusion continues during effective CPR. Therefore, ESS, rather than secondary hypothyroidism, may occur during shorter cardiac arrest events. However, in patients with longer durations of resuscitation, a clinical picture resembling that of secondary hypothyroidism may be observed [30]. In our study, TFT findings in the patients with longer arrest intervals were more impaired.
There are some differences between our study and some others investigating thyroid function in cardiac arrest patients. Regardless of resuscitation success, Longstreth and coworkers [24] observed low T3 and T4 levels and high TSH levels in patients with out-of-hospital cardiac arrest. They stated that these alterations in thyroid hormones may play a role in cardiac arrest aetiology and prognosis. Wortsman and coworkers [25] reported significantly depressed T3 and T4 levels. Likewise, T3, free T3, T4 and free T4 levels were reported to have decreased in animal studies [26,27]. However, when all patients are considered, our study demonstrated significantly lower T3 and free T3 in the cardiac arrest group, but no significant changes in T4, free T4 and TSH levels. However, the lower T4 levels observed in subgroup analyses in patients with longer resuscitation periods is consistent with those studies. Meanwhile, one of our patients in a vegetative state had lower T4 and free T4 values, as well as lower T3 and free T3. Therefore, we may conclude that T4 levels decrease, along with the decrease in the active hormone T3 in association with impairment in the hypothalamus–hypophysis–thyroid axis, particularly in patients with longer resuscitation periods. In 42% of pituitary apoplexy cases of various causes (haemorrhage, radiation, intracranial hypertension, etc.), secondary hypothyroidism developed [30]. The length of time in resuscitation may be one of the reasons for the different findings observed in the present study. Furthermore, our study group was homogenous because it comprised patients with cardiac arrest induced by ACS. Lack of a homogenous population in previous studies might have led to inconsistencies between the studies.
ESS may be observed in different forms. A milder form of ESS may be observed with only a decrease in T3, as in uncomplicated AMI, and a T4 decrease accompanying decreased T3 levels may also be observed, as was the case in cardiac arrest patients with longer CPR sessions in the present study. It is known that this condition is associated with increased mortality. Rarely, an increase in T4 may also be observed [13,14]. Moreover, it is known that an increase occurs normally at the level of reverse T3 in ESS, although we have not measured it. The cause of the decreased T3 in ESS has not been established. It has been attributed to various parameters, including test artifacts, inhibitors of T4 and T3 binding to proteins, decreased 5'-deiodinase activity and circulating cytokines. It is known that inflammation plays a critical role in the pathophysiology of the ESS that occurs in AMI. In particular, interleukin-6 plays a major role in the development of this syndrome. It inhibits conversion of T4 to T3 by inhibiting mRNA expression or by blocking 5'-deiodinase activity. This inhibition occurs both in the pituitary–thyroid axis and in peripheral transformation of the thyroid hormone [15,16]. Furthermore, Fliers and coworkers [17] reported a strong correlation between thyroid-releasing hormone gene expression and serum T3 and TSH concentrations in patients with various degrees of ESS. It is not known whether different mechanisms are involved in the changes that occur in TFTs during cardiac arrest.
The changes observed in thyroid function in the AMI group were characterized by a milder form of ESS and were consistent with previous studies [12,15,19]. However, Pavlou and coworkers [18] reported depressed T3, T4, free T3, free T4 and TSH serum levels in complicated AMI. Moreover, those authors maintained that ESS occurred both in AMI and in unstable angina pectoris, and they had suggested an association between the drop in T3 and cardiac damage.
Although downregulation of thyroid hormones occurring both in cardiac arrest and AMI groups may be regarded as an adaptive measure to decrease the cardiac workload and conserve energy during acute ischaemia, this effect continues in an unstable manner that then becomes maladaptive [19]. It is known that thyroid hormones have beneficial effects on cardiac contractility, output, systemic vascular resistance and diastolic functions [20-22]. Changes in thyroid hormones that occur because of cardiac arrest or AMI lead to critical haemodynamic alterations in the cardiovascular system by increasing the vascular resistance and decreasing cardiac output [20-23]. In particular, the decrease in active hormone T3 leads to further impairment in cardiac functions. Iervasi and coworkers [23] reported low serum T3 levels as an independent predictor of mortality in patients with cardiovascular disease. Alterations in TFTs are more marked in seriously ill patients [24-27]. In the present study the TFT findings in those who died within the first 2 months deteriorated more than did those in survivors.
Thyroid hormone replacement therapy has been considered as a result of favourable changes that occur in cardiac functions and cardiac gene expression following T3 administration in patients with ESS. Whitesall and coworkers [33] reported that T3 replacement did not have positive effects on cardiac function in dogs, but several previously conducted studies demonstrated that T3 replacement improved left ventricular function and normalized T3-responsive gene expression [26,27,34-39]. Similarly, increased LVEF values as a result of T3 administration following AMI was reported in animal studies [26,27,35]. Moreover, it was observed in open heart surgery that T3 improved haemodynamic parameters [36,39]. Left ventricular function is among the leading indicators of prognosis following AMI [40]. Furthermore, cardiogenic shock occurring in cardiac arrest and AMI patients is a critical predictor of mortality [41]. Cardiac output decreases significantly because of shock, and if thyroid dysfunction accompanies this then further functional impairments can be expected. Taniguchi and coworkers [8] established in donors with brain death that administration of T3 along with cortisol increased blood pressure and had a favourable, stabilizing effect on cardiac function. These studies show T3 to be a potential therapeutic approach to improving left ventricular function in ESS [26,27,34-41]. Nevertheless, large-scale studies of T3 therapy are required in the setting of haemodynamic instability following cardiac arrest and AMI. One of the limitations of the present study was the fact that some of our patients were administered drugs that could alter TFTs. However, a previous study in AMI patients [18] reported that β-blockers and thrombolytic therapy did not alter thyroid function. Only four patients were administered dopamine. Furthermore, we were unable to document ischaemia of the hypophysis or hypothalamus. Therefore, more studies are required to establish the extent of ischaemia of the hypophysis and hypothalamus in patients undergoing CPR and to investigate its impact on thyroid hormones. Another limitation of the study is that we did not measure the level of reverse T3 – an inactive metabolite with prognostic value.
Conclusion
TFTs are significantly altered in cardiac arrest induced by ACS. The changes in TFTs are even more pronounced in patients with longer periods of resuscitation. The changes in the surviving patients are characterized by ESS and improve by 2 months in patients who have not progressed to a vegetative state. Large-scale studies in cardiac arrest are required to demonstrate the course of TFTs, including measurement at 24 hours and of reverse T3 levels.
Key messages
• Thyroid function tests are significantly altered in cardiac arrest induced by ACS in acute stage.
• The changes in TFTs are even more pronounced in patients with longer periods of resuscitation.
• The changes in the surviving patients are characterized by euthyroid sick syndrome.
• These changes in acute stage improve dramatically by the end of the second month.
Abbreviations
ACS = acute coronary syndrome; AMI = acute myocardial infarction; CK-MB = creatine kinase MB isoenzyme; CPR = cardiopulmonary resuscitation; ESS = euthyroid sick syndrome; ICU = intensive care unit; LVEF = left ventricular ejection fraction; T3 = tri-iodothyronine; T4 = thyroxine; TFT = thyroid function test; TSH = thyroid-stimulating hormone.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
KI created and designed the study, drafted the manuscript, performed the statistical analysis and interpretation of data, and revised the manuscript. GO was involved in the collection, statistical analysis and interpretation of the data. ZA and TT conducted patient monitoring and data collection. NT contributed to the design and the coordination of the study as well as interpretation of the data. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 T3 and FT3 levels in the CA group had increased by the end of month 2.
Table 1 Patient characteristics
CA (a) AMI (b) Control (c) P
Number 50 31 40 -
Age (years) 59 ± 8 57 ± 9 58 ± 6 NS
Sex (male/female) 35/15 25/6 28/7 -
LVEF (%) 44.1 ± 8.2** 48.2 ± 8.6 65.9 ± 3.7* *c versus a, b **a versus b
Peak troponin I (μg/ml) 29.9 ± 26.1* 6.7 ± 1.6* < 0.01 *a versus b, c b versus c
Peak CK-MB (IU/l) 228.7 ± 147.4** 170.5 ± 61.2 14.6 ± 4.1* * c versus a, b **a versus b
*P < 0.0001, **P < 0.05. AMI, acute myocardial infarction; CA, cardiac arrest; CK-MB, creatine phosphokinase MB isoenzyme; LVEF, left ventricular ejection fraction; NS, not significant.
Table 2 Thyroid hormones and thyroid-stimulating hormone levels in the controls and cardiac arrest (day 3) and acute myocardial infarction (day 3) patients
CA day 3 (a) AMI day 3 (b) Control (c) P
Number 50 31 40 -
T3 (ng/ml) 0.83 ± 0.3* 1.12 ± 0.24 1.32 ± 0.28** *a versus b, c **b versus c
Free T3 (ng/ml) 0.19 ± 0.11* 0.27 ± 0.06 0.32 ± 0.06 *a versus b, c
T4 (μg/dl) 7.6 ± 2.3 8.3 ± 1.6 8.4 ± 1.8 NS
Free T4 (ng/dl) 1.21 ± 0.5 1.35 ± 0.2 1.28 ± 0.2 NS
TSH (μIU/ml) 1.22 ± 0.6 1.31 ± 0.8 1.2 ± 0.5 NS
*P < 0.0001, **P < 0.01. AMI, acute myocardial infarction; CA, cardiac arrest; NS, not significant; T3, tri-iodothyronine; T4, thyroxine; TSH, thyroid-stimulating hormone.
Table 3 Thyroid hormone and thyroid-stimulating hormone values for cardiac arrest and acute myocardial infarction groups at day 3 and month 2
CA day 3 (a) CA month 2 (b) AMI day 3 (c) AMI month 2 (d) P
Number 50 28 31 30
T3 (ng/ml) 0.83 ± 0.3* 1.15 ± 0.24 1.12 ± 0.24 1.18 ± 0.23 *a versus b
Free T3 (ng/ml) 0.19 ± 0.11* 0.29 ± 0.09 0.27 ± 0.06 0.29 ± 0.05 *a versus b
T4 (μg/dl) 7.62 ± 2.34 8.24 ± 2.4 8.27 ± 1.52 8.47 ± 1.5 NS
Free T4 (ng/dl) 1.23 ± 0.46 1.25 ± 0.27 1.35 ± 0.2 1.37 ± 1.65 NS
TSH (μIU/ml) 1.22 ± 0.58 1.25 ± 0.48 1.31 ± .0.83 1.27 ± 0.82 NS
*P < 0.0001. AMI, acute myocardial infarction; CA, cardiac arrest; NS, not significant; T3, tri-iodothyronine; T4, thyroxine; TSH, thyroid-stimulating hormone.
Table 4 Day 3 values for cardiac arrest subjected to cariopulmonary resuscitation alone and cardioversion alone
CPR CV P
Number 28 22 -
T3 (ng/ml) 0.73 ± 0.24 0.94 ± 0.29 <0.006
Free T3 (ng/ml) 0.16 ± 0.09 0.23 ± 0.05 <0.02
T4 (μg/dl) 7.23 ± 2.34 8.1 ± 2.28 NS
Free T4 (ng/dl) 1.15 ± 0.4 1.29 ± 0.5 NS
TSH (μIU/ml) 1.09 ± 0.5 1.38 ± 0.6 NS
Troponin I (μg/ml) 37.3 ± 28.9 20.5 ± 18.7 <0.03
CK-MB (IU/l) 271.8 ± 161.3 173.8 ± 107.7 <0.02
LVEF (%) 42.1 ± 7.9 46.8 ± 7.8 <0.05
CPR, cardiopulmonary resuscitation; CK-MB, creatine kinase MB isoenzyme; CV, cardioversion; LVEF, left ventricular ejection fraction; NS, not significant; T3, tri-iodothyronine; T4, thyroxine; TSH, thyroid-stimulating hormone.
Table 5 Thyroid hormone and thyroid-stimulating hormone values for the control group and acute myocardial infarction group on days 1 and 3
AMI day 1 (n = 31; a) AMI day 3 (n = 31; b) Control (n = 40; c) P
T3 (ng/ml) 1.23 ± 0.25 1.12 ± 0.24* 1.32 ± 0.28 *b versus c
Free T3 (ng/ml) 0.31 ± 0.06† 0.27 ± 0.06‡ 0.32 ± 0.06 †a versus b ‡b versus c
T4 (μg/dl) 8.4 ± 1.7 8.3 ± 1.6 8.4 ± 1.8 NS
Free T4 (ng/dl) 1.38 ± 0.2 1.35 ± 0.2 1.28 ± 0.2 NS
TSH (μIU/ml) 1.35 ± 0.9 1.31 ± 0.8 1.2 ± 0.5 NS
*P = 0.002, †P < 0.05, ‡P = 0.003. AMI, acute myocardial infarction; CA, cardiac arrest; NS, not significant; T3, tri-iodothyronine; T4, thyroxine; TSH, thyroid-stimulating hormone.
Table 6 LVEF, TFTs, troponin and CK-MB levels in the cardiac arrest group, subdivided into those who died and those who survived the first 2 months
CA survivors CA died P
Number 28 22 -
T3 (ng/ml) 0.9 ± 0.31 0.72 ± 0.18 0.02
Free T3 (ng/ml) 0.22 ± 0.12 0.15 ± 0.08 0.03
T4 (μg/dl) 8.1 ± 2.5 7.02 ± 2 NS
Free T4 (ng/dl) 1.27 ± 0.5 1.14 ± 0.5 NS
TSH (μIU/ml) 1.35 ± 0.5 1.05 ± 0.6 NS
Troponin I (μg/ml) 15.2 ± 9.8 48.6 ± 28.5 <0.0001
CK-MB (IU/l) 148.7 ± 86 330.5 ± 147.7 <0.0001
LVEF (%) 48.4 ± 7.5 38.7 ± 5.3 <0.0001
AMI, acute myocardial infarction; CA, cardiac arrest; CK-MB, creatine kinase MB isoenzyme; LVEF, left ventricular ejection fraction; NS, not significant; T3, tri-iodothyronine; T4, thyroxine; TSH, thyroid-stimulating hormone.
==== Refs
Goldstein S Landis JR Leighton R Ritter G Vasu CM Lantis A Serokman R Characteristics of the resuscitated out-of hospital cardiac arrest victim with coronary heart disease Circulation 1981 64 977 984 7285312
Ornato JP Peberty MA Tadler SC Strobos NC Factors associated with the occurence of cardiac arrest during hospitalization for acute myocardial infarction in the second national registry of myocardial infarction in the US Resuscitation 2001 48 117 123 11426473 10.1016/S0300-9572(00)00255-0
Fabricius-Bjerre N Astvad K Kjaerulff J Cardiac arrest following acute myocardial infarction: a study of 285 cases from three medical departments using a joint acute admission section containing a coronary care unit Acta Med Scand 1974 195 261 265 4821295
De Vos R Quality of life after cardiopulmonary resuscitation Resuscitation 1997 35 231 236 10203401 10.1016/S0300-9572(97)00068-3
Anonymous Part 1: Introduction to the International Guidelines 2000 for CPR and ECC. A consensus on Science: European Resuscitation Council Resuscitation 2000 46 3 15 10978786 10.1016/S0300-9572(00)00269-0
Robertson C Holmberg S Compression techniques and blood flow during cardiopulmonary resuscitation. A statement for the Advanced Life Support Working Party of the European Resuscitation Council Resuscitation 1992 24 123 132 1335603 10.1016/0300-9572(92)90018-8
Maiese K Caronna JJ Aminoff ME Neurological complications of cardiac arrest Neurology and General Medicine 1989 New York: Churchill Livingstone 145 157
Taniguchi S Kitamura S Kawachi K Doi Y Aoyama N Effects of hormonal supplements on the maintenance of cardiac function in potential donor patients after cerebral death Eur J Cardiothorac Surg 1992 6 96 101 1581088 10.1016/1010-7940(92)90082-9
Chopra IJ Clinical review 86: Euthyroid sick syndrome: is it a misnomer? J Clin Endocrinol Metab 1997 82 329 334 9024211 10.1210/jc.82.2.329
Docter R Krenning EP de Jong M Henneman G The sick euthyroid syndrome: changes in thyroid hormone serum parameters and hormone metabolism Clin Endocrinol (Oxf) 1993 39 499 518 8252737
Gomberg-Maitland M Frishman WH Thyroid hormone and cardiovascular disease Am Heart J 1998 135 187 196 9489964
Eber B Schumacher M Langsteger W Zweiker R Fruhwald FM Pokan R Gasser R Eber O Klein W Changes in thyroid hormone parameters after acute myocardial infarction Cardiology 1995 86 152 156 7728806
De Groot LJ Dangerous dogmas in medicine: the nonthyroidal illness syndrome J Clin Endocrinol Metab 1999 84 151 164 9920076 10.1210/jc.84.1.151
Chopra IJ Hershman JM Pardridge WM Nicoloff JT Thyroid function in nonthyroidal illness Ann Intern Med 1983 98 946 957 6407376
Kimura T Kanda T Kotajima N Kuwabara A Fukumura Y Kobayashi I Involvement of circulating interleukin-6 and its receptor in the development of euthyroid sick syndrome in patients with acute myocardial infarction Eur J Endocrinol 2000 143 179 184 10913935 10.1530/eje.0.1430179
Santini F Chopra IJ A radioimmunoassay of rat type I iodothyronine 5'-monodeiodinase Endocrinology 1992 131 2521 2526 1446594 10.1210/en.131.6.2521
Fliers E Guldenaar SE Wiersinga WM Swaab DF Decreased hypothalamic thyrotropin-releasing hormone gene expression in patients with nonthyroidal illness J Clin Endocrinol Metab 1997 82 4032 4036 9398708 10.1210/jc.82.12.4032
Pavlou HN Kliridis PA Panagiotopoulos AA Goritsas CP Vassilakos PJ Euthyroid sick syndrome in acute ischemic syndromes Angiology 2002 53 699 707 12463624
Friberg L Drvota V Bjelak AH Eggertsen G Ahnve S Association between increased levels of reverse triiodothyronine and mortality after acute myocardial infarction Am J Med 2001 111 699 703 11747849 10.1016/S0002-9343(01)00980-9
Biondi B Fazio S Palmieri EA Carella C Panza N Cittadini A Bone F Lombardi G Sacca L Left ventricular diastolic dysfunction in patients with subclinical hypothyroidism J Clin Endocrinol Metab 1999 84 2064 2067 10372711 10.1210/jc.84.6.2064
Hamilton M Prevalence and clinical implications of abnormal thyroid hormone metabolism in advanced heart failure Ann Thorac Surg 1993 56 1 Suppl S48 S52 8333797
Klein I Thyroid hormone and the cardiovascular system Am J Med 1990 88 631 637 2189307 10.1016/0002-9343(90)90531-H
Iervasi G Pingitore A Landi P Raciti M Ripoli A Scarlattini M L'Abbate A Donato L Low T3 syndrome: a strong prognostic predictor of death in patients with heart disease Circulation 2003 107 708 713 12578873 10.1161/01.CIR.0000048124.64204.3F
Longstreth WT JrManowitz NR DeGroot LJ Siscovick DS Mayor GH Copass MK Weinmann S Cobb LA Plasma thyroid hormone profiles immediately following out-of-hospital cardiac arrest Thyroid 1996 6 649 653 9001202
Wortsman J Premachandra BN Chopra IJ Murphy JE Hypothyroxinemia in cardiac arrest Arch Intern Med 1987 147 245 248 3101625 10.1001/archinte.147.2.245
D'Alecy LG Thyroid hormone in neural rescue Thyroid 1997 7 115 124 9086579
Facktor MA Mayor GH Nachreiner RF D'Alecy LG Thyroid hormone loss and replacement during resuscitation from cardiac arrest in dogs Resuscitation 1993 26 141 162 8290809 10.1016/0300-9572(93)90174-O
Anonymous Myocardial infarction redefined: a consensus document of the Joint European Society of Cardiology/American College of Cardiology Committee for the redefinition of myocardial infarction Eur Heart J 2000 21 1502 1513 10973764 10.1053/euhj.2000.2305
Alpert JS Thygesen K Antman E Bassand JP Myocardial infarction redefined: a consensus document of the Joint European Society of Cardiology/American College of Cardiology Committee for the redefinition of myocardial infarction J Am Coll Cardiol 2000 36 959 969 10987628 10.1016/S0735-1097(00)00804-4
Schiller NB Shah PM Crawford M DeMaria A Devereux R Feigenbaum H Gutgesell H Reichek N Sahn D Schnittger I Recommendations for quantitation of the left ventricle by two-dimensional echocardiography. American Society of Echocardiography Committee on Standards, Subcommittee on Quantitation of Two-Dimensional Echocardiograms J Am Soc Echocardiogr 1989 2 358 367 2698218
Veldhuis JD Hammond JM Endocrine function after spontaneous infarction of the human pituitary: report, review, and reappraisal Endocr Rev 1980 1 100 107 6785084
Baron JC Levasseur M Mazoyer B Legault-Demare F Mauguiere F Pappata S Jedynak P Derome P Cambier J Tran-Dinh S Thalamocortical diaschisis: positron emission tomography in humans J Neurol Neurosurg Psychiatry 1992 55 935 942 1431957
Toyoda K Fujii K Ibayashi S Nagao T Kitazono T Fujishima M Role of nitric oxide in regulation of brain stem circulation during hypotension J Cereb Blood Flow Metab 1997 17 1089 1096 9346434 10.1097/00004647-199710000-00011
Whitesall SE Mayor GH Nachreiner RF Zwemer CF D'Alecy LG Acute administration of T3 or rT3 failed to improve outcome following resuscitation from cardiac arrest in dogs Resuscitation 1996 33 53 62 8959774 10.1016/S0300-9572(96)00985-9
Dyke CM Yeh T JrLehman JD Abd-Elfattah A Ding M Wechsler AS Salter DR Triiodothyronine-enhanced left ventricular function after ischemic injury Ann Thorac Surg 1991 52 14 19 2069445
Hsu R Huang TS Chen YS Chu SH Effect of triiodothyronine administration in experimental myocardial injury J Endocrinol Invest 1995 18 702 709 8719301
Novitzky D Cooper DK Barton CI Greer A Chaffin J Grim J Zuhdi N Triiodothyronine as an inotropic agent after open heart surgery J Thorac Cardiovasc Surg 1989 98 972 978 2682025
Moruzzi P Doria E Agostoni PG Capacchione V Sganzerla P Usefulness of L-thyroxine to improve cardiac, and exercise performance in idiopathic dilated cardiomyopathy Am J Cardiol 1994 73 374 378 8109552 10.1016/0002-9149(94)90011-6
Ojamaa K Kenessey A Shenoy R Klein I Thyroid hormone metabolism and cardiac gene expression after acute myocardial infarction in the rat Am J Physiol Endocrinol Metab 2000 279 E1319 E1324 11093920
Klemperer JD Zelano J Helm RE Berman K Ojamaa K Klein I Isom OW Krieger K Triiodothyronine improves left ventricular function without oxygen wasting effects after global hypothermic ischemia J Thorac Cardiovasc Surg 1995 109 457 465 7877306
Ornato JP Peberdy MA Tadler SC Strobos NC Factors associated with the occurrence of cardiac arrest during hospitalization for acute myocardial infarction in the second national registry of myocardial infarction in the US Resuscitation 2001 48 117 123 11426473 10.1016/S0300-9572(00)00255-0
Dickey W Adgey J Mortality within hospital after resuscitation from ventricular fibrillation outside hospital Br Heart J 1992 67 334 338 1389711
|
16137355
|
PMC1269452
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 9; 9(4):R416-R424
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3727
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37291613735010.1186/cc3729ResearchIntensive care unit delirium is an independent predictor of longer hospital stay: a prospective analysis of 261 non-ventilated patients Thomason Jason WW 1Shintani Ayumi 2Peterson Josh F 3Pun Brenda T 4Jackson James C 5Ely E Wesley [email protected] Attending Physician, Division of Allergy/Pulmonary/Critical Care Medicine, Vanderbilt University School of Medicine, Nashville, TN, USA2 Research Assistant Professor of Biostatistics and Medicine, Departments of Internal Medicine, Divisions of General Internal Medicine and Center for Health Services Research, Vanderbilt University School of Medicine, Nashville, TN, USA3 Assistant Professor of Medicine and Bioinformatics, Departments of Internal Medicine, Divisions of General Internal Medicine and Center for Health Services Research, Vanderbilt University School of Medicine, Nashville, TN, USA4 Clinical Assistant Professor of Nursing, Division of Allergy/Pulmonary/Critical Care Medicine, Vanderbilt University School of Medicine, Nashville, TN, USA and Center for Health Services Research, Vanderbilt University School of Medicine, Nashville, TN, USA5 Research Assistant Professor of Medicine and Psychiatry, Division of Allergy/Pulmonary/Critical Care Medicine, Vanderbilt University School of Medicine, Nashville, TN, USA and Center for Health Services Research, Vanderbilt University School of Medicine, Nashville, TN, USA6 Associate Professor of Medicine, Division of Allergy/Pulmonary/Critical Care Medicine and Center of Health Services Research, Associate Director of Research, VA Tennessee Valley Geriatric Research, Education and Clinical Center (CRECC), Vanderbilt University School of Medicine, Nashville, TN, USA2005 1 6 2005 9 4 R375 R381 8 4 2005 4 5 2005 Copyright © 2004 Thomason et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
Delirium occurs in most ventilated patients and is independently associated with more deaths, longer stay, and higher cost. Guidelines recommend monitoring of delirium in all intensive care unit (ICU) patients, though few data exist in non-ventilated patients. The study objective was to determine the relationship between delirium and outcomes among non-ventilated ICU patients.
Method
A prospective cohort investigation of 261 consecutively admitted medical ICU patients not requiring invasive mechanical ventilation during hospitalization at a tertiary-care, university-based hospital between February 2002 and January 2003. ICU nursing staff assessed delirium and level of consciousness at least twice per day using the Confusion Assessment Method for the ICU (CAM-ICU) and Richmond Agitation-Sedation Scale (RASS). Cox regression with time-varying covariates was used to determine the independent relationship between delirium and clinical outcomes.
Results
Of 261 patients, 125 (48%) experienced at least one episode of delirium. Patients who experienced delirium were older (mean ± SD: 56 ± 18 versus 49 ± 17 years; p = 0.002) and more severely ill as measured by Acute Physiology and Chronic Health Evaluation II (APACHE II) scores (median 15, interquartile range (IQR) 10–21 versus 11, IQR 6–16; p < 0.001) compared to their non-delirious counterparts. Patients who experienced delirium had a 29% greater risk of remaining in the ICU on any given day (compared to patients who never developed delirium) even after adjusting for age, gender, race, Charlson co-morbidity score, APACHE II score, and coma (hazard ratio (HR) 1.29; 95% confidence interval (CI) 0.98–1.69, p = 0.07). Similarly, patients who experienced delirium had a 41% greater risk of remaining in the hospital after adjusting for the same covariates (HR 1.41; 95% CI 1.05–1.89, p = 0.023). Hospital mortality was higher among patients who developed delirium (24/125, 19%) versus patients who never developed delirium (8/135, 6%), p = 0.002; however, time to in-hospital death was not significant the adjusted (HR 1.27; 95% CI 0.55–2.98, p = 0.58).
Conclusion
Delirium occurred in nearly half of the non-ventilated ICU patients in this cohort. Even after adjustment for relevant covariates, delirium was found to be an independent predictor of longer hospital stay.
See related commentary
==== Body
Introduction
Delirium is defined as an acute change or fluctuation in mental status plus inattention, and either disorganized thinking or an altered level of consciousness at the time of the evaluation [1,2]. Numerous studies have described the incidence, prevalence, and costly impact of delirium with regard to patients in nursing homes and hospital wards [3-7], but few prospective investigations have focused on cohorts treated specifically within the intensive care unit (ICU). Several studies have now confirmed that delirium occurs in 60% to 80% of mechanically ventilated patients [2,8-10], though two investigations found a lower prevalence in an ICU cohort with a lesser severity of illness [11,12]. Among ventilated patients, this condition is independently associated with untoward clinical outcomes [10,13], including higher mortality [10]. In fact, every day spent in delirium was associated with a 10% higher risk of death and worse long-term cognitive function [10].
Only 5% of 912 critical care professionals surveyed in 2001 and 2002 reported monitoring for ICU delirium [14], and yet the Society of Critical Care Medicine (SCCM) has recommended routine monitoring for delirium for all ICU patients [15]. Because many aspects of delirium in the ICU may be preventable and/or treatable (e.g., hypoxemia, electrolyte disturbances, sleep deprivation, overzealous use of sedative agents), routine daily delirium monitoring may be justified in non-ventilated ICU patients if adverse outcomes were demonstrated among delirious patients within this population.
Therefore, we undertook this investigation to determine the incidence of delirium among non-ventilated ICU patients and to determine the association between delirium and length of stay in the ICU, length of stay in the hospital, and in-hospital mortality.
Materials and methods
Patients
The institutional review board at Vanderbilt University Medical Center (Nashville, TN, USA) approved this observational cohort study [16] as Health Insurance Portability Accountability Act compliant, and informed consent was waived. Enrollment criteria included any patient aged 18 years or older who was admitted for more than 24 hours to the medical ICU of Vanderbilt University's 658-bed medical center, and who did not require invasive mechanical ventilation. During the 11-month study interval from 1 February 2002 to 7 January 2003, all of the 261 patients who met the inclusion criteria were enrolled in the study and followed until either death or hospital discharge. None of the patients in this cohort have been previously published in other peer-reviewed manuscripts.
Data collection and study design
Nursing staff assessed sedation level via the Richmond Agitation-Sedation Scale (RASS; see Additional file 1) [17,18] and delirium status via the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU; see Additional file 2) as described in previous literature [2,19] (downloadable materials and discussion also available at [20]). Of note, the CAM-ICU has been validated in both non-ventilated and ventilated patient assessments [2,19]. These data were recorded prospectively at least once per 12-hour shift as part of routine nursing care. The implementation of delirium monitoring in our institution took place through a year-long quality assurance program. During this time, the validity and inter-rater reliability of the RASS and CAM-ICU were very high [16] and consistent with our previous reports [2,18]. Specifically, the compliance was 90% in over 2,000 patient bedside observations and agreement with reference standard CAM-ICU raters was high (kappa = 0.80). Information collected prospectively at the time of enrollment included patient demographics, severity of illness using the Acute Physiology and Chronic Health Evaluation II (APACHE II) [21] score, and admission diagnoses. The Charlson Comorbidity Index, which represents the sum of a weighted index that takes into account the number and seriousness of pre-existing co-morbid conditions, was calculated using ICD-9 codes as per Deyo et al. [22]. The diagnostic categories for ICU admission were recorded by the patients' medical teams as the diagnostic category most representative of the reason for ICU admission. Because this was not an intervention study, no specific treatment(s) were given to patients who were identified as delirious. All therapies with regard to sedation and delirium were left to the discretion of the physician team caring for each patient.
Delirium in the ICU was the independent variable for this study and was classified as in previous reports [9,10]. Patients who scored positive for delirium by the CAM-ICU at any time while in the ICU were categorized as 'Ever Delirium'. All others were categorized as 'Never Delirium'. The three dependent variables included lengths of stay in the ICU and in the hospital, and in-hospital mortality.
Statistical analysis
Fisher's exact tests, exact chi-square tests, and Wilcoxon rank sum tests were used as appropriate to determine whether or not baseline features differed between those with and without delirium. Cox proportional hazards regression analyses [23] were used to assess the effects of delirium on ICU length of stay, hospital length of stay, and time to in-hospital mortality. In order to conduct the most robust analysis of the relationship between delirium and the outcome variables, delirium was included as a time-dependent incidence variable, and coded as 0 for all days prior to the first delirious event and as 1 thereafter. Coma status was also included in each model as a time-dependent covariate and was coded similarly. Other baseline covariates included in each model were age, gender, race, APACHE II score, and Charlson co-morbidity index. Because of the limited number of events for the time to in-hospital mortality analysis, and in order to avoid consequences of over-fitting that might have resulted from including each covariate separately, principal component analysis was used to pool the effects of age, gender, race, APACHE II score, and Charlson for the mortality analysis only. Time-to-event curves were created according to the methods of Kaplan and Meier [24], and were compared using log-rank tests. All statistical analyses were conducted using SAS Release 8.0.2 (SAS Institute, Cary, NC, USA).
Results
Baseline characteristics
Of the 261 patients enrolled in the study, 125 (48%) experienced delirium. One patient was excluded from analysis because of persistent coma throughout the entire hospital stay, negating any attempts to define the presence or absence of delirium. Baseline characteristics of the patients are presented in Table 1, with the cohort divided into two groups: Ever Delirium (n = 125) and Never Delirium (n = 135). There were no significant differences between the Ever Delirium and Never Delirium groups for gender, race, Charlson co-morbidity scores, or admission diagnoses. The Ever Delirium patients were significantly older (mean 56 versus 49 years of age, p = 0.002), and had higher APACHE II scores (median 15 versus 11, p < 0.001). Primary medical diagnoses were similar between the groups, with pulmonary (e.g., chronic obstructive pulmonary disease exacerbation), gastrointestinal (e.g., variceal hemorrhage), and metabolic (e.g., drug overdose, diabetic ketoacidosis) syndromes being the most common reasons for admission to the ICU.
Clinical outcomes and multivariable analysis results
Lengths of stay
Results indicate that the Ever Delirium group stayed in the ICU one day longer (median days 4; interquartile range (IQR) 3 to 5 versus 3; IQR 2 to 4) and in the hospital two days longer (median days 5; IQR 2 to 8 versus 3; IQR 2 to 6) than the Never Delirium group. A Kaplan-Meier plot for the probability of remaining in the ICU according to the clinical distinction of Ever Delirium vs Never Delirium is shown in Fig. 1. A Kaplan-Meier plot for the probability of remaining in the hospital for the same groups is shown in Fig. 2. As shown in Table 2, at any given time during their ICU stay, patients who experienced at least one episode of delirium had a 29% greater risk of remaining in the ICU even after adjusting for age, gender, race, Charlson co-morbidity score, APACHE II score, and coma (hazard ratio (HR) 1.29; 95% confidence interval (CI) 0.98–1.69, p = 0.07). Similarly, patients who experienced delirium had a 41% greater risk of remaining in the hospital after adjusting for the same covariates (HR 1.41; 95% CI 1.05–1.89, p = 0.023).
In-hospital mortality
Of the patients in the Ever Delirium group, 19% died versus 6% of the Never Delirium patients. A Kaplan-Meier plot for the probability of death according to the clinical distinction of Ever Delirium versus Never Delirium is shown in Fig. 3. Cox proportional hazards regression results indicated that delirium was not significantly associated with time to in-hospital mortality after controlling for coma status, age, gender, race, APACHE II score, and Charlson co-morbidity index (p = 0.58; Table 2).
Discussion
Delirium developed in approximately half of the patients in our cohort, and was associated with a one day longer stay in the ICU and a two day longer stay in the hospital. This is the first investigation to document the high prevalence of delirium among a strictly non-ventilated adult ICU cohort, and to reveal its associated negative clinical outcomes. Considering the rising overall resource use and economic burden of caring for critically ill patients [25-27], our finding that ICU delirium is an independent predictor of longer stay in the hospital is of particular relevance. These data lend support to the SCCM clinical practice guideline recommendation [15] for routine monitoring of delirium for all adult ICU patients using validated tools such as the CAM-ICU, which has been validated in ventilated and non-ventilated critically ill patients [2,19].
We did not find a significant independent relationship between delirium and mortality after adjusting for multiple covariates. This may simply be a type II error due to the limited number of events, and our study was not prospectively powered to determine a definitive relationship between delirium and mortality. Furthermore, because we only followed patients until hospital death or discharge, our mortality analysis was not as comprehensive as previous reports that followed patients for 6 to 12 months [10,28]. While these ICU patients had a lower severity of illness than those in prior ICU studies isolated to ventilated patients, the myriad of data in other non-ICU populations showing delirium to be associated with prolonged stay, greater dependency of care, subsequent institutionalization, and increased mortality [3,5-7,12,28-35] would cause one to pause before assuming that our study disproves such a consistently strong association.
The dangerous and costly considerations of prolonged ICU and hospital stays shown in this cohort warrant strong consideration by multidisciplinary ICU teams. Standardized clinical monitoring of brain function (both arousal level and delirium) is in keeping with the 'systems approach' to patient assessment. Because the development of delirium is associated with untoward outcomes, one author has questioned whether or not missing the diagnosis is a medical error [36]. Considering that symptoms of ICU delirium are largely hypo- rather than hyper-active [37,38], anything short of objectively looking for delirium will result in undetected brain dysfunction. Thus, the alternative to daily monitoring for delirium is to persist with the status quo in which an estimated 60% to 80% of delirium is missed in the absence of standardized monitoring [37-41].
The strengths of this report include the unique patient population (non-ventilated ICU patients), the large number of patients enrolled (n = 261), and the consecutive enrollment process that spanned nearly a year. All data were derived from sedation scoring and delirium assessments by the bedside nurses as part of a multidisciplinary approach to care within the ICU using well-validated tools (RASS and CAM-ICU) on a frequent basis (i.e., at least once every 12 hours). Previous studies regarding the incidence of delirium have used either q-24 hour or q-weekly assessments. Study personnel performed spot checks prospectively, accuracy was confirmed [16], and data were analyzed using robust statistical methods. In fact, rather than simple logistic regression, we chose the more sophisticated approach using time-to-event analysis with Cox regression and treated both delirium and death as time-dependent covariates.
Several limitations of this study warrant comment. First of all, we did not have a tool to stratify by the severity of delirium. If such a tool had been available and employed, we may have been better able to recognize patients who were at the highest risk for negative outcomes. Currently, no validated measure to stratify the severity of delirium exists, though work in this area is ongoing. Second, a recurrent limitation in all cohort studies is that there may be unknown covariates that influence outcomes. Third, this observational investigation was not designed to prove a cause-and-effect relationship between delirium and clinical outcomes. It is certainly true that the delirium group was older and had a higher severity of illness, though our multivariable analysis was specifically designed to take these covariates into account. Ultimately, further research incorporating a randomized, prospective clinical trial focused either upon the prevention or treatment of delirium will be necessary to confirm such a relationship. Data from other investigations, however, suggest that such a cause-and-effect between delirium and negative clinical outcomes exists. For example, in response to systemic infections and injury, brain dysfunction may ensue, which will then lead to the generation of a central nervous system inflammatory response of its own. This process involves the production of specific cytokines, cell infiltration, and tissue damage [42,43]. Additionally, activation of the central nervous system's immune response is accompanied by the peripheral production of tumor necrosis factor α, interleukin 1, and interferon δ [42,44-46] that can contribute to multiple organ dysfunction syndrome. It is plausible, therefore, that the delirium experienced among our patients is not only a marker of end-organ damage, but also acts directly as a promoter of other organ system dysfunction.
Conclusion
Nearly one out of every two non-ventilated adult ICU patients in our cohort experienced delirium. Even after adjustment for multiple covariates, delirium was associated with a longer ICU stay and was an independent predictor of a longer hospital stay. We believe that these data are clinically significant, reinforce the SCCM clinical practice guidelines for the delivery of sedation and analgesia calling for routine delirium monitoring of all patients (including those not on mechanical ventilation), and should stimulate future research in the field of delirium prevention and treatment.
Key messages
• Delirium is a form of brain dysfunction known to be associated with higher mortality, cost, and long-term cognitive impairment in mechanically ventilated adults.
• The SCCM guidelines for sedation and analgesia recommend that ICU teams routinely monitor all ICU patients (ventilated ornot) for delirium, though little data exist for the non-ventilated group.
• In this prospective cohort study, delirium was detected using the CAM-ICU, which has been validated for use in both ventilated and non-ventilated patients. We found that delirium occurred in one out of every two non-ventilated ICU patients.
• Even after adjustment for relevant covariates, delirium was found to be an independent predictor of longer hospital stay. While univariate analysis found an association with higher mortality, that association did not reach statistical significance in the multivariable analysis.
• This study lends clinical relevance to adoption of delirium monitoring in all ICU patients, both those on and off mechanical ventilation.
Abbreviations
APACHE II = Acute Physiology and Chronic Health Evaluation II; CAM-ICU = confusion assessment method for the ICU; CI = confidence interval; HR = hazard ratio; ICU = intensive care unit; IQR = interquartile range; RASS = Richmond Agitation-Sedation scale; SCCM = Society of Critical Care Medicine.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
Each author of this manuscript has: made substantial contributions to conception and design, acquisition of data, and the analysis or interpretation of data; been involved in drafting the article or revising it critically for important intellectual content; and given final approval of the submitted version to be published.
Supplementary Material
Additional File 1
A pdf file with the Richmond Agitation-Sedation Scale.
Click here for file
Additional File 2
A pdf file with the CAM-ICU Features and Descriptions
Click here for file
Acknowledgements
The authors would like to thank Gordon Bernard for his insight and helpful contributions, which guided us in our approach to this manuscript. We would also like to thank Meredith Gambrell for her extensive time and efforts in preparation of the manuscript. Most importantly, we would like to thank the dedicated and open-minded ICU staff, all of who strive daily to improve their care of critically ill patients. JWWT is supported by HL07123 from the National Heart Lung and Blood Institute, National Institute of Health. EWE is the Associate Director of Research for the VA Tennessee Valley Geriatric Research and Education Clinical Center (GRECC). He is a recipient of the Paul Beeson Faculty Scholar Award from the Alliance for Aging Research and is a recipient of a K23 from the National Institute of Health (#AG01023-01A1). No other financial support was provided to conduct this investigation.
Figures and Tables
Figure 1 Delirium versus ICU length of stay. This Kaplan-Meier plot shows the relationship between delirium and length of stay in the ICU by classification of Ever Delirium versus Never Delirium (p = 0.004, univariate analysis).
Figure 2 Delirium versus hospital length of stay. This Kaplan-Meier plot shows the relationship between delirium and hospital length of stay by classification of Ever Delirium versus Never Delirium (p < 0.001, univariate analysis).
Figure 3 Delirium versus in-hospital mortality. This Kaplan-Meier plot shows the relationship between delirium and in-hospital mortality by classification of Ever Delirium versus Never Delirium (p = 0.11, univariate analysis).
Table 1 Patient demographicsaa
Ever Delirium (n = 125) Never Delirium (n = 135) p-value
Characteristic
Mean age (± 1 SD; years) 56 (± 18) 49 (± 17) 0.002
Male 62 (50%) 67 (50%) 1.0
No. of Caucasians 99 (79%) 115 (85%) 0.25
APACHE II score, median (IQR) 15 (10–21) 11 (6–16) <0.001
Charlson co-morbidity index, median (IQR) 4 (2–7) 3 (1–6) 0.079
Diagnostic category for ICU admission (%)b
Pulmonary 29 40
Gastrointestinal 20 21
Metabolic 22 18
Cardiac 7 9
Hematology/oncology 5 4
Neurologic 5 3
Renal 9 2
Other 3 3
aOne patient of the 261 enrolled had persistent coma and was never able to be evaluated for delirium. This patient was not included in the tables or figures. bThe diagnostic categories for ICU admission were recorded by the patients' medical teams as the diagnostic category most representative of the reason for ICU admission. There was no statistically significant difference between the groups in terms of admission categories (p = 0.23). Acute Physiology and Chronic Health Evaluation II (APACHE II) is a severity of illness scoring system, and these data were calculated using the most abnormal parameters during the first 24 hours following admission to the intensive care unit. APACHE II scores range from 0 (best) to 71 (worst). The Charlson co-morbidity index represents the sum of a weighted index that takes into account the number and seriousness of pre-existing comorbidities. ICU, intensive care unit; SD, standard deviation.
Table 2 Clinical outcomes and multivariable analysis results
Ever Delirium (n = 125) Never Delirium (n = 135) Hazard ratioa (95% CI) p-valuea
LOS in ICUb 4 (3,5) 3 (2,4) 1.29 (0.98–1.69) 0.07
LOS in hospitalb 5 (2,8) 3 (2,6) 1.41 (1.05–1.89) 0.023
In-hospital mortalityc 24 (19%) 8 (6%) 1.27 (0.54–2.98) 0.58
aHazard ratios and p-values taken from multivariable Cox proportional hazards regression models adjusting for coma status, age, gender, race, APACHE II score, and Charlson co-morbidity index. bIntensive care unit (ICU) and hospital lengths of stay expressed as median days with interquartile ranges. cMortality expressed as n (%). CI, confidence interval; LOS, length of stay.
==== Refs
American Psychiatric Association Diagnostic and Statistical Manual of Mental Disorders 1987 Washington, DC: American Psychiatric Association
Ely EW Inouye SK Bernard GR Gordon S Francis J May L Truman B Speroff T Gautam S Margolin R Hart RP Dittus R Delirium in mechanically ventilated patients: validity and reliability of the confusion assessment method for the intensive care unit (CAM-ICU) J Am Med Assoc 2001 286 2703 2710 10.1001/jama.286.21.2703
Inouye SK Schlesinger MJ Lyndon TJ Delirium: a symptom of how hospital care is failing older persons and a window to improve quality of hospital care Am J Med 1999 106 565 573 10335730 10.1016/S0002-9343(99)00070-4
Morrrison RS Magaziner J Gilbert M Koval KJ McLaughlin MA Orosz G Strauss E Siu AL Relationship between pain and opioid analgesics on the development of delirium following hip fracture J Gerontol Med Sci 2003 58A 76 81
Kiely DK Bergmann MA Murphy KM Jones RN Orav EJ Marcantonio ER Delirium among newly admitted postactue facility patients: prevalence, symptoms, and severity Gerontol A Biol Sci Med Sci 2003 58 M441 M445
Hemert V Mast VD Hengeveld MW Excess mortality in general hospital patients with delirium: a five year follow up of 519 patients seen in psychiatric consultation J Psychosom Res 1994 38 339 346 8064651 10.1016/0022-3999(94)90038-8
Van Hemert AM Van Der Mast RC Hengeveld MW Vorstenbosch M Excess mortality in general hospital patients with delirium: a 5 year follow up of 519 patients seen in psychiatric consultation J Psychosomatic Res 1994 38 339 346 10.1016/0022-3999(94)90038-8
McNicoll L Pisani MA Zhang Y Ely EW Siegel MD Inouye SK Delirium in the intensive care unit: occurrence and clinical course in older patients J Am Geriatr Soc 2003 51 591 598 12752832 10.1034/j.1600-0579.2003.00201.x
Ely EW Gautam S Margolin R Francis J May L Speroff T Truman B Dittus R Bernard R Inouye SK The impact of delirium in the intensive care unit on hospital length of stay Intensive Care Med 2001 27 1892 1900 11797025 10.1007/s00134-001-1132-2
Ely EW Shintani A Truman B Speroff T Gordon SM Harrell FE JrInouye SK Bernard GR Dittus RS Delirium as a predictor of mortality in mechanically ventilated patients in the intensive care unit J Am Med Assoc 2004 291 1753 1762 10.1001/jama.291.14.1753
Bergeron N Dubois MJ Dumont M Dial S Skrobik Y Intensive Care Delirium Screening Checklist: evaluation of a new screening tool Intensive Care Med 2001 27 859 864 11430542 10.1007/s001340100909
Lin SM Liu CY Wang CH Lin HC Huang CD Huang PY Fang YF Shieh MH Kuo HP The impact of delirium on the survival of mechanically ventilated patients Crit Care Med 2004 32 2254 2259 15640638
Milbrandt EB Deppen S Harrison PL Shintani AK Speroff T Stiles RA Truman B Bernard GR Dittus RS Ely EW Costs associated with delirium in mechanically ventilated patients Crit Care Med 2004 32 955 962 AU: please provide the first 10 authors' names for this reference 15071384 10.1097/01.CCM.0000119429.16055.92
Ely EW Stephens RK Jackson JC Thomason J Truman B Bernard G Dittus R Current opinions regarding the importance, diagnosis, and management of delirium in the intensive care unit: A survey of 912 Healthcare professionals Crit Care Med 2004 32 106 112 14707567 10.1097/01.CCM.0000098033.94737.84
Jacobi J Fraser GL Coursin DB Riker RR Fontaine D Wittbrodt ET Chalfin DB Masica MF Bjerke HS Coplin WM Clinical practice guidelines for the sustained use of sedatives and analgesics in the critically ill adult Crit Care Med 2002 30 119 141 11902253 10.1097/00003246-200201000-00020
Truman B Shintani A Jackson J Peterson JF Thomason J Ely EW Implementation of the SCCM guidelines for sedation and delirium monitoring in the ICU Am J Respir Crit Care Med 2003 167 A969
Sessler CN Gosnell M Grap MJ Brophy GT O'Neal PV Keane KA Tesoro EP Elswick RK The Richmond Agitation-Sedation Scale: validity and reliability in adult intensive care patients Am J Respir Crit Care Med 2002 166 1338 1344 12421743 10.1164/rccm.2107138
Ely EW Truman B Shintani A Thomason JWW Wheeler AP Gordon S Francis J Speroff T Gautam S Margolin R Monitoring sedation status over time in ICU patients: the reliability and validity of the Richmond Agitation Sedation Scale (RASS) J Am Med Assoc 2003 289 2983 2991 10.1001/jama.289.22.2983
Ely EW Margolin R Francis J May L Truman B Dittus R Speroff T Gautam S Bernard G Inouye S Evaluation of delirium in critically ill patients: validation of the confusion assessment method for the intensive care unit (CAM-ICU) Crit Care Med 2001 29 1370 1379 11445689 10.1097/00003246-200107000-00012
ICU Delirium and Cognitive Impairment Study Group
Knaus WA Wagner DP Draper EA Zimmerman JE Bergner M Bastos PG Sirio CA Murphy DJ Lotring T Damiano A The APACHE III prognostic system: risk prediction of hospital mortality for critically ill hospitalized adults Chest 1991 100 1619 1636 1959406
Deyo RA Cherkin DC Ciol MA Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases J Clin Epidemiology 1992 45 613 619 10.1016/0895-4356(92)90133-8
Altman DG Andersen PK Bootstrap investigation of the stability of a cox regression model Stat Med 2003 8 771 783
Kaplan EL Meier P Nonparametric estimation from incomplete observations J Am Stat Assoc 1958 53 457 481
Halpern NA Pastores SM Greenstein RJ Critical care medicine in the United States 1985–2000: an analysis of bed numbers, use, and costs Crit Care Med 2004 32 1254 1259 15187502 10.1097/01.CCM.0000128577.31689.4C
Pronovost PJ Angus DC Dorman T Robinson KA Dremsizov TT Young TL Physician staffing patterns and clinical outcomes in critically ill patients J Am Med Assoc 2002 288 2151 2162 10.1001/jama.288.17.2151
Pronovost PJ Needham DM Waters H Birkmeyer CM Calinawan JR Birkmeyer JD Dorman T Intensive care unit physician staffing: financial modeling of the Leapfrog standard Crit Care Med 2004 32 1247 1253 15187501 10.1097/01.CCM.0000128609.98470.8B
McCusker J Cole M Abrahamowicz M Primeau F Belzile E Delirium predicts 12 month mortality Arch Intern Med 2002 162 457 463 11863480 10.1001/archinte.162.4.457
Francis J Kapoor WN Prognosis after hospital discharge of older medical patients with delirium J Am Geriatr Soc 1992 40 601 606 1587979
Inouye SK Bogardus ST JrCharpentier PA Leo-Summers L Acampora D Holford TR Cooney LM Jr A multicomponent intervention to prevent delirium in hospitalized older patients N Engl J Med 1999 340 669 676 10053175 10.1056/NEJM199903043400901
Lawlor PG Gagnon B Mancini IL Pereira JL Hanson J Suarez-Almazor ME Bruera ED Occurrence, causes, and outcome of delirium in patients with advanced cancer patients: a prospective study Arch Intern Med 2000 160 786 794 10737278 10.1001/archinte.160.6.786
Levkoff SE Evans DA Liptzin B Clearly PD Lipsitz LA Wetle TT Reilly CH Pilgrim DM Schor J Rowe J Delirium: The occurrence and persistence of symptoms among elderly hospitalized patients Arch Intern Med 1992 152 334 340 1739363 10.1001/archinte.152.2.334
Marcantonio ER Goldman L Mangione CM Ludwig LE Muraca B Haslauer CM Donaldson MC Whittemore AD Sugarbaker DJ Poss R A clinical prediction rule for delirium after elective noncardiac surgery J Am Med Assoc 1994 271 134 139 10.1001/jama.271.2.134
Rabins PV Folstein MF Delirium and dementia: diagnostic criteria and fatality rates Br J Psych 1982 140 149 153
Rockwood K Cosway S Carver D Jarrett P Stadnyk K Fisk J The risk of dementia and death after delirium Age Ageing 1999 28 551 556 10604507 10.1093/ageing/28.6.551
Sanders AB Missed delirium in older emergency department patients: a quality-of-care problem Ann Emerg Med 2002 39 338 341 11867994 10.1067/mem.2002.122273
Peterson JF Truman BL Shintani A Thomason JWW Jackson JC Ely EW The prevalence of hypoactive, hyperactive, and mixed type delirium in medical ICU patients J Am Geriatr Soc 2003 51 S174
Meagher DJ Hanlon DO Mahony EO Casey PR Trzepacz PT Relationship between symptoms and motoric subtype of delirium J Neuropsychiatry Clin Neurosci 2000 12 51 56 10678513
Meagher DJ Delirium: optimising management Br Med J 2001 322 144 149 11159573
Truman B Ely EW Monitoring delirium in critically ill patients: using the Confusion Assessment Method for the ICU Crit Care Nurse 2003 23 25 36 12725193
Camus V Burtin B Simeone I Schwed P Gonthier R Dubos G Factor analysis supports the evidence of existing hyperactive and hypoactive subtypes of delirium Intl J Geriatr Psych 2000 15 313 316 10.1002/(SICI)1099-1166(200004)15:4<313::AID-GPS115>3.0.CO;2-M
Perry VH Andersson B Gordon S Macrophages and inflammation in the central nervous system Trends Neurosci 1993 16 268 273 7689770 10.1016/0166-2236(93)90180-T
Rothwell NJ Luheshi G Toulmond S Cytokines and their receptors in the central nervous system: physiology, pharmacology and pathology Pharmacol Ther 1996 69 85 95 8984509 10.1016/0163-7258(95)02033-0
Woiciechowsky C Asudullah K Nestler D Eberhardt B Platzer C Schoning B Glockner F Lanksch WR Volk H Docke W Sympathetic activation triggers systemic interleukin-10 release in immunodepression induced brain injury Nat Med 1998 4 808 813 9662372 10.1038/nm0798-808
Woiciechowsky C Schoening B Daberkow N Asche K Stoltenberg G Lanksch WR Brain IL-1 beta induces local inflammation but systemic anti-inflammatory response through stimulation of both hypothalamic-pituitary-adrenal axis and sympathetic nervous system Brain Res 1999 816 563 571 9878881 10.1016/S0006-8993(98)01238-4
Nicholson TE Renton KW The role of cytokines in the depression of CYP1A activity using cultured astrocytes as an in vitro model of inflammation in the CNS Drug Metab Dispos 2002 30 42 46 11744610 10.1124/dmd.30.1.42
|
16137350
|
PMC1269454
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 1; 9(4):R375-R381
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3729
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37311613735210.1186/cc3731ResearchThe role of cardiac troponin I as a prognosticator in critically ill medical patients: a prospective observational cohort study King Daniel A [email protected] Shlomi [email protected] Victor [email protected] Leonid [email protected] Yaniv [email protected] Medical Intensive Care Unit, Soroka University Medical Center and the Faculty of Health Sciences, Ben-Gurion University, Beer Sheva, Israel2005 31 5 2005 9 4 R390 R395 13 2 2005 16 3 2005 30 3 2005 7 5 2005 Copyright © 2004 King et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
Myocardial injury is frequently unrecognized in intensive care unit (ICU) patients. Cardiac troponin I (cTnI), a surrogate of myocardial injury, has been shown to correlate with outcome in selected groups of patients. We wanted to determine if cTnI level measured upon admission is an independent predictor of mortality in a heterogeneous group of critically ill medical patients.
Methods
We conducted a prospective observational cohort study; 128 consecutive patients admitted to a medical ICU at a tertiary university hospital were enrolled. cTnI levels were measured within 6 h of admission and were considered positive (>0.7 ng/ml) or negative. A variety of clinical and laboratory variables were recorded.
Results
Both cTnI positive and negative groups were similar in terms of age, sex and pre-admission co-morbidity. In a univariate analysis, positive cTnI was associated with increased mortality (OR 7.0, 95% CI 2.44–20.5, p < 0.001), higher Acute Physiology and Chronic Health Evaluation (APACHE) II scores and a higher rate of multi-organ failure and sepsis. This association between cTnI and mortality was more pronounced among elderly patients (>65 years of age). Multivariate analysis controlling for APACHE II score revealed that elevated cTnI levels are not independently associated with 28-day mortality.
Conclusion
In critically ill medical patients, elevated cTnI level measured upon admission is associated with increased mortality rate. cTnI does not independently contribute to the prediction of 28-day mortality beyond that provided by APACHE II.
See related commentary
==== Body
Introduction
Assessing the severity of illness and outcome of critically ill patients is important as it influences management strategies and resource allocation. Historically, research aimed at determining factors associated with intensive care unit (ICU) mortality focused on individual risk factors and the development of multivariable prediction scores. These investigations consistently suggested the importance of organ system failure as strong predictors of both ICU and hospital mortality [1-4]. Over the past decade, several studies indicated that cardiac dysfunction is a frequent and important factor in determining the outcome of critically ill patients [5,6]. The pathophysiology of myocardial injury in critically ill patients is believed to be multifactorial, including the underlying disease process, hypoxemia and acidosis as well as therapeutic maneuvers [7,8]. It is estimated that as many as 15% of ICU admissions are complicated by some degree of myocardial injury and as many as 85% of patients with sepsis may have raised cardiac troponin [5,9]. Elevated serum levels of cardiac troponin I (cTnI), a myocardial regulatory protein of the thin actin filament, are considered highly sensitive and specific indicators of myocardial injury [10]. Serial measurement of cTnI is routinely used in the evaluation of patients with acute coronary syndromes (ACS) for diagnostic and prognostic purposes [10-12]. Several studies have assessed the prognostic value of elevated cTnI in critically ill patients without ACS. While some suggested that cTnI levels correlate with myocardial damage and poor outcome, others could not confirm this association [6,13-17]. Because cTnI elevation reflects organ failure (i.e. myocardial injury) its role as an additional marker of severity of illness and outcome is biologically plausible; however, limited information is available regarding the relative significance of cTnI elevation as an independent predictor of mortality in relation to the Acute Physiology And Chronic Health Evaluation (APACHE) II score. We hypothesized that elevated cTnI will not contribute to the mortality prediction provided by the multivariable APACHE II score. Therefore, we conducted a prospective cohort study in which the main purpose was to determine whether cTnI, measured upon admission, is an independent predictor of mortality in a heterogeneous group of critically ill medical patients.
Materials and Methods
Study location and population
The study was conducted within the medical ICU (MICU, eight beds) of Soroka University Medical Center, Beer-Sheva, Israel, a tertiary university hospital in Israel. All patients admitted to the MICU during a nine-month period (September 2002 to June 2003) were evaluated prospectively. The nursing staff and the physicians providing care for the patients in the MICU were completely blinded to the troponin results. A total of 128 consecutive patients were enrolled in this observational cohort study. All definitions were determined prospectively. The definitions used for sepsis, severe sepsis and organ failure were those used by the Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis (PROWESS) investigators [18]. Patients diagnosed with ACS, defined as unstable angina, typical chest pain, ischemic ECG changes or cardiogenic pulmonary edema were excluded as well as patients requiring chronic hemodialysis or patients who underwent major surgery during the month preceding admission (in order to exclude peri-operative myocardial injury). The Ethics Committee of the Soroka University Medical Center approved the study protocol prior to its initiation.
Data collection
One of the investigators who did not participate in routine patient care made daily rounds in the ICU recording relevant data from patient medical records and the hospital mainframe computer for reports of laboratory and microbiologic data. A complete history and physical examination was recorded on each patient enrolled in the study. APACHE II score was calculated based on the worst values for the first 24 h after ICU admission. Serum levels of cTnI were measured within 6 h of admission. The commercial assay AxSYM Troponin-I (Abbott laboratories, Abbott Park, Illinois, USA), a micro particle enzyme immunoassay, was used to determine cTnI levels. The assay characteristics were as follows: detection limit 0.3 ng/ml; analytical range 0–50 ng/ml; assay coefficient of variation (CV) range 5.2–7.8%. The cTnI cutoff is 0.7 ng/ml with a CV of 10%. Thus any cTnI blood level > 0.7 ng/ml was considered abnormal and indicative of myocardial injury. The primary end point was 28-day mortality. Secondary end points included days on mechanical ventilation, length of stay, and the number of failing organs.
Statistical analysis
Continuous variables' data are expressed as mean value ± SD. Bi-variate hypotheses involving continuous variables were tested with a t-test for independent variables with normal distribution and Mann-Whitney test for variables with abnormal distribution. Normality of the data was tested with a one-sample Kolmogorov-Smirnov test to indicate the appropriateness of parametric testing. Categorical variables are expressed as percentage; comparisons between groups were made using the chi square test. Logistic regression analysis was used to identify independent variables associated with death. Variables that were associated with mortality in univariate analysis were considered for inclusion in the model, whereas parameters already incorporated into the APACHE-II score, such as age, creatinine level and mean arterial pressure were not included. Cumulative survival curves were constructed using the Kaplan-Meier method and compared with the log-rank test. Results were considered significant at p < 0.05. Statistical analysis was performed with the SPSS 11.5 statistical package (SPSS Inc. Chicago, IL, USA).
Results
A total of 128 patients met inclusion criteria and were evaluated during the study period. Demographic and clinical characteristics of the study patients enrolled are shown in Table 1. The mean age of the study population was 53.9 ± 19 years and 68% were male. There was no significant difference between the cTnI positive and negative groups in terms of background illnesses; ischemic heart disease (20% versus 12%, p = 0.2), chronic obstructive pulmonary disease (COPD; 17% versus 12%, p = 0.43), diabetes mellitus (31% versus 17%, p = 0.08), arterial hypertension (34% versus 26%, p = 0.34) or malignancy (6% versus 3%, p = 0.61). Stratifying the patients according to the absence or presence of cTnI elevation revealed that of the various causes of admission, only sepsis was associated with elevated troponin (p = 0.008). Patients with elevated cTnI had a significantly higher APACHE II score (p < 0.001), required longer duration of mechanical ventilation (p = 0.004) and their mortality rate increased from 9.7% to 42.9% (OR 7.0, 95%CI 2.68–18.3, p < 0.001). Clinical variables upon admission, particularly vasopressor requirement, did not correlate with cTnI levels. Although creatinine levels were higher in the cTnI group, none of the patients required renal replacement therapy.
Table 2 shows demographic and clinical information of the study population stratified according to outcome. Patients who died had higher cTnI levels (p < 0.001), were significantly older (p = 0.001), had greater APACHE II scores (p = 0.001) and longer duration of mechanical ventilation (p < 0.001). In contrast, the cause of admission was not associated with differences in mortality rate. Of the clinical variables evaluated upon admission only mean arterial pressure (p = 0.006), creatinine level (p = 0.004) and vasopressor requirement were significantly associated with higher mortality (p = 0.02).
Among elderly (older than 65 years) patients (n = 49) with elevated cTnI levels, the 28-day mortality rate was 10/15 patients (66.7%), while the mortality rate among elderly patients with normal cTnI levels was only 4/34 patients (11.8%). This pattern was observed also in the younger patients, but to a lesser extent. Kaplan-Meier survival analysis in the younger and older age groups is shown in Fig. 1a,b. Log-rank tests for both groups were statistically significant: 0.04 and <0.001, respectively.
The results of the multivariate analysis (logistic-regression) are shown in Table 3. The two variables that were included in the final model were cTnI and APACHE II; both highly correlated with mortality in univariate analysis. Elevated cTnI levels were not found to be an independent predictor of mortality regardless of APACHE II score (OR 2.8, 95% CI 0.87–9.2, p = 0.085). Multivariate analysis of the subgroup of patients admitted without sepsis (n = 82) reveals that while APACHE II remained significantly associated with 28-day mortality (OR 1.2, 95% CI 1.03–1.28 per point increment), abnormal cTnI level was not (OR 1.2, 95% CI 0.21–7.1).
Discussion
We found that in critically ill medical patients, elevated cTnI is associated with increased mortality and longer duration of mechanical ventilation. cTnI does not, however, independently contribute to the prediction of 28-day mortality beyond that provided by APACHE II.
Cardiac troponin I and T are the most specific and sensitive laboratory markers of myocardial cell injury and may be elevated in patients presenting with many conditions other than acute coronary syndrome [9,11]. Elevated cTnI levels also correlate with decreased left ventricular function in both coronary and non-coronary patients [13,16]. Cardiac dysfunction during sepsis is fairly well documented and has been associated with poor prognosis [5]. Moreover, a small recent study evaluated the value of brain natriuretic peptide (BNP) plasma levels as a marker of systolic myocardial dysfunction during severe sepsis [19]. This study suggested that systolic dysfunction is present in 44% of patients with severe sepsis, BNP is useful in its detection and high plasma levels of BNP are associated with poor outcome [19]. It remains unclear though whether, in this context, elevated cTnI reflects reversible or irreversible myocardial damage [7,9]. Our data indicate that in patients over 65 there is a stronger correlation between elevated cTnI and mortality, which can probably be attributed to the extent, and possibly irreversibility, of myocardial damage in this age group. An interesting finding of this study was that most deaths among younger patients occurred within the first five days, whereas in the elderly group the majority of deaths (60%) occurred after this time frame. This data may suggest that younger patients who survive the initial insult do relatively well. Our study, however, was not designed nor powered to address the effect of age on outcome.
Several studies have addressed the prognostic value of elevated cTnI in non-coronary patients. In selected groups such as COPD and hemodyalisis patients, elevated cTnI correlated with poor outcome [20,21]. A study in emergency department patients has shown that there is a significant correlation between cTnI elevation and outcome. APACHE II is not provided, however, nor were the patients stratified by cause of admission. Therefore, no comparison between this study and ours could be performed [17]. Relos et al. [22], evaluating surgical ICU patients, suggested that moderate elevation of serum troponin I, which are below the threshold required to diagnose overt myocardial infarction, may reflect ongoing myocardial injury in the critically ill and are associated with a higher mortality rate and longer hospital and ICU length of stay. To the best of our knowledge, only one study suggested an independent predictive value of elevated cTnI after controlling for severity of illness assessed by APACHE II [23]. A strong correlation between mortality and elevated cTnI in critically ill medical patients without coronary disease was shown in this study. The sample size was rather small (58 patients), however, and the majority of patients had sepsis (88%), which limits the interpretation of these results. In contrast to these studies, Kollef et al. [6] suggested that serial measurements of cTnI do not independently contribute to the prediction of hospital mortality beyond that provided by clinically recognized cardiac dysfunction. Differences in design and patient mix preclude meaningful comparisons between this study and ours. Our observation that cTnI elevation is not an independent predictor of mortality is not surprising because troponin reflects a single system malfunction while the multivariable APACHE II reflects several highly relevant systems in the context of critically ill patients. It is, therefore, unlikely that a single assay will provide an independent additional value beyond that provided by APACHE II. Nonetheless, our finding that cTnI elevation is an important marker of severity of illness and is associated with high mortality rate is still clinically relevant, particularly in view of the fact that the Kaplan-Meier analysis indicates that the discriminative effect of cTnI elevation is evident from the first day.
The present study included a relatively small number of patients, limiting the significance of post-hoc subgroup analysis and our ability to identify other independent determinants of early mortality. The fact that the frequency of ischemic heart disease (IHD) was similar among cTnI positive patients and cTnI negative patients supports the assumption that the elevated cTnI in our study should not be attributed to ACS. As we did not systematically perform echocardiography or evaluation of coronary flow in these patients, more objective assessment of coronary anatomy and myocardial function is not available. Therefore, any correlation between cTnI levels, in these patients, and irreversible myocardial dysfunction or ACS remains deductive. As indicated earlier, however, elevated cTnI has been previously shown to correlate with left ventricular function. In our study, cTnI was sampled only once upon admission. Even though the time course and kinetics of cTnI and its relation to outcome may be of interest, the main purpose of our study was to determine whether early cTnI elevation is of clinically relevant importance.
Conclusion
We conclude that troponin elevation may be used as an early marker of severity of illness and outcome, particularly in older patients, but it is not an independent predictor of mortality. Additional larger prospective studies will be required to determine if a single serum marker, reflecting myocardial injury, could be established as an independent prognostic tool.
Key messages
• cTnI is a surrogate of myocardial injury in critically ill medical patients
• cTnI may be used as an early marker of outcome
• The correlation between elevated cTnI and mortality may be stronger among patients older than 65 years of age
• cTnI does not independently contribute to the prediction of 28-day mortality beyond that provided by APACHE II.
Abbreviations
ACS = acute coronary syndromes; APACHE = Acute Physiology and Chronic Health Evaluation; BNP = brain natriuretic peptide; COPD = chronic obstructive pulmonary disease; cTnI = cardiac troponin I; ICU = intensive care unit.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
DAK performed literature research, contributed to protocol planning, was a primary data gatherer and drafted the manuscript. SC participated in data collection and statistical analysis. VN performed statistical analysis. LB participated in data collection. YA supervised research, planned protocol and edited the article.
Figures and Tables
Figure 1 Kaplan-Meier curves showing mortality rates among patients. (a) 65 years of age and older or (b) younger than 65.
Table 1 Background characteristics of study population, according to elevated troponin result
All (n = 128)a cTnI positive (n = 35)a cTnI negative (n = 93)a p-value
Age (years) 53.9 ± 19 58.7 ± 18.6 52.1 ± 19 0.08
Older than 65 years 49 (38) 15 (43) 34 (37) 0.51
Male 68 (53.1) 19 (54.3) 49 (52.7) 0.87
APACHE II 15.3 ± 8.9 22.6 ± 10 12.7 ± 6.8 <0.001b
Cause of admission
Sepsis 46 (35.9) 19 (54.3) 27 (29) 0.008
Respiratory failure 35 (27.3) 6 (17.1) 29 (31.2) 0.112
Poisoning/drug overdose 10 (7.8) 3 (8.6) 7 (7.5) 0.665
GI hemorrhage 6 (4.7) 2 (5.7) 4 (4.3) 0.844
Miscellaneous 31 (24.2) 5 (14.3) 26 (28) 0.108
Parameters upon admission
Maximal temperature 37.9 ± 1.1 38.2 ± 1.3 37.8 ± 1 0.03b
Minimal mean arterial pressure 70.6 ± 21.7 63.2 ± 15.4 73.2 ± 23 0.02
White blood cell count (×1000) 14.9 ± 8.5 15.6 ± 7.8 14.5 ± 8.8 0.54
Creatinine (μmol/l) 142 ± 133 177 ± 106 124 ± 142 <0.001b
Use of pressors 48 (37.5) 17 (48.6) 31 (33.3) 0.11
28-day mortality 24 (18.8) 15 (42.9) 9 (9.7) <0.001
Length of stay (ICU) 3.3 ± 28.5 6.4 ± 6.6 2.2 ± 33.1 0.07b
Days on mechanical ventilation 3.7 ± 5.4 3.2 ± 5.1 5.4 ± 6.0 0.004b
aNumbers are n, percentages (in parentheses) or mean ± SD, as appropriate. bMan-Whitney test was applied due to abnormality of the data distribution. APACHE, Acute Physiology and Chronic Health Evaluation; ICU, intensive care unit. GI, Gastrointestinal hemorrhage
Table 2 Background characteristics of study population, according to outcome
All (n = 128)a Alive (n = 104)a Dead (n = 24)a p value
Age (years) 53.9 ± 19 51.6 ± 19.4 64 ± 13.6 0.001
Male 68 (53.1) 57 (54.8) 11 (45.8) 0.427
APACHE II 15.3 ± 8.9 13.6 ± 7.5 22.7 ± 10.9 <0.001b
Cause of admission
Sepsis 46 (35.9) 34 (32.7) 12 (50) 0.111
Respiratory failure 35 (27.3) 27 (26) 8 (33.3) 0.465
Poisoning / drug overdose 10 (7.8) 10 (9.6) 0 0.206
GI hemorrhage 6 (4.7) 5 (4.8) 1 (4.2) 1.0
Miscellaneous 31 (24.2) 28 (26.9) 3 (12.5) 0.137
Parameters upon admission
Maximal temperature 37.9 ± 1.1 37.8 ± 1.1 38.1 ± 1.4 0.111b
Minimal mean arterial pressure 70.6 ± 21.7 73.1 ± 22.1 59.5 ± 15.9 0.006
White blood cell count (×1000) 14.8 ± 8.5 14.7 ± 8.1 15.4 ± 10.2 0.723
Creatinine (μmol/l) 133 ± 133 133 ± 142 159 ± 88 0.004b
Use of pressors 48 (37.5) 34 (32.7) 14 (58.3) 0.02
cTnI > 0.7 35 (27.3) 10 (19.2) 15 (62.5) <0.001
Length of stay (ICU) 3.3 ± 28.5 2.7 ± 31.5 6.3 ± 5.9 0.045b
Days on mechanical ventilation 3.7 ± 5.4 3.1 ± 5.2 6 ± 5.7 <0.001b
aNumbers are n, percentages (in parentheses) or mean ± SD, as appropriate. bMan-Whitney test was applied due to abnormality of the data distribution. APACHE, Acute Physiology and Chronic Health Evaluation; ICU, intensive care unit. GI, Gastrointestinal hemorrhage
Table 3 Results of logistic regression analysis for mortality (n = 128)
Variable Coefficient Standard error OR 95% CI of OR p-value
cTnI 1.04 0.6 2.82 0.87–9.2 0.085
APACHE II 0.09 0.03 1.094 1.02–1.17 0.009
==== Refs
Knaus WA Mortality risk prediction in sepsis Crit Care Med 1995 23 1793 1794 7587252 10.1097/00003246-199510000-00038
Knaus WA Draper EA Wagner DP Zimmerman JE APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Knaus WA Harrell FE JrLynn J Goldman L Phillips RS Connors AF JrDawson NV Fulkerson WJ JrCaliff RM Desbiens N The SUPPORT prognostic model. Objective estimates of survival for seriously ill hospitalized adults. Study to understand prognoses and preferences for outcomes and risks of treatments Ann Intern Med 1995 122 191 203 7810938
Knaus WA Wagner DP Draper EA Zimmerman JE Bergner M Bastos PG Sirio CA Murphy DJ Lotring T Damiano A The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalized adults Chest 1991 100 1619 1636 1959406
Guest TM Ramanathan AV Tuteur PG Schechtman KB Ladenson JH Jaffe AS Myocardial injury in critically ill patients. A frequently unrecognized complication JAMA 1995 273 1945 1949 7783306 10.1001/jama.273.24.1945
Kollef MH Ladenson JH Eisenberg PR Clinically recognized cardiac dysfunction: an independent determinant of mortality among critically ill patients. Is there a role for serial measurement of cardiac troponin I? Chest 1997 111 1340 1347 9149592
Wu AH Increased troponin in patients with sepsis and septic shock: myocardial necrosis or reversible myocardial depression? Intensive Care Med 2001 27 959 961 11497152 10.1007/s001340100970
Parker MM Myocardial dysfunction in sepsis: injury or depression? Crit Care Med 1999 27 2035 2036 10507643 10.1097/00003246-199909000-00059
Ammann P Pfisterer M Fehr T Rickli H Raised cardiac troponins BMJ 2004 328 1028 1029 15117768 10.1136/bmj.328.7447.1028
Antman EM Tanasijevic MJ Thompson B Schactman M McCabe CH Cannon CP Fischer GA Fung AY Thompson C Wybenga D Braunwa E Cardiac-specific troponin I levels to predict the risk of mortality in patients with acute coronary syndromes N Engl J Med 1996 335 1342 1349 8857017 10.1056/NEJM199610313351802
Hamm CW Risk stratifying acute coronary syndromes: gradient of risk and benefit Am Heart J 1999 138 S6 S11 10385785 10.1053/hj.1999.v138.99081
Polanczyk CA Lee TH Cook EF Walls R Wybenga D Printy-Klein G Ludwig L Guldbrandsen G Johnson PA Cardiac troponin I as a predictor of major cardiac events in emergency department patients with acute chest pain J Am Coll Cardiol 1998 32 8 14 9669242 10.1016/S0735-1097(98)00176-4
Fernandes CJ JrAkamine N Knobel E Cardiac troponin: a new serum marker of myocardial injury in sepsis Intensive Care Med 1999 25 1165 1168 10551977 10.1007/s001340051030
Arlati S Brenna S Prencipe L Marocchi A Casella GP Lanzani M Gandini C Myocardial necrosis in ICU patients with acute non-cardiac disease: a prospective study Intensive Care Med 2000 26 31 37 10663277 10.1007/s001340050008
Thiru Y Pathan N Bignall S Habibi P Levin M A myocardial cytotoxic process is involved in the cardiac dysfunction of meningococcal septic shock Crit Care Med 2000 28 2979 2983 10966282 10.1097/00003246-200008000-00049
Turner A Tsamitros M Bellomo R Myocardial cell injury in septic shock Crit Care Med 1999 27 1775 1780 10507597 10.1097/00003246-199909000-00012
Wright RS Williams BA Cramner H Gallahue F Willmore T Lewis L Ladenson JH Jaffe AS Elevations of cardiac troponin I are associated with increased short-term mortality in noncardiac critically ill emergency department patients Am J Cardiol 2002 90 634 636 12231092 10.1016/S0002-9149(02)02570-5
Bernard GR Vincent JL Laterre PF LaRosa SP Dhainaut JF Lopez-Rodriguez A Steingrub JS Garber GE Helterbrand JD Ely EW Efficacy and safety of recombinant human activated protein C for severe sepsis N Engl J Med 2001 344 699 709 11236773 10.1056/NEJM200103083441001
Charpentier J Luyt CE Fulla Y Vinsonneau C Cariou A Grabar S Dhainaut JF Mira JP Chiche JD Brain natriuretic peptide: A marker of myocardial dysfunction and prognosis during severe sepsis Crit Care Med 2004 32 660 665 15090944 10.1097/01.CCM.0000114827.93410.D8
Baillard C Boussarsar M Fosse JP Girou E Le Toumelin P Cracco C Jaber S Cohen Y Brochard L Cardiac troponin I in patients with severe exacerbation of chronic obstructive pulmonary disease Intensive Care Med 2003 29 584 589 12589528
Farkouh ME Robbins MJ Urooj Zafar M Shimbo D Davidson KW Puttappa R Winston J Halperin JL Epstein EM Patel M Association between troponin I levels and mortality in stable hemodialysis patients Am J Med 2003 114 224 226 12637137 10.1016/S0002-9343(02)01482-1
Relos RP Hasinoff IK Beilman GJ Moderately elevated serum troponin concentrations are associated with increased morbidity and mortality rates in surgical intensive care unit patients Crit Care Med 2003 31 2598 2603 14605530 10.1097/01.CCM.0000089931.09635.D2
Ammann P Maggiorini M Bertel O Haenseler E Joller-Jemelka HI Oechslin E Minder EI Rickli H Fehr T Troponin as a risk factor for mortality in critically ill patients without acute coronary syndromes J Am Coll Cardiol 2003 41 2004 2009 12798573 10.1016/S0735-1097(03)00421-2
|
16137352
|
PMC1269455
|
CC BY
|
2021-01-04 16:04:55
|
no
|
Crit Care. 2005 May 31; 9(4):R390-R395
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3731
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37351613735410.1186/cc3735ResearchShort-term effects of positive end-expiratory pressure on breathing pattern: an interventional study in adult intensive care patients Haberthür Christoph [email protected] Josef [email protected] Assistant Professor and head of Surgical Intensive Care Medicine, Department of Anaesthesia, Kantonsspital Luzern, Switzerland2 Professor in Biomedical Engineering, Section of Experimental Anaesthesiology, Department of Anaesthesia and Critical Care Medicine, University of Freiburg, Germany2005 9 6 2005 9 4 R407 R415 13 1 2005 16 2 2005 18 4 2005 11 5 2005 Copyright © 2005 Haberthür and Guttmann; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Positive end-expiratory pressure (PEEP) is used in mechanically ventilated patients to increase pulmonary volume and improve gas exchange. However, in clinical practice and with respect to adult, ventilator-dependent patients, little is known about the short-term effects of PEEP on breathing patterns.
Methods
In 30 tracheally intubated, spontaneously breathing patients, we sequentially applied PEEP to the trachea at 0, 5 and 10 cmH2O, and then again at 5 cmH2O for 30 s each, using the automatic tube compensation mode.
Results
Increases in PEEP were strongly associated with drops in minute ventilation (P < 0.0001) and respiratory rate (P < 0.0001). For respiratory rate, a 1 cmH2O change in PEEP in either direction resulted in a change in rate of 0.4 breaths/min. The effects were exclusively due to changes in expiratory time. Effects began to manifest during the first breath and became fully established in the second breath for each PEEP level. Identical responses were found when PEEP levels were applied for 10 or 60 s. Post hoc analysis revealed a similar but stronger response in patients with impaired respiratory system compliance.
Conclusion
In tracheally intubated, spontaneously breathing adult patients, the level of PEEP significantly influences the resting short-term breathing pattern by selectively affecting expiratory time. These findings are best explained by the Hering–Breuer inflation/deflation reflex.
==== Body
Introduction
Pulmonary stretch receptors affect the resting respiratory pattern by vagal afferents. Increased stretch receptor activity shortens the duration of inspiration in animals [1], human newborns, and children [2-9]. This reflex is well known as the Hering–Breuer inflation reflex [10]. Increased stretch receptor activity also prolongs expiration, and maintained inflations – if sufficiently large – can produce apnoea for a considerable period of time. In humans the inflation/deflation reflex is substantially weaker than in most animal species [1,8]. In adult humans, the inflation/deflation reflex has been found to become apparent if the inflation volume exceeds a critical threshold of about 1 l [3,9,11]. Recent work revealed that the inflation/deflation reflex is also operative in normal tidal breathing if subject is sleeping [12,13] or under slight sedation [14]. Moreover, Tryfon and coworkers [14] demonstrated that raising the level of continuous positive airway pressure (or positive end-expiratory pressure [PEEP]) significantly prolongs expiration in parallel with a PEEP-related increase in functional residual capacity. However, the findings of those studies [12-14] became apparent only after an inspiratory hold manoeuvre and are therefore beyond the limits of physiological or common clinical conditions.
Our aim in the present study was to investigate whether these effects also become apparent during standard clinical conditions. To this end, we repeatedly applied a pattern of different levels of PEEP in a heterogeneous population of tracheally intubated patients during unsupported spontaneous breathing. Using this approach we were able to demonstrate that increased PEEP was strongly associated with prolongation of the expiratory time (resulting in a fall in both respiratory rate and minute ventilation [VE]), whereas tidal volume (VT) and inspiratory time were not significantly affected.
Materials and methods
Patients
Thirty patients were investigated during weaning from mechanical ventilation. The characteristics of the patients and their underlying diseases are summarized in Table 1. Sedation therapy had been discontinued in 21 (70%) patients 6 ± 7 hours before the start of the study. At entrance into the study, all patients were either awake or easy to arouse, corresponding to a Ramsey sedation score of 2 or 3 [15]. Patients were breathing spontaneously in the pressure support ventilation (PSV) mode with an inspiratory pressure assist of 6–12 cmH2O above a PEEP of 5–10 cmH2O. For a patient to be enrolled in the study, the fractional inspired oxygen (FiO2) required to maintain their arterial oxygen tension and arterial oxygen saturation above 10 kPa (75 mmHg) and 92%, respectively, had to be 50% or less. Furthermore, the patients had to be able to maintain blood gas values within normal ranges, to exhibit no severe coughing or repeated swallowing, or obvious signs of discomfort, and to have stable cardiovascular and respiratory conditions (i.e. absence of systematic changes to ventilatory pattern, FiO2 or PEEP requirements, or haemodynamic conditions). Of these patients, 19 (63%) could be successfully extubated within 1–24 hours after termination of the study. The remaining 11 patients (37 %) had to be kept on the ventilator for a mean of 3 ± 2 days for prolonged weaning from mechanical ventilation for various reasons.
Study design
The study protocol was approved by the ethics committee of our institution, and informed consent was obtained either from the patient (in the case of elective surgery) or from the patient's next of kin. After study inclusion criteria had been fulfilled, patients were connected to a modified study ventilator [16], in which the level of PEEP was set to 5 cmH2O. If necessary, FiO2 was adjusted to maintain oxygenation within comfortable ranges (i.e. arterial oxygen tension and arterial oxygen sturation above 11 kPa [82.5 mmHg] and 94%, respectively). After 20-min period to stabilize the temperature and humidity of the system and to allow the patient to become accustomed to the ventilator, the ventilatory mode was switched from conventional PSV to the automatic tube compensation mode [16,17]. We used automatic tube compensation mode (i.e. tracheal continuous positive airway pressure) instead of conventional PSV mode in order to eliminate the influence of flow-dependent pressure drop across the endotracheal tube (ETT) [17,18].
The study ventilator was driven by an external control unit, which – in accordance with the study protocol – automatically changes the PEEP level in synchrony with the patient's breathing pattern (i.e. upward steps in PEEP during inspiration and downward steps during expiration). Even though this helps to minimize interference with the patient's breathing pattern, the duration of expiration of the first breath after PEEP adjustment was shortened and prolonged, respectively, by upward steps and downward steps in PEEP. Consequently, these breaths had to be rejected from our analyses.
Instead of airway pressure, the tracheal pressure was the target for PEEP adjustment. To this end, the tracheal pressure was continuously calculated [19] and repeated spot-check measurements between investigations guaranteed for quality of the calculation procedure [20]. The study protocol includes three levels of PEEP that were applied in the following sequence: 0, 5 and 10 cmH2O, and the 5 cmH2O again. Each level of PEEP was applied for 30 s (Fig. 1). To minimize time-dependent biases, the PEEP sequence was applied five times in a row. Additionally, we applied the study protocol for intervals of 10 s (n = 25) and 60 s (n = 18).
Measurements
The flow (V') was measured using a heated Fleisch No. 2 pneumotachograph (Metabo, Epalinges, Switzerland), which was placed at the proximal end of the ETT. Airway pressure was measured between the pneumotachograph and the outer end of the ETT. Spot-check measurement of tracheal pressure were taken by introducing a pressure measuring catheter into the trachea via the lumen of the ETT, as described in detail elsewhere [20]. The pressure transducers used to measure airway and tracheal pressures (32NA-005D; ICsensor, Milpitas, CA, USA) and the differential pressure transducer for measuring V' (CPS 10; Hoffrichter, Schwerin, Germany) were placed a small distance from the patient (20 cm) to achieve a good signal quality and short response time. Expiratory carbon dioxide was measured using a device (CO2 Analyzer 930; Siemens-Elema, Solna, Sweden) that was previously calibrated with a carbon dioxide concentration of 8.4%. Measured signals were digitized with 12-bit precision and stored at a rate of 100 Hz in a personal computer for further analyses. The mechanics of the respiratory system (i.e. compliance, resistance and intrinsic PEEP) were calculated from tracings originating from a preceding period of controlled mechanical ventilation [21,22], in which resistance is depicted as pure airway resistance (i.e. without the resistance of the ETT).
Analysis
Stored data were analyzed on a breath-by-breath basis after having rejected the breaths that were affected by transitions in PEEP level. We determined inspiratory time, expiratory time and respiratory rate by means of the flow signal, in which a combined criterion of flow and expiratory carbon dioxide concentration allowed us to differentiate between inspiration and expiration accurately [23]. VT was calculated by numerical integration of V'. VE was calculated by multiplying respiratory rate by VT. In 10 out of 30 patients, the exhaled carbon dioxide volume per minute was derived from the additive compound of flow and carbon dioxide samples during expiration. For subsequent statistical analysis, data for corresponding PEEP levels and identical time cues were averaged.
In a post hoc analysis we investigated the PEEP-related effect on breathing pattern in patients without (control group; n = 17) and with impaired lung mechanics (impaired group; n = 13). Allocation was roughly based on respiratory system compliance above or below 50 ml/cmH2O (Table 2). Because at the time of investigation pure airway resistance (i.e. without ETT resistance) was below 4 cmH2O·s/l in each patient, allocation according to airway resistance was not feasible.
Statistical analysis was performed using SYSTAT, version 5.2 (L. Wilkinson, M.A. Hill, E. Vang, Evanston, IL, USA). Differences in all parameters of interest were assessed by an analysis of variance suitable for repeated measures with four repeated within factors (i.e. the PEEP levels) and either no between factor (main analysis) or one between factor (group) in the post hoc analysis. Significance was expressed as the arithmetic mean of Greenhouse–Geisser's and Huynh–Feldt's adjusted P values. If the overall model showed significant results, then changes between PEEP levels were calculated. For baseline group differences in the post hoc analysis, the Kruskal–Wallis test was used to compare continuous variables and the χ2 test was used to compare categorical data. For all tests a two-sided α level of P < 0.025 was considered statistically significant. All data are presented as mean ± standard deviation, unless otherwise stated.
Results
For the investigation with 30 s intervals, a total of 175 ± 88 breaths per patient were eligible for statistical analysis. In each patient the increase in PEEP from 0 to 10 cmH2O was associated with a fall in VE from on average 11.6 ± 3.0 l/min to 10.0 ± 2.0 l/min (P < 0.0001). Whereas VT was unaffected by the increase in PEEP (545 ± 184 ml versus 550 ± 163 ml at PEEP 0 and 10 cmH2O, respectively; P = 0.571), the fall in VE was due to a decrease in respiratory rate from 23.6 ± 9.6 breaths/min at zero PEEP to 19.9 ± 7.3 breaths/min at 10 cmH2O PEEP (P = 0.001). The decrease in respiratory rate was due to a significant increase in expiratory time (from 2111 ± 893 ms at zero PEEP to 2599 ± 1047 ms at 10 cmH2O PEEP; P < 0.0001), whereas inspiratory time remained unaffected (869 ± 321 ms and 880 ± 312 ms at PEEP 0 and 10 cmH2O, respectively; P = 0.116). Expiratory carbon dioxide volume significantly decreased with increasing levels of PEEP (P < 0.01), whereas the end-expiratory carbon dioxide did not differ significantly.
For the variables of interest, the differences between all levels of PEEP are shown in Table 3; Fig. 2 shows the corresponding percentage changes. Figures 3 and 4 show the PEEP-related effects on VE and expiratory time in individual patients. An identical pattern was found when the PEEP pattern was applied for intervals of 10 s or 60 s (Fig. 5). Figure 6 shows changes in breathing pattern in the first, second and third breaths after steps of PEEP in either direction. Changes in breathing pattern had begun to manifest during the first breath and became fully established in the second breath after both upward and downward steps in PEEP.
Post hoc analysis
In patients with decreased respiratory system compliance, respiratory rate and VE were significantly higher and inspiratory time, expiratory time and VT were significantly smaller as compared with the corresponding parameters in the control group. Irrespective of these group differences, the PEEP-related effect on breathing pattern was the same in both groups (i.e. expiratory time increased, VE and respiratory rate decreased, and inspiratory time and VT remained unaffected by increases in the level of PEEP, and vice versa). Effects were more pronounced in patients with impaired respiratory system compliance than in the control group (P < 0.025). In both groups, the changes became fully established at the latest within the second breath after a change in PEEP.
Discussion
The findings of this clinical study, conducted in a heterogeneous population of adult intensive care patients, indicate that the level of PEEP significantly influences resting short-term breathing patterns by selectively affecting the duration of expiration. Thus, a reduction in PEEP is paralleled by an increase in respiratory rate and subsequently in VE, and vice versa. According to our findings, the magnitude of the effect is about 0.4 breaths/min per 1 cmH2O change in PEEP in either direction, and so it is about 10 times smaller than the effect found in anaesthetized animals [24]. Our findings are in accordance with the results of other studies in adult humans at normal tidal breathing [12-14], in which the findings were attributed to the Hering–Breuer inflation/deflation reflex [14]. In those studies, however, the effect became apparent only with an inspiratory hold manoeuvre. In contrast, we were able to show this effect also at normal tidal breathing under quite common clinical conditions (i.e. without any dedicated respiratory manoeuvre). Furthermore, we were also able to demonstrate that the PEEP-related effect had already begun to manifest during the first breath and became fully established in the second breath after adjustment to PEEP level. Consistent with these findings, we could not find any substantial difference when the study protocol was applied with PEEP durations as short as 10 s and up to 60 s.
The PEEP-related effect upon breathing pattern found in our heterogeneous study population was not only significant but also substantial. However, whether these findings are of clinical importance remains unclear. On the one hand, this is because in our study we focused on short-term rather than on long-term effects. From a theoretical point of view, the short-term responses seen here might simply be offset by a subsequent shift in blood carbon dioxide concentration beyond the time frame of 60 s. Such an hypothesis is strongly supported by the slight but significant decrease in exhaled carbon dioxide volume with increasing levels of PEEP (as a result of the PEEP-related drop in VE).
To investigate whether the PEEP-related short-term effect is only transient, we applied our study protocol to 10 of the 28 patients for durations of 180 and 300 s per level of PEEP. Unfortunately, the obtained data were inconsistent for the vast majority of these patients. This was due to the increased rate of artifacts (mainly due to restlessness during awakening) associated with the prolonged duration of investigation. Consequently, because our study was not designed to be a long-term investigation, we are unable to draw any conclusions on whether the observed short-term responses are preserved over time or whether they are offset by slowly adapting reflexes or behavioural responses over a longer period of observation. On the other hand, the PEEP-related short-term effect is sufficiently substantial that it could be detected by attentive clinicians, and so it could be used to adapt ventilatory settings for weaning from mechanical ventilation. Furthermore, short-term effects are of the utmost importance with respect to the automated weaning procedures and closed-loop ventilation strategies that now are increasingly being applied worldwide [25].
The question then arises as to the mechanism by which changes in PEEP can affect the breathing pattern, as found in this study. Based on findings in animals and humans, it is likely that the PEEP-related effect is best explained by the Hering–Breuer inflation/deflation reflex [3,8,10,26-28]. However, other mechanisms are also possible. First, the effect might be due to behavioural responses. At study entrance, all patients were either awake or easy to arouse because sedation had been discontinued 6 ± 7 hours before the start of the study in most patients and continued at a rather slight level in a few patients. Nevertheless, the PEEP-related responses were quite uniform; for example, they became fully established within the first two breaths after PEEP adjustment in virtually all patients. Furthermore, we did not find any difference in the PEEP-related effect between patients off and those who were still on (slight) sedation. In summary, behavioural responses are unlikely, although they cannot be ruled out.
Second, the PEEP-related effect could also be due to responses of the chemical feedback system. Based on both alveolar gas exchange (and thus on functional residual capacity as a function of PEEP) and the circulatory time (being approximately 6–7 s at normal conditions), any change in alveolar gas composition will influence the response of peripheral (mainly oxygen) and central (mainly carbon dioxide) chemoreceptors [3,8,24]. In the absence of lung overdistension the PEEP-related increase in end-expiratory lung volume would ameliorate gas exchange (if there were any effect on gas exchange at all), which then might result in a compensatory downregulation in ventilation. Effects will be stronger in patients with normal than in those with decreased lung compliance. Although such a possibility would fit well the findings of our study, the slight but significant decrease in expired volume of carbon dioxide with increasing levels of PEEP (resulting from the PEEP-related decrease in VE) does not support this hypothesis. In addition, if predominantly chemical feedback responses were at work, then different responses would be expected for the different periods of time for which the PEEP levels were applied (i.e. slight responses with durations of 10 s but stronger responses with durations of 60 s). Finally, the occurrence of the effect within the first breath after PEEP adjustment could hardly be accounted for by chemical feedback mechanisms.
A third alternative explanation for the observed PEEP-related effect might relate to persistent inspiratory muscle activity during exhalation [29,30]. The presence of persistent inspiratory muscle activity would prolong the expiratory time either independent of or in addition to the reflex-related response. The effect of persistent inspiratory muscle activity would manifest as a depressing influence on expiratory flow rate. However, because expiratory flow (i.e. peak flow rate and expiratory volume) was unaffected by the level of PEEP (Table 3), any hypothesis based on persistent inspiratory muscle activity as the source of the observed PEEP-related effect must be rejected.
In conclusion, the PEEP-related short-term effect on breathing pattern found in the present is best explained by neuronal reflex mechanisms (i.e. the Hering–Breuer inflation/deflation reflex).
In the post hoc analysis we found a similar response to changes of PEEP in patients with normal and those with decreased respiratory system compliance. For the latter, however, the effects were stronger. Consistent with our results, Tryfon and coworkers [14] found a less sensitive reflex in patients with supranormal compliance (i.e. patients with chronic obstructive pulmonary disease) as compared with control individuals (i.e. with relatively lower compliance) or patients with interstitial fibrosis (i.e. with decreased compliance). At first glance these findings are surprising, because with higher respiratory system compliance alterations in PEEP should result in stronger changes in lung volume, and so stronger PEEP-related effects on breathing pattern are anticipated in this setting. The findings of our and Tryfon's study refute this, which might be related to the expiratory muscle recruitment that has been found in awake (but not sleeping) subjects in response to increased end-expiratory lung volume [29,31,32]. Expiratory muscle recruitment due to a PEEP-related increase in end-expiratory lung volume would be expected predominantly at normal rather than at decreased respiratory system compliance. The finding of attenuated PEEP-related effects in patients with normal compliance fit well with this hypothesis. Even if some of our patients were under slight sedation (but easy to awake) during the investigation, the hypothesis might hold true because we did not find any difference in PEEP-related effect between wakeful patients and those under slight sedation.
An alternative hypothesis for our unexpected results centres on the potential occurrence of intrinsic PEEP during the investigation. If this is the case, then intrinsic PEEP would have occurred predominantly in patients with normal rather than in those with decreased respiratory system compliance. Consequently, the increase in external PEEP would have reduced the work of breathing, and thus would have attenuated changes in the PEEP-related effect on breathing pattern predominantly in patients with normal respiratory system compliance (as found in our and Tryfon's study). However, there was no evidence for the occurrence of intrinsic PEEP, at least during the preceding period of controlled mechanical ventilation, in which VE was at a similar level as during the observational period of the study. In addition, careful examination of the expiratory flow pattern during the investigation did not suggest expiratory flow limitation as an indirect sign of the occurrence of intrinsic PEEP.
Conclusion
In tracheally intubated, spontaneously breathing adult patients, the level of PEEP significantly influences the resting short-term breathing pattern by selectively affecting expiratory time. The mechanism is probably based on the Hering–Breuer inflation/deflation reflex. Further studies are needed to address counteracting behavioural and/or slow responses of chemical respiratory control, and therefore to elucidate the clinical importance of our findings.
Key messages
• In spontaneous breathing patients upwards steps in PEEP significantly decrease respiratory rate and minute ventilation whereas downward steps have just opposite effects
• The PEEP-related effects are exclusively due to alteration of the expiratory time
• Effects become fully established within the first two breaths after PEEP adjustment and went on for minimally one minute
• Findings of this study and theoretical considerations strongly suggest a reflex related response by the Hering-Breuer inflation/deflation reflex
Abbreviations
ETT = endotracheal tube; FiO2 = fractional inspired oxygen; PEEP = positive end-expiratory pressure; PSV = pressure support ventilation; VE = minute ventilation; VT = tidal volume.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
CH designed the study, carried out the measurements, performed the statistical analysis, and drafted the manuscript. JG conceived the study, and participated in its design and helped to draft the manuscript. All authors read and approved the final manuscript.
Acknowledgements
The study was supported by a grant from the 'Stiftung Krokus', Basel, Switzerland.
Figures and Tables
Figure 1 Time course of tracheal pressure during steps in PEEP. The examination was performed while the patient was breathing spontaneously at (tracheal) continuous positive airway pressure by means of the automatic tube compensation mode. Starting from 5 cmH2O, the level of PEEP was changed to 0, 5, 10 and (again) 5 cmH2O for durations of 30 s each for five consecutive runs. Note that the changes in PEEP level were in synchrony with the patient's breathing pattern (i.e. upward steps during inspiration and downward steps during expiration). PEEP, positive end-expiratory pressure; Ptrach, tracheal pressure.
Figure 2 Short-term effects of PEEP on breathing pattern. The increase in PEEP is paralleled by an increase in expiratory time with a concomitant fall in both respiratory rate and minute ventilation. The decrease in PEEP has an opposite effect. Tidal volume and inspiratory time were not significantly affected by changes in PEEP in either direction. Results are outlined as averaged percentage changes from zero PEEP and are expressed as mean ± 1 standard error of the mean (SEM). *P < 0.025, versus zero PEEP. PEEP, positive end-expiratory pressure.
Figure 3 Short-term effects of PEEP on VE in the individual patient. Results are from the investigation with PEEP steps of 30 s duration, and each coloured line shows findings for a different patient. Although VE is dispersed over a wide range, its behaviour was similar between patients (i.e. VE decreased with increasing levels of PEEP, and vice versa). PEEP, positive end-expiratory pressure; VE, minute ventilation.
Figure 4 Short-term effects of PEEP on expiratory time in the individual patient. Results are from the investigation with PEEP steps of 30 s duration, and each coloured line shows findings for a different patient. Although Tex is dispersed over a wide range, its behaviour was similar between patients (i.e. Tex increased with the increasing levels of PEEP, and vice versa). PEEP, positive end-expiratory pressure; Tex, expiratory time.
Figure 5 Effects on breathing pattern brought on by steps in PEEP of different duration. Effects are shown on VE, which behaved similarly whether the PEEP steps were applied for 10, 30, or 60 s. *P < 0.025, versus zero PEEP. PEEP, positive end-expiratory pressure; VE, minute ventilation.
Figure 6 Changes of breathing pattern in the first, second and third breath after steps of PEEP. Results are from the investigation with PEEP steps of 30 s duration. The shaded areas represent averaged values of all breath for the corresponding PEEP level; open bars indicate the first, second, and third breath after changes of PEEP. Values are expressed as means ± 1 standard error of the mean (SEM). Note that changes of breathing pattern were beginning to manifest within the first breath and became fully established in the second breath after both upward and downward steps in PEEP. *P < 0.025, versus mean values within the corresponding PEEP level. BTPS, body temperature pressure, (water damp) saturated; PEEP, positive end-expiratory pressure; rr, respiratory rate; Tex, expiratory time; Tin, inspiratory time; VE, minute ventilation.
Table 1 Patient characteristics
Patient Sex Age (years) Reason for intubation Underlying disease Duration MV (days) BSA (m2) PaO2/FiO2 (kPa/fraction) Crs (ml/cmH2O) Raw (cmH2O·s/l)
BP m 57 Coma Liver cirrhosis 3 2.1 27 65 1.1
BW m 68 ACBG CAD <1 1.96 36 77 1.8
BP f 69 ACBG CAD 2 1.82 29 54 2.5
CA f 56 Variceal bleeding Liver cirrhosis <1 1.87 39 88 1.2
FW m 58 ACBG CAD <1 2.06 30 71 2.1
FA f 65 Coma Diabetes mellitus 2 1.81 50 95 2.2
HH m 70 ACBG CAD <1 1.74 45 67 1.8
JH m 69 ACBG CAD <1 2 30 55 2.0
MG m 65 Valve replacement Aortic stenosis <1 1.92 30 64 1.9
PB m 60 ACBG CAD <1 2.2 45 90 1.9
PJ m 66 ACBG CAD <1 1.93 47 73 2.3
SH f 65 ACBG CAD <1 1.86 44 70 2.8
SE m 78 Coma Poisoning 6 2.05 30 83 1.9
PW m 55 Valve replacement Aortic regurgitation <1 1.92 43 62 1.7
HM f 67 ACBG CAD <1 1.84 33 90 2.1
HN m 65 Valve replacement Aortic stenosis <1 1.78 56 57 2.3
BR f 66 ACBG CAD <1 1.65 31 90 1.5
CF m 33 ARI Pneumonia 5 1.81 13 33 1.9
DM f 76 ARI Pneumonia 7 1.9 31 29 2.5
EW m 53 Resuscitation CAD, MI 2 1.94 19 47 2.1
FA m 54 ARI Pneumonia 12 2.11 25 35 3.0
RS m 53 ARI/ARDS Pneumonia 17 2.04 19 21 1.7
MA m 20 ARI/ARDS Pneumonia 13 1.88 23 35 1.4
PP m 31 Resuscitation CAD, AMI 6 2.15 23 41 2.3
PU m 51 ARI Pneumonia 13 1.97 19 48 2.6
SM m 70 ARI Pneumonia 4 1.87 14 19 2.1
SE f 63 ARI/ARDS Pneumonia 28 1.92 12 25 2.4
OE f 63 ARI/ARDS Pneumonia 22 1.92 23 31 1.8
TR f 68 Coma CNSj haemorrhage 2 1.89 45 47 2.9
WB f 53 ARI/ARDS Pneumonia 10 1.34 32 28 2.7
ACBG, aorto-coronary bypass grafting; AMI, acute myocardial infarction; ARDS, acute respiratory distress syndrome; ARI, acute respiratory insufficiency; BSA, body surface area; CAD, coronary artery disease; CNS, central nervous system; Crs, compliance of the respiratory system; FiO2, fractional inspired oxygen; MV, mechanical ventilation; PaO2, arterial oxygen tension; Raw, airway resistance (beyond the resistance of the endotracheal tube).
Table 2 Baseline characteristics: control versus impaired respiratory system mechanics
Parameter Control Impaired P
Number 17 13 -
Age (years ± SD) 65 ± 17 53 ± 13 <0.015
Sex (m/f) 11/6 8/5 NS
BSA (m2) 1.90 ± 0.20 1.91 ± 0.14 NS
PaO2/FiO2 (kPa/fraction) 38 ± 9 23 ± 9 <0.011
Crs (ml/cmH2O) 74 ± 14 34 ± 10 <0.001
Raw [cmH2O·s/l) 0.9 ± 0.4 1.2 ± 0.5 NS
Intrinsic PEEP (cmH2O) 0.1± 0.11 0.1 ± 0.2 NS
Shown are the baseline characteristics of patients without (control) and those with impaired respiratory system mechanics (impaired) according to compliance below or above 50 mL/1 cmH2O at the time of study entrance (post hoc analysis). BSA, body surface area; Crs, compliance of the respiratory system; FiO2, fractional inspired oxygen; Intrinsic PEEP, intrinsic positive end-expiratory pressure (in addition to external applied PEEP); PaO2, arterial oxygen tension; Raw, airway resistance (not included the resistance of the endotracheal tube).
Table 3 Changes in breathing pattern produced by gradual changes in PEEP
Variable of interest Changes in PEEP (cmH2O)
from 0 to 5 from 5 to10 from 10 to 5 from 5 to 0 P
Tex (ms) 229 ± 326 260 ± 400 -271 ± 282 -218 ± 249 <0.001
rr (breaths/min) -2.6 ± 4.4 -1.1 ± 3.5 1.3 ± 2.4 2.4 ± 4.1 <0.001
VE (l/min) -0.5 ± 1.9 -1.2 ± 1.1 1.1 ± 1.6 0.4 ± 1.3 <0.001
Tin (ms) 14 ± 73 -3 ± 54 7 ± 57 -18 ± 61 NS
VT (ml BTPS) 26 ± 93 -21 ± 40 45 ± 97 -50 ± 41 NS
ee-CO2 (%) 0.0 ± 0.2 0.2 ± 0.2 -0.1 ± 0.2 -0.1 ± 0.2 NS
Vex-CO2 (ml/min) -27 ± 73 -82 ± 20 67 ± 99 42 ± 106 <0.01
V'peak.ex (l/s) 0.08 ± 0.08 0.02 ± 0.09 0.00 ± 0.09 -0.10 ± 0.08 NS
VT,ex (ml BTPS) 18 ± 89 -18 ± 42 49 ± 96 -50 ± 40 NS
ee-CO2, end-expiratory carbon dioxide concentration; rr, respiratory rate; Tex, expiratory time; Tin, inspiratory time; VE, minute ventilation; Vex-CO2, exhaled carbon dioxide volume (derived from 10 patients); V'peak.ex, peak expiratory flow rate; VT, tidal volume; VT,ex, expired volume per breath.
==== Refs
Issa FG Porostocky S Effect of sleep on changes in breathing pattern accompanying sigh breaths Respir Physiol 1993 93 175 187 8210757 10.1016/0034-5687(93)90004-T
Rabbette PS Costeloe KL Stocks J Persistence of the Hering-Breuer reflex beyond the neonatal period J Appl Physiol 1991 71 474 480 1938718
Coleridge HM Coleridge JCG Reflexes evoked from tracheobronchial tree and lungs Handbook of Physiology, Section 3: the Respiratory System, Control of Breathing, Part 1 1986 II Baltimore, MA: The Williams & Wilkins Company 395 430
Martin RJ Nearman HS Katona PG Klaus MH The effect of a low continuous positive airway pressure on the reflex control of respiration in the preterm infant J Pediatr 1977 90 976 981 323448
Griffin F Greenough A Naik S The Hering-Breuer reflex in ventilated children Respir Med 1996 90 463 466 8869439 10.1016/S0954-6111(96)90172-9
Greenough A Pool J Hering-Breuer reflex in young asthmatic children Pediatr Pulmonol 1991 11 345 349 1758760
Stanley NN Altose MD Cherniack NS Fishman AP Changes in the strength of lung inflation reflex during prolonged inflation J Appl Physiol 1975 38 474 480 168175
Widdicombe JG Widdicombe JG Reflex control of breathing MTP International Review of Science, Physiology Series One, Respiratory Physiology 1974 3 London: Butterworth 273 301
Rabatte PS Stocks J Influence of volume dependency and timing of airway occlusions on the Hering-Breuer reflex in infants J Appl Physiol 1998 85 2033 2039 9843523
Hering E Breuer J Self-control of breathing by the vagus nerve [in German] Sitzber Deut Akad Wiss Wien 1868 57 672 690
Hamilton RD Winning AJ Horner RL Guz A The effect of lung inflation on breathing in man during wakefulness and sleep Respir Physiol 1988 73 145 154 3420318 10.1016/0034-5687(88)90062-X
Graves C Glass L Laporta D Meloche R Grassino A Respiratory phase locking during mechanical ventilation in anesthetized human subjects Am J Physiol 1986 250 R902 R909 3706575
Simon PM Zurob AS Wies WM Leiter JC Hubmayr RD Entrainment of respiration in humans by periodic lung inflations. Effect of state and CO2 Am J Respir Crit Care Med 1999 160 950 960 10471624
Tryfon S Kontakiotis T Mavrofridis E Patakas D Hering-Breuer reflex in normal adults and in patients with chronic obstructive pulmonary disease and interstitial fibrosis Respiration 2001 68 140 144 11287827 10.1159/000050483
Ramsay MAE Savege TM Simpson BRJ Goodwin R Controlled sedation with alphaxalone-alphadalone Br Med J 1974 2 656 659 4835444
Fabry B Guttmann J Eberhard L Wolff G Automatic compensation of endotracheal tube resistance in spontaneously breathing patients Technol Health Care 1994 1 281 291
Elsasser S Guttmann J Stocker R Mols G Haberthür C Accuracy of automatic tube compensation in new-generation mechanical ventilators Crit Care Med 2003 31 2619 2626 14605533 10.1097/01.CCM.0000094224.78718.2A
Haberthür C Fabry B Stocker R Ritz R Guttmann J Additional inspiratory work of breathing imposed by tracheostomy tubes and non-ideal ventilator properties in critically ill patients Intensive Care Med 1999 25 514 519 10401948 10.1007/s001340050890
Guttmann J Eberhard L Fabry B Bertschmann W Wolff G Continuous calculation of intratracheal pressure in tracheally intubated patients Anesthesiology 1993 79 503 513 8363076
Haberthür C Lichwarck-Aschoff M Guttmann J Guttmann J Continuous monitoring of tracheal pressure including spot-check of endotracheal tube resistance Technol Health Care 2003 11 413 424 14757920
Guttmann J Eberhard L Fabry B Zappe D Bernhard H Lichtwarck-Aschoff M Adolph M Wolff G Determination of volume-dependent respiratory system mechanics in mechanically ventilated patients using the new SLICE method Technol Health Care 1994 2 175 191
Eberhard L Guttmann J Wolff G Bertschmann W Minzer A Kohl HJ Zeravik J Adolph M Eckart J Intrinsic PEEP monitored in the ventilated ARDS patient with a mathematical method J Appl Physiol 1992 73 479 485 1399969
Brunner JX Wolff G Pulmonary Function Indices in Critical Care Patients 1988 Heidelberg: Springer
Nilsestuen JO Coon RL Woods M Kampine JP Location of lung receptors mediating the breathing frequency response to pulmonary CO2 Respir Physiol 1981 45 343 355 6800008 10.1016/0034-5687(81)90017-7
Iotti GA Closed-loop control mechanical ventilation Respir Care N Am 2001 7 341 522 10.1016/S1078-5337(05)70040-X
Cunningham DJC Robbins PA Wolff CB Integration of respiratory responses to changes in alveolar partial pressures of CO2 and O2 and in arterial pH Handbook of Physiology, Section 3: the Respiratory System, Control of Breathing, Part 2 1986 II Baltimore, MA: The Williams & Wilkins Company 475 528
BuSha BF Judd BG Manning HL Simon PM Searle BC Daubenspeck JA Leiter JC Identification of respiratory vagal feedback in awake normal subjects using pseudorandom unloading J Appl Physiol 2001 90 2330 2340 11356800
Grunstein MM Wyszogrodinski I Milic-Emili J Regulation of frequency and depth of breathing during expiratory threshold loading in cats J Appl Physiol 1975 38 869 874 1126897
Yan S Kayser B Tobiasz M Sliwinski P Comparison of static and dynamic intrinsic positive end-expiratory pressure using the Campbell diagram Am J Respir Crit Care Med 1996 154 938 944 8887589
Zakynthinos SG Vassilakopoulos T Zakynthinos E Mavrommatis A Roussos C Contribution of expiratory muscle pressure to dynamic intrinsic positive end-expiratory pressure: validation using the Campbell diagram Am Respir Crit Care Med 2000 162 1633 1644
Road JD Leevers AM Inspiratory and expiratory muscle function during continuous positive airway pressure in dogs J Appl Physiol 1990 68 1092 1100 2140347
Wakai Y Welsh MM Leevers AM Road JD Expiratory muscle activity in the awake and sleeping human during lung inflation and hypercapnia J Appl Physiol 1992 72 881 887 1533214
|
16137354
|
PMC1269457
|
CC BY
|
2021-01-04 16:04:55
|
no
|
Crit Care. 2005 Jun 9; 9(4):R407-R415
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3735
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37361613735810.1186/cc3736ResearchPneumothorax and mortality in the mechanically ventilated SARS patients: a prospective clinical study Kao Hsin-Kuo [email protected] Jia-Horng 2Sung Chun-Sung 3Huang Ying-Che 3Lien Te-Cheng [email protected] Attending physician, Department of Respiratory Therapy, Taipei Veterans General Hospital; Department of Medicine, Taoyuan Veterans Hospital; National Yang-Ming University School of Medicine, Taipei, Taiwan2 Attending physician and Chief of Department, Department of Respiratory Therapy, Taipei Veterans General Hospital; National Yang-Ming University School of Medicine, Taipei, Taiwan3 Attending physician, Department of Anesthesiology, Taipei Veterans General Hospital; National Yang-Ming University School of Medicine, Taipei, Taiwan4 Attending physician, Department of Respiratory Therapy, Taipei Veterans General Hospital; National Yang-Ming University School of Medicine, Taipei, Taiwan2005 22 6 2005 9 4 R440 R445 16 3 2005 22 4 2005 27 4 2005 12 5 2005 Copyright © 2005 Kao et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Pneumothorax often complicates the management of mechanically ventilated severe acute respiratory syndrome (SARS) patients in the isolation intensive care unit (ICU). We sought to determine whether pneumothoraces are induced by high ventilatory pressure or volume and if they are associated with mortality in mechanically ventilated SARS patients.
Methods
We conducted a prospective, clinical study. Forty-one mechanically ventilated SARS patients were included in our study. All SARS patients were sedated and received mechanical ventilation in the isolation ICU.
Results
The mechanically ventilated SARS patients were divided into two groups either with or without pneumothorax. Their demographic data, clinical characteristics, ventilatory variables such as positive end-expiratory pressure, peak inspiratory pressure, mean airway pressure, tidal volume, tidal volume per kilogram, respiratory rate and minute ventilation and the accumulated mortality rate at 30 days after mechanical ventilation were analyzed. There were no statistically significant differences in the pressures and volumes between the two groups, and the mortality was also similar between the groups. However, patients developing pneumothorax during mechanical ventilation frequently expressed higher respiratory rates on admission, and a lower PaO2/FiO2 ratio and higher PaCO2 level during hospitalization compared with those without pneumothorax.
Conclusion
In our study, the SARS patients who suffered pneumothorax presented as more tachypnic on admission, and more pronounced hypoxemic and hypercapnic during hospitalization. These variables signaled a deterioration in respiratory function and could be indicators of developing pneumothorax during mechanical ventilation in the SARS patients. Meanwhile, meticulous respiratory therapy and monitoring were mandatory in these patients.
==== Body
Introduction
Severe acute respiratory syndrome (SARS) is a transmissible pulmonary infection caused by a novel coronavirus [1,2]. About 20 to 30% of SARS patients may progress to severe hypoxemic respiratory failure that requires mechanical ventilation and intensive care unit (ICU) admission [3-6]. Pneumothorax, a major and potentially lethal complication of SARS and mechanical ventilation, often complicates the management of mechanically ventilated patients, and would be especially hazardous for patients in an individually isolated SARS ICU. Peiris et al. identified a high incidence of pneumomediastinum (12%) in a general population of SARS patients [3]. In addition, Lew and Fowler also observed a high incidence of pneumothorax (20 to 34%) in mechanically ventilated SARS patients [6,7]. However, no further investigations have assessed the risk factors of pneumothorax in the mechanically ventilated SARS patients.
Patients with acute respiratory distress syndrome (ARDS) and acute lung injury (ALI) [8] developing pneumothorax have been extensively studied. Previous studies have found that high inspiratory airway pressure and positive end-expiratory pressure (PEEP) were correlated with barotraumas [9-11]. Eisner et al. analyzed a cohort of 718 patients with ALI/ARDS and revealed that higher PEEP was related to an increased risk of barotraumas [12]. However, others were unable to identify any relationship between barotrauma and high ventilatory pressure or volume in patients with early ARDS [13-15]. Therefore, the relationship between airway pressure or volume and the development of barotraumas remains uncertain.
To our knowledge, there is no study on the risk factors of pneumothorax in mechanically ventilated SARS patients. To address this issue, we performed a prospective study to determine whether pneumothorax was produced by high ventilatory pressure or volume, and if it was associated with an increased mortality rate at 30 days after mechanical ventilation.
Materials and methods
This study included patients with SARS who were admitted to an isolation ICU at Taipei Veterans General Hospital. All patients satisfied the WHO case definition for SARS [16]. The research ethics board approved the study and we enrolled 41 patients with SARS who received mechanical ventilation between 14 May 2003 and 18 July 2003. Patients with pre-existing pneumothorax or chest tube thoracostomy were excluded. The primary study outcome variable was defined as radiographic evidence of new-onset pneumothorax at 30 days after ventilator use. Patients were censored at the first pneumothorax event, at the time of death, liberation from mechanical ventilation or discharge from the SARS ICU. Patients receiving mechanical ventilation were sedated with midazolam or propofol to facilitate mechanical ventilation; meanwhile, the sedatives were adjusted according to the Ramsay sedation score. Moreover, atracurium was used for neuromuscular paralysis to facilitate patient-ventilator synchrony in some patients. The dosage of atracurium was adjusted by peripheral nerve stimulator. When the patient was ready for weaning according to defined criteria, sedation and/or neuromuscular paralysis were discontinued.
Patient sex, age, actual body weight, APACHE II score and pre-existing comorbidities were recorded at entry. The PaO2/FiO2 ratio, PaO2, PaCO2, FiO2 and lung injury score [17] were recorded on ICU admission and daily during hospitalization. Ventilatory variables including PEEP, peak inspiratory pressure (PIP), mean airway pressure (MAP), tidal volume, tidal volume per kilogram, respiratory rate and minute ventilation were recorded at least once a day during the period of mechanical ventilation. When pneumothorax occurred, the highest pressure or volume of mechanical ventilation before the onset of pneumothorax were most likely to be the cause of pneumothorax [14]. Therefore, we compared the highest value of pressure and volume within a 24-hour period before the event in the patients with pneumothorax, with the overall values during mechanical ventilation in patients without pneumothorax.
Data were presented as mean ± standard deviation. The Mann-Whitney U test was used to compare data between patients with and without pneumothorax. We compared risk factors associated with the development of pneumothorax by Fisher's exact test for categorical variables. Non-parametric tests were chosen because of the small sample size in the pneumothorax group. Kaplan-Meier survival curves were compared by using the log-rank test. A p value of less than 0.05 was considered to indicate statistical significance. We used SPSS software (v10.0) for all analyses.
Results
Demographic and clinical characteristics are shown in Table 1. Of the 41 patients, the male-to-female ratio was 1:0.37 and mean age was 75.4 years. Five patients developed pneumothorax and the incidence of pneumothorax was 12%. The mean time to the development of pneumothorax was 8.0 ± 4.4 days after ventilator use. Of the patients, 28 (68%) met the criteria for either ALI or ARDS. Patients with pneumothorax were significantly associated with higher respiratory rate on admission, and more pronounced hypoxemia with lower PaO2/FiO2 ratio and higher PaCO2 during hospitalization.
Table 2 compares ventilator variables according to the presence or absence of pneumothorax. There were no significant differences in any pressure or volume between the patients with and without pneumothorax.
The overall survival rate was 59% at 30 days after mechanical ventilation. The relationship between pneumothorax and the probability of survival is shown in Fig. 1. There were no significant differences between the patients with and without pneumothorax.
Discussion
In the present study, we focused on the mechanically ventilated SARS patients and analyzed the risk factors of pneumothorax. Our study demonstrated that mechanically ventilated SARS patients with higher baseline respiratory rate, lower PaO2/FiO2 ratio, and higher PaCO2 during hospitalization were at a greater risk of developing pneumothorax. There were no significant differences in pressure, volume and mortality rate between the patients without and with pneumothorax. Barotrauma is a common complication in patients with SARS. The previous study by Peiris identified a high incidence of pneumomediastinum (12%) in a general population of SARS patients [3]. Choi et al. had also shown that subcutaneous emphysema, pneumothorax and pneumomediastinum were detected in six SARS patients (2.2%) who had not received positive-pressure ventilation [18].
In our study, the incidence of pneumothorax in mechanically ventilated SARS patients was lower than previous studies (12% versus 20 to 34%) [6,7]. The incidence of barotrauma in patients with ALI/ARDS varies widely. In most recent studies, it has ranged from 5 to 15% [12,14,19]. Gammon and colleagues have shown that the presence of ARDS is the major independent risk factor of barotraumas [13,20]. This may explain the lower incidence of pneumothorax in our study since the proportion of our patients with ALI/ARDS (68%) is lower than the other studies [6,7].
Another important finding in our study was the lack of correlation between ventilator variables and the presence of pneumothorax. Our results agreed with most of the previous studies that were done on ARDS patients. In the ARDS Network randomized controlled trial, low tidal volume ventilation decreased mortality without influencing the incidence of barotraumas [19]. In patients with sepsis-induced ARDS, there were no significant correlations between the ventilatory parameters and the development of pneumothorax or another air leak [14]. These authors suggested that barotrauma was more related to the underlying process than to the ventilator settings [14,15].
We found that the mechanically ventilated SARS patients with pneumothorax had a significant baseline tachypnea. Additionally, patients with a higher respiratory rate on admission also showed a trend of higher respiratory rate during hospitalization. (p = 0.06). Tachypnea on admission probably reflected the increased severity of the underlying disease [21], which may directly lead to a higher incidence of pneumothorax. There was also a higher risk of auto-PEEP in patients with tachypnea due to insufficient expiratory time, which may also contribute to the development of pneumothorax. However, auto-PEEP was not recorded in this study.
In our study, SARS patients with pneumothorax had a higher PaCO2 during hospitalization. Gattinoni et al. also observed a similar finding in ARDS patients with pneumothorax [11]. Increased dead space and cystic changes of lung parenchyma due to worsening underlying disease played a major role in patients with hypercapnia. This mechanism is further supported by a thin-section computed tomographic study that was done by Joynt and colleagues on the late stage of ARDS (more than 2 weeks after onset) caused by SARS [22]. They found that severe SARS-induced ARDS might independently result in cyst formation. In our study, patients with pneumothorax were also associated with a more pronounced hypoxemia, with lower PaO2/FiO2 during hospitalization compared with those without pneumothorax (65.8 versus 210.1). Oxygen-diffusing impairment and ventilation-perfusion maldistribution may play a role in developing hypoxemia in the mechanically ventilated SARS patient. A decrease in PaO2/FiO2 and increase in PaCO2 may be considered as a deterioration of respiratory condition in a patient with ALI/ARDS. The presence of pneumothorax together with hypoxemia/hypercapnia may indicate worsening of the underlying disease. This is supported by the large difference in APACHE II (26.0 ± 11.8 versus 20.7 ± 6.6) and ALI (2.51 ± 0.29 versus 1.59 ± 1.10) scores between patients with and without pneumothorax in this study, although these did not reach statistical significance.
In our study, the mortality rate was not significantly increased in patients with pneumothorax. In other studies on ALI/ARDS, the mortality directly attributable to barotrauma was low [12,14,23]. The mortality rate was 41% in our study, which was higher than the 26% from the results of five cohort studies [2-4,24,25]. Older age and more comorbidities may be the major causes. Age and coexisting illness, especially diabetes mellitus and heart disease, were consistently found to be independent prognostic factors for the risk of death and the need for intensive care in SARS patients [3-5,26,27].
There are several limitations to our study. Data were recorded once daily in individual isolation rooms and may have missed transient elevations in airway pressure/volume that could have led to alveolar disruption and pneumothorax. Secondly, we selected parameters that were easily measured and were previously shown or theorized to contribute to alveolar disruption, including ventilator variables and high-risk disease states. However, it is possible that an important variable such as plateau pressure was omitted from this analysis. Thirdly, there were only 41 mechanically ventilated SARS patients in our study. A study with a larger sample size may demonstrate statistical significance. The above factors are likely to cloud the relationship between the ventilatory variables and the occurrence of barotrauma.
Conclusion
The analysis of pneumothorax in mechanically ventilated SARS patients indicates that the patients with higher respiratory rates on admission, and lower PaO2/FiO2 ratio and higher PaCO2 during hospitalization had a greater risk of pneumothorax. The correlation between the clinical characteristics and pneumothorax may be considered as a deterioration of respiratory function in mechanically ventilated SARS patients developing pneumothorax. Pneumothorax in mechanically ventilated SARS patients may be an indicator of worsening underlying lung disease.
Key messages
• There were no significant differences in pressure, volume and mortality rate between the mechanically ventilated SARS patients without or with pneumothorax.
• Mechanically ventilated SARS patients with higher baseline respiratory rate, lower PaO2/FiO2 ratio, and higher PaCO2 during hospitalization were at a greater risk of developing pneumothorax.
• The correlation between the clinical characteristics and pneumothorax may be considered as a deterioration of respiratory function in mechanically ventilated SARS patients developing pneumothorax.
Abbreviations
ALI = acute lung injury; APACHE = Acute Physiology and Chronic Health Evaluation; ARDS = acute respiratory distress syndrome; FiO2 = fraction of inspired oxygen; MAP = mean airway pressure; ICU = intensive care unit; PEEP = positive end-expiratory pressure; PIP = peak inspiratory pressure, SARS = severe acute respiratory syndrome.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
T-CL participated in the design of the study and performed the statistical analysis. H-KK made contributions to the collection, analysis and interpretation of data. J-HW, C-SS and Y-CH made contributions to the design of the study and performed the statistical analysis.
Acknowledgements
The authors thank all health care workers of isolation SARS ICU in the Taipei Veterans General Hospital.
Figures and Tables
Figure 1 Kaplan-Meier curve of the probability of survival over time for mechanically ventilated SARS patients. (p = 0.11).
Table 1 Demographic and clinical characteristics according to the presence or absence of pneumothorax
Variable No pneumothorax Pneumothorax p value
Number of patients (%) 36 (88) 5 (12)
Gender (male/female) 26/10 4/1 1
Age, years 76.3 ± 10.4 68.8 ± 18.0 0.46
Body weight, kg 58.5 ± 12.4 57.0 ± 18.2 0.98
APACHE II score 20.7 ± 6.6 26.0 ± 11.8 0.41
Pre-existing comorbidities
Chronic renal insufficiency 4 0 1
Congestive heart failure 9 2 0.59
Diabetes mellitus 15 2 1
Chronic obstructive pulmonary disease 5 0 1
Pulmonary tuberculosis 2 2 0.06
Cerebrovascular disease 17 1 0.37
On ICU admission
Baseline lung injury score 1.27 ± 1.04 1.59 ± 0.59 0.35
Baseline respiratory rate 25.32 ± 7.53 36.00 ± 5.10 0.006
Baseline PaO2/FiO2 ratio 289.9 ± 172.9 272.6 ± 140.8 0.87
Baseline PaCO2 35.7 ± 9.3 49.4 ± 23.0 0.20
During hospitalization
Highest lung injury score 1.59 ± 1.10 2.51 ± 0.29 0.09
Highest respiratory rate 34.65 ± 5.19 40.80 ± 7.08 0.06
Lowest PaO2/FiO2 ratio 210.1 ± 123.8 65.8 ± 24.3 0.02
Highest PaCO2 49.9 ± 17.4 80.1 ± 12.3 0.004
ALI/ARDS (%) 24 (66%) 4 (80%) 1
Liberation from ventilator (%) at 30 days 11(31) 0 0.29
Data are presented as mean ± standard deviation. ALI, acute lung injury; APACHE, Acute Physiology and Chronic Health Evaluation; ARDS, acute respiratory distress syndrome; FiO2, fraction of inspired oxygen; ICU, intensive care unit; PEEP, positive end-expiratory pressure.
Table 2 The ventilator variables according to the presence or absence of pneumothorax
Variables No pneumothorax Pneumothorax p
Ventilatory pressure, cmH2O, or volume
positive end-expiratory pressure 7.94 ± 4.38 8.2 ± 2.0 0.54
peak inspiratory pressure 34.78 ± 6.80 33.8 ± 3.76 0.73
mean airway pressure 18.75 ± 4.89 20.8 ± 1.78 0.17
tidal volume, ml 761.02 ± 128.87 733.8 ± 154.0 0.43
tidal volume/kg, ml 12.32 ± 2.71 12.54 ± 3.34 0.97
Minute ventilation, l (on ICU admission) 10.40 ± 3.00 11.38 ± 2.84 0.34
Minute ventilation, l (during hospitalization) 15.33 ± 4.68 12.93 ± 4.10 0.26
Data are presented as mean ± standard deviation.
==== Refs
Ksiazek TG Erdman D Goldsmith CS Zaki SR Peret T Emery S Tong S Urbani C Comer JA Lim W A novel coronavirus associated with severe acute respiratory syndrome N Engl J Med 2003 348 1953 1966 12690092 10.1056/NEJMoa030781
Rota PA Oberste MS Monroe SS Nix WA Campagnoli R Icenogle JP Penaranda S Bankamp B Maher K Chen MH Characterization of a novel coronavirus associated with severe acute respiratory syndrome Science 2003 300 1394 1399 12730500 10.1126/science.1085952
Peiris JS Chu CM Cheng VC Chan KS Hung IFN Poon LLM Law KI Tang BSF Hon TYW Chan CS Clinical progression and viral load in a community outbreak of coronavirus-associated SARS pneumonia: a prospective study Lancet 2003 361 1767 1772 12781535 10.1016/S0140-6736(03)13412-5
Lee N Hui D Wu A Chan P Cameron P Joynt G Ahuja A Yung MY Leung CB To KF A major outbreak of severe acute respiratory syndrome in Hong Kong N Engl J Med 2003 348 1986 1994 12682352 10.1056/NEJMoa030685
Booth CM Matukas LM Tomlinson GA Rachlis AR Rose DB Dwosh HA Walmsley SL Mazzulli T Avendano M Derkach P Clinical features and short-term outcomes of 144 patients with SARS in the greater Toronto area JAMA 2003 289 2801 2809 12734147 10.1001/jama.289.21.JOC30885
Lew TWK Kwek TK Tai D Earnest A Loo S Singh K Kwan KM Chan Y Yim CF Bek SL Acute respiratory distress syndrome in critically ill patients with severe acute respiratory syndrome JAMA 2003 290 374 380 12865379 10.1001/jama.290.3.374
Fowler RA Lapinsky SE Hallett D Detsky AS Sibbald WJ Slutsky AS Stewart TE the Toronto SARS Critical Care Group Critically ill patients with severe acute respiratory syndrome JAMA 2003 290 367 373 12865378 10.1001/jama.290.3.367
Bernard GR Artigas A Brigham KL Carlet J Falke K Hudson L Lamy M Legall J Morris A Spragg R The American-European consensus conference on ARDS: definitions, mechanisms, relevant outcomes and clinical trial coordination Am J Respir Crit Care Med 1994 149 818 824 7509706
Petersen GW Baier H Incidence of pulmonary barotraumas in a medical ICU Crit Care Med 1983 11 67 69 6337021
Schnapp LM Chin DP Szaflarski N Matthay MA Frequency and importance of barotraumas in 100 patients with acute lung injury Crit Care Med 1995 23 272 278 7867352 10.1097/00003246-199502000-00012
Gattinoni L Bombino M Pelosi P Lissoni A Pesenti A Fumagalli R Tagliabue M Lung structure and function in different stages of severe adult respiratory distress syndrome JAMA 1994 271 1772 1779 8196122 10.1001/jama.271.22.1772
Eisner MD Thompson BT Schoenfeld D Anzueto A Matthay MA the Acute Respiratory Distress Syndrome Network Airway pressures and early barotraumas in patients with acute lung injury and acute respiratory distress syndrome Am J Respir Crit Care Med 2002 165 978 982 11934725
Gammon RB Shin MS Groves RH JnrHardin JM Hsu C Buchalter SE Clinical risk factors for pulmonary barotraumas: a multivariate analysis Am J Respir Crit Care Med 1995 152 1235 1240 7551376
Weg JG Anzueto A Balk RA Wiedemann HP Pattishall EN Schork MA Wagner LA The relation of pneumothorax and other air leaks to mortality in the acute respiratory distress syndrome N Engl J Med 1998 338 341 346 9449726 10.1056/NEJM199802053380601
Boussarsar M Thierry G Jaber S Roudot-Thoraval F Lemaire F Brochard L Relationship between ventilatory settings and barotraumas in the acute respiratory distress syndrome Intensive Care Med 2002 28 406 413 11967593 10.1007/s00134-001-1178-1
WHO Case definition for surveillance of severe acute respiratory syndrome SARS 2002
Murray JF Matthay MA Luce JM Flick MR An expanded definition of the adult respiratory distress syndrome Am Rev Respir Dis 1988 138 720 723 3202424
Choi KW Chau TN Tsang O Tso E Chiu MC Tong WL Lee PO Ng TK Ng WF Lee KC the Princess Margaret Hospital SARS Study Group Outcomes and prognostic factors in 267 patients with severe acute respiratory syndrome in Hong Kong Ann Intern Med 2003 139 715 723 14597455
Ventilation with lower tidal volumes as compared with traditional tidal volumes for acute lung injury and the acute respiratory distress syndrome. The Acute Respiratory Distress Syndrome Network. N Engl J Med 2000 342 1301 1308 10793162 10.1056/NEJM200005043421801
Gammon RB Shin MS Buchalter SE Pulmonary barotraumas in mechanical ventilation. Patterns and risk factors Chest 1992 102 568 572 1643949
Knaus WA Draper EA Wagner DP Zimmerman JE APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Joynt GM Antonio GE Lam P Wong KT Li T Gomersall CD Ahuja AT Late stage adult respiratory distress syndrome caused by severe acute respiratory syndrome: abnormal findings at thin-section CT Radiology 2004 230 339 346 14752179
DiRusso SM Nelson LD Safcsak K Miller RS Survival in patients with severe adult respiratory distress syndrome treated with high-level positive end-expiratory pressure Crit Care Med 1995 23 1485 1496 7664550 10.1097/00003246-199509000-00008
Tsang KW Ho PL Ooi GC Yee WK Wang T Chan-Yeung M Lam WK Seto WH Yam LY Cheung TM A cluster of cases of severe acute respiratory syndrome in Hong Kong N Engl J Med 2003 348 1977 1985 12671062 10.1056/NEJMoa030666
Peiris JSM Lai ST Poon LLM Guan Y Yam LYC Lim W Nicholls J Yee WKS Yan WW Cheung MT Coronavirus as a cause of severe acute respiratory syndrome Lancet 2003 361 1319 1325 12711465 10.1016/S0140-6736(03)13077-2
Tsui PT Kwok ML Yuen H Lai ST Severe acute respiratory syndrome: clinical outcome and prognostic correlates Emerg Infect Dis 2003 9 1064 1069 14519241
Chan JW Ng CK Chan YH Mok TY Lee S Chu SYY Law WL Lee MP Li PCK Short term outcome and risk factors for adverse clinical outcomes in adult with severe acute respiratory syndrome (SARS) Thorax 2003 58 686 689 12885985 10.1136/thorax.58.8.686
|
16137358
|
PMC1269458
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 22; 9(4):R440-R445
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3736
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37371613735710.1186/cc3737ResearchHigh frequency oscillatory ventilation compared with conventional mechanical ventilation in adult respiratory distress syndrome: a randomized controlled trial [ISRCTN24242669] Bollen Casper W [email protected] Well Gijs Th J [email protected] Tony 3Beale Richard J [email protected] Sanjoy 5Findlay George [email protected] Mehran [email protected] Jean-Daniel [email protected] Norbert [email protected] Cuno SPM [email protected] Vught Adrianus J [email protected] Fellow, Intensive Care, University Medical Centre Utrecht, The Netherlands2 Paediatrician, University Medical Centre Utrecht, The Netherlands3 Intensivist, St Thomas Hospital, London, UK4 Head, Intensive Care, St Thomas Hospital, London, UK5 Intensivist, University Hospital of Wales, Cardiff, UK6 Intensivist, Hopital Cochin, Paris, France7 Intensivist, University Hospital Mainz, Germany8 Clinical Epidemiologist, University Medical Centre Utrecht, The Netherlands9 Head, Intensive Care University Medical Centre Utrecht, The Netherlands2005 21 6 2005 9 4 R430 R439 19 12 2004 17 1 2005 22 4 2005 12 5 2005 Copyright © 2005 Bollen et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
To compare the safety and efficacy of high frequency oscillatory ventilation (HFOV) with conventional mechanical ventilation (CV) for early intervention in adult respiratory distress syndrome (ARDS), a multi-centre randomized trial in four intensive care units was conducted.
Methods
Patients with ARDS were randomized to receive either HFOV or CV. In both treatment arms a priority was given to maintain lung volume while minimizing peak pressures. CV ventilation strategy was aimed at reducing tidal volumes. In the HFOV group, an open lung strategy was used. Respiratory and circulatory parameters were recorded and clinical outcome was determined at 30 days of follow up.
Results
The study was prematurely stopped. Thirty-seven patients received HFOV and 24 patients CV (average APACHE II score 21 and 20, oxygenation index 25 and 18 and duration of mechanical ventilation prior to randomization 2.1 and 1.5 days, respectively). There were no statistically significant differences in survival without supplemental oxygen or on ventilator, mortality, therapy failure, or crossover. Adjustment by a priori defined baseline characteristics showed an odds ratio of 0.80 (95% CI 0.22–2.97) for survival without oxygen or on ventilator, and an odds ratio for mortality of 1.15 (95% CI 0.43–3.10) for HFOV compared with CV. The response of the oxygenation index (OI) to treatment did not differentiate between survival and death. In the HFOV group the OI response was significantly higher than in the CV group between the first and the second day. A post hoc analysis suggested that there was a relatively better treatment effect of HFOV compared with CV in patients with a higher baseline OI.
Conclusion
No significant differences were observed, but this trial only had power to detect major differences in survival without oxygen or on ventilator. In patients with ARDS and higher baseline OI, however, there might be a treatment benefit of HFOV over CV. More research is needed to establish the efficacy of HFOV in the treatment of ARDS. We suggest that future studies are designed to allow for informative analysis in patients with higher OI.
See related commentary
==== Body
Introduction
Mechanical ventilation of patients with adult respiratory distress syndrome (ARDS) may cause lung injury and, subsequently, multi-organ failure [1]. Multi-organ failure is a major cause of death in ARDS [2]. In particular, repetitive opening and closure of alveoli with significant shear forces exerted to the alveolar walls and over-distension of alveoli and small airways are thought to be main factors leading to ventilator induced lung injury. Lung protective ventilation strategies with low tidal volumes and high end-expiratory pressures are used to prevent ventilator induced lung injury [3]. In high frequency oscillatory ventilation (HFOV), extremely small tidal volumes are combined with a high mean airway pressure to prevent atelectasis and at the same time limit peak inspiratory pressures. HFOV is suggested, by some, to be the theoretically most optimal form of lung protective ventilation [4]. The role of HFOV in ARDS, however, has to be established yet.
Most studies comparing HFOV with conventional mechanical ventilation (CV) have been performed in premature neonatal patients [5]. The routine use of HFOV as an elective treatment in premature neonates with respiratory distress is equivocal. In a recent paper we have argued that improvements in CV strategies have diminished the relative benefit of HFOV [6]. There is much less evidence in adult and paediatric patients. Three non-randomized prospective trials and no more than two randomized controlled trials in patients with ARDS have been published to establish the safety and efficacy of HFOV [7-11]. In these trials, the oxygenation index (OI), a cost benefit ratio of inspired oxygen times airway pressure divided by arterial oxygen pressure (OI = FiO2 × MAP × 100)/paO2), was an important predictor of mortality.
We performed a randomized controlled trial designed to test the safety and efficacy of HFOV as a primary mode of ventilation in ARDS patients compared with CV. This study was prematurely terminated because of a low inclusion rate and the completion of a similar trial [7]. We compared survival without supplemental oxygen or on ventilator, mortality, therapy failure and crossover.
Materials and methods
Between October 1997 and March 2001 61 patients were enrolled in a randomized controlled trial comparing HFOV with CV in patients with ARDS to detect differences in mortality, therapy failure and ventilatory support at 30 days. This study was conducted in intensive care units in London, Cardiff, Paris and Mainz. Patients with ARDS and a bodyweight greater than 35 kg were randomized to receive either HFOV or CV. ARDS was defined as the pressure of arterial oxygen divided by the fraction of inspired oxygen (paO2/FiO2) < 200 mmHg, radiographic evidence of bilateral infiltrates on chest X-ray and no evidence of atrial hypertension. Patients with a non-pulmonary terminal disease, severe chronic obstructive pulmonary disease or asthma and grade 3 or 4 air-leak were excluded. Patients with FiO2 > 0.80 for 48 h or more than 10 days of mechanical ventilation before meeting the entry criteria were excluded as well. Randomization was by a sequentially numbered computerized randomization algorithm. The allocation to treatment was concealed until study entry. This study was approved by the ethical committee board of all participating institutions and was in compliance with the Helsinki Declaration. Informed consent was obtained from next of kin of patients prior to study entry.
The general physiological targets for the two ventilator arms were similar. The oxygenation goal was to maintain an O2 saturation ≥ 88% or paO2 > 60 mmHg with a FiO2 < 0.6. The ventilatory goal was to establish an arterial pH > 7.20 and a HCO3 > 19 mmol/l while minimizing peak inspiratory pressures irrespectively of arterial carbon dioxide pressure (paCO2). The priority in both treatment arms was to maintain lung volume by first weaning FiO2 to < 0.60 after which mean airway pressure and FiO2 were given equal priority for reduction. Patients were crossed over to the alternative ventilator in case of therapy failure: intractable hypotension despite maximum support (RR mean < 60 mmHg for > 4 h or < 50 mmHg for > 1 h); intractable respiratory acidosis (pH 7.20 at HCO3 > 19 mmol/l for > 6 h); oxygenation failure (rising OI of more than two times since study entry or OI > 42 after 48 h; OI = (FiO2 × MAP × 100)/paO2)); and grade 4 air leak (air leak with multiple recurrences (> 4); air leak requiring more than two chest tubes per hemithorax; air leak continuing longer than 120 h; or pneumopericardium or pneumoperitoneum). Patients could be withdrawn from the study treatment for the following reasons: withdrawal of consent; weaned from mechanical ventilation; death or treatment failure after crossover.
In the CV treated group, patients were treated with time cycled pressure controlled ventilation. Respiratory rate to achieve low tidal volumes was free up to 60/minute. Maximum peak inspiratory pressure was limited to 40 cmH2O. To minimize the inspiratory pressures, an arterial pH > 7.20 was acceptable irrespectively of the level of paCO2. Positive end-expiratory pressure was advocated up to 15 cmH2O. An inspiratory:expiratory ratio up to 2:1 could be used to achieve adequate oxygenation. Otherwise, the patient was crossed over to HFOV as indicated above. More detailed ventilation procedures and methods of weaning were according to standard protocols of the investigating centres.
Patients in the HFOV group were ventilated with the SensorMedics 3100B ventilator (SensorMedics, Bilthoven, the Netherlands). A high lung volume strategy was used as has been previously described [12]. HFOV was started with continuous distending pressure (CDP) at 5 cm H2O higher than mean airway pressure (MAP) on CV and then adjusted to achieve and maintain optimal lung volume. Therefore, initially, CDP was increased until an O2 saturation > 95% was achieved. CDP was not decreased until FiO2 < 0.60 was feasible applying the general physiological targets mentioned earlier. Pulmonary inflation was checked by chest X-rays if increasing CDP did not result in O2 saturation > 88%. Frequency was initially set at 5 Hz with an inspiratory time of 33%. Delta P was adjusted according to paCO2 and chest wall vibrations. If ventilation did not improve despite a maximum Delta P, the frequency could be lowered. Weaning was instigated if paO2 > 60 mmHg at FiO2 < 0.40 and suction was well tolerated by decreasing Delta P and CDP to continuous positive airway pressure level. Ventilator weaning was continued on CV according to the standard protocol of the unit.
Measurements
Assessment of the principal outcomes and repeated measurements was not blinded. The principal outcomes consisted of: cumulative survival without mechanical ventilation or oxygen dependency at 30 days; mortality at 30 days; therapy failure; crossover rate; and persisting pulmonary problems defined as oxygen dependency or still being on a ventilator at 30 days. Data collection began one hour following randomization for the conventionally treated patients and at the initiation of HFOV for the HFOV treated patients. The time period on CV prior to the study, ET tube length and diameter, air leak score, Acute Physiologic and Chronic Health Evaluation (APACHE) II score at admission, arterial blood gases, ventilator settings and cardiovascular measurements were recorded. Arterial blood gases, ventilator settings, heart rate, blood pressure and cardiac output, if available, were registered after study entry or crossover and every eight hours for four days on the assigned ventilator. Ventilator settings and blood gases were recorded for every change of ventilator settings during the first three days of treatment.
Statistical analysis
In analyses of primary outcomes, the intention to treat principle was used. Based on a projected survival without mechanical ventilation or oxygen dependency in the control group of 25%, an increase to 51% in the HFOV group would be detectable with 106 patients (alpha of 0.05, power of 0.80) [9]. Univariate logistic regression analysis was used to calculate differences in 30 day survival without mechanical ventilation or oxygen dependency, mortality, crossover, therapy failure and incidence of supplemental oxygen dependency or mechanical ventilation at 30 days. Cox proportional hazard analysis was conducted to detect differences in mortality. The proportionality assumption was graphically tested using log minus log plots. Multivariate logistic regression and Cox proportional hazard analysis for mortality were used to adjust in case of post-randomization differences in a priori defined pre-treatment conditions (dummy variables for study site, OI, ventilatory index (ventilatory index = (peak inspiratory pressure (mmHg) × respiratory rate × paCO2 (mmHg))/1000), APACHE II score, age and weight). Furthermore, we looked at the relation between the OI response and mortality. Average values and standard errors of respiratory and circulatory parameters were calculated for days 1, 2, 3, and 4 of the study. Significant differences between treatment groups were tested by a general linear mixed model analysis. P-values were calculated 2-sided. All analyses were conducted using SPSS 12.0.1 for Windows software (SPSS Inc., Chicago, Illinois, U.S.).
Results
The study was stopped prematurely after inclusion of 61 patients because of a low inclusion rate and the completion of another trial comparing HFOV with CV in patients with ARDS [7]. Of the 61 patients, 37 were randomized to receive HFOV and 24 to receive CV. Follow up time to 30 days was incomplete in seven patients (five HFOV and two CV).
The baseline OI at study entry was higher in the HFOV group than in the CV group, (25 versus 18; Table 1). Patients were comparable for age and APACHE II score. The youngest patient was 17 years and the oldest patient was 77 years. The female:male ratio was lower in the HFOV group than in the CV group (0.24 versus 0.42). The majority of patients (80%) were diagnosed with sepsis or pneumonia. Prior to randomization, patients were ventilated with an average tidal volume of 9.3 ml/kg ideal bodyweight in the HFOV group and 8.4 ml/kg ideal bodyweight in the CV group. (Ideal body weight was calculated as: males, weight = 50 + 0.91 × (height in centimetres – 152.4); females, weight = 45 + 0.91 × (height in centimetres – 152.4)). Peak inspiratory pressures were comparable for both treatment groups. In one case, the limitation of 40 mmHg for peak inspiratory pressures was violated in the CV group. There were no major differences between treatment groups in mean airway pressures or peak end-expiratory pressures. Blood gas results prior to randomization showed a lower arterial oxygen saturation and paO2 in the HFOV group compared with the CV group.
The primary outcomes are presented in Table 2. There was no difference in cumulative survival without oxygen dependency or still on mechanical ventilation at 30 days between HFOV and CV. Mortality at 30 days did not differ significantly between HFOV and CV. An important cause of death was withdrawal of treatment (10 cases in 24 deaths). None of the deaths were directly related to the assigned therapy. Figure 1 shows a nearly identical cumulative survival of the HFOV group and the CV group corrected for the baseline covariates; study site, OI, ventilatory index, APACHE II score, age and weight. The survival curves of the duration of ventilation were virtually identical for the HFOV group and the CV group (data not shown). The median duration of ventilation was 20 days (± 6 SD) for HFOV and 18 days (± 5 SD) in the CV treatment group.
Treatment failure occurred in 10 patients (27%) in the HFOV group compared with five patients (21%) in the CV group. Seven patients (19%) treated with HFOV crossed over to CV; in the CV group four patients (17%) were switched to HFOV. Of the four patients that crossed over in the CV group, two patients died and one patient was on supplemental oxygen therapy at 30 days. In the HFOV group, five patients that crossed over died and two patients were still on ventilator or needed extra oxygen. The occurrence of being on oxygen or mechanical ventilation at 30 days in survivors was equal between HFOV and CV.
Ventilatory settings and blood gas results at days 1, 2, 3 and 4 of the study are shown in Table 3. Patients with HFOV were ventilated with higher mean airway pressures than patients on CV (p = 0.03). FiO2 was also higher in the HFOV group compared with the CV group. This difference between the treatment groups was not significant (p = 0.33). Results of blood gases were comparable between the two treatment groups including all patients. Patients that crossed over in the CMV group had significantly lower pH than patients who did not cross over in the CMV group (p = 0.02). This difference, however, was not found between patients who did and did not cross over in the HFOV group (p = 0.56). The OI, on the other hand, was higher in both patients that crossed over in the CMV group and patients that crossed over in the HFOV group compared with patients that did not cross over (p = 0.07 and p = 0.05, respectively).
Systolic arterial blood pressure and mean arterial blood pressure were higher in the HFOV treated patients compared with CV treated patients (p = 0.06 versus p = 0.07). Cardiac output was comparable between the two treatment groups (data not shown).
The OI response in all patients treated with either HFOV or CV did not differ significantly between survivors and non-survivors (Figure 2). The OI response from day 1 to day 2 was significantly larger in HFOV than in CV treated patients (p < 0.01). Within treatment groups there was a significant difference in initial OI between survivors and non-survivors in CV treated patients, but OI response to treatment did not differentiate between survivors and non-survivors in CV treated patients. In the HFOV treated patients there was no difference in the baseline OI, nor was there a difference in OI response between survivors and non-survivors.
The results of a post hoc analysis are shown in Figure 3. Adjusted odds ratios for mortality were calculated for samples of the study population including patients with progressively higher baseline OI prior to randomization. This suggested that, in patients with a higher baseline OI, the effect of treatment with HFOV was relatively better compared with CV. OI was evaluated as an interaction term in a Cox Proportional Hazard model with treatment, age and OI as explanatory variables. The likelihood ratio test comparing the reduced (no-interaction) with the full (interaction) model showed a p-value of 0.048.
Discussion
No significant differences between HFOV and CV were observed, but this trial only had power to detect major differences in mortality or survival without oxygen dependency or on ventilator. Furthermore, 11 of 61 patients were crossed over to a different treatment arm; this also diminished the power to detect potential treatment differences. A post hoc analysis, however, suggested that in patients with a higher baseline OI, HFOV may be more effective than CV.
This trial was stopped because of a low inclusion rate and the completion of another similar trial [7]. The low inclusion rate was not because of competing trials but probably due to the limited number of investigators (four centres compared with nine centres in the study by Derdak et al.). The number of patients included in the two treatment arms differed considerably. This misbalance was due to stopping the trial early. There were no protocol violations. Furthermore, baseline OI at study entry was higher in the HFOV group than in the CV group. The OI has been recognized as an important prognostic determinant of mortality [13].
HFOV was started early in the course of ARDS. Patients were ventilated on HFOV according to the open lung concept. This resulted in significantly higher mean airway pressures compared with CV ventilated patients. This mainly determined the higher OI in the HFOV group during the first days. FiO2 and paO2 values were similar between HFOV and CV patients. Potential theoretical risks of HFOV therapy, overdistension of the pulmonary system leading to barotrauma or cardiovascular compromise, packing of mucus leading to ineffective ventilation or blocking of the endotracheal tube were not encountered. None of the HFOV ventilated patients developed necrotizing tracheobronchitis.
Patients in the CV group were ventilated following a lung protective strategy targeted to minimizing tidal volumes. The tidal volumes per kg ideal bodyweight that were used in this study were higher than tidal volumes used in studies of lung protective ventilation strategies [14]. On the other hand, tidal volumes in our study were significantly lower than tidal volumes that were found to be harmful in those studies. Peak inspiratory pressures were limited to 40 cmH2O in the CV group. This restriction was violated in only one case. Nine patients were ventilated with pressures above 35 cmH2O. Furthermore, the overall mortality and survival without mechanical ventilation or oxygen dependency at 30 days did not suggest that the ventilation treatment in the CV group was suboptimal.
The OI represents the pressure and oxygen cost for oxygenation. It has been regarded as a marker of lung injury and prognostic indicator of treatment success [15]. In CV treated patients there was a significant difference in baseline OI between survivors and non-survivors. Baseline OI did not, however, differentiate between survivors and non-survivors in HFOV treated patients. Although in some studies OI response to treatment was a predictor of outcome [7,9], we could not reproduce this relation. A possible explanation could be that fewer numbers of patients were included in our analysis. Also, we used a different time window; we compared OI on a daily basis whereas in a study by Derdak et al. [7] OI was compared every 4 h. In that study, OI response was maximally different at 16 h [7]. In our study, OI response only differed significantly between HFOV and CV treated patients. This difference for the most part could be explained by the higher mean airway pressures used in the HFOV group.
A post hoc analysis suggested that baseline OI could be an important effect modifier of the relative treatment effect of HFOV compared with CV. We hypothesize that within the pressure-ventilation curve there is a safe window between under-inflation with atelectasis and shear stress and over-inflation with barotrauma [4,16]. In patients with ARDS with higher OI, this safe window possibly becomes too small for CV to prevent ventilator induced lung injury. This concept is supported by animal experiments where addition of positive end-expiratory pressure (PEEP) resulted in additional over-inflation contributing to ventilator-induced lung injury [17]. The combination of high levels of PEEP and over-distension are directly reflected in the OI. HFOV seemed to offer an advantage over CV only in patients with a higher initial OI. This is in accordance with observational studies that showed that better survival rates in more severe ARDS with higher OI was associated with HFOV treatment [11,18]. In fact, HFOV has been recommended in patients who require high mean airway pressure and FiO2 exceeding 60% corresponding to an OI > 20 when paO2 = 60 mmHg [12]. Because these findings result from a post hoc analysis, however, they can only be regarded as hypothesis generating still to be confirmed.
Previous trials did not show a significant difference in mortality in patients with ARDS between HFOV and CV [19]. In our trial, mortality in the HFOV group was similar to mortality reported in the previous trials, but mortality in the CV group was considerably less, in accordance with the imbalance in prognostic indicators at baseline.
More evidence is needed to confirm a beneficial effect of HFOV over CV in the treatment of ARDS. Our results and those from previous trials seem promising but could depend on other criteria to select patients with ARDS that benefit from HFOV compared with CV. One of these criteria could be OI. Therefore, we believe that in future research comparing HFOV with CV as early treatment of ARDS, it is important to focus on patients with higher levels of baseline OI. As treatment differences will be smaller than our prior estimate was, larger trials are needed. We do not think that OI response can be used as an alternative outcome measurement for treatment success or failure.
Conclusion
In this study, we were not able to find significant differences in efficacy or safety between HFOV and CV as early treatment of ARDS. A post hoc analysis suggested that HFOV could prevent mortality compared with CV in patients with a higher baseline OI. Therefore, it is important in future studies to enable informative analysis of patients with higher baseline OI. To achieve sufficient power to detect possible important treatment differences in subgroups of patients with higher OI, larger multi-centre trials are warranted.
Key messages
• This study was not powered to show significant differences in efficacy or safety between HFOV and CV as early treatment of ARDS.
• However, a post hoc analysis suggested a better treatment effect of HFOV compared with CV in patients with higher baseline OI.
• Future studies should be designed to allow for informative analysis in patients with higher OI.
Abbreviations
ARDS = adult respiratory distress syndrome; CDP = continuous distending pressure; CI = confidence interval; CV = conventional mechanical ventilation; FiO2 = fraction of inspired oxygen; HFOV = high frequency oscillatory ventilation; MAP = mean airway pressure; OI = oxygenation index; OR = odds ratio; paCO2 = pressure of arterial carbon dioxide; paO2 = pressure of arterial oxygen; PEEP = positive end-expiratory pressure; SaO2 = arterial oxygen saturation.
Competing interests
Supported in part by SensorMedics Corporation, which also provided use of the 3100B high-frequency oscillatory ventilators. None of the study investigators have a financial interest in SensorMedics Corporation. The authors declare that they have no competing interests.
Authors' contributions
AJvV initiated the study, participated in its design and coordination and helped to draft the manuscript. CWB, CSPMU and GTJvW performed the statistical analyses and wrote the manuscript. TS, RJB, SS, GF, MM, JC and NW participated in its design and conducted the study. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 Cumulative mortality incidence for high frequency oscillatory ventilation (HFOV) versus conventional mechanical ventilation (CV). Curves are estimates of cumulative risk corrected for study site, baseline oxygenation index and ventilatory index, APACHE II score, age and weight.
Figure 2 Oxygenation index (OI) in survivors versus non-survivors and high frequency oscillatory ventilation (HFOV) versus conventional mechanical ventilation (CV). OIs are represented by diamonds as means with bars as 95% confidence intervals (CI). Reported p-values for baseline OI are corrected for study site, ventilatory index, APACHE II score, age and weight. The baseline OI did not significantly predict mortality in all patients or in HFOV (p = 0.06 and p = 0.41, respectively). §Baseline OI was significantly different between survivors and non-survivors in the CV group (p = 0.04). Significant differences between OI responses were calculated by linear mixed model analyses. #Significant difference in OI response between HFOV and CV (p = < 0.01). OI response did not differentiate between survivors and non-survivors in all patients or in CV and HFOV separately (p = 0.28, p = 0.12 and p = 0.95, respectively).
Figure 3 Post hoc analysis of the treatment effect on mortality relative to baseline oxygenation index (OI). On the y-axis the odds ratio of mortality (OR) adjusted for study site, OI, ventilatory index, APACHE II score, age and weight is presented by diamonds and 95% confidence intervals by bars. On the x-axis the different analyses are depicted including patients with increasing levels of initial OI at study entry. N denotes the number of patients in each subgroup. CI, confidence interval; CMV, conventional mechanical ventilation; HFOV, high frequency oscillatory ventilation.
Table 1 Patient characteristics at study entry
HFOV CV
N 37 24
Female:male ratio 9/28 (24%) 10/14 (42%)
Mean age (years) 81.0 ± 20.5 81.7 ± 12.5
Weight 50.7 ± 17.4 55.4 ± 12.8
APACHE II score 21.1 ± 7.6 20.1 ± 9.3
Diagnosis (%)
Trauma 1 (3) 2 (9)
Sepsis 25 (68) 13 (57)
Pneumonia 8 (22) 3 (13)
Other 3 (8) 5 (22)
Site (%)
United Kingdom 24 (65) 15 (63)
France 5 (21) 7 (19)
Germany 4 (17) 6 (16.2)
Ventilation time prior to study (days) 2.1 ± 2.6 1.5 ± 1.8
Oxygenation index 25.2 ± 13.0 18.0 ± 7.4
Ventilatory index 33.8 ± 20.4 30.3 ± 12.5
Respiratory rate (per min) 18.1 ± 4.1 17.8 ± 4.6
Tidal volume(ml) 618.4 ± 142.6 549.7 ± 130
Tidal volume per ideal bodyweight (ml/kg) 9.3 ± 2.2 8.4 ± 2.0
Peak inspiratory pressure (cmH2O) 33.1 ± 6.8 32.3 ± 5.4
Positive end-expiratory pressure (cmH2O) 13.9 ± 3.8 12.9 ± 3.2
Mean airway pressure (cmH2O) 21.5 ± 5.4 21.0 ± 5.1
FiO2 0.84 ± 0.19 0.76 ± 0.19
pH 7.3 ± 0.13 7.3 ± 0.11
paCO2 (mmHg) 53.5 ± 17.3 52.2 ± 11.9
paO2 (mmHg) 80.8 ± 24.1 93.3 ± 24.5
SaO2 (percentage) 90.8 ± 6.4 94.3 ± 3.1
Heart rate 109.8 ± 23.7 111.2 ± 29.5
Mean arterial pressure (cmH2O) 75.3 ± 13.1 72.2 ± 14.1
Central venous pressure (cmH2O) 13.5 ± 4.2 13.8 ± 4.9
Values are presented as means with standard deviations. APACHE II, Acute Physiologic and Chronic Health Evaluation II; CV, conventional mechanical ventilation; FiO2, fraction of inspired oxygen; HFOV, high frequency oscillatory ventilation; OI, oxygenation index; paO2, pressure of arterial oxygen, paCO2, pressure of arterial carbon dioxide; SaO2, arterial oxygen saturation.
Table 2 Primary outcomes
Unadjusted Adjusted
HFOV CV p-value OR 95% CI OR 95% CI
N 37 24
Survival without supplemental oxygen or on ventilator 12 (32%) 9 (38%) 0.79 0.80 0.27–2.53 0.80 0.22–2.97
Mortality 16 (43%) 8 (33%) 0.59 1.52 0.45–2.59 1.15 0.43–3.10
Circulatory failure 6 2
Cardiac arrhythmia 3 1
Brain death 0 2
Withdrawal of life support 7 3
Therapy failure 10 (27%) 5 (21%) 0.76 1.41 0.41–4.78 1.35 0.35–5.22
Hypotension 4 1
Acidosis 1 1
Oxygenation 4 2
Air leak 1 1
Cross-over 7 (19%) 4 (17%) 0.82 1.17 0.30–4.51 0.62 0.12–3.19
Supplemental oxygen or on ventilator at 30 days 9 (24%) 7 (29%) 0.96 0.96 0.26–3.58 0.67 0.12–3.84
Values between brackets are percentages of N (number of patients included in the analyses) except for CLD (Chronic Lung Disease) that has the number of survivors in the denominator. CI, confidence interval; OR, odds ratio unadjusted and adjusted for study site, OI, ventilatory index, APACHE II score, age and weight.
Table 3 Ventilatory conditions
HFOV CV
Cross-over No (30) Yes (7) No (20) Yes (4)
Day 1 N = 28 N = 7 (7 HFOV) N = 19 N = 4 (4 CV)
Peak inspiratory pressure (cmH2O) 32 ± 4.2 35 ± 6.9
Positive end-expiratory pressure (cmH2O) 14 ± 2.1 12 ± 4.5
Mean airway pressure (cmH2O) 30 ± 5.6a 32 ± 6.3a 22 ± 3.2 22 ± 6.1
Tidal volume per ideal bodyweight (ml/kg) 9 ± 1.7 8 ± 0.7
Frequency (HFOV, Hz; CV, breaths/min) 5 ± 0.5 5 ± 0.9 17.3 ± 3 17.3 ± 6
Delta P (cmH2O) 63 ± 14 70 ± 12.1
FiO2 0.78 ± 0.19 0.82 ± 0.12 0.68 ± 0.12 0.78 ± 0.21
pH 7.32 ± 0.08 7.31 ± 0.11 7.34 ± 0.08 7.22 ± 0.07b
pCO2 (mmHg) 49 ± 11.3 57 ± 13 48 ± 9 52 ± 15.8
pO2 (mmHg) 126 ± 79.2 93 ± 37.1 98 ± 26.6 99 ± 25
SaO2 (percentage) 95 ± 3 90 ± 10.7 96 ± 2.4 94 ± 4.5
Oxygenation index 26 ± 16 31 ± 8.3c 17 ± 7.5 19 ± 11.2c
Day 2 N = 27 N = 7 (6 HFOV) N = 19 N = 4 (2 CV)
Peak inspiratory pressure (cmH2O) 25 ± 6.7 36 ± 7.2 31 ± 4.5 30 ± 2.6
Positive end-expiratory pressure (cmH2O) 11 ± 1.2 15 ± 1.9 14 ± 2.7 12 ± 4.7
Mean airway pressure (cmH2O) 28 ± 6.7a 29 ± 4.3a 21 ± 2.3 22 ± 9.1
Tidal volume per ideal bodyweight (ml/kg) 9 ± 1.6 10 ± 1.9 8 ± 1.6 8 ± 1
Frequency (HFOV, Hz; CV, breaths/min) 5.0 ± 0.4 4.8 ± 1.1 17.4 ± 2.6 17.2 ± 1.2
Delta P (cmH2O) 64 ± 14.5 73 ± 14.8 70 ± 13.8
FiO2 0.55 ± 0.17 0.57 ± 0.14 0.53 ± 0.12 0.76 ± 0.20
pH 7.36 ± 0.07 7.35 ± 0.04 7.38 ± 0.06 7.22 ± 0.08b
pCO2 (mmHg) 45 ± 9 51 ± 8.9 46 ± 8.3 53 ± 8.5
pO2 (mmHg) 96 ± 21 83 ± 12.4 100 ± 27 87 ± 41.8
SaO2 (percentage) 95 ± 2.1 94 ± 1.9 96 ± 1.8 87 ± 16.1
Oxygenation index 17 ± 10.2 21 ± 8.2c 12 ± 3.6 22 ± 10.5c
Day 3 N = 23 N = 7 (4 HFOV) N = 19 N = 4 (2 CV)
Peak inspiratory pressure (cmH2O) 21 ± 3.1 32 ± 12 30 ± 4 27 ± 6
Positive end-expiratory pressure (cmH2O) 9 ± 3 10 ± 4.3 13 ± 2.8 11 ± 5.7
Mean airway pressure (cmH2O) 23 ± 7.1a 25 ± 6.9a 20 ± 2.8 24 ± 2.3
Tidal volume per ideal bodyweight (ml/kg) 9 ± 1.5 9 ± 3.5 9 ± 1.6 7 ± 1.6
Frequency (HFOV, Hz; CV, breaths/min) 5.0 ± 0.4 4.6 ± 0.5 18.8 ± 6.5 19.9 ± 5.8
Delta P (cmH2O) 66 ± 12.4 66 ± 19.1 67 ± 0.7
FiO2 0.46 ± 0.13 0.55 ± 0.15 0.46 ± 0.11 0.65 ± 0.26
pH 7.39 ± 0.06 7.37 ± 0.06 7.39 ± 0.06 7.33 ± 0.1b
pCO2 (mmHg) 45 ± 10.4 47 ± 12.9 48 ± 9 47 ± 12.6
pO2 (mmHg) 89 ± 19.7 86 ± 46.2 91 ± 13.7 89 ± 22.4
SaO2 (percentage) 94 ± 6.7 89 ± 14.1 96 ± 1.9 95 ± 2.4
Oxygenation index 14 ± 7.2 19 ± 9.3c 11 ± 3.7 20 ± 12.3c
Day 4 N = 22 N = 7 (3 HFOV) N = 19 N = 2 (0 CV)
Peak inspiratory pressure (cmH2O) 25 ± 8 31 ± 6.9 28 ± 6.9
Positive end-expiratory pressure (cmH2O) 9 ± 4.6 11 ± 4.2 11 ± 3.2
Mean airway pressure (cmH2O) 22 ± 7.8a 24 ± 6.2a 17 ± 5.6 24 ± 3.2
Tidal volume per ideal bodyweight (ml/kg) 10 ± 2.4 7 ± 3.1 8 ± 2.2
Frequency (HFOV, Hz; CV, breaths/min) 5.0 ± 0.3 4.3 ± 0.6 17.9 ± 5.3
Delta P (cmH2O) 57 ± 11.4 70 ± 11.8 48 ± 14.8
FiO2 0.45 ± 0.11 0.57 ± 0.18 0.45 ± 0.11 0.51 ± 0.12
pH 7.42 ± 0.14 7.37 ± 0.1 7.43 ± 0.12 7.45 ± 0.06b
pCO2 (mmHg) 43 ± 12.3 46 ± 7.5 41 ± 10.3 44 ± 11.1
pO2 (mmHg) 85 ± 22.3 84 ± 30.5 87 ± 27.4 74 ± 23.7
SaO2 (percentage) 89 ± 15.3 90 ± 14.1 89 ± 17.2 84 ± 20
Oxygenation index 12 ± 5.6 18 ± 7.9c 10 ± 4.3 19 ± 9.5c
The columns represent the treatment allocation. Measurements were made day 1, 2, 3 and 4 of the study. Peak inspiratory pressure, positive end-expiratory pressure and tidal volume per ideal bodyweight were measured in high frequency oscillatory ventilation (HFOV) after crossover to conventional mechanical ventilation (CV). Values are presented as means with standard deviations. aHigher mean airway pressures in HFOV compared with CV (p = 0.03). bSignificantly lower pH in patients that cross over in the CV group (p = 0.017). cHigher OI in patients that crossed over compared with patients that did not cross over (p = 0.07 and p = 0.05, respectively). FiO2, fraction of inspired oxygen; paCO2, pressure of arterial carbon dioxide; paO2, pressure of arterial oxygen; SaO2, arterial oxygen saturation.
==== Refs
Frank JA Matthay MA Science review: mechanisms of ventilator-induced injury Crit Care 2003 7 233 241 12793874 10.1186/cc1829
Rubenfeld GD Epidemiology of acute lung injury Crit Care Med 2003 31 S276 S284 12682453 10.1097/01.CCM.0000057904.62683.2B
Brower RG Rubenfeld GD Lung-protective ventilation strategies in acute lung injury Crit Care Med 2003 31 S312 S316 12682458 10.1097/01.CCM.0000057909.18362.F6
Froese AB High-frequency oscillatory ventilation for adult respiratory distress syndrome: let's get it right this time! Crit Care Med 1997 25 906 908 9201040 10.1097/00003246-199706000-00004
Henderson-Smart D Bhuta T Cools F Offringa M Elective high frequency oscillatory ventilation versus conventional ventilation for acute pulmonary dysfunction in preterm infants Cochrane Database Syst Rev 2003 4 CD000104 14583909
Bollen CW Uiterwaal CS van Vught AJ Cumulative metaanalysis of high-frequency versus conventional ventilation in premature neonates Am J Respir Crit Care Med 2003 168 1150 1155 14607823 10.1164/rccm.200306-721CP
Derdak S Mehta S Stewart TE Smith T Rogers M Buchman TG Carlin B Lowson S Granton J High-frequency oscillatory ventilation for acute respiratory distress syndrome in adults: a randomized, controlled trial Am J Respir Crit Care Med 2002 166 801 808 12231488 10.1164/rccm.2108052
David M Weiler N Heinrichs W Neumann M Joost T Markstaller K Eberle B High-frequency oscillatory ventilation in adult acute respiratory distress syndrome Intensive Care Med 2003 29 1656 1665 12897997 10.1007/s00134-003-1897-6
Arnold JH Hanson JH Toro-Figuero LO Gutierrez J Berens RJ Anglin DL Prospective, randomized comparison of high-frequency oscillatory ventilation and conventional mechanical ventilation in pediatric respiratory failure Crit Care Med 1994 22 1530 1539 7924362
Mehta S Lapinsky SE Hallett DC Merker D Groll RJ Cooper AB MacDonald RJ Stewart TE Prospective trial of high-frequency oscillation in adults with acute respiratory distress syndrome Crit Care Med 2001 29 1360 1369 11445688 10.1097/00003246-200107000-00011
Fort P Farmer C Westerman J Johannigman J Beninati W Dolan S Derdak S High-frequency oscillatory ventilation for adult respiratory distress syndrome – a pilot study Crit Care Med 1997 25 937 947 9201044 10.1097/00003246-199706000-00008
Derdak S High-frequency oscillatory ventilation for acute respiratory distress syndrome in adult patients Crit Care Med 2003 31 S317 S323 12682459 10.1097/01.CCM.0000057910.50618.EB
Sarnaik AP Meert KL Pappas MD Simpson PM Lieh-Lai MW Heidemann SM Predicting outcome in children with severe acute respiratory failure treated with high-frequency ventilation Crit Care Med 1996 24 1396 1402 8706497 10.1097/00003246-199608000-00020
Ventilation with lower tidal volumes as compared with traditional tidal volumes for acute lung injury and the acute respiratory distress syndrome. The Acute Respiratory Distress Syndrome Network. N Engl J Med 2000 342 1301 1308 10793162 10.1056/NEJM200005043421801
van Genderingen HR van Vught JA Jansen JR Duval EL Markhorst DG Versprille A Oxygenation index, an indicator of optimal distending pressure during high-frequency oscillatory ventilation? Intensive Care Med 2002 28 1151 1156 12185440 10.1007/s00134-002-1368-5
Lachmann B Open up the lung and keep the lung open Intensive Care Med 1992 18 319 321 1469157 10.1007/BF01694358
Ricard JD Dreyfuss D Saumon G Ventilator-induced lung injury Eur Respir J Suppl 2003 42 2s 9s 12945994 10.1183/09031936.03.00420103
Mehta S Granton J MacDonald RJ Bowman D Matte-Martyn A Bachman T Smith T Stewart TE High-frequency oscillatory ventilation in adults: the Toronto experience Chest 2004 126 518 527 15302739 10.1378/chest.126.2.518
Wunsch H Mapstone J High-frequency ventilation versus conventional ventilation for treatment of acute lung injury and acute respiratory distress syndrome Cochrane Database Syst Rev 2004 CD004085 14974056
|
16137357
|
PMC1269459
|
CC BY
|
2021-01-04 16:04:55
|
no
|
Crit Care. 2005 Jun 21; 9(4):R430-R439
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3737
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37411613735610.1186/cc3741ResearchLactate concentration gradient from right atrium to pulmonary artery Gutierrez Guillermo [email protected] Lakhmir S [email protected] Michael G [email protected] Nevin M 4Zia Hasan 51 Professor of Medicine and Anesthesiology, Pulmonary and Critical Care Medicine Division and Department of Medicine, The George Washington University Medical Center Washington, DC, USA2 Assistant Professor of Anesthesiology and Medicine, Critical Care Medicine Division, Department of Anesthesiology, The George Washington University Medical Center Washington, DC, USA3 Associate Professor Anesthesiology, Critical Care Medicine Division, Department of Anesthesiology, The George Washington University Medical Center Washington, DC, USA4 Clinical Professor of Surgery, Cardio-Thoracic Critical Care, Department of Surgery, The George Washington University Medical Center Washington, DC, USA5 Senior Resident, Division of General Surgery, Department of Surgery, The George Washington University Medical Center Washington, DC, USA2005 10 6 2005 9 4 R425 R429 26 4 2005 9 5 2005 16 5 2005 20 5 2005 Copyright © 2005 Gutierrez et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
We compared simultaneous measurements of blood lactate concentration ([Lac]) in the right atrium (RA) and in the pulmonary artery (PA). Our aim was to determine if the mixing of right atrial with coronary venous blood, having substantially lower [Lac], results in detectable decreases in [Lac] from the RA to the PA.
Methods
A prospective, sequential, observational study was conducted in a medical-surgical intensive care unit. We enrolled 45 critically ill adult individuals of either sex requiring pulmonary artery catheters (PACs) to guide fluid therapy. Immediately following the insertion of the PAC, one paired set of blood samples per patient was drawn in random order from the PAC's proximal and distal ports for measurement of hemoglobin concentration, O2 saturation (SO2) and [Lac]. We defined Δ[Lac] as ([Lac]ra - [Lac]pa), ΔSO2 as (SraO2 - SpaO2) and the change in O2 consumption (ΔVO2) as the difference in systemic VO2 calculated using Fick's equation with either SraO2 or SpaO2 in place of mixed venous SO2. Data were compared by paired Student's t-test, Spearman's correlation analysis and by the method of Bland and Altman.
Results
We found SraO2 > SpaO2 (74.2 ± 9.1 versus 69.0 ± 10.4%; p < 0.001) and [Lac]ra > [Lac]pa (3.9 ± 3.0 versus 3.7 ± 3.0 mmol.l-1; p < 0.001). Δ[Lac] correlated with ΔVO2 (r2 = 0.34; p < 0.001).
Conclusion
We found decreases in [Lac] from the RA to PA in this sample of critically ill individuals. We conclude that parallel decreases in SO2 and [Lac] from the RA to PA support the hypothesis that these gradients are produced by mixing RA with coronary venous blood of lower SO2 and [Lac]. The present study is a preliminary observation of this phenomenon and further work is needed to define the physiological and clinical significance of Δ[Lac].
See related commentary
==== Body
Introduction
Pulmonary artery (PA) blood comprises the mixed venous effluent from all organs, with the notable exception of the lungs. PA O2 saturation (SpaO2) has been promoted as an index of tissue oxygenation [1,2] because it is thought to be related to the average end capillary blood PO2 [3].
In a prior study [4], we measured the O2 saturation (SO2) of right atrial blood (SraO2) and SpaO2 in samples drawn from the proximal and distal ports of PA catheters (PACs) placed in critically ill patients. We noted that SpaO2 was consistently lower than SraO2 by approximately 5%. Others have noted a similar step-down in O2 saturation from the right atrium (RA) to the PA [5,6], and continuous measurements in critically ill patients have shown a similar difference between SpaO2 and central venous (CV) O2 saturation (ScvO2) of approximately 7% [7].
The RA to PA O2 saturation gradient (defined as ΔSO2 = SraO2 - SpaO2) is likely the result of mixing atrial blood with highly desaturated blood entering the right heart chambers from the coronary veins. This includes blood flowing from the coronary sinus, the great cardiac vein and other major epicardial veins.
As a result of myocardial lactate extraction from the coronary circulation, the CV lactate concentration ([Lac]cv) is the lowest of any venous blood [8,9]. In the present study we compare blood lactate concentration ([Lac]) in paired samples drawn from the proximal and distal ports of PACs placed in critically ill patients ([Lac]ra and [Lac]pa) to establish whether we could also detect a decreasing lactate concentration gradient from right atrium to pulmonary artery (Δ[Lac] = [Lac]ra - [Lac]pa).
Methods
This was a prospective, sequential study performed in the George Washington University Hospital intensive care unit. The George Washington University Institutional Review Board approved the study and informed consent was obtained from the patient or from the next of kin.
The data presented were culled from a subset of patients enrolled in a previous study [4]. We enrolled individuals older than 18 years of age of either sex in whom their physicians determined that a PAC was required to guide fluid therapy. Enrollment in the study occurred at the time the patient or the nearest relative consented to the introduction of the PAC. On the basis of their medical history, we excluded patients with uncorrected valvular incompetence, intra-cardiac shunting or those who required insertion of the pulmonary artery catheter through the femoral vein.
A 7.5 French, 5 lumen, 110 cm length, PAC with the right atrial lumen positioned 30 cm from the tip (Edwards Lifesciences, Irvine, CA, USA) was inserted through the internal jugular vein or the subclavian vein using a percutaneous sheath introducer (8.5 French; Edwards Lifesciences). The insertion technique is described elsewhere [4]. Care was taken to place the distal port catheter in the PA and the proximal port in the RA.
Immediately after the insertion of the PA catheter, each patient had one set of paired blood samples drawn in rapid succession, and in random order, from the proximal and distal port. We took proximal port blood to be representative of RA blood, whereas distal port blood was considered to be PA blood. The first 2 ml of blood drawn for each sample were discarded to prevent contamination with flushing fluid. Blood samples were drawn with the catheter balloon deflated to avoid contamination of the distal port sample with pulmonary capillary blood. Arterial O2 saturation was determined from a previously in vivo calibrated pulse oximeter.
Blood samples were placed on ice and taken to a central laboratory for measurement of [Lac] (Ektachem 950 IRC Chemistry Analyzer with a Vitros Products lactate slide, Ortho-Clinical Diagnostic, Inc., Rochester, NY, USA), hemoglobin concentration ([Hb]) and O2 saturation (ABL700 Radiometer America Inc., Westlake, OH, USA). We measured cardiac output (CO) by the thermodilution method as the average of three sequential determinations.
Systemic O2 delivery (DO2), O2 consumption (VO2), O2 extraction ratio (ERO2), double product (DP; heart rate (HR) × mean arterial pressure (MAP)) and left ventricular stroke work index (LVSWI) were computed using standard formulae. We defined ΔVO2 as the difference in systemic VO2 calculated with Fick's equation with either SpaO2 or SraO2 in place of the mixed venous SO2 (SvO2); ΔVO2 = Qpa × 13.4 × [Hb] × (SraO2 - SpaO2) ml.min-1.
Paired Student's t-test was used to compare atrial to PA measurements. [Lac]ra and [Lac]pa were compared by Spearman's correlation analysis [10]. The method of Bland and Altman [11] was used to investigate the effect of lactate concentration on the differences between paired observations. The relationships between Δ[Lac] and ΔSO2, ΔVO2 and other hemodynamic parameters were analyzed by Spearman's correlation analysis. Data are shown as mean ± SD with p < 0.05 denoting a significant difference.
Results
We enrolled 45 patients in the study, including 18 women. The study group was composed of 31 post-operative patients (26 post-cardiac surgery), 11 patients in septic shock from various medical conditions, 2 patients with severe gastrointestinal bleeding and 1 patient in congestive heart failure. Demographic and hemodynamic parameters for the group are listed in Table 1.
The mean SO2 and lactate concentrations for RA and PA blood samples are shown in Table 2. SraO2 was greater than SpaO2 (p < 0.001), with ΔSO2 = 5.2 ± 4.8%. [Lac]ra was greater than [Lac]pa (p < 0.001), with Δ[Lac] = 0.2 ± 0.2 mmol.l-1.
Shown in Fig. 1 is a Bland-Altman plot comparing [Lac]ra and [Lac]pa. There was a bias towards greater [Lac]ra of 0.2 mmol.l-1 (p < 0.001) with a 95% confidence interval for the population of -0.15 to 0.56 mmol.l-1. There was no discernable relationship between [Lac]ra and Δ[Lac] (r2 = 0.03; p = 0.33), indicating that Δ[Lac] was not a concentration dependent phenomenon. Moreover, we found no significant relationships between [Lac]ra and SraO2 or between [Lac]pa and SpaO2.
There was a significant relationship between Δ[Lac] and ΔVO2 (Δ[Lac] mmol.l-1 = 0.0026 ΔVO2 ml.min-1 + 0.0975; r2 = 0.34; p < 0.0001) with a standard error of the estimate of 0.15 mmol.l-1 (Fig. 2). There were no significant correlations between Δ[Lac] and cardiac index, DP, LVSWI, DO2, VO2 or ERO2.
Discussion
We detected a decreasing Δ[Lac] when comparing paired blood samples drawn from the proximal and distal ports of PACs. We also noted Δ[Lac] correlated with ΔVO2. To our knowledge, these novel findings have not been reported elsewhere.
Only one other study in the literature has compared central venous [Lac] to [Lac]pa. This study found no differences in [Lac], although it was biased by the use of multiple blood samples (n = 50) drawn from 12 critically ill patients [12]. Our study used only one comparison per subject, which perhaps may explain the difference in results.
We used a standard clinical laboratory instrument to measure [Lac] having a 95% precision of ± 0.1 mmol.l-1. Even assuming a worst case scenario of a systematic instrument bias of -0.1 mmol.l-1, the difference in [Lac] between RA and PA would have remained statistically significant.
The declining [Lac] gradient from RA to PA is likely the result of mixing RA blood with blood of lower [Lac] emanating from the coronary venous system. Lactate oxidation accounts for 10% to 20% of total myocardial aerobic energy production [13], a proportion that increases substantially in sepsis [14]. As a result of myocardial lactate extraction, coronary venous [Lac] is substantially lower than arterial [Lac] and is the lowest of all venous effluents [15]. The dilution of RA blood by coronary venous blood of lower [Lac] is a plausible explanation for the small but detectable difference in [Lac] from RA to PA.
Since RA blood is the mixture of superior vena cava and inferior vena cava (IVC) blood, the possibility exists that these blood streams had not thoroughly mixed at the proximal PAC sampling port. In this case, one could expect further mixing to occur between IVC and RA blood while flowing into the pulmonary artery. Our results do not support this hypothesis. Direct measurements in humans show that IVC blood has the highest [Lac] of any major vein [9] and further mixing of RA with IVC blood would have produced higher, not lower, [Lac]pa. A factual resolution of this question can only be achieved by direct measurement of [Lac] from IVC to PA.
Only three individuals in our group had [Lac]ra < [Lac]pa. These patients had no distinguishing features to help us differentiate them from others in the group. It is possible that accidental mislabeling of the samples may have accounted for a negative Δ[Lac] but we think it unlikely, given the care taken with the labeling and measuring of the samples. Another possibility is that these individuals experienced myocardial ischemia, a condition associated with an upsurge in glucose metabolism and net lactate release by the heart [17-19]. Myocardial lactate release, as opposed to the normal state of myocardial uptake, would have resulted in [Lac]ra < [Lac]pa.
Others have noted a linear relationship between myocardial O2 consumption (MVO2) and myocardial lactate uptake, reflecting the O2 cost of lactate utilization by the heart [14]. We did not measure MVO2 directly but calculated ΔVO2, a parameter denoting the difference in systemic VO2 prior to and immediately after entry of myocardial effluent blood into the venous circulation. As such, ΔVO2 bears a direct relationship to MVO2. We noted a linear relationship between ΔVO2 and Δ[Lac] (Fig. 2) similar to that described between MVO2 and myocardial lactate uptake. This finding suggests that Δ[Lac] also may be related, in a yet to be established fashion, to MVO2.
Conclusion
We found decreases in [Lac] from RA to PA in this sample of critically ill individuals. We conclude that parallel decreases in SO2 and [Lac] from RA to PA support the hypothesis that these gradients are produced by mixing RA with coronary venous blood of lower SO2 and [Lac]. The present study is a preliminary observation of this phenomenon and further work is needed to define the physiological and clinical significance of Δ[Lac].
Key messages
• Oxygen and lactate concentrations are lower in PA blood than in RA blood.
• The oxygen and lactate concentration gradients from RA to PA are likely the result of mixing atrial with coronary venous blood.
• The possibility exists that these concentration gradients may reflect changes in myocardial energy requirements.
Abbreviations
CV = coronary venous; CVP = central venous pressure; DO2 = systemic O2 delivery; DP = double product; ERO2 = oxygen extraction ratio; [Hb] = hemoglobin concentration; HR = heart rate; IVC = inferior vena cava; Δ[Lac] = lactate concentration gradient from right atrium to pulmonary artery; [Lac] = blood lactate concentration; LVSWI = left ventricular stroke work index; MAP = mean arterial pressure; MPP = mean pulmonary pressure; MVO2 = myocardial O2 consumption; PA = pulmonary artery; PAC = pulmonary artery catheter; PAOP = pulmonary artery occlusion pressure; RA = right atrium; SO2 = O2 saturation; ΔSO2 = O2 saturation gradient from right atrium to pulmonary artery; SVRI = systemic vascular resistance index; VO2 = O2 consumption.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
GG conceived the study, participated in its design, performed statistical analysis and drafted the manuscript. LSC and HZ participated in the design of the study, collected data and helped to draft the manuscript. MGS and NMK conducted the study, collected data and helped to draft the manuscript. All authors read and approved the final manuscript.
Acknowledgements
The George Washington University Medical Center Department of Anesthesiology Research Fund financed the study in its entirety. Preliminary results of the study were presented in abstract form at the 2003 American Thoracic Society International Conference, Seattle, WA, USA.
Figures and Tables
Figure 1 Bland-Altman plot comparing [Lac]ra and [Lac]pa. Bias 0.21 mmol.L-1 with a 95% confidence interval for the population of -0.15 to 0.56 mmol.L-1.
Figure 2 Linear correlation of Δ[Lac] to ΔVO2. The latter represents the difference in VO2 calculated using either SraO2 or SpaO2 in place of mixed venous SO2 in the Fick's Equation (Δ[Lac] mmol.L-1 = 0.0026 ΔVO2 ml.min-1 + 0.0975; r2 = 0.34; p < 0.0001). Standard error of the estimate 0.15 mmol.L-1.
Table 1 Study population demographic and hemodynamic parameters
Patient parameters (n = 45) Mean ± SD
Age (years) 57.6 ± 13.2
APACHE II score 13.8 ± 6.0
HR (bpm) 92.1 ± 16.5
MAP (mmHg) 81.8 ± 13.0
MPP (mmHg) 27.6 ± 9.9
PAOP (mmHg) 18.6 ± 7.0
CVP (mmHg) 15.0 ± 6.1
Cardiac output (ml.min-1) 6.1 ± 2.6
Cardiac Index (ml.min-1.m-2) 3.3 ± 1.5
LVSWI (g.m.m-2.beat) 39.7 ± 12.6
DP (mmHg.beat.min-1) 7694 ± 1944
SVRI (dynes.sec.m-5) 2002 ± 1316
Hemoglobin (g.dl-1) 10.8 ± 2.0
CVP, central venous pressure; DP, double product (HR × MAP); HR, heart rate; LVSWI, left ventricular stroke work index; MAP, mean arterial pressure; MPP, mean pulmonary pressure; PAOP, pulmonary artery occlusion pressure; SVRI, systemic vascular resistance index.
Table 2 O2 saturation and lactate concentration of paired RA and PA blood samples
RA blood PA blood Gradient (Δ)
O2 saturation (%) 74.2 ± 9.1 (53.1, 94.3) 69.0 ± 10.4a (47.3, 90.5) 5.2 ± 4.8 (-8.1, 14.9)
Lactate concentration (mmol.l-1) 3.9 ± 3.0 (0.6, 11.7) 3.7± 3.0a (0.3, 11.9) 0.2 ± 0.2 (-0.3, 0.7)
aP < 0.001 when comparing atrial to mixed venous blood by paired t-test. Mean ± SD; range shown in parenthesis; n = 45. RA, right atrium; PA, pulmonary artery.
==== Refs
Vincent JL The relationship between oxygen demand, oxygen uptake, and oxygen supply Intensive Care Med 1990 16 suppl 2 S145 S148 2289980 10.1007/BF01785244
Gutierrez G Wulf-Gutierrez M Reines HD Monitoring oxygen transport and tissue oxygenation Curr Opin Anes 2004 17 107 117 10.1097/00001503-200404000-00004
Schumacker PT Long GR Wood LD Tissue oxygen extraction during hypovolemia: role of hemoglobin P50 J Appl Physiol 1987 62 1801 1807 3597253
Chawla LS Zia H Gutierrez G Katz NM Seneff MG Shah M Lack of equivalence between central and mixed venous oxygen saturation Chest 2004 126 1891 1896 15596689 10.1378/chest.126.6.1891
Berridge JC Influence of cardiac output on the correlation between mixed venous and central venous oxygen saturation Br J Anaesth 1992 69 409 410 1419454
Edwards JD Mayall RM Importance of the sampling site for measurement of mixed venous oxygen saturation in shock Crit Care Med 1998 26 1356 1360 9710094 10.1097/00003246-199808000-00020
Reinhart K Kuhn HJ Hartog C Bredle DL Continuous central venous and pulmonary artery oxygen saturation monitoring in the critically ill Intensive Care Med 2004 30 1572 1578 15197435 10.1007/s00134-004-2337-y
Wolfhard UF Brinkmann M Splittgerber FH Knocks M Sack S Piotrowski JA Schieffer M Gunnicker M Myocardial lactate extraction during repeated fibrillation/defibrillation episodes in defibrillator implantation testing Pacing Clin Electrophysiol 1998 21 1795 1801 9744445
Waldau T Larsen VH Bonde J Fogh-Andersen N Lactate, pH, and blood gas analysis in critically ill patients Acta Anaesthesiol Scand Suppl 1995 107 267 271 8599289
Zar JH Simple Linear Correlation Biostatistical Analysis 1999 4 Prentice-Hall, Inc; Upper Saddle River, NJ 377 383
Bland JM Altman D Statistical methods for assessing agreement between two methods of clinical measurement Lancet 1986 1 307 310 2868172
Weil MH Michaels S Rackow EC Comparison of blood lactate concentrations in central venous, pulmonary artery, and arterial blood Crit Care Med 1987 15 489 490 3568712
Abel ED Glucose transport in the heart Front Biosci 2004 9 201 215 14766360
Dhainaut JF Huyghebaert MF Monsallier JF Lefevre G Dall'Ava-Santucci J Brunet F Villemant D Carli A Raichvarg D Coronary hemodynamics and myocardial metabolism of lactate, free fatty acids, glucose, and ketones in patients with septic shock Circulation 1987 75 533 541 3815765
Stanley WC Chandler MP Energy metabolism in the normal and failing heart: potential for therapeutic interventions Heart Fail Rev 2002 7 115 130 11988636 10.1023/A:1015320423577
Crittenden MD Intraoperative metabolic monitoring of the heart: I. Clinical assessment of coronary sinus metabolites Ann Thorac Surg 2001 72 S2220 S2226 11789845 10.1016/S0003-4975(01)03296-9
Stanley WC Lopaschuk GD Hall JL McCormack JG Regulation of myocardial carbohydrate metabolism under normal and ischaemic conditions. Potential for pharmacological interventions Cardiovasc Res 1997 33 243 257 9074687 10.1016/S0008-6363(96)00245-3
Peuhkurinen K Ikaheimo M Airaksinen J Huikuri H Linnaluoto M Takkunen J Changes in myocardial energy metabolism in elective coronary angioplasty Cardiovasc Res 1991 25 158 163 1742766
Foltz WD Merchant N Downar E Stainsby JA Wright GA Coronary venous oximetry using MRI Magn Reson Med 1999 42 837 848 10542342 10.1002/(SICI)1522-2594(199911)42:5<837::AID-MRM3>3.0.CO;2-8
|
16137356
|
PMC1269463
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jun 10; 9(4):R425-R429
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3741
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37491613735910.1186/cc3749ResearchQuality of interhospital transport of critically ill patients: a prospective audit Ligtenberg Jack JM [email protected] L Gert [email protected] Ymkje [email protected] der Werf Tjip S [email protected] John HJM [email protected] Jaap E [email protected] Jan G [email protected] Internist-intensivist, Intensive and Respiratory Care Unit (ICB), Department of Internal Medicine, University Medical Center Groningen, Groningen, The Netherlands2 Intensive Care Nurse, Intensive and Respiratory Care Unit (ICB), Department of Internal Medicine, University Medical Center Groningen, Groningen, The Netherlands3 Senior resident, Department of Internal Medicine, University Medical Center Groningen, Groningen, The Netherlands4 Pulmonologist-intensivist, Intensive and Respiratory Care Unit (ICB), Department of Internal Medicine, University Medical Center Groningen, Groningen, The Netherlands5 Anesthesiologist-intensivist, Intensive and Respiratory Care Unit (ICB), Department of Internal Medicine, University Medical Center Groningen, Groningen, The Netherlands2005 1 7 2005 9 4 R446 R451 19 1 2005 2 3 2005 23 5 2005 2 6 2005 Copyright © 2005 Ligtenberg et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
The aim of transferring a critically ill patient to the intensive care unit (ICU) of a tertiary referral centre is to improve prognosis. The transport itself must be as safe as possible and should not pose additional risks. We performed a prospective audit of the quality of interhospital transports to our university hospital-based medical ICU.
Methods
Transfers were undertaken using standard ambulances. On departure and immediately after arrival, the following data were collected: blood pressure, heart rate, body temperature, oxygen saturation, arterial blood gas analysis, serum lactic acid, plasma haemoglobin concentration, blood glucose, mechanical ventilation settings, use of vasopressor/inotropic drugs, and presence of venous and arterial catheters. Ambulance personnel completed forms describing haemodynamic and ventilatory data during transport. Data were collected by our research nurse and analyzed.
Results
A total of 100 consecutive transfers of ICU patients over a 14-month period were evaluated. Sixty-five per cent of patients were mechanically ventilated; 38% were on vasoactive drugs. Thirty-seven per cent exhibited an increased number of vital variables beyond predefined thresholds after transport compared with before transport; 34% had an equal number; and 29% had a lower number of vital variables beyond thresholds after transport. The distance of transport did not correlate with the condition on arrival. Six patients died within 24 hours after arrival; vital variables in these patients were not significantly different from those in patients who survived the first 24 hours. ICU mortality was 27%. Adverse events occurred in 34% of transfers; in 50% of these transports, pretransport recommendations given by the intensivist of our ICU were ignored. Approximately 30% of events may be attributed to technical problems.
Conclusion
On aggregate, the quality of transport in our catchment area carried out using standard ambulances appeared to be satisfactory. However, examination of the data in greater detail revealed a number of preventable events. Further improvement must be achieved by better communication between referring and receiving hospitals, and by strict adherence to checklists and to published protocols. Patients transported between ICUs are still critically ill and should be treated as such.
See related commentary
==== Body
Introduction
Transfer of critically ill patients to the intensive care unit (ICU) of a tertiary referral centre is intended to improve prognosis. The transport itself must be as safe as possible and should not pose additional risks to the patient. Circulatory or ventilatory problems may arise in the ambulance as well as during transportation inside the hospital [1-3]. Monitoring capabilities are limited during transportation, and fewer (and less skilled) 'hands' are available on the road as compared with the ICU environment.
We accept 90–100 patient transfers each year from the ICUs of other hospitals. In our region patients are generally transported in standard ambulances; it is the responsibility of the referring hospital to ensure that a safe transfer takes place. We conducted a prospective audit of the quality of these transports, addressing the following questions. Did vital variables, documented before transfer, pass critical thresholds during transportation? Were changes in vital variables dependent on duration and/or length of transport? Did patients dying during or shortly after transfer have other vital variables before or during transport than did patients who survived the first 24 hours? Would it be possible to predict which patients will not benefit at all from transfer? Finally, what was the frequency of adverse incidents related to transfer?
We anticipated that answers to these questions would help us to decide whether an upgrade to transports in our area is needed, for example putting a mobile intensive care unit (ICU) from our university hospital into action.
Materials and methods
A communication was send to the referring hospitals and ambulance services explaining the aims of the study. We clarified the protocol at different locations. Once these services and hospitals had agreed to participate, the study was started. In all cases, patients were transferred after telephone consultation with the supervising staff member of our ICU, who authorized the admission. The referring hospital was advised to stabilize the patient as much as possible and to send a skilled physician with the patient. Predefined study variables in 100 consecutive ICU transports were recorded just before departure and immediately after arrival. The following data were collected: blood pressure, heart rate, body temperature, oxygen saturation, arterial blood gases, serum lactic acid, plasma haemoglobin concentration, blood glucose, mechanical ventilation settings, use of vasopressor/inotropic medication, and presence of venous and arterial catheters. Ambulance personnel completed forms describing haemodynamic and ventilatory data during transport. Blood sampling and data acquisition on arrival were performed with the patient still on the ambulance stretcher (if this was considered safe), before changes to the 'on the road' therapy were instituted. Immediately thereafter, the patient was moved to the ICU bed and connected to the ICU ventilator. Data were instantly noted on a simple data sheet, and checked and collected by our research nurse.
The local medical ethics committee was informed and approved the design of our study.
Statistical analysis
We tested, for each parameter, whether the value beyond a predefined threshold on departure differed from the value beyond the threshold on arrival, using the the McNemar test. This test is typically used in a repeated measures situation, in which each subject's response is elicited twice, once before and once after a specified event (in this case transfer) occurs. For calculations of the change in number of vital variables beyond the threshold occurring as a result of transport, variables were included – when available – both from before and after transport. For each individual the total number of vital variables beyond threshold before and after transport was determined. Whether there was a difference between these two time points was tested using the Wilcoxon signed ranks test. P < 0.05 was considered statistically significant. Data were analyzed using SPSS for Windows version 12.0 (SPSS Inc, Chicago, IL, USA).
Results
In total, 100 consecutive transfers of ICU patients were evaluated over a 14-month period.
Transport characteristics
Patient transport characteristics are summarized in Table 1. Most transfers (96) were from 18 regional hospitals in the north eastern part of The Netherlands; four were from four ICUs located elsewhere in The Netherlands. Three ICUs transferred 10 or more patients; five ICUs transferred five to nine patients; and the others transferred between one and four patients. More than one-third arrived during the night shift (i.e. between 17.00 and 08.00 hours). An ICU nurse was present in 23% of transports and a physician in 57%. Blood was drawn within 6.4 ± 9 min (mean ± standard deviation) after arrival at our ICU.
Diagnoses
Respiratory problems (e.g. oxygenation problems during mechanical ventilation or weaning difficulties) were the most common reason for transfer (Table 2). Severe multiple organ failure and sepsis together were responsible for 35% of transfers, in part because of a need for renal replacement therapy. Four patients with gastrointestinal tract bleeding were transferred because trained interventional endoscopists were not available in the admission hospital. Shortage of ICU capacity was cited as the reason for transport on only a few occasions. A diagnostic problem existed in more than 30% of cases.
Patient characteristics on arrival
The characteristics of patients on arrival are summarized in Table 3. Sixty-five per cent of patients were mechanically ventilated and 38% were on vasoactive drugs.
Vital parameters
Variables on departure and arrival are summarized in Table 4. The percentage of patients arriving with values beyond predefined critical thresholds is shown. We defined these thresholds as clinically relevant deteriorations, based on thresholds cited in the literature (e.g. the haemoglobin threshold cited in the study by Hebert and coworkers [4] and in the TRICC (Transfusion requirements in critical care) trial [5] or thresholds used in clinical practice in The Netherlands. The number of patients in whom a critical threshold was reached during transport was calculated (with normal values on departure but values beyond critical thresholds at arrival indicating a worsening in the patient's condition during transfer).
The median number of variables beyond threshold was 2 (of the 12 mentioned in Table 4), both before and after transfer. The maximum number of variables beyond threshold in one patient was 6 (after transfer). Thirty-seven per cent of patients exhibited an increased number of variables beyond threshold after transport as compared with before transport (23% had one parameter more after transport, 9% had two more, 3% had three more, 1% had four more and 1% had five more); 34% had an equal number beyond threshold before and after transport; and 29% had a lower number beyond threshold after transport. These differences were not statistically significant (P = 0.182, by Wilcoxon signed rank test).
Patients in whom there was a greater number of variables beyond threshold after transport than before transport did not have a longer transportation time than did the other patients (median transport time: 40 min versus 38 min, respectively; P = 0.76, by Mann–Whitney U-test). Six patients (6%) died within 24 hours after arrival. These patients had a median of 2.5 variables beyond threshold at departure, and a median of 3 vital parameters beyond threshold at arrival. For the patients who did not die within 24 hours there were 2 variables beyond threshold both before and after transport (P = 0.74, by Mann–Whitney U-test). ICU mortality was 27%.
Adverse events
Events were recorded in 34 out of 100 transfers (examples of such events are sumarized in Table 5). The impact of various events was graded as follows: grade 1 = deviation from ambulance guidelines/local protocols/advice from tertiary centre; grade 2 = of vital importance – immediate action needed on arrival; and grade 3 = of vital importance – immediate action needed on arrival – event probably avoidable.
In summary, adverse events occurred in 34% of transfers. In 50% of these transports recommendations for safe transport of the patient, given by the intensivist of our ICU, were ignored. We estimate that 70% of events could have been avoided by better preparation for the transfer. Approximately 30% of events could be attributed to technical problems during transport; some of these could have been prevented (e.g. shortage of oxygen on the road).
Discussion
In this prospective study, changes in parameters for major worsening during transport never achieved statistical significance (Table 4). Based on this, one could conclude that the quality of transport in our catchment area, carried out with standard ambulances, is sufficient. However, evaluation of data for individual patients showed some serious deteriorations during transport; in 34% of transports events occurred, some of which had vital complications. Of course, this could be because the patients were critically ill. Another cause of deterioration could be the occurrence of adverse events during transport. In 50% of the transports with events, recommendations for transport of the patient – given by the intensivist of our ICU – were ignored. We estimate that 70% of events could have been avoided by better preparation and communication before transfer. This may represent poor clinical care by the referring centres, perhaps caused by underestimation of the risks associated with ICU tranfers or overestimation of the skills of ambulance personnel. Another reason may be that we had no standard protocol for giving feedback to the referring physician after the transfer had been performed. Thirty per cent of events could be attributed to technical problems during transport.
ICU mortality in our study group was 27%; because 73% survives one could state, that our way of selecting patients for referral is adequate and that these 73% benefit from admission to our ICU. However, no data are available on mortality in comparable patients in our region, who were not transported. The APACHE II score on arrival did not differ from our average ICU population; however, the APACHE score of the study group may be affected by prior stabilisation during admission in the referring hospital. Mortality of our total ICU population is 18%; this illustrates the fact that APACHE score may not be an adequate instrument to predict mortality in transferred patients from other hospitals [6]. Durairaj et al -in a large study in > 3000 transferred ICU patients – used diagnostic category and comorbidity scores, which showed a better correlation with morbidity and mortality in transferred patients [7].
We are concerned about the observed lack of preparation before transfer of patients. Although we consistently advised that a skilled physician accompany the patient, a number of patients arrived without a doctor. Simply adherence to existing ambulance checklists would have avoided a few events, for example equipment failures, incomplete supplies, shortage of oxygen or batteries, and drug administration errors.
We do not know whether a special retrieval team using a mobile ICU would improve the quality of transports in our catchment area. Several positive experiences with special retrieval teams have been reported [8,9]. In a study conducted by Bellingan and coworkers [7] transports by a specialist retrieval team, compared with standard ambulance transport with a doctor from the referring hospital, resulted in more stabile transports and a reduction in mortality during the first 12 hours from 7.7% to 3%. ICU mortality was not significantly different (35% versus 28%). It seems logical to use a specialist team and a mobile ICU for transport of more severely ill patients [10], but we were unable to find reports of pretransport parameters that could predict which patients will deteriorate during transfer and who may benefit from a retrieval ream with a mobile ICU.
Guidelines for safe ICU transfers have been reported by The Netherlands Society of Intensive Care, among others [11-13], and there are new regional ambulance guidelines that require a skilled physician to accompany each ventilated patient; if this is not possible the patient will not be transported. Based on our own findings and the new guidelines, our recommendations to the referring centre for the transfer of patients have become more strict.
Conclusion
Before deciding to transport a critically ill patient, it must be borne in mind that such a transfer has its own risks. Such risks will become more prominent in the near future because of the tendency to centralize advanced health support to a few regional centres. Further improvement must be achieved by better communication between the referring and receiving hospital before transport is initiated, and by strict adherence to checklists and to published guidelines. Patients transported between ICUs are still critically ill patients and should be treated as such. Whether these measures will render the use of a mobile ICU in our area unnecessary is not yet known.
Key messages
• Interhospital transfer of critically ill patients must be as safe as possible and should not pose additional risks.
• Seventy per cent of events that occurred could have been avoided by better preparation for the transfer.
• Further improvement may be achieved by better communication between the referring and receiving hospital before the transport is initiated, and by strict adherence to checklists and published guidelines.
Abbreviations
APACHE = Acute Physiology and Chronic Health Evaluation; ICU = intensive care unit.
Competing interests
The author(s) declare that they have no competing interests.
Acknowledgements
We thank the referring hospitals, ambulance personnel and intensive care nurses of our ICU for their enthusiastic cooperation.
Figures and Tables
Table 1 Transport characteristics
Characteristic Value
Distance (km) 57 ± 43
Transporting time (min) 47 ± 30
Arrival 17.00–08.00 hours (%) 37
Ambulance nurse (%) 100
+ ICU nurse (%) 23
+ Physician (and ICU nurse) (%) 57
Blood sampling after arrival (min) 6.4 ± 9
Data are given as mean ± standard deviation or as percentage.
Table 2 Transfer diagnosis
Reason for transfer %
Respiratory problems 32
Multiple organ failure 25
Sepsis 10
Cardiac 8
(Neuro)surgical problems 8
Gastrointestinal bleeding 4
Intoxication 4
Other diagnosis 9
'Other diagnoses' include end-stage liver failure (n = 1), HELLP (haemolysis–elevated liver enzymes–low platelets) syndrome (n = 2), microangiopathic thrombotic syndrome (thrombotic thrombocytopenic purpura/haemolytic uremic syndrome; n = 2), Wegener's granulomatosis (n = 2), and pulmonary embolism (n = 2).
Table 3 Characteristics of patients on arrival
Characteristic Value
Age (years) 54.7 ± 1.7
Sex (female/male) 49/51
Mechanically ventilated 65
Oxygen
Mask 14
Nasal 21
Central venous line 47
Intra-arterial catheter 72
Peripheral venous line 96
Vasopressor/inotropic drugs 38
APACHE II score 12.6 ± 0.7
Data are expressed as a percentage (%) or as mean ± standard deviation. APACHE, Acute Physiology and Chronic Health Evaluation.
Table 4 Variables on departure and arrival
Variable Departure (mean ± SD) Arrival (mean ± SD) Arrival (min–max) Critical threshold Beyond threshold on departure (%) Beyond threshold during transport (%) Beyond threshold on arrival (%) P
Arterial pH 7.35 ± 0.17 7.36 ± 0.13 6.98–7.57 <7.20 7 - 13 0.18
Oxygen saturation (%) 94.2 ± 7.0 93.9 ± 7.3 68–100 <90 13 20 16 0.58
PCO2 (kPa) 13.5 ± 9.9 18.3 ± 14.7 4.30–71.8 <8 12 - 16 0.45
PCO2 (kPa) 5.6 ± 1.8 5.9 ± 2.3 2.80–12.91 >6.0 35 - 35 1.00
SBP (mmHg) 120.8 ± 22.8 126.9 ± 30 60–210 <90, >180 7, 0 14, 3 11, 3 0.09
DBP (mmHg) 64.4 ± 15.9 68.43 ± 18.4 37–145 <50, >110 14, 1 14, 3 14, 2 1.00
Heart rate (beats/min) 103.5 ± 23.8 103.9 ± 23.3 50–160 <50, >120 0, 30 2, 31 0, 28 0.82
Temperature (°C) 37.7 ± 1.3 37.04 ± 1.3 32.0–40.4 <36.0 8 - 12 0.29
Lactate (mmol/l) 1.9 ± 2.5 2.1 ± 1.9 0.6–13.2 >3.0 10 - 8 1.00
Haemoglobin (mmol/l; g/dl × 1.6) 6.9 ± 1.5 6.6 ± 1.4 3.4–10.0 <5.0 6 - 11 0.06
Glucose (mmol/l) 9.0 ± 5.0 7.9 ± 4.3 3.4–25.5 <4.0, >12.0 15 - 15 1.00
HCO3-(mmol/l) 23.2 ± 6.3 23.9 ± 7.0 5.1–44.0 <20 36 - 28 0.09
*P values were calculated using the McNemar test. -, not measured during transport; DBP, diastolic blood pressure; PCO2, partial carbon dioxide tension; PO2, partial oxygen tension; SBP, systolic blood pressure.
Table 5 Examples of recorded adverse events
Transfer characteristics Adverse event Severity (grade 1–3)a
Pulmonary embolus PaO2 on departure 4.2 kPa; not intubated; PaO2 on arrival 4.7 kPa 3
Oesophageal bleeding Only one peripheral intravenous line; no accompanying physician; active bleeding; PaO2 on arrival 6.7 kPa 3
Sepsis, rhabdomyolysis RI; shock on arrival 2
Imminent RI; Wegener's granulomatosis No blood pressure measured on the road (160 km); PaO2 on arrival 6.7 kPa, SaO2 86% 1
Pulmonary embolus No accompanying physician; RI on arrival 2
ARDS, MOF SaO2 93% at departure, 69% on arrival 3
Streptococcal pneumonia/sepsis; imminent RI Not intubated (despite advice); norepinephrine via peripheral intravenous line 3
Sleep apnoea syndrome; RI PaO2 on departure 6.9 kPa; during transport SaO2↓ 74% and cardiac ischaemia; no physician 3
Haemorragic shock; mechanical ventilation No accompanying physician; active bleeding (3 units packed cells on the road); oxygenation problems 3
Infectious endocarditis; mechanical ventilation No physician; haemodynamically unstable on the road 1
Septic shock; imminent RI Not intubated (despite advice); RI on arrival 3
Septic shock; MOF Norepinephrine via peripheral intravenous line 2
Suicide attempt (benzodiazepine) Deep coma; not intubated; apnoea en route; cyanotic on arrival 3
Postsurgical; mechanical ventilation Oxygen supply breakdown before arrival 3
COPD, pneumonia Shortage of oxygen before arrival 3
Haemodyalisis postsurgical No blood pressure measured on the road 1
Active bleeding digestive tract Only one peripheral intravenous line 1
ARDS; mechanical ventilation Ambulance breakdown; 40 min delay 1
aGrades of severity: grade 1 = deviation from guidelines/protocol; grade 2 = of vital importance – immediate action needed on arrival; and grade 3 = of vital importance – immediate action needed on arrival – event probably avoidable. ARDS, acute respiratory distress syndrome; MOF, multiple organ failure; PaO2, arterial oxygen tension; RI, respiratory insufficiency (imminent need for mechanical ventilation); SaO2, arterial oxygen saturation.
==== Refs
Waydhas C Schneck G Duswald KH Deterioration of respiratory function after intra-hospital transport of critically ill surgical patients Intensive Care Med 1995 21 784 789 8557864 10.1007/BF01700959
Kreeftenberg HG JrLigtenberg JJ Arnold LG van der Werf TS Tulleken JE Zijlstra JG Condition on arrival of transferred critically ill patients Neth J Med 2000 57 180 184 11063863 10.1016/S0300-2977(00)00055-3
Beckmann U Gillies DM Berenholtz SM Wu AW Pronovost P Incidents relating to the intra-hospital transfer of critically ill patients. An analysis of the reports submitted to the Australian Incident Monitoring Study in Intensive Care Intensive Care Med 2004 30 1579 1585 14991102 10.1007/s00134-004-2177-9
Hebert PC Wells G Blajchman MA Marshall J Martin C Pagliarello G Tweeddale M Schweitzer I Yetisir E A multicenter, randomized, controlled clinical trial of transfusion requirements in critical care (vol 340, pg 409, 1999) N Engl J Med 1999 340 1056 10.1056/NEJM199902113400601
Hebert PC Transfusion requirements in critical care (TRICC): a multicentre, randomized, controlled clinical study. Transfusion Requirements in Critical Care Investigators and the Canadian Critical care Trials Group Br J Anaesth 1998 81 Suppl 1 25 33 10318985
Escarce JJ Kelley MA Admission source to the medical intensive care unit predicts hospital death independent of APACHE II score JAMA 1990 264 2389 2394 2231994 10.1001/jama.264.18.2389
Durairaj L Will JG Torner JC Doebbeling BN Prognostic factors for mortality following interhospital transfers to the medical intensive care unit of a tertiary referral center Crit Care Med 2003 31 1981 1986 12847392 10.1097/01.CCM.0000069730.02769.16
Bellingan G Olivier T Batson S Webb A Comparison of a specialist retrieval team with current United Kingdom practice for the transport of critically ill patients Intensive Care Med 2000 26 740 744 10945392 10.1007/s001340051127
Gebremichael M Borg U Habashi NM Cottingham C Cunsolo L McCunn M Reynolds HN Interhospital transport of the extremely ill patient: the mobile intensive care unit Crit Care Med 2000 28 79 85 10667503 10.1097/00003246-200001000-00013
Manji M Bion JF Transporting critically ill patients Intensive Care Med 1995 21 781 783 8557863 10.1007/BF01700958
Lieshout van EJ Guideline for the transport of ICU patients NVIC Monitor 2001 6 22 25
Warren J Fromm RE JrOrr RA Rotello LC Horst HM Guidelines for the inter- and intrahospital transport of critically ill patients Crit Care Med 2004 32 256 262 14707589 10.1097/01.CCM.0000104917.39204.0A
Shirley PJ Bion JF Intra-hospital transport of critically ill patients: minimising risk Intensive Care Med 2004 30 1508 1510 15197442 10.1007/s00134-004-2293-6
|
16137359
|
PMC1269465
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jul 1; 9(4):R446-R451
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3749
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37541613736010.1186/cc3754ResearchIntra-abdominal hypertension in patients with severe acute pancreatitis De Waele Jan J [email protected] Eric 1Blot Stijn I 2Decruyenaere Johan 3Colardyn Francis 41 Intensivist, Intensive care unit, Ghent University Hospital, Gent, Belgium2 Professor, Intensive care unit, Ghent University Hospital, Gent, Belgium3 Professor and Head, Intensive care unit, Ghent University Hospital, Gent, Belgium4 Professor and Chief Executive Officer, Ghent University Hospital, Gent, Belgium2005 6 7 2005 9 4 R452 R457 25 3 2005 24 4 2005 3 6 2005 6 6 2005 Copyright © 2005 De Waele et al., licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is cited.
Introduction
Abdominal compartment syndrome has been described in patients with severe acute pancreatitis, but its clinical impact remains unclear. We therefore studied patient factors associated with the development of intra-abdominal hypertension (IAH), the incidence of organ failure associated with IAH, and the effect on outcome in patients with severe acute pancreatitis (SAP).
Methods
We studied all patients admitted to the intensive care unit (ICU) because of SAP in a 4 year period. The incidence of IAH (defined as intra-abdominal pressure ≥ 15 mmHg) was recorded. The occurrence of organ dysfunction during ICU stay was recorded, as was the length of stay in the ICU and outcome.
Results
The analysis included 44 patients, and IAP measurements were obtained from 27 patients. IAH was found in 21 patients (78%). The maximum IAP in these patients averaged 27 mmHg. APACHE II and Ranson scores on admission were higher in patients who developed IAH. The incidence of organ dysfunction was high in patients with IAH: respiratory failure 95%, cardiovascular failure 91%, and renal failure 86%. Mortality in the patients with IAH was not significantly higher compared to patients without IAH (38% versus 16%, p = 0.63), but patients with IAH stayed significantly longer in the ICU and in the hospital. Four patients underwent abdominal decompression because of abdominal compartment syndrome, three of whom died in the early postoperative course.
Conclusion
IAH is a frequent finding in patients admitted to the ICU because of SAP, and is associated with a high occurrence rate of organ dysfunction. Mortality is high in patients with IAH, and because the direct causal relationship between IAH and organ dysfunction is not proven in patients with SAP, surgical decompression should not routinely be performed.
==== Body
Introduction
Despite recent advances in the management of patients, such as early enteral nutrition and withholding surgery until proven infection of pancreatic necrosis, severe acute pancreatitis (SAP) remains a disease with an unpredictable clinical course and significant morbidity and mortality [1]. Infection still remains the most feared complication, but also the presence of organ dysfunction is increasingly recognized as an important risk factor for mortality in patients with severe disease [2-4].
Intra-abdominal hypertension (IAH) has been recognized as a cause of organ dysfunction in critically ill patients, with respiratory and renal dysfunction often most prominent [5]. This syndrome, referred to as the abdominal compartment syndrome, has most extensively been described in patients who underwent emergency abdominal surgery or after abdominal trauma, but also in patients with non-abdominal diseases such as burns [6] and massive fluid resuscitation [7].
Some recent studies [8,9] suggest that IAH is a frequent finding in SAP patients. The clinical relevance of this remains unclear, although Pupelis et al. [9] found a relation between elevated intra-abdominal pressure (IAP; above 25 mmHg) and persistent subsequent organ dysfunction. Tao et al. [10] described a high incidence of IAH in patients with early SAP, but lack of a definition of IAH and methodological issues make interpretation of these data difficult [10].
The levels at which elevated IAP can cause organ dysfunction are lower than in the study by Pupelis et al. Values of as low as 15 mmHg may result in clinically significant organ damage [11-13], but clinical significance of this lower threshold in patients with SAP remains to be determined.
The aim of this analysis was to study patient factors associated with the development of IAH. Furthermore, we studied the incidence of organ failure in patients with SAP and IAH, and the association of the presence of IAH and outcome.
Materials and methods
Patients
We studied all patients admitted because of SAP to the intensive care unit (ICU) of the Ghent University Hospital (Gent, Belgium) between January 2000 and March 2004. SAP was defined according to the criteria described by the International Symposium on Acute Pancreatitis [14]. Patients names were retrieved from the hospital registry using ICD code 577.0 (acute pancreatitis), and files were reviewed retrospectively. Patients who were referred from other hospitals later than 7 days after the start of SAP were excluded. The study was approved by the local ethical committee.
Preoperative data collected included age, gender, etiology of SAP, C-reactive protein level, Ranson score and Acute Physiology And Chronic Health Evaluation (APACHE) II score [15] on admission and C-reactive protein at 48 h after admission.
Data acquisition
IAP values were measured every 8 h when IAP was below 15 mmHg, and every 4 h when above 15 mmHg, and were retrieved from the patients file. IAP was measured using the transvesical route, as described by Cheatham et al. [16], after instillation of 50 ml of saline in the bladder. IAP measurements were obtained from patients when multiple intra-abdominal fluid collections were present on CT scan on admission, or when there was the clinical suspicion of IAH. These clinical indications included oliguria, hypoxia, abdominal distension, and severe abdominal pain. The incidence of IAH (defined as IAP ≥ 15 mmHg) was recorded, as was the maximal IAP value obtained during ICU stay, and the duration of IAP levels ≥ 15 mmHg.
The occurrence rate of organ dysfunction during ICU stay was recorded and defined as: cardiovascular, hypotension requiring vasoactive medication (epinephrine, norepinephrine, dobutamine at any dose, or dopamine at doses above 2 mcg/kg/min); renal, serum creatinine above 2.0 mg/dl; pulmonary, the need for mechanical ventilation or PaO2/FiO2 ratio < 300. Mortality was defined as in-hospital mortality.
Interventions to alleviate IAH were recorded, as were complications of these interventions. Decompressive laparotomy was considered when rapidly deteriorating, therapy resistant multiple organ dysfunction was present in the first days after admission, and decided on a patient to patient basis.
Statistical analysis
Statistical analysis was performed using SPSS for Windows 11.0.1 ® (SPSS, Chicago, IL, USA). Continuous variables were compared using the Mann Whitney U-test. Categorical data were compared using the Chi-square or Fisher Exact test. Continuous data are expressed as mean (standard deviation) if the data were normally distributed, or median (interquartile range) if the distribution was not normal. Categorical data are reported as n (%). Pearson correlation coefficient between maximal IAP and APACHE II score was calculated. Mean IAP values from day 1 to 7 were compared using the Friedman test. A double sided p-value of less than 0.05 was considered statistically significant.
Results
General
Forty-four patients were admitted to the ICU because of SAP during the study period. Mean age was 57 years (15.8) and 27 were male (61%). The etiology of acute pancreatitis was biliary tract stones in 19 patients, alcohol intake in 12, hyperlipemia in 4, and trauma in 2. In 7 patients, the cause of pancreatitis could not be determined. Mean Ranson score of the patients was 5.5 (2.6), mean APACHE II score was 18 (9.2).
IAP monitoring
IAP measurements were obtained from 27 patients, but in the remaining 17 patients IAP was not measured. Of the 27 patients, 21 developed IAH (78%). In 12 patients, IAP monitoring was available from the first day of admission to the ICU. In these 12 patients, IAH developed after a median of 1 day after admission to the hospital and mean IAP increased from 16 at day 1 to 22 mmHg the day after, and remained elevated. There was a trend towards a significant difference between the mean IAP values during the first week of admission (p = 0.12) (Fig. 1).
The maximum IAP in patients with IAH averaged 27 (7.8) mmHg. In patients who did not undergo abdominal decompression (n = 17), IAH persisted for a median of 6 days (interquartile range 3–8). Maximal IAP correlated significantly with APACHE II score (correlation coefficient 0.60, p < 0.002) (Fig. 2).
Factors associated with IAH
In univariate analysis, the APACHE II and Ranson scores on admission were higher in patients who developed IAH (Table 1). Age, gender, cause of pancreatitis and C-reactive protein levels at 48 h were not significantly different between the two groups. Pancreatic necrosis was documented in all but one patient who developed IAH, whereas only three patients without IAH had pancreatic necrosis; the other three patients had pancreatic oedema and peripancreatic fluid collections on CT scan.
Organ dysfunction, surgical interventions and outcome
The incidence of organ dysfunction was higher in patients with IAH compared to patients without IAH (Table 1). Thirteen patients with IAH were treated with renal replacement therapy compared to none in the patients without IAH. Duration of mechanical ventilation was maintained for 15 (12.6) days in patients with IAH.
Surgical treatment was more frequent in patients with IAH. Of the 21 patients with IAH, 9 were treated surgically, whereas no patient in the non-IAH group needed surgery (p = 0.07). The indication for surgery was abdominal compartment syndrome in four patients and infected pancreatic necrosis in five patients. Abdominal decompression was performed surgically through a midline laparotomy. In four patients, a temporary abdominal closure system was used because of abdominal compartment syndrome with IAP ranging from 25 to 45 mmHg. IAP decreased in all patients (Fig. 3). In one patient, necrosectomy was performed at the time of decompression. Three of these patients died early in the postoperative course. The cause of death was uncontrollable retroperitoneal bleeding in two patients, and further deterioration of organ dysfunction in another patient. Patients with IAH stayed significantly longer in the ICU and in the hospital than patients without IAH (Table 1).
Mortality in the patients with IAH was not significantly higher than in patients without IAH (8/21 (38%) versus 1/6 (16%), p = 0.63). The non-IAH patient died after therapy was withdrawn early after ICU admission because of a concomitant advanced brain tumour. Four patients with IAH died within 4 days after hospital admission, three of them within 24 h after surgical decompression. The four other patients with IAH with fatal outcomes died on day 12, 26, 35 and 38 because of persistent organ dysfunction, in association with infected pancreatic necrosis in three patients.
Discussion
In this cohort of patients admitted to the ICU because of SAP, the incidence of IAH was 51%. When only patients in who IAP monitoring was performed are considered, the incidence reached 78%, but this might be an overestimation as IAP measurement was not performed routinely and was based upon clinical suspicion for IAH. Also, IAH developed early in the course of the disease; in the majority of the patients in whom IAP monitoring was available from the day of admission, IAH developed within 24 h after ICU admission.
Although the difference in IAP during the first week was not significant, there seem to be three time frames early in the course of the disease. At day 1 the IAP was already elevated, and it then increased to the maximal level at day 2 and remained elevated until day four after admission. IAH in patients with SAP seems to be an early event.
Maximal IAP values were well above the 15 mmHg threshold used for the definition of IAH, and were as high as 25 to 40 mmHg in some patients, including the four patients who underwent abdominal decompression for abdominal compartment syndrome. These high values of IAP may be an explanation for the high incidence of organ failure in these patients, as all patients with IAH developed at least one organ failure, and the majority two or more.
SAP patients develop IAH for several reasons. Pancreatic or retroperitoneal inflammation is the most obvious reason in the early course of the disease. Aggressive fluid resuscitation, resulting in generalized and visceral odema in particular, will add to the intra-abdominal volume during the first days of severe disease. Furthermore, paralytic ileus and peripancreatic acute fluid collections can also increase IAP.
From the APACHE II and Ranson scores of the patients, it seems that the more severe the disease, the higher the likelihood to develop IAH. But IAH itself may be an early predictor of severe disease, as elevated IAP seems to occur early in the course of the disease. IAH may even contribute to disease severity in patients with SAP, but the exact role remains to be determined. Elevated IAP causes intestinal hypoperfusion even at levels as low as 8 to 12 mmHg [12]. In the setting of SAP, pancreatic perfusion may also be affected, and possibly IAH may contribute to the development of pancreatic hypoperfusion and eventually pancreatic necrosis. The observation of increased bacterial translocation in patients with IAH and abdominal compartment syndrome [17] may also apply to patients with SAP. Animal studies have shown an increased rate of bacterial translocation in acute pancreatitis [18], but the role of IAH in this remains to be elucidated.
Patients with IAH had necrosis more often and were operated on more often. This resulted in a longer ICU and hospital stay for these patients. Surgical decompression was performed in four patients with IAP levels above 25 mmHg and severe organ dysfunction, but only one patient survived. The three other patients succumbed early after decompression, two patients from hemorrhagic shock and one from further deteriorating multiple organ dysfunction syndrome. The necrosectomy that was performed in the first patient treated with abdominal decompression possibly played a role in the hemorrhagic shock and deterioration early after surgery. Necrosectomy was not applied to subsequent patients who underwent decompression. In the second patient who died of uncontrollable bleeding from the retroperitoneum, the bleeding itself may have played a role in the development of IAH. At laparotomy, there was a large retroperitoneal haematoma, with active bleeding, possibly caused by an eroded vessel or pseudoaneurysm. Due to the profound bleeding, no clear cause could be identified and, unfortunately, the family of the patient refused a post mortem examination.
This experience in our four patients has tempered our initial enthusiasm for decompression in patients with IAH and SAP [19]. Other authors also reported poor survival rates after surgical decompression in patients with SAP [8]. Patient selection may, however, bias the results of decompression, as only patients with uncontrollable organ dysfunction have been considered candidates for decompression in our unit, and also the timing of surgical decompression may play a crucial role.
There has been a recent trend towards postponing surgery in patients with SAP because early surgical intervention was associated with an increased mortality rate [20,21]. This could also be concluded from our limited number of patients who died shortly after surgery, but it should be considered that the strategy of early intervention in SAP without infection, where the retroperitoneum is debrided, differs substantially from a procedure in which the abdomen is opened, but the retroperitoneum is left untouched. Moreover, in one of the patients that was decompressed and debrided in our study, an uncontrollable haemorrhage from the retroperitoneum occurred, and the patient died a few hours later.
Little can be concluded from this study as to the usefulness of early debridement but, in our experience, the absence of infection, increased age and acute renal failure were associated with an increased mortality in a series of patients who were treated surgically for severe acute pancreatitis; the timing of the surgical intervention itself had no effect on this [22].
Conclusion
Severity of disease predisposes for IAH in patients with SAP. The occurrence rate of IAH is high, and IAH is associated with organ dysfunction in the majority of patients. Mortality is high in patients with IAH, but it is not clear if surgical decompression in these patients is advantageous.
Key messages
• IAH is a frequent finding in critically ill patients with SAP
• The maximum IAP is related to the severity of illness
• Organ dysfunction is present in patients with moderately increased IAPs
• Abdominal decompression was associated with a 75% mortality in this study
Abbreviations
APACHE = Acute Physiology And Chronic Health Evaluation; IAH = intra-abdominal hypertension; IAP = intra-abdominal pressure; ICU = intensive care unit; SAP = severe acute pancreatitis.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
JDW and EH conceived and designed the study. Acquisition of a substantial portion of data was done by JDW. Analysis and interpretation of data was performed by JDW, EH and SB. JDW and SB drafted the manuscript. FC, JDC and EH critically revised the manuscript for important intellectual content. EH and SB supplied statistical expertise. FC supervised and was responsible overall for all aspects of the study.
Acknowledgements
This study was supported by a clinical doctoral grant of the Fund for Scientific Research, Flanders, Belgium (FWO, Vlaanderen).
Figures and Tables
Figure 1 Evolution of intra-abdominal pressure (IAP) in the first week after admission. Mean IAP with 95% confidence interval (CI).
Figure 2 Correlation between maximal intra-abdominal pressure and APACHE II score in patients with severe acute pancreatitis.
Figure 3 Effect of surgical decompression on intraabdominal pressure.
Table 1 Characteristics of patients who did or did not develop intra-abdominal hypertension during ICU stay (n = 27)
Characteristic IAH (n = 21) Non-IAH (n = 6) p-value
Age 53 (45–68) 46 (26–76) 0.629
Male gender 15 (71%) 4 (67%) 1.000
APACHE II score 21 (15–28) 10 (8–11) 0.005
Ranson score 7 (6–8) 3 (1–5) 0.014
Etiology of pancreatitis 0.552
Biliary 7 (33%) 4 (67%)
Alcohol 8 (38%) 2 (33%)
Hyperlipemia 3 (14%)
Trauma 1 (5%)
Unknown 2 (10%)
CRP level 48 h after admission (mg/dL) 34 (19–40) 34 (26–39) 0.521
Pancreatic necrosis 20 (95%) 3 (50%) 0.025
Surgical management 9 (43%) 0 (0%) 0.070
Infected pancreatic necrosis 5 (24%) 0 (0%) 0.555
Organ dysfunction
Pulmonary failure 20 (95%) 2 (33%) 0.004
Cardiovascular failure 19 (91%) 1 (17%) 0.001
Renal failure 18 (86%) 1 (17%) 0.004
LOS ICU (days) 21 (10–37) 3 (1–5) 0.003
LOS hospital (days) 42 (20–90) 12 (3–14) 0.015
APACHE, Acute Physiology And Chronic Health Evaluation; CRP, C-reactive protein; IAH, intra-abdominal hypertension; ICU, intensive care unit; LOS, length of stay.
==== Refs
Wilmer A ICU management of severe acute pancreatitis Eur J Intern Med 2004 15 274 280 15450983 10.1016/j.ejim.2004.06.004
Khan AA Parekh D Cho Y Ruiz R Selby RR Jabbour N Genyk YS Mateo R Improved prediction of outcome in patients with severe acute pancreatitis by the APACHE II score at 48 hours after hospital admission compared with the APACHE II score at admission Arch Surg 2002 137 1136 1140 12361419 10.1001/archsurg.137.10.1136
Buter A Imrie CW Carter CR Evans S McKay CJ Dynamic nature of early organ dysfunction determines outcome in acute pancreatitis Br J Surg 2002 89 298 302 11872053 10.1046/j.0007-1323.2001.02025.x
Dugernier T Reynaert M Laterre PF Early multi-system organ failure associated with acute pancreatitis: a plea for a conservative therapeutic strategy Acta Gastroenterol Belg 2003 66 177 183 12891929
Sugrue M Jones F Deane SA Bishop G Bauman A Hillman K Intra-abdominal hypertension is an independent cause of postoperative renal impairment Arch Surg 1999 134 1082 1085 10522851 10.1001/archsurg.134.10.1082
Wittman DH Iskander GA The compartment syndrome of the abdominal cavity: a state of the art review J Intensive Care Med 2000 15 201 220 10.1046/j.1525-1489.2000.00201.x
Balogh Z McKinley BA Cocanour CS Kozar RA Valdivia A Sailors RM Moore FA Supranormal trauma resuscitation causes more cases of abdominal compartment syndrome Arch Surg 2003 138 637 642 12799335 10.1001/archsurg.138.6.637
Gecelter G Fahoum B Gardezi S Schein M Abdominal compartment syndrome in severe acute pancreatitis: an indication for a decompressing laparotomy? Dig Surg 2002 19 402 404 discussion 404–405 12435913 10.1159/000065820
Pupelis G Austrums E Snippe K Berzins M Clinical significance of increased intraabdominal pressure in severe acute pancreatitis Acta Chir Belg 2002 102 71 74 12051093
Tao HQ Zhang JX Zou SC Clinical characteristics and management of patients with early acute severe pancreatitis: experience from a medical center in China World J Gastroenterol 2004 10 919 921 15040047
Mertens zur Borg I Lim A Verbrugge SJC Ijzermans JNM Klein J Effect of intraabdominal pressure elevation and positioning on hemodynamic responses during carbon dioxide pneumoperitoneum for laparoscopic donor nephrectomy: A prospective controlled clinical study Surg Endosc 2004 18 919 923 15108115 10.1007/s00464-003-8817-2
Schwarte LA Scheeren TW Lorenz C De Bruyne F Fournell A Moderate increase in intraabdominal pressure attenuates gastric mucosal oxygen saturation in patients undergoing laparoscopy Anesthesiology 2004 100 1081 1087 15114204 10.1097/00000542-200405000-00009
Malbrain ML Is it wise not to think about intraabdominal hypertension in the ICU? Curr Opin Crit Care 2004 10 132 145 15075724 10.1097/00075198-200404000-00010
Bradley EL 3rd A clinically based classification system for acute pancreatitis. Summary of the International Symposium on Acute Pancreatitis, Atlanta, GA, September 11 through 13, 1992 Arch Surg 1993 128 586 590 8489394
Knaus WA Draper EA Wagner DP Zimmerman JE APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Cheatham ML Safcsak K Intraabdominal pressure: a revised method for measurement J Am Coll Surg 1998 186 594 595 9583702 10.1016/S1072-7515(98)00122-7
Doty JM Oda J Ivatury RR Blocher CR Christie GE Yelon JA Sugerman HJ The effects of hemodynamic shock and increased intra-abdominal pressure on bacterial translocation J Trauma 2002 52 13 17 11791046
Cicalese L Sahai A Sileri P Rastellini C Subbotin V Ford H Lee K Acute pancreatitis and bacterial translocation Dig Dis Sci 2001 46 1127 1132 11341659 10.1023/A:1010786701289
De Waele JJ Hesse UJ Life saving abdominal decompression in a patient with severe acute pancreatitis Acta Chir Belg 2005 105 96 98 15790212
Hartwig W Maksan SM Foitzik T Schmidt J Herfarth C Klar E Reduction in mortality with delayed surgical therapy of severe pancreatitis J Gastrointest Surg 2002 6 481 487 12023003 10.1016/S1091-255X(02)00008-2
Mier J Leon EL Castillo A Robledo F Blanco R Early versus late necrosectomy in severe necrotizing pancreatitis Am J Surg 1997 173 71 75 9074366 10.1016/S0002-9610(96)00425-4
De Waele JJ Hoste E Blot SI Hesse U Pattyn P de Hemptinne B Decruyenaere J Vogelaers D Colardyn F Perioperative factors determine outcome after surgery for severe acute pancreatitis Crit Care 2004 8 R504 511 15566598 10.1186/cc2991
|
16137360
|
PMC1269467
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jul 6; 9(4):R452-R457
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3754
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37601613736210.1186/cc3760ResearchValidation of a method to partition the base deficit in meningococcal sepsis: a retrospective study O'Dell Ellen [email protected] Shane M [email protected] Andrew [email protected] Jo 3Murdoch Ian A [email protected] Fellow, Department of Paediatric Intensive Care, Guy's and Saint Thomas' Hospitals, London, UK2 Consultant, Department of Paediatric Intensive Care, Guy's and Saint Thomas' Hospitals, London, UK3 Resident, Department of Paediatric Intensive Care, Guy's and Saint Thomas' Hospitals, London, UK4 Consultant, Department of Paediatric Intensive Care, Guy's and Saint Thomas' Hospitals, London, UK2005 8 7 2005 9 4 R464 R470 26 2 2005 14 4 2005 18 5 2005 10 6 2005 Copyright © 2005 O'Dell et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
The base deficit is a useful tool for quantifying total acid–base derangement, but cannot differentiate between various aetiologies. The Stewart–Fencl equations for strong ions and albumin have recently been abbreviated; we hypothesised that the abbreviated equations could be applied to the base deficit, thus partitioning this parameter into three components (the residual being the contribution from unmeasured anions).
Methods
The two abbreviated equations were applied retrospectively to blood gas and chemistry results in 374 samples from a cohort of 60 children with meningococcal septic shock (mean pH 7.31, mean base deficit -7.4 meq/L). Partitioning required the simultaneous measurement of plasma sodium, chloride, albumin and blood gas analysis.
Results
After partitioning for the effect of chloride and albumin, the residual base deficit was closely associated with unmeasured anions derived from the full Stewart–Fencl equations (r2 = 0.83, y = 1.99 – 0.87x, standard error of the estimate = 2.29 meq/L). Hypoalbuminaemia was a common finding; partitioning revealed that this produced a relatively consistent alkalinising effect on the base deficit (effect +2.9 ± 2.2 meq/L (mean ± SD)). The chloride effect was variable, producing both acidification and alkalinisation in approximately equal proportions (50% and 43%, respectively); furthermore the magnitude of this effect was substantial in some patients (SD ± 5.0 meq/L).
Conclusion
It is now possible to partition the base deficit at the bedside with enough accuracy to permit clinical use. This provides valuable information on the aetiology of acid–base disturbance when applied to a cohort of children with meningococcal sepsis.
==== Body
Introduction
Metabolic acidosis is a common biochemical finding in critically ill patients [1]. The prognostic significance of this entity is recognised in many mortality risk scores, in which the predicted risk increases in proportion to the degree of acidosis [2-4]. The commonest bedside tool for quantifying a metabolic acidosis is the base deficit [5]. Although the base deficit is an accurate measure of total acute acid–base derangement, it cannot delineate the various aetiologies that can contribute to an acidosis [6,7]. Broadly speaking, these include tissue acids (which dissociate into lactate and other, unmeasured anions), hyperchloraemia (a 'normal anion gap' acidosis), and weak acids (traditionally known as buffers, of which albumin is the most important). It is not uncommon for the three aetiologies to coexist in the critically ill patient; furthermore, the relative contribution from each can vary with time [8,9]. The cause, treatment, and perhaps prognostic significance of each of these aetiologies differ; a tool to partition the base deficit for each component would therefore be useful [10].
Recent insights into acid–base physiology (the Stewart–Fencl approach) have provided a method of quantifying each component of the acid–base status [6,7,11]. However, the necessary physicochemical calculations are cumbersome and require the simultaneous measurement of many biochemical variables. Two abbreviated versions of the Stewart–Fencl equations have recently been derived: one for albumin, the other for chloride [12,13]. We hypothesised that, by applying these to the base deficit, the residual would reflect the acidifying effect of unmeasured anions, thus partitioning the base deficit into its three components. Our secondary hypothesis was that the loss of accuracy as a consequence of applying these abbreviated formulae to the base deficit would not be great enough to compromise clinical validity. We have investigated this retrospectively in a cohort of 60 children with meningococcal septic shock. This patient group was chosen for two reasons: metabolic acidosis is a common occurrence in itself, and so are derangements in all three components contributing to the acidosis.
Methods
The study was approved by the Institutional Ethics Committee, which waived the need for informed consent.
Patients
We examined data retrospectively from 68 consecutive patients with meningococcal sepsis admitted to the paediatric intensive care unit from January 2001 to June 2003. Cases were identified from the departmental database. Patients with meningococcal meningitis without septic shock were excluded. Septic shock was defined as the need for more than 40 ml/kg of fluid resuscitation within 4 hours of presentation to hospital or the requirement for inotropic medication [14]. All blood samples taken during the first 72 hours of admission, in which a full chemistry profile was measured simultaneously with arterial blood gas analysis, were analysed.
After exclusion of those without septic shock, full data were available for 374 blood samples from 60 patients (giving a median of six samples per patient). Patient demographics were as follows: median (interquartile) age 2 years (0.8 to 9.5), weight 13 kg (10 to 19), Paediatric Index of Mortality version 2 (PIM2)-derived mortality risk 11.0% (6 to 16), crude mortality 10.0% (PIM2-predicted death rate 13.8%). In addition, 88% of patients required mechanical ventilation, 82% received inotropes, and the amount of fluid administered in the first 24 hours after admission was 158 ± 65 ml/kg (mean ± SD).
Blood chemistry analysis
Arterial blood gases and chemistry were measured with the Instrumentation Laboratory 1640 blood gas analyser (Lexington, MA, USA) and Synchron LX20 (Beckman Coulter Inc., High Wycombe, Buckinghamshire, UK) The total base deficit (BDtot) was calculated according to the algorithm for 'standard base deficit' in the blood gas analyser, which necessitated the concurrent measurement of haemoglobin. Plasma albumin was measured by binding of bromocresol green dye, and whole blood lactate by the enzymatic method (YSI 2300 STAT plus analyser; Yellow Springs Instruments, Yellow Springs, OH, USA). The precision values for the above analysers were as follows: pH 0.009 to 0.005 (at pH values of 7.15 and 7.66 respectively), pCO2 2.74 to 2.78%, ion-specific electrodes all less than 2%, albumin 6.2% and 3.0% (at albumin concentrations of 13 and 37 g/L, respectively), and lactate 2.0%.
Formulae to partition the base deficit
BDtot is influenced by three factors: weak acids, of which albumin is dominant (BDalb); strong ions, of which chloride concentration (relative to sodium) is the most important (BDcl); and net unmeasured anions from tissue acids (BDUMA). Blood lactate can be considered as either a strong ion (if measured) or an unmeasured anion (if unmeasured). For the purposes of this study we designated lactate as an unmeasured anion.
These three factors can exert an acidifying (hyperalbuminaemia, hyperchloraemia, excess unmeasured anions) or an alkalinising (hypoalbuminaemia, hypochloraemia, excess unmeasured cations) effect on the total base deficit, according to their net charge [6,7,11].
If formulae quantifying the effect of albumin and chloride on the base deficit are applied and subtracted from the total base deficit, the remainder should equal the contribution from net unmeasured anions (or cations if the residual base deficit is positive), so that
BDtot - BDalb - BDCl = BDUMA
If this method is accurate, the residual (BDUMA) should therefore approximate unmeasured anions calculated from the Stewart–Fencl equations. After considering the precision of the blood gas and chemistry analysers, the expected loss of accuracy due to the abbreviated nature of the base deficit equations, and the normal value for the Stewart–Fencl strong ion gap (see below), we set an a priori limit of 3 meq/L for the standard error of the estimate (SEE).
The formulae for base deficit used in this study have been derived elsewhere [12,13], and are as follows:
BDalb = [42 - albumin (g/L)] × 0.25
BDCl = [Na+] - [Cl-] - 32
A full explanation of the Stewart–Fencl methodology is reported elsewhere [6,7,11,15]; however, a brief explanation is included in Additional file 1. The equations are as follows:
Unmeasured anions (strong ion gap) = strong ion difference (measured) - strong ion difference (effective)
Strong ion difference (measured) = [Na+ + K+ + Ca2+ + Mg2+] - [Cl- ] (all in meq/L)
Strong ion difference (effective) = (12.2 × pCO2/(10-pH)) + 10 × [Alb] (g/L) × (0.123 × pH - 0.631) + [PO42-] (mmol/L) × (0.309 × pH - 0.469).
Statistical analysis
Data are reported as means and SD. Agreement between unmeasured anions calculated from the base deficit (BDUMA) and Stewart–Fencl methods (strong ion gap) was assessed by linear regression with the use of the ordinary least-squares method (Microsoft Excel).
Results
Acid–base and biochemical results are shown in Table 1. A significant metabolic acidosis was seen for the group as a whole (mean pH 7.31, BDtot -7.4). The unmeasured anion-related base deficit was greater than the total base deficit; this was predominantly due to the alkalinising effect of hypoalbuminaemia (mean albumin effect on base deficit +2.9; Table 1). This is also shown in the histogram for BDalb (Fig. 1a), showing an alkalinising effect in 91% of samples. The influence of chloride was variable, producing both acidifying and alkalinising effects in approximately equal proportions (50% and 43%, respectively; Fig. 1b). It is also notable that the range of chloride effect on base deficit was more extreme than that for albumin (SD 5.0 versus 2.2; Table 1 and Fig. 1).
Not surprisingly, BDtot was weakly associated with Stewart–Fencl unmeasured anions (r2 = 0.27; Fig. 2a). However, after adjustment for chloride and albumin, BDUMA showed a strong, linear association with Stewart–Fencl unmeasured anions (r2 = 0.83; Fig. 2b). The full regression equation was Stewart–Fencl-derived UMA = 1.99 - (0.87 × BDUMA), SEE 2.29 meq/L.
Finally, it is noted that the regression analysis used multiple samples taken from each patient; thus each measurement cannot be considered truly independent in a statistical sense (even though an individual patient's base deficit may have changed markedly over the 72 hours after admission). In this setting, multiple measurements taken on an atypical patient could potentially bias the regression analysis. To investigate this we reanalysed the data in two ways. First, a standardised residual plot was inspected, which did not reveal obvious deviation from normality among the residuals, nor any extreme outliers. Second, we repeated the regression analysis, using one measurement only per patient (n = 60). Measurements were chosen by means of the random number generator in Excel using a uniform distribution, whereby the sample with the largest assigned random number from each patient's group of samples was chosen. This produced remarkably similar results: r2 = 0.854, Stewart–Fencl-derived UMA = 2.39 - (0.871 × BDUMA), SEE 2.17 meq/L. We therefore conclude that the above approach is valid.
Discussion
These findings demonstrate that the base deficit can be partitioned at the bedside by the application of two simple formulae, requiring the measurement of plasma sodium, chloride and albumin concurrently with the arterial blood gas. This was validated by comparing the unmeasured anion portion of the base deficit with that calculated from the Stewart–Fencl equations, yielding a high coefficient of determination (r2 = 0.83). However, to assess whether this model is accurate enough for clinical use, we must consider three other aspects of the regression analysis, namely the SEE (2.29 meq/L), the slope (-0.87) and the intercept (1.99 meq/L).
Inspection of the residual plots (data not shown) did not reveal an unusual pattern; furthermore, the residuals seemed normally distributed with consistent variance. Thus we can say that about 95% of the time the true unmeasured anions will lie within ± 4.5 meq/L of that estimated by the partitioned base deficit (1.96 × standard error).
The sources of this error are threefold, including both the abbreviated albumin and chloride equations and the fact that phosphate, the other major weak acid, is not accounted for. Albumin charge is a function of pH [7,11,16,17]; the albumin error will therefore increase with the degree of acidosis (both respiratory and metabolic).
However, this effect is not large; Story has estimated a typical error to be less than ± 1 meq/L [12], which is confirmed in the present study (data not shown). The abbreviated chloride equation does not account for the effect of variation in other cations (including potassium, calcium and magnesium). Lastly, omitting phosphate from the base deficit equations results in an error in unmeasured anion estimation of 1.8 meq/L for every 1 mmol/L change in phosphate concentration (again, this will alter slightly depending upon pH).
The slope of the regression equation (-0.87) can be interpreted as meaning that a decrease in base deficit of -10 meql/L will represent an increase of 8.7 meq/L in unmeasured anions; in essence this is close enough to be considered as an inverse equimolar relationship. It is also notable that the intercept, which occurs when the base deficit is equal to zero, yields an estimated unmeasured anion value of 2 meq/L, which is consistent with the normal value for this parameter [17].
In summary, we feel that the properties described above permit bedside partitioning of the base deficit, provided that the user is aware of the limitations and sources of error. Partitioning provides the clinician with valuable information about the aetiology of an acidosis, which can have implications for treatment and prognosis. Examples are outlined below.
Underestimation of tissue acidosis: critically ill patients have a high incidence of hypoalbuminaemia that produces an alkalinising effect, masking the true degree of 'tissue acidosis' [10,17,18]. In addition, several authors have documented relative hypochloraemia when tissue acidosis occurs, postulating that this represents a compensatory mechanism [8,19]. Neither phenomenon will be apparent from an unpartitioned base deficit.
Recognition of iatrogenic causes of an acidosis: the use of albumin solutions for resuscitation is common in paediatrics, and may become more widespread in adult practice since the publication of recent safety data [20]. Albumin-based fluids can propagate an acidosis by two mechanisms: they increase plasma albumin concentration, and most contain an abundant source of chloride. The latter mechanism is common to any fluid containing a high concentration of chloride relative to sodium (for example 0.9% saline) [21-25]. Persistent acidosis in this setting might be interpreted erroneously as being due to tissue hypoperfusion from inadequate resuscitation, provoking an unnecessary escalation of therapy.
Prognosis: the prognostic value of an acidosis related to unmeasured anions is uncertain, with studies producing conflicting results [26-30]; this may be due to the variable aetiology and composition of unmeasured anions (such as ketones, organic acids, sulphate and acetate). Conversely, several studies have suggested that a hyperchloraemic acidosis may carry a more favourable prognosis [27,28,31].
A potential criticism of the partitioning approach is that it may offer the same information as the anion gap. This is partly true, provided that the anion gap is corrected for albumin [17,18]. However, the anion gap alone cannot diagnose a mixed acidosis (unmeasured anion plus hyperchloraemic), which occurs frequently in critically ill patients. By partitioning the base deficit, we are in effect combining these two parameters into a single measurement that contains both quantitative and qualitative information.
This study did not attempt to address the role of lactate, but merely sought to validate a method for partitioning the base deficit. The prognostic and therapeutic value of lactate measurement is well established [27,32,33]; this anion is routinely measured as a point-of-care test in many critical care units. We suggest that lactate measurement is complementary to the partitioned base deficit approach, providing a method of further subdividing BDUMA into lactate and non-lactate components. This is important, because the two components are not tightly correlated [8,9].
Conclusion
In summary, we have validated two simple equations that permit partitioning of the base deficit into three components (chloride, albumin and unmeasured anions), providing for a more detailed bedside analysis of acid–base disturbances. We have found this to be useful in everyday clinical practice.
Key messages
• It is possible, by application of two simple equations, to partition the base deficit into three components: chloride, albumin and unmeasured anions.
• This requires simultaneous measurement of an arterial blood gas, and venous plasma sodium, chloride and albumin.
• Agreement between unmeasured anions calculated from the partitioned base deficit and from the full Stewart–Fencl equations produces good agreement (r2 = 0.83) in a cohort of patients with meningococcal sepsis.
• The partitioned base deficit reveals a predominantly alkalinising effect of albumin in this group (effect +2.9 ± 2.2 meq/L (mean ± SD)).
• The effect of chloride on the base deficit was more variable, producing significant acidifying and alkalinising effects in almost equal measure (effect -0.5 ± 5.0 meq/L (mean ± SD)).
Abbreviations
BDalb = base deficit due to albumin; BDCl = base deficit due to chloride; BDtot = total base deficit; BDUMA = base deficit due to unmeasured anions; PIM2 = Paediatric Index of Mortality version 2; SEE = standard error of the estimate.
BDalb = base deficit due to albumin; BDCl = base deficit due to chloride; BDtot = total base deficit; BDUMA = base deficit due to unmeasured anions; PIM2 = Paediatric Index of Mortality version 2; SEE = standard error of the estimate.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
EOD performed data collection, preliminary data analysis and co-wrote the first draft of the manuscript. SMT conceived the study, performed data analysis and co-wrote the first draft of the manuscript. AD participated in the design of the study, derived one of the base deficit formulae and advised on data analysis. JA performed data collection. IAM supervised the project and participated in study design. All authors read and approved the final manuscript.
Supplementary Material
Additional File 1
A Word document describing Stewart's physiochemical approach to acid–base balance.
Click here for file
Figures and Tables
Figure 1 Histograms demonstrating the effect of (a) albumin and (b) chloride on total base deficit for all blood samples (n = 374).
Figure 2 Scatter plots showing relationship between Stewart–Fencl derived unmeasured anions and base deficit. (a) unpartitioned (total) base deficit and (b) partitioned (unmeasured anion fraction) base deficit.
Table 1 Acid–base and biochemical parameters for all samples (n = 374)
Variable Unit Mean SD
PH - 7.31 0.09
pCO2 Kpa 4.6 1.1
BDtot meq/L -7.4 4.7
BDalb meq/L +2.9 2.2
BDCl meq/L +0.16 5.0
BDUMA meq/L -10.0 5.7
Sodium meq/L 140 5.1
Chloride meq/L 109 6.9
Albumin g/L 30.5 8.8
Lactate meq/L 2.5 2.3
Strong ion gap meq/L 10.7 5.5
BDalb, base deficit due to albumin; BDCl, base deficit due to chloride; BDtot, total base deficit; BDUMA, base deficit due to unmeasured anions.
==== Refs
Gauthier PM Szerlip HM Metabolic acidosis in the intensive care unit Crit Care Clin 2002 18 289 308 12053835 10.1016/S0749-0704(01)00012-4
Slater A Shann F Pearson G Paediatric Index of Mortality (PIM) Study Group PIM2: a revised version of the Paediatric Index of Mortality Intensive Care Med 2003 29 Paediatric Index of Mortality (PIM) Study Group: 278 285 12541154
Pollack MM Patel KM Ruttimann UE PRISM III: an updated Pediatric Risk of Mortality score Crit Care Med 1996 24 743 752 8706448 10.1097/00003246-199605000-00004
Carrol ED Riordan FA Thomson AP Sills JA Hart CA The role of the Glasgow meningococcal septicaemia prognostic score in the emergency management of meningococcal disease Arch Dis Child 1999 81 281 282 10532936
Siegemund M van Bommel J Ince C Assessment of regional tissue oxygenation Intensive Care Med 1999 25 1044 1060 10551958 10.1007/s001340051011
Kellum JA Metabolic acidosis in the critically ill: lessons from physical chemistry Kidney Int Suppl 1998 66 S81 S86 9573580
Gilfix BM Bique M Magder S A physical chemical approach to the analysis of acid-base balance in the clinical setting J Crit Care 1993 8 187 197 8305955 10.1016/0883-9441(93)90001-2
Durward A Skellett S Mayer A Taylor D Tibby SM Murdoch IA The value of the chloride: sodium ratio in differentiating the aetiology of metabolic acidosis Intensive Care Med 2001 27 828 835 11430538 10.1007/s001340100915
Moviat M van Haren F van der Hoeven H Conventional or physicochemical approach in intensive care unit patients with metabolic acidosis Crit Care 2003 7 R41 R45 12793889 10.1186/cc2184
Fencl V Jabor A Kazda A Figge J Diagnosis of metabolic acid-base disturbances in critically ill patients Am J Respir Crit Care Med 2000 162 2246 2251 11112147
Stewart PA Modern quantitative acid-base chemistry Can J Physiol Pharmacol 1983 61 1444 1461 6423247
Story DA Morimatsu H Bellomo R Strong ions, weak acids and base excess: a simplified Fencl-Stewart approach to clinical acid-base disorders Br J Anaesth 2004 92 54 60 14665553 10.1093/bja/aeh018
Taylor D Durward A Tibby SM Thorburn K Holton F Johnstone IC Murdoch IA Pitfalls of traditional acid base analysis in diabetic ketoacidosis [abstract] Pediatr Crit Care Med 2004 5 s311
Goldstein B Giroir B Randolph A International pediatric sepsis consensus conference: definitions for sepsis and organ dysfunction in pediatrics Pediatr Crit Care Med 2005 6 2 8 15636651 10.1097/01.PCC.0000149131.72248.E6
Fencl V Leith DE Stewart's quantitative acid-base chemistry: applications in biology and medicine Respir Physiol 1993 91 1 16 8441866 10.1016/0034-5687(93)90085-O
Figge J Rossing TH Fencl V The role of serum proteins in acid-base equilibria J Lab Clin Med 1991 117 453 467 2045713
Wilkes P Hypoproteinemia, strong-ion difference, and acid-base status in critically ill patients J Appl Physiol 1998 84 1740 1748 9572825
Durward A Mayer A Skellett S Taylor D Hanna S Tibby SM Murdoch IA Hypoalbuminaemia in critically ill children: incidence, prognosis, and influence on the anion gap Arch Dis Child 2003 88 419 422 12716714 10.1136/adc.88.5.419
Funk GC Zauner C Bauer E Oschatz E Schneeweiss B Compensatory hypochloraemic alkalosis in diabetic ketoacidosis Diabetologia 2003 46 871 873 12802497 10.1007/s00125-003-1119-3
Finfer S Bellomo R Boyce N French J Myburgh J Norton R SAFE Study Investigators A comparison of albumin and saline for fluid resuscitation in the intensive care unit N Engl J Med 2004 350 2247 2256 SAFE Study Investigators: 15163774 10.1056/NEJMoa040232
Kellum J Bellomo R Kramer DJ Pinsky MR Etiology of metabolic acidosis during saline resuscitation in endotoxaemia Shock 1998 9 364 368 9617887
Schiengraber S Rehm M Sehmisch C Finsterer U Rapid saline infusion produces hyperchloraemic metabolic acidosis in patients undergoing gynaecological surgery Anaesthesiology 1999 90 1265 1270 10.1097/00000542-199905000-00007
McFarlane C Lee A A comparison of Plasmalyte 148 and 0.9% saline for intra-operative fluid replacement Anaesthesia 1994 49 779 781 7978133
Moon PF Kramer GC Hypertonic saline dextran resuscitation from haemorrhagic shock induces transient mixed acidosis Crit Care Med 1995 23 323 331 7532561 10.1097/00003246-199502000-00019
Skellett S Mayer A Durward A Tibby SM Murdoch IA Chasing the base deficit: hyperchloraemic acidosis following 0.9% saline fluid resuscitation Arch Dis Child 2000 83 514 516 11087291 10.1136/adc.83.6.514
Balasubramanyan N Havens PL Hoffman GM Unmeasured anions identified by the Fencl-Stewart method predict mortality better than base excess, anion gap, and lactate in patients in the pediatric intensive care unit Crit Care Med 1999 27 1577 1581 10470767 10.1097/00003246-199908000-00030
Hatherill M Waggie Z Purves L Reynolds L Argent A Mortality and the nature of metabolic acidosis in children with shock Intensive Care Med 2003 29 286 291 12594588 10.1007/s00134-003-1888-7
Durward A Tibby SM Skellett S Austin C Anderson D Murdoch IA The strong ion gap predicts mortality in children following cardiopulmonary bypass surgery Pediatr Crit Care Med 2005 6 281 285 15857525 10.1097/01.PCC.0000163979.33774.89
Rocktaeschel J Morimatsu H Uchino S Bellomo R Unmeasured anions in critically ill patients: can they predict mortality? Crit Care Med 2003 31 2131 2136 12973170 10.1097/01.CCM.0000079819.27515.8E
Cusack RJ Rhodes A Lochhead P Jordan B Perry S Ball JA Grounds RM Bennett ED The strong ion gap does not have prognostic value in critically ill patients in a mixed medical/surgical adult ICU Intensive Care Med 2002 28 864 869 12122523 10.1007/s00134-002-1318-2
Brill SA Stewart TR Brundage SI Schreiber MA Base deficit does not predict mortality when secondary to hyperchloremic acidosis Shock 2002 17 459 462 12069180 10.1097/00024382-200206000-00003
Hatherill M McIntyre AG Wattie M Murdoch IA Early hyperlactataemia in critically ill children Intensive Care Med 2000 26 314 318 10823388 10.1007/s001340051155
De Backer D Lactic acidosis Minerva Anestesiol 2003 69 281 284 12766720
|
16137362
|
PMC1269470
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jul 8; 9(4):R464-R470
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3760
|
oa_comm
|
==== Front
Crit CareCritical Care1364-85351466-609XBioMed Central London cc37651613736110.1186/cc3765ResearchPerformance of prognostic models in critically ill cancer patients – a review den Boer Sylvia [email protected] Keizer Nicolette F [email protected] Jonge Evert [email protected] Intensivist, Department of Intensive Care, Academic Medical Center, Universiteit van Amsterdam, Amsterdam, Netherlands2 Informatician, Department of Medical Informatics, Academic Medical Center, Universiteit van Amsterdam, Amsterdam, Netherlands2005 8 7 2005 9 4 R458 R463 27 4 2005 26 5 2005 2 6 2005 16 6 2005 Copyright © 2005 den Boer et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
Prognostic models, such as the Acute Physiology and Chronic Health Evaluation (APACHE) II or III, the Simplified Acute Physiology Score (SAPS) II, and the Mortality Probability Models (MPM) II were developed to quantify the severity of illness and the likelihood of hospital survival for a general intensive care unit (ICU) population. Little is known about the performance of these models in specific populations, such as patients with cancer. Recently, specific prognostic models have been developed to predict mortality for cancer patients who are admitted to the ICU. The present analysis reviews the performance of general prognostic models and specific models for cancer patients to predict in-hospital mortality after ICU admission.
Methods
Studies were identified by searching the Medline databases from 1994 to 2004. We included studies evaluating the performance of mortality prediction models in critically ill cancer patients.
Results
Ten studies were identified that evaluated prognostic models in cancer patients. Discrimination between survivors and non-survivors was fair to good, but calibration was insufficient in most studies. General prognostic models uniformly underestimate the likelihood of hospital mortality in oncological patients. Two versions of a specific oncological scoring systems (Intensive Care Mortality Model (ICMM)) were evaluated in five studies and showed better discrimination and calibration than the general prognostic models.
Conclusion
General prognostic models generally underestimate the risk of mortality in critically ill cancer patients. Both general prognostic models and specific oncology models may reliably identify subgroups of patients with a very high risk of mortality.
==== Body
Introduction
Advances in oncological and supportive care have led to improved prognosis and extension of survival time in cancer patients. However, such advances have often been achieved through aggressive therapies and support, at high expense. Some of these patients require admission to the intensive care unit (ICU) for acute concurrent illness, postoperative care, or complications of their cancer or its therapy. Recent studies [1-6] suggest that mortality of cancer patients in the ICU is comparable with that of patient groups suffering from other severe diseases, but others reported a poor prognosis with much higher mortality rates [7]. Efforts have been made to identify parameters that are associated with poor prognosis and to develop scoring models for predicting hospital mortality at ICU admission of cancer patients.
Different prognostic systems, such as the Acute Physiology and Chronic Health Evaluation (APACHE) II or III [8,9], the Simplified Acute Physiology Score (SAPS) II [10], and the Mortality Probability Models (MPM) II [11], have been developed to predict the outcome of critically ill patients admitted to the ICU. Although these models perform well in predicting the mortality of the general ICU patient population, they may well under- or overestimate mortality in selected patient subpopulations that were not well represented in the original cohort on which the model was developed. Therefore, new models were designed for specific populations, such as cancer patients [12,13]. By including variables specific to oncology such as disease progression/recurrence, performance status and type of treatment, they were expected to perform better than the general models. The aim of this review is to evaluate the performance of the general severity-of-illness scores (APACHE II and III, SAPS II, MPM II) and the specific oncological scoring systems in cancer patients requiring admission to the ICU.
Methods and materials
Sources and selection criteria
The information in this review is based on results of a Medline search for recent studies published between 1994 and 2004. The key words used included "Severity of illness scores", "Acute Physiology and Chronic Health Evaluation (APACHE)", "Simplified Acute Physiology Score (SAPS)", "Mortality Probability Model II", "Cancer", "Oncology", "Critical care", "Prognosis and outcome" and "Hospital mortality". Based on the title and abstract of the publication, we selected English-language articles containing information on the performance of prognostic models in cancer patients admitted to ICUs. The references of all selected reports were cross-checked for other potentially relevant articles. It was envisaged that the studies would be too heterogeneous to combine for a formal meta-analysis and therefore a narrative synthesis was undertaken.
Results
Performance of the prognostic models
Although several measures exist for evaluating the performance of prognostic models, all identified studies used receiver operating characteristic (ROC) curves and the area under the curve (AUC) [14] to evaluate discrimination and the Hosmer-Lemeshow goodness-of-fit H- or Ĉ-statistics [15] to evaluate the calibration of the prognostic models.
'Discrimination' refers to a model's ability to distinguish survivors from non-survivors. The AUC represents the probability that a patient who died had a higher predicted probability of dying than a patient who survived. An AUC of 0.5 indicates that the model does not predict better than chance. The discrimination of a prognostic model is considered perfect if AUC = 1, good if AUC >0.8, moderate if AUC is 0.6 to 0.8, and poor if AUC <0.6 [16]. The AUC of a model gives no indication of how close the predicted probabilities are to the observed outcome. To take this aspect of a model's performance into account, we have to look at the calibration and accuracy of the prognostic models.
'Calibration' refers to the agreement between predicted probabilities and the 'true probabilities'. Of course, the true probability of a patient's outcome is not known, otherwise there would be no need to develop prognostic models. However, the true probabilities can be approximated by taking the mean of the observed outcomes within predefined groups of patients. The selected studies used Hosmer-Lemeshow H- or Ĉ-statistics. Both H- and Ĉ-statistics compare the observed mortality in a group with the predicted mortality of that group. A disadvantage of the Hosmer-Lemeshow tests is that the value of the statistic is sensitive to the choice of the cut-off points that define the groups. The H- and Ĉ-statistics differ in the way the groups of patients are composed [15]. Grouping for the H-statistic is based on partitioning of the probability interval (0–1) into ten equally sized ranges. The Ĉ goodness-of-fit statistic sorts observations according to their expected probability and partitions the observations into ten groups of equal size. A high H or Ĉ relates to a small p value, implying significant difference between observed and predicted mortality, and thus indicates a lack of fit of the model. It is a generally known weakness of the Hosmer-Lemeshow goodness-of-fit statistics that the sample size has a major influence on the measured calibration. Using small samples will result in an apparently good fit, using large samples will result in an apparently poor fit [17,18].
Discrimination and calibration describe the overall predictive power of a model. This is important when analyzing the mortality risk of a population, for example, to determine performance of an ICU by measuring the standardized mortality ratio (SMR: observed mortality divided by predicted mortality) as mortality adjusted for severity-of-illness. When caring for an patient, it is more important that a model can reliably predict the likelihood of an outcome of an individual patient; this is called 'accuracy'. Accuracy refers to the difference between predictions and observed outcomes at the level of individuals. The mean squared error (MSE), also called Brier score, is an example of an accuracy measure [19]. None of the selected studies evaluated accuracy measures. However, some studies notify that when caring for an individual patient, it is more important that a model can reliably identify patients with a very high risk of dying. Therefore, they evaluated the performance of prognostic models at specific cut-off points, dividing high-risk patients from low-risk patients.
Discrimination
Nine studies reported on discriminating ability of general prognostic models in ICU patients with cancer (Table 1) [12,16,20-26]. The APACHE II score was evaluated in six studies and showed poor to good discriminating value with areas under the ROC curve between 0.60 and 0.78 [16,20-22,25,26]. Discrimination of the SAPS II model was fair to good with areas under the ROC curve between 0.67 and 0.83 [20-23,25,26]. The MPM II model showed poor discrimination in one study [12], but good discrimination in another [26]. Discrimination of all models differed importantly between studies. All models showed fair to good discrimination in the study by Soares [26] and poor discrimination in the study by Sculier [20]. This may be related to differences in casemix of patients; whereas most patients in the latter trial had metastatic or disseminated haematological disease, most patients in the study by Soares had locoregional cancer or cancer in remission only.
In 1998, Groeger and others developed a model specific for cancer patients [12]. It included physiological data, disease-related variables (allogeneic bone marrow transplantation and recurrent or progressive cancer), and performance status before hospitalization. Tested on an independent set of patients, the model showed good discriminating power with an area under the ROC curve of 0.81. Good discriminating ability was confirmed in two other studies [22,26]. In 2003, Groeger et al. developed another specific model with good discriminating performance (AUC = 0.82) that predicts in-hospital mortality in cancer patients at 72 h after ICU admission [13].
Calibration
As shown in Table 1, most studies showed poor calibration for APACHE II and III, SAPS II and MPM II [12,20,22,26]. The studies that have good H- or Ĉ-statistics (p>0.05) are very small. Poor calibration resulted in a uniform underestimation of the mortality risk using the general prognostic models. Hence, the observed mortality was uniformly higher than the predicted mortality (SMR>1). In contrast with the general prognostic models, the specific models for cancer patients showed good calibration with SMR of 1.0 to 1.05 [12,13,21] or poor calibration resulting in a uniform overestimation of mortality risk (SMR 0.75 [26]).
Identification of subgroups at (very) high risk
It may be particularly important to identify patients at very high risk of dying. Patients with limited life expectancy do not necessarily prefer life-extending treatment over care focused on relieving pain and discomfort. The willingness to receive life-sustaining treatment depends on the burden of treatment, the outcome and the likelihood of the outcome [27]. In patients aged 65 years and older, the willingness to receive cardiopulmonary resuscitation if they suffered a cardiac arrest decreased from 41 to 22% after learning the probability of survival (10 to 17%) [28].
Although none of the selected studies report on accuracy measures, some studies report on the performance of prognostic models at specific cut-off points for predicted mortality. Results are summarized in Table 2. High predicted mortality by the general prognostic models as well as the specific ICU Cancer Mortality Model (ICCM) was associated with very high observed mortality rates. For example, in a study by Staudinger and others, 7% of the studied population had more than 79 APACHE III points, and all of these patients died before hospital discharge [24]. In another study by Sculier and others, 5.4% of patients had a predicted mortality of >70% according to the APACHE II model. In these patients, the actual observed mortality was 86% [20]. Thus, in a limited number of cancer patients, a very high mortality chance after ICU admission may be predicted. It may be speculated that some patients would prefer not to undergo intensive care treatment if their predicted mortality is very high, for example, more than 80 or 90%. Providing prognostic information to patients, their relatives and physicians could help to provide intensive care that is more in accordance to patients' own preferences.
Discussion
The general prognostic models for ICU patients generally underestimate the risk of dying for cancer patients admitted to ICUs. This is important when interpreting SMRs of different ICUs, since ICUs with relatively more cancer patients will have a higher SMR. However, these models are able to identify subgroups of patients with a very high mortality risk. Thus, they may have a role in giving information about the prognosis to patients and their relatives.
Only a few models exist that were specifically designed for cancer patients and which include data about type and stage of cancer, and functional status of patients. These models showed better discrimination and calibration than the general models. Thus, they may have a role in comparing SMRs of cancer patient populations in the ICU. However, they have been validated in relatively few patients and new larger studies are required to confirm the value of these models. Most studies had a retrospective design and limited number of patients, and the moderate differences among the scoring systems do not allow conclusion of the superiority of one of them. Because of large variations in their design (type of patients, observed mortality (33 to 60%), mix of H- and Ĉ-statistics), it is difficult to perform meaningful comparisons between them. Different casemixes, national or regional patient populations and critical care management can lead to different outcomes.
A limitation of all models is the fact that they do not take into account that better treatments may become available and that prognosis may improve over time. Indeed, it has been shown that survival of patients after haematopoietic stem cell transplantation who received mechanical ventilation, improved from lower than 10% in the period before 1990 to 25 to 50% in the period 1994 to 2000 [16]. Thus, prognostic information should be interpreted cautiously. Nevertheless, patients and physicians need optimal information about the likelihood of a beneficial outcome of intensive care treatment. Prognostic models do at least contribute to this information.
Conclusion
The general prognostic models for ICU patients generally underestimate the risk of dying for cancer patients admitted to ICUs. Models specific for cancer patients show better calibration and discrimination than the general models. Both general models and specific oncology models may reliably identify subgroups of patients with a very high mortality risk and thus may be useful to inform patients and their relatives about the likelihood of a beneficial outcome.
Key messages
• General prognostic models for ICU patients, such as APACHE II or SAPS II, tend to underestimate the risk of dying for patients with cancer admitted to ICUs.
• Prognostic models specifically designed for ICU patients with cancer show better calibration and discrimination than the general models.
• Both general models and specific oncology models reliably identify subgroups of patients with a very high risk of dying.
Abbreviations
APACHE = Acute Physiology and Chronic Health Evaluation; AUC, area under the curve; ICCM = ICU Cancer Mortality Model; ICU, intensive care unit; MPM = Mortality Probability Model; ROC = receiver operating curve; SAPS = Simplified Acute Physiology Score; SMR = standardized mortality rate.
Competing interests
The authors declare that they do not have competing interests.
Authors' contributions
SdB and EdJ acquired and interpreted the data. NFdK interpreted the data. All authors contributed in preparing the manuscript. All authors read and approved the final manuscript.
Figures and Tables
Table 1 Overall predictive performance of prognostic models in ICU patients with cancer
Mortality%
Study N Solid/metastatic/haematological malignancies Prognostic score ROC Hosmer-Lemeshow goodness-of-fit H or Ĉ test p value Pred Obs SMR
Sculier [20] 261 APACHE II 0.60 0.001 (H) 26.5 33 1.25
solid 77% SAPS II 0.67 0.001 26.1 1.26
metastatic 62%
haematological 23%
Groeger [12] 805 MPM II 0.63 <0.001 (Ĉ) 22 41 1.86
230 ICMM 0.81 0.310 ng 1.02
solid and haematological
Groeger [13] 415 ICMM at 72 h 0.82 0.354 (Ĉ) ng ng 1.0
solid and haematological
Schellongowki [21] 242 APACHE II 0.78 0.058 (Ĉ) ng 44 1.05
solid 45% SAPS II 0.83 0.066 ng
haematological 55% ICMM (2) 0.70 0.115 42
Berghmans [22] 247 APACHE II 0.65 0.002 (H) 32 34 1.06
solid 80.5% SAPS II 0.72 <0.0001 24 1.42
metastatic 62% ICMM (2) 0.79 0.060 28 1.21
haematological 19.5%
Guiguet [23] 94 SAPS II 0.78 0.750 (H) ng 60 1
solid 44.7%
haematological 55.3%
Staudinger [24] 414 APACHE III 0.75 ng ng 47
solid 42%
haematological 58%
Benoit [25] 124 SAPS II 0.77 0.60 (Ĉ) ng 54
haematological 100% APACHE II 0.71 0.39 ng
Afessa [16] 112 APACHE III 0.70 0.584 (H) 42 46 1.09
haematological 100% APACHE II ng ng 44 1.03
Soares [26] 542 SAPS II 0.82 <0.001 (H) 47.9 58.7 1.23
solid 88.8% APACHE III 0.81 <0.001 42.6 1.38
haematological 11.2% ICMM (2) 0.80 <0.001 78.7 0.75
excluding scheduled MPM II 24 h 0.79 <0.001 37.7 1.56
surgical patients APACHE II 0.75 <0.001 38.2 1.54
MPM II 0 h 0.73 <0.001 25.0 2.35
Shown are areas under ROC, p value belonging to Hosmer-Lemeshow goodness-of-fit H- or Ĉ- statistics and SMRs for individual mortality prediction models. APACHE, Acute Physiology and Chronic Health Evaluation; ICMM, ICU Cancer Mortality Model; MPM, Mortality Probability Model; ROC, receiver operating curve; SAPS, Simplified Acute Physiology Score; SMR, standardized mortality rate; ng, not given.
Table 2 Positive predictive value of prognostic models at specific cut-off values for predicted mortality by severity-of illness models
Study Prognostic model and cut-off probability of mortality Observed Mortality, %
Sculier [20] APACHE II >70% 86
SAPS II >60% 67
Berghmans [22] APACHE II >60% 78
SAPS II >60% 86
ICMM >40% 67
Staudinger [24] APACHE III ≥ 79* 100
Kroschinsky [29] SAPS II ≥ 80* 95
Groeger [13] ICCM72 >70% 83
Groeger [12] ICCM >80% 91
*Severity of illness scoring points, not predicted mortality. APACHE, Acute Physiology and Chronic Health Evaluation; ICCM, ICU Cancer Mortality Model; ICCM72, ICU Cancer Mortality Model at 72 h; SAPS, Simplified Acute Physiology Score.
==== Refs
Baumann WR Jung RC Koss M Boylen CT Navarro L Sharma OP Incidence and mortality of adult respiratory distress syndrome: a prospective analysis from a metropolitan hospital Crit Care Med 1986 14 1 4 3484443
Sussman NL Lake JR Treatment of hepatic failure 1996: current concepts and progress toward liver dialysis Am J Kidney Dis 1996 27 605 621 8629619
Krafft P Fridrich P Pernerstorfer T Fitzgerald RD Koc D Schneider B Hammerle AF Steltzer H The acute respiratory distress syndrome: definitions, severity and clinical outcome. An analysis of 101 clinical investigations Intensive Care Med 1996 22 519 529 8814466
Marik PE Craft M An outcomes analysis of in-hospital cardiopulmonary resuscitation: the futility rationale for do not resuscitate orders J Crit Care 1997 12 142 146 9328854 10.1016/S0883-9441(97)90044-7
The British Thoracic Society Research Committee and The Public Health Laboratory Service The aetiology, management and outcome of severe community-acquired pneumonia on the intensive care unit Respir Med 1992 86 7 13 1565823
Massion PB Dive AM Doyen C Bulpa P Jamart J Bosly A Installe E Prognosis of hematologic malignancies does not predict intensive care mortality Crit Care Med 2002 30 2260 2270 12394954 10.1097/00003246-200210000-00014
Sculier JP Markiewicz E Medical cancer patients and intensive care Anticancer Res 1991 11 2171 2174 1776858
Knaus W Draper E Wagner D Zimmerman J APACHE II: a severity of disease classification system Crit Care Med 1985 13 818 829 3928249
Knaus W Wagner D Draper E Zimmerman J Bergner M Bastos P Sirio C Murphy D Lotring T Damiano A Harrel F The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalised adults Chest 1991 100 1619 1636 1959406
Le Gall J-R Lemeshow SS Saulnier F A new Simplified Acute Physiology score (SAPS II) based on a European/North American multicenter study JAMA 1993 270 2957 2963 8254858 10.1001/jama.270.24.2957
Lemeshow S Teres D Klar J Avrunin JS Gehlbach SH Rapoport J Mortality Probability Models (MPMII) based on an international cohort of intensive care unit patients JAMA 1993 270 2478 2486 8230626 10.1001/jama.270.20.2478
Groeger JS Lemeshow S Price K Nierman DM White P JrKlar J Granovsky S Horak D Kish SK Multicenter outcome study of cancer patients admitted to the intensive care unit: a probability of mortality model J Clin Oncol 1998 16 761 770 9469368
Groeger JS Glassman J Nierman DM Wallace SK Price K Horak D Landsberg D Probability of mortality of critically ill cancer patients at 72 h of intensive care unit (ICU) management Support Care Cancer 2003 11 686 695 12905057 10.1007/s00520-003-0498-9
Hanley JA McNeil BJ The meaning and use of the area under a receiver operating characteristic (ROC) curve Radiology 1982 143 29 36 7063747
Hosmer DW Lemeshow S Applied logistic regression 2000 New York: John Wiley & Sons
Afessa B Tefferi A Dunn WF Litzow MR Peters SG Intensive care unit support and Acute Physiology and Chronic Health Evaluation III performance in haematopoietic stem cell transplant recipients Crit Care Med 2003 31 1715 1721 12794410 10.1097/01.CCM.0000065761.51367.2D
Zhu BP Lemeshow S Hosmer DW Klar J Avrunin J Teres D Factors affecting the performance of the models in the Mortality Probability Model II system and strategies of customization: a simulation study Crit Care Med 1996 24 57 63 8565539 10.1097/00003246-199601000-00011
Vergouwe Y Steyerberg EW Eijkemans R Habbema D Sample size considerations for the performance assessment of predictive models: A simulation study Control Clin Trials 2003 S43 S44
Brier G Verification of forecasts expressed in terms of probability Monthly Weather Rev 1950 78 1 3
Sculier JP Paesmans M Markiewicz E Berghmans T Scoring systems in cancer patients admitted for an acute complication in a medical intensive care unit Crit Care Med 2000 28 2786 2792 10966251 10.1097/00003246-200008000-00018
Schellongowski P Benesch M Lang T Traunmuller T Zauner C Laczika K Locker GJ Frass M Staudinger T Comparison of three severity scores for critically ill cancer patients Intensive Care Med 2004 30 430 436 14598029 10.1007/s00134-003-2043-1
Berghmans T Paesmans M Sculier JP Is a specific oncological scoring system better at predicting the prognosis of cancer patients admitted for an acute medical complication in an intensive care unit than general gravity scores? Support Care Cancer 2004 12 234 239 14740281 10.1007/s00520-003-0580-3
Guiguet M Blot F Escudier B Antoun S Leclercq B Nitenberg G Severity-of-illness scores for neutropenic cancer patients in an intensive care unit: Which is the best predictor? Do multiple assessment times improve the predictive value? Crit Care Med 1998 26 488 493 9504577 10.1097/00003246-199803000-00020
Staudinger T Stoiser B Mullner M Locker GJ Laczika K Knapp S Burgmann H Wilfing A Kofler J Thalhammer F Frass M Outcome and prognostic factors in critically ill cancer patients admitted to the intensive care unit Crit Care Med 2000 28 1322 1328 10834673 10.1097/00003246-200005000-00011
Benoit D Vandewoude K Decruyenaere J Hoste E Colardyn F Outcome and early prognostic indicators in patients with a haematological malignancy admitted to the intensive care unit for a life-threatening complication Crit Care Med 2003 31 104 112 12545002 10.1097/00003246-200301000-00017
Soares M Fontes F Dantas J Gadelha D Cariello P Nardes F Amorim C Toscano L Rocco J Performance of six severity-of-illness scores in cancer patients requiring admission to the intensive care unit: a prospective observational study Crit Care 2004 8 R194 R203 15312218 10.1186/cc2870
Fried T Bradley E Towle V Allore H Understanding the treatment preferences of seriously ill patients N Engl J Med 2002 346 1061 1066 11932474 10.1056/NEJMsa012528
Murphy D Burrows D Santilli S Kemp A Tenner S Kreling B Teno J The influence of the probability of survival on patients' preferences regarding cardiopulmonary resuscitation N Engl J Med 1994 330 545 549 8302322 10.1056/NEJM199402243300807
Kroschinsky F Weise M Illmer T Haenel M Bornhaeuser M Hoeffken G Ehninger G Schuler U Outcome and prognostic features of intensive care unit treatment in patients with haematological malignancies Intensive Care Med 2002 28 1294 1300 12209280 10.1007/s00134-002-1420-5
|
16137361
|
PMC1269472
|
CC BY
|
2021-01-04 16:04:54
|
no
|
Crit Care. 2005 Jul 8; 9(4):R458-R463
|
utf-8
|
Crit Care
| 2,005 |
10.1186/cc3765
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460010.1371/journal.pgen.0010048plge-01-04-05Research ArticleCytoskeletal Rearrangements in Synovial Fibroblasts as a Novel Pathophysiological Determinant of Modeled Rheumatoid Arthritis Expression Profiling of Modeled ArthritisAidinis Vassilis 1*Carninci Piero 2Armaka Maria 1Witke Walter 3Harokopos Vaggelis 1Pavelka Norman 4Koczan Dirk 5Argyropoulos Christos 6Thwin Maung-Maung 7Möller Steffen 5Kazunori Waki 2Gopalakrishnakone Ponnampalam 7Ricciardi-Castagnoli Paola 4Thiesen Hans-Jürgen 5Hayashizaki Yoshihide 2Kollias George 11 Institute of Immunology, Alexander Fleming Biomedical Sciences Research Center, Athens, Greece
2 RIKEN Genomic Sciences Center, Yokohama, Japan
3 Mouse Biology Programme, EMBL, Monterotondo, Italy
4 Department of Biotechnology and Bioscience, University of Milano-Bicocca, Milan, Italy
5 Institute of Immunology, University of Rostock, Rostock, Germany
6 Laboratory of Medical Physics, University of Patras, Patras, Greece
7 Department of Anatomy, National University of Singapore, Singapore
Roopenian Derry C EditorThe Jackson Laboratory, United States of America* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 1 4 e4822 4 2005 14 9 2005 Copyright: © 2005 Aidinis et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Rheumatoid arthritis is a chronic inflammatory disease with a high prevalence and substantial socioeconomic burden. Despite intense research efforts, its aetiology and pathogenesis remain poorly understood. To identify novel genes and/or cellular pathways involved in the pathogenesis of the disease, we utilized a well-recognized tumour necrosis factor-driven animal model of this disease and performed high-throughput expression profiling with subtractive cDNA libraries and oligonucleotide microarray hybridizations, coupled with independent statistical analysis. This twin approach was validated by a number of different methods in other animal models of arthritis as well as in human patient samples, thus creating a unique list of disease modifiers of potential therapeutic value. Importantly, and through the integration of genetic linkage analysis and Gene Ontology–assisted functional discovery, we identified the gelsolin-driven synovial fibroblast cytoskeletal rearrangements as a novel pathophysiological determinant of the disease.
Synopsis
Rheumatoid arthritis (RA) is a chronic destructive disease that affects 1–3% of the general population, exacting substantial personal, social, and economic costs. Current treatments alleviate the symptoms and offer immediate relief for many patients but do not cure the disease. While the cause of the disease remains poorly understood, the completion of the Human Genome Project and the emergence of functional genomics and high-throughput technologies offer intriguing new possibilities. For example, expression profiling creates a molecular fingerprint of the disease status by quantifying the expression levels of thousand of genes simultaneously. Similarly, reverse genetics (the genetic modification of a particular gene in search of its function) allow for the creation of animal models of disease. To discover novel genes and/or cellular pathways involved in the development of the disease, the authors used two methods in an animal model of RA for large-scale expression profiling. They identified a large number of genes and molecular processes that are deregulated in the disease. Using this information, the authors described pathophysiologic determinants of RA and created a unique list of disease modifiers of potential therapeutic value.
Citation:Aidinis V, Carninci P, Armaka M, Witke W, Harokopos V, et al (2005) Cytoskeletal rearrangements in synovial fibroblasts as a novel pathophysiological determinant of modeled rheumatoid arthritis. PLoS Genet 1(4): e48.
==== Body
Introduction
Rheumatoid arthritis (RA) is a chronic destructive arthropathy with a prevalence of 1–3% and substantial personal, social, and economic costs. It is characterized by prolonged inflammation of the joints, eventually leading to destruction of the cartilage and bone. Inflammation is initially localized in the synovial lining, a monolayer of synovial cells that lines diarthroidal joints. In RA, the synovial lining becomes markedly thickened due to synovial cell proliferation and infiltration by inflammatory cells. This proliferative mass, the pannus, invades and destroys articular cartilage and bone, leading to irreversible destruction of joint structure and function [1]. Current therapies of RA rely mainly on symptomatic treatment with nonsteroidal antiinflammatory drugs and/or with disease-modifying antirheumatic drugs. However, even the best available treatments (such as targeting tumour necrosis factor [TNF] and TNF signalling) do not cure the disease and do not even sufficiently retard progression in the majority of the patients, while they often exhibit adverse side effects [2].
Despite intense efforts, the aetiology and pathogenesis of RA remain poorly understood. Traditional research paradigms for RA have implicated a variety of mechanisms that contribute to the initiation and perpetuation of synovial inflammation, including autoantibodies and immune complexes, T cell-mediated antigen-specific responses, persistence of cytokine networks and other proinflammatory molecules, genetic bias and sex predisposition, and tumour-like behaviour of the arthritic synovium [3]. Animal models of RA share many clinical features with the human disease and hence constitute valuable tools in deciphering the pathogenic mechanisms that govern disease activation and perpetuation [4]. Among them, the TNF-transgenic (TNF-Tg) mouse [5] has been instrumental in demonstrating the role of TNF in the development of the disease and foreshadowed the introduction and success of anti-TNF therapies that transformed the effective management of the disease [6]. In this model, chronic overexpression of human TNF results in a chronic, erosive, symmetric polyarthritis, with 100% phenotypic penetrance, timed disease onset, and progressive histological symptoms that closely resemble human RA [5–7].
To gain further insights into the pathophysiology of the disease and to discover genes and/or pathways involved in its pathogenesis, we have utilised the TNF-Tg animal model of RA for large-scale expression profiling with both subtractive libraries and oligonucleotide microarray hybridizations. Differential expression was validated by a number of methods, in both mouse and human patient samples, thus creating a unique database of potential disease modifiers and therapeutic targets. Moreover, in an attempt to discover deregulated cellular functions based on functional annotations of deregulated genes, we identified the gelsolin-driven synovial fibroblast cytoskeleton rearrangement as a pathophysiological determinant of the disease.
Results
To discover genes and cellular pathways that participate in the pathogenesis of RA on a large scale, we used a twin high-throughput approach, comprising two entirely different methodologies governed by different constraints and analyzed by different statistics. Total RNA samples were extracted from whole-joint (WJ) and synovial fibroblast (SF) ex vivo cultures isolated from 6-wk-old mice with RA (Tg197, hTNF
+/−; disease severity index 3; [5]) and their normal (wild-type [WT]) littermates. Each sample (out of four: RA SF, WT SF, RA WJ, and WT WJ) consisted of equimolar amounts of RNA pooled from four (two male and two female) mice (16 mice altogether), to some extent equalizing biological diversity. Biological replicates (different extractions from independent mice, employing the same extraction protocols and pooling strategy) were used for both the creation of subtractive cDNA libraries and the hybridization of oligonucleotide microarrays.
Subtractive cDNA Libraries and Large-Scale Sequencing
Four different full-length cDNA libraries (RA SF, WT SF, RA WJ, and WT WJ) were normalized and subtracted to each other, as outlined in Figure S1A. Due to experimental design, the resulting subtracted libraries (L0, L1, L2, and L7) contained cDNAs from the tester cDNA library only and therefore constituted libraries of up-regulated genes in the corresponding tester library (or of down-regulated genes in the driver library). From the subtractive libraries, 27,511 cDNAs/clones were sequenced and clustered to 9,176 clusters/genes, as summarized in Figure S1B. Each gene was then annotated through BLAST homology searches at Unigene, FANTOM, and SWISSPROT databases.
In summary, among the 9,176 genes found, 7,977 corresponded to known genes, while 1,199 had sequences not reported previously. Each gene was represented by a different number of clones in almost all libraries, directly proportional to subtraction efficiency and transcript abundance. The relative distribution of each gene in each library is the true measure of differential expression, which can be obscured by sampling errors arising by chance when the clones are selected. Therefore, in order to identify the truly differentially expressed genes, a likelihood value (R) was assigned to each gene from pairwise comparisons of the relative libraries (SF/L0L1 and WJ/L2L7). The statistical significant thresholds were then calculated (Figure S2), and two significance levels were selected (summarized in Table 1): a very high one (99.99%; p ≤ 0.0001) to report the results independently, and a lower one (99%; p ≤ 0.01) for comparison with the corresponding results from the DNA microarray hybridizations. Known (by a Unigene cluster ID) differentially expressed genes at 99.99% significance level are presented in Table S1.
Table 1 Summary Results of Subtractive Libraries and Microarray Hybridizations
Table 2 Summary Results of Oligonucleotide Microarray Hybridizations
Table 3 Differentially Expressed Genes in Arthritic WJs or SFs Commonly Identified by Subtractive Libraries and Oligonucleotide DNA Microarray Hybridizations
Oligonucleotide Microarray Hybridizations
Fluorescently labeled cRNA probes made from similar (biological replicates, see above) samples of total RNA (RA SF, WT SF, RA WJ, and WT WJ) were utilized to hybridize in duplicate the Affymetrix Mu11K oligonucleotide DNA chipset (eight chipsets, 16 chips total). Furthermore, similar samples (equimolar amounts of RNA pooled from four—two male and two female—6-wk-old arthritic mice; severity index 3) of total RNA (from SF and WJ) from another animal model of arthritis (spontaneous, knock-in, TNF
ΔARE+/-) [8] were used for additional chip hybridizations (four chipsets, eight chips total). The MIAME-compliant [9] microarray data (see the ArrayExpress database [10], accession number E-MEXP-255) were normalized and analyzed as outlined in Figure S3 and described in detail in Materials and Methods and Figure S4. Differentially expressed genes (DEGs) were selected, utilizing a sample-specific fold-change model (FCM) [11] at different significance levels (90%, p ≤ 0.1; 95%, p ≤ 0.05), where selected DEGs have always an observed fold-change higher than the expected fold-change. The results for both animal models are summarized in Table 2. Sample-specific DEGs common to both animal models (significance: 95%, p ≤ 0.05) are presented in Table S2.
Cross-Platform Comparison and Validation
Both differential expression analysis methods presented above produced lists of deregulated genes of high statistical significance. To validate the results independently and to avoid performing numerous RT-PCRs, differential expression results by both platforms of analysis were compared to each other for the same animal model (Tg197, hTNF
+/-), sample type (WJ, SF), and direction of deregulated expression (up or down). The comparison, at significance levels 99% for the libraries and 90% for the microarrays (selected based on similar output gene numbers), was performed through the NetAffx database (http://www.affymetrix.com). Although few examples of expression profiling cross-platform overlaps have been reported [12,13], in this study 46 genes (15 for SF, 31 for WJ) were commonly predicted as up- or down-regulated by both methods (combined p-value of 0.001) (Table 3).
To verify the validity of the twin high-throughput approach and to prove that the reported gene list is self-validated, a number of the predicted deregulated genes were confirmed with different methods. Expression profiling in the knock-in disease model (TNF
ΔARE+/-) confirmed 40 of the genes (Table 3). Automated literature text mining (Biolab Experiment Assistant, BIOVISTA, Greece) identified 20 of the genes previously associated with RA (Figure S5; summarized in Table 3). In addition, 11 representative genes (SF: Gsn, Aqp1, mglap, cdc42 hom, and Hp; WJ: Marco, Hp, CD14, Mb, Bsg, Ptgis) were further confirmed by semi-quantitative (SQ) RT-PCR (Figure S6A–S6C; summarized in Table 3). All SQ-RT-PCRs were performed in the linear range of the reaction (at three different concentrations normalized against housekeeping genes) in biological replicates (different extractions from independent mice, employing the same extraction protocols and pooling strategy) of the samples used for both the subtractive libraries and the microarray hybridizations. Representative genes were selected on the basis of (1) different sample source (six from WJs and five from SFs), (2) different direction and degree of deregulated expression (six up-regulated and five down-regulated), and (3) biological interest and potential follow-up. Moreover, two of them (SF: Gsn, Aqp1) were also confirmed by real-time RT-PCR (Figure S6D; summarized in Table 3).
In an attempt to combine gene expression analysis with genetic linkage analysis, all differentially expressed genes were mapped to the chromosomes together with the known quantitative trait loci (QTL, chromosomal regions/genes segregating with a quantitative trait) for an induced animal model of arthritis, collagen-induced arthritis (CIA) [14]. As graphically represented in Figure 1 (and summarized in Table 3), eight genes mapped to CIA QTL (WJ: Ctss, Pitpnm, Ncf1, Psmb8, and Siat8e, SF: Rab14, Aqp1, and Gsn).
Figure 1 Chromosomal Localization of Identified Deregulated Genes in the RA Animal Model, Together with the QTL for CIA
Red and green arrowheads indicate up- and down-regulated genes, respectively. Black and blue lettering refers to deregulated genes in WJs and SFs, respectively.
The expression level of seven of the genes found deregulated in the arthritic mice and confirmed by RT-PCR in mouse samples, was also examined in human patients' RNA samples with real-time RT-PCR analysis. Due to the lack of normal (WT) human synovium samples, we compared the expression of 19 RA samples with eight osteoarthritis samples as controls, a consensus strategy for differential expression analysis in arthritis [15,16]. As shown in Figure 2 and summarized in Table 3, the deregulated expression of six (out of seven assayed) of the genes was confirmed in humans as well, including four with high statistical significance (Gsn, Aqp1, Bsg, and Mb), thus extending the validity and utility of the mouse model-generated deregulated gene list to the human disease.
Figure 2 Confirmation of Deregulated Expression in Human Patient Samples
Quantitative RT-PCR for the indicated genes for 19 RA and eight osteoarthritis (OA) samples. Values were normalized to the expression of B2m and were expressed as expression index. Similar results were obtained upon normalization to L32. Nonparametric Mann-Whitney statistical tests were used to derive p-values.
Arthritic Synovial Fibroblasts Have a Rearranged Cytoskeleton
The twin high-throughput expression profiling approach described above yielded a large number of disease-implicated, deregulated genes of high statistical significance. Furthermore, to (1) prove the validity and extend the utility of the expression data analysis even further, (2) infer deregulated biological functions from the gene expression data, and (3) define functional criteria for further gene selection, the selected genes were annotated in the form of the Gene Ontology (GO) [17] term “biological process.” GO term frequencies in the selected gene lists were then calculated, and their statistical significance was estimated. As shown in Table S3, predicted deregulated functions in SFs include, as expected, collagen catabolism, complement activation, and immune and stress responses. Interestingly, five out of 26 significantly (p < 0.01) deregulated GO functions concerned (directly or indirectly) the actin cytoskeleton, suggesting that arthritic SFs have a rearranged actin cytoskeleton. In order to confirm the prediction, F-actin was visualized in vitro on arthritic as well as WT SFs (both primary and immortalized) in mice. As is evident from Figure 3A, arthritic SFs exhibit pronounced stress fibers, thus validating the in silico, expression-based hypothesis.
Figure 3 Mouse Arthritic SFs have a Rearranged Cytoskeleton and Increased ECM Adhesion
(A) Immunofluoresence of arthritic (RA) or normal (WT) primary and immortalized synovial fibroblasts (pSFs and iSFs respectively) for F-actin.
(B) Adhesion assays of arthritic (RA) or normal (WT) primary and immortalized SFs on purified ECM components as described in Materials and Methods. Error bars indicate standard deviation of triplicate samples from their mean value.
(C) Transmission electron microscopy (TEM, magnification 5,000×) of SFs from ankle joints isolated from WT and arthritic mice. Arrowheads indicate dilation r-ER, while arrows point to swollen mitochondria with distorted cristae. N, nuclei.
Stress fibers within fibroblasts allow them to exert tension on the extracellular matrix (ECM) surrounding them—an essential process in wound healing. It is well understood that differences in the actin cytoskeleton reflect altered ECM attachment properties and/or vice versa. Indeed, arthritic SFs were shown to adhere to different proteins of the ECM (fibronectin, vitronectin, laminin, and collagen) with increased affinity in vitro (Figure 3B).
Attachment to the ECM (and associated cytoskeletal changes), mediated mainly through engagement and clustering of transmembrane integrin molecules, largely define cell shape and morphology, as well as their behavior and fate. Increased adhesion to the ECM is expected to lead to a more elongated shape. In order to confirm the increased adhesion of the arthritic fibroblasts in vivo, we examined ankle joints from arthritic or WT littermate mice on an ultrastructural level with transmission electron microscopy (Figure 3C). In WT mice, SFs contained prominent nuclei, abundant rough endoplasmic reticulum (r-ER), and mitochondria of different shapes and sizes. In contrast, remarkable modification of the SFs was noticed in the joints of the arthritic mice, where most randomly flattened cells had an elongated shape characterized by dilation of the r-ER and by swollen mitochondria with distorted cristae.
Therefore, it seems that one of the pathogenic mechanisms in RA is the promotion of actin polymerization and rearrangement of the actin cytoskeleton. Gelsolin (Gsn) is a gene that maps in one of the CIA QTLs (see Figure 1), and its expression was found down-regulated in arthritic SFs by both subtractive libraries and microarray hybridizations (see Table 3). Its deregulated expression was confirmed by real-time RT-PCR in both mouse (see Figure S6D) and human samples (Figure 2), as well as by Western blot in immortalized SFs (see Figure S6E). The gene encodes an actin-binding protein with filament-severing properties [18], and Gsn
−/− fibroblasts have excessive actin stress fibers [18], very similar to the ones observed in arthritic SFs (see Figure 3A). In order to prove the involvement of cytoskeletal organization in the pathogenesis of RA and to highlight the role of gelsolin in it, the arthritic mice (Tg197, hTNF
+/−) were mated with the gsn knockout mice (gsn
−/−) [18]. Knocking out gsn expression from SFs should promote RA pathogenesis by inhibiting the severing activity of gelsolin. Indeed, and as shown in Figure 4, abolishing gsn expression resulted in hyperplasia of the synovial membrane and exacerbation of the disease.
Figure 4 Knocking out Gelsolin Expression Results in Disease Exacerbation
(A) Histopathological scores. * p = 0.025 n = 8–11.
(B) Representative histopathological analysis (stained with haematoxylin-eosin) of arthritic joints. Shown images were assembled from multiple overlapping sections. Arrows indicate the synovial membrane.
Discussion
Twin Expression Profiling, Statistical Analysis, and Validation
Expression profiling, the relative quantification of the expression levels of thousands of genes simultaneously, is one of the most promising approaches for understanding mechanisms of differentiation, development, and disease. However, the small number of samples usually employed substantially limits statistical analysis and precludes application of complex multivariate methods, which would be more appropriate for polygenic diseases such as RA. As a consequence, results should be confirmed by an independent biological method, which largely diminishes the high-throughput nature and discovery rate of any given approach.
To overcome these problems and increase the discovery rate, we used a twin high-throughput approach composed of subtractive cDNA libraries and oligonucleotide DNA microarray hybridizations. Both methodologies are governed by different constraints: random chance for the libraries' robotic clone selection and chip design for the microarrays. Consequently, both methodologies are completely uncorrelated towards the propagation of error. Therefore, the intersection of their statistically significant deregulated gene lists is expected to have a very low false discovery rate. To illustrate this and to exhibit the validity of the twin high-throughput approach, a number of representative genes from the commonly selected list was confirmed by a number of different methods, such as automated literature search and SQ and quantitative RT-PCR in both mouse and human samples. Therefore, we have shown that coupling two different high-throughput approaches largely decreases the need for independent confirmation and consequently increases the number of likely deregulated genes.
All differentially expressed genes were mapped to the chromosomes, together with the known QTL for an animal model of arthritis, CIA [14]. Eight deregulated genes mapped to CIA QTL (WJ: Ctss, Pitpnm, Ncf1, Psmb8, and Siat8e; SF: Rab14, Aqp1, and Gsn; see Figure 1), suggesting that these genes may have a dominant influence in arthritic processes, irrespective of the inciting stimulus (autoimmune or inflammatory). To that end, it was recently reported that a naturally occurring polymorphism of Ncf1 (an NADPH oxidase subunit; QTL Cia 14) regulates arthritis severity [19] and is currently being commercially exploited in a drug discovery programme. Similarly, knocking out Aqp1 (which encodes a water channel protein; QTL Cia 6) expression revealed its fundamental role in cell migration—central to wound healing and tumour spread [20]. Interestingly, several of the genes cluster together in adjacent regions of the chromosomes (4, 7, 10, 11 and 13). These loci could define new QTL or imply common regulatory control at the chromatin level.
Actin Cytoskeleton and RA Pathogenesis
For polygenic diseases such as RA, knowledge about concerted gene functions or cellular processes might provide valuable clues and help to prioritize targets. In this context, we searched for deregulated processes rather than genes, based on functional annotations of deregulated genes, as these are formalized through the controlled vocabulary of the Gene Ontology Consortium [17]. GO term frequencies were calculated in the selected gene list and their statistical significance was estimated. The resulting list (Table S3) includes a number of expected functions that encompass accumulated knowledge about the pathogenesis of the disease, such as collagen catabolism, complement activation, and immune and stress responses. More importantly, five out of 26 predicted deregulated GO functions concerned (directly or indirectly) the actin cytoskeleton, thus forming a valid hypothesis to be explored. Accordingly, F-actin stress fibers were visualized in vitro, in arthritic and WT SFs, both primary and immortalized. As is evident in Figure 3A, there was a striking difference, with the arthritic SFs exhibiting pronounced stress fibers. Notably, stress fibers appear in differentiated fibroblasts called myofibroblasts, specialized contractile fibroblasts with an important role in establishing tension during wound healing and pathological contracture [21,22]. This is the first direct indication of the possible presence of myofibroblasts in the arthritic synovium, although it has been previously reported that transforming growth factor-b1 and interleukin-4 can induce a myofibroblastic phenotype to SFs in vitro [23]. Consistent with this notion, we observed that arthritic SFs have more pronounced focal adhesion kinase-positive islands (unpublished data), a prominent feature of myofibroblasts [22]. The actin cytoskeleton interacts bidirectionally with the ECM through receptors (mainly integrins) that possess extracellular binding sites for laminin, collagen, fibronectin, and other ECM components. The formation of intimate, extensive adhesive contacts between cells and ECM results from cooperation between adhesive systems and the actin cytoskeleton and the generation of force across regions of the cell [24]. In this context, myofibroblasts are thought to exert increased tension to the substratum through their increased adhesive capacity, which results in a more flattened, elongated cell shape [22]. Accordingly, we have observed that arthritic SFs adhere to various ECM components with increased affinity in vitro (see Figure 3B), resulting in a more elongated shape in vivo (see Figure 3C), further corroborating the possible existence of myofibroblasts in the arthritic synovium. In accord with this result, the possible presence of myofibroblasts in the human arthritic synovium was recently implied, by immunohistochemical analysis [25].
While the existence of myofibroblasts and their role in the pathogenesis of RA remain to be further explored, the fact remains that the reorganization of the actin cytoskeleton and the associated deregulation of ECM adhesion seem to be an intrinsic property of arthritic SFs. To this end, it was recently reported that the prevalence of specific autoantibodies against cytoskeletal antigens is elevated in patients with RA [26]. Autoantibodies serve as important serological markers in the diagnosis of various autoimmune and connective tissue diseases, including RA [27,28]. A large number of RA-specific autoantibodies of high diagnostic value are directed against components of the cytoskeleton: anti-vimentin, anti-keratin, and anti-filaggrin [28]. Filaggrin is a keratin cross-linker, an intermediate filament-aggregating protein, that can affect other cytoskeletal elements, including actin microfilaments, by a mechanism similar to actin filament severing by gelsolin [29]. Most of the above-mentioned autoantibodies recognize the citrullinated form of cytoskeletal proteins [30]. Since citrullination of proteins is not specific for RA, our results may provide the molecular basis, by a yet unknown mechanism, for the presence of anti-cytoskeleton antibodies in RA.
Recent experiments have shown that the cytoskeleton plays a critical role in the regulation of various cellular processes linked to cell transformation and tumorigenesis, such as contact inhibition and anchorage-independent cell growth [31]. Accumulated evidence suggesting that arthritic SFs also exhibit characteristic of transformed cells led to the working hypothesis that the arthritic synovium is a locally invasive tumour [32]. The rearranged cytoskeleton in arthritic SFs therefore reinforces the concept of a transformed-like character of the SF and opens up new directions in the pharmacological treatment of RA.
Gelsolin is an actin-binding protein [33] that has been implicated, among others, in the transduction of signals into dynamic rearrangements of the cytoskeletal architecture. In the presence of calcium, gelsolin severs preexisting actin filaments and caps them, thereby preventing monomer addition to their fast-growing ends. The barbed end cap is highly stable, even in the absence of calcium, unless displaced by interactions with regulatory phospholipids such as phosphatidylinositol-4,5-bisphosphate. In the presence of a large pool of profilin (another actin-binding protein), or under depolymerizing conditions, these gelsolin-capped ends allow the disassembly of populations of actin filaments by subunit loss from the pointed ends [34]. Gsn
−/− fibroblasts were found to have excessive stress fibers in vitro [18], similar to the ones observed in arthritic SFs, where gelsolin was found to be down-regulated by a variety of methods. Knocking out Gsn expression from the arthritic mice resulted in exacerbation of the disease (Figure 4), therefore proving the participation of the actin cytoskeleton rearrangement in the pathophysiology of the disease. Extending the similarities of the arthritic synovium with tumours, gelsolin was found to be one of the most strikingly down-regulated markers upon malignant transformation of fibroblasts by Ras [35], while overexpression of a gelsolin mutant was shown to suppress Ras-induced transformation [36]. Its expression was undetectable or reduced in a majority of human gastric, bladder, lung, colon, and breast tumours [37–39].
Conclusion
We have shown that the combination of sequencing of subtractive cDNA libraries and microarray expression analysis is highly reliable and yields self-validated targets and, when integrated with other functional genomics approaches such as genetic linkage and GO-assisted functional discovery, can provide novel insights into the pathophysiology of RA. As an example, we have investigated one of the predicted deregulated cellular processes, cytoskeletal organization, and one of the predicted deregulated genes involved in these processes, gelsolin, to show that gelsolin-driven actin cytoskeleton rearrangement is a novel pathophysiological determinant of RA.
Materials and Methods
Animals.
All mice were bred at the animal facilities of the Alexander Fleming Biomedical Sciences Research Center under specific pathogen-free conditions, in compliance with the Declaration of Helsinki principles. Mice were housed at 20–22 °C, 55 ± 5% humidity, and a 12-h light-dark cycle; water was given ad libitum. “Arthritic” transgenic mice (Tg197, hTNF +/− maintained on a mixed CBA × C57BL/6 genetic background for over 20 generations), carried the human TNF gene where the 3′ UTR was replaced by the corresponding one from b-globin [5]. “Arthritic” knockin mice (TNF ΔARE/+ maintained on a 129/C57BL6 background for over 20 generations) expressed the endogenous mTNF gene, where 69 bp encompassing the TNF ARE (AU-rich elements) at the 3′ UTR have been deleted, resulting in increased message stability and translational efficiency [8].
Cell isolation and culture.
SFs (VCAM+) were isolated from 6- to 8-wk-old mice essentially as previously described [7]. Fibroblasts were selected by continuous culturing for at least 21 d and a minimum of four passages. No macrophage markers could be detected by FACS analysis. Cells were grown at 37 °C, 5% CO2 in complete DMEM medium (Gibco-BRL, San Diego, California, United States) supplemented with 10% FBS and 100 U/ml of penicillin-streptomycin. The creation and culture conditions (33 °C, 10 U/ml of murine rIFN-γ) of conditionally immortalized SF cell lines has been described previously [7].
RNA extraction.
Total RNA was extracted from subconfluent (70–80%) cultured SFs (primary or conditionally immortalized) with the RNAwiz reagent (Ambion, Austin, Texas, United States), followed by single passage through an RNeasy column (QIAGEN, Hilden, Germany) according to manufacturers' instructions. Total RNA was extracted from WJs using the guanidinium isothiocynate-acid phenol protocol [40], followed by a single passage through an RNeasy column. RNA integrity was assessed by electrophoresis on denaturing 1.2% agarose-formaldehyde gels. RNA quantity and quality were calculated based on OD readings at 260/280 nm.
Generation of full-length cDNA libraries.
First-strand cDNA synthesis was performed using SUPERSCRIPT II (Life technologies/Invitrogen) at 56 °C in the presence of trehalose and sorbitol. The cap structure at the 5′ end of mRNAs was biotinylated, and full-length cDNAs were selected, after RNAse I treatment, using streptavidin-coated beads [41]. Second-strand synthesis was primed by the single-strand linker ligation method (SSLLM), where a double-stranded DNA linker (with random 6-bp, dN6 or dGN5, 3′ overhangs) is ligated to the single-stranded cDNA [42]. The second strand is made by primer extension using mixtures of long-range thermostable polymerases, followed by restriction digestion (BamHI and XhoI) and ligation to the λ-FLCI phagemid vector [42,43]. After packaging, cDNA libraries were amplified on solid phase, as previously described [44].
cDNA library normalization and subtraction.
The procedure is outlined in Figure S1A. Amplified phagemid λ-FLCI cDNA libraries were used to infect the Cre-expressing bacterial cell line BNN-132 and the excised plasmids were isolated [43]. Double-stranded plasmid DNA was nicked by the site-specific (f1 origin) endonuclease GeneII and converted to circular single-stranded form by digestion with exonuclease III [44]. Circular, single-stranded plasmid cDNA libraries (tester libraries) were then subjected to a single normalization-subtraction step, with PCR-derived single strand antisense DNA drivers produced from these libraries [44]. Normalization refers to low CoT (reassociation rate) hybridization (CoT = 2) with drivers produced from the “self” library, and it aims to decrease the representation of highly expressed mRNAs. Subtraction refers to high CoT hybridization (CoT = 100~200) with drivers produced from different libraries, and it aims to remove mRNAs common in both populations. Hybridized double-stranded cDNAs were removed with two passes through a hydroxyapatite column. Nonhybridized, single stranded cDNAs were converted to double stranded and were subsequently electroporated into DH10b bacterial cells, where only tester cDNAs are able to propagate (due to the presence of a replication origin and antibiotic resistance) [44].
High-throughput sequencing and sequence analysis.
Colony picking, cDNA sequencing and sequence analysis were performed essentially as previously described [45–47]. Sequences were filtered for primer and vector sequences [47] and masked for rodent-specific and mammalian-wide repeats with RepeatMasker (http://www.phrap.org). EST clustering was performed with stackPACK and d2_cluster (word size 6, similarity cutoff 0.98, minimum sequence size 50, window size 200) [48,49]. Homology searches with known genes were performed with BLAST [50,51] in the Unigene (http://www.ncbi.nlm.nih.gov/UniGene), FANTOM2 (http://fantom2.gsc.riken.go.jp) [52], and SWISSPROT (http://www.ebi.ac.uk/swissprot/ ) databases. BLAST results were associated with GO terms (http://www.geneontology.org) at http://source.stanford.edu/cgi-bin/sourceSearch (for Unigene), ftp://ftp.ebi.ac.uk/pub/databases/GO/goa/ SPTR/gene_association.goa_sptr.gz (for SWISSPROT), and http://fantom2.gsc.riken.go.jp (for FANTOM2). The detailed results (including clone numbers, clusters/genes, BLAST results and E values, accession numbers, and GO assignments) can be found in the corresponding author's Web site, at http://www.fleming.gr/en/investigators/Aidinis/data.html. Most of the sequencing data, have been already submitted to public databases [46], in the context of the ongoing FANTOM (Functional annotation of the mouse) project [52,53].
Differential gene expression statistical analysis of subtractive libraries.
The differential gene expression/abundance among the different subtracted cDNA libraries was calculated with a single statistical test designed especially for this purpose [54]. Essentially, the formula is the entropy of a partitioning of genes among cDNA libraries and is described by an R-value, which is the log likelihood ratio statistic, that follows (2Rj) an asymptotic χ2 distribution [54]. The formula for the statistic Rj for the jth gene is given by the expression:
where
where m is the number of cDNA libraries, N
i is the total number of clones sequenced in the ith library, χi,j the counts (transcript copies) of the jth gene in the ith library, and fi is the frequency of gene transcripts copies of the jth gene in all the libraries.
The significance of the R-statistic was established by utilizing a resampling method that tries to establish an “optimal” cutoff value using simulated library datasets based on the observed counts. The method consists of the following steps. (1) For every gene, the common gene transcript frequency fj is calculated with the help of Equation 2. This number corresponds to the expected frequency of gene transcripts under the null hypothesis of no difference in the abundance of gene j across all libraries. (2) The parameters of the Poisson distributions giving the sampling distribution of clone abundance for each gene and library are calculated under the null hypothesis. This parameter is equal to Ni × fj for gene j in library i, where Ni is the total number of clones (taking into account all genes) sequenced in library i and fj was calculated in step 1. This value is equal to the expected absolute number of clones of the jth gene in the ith library under the null hypothesis. (3) Then, for each library compared, and each gene, a random number is generated from the Poisson distribution of step 2. This number is a simulated count compatible with the null hypothesis, i.e., that the gene frequency for a particular gene is the same across all libraries. This step uses the random number-generating function of the computer to create an artificial dataset corresponding to the actual experiment of library creation and analysis for each of the libraries compared under the null hypothesis. (4) For gene j in the artificial dataset of step 3, the test statistic R is calculated by substituting for χi,j the random number generated in step 3, for fj the frequency calculated in step 1, and for Ni the total number of clones sequenced. (5) The R-values calculated in step 4 are sorted in descending order, and for a range of values for specificity (i.e., the true negative rate), the corresponding R-value is found. These are by definition true negatives, since they were obtained from libraries under the null hypothesis. (6) Steps 3–5 are repeated 1,000 times. The resulting data allowed us to construct the histogram shown in Figure S2A. These histograms depict the distribution of the R-value as a function of the required true negative rate cutoff. The mean of this (empirical) distribution is used as the R-value cutoff for the analysis of the experimental dataset. (7) The experimental dataset R-values are computed for all the comparative libraries using the observed clone counts. Genes with R-values greater than the previously established cutoff are considered to be differentially regulated between the libraries. Calculations were performed in the Computer Algebra System Mathematica 4.2 (http://www.wri.com) with the MathStatica add-on [55].
High density oligonucleotide array hybridization.
cRNA probes were generated from 5 μg of total RNA and hybridized to the Mu11K (subunits A and B) chip set according to the manufacturer's instructions (Affymetrix, Santa Clara, California, United States) and as previously described [7]. The chip set is designed to collectively recognize 13,179 distinct murine transcripts, where the expression level of any given gene is interrogated by 40 oligonucleotides, 20 with a perfect match sequence and 20 carrying a single mismatch. Hybridized chips were washed and stained in a Fluidics Station (Affymetrix) following standard protocols; subsequently, fluorescence intensities were read by an Affymetrix scanner. All MIAME-compliant microarray data can be downloaded from the ArrayExpress database (accession number E-MEXP-255), as well as from the corresponding author's web site (http://www.fleming.gr/en/investigators/Aidinis/data.html).
Microarray data preprocessing and normalization.
The procedure is outlined in Figure S3. Low-level analysis of the resulting scanned image (as background subtraction and computation of individual probe cell intensities) was automatically performed by MicroArray Suite 5.0 (MAS5) software (Affymetrix). Briefly, the quantitative level of expression (signal value), as well a qualitative measure of expression (detection call), of any given gene is calculated by proprietary Affymetrix algorithms from the combined, background-adjusted hybridization intensities of 20 perfect match and 20 single mismatch oligonucleotides (probe set). All signal values from all chips in the experiment were scaled to reach target intensity (TGT) of 500 or 2,500, following Affymetrix recommendations for the individual chip sets used. Scaled values were then submitted to a normalization step that is intended to force each chip subunit's signal distribution to have an identical overall shape across chips of the same subunit. Gene expression values were reorganized into two distinct data matrices, one for each chip set subunit (A,B), where rows and columns represented genes and chips, respectively. The columns of the data matrix were normalized to have the same quantiles using BioConductor [56]. Each quantile of each column was set to the mean of that quantile across arrays (see Figure S4A). The rows of the two data matrices were then merged, and the columns of the resulting data matrix were split according to the corresponding animal model of RA (Tg197, TNFΔARE) and to the type of sample (SF or WJ), thus obtaining a total of four distinct data matrices. Finally, genes that were not called “present” (detection call) or that had a normalized signal lower than one-fifth of the TGT in at least one sample were not further considered. Subsequent analysis was therefore performed on the remaining 8,021 probe sets.
Microarray data analysis.
Comparisons of Affymetrix chips hybridizations and real-time PCR have indicated that chip analyses are accurate and reliable, and that they underestimate differences in gene expression [57]. Nevertheless, in order to assess the system's performance, the reproducibility of the (technical) duplicate samples was examined, in terms of Pearson correlation coefficient of normalized signals, as well as detection concordance (expressed as the percentage of probe sets that were consistently detected with a “present” or an “absent” call in both chips). As can be seen in Figure S4B, Affymetrix data were highly reproducible and reliable.
As previously reported [58,59], the significance of fold-changes, which are commonly used as a measure of differential expression, is highly dependent on the expression level of the corresponding genes. Therefore, in order to identify genes that are significantly differentially expressed, we decided not to use a fixed fold-change threshold, but instead to model intensity-dependent fold-changes between replicated chips using a variant of a previously described approach [60]. Briefly, we first calculated fold-changes in pairwise comparisons of replicated chips representing the transcriptome of the same sample type (SF or WJ) measured (under the same experimental condition using the following definition:
where x
i was the normalized expression level of a given gene in the ith chip. Genes were then ranked according to the minor of their two expression values—i.e., min(x1, x2)—and the overall expression range was partitioned into ten intervals, each containing an equal number of probe sets. In each partition the median of all min(x1, x2) values as well as the 90th and the 95th percentile fold-change were determined, thus obtaining ten distinct modeling points for each sample type (SF or WJ) and for each of the two significance levels (90% or 95%). Based on the observation that measurement variability of high-density oligonucleotide microarrays depends on signal intensity following a power law [11], a continuous FCM was derived from the ten modeling points using a least squares linear fit in log-log plots. The following modeling parameters have been obtained for the two FCMs (see Figure S4C): a slope of −0.327 and an intercept of 2.356 for the SF FCM, and a slope of −0.346 and an intercept of 2.678 for the WJ FCM. This means that WJs are, as expected, intrinsically more variable than SFs. Using slopes (a90% or a95%) and intercepts (b90% or b95%) of the resulting regression curves, we could obtain for each given sample type (SF or WJ) and for any given minimum expression level both a 90% and a 95% significance threshold:
For each of the four data matrices (i.e., Tg197/SF, TNFΔARE/SF, Tg197/WJ, and TNFΔARE/WJ), an observed average fold-change between experimental conditions (diseased or normal) was calculated for each gene in the following way:
where μi is the average expression level of a given gene in the ith experimental condition. If no replicates were available, the single expression value of the given gene was used instead of the average. For each gene, this observed fold-change was then compared to the fold-change that was expected at a 90% (or 95%) significance level using the corresponding “sample type”-specific FCM, given the observed values of μi of that gene:
Finally, genes with FCobs > FC90% (or > FC95%) were selected as differentially expressed at a significance level of 90% (or 95%, respectively).
Visualization of QTL and gene-expression data.
This was performed with Expression view [61] at http://ensembl.pzr.uni-rostock.de/Mus_musculus/expressionview. QTL data were derived from Serrano-Fernandez et al. [62].
Functional clustering and determination of statistical significance.
Biological annotation, in the form of GO “Biological process” term [17], for each of the genes (probe sets) in Table 3, was obtained from the NetAffx portal (http://www.affymetrix.com). We then calculated the observed GO term frequencies, as means to discover deregulated functions (Table S3). The statistical significance of GO term frequencies was determined essentially as has been previously described [63]. Briefly, the hypergeometric distribution was used to obtain the chance probability of observing a given number of genes annotated in NetAffx with a particular GO term and then calculating appropriate p-values. More specifically, the probability of observing at least k probe sets annotated with a particular GO term within a list of selected probe sets of size n was calculated [63] as:
where f was the total number of genes within a functional category, and g was the total number of probe sets on the chip set (13,179).
Adhesion assays.
These assays were performed on Cytometrix adhesion strips (Chemicon, Temecula, California, United States) coated with human fibronectin, vitronectin, laminin, and collagen I, according to the manufacturer's instructions. Briefly, cells (in triplicates) were allowed to adhere to the above-mentioned substrates, and unbound cells were removed with sequential washes with PBS containing Ca++ and Mg++. Adhered cells were then stained with crystal violet, solubilized, and their absorbance determined at 570 nm.
Immunofluorescence.
Cells were fixed using 4% paraformaldehyde in PBS and stained by standard methods. For visualizing F-actin cells were stained with Alexa594-phalloidin (Molecular Probes, Eugene, Oregon, United States). P-tyrosine was detected using the 4G10 antibody (Upstate Biotechnology, Waltham, Massachusetts, United States), focal adhesion kinase was detected using a monoclonal antibody (clone 77, BD Transduction Laboratories, Lexington, Kentucky, United States). Antibodies against gelsolin were raised by immunizing rabbits with recombinant mouse gelsolin; immune serum was used at 1:500 dilution. For Western blot analysis, 10 μg of total cell protein was separated by SDS-PAGE, transferred to Immobilon-P membrane, probed for gelsolin, and re-probed for actin as an internal loading control.
RT-PCR.
Total RNA was extracted from SF and WJ tissue using TRIzol reagent (Life Technologies, Rockville, Maryland, United States) in accordance with the manufacturer's instructions. RNA yield and purity were determined spectrophotometrically at 260/280 nm. First-strand cDNA synthesis was performed using the MMLV reverse transcriptase and oligo-dT15 (Promega, Madison, Wisconsin, United States).
SQ PCR was performed by 20–25 cycles of denaturation at 95 °C for 30 s, annealing at 57–62 °C (depending on the Tm of each individual set of primers) for 30 s, and extension at 72 °C for 1 min, in a final volume of 20 μl. The products were separated by electrophoresis on 1.5% agarose gel and stained with ethidium bromide. Product intensity was quantified with GelWorks 1D Advanced (v. 4.01) and normalized to the intensity of B2m and/or L32. The primers were selected to span two exons, while the two control primers were chosen from the Primer Bank database (http://pga.mgh.harvard.edu/primerbank/). Primer sequences (listed in the 5′ to 3′ direction, and designated as s, sense, and as, antisense) and product sizes (in bp) were as follows: Aqp1 (s, TCACCCGCAACTTCTCAAAC; as, AGCTCTGAGACCAGGAAACA, 400), Bsg (s, ATGAGAAGAGGCGGAAGCCA; as, CCACTCCACAGGGCTGTAGT, 426), Cd14 (s, CAATCCTGAATTGGGCGAGA; as, CGAGTGGGATTCAGAGTCCA, 400), Cdc42 (s, AAGTGGCCCAGATCCTGGAA; as, AGCACTGCACTTTTGGGGTT, 380), Gsn (s, TGCAGGAAGACCTGGCTACT; as, ATGGCTTGGTCCTTACTCAG, 300), Hp (s, GAAGCAATGGGTGAACACAG; as, GGGGTGGAGAACGACCTTCT, 331), Marco (s, CACAGGAATTCAAGGACAGA; as, ATTGTCCAGCCAGATGTTCC, 397), Mglap (s, CAGTCCCTTCATCAACAGGA; as, CTGCAGGAGATATAAAACGA, 274), Mb (s, TCACACGCCACCAAGCACAA; as, TGGGCTCCAGGGTAACACTG, 354), Ptgis (s, TCACAGATGACCACACTCCC; as, GCAGTAGGACGACAAATTGT, 403), B2M (s, TTCTGGTGCTTGTCTCACTGA; as, CAGTATGTTCGGCTTCCCATTC, 104), and L32 (s, TTAAGCGAAACTGGCGGAAAC; as, TTGTTGCTCCCATAACCGATG, 100).
Quantitative real-time PCR was performed using the iCycler iQ Real-Time detection system and the IQ SYBR Green Supermix (Bio-Rad Laboratories, Hercules, California, United States), according to the manufacturer's instructions, for one cycle of 94 °C for 4 min, and 40 cycles of denaturation at 95 °C for 50 s and annealing at 57–62 °C for 50 s. Primers were chosen from exons separated by large introns, and the PCR quality and specificity was verified by melting curve analysis and gel electrophoresis. Values were normalized to B2m and/or L32 (using the same primers as for the SQ PCR). Mouse (m) and human (h) primer sequences and expected lengths were as follows (listed in the 5′ to 3′ direction, and designated as s, sense, and as, antisense): mAqp 1 (s, TCACCCGCAACTTCTCAAAC; as, TCATGCGGTCTGTGAAGTCG, 123), mGsn (s, TGCAGGAAGACCTGGCTACT; as, TCGATGTACCGCTTAGCAGA, 130), hAqp 1 (s, CTCCCTGACTGGGAACTCG; as, GGGCTACAGAGAGGCCGAT, 182), hBsg (s, TTCCTGGGCATCGTGGCTGA; as, GCGGACGTTCTTGCCTTTGT, 159), hCd14 (s, CGGCGGTCTCAACCTAGAG; as, GCCTACCAGTAGCTGAGCAG, 142), hCdc42 (s, AATTGATCTCAGAGATGACC; as, TTTAGGCCTTTCTGTGTAAG, 150), hGsn (s, GGTGTGGCATCAGGATTCAAG; as, TTTCATACCGATTGCTGTTGGA,199), hMarco (s, TGGGACGAGATGGAGCAAC; as, CCCTTAGTTCCAGTTTCCCCTT, 193), hMb (s, TTGGTGCTGAACGTCTGGG; as, CTGTGCCAGGGGCTTAATCTC, 249), hB2M (s, CTGAAAAAGATGAGTATGCC; as, ACCCTACATTTTGTGCATAA, 202) and hL32 (s, TTAAGCGTAACTGGCGGAAAC; as, GAGCGATCTCGGCACAGTAA, 210).
Cycle threshold (Ct; the first cycle in which amplification can be detected) values were obtained from the iCycler iQ software for each gene of interest (GOI) and control housekeeping genes (HKG; L32 and/or b2m), together with amplification efficiencies (η; 80–120%). For the mouse samples, we calculated the relative expression of the samples to WT controls as reference samples using the gene expression-relative quantification Microsoft Excel add-on macro (Bio-Rad) that utilizes the following formulas: relative expression = 2−(S ΔCt-R ΔCt), where ΔCt = GOI Ct − HKG Ct. For the human samples, Ct values were converted to concentration values (ng/ml) by utilizing the standard curve made by serial dilutions (in duplicates) of a reference sample. Values were normalized to the corresponding values of the reference (housekeeping) gene(s) and presented (in logarithmic scale for visualization purposes) as expression index.
Arthritic score and histopathology.
Paraffin-embedded joint tissue samples were sectioned and stained with haematoxylin and eosin. Arthritic histopathology was assessed (in a blinded fashion) separately for synovial hyperplasia, existence of inflammatory sites, cartilage destruction, and bone erosion using a semiquantitative (0–5) scoring as described previously [64].
Transmission electron microscopy.
Ankle joints (dissected from the right hind leg of each mouse—three Tg197 and three WT) were split open longitudinally through the midline between the tibia and the talus, and were pre-fixed with 2.5% glutaraldehyde in PBS (pH 7.4) overnight. After post-fixing with 1% osmium tetroxide in PBS for 2 h, the samples were dehydrated in a graded series of ethanol and processed into Araldite. Semithin sections (1.0 μ) were cut and stained with methylene blue to observe the orientation under the light microscope. Ultrathin sections (80–90 nm) were then cut with an ultramicrotome (Riechert-Jung Ultracut E, Leica, Wetzlar, Germany), mounted on copper grids, counterstained with uranyl acetate and lead citrate, and evaluated by electron microscope (CM120 Biotwin, FEI Company, Hillsoro, Oregon, United States).
Supporting Information
Figure S1 Subtractive cDNA Libraries and Large-Scale Sequencing
(A) Outline of experimental strategy for the preparation of subtracted cDNA libraries and analysis of differential expression.
(B) Summary of normalized and subtracted cDNA libraries and sequencing results.
(1.2 MB PDF)
Click here for additional data file.
Figure S2 Calculation of the Statistically Significant Thresholds of the R-Statistic at Different Specificities
(A) Distribution of the R-statistic for the indicated specificities and pairwise library comparisons.
(B) Tabulated R-value cutoffs for the indicated specificities and pairwise library comparisons.
(1.4 MB PDF)
Click here for additional data file.
Figure S3 Microarray Data Normalization and Analysis Outline
(775 KB PDF)
Click here for additional data file.
Figure S4 Microarray Data Normalization, Evaluation, and Statistical Selection
(A) Quantile normalization on separated oligonucleotide chip subunits.
(B) Reproducibility of technical duplicate samples in Affymetrix hybridizations.
(C) “Sample type”-specific FCMs derived from replicated chips.
(2.2 MB PDF)
Click here for additional data file.
Figure S5 Deregulated Genes Previously Reported to Be Associated with RA
The associations were identified through the Biolab Experiment Assistant text-mining software.
(A) PubMed identification numbers of corresponding publications.
(B) Schematic representation of text mining results. Red and green indicate up-regulated and down-regulated genes, respectively; black and blue indicate WJ and SF, respectively; numerical values indicate number of PubMed references.
(1.5 MB PDF)
Click here for additional data file.
Figure S6 Validation of Expression Profiling Results
(A–C) SQ RT-PCR of the indicated genes from different cDNA amounts of arthritic or WT WJs, primary SFs (pSF) and immortalized SFs (iSF). Ethidium bromide-stained PCR products were quantified with GelWorks 1D Advanced (v. 4.01) software and were normalized against the expression of B2M.
(D) Real-time RT-PCR of the indicated genes from the indicated samples.
(E) Western blot of whole cell extracts (two independent preparations) from arthritic and WT immortalized SFs probed with antibodies to gelsolin and actin. Gelsolin immunostaining was quantified with GelWorks 1D Advanced (v. 4.01) software and normalized against the corresponding actin intensities.
(2.0 MB PDF)
Click here for additional data file.
Table S1 Differential Expression Results from Subtractive Libraries at 99.99% Specificity
(32 KB XLS)
Click here for additional data file.
Table S2 Differentially Expressed Genes Common in Two Animal Models of RA from Oligonucleotide DNA Microarray Hybridizations at 95% Specificity
(59 KB XLS)
Click here for additional data file.
Table S3 GO “Biological Process” Term Frequencies for the Differentially Expressed Genes in Arthritic SFs and WJs, Selected from Both Subtractive Libraries and Microarrays
(24 KB XLS)
Click here for additional data file.
This work was supported by: European Commission grant (QLG1-CT-2001–01407) to GK, VA, and HJT; Research Grant for Advanced and Innovational Research Program in Life Science to YH; Core Research for Evolutional Science and Technology grant of Japan Science and Technology Corporation to YH; and Greek Ministry of Development/General Secretariat of Research grant (GGET/IAP 041) to VA. Clinical samples (synovial tissue samples) were provided by the BMBF-Leitprojekt “Proteom-Analyse des Menschen” (FKZ01GG9831–44). Patients informed consent was obtained according to the national ethical rules, as these were implemented by the corresponding local ethic commission.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. VA and GK conceived and designed the experiments. VA, PC, MA, WW, VH, DK, and MMT performed the experiments. VA, VH, NP, CA, MMT, SM, and WK analyzed the data. VA, WW, PG, PRC, HJT, YH, and GK contributed reagents/materials/analysis tools. VA wrote the paper.
Abbreviations
CIAcollagen-induced arthritis
DEGdifferentially expressed gene
ECMextracellular matrix
FCMfold-change model
GOgene ontology
RArheumatoid arthritis
r-ERrough endoplasmic reticulum
SFsynovial fibroblasts
SQsemiquantitative
Tgtransgenic
TNFtumour necrosis factor
Qquantitative
QTLquantitative trait loci
WJwhole joint
WTwild-type
==== Refs
References
Choy EH Panayi GS 2001 Cytokine pathways and joint inflammation in rheumatoid arthritis N Engl J Med 344 907 916 11259725
Smolen JS Steiner G 2003 Therapeutic strategies for rheumatoid arthritis Nat Rev Drug Discov 2 473 488 12776222
Firestein GS 2003 Evolving concepts of rheumatoid arthritis Nature 423 356 361 12748655
Lindqvist AKB Bockermann R Johanson ACM Nandakumar KS Johannensson M 2002 Mouse models for rheumatoid arthritis Trends Genet 18 S7 S13
Keffer J Probert L Cazlaris H Georgopoulos S Kaslaris E 1991 Transgenic mice expressing human tumour necrosis factor: A predictive genetic model of arthritis EMBO J 10 4025 4031 1721867
Li P Schwarz EM 2003 The TNF-alpha transgenic mouse model of inflammatory arthritis Springer Semin Immunopathol 25 19 33 12904889
Aidinis V Plows D Haralambous S Armaka M Papadopoulos P 2003 Functional analysis of an arthritogenic synovial fibroblast Arthritis Res Ther 5 R140 157 12723986
Kontoyiannis D Pasparakis M Pizarro TT Cominelli F Kollias G 1999 Impaired on/off regulation of TNF biosynthesis in mice lacking TNF AU-rich elements: Implications for joint and gut-associated immunopathologies Immunity 10 387 398 10204494
Brazma A Hingamp P Quackenbush J Sherlock G Spellman P 2001 Minimum information about a microarray experiment (MIAME)—Toward standards for microarray data Nat Genet 29 365 371 11726920
Brazma A Parkinson H Sarkans U Shojatalab M Vilo J 2003 ArrayExpress—A public repository for microarray gene expression data at the EBI Nucleic Acids Res 31 68 71 12519949
Pavelka N Pelizzola M Vizzardelli C Capozzoli M Splendiani A 2004 A power law global error model for the identification of differentially expressed genes in microarray data BMC Bioinformatics 5 203 15606915
Irizarry RA Warren D Spencer F Kim IF Biswal S 2005 Multiple-laboratory comparison of microarray platforms Nat Methods 2 345 350 15846361
Marshall E 2004 Getting the noise out of gene arrays Science 306 630 631 15499004
Trentham DE Townes AS Kang AH 1977 Autoimmunity to type II collagen an experimental model of arthritis J Exp Med 146 857 868 894190
Neumann E Kullmann F Judex M Justen HP Wessinghage D 2002 Identification of differentially expressed genes in rheumatoid arthritis by a combination of complementary DNA array and RNA arbitrarily primed-polymerase chain reaction Arthritis Rheum 46 52 63 11817609
van der Pouw Kraan TC van Gaalen FA Kasperkovitz PV Verbeet NL Smeets TJ 2003 Rheumatoid arthritis is a heterogeneous disease: Evidence for differences in the activation of the STAT-1 pathway between rheumatoid tissues Arthritis Rheum 48 2132 2145 12905466
Ashburner M Ball CA Blake JA Botstein D Butler H 2000 Gene ontology: Tool for the unification of biology. The Gene Ontology Consortium Nat Genet 25 25 29 10802651
Witke W Sharpe AH Hartwig JH Azuma T Stossel TP 1995 Hemostatic, inflammatory, and fibroblast responses are blunted in mice lacking gelsolin Cell 81 41 51 7720072
Olofsson P Holmberg J Tordsson J Lu S Akerstrom B 2003 Positional identification of Ncf1 as a gene that regulates arthritis severity in rats Nat Genet 33 25 32 12461526
Saadoun S Papadopoulos MC Hara-Chikuma M Verkman AS 2005 Impairment of angiogenesis and cell migration by targeted aquaporin-1 gene disruption Nature 434 786 792 15815633
Gabbiani G 2003 The myofibroblast in wound healing and fibrocontractive diseases J Pathol 200 500 503 12845617
Tomasek JJ Gabbiani G Hinz B Chaponnier C Brown RA 2002 Myofibroblasts and mechano-regulation of connective tissue remodelling Nat Rev Mol Cell Biol 3 349 363 11988769
Mattey DL Dawes PT Nixon NB Slater H 1997 Transforming growth factor beta 1 and interleukin 4 induced alpha smooth muscle actin expression and myofibroblast-like differentiation in human synovial fibroblasts in vitro: Modulation by basic fibroblast growth factor Ann Rheum Dis 56 426 431 9486005
Gumbiner BM 1996 Cell adhesion: The molecular basis of tissue architecture and morphogenesis Cell 84 345 357 8608588
Kasperkovitz PV Timmer TC Smeets TJ Verbeet NL Tak PP 2005 Fibroblast-like synoviocytes derived from patients with rheumatoid arthritis show the imprint of synovial tissue heterogeneity: Evidence of a link between an increased myofibroblast-like phenotype and high-inflammation synovitis Arthritis Rheum 52 430 441 15692990
Shrivastav M Mittal B Aggarwal A Misra R 2002 Autoantibodies against cytoskeletal proteins in rheumatoid arthritis Clin Rheumatol 21 505 510 12447636
von Muhlen CA Tan EM 1995 Autoantibodies in the diagnosis of systemic rheumatic diseases Semin Arthritis Rheum 24 323 358 7604300
van Boekel MA Vossenaar ER van den Hoogen FH van Venrooij WJ 2002 Autoantibody systems in rheumatoid arthritis: Specificity, sensitivity and diagnostic value Arthritis Res 4 87 93 11879544
Presland RB Kuechle MK Lewis SP Fleckman P Dale BA 2001 Regulated expression of human filaggrin in keratinocytes results in cytoskeletal disruption, loss of cell-cell adhesion, and cell cycle arrest Exp Cell Res 270 199 213 11640884
Vossenaar ER van Venrooij WJ 2004 Citrullinated proteins: Sparks that may ignite the fire in rheumatoid arthritis Arthritis Res Ther 6 107 111 15142259
Pawlak G Helfman DM 2001 Cytoskeletal changes in cell transformation and tumorigenesis Curr Opin Genet Dev 11 41 47 11163149
Firestein GS 1996 Invasive fibroblast-like synoviocytes in rheumatoid arthritis. Passive responders or transformed aggressors? Arthritis Rheum 39 1781 1790 8912499
dos Remedios CG Chhabra D Kekic M Dedova IV Tsubakihara M 2003 Actin binding proteins: Regulation of cytoskeletal microfilaments Physiol Rev 83 433 473 12663865
McGough AM Staiger CJ Min JK Simonetti KD 2003 The gelsolin family of actin regulatory proteins: Modular structures, versatile functions FEBS Lett 552 75 81 14527663
Vandekerckhove J Bauw G Vancompernolle K Honore B Celis J 1990 Comparative two-dimensional gel analysis and microsequencing identifies gelsolin as one of the most prominent downregulated markers of transformed human fibroblast and epithelial cells J Cell Biol 111 95 102 2164032
Mullauer L Fujita H Ishizaki A Kuzumaki N 1993 Tumor-suppressive function of mutated gelsolin in ras-transformed cells Oncogene 8 2531 2536 8395682
Asch HL Head K Dong Y Natoli F Winston JS 1996 Widespread loss of gelsolin in breast cancers of humans, mice, and rats Cancer Res 56 4841 4845 8895730
Tanaka M Mullauer L Ogiso Y Fujita H Moriya S 1995 Gelsolin: A candidate for suppressor of human bladder cancer Cancer Res 55 3228 3232 7614452
Dosaka-Akita H Hommura F Fujita H Kinoshita I Nishi M 1998 Frequent loss of gelsolin expression in non-small cell lung cancers of heavy smokers Cancer Res 58 322 327 9443412
Chomczynski P Sacchi N 1987 Single-step method of RNA isolation by acid guanidinium thiocyanate-phenol-chloroform extraction Anal Biochem 162 156 159 2440339
Carninci P Hayashizaki Y 1999 High-efficiency full-length cDNA cloning Methods Enzymol 303 19 44 10349636
Shibata Y Carninci P Watahiki A Shiraki T Konno H 2001 Cloning full-length, cap-trapper-selected cDNAs by using the single-strand linker ligation method Biotechniques 30 1250 1254 11414214
Carninci P Shibata Y Hayatsu N Itoh M Shiraki T 2001 Balanced-size and long-size cloning of full-length, cap-trapped cDNAs into vectors of the novel lambda-FLC family allows enhanced gene discovery rate and functional analysis Genomics 77 79 90 11543636
Hirozane-Kishikawa T Shiraki T Waki K Nakamura M Arakawa T 2003 Subtraction of cap-trapped full-length cDNA libraries to select rare transcripts Biotechniques 35 510 516, 518 14513556
Shibata K Itoh M Aizawa K Nagaoka S Sasaki N 2000 RIKEN integrated sequence analysis (RISA) system—384-format sequencing pipeline with 384 multicapillary sequencer Genome Res 10 1757 1771 11076861
Carninci P Waki K Shiraki T Konno H Shibata K 2003 Targeting a complex transcriptome: The construction of the mouse full-length cDNA encyclopedia Genome Res 13 1273 1289 12819125
Konno H Fukunishi Y Shibata K Itoh M Carninci P 2001 Computer-based methods for the mouse full-length cDNA encyclopedia: Real-time sequence clustering for construction of a nonredundant cDNA library Genome Res 11 281 289 11157791
Miller RT Christoffels AG Gopalakrishnan C Burke J Ptitsyn AA 1999 A comprehensive approach to clustering of expressed human gene sequence: The sequence tag alignment and consensus knowledge base Genome Res 9 1143 1155 10568754
Burke J Davison D Hide W 1999 d2_cluster: A validated method for clustering EST and full-length cDNAsequences Genome Res 9 1135 1142 10568753
Altschul SF Gish W Miller W Myers EW Lipman DJ 1990 Basic local alignment search tool J Mol Biol 215 403 410 2231712
Altschul SF Madden TL Schaffer AA Zhang J Zhang Z 1997 Gapped BLAST and PSI-BLAST: A new generation of protein database search programs Nucleic Acids Res 25 3389 3402 9254694
Okazaki Y Furuno M Kasukawa T Adachi J Bono H 2002 Analysis of the mouse transcriptome based on functional annotation of 60,770 full-length cDNAs Nature 420 563 573 12466851
Kawai J Shinagawa A Shibata K Yoshino M Itoh M 2001 Functional annotation of a full-length mouse cDNA collection Nature 409 685 690 11217851
Stekel DJ Git Y Falciani F 2000 The comparison of gene expression from multiple cDNA libraries Genome Res 10 2055 2061 11116099
Rose C Smith MD 2002 Mathematical statistics with Mathematica New York Springer Verlag 496 p.
Ihaka R Gentleman R 1996 A language for data analysis and graphics J Comp Graph Stat 5 299 314
Yuen T Wurmbach E Pfeffer RL Ebersole BJ Sealfon SC 2002 Accuracy and calibration of commercial oligonucleotide and custom cDNA microarrays Nucleic Acids Res 30 e48 12000853
Tsien CL Libermann TA Gu X Kohane IS 2001 On reporting fold differences Pac Symp Biocomput 2001 496 507
Tu Y Stolovitzky G Klein U 2002 Quantitative noise analysis for gene expression microarray experiments Proc Natl Acad Sci U S A 99 14031 14036 12388780
Mutch DM Berger A Mansourian R Rytz A Roberts MA 2002 The limit fold change model: A practical approach for selecting differentially expressed genes from microarray data BMC Bioinformatics 3 17 12095422
Fischer G Ibrahim SM Brockmann GA Pahnke J Bartocci E 2003 Expressionview: Visualization of quantitative trait loci and gene-expression data in Ensembl Genome Biol 4 R77 14611663
Serrano-Fernandez P Ibrahim SM Godde R Epplen J Moller S 2004 Intergenomic consensus in multifactorial inheritance loci: The case of multiple sclerosis Genes Immun 5 615 620 15573086
Tavazoie S Hughes JD Campbell MJ Cho RJ Church GM 1999 Systematic determination of genetic network architecture Nat Genet 22 281 285 10391217
Mould AW Tonks ID Cahill MM Pettit AR Thomas R 2003
Vegfb gene knockout mice display reduced pathology and synovial angiogenesis in both antigen-induced and collagen-induced models of arthritis Arthritis Rheum 48 2660 2669 13130487
|
16254600
|
PMC1270006
|
CC BY
|
2021-01-05 08:00:25
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e48
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010048
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460110.1371/journal.pgen.0010050plge-01-04-07Research ArticleThe Axon Guidance Receptor Gene ROBO1 Is a Candidate Gene for Developmental Dyslexia ROBO1 in DyslexiaHannula-Jouppi Katariina 1Kaminen-Ahola Nina 1Taipale Mikko 12Eklund Ranja 1Nopola-Hemmi Jaana 13Kääriäinen Helena 45Kere Juha 16*1 Department of Medical Genetics, University of Helsinki, Finland
2 European Molecular Biology Laboratory, Gene Expression Programme, Heidelberg, Germany
3 Department of Pediatrics, Jorvi Hospital, Espoo, Finland
4 Department of Medical Genetics, The Family Federation of Finland, Helsinki, Finland
5 Department of Medical Genetics, University of Turku, Turku, Finland
6 Department of Biosciences at Novum and Clinical Research Centre, Karolinska Institutet, Stockholm, Sweden
Flint Jonathan EditorUniversity of Oxford, United Kingdom* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 21 9 2005 1 4 e507 4 2005 21 9 2005 Copyright: © 2005 Hannula-Jouppi et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Dyslexia, or specific reading disability, is the most common learning disorder with a complex, partially genetic basis, but its biochemical mechanisms remain poorly understood. A locus on Chromosome 3, DYX5, has been linked to dyslexia in one large family and speech-sound disorder in a subset of small families. We found that the axon guidance receptor gene ROBO1, orthologous to the Drosophila roundabout gene, is disrupted by a chromosome translocation in a dyslexic individual. In a large pedigree with 21 dyslexic individuals genetically linked to a specific haplotype of ROBO1 (not found in any other chromosomes in our samples), the expression of ROBO1 from this haplotype was absent or attenuated in affected individuals. Sequencing of ROBO1 in apes revealed multiple coding differences, and the selection pressure was significantly different between the human, chimpanzee, and gorilla branch as compared to orangutan. We also identified novel exons and splice variants of ROBO1 that may explain the apparent phenotypic differences between human and mouse in heterozygous loss of ROBO1. We conclude that dyslexia may be caused by partial haplo-insufficiency for ROBO1 in rare families. Thus, our data suggest that a slight disturbance in neuronal axon crossing across the midline between brain hemispheres, dendrite guidance, or another function of ROBO1 may manifest as a specific reading disability in humans.
Synopsis
Dyslexia, or specific reading disability, is a common learning disorder with a complex, partially genetic basis. A number of chromosomal regions harboring genes involved in dyslexia have been identified, and in this study the authors describe a candidate gene from one such locus, called DYX5, on Chromosome 3. The authors show that an axon guidance receptor gene, ROBO1, is disrupted by a chromosomal translocation in one dyslexic individual; furthermore, this study shows that the expression of ROBO1 is reduced on chromosomes from dyslexics in a large pedigree in which dyslexia has been linked to DYX5. ROBO1 has a role in regulating axon crossing across the midline between brain hemispheres and guidance of neuronal dendrites. As suggested by these findings, dyslexia may be caused in rare families by a small change in the expression of ROBO1, such as loss of one functional copy. Thus, ROBO1 is a candidate for a dyslexia susceptibility gene.
Citation:Hannula-Jouppi K, Kaminen-Ahola N, Taipale M, Eklund R, Nopola-Hemmi J, et al. (2005) The axon guidance receptor gene ROBO1 is a candidate gene for developmental dyslexia. PLoS Genet 1(4): e50.
==== Body
Introduction
Dyslexia refers to a difficulty in reading and writing despite normal intelligence, senses, and adequate education. The primary difficulty lies in phonological processing, rapid naming, and the recognition of phonemes [1]. Dyslexia is a common disorder, affecting 3% to 10% of the population [2]. Familial occurrence has been reported and twin and family studies indicate a strong genetic component in its etiology [3]. So far, genome-wide screens have linked dyslexia to loci on Chromosomes 1p34–36 (DYX8), 2p16-p15 (DYX3), 3p12-q13 (DYX5), 6p21.3 (DYX2), 11p15.5 (DYX7), 15q21 (DYX1), 18p11.2 (DYX6), Xq27.3 (DYX9), and 7q32 (http://www.ncbi.nlm.nih.gov/omim)[4–12]. The first dyslexia candidate susceptibility gene, DYX1C1 at 15q21, was recently identified, and a region corresponding to DYX2 has been narrowed down to a small number of candidate genes [12–16].
We ascertained earlier a large four-generation family including 21dyslexic individuals available to the study with severe dyslexia segregating in a dominant fashion, and mapped the susceptibility gene by a genome-wide scan to a region on Chromosome 3 that was named the DYX5 locus [17]. A 20-cM long haplotype was shared between 19 to 21 dyslexic individuals, giving statistically significant support for the mapping of the gene, but insufficient resolution to identify the gene [17]. Recently, speech-sound disorder was studied in a large set of 77 families from the United States, with results supporting association of this disorder to the DYX5 locus in the majority of families [18]. Phonological processing was the common phenotypic component found deficient in both studies [17–19].
We report here the localization of a translocation in an individual with dyslexia and a translocation t(3;8)(p12;q11) to the DYX5 region and specifically to the first intron of ROBO1. ROBO1 is a neuronal axon guidance receptor gene involved in brain development, and thus an attractive candidate gene for dyslexia susceptibility [20–22]. Furthermore, in the large pedigree with 19 dyslexic individuals genetically linked to DYX5 and more specifically, a specific and rare haplotype of ROBO1, the expression of ROBO1 from this haplotype was absent or attenuated in affected individuals. We conclude that dyslexia may be caused by partial haplo-insufficiency for ROBO1. In addition, we searched for novel exons and splice variants of ROBO1 that may help in understanding the apparent phenotypic discrepancy between heterozygous Dutt1and Robo1 knockout mice that develop lung cancers and lymphomas and humans whose developmental phenotypic consequence appears to be dyslexia.
Results
We identified a patient who was diagnosed with both dyslexia and a translocation t(3;8)(p12;q11) involving the DYX5 region. He came to our attention initially because of infertility. He had three siblings, one of whom was also diagnosed with dyslexia, but two were of subnormal intelligence and thus undefined for the dyslexia phenotype. Detailed phenotypic comparison was not possible, because the family members were not available for retesting. The index patient was the only translocation carrier among the siblings and likely to possess a de novo translocation (see Materials and Methods for details of the family). In spite of the discrepancy between the translocation status and apparent concordance for dyslexia between two siblings, we decided to map the translocation breakpoint in the hope of gaining insight to a possible candidate gene for further evaluation in the genetically informative large family [17,19]. We used fluorescence in situ hybridization to narrow down the breakpoint until a probe bacterial artificial chromosome (BAC) clone RP11-143B12 hybridized to both der(3) chromosomes as well as the normal Chromosome 3 (Figure 1A). The genomic sequence of clone RP11-143B12 corresponds to a Chromosome 3 scaffold sequence on the Celera public database (http://public.celera.com/cds/login.cfm) and to BACs AC117479 (43316–43873 base pairs [bp]) and AC117461 (50209–50745 bp) in the National Center for Biotechnology database (http://www.ncbi.nlm.nih.gov) (Figure 1C). Further identification of the breakpoint was conducted by Southern hybridization with PCR-amplified non-repetitive genomic DNA fragment probes from the clone. A 971-bp probe revealed DNA rearrangements, refining the breakpoint to a 4.7-kilobase (kb) interval on the Celera scaffold sequence or a 4.0-kb interval corresponding to nucleotides 57269–61233 of BAC AC117479 (Figure 1B and 1D). Surprisingly, the translocation breakpoint localized within the ortholog of the Drosophila roundabout (robo) gene, ROBO1, alternatively spliced and also named DUTT1 (Deleted in U Twenty Twenty), disrupts ROBO1 between exons 1 and 2.
Figure 1 Delineation of Translocation Breakpoint Region and ROBO1 Structure
(A) Fluorescence in situ hybridization with BAC clone RP11-143B12 as a probe, showing hybridization signals in Chromosome 3 (upward arrow), der(3), and der(8) (horizontal arrows).
(B) Southern hybridization with a probe derived from RP11-143B12 shows genomic rearrangements (arrowheads) with five restriction enzymes in translocation patient (P) compared to the control sample.
(C) A gene map of Chromosome 3p13-3q11.1 showing the cytogenetic localization of the translocation breakpoint (black bar). An arrow indicates the direction of ROBO1 transcription. Localization of the translocation breakpoint (square bracket) to BACs AC117479 and AC117461.
(D) Splice variants and exon structure of ROBO1 (exons numbered from 1–30). Novel exons a and 7b and additional sequence to exon 1 are indicated in solid black. Exons unique to ROBO1 (hatched black) and DUTT1 (hatched grey) and common to both ROBO1 and DUTT1 (solid grey) are indicated. Corresponding BACs to exons are shown below. The translocation disrupting ROBO1 between exons 1 and 2 in AC117479 is shown by vertical grey bar. Dotted lines indicate DUTT1 variants. Novel splice variants are shown by grey lines and numbered (1), exclusion of exon 2 (89–169 of AF040990) (2), exclusion of DUTT1 exon 2 (1019–1345 of Z95705) (3), exclusion of exon 19 (2813–2829 of AF040990) (4), initial 165 bp of exon 22 (3037–3201 of AF040990) (5), 905 bp from exons 24–28 (3603–4508 of AF040990) (6), 878 bp from exons 25–28 (3641–4528 of AF040990) and (7), exclusion of exon 29 (4745–4939 of AF040990).
Because of its known function in neuronal axon guidance in the developing brain, ROBO1 was a plausible candidate gene for dyslexia susceptibility [20–22]. We sequenced its exons, splice sites, 1 kb of ROBO1 promoter region upstream of exon 1, and the extended 3′ UTR region of ROBO1 variant 2 from the genomic DNA of initially one dyslexic individual and his parents (father dyslexic, mother unaffected) from the large linkage family (Figure 2A). All exons were also sequenced from the cDNA of another dyslexic individual from the extended family. Comparison of the sequences to ROBO1and ROBO1 variant 2 sequences revealed altogether seven sequence variations, two of them previously known [23]. All of the observed changes were confirmed in three additional pedigree members (dyslexic father, son, and unaffected mother) by sequencing. Dyslexic individuals had two silent single nucleotide polymorphisms (SNPs) in ROBO1 exons 12 and 18 (1741G > A, 2794C > A; numbering according to ROBO1), an exonic 3-bp deletion and insertion polymorphism (DIP6203–6205; numbering for ROBO1 variant 2), four SNPs in 3′ UTR (UTR, 6227C > A, 6483T > A, 6651T > A, 6923T > G; numbering for ROBO1 variant 2), and four intronic SNPs (intron 2: 59567 and intron 7: 1451; numbering for BAC RP11-588D3; intron 25: 16181 and 16198; numbering for BAC RP11-26M20) (Figure 2B).
Figure 2 Analysis of ROBO1 in the Large Family Linked to DYX5
(A) An abridged pedigree of the family linked to DYX5 [17,19]. Numbers refer to samples studied for ROBO1 expression (C and D). A dot indicates carriers of the dyslexia-linked haplotype [17] and circled dots indicate individuals genotyped for all markers (B). Diamonds denote individuals genotyped for all markers, but not sharing the haplotype. Affected individuals are shaded black and unverified dyslectics are shaded gray.
(B) Markers (right) and alleles (left) that define the haplotype linked to dyslexia (A). The bar indicates the extent of the ROBO1 haplotype carried by patients marked with a dot (A).
(C) Sequencing of cDNA reveals absent or attenuated expression (p < 0.017 for all measurements) of the ROBO1 allele (SNP 6483A > T indicated by arrows) encoded by the dyslexia-linked haplotype as compared to genomic sequence. In the control, both alleles show equal allelic ratios in genomic and cDNA. Patient numbers refer to (A).
(D) Attenuation of ROBO1 mRNA expression from the dyslexia-associated allele. Allelic expression of ROBO1 was assessed by sequencing the SNP 6483 (A/T) as in (C). Allelic ratios were assessed by five to six replicated sequencing tracings in four controls (21 data points) and four dyslexic individuals (24 data points). The results are expressed as the mRNA level of the dyslexia-associated allele as compared to the corresponding allele mRNA level in controls. Data are shown as mean ± 1 standard error of the mean (bars). Significance was assessed by two-tailed t test.
Genotyping of the exonic SNPs in ten additional family members and two unrelated non-dyslexic individuals confirmed that a specific SNP haplotype segregated with dyslexia, consistent with the previously observed linkage, but revealed that none of the polymorphisms was uniquely observed only in dyslexic individuals. For example, the DIP6203–6205*GAT+ allele had an allele frequency of 22% (94/434 among healthy control participants' chromosomes) and did not show significant association to dyslexia in our replication sample set of 96 dyslexic individuals from other families. None of the four observed intronic SNPs produced alternative splice variants. Because the ROBO1 gene spans about 990 kb of genomic DNA and contains altogether over 2,200 intronic SNPs (according to the NCBI SNP database), their exhaustive listing in our family members was impractical. In accordance with our previous genome scan on dyslexia in Finland, suggesting that DYX5 locus is not involved in most families [9], the same haplotype as in this large family was not observed in other, unrelated families (unpublished data).
The silent and 3′ UTR SNPs provided assays to study the transcription of ROBO1. Our rationale was to measure whether both alleles of ROBO1 were equally transcribed in dyslexic individuals segregating the dominant susceptibility haplotype. Comparison of genomic and cDNA samples from four dyslexic individuals showed that ROBO1 mRNA was only weakly or not at all transcribed from the allele that segregated with dyslexia (Figure 2C), whereas in non-readers' lymphocytes, as well as control brain RNA, biallelic expression was consistently observed. Of note, there was considerable variation between individuals, suggesting that the regulation of expression is complex. The SNP 2794C > A was heterozygous in one patient, 6483T > A in four, 6651T > A in four, and 6923T > G in one, and combining all results, the expression was significantly attenuated for the dyslexia-linked allele as measured by allelic peak heights (p = 0.017 by two-tailed t test). To verify this initial analysis, we repeated the assay for all four dyslexic and four control participants by sequencing the SNP 6483T > A again. The results from five to six replicated sequencing assays for each participant are shown in Figure 2D. By the repeated measurements, the mean expression level of the dyslexia-associated allele in dyslexic participants was 66% of the same allele in controls (p < 0.0004 by two-tailed t test). To exclude the possibility that an individual SNP behaved aberrantly in the analysis, we also sequenced the SNP 6651T > A, yielding similar findings, and both SNP assays combined, the observation of allelic imbalance in cases versus controls was highly significant (p < 0.00005 by t test). As no SNP was specific for the ROBO1 or DUTT1 transcript only, we cannot assess isoform-specific down-regulation in the large family. In the translocation patient, two SNPs were heterozygous, both in the region corresponding to exons common to both ROBO1 and DUTT1 transcripts (6651T > A and 6923T > G). These SNPs revealed two alleles present in cDNA in the translocation patient, suggesting that DUTT1 might be biallelically expressed even though the genomic structure of ROBO1 was disrupted by translocation in one chromosome.
To study the possibility that the suppression of expression involved other genes than ROBO1 in the dyslexia susceptibility haplotype, we genotyped known SNPs in the nearby genes GBE1 (341C/G, 646A/G, 1597A/G, 1794C/T, 2349T/G, 2363A/G, 2761A/T) and HTR1F (528C/T, 783T/A) in the four dyslexic individuals of the large family (Figure 1C). Heterozygosity was detected for the GBE1 SNPs 2363A > G and 646A > G in three patients. For these polymorphisms, normal biallelic expression was observed in all three patient samples in contrast to the finding with ROBO1, suggesting that transcription of ROBO1 was specifically silenced. Heritable variation in allelic expression levels has previously been documented for several genes, and might conceivably arise by a number of different mechanisms, such as variation in enhancer and suppressor elements, splicing efficiency, transcript stability, or epigenetic modifications [24]. Two other positional candidate genes, DRD3 and 5HT1F, mapped outside the shared haplotype [17].
Seemingly silent exonic and intronic polymorphisms may induce disease related splice variants [25]. Thus, we studied the possibility of ROBO1 alternative splicing by RT-PCR of all ROBO1 and DUTT1 exons from a dyslexic individual in the linkage family, an unrelated healthy control, and adult human brain cDNA. Seven novel splice variants were detected. Their cloning and sequencing revealed the exclusion of exons 2 (88 bp, 89–169 of ROBO1), 19 (27 bp, 2813–2829 of ROBO1), and 29 (196 bp 4745–4939 of ROBO1) entirely and exclusively of DUTT1 exon 2 (346 bp 1019–1345 of DUTT1); the initial 165 bp of exon 22 (3037–3201 of ROBO1); 905 bp ranging from exons 24 to 28 (3603–4508 of ROBO1); and 878 bp ranging from exons 25 to 28 (3641–4528 of ROBO1) (Figure 1D). No splice variants were uniquely different between dyslexic and control individuals; however, quantitative differences could not be reliably assessed. Comparison of the genomic and cDNA sequences for DUTT1 in several individuals suggested that exon 7 of DUTT1 is not colinear with genomic sequence. Instead, DUTT1 bases 1891–1900 (gttgggtct: valine, glycine, and serine), in the beginning of DUTT1 exon 7 (ROBO1 exon 8) belong to a new short exon, marked exon 7b (Figure 1D) corresponding to bases 5987–5995 of BAC RP11-588D3. These bases have previously been reported as part of the DUTT1 gene but they are not included in the ROBO1 cDNA sequence [23]. In all individuals sequenced, the cDNA sequence included the new exon 7b, indicating that it is included in the major splice form in at least brain and lymphoblast RNA.
As the known ROBO1 sequence AF040990 starts from the transcription initiation site, we sought to determine the ROBO1 5′ UTR sequence by a BLAST search for expressed sequence tags homologous for the 5′ ROBO1 region. The expressed sequence tag AW450262, homologous to ROBO1 exons 1 and 2, indicated an additional site (referred to as ROBO1 exon a) upstream on BAC AC125624 (bases 28508–28470) (Figure 1D). RT-PCR with an initial primer in the exon a sequence and primers in ROBO1 exons 1 and 4 revealed an additional 52 bp (bases 34119–34168 of BAC AC125815) 5′ of the transcription initiation site on ROBO exon 1 and also confirmed the presence of the novel exon a on the BAC AC125624. Additional primers were designed 5′ to the novel exon a, and RT-PCR performed similarly as above showed the a exon to span at least 129 bp (bases 28593–28466 of BAC AC125624). 5′ rapid amplification of cDNA ends (RACE) revealed additional 326 bps, stretching the exon to 28919–28466 on the BAC AC125624.
A significant fraction of human genes has been under positive Darwinian selection since the common ancestor of humans and chimpanzees [26]. For example, FOXP2, the gene implicated in speech and language, has undergone a selective sweep during human evolution [27]. Therefore, we sequenced ROBO1 from chimpanzee, pygmy chimpanzee, gorilla, and orangutan, and used the rat ROBO1 sequence as the out-group. We used likelihood ratio test to analyze variation in the selective pressure in ROBO1 sequence in the different lineages. Non-synonymous and synonymous (dN and dS) ratio was smaller than 1 in all lineages, implicating purifying Darwinian selection. However, the likelihood ratio test rejected the null hypothesis of fixed dN/dS ratio in all lineages. A model in which omega value was higher in lineages leading to humans, chimpanzees, and gorillas was significantly better than a free-ratio model (p < 0.001) (Figure 3 and Tables S1 and S2). This suggests that the selective pressure for ROBO1 gene has changed 12 to16 million years ago, after the divergence of the orangutan branch.
Figure 3 Coding Changes of ROBO1 during Primate Evolution
Phylogenetic tree of ROBO1 protein evolution in hominoids. Rat was used as the out-group in sequence comparisons. dN/dS ratios of the branches were calculated with the Codeml program, assuming a freely varying ratio. A model in which omega value was higher in lineages leading to humans, chimpanzees, and gorillas was significantly better than a free-ratio model (p < 0.001).
Discussion
Our data suggested that two functional copies of ROBO1 are required in brain development to acquire normal reading ability, and partial haplo-insufficiency for ROBO1 may predispose humans to specific dyslexia. robo was originally identified in Drosophila in a search for genes controlling the midline crossing of axons; in mutant robo embryos, axons cross and recross across the midline too many times [20–22]. The human ortholog of robo, ROBO1 (also named DUTT1), was identified as a potential tumor-suppressor gene in a small-cell lung cancer cell line [28]. ROBO1 and DUTT1 are presumed to be alternative splice variants with different initial exons and initiation codons and may thus have in part distinct functions [23]. Homozygous Robo1/Dutt1 knockout mice are embryonically lethal, but heterozygous mice were found to develop lymphomas and lung adenocarcinomas at high frequency [29]. The human ROBO1/DUTT1 locus has been found deleted in a child with developmental delay and congenital anomalies but without cancers [30]. These observations seem to pose a dilemma for understanding the functions of ROBO1 in different species.
robo encodes a transmembrane receptor that belongs to the immunoglobulin superfamily. It consists of immunoglobulin domains, three fibronectin domains, a transmembrane domain, and a long intracellular region with no recognizable motifs, but four proline rich repeats, which are suggested to act with enabled, abelson, SH3 binding proteins, and other downstream signaling molecules [20,31]. As shown by our data, ROBO1 undergoes alternative splicing in a more complex manner than previously appreciated and the functions of these alternative splice variants are unknown. It is possible that still different promoters and splice variants in different tissues cause alternative phenotypes. The translocation in our patient disrupts specifically ROBO1, but not DUTT1, and therefore does not contradict the previous observations of DUTT1 as a tumor suppressor gene. We thus hypothesize that the two-splice variants in humans are associated with different key functions in different tissues: ROBO1 might correspond more closely to functions in the human brain that are modeled by neuronal functions in the fruit fly, and DUTT1 functions appear to correspond to the mouse model of lung tumorigenesis.
In the fruit fly, robo is a receptor for secreted repellent slit proteins and acts as a gatekeeper of axonal crossing on the left-right axis. Activation of robo makes axons indifferent to the chemoattractant netrin, a ligand of the DCC receptor [32]. Netrin-dependent activation of the DCC receptor increases transient phosphorylation of ERK1, ERK2, and the transcription factor ELK1. The activation of the ERK pathway has been suggested as necessary for experience-dependent plasticity and for long-term potentiation of synaptic transmission in visual cortex development in the rat [33]. Interestingly, a possible ELK1 binding site was altered and associated with dyslexia in DYX1C1, a candidate gene for dyslexia susceptibility on Chromosome 15q21, and ELK1 has been implicated in learning in the rat [13,34–36]. A suggested functional role for DYX1C1 remains unconfirmed, because its associations have not been unambiguously replicated, leaving the relevance of the ELK1 binding site open [37–42]. Slit/Robo/DCC signaling has also been implicated in cortical dendritic guidance and development [43,44]. Furthermore, a recently identified dyslexia susceptibility locus on Xq27.3 includes the SLITRK2 and SLITRK4 genes, which belong to the SLIT and NTRK-like family of genes involved in mouse neurite outgrowth and show high homology to Slit proteins [10,45,46]. Thus, the identification of ROBO1 as a susceptibility gene in dyslexia may implicate a key developmental pathway in which slight disturbances may lead to specific reading disability.
Genomic sequences and predicted transcripts for ROBO1 in four apes revealed a high level of variation between the related species and humans (Figure 3). We detected seven amino acid changes between human and chimpanzee and 20 between humans and orangutan. An analysis of dN/dS substitutions revealed that the selective pressure on ROBO1 has changed after the divergence of the orangutan branch. Although according to stringent criteria, only dN/dS ratios higher than 1 are regarded as signs of positive selection; it has been shown that genes expressed in the brain are under stronger selective pressure than genes expressed, for example, in the liver [47]. In addition, in primates many brain-expressed genes show significantly higher dN/dS ratios than housekeeping genes when compared to the rodent counterparts, even though the dN/dS ratios in primates are still well below 1. Interestingly, it was recently shown that the evolution of SLIT1, a ligand for ROBO1, has been significantly faster in primates than in rodents [48]. Also other proteins involved in axonal path-finding, such as SEMA4F and EPHA6, were shown to have undergone adaptive evolution [48]. We propose that ROBO1 might have undergone rapid changes in the recent primate evolution that may be related to its largely uncharacterized functions in the human brain.
Taken together, these results implicate a well-known pathway of neuronal development in a highly specific cognitive function in humans. This function attributed here to the ROBO1 transcript variant including newly discovered exons is distinct from the role that has been suggested in lung tumorigenesis for the alternatively spliced transcript variant DUTT1. This insight may open new visions for understanding complex brain processes and provide a framework for building testable hypotheses for the biology of reading.
Materials and Methods
Patients.
A multiplex four-generation family with severe dyslexia segregating in a dominant fashion (Figure 2A) has been previously studied for genetic linkage and phenotype [17,19]. Dyslexia was diagnosed in 27 out of 74 family members by thorough testing including an intelligence test, a Finnish reading and writing test for adults and for children according to their school grade, and a neuropsychological test battery [49–52]. Detailed psychological evaluation of this family has been reported elsewhere [19].
A dyslexic individual with a balanced reciprocal translocation t (3;8) (p12;q11) came to our attention because of infertility and was diagnosed with oligoteratozoospermia. He has three siblings, and all four children have been neuropsychologically evaluated at a specialist hospital. However, the family members were not available for retesting and thus our data are based on the clinical records. The translocation carrier and his sister were diagnosed with severe dyslexia while the other two siblings had subnormal intelligence, but not dyslexia. The mother was reported as a good reader, but no information on reading performance was available on the deceased father. The other three siblings have a normal karyotype, whereas the parents were not available for karyotyping. Because the index case presented with infertility, but no such history or miscarriages were recorded for his mother, it is likely that the translocation had arisen de novo. Thus, the translocation and dyslexia did not apparently cosegregate in two siblings, but as dyslexia is a complex phenotype, no inference is conclusive to either reject or confirm a possible causal association of the translocation with dyslexia in the index case.
For association studies, dyslexic and non-dyslexic individuals were recruited from 23 unrelated families and 33 unrelated dyslexic and non-dyslexic couples from the Department of Pediatric Neurology at the Hospital for Children and Adolescents, University of Helsinki, and the Child Research Centre, Jyväskylä, Finland. Additional population controls consisted of 100 anonymous blood donors. The diagnosis and degree of dyslexia were determined by Finnish reading and spelling tests designed for children and adults [49,50]. Intelligence was estimated by Wechsler tests for adults (WAIS-R) or for children (WISC-R) [51,52]. The diagnostic criteria for dyslexia included normal performance intelligence quotient (PIQ > 85) and remarkable deviation (depending on age, at least two years) in reading skills. This study has been approved by the ethical review board of the Helsinki University Central Hospital, and informed consent was obtained from the participants.
Fluorescence in situ and Southern hybridization.
YAC clones A136E9 (Washington University, St. Louis, Missouri, United States), 34FC9, 15AC10, 39F13, 2DG9, 25DH8, 35AH8 (ICI/Zeneca), 912A11, 934E8, 422A6, 959F5, 650C2, 938D4, and BAC RP11-143B12 were used as probes in fluorescence in situ hybridization experiments. The genomic sequence of clone RP11-143B12 was obtained by BLAST search with the BAC 5′ and 3′ends (AQ373182, T7, and AQ373179 Sp6, respectively) to a Chromosome 3 scaffold sequence on the Celera public database (http://public.celera.com/cds/login.cfm) and to BACs in the National Center for Biotechnology CBI database (http://www.ncbi.nlm.nih.gov/).
Southern hybridization probes were PCR-amplified genomic fragments from non-repetitive regions on the BAC clone RP11-143B12. The repeats in the clone sequence were detected by RepeatMasker (http://repeatmasker.genome.washington.edu/cgi-bin/RepeatMasker) and PCR primers were designed by Primer3 (http://frodo.wi.mit.edu/cgi-bin/primer3/primer3_www.cgi) to encompass non-repetitive segments of 700–1,000 bp (primer sequences available from authors on request). PCR assays were performed under standard conditions and 10 ng of the purified PCR products (Qiagen PCR purification kit) (Qiagen, Valencia, California, United States) and 10 ng of the purified probes were labeled with [a-32P] dCTP (Rediprime DNA Labeling System, Amersham Biosciences, Little Chalfont, United Kingdom). Southern blotting and hybridizations were performed by standard protocols with seven μg of DNA from the translocation patient and a healthy control individual digested in separate reactions with BamHI, BglII, EarI, EcoRI, HaeII, HindIII, NcoI, and PstI (New England Biolabs, Beverly, Massachusetts, United States).
Polymorphism screening of ROBO1 and expression analysis.
All ROBO1 and DUTT1 exons were PCR-amplified and sequenced from genomic DNA and the cDNA of selected individuals from the large linkage pedigree (Figure 2A). In addition, the novel exonic sequences were identified and 2 kb of ROBO1 promoter region upstream of the novel exon a and the 3′ UTR of ROBO1 variant 2 were sequenced. BACs corresponding to exons were identified through BLAST searches. Primate DNA samples were obtained from the Coriell Institute (Camden, New Jersey, United States) (Primate Panel PRP00001) and orthologs of ROBO1 were sequenced directly after PCR with human-specific primers.
RNA was extracted from EBV transformed lymphocyte cell lines from four dyslexic and four normal readers by Ficoll gradient centrifugation (Qiagen Rneasy purification kit) and RT-PCR was used to amplify cDNA segments containing heterozygous SNPs in genomic DNA. As controls, we used genomic DNA samples from the same individuals as well as brain mRNA (Clontech, Palo Alto, California, United States). All sequencing was performed using dye-terminator chemistry and automated sequencers (ABI, Columbia, Maryland, United States).
To assess the allele-specific expression, we followed a standard method [53]. The assay is based on the comparison of allelic peak heights (in arbitrary units) in cDNA sequence (after RT-PCR) and genomic sequence from each individual. An allelic ratio is calculated for each sequence (e.g., height of allele A per height of allele C). Because the allelic ratio in genomic sequence is by definition 1 (each allele is present as one copy per diploid genome), but the actual value may differ from 1 (because of chemical properties of the sequencing reactions), the cDNA allelic ratio values are normalized by dividing by the genomic allelic ratio in each experiment. To assess whether the normalized cDNA allelic ratios differed in dyslexic patients as compared with controls, the values from replicated experiments were compared between the groups by two-tailed t-test. To estimate the degree of attenuation of one allele in dyslexic patients, the average cDNA allelic ratio in dyslexic patients was divided by the average cDNA allelic ratio in controls. Standard deviation of the measurements was calculated on replicated experiments.
5′ RACE.
5′ RACE was performed on human brain RNA using the SMART RACE cDNA Amplification kit (Clontech). 5′ cDNA ends were amplified with the Universal Primer A and a specific Robo 5′ RACE-R1 (gcagacgcagccctgcaacttt) primer, followed by nested PCR with the Nested Universal Primer A. PCR products were purified and directly sequenced (ABI).
Evolutionary analysis of ROBO1 sequence.
Likelihood ratio test was performed with the Codeml program of the PAML package [26].
Supporting Information
Table S1 Comparison of ROBO1 between Human and Four Non-Human Primates
Nucleic acid (according to AF040990 exons 1 to 29, NM_133631 exon 30) and amino acid changes are shown for each exon of ROBO1 in comparison to the corresponding human BAC sequence; + indicates the presence of a change in a non-human species. Amino acid changes are shaded. No differences were observed for DUTT1 exon 1.
(311 KB DOC)
Click here for additional data file.
Table S2 Analysis of ROBO1 Evolution with PAML
Likelihood values and parameter estimates under different models
(19 KB XLS)
Click here for additional data file.
Accession Numbers
The GenBank (http://www.ncbi.nlm.nih.gov/geo) accession numbers for genes discussed in this paper are DUTT1 (Z95705), GBEI (NM_000158), HTRIF (NM_000866), human homolog 1of the Drosophila roundabout gene, ROBO1 (AF040990), the first intron of ROBO1 (NM_002941), ROBO1 variant 2 (NM_133631), Rat ROBO1 (NM_022188), Homo sapiens clone sequences for BACs RP11-588D3 (AC055731) and BAC RP11-26M20 (AC106720).
This study was supported by Sigrid Jusélius Foundation, Academy of Finland, and the Technology Development Agency of Finland (Tekes). JK is a member of Biocentrum Helsinki and Center of Excellence for Disease Genetics at University of Helsinki.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. KHJ, NKA, MT, and JK conceived and designed the experiments. KHJ, NKA, MT, and RE performed the experiments. KHJ, NKA, MT, and JK analyzed the data. JNH and HK contributed reagents/materials/analysis tools. KHJ, NKA, and JK wrote the paper.
A previous version of this article appeared as an Early Online Release on September 21, 2005 (DOI: 10.1371/journal.pgen.0010050.eor).
Abbreviations
BACbacterial artificial chromosome
bpbase pair
dN/dSnon-synonymous and synonymous
kbkilobase
RACErapid amplification of cDNA ends
SNPsingle nucleotide polymorphism
==== Refs
References
Shaywitz SE 1998 Dyslexia N Engl J Med 338 307 312 9445412
Pennington BF 1990 Annotation: The genetics of dyslexia J Child Psychol Psychiatry 31 193 201 2179251
DeFries JC Fulker DW LaBuda MC 1987 Evidence for a genetic etiology in reading disability of twins Nature 329 537 539 3657975
Rabin M Wen XL Hepburn M Lubs HA 1993 Suggestive linkage of developmental dyslexia to chromosome 1p34–p36 Lancet 342 178
Cardon LR Smith SD Fulker DW Kimberling WJ Pennington BF 1994 Quantitative trait locus for reading disability on chromosome 6 Science 266 276 279 7939663
Grigorenko EL Wood FB Meyer MS Hart LA Speed WC 1997 Susceptibility loci for distinct components of developmental dyslexia on chromosomes 6 and 15 Am J Hum Genet 60 27 39 8981944
Fagerheim T Raeymaekers P Tonnessen FE Pedersen M Tranebjaerg L 1999 A new gene (DYX3) for dyslexia is located on chromosome 2 J Med Genet 36 664 669 10507721
Fisher SE Francks C Marlow AJ MacPhie IL Newbury DF 2002 Independent genome-wide scans identify a chromosome 18 quantitative-trait locus influencing dyslexia Nature Genet 30 86 91 11743577
Kaminen N Hannula-Jouppi K Kestila M Lahermo P Muller K 2003 A genome scan for developmental dyslexia confirms linkage to chromosome 2p11 and suggests a new locus on 7q32 J Med Genet 40 340 345 12746395
de Kovel CG Hol FA Heister JG Willemen JJ Sandkuijl LA 2004 Genome-wide scan identifies susceptibility locus for dyslexia on Xq27 in an extended Dutch family J Med Genet 41 652 657 15342694
Hsiung GY Kaplan BJ Petryshen TL Lu S Field L 2004 A dyslexia susceptibility locus (DYX7) linked to dopamine D4 receptor (DRD4) region on chromosome 11p15.5 Am J Med Genet 125B 112 119 14755455
Fisher SE DeFries JC 2002 Developmental dyslexia: Genetic dissection of a complex cognitive trait Nat Rev Neurosc 3 767 780
Taipale M Kaminen N Nopola-Hemmi J Haltia T Myllyluoma B 2003 A candidate gene for developmental dyslexia encodes a nuclear tetratricopeptide repeat domain protein dynamically regulated in brain Proc Natl Acad Sci U S A 100 11553 11558 12954984
Cope N Harold D Hill G Moskvina V Stevenson J 2005 Strong evidence that KIAA0319 on chromosome 6p is a susceptibility gene for developmental dyslexia Am J Hum Genet 76 581 591 15717286
Deffenbacher KE Kenyon JB Hoover DM Olson RK Pennington BF 2004 Refinement of the 6p21.3 quantitative trait locus influencing dyslexia: Linkage and association analyses Hum Genet 115 128 138 15138886
Francks C Paracchini S Smith SD Richardson AJ Scerri TS A 77-kilobase region of chromosome 6p22.2 is associated with dyslexia in families from the United Kingdom and from the United States Am J Hum Genet 75 1046 1058
Nopola-Hemmi J Myllyluoma B Haltia T Taipale M Ollikainen V 2001 A dominant gene for developmental dyslexia on Chromosome 3 J Med Genet 38 658 664 11584043
Stein CM Schick JH Gerry Taylor H Shriberg LD Millard C 2004 Pleiotropic effects of a chromosome 3 locus on speech-sound disorder and reading Am J Hum Genet 74 283 297 14740317
Nopola-Hemmi J Myllyluoma B Voutilainen A Leinonen S Kere J 2002 Familial dyslexia: Neurocognitive and genetic correlation in a large Finnish family Dev Med Child Neurol 44 580 586 12227612
Kidd T Brose K Mitchell KJ Fetter RD Tessier-Lavigne M 1998 Roundabout controls axon crossing of the CNS midline and defines a novel subfamily of evolutionarily conserved guidance receptors Cell 92 205 215 9458045
Kidd T Bland KS Goodman CS 1999 Slit is the midline repellent for the robo receptor in Drosophila
Cell 96 785 794 10102267
Seeger M Tear G Ferres-Marco D Goodman CS 1993 Mutations affecting growth cone guidance in Drosophila: Genes necessary for guidance toward or away from the midline Neuron 10 409 426 8461134
Dallol A Forgacs E Martinez A Sekido Y Walker R 2002 Tumour specific promoter region methylation of the human homologue of the Drosophila Roundabout gene DUTT1 (ROBO1) in human cancers Oncogene 21 3020 3028 12082532
Yan H Yuan W Velculescu VE Vogelstein B Kinzler KW 2002 Allelic variation in human gene expression Science 297 1143 12183620
Pagani F Baralle FE 2004 Genomic variants in exons and introns: Identifying the splicing spoilers Nature Rev 5 389 396
Clark AG Glanowski S Nielsen R Thomas PD Kejariwal A 2003 Inferring non-neutral evolution from human-chimp-mouse orthologous gene trios Science 302 1960 1963 14671302
Enard W Przeworski M Fisher SE Lai CS 2002 Molecular evolution of FOXP2, a gene involved in speech and language Nature 418 869 872 12192408
Sundaresan V Chung G Heppell-Parton A Xiong J Grundy C 1998 Homozygous deletions at 3p12 in breast and lung cancer Oncogene 17 1723 1729 9796701
Xian J Aitchison A Bobrow L Corbett G Pannell R 2004 Targeted disruption of the 3p12 gene, Dutt1/Robo1, predisposes mice to lung adenocarcinomas and lymphomas with methylation of the gene promoter Cancer Res 64 6432 6437 15374951
Petek E Windpassinger C Simma B Mueller T Wagner K 2003 Molecular characterisation of a 15 Mb constitutional de novo interstitial deletion of chromosome 3p in a boy with developmental delay and congenital anomalies J Hum Genet 48 283 287 12836054
Bashaw GJ Kidd T Murray D Pawson T Goodman CS 2000 Repulsive axon guidance: Abelson and enabled play opposing roles downstream of the Roundabout receptor Cell 101 703 715 10892742
Stein E Tessier-Lavigne M 2001 Hierarchical organization of guidance receptors: Silencing of netrin attraction by slit through a Robo/DCC receptor complex Science 291 1928 1938 11239147
Di Cristo G Berardi N Cancedda L Pizzorusso T Putignano E 2001 Requirement of ERK activation for visual cortical plasticity Science 292 2337 2340 11423664
Sgambato V Vanhoutte P Pages C Rogard M Hipskind R 1998 In vivo expression and regulation of Elk-1, a target of the extracellular regulated kinase signaling pathway, in the adult rat brain J Neurosci 18 214 226 9412502
Cammarota M Bevilaqua LR Ardenghi P Paratcha G Levi de Stein M 2000 Learning-associated activation of nuclear MAPK, CREB, and Elk-1, along with Fos production, in the rat hippocampus after a one-trial avoidance learning: Abolition by NMDA receptor blockade Brain Res Mol Brain Res 76 36 46 10719213
Berman DE 2003 Modulation of taste-induced Elk-1 activation by identified neurotransmitter systems in the insular cortex of the behaving rat Neurobiol Learn Mem 79 122 126 12482687
Chapman NH Igo RP Thomson JB Matsushita M Brkanac Z 2004 Linkage analyses of four regions previously implicated in dyslexia: Confirmation of a locus on chromosome 15q Am J Med Genet B Neuropsychiatr Genet 131B 67 75 15389770
Scerri TS Fisher SE Francks C MacPhie IL Paracchini S 2004 Putative functional alleles of DYX1C1 are not associated with dyslexia susceptibility in a large sample of sibling pairs from the UK J Med Genet 41 853 857 15520411
Wigg KG Couto JM Feng Y Anderson B Cate-Carter TD 2004 Support for EKN1 as the susceptibility locus for dyslexia on 15q21 Mol Psychiatry 9 1111 1121 15249932
Cope NA Hill G van den Bree M Harold D Moskvina V 2005 No support for association between dyslexia susceptibility 1 candidate 1 and developmental dyslexia Mol Psychiatry 10 237 238 15477871
Marino C Giorda R Lorusso ML Vanzin L Salandi N 2005 A family-based association study does not support DYX1C1 on 15q21.3 as a candidate gene in developmental dyslexia Eur J Hum Genet 13 491 499 15702132
Meng H Hager K Held M Page GP Olson RK 2005 TDT-association analysis of EKN1 and dyslexia in a Colorado twin cohort Hum Genet Epub ahead of print. PMID: 16133186
Whitford KL Marillat V Stein E Goodman CS Tessier-Lavigne M 2002 Regulation of cortical dendrite development by Slit-Robo interactions Neuron 33 47 61 11779479
Furrer MP Kim S Wolf B Chiba A 2003 Robo and Frazzled/DCC mediate dendritic guidance at the CNS midline Nat Neurosci 6 223 230 12592406
Aruga J Yokota N Mikoshiba K 2003a Human SLITRK family genes: Genomic organization and expression profiling in normal brain and brain tumor tissue Gene 315 87 94 14557068
Aruga J Mikoshiba K 2003b Identification and characterization of Slitrk, a novel neuronal transmembrane protein family controlling neurite outgrowth Mol Cell Neurosci 24 117 129 14550773
Duret L Mouchiroud D 2000 Determinants of substitution rates in mammalian genes: Expression pattern affects selection intensity but not mutation rate Mol Biol Evol 17 68 70 10666707
Dorus S Vallender EJ Evans PD Anderson JR Gilbert SL 2004 Accelerated evolution of nervous system genes in the origin of Homo sapiens
Cell 119 1027 1040 15620360
Häyrinen T Serenius-Sirve S Korkman M 1999 Reading and writing test designed for and normated in Finnish elementary school (in Finnish) Psykologien kustannus Oy, Helsinki.
Leinonen S Müller K Leppänen P Aro M Ahonen T 2001 Heterogeneity in adult dyslexic readers: Relating processing skills to the speed and accuracy of oral text reading Read Writ Interdisc J 14 265 296
Wechsler D 1992 Wechsler adult intelligence scale revised (WAIS-R) Psykologien kustannus Oy and The psychological corporation USA, Helsinki.
Wechsler D 1984 Wechsler intelligence scale for children revised (WISC-R) Psykologien kustannus Oy and The psychological corporation USA, Helsinki.
Pastinen T Sladek R Gurd S Sammak A Ge B 2003 A survey of genetic and epigenetic variation affecting human gene expression Physiol Genomics 16 184 193
|
16254601
|
PMC1270007
|
CC BY
|
2021-01-05 08:00:24
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e50
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010050
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460210.1371/journal.pgen.0010053plge-01-04-06Research ArticleGli2 and Gli3 Localize to Cilia and Require the Intraflagellar Transport Protein Polaris for Processing and Function Cilia and Gli Processing and FunctionHaycraft Courtney J 1Banizs Boglarka 1Aydin-Son Yesim 2Zhang Qihong 1Michaud Edward J 23Yoder Bradley K 1*1 Department of Cell Biology, University of Alabama, Birmingham, Alabama, United States of America
2 University of Tennessee Oak Ridge National Laboratory Graduate School of Genome Science and Technology, University of Tennessee, Knoxville, Tennessee, United States of America
3 Life Sciences Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, United States of America
Barsh Gregory EditorStanford University School of Medicine, United States of America* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 26 9 2005 1 4 e5331 8 2005 26 9 2005 This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.2005This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.Intraflagellar transport (IFT) proteins are essential for cilia assembly and have recently been associated with a number of developmental processes, such as left–right axis specification and limb and neural tube patterning. Genetic studies indicate that IFT proteins are required for Sonic hedgehog (Shh) signaling downstream of the Smoothened and Patched membrane proteins but upstream of the Glioma (Gli) transcription factors. However, the role that IFT proteins play in transduction of Shh signaling and the importance of cilia in this process remain unknown. Here we provide insights into the mechanism by which defects in an IFT protein, Tg737/Polaris, affect Shh signaling in the murine limb bud. Our data show that loss of Tg737 results in altered Gli3 processing that abrogates Gli3-mediated repression of Gli1 transcriptional activity. In contrast to the conclusions drawn from genetic analysis, the activity of Gli1 and truncated forms of Gli3 (Gli3R) are unaffected in Tg737 mutants at the molecular level, indicating that Tg737/Polaris is differentially involved in specific activities of the Gli proteins. Most important, a negative regulator of Shh signaling, Suppressor of fused, and the three full-length Gli transcription factors localize to the distal tip of cilia in addition to the nucleus. Thus, our data support a model where cilia have a direct role in Gli processing and Shh signal transduction.
Synopsis
Cilia are small projections extending from the surface of most cells. Research has shown that they are important in diseases such as cystic kidney diseases as well as during the development of many tissues including the limb. More recently, proteins such as Polaris, which is required to build cilia, have been demonstrated to be essential for the regulation of Sonic hedgehog (Shh) signaling, although the mechanism has remained elusive. Precise regulation of Shh signal transduction is important for the proper development of many tissues. Excessive activation of the Shh pathway results in severe developmental defects and has been implicated in certain types of cancer. In the limb, Shh signaling is involved in digit development, and excess signaling leads to the formation of extra digits. The main targets of Shh signaling are the Glioma (Gli) family of transcription factors, and Gli3 has been shown to be processed to a shortened repressor form when Shh signaling is repressed. The localization of the Gli transcription factors and Suppressor of fused, a protein involved in the regulation of Gli protein function, to cilia suggests that the cilia may be an important site for regulation of Shh signal transduction by modulating Gli protein function.
Citation:Haycraft CJ, Banizs B, Aydin-Son Y, Zhang Q, Michaud EJ, et al. (2005) Gli2 and Gli3 localize to cilia and require the intraflagellar transport protein Polaris for processing and function. PLoS Genet 1(4): e53.
==== Body
Introduction
Cilia are microtubule-based organelles that protrude from the surface of most cells in the mammalian body and are formed through a conserved process termed intraflagellar transport (IFT) [1]. Polaris, the protein encoded by Tg737, is a core component of the mammalian IFT machinery and is required for the formation of all cilia and flagella [2,3]. Mice homozygous for the hypomorphic Tg737orpk allele exhibit phenotypes in many tissues including the formation of cysts in the kidney, liver, and pancreas, hydrocephalus, and skeletal patterning defects that include extra molar teeth, cleft palate, and preaxial polydactyly [2,4–6]. While Tg737orpk mutants are viable, complete loss of Tg737 function in Tg737
Δ2–3β-gal mutants results in midgestation lethality, randomization of the left–right body axis, neural tube closure and patterning defects, and formation of eight to ten unpatterned digits per limb [3]. In Tg737orpk and Tg737
Δ2–3β-gal mutant mice, cilia are severely malformed or absent, respectively, suggesting that this organelle is required for normal development and patterning of many tissues in the mammalian body [2,3,5].
The mammalian limb is patterned through the interaction of three main signaling centers [7]. The apical ectodermal ridge is necessary for proper limb outgrowth and proximal–distal length while the surface ectoderm regulates dorsal–ventral patterning. The zone of polarizing activity, located in the posterior mesenchyme, is involved in anterior–posterior patterning including the formation of five digits per limb. Sonic hedgehog (Shh) is secreted by cells in the zone of polarizing activity, and many polydactyl mutants in the mouse have been shown to have ectopic expression of either Shh or genes activated by Shh. The main targets of Shh signaling are the Glioma (Gli) transcription factors [8]. Three Gli transcription factors (Gli1, Gli2, and Gli3) have been identified in mammals. Gli3 exists as a full-length “activator” (Gli3A) that is proteolytically processed into a smaller form with repressor activity (Gli3R) in the absence of Shh ligand [9]. Binding of Shh to its receptor Patched1 (Ptch1) leads to the derepression of Smoothened and blocks processing of the Gli3 transcription factor.
While mutations in Gli1 or Gli2 alone have no affect on digit patterning, loss of one or both alleles of Gli3 in Gli3Xt-J mice is sufficient to produce ectopic digits [10,11]. The severe polydactyly in Gli3Xt-J homozygous mutants is associated with ectopic expression of Shh and its target genes [12]; however, the loss of Shh in Shh−/−;Gli3Xt-J double mutants results in identical digit patterning defects as seen in Gli3Xt-J mutants alone, leading to the hypothesis that a pentadactyl restraint is imposed on the limb by Shh counteraction of Gli3 repressor activity [13,14].
Another important component of the Shh signaling pathway involved in Gli protein regulation is Suppressor of fused (Sufu). Sufu is a negative regulator of Shh signaling that interacts with all three Gli proteins and mediates their nuclear export in the absence of Shh [15–17]. In Drosophila, Su(fu) is thought to link Ci (Gli homolog) to the ubiquitin proteasome required for conversion of Ci to the small repressor form, as well as retain unprocessed Ci in the cytoplasm in the absence of ligand; however, this has not yet been demonstrated for the mammalian pathway [8].
Previously, we demonstrated that partial disruption of IFT function in Tg737orpk mutants results duplication of digit I, while complete loss of IFT in Tg737
Δ2–3β-gal mutants leads to the formation of up to ten unpatterned digits per limb [6]. Despite the formation of excess digits and a known involvement of Shh in ectopic digit formation in many mouse models [7], Shh expression is not altered in either Tg737orpk or Tg737
Δ2–3β-gal mutants [6]. In addition, no alterations in Ptch1 expression in Tg737orpk hypomorphic mutants is observed. Recent evidence from the labs of Anderson and Niswander has shown that IFT proteins, including Tg737/Polaris, are essential for Shh signaling in both neural tube and limb patterning at the level of Gli3 processing, although the connection between IFT and this signaling pathway remains enigmatic [18–20]. Previous models have speculated that there is a cytosolic role for the IFT proteins or that the cilia generate a signal affecting the activity of the Gli transcription factors. However, in this report we provide evidence supporting the hypothesis that there is a direct role for cilia in Shh signal transduction. Despite normal Shh expression in homozygous Tg737
Δ2–3β-gal mutant limbs [6], expression of the Shh downstream targets such as Ptch1 and Gli1 is lost. Furthermore, cells isolated from mutant limb buds are unable to respond to ShhN conditioned medium (ShhN-CM) in vitro. Even though mutant cells are unable to respond to ShhN-CM, the pathway can be activated by exogenously expressed Gli1; however, Gli2 and full-length Gli3 were found to be inactive in the absence of Polaris. Although full-length Gli3 was nonfunctional in Tg737
Δ2–3β-gal mutant cells, expression of a processed form of Gli3 (Gli3R) acts as a potent repressor of Gli1-mediated induction of Ptch1 expression. Additionally, the amount of full-length Gli3 was markedly increased in Tg737
Δ2–3β-gal mutant embryos relative to wild-type controls, suggesting that loss of Polaris results in inefficient Gli3 processing. Most importantly, the data indicate that all three full-length Gli proteins along with Sufu colocalize to the distal tips of cilia in primary limb bud cells. Together our data support a direct role for cilia in the mammalian hedgehog signaling pathway and raise the intriguing possibility that the tip of the cilium is a specialized domain in which proteolytic machinery is concentrated for processing and regulating the activity of Shh signal transduction.
Results
Cilia Are Present on Both Ectoderm and Mesenchyme Cells of the Limb Bud
To determine whether cilia are present on the developing mouse limb bud, we conducted electron microscopic analysis of embryonic day 11.5 (E11.5) limb buds. Using transmission electron microscopy, cilia were found on the mesenchyme. These cilia have a 9 + 0 microtubule structure, were frequently found in depressions in the cell membrane, and were always closely associated with the Golgi apparatus (Figure 1A–1C). In addition, small vesicular structures were frequently detected near the base of the cilium. Using scanning electron microscopy, we also determined that most, if not all, ectodermal cells exhibit a single cilium (Figure 1D and 1E).
Figure 1 Cilia Are Present on Both Mesenchymal and Ectodermal Cells of the Developing Limb
(A–C) Transmission electron micrographs of limb bud mesenchyme show cilia (arrows) closely associated with the Golgi apparatus (“G”). The cilia exhibited a 9 + 0 structure (C) and are often found in deep depressions in the membranes (B). Frequently, small vesicles are observed fusing or budding with the surrounding membrane (arrowheads in [B] and [C]).
(D and E) Scanning electron micrographs of the limb ectoderm show a single cilium (arrows) on nearly all ectodermal cells.
(F and G) Immunolocalization of Polaris (red) and acetylated α-tubulin (green) in frozen sections of limb buds shows that Polaris concentrates at the base and tip of the axoneme in both mesenchymal (F) and ectodermal (G) cells. Nuclei are blue.
(H and I) In primary cultures of cells isolated from E11.5 limb buds, cilia (arrow in H) are also present when visualized with anti-acetylated α-tubulin (green) and anti-Polaris (red) antisera (H). Cilia are absent on cells isolated from Tg737
Δ2–3β-gal mutant limb buds (I); however, the stabilized microtubules were still evident around the basal body region (arrow). The nuclear staining for Polaris is present in the Tg737
Δ2–3β-gal cells, indicating that it is nonspecific. Nuclei are blue.
To further confirm the presence of cilia in the limb bud, we conducted immunofluorescence analysis of frozen sections using anti-acetylated α-tubulin, which recognizes stabilized microtubules including the cilium axoneme, and anti-Polaris antiserum. The data indicate that Polaris concentrates at the base and distal tip of cilia on both ectodermal and mesenchymal cells as well as in a punctuate pattern overlapping that of acetylated α-tubulin in the axoneme (Figure 1F–1H). In primary cultures of limb bud cells, cilia were found on most cells when visualized with anti-acetylated α-tubulin and anti-Polaris antibodies (Figure 1H). In contrast, the cilia were completely absent from cells isolated from Tg737
Δ2–3β-gal mutants (Figure 1I). Domains of stabilized microtubules were still present around the microtubule organizing center (MTOC) from which the cilia would have emerged.
The Hedgehog Signaling Pathway Is Repressed in Tg737
Δ2–3β-gal Mutants
In agreement with previous data in the limb and neural tube [18–20], there was no significant expression of two downstream targets of Shh signaling, Ptch1 and Gli1, in Tg737
Δ2–3β-gal null mutant limb buds (Figure 2A and 2B). These data suggest that despite normal Shh expression in these mutants, Shh release or reception is impaired because of the loss of Polaris. These results confirm our assessment that the IFT mutant limb phenotype is not due to ectopic activation of the Shh pathway and that the phenotype in Tg737
Δ2–3β-gal mutants resembles that of Gli3 −/−;Shh−/− embryos [13,14].
Figure 2 Shh Signaling Is Defective in Tg737
Δ2–3β-gal Mutants
(A and B) In situ hybridization analysis of Ptch1 (A) and Gli1 (B) expression indicates that they are not expressed in the posterior limb buds of Tg737
Δ2–3β-gal mutant embryos (E10.5; right panels) as they are in wild-type controls (E10.5, left panels).
(C) Incubation of wild-type limb bud cells with ShhN-CM results in upregulation of Gli1 and Ptch1 expression (left lanes) compared to vector conditioned medium, whereas no increase is seen in cells isolated from Tg737
Δ2–3β-gal mutant limb buds (right lanes). The relative levels of induction standardized to actin are indicated below each lane.
Cells Lacking Polaris Are Unable to Respond to ShhN
To test whether Polaris is required for Shh reception, we isolated cells from Tg737
Δ2–3β-gal mutant and wild-type limb buds (E11.5) and cultured the cells in ShhN-CM. The ability of the cells to respond to ShhN was determined by induction of Ptch1 and Gli1 expression using semi-quantitative reverse transcription PCR (RT-PCR). While robust response to ShhN-CM was seen in wild-type cells, cells lacking Tg737 showed no increase in the expression of Ptch1 or Gli1 relative to control treated cells (Figure 2C). These data indicate that Polaris is required in Shh responding cells to activate the Shh signaling pathway in the presence of ligand.
Loss of Polaris Results in Altered Gli Activity and Processing
Genetic studies have indicated that IFT function is required for Shh signaling downstream of Ptch1, possibly at the level of Gli function [18–20]. To further explore the connection between Gli activity and Polaris, we used adenoviruses [21] to express the full-length Gli proteins in Tg737 null cells. Previous results have shown that ectopic expression of Gli1 and Gli2 can induce transcription of Shh target genes while Gli3 has been shown to inhibit Gli1-mediated transcription [8,21,22]. As seen in wild-type cells, infection of Tg737
Δ2–3β-gal primary limb cells with full-length Gli1 resulted in increased transcription of Ptch1 compared to infection with green fluorescent protein (GFP)–only virus (Figure 3A). This indicates that Polaris function is not required for Gli1-mediated pathway activation. However, infection of Tg737
Δ2–3β-gal primary limb bud cells with Gli2-expressing virus failed to induce Ptch1 transcription (Figure 3B) suggesting that Gli2 function requires the activity of Polaris. It is unclear at this time whether the loss of Gli2 function in Tg737
Δ2–3β-gal mutants is due to a requirement of Polaris for Gli2 stability or other post-translational regulation. As seen in previous studies, infection of cells with the full-length form of Gli3 was able to repress Gli1-mediated transcription when coexpressed in wild-type cells [22]. However, in cells lacking Polaris, full-length Gli3 failed to repress pathway activation by Gli1, as evidenced by increased Ptch1 expression (Figure 3A).
Figure 3 Gli2 and Full-Length Gli3 Function Is Disrupted in Tg737
Δ2–3β-gal Mutant Cells
(A) Infection of primary limb bud cells (E11.5) with Gli1::GFP expressing adenovirus induces increased Ptch1 transcription in wild-type cells when compared to infection with GFP-only virus (GFP). Coinfection of wild-type cells with Gli1::GFP and Gli3::GFP results in a decrease in the level of Ptch1 expression when compared to cells infected with Gli1::GFP only. As seen in wild-type cells, infection of Tg737
Δ2–3β-gal mutants with Gli1::GFP induced Ptch1 expression. However, full-length Gli3::GFP was unable to suppress Gli1::GFP-mediated induction of Ptch1 in the absence of Polaris (Tg737
Δ2–3β-gal). No expression was seen in controls without reverse transcriptase (−RT).
(B) Infection of wild-type cells with a Gli2::GFP expressing adenovirus induced Ptch1 expression; however, in Tg737
Δ2–3β-gal primary limb bud cells, infection with the Gli2::GFP expressing adenovirus failed to induce the pathway, when compared to infection with GFP-only virus (GFP, right lanes).
(C) Western blot analysis of proteins isolated from whole E11.5 wild-type embryos (left lane) shows that Gli3 is predominantly found in the processed repressor form (Gli3R). While some Gli3R is evident in the mutant samples, a large proportion of Gli3 remains unprocessed (Gli3A) in Tg737
Δ2–3β-gal mutants (right lane).
(D) Coinfection of wild-type or Tg737
Δ2–3β-gal mutant cells with Gli1::GFP and a truncated Gli3R::GFP indicates that Gli3R is able to repress Gli1-mediated induction of Ptch1.
Numbers below each lane in (A), (B), and (D) indicate the expression level of Ptch1 relative to the actin control for the experiment shown.
Gli3 Processing Is Inhibited by Loss of Polaris
The above data raised the possibility that loss of Polaris impaired the conversion of the full-length Gli3 to the truncated repressor form. To determine if this was the case, we examined the levels of full-length and processed forms of Gli3 in wild-type and Tg737
Δ2–3β-gal whole embryos (E11.5) by Western blot analysis using Gli3 antiserum (gift of B. Wang). In agreement with previously published results [19,20], there was a marked increase in the ratio of the full-length Gli3 to the processed form of Gli3R in Tg737 mutants (Figure 3C), although some Gli3R is clearly evident. Together, these data suggest that Polaris is required for efficient processing of Gli3.
To determine if the loss of Gli3-mediated repression in Tg737
Δ2–3β-gal mutants was due to defects in processing of full-length Gli3 to the repressor form, we infected primary cells with a truncated form of Gli3 (Gli3R) and analyzed the effect on Gli1-mediated transcription. In both wild-type and Tg737
Δ2–3β-gal mutant cells, the processed form of Gli3R was able to function as a potent repressor of Gli1-induced transcription of Ptch1 (Figure 3D). These results indicate that the loss of Gli3 activity observed with the full-length form is due to a defect in processing and not a loss of repressor activity or trafficking to the nucleus.
Partial Loss of Polaris Function Exacerbates the Phenotype of Gli3 Heterozygous Mutants
We predicated that if Polaris is required for proper Gli3 processing, partial loss of Polaris function as seen with the Tg737orpk hypomorphic allele would exacerbate the phenotype of Gli3 heterozygous mice and cause a phenotype that is more reminiscent of Gli3 null mutants. To evaluate this possibility, we crossed Tg737orpk/+ heterozygous mice with Gli3Xt-J/+;Tg737orpk/+ compound heterozygotes and correlated the resulting phenotypes with the genotypes of the embryos. Heterozygous Gli3Xt-J/+ mice are viable and exhibit a single additional preaxial digit similar to that seen in homozygous Tg737orpk/orpk mutants [6,12,23]. In contrast, homozygous Gli3Xt-J/Xt-J mutants are nonviable and have 8–10 nonpatterned digits per limb, as is also seen in Tg737
Δ2–3β-gal null mutants [12]. Intriguingly, no viable Gli3Xt-J/+;Tg737orpk/orpk offspring were obtained (0/64 pups; seven litters). Analysis at earlier developmental stages indicated that the Gli3Xt-J/+;Tg737orpk/orpk mice die during gestation with severe developmental abnormalities including 6–9 digits per limb, exencephaly, abdominal closure defects, and edema (Figure 4A–4E; data not shown). These phenotypes are not characteristic of Gli3Xt-J/+ heterozygous or Tg737orpk/orpk homozygous mice alone but are seen in Gli3Xt-J/Xt-J homozygous mutants.
Figure 4
Gli3Xt-J/+;Tg737orpk/orpk Embryos Resemble Gli3Xt-J/Xt-J Null Embryos and Are Responsive to ShhN-CM
(A) Example of exencephaly observed in Gli3Xt-J/+;Tg737orpk/orpk embryos that is never observed in Gli3Xt-J/
+ or Tg737orpk/orpk mutants alone.
(B–E) Functional interaction of Gli3 and Tg737 in digit development. Whereas Gli3Xt-J/+ (C) and Tg737orpk/orpk (D) embryos each develop one extra preaxial digit (asterisks), Gli3Xt-J/+;Tg737orpk/orpk embryos (E) develop multiple ectopic digits compared to wild-type embryos (B). Anterior is to the top.
(F) Incubation of cells from Gli3+/+;Tg737+/+, Gli3Xt-J/+;Tg737+/+, and Gli3Xt-J/+;Tg737orpk/orpk mutant mice with ShhN-CM resulted in increased expression of Gli1 when compared to cells from the same embryo treated with control medium, as determined by quantitative RT-PCR analysis. The results are reported for four littermates of the indicated genotypes. Each sample was analyzed in duplicate, and results are reported as the average fold increase. Unlike Tg737
Δ2–3β-gal (null) mutants, which are nonresponsive to ShhN-CM, the Gli3Xt-J/+;Tg737orpk/orpk samples are able to respond and activate the pathway, indicating that the Gli3Xt-J/+;Tg737orpk/orpk phenotype does not resemble that of Tg737
Δ2–3β-gal (null) mutants but rather that of Gli3Xt-J/ Xt-J.
While both Tg737
Δ2–3β-gal and Gli3Xt-J mutant mice have similar digit patterning phenotypes, Tg737
Δ2–3β-gal mutants do not express Ptch1 or Gli1 in the posterior limb bud while in Gli3Xt-J mutants the expression domains of Ptch1 or Gli1 are expanded. To determine if the limb patterning defects in Gli3Xt-J/+;Tg737orpk/orpk mutants were due to a loss of Shh responsiveness, as seen in Tg737
Δ2–3β-gal mutants, or more closely resembled the phenotype of Gli3Xt-J mutants, we tested primary cells derived from these embryos for their ability to induce Gli1 expression in response to ShhN-CM. In contrast to Tg737
Δ2–3β-gal mutants, Gli3Xt-J/+;Tg737orpk/orpk cells responded to ShhN-CM with increased expression of Gli1 when analyzed by quantitative RT-PCR, although at reduced levels compared to wild-type cells (Figure 4F). In agreement with the quantitative RT-PCR data, Gli1 expression was observed in the posterior region of all embryos by in situ hybridization (data not shown). No overt or consistent differences were evident in Gli3Xt-J/+ or Gli3Xt-J/+;Tg737orpk/orpk embryos when compared to wild-type (data not shown). Together these data indicate that the Gli3Xt-J/+;Tg737orpk/orpk phenotype more closely resembles that of Gli3 homozygotes than that of Tg737 null mutants, which are nonresponsive to ShhN-CM.
Exogenously Expressed Components of the Shh Signaling Pathway Localize to the Cilia
While it is known that IFT is required for Shh signaling [18,19], it remains unclear whether this is due to a requirement for cilia in Shh pathway activation, production of a secondary signal by cilia, or a novel non-ciliary role for IFT. To begin distinguishing between these possibilities, we evaluated the subcellular localization of several key proteins involved in the Shh signaling pathway, including Gli1, Gli2, Gli3, Gli3R, and Sufu relative to cilia and the IFT protein Polaris.
In the case of the Gli1, Gli2, Gli3, and Gli3R proteins, localization was determined by infection of primary Tg737
Δ2–3β-gal mutant and wild-type limb bud cells with adenoviral vectors that express the full-length Gli proteins or the truncated Gli3R fused to GFP [21]. Infections were performed such that greater than 75% of the cells expressed GFP. For these studies, we focused on cells that had low levels of exogenous expression to minimize any effects that overexpression may have on protein localization. For all three full-length Gli proteins, expression was detected in the nucleus in cells that expressed high levels of GFP (Figure 5; data not shown), as reported previously [22]. However, we also detected a small domain of GFP in all cells expressing the tagged protein that was located near the cilium axoneme as visualized with anti-acetylated α-tubulin antibodies (Figure 5A–5C). The GFP signals failed to colocalize with γ-tubulin (basal body marker), indicating that the Gli::GFP proteins do not localize to the basal body at the base of the cilia (Figure 5E; data not shown). Rather, the GFP signal was found to colocalize with a subdomain of Polaris (Figure 5F and 5G; data not shown). The colocalization of Gli::GFP with a domain of Polaris, but not with γ-tubulin, indicates that the full-length Gli proteins concentrate at the tip of the cilium but not at the base. Treatment of infected cells with ShhN-CM did not alter the distribution of Gli1, Gli2, or Gli3 at the distal tips of cilia (data not shown). However, it may be difficult to assess any changes in localization since GFP is fused to the C-terminus of the Gli proteins. Thus, processing that occurs in the case of full-length Gli3 would remove the GFP tag and prevent visualization of the truncated N-terminal form of the protein that traffics to the nucleus.
Figure 5 GFP-Tagged Gli Proteins Localize to the Distal Tip of the Cilium in Primary Limb Cell Cultures
(A–D) Cells were isolated from limb buds of wild-type embryos at E11.5 and infected with the indicated adenovirus. All three full-length GFP-tagged Gli proteins (green) localize to a domain in the cilium axoneme, which is visualized with anti-acetylated α-tubulin staining (red). In contrast, Gli3R::GFP is restricted to the nucleus and is not detected in this domain (D).
(E) The full-length Gli::GFP proteins (Gli2::GFP shown here) do not colocalize with the basal body at the base of the cilium, which is visualized with anti-γ-tubulin staining (red), indicating that the full-length Gli proteins localize to the tips of the cilia.
(F and G) Gli2::GFP (F) and Gli3::GFP (G) colocalize with a subdomain of Polaris (red) at the distal tip of the cilium.
(H–K) In Tg737
Δ2–3β-gal mutant limb bud cells, the GFP-tagged Gli1 (H), Gli2 (I), and Gli3 (J) proteins localize to the nucleus and in the region of stabilized microtubules around the MTOC marked by anti-acetylated α-tubulin. In contrast, the processed form of Gli3 (Gli3R::GFP) (K) is detected only in the nucleus.
Insets in all panels show the GFP (green) and nuclear (blue) staining only for the indicated cilium (arrow) or region (box).
In contrast to the localization observed for the three full-length Gli proteins, Gli3R::GFP was detected predominantly in the nucleus. We could detect no GFP signal at the distal tip of cilia, suggesting that after processing, Gli3R is released from the cilia or that the cilia targeting domain is located in the C-terminus of Gli3 (Figure 5D). These possibilities are currently being explored.
In Tg737
Δ2–3β-gal mutant cells that lack cilia, the Gli::GFP fusion proteins were seen in the nucleus, as observed in wild-type samples. Additionally, Gli1::GFP, Gli2::GFP, and Gli3::GFP were localized around the MTOC, where the cilia would have formed (Figure 5H–5J). In contrast, Gli3R was present mainly in the nucleus and was not detected around the MTOC in either the wild-type or Tg737
Δ2–3β-gal mutant cells (Figure 5D and 5K). The nuclear localization of the Gli::GFP proteins in Tg737
Δ2–3β-gal mutants suggests that Polaris is not required for nuclear import of the Gli transcription factors.
Endogenous Gli3 and Sufu Localize to the Tip of Cilia
To confirm the localization of GFP-tagged Gli3 at the tip of cilia, and to determine if this was the full-length Gli3 protein or only the GFP-tagged C-terminus, we conducted immunofluorescence analysis of endogenous Gli3 in noninfected primary limb cells using Gli3 antisera generated against the N-terminus of the protein (Figure 6). The data indicate that, as seen with exogenously expressed Gli3::GFP, endogenous Gli3 was concentrated at the tip of cilia (Figure 6A; data not shown). Since the Gli3 antiserum recognizes the N-terminus of Gli3, and GFP is fused to the C-terminus in Gli3::GFP virus, the data suggest that it is the full-length form of Gli3 that localizes to the cilium tip.
Figure 6 Endogenous Sufu and Gli3 Localize to the Distal Tip of the Cilium in Wild-Type Primary Limb Cell Cultures
Sufu (green) and endogenous Gli3 (red) concentrate in the same domain in cultured wild-type limb bud cells (A). As shown for the full-length Gli::GFP proteins, endogenous Sufu does not colocalize with γ-tubulin (red) (B), but is concentrated in a domain at the distal end of the acetylated α-tubulin staining (red) (C). Sufu also partially overlaps with a domain of Polaris (red) (D) in cultured wild-type limb bud cells. Pre-incubation of anti-Sufu antiserum with the immunizing peptide (E), but not with a nonspecific peptide (F), blocks staining at the distal tip of the cilium (anti-Polaris, red; anti-Sufu, green). Inset in all panels shows Gli3 (A) or Sufu (B–F) staining only in the indicated cilium (arrow).
Since all three Gli proteins localize to the tip of the cilium and to the nucleus, and since Sufu has been shown to directly interact with the Gli proteins [17], we predicted that Sufu would also be present in these two regions of the cell. To explore this possibility, we analyzed the localization of endogenous Sufu by immunofluorescence in primary limb bud cultures. The data confirm that endogenous Sufu colocalized with endogenous Gli3 and with the Gli::GFP fusion proteins, and partially overlapped with Polaris (Figure 6A and 6D; data not shown). It was also present adjacent to acetylated α-tubulin in the cilium axoneme and did not localize with γ-tubulin at the basal body, again indicating that these proteins concentrate at the tip of the cilium (Figure 6B and 6C). A low level of Sufu was also evident in the nucleus but was in a different plane of focus when the Sufu image for Figure 6 was captured. In addition, Sufu was detected in the cytosol; however, whereas the nuclear and cilium tip signals were blocked by preincubation with the immunizing peptide, the cytoplasmic signal remained unchanged (Figure 6E and 6F), suggesting that it may be nonspecific.
Discussion
Cilia are expressed on many different cell types in the mammalian body. They are formed and maintained by a highly conserved process termed IFT, but they perform diverse functions on various cell types [1]. In mammals, cilia have been demonstrated to play a critical role in developmental processes—from left–right axis specification and skeletal patterning to normal kidney, pancreas, and liver physiology—as well as disease processes [2–4,6,24]. Recent data have demonstrated that the IFT proteins, which are necessary for cilia formation, are also required for proper limb and neural tube patterning [6,18–20]. Furthermore, the IFT proteins have been shown to function as part of the Shh signal transduction pathway and in regulating Gli activity [19,20]. Here we show that Gli2 and full-length Gli3 function are disrupted in the Tg737
Δ2–3β-gal cilium mutants. In contrast, our data indicate that Gli1 and a processed form of Gli3 (Gli3R) are able to induce or repress the Shh pathway, respectively, regardless of the presence or absence of Polaris. Most important, all three full-length Gli::GFP proteins, as well as endogenous Gli3 and Sufu, localize to the distal tip of cilia in primary limb bud cell cultures, supporting a direct role for cilia in regulating Shh signal transduction.
Mice homozygous for the hypomorphic Tg737orpk allele have preaxial polydactyly, but no alterations in expression of Shh or its downstream targets are evident [6]. Preaxial polydactyly is also seen in mice heterozygous for the Xt-J allele of Gli3 [10]. Interestingly, Gli3Xt-J/+;Tg737orpk/orpk mice develop multiple ectopic digits on all four limbs and die during gestation. This phenotype is reminiscent of both Gli3Xt-J and Tg737
Δ2–3β-gal homozygous mutants. However, unlike cells from Tg737
Δ2–3β-gal mutants, cells from Gli3Xt-J/+;Tg737orpk/orpk mice are able to induce Gli1 transcription in response to ShhN-CM. These data suggest that the hypomorphic Tg737orpk allele further disrupts Gli3 function and converts the Gli3Xt-J/+ heterozygotes to a phenotype more similar to that seen in the Gli3 null mutants.
Previous genetic studies on IFT mutant mice have indicated that IFT proteins function in Shh signal transduction at a step downstream from the membrane proteins Ptch and Smoothened and are required for all Gli function [18–20]. In contrast, our data suggest Polaris is required for specific Gli functions since Gli1 and truncated Gli3 (Gli3R) are active in the absence of Polaris while Gli2 and full-length Gli3 are not. These data suggest that the defects observed when mutant cells are infected with full-length Gli3 are likely due to loss of processing of Gli3 to form the repressor, but not loss of repressor activity. While processing of Gli2 remains controversial, Gli2 function is disrupted in Tg737
Δ2–3β-gal mice, suggesting that IFT or cilia are required for some aspect of Gli2 regulation or activity. This does not appear to involve translocation of the Gli proteins to the nucleus since they are all detectable in the nucleus of Tg737
Δ2–3β-gal mutant cells.
In Drosophila, Su(fu) is involved in the negative regulation of Ci (Gli homolog) transcriptional activity by sequestering it in the cytoplasm and targeting it for proteolytic processing to produce a transcriptional repressor [8]. Whether the role of Sufu in targeting Gli proteins for proteolytic processing is conserved in mammals has yet to be determined. Mammalian Sufu has been shown to interact with all three Gli proteins through a conserved SYGH motif in the N-terminus of the Gli proteins in addition to a region in the C-terminus of Gli1, and negatively regulates Gli1 transcriptional activity [16,17]. The colocalization of Sufu and the Gli proteins to the tip of the cilium, along with a requirement for IFT in proper Gli3 processing, suggests that mammalian Sufu may have a similar role in Gli regulation and, furthermore, that proteolytic processing may occur at the tip of the cilium.
Huangfu et al. [18] proposed two possible models for how IFT may regulate Shh signaling, one suggesting the involvement of a cilia-derived signal that is required for Shh pathway activation and a second model in which IFT has two separate functions, one in ciliogenesis and a second one in intracellular transport. While testing these models is hindered by our inability to specifically disrupt cilia formation without also perturbing IFT, the data presented here support a direct role for cilia in Shh pathway regulation. This is based on the localization of multiple components of the Shh pathway in the cilia of wild-type cells, and on altered Gli3 processing and impaired Gli2 function detected in cells lacking this organelle. While we cannot conclusively rule out a non-ciliary function of IFT, we propose that IFT functions to direct and concentrate the Gli proteins, Sufu, and possibly the proteolytic machinery needed for efficient processing of the Gli proteins to a domain located at the distal tips of cilia. In the absence of the cilium, the Gli proteins localize diffusely around the basal body region and fail to undergo normal processing, resulting in their impaired activity.
Materials and Methods
Mouse strains and methods.
Tg737
Δ2–3β-gal, Tg737orpk, and Gli3Xt-J mice have been previously described [3,4,23]. Tg737
Δ2–3β-gal mice were maintained on a mixed FVB × BALB/c background and genotyped as described [3]. Tg737orpk mice were maintained on an FVB background and genotyped as described [3]. Gli3Xt-J mice were maintained on a C57BL/6 background and were genotyped as described [3,25]. Analysis of phenotypes for Gli3Xt-J;Tg737orpk mice was performed on pups or embryos from F1 intercrosses. For staged embryos, noon of the day of the vaginal plug appeared was considered E0.5.
In situ hybridization analysis was performed according to standard protocols [26]. Ptch1 and Gli1 probes were previously described [27,28].
Skeletal stains were performed as described [29].
Cell culture.
ShhN-CM and vector-only control conditioned medium were generated as previously described [30]. For induction assays, cells were cultured in a 1:1 mixture of conditioned medium and DMEM + 15% FBS overnight prior to RNA isolation.
Embryos were isolated and identified by phenotype (Tg737
Δ2–3β-gal mutants) or by PCR using DNA isolated from yolk sacs. Limb buds for cell culture experiments were removed and treated with 0.25% trypsin in PBS for 15 min at room temperature. Following trypsin treatment, cells were mechanically dissociated and FBS was added to 10%. Cells were collected by centrifugation and plated with DMEM + 15% FBS.
Gli1::GFP, Gli2::GFP, and Gli3::GFP adenovirus constructs encoding C-terminal GFP fusions have been previously described [21]. The Gli3R::GFP adenovirus was generated by replacing the full-length cDNA in the Gli3::GFP vector with the coding region corresponding to amino acids 1–677. This truncated form of Gli3 (Gli3R) has been previously shown to act as a constitutive repressor of Shh signaling [31]. All infections were optimized to produce greater than 75% infection and were done at least three times. To determine if expression levels were consistent between samples, we examined the level of expression of the Gli::GFP genes by RT-PCR using primers specific for the GFP coding region. Similar levels of expression were seen in wild-type and mutant samples under the same infection conditions. All localization data shown are representative of the pattern observed in the majority of the cells from all experiments. An identical pattern of localization was observed in the IMCD mouse kidney cell line. No specific localization of GFP was found for cells infected with GFP control virus only.
RNA isolation and RT-PCR.
RNA was isolated using TRIzol (Invitrogen, Carlsbad, California, United States) according to the manufacturer's instructions. Reverse transcription was performed using SuperScript II reverse transcriptase (Invitrogen) according to the manufacturer's instructions. Equal amounts of cDNA were used as templates for PCR with Taq Polymerase (Brinkman Instruments, Westbury, New York, United States) according to the manufacturer's instructions. Relative expression levels were calculated by comparing the intensity of the Ptch1 or Gli1 PCR product to the actin PCR product in the same reaction using LabWorks 4.0 software (UVP, Upland, California, United States). Primer sequences are available upon request.
Quantitative RT-PCR measurement was performed using the SmartCycler machine (Cepheid, Sunnyvale, California, United States). The TaqMan primer and probe sets, for Gli1 and 18S rRNA (TaqMan Assays-on-Demand Products), were purchased from Applied Biosystems (Foster City, California, United States). The 18S rRNA gene was used as an internal control. The threshold cycle (C
T) for Gli1 was first normalized to the corresponding 18S rRNA C
T. Relative fold differences were then determined using the 2(–Δ,Δ[C
T
]) method [32] by comparing the expression levels in ShhN-CM-induced cells to their vector conditioned medium controls. No significant difference in basal Gli1 expression in vector conditioned medium–treated cells was evident between wild-type and Gli3Xt-J/+ samples.
Immunofluorescence.
For analysis of cilia in vivo, limb buds were dissected from wild-type embryos (E10.5), embedded in OCT, and snap frozen. Sections of 20 μm were cut and stained as previously described [33] using 0.2% Triton X-100 for permeabilization. Cultured primary cells were fixed, permeabilized, and stained using an identical procedure. Anti-Polaris polyclonal antibody was generated by Sigma-Genosys (The Woodlands, Texas, United States) and screened for specificity by Western blot analysis and immunofluorescence. The antiserum recognized a single band of the correct size in wild-type samples by Western blot analysis; this band was absent in Tg737
Δ2–3β-gal samples. Only faint nuclear staining was observed in Tg737
Δ2–3β-gal primary cells by immunofluorescence, whereas localization to cilia was additionally observed in wild-type samples. Sufu antibody and blocking peptide were obtained from Santa Cruz Biotechnology (Santa Cruz, California, United States). Identical staining with the Sufu antibody was seen in the IMCD mouse kidney cell line. Affinity-purified anti-Gli3 antiserum was provided by B. Wang and used as previously described [9]. Acetylated α-tubulin and γ-tubulin antibodies were obtained from Sigma (St. Louis, Missouri, United States). Gli::GFP fusion proteins were visualized using the GFP tag encoded by the adenovirus constructs. Nuclei were stained with Hoechst 33258 (Sigma).
All fluorescence imaging was performed on a Nikon (Tokyo, Japan) TE200 Eclipse inverted epifluorescence microscope equipped with a CoolSnap HQ cooled CCD camera (Roper Scientific, Trenton, New Jersey, United States) and MetaMorph imaging software (Molecular Devices, Downington, Pennsylvania, United States). All filters and shutters were computer driven.
Electron microscopy.
Wild-type E11.5 embryos were isolated in PBS and fixed in 2.5% gluteraldehyde in 0.1 M cacodylate buffer for 90 min at room temperature. Samples were washed in three changes of 0.1 M cacodylate buffer, fixed in 1% OsO4 for 60 min at room temperature, washed in thee changes of 0.1 M cacodylate buffer, and dehydrated through a graded ethanol series. Samples were processed for scanning or transmission electron microscopy by the University of Alabama at Birmingham High Resolution Imaging Facility using standard procedures. Scanning electron microscopy samples were imaged on an ISI SX-40 scanning electron microscope (Topcon Technologies, Paramus, New Jersey, United States). Transmission electron microscopy sections were imaged on a Zeiss (Oberkochen, Germany) EM 10C transmission electron microscope.
We thank C. Chang, G. Marqués, and S. Nozell for critical reading of the manuscript, B. Wang for Gli3 antiserum, D. Robbins for the ShhN expression construct, C.-M. Fan for the Gli::GFP adenovirus constructs, M. Croyle for technical assistance, and the members of our laboratories for helpful discussions. We thank the University of Alabama at Birmingham High Resolution Imaging Facility for assistance with electron microscopy sample preparation and analysis. This work was supported by a March of Dimes Grant (1-FY04–95) to BKY. Additional support was provided to CJH through a National Institutes of Health T32 training grant (DK07545) to D. Benos. YAS and EJM received support from the Office of Biological and Environmental Research, US Department of Energy, under contract DE-AC05-00OR22725 with UT-Battelle.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. CJH and BKY conceived and designed the experiments. CJH, BB, YAS, QZ, and BKY performed the experiments. CJH, YAS, EJM, and BKY analyzed the data. CJH and BKY wrote the paper.
A previous version of this article appeared as an Early Online Release on September 26, 2005 (DOI: 10.1371/journal.pgen.0010053.eor).
Abbreviations
E[number]embryonic day [number]
GFPgreen fluorescent protein
GliGlioma
IFTintraflagellar transport
MTOCmicrotubule organizing center
PtchPatched
RT-PCRreverse transcription PCR
ShhSonic hedgehog
ShhN-CMShhN conditioned medium
SufuSuppressor of fused
==== Refs
References
Scholey JM 2003 Intraflagellar transport Annu Rev Cell Dev Biol 19 423 443 14570576
Pazour GJ Dickert BL Vucica Y Seeley ES Rosenbaum JL 2000 Chlamydomonas IFT88 and its mouse homologue, polycystic kidney disease gene Tg737, are required for assembly of cilia and flagella J Cell Biol 151 709 718 11062270
Murcia NS Richards WG Yoder BK Mucenski ML Dunlap JR 2000 The Oak Ridge Polycystic Kidney (orpk) disease gene is required for left-right axis determination Development 127 2347 2355 10804177
Moyer JH Lee-Tischler MJ Kwon HY Schrick JJ Avner ED 1994 Candidate gene associated with a mutation causing recessive polycystic kidney disease in mice Science 264 1329 1333 8191288
Yoder BK Tousson A Millican L Wu JH Bugg CE Jr 2002 Polaris, a protein disrupted in orpk mutant mice, is required for assembly of renal cilium Am J Physiol Renal Physiol 282 F541 F552 11832437
Zhang Q Murcia NS Chittenden LR Richards WG Michaud EJ 2003 Loss of the Tg737 protein results in skeletal patterning defects Dev Dyn 227 78 90 12701101
Tickle C 2003 Patterning systems—From one end of the limb to the other Dev Cell 4 449 458 12689585
Ingham PW McMahon AP 2001 Hedgehog signaling in animal development: Paradigms and principles Genes Dev 15 3059 3087 11731473
Wang B Fallon JF Beachy PA 2000 Hedgehog-regulated processing of Gli3 produces an anterior/posterior repressor gradient in the developing vertebrate limb Cell 100 423 434 10693759
Mo R Freer AM Zinyk DL Crackower MA Michaud J 1997 Specific and redundant functions of Gli2 and Gli3 zinc finger genes in skeletal patterning and development Development 124 113 123 9006072
Park HL Bai C Platt KA Matise MP Beeghly A 2000 Mouse Gli1 mutants are viable but have defects in SHH signaling in combination with a Gli2 mutation Development 127 1593 1605 10725236
Masuya H Sagai T Wakana S Moriwaki K Shiroishi T 1995 A duplicated zone of polarizing activity in polydactylous mouse mutants Genes Dev 9 1645 1653 7628698
Litingtung Y Dahn RD Li Y Fallon JF Chiang C 2002 Shh and Gli3 are dispensable for limb skeleton formation but regulate digit number and identity Nature 418 979 983 12198547
te Welscher P Zuniga A Kuijper S Drenth T Goedemans HJ 2002 Progression of vertebrate limb development through SHH-mediated counteraction of GLI3 Science 298 827 830 12215652
Stone DM Murone M Luoh S Ye W Armanini MP 1999 Characterization of the human suppressor of fused, a negative regulator of the zinc-finger transcription factor Gli J Cell Sci 112 4437 4448 10564661
Ding Q Fukami S Meng X Nishizaki Y Zhang X 1999 Mouse suppressor of fused is a negative regulator of sonic hedgehog signaling and alters the subcellular distribution of Gli1 Curr Biol 9 1119 1122 10531011
Dunaeva M Michelson P Kogerman P Toftgard R 2003 Characterization of the physical interaction of Gli proteins with SUFU proteins J Biol Chem 278 5116 5122 12426310
Huangfu D Liu A Rakeman AS Murcia NS Niswander L 2003 Hedgehog signalling in the mouse requires intraflagellar transport proteins Nature 426 83 87 14603322
Liu A Wang B Niswander LA 2005 Mouse intraflagellar transport proteins regulate both the activator and repressor functions of Gli transcription factors Development 132 3103 3111 15930098
Huangfu D Anderson KV 2005 Cilia and Hedgehog responsiveness in the mouse Proc Natl Acad Sci U S A 102 11325 11330 16061793
Buttitta L Mo R Hui CC Fan CM 2003 Interplays of Gli2 and Gli3 and their requirement in mediating Shh-dependent sclerotome induction Development 130 6233 6243 14602680
Ruiz i Altaba A 1999 Gli proteins encode context-dependent positive and negative functions: Implications for development and disease Development 126 3205 3216 10375510
Hui CC Joyner AL 1993 A mouse model of Greig cephalopolysyndactyly syndrome: The extra-toesJ mutation contains an intragenic deletion of the Gli3 gene Nat Genet 3 241 246 8387379
Zhang Q Davenport JR Croyle MJ Haycraft CJ Yoder BK 2005 Disruption of IFT results in both exocrine and endocrine abnormalities in the pancreas of Tg737(orpk) mutant mice Lab Invest 85 45 64 15580285
Maynard TM Jain MD Balmer CW LaMantia AS 2002 High-resolution mapping of the Gli3 mutation extra-toes reveals a 51.5-kb deletion Mamm Genome 13 58 61 11773971
Wilkinson DG 1992 In situ hybridization: A practical approach New York IRL Press 163 p.
Goodrich LV Johnson RL Milenkovic L McMahon JA Scott MP 1996 Conservation of the hedgehog/patched signaling pathway from flies to mice: Induction of a mouse patched gene by Hedgehog Genes Dev 10 301 312 8595881
Hui CC Slusarski D Platt KA Holmgren R Joyner AL 1994 Expression of three mouse homologs of the Drosophila segment polarity gene cubitus interruptus, Gli, Gli-2, and Gli-3, in ectoderm- and mesoderm-derived tissues suggests multiple roles during postimplantation development Dev Biol 162 402 413 8150204
McLeod MJ 1980 Differential staining of cartilage and bone in whole mouse fetuses by alcian blue and alizarin red S Teratology 22 299 301 6165088
Zeng X Goetz JA Suber LM Scott WJ Jr Schreiner CM 2001 A freely diffusible form of Sonic hedgehog mediates long-range signalling Nature 411 716 720 11395778
Meyer NP Roelink H 2003 The amino-terminal region of Gli3 antagonizes the Shh response and acts in dorsoventral fate specification in the developing spinal cord Dev Biol 257 343 355 12729563
Livak KJ Schmittgen TD 2001 Analysis of relative gene expression data using real-time quantitative PCR and the 2(-delta delta C(T)) method Methods 25 402 408 11846609
Taulman PD Haycraft CJ Balkovetz DF Yoder BK 2001 Polaris, a protein involved in left-right axis patterning, localizes to basal bodies and cilia Mol Biol Cell 12 589 599 11251073
|
16254602
|
PMC1270009
|
CC0
|
2021-01-05 08:00:57
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e53
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010053
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460310.1371/journal.pgen.0010054plge-01-04-10ReviewPerspectives on Human Genetic Variation from the HapMap Project McVean Gil *Spencer Chris C. A Chaix Raphaelle Gil McVean, Chris C. A. Spencer, and Raphaelle Chaix are in the Department of Statistics, University of Oxford, Oxford, United Kingdom.
*To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 1 4 e54Copyright: © 2005 McVean et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.ABSTRACT
The completion of the International HapMap Project marks the start of a new phase in human genetics. The aim of the project was to provide a resource that facilitates the design of efficient genome-wide association studies, through characterising patterns of genetic variation and linkage disequilibrium in a sample of 270 individuals across four geographical populations. In total, over one million SNPs have been typed across these genomes, providing an unprecedented view of human genetic diversity. In this review we focus on what the HapMap project has taught us about the structure of human genetic variation and the fundamental molecular and evolutionary processes that shape it.
Citation:McVean G, Spencer CCA, Chaix R (2005) Perspectives on human genetic variation from the HapMap project. PLoS Genet 1(4): e54.
==== Body
Introduction
In human genetics, association studies aim to identify loci that contribute to disease susceptibility by comparing patterns of genetic variation between people with a disease (cases) and those without (controls) [1]. Without any prior knowledge about which genes are likely to be important, the researcher faces the expensive possibility of trying to look at all the 10 million or so polymorphic sites in the genome where the less common allele has a frequency of at least 1%, not to mention polymorphic inversions, duplications, microsatellites, and other forms of heritable variation. However, in recent years a number of empirical studies have revealed a structure to human genetic variation that could dramatically reduce the cost of association studies [2–9]. In particular, alleles at nearby loci often show strong statistical association (known as linkage disequilibrium [LD]). Coupled with observations that human recombination is concentrated into short (1–2 kb) hotspots that occur every 100–200 kb [10–12], and that these recombination hotspots are often coincident with a breakdown of allelic association [10], efficient genome-wide association studies became a possibility [13] because a few markers within each domain of strong association can be used to tag nearby variation. Here we use the term “tag” to imply that statistical tests for association carried out by using selected marker loci are as powerful (or nearly so) as if all single nucleotide polymorphisms (SNPs) were included.
However, in order to define efficient markers for subsequent studies, local knowledge of the structure of genetic variation across the genome is required. Choosing SNPs at set intervals across the genome, as one might in linkage studies, will fail to capture local patterns of allelic association and will consequently fail to tag efficiently. For this reason, the International HapMap Project was founded in 2002, with the goal of mapping the structure of allelic association across the human genome [14]. With the participation of funding agencies, academic research centres, and industrial partners in many countries, the initial aim was to genotype one SNP every 5 kb in the human genome across 270 individuals from four geographical populations. These individuals are 30 mother–father–offspring trios from the Yoruba people of Ibidan Peninsula in Nigeria (referred to as YRI), 30 such trios from the CEPH project in Utah (CEU), 45 unrelated individuals from the Han Chinese population of Beijing (CHB), and 45 unrelated individuals of Japanese ancestry from the Tokyo area (JPT) (for many analyses the CHB and JPT samples are combined within a single “analysis panel”). This project, referred to as the Phase 1 HapMap, is now complete, and the data, with associated summaries and query-based tools, are available online at http://www.hapmap.org with an accompanying manuscript published in Nature [15]. Further phases of the project, involving the typing of nearly 4 million SNPs across the same samples, and SNPs in a limited set of regions across multiple other population samples, are also under way.
What have we learnt from the project? For the medical geneticist the good news is that whole-genome association studies are still looking feasible. Technologies that provide high-throughput whole-genome genotyping of a few hundred thousand well-chosen SNPs should provide adequate power in most populations to detect single-locus associations for SNPs of moderate frequency and relative risk (we are being deliberately vague because the exact details depend on sample size and disease parameters [13]). Of course, not all complex diseases will have such an obvious genetic aetiology, and efforts to look for rare SNP effects [16], genetic interactions [17], or genotype-by-environment interactions [18] in candidate regions will no doubt also be fruitful. Furthermore, the design and analysis of association studies is still very much an area of active research that will only really be understood when large-scale association studies start becoming a reality.
However, while the use of the HapMap data for future association studies is the primary goal of the project, it also provides an unprecedented view of human genetic diversity that has provided novel insight into many other areas of biological interest. These include the distribution of recombination hotspots and coldspots, the effects of natural selection, and how these forces and others interact to shape human genetic variation. Our personal understanding of LD and how it relates to the underlying evolutionary and molecular forces has changed enormously through staring hard at more than a quarter of a billion genotypes. Therefore, what we are setting out to present in this review is a highly subjective set of observations made from the HapMap data that reflect what we have learned about the structure of human genetic variation.
Understanding the Structure of Human Genetic Variation
Every chromosome carries a unique combination of alleles that is known as a haplotype. However, within regions of about 500 kb and less it is possible to find combinations of SNPs that are found in multiple unrelated individuals. Such “blocks” point to regions that have not been broken up by recombination and are often separated from each other by short regions where there is evidence for considerable recombination (recombination hotspots). These observations led to the idea of the human genome as a colourful mosaic of haplotype blocks delimited by recombination hotspots [19]. While this model is helpful in conveying the broad nature of human genetic variation, it fails to capture the true complexity. In this section we discuss four observations arising from analysis of the HapMap data that help to provide a more complete picture of the nature and causes of LD and genetic variation.
In non-recombining regions, the genealogical tree determines the strength of LD.
Recombination acts to break down associations between alleles that arise because new mutations appear on a single genetic background. As we might expect, associations between alleles at loci separated by considerable genetic distances show consistently low levels of LD as measured by any statistic. However, and perhaps surprisingly, the converse is not true. Certain statistics of LD, and in particular the degree of statistical association between alleles as measured by the square of the correlation coefficient, r
2 [20], can take low values even in regions of low or no recombination (r
2 is the most relevant measure of LD for association studies because of the one-to-one relationship between r
2 and the relative power of statistical tests at a marker locus compared to the causative locus [21]).
Why can LD be low even in non-recombining regions? When there is no recombination, all parts of the sequence share the same genealogical tree. So in terms of determining the strength of associations, what is important is where mutations appear in this tree (Figure 1). Two mutations that occur on the same branch of the genealogy will be present on the same chromosomes and, hence, will be in complete association. In contrast, two mutations that occur in completely different parts of the tree will occur in different chromosomes, and may only be weakly associated. This is really just another way of saying that the r
2 measure of LD is dependent on allele frequencies [22], but it has important consequences for association studies because the genealogical history of chromosomes taken from different parts of the world (or even repeat samples from the same places) are likely to be different.
Figure 1 The Relationship between Genealogical History and Allelic Association
The upper part of the figure represents the genealogy for the 13 haplotypes observed in a 40-kb region of Chromosome 1 (between SNPs rs12085605 and rs932087) where there is no evidence for recombination (for no pair of SNPs are all four possible combinations of alleles observed), with the location of polymorphic mutations indicated by circles. The lower part of the figure indicates the relative frequency of each haplotype in the sample from each of the three panels (in greyscale, with white indicating 0% and black indicating 100%). The dotted line in the genealogy indicates a branch of the tree that is not present in the CEU sample and whose removal results in perfect association between SNPs rs12085824 and rs11205476.
We can see this effect in the example shown in Figure 1, a 40-kb region of chromosome 1. Here, we find 17 SNPs that show no evidence for recombination and result in 13 unique haplotypes that can be related to each other through a perfect phylogeny (i.e., there is no need to invoke repeat or back mutation). As one might expect, we observe differences in haplotype frequencies between panels, with the majority of haplotypes being found in only one panel (seven haplotypes are present in one panel only, compared to three being found in all). The difference in haplotype distribution leads to differences in allelic association; for example, SNPs rs12085824 and rs11205476 are in complete association in CEU (r
2 = 1), in strong association in CHB + JPT (r
2 = 0.88), and only moderately associated in YRI (r
2 = 0.58). More importantly, there is a clade of the genealogy (represented by the dotted line) that is not represented in the CEU sample (though it might be found with deeper sampling). Without this clade, the two SNPs effectively occur on the same branch and are therefore in complete association. The practical implication of this observation is that tagging choices may well be population specific, even in regions of low or no recombination. However, another more exciting possibility is that such differences between populations in genealogical trees constructed from non-recombining regions across the whole genome will provide novel insights into the demographic history of modern humans.
High-frequency haplotypes can cross recombination hotspots.
As stated above, within a population, associations between alleles separated by large genetic (recombination) distances are consistently low. But how large a distance is large? For example, is a single recombination hotspot sufficient to break down all associations? Put another way, if we are interested in tagging variation, should we break the genome into regions separated by recombination hotspots, or can tagging across hotspots ever be effective?
The answer is fairly straightforward. Recombination hotspots are rarely strong enough to remove all allelic association across them. Often, and particularly in the CEU, CHB, or JPT population samples, we find common haplotypes (at frequency of 10% and higher) that span recombination hotspots. Figure 2 demonstrates the relationship between common haplotypes and recombination rates in the ENCODE region on Chromosome 7q31.33 (data from [15]). As might be expected, haplotypes are considerably longer in CEU and CHB + JPT than in the YRI sample, reflecting the effect historical bottlenecks can have in reducing haplotype diversity and creating large haplotypes that take many hundreds of generations to be broken up by recombination. What is striking is that only one hotspot out of the six identified in the region is sufficiently hot to break all common haplotypes. Actually, we should not be particularly surprised by this result. At the hottest recombination hotspot identified across the autosomes, we would expect only one cross-over event in 114 meioses (a genetic distance of 0.9 cM), and at the “average” hotspot we would expect a recombination event every 1,300 meioses (0.075 cM).
Figure 2 Patterns of Haplotype Structure and Recombination in the HapMap ENCODE Region on Chromosome 7q31.33
The estimated recombination rate (in centimorgans per megabase) is shown as a dark blue line, with statistically significant recombination hotspots (see [15] for details) as grey lines. For each analysis panel, each non-redundant haplotype with a frequency of at least 10% is represented by a horizontal line between the starting and ending SNPs (see [15] for details of methodology); the vertical height of these lines is arbitrary. Note that only one of the six hotspots is sufficiently strong to break all common haplotypes.
Untaggable SNPs typically, but not always, occur in recombination hotspots.
No matter how hard you try, for certain SNPs there is just no other variant in the human genome that is in sufficient association to work effectively as its tag. Such “untaggable” SNPs are only problematic for association studies if you don't know where they are (otherwise they can just be included in genotyping studies). However, because even Phase II of the HapMap project will not type every SNP in the genome, it is important to learn about the distribution of such SNPs. In particular, can we predict where they might occur?
To answer this question we need to turn to the HapMap ENCODE project. This refers to a study within the project that resequenced 500 kb from each of ten ENCODE regions in 16 chromosomes from each analysis panel (i.e., a total of 48 chromosomes), followed by genotyping of all identified SNPs in the entire HapMap sample. While this does not provide complete ascertainment, it is expected to have identified almost all common (minor allele frequency > 5%) SNPs in each region (the average density of common SNPs is 1.5 per kilobase). A very high proportion of all common SNPs have at least one highly efficient potential tag (r
2 ≥ 0.8; 92% in CEU, 90% in CHB + JPT, and 80% in YRI), and the figures get better if you allow for less-efficient tagging and/or a higher threshold on minor allele frequency. However, across the ten ENCODE regions, a handful of really high frequency SNPs (minor allele frequency > 25%) have no tags at all (maximum r
2 < 0.2; 11/3,261 in CEU, 12/3,270 in CHB + JPT, and 20/2,961 in YRI).
What might cause a really common SNP to be untaggable? One obvious possibility is that these SNPs lie in the middle of recombination hotspots. Figure 3 shows the location of the untaggable SNPs in two of the ENCODE regions, along with the estimated recombination rate profile. In the region on Chromosome 2q37.1, all untaggable SNPs fall in the middle of recombination hotspots. This is also true for two of the four untaggable SNPs in the region on Chromosome 7q31.33, but we need a different explanation for the other two in this region. One possibility is just chance. As seen above, even if there is no recombination, genealogical structure can lead to differences in allelic association between populations, and neither of these untaggable SNPs is completely untaggable in all populations. It is also possible that these SNPs might be hypermutable sites (such as methylated cytosine–guanine dinucleotides), or that they are hotspots of gene conversion, or that they have a high error rate (all of which would lead to low allelic association). Whatever the cause, the conclusion is that untaggable SNPs, while concentrated in recombination hotspots, are not restricted to them.
Figure 3 The Relationship between Recombination Rate, Recombination Hotspots, and the Location of Untaggable SNPs
For two HapMap ENCODE regions the estimated recombination rate (dark blue line) and the location of statistically significant hotspots (grey lines) are shown along with the location of SNPs that are untaggable in the YRI (green) CEU (red), or CHB + JPT (purple) panels. Note that most, but not all, untaggable SNPs occur in recombination hotspots.
Regions of unusual genetic variation point to interesting biological features.
There is great heterogeneity across the genome in terms of patterns of genetic variation. Some of this heterogeneity is due to variation in factors such as mutation rate and recombination rate. Some of this heterogeneity arises because of the stochastic properties of mutation and genealogical history. But there are also other forces such as natural selection and genomic features such as inversions that may influence local patterns of variation. How can we look for the effects of such factors? There are two approaches. Either we can try to predict what we would expect to observe under models with and without such effects [23,24], or we can simply look at the empirical distribution of statistics of genetic variation and take as candidate regions those showing extreme or unusual patterns. The difficulty of the first approach is that accurately modelling human variation (and SNP ascertainment) is probably impossible. The difficulty of the latter approach is that there is no guarantee that empirically unusual patterns point to biologically interesting features.
However, it is possible to validate empirical approaches by asking whether regions where independent evidence points to biological interest are outliers in terms of genetic variation (or alternatively identify the statistics that identify such regions as unusual). The good news is that several genes or features for which biological interest is known do stand out as being unusual in the HapMap data in some sense. For example, the lactase gene (LCT, associated with lactose tolerance) has one of the highest relative extended haplotype homozygosity (rEHH [25]) scores in the CEU population, as does the beta-globin gene (HBB, associated with protection against Plasmodium falciparum malaria) in the YRI population. The HLA region (associated with resistance to multiple infectious diseases [26]) is one of a handful of gene clusters across the genome where there are haplotypes at frequencies of 1% across the combined population sample that span over 500 SNPs and more than 1cM. The known polymorphic inversion on Chromosome 17 [27] stands out as having the greatest number of SNPs in complete association (66 SNPs with r
2 = 1 in Phase I HapMap) in the entire genome, and there are only 33 nonsynonymous SNPs across the Phase I HapMap that show as much population differentiation as the SNP rs12075 typed in the Duffy gene (FY, associated with protection against P. vivax malaria). The implication of these findings is that other genomic regions with similarly unusual patterns of variation are candidates for biologically interesting loci. Of course, some may have such extreme statistics purely by chance, and genotyping projects are likely to miss certain features (such as high or low genetic diversity and rare mutations) that are informative about other biologically interesting loci.
Another question we can ask is whether genes previously reported as showing evidence for the action of historical selection (because they do not conform to the expectations of statistical models that assume neutrality) are also unusual within the empirical, genome-wide distribution. Table 1 shows the value of two selection statistics (Tajima's D [28] and Fay and Wu's H [29]) that are commonly used to infer the action of historical selection from genetic variation for 19 genes computed from the HapMap data (in 30-SNP windows around the midpoint of each gene). Because of the ascertainment bias in the frequencies of SNPs chosen for genotyping, we do not expect either statistic to follow the standard neutral distribution. However, we can ask whether these genes fall within the tails of the empirical distribution (computed from regions at least 30 kb from known genes) or within the tails of the empirical distribution of regions matched for local recombination rate (the variance of selection statistics is influenced by recombination rate such that more extreme values are expected in regions of low recombination [30]).
Table 1 Selection Statistics in the HapMap Data for Genes Reported to Have Experienced Recent Adaptive Evolution
Of the 19 genes with previous evidence for historical selection, 12 show an unusual pattern of genetic variation in at least one population (defined as having a value lying in either the bottom 5% or top 5% of empirical values). Superficially, this result suggests that statistical tests based on rejecting a simple population genetics model are effective at detecting genes of interest. However, for 114 tests, we might expect 11 to lie in either the top or bottom 5% of observations, compared to the 17 observed. Another concern is that genes of known functional and selective importance, such as Duffy and CD40 ligand, do not fall in the tails of the empirical distribution of Tajima's D and Fay and Wu's H statistics and others, such as MMP3, hemochromatosis (HFE), and aldehyde dehydrogenase 2 (ALDH2) show patterns that are unusual, but not indicative of the action of recent selective sweeps.
There are two main conclusions from these analyses. First, that biologically interesting loci often do have unusual patterns of genetic variation, but that there is no single way of measuring “unusual” that is uniformly powerful for detecting the action of natural selection. Second, that rejection of neutral evolutionary models is no guarantee that the locus is unusual when compared to the rest of the genome. One of the great strengths of the HapMap data is that they will provide an alternative, empirical basis on which to assess how unusual the pattern of variation is at a given locus. However, it will still be many years before we know how reliable “looking in the tails” is as an approach to identifying genes of selective and functional importance.
Conclusions
Integrating our knowledge about gene function, genome structure, chromatin organisation, recombination rate, mutation processes, and evolutionary history to provide a coherent understanding of the structure of the human genome and human genetic variation is a task that is just starting. It is also a task that has been greatly aided by the HapMap project with its unprecedented view of SNP variation, and there is no doubt that researchers will be uncovering fascinating patterns in the data for years to come. As the subsequent phases of the project progress, we can also expect to gain an even more detailed view of the differences between our genomes and the evolutionary and biological forces that have made us.
We wish to thank the participants of the International HapMap Project and members of the Analysis Group in particular. Also, we thank Daniel Wilson and Niall Cardin for comment and discussions of the manuscript. The work of the Oxford Statistics Department within the HapMap project was funded by grants from National Institutes of Health and The SNP Consortium. RC is funded by the Fyssen Foundation.
Abbreviations
LDlinkage disequilibrium
SNPsingle nucleotide polymorphism
==== Refs
References
Hirschhorn JN Daly MJ 2005 Genome-wide association studies for common diseases and complex traits Nat Rev Genet 6 95 108 15716906
Taillon-Miller P Bauer-Sardina I Saccone NL Putzel J Laitinen T 2000 Juxtaposed regions of extensive and minimal linkage disequilibrium in human Xq25 and Xq28 Nat Genet 25 324 328 10888883
Abecasis GR Noguchi E Heinzmann A Traherne JA Bhattacharyya S 2001 Extent and distribution of linkage disequilibrium in three genomic regions Am J Hum Genet 68 191 197 11083947
Daly MJ Rioux JD Schaffner SF Hudson TJ Lander ES 2001 High-resolution haplotype structure in the human genome Nat Genet 29 229 232 11586305
Patil N Berno AJ Hinds DA Barrett WA Doshi JM 2001 Blocks of limited haplotype diversity revealed by high-resolution scanning of human chromosome 21 Science 294 1719 1723 11721056
Reich DE Cargill M Bolk S Ireland J Sabeti PC 2001 Linkage disequilibrium in the human genome Nature 411 199 204 11346797
Dawson E Abecasis GR Bumpstead S Chen Y Hunt S 2002 A first-generation linkage disequilibrium map of human chromosome 22 Nature 418 544 548 12110843
Gabriel SB Schaffner SF Nguyen H Moore JM Roy J 2002 The structure of haplotype blocks in the human genome Science 296 2225 2229 12029063
Phillips MS Lawrence R Sachidanandam R Morris AP Balding DJ 2003 Chromosome-wide distribution of haplotype blocks and the role of recombination hot spots Nat Genet 33 382 387 12590262
Jeffreys AJ Kauppi L Neumann R 2001 Intensely punctate meiotic recombination in the class II region of the major histocompatibility complex Nat Genet 29 217 222 11586303
Crawford DC Bhangale T Li N Hellenthal G Rieder MJ 2004 Evidence for substantial fine-scale variation in recombination rates across the human genome Nat Genet 36 700 706 15184900
McVean GA Myers SR Hunt S Deloukas P Bentley DR 2004 The fine-scale structure of recombination rate variation in the human genome Science 304 581 584 15105499
Wang WY Barratt BJ Clayton DG Todd JA 2005 Genome-wide association studies: Theoretical and practical concerns Nat Rev Genet 6 109 118 15716907
The International HapMap Consortium 2003 The International HapMap Project Nature 426 789 796 14685227
The International HapMap Consortium 2005 A haplotype map of the human genome Nature In press.
Lin S Chakravarti A Cutler DJ 2004 Exhaustive allelic transmission disequilibrium tests as a new approach to genome-wide association studies Nat Genet 36 1181 1188 15502828
Marchini J Donnelly P Cardon LR 2005 Genome-wide strategies for detecting multiple loci that influence complex diseases Nat Genet 37 413 417 15793588
Hunter DJ 2005 Gene-environment interactions in human diseases Nat Rev Genet 6 287 298 15803198
Paabo S 2003 The mosaic that is our genome Nature 421 409 412 12540910
Hill WG Robertson A 1968 Linkage disequilibrium in finite populations Theor Appl Genet 38 226 231 24442307
Pritchard JK Przeworski M 2001 Linkage disequilibrium in humans: Models and data Am J Hum Genet 69 1 14 11410837
Hedrick PW 1987 Gametic disequilibrium measures: Proceed with caution Genetics 117 331 341 3666445
Nielsen R 2001 Statistical tests of selective neutrality in the age of genomics Heredity 86 641 647 11595044
Tishkoff SA Verrelli BC 2003 Patterns of human genetic diversity: Implications for human evolutionary history and disease Annu Rev Genomics Hum Genet 4 293 340 14527305
Sabeti PC Reich DE Higgins JM Levine HZ Richter DJ 2002 Detecting recent positive selection in the human genome from haplotype structure Nature 419 832 837 12397357
Cooke GS Hill AV 2001 Genetics of susceptibility to human infectious disease Nat Rev Genet 2 967 977 11733749
Stefansson H Helgason A Thorleifsson G Steinthorsdottir V Masson G 2005 A common inversion under selection in Europeans Nat Genet 37 129 137 15654335
Tajima F 1989 Statistical method for testing the neutral mutation hypothesis by DNA polymorphism Genetics 123 585 595 2513255
Fay JC Wu CI 2000 Hitchhiking under positive Darwinian selection Genetics 155 1405 1413 10880498
Wall J 1999 Recombination and the power of statistical tests of neutrality Genet Res 74 65 79
Evans PD Anderson JR Vallender EJ Gilbert SL Malcom CM 2004 Adaptive evolution of ASPM, a major determinant of cerebral cortical size in humans Hum Mol Genet 13 489 494 14722158
Kouprina N Pavlicek A Mochida GH Solomon G Gersch W 2004 Accelerated evolution of the ASPM gene controlling brain size begins prior to human brain expansion PLoS Biol 2 e126. DOI: 10.1371/journal.pbio.0020126 15045028
Zhang J 2003 Evolution of the human ASPM gene, a major determinant of brain size Genetics 165 2063 2070 14704186
Hamblin MT Thompson EE Di Rienzo A 2002 Complex signatures of natural selection at the Duffy blood group locus Am J Hum Genet 70 369 383 11753822
Bersaglieri T Sabeti PC Patterson N Vanderploeg T Schaffner SF 2004 Genetic signatures of strong recent positive selection at the lactase gene Am J Hum Genet 74 1111 1120 15114531
Schliekelman P Garner C Slatkin M 2001 Natural selection and resistance to HIV Nature 411 545 546 11385558
Osier MV Pakstis AJ Soodyall H Comas D Goldman D 2002 A global perspective on genetic variation at the ADH genes reveals unusual patterns of linkage disequilibrium and diversity Am J Hum Genet 71 84 99 12050823
Toomajian C Ajioka RS Jorde LB Kushner JP Kreitman M 2003 A method for detecting recent selection in the human genome from allele age estimates Genetics 165 287 297 14504236
Enard W Przeworski M Fisher SE Lai CS Wiebe V 2002 Molecular evolution of FOXP2, a gene involved in speech and language Nature 418 869 872 12192408
Wooding S Kim UK Bamshad MJ Larsen J Jorde LB 2004 Natural selection and molecular evolution in PTC, a bitter-taste receptor gene Am J Hum Genet 74 637 646 14997422
Slatkin M Bertorelle G 2001 The use of intraallelic variability for testing neutrality and estimating population growth rate Genetics 158 865 874 11404347
Wang YQ Su B 2004 Molecular evolution of microcephalin, a gene determining human brain size Hum Mol Genet 13 1131 1137 15056608
Evans PD Anderson JR Vallender EJ Choi SS Lahn BT 2004 Reconstructing the evolutionary history of microcephalin, a gene controlling human brain size Hum Mol Genet 13 1139 1145 15056607
Stajich JE Hahn MW 2005 Disentangling the effects of demography and selection in human history Mol Biol Evol 22 63 73 15356276
Wang E Ding YC Flodman P Kidd JR Kidd KK 2004 The genetic architecture of selection at the human dopamine receptor D4 (DRD4) gene locus Am J Hum Genet 74 931 944 15077199
Rockman MV Hahn MW Soranzo N Loisel DA Goldstein DB 2004 Positive selection on MMP3 regulation has shaped heart disease risk Curr Biol 14 1531 1539 15341739
Pagnier J Mears JG Dunda-Belkhodja O Schaefer-Rego KE Beldjord C 1984 Evidence for the multicentric origin of the sickle cell hemoglobin gene in Africa Proc Natl Acad Sci U S A 81 1771 1773 6584911
Oota H Pakstis AJ Bonne-Tamir B Goldman D Grigorenko E 2004 The evolution and population genetics of the ALDH2 locus: Random genetic drift, selection, and low levels of recombination Ann Hum Genet 68 93 109 15008789
Harding RM Healy E Ray AJ Ellis NS Flanagan N 2000 Evidence for variable selective pressures at MC1R Am J Hum Genet 66 1351 1361 10733465
Huttley GA Easteal S Southey MC Tesoriero A Giles GG 2000 Adaptive evolution of the tumour suppressor BRCA1 in humans and chimpanzees. Australian Breast Cancer Family Study Nat Genet 25 410 413 10932184
Saunders MA Slatkin M Garner C Hammer MF Nachman MW 2005 The span of linkage disequilibrium caused by selection on G6PD in humans Genetics E-pub ahead of print.
|
16254603
|
PMC1270010
|
CC BY
|
2021-01-05 07:59:43
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e54
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010054
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460410.1371/journal.pgen.0010055plge-01-04-08Research ArticleGain-of-Function Screen for Genes That Affect Drosophila Muscle Pattern Formation Screen for Genes Affecting Muscle PatternStaudt Nicole 1Molitor Andreas 12Somogyi Kalman 3Mata Juan 3Curado Silvia 3Eulenberg Karsten 2Meise Martin 2Siegmund Thomas 2Häder Thomas 2Hilfiker Andres 2Brönner Günter 2Ephrussi Anne 3Rørth Pernille 3Cohen Stephen M 3Fellert Sonja 1Chung Ho-Ryun 1Piepenburg Olaf 1Schäfer Ulrich 1Jäckle Herbert 1Vorbrüggen Gerd 1*1 Max Planck Institut für biophysikalische Chemie, Göttingen, Germany
2 DeveloGen, Göttingen, Germany
3 Developmental Biology Unit, European Molecular Biology Laboratory, Heidelberg, Germany
Mackay Trudy EditorNorth Carolina State University, United States of America* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 1 4 e557 7 2005 29 9 2005 Copyright: © 2005 Staudt et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.This article reports the production of an EP-element insertion library with more than 3,700 unique target sites within the Drosophila melanogaster genome and its use to systematically identify genes that affect embryonic muscle pattern formation. We designed a UAS/GAL4 system to drive GAL4-responsive expression of the EP-targeted genes in developing apodeme cells to which migrating myotubes finally attach and in an intrasegmental pattern of cells that serve myotubes as a migration substrate on their way towards the apodemes. The results suggest that misexpression of more than 1.5% of the Drosophila genes can interfere with proper myotube guidance and/or muscle attachment. In addition to factors already known to participate in these processes, we identified a number of enzymes that participate in the synthesis or modification of protein carbohydrate side chains and in Ubiquitin modifications and/or the Ubiquitin-dependent degradation of proteins, suggesting that these processes are relevant for muscle pattern formation.
Synopsis
Muscle pattern formation during embryogenesis requires the activity of a distinct network of genes. In the model organism Drosophila, this process involves the determination of stem-cell-like muscle founder cells, their differentiation, and their attraction to tendon-like epidermal cells, termed apodemes, to which the muscles attach. In order to systematically identify genes involved in these processes, a collection of fruit fly strains was generated that can be used for the ectopic expression of more than 3,700 individual fruit fly genes in a spatiotemporally restricted manner. In order to address muscle pattern formation, the collection was used to express the genes in the developing apodemes and in a series of distinct epidermal cells that serve as migration substrate for developing muscles towards the apodemes. In addition to already known factors, some 60 novel gene activities were found to interfere under these circumstances with the formation of the muscle pattern. In addition to providing a most valuable tool for the Drosophila community of researchers, the results provide a framework for a detailed analysis of the gene network and insight into molecular mechanisms underlying embryonic muscle pattern formation.
Citation:Staudt N, Molitor A, Somogyi K, Mata J, Curado S, et al. (2005) Gain-of-function screen for genes that affect Drosophila muscle pattern formation. PLoS Genet 1(4): e55.
==== Body
Introduction
Whole genome sequences of many animals are now known, including those of Caenorhabditis elegans, human, mouse, and Drosophila melanogaster (see for example [1–4]). The task now facing biologists is to discover the functions of the annotated genes within the genomes. For some organisms, such as C. elegans, it is possible to adopt a systematic approach to ablate gene function by, for example, the RNA interference technique [5,6]. For Drosophila melanogaster, a widespread analysis of gene function has been undertaken by systematic EMS mutagenesis and transposon tagging approaches using the P-element [7,8]. However, since two-thirds of genes in D. melanogaster and in C. elegans cause only subtle or even unscorable mutant phenotypes [9], a complementary approach was used. This approach is based on conditional overexpression of genes in order to generate gain-of-function phenotypes. It involves upstream activating sequences (UASs) of yeast provided through a recombinant transposon insertion, termed EP-element [10]. The inserted UAS can be used to transcriptionally activate an endogenous gene next to the insertion site by the transgene-dependent expression of the yeast transcription factor GAL4 under the control of a constitutively active promoter or spatiotemporally regulated enhancer elements [10,11].
Here we describe a newly generated EP-element library composed of more than 3,700 unique insertion sites and their location within the D. melanogaster genome. We employed this and previously constructed EP-element libraries for a systematic gain-of-function screen to identify gene activities that interfere with the proper development of the segmentally repeated muscle pattern. We designed a GAL4-driver to express endogenous genes in single epidermal cell rows, one anterior and one posterior to the embryonic engrailed expression domain [12], asking whether misexpression of genes in these locations alters the identity and/or the spatial cues of cells and thereby interferes with the genetically controlled migration and pathfinding ability of myotubes, as well as their anchoring properties (reviewed in [13]).
Muscle pattern formation is a stereotyped and segmentally repeated developmental process. Once muscle founder cells are born and determined, they grow by fusion with undetermined muscle cells. The resulting myotubes extend via the growth of cone-like tips along the inner surface of epidermal cells (reviewed in [13]), which serve as a migration substrate towards a distinct set of tendon-like segment border cells, termed apodemes, to which myotubes finally attach [13,14]. stripe, which encodes a zinc-finger-type transcription factor, is essential for apodeme cell formation at the segment borders [15]. In stripe mutants, myotubes soon fail to be properly guided, indicating that developing apodeme cells not only serve as attachment sites but also provide guiding cues for the migrating myotubes [15,16]. In addition, myotube guidance is also controlled by the myotube-expressed gene grip [17], by Slit/Robo signaling [18,19], and the attachment to apodemes involves the atypical receptor tyrosine kinase Derailed [20] as well as fibroblast growth factor signaling activity [21]. These and other results [13] have established that myotube guidance and attachment are controlled by interactions between epidermal and muscle cells and that interfering with their interactions causes scorable effects on the stereotyped muscle pattern. Here, we describe a systematic gain-of-function screen towards identifying gene activities that can interfere with the formation of the proper muscle pattern in the D. melanogaster embryo by using the specially designed UAS/GAL4 misexpression system.
Results/Discussion
Generation of an EP-Element Insertion Collection
We generated a novel EP-element collection for D. melanogaster. It contains single insertions bearing GAL4-dependent UASs of yeast at their ends [22,23]. Genes properly oriented with respect to the UAS sequences can therefore be conditionally expressed via transgene-derived GAL4 activity [10].
Of more than 13,800 individual EP-element lines initially generated, the insertion sites of more than 11,700 individual lines were determined by a combined PCR/sequencing approach [22]. Among these insertion sites, we identified a total of 3,707 unique EP-element insertion sites within the D. melanogaster genome. Their location and the orientation of the EP-elements are summarized in Tables S1 and S2. Using this collection, roughly a quarter of the D. melanogaster genes [2] can be activated by transgene-derived GAL4 activity that is driven by constitutively active promoters or cell-specific enhancer elements [11]. In addition, a sizable portion of the EP-elements are in reverse orientation with respect to genes (that is, there are no other annotated transcription units within a range of 10 kb of genomic DNA [2]), suggesting that the activity of these genes would likely be knocked down in response to transgene-derived GAL4 activation (see Tables S1 and S2).
Generation of a GAL4-Driver Line Causing Epidermal Stripe Expression
In order to perform a large-scale gain-of-function screen for gene activities that interfere with D. melanogaster muscle pattern formation during embryogenesis, we designed a GAL4-driver that allowed the misexpression of EP-targeted genes in presumptive apodeme precursors at the segment border and in an ectopic array of intrasegmental cell rows within the epidermis of the embryo. We made use of the sr239 enhancer element of the stripe gene, which drives gene expression in a single cell row posterior to the engrailed expression domain. These cells correspond to dorsal and lateral apodeme precursor cells at a midstage of embryogenesis (stage 12) [12]. We fused this element with the GAL4 coding region to express UAS-targeted genes in a subset of apodeme cells. In addition, we used the sr239Δpan enhancer [12], termed srmod, to drive GAL4 expression in apodeme cells and in a subset of epidermal cells anterior to the engrailed expression domain (Figure 1A and 1B) that serve as a migration substrate for myotubes. We expected these tools to facilitate the identification of genes whose activities interfere with processes such as myotube guidance or muscle attachment when expressed in response to one or both of the GAL4-drivers.
Figure 1 Expression Pattern of the srmodGAL4-Driver and Induced Muscle Pattern Defects in Response to EP Targeted Endogenous D. melanogaster Genes
(A) Expression pattern of srmodGAL4 driving a UAS-lacZ transgene in a stage 15 embryo (lateral view). Note the expression of β-Galactosidase in a segmentally repeated pattern of segment border cells, which appear as stripes, and in an array of partially interrupted intersegmental cell rows between them.
(B) Enlarged lateral area of a stage 16 embryo stained with anti-MHC antibodies (green) to visualize the muscle pattern and anti-β-Galactosidase antibodies (red) to visualize the apodemes and intrasegmental epidermal cells that express the marker gene in response to srmodGAL4 activity (red).
(C and D) Muscle pattern (anti-MHC staining) of wild-type stage 16 embryos (C) and enlarged lateral area (D) outlined in (C).
(E–J) Examples of muscle phenotypes (enlargements as in [D]) in response to srmodGAL4-driven misexpression of EP-targeted genes. Note that expression of CG9742 (HD10913) induces abnormal attachment of LT4 and LT5 muscles ([E], arrowheads), adk1 (HD32155) affects the shape and attachment of all LT muscles ([F], arrowheads), parg (HD10914) causes an abnormal shape and attachment of all lateral muscles (G), and CG32436 (HD35012) impairs the fusion of myoblasts (see unfused MHC-stained myoblasts) ([H], arrowheads), whereas CG4963 (HD35059) (I) and CG31710 (EP2160) (J) affect determination, growth, and attachment of many muscles at the same time.
Screening for Genes That Interfere with Muscle Pattern Formation
To activate misexpression of endogenous EP-targeted genes, we crossed females bearing the srmodGAL4-driver with male individuals from about 4,500 lines of the new (Table S1) and a previous EP-element collection [10] and asked whether the misexpression causes a lethal phenotype, knowing that impairing the stereotyped muscle pattern prevents the hatching of fully differentiated larvae from the egg shell [15,16]. In the next step, we examined whether muscle pattern defects can be observed after staining of the fully differentiated but unhatched embryos with anti–Myosin heavy chain (MHC) antibodies. To distinguish between interfering gene activities that were induced in the segment border apodeme cells and those that were derived from the intervening epidermal cells, we performed corresponding crosses using srGAL4-bearing instead of srmodGAL4-bearing females.
We identified an initial set of 78 EP-element lines (1.7%) that caused a specific srmodGAL4-dependent muscle pattern phenotypes. To confirm that the observed phenotypes in the embryonic muscle pattern originated from misexpression of a given gene, we tested whether (i) the phenotype could be reverted by the precise excision of the EP-element, (ii) the potential target gene was expressed in a GAL4-dependent fashion (this was tested using in situ hybridization or antibody staining for product detection), (iii) the muscle pattern defects also occurred by over-expression of corresponding cDNA from UAS-dependent transgenes, or (iv) whether misexpression of the same transcription unit by a different EP-element insertion caused a similar phenotype.
The strength and penetrance of the misexpression muscle pattern phenotypes were variable (compare Figure 1C and 1D with 1E–1J). We found embryos in which only single muscle fibers were abnormally attached to apodemes (Figure 1E and 1F), embryos in which most muscles of the dorsal and lateral region of the embryo were abnormally shaped and attached to ectopic epidermal sites (Figure 1G), and embryos in which the early processes of myogenesis were aberrant, as concluded from impaired myoblast fusions (Figure 1H) and muscle misdetermination (Figure 1I and 1J). The different defects suggest that activities derived from the misexpressed genes can interfere with cell determination as well as guiding and targeting events during muscle pattern formation. In some cases, the defects observed were not restricted to dorsal and lateral muscles but also included muscles in the ventral region of the embryo, where few epidermal cells express srmod-dependent GAL4 activity (not shown).
Gene Activities That Interfere with Muscle Growth and Attachment
Of the initially identified 78 EP-element insertions, 66 GAL4-driven transcription units could be unambiguously identified to be the cause of the gain-of-function phenotypes (Table S3). Of those, ten transcription units were expressed in anti-sense orientation, implying that misexpression of transcripts in reverse orientation is likely to cause a knock-down phenotype. Analysis of the expression patterns of some of the anti-sense-tagged candidates indicated that the transcripts are expressed ubiquitously or accumulate at the segment border (see below). Thus, GAL4-driven misexpression may result in reduced gene activity. Fifty-six transcription units were in sense orientation, suggesting misexpression phenotypes in response to GAL4-drivers. Computer assisted analysis of the products of the targeted transcription units revealed that many of the candidates with known or predicted functions encode for membrane-associated or -secreted factors as well as for components known to be involved in protein modification and degradation (Figure 2).
Figure 2 Schematic Representation of the Classification of the 66 Identified Candidate Genes into Functional Groups
The affiliation of the genes products is indicated by the color and the size of the fragments represents the quantitative distribution. cytsk., cytoskeleton; nucl. acid bdg., nucleic acid binding; prot. mod. + degrade., protein modification or degradation; secr. + membrane assoc., secreted or membrane-associated factors; transp. + carrier, transporter or carrier; unknown fct., unknown function.
Genes coding for membrane associated and secreted factors.
Thirteen genes encode proteins that contain diagnostic domains for membrane association or secretion. This group includes Tetraspanin Tsp42Ee (CG10106), one protein with three transmembrane domains (CG9030), and five factors with a single transmembrane domain that is typical for receptor-type proteins. This last group includes Toll, a receptor that participates in dorsoventral patterning of the embryo and innate immune response, and Syndecan (Sdc), a heparan sulfate proteoglycan (HSPG) that participates in Slit/Robo signaling [24,25]. Furthermore, membrane-associated factors were identified including CG33207/Pxb, which functions as an attenuator for hedgehog signaling [26], and the polychaetoid protein, a guanylate kinase at the adherens junctions that participates in JNK signaling [27].
The identification of a subset of transmembrane proteins in which two out of four proteins (Sdc and Toll) are already known to participate in muscle pattern formation [25,28] provides trust that other identified genes that code for membrane associated and secreted factors with unknown functions may also take part in the process. These uncharacterized factors include CG14052, CG6301, and CG17368, which encode small proteins containing an N-terminal signal peptide, implying that they represent secreted factors for which functions need to be established.
Protein modification and degradation.
A total of eight genes encode for factors involved in protein modification and degradation. Three genes encode components of the Ubiquitin pathway including uba1 and effete, which encode E1 and E2 enzymes, respectively, as well as CG11033, which codes for an uncharacterized F-Box protein. F-Box proteins are required for target protein binding and for Ubiquitin transfer by the E2/E3 complex. Both effete and uba1 have been shown to participate in neurogenesis [29,30]. uba1 was initially found in a gain-of-function screen for genes involved in motor axon guidance [31].
Of the subset of proteins modifying enzymes, five play a role in modifying carbohydrate side chains of peptides. Of those, sulfateless encodes a heparan sulfate–glucosamine-N-sulfotransferase required for Decapentaplegic, Hedgehog, and fibroblast growth factor signaling [32]. The finding of a HSPG-modifying enzyme is consistent with the concurrent identification of the HSPG Sdc (see above), already known to affect muscle guidance [25,33]. We identified also a second sulfotransferase (CG32629/CG32632 fusion) and two genes that code for proteins that modify extracellular carbohydrates (CG31973 and gnbp3). The finding of several enzymes involved in carbohydrate side chain synthesis and modification suggests that they play a role not only in axon guidance [34] but also in muscle guidance and/or apodeme targeting. In addition, the identification of several Ubiquitin pathway components implicated in protein degradation suggests a role also of this process in muscle pattern formation.
Transcription factors and RNA binding proteins.
Only six potential or known DNA or RNA binding factors were identified. This result suggests that only a comparatively small number of transcription factors can interfere with the functional development of apodeme cells in a manner recently shown for the zinc finger protein encoded by stripe [15,16]. Interestingly, the identified transcription factors also include two zinc finger proteins, encoded by schnurri and escargot (esg), that have been shown to act in the formation of the tracheal system [35–37]. Esg is involved in Cadherin-mediated adhesion [37]. Thus, its misexpression may cause abnormal adhesion of muscles when esg-expressing epidermal cells are provided as a substrate. schnurri activity is required to properly mediate TGFβ signaling [35]. Its ectopic expression may therefore cause an improper signaling readout that impairs myotube outgrowth and/or muscle attachment.
Cytoskeleton factors.
Three genes code for known cytoskeleton binding proteins such as Katanin80, a WD40 domain microtubule binding protein of the Katanin complex involved in micotubule severing. In addition, we found chickadee, which was identified twice by independent EP-element insertions in this screen. chickadee protein is involved in Actin filament organization and contains a phosphatidylinositol-4,5-bisphophate binding motif. This motif is noteworthy with respect to rdgBβ, which codes for a phosphatidylinositol transfer protein coupling phosphatidylinositol delivery and phosphatidylinositol-4,5-bisphophate synthesis relevant for cell–cell signaling processes (reviewed in [38]) and which was also identified in the screen.
Factors involved in cell cycle control and biosynthesis.
Seven factors involved in central steps of biosynthesis were identified. They include the ribosomal protein RpL18A and the polyadenylation binding protein Pabp2. In addition, cell cycle control genes such as the D. melanogaster CDC25 homolog twine and two cycline genes were found. Interference of overexpressed general biosynthesis factors and cell cycle control genes can be explained if they would alter proper epidermal cell differentiation, patterns of cell death, and/or patterns of cell divisions. In these cases, gene expression could impair processes required to maintain or generate properly differentiated epidermal cells that serve as substrate for the outgrowing myotubes or provide spatial cues relevant for this process.
Gene Expression Patterns
We examined the expression pattern of a total of 46 of the identified genes. This criterion for validation of potential gene functions for muscle guidance and attachment control included whole mount in situ hybridization using anti-sense RNA probes prepared from respective cDNAs or genomic fragments covering parts of the identified candidates as well as information available from a D. melanogaster database [39].
The majority of genes are expressed in patterns that could not be directly correlated with muscle pattern formation. However, most were either ubiquitously expressed or they were maternally contributed, and transcripts are present in eggs and during early embryogenesis. Yet, about one-third of the genes were expressed in spatiotemporal patterns in the epidermis during the stage when myotube migration takes place. Eight of these genes were expressed in the apodeme precursor cells of wild-type embryos, including seven of the 12 genes that encode cell surface proteins or secreted factors. Examples of the gene expression pattern are shown Figure 3.
Figure 3 Expression Patterns of Genes that Cause a Gain-of-Function Muscle Phenotype
Lateral views of embryos at stage 11 (M), stage 13 (A, C, E, G, I, K, and N), and stage 16 (B, D, F, H, J, and L) that were stained with transcript-specific anti-sense RNA probes or with anti-Toll antibodies (A and B). Note the expression of Toll (A and B) in segment border cells, sdc (LD08230) (C and D) in trachea, segment border cells, and the differentiated apodemes, CG3563 (LD15689) (E and F) in the apodeme precusor cells at the segment border, CG13913 (RE53394) (G and H) and CG5008/gnbp3 (SD21560) (I and J) in a subset of apodeme precursors and cells of the epidermis, CG14713/14714 transcripts (AT17253) (K and L) in the dorsal and ventral epidermis around the segment border, and pxb (SD26190) (M and N) in intrasegmental epidermal stripes.
The fraction of genes that are expressed in apodeme cells at the stage when they are targeted by the muscles includes Toll, sdc, CG3563, CG13913, and gnbp3 (Figure 3A–3J). The expression patterns of Toll and Sdc have previously been described. Toll is expressed in a subset of the developing apodeme cells and participates in muscle pattern formation [28]. Sdc is expressed in the mesoderm, the tracheal system, the axons of the central nervous system and in the differentiated apodemes. In sdc mutants, muscles fail to respect the ventral midline as a migration border, cross the border, and subsequently attach to apodemes at the other side of the midline [25]. Other genes, such as CG14713 and pxb/AT17253, are expressed in the intrasegmental region of the epidermis that is crossed by the migrating myotubes (Figure 3K–3N).
It is noteworthy that the set of identified genes also includes genes that are normally expressed in cells of other tissues or organs whose development involves migratory processes of cells or groups of cells. These include the developing tracheal system, germ line precursor cells, the midgut, and the nervous system. Thus, although the expression patterns exclude a role for these genes during the normal process of muscle pattern formation, they could play a direct or indirect role in guiding migrating cells in regions of the wild-type embryo where they are normally expressed. Preliminary results with a gene specifically expressed in germ line precursor cells supports this proposal (G.V., unpublished data).
Gene Activities Required for Muscle Pattern Formation
Muscle pattern phenotypes of Toll, gut feeling (oda), and sulfateless mutant embryos have already been described [28,32,40]. In order to test whether other genes that were identified in the gain-of-function screen also caused a loss-of-function phenotype, we examined the muscle pattern of loss-of-function mutants that were described in a context different from muscle pattern formation. Figure 4 shows two examples of the analysis, indicating that esg (compare Figure 4A with 4B) and sdc (compare Figure 4A and 4D with 4C, 4E, and 4F) loss-of-function mutant embryos develop variable muscle pattern defects that include the absence of lateral transverse muscles, loss of muscle fibers, and abnormally shaped muscles. Since sdc and esg are expressed in the epidermis, we used expression of delilah, a marker for the muscle attachment sites [41], to examine whether an altered pattern of attachment sites is a likely cause of the muscle pattern defects. No pattern defects were observed in sdc and esg mutants (Figure 4G–4I). Thus, the muscle pattern defects observed with both the gain-of-function and loss-of-function mutants are consistent with the argument that the gene esg participates in the regulation of adhesion processes, as previously proposed for esg function during tracheal system development [37], and that sdc is required for early Slit/Robo-signaling-dependent muscle guidance, as described recently (Figure 4E; [25]). Our results also show that sdc-dependent Slit signaling serves as a muscle attractant during a late phase of muscle guidance [19], since abnormal muscle elongations are observed in fully developed but unhatched sdc mutant larvae (Figure 4C and 4F).
Figure 4 Muscle Pattern Defects in esg and sdc Mutants
Muscle pattern of three segments of oreR (A, D, and G), esg L2 (B and H), and sdc 23 mutant embryos (C, E, F, and I) after staining with anti-MHC antibodies or using a delilah transcript-specific anti-sense RNA probe (G–H). Lateral (A–C and G–I) and ventral views (D–F) of embryos at stage 14 (E) and stage 16 (A–D and F–I). esg mutant embryos show variable muscle pattern defects with muscles absent ([B], arrowheads). In sdc mutant embryos few muscles cross the ventral midline in a position dorsal to the central nervous system ([E], arrowheads), and they show disruptions of the pattern in the ventral region ([C], arrowheads). The typical “finger-type pattern” of the ventral muscles of wild-type embryos (D) is unordered in sdc 23 mutant embryos, with ventral muscles aligning in parallel with the anterio-posterior axis, ignoring the segment border attachment ([F], arrowheads). Also shown is the pattern of epidermal muscle attachment sites (delilah marker gene expression) in wild-type (G), esg (H), and sdc (I) mutant embryos. Note that the pattern is unchanged in the mutants.
Conclusion
We identified a series of genes whose activity impairs muscle pattern formation when misexpressed in a defined pattern of epidermal cells that represent the migration substrate and/or the attachment sites for the outgrowing myotubes and muscle fibers. The 66 identified candidate genes were selected from an EP-element insertion library composed of more than 4,500 individual lines. This number suggests that about 1.5% of D. melanogaster genes can affect muscle pattern formation when expressed in cells that are contacted by myotubes or muscles. Although the screening system can certainly be regarded as artificial, it nevertheless identified genes such as Toll, gut feeling, and sulfateless that have been previously implicated in muscle pattern formation because the corresponding loss-of-function mutations cause variable muscle phenotypes [28,32,40]. In addition, it identified genes whose products are known to participate in cell migration and/or cell targeting processes in the embryo. These genes include esg and sdc [25,37], and, as shown in Figure 4, loss-of-function mutations in these two genes cause a defective muscle pattern in the embryo, indicating that the activity of these genes is essential for embryonic muscle development.
It is interesting to note that the misexpression screen identified, in addition to the HSPG Sdc, a number of other muscle-pattern-disturbing genes that code for factors known to participate in carbohydrate side chain synthesis or side chain modification. Thus, these enzymes are likely to participate in communication events between muscle and epidermal cells, processes that may also involve signaling molecules in addition to Slit [18,19]. Similarly, the independent identification of three components of the Ubiquitin system suggests that Ubiqitin modifications of proteins or their stability play a role in muscle pattern formation. The plethora of factors identified here open a new avenue towards a detailed functional analysis of processes underlying the interplay of myotubes, their epidermal migration substrate, and the specialized segmental border cells to which myofibers ultimately attach. In addition, they can be used towards developing an understanding of migratory processes in other developmental processes of D. melanogaster and, in view of the conservation of the genes identified here, possibly also in other species including mammals.
Materials and Methods
Genetics and expression detection.
To generate novel EP-element integration lines we used two different EP-elements. EPg was modified to function in the female germline and contains the white+ gene as a selectable marker in white mutant individuals [23]. The second EP-element, P{Mae-UAS.6.11}, contains the yellow gene as a corresponding marker [42]. More than 8,500 independent EPg insertions were generated using a jump-starter line from an EPg insert on a CyO chromosome (EPg4–38), and 5,100 independent lines were established using P{Mae-UAS.6.11}. Chromosomes bearing an EP-element integration were kept either as homozygous lines or in trans to a corresponding balancer chromosome. EP-element-bearing males were crossed with srmodGAL4-bearing females, and their F1 offspring were screened for lethality. In case of lethality, candidates were crossed with srmodGAL4- and srGAL4-drivers and their F1 offspring were examined after staining with anti-MHC antiserum (kindly provided by D. Kiehart) using the staining protocol previously described [12]. Anti-sense DIG-labeled RNA probes were prepared and whole mount in situ hybridization was performed as described [12].
Molecular analysis.
The EPg and the pMae elements have been previously described [23,42]. The srGAL4 and the srmodGAL4 lines were obtained by cloning the KpnI and XbaI of the sr239 and sr239Δpan DNA as described in [12] into the p221 vector (kindly provided by C. Klämbt). In order to determine the EP-element integration sites within genomic DNA, we performed inverse PCR as described on the Berkeley Drosophila Genome Project Web page (http://www.fruitfly.org/) with overnight digestion by either MaeI or Csp6I. Fragments were amplified for the 5′-end of P{EP,y+}; the primers used for the pMae were 5′-CAGCTGCGCTTGTTTATTTGC-3′ (forward) and 5′-TGGGAATTCGTTAACAGATCCAC-3′ (reverse), and for the EPg were Pw new up (CAG CCG AAT TAA TTC TAG TTC CAG TGA A) and Pw new low (ACT TCG GCA CGT GAA TTA ATT TTA CTC C). The amplified DNA was sequenced and used to determine the insertion site (see Table S1).
Supporting Information
Table S1 Description of the Tested Insertion Lines
The lines are ordered according to their names (Line). DG-EP denotes lines that carry the P{Mae-UAS.6.11}-element whereas HD-EP stands for lines that are generated by mobilization of the modified EPg-element. The chromosome arm (Arm), orientation (Strand), and position according to D. melanogaster genomic sequence release 3 (Position) are indicated. The 5′ sequence tag (forward strand) for each insertion line is listed (Sequence).
(607 KB XLS)
Click here for additional data file.
Table S2 Insertion Lines Available from the Bloomington Stock Center
The lines are ordered according to the name under which they will be kept in the Bloomington Stock Center (Line). The former name as used in Table 1 (Old Name) is also listed. The chromosome arm (Arm) as well as the position according to D. melanogaster genomic sequence release 4.1 (Coordinate) is shown. The next gene (Gene) with its extension (5′ Gene and 3′ Gene, respectively) and orientation (Strand) as well as the relative position of the insert to the gene (Position) is indicated. The 5′ sequence tag for each insertion line is listed under Sequence.
(152 KB XLS)
Click here for additional data file.
Table S3 Identified Candidate Genes
The candidate genes are arranged into groups by their proposed biological function. For each candidate, the CG number (according to the FlyBase [http://flybase.bio.indiana.edu/]), the gene synonym, the EP number, the orientation of the expressed transcript, predicted protein domains, the biological process, the expression pattern, the criteria for the validation of candidate genes, and a description of the gain-of-function muscle phenotype are listed. The wild-type expression patterns are based on the Berkeley Drosophila Genome Project in situ expression data [39], Berkeley Drosophila Genome Project CHIP-expression data, or in situ hybridization using either genomic fragments (gen frag) or Ests. In this case the name of the Est used is listed. Abbreviations used to describe the expression are as follows: Ap, muscle attachment sites; Br, brain; Ep, epidermis; Fb, fatbody; Gc, garland cells; Go, gonads; He, heart; Hg, hindgut; mat, maternal expression; Md, early mesoderm; Mg, midgut; Ml, ventral midline; Mu, muscles; Nb, neuroblasts; Pc, pericardial cells; Sb, epidermal segment border; Sg, salivary glands; Tp, tracheal placodes; Ts, tracheal system; zyg, zygotic expression. The criteria used to validate the identified candidate genes were (1) reversion of the EP-element, (2) induction of expression, (3) similar phenotype induced by an UAS cDNA transgene, (4) additional EP-element, and (5) published data. Anti-sense candidates were only considered in the absence of a gene in sense orientation within 10 kbp downstream of the EP-element (6).
(180 KB DOC)
Click here for additional data file.
We thank our colleagues in the lab for their various contributions and critical discussions, G. Dowe for sequencing, and U. Jahns-Meyer for transformations. Furthermore, we are grateful to Reinhard Schuh and Christian Klämbt for fly stocks and for support. This research was supported by the Max-Planck-Society, the SFB271, and the Graduiertenkolleg “Molekulare Genetik der Entwicklung” of the Deutsche Forschungsgemeinschaft (GV and HJ).
Competing interests. The authors have declared that no competing interests exist.
Author contributions. KE, MM, AH, AE, PR, SMC, HJ, and GV conceived and designed the experiments. NS, AM, KS, JM, SC, AH, GB, AE, PR, SF, OP, US, and GV performed the experiments. TS, TH, HRC, US, and GV analyzed the data. GV wrote the paper.
Abbreviations
esg
escargot
HSPGheparan sulfate proteoglycan
MHCMyosin heavy chain
SdcSyndecan
UASupstream activating sequence
==== Refs
References
C. elegans Sequencing Consortium (1998) Genome sequence of the nematode C. elegans: A platform for investigating biology. Science 282: 2012–2018.
Adams MD Celniker SE Holt RA Evans CA Gocayne JD 2000 The genome sequence of Drosophila melanogaster
Science 287 2185 2195 10731132
Lander ES Linton LM Birren B Nusbaum C Zody MC 2001 Initial sequencing and analysis of the human genome Nature 409 860 921 11237011
Venter JC Adams MD Myers EW Li PW Mural RJ 2001 The sequence of the human genome Science 291 1304 1351 11181995
Fraser AG Kamath RS Zipperlen P Martinez-Campos M Sohrmann M 2000 Functional genomic analysis of C. elegans chromosome I by systematic RNA interference Nature 408 325 330 11099033
Kamath RS Fraser AG Dong Y Poulin G Durbin R 2003 Systematic functional analysis of the Caenorhabditis elegans genome using RNAi Nature 421 231 237 12529635
Bellen HJ Levis RW Liao G He Y Carlson JW 2004 The BDGP gene disruption project: Single transposon insertions associated with 40% of Drosophila genes Genetics 167 761 781 15238527
Thibault ST Singer MA Miyazaki WY Milash B Dompe NA 2004 A complementary transposon tool kit for Drosophila melanogaster using P and piggyBac Nat Genet 36 283 287 14981521
Dow JA 2003 The Drosophila phenotype gap—And how to close it Brief Funct Genomic Proteomic 2 121 127 15239933
Rørth P 1996 A modular misexpression screen in Drosophila detecting tissue-specific phenotypes Proc Natl Acad Sci U S A 93 12418 12422 8901596
Brand AH Perrimon N 1993 Targeted gene expression as a means of altering cell fates and generating dominant phenotypes Development 118 401 415 8223268
Piepenburg O Vorbrüggen G Jäckle H 2000
Drosophila segment borders result from unilateral repression of hedgehog activity by wingless signaling Mol Cell 6 203 209 10949042
Schnorrer F Dickson BJ 2004 Muscle building; mechanisms of myotube guidance and attachment site selection Dev Cell 7 9 20 15239950
Volk T 1999 Singling out Drosophila tendon cells: A dialogue between two distinct cell types Trends Genet 15 448 453 10529807
Frommer G Vorbrüggen G Pasca G Jäckle H Volk T 1996 Epidermal egr-like zinc finger protein of Drosophila participates in myotube guidance EMBO J 15 1642 1649 8612588
Vorbrüggen G Jäckle H 1997 Epidermal muscle attachment site-specific target gene expression and interference with myotube guidance in response to ectopic stripe expression in the developing Drosophila epidermis Proc Natl Acad Sci U S A 94 8606 8611 9238024
Swan LE Wichmann C Prange U Schmid A Schmidt M 2004 A glutamate receptor-interacting protein homolog organizes muscle guidance in Drosophila
Genes Dev 18 223 237 14729572
Kidd T Bland KS Goodman CS 1999 Slit is the midline repellent for the robo receptor in Drosophila
Cell 96 785 794 10102267
Kramer SG Kidd T Simpson JH Goodman CS 2001 Switching repulsion to attraction: Changing responses to slit during transition in mesoderm migration Science 292 737 740 11326102
Callahan CA Bonkovsky JL Scully AL Thomas JB 1996 derailed is required for muscle attachment site selection in Drosophila
Development 122 2761 2767 8787750
Stathopoulos A Tam B Ronshaugen M Frasch M Levine M 2004 pyramus and thisbe: FGF genes that pattern the mesoderm of Drosophila embryos Genes Dev 18 687 699 15075295
Beinert N Werner M Dowe G Chung HR Jäckle H 2004 Systematic gene targeting on the X chromosome of Drosophila melanogaster
Chromosoma 113 271 275 15480728
Mata J Curado S Ephrussi A Rørth P 2000 Tribbles coordinates mitosis and morphogenesis in Drosophila by regulating string/CDC25 proteolysis Cell 101 511 522 10850493
Johnson KG Ghose A Epstein E Lincecum J O'Connor MB 2004 Axonal heparan sulfate proteoglycans regulate the distribution and efficiency of the repellent slit during midline axon guidance Curr Biol 14 499 504 15043815
Steigemann P Molitor A Fellert S Jäckle H Vorbrüggen G 2004 Heparan sulfate proteoglycan syndecan promotes axonal and myotube guidance by slit/robo signaling Curr Biol 14 225 230 14761655
Inaki M Kojima T Ueda R Saigo K 2002 Requirements of high levels of Hedgehog signaling activity for medial-region cell fate determination in Drosophila legs: Identification of pxb, a putative Hedgehog signaling attenuator gene repressed along the anterior-posterior compartment boundary Mech Dev 116 3 18 12128201
Takahashi K Matsuo T Katsube T Ueda R Yamamoto D 1998 Direct binding between two PDZ domain proteins Canoe and ZO-1 and their roles in regulation of the jun N-terminal kinase pathway in Drosophila morphogenesis Mech Dev 78 97 111 9858699
Halfon MS Hashimoto C Keshishian H 1995 The Drosophila toll gene functions zygotically and is necessary for proper motoneuron and muscle development Dev Biol 169 151 167 7750635
Norga KK Gurganus MC Dilda CL Yamamoto A Lyman RF 2003 Quantitative analysis of bristle number in Drosophila mutants identifies genes involved in neural development Curr Biol 13 1388 1396 12932322
Watts RJ Hoopfer ED Luo L 2003 Axon pruning during Drosophila metamorphosis: Evidence for local degeneration and requirement of the ubiquitin-proteasome system Neuron 38 871 885 12818174
Kraut R Menon K Zinn K 2001 A gain-of-function screen for genes controlling motor axon guidance and synaptogenesis in Drosophila
Curr Biol 11 417 430 11301252
Lin X Buff EM Perrimon N Michelson AM 1999 Heparan sulfate proteoglycans are essential for FGF receptor signaling during Drosophila embryonic development Development 126 3715 3723 10433902
Spring J Paine-Saunders SE Hynes RO Bernfield M 1994
Drosophila syndecan: Conservation of a cell-surface heparan sulfate proteoglycan Proc Natl Acad Sci U S A 91 3334 3338 8159748
Holt CE Dickson BJ 2005 Sugar codes for axons? Neuron 46 169 172 15848796
Grieder NC Nellen D Burke R Basler K Affolter M 1995 Schnurri is required for Drosophila Dpp signaling and encodes a zinc finger protein similar to the mammalian transcription factor PRDII-BF1 Cell 81 791 800 7774018
Samakovlis C Manning G Steneberg P Hacohen N Cantera R 1996 Genetic control of epithelial tube fusion during Drosophila tracheal development Development 122 3531 3536 8951068
Tanaka-Matakatsu M Uemura T Oda H Takeichi M Hayashi S 1996 Cadherin-mediated cell adhesion and cell motility in Drosophila trachea regulated by the transcription factor Escargot Development 122 3697 3705 9012491
Allen-Baume V Segui B Cockcroft S 2002 Current thoughts on the phosphatidylinositol transfer protein family FEBS Lett 531 74 80 12401207
Tomancak P Beaton A Weiszmann R Kwan E Shu S 2002 Systematic determination of patterns of gene expression during Drosophila embryogenesis Genome Biol 3 RESEARCH0088 12537577
Salzberg A Golden K Bodmer R Bellen HJ 1996 gutfeeling, a Drosophila gene encoding an antizyme-like protein, is required for late differentiation of neurons and muscles Genetics 144 183 196 8878684
Armand P Knapp AC Hirsch AJ Wieschaus EF Cole MD 1994 A novel basic helix-loop-helix protein is expressed in muscle attachment sites of the Drosophila epidermis Mol Cell Biol 14 4145 4154 8196652
Crisp J Merriam J 1997 Efficiency of an F1 selection screen in a pilot two-component mutagenesis involving Drosophila melanogaster misexpression phenotypes Drosoph Inf Serv 80 90 92
|
16254604
|
PMC1270011
|
CC BY
|
2021-01-05 08:00:58
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e55
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010055
|
oa_comm
|
==== Front
PLoS GenetPLoS GenetpgenplgeplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, USA 1625460510.1371/journal.pgen.0010056plge-01-04-09Research ArticleDiscovery of Human Inversion Polymorphisms by Comparative Analysis of Human and Chimpanzee DNA Sequence Assemblies Discovery of Human Inversion VariantsFeuk Lars 12MacDonald Jeffrey R 1Tang Terence 1Carson Andrew R 12Li Martin 1Rao Girish 1Khaja Razi 1Scherer Stephen W 12*1 The Centre for Applied Genomics, Department of Genetics and Genomic Biology, The Hospital for Sick Children, Toronto, Ontario, Canada
2 Department of Molecular and Medical Genetics, University of Toronto, Ontario, Canada
Trask Barbara EditorFred Hutchinson Cancer Research Center, United States of America* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 29 9 2005 1 4 e561 8 2005 29 9 2005 Copyright: © 2005 Feuk et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.With a draft genome-sequence assembly for the chimpanzee available, it is now possible to perform genome-wide analyses to identify, at a submicroscopic level, structural rearrangements that have occurred between chimpanzees and humans. The goal of this study was to investigate chromosomal regions that are inverted between the chimpanzee and human genomes. Using the net alignments for the builds of the human and chimpanzee genome assemblies, we identified a total of 1,576 putative regions of inverted orientation, covering more than 154 mega-bases of DNA. The DNA segments are distributed throughout the genome and range from 23 base pairs to 62 mega-bases in length. For the 66 inversions more than 25 kilobases (kb) in length, 75% were flanked on one or both sides by (often unrelated) segmental duplications. Using PCR and fluorescence in situ hybridization we experimentally validated 23 of 27 (85%) semi-randomly chosen regions; the largest novel inversion confirmed was 4.3 mega-bases at human Chromosome 7p14. Gorilla was used as an out-group to assign ancestral status to the variants. All experimentally validated inversion regions were then assayed against a panel of human samples and three of the 23 (13%) regions were found to be polymorphic in the human genome. These polymorphic inversions include 730 kb (at 7p22), 13 kb (at 7q11), and 1 kb (at 16q24) fragments with a 5%, 30%, and 48% minor allele frequency, respectively. Our results suggest that inversions are an important source of variation in primate genome evolution. The finding of at least three novel inversion polymorphisms in humans indicates this type of structural variation may be a more common feature of our genome than previously realized.
Synopsis
Chimpanzee is the closest relative to humans having DNA sequences about 98% identical to each other. Small DNA sequence changes and probably more importantly, larger structural changes of chromosomes, led to the divergence of the two species some 6 million years ago. Until recently, there were ten structural differences visible under the microscope between chimpanzee and human, and nine of these were inversions of DNA. Through computational comparisons of genome sequences, the current study identifies another 1,576 putative inversion events. Thirty-three of these were larger than 100,000 base pairs in size and 29 intersect genes, prioritizing them for evolutionary studies. Twenty-three of the inversions have been confirmed experimentally with the largest being 4.3 million base pairs in size on human Chromosome 7. Surprisingly, three of the “inverted” regions were found to be variable in their orientation in the human population (in some cases the inversion was in the ancestral orientation found in chimpanzee). These observations indicate the human genome is still evolving in structure. Moreover, since such variable inversions have been shown to predispose to other (sometimes deleterious) changes in chromosomes, the new data delineate potential disease-associated genes.
Citation:Feuk L, MacDonald JR, Tang T, Carson AR, Li M, et al. (2005) Discovery of human inversion polymorphisms by comparative analysis of human and chimpanzee DNA sequence assemblies. PLoS Genet 1(4): e56.
==== Body
Introduction
Humans and chimpanzees diverged approximately 6 million years ago, making the chimpanzee the closest extant relative to modern humans. The characterization of sequence changes both at the nucleotide and the structural level is therefore important for the understanding of primate evolution, including human-specific traits. At the nucleotide level, the identity of the genomes has been estimated to be 98% to 99% [1–5], excluding insertions and deletions and other small rearrangements. The chimpanzee Chromosome 22 (PTR22), which is orthologous to human Chromosome 21(HSA21), was the first to be sequenced and the majority of the rest of the genome is represented as a draft assembly [2,6]. The exact nucleotide substitution rate for the alignment of these sequences is 1.23% (excluding insertions and deletions)[2]. Taking insertion and deletion events into account, the sequence identity has been estimated to be about 95% [7].
In addition to nucleotide level changes, large structural rearrangements have also occurred between the species and they are discernable through comparison of the G-banded karyotypes. The most obvious structural difference between the human and chimpanzee genomes is the fusion of two acrocentric chromosomes creating human Chromosome 2. This results in a lower total chromosome number in humans (22, XY versus 23, XY). In addition, there are nine visible pericentric inversions affecting Chromosomes 1, 4, 5, 9, 12, 15, 16, 17, and 18 [8]. Of these rearrangements, only the fusion creating Chromosome 2 and the inversions on Chromosome 1 and 18 are specific to the human lineage, while the remaining changes have occurred in the chimpanzee lineage.
Early comparative studies between the human and chimpanzee genomes focused mainly on localized sequencing efforts and characterization of karyotypically visible chromosomal rearrangements. More recently, a number of studies have been performed with the goal of characterizing loss and gain of submicroscopic regions of DNA using comparative genomic hybridization [9,10]. The results reveal that copy number differences are abundant between the human and chimpanzee genomes, which agree with studies of segmental duplications in the genomes of several species [11–13]. These latter studies show that there is a higher incidence of segmental duplications in the human genome than in the mouse or rat genomes, indicating that increases in sequence copy number are more common in recent primate evolution [14]. The high frequency of copy number differences between humans and chimpanzees are also consistent with the findings that these types of structural variants are present as a common type of polymorphism in the human genome [15–19].
Although recent technological advances allow for detection of most types of genomic variation, limitations in available methodology have prevented the genome-wide discovery of balanced rearrangements such as inversions. Nonetheless, the fact that nine known cytogenetically visible inversion events distinguish the human and chimpanzee genomes indicates that these may have been a common form of structural rearrangement during primate evolution. The comparative study of human Chromosome 21 and chimpanzee Chromosome 22 did not assess the extent of inversion events between the two species [6]. The recent publication of the chimpanzee genome and accompanying comparative analysis of structural rearrangements did not address inversion events beyond those that are visible in the karyotype [2,16].
Characterization of inversion events between humans and chimpanzees are important because inversions can affect the expression of genes adjacent to the breakpoints, or directly interrupt genes spanning the breakpoints. Large inversions have also been proposed to be a direct driving force in speciation [20] and have been shown to suppress recombination [21]. It is also important to investigate inversion events between humans and chimpanzees as the frequency of such events can provide an indication as to what extent inversion variants exist as polymorphisms in the human population. Since inversion polymorphisms are difficult to detect, there has not been, until very recently, an estimate of their occurrence in the human genome. By mapping fosmid ends to the reference genome sequence, Tuzun et al. identified 56 putative inversion breakpoints in a single individual (inversion breakpoint pairs cannot be identified unambiguously using this approach). This implies that inversion polymorphisms are much more common than previously assumed.
Using the draft sequence of Pan troglodytes (chimpanzee), we have used alignments between the human and chimpanzee genomes to identify regions of inverted orientation. Through this computational approach, there is no theoretical limit to the size resolution of inversion regions that can be identified and resolved. It is, therefore, possible to equally identify cytogenetically detectable, as well as nucleotide level inversions, with a resolution down to the breakpoint sequence itself. Of the 1,576 computationally predicted inversions, 23 have now been confirmed experimentally. Screening in human control individuals also revealed three regions to be polymorphic in the human population. Our data indicate that inversions have occurred frequently in recent primate evolution, and both computational analysis and experimental data support the observation that inversion polymorphisms may be common in the human genome.
Results
Computational Analysis for Identification of Putative Inversions
Net alignments of the human genome (assembly Build 35, hg_17) with the chimpanzee draft genome (assembly Build 1) were downloaded from the University of California, Santa Cruz (http://genome.ucsc.edu/). All alignments of inverted orientation more than 20 base pair (bp) in length were identified. Segmental duplication and repeats were identified as a source of non-syntenic alignments and all alignments with a repeat or duplication content of more than 90%, as well as regions where the net alignment and the reciprocal best-hits between human and chimpanzee sequences disagreed, were excluded. After filtering, 1,576 putative inversions were identified between the two genomes (Figure 1 and Table S1). In total, these regions cover more than 154 mega-bases (Mb) of DNA. In addition to the National Center for Biotechnology (NCBI) assembly Build 35 of the human genome, there is an independent assembly of human Chromosome 7 (CRA_TCAGchr7v2 from http://www.chr7.org) [22]. Using this assembly of Chromosome 7, another two regions of inverted orientation compared to the chimpanzee Chromosome 6 were identified.
Figure 1 Genome-Wide Distribution of the 1,576 Putative Inversions Identified between the Human and Chimpanzee Assemblies
Human chromosomes are shown to the left and the syntenic chimpanzee chromosome to the right. Each red line corresponds to an inversion between the human and chimpanzee assemblies. Regions larger than 100 kb are represented with multiple lines. These include the large inversions on human Chromosome 4, 5, 15, and 18, while those on human Chromosomes 9, 12, and 17 were not identified. The karyotypically visible pericentric inversions on Chromosome 1 and 16 have not been described at the molecular level.
The inverted regions identified are distributed amongst and throughout the human chromosomes (Figure 2A), with the highest content being on the X chromosome. Inversions range in size from 23 bp to 62 Mb, with the largest regions representing the karyotypically visible pericentric inversions (Figure 2B). In total, 33 inversions larger than 100 kb were identified. Seven of the nine known pericentric inversions have been characterized at the molecular level, and four of these seven regions were identified in our analysis (on human Chromosomes 4, 5, 15, and 18; Tables S3 and S4). The previously characterized inversions on Chromosomes 9, 10, and 19 [8] are not present in the correct orientation in the current draft of the chimpanzee assembly, presumably since these chromosomes still contain errors due to the draft nature of the sequence. A summary of all regions more than 25 kb and a comparison to previously published data is shown in Table S2.
Figure 2 Size Distribution and Chromosomal Distribution of Putative Inversions
(A) Size distribution of inversions. The size of the inversion regions identified range from 23 bp to 62 Mb, but more than half of all regions identified are less then 250 bp. The algorithm used to create the net alignments is more prone to make errors and assign random orientation to very short regions (Table S1 and Figure S1). However, we did not see this trend in the regions chosen for experimental validation. Thirty-three of the regions identified were larger than 100 kb in size.
(B) Chromosomal distribution of inversion regions. The autosomal chromosomes have a distribution of inversions roughly correlated to the size of the chromosome, except for Chromosome 19 which carries approximately the same number as Chromosomes 1 to 4. The X chromosome also shows an increase of inversions compared to autosomes of corresponding size.
Experimental Validation
Twenty-seven regions were selected for experimental validation (Table 1). These included five inversions larger than 500 kb, as well as the two regions that differ between the two human Chromosome 7 assemblies. The remaining regions were chosen to represent inversions of varying length. The initial phase of the project involved only Chromosome 7, and the selection of regions chosen for experimental validation is therefore biased towards this chromosome. The five largest inversions, 4.3 Mb at 7p22, 1.4 Mb at 2p25, 730 kb at 7p22, 680 kb at 19q13, and 670 kb at 7p12, were examined using three-color interphase fluorescence in situ hybridization (FISH) with bacterial artificial chromosomes (BACs) and fosmid probes. FISH experiments were performed using cell lines from human, chimpanzee, and gorilla. Gorilla was included as an out-group to determine in which lineage the inversion event occurred. Four of the five regions investigated by FISH were confirmed to be inverted between human and chimpanzee (Table 1), while the chimpanzee assembly did not match our results for a 1.4-Mb region on Chromosome 2. In three cases, the orientation of the region in gorilla matched that of the chimpanzee, indicating that the inversion is specific to the human lineage. The 4.3-Mb inversion at 7p14 is the largest inversion in this dataset not previously described in literature (Figure 3A). The inversion is almost entirely contained within the 7p14.1 band on G-banded chromosomes, which may explain why it was not detected in previous cytogenetic studies.
Table 1 Regions That Were Tested Experimentally
Figure 3 FISH Confirmation of Inversions
(A) Three-color interphase FISH targeting the largest novel inversion between human and chimpanzee identified in this study. The probe order based on the human assembly is RP11-91E16 (red), RP11-321C5 (yellow), and RP11-81F19 (green). The result for human interphase testing is shown to the left and shows the expected the probe order red-yellow-green. The result for chimpanzee and gorilla displays the inverted probe order, red-green-yellow, using identical probes. For this region, each of ten human controls showed the same probe order.
(B) Results showing an interphase nucleus from a human control polymorphic for the 730-kb inversion at 7p22. The probe order is red-yellow-green in the human assembly, and red-green-yellow in the chimpanzee assembly. The probe order for gorilla matches that of the chimpanzee.
The remaining 22 experimentally tested regions were investigated by PCR followed by DNA sequencing. Nineteen of these regions were shown to be inverted in chimpanzee compared to human. Of these, ten had the same orientation in gorilla and chimpanzee, while nine regions were the same in gorilla and human (Table 1). All regions that were inverted between human and chimpanzee were further tested in two additional chimpanzees, one additional gorilla, one orangutan, and one macaque. These results were also consistent, with the lower primates matching the orientation found in gorilla.
Correlation with Genomic Features
Both inversions and copy number polymorphisms in the human genome show a strong correlation with regions containing segmental duplications [15,18,19,23]. Analyses of correlations between inversion regions and segmental duplications were, therefore, performed for all inversion regions in the human and chimpanzee genomes, respectively (Table S2). In both species there is a highly significant increase of segmental duplications around inversion breakpoints as compared to the genome-wide average (p < 0.0001 in both genomes). In the human genome, 75% of the 66 inversions more than 25 kb are flanked on one or both sides by segmental duplications. For 13 regions the flanking duplications are highly identical (96.6% average identity) and nine of these regions are of inverted orientation, which may explain the mechanism by which the inversion occurred.
Of the 1,576 putative inversion regions we detected computationally, 151 overlap RefSeq genes in the human genome assembly (Table S5), and 39 of these have one or more genes entirely contained within the inversion segment. Moreover, 83 inversions are contained within a gene (Table S6), and the remaining 29 regions have a breakpoint that intersects a gene, prioritizing them as candidates for biological and evolutionary studies.
Identification of Human Polymorphic Inversions
To confirm the orientation in human samples, all 23 experimentally validated inversions between human and chimpanzee were interrogated in ten unrelated individuals from the Centre d'etude du polymorphisme humain (CEPH) collection. The results for 20 of the 23 regions confirmed the orientation found in the initial human cell line. Three regions, however, were discovered to be polymorphic, with one allele matching the human assembly and the other allele matching the chimpanzee assembly. The first region, a 730- kb interval at human 7p22, was identified by interphase FISH, and two out of 20 individuals (10%) were found to be heterozygous (Figure 3B). The region is flanked on both sides by segmental duplications of high-sequence identity (Figure 4A). Detailed sequence analysis of these segmental duplications show several independent elements with both intra- and inter-chromosomal distribution patterns. One of these segmental duplications is present as a pair of duplicons on each side of the inverted region. The duplicons are of inverted orientation and extend for ~100 kb with an average sequence identity of 99%. The flanking duplications may be an indication that this inversion is a recurrent variant. In order to establish whether the inversion was a de novo event, both parents of one of the inversion carriers were tested. The results show that the variant was inherited from the mother, who was also a carrier for the inversion. The inversion region encompasses more than 15 RefSeq genes, including the PMS2 gene, known to be involved in colorectal cancer [24].
Figure 4 Overview of Polymorphic Regions
(A) Overview of the region at 7p22 harbouring a 730-kb inversion variant. Each side of the inversion is flanked by highly identical segmental duplications of inverted orientation extending for ~100 kb with an average identity of 99%. Blue bars indicate that the duplications are intra-chromosomal, while green bars harbour both intra- and inter-chromosomal duplications. It is currently not clear exactly where within these segmental duplications the breakpoints occur. The region is comparatively gene-rich and provides an interesting target for diseases with linkage to this region.
(B) Inversion polymorphism at 7q11. The inversion (shown in red) also led to a deletion of 5 kb (blue). The inversion and deletion variant is now the major allele. Two SNPs in perfect linkage disequilibrium (LD) with the inversion are also shown. There are no genes overlapping this inversion variant.
(C) Inversion polymorphism at 16q24. This 1-kb inversion may have been induced by the flanking ALU repeats. The inversion indicated by the net alignment between human and chimpanzee is shown in blue. Experimental data show that the actual inversion (red) is approximately 400 bp longer than indicated by the net alignment. PCR results for two CEPH families are shown to the right. The top PCR was designed for the Build 35 assembly (652 bp) and the lower PCR was designed for the chimpanzee sequence (900 bp). The variant is inherited and shows the expected pattern of inheritance.
The second inversion polymorphism, a 13-kb fragment at 7q11, corresponds to one of the regions that was also found to differ between the two human Chromosome 7 assemblies. The region is approximately 18 kb in size in the chimpanzee assembly, 18 kb in the National Center for Biotechnology assembly, and 13 kb in the CRA_TCAGchr7v2 assembly. The difference between the human assemblies is a 13-kb inversion associated with a 5-kb deletion (Figure 4B). There are no annotated genes in this region. Of the ten individuals tested initially, four were found to be homozygous for the inversion and deletion 13-kb region. The variant was found to be stably inherited as a polymorphism in a three-generation pedigree. Upon examination of the block pattern in the HapMap analysis [25] of this region, it was found that linkage disequilibrium with adjacent markers was very high. Eight CEPH samples that are part of the HapMap sample were then chosen based on their genotypes for five single nucleotide polymorphism (SNP) markers (rs1464853, rs1464851, rs1525303, rs1525287, and rs1568868) overlapping or flanking the inversion variant. In this small sample there was perfect linkage disequilibrium between the inversion variant and three SNP markers (Table 2). In fact, in the CEPH HapMap samples, marker rs152303, which is located within the inverted region, acts as a perfect surrogate marker for the inversion. Using the HapMap data we estimated the allele frequency for the inversion variant in CEPH samples of European ancestry (Table 3). The data show that the minor allele (18-kb allele, 30% frequency) matches the orientation of the chimpanzee genome and is represented in the National Center for Biotechnology Build 35 assembly, while the major allele (13-kb inversion and deletion allele) is represented in the CRA_TCAGchr7v2 assembly.
Table 2 LD with 18-kb Inversion and Deletion Variant
Table 3 Frequency of the Inversion in the CEPH HapMap Samples Based on LD with Surrounding Markers
The third inversion polymorphism example is a 1,065-bp fragment located on Chromosome 16q24 (Figure 4C), with a minor allele frequency of 48% in 12 CEPH controls. The breakpoints occur within Alu repeats flanking the inversion on both sides. In this case the inversion was found to be 403 bp longer than indicated by the net alignment between human and chimpanzee. There are no genes overlapping this inversion and we found no evidence for linkage disequilibrium between surrounding markers in the HapMap data.
Discussion
We describe the first comprehensive high-resolution study of inversions in recent primate evolution. With our experimentally verified data alone we more than double the catalogue of inversions that exist between the human and chimpanzee genomes. Importantly, we also identify another 1,549 putative inversion regions that can now be assessed experimentally. Taken together, our data indicate that inversions have been a frequent rearrangement in the evolution of the human genome and, as such, find that more than half of the experimentally validated inversions are specific to the human lineage. These results, in conjunction with the recent data from Tuzun et al., indicate that polymorphic inversions along chromosomes are common in the human population.
One of the limitations of the present study is that the chimpanzee assembly is currently a low-coverage draft sequence. This decreases the accuracy of the genome alignments. We therefore had to apply rigorous filtering criteria to reduce the vast number of false-positive inversion alignments. With a starting set of more than 6,000 regions from the raw alignment, only 1,576 regions remained after filtering. A separate source of chimpanzee genomic sequence is available through the fully sequenced BACs, which are currently not incorporated in the chimpanzee draft assembly. Analyses of the overlap between our computationally predicted regions and the BAC sequences supported 11 out of 60 regions that could be unambiguously mapped. As this analysis indicates, there is still a large fraction of false-positive inversions among these 1,576 regions. Some regions are also inherently difficult to assemble properly. One example is the 1.4-Mb region on Chromosome 2, where the chimpanzee assembly orientation was not confirmed by our FISH analysis during experimental validation. The region is flanked by gaps in both the human and chimpanzee assemblies, and we note that the orientation in the human assembly for this region was inverted between the two most recent assemblies of the human genome. With the currently available draft assembly for the chimpanzee, it is difficult to obtain a dataset devoid of false-positive inverted alignments. It is important to point out that the computational analysis provides a list of putative regions that are relevant for further studies and not a finalized list of inversion between the human and chimpanzee genomes. The release of the complete chimpanzee sequence in the future will greatly facilitate this type of approach and allow for a more accurate and complete identification of structural variation between the human and chimpanzee genomes.
The putative inversions are distributed throughout the genome with no obvious bias for specific regions of the chromosomes. A large fraction of the putative inversions are smaller than 250 bp in length. It is important to point out that net alignment programs are more likely to randomly assign orientation to very short alignments and the over-representation of very small inversions may therefore reflect a higher false-positive rate. In order to try and address this we performed an analysis of the percent identity between human and chimpanzee sequences for each of the alignments (Table S1). The results show that there is a lower average identity for regions less than 1 kb in size (Figure S1). The percent identity can be used as a quality measure for the alignments.
Our results show that there is a significant correlation between inversion events and segmental duplications. This was to be expected as segmental duplications have been shown to be associated both with copy number changes between human and chimpanzee [9,10] and with copy number variants in the human population [15,18,19]. This correlation is important for our understanding of the mechanisms underlying structural variation in the human genome. We also observe that the breakpoint of the 730-kb inversion on 7p22, which was found to be polymorphic in the human population, maps to exactly the same region as the breakpoint for the pericentric inversion that occurred after the divergence of higher primates from the orangutan [26]. This further supports the notion that chromosomal breakpoints have been reused throughout mammalian evolution [27].
The most interesting finding in this study is that three of the inversions were identified to be polymorphic in the human population. It would be expected that a certain fraction of the differences found between the human and chimpanzee assemblies are polymorphic in one of the two species, but perhaps not to the extent (13%) observed in this study. However, our selection of regions chosen for experimental validation was biased in that it included regions found to differ between the two human assemblies for Chromosome 7, thus increasing the chance to find regions that are polymorphic in the human population. Prior to the recent results by Tuzun et al., only a handful of inversions in the human genome were described (see The Database of Genomic Variants; http://projects.tcag.ca/variation/).
Characterization of inversion variants is important as they can instigate illegitimate recombination events leading to chromosome deletion in the off-spring of carriers. For example, a significantly higher incidence of inversions has been found in parents of patients in some microdeletion syndromes, including Williams-Beuren syndrome [28], Angelman syndrome[29], and Sotos syndrome [30]. Inversion may also act as suppressors of recombination, as was recently shown for a 900-kb inversion polymorphism on human Chromosome 17, which was also found to be positively selected for in several European populations [21]. Population studies are required to determine if the inversion polymorphisms identified here are recurrent or have increased in frequency after a single mutation event. The inversion and deletion polymorphism on 7q11 is unlikely to be a recurrent event as it is in perfect linkage disequilibrium with surrounding markers on a specific haplotype background.
In conclusion, we show that there have been a substantial number of inversion rearrangements in the human genome since the divergence from the chimpanzee. This finding indicates that inversion variants are likely to be abundant in the human genome, and this notion is further supported by the fact that three of the regions investigated in detail in this study were polymorphic in unrelated human samples. Many previously identified inversion variants have been linked to susceptibility to disease or increased risk for disease (usually via microdeletion) in the off-spring. Further studies aimed at understanding the impact the inversions identified in this study may have on gene expression and disease are now possible and are under way.
Materials and Methods
Identification of putative inversions.
The net human and chimpanzee alignments were downloaded from the University of California at Santa Cruz Web site (http://genome.ucsc.edu/). The net alignments were derived from BLASTZ alignments [31] generated by comparing the November, 2003 chimpanzee (panTro1) genome assembly and the May, 2004 (hg17) human genome assembly (http://genome.ucsc.edu/cgi-bin/hgTrackUi?hgsid=59218717&g=netPanTro1). The dataset was filtered to extract significant matches found to represent putative inversions. Matches to random chromosomes were removed, and only matches to syntenic chromosomes were maintained [8]. To reduce the number of false-positives, only those alignments better than or equal to level three were kept. This would preclude identification of inversions within inversions, but was required to reduce the number of potential artefacts. All inversion sequences were lower-case masked for highly repetitive elements by RepeatMasker (A.F. Smit and P. Green, unpublished data), and segmental duplications were downloaded from the human genome segmental duplication database (http://projects.tcag.ca/humandup) [32]. Net alignments with repeats or duplications that comprised greater than 90% of the inverted sequence were removed. The best reciprocal chain alignments from UCSC were obtained (panTro1.rbest.chain) to further refine the dataset to obtain the best set of inversions. In-house Perl scripts were developed to construct the rbest gapped alignment and this was used to filter out additional non-syntenic matches. The analysis of sequence identity for inverted regions was calculated as percent match, defined as the number of matching nucleotides within the inversion divided by the length of the alignment excluding insertions and deletions. To visualize the location and distribution of inversions, the data were converted and displayed using the publicly available visualization tool, GenomePixelizer [33] (http://www.atgc.org/GenomePixelizer/GenomePixelizer_Welcome.html).
Correlation with segmental duplications and genes.
To determine the association between large inversions and flanking segmental duplications, the proximal and distal breakpoints of the putative inversions larger than 25 kb were scanned for the presence of segmental duplications [32]. A window of 25 kb (± 12.5 kb from each breakpoint) was examined for the presence or absence of duplications in the human and chimpanzee genomes (http://projects.tcag.ca/xenodup). These results were then compared with the genome-wide average derived from all 25-kb windows in the genome. A chi squared test was performed to determine the significance of the relationship.
The association between inversions and genes was examined to determine if any genes may have been interrupted by an inversion event. The current RefSeq dataset was downloaded from the UCSC Web site. In-house Perl scripts were developed to compare the location of genes and inversions. Three classes of relationships were described; genes spanning one inversion breakpoint, genes spanning both inversion breakpoints, and genes contained within an inversion.
We obtained the November, 2003 chimpanzee genome assembly through the UCSC Human Genome Browser. All chromosome sequences were lower-case masked for highly repetitive elements by RepeatMasker (A.F. Smit and P. Green, unpublished data). Each of the 26 masked chromosome sequences (including one unmapped chromosome sequence “ChrUn”) was compared against itself by chromosome-wide megaBLAST to detect intra-chromosomal segmental duplications. All possible pair-wise comparisons with each of the other 25 chromosomes were performed to detect inter-chromosomal segmental duplications. All BLAST results were subsequently filtered to eliminate low-quality and fragmented alignments according to methods previously described [32]. The data are displayed in the non-human genome segmental duplication database (http://projects.tcag.ca/xenodup).
Analysis of chimpanzee BAC clones.
To obtain additional support for putative inversions, fully sequenced chimpanzee BAC clones were mapped to the human genome assembly by BLAST, and those that overlapped breakpoints of inversions were detected. A sequence comparison of the chimpanzee clone and human DNA at the inversion breakpoint was performed. The inversion sequence along with 1 kb of flanking DNA sequence from the human genome was compared by BLAST with the entire chimpanzee clone and each segment (flanking proximal, inversion, flanking distal) was scored as + or − in orientation. Therefore a +/−/+ or −/+/− match was taken to confirm the inversion. Entries for which a significant match was not found for any of the three segments were analyzed manually.
PCR and sequencing.
A total of three oligonucleotide primers were designed for each inversion region with one primer within the inversion region based on the human orientation, one primer within the inversion region based on the chimpanzee orientation, and one primer outside the inversion region. All primers were optimized using a gradient hybridization temperature from 56–61 °C. The PCR cycling conditions were 95 °C for 5 min followed by 38 cycles of (95 °C 15 s, optimized hybridization temperature 30 s, 72 °C 60 s per kb product length) and a final extension 72 °C for 5 min. PCR primers and optimized hybridization temperatures are available upon request.
Fluorescence in situ hybridization.
FISH was used to score inversion both between humans and chimpanzees, as well as for scoring of inversion polymorphism in humans. Each region was interrogated using three- color interphase FISH with two probes within the putative inversion region and a reference probe outside. BAC clones were used as probes for four of the region, while fosmid clones were used for the 7q12 region. Each clone was first tested on DAPI stained metaphase chromosomes to ensure that each probe mapped uniquely to the correct chromosomal location. At least 100 interphase nuclei were scored for each probe set. The polymorphic region on 7p22 was further validated with a set of independent probes. The probes used for FISH are shown in Table S7. Interphase nuclei were hybridized as previously described [28].
Cell lines and DNA samples.
Human CEPH lymphocyte cell lines were obtained from Coriell Cell Repositories (Coriell, Camden, New Jersey, United States) and primate lymphocyte cell and fibroblast lines were obtained from The European Collection of Cell Cultures. Genomic DNA was extracted from cell lines. The ten CEPH individuals used for FISH screening were: GM10859, GM7057, GM06990, GM10858, GM10832, GM13114, GM13180, GM13181, GM10835, and GM10834. The chimpanzee and gorilla cell lines used for FISH analysis were ECACC cell lines #89072704 and #89072703. DNA from healthy controls of European ancestry was used for PCR based assays.
Supporting Information
Figure S1 Sequence Identity for Putative Inversion
Distribution of percent match between human and chimpanzee sequences for inverted regions. This distribution indicates that regions less than 1 kb in size are more likely to contain false-positive inversions. The percent match for each region is shown is Table S1 and can be viewed as a quality measure for the underlying alignment.
(28 KB PDF)
Click here for additional data file.
Table S1 All Putative 1,576 Inversion Regions between the Human and Chimpanzee Genomes
(116 KB PDF)
Click here for additional data file.
Table S2 Inversions More than 25 kb
(89 KB XLS)
Click here for additional data file.
Table S3 Previously Published Inversions between the Human and Chimpanzee Genomes
(24 KB XLS)
Click here for additional data file.
Table S4 Comparison to Previously Published Inversions
(24 KB XLS)
Click here for additional data file.
Table S5 Inversions Overlapping Genes
(20 KB XLS)
Click here for additional data file.
Table S6 Inversions Contained within Genes
(28 KB XLS)
Click here for additional data file.
Table S7 Probes Used for FISH Experiments
(19 KB XLS)
Click here for additional data file.
The authors would like to thank MaryAnn George for help with karyotyping of primate cell lines, and the Chimpanzee Sequencing and Analysis Consortium for generating the chimpanzee genome assembly. This study was supported by Genome Canada/Ontario Genomics Institute, The Centre for Applied Genomics, and The Hospital for Sick Children Foundation, and the McLaughlin Centre for Molecular Medicine. LF is supported by the Swedish Medical Research Council, and SWS is an Investigator of the Canadian Institutes of Health Research and an International Scholar of the Howard Hughes Medical Institute.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. LF, JRM, and SWS conceived and designed the experiments. LF, TT, ML, GR, and RK performed the experiments. LF, JRM, TT, ARC, ML, GR, RK, and SWS analyzed the data. SWS contributed reagents/materials/analysis tools. LF, JRM, and SWS wrote the paper.
A previous version of this article appeared as an Early Online Release on September 29, 2005 (DOI: 10.1371/journal.pcbi.0010055.eor).
Abbreviations
BACbacterial artificial chromosome
bpbase pair
CEPHCentre d'etude du polymorphisme humain
FISHfluorescence in situ hybridization
kbkilobase
LDlinkage disequilibrium
Mbmega-base
==== Refs
References
Fujiyama A Watanabe H Toyoda A Taylor TD Itoh T 2002 Construction and analysis of a human-chimpanzee comparative clone map Science 295 131 134 11778049
Li WH Saunders MA 2005 Initial sequence of the chimpanzee genome and comparison with the human genome Nature 437 69 87 16136131
Ebersberger I Metzler D Schwarz C Paabo S 2002 Genome-wide comparison of DNA sequences between humans and chimpanzees Am J Hum Genet 70 1490 1497 11992255
Hacia JG Makalowski W Edgemon K Erdos MR Robbins CM 1998 Evolutionary sequence comparisons using high-density oligonucleotide arrays Nat Genet 18 155 158 9462745
Thomas JW Touchman JW Blakesley RW Bouffard GG Beckstrom-Sternberg SM 2003 Comparative analyses of multi-species sequences from targeted genomic regions Nature 424 788 793 12917688
Watanabe H Fujiyama A Hattori M Taylor TD Toyoda A 2004 DNA sequence and comparative analysis of chimpanzee chromosome 22 Nature 429 382 388 15164055
Britten RJ 2002 Divergence between samples of chimpanzee and human DNA sequences is 5%, counting indels Proc Natl Acad Sci U S A 99 13633 13635 12368483
Yunis JJ Prakash O 1982 The origin of man: A chromosomal pictorial legacy Science 215 1525 1530 7063861
Fortna A Kim Y MacLaren E Marshall K Hahn G 2004 Lineage-specific gene duplication and loss in human and great ape evolution. PLoS Biol 2: E207. DOI: 10.1371/journal.pbio.0020207
Locke DP Segraves R Carbone L Archidiacono N Albertson DG 2003 Large-scale variation among human and great ape genomes determined by array comparative genomic hybridization Genome Res 13 347 357 12618365
Tuzun E Bailey JA Eichler EE 2004 Recent segmental duplications in the working draft assembly of the brown Norway rat Genome Res 14 493 506 15059990
Cheung J Wilson MD Zhang J Khaja R MacDonald JR 2003 Recent segmental and gene duplications in the mouse genome Genome Biol 4 R47 12914656
Armengol L Pujana MA Cheung J Scherer SW Estivill X 2003 Enrichment of segmental duplications in regions of breaks of synteny between the human and mouse genomes suggest their involvement in evolutionary rearrangements Hum Mol Genet 12 2201 2208 12915466
Bailey JA Baertsch R Kent WJ Haussler D Eichler EE 2004 Hotspots of mammalian chromosomal evolution Genome Biol 5 R23 15059256
Iafrate AJ Feuk L Rivera MN Listewnik ML Donahoe PK 2004 Detection of large-scale variation in the human genome Nat Genet 36 949 951 15286789
Cheng Z Ventura M She X Khaitovich P Graves T 2005 A genome-wide comparison of recent chimpanzee and human segmental duplications Nature 437 88 93 16136132
Sebat J Lakshmi B Troge J Alexander J Young J 2004 Large-scale copy number polymorphism in the human genome Science 305 525 528 15273396
Sharp AJ Locke DP McGrath SD Cheng Z Bailey JA 2005 Segmental duplications and copy-number variation in the human genome Am J Hum Genet 77 78 88 15918152
Tuzun E Sharp AJ Bailey JA Kaul R Morrison VA 2005 Fine-scale structural variation of the human genome Nat Genet 37 727 732 15895083
Navarro A Barton NH 2003 Chromosomal speciation and molecular divergence—accelerated evolution in rearranged chromosomes Science 300 321 324 12690198
Stefansson H Helgason A Thorleifsson G Steinthorsdottir V Masson G 2005 A common inversion under selection in Europeans Nat Genet 37 129 137 15654335
Scherer SW Cheung J MacDonald JR Osborne LR Nakabayashi K 2003 Human chromosome 7: DNA sequence and biology Science 300 767 772 12690205
Fredman D White SJ Potter S Eichler EE Dunnen JT 2004 Complex SNP-related sequence variation in segmental genome duplications Nat Genet 36 861 866 15247918
Papadopoulos N Nicolaides NC Wei YF Ruben SM Carter KC 1994 Mutation of a mutL homolog in hereditary colon cancer Science 263 1625 1629 8128251
The International HapMap Consortium 2003 The International HapMap Project Nature 426 789 796 14685227
Muller S Finelli P Neusser M Wienberg J 2004 The evolutionary history of human chromosome 7 Genomics 84 458 467 15498453
Pevzner P Tesler G 2003 Human and mouse genomic sequences reveal extensive breakpoint reuse in mammalian evolution Proc Natl Acad Sci U S A 100 7672 7677 12810957
Osborne LR Li M Pober B Chitayat D Bodurtha J 2001 A 1.5 million-base pair inversion polymorphism in families with Williams-Beuren syndrome Nat Genet 29 321 325 11685205
Gimelli G Pujana MA Patricelli MG Russo S Giardino D 2003 Genomic inversions of human chromosome 15q11-q13 in mothers of Angelman syndrome patients with class II (BP2/3) deletions Hum Mol Genet 12 849 858 12668608
Visser R Shimokawa O Harada N Kinoshita A Ohta T 2005 Identification of a 3.0-kb major recombination hotspot in patients with sotos syndrome who carry a common 1.9-Mb microdeletion Am J Hum Genet 76 52 67 15580547
Schwartz S Kent WJ Smit A Zhang Z Baertsch R 2003 Human-mouse alignments with BLASTZ Genome Res 13 103 107 12529312
Cheung J Estivill X Khaja R MacDonald JR Lau K 2003 Genome-wide detection of segmental duplications and potential assembly errors in the human genome sequence Genome Biol 4 R25 12702206
Kozik A Kochetkova E Michelmore R 2002 GenomePixelizer—a visualization program for comparative genomics within and between species Bioinformatics 18 335 336 11847088
|
16254605
|
PMC1270012
|
CC BY
|
2021-01-05 08:00:24
|
no
|
PLoS Genet. 2005 Oct 28; 1(4):e56
|
utf-8
|
PLoS Genet
| 2,005 |
10.1371/journal.pgen.0010056
|
oa_comm
|
==== Front
Epidemiol Perspect InnovEpidemiologic perspectives & innovations : EP+I1742-5573BioMed Central London 1742-5573-2-91618535410.1186/1742-5573-2-9MethodologyAn easy approach to the Robins-Breslow-Greenland variance estimator Silcocks Paul [email protected] Trent Research & Development Support Unit, Medical School, Queen's Medical Centre, Nottingham, NG7 2UH UK2005 26 9 2005 2 9 9 26 4 2005 26 9 2005 Copyright © 2005 Silcocks; licensee BioMed Central Ltd.2005Silcocks; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The Mantel-Haenszel estimate for the odds ratio (and its logarithm) in stratified case control studies lacked a generally acceptable variance estimate for many years. The Robins-Breslow-Greenland estimate has met this need, but standard textbooks still do not provide an explanation of how it is derived. This article provides an accessible derivation which demonstrates the link between the Robins-Breslow-Greenland estimate and the familiar Woolf estimate for the variance of the log odds ratio, and which could easily be included in Masters level courses in epidemiology. The relationships to the unconditional and conditional maximum likelihood estimates are also reviewed.
Odds ratioVarianceRobins-Breslow-GreenlandRBG estimate
==== Body
Introduction
The Mantel-Haenszel (MH) estimate for the summary odds ratio across several 2 × 2 tables, ψMH, was proposed in 1959 [1]. Over twenty years later the lack of a robust estimate for its variance was still being noted [2], yet only a few years afterwards, Robins, Breslow and Greenland introduced their now generally-accepted variance estimator [3] for the Mantel-Haenszel log-odds ratio (denoted by the RBG estimate). This replaced estimation of confidence limits based on the unsatisfactory test-based procedure of Miettinen or the computationally intensive Cornfield type limits which had hitherto been used.
While a useful review of Mantel-Haenszel methods has been published, including some aspects of the historical development towards the RBG estimator [4] the formal derivations by Robins, Breslow & Greenland [3] and Phillips & Holland [5] are not, in the view of this author, easily comprehended. The former omits steps in the argument, while the latter appeals to descending factorial powers. Possibly it is no surprise that even modern textbooks [6,7] merely state the RBG formula without deriving it.
While other variance estimators exist, some are ad hoc, such as the application of the cohort study formula to case-control data suggested by Clayton and Hills [8], only apply to the large few strata case [9] or are closely related to the RBG estimator [10]. One rather different exception is Sato's formula [11] but this procedure gives confidence limits directly in the odds ratio scale.
It is the intention of this article to present an informal derivation of the RBG estimator as an extension of the familiar variance formula of Woolf [12], and which could readily be included in standard textbooks of epidemiology or biostatistics. I will describe this from the perspective of a case-control study.
Analysis
How does the Mantel-Haenszel estimate arise?
Consider a stratified case-control study for which the ith of k independent tables is:
Neglecting constants, the unconditional likelihood for the ith table is:
where in the ith table θi = probability of exposure if a case and øi = probability of exposure if a control.
The maximum likelihood estimate (MLE) for øi is given by bi/(bi + di) and if we re-parameterise θi as ψøi/ [ψøi + (1 - øi)], where ψ is the odds ratio (assumed common to all tables), the contribution to the overall log likelihood made by terms involving ψ is:
∑ ai ln {ψøi/[ψøi + (1 - øi)]} + ci ln {1/[ψøi + (1 - øi)]}.
Differentiating with respect to ψ and equating to zero, and rearranging (noting that ai + ci = n1i)
we obtain:
∑ ai ln {ψøi/[ψøi + (1 - øi)]} + ci ln {1/[ψøi + (1 - øi)]}.
i.e., ∑ {ai - n1iψøi/[ψøi + (1 - øi)]} = 0
i.e., ∑ {[ψaiøi + ai - aiøi - n1iψøi]/[ψøi + (1 - øi)]} = 0.
This must be solved numerically to obtain the MLE for ψ, but if the denominators do not vary too much across the tables we merely have to solve:
∑ [ψaiøi + ai - aiøi - n1iψøi] = 0
i.e., ∑ [ψ(ai - n1i)øi + ai (1 - øi)] = 0
or, ∑ ai (1 - øi) = ∑ ψ(n1i - ai)øi
giving, ∑ ai (1 - øi) = ψ ∑ (n1i - ai)øi
and since, ψi = bi/(bi + di) = bi/n0i
This can be used as a first approximation to find the MLE (if there is only one table then ψ is the unconditional MLE = ad/bc). Now in stratified case-control studies with a constant ratio, r, of controls to cases, the total number of subjects in each stratum is given by ni = n0i (1 + r), so n0i = ni/(1 + r). A constant r will be achieved by design if there is caliper matching; otherwise – as with a post-stratified analysis – this will be only approximately true. The term (1 + r) can then be cancelled and we are left with:
The MH estimator is therefore a first approximation to the unconditional MLE in the large strata case with a constant control:case ratio across strata. However the MH estimator actually coincides with the conditional MLE for the matched pairs design, as outlined, for example on page 164 of Breslow & Day [2].
The sensitivity to variation in the øi and constancy of the control:case ratio is not high, as shown by the data in Table 1. In a sense this would be expected because for the most sparse (e.g., pair-matched) data the control:case ratio will be constant, and while the øi then have maximum variance – being only 0 and 1, the MH estimate coincides with the conditional MLE. Conversely, for large strata the control:case ratio will vary, but the variance of the øi will be less and the MH estimate will then approximate the unconditional MLE.
Table 1 Simulated case-control data with true odds ratio = 5
Case Control ø Controls:cases
36 97 0.58 4
6 71
42 168
Ca Co
41 79 0.94 2
1 5
42 84
Ca Co
2 1 0.02 2
26 55
28 56
Ca Co
19 25 0.30 3
9 59
28 84
Ca Co
20 41 0.37 4
8 71
28 112
Ca Co
30 21 0.62 1
4 13
34 34
Ca Co
22 26 0.46 2
6 30
28 56
Odds ratio estimates (Stata v7.0):
Mantel-Haenszel 4.38 (95% CI 2.85 to 6.72)
Conditional MLE 4.36 (95% CI 2.85 to 6.67)
Unconditional MLE 4.42 (95% CI 2.88 to 6.78)
Deriving the variance of the Mantel-Haenszel estimate
Consider again the ith 2 × 2 table, giving the frequencies in each cell:
For odds ratio , estimated for a single table by the cross-product ratio aidi/bici, application of the delta method gives Woolf's logit-based formula [8]:
with ni = ai + bi + ci + di and,
The delta method is a widely used procedure in statistics when an approximation is needed for the variance of a function of a variable whose variance is known. In this instance the variable with known variance is a proportion p, and the function is the logit. The basic delta method formula is: var(y) ≈ (dy/dx)2 var(x) from which, if y = logit(p = ln[p/(1 - p)],
var(y) ≈ (1/p + 1/(1 - p))2 p(1 - p)/n
= (1/p + 1/(1 - p))1/n
= (1/a + 1/b)
if p = a/n and n = a + b.
Here we have two independent proportions (the proportion of cases and controls exposed) and Woolf's formula is obtained by estimating the variances of the separate logits and adding them.
For k such 2 × 2 tables, each representing a separate stratum, the Mantel-Haenszel pooled estimate of the common odds ratio ψ is given by:
Hence ψMH is a weighted average of the stratum-specific odds ratios. The weights approximate the inverse of the variance of each i if the true value of ψ = 1. Note that the assumption here of a common odds ratio is not required for the Mantel-Haenszel test.
To derive the variance, in addition to the approximation involved in application of the delta rule, an assumption is also made that each stratum-specific odds ratio is close enough to the Mantel-Haenszel pooled estimate to permit terms like aidi/bici to be replaced by ψMH.
We then proceed by obtaining an approximation which avoids zeros in the formula for var[ln()]. The motivation for this can be seen by comparing the weights for ψMH – which are unaffected by zeros except for deleting such strata – whereas if Woolf's variances were used, the result would be indeterminate if cells with zeros were present.
Taking the weights as constant,
Assuming a common odds ratio ψ, estimated by ψMH, this can be written as:
Leading to a formula suggested by Hauck [9]:
As mentioned above, a problem with this formula is that it fails if cell entries are zero. However we can proceed further by re-writing the formula as:
On substituting 1/ψMH for (bici/aidi):
Now if the rows of the 2 × 2 table are interchanged, the variance stays the same. But a similar argument to that above leads to:
(Note that the new odds ratio formed by exchanging rows is just 1/ψMH.) "The" variance, V, of ln(ψMH) is therefore taken to be the mean of the two estimates [13] as follows:
Let R = ∑ (aidi/ni) and S = ∑ (bici/ni). On substituting into the two variance formulae:
Next, divide the top and bottom by S2 and move the term outside the brackets to obtain:
which is eq. 9 in Phillips & Holland [5].
If we now put
Pi = (ai + di)/ni and Qi = (bi + ci)/ni with Ri = aidi/ni and Si = bici/ni
then
which on multiplying out the brackets, rearranging and noting that R/S = ψMH, gives:
This is the RBG formula!
When there is only one stratum, this reduces to (1/a + 1/b + 1/c + 1/d) which is the familiar logit based formula of Woolf and which approaches 0 as the sample size increases, assuming a finite true odds ratio. Clearly as the RBG variance estimate is a finite sum of such estimators the RBG estimate will also approach 0, for large strata.
The RBG estimator was derived above on the assumption that the stratum-specific odds ratio estimates could be liberally replaced by the common value, in turn estimated by ψMH; both assumptions are reasonable with large samples per stratum. However, the success of the RBG formula derives from its being applicable also to the sparse data case.
To see this, consider a matched-pairs case control study. The capital letters denote the frequency of case-control pairs.
In such a study each stratum has only two observations. The table can be decomposed into four types of "unmatched" table according to the exposure category of the case and the control, the frequency of each type being given by the frequency of the corresponding case-control pairs:
Only the B such tables with ai = di = 1 and the C such tables with bi = ci = 1 contribute to the estimate of the odds ratio. Note that these are disjoint sets of tables.
Under these circumstances: ψMH = B/C which coincides with the conditional MLE and:
a) the middle term of the RBG formula vanishes because if bici = 1 then (ai + di) = 0, and if aidi= 1 then (bi+ ci) = 0
b) R = ∑ aidi/ni = B/2 & S = ∑ bici/ni = C/2
c) There are B terms in which aidi (ai + di) = 2 C terms in which bici (bi + ci) = 2
giving:
V = B/B2 + C/C2 = 1/B + 1/C
This is not only the familiar logit based formula for the variance of the log odds ratio for matched pairs, but is also the variance of the conditional maximum likelihood estimate. This is asymptotically consistent from the general properties of a MLE (and it's easy to see that as the number of tables increases, V → 0).
In other words, the RBG formula, though derived here without assuming validity in the sparse case, does in fact possess this property.
Table 1 shows how closely the conditional maximum likelihood estimate, unconditional maximum likelihood estimate, and MH estimate agree, despite varying øi and control:case ratio.
Conclusion
The Mantel-Haenszel estimate of the odds ratio approximates the maximum likelihood estimate for large, few strata and coincides with the conditional maximum likelihood estimate for the sparse data (matched pairs) case.
The RBG formula is the estimator of choice for the variance of the Mantel-Haenszel log-odds-ratio because it applies both in the large few strata case and in the many sparse strata case (as in matched pairs analysis), when the RBG variance estimate actually coincides with the conditional maximum likelihood variance estimate.
Moreover the RBG formula reduces to familiar standard forms for a single stratum and for matched pairs.
Formal derivation of the RBG formula is tricky but an informal, accessible derivation is possible as outlined above, which uses nothing more advanced than the delta method for approximating a variance.
Competing interests
The author(s) declare that they have no competing interests.
==== Refs
Mantel N Haenszel W Statistical aspects of the analysis of data from retrospective studies of disease JNCI 1959 22 719 748 13655060
Breslow NE Day NE Statistical methods in cancer research, Volume 1 – the analysis of case-control studies 1980 Lyons: International Agency for Research on Cancer
Robins J Breslow N Greenland S Estimators of the Mantel-Haenszel variance consistent in both sparse data and large-strata limiting models Biometrics 1986 42 311 323 3741973
Kuritz SJ Landis JR Koch GG A general overview of Mantel-Haenszel methods: applications and recent developments Ann Rev Public Health 1988 9 123 160 3288229 10.1146/annurev.pu.09.050188.001011
Phillips A Holland PW Estimators of the variance of the Mantel-Haenszel log-odds-ratio estimate Biometrics 1987 43 425 431
Armitage P Berry G Matthews JNS Statistical methods in medical research 2002 4 Oxford: Blackwell Science
Fleiss JL Levin B Paik MC Statistical Methods for Rates & Proportions 2003 Chichester: John Wiley
Clayton D Hills M Statistical methods in epidemiology 1995 Oxford: Oxford University Press
Hauck WW The large-sample variance of the Mantel-Haenszel estimator of a common odds ratio Biometrics 1979 35 817 819
Flanders WD A new variance estimator for the Mantel-Haenszel odds ratio Biometrics 1985 41 637 642
Sato T Confidence limits for the Common Odds Ratio Based on the Asymptotic Distribution of the Mantel-Haenszel Estimator Biometrics 1990 46 71 80
Woolf B On estimating the relationship between blood group and disease Human Genet 1955 19 251 253
Ury HK Hauck's approximate large-sample variance of the Mantel-Haenszel estimator [letter] Biometrics 1982 38 1094 1095
|
16185354
|
PMC1270683
|
CC BY
|
2021-01-04 16:36:38
|
no
|
Epidemiol Perspect Innov. 2005 Sep 26; 2:9
|
utf-8
|
Epidemiol Perspect Innov
| 2,005 |
10.1186/1742-5573-2-9
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r641608684610.1186/gb-2005-6-8-r64ResearchGenomic mapping of RNA polymerase II reveals sites of co-transcriptional regulation in human cells Brodsky Alexander S [email protected] Clifford A [email protected] Ian A [email protected] Giles [email protected] Benjamin J [email protected] Xiaole S [email protected] Edward A [email protected] Pamela A [email protected] Department of Systems Biology, Harvard Medical School and Department of Cancer Biology, Dana-Farber Cancer Institute, 44 Binney St, Boston, MA 02115, USA2 Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute and Harvard School of Public Health, Boston, MA 02155, USA3 Department of Medicine, Harvard Medical School and Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA 02115, USA2005 15 7 2005 6 8 R64 R64 4 1 2005 7 4 2005 17 6 2005 Copyright © 2005 Brodsky et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Determination of the distribution of RNA Polymerase II within regions of the human genome identifies novel sites of transcription and suggests that a major factor of transcription elongation control in mammals is the coordination of transcription and pre-mRNA processing to define exons.
Background
Transcription by RNA polymerase II is regulated at many steps including initiation, promoter release, elongation and termination. Accumulation of RNA polymerase II at particular locations across genes can be indicative of sites of regulation. RNA polymerase II is thought to accumulate at the promoter and at sites of co-transcriptional alternative splicing where the rate of RNA synthesis slows.
Results
To further understand transcriptional regulation at a global level, we determined the distribution of RNA polymerase II within regions of the human genome designated by the ENCODE project. Hypophosphorylated RNA polymerase II localizes almost exclusively to 5' ends of genes. On the other hand, localization of total RNA polymerase II reveals a variety of distinct landscapes across many genes with 74% of the observed enriched locations at exons. RNA polymerase II accumulates at many annotated constitutively spliced exons, but is biased for alternatively spliced exons. Finally, RNA polymerase II is also observed at locations not in gene regions.
Conclusion
Localizing RNA polymerase II across many millions of base pairs in the human genome identifies novel sites of transcription and provides insights into the regulation of transcription elongation. These data indicate that RNA polymerase II accumulates most often at exons during transcription. Thus, a major factor of transcription elongation control in mammalian cells is the coordination of transcription and pre-mRNA processing to define exons.
==== Body
Background
Transcriptional and post-transcriptional regulation of gene expression intersect at RNA polymerase II. The rate of polymerase II movement is altered by loading of transcription factors at the promoter, chromatin structure, pre-mRNA processing, elongation control and termination [1-3]. Thus, polymerase II accumulates at promoters as well as at different locations across a particular gene [4], but the general patterns across many different genes have yet to be explored. Numerous factors such as histones, post-translation modifying enzymes, and RNA-binding proteins regulate these processes [1,3]. One key determinant of transcription is the phosphorylation state of the carboxy-terminal domain (CTD) of polymerase II [5,6] which becomes hyperphosphorylated during transcription elongation [4,6-9]. Much of our understanding of transcription elongation comes from work in prokaryotes and yeast where most genes are intronless [1,3]. Transcription and pre-mRNA processing are coordinated, as the two processes affect the efficiency of each other [2,10]. The spatial patterns of the different phosphorylation states of polymerase II across genes remains poorly understood in mammalian systems.
Results and discussion
To explore the range of locations where polymerase II accumulates across the genome, we performed chromatin immunoprecipitation (ChIP) from HeLa S3 cells, and profiled the purified DNA using an oligonucleotide-tiled microarray interrogating the Encyclopedia of DNA Elements (ENCODE) regions [11] covering 471 known genes. Two antibodies were used, 8WG16 and 4H8, which recognize the hypophosphorylated (PolIIa) or a phosphorylation-independent state of the CTD of polymerase II (PolII), respectively. Thus, the 4H8 antibody is recognizing the total polymerase II population. Isolated DNA was amplified using a multiple displacement amplification (MDA) strategy (see Materials and methods) [12].
To identify sites of enrichment, we used a non-parametric approach generalizing the Wilcoxon signed-rank test [13]. Signals across 1,000 nucleotides were used to determine a p-value for each probe. Probes were filtered for uniqueness within the bandwidth. Probes with p-values below 10-4 were selected for further analysis because this threshold has a low false-positive rate as determined by PCR analysis (Figure 1). With these parameters, the hypophosphorylated-specific anti-PolIIa antibody reveals 102 occupied sites, whereas the phosphorylation-independent antibody shows 550 sites (Table 1).
RNA polymerase II has distinct landscapes across each gene. Figure 2 shows representative genes with polymerase enrichments. PolIIa is highly enriched at transcription initiation sites. On the other hand, PolII shows gene-specific landscapes with the strongest enrichments at exons within actively transcribed loci. Active genes reveal lower p-values across the gene compared with intergenic or inactive genes (compare Figure 2a and 2b), indicating a relative absence of polymerase II from the nontranscribed regions. Some smaller genes with high exon density, such as SF1, reveal significant polymerase signal across almost the entire locus (Figure 2a). Distinct accumulations are observed with significant p-values around exons for both SF1 and KIAA1932. In the KIAA1932 gene, PolII is enriched at a subset of constitutively and alternatively spliced exons (Figure 2c). For some genes, RNA polymerase II is enriched at relatively few locations within the gene (Figure 2d).
An important question is to determine if the polymerase II sites are indicative of active transcription. We addressed this in multiple ways. First, microarray expression profiling of the mRNA with Affymetrix U133 Plus 2 chips confirms that many of the RNA polymerase II-associated genes are actively expressed in HeLa cells, as seen in a plot of mRNA expression level versus p-value in Figure 3. Genes with significant RNA polymerase II enrichment are biased towards genes with higher mRNA levels. Figure 3 also shows that some genes have apparently high mRNA levels but no significant levels of PolII or PolIIa. This could be due to very low transcription levels but high mRNA stability. Second, we measured RNA from the same HeLa cells on the ENCODE tiled arrays. We observe that 34% of the PolII sites overlap with RNA signal (compared to approximately 8% expected at random) and 50% of the PolII locations are within 1 kb of some RNA signal (compared to 13% expected at random). Many sites where small pieces of RNA are synthesized, such as small exons, may be missed as a result of the spacing of the oligonucleotide probes and the imperfect nature of the probes. Third, many of the PolII and PolIIa sites overlap with annotated expressed sequence tags (ESTs) and mRNAs. Eighty-seven percent of the PolII-enriched and 88% of the PolIIa-enriched locations overlap with EST regions, compared to 31% and 44% expected at random, respectively. Lastly, reverse transcriptase PCR checks of KIAA1932 and DKC1 indicate that these genes are being expressed (data not shown). These data suggest that RNA polymerase II sites are biased towards regions of active transcription and that determining sites of enrichment of RNA polymerase II is an indicator of transcription.
Levels of RNA polymerase II enrichment at internal exons can vary between genes. To examine whether these patterns are influenced by expression levels, two categories were created: genes with multiple PolII enrichments at internal exons; and genes with PolII at one or zero internal exons. When compared to the mRNA levels, there is no significant difference between the two categories, suggesting that the number of PolII sites across the gene does not vary significantly with RNA levels. Genes with observable PolII enrichment at internal exons are correlated with higher mRNA levels on the expression array. This is consistent with reports proposing the use of PolII ChIP to monitor gene expression [14]. Therefore, the number of PolII sites at internal exons may reflect different levels of transcription elongation control and not just the sensitivity of the experiment.
Distinct from the hypophosphorylation-specific antibody, the phosphorylation-independent antibody reveals diverse enrichment locations for PolII. In total, 74% of the identified PolII locations are near an annotated knownGene, RefSeq, or genscan exon as summarized in Table 1 (see Additional data file 2 for a list of PolII genscan exon locations). Unlike PolIIa, PolII sites are distributed between the 5' and 3' ends of genes, with a slight bias towards terminal exons over initiating exons (Figure 4). This is probably reflecting the stalling of PolII during the coupled processes of transcription termination and 3'-end processing [15]. For some genes, significant PolII signal is observed more than 1 kb past the terminal exon, which might indicate transcription of the longer pre-mRNA before 3'-end cleavage and polyadenylation [16]. Figure 5 shows two representative genes with significant PolII enrichment past the terminal exon.
Most of the hypophosphorylated PolIIa locations at internal exons also overlap a transcription initiation site, as the internal exon in question is often the second exon in the gene. Only two enrichment sites overlap with an internal exon without also being near the first exon of a transcript. One of these is at a CpG island in the MCF2L gene and the other may be an alternative transcription initiation site as annotated in the HG17 assembly at the beginning of the ITGB4BP gene. To classify the remaining sites within introns or in intergenic regions, enrichment sites were compared to other gene databases. As summarized in Table 1, four PolIIa sites are in introns, but three of these are within resolution of annotated or predicted exons, leaving only one location not overlapping an exon of some kind. There are 28 hypophosphorylated polymerase sites not in a RefSeq gene region. After following a similar filtering approach, only 14 sites remain that are not near a putative exon. Thus, only 14% of PolIIa-enriched locations do not overlap with a known exon or actively transcribed region. Additional data file 2 lists PolIIa sites at predicted exons that are probably newly identified transcription initiation locations in HeLa cells. Figure 5 shows two examples of PolII and RNA signal at new sites of transcription. From the pattern of enrichments it is probable that many of these predicted exons are real and are transcription initiation locations, given the observed strong bias of the 8WG16 antibody for transcription initiation locations in well annotated genes.
To determine the generality of these observations, all RNA polymerase II occupancy sites were compared with the known genes and RefSeq databases, version HG16. PolIIa is highly enriched for the first exons around transcription initiation sites (Figure 4) representing 77 of 551 known genes in HG16 on the array (see Additional data file 1 for the entire lists).
Elongation control is a common transcriptional regulation mechanism believed to affect a wide range of functional gene classes [1]. In particular, RNA polymerase II pausing has been proposed to be associated with alternative splicing, [2]. To determine if there is a bias for alternative exons, we counted all the annotated alternatively spliced exons in the knownGene database and determined the distribution of PolII enrichment locations on them. PolII is enriched at 57% of the annotated alternatively spliced exons of the active genes compared to 37% of annotated actively transcribed constitutively expressed exons. We also examined the distribution of all PolII p-values on different types of exons. Each exon was mapped to the smallest p-value ChIP-enriched site that overlaps the exon. The cassette exons are found to be more significantly associated with smaller p-values compared to constitutively expressed exons according to the two-sample Kolmogorov-Smirnov test with a two sided p-value of less than 0.0035.
One attractive hypothesis is that sites of exon enrichment may reflect weaker splice sites where PolII stalls during splice site recognition. Using two different empirical methods to estimate splice site strength, no significant differences are observed between the exons overlapping PolII and those that do not [17,18]. Alternatively, some of the annotated constitutively expressed exons may actually be subject to alternative splicing decisions. Kampa et al. suggest that the levels of alternative splicing are much higher than commonly believed and annotated in the human genome from their examination of expression on tiled arrays [19]. Consistent with these findings, RNA polymerase II sites may be predicting which exons are being co-transcriptionally alternatively spliced.
To determine if there is any pattern for the 120 PolII enrichment sites that are in RefSeq introns, we compared these sites to knownGene, genscan, geneid, and sgpGene databases and find 31 within resolution of putative exons. Of the remaining 89, 57 are in genes with PolII enrichment sites that also overlap exons, suggesting that they are actively transcribed genes. No clear intronic positional bias is observed.
In conclusion, we have identified new sites of RNA polymerase II accumulation across hundreds of genes in mammalian cells. The large majority of polymerase II-enriched locations are at actively transcribed exons with a bias towards annotated alternatively spliced exons. Many of the PolII sites at annotated constitutively expressed exons may be sites of alternative splicing. Whatever the eventual splicing decision, these observations suggest that events around exons slow transcription elongation. A recent study suggests that even general splicing factors may slow elongation [20]. Stalling of RNA polymerase II near exons may function to slow RNA synthesis in order to wait for the competition of myriad splicing signals to be resolved in order to define the exon [21,22]. These ChIP data identify where these states of RNA polymerase II are localizing across the ENCODE regions.
Across genes, these data are consistent with the hypothesis of transcriptional pausing at particular locations. Alternatively, it is possible that RNA polymerase II is rearranging during transcription such that the epitope is only accessible around exons. Thus, the conformation of polymerase II may be changing and not the transcription rate. Nonetheless, it is interesting that the majority of observable elongating polymerase II accumulates around exons, suggesting that a major feature of transcription elongation control is coupling to pre-mRNA processing.
These observations differ from those observed in intronless genes typically found in prokaryotes and yeast where a more uniform PolII enrichment is observed across genes [16]. What appears to be conserved is PolII accumulation in coding regions compared to intronic regions. These data highlight the complexity and gene-specific nature of transcription regulation not only at transcription initiation and termination locations but at specific exons. Together, these observations suggest that a major feature of transcription elongation control in mammalian cells is exon definition. Thus, these data provide new insights into the coordination of transcription and pre-mRNA processing in mammalian cells.
Materials and methods
Chromatin immunoprecipitation and DNA amplification
Chromatin immunoprecipitations (ChIP) were performed as described with the following modifications [23]. HeLa S3 cells were first crosslinked with dimethyl adipimidate (DMA) (Pierce) for 10 min, washed with PBS and then crosslinked with formaldehyde for 10 min. Cells were collected, lysed, and chromatin was sheared by sonication to an average length of 1 kb as determined after RNase treatment of the samples on an agarose gel. Chromatin was prepared from four independently grown batches of cells and pooled to generate three replicate immunoprecipitations (IP) and six input samples. Briefly, 8WG16 (Covance) and 4H8 (AbCam) antibodies were incubated with a 50:50 mix of Dynal protein A/G beads for more than 16 h at 4°C in PBS with 5 mg/ml BSA. After washing in PBS, beads with bound antibody were incubated with chromatin from approximately 2 × 107 cells for more than 16 h at 4°C. Beads were washed eight times with RIPA buffer (50 mM HEPES pH 7.6, 1 mM EDTA, 0.7% DOC, 1% IGEPAL, 0.5 M LiCl) before DNA was eluted at 65°C in TE/1% SDS. Crosslinks were reversed by incubating at 65°C for more than 12 h followed by proteinase K treatment, phenol extraction and RNase treatment. Isolated DNA was then amplified isothermally using random nonamer primers and Klenow polymerase (Invitrogen) for more than 4 h, yielding approximately 2 μg of DNA per IP. DNA was prepared and hybridized on Affymetrix ENCODE oligonucleotide tiled arrays using the fragmentation, hybridization, staining and scanning procedure described by Kennedy et al. [24]. Affymetrix ENCODE microarrays have interrogating 25mer oligonucleotide probes tiled every 20 bp on average. A sample of chromatin was set aside before IP and used to represent the input DNA.
Tiled array analysis
Quantile normalization was used to make the distribution of probe intensities the same for all arrays [25]. In the case of the Affymetrix GTRANS software quantile normalization is used within treatment and control replicate sets. Non-parametric methods based on ranks were used to identify ChIP-enriched regions. These methods make mild assumptions about the data distributions and are insensitive to outlying observations. A p-value was calculated for every assay probe on the array. The set of probes used in the calculation of this p-value was defined by a bandwidth parameter b. All probes centered on the chromosome at positions less than b bases 5' or 3' of the given probe position are included in this set.
The Wilcoxon rank sum test [26], also known as the Mann-Whitney U test, is the basis of the p-value statistic computed by the Affymetrix GTRANS software. The control and treatment observation sets are, respectively, the sets of normalized control and normalized treatment intensities from all replicates and all probes within the bandwidth. The null hypothesis is that the treatment set mean is no larger than that of the control set.
To take into account probe-to-probe variability we used a generalization of the Wilcoxon signed-rank test for blocked data. All input and IP normalized, sign(PM-MM)max(1,|PM-MM|) intensities (where PM are perfect match and MM are mismatched probes) interrogating the same chromosomal location were assigned to the same block. Aligned observations were derived by subtracting the median normalized intensity for a given block from each observation in that block. All aligned observations within the bandwidth were ranked. A statistic W was defined as the sum of the ranks of the aligned IP observations. A p-value was derived from W, based on the joint null distribution of the aligned input and IP ranks. The analyses depend on the assumption that probes are independent. Probes were mapped to the genomic coordinates to ensure that no probe mapped to more than one location in any 1,000-bp window and that no two probes map to the same genomic location.
RNA arrays
RNA samples were isolated from HeLa S3 cells and purified with trizol (Invitrogen) and RNeasy (Qiagen). RNA was amplified and hybridized to Affymetrix U133 Plus 2 arrays using standard methods. Three biological replicates were quantile normalized. Gene expression was indicated by the median of PM-MM values over all probes. The hypothesis of difference in gene expression between groups of genes, based on median PM-MM, was tested using the Wilcoxon rank sum statistic. For hybridization to the ENCODE tiled array, RNA was similarly isolated and double-stranded cDNA was generated using Invitrogen Superscript cDNA synthesis kit. cDNA (1-1.5 μg) was hybridized to the tiled array. Three biological replicates were performed for each RNA array.
Genomic annotation
Sites were determined to be near a genomic annotation if they were within the apparent 1,000 bp resolution. Sites shorter than 1,000 bp were scaled in size to include 1,000 bp around the center of the site. Sites that were longer than 1,000 bp used the data-determined length for their resolution size. Databases were downloaded from the University of California at Santa Cruz (UCSC) Golden Path Genome Browser and loaded into a local MySQL database. Exons were compared and classified as one or more of the following: start, terminal, alternatively spliced, constitutive or cassette. Because the arrays were designed using the HG15 assembly, the data were compared to this version of the human genome unless otherwise noted. The active gene list was defined as those with PolIIa at the first exon of the gene.
Real-time PCR
PCR primer pairs were designed to amplify 100-bp fragments from selected genomic regions (see Additional data file 8). Each real-time PCR reaction contained 50 nM primers, approximately 1 ng DNA and 1 × ABI SYBR PCR reaction mix. A fluorescence value proportional to the initial quantity of target DNA was calculated by a log-linear regression analysis for each quadruplicate amplification curve [27]. We normalized this value to an input chromatin sample, then normalized this ratio to a reference gene, PAPT, which is not expressed in HeLa cells, to calculate a relative enrichment value for the target ((TargetIP)/(TargetInp))/((PAPTIP)/(PAPTInput)).
Data availability
All data is present at Gene Expression Omnibus (GEO) at accession number GSE2735.
Additional data files
The following additional data are available with the online version of this paper. Additional data file 1 is a table listing PolIIa annotated to refGene. Additional data file 2 is a table listing PolIIa annotated to known genes. Additional data file 3 is a table listing PolIIa annotated to RefSeq. Additional data file 4 is a table listing PolII annotated to known genes. Additional data file 5 is a table listing PolII annotated to genscan exons. Additional data file 6 is a table listing knownGene and RefSeq populations on the ENCODE array. Additional data file 7 is a table listing the PolIIa-defined active gene list. Additional data file 8 is the PCR primer list and annotation.
Supplementary Material
Additional File 1
A table listing PolIIa annotated to refGene.
Click here for file
Additional File 2
A table listing PolIIa annotated to known genes.
Click here for file
Additional File 3
A table listing PolIIa annotated to RefSeq.
Click here for file
Additional File 4
A table listing PolII annotated to known genes.
Click here for file
Additional File 5
A table listing PolII annotated to genscan exons.
Click here for file
Additional File 6
A table listing knownGene and RefSeq populations on the ENCODE array.
Click here for file
Additional File 7
A table listing the PolIIa-defined active gene list.
Click here for file
Additional File 8
The PCR primer list and annotation.
Click here for file
Acknowledgements
We thank Pamela Hollasch, Maura Berkeley and the DFCI Affymetrix core for all their assistance, and Jason Carroll and Jessica Hurt for critical reading of the manuscript. We thank Adnan Derti for trying some splice-site strength analysis. This work was funded by a NHGRI K22 career award, HG02488-01A1 (A.S.B.), and a DOD grant DAMD17-02-0364 (P.A.S.).
Figures and Tables
Figure 1 Enrichment of selected genomic regions in ChIP. (a) PolII ChIP; (b) PolIIa ChIP. Real-time PCR relative enrichment ratios for selected regions are found to be enriched more often with p-values below 10-4. These regions include both intra- and intergenic locations as listed in Additional data file 8.
Figure 2 RNA polymerase II shows a variety of gene-specific enrichment patterns. Graphs plot 10log(p-value) mapped to chromosome position with the significant p-values greater than 40 indicated by the rectangle blocks below the graph. Values are plotted at every probe location. Flat lines indicate weak p-values and gaps indicate the absence of probes. The high density of probes across these genes suggest that the observed patterns are not due to probe bias. A scale bar is shown for each panel to reflect the different gene lengths displayed. RefSeq genes and knownGenes are annotated in green and blue, respectively, with thick bars representing exons and thin lines introns. Genes above the white bar are ordered 5' to 3', whereas those below the white bar are 3' to 5'. (a) On the highly expressed SF1 gene, PolIIa localizes to the first exon only. PolII accumulates across the gene with a distinctive pattern. (b) No significant signal is observed across the inactive NRXN2 locus which is near SF1 on chromosome 11. Graphs are plotted on the same scale as (a). (c) The moderately expressed gene KIAA1932 also reveals distinct accumulations across the gene. The red box highlights alternatively spliced exons. At the 3' end of the gene, some PolIIa signal is observed, probably indicative of the expression of a small gene antisense to KIAA1932. (d) Another commonly observed pattern is exemplified by the EHD1 gene. Both anti-polymerase antibodies recognize the first exon, but no other significant signal is observed across the gene until the 3' end.
Figure 3 Different RNA polymerase states show distinct exon biases. Pie charts representing the percentage of exons in each category at RNA polymerase enrichment locations. These include exons from enrichment locations that include more than one exon. PolIIa is strongly biased towards transcription initiation locations. Most of the internal exons are second exons overlapping with first exons. The phosphorylation-independent antibody recognizes PolII at both transcription initiation and termination locations with a slight bias towards termination locations.
Figure 4 Low p-value PolII and PolIIa enrichments are biased towards higher mRNA levels. The plot depicts the observed intensity from Affymetrix U133 Plus 2 chips compared with different p-values of PolII (white) and PolIIa (gray). Some genes with no significant PolII enrichment show high levels of observed intensity.
Figure 5 PolII enrichment is not always within annotated gene boundaries. Views are from the UCSC Genome Browser genome version HG16. PolIIa is in black and PolII is in blue with four rows for each, representing the data at different p-values: p < 10-5, p < 10-4, p < 10-3, and p < 10-2 from top to bottom. RNA signal in red. (a, b) PolII extending beyond the 3' end of the annotated gene. (c, d) PolII signal in putative intergenic regions with observed RNA signal also observed in the vicinity; (d) covers chromosome 11, positions 285,000-290,000. These regions are conserved and are also near predicted genscan exons. These novel sites not in the gene regions were confirmed by PCR.
Table 1 Summary of RNA polymerase II locations
Sites Pol IIa Pol II
Total sites 102 550
RefSeq total exons 70 289
RefSeq first exons 63 75
RefSeq terminal exons 2 91
RefSeq internal exons 5 123
RefSeq introns 4 120
knownGene exon 0 5
genscan exon 1 23
geneid or sgpGene 0 3
Active gene introns 2 57
Inactive introns 1 32
No RefSeq overlap 28 141
knownGene total exons 5 38
knownGene first exon 5 13
knownGene terminal exon 0 4
knownGene internal exon 0 21
No RefSeq or knownGene 23 90
genscan exons 7 43
geneid or sgpGene 2 6
The order indicates the flowchart of filtering through the different databases. Enrichment sites were first compared to the RefSeq database. Sites that are not near exons were then divided into two categories: locations that are in RefSeq introns; and locations that are not in a RefSeq gene. The latter are then compared with knownGene and predicted gene databases. For both RNA polymerase II phosphorylation states, the large majority of sites are near an exon.
==== Refs
Arndt KM Kane CM Running with RNA polymerase: eukaryotic transcript elongation. Trends Genet 2003 19 543 550 14550628 10.1016/j.tig.2003.08.008
Kornblihtt AR de la Mata M Fededa JP Munoz MJ Nogues G Multiple links between transcription and splicing. RNA 2004 10 1489 1498 15383674 10.1261/rna.7100104
Sims RJ 3rdBelotserkovskaya R Reinberg D Elongation by RNA polymerase II: the short and long of it. Genes Dev 2004 18 2437 2468 15489290 10.1101/gad.1235904
Cheng C Sharp PA RNA polymerase II accumulation in the promoter-proximal region of the dihydrofolate reductase and gamma-actin genes. Mol Cell Biol 2003 23 1961 1967 12612070 10.1128/MCB.23.6.1961-1967.2003
Dahmus ME Reversible phosphorylation of the C-terminal domain of RNA polymerase II. J Biol Chem 1996 271 19009 19012 8759772
Komarnitsky P Cho EJ Buratowski S Different phosphorylated forms of RNA polymerase II and associated mRNA processing factors during transcription. Genes Dev 2000 14 2452 2460 11018013 10.1101/gad.824700
Boehm AK Saunders A Werner J Lis JT Transcription factor and polymerase recruitment, modification, and movement on dhsp70 in vivo in the minutes following heat shock. Mol Cell Biol 2003 23 7628 7637 14560008 10.1128/MCB.23.21.7628-7637.2003
Kim M Ahn SH Krogan NJ Greenblatt JF Buratowski S Transitions in RNA polymerase II elongation complexes at the 3' ends of genes. EMBO J 2004 23 354 364 14739930 10.1038/sj.emboj.7600053
Ahn SH Kim M Buratowski S Phosphorylation of serine 2 within the RNA polymerase II C-terminal domain couples transcription and 3' end processing. Mol Cell 2004 13 67 76 14731395 10.1016/S1097-2765(03)00492-1
Hirose Y Tacke R Manley JL Phosphorylated RNA polymerase II stimulates pre-mRNA splicing. Genes Dev 1999 13 1234 1239 10346811
The ENCODE (ENCyclopedia Of DNA Elements) Project. Science 2004 306 636 640 15499007 10.1126/science.1105136
Dean FB Hosono S Fang L Wu X Faruqi AF Bray-Ward P Sun Z Zong Q Du Y Du J Comprehensive human genome amplification using multiple displacement amplification. Proc Natl Acad Sci USA 2002 99 5261 5266 11959976 10.1073/pnas.082089499
Cawley S Bekiranov S Ng HH Kapranov P Sekinger EA Kampa D Piccolboni A Sementchenko V Cheng J Williams AJ Unbiased mapping of transcription factor binding sites along human chromosomes 21 and 22 points to widespread regulation of noncoding RNAs. Cell 2004 116 499 509 14980218 10.1016/S0092-8674(04)00127-8
Sandoval J Rodriguez JL Tur G Serviddio G Pereda J Boukaba A Sastre J Torres L Franco L Lopez-Rodas G RNAPol-ChIP: a novel application of chromatin immunoprecipitation to the analysis of real-time gene transcription. Nucleic Acids Res 2004 32 e88 15247321 10.1093/nar/gnh091
Enriquez-Harris P Levitt N Briggs D Proudfoot NJ A pause site for RNA polymerase II is associated with termination of transcription. EMBO J 1991 10 1833 1842 2050120
Kim M Krogan NJ Vasiljeva L Rando OJ Nedea E Greenblatt JF Buratowski S The yeast Rat1 exonuclease promotes transcription termination by RNA polymerase II. Nature 2004 432 517 522 15565157 10.1038/nature03041
Shapiro MB Senapathy P RNA splice junctions of different classes of eukaryotes: sequence statistics and functional implications in gene expression. Nucleic Acids Res 1987 15 7155 7174 3658675
Zhang MQ Marr TG A weight array method for splicing signal analysis. Comput Appl Biosci 1993 9 499 509 8293321
Kampa D Cheng J Kapranov P Yamanaka M Brubaker S Cawley S Drenkow J Piccolboni A Bekiranov S Helt G Novel RNAs identified from an in-depth analysis of the transcriptome of human chromosomes 21 and 22. Genome Res 2004 14 331 342 14993201 10.1101/gr.2094104
Ujvari A Luse DS Newly Initiated RNA encounters a factor involved in splicing immediately upon emerging from within RNA polymerase II. J Biol Chem 2004 279 49773 49779 15377657 10.1074/jbc.M409087200
Roberts GC Gooding C Mak HY Proudfoot NJ Smith CW Co-transcriptional commitment to alternative splice site selection. Nucleic Acids Res 1998 26 5568 5572 9837984 10.1093/nar/26.24.5568
Robson-Dixon ND Garcia-Blanco MA MAZ elements alter transcription elongation and silencing of the fibroblast growth factor receptor 2 exon IIIb. J Biol Chem 2004 279 29075 29084 15126509 10.1074/jbc.M312747200
Ren B Cam H Takahashi Y Volkert T Terragni J Young RA Dynlacht BD E2F integrates cell cycle progression with DNA repair, replication, and G(2)/M checkpoints. Genes Dev 2002 16 245 256 11799067 10.1101/gad.949802
Kennedy GC Matsuzaki H Dong S Liu WM Huang J Liu G Su X Cao M Chen W Zhang J Large-scale genotyping of complex DNA. Nat Biotechnol 2003 21 1233 1237 12960966 10.1038/nbt869
Bolstad BM Irizarry RA Astrand M Speed TP A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 2003 19 185 193 12538238 10.1093/bioinformatics/19.2.185
Hollander M Wolfe DA Nonparametric Statistical Methods 1999 2nd New York: John Wiley
Ostermeier GC Liu Z Martins RP Bharadwaj RR Ellis J Draghici S Krawetz SA Nuclear matrix association of the human beta-globin locus utilizing a novel approach to quantitative real-time PCR. Nucleic Acids Res 2003 31 3257 3266 12799453 10.1093/nar/gkg424
|
16086846
|
PMC1273631
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 15; 6(8):R64
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r64
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r641608684610.1186/gb-2005-6-8-r64ResearchGenomic mapping of RNA polymerase II reveals sites of co-transcriptional regulation in human cells Brodsky Alexander S [email protected] Clifford A [email protected] Ian A [email protected] Giles [email protected] Benjamin J [email protected] Xiaole S [email protected] Edward A [email protected] Pamela A [email protected] Department of Systems Biology, Harvard Medical School and Department of Cancer Biology, Dana-Farber Cancer Institute, 44 Binney St, Boston, MA 02115, USA2 Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute and Harvard School of Public Health, Boston, MA 02155, USA3 Department of Medicine, Harvard Medical School and Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA 02115, USA2005 15 7 2005 6 8 R64 R64 4 1 2005 7 4 2005 17 6 2005 Copyright © 2005 Brodsky et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Determination of the distribution of RNA Polymerase II within regions of the human genome identifies novel sites of transcription and suggests that a major factor of transcription elongation control in mammals is the coordination of transcription and pre-mRNA processing to define exons.
Background
Transcription by RNA polymerase II is regulated at many steps including initiation, promoter release, elongation and termination. Accumulation of RNA polymerase II at particular locations across genes can be indicative of sites of regulation. RNA polymerase II is thought to accumulate at the promoter and at sites of co-transcriptional alternative splicing where the rate of RNA synthesis slows.
Results
To further understand transcriptional regulation at a global level, we determined the distribution of RNA polymerase II within regions of the human genome designated by the ENCODE project. Hypophosphorylated RNA polymerase II localizes almost exclusively to 5' ends of genes. On the other hand, localization of total RNA polymerase II reveals a variety of distinct landscapes across many genes with 74% of the observed enriched locations at exons. RNA polymerase II accumulates at many annotated constitutively spliced exons, but is biased for alternatively spliced exons. Finally, RNA polymerase II is also observed at locations not in gene regions.
Conclusion
Localizing RNA polymerase II across many millions of base pairs in the human genome identifies novel sites of transcription and provides insights into the regulation of transcription elongation. These data indicate that RNA polymerase II accumulates most often at exons during transcription. Thus, a major factor of transcription elongation control in mammalian cells is the coordination of transcription and pre-mRNA processing to define exons.
==== Body
Background
Transcriptional and post-transcriptional regulation of gene expression intersect at RNA polymerase II. The rate of polymerase II movement is altered by loading of transcription factors at the promoter, chromatin structure, pre-mRNA processing, elongation control and termination [1-3]. Thus, polymerase II accumulates at promoters as well as at different locations across a particular gene [4], but the general patterns across many different genes have yet to be explored. Numerous factors such as histones, post-translation modifying enzymes, and RNA-binding proteins regulate these processes [1,3]. One key determinant of transcription is the phosphorylation state of the carboxy-terminal domain (CTD) of polymerase II [5,6] which becomes hyperphosphorylated during transcription elongation [4,6-9]. Much of our understanding of transcription elongation comes from work in prokaryotes and yeast where most genes are intronless [1,3]. Transcription and pre-mRNA processing are coordinated, as the two processes affect the efficiency of each other [2,10]. The spatial patterns of the different phosphorylation states of polymerase II across genes remains poorly understood in mammalian systems.
Results and discussion
To explore the range of locations where polymerase II accumulates across the genome, we performed chromatin immunoprecipitation (ChIP) from HeLa S3 cells, and profiled the purified DNA using an oligonucleotide-tiled microarray interrogating the Encyclopedia of DNA Elements (ENCODE) regions [11] covering 471 known genes. Two antibodies were used, 8WG16 and 4H8, which recognize the hypophosphorylated (PolIIa) or a phosphorylation-independent state of the CTD of polymerase II (PolII), respectively. Thus, the 4H8 antibody is recognizing the total polymerase II population. Isolated DNA was amplified using a multiple displacement amplification (MDA) strategy (see Materials and methods) [12].
To identify sites of enrichment, we used a non-parametric approach generalizing the Wilcoxon signed-rank test [13]. Signals across 1,000 nucleotides were used to determine a p-value for each probe. Probes were filtered for uniqueness within the bandwidth. Probes with p-values below 10-4 were selected for further analysis because this threshold has a low false-positive rate as determined by PCR analysis (Figure 1). With these parameters, the hypophosphorylated-specific anti-PolIIa antibody reveals 102 occupied sites, whereas the phosphorylation-independent antibody shows 550 sites (Table 1).
RNA polymerase II has distinct landscapes across each gene. Figure 2 shows representative genes with polymerase enrichments. PolIIa is highly enriched at transcription initiation sites. On the other hand, PolII shows gene-specific landscapes with the strongest enrichments at exons within actively transcribed loci. Active genes reveal lower p-values across the gene compared with intergenic or inactive genes (compare Figure 2a and 2b), indicating a relative absence of polymerase II from the nontranscribed regions. Some smaller genes with high exon density, such as SF1, reveal significant polymerase signal across almost the entire locus (Figure 2a). Distinct accumulations are observed with significant p-values around exons for both SF1 and KIAA1932. In the KIAA1932 gene, PolII is enriched at a subset of constitutively and alternatively spliced exons (Figure 2c). For some genes, RNA polymerase II is enriched at relatively few locations within the gene (Figure 2d).
An important question is to determine if the polymerase II sites are indicative of active transcription. We addressed this in multiple ways. First, microarray expression profiling of the mRNA with Affymetrix U133 Plus 2 chips confirms that many of the RNA polymerase II-associated genes are actively expressed in HeLa cells, as seen in a plot of mRNA expression level versus p-value in Figure 3. Genes with significant RNA polymerase II enrichment are biased towards genes with higher mRNA levels. Figure 3 also shows that some genes have apparently high mRNA levels but no significant levels of PolII or PolIIa. This could be due to very low transcription levels but high mRNA stability. Second, we measured RNA from the same HeLa cells on the ENCODE tiled arrays. We observe that 34% of the PolII sites overlap with RNA signal (compared to approximately 8% expected at random) and 50% of the PolII locations are within 1 kb of some RNA signal (compared to 13% expected at random). Many sites where small pieces of RNA are synthesized, such as small exons, may be missed as a result of the spacing of the oligonucleotide probes and the imperfect nature of the probes. Third, many of the PolII and PolIIa sites overlap with annotated expressed sequence tags (ESTs) and mRNAs. Eighty-seven percent of the PolII-enriched and 88% of the PolIIa-enriched locations overlap with EST regions, compared to 31% and 44% expected at random, respectively. Lastly, reverse transcriptase PCR checks of KIAA1932 and DKC1 indicate that these genes are being expressed (data not shown). These data suggest that RNA polymerase II sites are biased towards regions of active transcription and that determining sites of enrichment of RNA polymerase II is an indicator of transcription.
Levels of RNA polymerase II enrichment at internal exons can vary between genes. To examine whether these patterns are influenced by expression levels, two categories were created: genes with multiple PolII enrichments at internal exons; and genes with PolII at one or zero internal exons. When compared to the mRNA levels, there is no significant difference between the two categories, suggesting that the number of PolII sites across the gene does not vary significantly with RNA levels. Genes with observable PolII enrichment at internal exons are correlated with higher mRNA levels on the expression array. This is consistent with reports proposing the use of PolII ChIP to monitor gene expression [14]. Therefore, the number of PolII sites at internal exons may reflect different levels of transcription elongation control and not just the sensitivity of the experiment.
Distinct from the hypophosphorylation-specific antibody, the phosphorylation-independent antibody reveals diverse enrichment locations for PolII. In total, 74% of the identified PolII locations are near an annotated knownGene, RefSeq, or genscan exon as summarized in Table 1 (see Additional data file 2 for a list of PolII genscan exon locations). Unlike PolIIa, PolII sites are distributed between the 5' and 3' ends of genes, with a slight bias towards terminal exons over initiating exons (Figure 4). This is probably reflecting the stalling of PolII during the coupled processes of transcription termination and 3'-end processing [15]. For some genes, significant PolII signal is observed more than 1 kb past the terminal exon, which might indicate transcription of the longer pre-mRNA before 3'-end cleavage and polyadenylation [16]. Figure 5 shows two representative genes with significant PolII enrichment past the terminal exon.
Most of the hypophosphorylated PolIIa locations at internal exons also overlap a transcription initiation site, as the internal exon in question is often the second exon in the gene. Only two enrichment sites overlap with an internal exon without also being near the first exon of a transcript. One of these is at a CpG island in the MCF2L gene and the other may be an alternative transcription initiation site as annotated in the HG17 assembly at the beginning of the ITGB4BP gene. To classify the remaining sites within introns or in intergenic regions, enrichment sites were compared to other gene databases. As summarized in Table 1, four PolIIa sites are in introns, but three of these are within resolution of annotated or predicted exons, leaving only one location not overlapping an exon of some kind. There are 28 hypophosphorylated polymerase sites not in a RefSeq gene region. After following a similar filtering approach, only 14 sites remain that are not near a putative exon. Thus, only 14% of PolIIa-enriched locations do not overlap with a known exon or actively transcribed region. Additional data file 2 lists PolIIa sites at predicted exons that are probably newly identified transcription initiation locations in HeLa cells. Figure 5 shows two examples of PolII and RNA signal at new sites of transcription. From the pattern of enrichments it is probable that many of these predicted exons are real and are transcription initiation locations, given the observed strong bias of the 8WG16 antibody for transcription initiation locations in well annotated genes.
To determine the generality of these observations, all RNA polymerase II occupancy sites were compared with the known genes and RefSeq databases, version HG16. PolIIa is highly enriched for the first exons around transcription initiation sites (Figure 4) representing 77 of 551 known genes in HG16 on the array (see Additional data file 1 for the entire lists).
Elongation control is a common transcriptional regulation mechanism believed to affect a wide range of functional gene classes [1]. In particular, RNA polymerase II pausing has been proposed to be associated with alternative splicing, [2]. To determine if there is a bias for alternative exons, we counted all the annotated alternatively spliced exons in the knownGene database and determined the distribution of PolII enrichment locations on them. PolII is enriched at 57% of the annotated alternatively spliced exons of the active genes compared to 37% of annotated actively transcribed constitutively expressed exons. We also examined the distribution of all PolII p-values on different types of exons. Each exon was mapped to the smallest p-value ChIP-enriched site that overlaps the exon. The cassette exons are found to be more significantly associated with smaller p-values compared to constitutively expressed exons according to the two-sample Kolmogorov-Smirnov test with a two sided p-value of less than 0.0035.
One attractive hypothesis is that sites of exon enrichment may reflect weaker splice sites where PolII stalls during splice site recognition. Using two different empirical methods to estimate splice site strength, no significant differences are observed between the exons overlapping PolII and those that do not [17,18]. Alternatively, some of the annotated constitutively expressed exons may actually be subject to alternative splicing decisions. Kampa et al. suggest that the levels of alternative splicing are much higher than commonly believed and annotated in the human genome from their examination of expression on tiled arrays [19]. Consistent with these findings, RNA polymerase II sites may be predicting which exons are being co-transcriptionally alternatively spliced.
To determine if there is any pattern for the 120 PolII enrichment sites that are in RefSeq introns, we compared these sites to knownGene, genscan, geneid, and sgpGene databases and find 31 within resolution of putative exons. Of the remaining 89, 57 are in genes with PolII enrichment sites that also overlap exons, suggesting that they are actively transcribed genes. No clear intronic positional bias is observed.
In conclusion, we have identified new sites of RNA polymerase II accumulation across hundreds of genes in mammalian cells. The large majority of polymerase II-enriched locations are at actively transcribed exons with a bias towards annotated alternatively spliced exons. Many of the PolII sites at annotated constitutively expressed exons may be sites of alternative splicing. Whatever the eventual splicing decision, these observations suggest that events around exons slow transcription elongation. A recent study suggests that even general splicing factors may slow elongation [20]. Stalling of RNA polymerase II near exons may function to slow RNA synthesis in order to wait for the competition of myriad splicing signals to be resolved in order to define the exon [21,22]. These ChIP data identify where these states of RNA polymerase II are localizing across the ENCODE regions.
Across genes, these data are consistent with the hypothesis of transcriptional pausing at particular locations. Alternatively, it is possible that RNA polymerase II is rearranging during transcription such that the epitope is only accessible around exons. Thus, the conformation of polymerase II may be changing and not the transcription rate. Nonetheless, it is interesting that the majority of observable elongating polymerase II accumulates around exons, suggesting that a major feature of transcription elongation control is coupling to pre-mRNA processing.
These observations differ from those observed in intronless genes typically found in prokaryotes and yeast where a more uniform PolII enrichment is observed across genes [16]. What appears to be conserved is PolII accumulation in coding regions compared to intronic regions. These data highlight the complexity and gene-specific nature of transcription regulation not only at transcription initiation and termination locations but at specific exons. Together, these observations suggest that a major feature of transcription elongation control in mammalian cells is exon definition. Thus, these data provide new insights into the coordination of transcription and pre-mRNA processing in mammalian cells.
Materials and methods
Chromatin immunoprecipitation and DNA amplification
Chromatin immunoprecipitations (ChIP) were performed as described with the following modifications [23]. HeLa S3 cells were first crosslinked with dimethyl adipimidate (DMA) (Pierce) for 10 min, washed with PBS and then crosslinked with formaldehyde for 10 min. Cells were collected, lysed, and chromatin was sheared by sonication to an average length of 1 kb as determined after RNase treatment of the samples on an agarose gel. Chromatin was prepared from four independently grown batches of cells and pooled to generate three replicate immunoprecipitations (IP) and six input samples. Briefly, 8WG16 (Covance) and 4H8 (AbCam) antibodies were incubated with a 50:50 mix of Dynal protein A/G beads for more than 16 h at 4°C in PBS with 5 mg/ml BSA. After washing in PBS, beads with bound antibody were incubated with chromatin from approximately 2 × 107 cells for more than 16 h at 4°C. Beads were washed eight times with RIPA buffer (50 mM HEPES pH 7.6, 1 mM EDTA, 0.7% DOC, 1% IGEPAL, 0.5 M LiCl) before DNA was eluted at 65°C in TE/1% SDS. Crosslinks were reversed by incubating at 65°C for more than 12 h followed by proteinase K treatment, phenol extraction and RNase treatment. Isolated DNA was then amplified isothermally using random nonamer primers and Klenow polymerase (Invitrogen) for more than 4 h, yielding approximately 2 μg of DNA per IP. DNA was prepared and hybridized on Affymetrix ENCODE oligonucleotide tiled arrays using the fragmentation, hybridization, staining and scanning procedure described by Kennedy et al. [24]. Affymetrix ENCODE microarrays have interrogating 25mer oligonucleotide probes tiled every 20 bp on average. A sample of chromatin was set aside before IP and used to represent the input DNA.
Tiled array analysis
Quantile normalization was used to make the distribution of probe intensities the same for all arrays [25]. In the case of the Affymetrix GTRANS software quantile normalization is used within treatment and control replicate sets. Non-parametric methods based on ranks were used to identify ChIP-enriched regions. These methods make mild assumptions about the data distributions and are insensitive to outlying observations. A p-value was calculated for every assay probe on the array. The set of probes used in the calculation of this p-value was defined by a bandwidth parameter b. All probes centered on the chromosome at positions less than b bases 5' or 3' of the given probe position are included in this set.
The Wilcoxon rank sum test [26], also known as the Mann-Whitney U test, is the basis of the p-value statistic computed by the Affymetrix GTRANS software. The control and treatment observation sets are, respectively, the sets of normalized control and normalized treatment intensities from all replicates and all probes within the bandwidth. The null hypothesis is that the treatment set mean is no larger than that of the control set.
To take into account probe-to-probe variability we used a generalization of the Wilcoxon signed-rank test for blocked data. All input and IP normalized, sign(PM-MM)max(1,|PM-MM|) intensities (where PM are perfect match and MM are mismatched probes) interrogating the same chromosomal location were assigned to the same block. Aligned observations were derived by subtracting the median normalized intensity for a given block from each observation in that block. All aligned observations within the bandwidth were ranked. A statistic W was defined as the sum of the ranks of the aligned IP observations. A p-value was derived from W, based on the joint null distribution of the aligned input and IP ranks. The analyses depend on the assumption that probes are independent. Probes were mapped to the genomic coordinates to ensure that no probe mapped to more than one location in any 1,000-bp window and that no two probes map to the same genomic location.
RNA arrays
RNA samples were isolated from HeLa S3 cells and purified with trizol (Invitrogen) and RNeasy (Qiagen). RNA was amplified and hybridized to Affymetrix U133 Plus 2 arrays using standard methods. Three biological replicates were quantile normalized. Gene expression was indicated by the median of PM-MM values over all probes. The hypothesis of difference in gene expression between groups of genes, based on median PM-MM, was tested using the Wilcoxon rank sum statistic. For hybridization to the ENCODE tiled array, RNA was similarly isolated and double-stranded cDNA was generated using Invitrogen Superscript cDNA synthesis kit. cDNA (1-1.5 μg) was hybridized to the tiled array. Three biological replicates were performed for each RNA array.
Genomic annotation
Sites were determined to be near a genomic annotation if they were within the apparent 1,000 bp resolution. Sites shorter than 1,000 bp were scaled in size to include 1,000 bp around the center of the site. Sites that were longer than 1,000 bp used the data-determined length for their resolution size. Databases were downloaded from the University of California at Santa Cruz (UCSC) Golden Path Genome Browser and loaded into a local MySQL database. Exons were compared and classified as one or more of the following: start, terminal, alternatively spliced, constitutive or cassette. Because the arrays were designed using the HG15 assembly, the data were compared to this version of the human genome unless otherwise noted. The active gene list was defined as those with PolIIa at the first exon of the gene.
Real-time PCR
PCR primer pairs were designed to amplify 100-bp fragments from selected genomic regions (see Additional data file 8). Each real-time PCR reaction contained 50 nM primers, approximately 1 ng DNA and 1 × ABI SYBR PCR reaction mix. A fluorescence value proportional to the initial quantity of target DNA was calculated by a log-linear regression analysis for each quadruplicate amplification curve [27]. We normalized this value to an input chromatin sample, then normalized this ratio to a reference gene, PAPT, which is not expressed in HeLa cells, to calculate a relative enrichment value for the target ((TargetIP)/(TargetInp))/((PAPTIP)/(PAPTInput)).
Data availability
All data is present at Gene Expression Omnibus (GEO) at accession number GSE2735.
Additional data files
The following additional data are available with the online version of this paper. Additional data file 1 is a table listing PolIIa annotated to refGene. Additional data file 2 is a table listing PolIIa annotated to known genes. Additional data file 3 is a table listing PolIIa annotated to RefSeq. Additional data file 4 is a table listing PolII annotated to known genes. Additional data file 5 is a table listing PolII annotated to genscan exons. Additional data file 6 is a table listing knownGene and RefSeq populations on the ENCODE array. Additional data file 7 is a table listing the PolIIa-defined active gene list. Additional data file 8 is the PCR primer list and annotation.
Supplementary Material
Additional File 1
A table listing PolIIa annotated to refGene.
Click here for file
Additional File 2
A table listing PolIIa annotated to known genes.
Click here for file
Additional File 3
A table listing PolIIa annotated to RefSeq.
Click here for file
Additional File 4
A table listing PolII annotated to known genes.
Click here for file
Additional File 5
A table listing PolII annotated to genscan exons.
Click here for file
Additional File 6
A table listing knownGene and RefSeq populations on the ENCODE array.
Click here for file
Additional File 7
A table listing the PolIIa-defined active gene list.
Click here for file
Additional File 8
The PCR primer list and annotation.
Click here for file
Acknowledgements
We thank Pamela Hollasch, Maura Berkeley and the DFCI Affymetrix core for all their assistance, and Jason Carroll and Jessica Hurt for critical reading of the manuscript. We thank Adnan Derti for trying some splice-site strength analysis. This work was funded by a NHGRI K22 career award, HG02488-01A1 (A.S.B.), and a DOD grant DAMD17-02-0364 (P.A.S.).
Figures and Tables
Figure 1 Enrichment of selected genomic regions in ChIP. (a) PolII ChIP; (b) PolIIa ChIP. Real-time PCR relative enrichment ratios for selected regions are found to be enriched more often with p-values below 10-4. These regions include both intra- and intergenic locations as listed in Additional data file 8.
Figure 2 RNA polymerase II shows a variety of gene-specific enrichment patterns. Graphs plot 10log(p-value) mapped to chromosome position with the significant p-values greater than 40 indicated by the rectangle blocks below the graph. Values are plotted at every probe location. Flat lines indicate weak p-values and gaps indicate the absence of probes. The high density of probes across these genes suggest that the observed patterns are not due to probe bias. A scale bar is shown for each panel to reflect the different gene lengths displayed. RefSeq genes and knownGenes are annotated in green and blue, respectively, with thick bars representing exons and thin lines introns. Genes above the white bar are ordered 5' to 3', whereas those below the white bar are 3' to 5'. (a) On the highly expressed SF1 gene, PolIIa localizes to the first exon only. PolII accumulates across the gene with a distinctive pattern. (b) No significant signal is observed across the inactive NRXN2 locus which is near SF1 on chromosome 11. Graphs are plotted on the same scale as (a). (c) The moderately expressed gene KIAA1932 also reveals distinct accumulations across the gene. The red box highlights alternatively spliced exons. At the 3' end of the gene, some PolIIa signal is observed, probably indicative of the expression of a small gene antisense to KIAA1932. (d) Another commonly observed pattern is exemplified by the EHD1 gene. Both anti-polymerase antibodies recognize the first exon, but no other significant signal is observed across the gene until the 3' end.
Figure 3 Different RNA polymerase states show distinct exon biases. Pie charts representing the percentage of exons in each category at RNA polymerase enrichment locations. These include exons from enrichment locations that include more than one exon. PolIIa is strongly biased towards transcription initiation locations. Most of the internal exons are second exons overlapping with first exons. The phosphorylation-independent antibody recognizes PolII at both transcription initiation and termination locations with a slight bias towards termination locations.
Figure 4 Low p-value PolII and PolIIa enrichments are biased towards higher mRNA levels. The plot depicts the observed intensity from Affymetrix U133 Plus 2 chips compared with different p-values of PolII (white) and PolIIa (gray). Some genes with no significant PolII enrichment show high levels of observed intensity.
Figure 5 PolII enrichment is not always within annotated gene boundaries. Views are from the UCSC Genome Browser genome version HG16. PolIIa is in black and PolII is in blue with four rows for each, representing the data at different p-values: p < 10-5, p < 10-4, p < 10-3, and p < 10-2 from top to bottom. RNA signal in red. (a, b) PolII extending beyond the 3' end of the annotated gene. (c, d) PolII signal in putative intergenic regions with observed RNA signal also observed in the vicinity; (d) covers chromosome 11, positions 285,000-290,000. These regions are conserved and are also near predicted genscan exons. These novel sites not in the gene regions were confirmed by PCR.
Table 1 Summary of RNA polymerase II locations
Sites Pol IIa Pol II
Total sites 102 550
RefSeq total exons 70 289
RefSeq first exons 63 75
RefSeq terminal exons 2 91
RefSeq internal exons 5 123
RefSeq introns 4 120
knownGene exon 0 5
genscan exon 1 23
geneid or sgpGene 0 3
Active gene introns 2 57
Inactive introns 1 32
No RefSeq overlap 28 141
knownGene total exons 5 38
knownGene first exon 5 13
knownGene terminal exon 0 4
knownGene internal exon 0 21
No RefSeq or knownGene 23 90
genscan exons 7 43
geneid or sgpGene 2 6
The order indicates the flowchart of filtering through the different databases. Enrichment sites were first compared to the RefSeq database. Sites that are not near exons were then divided into two categories: locations that are in RefSeq introns; and locations that are not in a RefSeq gene. The latter are then compared with knownGene and predicted gene databases. For both RNA polymerase II phosphorylation states, the large majority of sites are near an exon.
==== Refs
Arndt KM Kane CM Running with RNA polymerase: eukaryotic transcript elongation. Trends Genet 2003 19 543 550 14550628 10.1016/j.tig.2003.08.008
Kornblihtt AR de la Mata M Fededa JP Munoz MJ Nogues G Multiple links between transcription and splicing. RNA 2004 10 1489 1498 15383674 10.1261/rna.7100104
Sims RJ 3rdBelotserkovskaya R Reinberg D Elongation by RNA polymerase II: the short and long of it. Genes Dev 2004 18 2437 2468 15489290 10.1101/gad.1235904
Cheng C Sharp PA RNA polymerase II accumulation in the promoter-proximal region of the dihydrofolate reductase and gamma-actin genes. Mol Cell Biol 2003 23 1961 1967 12612070 10.1128/MCB.23.6.1961-1967.2003
Dahmus ME Reversible phosphorylation of the C-terminal domain of RNA polymerase II. J Biol Chem 1996 271 19009 19012 8759772
Komarnitsky P Cho EJ Buratowski S Different phosphorylated forms of RNA polymerase II and associated mRNA processing factors during transcription. Genes Dev 2000 14 2452 2460 11018013 10.1101/gad.824700
Boehm AK Saunders A Werner J Lis JT Transcription factor and polymerase recruitment, modification, and movement on dhsp70 in vivo in the minutes following heat shock. Mol Cell Biol 2003 23 7628 7637 14560008 10.1128/MCB.23.21.7628-7637.2003
Kim M Ahn SH Krogan NJ Greenblatt JF Buratowski S Transitions in RNA polymerase II elongation complexes at the 3' ends of genes. EMBO J 2004 23 354 364 14739930 10.1038/sj.emboj.7600053
Ahn SH Kim M Buratowski S Phosphorylation of serine 2 within the RNA polymerase II C-terminal domain couples transcription and 3' end processing. Mol Cell 2004 13 67 76 14731395 10.1016/S1097-2765(03)00492-1
Hirose Y Tacke R Manley JL Phosphorylated RNA polymerase II stimulates pre-mRNA splicing. Genes Dev 1999 13 1234 1239 10346811
The ENCODE (ENCyclopedia Of DNA Elements) Project. Science 2004 306 636 640 15499007 10.1126/science.1105136
Dean FB Hosono S Fang L Wu X Faruqi AF Bray-Ward P Sun Z Zong Q Du Y Du J Comprehensive human genome amplification using multiple displacement amplification. Proc Natl Acad Sci USA 2002 99 5261 5266 11959976 10.1073/pnas.082089499
Cawley S Bekiranov S Ng HH Kapranov P Sekinger EA Kampa D Piccolboni A Sementchenko V Cheng J Williams AJ Unbiased mapping of transcription factor binding sites along human chromosomes 21 and 22 points to widespread regulation of noncoding RNAs. Cell 2004 116 499 509 14980218 10.1016/S0092-8674(04)00127-8
Sandoval J Rodriguez JL Tur G Serviddio G Pereda J Boukaba A Sastre J Torres L Franco L Lopez-Rodas G RNAPol-ChIP: a novel application of chromatin immunoprecipitation to the analysis of real-time gene transcription. Nucleic Acids Res 2004 32 e88 15247321 10.1093/nar/gnh091
Enriquez-Harris P Levitt N Briggs D Proudfoot NJ A pause site for RNA polymerase II is associated with termination of transcription. EMBO J 1991 10 1833 1842 2050120
Kim M Krogan NJ Vasiljeva L Rando OJ Nedea E Greenblatt JF Buratowski S The yeast Rat1 exonuclease promotes transcription termination by RNA polymerase II. Nature 2004 432 517 522 15565157 10.1038/nature03041
Shapiro MB Senapathy P RNA splice junctions of different classes of eukaryotes: sequence statistics and functional implications in gene expression. Nucleic Acids Res 1987 15 7155 7174 3658675
Zhang MQ Marr TG A weight array method for splicing signal analysis. Comput Appl Biosci 1993 9 499 509 8293321
Kampa D Cheng J Kapranov P Yamanaka M Brubaker S Cawley S Drenkow J Piccolboni A Bekiranov S Helt G Novel RNAs identified from an in-depth analysis of the transcriptome of human chromosomes 21 and 22. Genome Res 2004 14 331 342 14993201 10.1101/gr.2094104
Ujvari A Luse DS Newly Initiated RNA encounters a factor involved in splicing immediately upon emerging from within RNA polymerase II. J Biol Chem 2004 279 49773 49779 15377657 10.1074/jbc.M409087200
Roberts GC Gooding C Mak HY Proudfoot NJ Smith CW Co-transcriptional commitment to alternative splice site selection. Nucleic Acids Res 1998 26 5568 5572 9837984 10.1093/nar/26.24.5568
Robson-Dixon ND Garcia-Blanco MA MAZ elements alter transcription elongation and silencing of the fibroblast growth factor receptor 2 exon IIIb. J Biol Chem 2004 279 29075 29084 15126509 10.1074/jbc.M312747200
Ren B Cam H Takahashi Y Volkert T Terragni J Young RA Dynlacht BD E2F integrates cell cycle progression with DNA repair, replication, and G(2)/M checkpoints. Genes Dev 2002 16 245 256 11799067 10.1101/gad.949802
Kennedy GC Matsuzaki H Dong S Liu WM Huang J Liu G Su X Cao M Chen W Zhang J Large-scale genotyping of complex DNA. Nat Biotechnol 2003 21 1233 1237 12960966 10.1038/nbt869
Bolstad BM Irizarry RA Astrand M Speed TP A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 2003 19 185 193 12538238 10.1093/bioinformatics/19.2.185
Hollander M Wolfe DA Nonparametric Statistical Methods 1999 2nd New York: John Wiley
Ostermeier GC Liu Z Martins RP Bharadwaj RR Ellis J Draghici S Krawetz SA Nuclear matrix association of the human beta-globin locus utilizing a novel approach to quantitative real-time PCR. Nucleic Acids Res 2003 31 3257 3266 12799453 10.1093/nar/gkg424
|
16086847
|
PMC1273632
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 27; 6(8):R65
|
latin-1
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r65
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r661608684810.1186/gb-2005-6-8-r66ResearchEvolution of selenium utilization traits Romero Héctor [email protected] Yan [email protected] Vadim N [email protected] Gustavo [email protected] Laboratorio de Organización y Evolución del Genoma, Dpto. de Biología Celular y Molecular, Instituto de Biología, Facultad de Ciencias, Iguá 4225, Montevideo, CP 11400, Uruguay2 Escuela Universitaria de Tecnología Médica, Facultad de Medicina, Piso 3 Hospital de Clínicas, Avda. Italia s/n, Montevideo, CP 11600, Uruguay3 Department of Biochemistry, University of Nebraska, Lincoln, NE 68588-0664, USA4 Cátedra de Inmunología, Facultad de Química/Ciencias, Instituto de Higiene, Avda. A. Navarro 3051, Montevideo, CP 11600, Uruguay2005 27 7 2005 6 8 R66 R66 20 4 2005 7 6 2005 27 6 2005 Copyright © 2005 Romero et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Completely sequenced genomes were analyzed for occurrence of SelA, B, C, D and ybbB genes. SelB and SelC were found to be signatures for the Sec decoding trait, while SelD defines the overall selenium utilization.
Background
The essential trace element selenium is used in a wide variety of biological processes. Selenocysteine (Sec), the 21st amino acid, is co-translationally incorporated into a restricted set of proteins. It is encoded by an UGA codon with the help of tRNASec (SelC), Sec-specific elongation factor (SelB) and a cis-acting mRNA structure (SECIS element). In addition, Sec synthase (SelA) and selenophosphate synthetase (SelD) are involved in the biosynthesis of Sec on the tRNASec. Selenium is also found in the form of 2-selenouridine, a modified base present in the wobble position of certain tRNAs, whose synthesis is catalyzed by YbbB using selenophosphate as a precursor.
Results
We analyzed completely sequenced genomes for occurrence of the selA, B, C, D and ybbB genes. We found that selB and selC are gene signatures for the Sec-decoding trait. However, selD is also present in organisms that do not utilize Sec, and shows association with either selA, B, C and/or ybbB. Thus, selD defines the overall selenium utilization. A global species map of Sec-decoding and 2-selenouridine synthesis traits is provided based on the presence/absence pattern of selenium-utilization genes. The phylogenies of these genes were inferred and compared to organismal phylogenies, which identified horizontal gene transfer (HGT) events involving both traits.
Conclusion
These results provide evidence for the ancient origin of these traits, their independent maintenance, and a highly dynamic evolutionary process that can be explained as the result of speciation, differential gene loss and HGT. The latter demonstrated that the loss of these traits is not irreversible as previously thought.
==== Body
Background
Selenium (Se) is an essential trace element for numerous organisms that belong to the three domains of life. The most relevant biological form of selenium is the rare amino acid selenocysteine (Sec), the selenium analog of cysteine (Cys). Sec is co-translationally incorporated into protein [1-3]. In functionally characterized selenoproteins, Sec is the catalytic group in the active site and is directly involved in redox catalysis. It is thought that Sec confers a functional advantage over cysteine at these active sites, increasing the catalytic efficiency of the enzymes [4]. Despite this selective advantage, the set of selenoproteins in any given organism is small [5,6].
Sec is inserted into selenoproteins at in-frame UGA codons (usually termination codons) by tRNASec (SelC) [2,7]. Interpretation of UGA as Sec requires translational reprogramming, which is provided by the Sec insertion sequence (SECIS) element, a cis-acting stem-loop structure present in the selenoprotein mRNA [2]. The decoding of Sec in bacteria also involves a Sec-specific elongation factor (SelB) which binds GTP, the SECIS element and the tRNASec [8,9]. In eukaryotes, this function is carried out by two proteins: EF-Sec and SECIS-binding protein (SBP2). EF-Sec is a Sec-specific elongation factor, distantly related to bacterial SelB, that binds GTP, tRNASec and SBP2; this latter protein, in turn, binds the SECIS element [10]. Sec synthesis is the other part of the metabolic pathway required for biosynthesis of selenoproteins. It takes place on tRNASec, which is first aminoacylated with serine (by a canonical seryl-tRNA synthetase) and then modified to selenocysteinyl-tRNA, in the reaction that uses selenophosphate as the selenium donor [9]. In the Bacteria, this reaction is catalyzed by Sec synthase (SelA). The functional equivalent of SelA in Archaea and Eukarya has not been described. A phosphoseryl-tRNASec kinase (PSTK) has been recently identified only in eukaryotic and archaeal Sec-incorporating organisms [11]. It has been suggested that this protein can play a role in Sec biosynthesis and/or regulation. The synthesis of selenophosphate is catalyzed by selenophosphate synthetase (SelD) from ATP and selenide in both prokaryotes and eukaryotes.
Selenophosphate has also been described as a precursor for the last step of the synthesis of the modified base 5-methylaminomethyl-2-selenouridine in the wobble position of the anticodons of Lys, Glu and Gln tRNAs [12], and this reaction was reported to be catalyzed by YbbB in Escherichia coli [13]. The function of this modified base is not known.
Thus, considerable efforts in recent years have been made to elucidate molecular details of Sec decoding in the three domains of life. In addition, the selenoproteome of several species has been the subject of intensive research [5,6,14-16]. Despite this progress, fundamental issues relating to the evolution of Sec utilization remain unclear. On the basis of the complexity and similarity of the Sec-insertion mechanisms in different organisms, it has been proposed that the Sec-decoding trait arose once, before the division of the three domains of life, and was subsequently lost in some lineages. It is also thought that once an organism has lost the Sec-insertion system, it cannot re-emerge. Whether the Sec biosynthesis and insertion pathway evolved before the fixation of the genetic code or whether this was a late addition is not known [17-19].
Here we provide a map of Sec-incorporating and selenouridine-utilizing organisms within the tree of life, based on the analysis of completely sequenced genomes. From phylogenetic analysis of all components of the Sec-decoding machinery, we present clear evidence for the loss of the trait in many lineages at different taxonomic levels, and examples of acquisition of the trait by horizontal gene transfer (HGT). In addition, we describe and explain the maintenance of selenophosphate synthetase in non-Sec-incorporating organisms, and use this information to define a selenouridine-utilization trait as well as a general selenium-utilization trait. We find that the 2-selenouridine pathway can also be acquired by HGT. These data suggest that the loss of selenium utilization is not irreversible.
Results
A map of selenium utilization within the tree of life
Figure 1 displays a phylogenetic tree, based on rRNA, of the 155 species whose entire genomes have been sequenced (see Materials and methods for the rationale behind the use of rRNA and other alternatives). The criteria for the occurrence of the Sec-decoding trait included the presence of known genes involved in Sec decoding (that is, selA, selB, selC, selD), and at least one gene encoding a known selenoprotein in the genome, inferred by the presence of a UGA codon within a coding region (at the location corresponding to Cys in homologs) followed by a downstream SECIS element. Using these criteria, a total of 29 bacterial and three archaeal species were found to be Sec decoding. These criteria were in agreement with experimental evidence when available. A map of selA, selB, selC and selD within the species tree is provided in Figure 1.
Despite the bias among the prokaryotic genomes so far sequenced [20], in which proteobacteria are over-represented and some phyla are not represented at all, the taxa distribution of Sec incorporation revealed interesting features of this trait. First, the trait is widely distributed and present in numerous bacterial phyla (Proteobacteria, Firmicutes, Spirochaetes, Actinobacteria, Aquificae). Second, we observed the presence and absence of the trait in taxa within monophyletic groups. This phenomenon takes place within clades at different evolutionary levels, namely phylum, class, order, family, genera, and even species and is illustrated in Figure 1. Considering the genus level, we observed this phenomenon within Pseudomonas, Treponema, Clostridium and Yersinia. The more revealing case of this absence/presence pattern is that of the KIM, biovar Mediaevalis and CO92 strains of Yersinia pestis. Whereas the KIM strain possesses a functional Sec-decoding machinery, the CO92 and Mediaevalis strains carry a confirmed sequenced selB pseudogene, whose coding region is disrupted by a frameshift [21]. Furthermore, the CO92 and the Mediaevalis strains possess selA, selC and selD, indicating that the loss of the ability to decode Sec is very recent. Moreover, at position 203 of the α-subunit of formate dehydrogenase type O there is a UGA codon in the three strains, which is decoded as Sec by the KIM strain, but could not be decoded as such in the CO92 and the Mediaevalis strains. The three strains also possess formate dehydrogenase type H with a Cys-containing α-subunit (fdhH).
SelB and selC can be considered as the gene signature of organisms able to decode Sec: their presence in genomes always coincides with that of selA (excluding archaeal and eukaryote domains), selD, and selenoproteins. A putative ortholog of selA is present in Helicobacter pylori (a Sec-non-incorporating organism). The presence of this protein in the two strains of this species is intriguing and raises the question of whether this protein serves a different function or is just a remnant of the Sec-decoding machinery. The case of selD is different, because it is present in several species that lack other genes necessary for Sec decoding. SelD orthologs are indicated in red in Figure 1.
Bacteria possessing selD but not the Sec-decoding trait include Bordetella bronchiseptica, Pseudomonas syringae, Porphyromonas gingivalis, Nitrosomonas europaea, Bdellovibrio bacteriovorus and Enterococcus faecalis. In addition, two cyanobacteria - Prochlorococcus marinus and Synechococcus species - possess a putative selD homolog with a 320-amino-acid amino-terminal extension with similarity to NADH dehydrogenases. All selDs from non-Sec-incorporating bacteria, excluding those of cyanobacteria, are likely to be 'true orthologs' to selD from Sec-incorporating bacteria because the topology of the selD phylogeny parallels the topology of species for both Sec-incorporating and non-Sec-incorporating species (Figure 2d), and because the sequence signatures of bacterial selD are present and are of similar length (see Additional data files). Although it is difficult to sketch an evolutionary history for the selD from cyanobacteria, it is clear that these proteins have many features of selDs and could be viewed as true selenophosphate synthetases (see below).
The fact that selenophosphate is also the precursor for the synthesis of 2-selenouridine [12], a modified base that is present at the wobble position of Lys, Glu and Gln tRNAs, suggests that selD may have been maintained in these organisms to generate selenophosphate for 2-selenouridine synthesis. Thus, we investigated the distribution of ybbB, a gene encoding the catalyst of the last step of 2-selenouridine synthesis [13], and its association with selD.
A search across genomes for ybbB (indicated in gray in Figure 1) revealed that six out of seven selD-containing and non-Sec-decoding species also contained ybbB. In addition, all ybbB-containing organisms also possess selD, including cyanobacteria. Furthermore, in most of these species, except P. gingivalis and cyanobacteria, both genes are located contiguously and arranged in an operon (Figure 3), as has been previously suggested [13]. This gene organization is also seen in some species that incorporate Sec and possess ybbB: selD is contiguous to ybbB in some genomes, but is rarely contiguous to the selA-selB operon (Figure 3).
The analysis of selA, selB, selC, selD and ybbB genes also revealed that, within the set of species that incorporate Sec, many, but not all, organisms, possess ybbB and vice versa. In other words, the set of species that incorporate Sec into protein overlaps with, but is different from, the set of species that possess ybbB (Figure 1). It is important to note that a low-identity homolog to bacterial ybbB is present in Methanococcus jannaschii and Methanopyrus kandleri, and absent in other archaea, suggesting that this base modification might not be unique to bacteria.
Finally, we investigated the presence of additional genes linked to the selenouridine synthesis trait by searching genomes for genes that occur in organisms possessing ybbB and are absent in organisms lacking ybbB. This search did not identify any additional gene associated with this trait. Thus, the overall analysis allows us to corroborate that the two products of these genes form a pathway with 2-selenouridine in the tRNA as the final product. However, only ybbB is the gene signature of this trait. On the other hand, the dual use of selenophosphate (for Sec decoding and 2-selenouridine biosynthesis) makes selD a signature of a broader trait of selenium utilization, and our data suggest that both Sec decoding and selenouridine traits are independently maintained, but both require selD.
Phylogeny of selA, selB, selC, selD and ybbB: evidence of horizontal gene transfer (HGT) of Sec-decoding and selenouridine synthesis traits
The phylogenies of selA, selB, selC, selD and ybbB shown in Figure 2 are neither mutually coherent nor match the 'species tree' (Figure 1). This does not necessarily imply an error in the phylogenetic reconstruction since the evolutionary history of each gene could be different. Many nodes are mutually consistent across different methods and have high statistical support. Certain anomalous situations occur with distantly related organisms (deep nodes), which could be due to the limitations of these analyses. However, some of the inconsistencies may be considered as 'genuine' and raise HGT as the most likely alternative explanation.
A striking observation is the clustering of P. profundum (a γ-proteobacterium) with T. denticola (a spirochete) at a basal position of the selA, selB and selC trees. This topology is consistent in various phylogenetic reconstruction methods and has high statistical support in all cases. The congruence of the trees sustains the idea that these genes were horizontally transferred to P. profundum. Several facts provide further support for this proposition. The P. profundum genome encodes four selenoproteins: two glycine reductases A, one glycine reductase B and selenophosphate synthetase. This selenoproteome is entirely distinct from that of γ-proteobacteria and very similar to that of T. denticola, which consists of glycine reductase A, two glycine reductases B, selenophosphate synthetase, glutathione peroxidase and thioredoxin. Furthermore, glycine reductase is absent in every other proteobacterial genome. In addition, P. profundum is the single prokaryotic genome that has two selDs: one encodes a Sec-containing isoform that is located next to the selAB operon, on chromosome II; the second encodes a Cys-containing enzyme that is adjacent to ybbB on chromosome I. The phylogeny of selD places the Cys isoform within the γ-proteobacterial clade as expected according to the organismal phylogeny, whereas the Sec isoform does not cluster with T. denticola or with γ-proteobacteria. Altogether, these results indicate that it is highly unlikely that P. profundum has acquired the Sec-decoding trait by vertical descent, raising HGT as the obvious alternative. In addition, we analyzed the codon usage of selA and selB, looking for an anomalous pattern, using the method described by García-Vallvé [22]. These genes do not display biased values of codon usage with respect to the rest of the genes. This result could indicate that P. profundum has already adapted the codon usage of these genes to its internal values. A recent paper suggested that the compatible codon usage between foreign genes and recipient genomes increases the probability of HGT [23]. Since T. denticola and P. profundum do not share the same environment, it is likely that the HGT took place from a species of spirochete, a bacterial phylum exhibiting great variability in habitat and physiology [24].
An additional incongruence is observed when the selA, selB, selC and the species trees are compared among Pseudomomas spp. (γ-proteobacteria), Sinorhizobium meliloti (α-proteobacterium) and Burkholderia spp. (β-proteobacteria) (Figure 2). The evolutionary history of these genes is, however, difficult to solve.
Conflicts relating to the selenouridine synthesis trait were also observed. The consistent and statistically supported cluster between Bordetella bronchiseptica (a β-proteobacterium) and Pseudomonas spp., within the γ-proteobacteria clade in both selD and ybbB gene trees, strongly suggests an event of HGT from Pseudomonas spp. to B. bronchiseptica. A situation that cannot be explained by vertical descent is also the cluster of Nitrosomonas europae (a β-proteobacterium) and Bdellovibrio bacteriovorus (δ-proteobacteria) in selD and ybbB phylogenies (Figure 2). The location of ybbB and selD genes also supports this possibility: while arranged in an operon in N. europeae and B. bacteriovorus; they are distant in the genomes of the other δ-proteobacteria (Figure 3). Furthermore, B. bacteriovorus is a predatory bacterium with a multiplication phase within many Gram-negative bacteria [25]. Thus, the ready access to the prey's genetic information and vice versa might be a possible explanation for this HGT event.
Discussion
The distribution of the Sec-decoding trait within the 'species tree' prompts the question of how it evolved. A supported conclusion from our data is the common origin of selA, selB, selC and selD in the domain Bacteria. This is based on the absence of close paralogs for Sec-decoding genes in bacteria, the high bootstrap value for the bacterial node in all phylogenies, and the presence of bacterial sequence signatures in selA, selB, selC and selD sequences (see Additional data files). The phylogenies of selB, selC and selD also indicate that the archaeal and eukaryal Sec-decoding genes cluster together. This is further supported by the similar overall organization of the Sec-decoding machinery in Archaea and Eukarya [26-28]. The emergence of the Sec-decoding trait before the division of the three domains has been previously postulated [18,29]. The evolution of the Sec insertion system only once is certainly the most parsimonious evolutionary scenario. However, this does not necessarily imply that every gene involved in Sec-decoding has a common origin. This is exemplified by selA: no clear ortholog has been found in Archaea and Eukarya. This suggests that the mechanism of Sec biosynthesis and insertion could have been adjusted during evolution.
Assuming the common origin of the Sec-decoding trait, it is possible to sketch a scenario compatible with our results in order to explain the pattern of presence/absence of the Sec-decoding trait. We propose that this pattern is the result of two mechanisms, primarily speciation and differential gene loss, with some contribution from HGT. Regarding the selenouridine synthesis trait, the results also suggest a common origin in the bacterial domain, as well as the possibility that 2-selenouridine pathway can be acquired by HGT.
An important issue in the evolution of Se utilization traits relates to the selective forces operating to maintain, loose or acquire the traits. Although it is not possible to draw conclusions, the search for a common biochemical, physiological or ecological trait in organisms possessing/lacking either or both traits provides interesting clues. The analysis of the prokaryotic selenoproteome revealed that formate dehydrogenase is present in most organisms capable of Sec decoding, exceptions being T. denticola, P. profundum, Clostridium perfringens and Thermoanerobacter tengcongensis [6]. Formate dehydrogenase plays a key role in anaerobic respiration. Indeed, most of these species are obligatory anaerobes or facultative aerobes; the sole exception was S. meliloti, a symbiotic nitrogen-fixing obligatory aerobe that lives in the oxygen-limited environment of the nodule [30]. Formate dehydrogenase is the single Sec-containing polypeptide encoded in the Sinorhizobium meliloti genome [6,30], suggesting that the presence of the trait may be important for respiration under conditions of restricted oxygen supply. On the other hand, glycine reductase is present in T. denticola, P. profundum and T. tengcongensis and several species of the genera Clostridium except C. perfringens. Glycine reductase is an enzymatic complex that allows certain anaerobic bacteria to conserve energy via a soluble substrate level phosphorylation system [31]. Sec is more reactive than Cys by virtue of the lower pKa and higher nucleophilicity of selenol group compared to that of the thiol group [12], and can increase the pH range at which certain enzymes are active [32]. This might have conferred a selective advantage improving catalytic efficiency of proteins.
Regarding selective forces operating on the evolution of the selenouridine synthesis trait, we begin from the fact that synthesis of 2-selenouridine is carried out exclusively at the wobble position (first of the anticodon) of the tRNAs for lysine, glutamate and glutamine (the only amino acids encoded by twofold purine-ending codons). Several modifications of this base have been reported to be essential for correct decoding; thiouridine, in particular, would convert the base into an ionized form that would favor pairing with A and G, and avoid pairing with U or C, contributing to the discrimination of twofold codons ending in purine from those ending in pyrimidine [33]. The low pKa value of 2-selenouridine of these tRNAs would be consistent with this argument and it has been suggested that this would also favor base-pairing with G [34]. Thus, we postulate that selenium modification of tRNAs matching twofold codons might be a refinement in the base discrimination at the wobble position. The interaction of the first base of the anticodon with the third base of the codon plays an important role in the efficiency and accuracy of the translation process, suggesting that this base modification could be linked to certain aspects of codon usage. In any case, it should be stressed that ybbB null E. coli has no apparent phenotypic differences to wild type-E. coli and does not alter nonsense suppression phenotype [13].
One of the driving forces for the loss of the traits probably relates to the variability of selenium abundance in the environment. The absolute dependence of organisms on Se can compromise their existence if dietary Se becomes limiting. In these situations, enzymes containing Sec as catalytic residues could have evolved into Cys-containing proteins or, alternatively, both Sec-containing and Cys-containing forms could be maintained. This latter case is exemplified by the genome of M. maripaludis, which encodes several Sec-containing proteins and also homologs that contain cysteine in place of Sec. In a medium that contains adequate amounts of selenium, this organism represses the synthesis of the cysteine homologs, but this repression is not observed in a mutant with disrupted selB [35], suggesting that the cysteine homologs are a backup system in case of selenium scarcity. Nevertheless, the existence of organisms carrying only one of the selenium-utilization traits suggests that selenium availability might not be the sole factor involved in the loss of either trait. It is also possible that the higher reactivity of selenium over sulfur in biological molecules might have had a role in counterselecting the pervasive use of Sec and/or selenouridine in living systems.
Conclusion
This paper provides an organismal map for Sec-decoding and 2-selenouridine synthesis traits within the tree of life, and defines selB and selC as the gene signature of the Sec-decoding trait, ybbB as the gene signature of selenouridine synthesis, with selD defining overall selenium utilization. We show that the set of species that incorporate Sec overlaps with, yet is distinct from, the set of species that synthesize 2-selenouridine, and our data suggest that Sec decoding and 2-selenouridine traits can be independently maintained, and both require selD.
Analysis of the phylogenies of the Sec-decoding and 2-selenouridine synthesis genes provides evidence for the ancient origin of these traits and demonstrates that their evolution is a highly dynamic process that occurs at different evolutionary levels, namely phylum, class, order, family, genera, and even species. We show that this process can be explained as the result of speciation and differential gene loss, and provide conclusive evidence that the loss of these traits is not irreversible as previously thought, and that entire sets of genes can be acquired by HGT. It is striking that the genetic code of an organism and the amino-acid repertoire can be 'laterally' expanded.
The study of selenium-utilization traits, which directly associate protein synthesis with a discrete set of genes, can contribute to the understanding of basic questions regarding the evolution of the genetic code and the translation machinery.
Materials and methods
Sequences of selA, selB, selC, selD and ybbB
Complete genome sequences of 194 prokaryotes were retrieved from GenBank [36] as of 20 October 2004, representing 151 species.
Annotated sequences corresponding to selA, selB, selD and ybbB prokaryotic genes were retrieved from GenBank, and used as queries to perform local BLAST searches across a database generated with the 194 genomes. For selA, selD and ybbB, hits with an e-value below e-15 were recovered; for selB, the cutoff e-value was e-30 to decrease the number of hits corresponding to other translation factors. A total of 242 selB, 48 selA, 47 selD and 25 ybbB sequences were recovered. The sequences were aligned using ClustalW [37], a raw phylogenetic analysis was conducted and clear nonorthologous sequences were discarded. This dataset was manually curated, and the number of sequences was reduced to 29 selA, 32 selB, 41 selD, and 21 ybbB sequences. These datasets were used as queries for BLAST searches but no new sequences were identified. Finally, only one sequence by strain was included and the set was supplemented with sequences of three representative eukaryotes (Caenorhabditis elegans, Drosophila melanogaster and Homo sapiens).
Most sequences of selC were retrieved from GenBank or identified using tRNAscan software with default parameters [38,39]. The sequences of Wolinella succinogenes, Helicobacter hepaticus, Burkholderia mallei and Thermoanaerobacter tencongensis, were found changing the parameters to 'Cove-only search mode' and lowering the tRNA Cove cutoff score to 6.
All sequences are provided aligned in the Additional data files.
Alignments and phylogenetic reconstruction of gene trees
Several phylogenetic gene trees were built using different inference methods performed on different sequence alignments. Sequences were aligned with T-coffee version 1.37 [40], ClustalW 1.8 and Dialign-2 [41] using different parameters. The 'score' option of the T-coffee software was enabled to assess alignment quality. The alignments were then visually inspected, compared and uncertain sites were removed. In another approach, we applied the g-blocks software [42] to remove unstable blocks with 2 different sets of parameters. The final alignment sets were the following: i) raw alignments using each software with two different sets of parameters ii) 'sub-alignments' obtained removing the unstable blocks from the raw alignments using g-blocks software, and iii) 'sub-alignment' obtained using the '-score' option of T-coffee for evaluation of the alignment, then the low scoring regions were removed manually.
For each of these alignments, we applied several phylogenetic reconstruction methods including Neighbor Joining using MEGA software [43], Maximum Likelihood (ML) using phyML 2.4 software [44] and Bayesian approaches using MrBayes 3.0b4 [45]. For each of these methods, different transition matrices (WAG and JTT) and evolutionary models were tested. In total, more than 80 trees were analyzed for each gene. The gene trees presented in Figure 2 were built using the T-coffee alignment evaluated with the '-score' option and manually refined. The ML and Bayesian trees were built using WAG matrix and gamma+invar model of evolutionary change. In the ML method, the assessment of node reliability was done using 100 bootstrap replicates. In the case of Bayesian analyses, four heated Markov chains were started from random trees and run for 1,000,000 generations each. Chains were sampled every 500 generations to assure independence. Sample points prior to reach stationary (200) were discarded as 'burn-in'.
Almost all trees yielded similar topologies and, more important, all of them supported the conclusions. In particular, the HGT results were reproduced with any of the alignments and phylogenetic trees.
Species tree
Different species trees were initially constructed, based on small-subunit (SSU) rRNA, EF-Tu/EF-1α (a highly conserved translation elongation factor present in all organisms), and a concatenated set of 9 ortholog sequences present in all prokaryotes. A set of aligned SSU-rRNA sequences was retrieved from the Ribosomal Database Project (RDP) [46], release 2.1, missing sequences were retrieved from Genbank and aligned against the set from the RDP using the profile option of ClustalW.
EF-Tu and EF-1α were recovered using a similar approach to that described for selA, selB, selD and ybbB, and aligned using T-coffee. An all-against-all BLAST search was performed sequentially and best reciprocal hits were identified as putative orthologs. A set of nine genes was obtained. These sequences were aligned with ClustalW and concatenated. In all cases we used the neighbor-joining (NJ) algorithm to build the trees from different distance matrices using MEGA software [43]. In the case of SSU-rRNA, the Tamura-Nei (TN93) distance with pairwise-deletion was calculated. For the amino-acid alignments we use the JTT transition matrix.
The different 'species trees' display some discrepancies. In any case, the conclusions drawn are maintained with any of the above mentioned 'species tree'. The SSU-rRNA was finally adopted because it is by large the commonest used, and trees inferred for this gene are sound descriptors of the general evolutionary history of prokaryotes. This tree also recovers the major groups described in Bergey's Manual [47].
Codon bias analyses
Codon bias was evaluated according to the method described in [22]. This method uses the Mahalanobis distance measure for detecting outliers in a multivariate distribution.
Search for additional genes linked to Se-U trait
To study the possible association of a certain gene with ybbB we run an all-against-all BLAST search with an e-value threshold of e-10 among organisms carrying ybbB, to pick up homologs present in all these genomes. Then we used this set of genes to run a new BLAST search against a control set of closely related species lacking a ybbB homolog. This search detected no gene. When we excluded Sec-decoding species from the control set, we were able to recover a single gene: selD.
Additional data files
Additional data are available with the online version of this paper. Additional data file 1 is a table containing the gene locations of selA, selB, selC selD and ybbB in the genomes analyzed in this work. Additional data file 2 contains the sequence alignments of selA, selB, selC selD and ybbB of the genomes analyzed in this work.
Supplementary Material
Additional data file 1
A table containing the gene location of selA, selB, selC selD and ybbB in the genomes analyzed in this work
Click here for file
Additional data file 2
Sequence alignments of selA, selB, selC selD and ybbB of the genomes analyzed in this work.
Click here for file
Acknowledgements
This work was supported by Fogarty International Research Collaboration Award TW006959 and NIH GM061603 grant to V.N.G. We thank Dr Alexey Lobanov for help in identifying tRNASec sequences and Héctor Musto (Universidad de la República, Uruguay) for critical reading of the manuscript. We also thank the faculty, teaching assistants and students of the 'Workshop on Molecular Evolution' 2004 at the Marine Biological Laboratory, Woodshole, attended by HR, for valuable general discussion.
Figures and Tables
Figure 1 Distribution of selenium-utilization traits. The figure depicts the species tree for all organisms completely sequenced so far, based on the phylogenetic reconstruction using the small subunit rRNA sequences and is in good agreement with other consensus phylogenetic trees. Species able to decode Sec are those possessing selA (yellow) (excluding Archaea), selB (blue), selC (green) and selD (red). The presence of ybbB (gray) and selD indicates the ability to synthesize 2-selenouridine.
Figure 2 Phylograms of selA, selB, selC, selD and ybbB. Phylograms were inferred using phyML2.4 from curated T-Coffee alignments. The values above and below each branch indicate boostrap values (>70) of maximum likelihood analysis and posterior probabilities (>0.90) of Bayesian analysis respectively. In all trees γ-proteobacteria are highlighted in red, β-proteobacteria in blue and Sinorhizobium meliloti (α-proteobacteria) in green. Red circles denote putative horizontal gene transfer events.
Figure 3 Genome location of selA, selB, selC, selD and ybbB. Each bar represents one replicon of a species. On the vertical axis the species name, phylum, and domain are specified. The horizontal axis corresponds to the replicon size. Location of selA (yellow), selB (blue), selC (green), selD (red) and ybbB (black) is indicated; arrows denote direction of transcription.
==== Refs
Low SC Berry MJ Knowing when not to stop: selenocysteine incorporation in eukaryotes. Trends Biochem Sci 1996 21 203 208 8744353 10.1016/0968-0004(96)10025-6
Hatfield DL Gladyshev VN How selenium has altered our understanding of the genetic code. Mol Cell Biol 2002 22 3565 3576 11997494 10.1128/MCB.22.11.3565-3576.2002
Driscoll DM Copeland PR Mechanism and regulation of selenoprotein synthesis. Annu Rev Nutr 2003 23 17 40 12524431 10.1146/annurev.nutr.23.011702.073318
Lee SR Bar-Noy S Kwon J Levine RL Stadtman TC Rhee SG Mammalian thioredoxin reductase: oxidation of the C-terminal cysteine/selenocysteine active site forms a thioselenide, and replacement of selenium with sulfur markedly reduces catalytic activity. Proc Natl Acad Sci USA 2000 97 2521 2526 10688911 10.1073/pnas.050579797
Kryukov GV Castellano S Novoselov SV Lobanov AV Zehtab O Guigo R Gladyshev VN Characterization of mammalian selenoproteomes. Science 2003 300 1439 1443 12775843 10.1126/science.1083516
Kryukov GV Gladyshev VN The prokaryotic selenoproteome. EMBO Rep 2004 5 538 543 15105824 10.1038/sj.embor.7400126
Leinfelder W Zehelein E Mandrand-Berthelot MA Bock A Gene for a novel tRNA species that accepts L-serine and cotranslationally inserts selenocysteine. Nature 1988 331 723 725 2963963 10.1038/331723a0
Forchhammer K Rucknagel KP Bock A Purification and biochemical characterization of SELB, a translation factor involved in selenoprotein synthesis. J Biol Chem 1990 265 9346 9350 2140572
Bock A Biosynthesis of selenoproteins - an overview. Biofactors 2000 11 77 78 10705967
Copeland PR Driscoll DM Purification, redox sensitivity, and RNA binding properties of SECIS-binding protein 2, a protein involved in selenoprotein biosynthesis. J Biol Chem 1999 274 25447 25454 10464275 10.1074/jbc.274.36.25447
Carlson BA Xu XM Kryukov GV Rao M Berry MJ Gladyshev VN Hatfield DL Identification and characterization of phosphoseryl-tRNA[Ser]Sec kinase. Proc Natl Acad Sci U S A 2004 101 12848 12853 15317934 10.1073/pnas.0402636101
Stadtman TC Selenocysteine. Annu Rev Biochem 1996 65 83 100 8811175 10.1146/annurev.bi.65.070196.000503
Wolfe MD Ahmed F Lacourciere GM Lauhon CT Stadtman TC Larson TJ Functional diversity of the rhodanese homology domain: the Escherichia coli ybbB gene encodes a selenophosphate-dependent tRNA 2-selenouridine synthase. J Biol Chem 2004 279 1801 1809 14594807 10.1074/jbc.M310442200
Castellano S Morozova N Morey M Berry MJ Serras F Corominas M Guigo R In silico identification of novel selenoproteins in the Drosophila melanogaster genome. EMBO Rep 2001 2 697 702 11493597 10.1093/embo-reports/kve151
Lescure A Gautheret D Carbon P Krol A Novel selenoproteins identified in silico and in vivo by using a conserved RNA structural motif. J Biol Chem 1999 274 38147 38154 10608886 10.1074/jbc.274.53.38147
Martin-Romero FJ Kryukov GV Lobanov AV Carlson BA Lee BJ Gladyshev VN Hatfield DL Selenium metabolism in Drosophila: selenoproteins, selenoprotein mRNA expression, fertility, and mortality. J Biol Chem 2001 276 29798 29804 11389138 10.1074/jbc.M100422200
Jukes TH Genetic code 1990. Outlook. Experientia 1990 46 1149 1157 2147658
Bock A Forchhammer K Heider J Baron C Selenoprotein synthesis: an expansion of the genetic code. Trends Biochem Sci 1991 16 463 467 1838215 10.1016/0968-0004(91)90180-4
Gladyshev VN Kryukov GV Evolution of selenocysteine-containing proteins: significance of identification and functional characterization of selenoproteins. Biofactors 2001 14 87 92 11568444
Venter JC Remington K Heidelberg JF Halpern AL Rusch D Eisen JA Wu D Paulsen I Nelson KE Nelson W Environmental genome shotgun sequencing of the Sargasso Sea. Science 2004 304 66 74 15001713 10.1126/science.1093857
Deng W Burland V Plunkett G 3rdBoutin A Mayhew GF Liss P Perna NT Rose DJ Mau B Zhou S Genome sequence of Yersinia pestis KIM. J Bacteriol 2002 184 4601 4611 12142430 10.1128/JB.184.16.4601-4611.2002
Garcia-Vallve S Romeu A Palau J Horizontal gene transfer in bacterial and archaeal complete genomes. Genome Res 2000 10 1719 1725 11076857 10.1101/gr.130000
Medrano-Soto A Moreno-Hagelsieb G Vinuesa P Christen JA Collado-Vides J Successful lateral transfer requires codon usage compatibility between foreign genes and recipient genomes. Mol Biol Evol 2004 21 1884 1894 15240837 10.1093/molbev/msh202
Seshadri R Myers GS Tettelin H Eisen JA Heidelberg JF Dodson RJ Davidsen TM DeBoy RT Fouts DE Haft DH Comparison of the genome of the oral pathogen Treponema denticola with other spirochete genomes. Proc Natl Acad Sci USA 2004 101 5646 5651 15064399 10.1073/pnas.0307639101
Rendulic S Jagtap P Rosinus A Eppinger M Baar C Lanz C Keller H Lambert C Evans KJ Goesmann A A predator unmasked: life cycle of Bdellovibrio bacteriovorus from a genomic perspective. Science 2004 303 689 692 14752164 10.1126/science.1093027
Fagegaltier D Hubert N Yamada K Mizutani T Carbon P Krol A Characterization of mSelB, a novel mammalian elongation factor for selenoprotein translation. EMBO J 2000 19 4796 4805 10970870 10.1093/emboj/19.17.4796
Hubert N Sturchler C Westhof E Carbon P Krol A The 9/4 secondary structure of eukaryotic selenocysteine tRNA: more pieces of evidence. RNA 1998 4 1029 1033 9740122 10.1017/S1355838298980888
Foster CB Selenoproteins and the metabolic features of the archaeal ancestor of eukaryotes. Mol Biol Evol 2005 22 383 386 15483329 10.1093/molbev/msi007
Rao M Carlson BA Novoselov SV Weeks DP Gladyshev VN Hatfield DL Chlamydomonas reinhardtii selenocysteine tRNA[Ser]Sec. RNA 2003 9 923 930 12869703 10.1261/rna.5510503
Barnett MJ Fisher RF Jones T Komp C Abola AP Barloy-Hubler F Bowser L Capela D Galibert F Gouzy J Nucleotide sequence and predicted functions of the entire Sinorhizobium meliloti pSymA megaplasmid. Proc Natl Acad Sci USA 2001 98 9883 9888 11481432 10.1073/pnas.161294798
Andreesen JR Glycine reductase mechanism. Curr Opin Chem Biol 2004 8 454 461 15450486 10.1016/j.cbpa.2004.08.002
Gromer S Johansson L Bauer H Arscott LD Rauch S Ballou DP Williams CH JrSchirmer RH Arner ES Active sites of thioredoxin reductases: why selenoproteins? Proc Natl Acad Sci USA 2003 100 12618 12623 14569031 10.1073/pnas.2134510100
Takai K Yokoyama S Roles of 5-substituents of tRNA wobble uridines in the recognition of purine-ending codons. Nucleic Acids Res 2003 31 6383 6391 14602896 10.1093/nar/gkg839
Ching WM Characterization of selenium-containing tRNAGlu from Clostridium sticklandii. Arch Biochem Biophys 1986 244 137 146 2418784 10.1016/0003-9861(86)90102-5
Rother M Mathes I Lottspeich F Bock A Inactivation of the selB gene in Methanococcus maripaludis: effect on synthesis of selenoproteins and their sulfur-containing homologs. J Bacteriol 2003 185 107 114 12486046 10.1128/JB.185.1.107-114.2003
GenBank
Thompson JD Higgins DG Gibson TJ CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Nucleic Acids Res 1994 22 4673 4680 7984417
Lowe TM Eddy SR tRNAscan-SE: A program for improved detection of transfer RNA genes in genomic sequence. Nucl Acids Res 1997 25 955 964 9023104 10.1093/nar/25.5.955
tRNAscan
Notredame C Higgins DG Heringa J T-Coffee: A novel method for fast and accurate multiple sequence alignment. J Mol Biol 2000 302 205 217 10964570 10.1006/jmbi.2000.4042
Morgenstern B DIALIGN 2: improvement of the segment-to-segment approach to multiple sequence alignment. Bioinformatics 1999 15 211 218 10222408 10.1093/bioinformatics/15.3.211
Castresana J Selection of conserved blocks from multiple alignments for their use in phylogenetic analysis. Mol Biol Evol 2000 17 540 552 10742046
Kumar S Tamura K Jakobsen IB Nei M MEGA2: molecular evolutionary genetics analysis software. Bioinformatics 2001 17 1244 1245 11751241 10.1093/bioinformatics/17.12.1244
Guindon S Gascuel O A simple, fast, and accurate algorithm to estimate large phylogenies by maximum likelihood. Syst Biol 2003 52 696 704 14530136 10.1080/10635150390235520
Ronquist F Huelsenbeck JP MrBayes 3: Bayesian phylogenetic inference under mixed models. Bioinformatics 2003 19 1572 1574 12912839 10.1093/bioinformatics/btg180
Cole JR Chai B Farris RJ Wang Q Kulam SA McGarrell DM Garrity GM Tiedje JM The Ribosomal Database Project (RDP-II): sequences and tools for high-throughput rRNA analysis. Nucleic Acids Res 2005 33 D294 D296 15608200 10.1093/nar/gki038
Taxonomic Outline of the Prokaryotes
|
16086848
|
PMC1273633
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 27; 6(8):R66
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r66
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r671608684910.1186/gb-2005-6-8-r67ResearchPatterns of intron sequence evolution in Drosophila are dependent upon length and GC content Haddrill Penelope R [email protected] Brian [email protected] Daniel L [email protected] Peter [email protected] Institute of Evolutionary Biology, School of Biological Sciences, University of Edinburgh, Edinburgh EH9 3JT, UK2 Section of Ecology, Behavior and Evolution, Division of Biological Sciences, University of California San Diego, La Jolla, CA 92093, USA2005 27 7 2005 6 8 R67 R67 4 3 2005 25 4 2005 29 6 2005 Copyright © 2005 Haddrill et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
An analysis of inter-specific divergence in 225 intron fragments in Drosophila melanogaster and D. simulans reveals a strongly negative correlation between intron length and divergence and intron divergence and GC content. This suggests that most intronic DNA is evolving under considerable constraint.
Background
Introns comprise a large fraction of eukaryotic genomes, yet little is known about their functional significance. Regulatory elements have been mapped to some introns, though these are believed to account for only a small fraction of genome wide intronic DNA. No consistent patterns have emerged from studies that have investigated general levels of evolutionary constraint in introns.
Results
We examine the relationship between intron length and levels of evolutionary constraint by analyzing inter-specific divergence at 225 intron fragments in Drosophila melanogaster and Drosophila simulans, sampled from a broad distribution of intron lengths. We document a strongly negative correlation between intron length and divergence. Interestingly, we also find that divergence in introns is negatively correlated with GC content. This relationship does not account for the correlation between intron length and divergence, however, and may simply reflect local variation in mutational rates or biases.
Conclusion
Short introns make up only a small fraction of total intronic DNA in the genome. Our finding that long introns evolve more slowly than average implies that, while the majority of introns in the Drosophila genome may experience little or no selective constraint, most intronic DNA in the genome is likely to be evolving under considerable constraint. Our results suggest that functional elements may be ubiquitous within longer introns and that these introns may have a more general role in regulating gene expression than previously appreciated. Our finding that GC content and divergence are negatively correlated in introns has important implications for the interpretation of the correlation between divergence and levels of codon bias observed in Drosophila.
==== Body
Background
Non-coding DNA makes up a large proportion of the genomes of most eukaryotes, yet little is known about its functional significance and the forces affecting its evolution. The identification of functional regions of the genome has tended to concentrate on coding DNA, yet the recent shift in focus towards non-coding DNA has revealed that introns and intergenic sequences may be subject to considerable levels of selective constraint, implying that they contain functional elements [1-6]. No consistent patterns have emerged from the relatively few studies that have thus far investigated levels of constraint on intron DNA sequences; some studies conclude that such DNA is evolving under little or no selective constraint, while others find considerable levels of constraint (for examples, see [3,7-10]). Moreover, the mode of evolution for such types of sequence is still unclear.
Several recent studies have attempted to estimate the proportion of sites within introns that is subject to selective constraint. For example, Jareborg et al. [11] estimate that 23% of intronic sites in mouse-rat genome comparisons are evolutionarily conserved. Similarly, Shabalina and Kondrashov [12] estimate (conservatively) that 17% of nucleotide sites within introns are selectively constrained between Caenorhabditis elegans and Caenorhabditis briggsae; this was at least in part due to their function in splicing, because constraint appeared to be higher at the edges of introns. Likewise, Bergman and Kreitman [3] estimate that 22-26% of non-coding sequences (intergenic and intronic) are highly constrained between Drosophila melanogaster and Drosophila virilis. In contrast to these studies, Halligan et al. [9] found that most intronic sites (excluding those necessary for correct splicing) in Drosophila were evolving approximately 17% faster than fourfold synonymous sites. They concluded that these sites were effectively evolving free from selective constraint. The discrepancies among previous studies suggest that no clear conclusions can yet be drawn regarding the levels of selective constraint in non-coding intronic DNA.
Intron size is one possible factor that may explain these conflicting results. Comeron and Kreitman [13] and others have noted an asymmetrical distribution of intron lengths in D. melanogaster; a large number of short introns clustered around a minimal intron length and a broader distribution of longer introns (median intron size of 86 base-pairs (bp), mean intron size of 1411 bp; [14]). Based on multi-species data for 15 introns (13 short and 2 long), Parsch [15] showed that there were significantly fewer substitutions per site in the two longer introns. He suggested that this pattern may be due to the presence of a greater number of regulatory elements that are subject to purifying selection in longer introns.
If regulatory elements occur frequently in introns, and these are of some minimal size, it follows that size may be an important factor in intron evolution. In agreement with this prediction, Marais et al. [16] noted a marginally significant (P = 0.03) negative correlation between intron divergence and size for first introns (but not other introns) in the dataset of Halligan et al. [9]. Marais et al. [16] suggested that this correlation between divergence and length may be expected for first introns because they are on average two times longer than other introns [17] and also tend to contain more known regulatory elements, at least in mammals [8]. Because the dataset used consisted mostly of short introns, it is unclear whether the pattern they observed is specific to first introns (due to an association between first introns and regulatory elements) and whether the relationship between divergence and size is primarily driven by the fact that first introns are longer. Here we revisit the relationship between intron length and evolutionary constraint (as measured by levels of divergence between D. melanogaster and D. simulans) by combining published data for 225 intron fragments sampled from a much broader distribution of intron lengths and positions within genes.
Results and discussion
Levels of divergence are correlated with intron length
We investigated levels of divergence at a total of 225 introns (a mixture of complete short introns and several hundred base-pair fragments of longer introns) scattered across the Drosophila genome. The relationship between intron length and nucleotide divergence for all complete introns and intron fragments surveyed is shown in Figure 1. A strongly negative correlation between intron length and divergence is apparent (Spearman correlation coefficient Rs = -0.388, P < 10-4). We also divided the data into two size classes based on the median intron size of 86 bp in Drosophila [14]; small (≤86 bp) introns and large (>86 bp) introns. The large intron class showed significantly lower divergences than the small intron class (Wilcoxon two-sample test statistic W = 17079.5, P < 10-4). The correlation between intron length and divergence is somewhat weaker, but still significant within the longer intron class (Rs = -0.278, P = 0.006).
It has been noted that introns harbouring regulatory elements tend to be first introns [6,8], and that first introns tend to be longer in Drosophila [17]. Thus a relationship between intron size and divergence might only be expected for first introns [16]. Indeed, previous studies have failed to find evidence of constraint outside first introns [16,18]. In Figure 1, we show that the strong correlation between divergence and intron length is not specific to first introns (first introns Rs = -0.451, P < 10-4; non-first introns Rs = -0.304, P < 10-4). Mean divergences were not significantly different between first and non-first introns when compared within short and long size classes (Table 1). These results suggest that regulatory elements may be common enough across all longer introns that constraint is independent of the position of an intron within a gene.
While this is strong evidence for evolutionary constraint on longer introns, short introns do not appear to evolve much more slowly than synonymous sites in Drosophila. To illustrate this, Figure 2 shows average divergence estimates (with two standard errors) for synonymous sites from 102 coding regions [19] compared to those for the small (≤86 bp) and large (>86 bp) size classes of introns. Average divergence at non-synonymous sites [19] is also shown for comparison. Synonymous site divergence is significantly higher than levels of divergence for large introns (Wilcoxon two-sample W = 7745.5, P < 10-4) but not small introns (Wilcoxon two-sample W = 15115.5, P = 0.617). This finding is consistent with the conclusions of Halligan et al. [9] that introns and synonymous sites evolve at similar rates, given that their dataset contained few long introns. One half of the introns in the genome are less than 86 base-pairs long, but these comprise only about 5% of total intronic DNA in the genome [14]. Thus, ironically, while the majority of introns in the Drosophila genome may be evolving under little or no selective constraint, most intronic DNA in the genome is likely to be evolving under considerable constraint.
Divergence and base composition of introns
Introns are more AT-rich than synonymous sites in Drosophila [20] (Table 1). Could lower levels of divergence then be an artefact of local GC content? There is a significantly negative relationship between divergence and GC content in the intron dataset (Rs = -0.345, P < 10-4) (Figure 3a), and a significantly positive relationship between intron length and GC content (Rs = 0.237, P < 10-3) (Figure 3b). The partial correlation coefficient for divergence versus length, controlling for GC content, is -0.132 (95% bootstrap confidence interval: -0.192/-0.089). The partial correlations for divergence versus GC content (controlling for length) and GC content versus length (controlling for divergence) were -0.292 (-0.410/-0.168) and 0.030 (-0.037/0.120), respectively. These results suggest that the relationship between intron length and divergence is not a confounding effect of GC content, despite the negative correlation between divergence and GC content.
Similar to the pattern we observe in introns, a negative association between synonymous site substitution rates and GC content at the third position of codons has previously been noted in Drosophila [21] and in mammals [22]. This pattern at synonymous sites has been cited as evidence of selection for codon usage bias, as preferred codons are usually GC rich [21,23]; however, selection on codon usage obviously cannot explain the same pattern in introns. The negative relationship between divergence and GC content in introns might instead reflect local variation in the extent of mutational rates or biases [22,24], or the effects of biased gene conversion favouring GC over AT, which mimics the effect of selection in favour of GC nucleotides [25].
The possible role of mutational bias can be examined using the following method. It follows from the standard model of drift and reversible mutation that, if AT mutates to GC at rate u and GC mutates to AT at rate ku the equilibrium frequency of GC for neutral sites (neglecting polymorphic sites) is approximated by p = 1/(1 + k), and the equilibrium rate of substitutions is K = 2uk/(1+k) [26,27]. This yields the relation K = 2u(1 - p), so that the equilibrium rate of substitution is negatively and linearly related to GC content. This formula predicts that the intercept (divergence at zero GC content) is equal to the absolute value of the slope, and so this hypothesis is testable. The regression coefficient of divergence on GC content in the complete dataset is -0.180 (-0.254/-0.106), and the corresponding intercept is 0.157 (0.115/0.163), which at first sight is consistent with the hypothesis that variation in the level of the mutational bias parameter, k, is sufficient to account for the relation between divergence and GC content.
The relationship between divergence and length, however, makes the above test problematic, in view of the wide variation in intron length. If only the 127 short introns (length ≤ 86 bp) are used, which are much more uniform in length, the regression of divergence on GC content is almost unchanged at -0.116 (-0.207/-0.023), and the intercept is 0.150 (0.142/0.162). Note, however, that there is a significant partial correlation of 0.166 (0.041/0.345) between GC content and length for short introns, but not for long introns, so there is still a residual relation between length and GC content in short introns. While we cannot rule out the possibility that biased gene conversion and/or selection in favour of GC versus AT explains the relationship between GC content and divergence, our analysis suggests that variation in mutational bias may be sufficient. If this process also explains the relationship between synonymous site divergence and GC content, tests for selection on codon bias based on negative correlations between codon bias and divergence (recently discussed by Bierne and Eyre-Walker [28] and Dunn et al. [29]) lose their force. These have been criticized on other theoretical grounds by Eyre-Walker and Bulmer [26].
The density of functional elements in introns
The correlation analyses strongly suggest that longer introns show lower levels of divergence, and that this is not simply caused by mutational rate differences related to GC content, although other sources of mutation rate differences cannot of course be ruled out. So why might longer introns be subject to higher levels of constraint? Introns are known to contain regulatory elements (for examples, see [30,31], and see [32] for a recent review of the mammalian literature), so it is possible that longer introns are more constrained because they contain more of these elements.
Are putative regulatory elements in longer introns discrete entities (such as clusters of binding sites for transcription factors), or is this regulatory function more diffuse? If intronic regulatory elements occur in clusters, surrounded by unconstrained regions, we might expect to find higher levels of divergence in the short, several hundred base-pair regions of very long introns (such as those surveyed here), compared to intermediate-sized introns, provided that they have similar total amounts of regulatory sequences. The rationale for this is that, if constrained regulatory elements are clustered into one region, short fragments of very long introns would be unlikely to coincide by chance with a functional element, whereas similarly sized regions from introns of intermediate length would be more likely to coincide with such elements. Such clustering is possible, given that transcription factor binding sites and regulatory elements can range in size from a few base-pairs up to several hundred base-pairs (for examples, see [33-36]). If the proportion of regulatory sequence is similar in long and intermediate introns, however, no difference in mean divergence is expected, but clustering would cause a higher variance in divergence in very long versus intermediate-length introns (after removing the binomial sampling variance). If regulatory elements in introns are widely dispersed, however, there is no reason to expect greater means or variances of divergence in fragments from very long introns. In fact, the mean divergence for the small number of intron fragments from introns longer than 4,500 bp is 0.054 (SE = 0.004, n = 9). This is significantly smaller than for the small (≤86 bp) intron class (mean divergence = 0.110, n = 127, Wilcoxon two-sample W = 252, P = 0.001) and marginally significantly lower than for introns of intermediate size (between 87 bp and 4,500 bp: mean divergence = 0.072, n = 89, W = 4494, P = 0.044). The non-binomial standard deviation in divergence is estimated to be 0.0056 for the very long introns, compared with 0.023 for the 38 intermediate-sized ones for which fragments at least 20 bp shorter than the introns were used for estimating divergence (this ensures that both classes represent samples rather than complete sequences). This is the opposite pattern to what is expected with strong clustering of regulatory sequences. Levels of constraint, and thus the density of putatively funtional regulatory elements, therefore appear to be relatively uniform across longer introns.
A uniform density of regulatory functions is unexpected if these often involve clusters of, for example, transcription factor binding sites. However, it might be expected, for example, if the regulatory functions of introns often involve the formation of complex secondary structures. Evidence suggesting that intron sequence and length affects the secondary structure of precursor messenger RNA (pre-mRNA) is accumulating. If this secondary structure plays a regulatory role, it is likely to be conserved. Several studies have found evidence for epistatic selection on introns to maintain pre-mRNA secondary structure [37-39], and there is also evidence for a functional role of RNA secondary structure in splicing [40,41] and gene expression [42,43]. For example, Chen and Stephan [44] found that mutations disrupting a hairpin structure in intron 1 of the D. melanogaster Adh gene reduce splicing efficiency and decrease production of the Adh protein. These authors show that compensatory mutations that restore the secondary structure result in a mutant indistinguishable from the wild type in splicing efficiency and protein production. A hairpin structure in the second intron of this gene also shows striking structural conservation across ten species in three sub-genera of Drosophila [45]. Our finding that the density of constrained sequences does not appear to be a function of intron length (within the long intron class) suggests that pre-mRNA secondary structure may be a more common mechanism mediating gene regulation than discrete regulatory elements such as intronic transcriptional enhancers.
Conclusion
Most introns in Drosophila are relatively short, but these short introns make up only a small fraction of total intronic DNA in the genome. We demonstrate that levels of selective constraint are higher with increasing intron length. Thus, while the majority of introns in the Drosophila genome may be evolving under little or no selective constraint, the majority of intronic DNA in the genome is likely to be evolving under considerable constraint. We also find that the density of functionally important elements within longer introns does not appear to depend on their length. This suggests that functional elements may be ubiquitous within longer introns and that these introns may have a more general role in regulating gene expression than previously appreciated, possibly via the formation of pre-mRNA secondary structures. This pattern contrasts with that found in mammals, where constraint does not appear to be a function of intron length [46] (A Kondrashov, personal communication). An unexpected corollary of our study is the finding of a negative correlation between divergence and GC content in introns. This finding implies that a similar pattern observed for synonymous sites in Drosophila may reflect mutational biases rather than selection for codon usage.
Materials and methods
Introns
We combined data from three recent studies of complete introns or several hundred base-pair fragments of longer introns located on the X chromosome of D. melanogaster. Halligan et al. [9] compiled previously published data for D. melanogaster and D. simulans sequences for each of 163 introns. We combined these data with introns surveyed in D. melanogaster and D. simulans by Glinka et al. [47]. All the Glinka et al. [47] intron fragments were compared to the DNA sequence of the D. melanogaster genome [48]. Ten of these intron fragments were removed from the analysis because they contained exonic or 5'/3' untranslated region sequences. The alignments for a further 12 of the Glinka et al. [47] fragments were trimmed to remove small quantities of exonic or untranslated region sequences. The final Glinka et al. [47] dataset used in the analysis therefore contained 53 intron fragments (details on request to PR Haddrill). To this we added nine more intron fragments surveyed by Haddrill et al. [49]. For consistency with Halligan et al. [9], we realigned these sequences with the program MCALIGN, using the insertion-deletion frequency model defined for Drosophila intronic DNA [50,51]. Divergence estimates per site and the GC content of introns were generated for each alignment (excluding the 6 bp/16 bp at the 5'/3' end of the intron, which include bases that are constrained because they are necessary for correct splicing) using the DnaSP software package (Version 4) [52], which corrects divergence values for multiple hits using the Jukes-Cantor equation [53]. The use of divergence as a proxy for constraint is appropriate, because the level of selective constraint in a sequence will directly affect the divergence between two species; highly constrained sequences will show little divergence, whereas sequences under little or no selective constraint will accumulate differences more rapidly. Sites overlapping alignment gaps were excluded from the count of total base-pairs. The total length of each intron was determined using the DNA sequence of the D. melanogaster genome [48]. The mean total intron length across the entire dataset was 936.5 bp and the mean length of the fragments of introns analyzed here was 230.2 bp.
Because we did not analyse the entire length of all of the introns included in this study, we were unable to investigate whether intron lengths vary substantially between D. melanogaster and D. simulans. Previous evidence suggests that intron lengths are unlikely to differ to any great extent between the two species, however, and that transitions between the short and long intron size class are rare [15,20,54].
Partial moment correlation coefficients and least-squares regression coefficients were calculated by the standard formulae, and their significance assessed by bootstrapping over loci 1,000 times to obtain their resampling distributions [55].
Coding regions
As a comparison for levels of divergence at intron sites, we used synonymous site divergences from 102 genes compiled by Betancourt and Presgraves [19]. Single-pass sequenced ESTs from this same study were not included in the analysis. Estimates of synonymous site divergences calculated using the Nei and Gojobori [56] correction were kindly provided by A Betancourt. Divergence estimates for synonymous sites based on D. melanogaster - D. simulans alignments for 35 additional X-linked coding regions were identical, and did not differ significantly from divergence estimates for fourfold degenerate sites (P Andolfatto, unpublished data). Several previous studies have documented a positive relationship between exon length and synonymous site divergences [57-59]. This relationship is in the opposite direction to that which would be expected if there were some (unknown) factor co-varying with gene length and neutral divergence that was responsible for the negative association between intron length and intron divergence. Non-synonymous site divergences from the same 102 genes compiled by Betancourt and Presgraves [19] (kindly provided by A Betancourt) were also used in Figure 2 for visual comparison with synonymous and intron sites; as expected, these are smaller than the other values, consistent with strong selection against most amino acid substitutions.
Effects of sex linkage
As our data come from three different sources, we investigated possible biases relating to how and why the data were collected. In particular, the studies of Haddrill et al. [49] and Glinka et al. [47] surveyed intron fragments from longer introns on the X chromosome, whereas the data of Halligan et al. [9] contains mostly short introns from all chromosomes. We note a significant difference between autosomal versus X-linked introns in both levels of divergence (Wilcoxon two-sample W = 13502.5, P = 0.006) and GC content (W = 13211.5, P = 0.005). When comparing within size classes (≤86 bp versus >86 bp), however, levels of divergence are not significantly different between autosomal and X-linked introns, and GC content is significantly different for the short intron class, but not the long intron class. The negative correlation between intron length and divergence holds for autosomal and X-linked introns separately (autosomes, Spearman Rs = -0.261, P = 0.006; X-linked, Spearman Rs = -0.403, P < 10-4) as does the negative relationship between GC content and divergence (autosomes, Spearman Rs = -0.281, P = 0.003; X-linked, Spearman Rs = -0.371, P < 10-4). The differences in levels of divergence and GC content between autosomal and X-linked introns, therefore, cannot explain the observed relationships between intron length versus divergence and GC content versus divergence.
Additional data files
The following additional data are available with the online version of this paper. Additional data file 1 is an Excel file listing all introns analyzed. Additional data files 2, 3 and 4 conatain alignments of the Glinka et al. [47], Haddrill et al. [49] and Halligan et al. [9] data, respectively. Additional data file 5 contains programs written to carry out partial moment correlations, least-squares regressions and bootstrapping procedures and the data used for these analyses.
Supplementary Material
Additional File 1
An Excel file listing all introns analyzed
Click here for file
Additional File 2
Alignments of the Glinka et al. [47] data
Click here for file
Additional File 3
Alignments of the Haddrill et al. [49] data
Click here for file
Additional File 4
Alignments of the Halligan et al. [9] data
Click here for file
Additional File 5
Programs written to carry out partial moment correlations, least-squares regressions and bootstrapping procedures and the data used for these analyses
Click here for file
Acknowledgements
We thank A Betancourt for providing divergence estimates for the Betancourt and Presgraves [19] dataset. We thank D Bachtrog, M Przeworski, K Dyer, F Kondrashov and D Presgraves for comments on the manuscript. This work was funded in part by a Biotechnology and Biological Sciences Research Council Grant (to PA and BC) and an AP Sloan Fellowship in Molecular and Computational Biology to PA. BC is supported by The Royal Society.
Figures and Tables
Figure 1 The relationship between intron length and the level of divergence between D. melanogaster and D. simulans for the combined dataset of 225 introns. A significantly negative correlation is found for all introns (Spearman correlation coefficient Rs = -0.388, P < 10-4), first introns (Rs = -0.451, P < 10-4) and non-first introns (Rs = -0.304, P < 10-4).
Figure 2 Mean divergences for non-synonymous sites, synonymous sites and both small and large introns. Mean levels of divergence between D. melanogaster and D. simulans for non-synonymous and synonymous sites of coding data, introns ≤86 bp and introns >86 bp. Error bars indicate two standard errors. Synonymous site divergence is significantly greater than large (Wilcoxon two-sample test statistic W = 7745.5, P < 10-4) but not small (W = 15115.5, P = 0.6173) intron divergences. Small intron divergence is significantly greater than large intron divergence (W = 17079.5, P < 10-4).
Figure 3 The relationship between intron fragment GC content and both divergence and length. (a) The relationship between GC content of intron fragments and divergence between D. melanogaster and D. simulans (Spearman correlation coefficient Rs = -0.345, P < 10-4). (b) The relationship between GC content of intron fragments and intron length (Rs = 0.237, P < 10-3).
Table 1 Mean divergence and GC content values for each class of DNA
Divergence GC Content
All Short* Long* All Short* Long*
Introns
All 0.093 (0.004) 0.110 (0.005) 0.070 (0.003) 0.357 (0.006) 0.345 (0.009) 0.371 (0.007)
First 0.101 (0.005) 0.114† (0.006) 0.072† (0.006) 0.361 (0.010) 0.352† (0.013) 0.383† (0.011)
Non-first 0.085 (0.005) 0.105† (0.009) 0.069† (0.004) 0.352 (0.007) 0.337† (0.012) 0.365† (0.008)
Synonymous sites 0.127 (0.019) 0.654 (0.014)
Values are mean (standard error). *Introns were divided into two classes based on the median intron length (86 bp) [14]: short, ≤86 bp; long, >86 bp. †Divergence and GC content values did not differ between first and non-first introns when compared within long and short size classes.
==== Refs
Hardison RC Conserved noncoding sequences are reliable guides to regulatory elements. Trends Genet 2000 16 369 372 10973062 10.1016/S0168-9525(00)02081-3
Clark AG The search for meaning in noncoding DNA. Genome Res 2001 11 1319 1320 11483570 10.1101/gr.201601
Bergman CM Kreitman M Analysis of conserved noncoding DNA in Drosophila reveals similar constraints in intergenic and intronic sequences. Genome Res 2001 11 1335 1345 11483574 10.1101/gr.178701
Shabalina SA Ogurtsov AY Kondrashov VA Kondrashov AS Selective constraint in intergenic regions of human and mouse genomes. Trends Genet 2001 17 373 376 11418197 10.1016/S0168-9525(01)02344-7
Dermitzakis ET Reymond A Lyle R Scamuffa N Ucla C Deutsch S Stevenson BJ Flegel V Bucher P Jongeneel CV Antonarakis SE Numerous potentially functional but non-genic conserved sequences on human chromosome 21. Nature 2002 420 578 582 12466853 10.1038/nature01251
Gaffney DJ Keightley PD Unexpected conserved non-coding DNA blocks in mammals. Trends Genet 2004 20 332 337 15262402 10.1016/j.tig.2004.06.011
Li W-H Graur D Fundamentals of Molecular Evolution 1991 Sunderland, Massachusetts: Sinauer
Majewski J Ott J Distribution and characterization of regulatory elements in the human genome. Genome Res 2002 12 1827 1836 12466286 10.1101/gr.606402
Halligan DL Eyre-Walker A Andolfatto P Keightley PD Patterns of evolutionary constraints in intronic and intergenic DNA of Drosophila. Genome Res 2004 14 273 279 14762063 10.1101/gr.1329204
Gibbs RA Weinstock GM Metzker ML Muzny DM Sodergren EJ Scherer S Scott G Steffen D Worley KC Burch PE Genome sequence of the Brown Norway rat yields insights into mammalian evolution. Nature 2004 428 493 521 15057822 10.1038/nature02426
Jareborg N Birney E Durbin R Comparative analysis of noncoding regions of 77 orthologous mouse and human gene pairs. Genome Res 1999 9 815 824 10508839 10.1101/gr.9.9.815
Shabalina SA Kondrashov AS Pattern of selective constraint in C. elegans and C. briggsae genomes. Genet Res Camb 1999 74 23 30
Comeron JM Kreitman M The correlation between intron length and recombination in Drosophila: dynamic equilibrium between mutational and selective forces. Genetics 2000 156 1175 1190 11063693
Yu J Yang Z Kibukawa M Paddock M Passey DA Wong GK-S Minimal introns are not "junk". Genome Res 2002 12 1185 1189 12176926 10.1101/gr.224602
Parsch J Selective constraints on intron evolution in Drosophila. Genetics 2003 165 1843 1851 14704170
Marais G Nouvellet P Keightley PD Charlesworth B Intron size and exon evolution in Drosophila. Genetics 2005 170 481 485 15781704 10.1534/genetics.104.037333
Duret L Why do genes have introns? Recombination might add a new piece to the puzzle. Trends Genet 2001 17 172 175 11275306 10.1016/S0168-9525(01)02236-3
Keightley PD Gaffney DJ Functional constraints and frequency of deleterious mutations in noncoding DNA of rodents. Proc Natl Acad Sci USA 2003 100 13402 13406 14597721 10.1073/pnas.2233252100
Betancourt AJ Presgraves DC Linkage limits the power of natural selection in Drosophila. Proc Natl Acad Sci USA 2002 99 13616 13620 12370444 10.1073/pnas.212277199
Akashi H Molecular evolution between Drosophila melanogaster and D. simulans: reduced codon bias, faster rates of amino acid substitution, and larger proteins in D. melanogaster. Genetics 1996 144 1297 1307 8913769
Moriyama EN Hartl DL Codon usage bias and base composition of nuclear genes in Drosophila. Genetics 1993 134 847 858 8349115
Filipski J Why the rate of silent codon substitutions is variable within a vertebrate's genome. J Theor Biol 1988 134 159 164 3244279
Akashi H Synonymous codon usage in Drosophila melanogaster: natural selection and translational accuracy. Genetics 1994 136 927 935 8005445
Wolfe K Sharp PM Li W-H Mutation rates differ among regions of the mammalian genome. Nature 1989 337 283 285 2911369 10.1038/337283a0
Nagylaki T Evolution of a finite population under gene conversion. Proc Natl Acad Sci USA 1983 80 6278 6281 6578508
Eyre-Walker A Bulmer M Synonymous substitution rates in enterobacteria. Genetics 1995 140 1407 1412 7498779
Sueoka N Directional mutation pressure, mutator mutations, and dynamics of molecular evolution. J Mol Evol 1993 37 137 153 8411203
Bierne N Eyre-Walker A The problem of counting sites in the estimation of the synonymous and nonsynonymous substitution rates: Implications for the correlation between the synonymous substitution rate and codon usage bias. Genetics 2003 165 1587 1597 14668405
Dunn KA Bielawski JP Yang ZH Substitution rates in Drosophila nuclear genes: Implications for translational selection. Genetics 2001 157 295 305 11139510
Lou L Bergson C McGinnis W Deformed expression in the Drosophila central nervous system is controlled by an autoactivated intronic enhancer. Nucleic Acids Res 1995 23 3481 3487 7567459
Bartoszewski S Gibson JB Regulation of the expression of the sn-glycerol-3-phosphate dehydrogenase gene in Drosophila melanogaster. Biochem Genet 1998 36 329 350 9919359 10.1023/A:1018745412966
Shabalina SA Spiridonov NA The mammalian transcriptome and the function of non-coding DNA sequences. Genome Biol 2004 5 105 15059247 10.1186/gb-2004-5-4-105
Berman BP Nibu Y Pfeiffer BD Tomancak P Celniker SE Levine M Rubin GM Eisen MB Exploiting transcription factor binding site clustering to identify cis-regulatory modules involved in pattern formation in the Drosophila genome. Proc Natl Acad Sci USA 2002 99 757 762 11805330 10.1073/pnas.231608898
Dermitzakis ET Bergman CM Clark AG Tracing the evolutionary history of Drosophila regulatory regions with models that identify transcription factor binding sites. Mol Biol Evol 2003 20 703 714 12679540 10.1093/molbev/msg077
Berezikov E Guryev V Plasterk RHA Cuppen E CONREAL: Conserved regulatory elements anchored alignment algorithm for identification of transcription factor binding sites by phylogenetic footprinting. Genome Res 2004 14 170 178 14672977 10.1101/gr.1642804
Bergman CM Carlson JW Celniker SE Drosophila DNase I footprint database: a systematic genome annotation of transcription factor binding sites in the fruitfly, Drosophila melanogaster. Bioinformatics 2005 21 1747 1749 15572468 10.1093/bioinformatics/bti173
Schaeffer SW Miller EL Estimates of linkage disequilibrium and the recombination parameter determined from segregating nucleotide sites in the alcohol dehydrogenase region of Drosophila pseudoobscura. Genetics 1993 135 541 552 8244013
Kirby DA Muse SV Stephan W Maintenance of pre-mRNA secondary structure by epistatic selection. Proc Natl Acad Sci USA 1995 92 9047 9051 7568070
Matzkin LM Eanes WF Sequence variation of alcohol dehydrogenase (Adh) paralogs in cactophilic Drosophila. Genetics 2003 163 181 194 12586706
Solnick D Alternative splicing caused by RNA secondary structure. Cell 1985 43 667 676 4075405 10.1016/0092-8674(85)90239-9
Leicht BG Muse SV Hanczyc M Clark AG Constraints on intron evolution in the gene encoding the Myosin alkali light chain in Drosophila. Genetics 1995 139 299 308 7535717
Liebhaber SA Cash F Eshleman SS Translation inhibition by an mRNA coding region secondary structure is determined by its proximity to the AUG initiation codon. J Mol Biol 1992 226 609 621 1507219 10.1016/0022-2836(92)90619-U
Carlini DB Chen Y Stephan W The relationship between third-codon position nucleotide content, codon bias, mRNA secondary structure and gene expression in the Drosophilid alcohol dehydrogenase genes Adh and Adhr. Genetics 2001 159 623 633 11606539
Chen Y Stephan W Compensatory evolution of a precursor messenger RNA secondary structure in the Drosophila melanogaster Adh gene. Proc Natl Acad Sci USA 2003 100 11499 11504 12972637 10.1073/pnas.1932834100
Stephan W Kirby DA RNA folding in Drosophila shows a distance effect for compensatory fitness interactions. Genetics 1993 135 97 103 8224831
Ogurtsov AY Sunyaev S Kondrashov AS Indel-based evolutionary distance and mouse-human divergence. Genome Res 2004 14 1610 1616 15289479 10.1101/gr.2450504
Glinka S Ometto L Mousset S Stephan W De Lorenzo D Demography and natural selection have shaped genetic variation in Drosophila melanogaster: A multi-locus approach. Genetics 2003 165 1269 1278 14668381
FlyBase: A database of the Drosophila genome
Haddrill PR Thornton KR Charlesworth B Andolfatto P Multilocus patterns of nucleotide variability and the demographic and selection history of Drosophila melanogaster populations. Genome Res 2005 15 790 799 15930491 10.1101/gr.3541005
Keightley PD Johnson T MCALIGN: stochastic alignment of noncoding DNA sequences based on an evolutionary model of sequence evolution. Genome Res 2004 14 442 450 14993209 10.1101/gr.1571904
MCALIGN for alignment of noncoding DNA
DnaSP Software
Jukes TH Cantor CR Munro HN Evolution of protein molecules. Mammalian Protein Metabolism III 1969 New York: Academic Press 21 132
Stephan W Rodriguez VS Zhou B Parsch J Molecular evolution of the Metallothionein gene Mtn in the melanogaster species group: results from Drosophila ananassae. Genetics 1994 138 135 143 8001781
Sokal RR Rohlf FJ Biometry 1995 San Francisco: WH Freeman
Nei M Gojobori T Simple methods for estimating the numbers of synonymous and nonsynonymous nucleotide substitutions. Mol Biol Evol 1986 3 418 426 3444411
Powell JR Moriyama EN Evolution of codon usage bias in Drosophila. Proc Natl Acad Sci USA 1997 94 7784 7790 9223264 10.1073/pnas.94.15.7784
Comeron JM Kreitman M Aguade M Natural selection on synonymous sites is correlated with gene length and recombination in Drosophila. Genetics 1999 151 239 249 9872963
Duret L Mouchiroud D Expression pattern and, surprisingly, gene length shape codon usage in Caenorhabditis, Drosophila and Arabidopsis. Proc Natl Acad Sci USA 1999 96 4482 4487 10200288 10.1073/pnas.96.8.4482
|
16086849
|
PMC1273634
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 27; 6(8):R67
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r67
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r681608685010.1186/gb-2005-6-8-r68ResearchThe Dictyostelium genome encodes numerous RasGEFs with multiple biological roles Wilkins Andrew [email protected] Karol [email protected] Derek J [email protected] Deenadayalan [email protected]üller Rolf [email protected] Paul R [email protected]öckner Gernot [email protected] Ludwig [email protected] Angelika A [email protected] Robert H [email protected] School of Biosciences, University of Birmingham, Birmingham B15 2TT, UK2 Genome Analysis, Institute for Molecular Biotechnology, Beutenbergstrasse 11, D-07745 Jena, Germany3 Department of Microbiology, La Trobe University, VIC 3086, Australia4 Centre for Biochemistry and Centre for Molecular Medicine Cologne, Medical Faculty, University of Cologne, Joseph-Stelzmann-Strasse 52, 50931 Cologne, Germany2005 28 7 2005 6 8 R68 R68 28 2 2005 9 5 2005 21 6 2005 Copyright © 2005 Wilkins et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
A survey of the Dictyostelium genome reveals at least 25 RasGEFs, all of which appear to be expressed at some point in development. Disruption of several of these novel RasGEFs reveals that many have clear phenotypes, suggesting that the unexpectedly large number of RasGEF genes reflects an evolutionary expansion of the range of Ras signaling.
Background
Dictyostelium discoideum is a eukaryote with a simple lifestyle and a relatively small genome whose sequence has been fully determined. It is widely used for studies on cell signaling, movement and multicellular development. Ras guanine-nucleotide exchange factors (RasGEFs) are the proteins that activate Ras and thus lie near the top of many signaling pathways. They are particularly important for signaling in development and chemotaxis in many organisms, including Dictyostelium.
Results
We have searched the genome for sequences encoding RasGEFs. Despite its relative simplicity, we find that the Dictyostelium genome encodes at least 25 RasGEFs, with a few other genes encoding only parts of the RasGEF consensus domains. All appear to be expressed at some point in development. The 25 genes include a wide variety of domain structures, most of which have not been seen in other organisms. The LisH domain, which is associated with microtubule binding, is seen particularly frequently; other domains that confer interactions with the cytoskeleton are also common. Disruption of a sample of the novel genes reveals that many have clear phenotypes, including altered morphology and defects in chemotaxis, slug phototaxis and thermotaxis.
Conclusion
These results suggest that the unexpectedly large number of RasGEF genes reflects an evolutionary expansion of the range of Ras signaling rather than functional redundancy or the presence of multiple pseudogenes.
==== Body
Background
Ras proteins are small GTPases that sit at the center of numerous signaling pathways in essentially all eukaryotes [1]. Their activity is controlled by which guanine nucleotide is bound. When it is GDP, the Ras proteins are inactive and do not bind to their targets. Guanine-nucleotide exchange factors (RasGEFs) catalyze the replacement of GDP with GTP [2]. This makes the Ras proteins active, and able to bind multiple activators and signal transducers. RasGEFs are therefore the initiators of Ras signaling, and understanding their behavior is the key to understanding Ras signaling.
Multiple roles of Ras pathways
The RAS gene was originally described as the cellular counterpart of a viral oncogene, v-ras [3]. The virally encoded protein, which is constitutively GTP-bound even in the absence of RasGEFs and is therefore constantly active [4], causes unchecked mitogenesis and proliferation in appropriate cell lines. Normal mammalian cells encode three different Ras families, Ha-Ras, Ki-Ras and N-Ras, all members of which are highly similar to one another. Examination of tumors from numerous patients has since confirmed that endogenous Ras has a key role in growth control - as many as 90% of pancreatic carcinomas contain a mutated Ras gene similar to v-ras.
The connection between Ras and growth has now been found to be far more complex. In primary cultures, expression of activated Ras causes apoptosis, not unrestricted growth, and activation of Ras has been shown to cause a range of effects including increased cell motility [5], macropinocytosis [6], and alterations in cell identity [7]. These changes are mediated by a range of downstream effectors, most important of which are the lipid kinase phosphatidylinositol 3-kinase (PI3K) and the protein kinase Raf [8,9].
RasGEFs were first identified in Saccharomyces cerevisiae, in which loss of the CDC25 gene was found to arrest growth by blocking Ras activation of adenylyl cyclase [10]. This was followed by the identification of Drosophila Son of sevenless (Sos) [11] and mammalian hSos1 [12], each of which contains a catalytic domain related to that in CDC25. Hundreds of RasGEFs are now known. All share a considerable stretch of homology, including at least two discrete domains - an amino-terminal domain of unclear function (although crystallographic evidence suggests a structural role [13]) and a carboxy-terminal one that mediates GTP-GDP exchange.
RasGEFs and signaling
In general, RasGEFs are now seen as signaling adaptors and integrators; they couple various signaling processes at the cell membrane to Ras and thus to changes inside the cell. The best understood signals to Ras derive from receptor tyrosine kinases (RTKs). When RTKs are stimulated by their ligands, they recruit adaptors such as Grb2 [14], which bind directly to RasGEFs. This recruitment localizes the RasGEFs to the membrane and thus brings them into proximity with Ras [15]. Other RasGEFs are activated by different signals, for example Ca2+ [16], but the underlying mechanism is thought to be similar. The domains that surround the RasGEF catalytic regions are therefore critical, as they mediate membrane localization and activation.
Several major families of RasGEFs can be found in the literature, classified by their domain structure. The most widely known is typified by the product of the Drosophila Sos gene [11]. Members contain one or two pleckstrin homology (PH) domains, implying upstream regulation by membrane phospholipids. They also contain a Dbl homology (DH) domain, which is also found in GEFs for the small GTPases Rho and Rac, although these domains' association with Rac signaling is far less clear than their proven roles as Ras regulators [17]. The EPAC family of GEFs contains cyclic nucleotide monophosphate binding domain (cNMP) motifs that bind cAMP and activate the GTPase Rap when cAMP is present [18]. Similarly, the Ras-GRF family of RasGEFs contain EF hands and activate Ras in response to calcium and diacylglycerol signaling [16]. A significant number of known RasGEF relatives have no obvious signaling domains. The best known of these is C3G, which is thought to activate the Ras-like GTPase Rap1 [19].
Ras pathways in Dictyostelium
The social ameba Dictyostelium discoideum uses Ras pathways to control multiple signaling processes including cell movement, polarity and cytokinesis, chemotaxis, macropinocytosis and multicellular development [20,21]. It is notable for the relatively large number of Ras subfamily members (there are 15 encoded in the genome, including 11 that most closely resemble Ras, three Rap and one Rheb [22]). All of the six so far studied appear to have nonredundant and important roles in cell physiology [23-26] (although six of the remaining ras genes are exceptionally similar [22]). This is the more surprising as Dictyostelium does not encode RTKs [27]. As described previously, RTKs are thought to be one of the major inputs for Ras signaling in mammalian cells. In the absence of RTKs, the best prospects for finding upstream regulators would appear to be through identification of adaptor proteins and binding partners. Such proteins are presumably responsible for recruiting RasGEFs to the membrane and thus controlling their activity. However, two of the four Dictyostelium RasGEFs characterized thus far (aimless [28] and rasgefB [29]) offer few clues to their regulation. Unlike nearly all other RasGEFs described in the literature, which contain a panoply of protein-interaction and regulatory domains including SH3, PH, IQ, and PDZ domains, neither Aimless nor RasGEFB contains recognizable signaling domains [28,29]. Two other RasGEF family members, GbpC and GbpD [30], on the other hand, contain multiple domains, including cGMP-binding domains, and in GbpC, a kinase domain, leucine-rich repeats, and a DEP domain. None of these domains have been shown to regulate GEF activity in Dictyostelium, although in the mammalian EPAC family a cAMP-binding cNMP domain is thought to regulate GEF activity [31]. It therefore seems plausible that cGMP regulates the activity of at least GbpC. The limited number of signaling domains in the other RasGEFs has been a serious obstacle to understanding Ras family signaling in Dictyostelium; to date, no binding proteins or signaling partners have been discovered.
In this paper we describe an unexpectedly large number - at least 25 - of predicted RasGEFs in the Dictyostelium genome. Several of these contain different known signaling domains and others contain none. This suggests an unprecedentedly complex and poorly understood network of Ras signaling in an organism whose signaling otherwise appears relatively simple.
Results
Identification of RasGEF genes
The assembly of the Dictyostelium genome is now complete, representing more than 99% of the genes [27]. This allowed us to estimate the total number of RasGEF sequences. We searched the assembled sequence using FASTA and TBLASTN programs and the RasGEF catalytic sequence (also known as the RasGEF domain) as bait. Approximately 30 sequences gave significant scores with these programs.
An additional domain, known as RasGEF_N (hereafter called RasGEF amino-terminal for clarity), is found in nearly all 'true' RasGEFs (proteins that clearly activate Ras in vivo) as well as in GEFs for a number of related proteins such as Rap1. After excluding sequences that did not contain complete copies of both RasGEF and RasGEF amino-terminal domains, we were left with 25 sequences encoding Dictyostelium RasGEFs. There were no obvious pseudogenes - all 25 sequences contain a clear open reading frame (ORF) and both domains were complete and uninterrupted.
Four of the RasGEF genes we found have been described previously. The aimless and rasgefB genes have roles in chemotaxis and endocytosis, respectively [28,29], while the gbpC and gbpD genes encode cGMP-binding proteins which are thought to couple intracellular cGMP to other signaling pathways [30]. For the sake of consistency, we have renamed these genes gefA, gefB, gefT and gefU.
Four further genes encode the carboxy-terminal RasGEF homology domain but not the RasGEF amino-terminal domain, and might therefore be expected to be specific for Ras-related proteins which are not part of the Ras family proper. We have named these gflA, gflB, gflC and gflD (full details can be found in Dictybase [32]). It is not yet clear whether these are likely to be RasGEFs with a subset of normal functions or GEFs for more distant relatives of Ras.
Sequence and domain analysis
As described earlier, the presence of signaling domains separate from the RasGEF and RasGEF amino-terminal domain has been a central part of the identification of Ras signaling pathways in higher eukaryotes [33]. However, half of the 25 genes we have identified contain no such clues. Figure 1 shows the domain structure of the predicted products from the 25 complete RasGEFs. Eight, including RasGEFB, contain no domains detected by SMART [34] or PFAM [35], other than the two RasGEF domains. A further five contain no additional domains other than the enigmatic LisH domain, whose function is thought to relate to motility and microtubule function but is poorly understood [36], and one contains only two F-boxes, motifs connected with ubiquitination and protein breakdown rather than with upstream signaling.
We were surprised to find that none of the Dictyostelium RasGEFs resembled members of known families with signaling motifs. Only one of the 25 predicted proteins, GefC, includes a DH/Rho-GEF domain like Sos family members [17]. However, unlike the Sos family, GefC does not contain a PH domain. Instead, its amino-terminal region contains three domains that resemble RCC1, a GEF for the small GTPase Ran [37]. Ran is involved in the control of nuclear transport and mitosis, and is unusual in that it does not contain lipid adducts and is therefore not located at the plasma membrane [38]. There is no clear precedent for a connection between Ras and RCC1 signaling.
The two previously described cGMP-binding proteins RasGEFT and RasGEFU (also known as cGMP-binding proteins C and D [30]) have cNMP domains, like the mammalian EPAC family. RasGEFU is somewhat similar to EPAC family members, but the three cNMP domains lie beyond the RasGEF amino-terminal and RasGEF domains, unlike the usual upstream location, and the G-protein-associated DEP domain is replaced by a GRAM membrane-localization domain. These large-scale changes make it seem unlikely that EPACs and Dictyostelium cGMP-binding RasGEFs are evolutionary orthologs. It seems more likely that convergent evolution has selected independent cyclic nucleotide-regulated GEFs. This is supported by phylogenetic analysis (see below), which failed to group RasGEFT and RasGEFU with human EPAC.
The third principal family of mammalian RasGEFs is the calcium-regulated Ras-GRFs [16]. These do not appear to be present in Dictyostelium at all. Neither the EF hands present in Ras-GRFs nor any other clear calcium-binding motifs are found in any of the 25 RasGEFs examined here.
The only protein described in the literature that resembles the Dictyostelium RasGEF homologs is the C3G RapGEF [19], but as this similarity is based on an absence of other defined signaling domains rather than any positive homology, it seems uninformative. In particular, the Dictyostelium gene products lack SH3-binding polyproline domains at their amino termini, which suggests that they have no strong evolutionary relationship with C3G. Again, phylogenetic analysis supports this view (see below).
Actin- and Rho-binding RasGEFs
Three members of the Dictyostelium RasGEF family have domains that suggest a direct link with the actin cytoskeleton. Two contain domains that are associated with direct binding to F-actin. RasGEFF contains three tandem kelch repeats, while RasGEFP contains a calponin homology (CH) domain. A third, RasGEFD, contains a RhoGAP homology domain. While most authors have associated this with inactivation of Rac- and Rho-family members, there is some evidence that some RhoGAP homology domains are found in downstream effectors. We were intrigued to note that Saccharomyces BEM2, a RhoGAP homolog that is required for cell polarity and normal actin dynamics [39], also appears to contain RasGEF and RasGEF amino-terminal domains that have not been described in the literature. This suggests that RasGEFD might perform a similar role in Dictyostelium to BEM2 in Saccharomyces, although no obvious phenotype was seen in growing gefD knockouts (see below).
Phylogenetic analysis
We constructed phylogenetic trees using the conserved RasGEF domains of all the Dictyostelium proteins, both with and without a selection of mammalian RasGEFs for comparison (Figure 2). To our surprise, few groupings were strongly supported during bootstrapping. In general, it is clear that RasGEFF, RasGEFO and RasGEFH are the least similar to other RasGEFs in Dictyostelium and elsewhere. RasGEFI and RasGEFJ are highly similar, suggesting they they arose from a relatively recent gene duplication. Two other pairings, RasGEFs B and V and RasGEFs R and S, were also supported in more than 50% of bootstraps. Other, larger groups that might subdivide the RasGEFs according to function were conspicuously poorly supported, suggesting that the diversification of RasGEF genes happened relatively early in the divergence of Dictyostelium. There is also no evidence that the division of mammalian RasGEFs into SOS, RasGRP, EPAC, C3G and RalGDS families had occurred when Dictyostelium diverged from the animal line.
Expression during growth and development
To determine whether the large number of Dictyostelium RasGEFs was connected with growth (the first role found for Ras pathways), or signaling during development, we determined the expression patterns of all of the 25 genes during growth and development. Initially we screened Northern blots; if a clear result was not obtained we used RT-PCR. We used probes generated from those clones represented in the Tsukuba cDNA project, and used PCR to make the remainder. Table 1 shows the results. All RasGEF genes are expressed at some stage in development; of the genes tested, gefG, of which transcripts could only be detected at 12 and 14 hours of development, was the most weakly expressed. This clearly implies that the large RasGEF family is not made up of pseudogenes or evolutionary relics, which are frequently not expressed.
In general, three patterns of expression were seen. The largest group of genes, which includes gefA, D, F, H, J and L, is expressed in growing cells and with relatively slight changes throughout development. A second group, comprising gefB and K, is expressed at low levels during growth with a sharp increase early in development, while the third, typified by gefC and E, is only expressed after about 12 h of development. These results are consistent with varied roles for RasGEFs in multiple aspects of the Dictyostelium life cycle.
Disruption of a sample of RasGEF genes
There are two obvious ways to explain the unexpectedly large number of RasGEF genes in Dictyostelium. The first would be that these genes are mainly redundant, and that each biological function of a RasGEF is mediated by several Dictyostelium genes. The other possibility is that Dictyostelium has greatly broadened the scope of Ras signaling at some stage in its evolution, and that most GEFs are required for a distinct signaling process. To distinguish between these possibilities, we disrupted a sample of RasGEF genes and searched for phenotypes. gefA (aimless) had already been disrupted [40] and gefB was disrupted during the initial part of this work [41]; each mutant has a clear phenotype. We further attempted to disrupt GEFs C, D, E, F, G, K and L. We were successful in all cases except gefF, which could not be disrupted even after several attempts. This suggests that gefF might have an important role during growth, but confirmation will require disruption in diploids [42].
Phenotypes of mutants
All of the mutants we obtained grow normally, with the exception of gefB mutants, which grow relatively normally on bacterial plates but are unable to grow in axenic culture due to a loss of fluid-phase endocytosis (Figure 3). The morphology of growing cells is apparently normal for each strain apart from gefB mutants, which appear flattened and polar (Figure 4) in a manner reminiscent of growing nonaxenic cells. Similarly, the chemotactic aggregation and development of all mutants examined here except gefB were apparently normal (Figure 5), forming morphologically normal slugs and fruiting bodies with the usual timing. As previously described, gefB mutants aggregate but form no slugs and make highly aberrant fruiting bodies [41]. Three RasGEF mutants described elsewhere show abnormal chemotaxis - gefA/aimless and gefT are both seriously defective [28,43], though gefA can be coaxed to make rather aberrantly shaped slugs, while gefU cells are hyperpolar and better at chemotaxis than wild type [43].
Slug movement phenotypes
While testing the late development of mutants, we observed an apparent lack of slug phototaxis in gefE mutants, despite morphologically normal slugs and fruiting bodies. We therefore assessed slug phototaxis in each of the mutants apart from gefB, which does not form slugs. Figure 6a shows the trails from a sample of slugs migrating towards a lateral light source. Mutants in gefC, gefD, gefG and gefK (not shown) exhibit normal phototaxis, but gefE and gefL mutants are plainly aberrant. The problems with the two strains appear to be different. gefE slugs migrate similar distances to the wild type, but far less accurately towards the light source, whereas gefL slugs are both less accurate and migrate shorter distances, a phenotype that is frequently observed in phototaxis mutants [44].
A more detailed analysis of slug photo- and thermotaxis is shown in Figure 6b and 6c. Both gefE and gefL mutant slugs are clearly poor in phototaxis assays at all cell densities, with gefE mutants the worst affected (Figure 6b, inset) - gefE mutant slugs are at least 20 times less phototactically accurate than wild type, though some orientation is clearly visible (Figure 5a) and measurable (Figure 6b, inset). This behavior is strikingly reminiscent of the phenotype of rasD mutants, which also move normally but with greatly reduced accuracy [25]. Taken together with the expression pattern of gefE, which closely resembles that of rasD, this suggests that RasGEFE is the GEF that causes the most RasD activation during phototaxis.
Unusually for slug phototaxis mutants, the gefE and gefL mutant slugs are also affected in thermotaxis, although the severity of the phenotypes is reversed. Slugs from both mutant strains show diminished but measurable thermotaxis, but gefL mutant slugs seem particularly insensitive to temperature gradients (Figure 6c).
Discussion
We initiated this work to assess whether the large number of RasGEFs in Dictyostelium were functionally redundant, or whether each had a discrete function. Functional redundancy would be caused by groups of RasGEFs having shared or poorly differentiated functions. In particular, one model for RasGEF action suggests that Ras signaling works as a complex network in which specific signals can be transduced by multiple RasGEFs, each of which can activate multiple different Ras proteins. If this model were correct, deletion of any one RasGEF (or indeed Ras) would cause very slight effects, with double and multiple mutants causing progressively more significant deficiencies in a range of different Ras pathways.
The work described in this paper, in agreement with previous work from our labs (for example [42]) and others [43], suggests that RasGEFs have relatively precisely defined roles. Of the RasGEF genes that have now been disrupted, six out of ten had clear phenotypes. We presume that a more complete study of the minute details of the life cycle would also reveal phenotypes in some of the remaining four. This would not be predicted if genetic redundancy was the rule. These results are somewhat distorted, because the RasGEFs were in general named in the order in which they were isolated. The majority of the genes we disrupted were identified relatively early in the lifetime of the cDNA project in Tsukuba, Japan, and therefore tend to be expressed at reasonably high levels. Later genes were identified by screening sequences provided by the genome project and are therefore likely to be either expressed at lower levels or under non-standard conditions, for example, environmental responses that are not seen under laboratory conditions. Even with this caution in mind, though, the clear phenotypes suggest that Dictyostelium uses relatively simple Ras pathways.
This is further supported by the similarities between specific Ras and GEF mutants. As previously described, Mutants in gefB and rasS behave similarly [23,29]. Likewise, in this work we show that gefE resembles rasD in both mutant phenotype and expression pattern. The aimless/gefA phenotype is also very similar to that of rasC [24,28], although one paper suggests a connection between rasG, gefA and the effector protein RIP3 [45]. Again, this suggests that Dictyostelium Ras pathways are relatively linear, with each GEF in general acting on a single Ras protein, rather than the networks that some might have expected.
This, of course, raises the question of what role is played by the other GEFs we describe. The rasB gene has not yet been disrupted, despite multiple attempts. It is thus not currently possible to discern its biological role. No RasGEF genes have been found to cause phenotypes that resemble those of rasG mutants. Finally, at least two other Ras-family proteins (RasX and Rap1) are likely to be activated by RasGEF family members [22]. In mammalian cells, the catalytic domains of RapGEFs are indistinguishable from RasGEFs, and frequently contain no obvious signaling domains much like several of the GEF genes described in this paper. We have connected three of the RasGEFs with specific Ras proteins, leaving 22 RasGEFs to couple to the smaller number of remaining Ras family members. This inequality implies that each Ras protein is likely to be activated by multiple RasGEFs, perhaps explaining the slight phenotypes seen in the unpaired RasGEF mutants.
We have not studied the gflA, gflB, gflC and gflD genes, which encode carboxy-terminal RasGEF homology domains without RasGEF amino-terminal domains, and might therefore be expected to be specific for Ras-related proteins that are not part of the Ras family proper. It is not entirely clear what this might mean for Dictyostelium. Two Ras proteins (RasG and RasD) are most similar to the canonical mammalian Ras families, and others (RasB, RasC, the as-yet unpublished RasX, and RasS) are less closely related to mammalian Ras while still plainly being part of the Ras family [22]. We suspect from the phenotypes of the gefB mutants that RasGEFB acts directly upon RasS. This would imply that the gfl genes act on more distant relatives of canonical Ras than RasS, which would be characterized as unusual small GTPases (for example the RasX family or RsmA-K [22]) rather than members of the Ras family proper.
Above all, the complexity in Ras pathways in Dictyostelium is unexpected for a relatively simple organism with no receptor tyrosine kinases [27]. It might be that the limited range of receptors was a driving force for diversification of Ras signaling during evolution, or it might be that a large number of RasGEFs are highly specialized for specific subsets of some complex role (detecting and integrating starvation signals and quorum factors, for example). A complete understanding will require further analysis of RasGEF genes, but above all a better knowledge of the range of signals that Dictyostelium cells use in their normal biological context.
Conclusion
The clear suggestion from our work is that the unexpectedly large number of Dictyostelium RasGEFs derive from an unusually diverse range of inputs to Ras pathways, rather than large-scale redundancy or multiple pseudogenes following gene duplications. Identifying the input signals, and the mechanisms by which they connect to RasGEF activation, will be a major future challenge for Dictyostelium biology.
Materials and methods
Identification of genes encoding RasGEFs
Known and previously identified rasGEF genes were used to perform TBLASTN searches against the whole dataset generated by the Dictyostelium Sequencing Consortium [46-49]. In addition, IPRscan results for the predicted proteome of the previously published chromosome 2 [50] were screened for motifs IPR001895 (RasGEF) and IPR000651 (RasGEF amino-terminal [51]). In addition, we applied hmmsearch (HMMer package [52]) to scan the protein-translated pre-draft genome assembly (ORFs expected to be about 30 amino acids) for Pfam motifs PF00617 (RasGEF) and PF00618 (RasGEF amino-terminal). Contig sequences generated as described in [53] were extended and verified as described [50] to obtain full and high-quality coverage of the genes. Gene models were predicted using geneid [54]. The Dictyostelium genome is housed and curated at dictyBase [32].
Domain analysis
The predicted complete amino-acid sequences were analyzed using the SMART program at the SMART website in Heidelberg [34]. In initial searches, borderline matches and matches from other libraries were included to ensure that important genes or domains were not excluded (in particular, several of the RasGEF amino-terminal domains are close to the borderline of significance, on either side). Images were copied directly from the SMART website.
Cell growth and development
D. discoideum AX3 and AX2 cells were either grown axenically in HL5 medium or on a bacterial food source at 22°C [55]. For bacterially grown cells, SM agar plates were inoculated with 105-106 Dictyostelium cells plus a suspension of Klebsiella in LB. To follow differentiation, cells growing exponentially from bacterial plates or axenic growth media were washed three times in KK2 (16.5 mM KH2PO4, 3.8 mM K2HPO4 pH 6.0) and plated on KK2 agar or nitrocellulose filters (Millipore). Transformation was performed by a modification of Howard et al. [56]; briefly, cells growing exponentially were mixed with 25 μg of linearized DNA and electroporated in a BioRad gene pulser at 1.0 or 1.1 V, 3 μF with a 5-ohm resistance in series. After 10 min incubation on ice, cells were placed at 22°C for 15 sec in the presence of 2 μl healing solution (100 mM MgCl2, 100 mM CaCl2) and then HL-5 added. Blasticidin-S (ICN) or G418 (Calbiochem) (10 μg/ml) was added 24 h after electroporation. After 7-8 days antibiotic selection, transformants were cloned on lawns of Klebsiella growing on SM agar.
Gene-disruption constructs
Genes were disrupted using the blasticidin cassette from pRHI100, a derivative of pBrsΔBam [57] with additional restriction sites. Clones for each RasGEF kindly provided by the Japanese cDNA database were cut with an appropriate restriction enzyme to generate a site near the middle of the cDNA, and the Bsr gene was inserted into the gap using a rapid ligation kit (NEB). The knockout construct was cut out of the vector (usually using SalI and NotI) and electroporated into axenic Dictyostelium as described above. Disrupted genes were identified using Southern blots with the entire cDNA used as a probe.
Northern blotting and RT-PCR
Development of D. discoideum cells, isolation of RNA at different time points of development (0, 4, 8, 12, 14 and 18 h) and Northern blotting was performed as described [58]. RNA (5 μg) from each time point was reverse transcribed with random prime hexanucleotides using M-MLV RNase H(-) reverse transcriptase according to the manufacturer's protocol (Promega). 1 μl of a 1:10 dilution of the cDNA was used in PCR reactions with gene-specific primers for the different RasGEFs. Primers of 30 bases for product sizes of ~500 bp were selected with the program GeneFisher [59]. PCR was performed at 94°C for 60 sec (denaturation), 55°C (RasGEFG, O, P and S) or 60°C (all others) for 45 sec (annealing), 68°C for 60 sec (elongation) and with 25, 30 (RasGEFO and P) or 40 (RasGEFG and S) cycles. Reactions were separated on agarose gels. RT-PCR and Northern blot results were scored according to the key in Table 1.
Phototaxis and thermotaxis
Qualitative phototaxis tests were performed as described previously [44] by using sterile spatula-style toothpicks to transfer cells to charcoal agar plates from the edges of colonies growing on Klebsiella aerogenes lawns. Phototaxis was scored after 48 h incubation at 21°C with a lateral light source. For quantitative phototaxis experiments, washed amebae were inoculated onto the centers of charcoal agarose plates (pH 6.5) at various densities and incubated with a lateral light source for 48 h at 21°C. For quantitative thermotaxis experiments, washed amebae were inoculated onto the centers of water agarose plates (~2.4 × 106 cells/cm2) and incubated for 72 h in darkness on a heat bar producing a 0.2°C/cm temperature gradient at the agarose surface. Arbitrary temperature units correspond to a temperature range of 14°C (T1) to 28°C (T8), as measured at the center of plates in separate calibration experiments. Slug trails were transferred to PVC disks, stained with Coomassie Blue, and digitized. Slug orientation was analyzed using directional statistics [60].
Acknowledgements
We gratefully acknowledge the continuing assistance of the members of the Dictyostelium genome consortium and the Dictyostelium cDNA project in Japan. Work on this project was funded by an MRC Senior Fellowship to R.H.I., by a BBSRC project grant and by the DFG.
Figures and Tables
Figure 1 Domain structure of predicted Dictyostelium RasGEFs. Predicted protein sequences of all complete RasGEFs predicted from the Dictyostelium genome were searched in the SMART 4.0 database [34]. Individual domains are labeled in the figure. Pink, segments of low compositional complexity determined by the SEG program. Green, potential coiled-coil regions determined by the Coils2 program. Blue bars, transmembrane segments as predicted by the TMHMM2 program.
Figure 2 Phylogenetic analysis of predicted Dictyostelium RasGEFs. The RasGEF domains of (a) the predicted Dictyostelium RasGEFs alone or (b) the predicted Dictyostelium RasGEFs with selected representatives of human RasGEF families were aligned using ClustalX 1.83. Domain boundaries were predicted by SMART [34]. Phylogenetic trees were constructed using 1,000 bootstraps, excluding all matches within gaps. Closed circles indicate nodes found in 100% of bootstraps; open circles indicate nodes found in ≥50% of bootstraps. All other nodes are found in <50% of bootstraps.
Figure 3 Growth of RasGEF mutants in axenic suspension. Vegetative AX3 and gef null cells were grown in axenic medium in shaken flasks. At the indicated times, samples were taken and counted in duplicate using a hemocytometer. Mutants in gefK and gefL were observed on different occasions but were similarly comparable to the parental strain.
Figure 4 Morphology of gef mutant cells. Vegetative AX3 and gef null cells were grown in axenic medium in tissue culture dishes for 24 h and visualized by phase-contrast microscopy. The white bar represents 20 μm. Mutants in gefK and gefL were observed on different occasions but were similarly comparable to the parental strain.
Figure 5 Morphology of gef mutant fruiting bodies. AX3 and gef null cells were developed at 3 × 106 cells/cm2 on nitrocellulose filters for 48 h then photographed using a dissecting microscope. The white bar represents 2 mm. Mutants in gefK and gefL were observed on different occasions but were similarly comparable to the parental strain.
Figure 6 Phototaxis of gef mutant slugs. (a) Traces of slug trails from slugs migrating towards a light source to the right of the figure. Slugs migrated on charcoal agar plates; trails were transferred to white filters and stained with Coomassie blue. (b) Accuracy of slug phototaxis at different starting cell densities. Different quantities of cells were deposited on a charcoal agarose plate and allowed to move for 48 h. Trails were visualized as above, and phototactic efficiency was measured as described in Materials and methods. The inset shows data for gefE and gefL mutants replotted with different axes to reveal significant, if diminished, phototaxis in each case. Other mutants behaved apparently normally (data not shown). (c) Dependence of slug thermotaxis on temperature. Cells were plated and thermotaxis measured as described in Materials and methods. Other mutants behaved apparently normally (data not shown).
Table 1 Expression of gef genes during growth and development
0 h 4 h 8 h 12 h 14 h 18 h
gefA (aleA)* +++ +++ +++ ND ND ND
gefB* + + ++ ++ ++ +
gefC* - - +++ +++ +++ ++
gefD* + + ++ ++ + +
gefE* - - - ++ + +
gefF* + ++ + + + +
gefG* - - - +/- +/- -
gefH* +++ +++ +++ +++ ++ +
gefI† + ++ ++ +++ ++ ++
gefJ1/J2* + + + + + +
gefK† + ++ ++ +++ ++ ++
gefL* ++ ++ ++ ND ++ ++
gefM† ++ ++ ++ +++ +++ ++
gefN† + ++ ++ +++ ++ ++
gefO† +/- +/- +/- + +/- +/-
gefP† +/- +/- +/- + +/- +/-
gefQ† ++ ++ +++ +++ ++ ++
gefR† + + ++ ++ ++ ++
gefS† +/- +/- +/- + + +/-
gefT† ++ +++ +++ +++ ++ +
gefU† + ++ +++ +++ ++ +
gefV† ++ ++ ++ +++ ++ ++
gefW† ++ ++ ++ +++ ++ ++
gefX† ++ ++ +++ +++ ++ ++
gefY† ++ ++ +++ +++ ++ ++
A summary of results from Northern blots, RT-PCR and published data. GEFs are indicated in the left column and hours of development in the top row. gefJ is encoded by two separate but near-identical genes which cannot be differentiated at this level. *Northern Blot; †RT-PCR. +++, strong band; ++, clear band; +, weak band; +/-, hardly detectable; -, no expression detected; ND, not done.
==== Refs
Hughes DA Control of signal transduction and morphogenesis by Ras. Semin Cell Biol 1995 6 89 94 7548847 10.1016/1043-4682(95)90005-5
Boguski MS McCormick F Proteins regulating Ras and its relatives. Nature 1993 366 643 654 8259209 10.1038/366643a0
Shih TY Weeks MO Young HA Scholnick EM Identification of a sarcoma virus-coded phosphoprotein in nonproducer cells tranformed by Kirsten or Harvey murine sarcoma virus. Virology 1979 96 64 79 223311 10.1016/0042-6822(79)90173-9
Der CJ Finkel T Cooper GM Biological and biochemical properties of human rasH genes mutated at codon 61. Cell 1986 44 167 176 3510078 10.1016/0092-8674(86)90495-2
Alexandrova AY Dugina VB Paterson H Bershadsky AD Vasiliev JM Motility of intracellular particles in rat fibroblasts is greatly enhanced by phorbol ester and by over-expression of normal p21N-ras. Cell Motil Cytoskeleton 1993 25 254 266 8221903 10.1002/cm.970250306
Bar-Sagi D Feramisco JR Induction of membrane ruffling and fluid-phase pinocytosis in quiescent fibroblasts by ras proteins. Science 1986 233 1061 1068 3090687
Yochem J Sundaram M Han M Ras is required for a limited number of cell fates and not for general proliferation in Caenorhabditis elegans. Mol Cell Biol 1997 17 2716 2722 9111342
Marais R Light Y Paterson HF Marshall CJ Ras recruits Raf-1 to the plasma membrane for activation by tyrosine phosphorylation. EMBO J 1995 14 3136 3145 7542586
Rodriguez-Viciana P Warne PH Dhand R Vanhaesebroeck B Gout I Fry MJ Waterfield MD Downward J Phosphatidylinositol-3-OH kinase as a direct target of Ras. Nature 1994 370 527 532 8052307 10.1038/370527a0
Broek D Toda T Michaeli T Levin L Birchmeier C Zoller M Powers S Wigler M The S. cerevisiae CDC25 gene product regulates the RAS/adenylate cyclase pathway. Cell 1987 48 789 799 3545497 10.1016/0092-8674(87)90076-6
Bonfini L Karlovich CA Dasgupta C Banerjee U The Son of sevenless gene product: a putative activator of Ras. Science 1992 255 603 606 1736363
Fath I Apiou F Schweighoffer F Chevallier-Multon MC Ciora T Dutrillaux B Tocque B Identification of two human homologs to Drosophila SOS (son of sevenless) localized on two different chromosomes. Nucleic Acids Res 1993 21 4398 8415003
Boriack-Sjodin PA Margarit SM Bar-Sagi D Kuriyan J The structural basis of the activation of Ras by Sos. Nature 1998 394 337 343 9690470 10.1038/28548
Buday L Downward J Epidermal growth factor regulates p21ras through the formation of a complex of receptor, Grb2 adapter protein, and Sos nucleotide exchange factor. Cell 1993 73 611 620 8490966 10.1016/0092-8674(93)90146-H
Aronheim A Engelberg D Li N al-Alawi N Schlessinger J Karin M Membrane targeting of the nucleotide exchange factor Sos is sufficient for activating the Ras signaling pathway. Cell 1994 78 949 961 7923364 10.1016/0092-8674(94)90271-2
Farnsworth CL Freshney NW Rosen LB Ghosh A Greenberg ME Feig LA Calcium activation of Ras mediated by neuronal exchange factor Ras-GRF. Nature 1995 376 524 527 7637786 10.1038/376524a0
Nimnual AS Yatsula BA Bar-Sagi D Coupling of Ras and Rac guanosine triphosphatases through the Ras exchanger Sos. Science 1998 279 560 563 9438849 10.1126/science.279.5350.560
de Rooij J Zwartkruis FJ Verheijen MH Cool RH Nijman SM Wittinghofer A Bos JL Epac is a Rap1 guanine-nucleotide-exchange factor directly activated by cyclic AMP. Nature 1998 396 474 477 9853756 10.1038/24884
Gotoh T Hattori S Nakamura S Kitayama H Noda M Takai Y Kaibuchi K Matsui H Hatase O Takahashi H Identification of Rap1 as a target for the Crk SH3 domain-binding guanine nucleotide-releasing factor C3G. Mol Cell Biol 1995 15 6746 6753 8524240
Chubb JR Insall RH Dictyostelium: an ideal organism for genetic dissection of Ras signaling networks. Biochim Biophys Acta 2001 1525 262 271 11257439
Wilkins A Insall RH Small GTPases in Dictyostelium: lessons from a social amoeba. Trends Genet 2001 17 41 48 11163921 10.1016/S0168-9525(00)02181-8
Insall R Gaudet P Weeks G Loomis WF, Kuspa A The small GTPase superfamily. Dictyostelium 2005 Norfold: Horizon Bioscience
Chubb JR Wilkins A Thomas GM Insall RH The Dictyostelium RasS protein is required for macropinocytosis, phagocytosis and the control of cell movement. J Cell Sci 2000 113 709 719 10652263
Lim CJ The Ras subfamily protein, RasC, is required for the aggregation of Dictyostelium discoideum. PhD thesis 2002 University of British Columbia
Wilkins A Khosla M Fraser DJ Spiegelman GB Fisher PR Weeks G Insall RH Dictyostelium RasD is required for normal phototaxis, but not differentiation. Genes Devel 2000 14 1407 1413 10837033
Tuxworth RI Cheetham JL Machesky LM Spiegelmann GB Weeks G Insall RH Dictyostelium RasG is required for normal motility and cytokinesis, but not growth. J Cell Biol 1997 138 605 614 9245789 10.1083/jcb.138.3.605
Eichinger L Pachebat JA Glockner G Rajandream MA Sucgang R Berriman M Song J Olsen R Szafranski K Xu Q The genome of the social amoeba Dictyostelium discoideum. Nature 2005 435 43 57 15875012 10.1038/nature03481
Insall RH Borleis J Devreotes PN The aimless RasGEF is required for processing of chemotactic signals through G-protein-coupled receptors in Dictyostelium. Curr Biol 1996 6 719 729 8793298 10.1016/S0960-9822(09)00453-9
Wilkins A Chubb J Insall RH A novel Dictyostelium RasGEF is required for normal endocytosis, cell motility and multicellular development. Curr Biol 2000 10 1427 1437 11102804 10.1016/S0960-9822(00)00797-1
Goldberg JM Bosgraaf L van Haastert PJM Smith JL Identification of four candidate cGMP targets in Dictyostelium. Proc Natl Acad Sci USA 2002 99 6749 6754 12011437 10.1073/pnas.102167299
de Rooij J Rehmann H van Triest M Cool RH Wittinghofer A Bos JL Mechanism of regulation of the Epac family of cAMP-dependent RapGEFs. J Biol Chem 2000 275 20829 20836 10777494 10.1074/jbc.M001113200
dictyBase
Quilliam LA Rebhun JF Castro AF A growing family of guanine nucleotide exchange factors is responsible for activation of Ras-family GTPases. Prog Nucleic Acid Res Mol Biol 2002 71 391 444 12102558
SMART
PFAM
Emes RD Ponting CP A new sequence motif linking lissencephaly, Treacher Collins and oral-facial-digital type 1 syndromes, microtubule dynamics and cell migration. Hum Mol Genet 2001 10 2813 2820 11734546 10.1093/hmg/10.24.2813
Kadowaki T Goldfarb D Spitz LM Tartakoff AM Ohno M Regulation of RNA processing and transport by a nuclear guanine nucleotide release protein and members of the Ras superfamily. EMBO J 1993 12 2929 2937 7687541
Rush MG Drivas G D'Eustachio P The small nuclear GTPase Ran: how much does it run? BioEssays 1996 18 103 112 8851043 10.1002/bies.950180206
Wang T Bretscher A The rho-GAP encoded by BEM2 regulates cytoskeletal structure in budding yeast. Mol Biol Cell 1995 6 1011 1024 7579704
Insall RH Borleis J Devreotes PN The aimless RasGEF is required for processing of chemotactic signals through G-protein-coupled receptors in Dictyostelium. Curr Biol 1996 6 719 729 8793298 10.1016/S0960-9822(09)00453-9
Wilkins A Chubb JR Insall RH A novel Dictyostelium RasGEF is required for normal endocytosis, cell motility and multicellular development. Curr Biol 2000 10 1427 1437 11102804 10.1016/S0960-9822(00)00797-1
King J Insall RH Parasexual genetics of Dictyostelium gene disruptions: identification of a ras pathway using diploids. BMC Genetics 2003 4 12 12854977 10.1186/1471-2156-4-12
Bosgraaf L Waijer A Engel R Visser AJ Wessels D Soll D van Haastert PJ RasGEF-containing proteins GbpC and GbpD have differential effects on cell polarity and chemotaxis in Dictyostelium. J Cell Sci 2005 118 1899 1910 15827084 10.1242/jcs.02317
Darcy PK Wilczynska Z Fisher PR Genetic analysis of Dictyostelium slug phototaxis mutants. Genetics 1994 137 977 985 7982578
Lee S Parent CA Insall R Firtel RA A novel Ras-interacting protein required for chemotaxis and cyclic adenosine monophosphate signal relay in Dictyostelium. Mol Biol Cell 1999 10 2829 2845 10473630
Dictyostelium genome analysis
Baylor Human Genome Sequencing Center
The Welcome Trust Sanger Institute
Dictyostelium discoideum genome project
Glockner G Eichinger L Szafranski K Pachebat JA Bankier AT Dear PH Lehmann D Baumgart C Parra G Abril JF Sequence and analysis of chromosome 2 of Dictyostelium discoideum. Nature 2002 418 79 85 12097910 10.1038/nature00847
InterPro
HMMER: sequence analysis using profile hidden Markov models
Kollmar M Glockner G Identification and phylogenetic analysis of Dictyostelium discoideum kinesin proteins. BMC Genomics 2003 4 47 14641909 10.1186/1471-2156-4-S1-S47
Szafranski K Lehmann R Parra G Guigo R Glockner G Gene organization features in A/T-rich organisms. J Mol Evol 2005 60 90 98 15696371 10.1007/s00239-004-0201-2
Sussman R Sussman M Cultivation of Dictyostelium discoideum in axenic culture. Biochem Biophys Res Commun 1967 29 53 55 6069704 10.1016/0006-291X(67)90539-6
Howard PK Ahern KG Firtel RA Establishment of a transient expression system for Dictyostelium discoideum. Nucl Acids Res 1988 16 2613 2623 3362676
Sutoh K A transformation vector for Dictyostelium discoideum with a new selectable marker bsr. Plasmid 1993 30 150 154 8234487 10.1006/plas.1993.1042
Knuth M Khaire N Kuspa A Lu SJ Schleicher M Noegel AA A novel partner for Dictyostelium filamin is an alpha-helical developmentally regulated protein. J Cell Sci 2004 117 5013 5022 15383615 10.1242/jcs.01366
Gene Fisher: interactive primer design
Fisher FR Smith E Williams KL An extracellular chemical signal controlling phototactic behavior by D. discoideum slugs. Cell 1981 23 799 807 7226230 10.1016/0092-8674(81)90444-X
|
16086850
|
PMC1273635
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 28; 6(8):R68
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r68
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r691608685110.1186/gb-2005-6-8-r69ResearchTandem repeat copy-number variation in protein-coding regions of human genes O'Dushlaine Colm T [email protected] Richard J [email protected] Stephen D [email protected] Denis C [email protected] Bioinformatics Core, Department of Clinical Pharmacology and Institute of Biopharmaceutical Sciences, Royal College of Surgeons in Ireland, 123 St Stephen's Green, Dublin 2, Ireland2005 28 7 2005 6 8 R69 R69 11 2 2005 31 5 2005 13 7 2005 Copyright © 2005 O'Dushlaine et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Tandem repeat polymorphisms in human proteins were characterized using the UniGene dataset. This analysis suggests that 1 in 20 proteins are likely to contain tandem repeat copy-number polymorphisms within coding regions; these were prevalent among protein-binding proteins.
Background
Tandem repeat variation in protein-coding regions will alter protein length and may introduce frameshifts. Tandem repeat variants are associated with variation in pathogenicity in bacteria and with human disease. We characterized tandem repeat polymorphism in human proteins, using the UniGene database, and tested whether these were associated with host defense roles.
Results
Protein-coding tandem repeat copy-number polymorphisms were detected in 249 tandem repeats found in 218 UniGene clusters; observed length differences ranged from 2 to 144 nucleotides, with unit copy lengths ranging from 2 to 57. This corresponded to 1.59% (218/13,749) of proteins investigated carrying detectable polymorphisms in the copy-number of protein-coding tandem repeats. We found no evidence that tandem repeat copy-number polymorphism was significantly elevated in defense-response proteins (p = 0.882). An association with the Gene Ontology term 'protein-binding' remained significant after covariate adjustment and correction for multiple testing. Combining this analysis with previous experimental evaluations of tandem repeat polymorphism, we estimate the approximate mean frequency of tandem repeat polymorphisms in human proteins to be 6%. Because 13.9% of the polymorphisms were not a multiple of three nucleotides, up to 1% of proteins may contain frameshifting tandem repeat polymorphisms.
Conclusion
Around 1 in 20 human proteins are likely to contain tandem repeat copy-number polymorphisms within coding regions. Such polymorphisms are not more frequent among defense-response proteins; their prevalence among protein-binding proteins may reflect lower selective constraints on their structural modification. The impact of frameshifting and longer copy-number variants on protein function and disease merits further investigation.
==== Body
Background
DNA tandem repeats are two or more adjacent and approximate copies of a sequence of nucleotides. The presence of tandem repeats has been associated with a number of diseases and phenotypic conditions. For instance, repeat polymorphisms in 5' and 3' regions are known to cause diseases such as Huntington's disease [1] and certain forms of Fragile X syndrome [2]. Other tandem repeat polymorphisms in noncoding regions are known to modify function through their impact on gene regulation [3,4]. These polymorphisms can arise from events such as unequal crossover, replication slippage or double-strand break repair [5-7].
Polymorphism of tandem repeats within protein-coding sequences is known to modulate disease risks and can effect changes in the protein products of genes, leading to diseases such as myotonic dystrophy [8]. A number of diseases caused by repeat polymorphism arise from the expansion of trinucleotide repeats [9]. Other longer repeat polymorphisms have been postulated to modify disease risk (for example, platelet glycoprotein Ib-α (GP1BA) repeat [10], the kringle repeat in apolipoprotein(a) (LPA) [11], and P-selectin ligand (SELPLG) repeat [12]).
While single-nucleotide polymorphisms (SNPs) are currently the subject of extensive research, tandem repeats can exhibit high levels of length polymorphism that will potentially alter protein function. In addition, the comparatively greater mutability of certain classes of tandem repeats may lead to a different spectrum of effects on function, as mildly deleterious variants of recent origin may not have had time to be eliminated. Previous studies [13,14] have predicted polymorphism using a minimum threshold of repeating units and a minimum homogeneity criteria. The threshold refers to the minimum number of repeat units needed for a locus to be scored as likely to contain polymorphism, and the homogeneity refers to percentage of nucleotides within a repeat that may deviate from the core repetitive unit. The criteria depended on the length of the repeat unit and were drawn from the literature on repeat polymorphisms. For instance, for a dinucleotide repeat block to be scored as a likely polymorphism, a threshold number of eight repeat units and a minimum homogeneity of 0.9 was required.
This approach was used to predict 11,265 potentially polymorphic tandem repeats and led to the proposal that 22% of UniGene [15] clusters contain at least one potentially polymorphic locus [14]. Of these, 8% were predicted to be in coding regions. If polymorphic, these loci could cause frameshift mutations, which would be likely to significantly alter the protein product. However, these studies only analyzed a single representative sequence from each UniGene cluster, and did not investigate the observed variability among all sequences within the cluster. Additional studies predicting potentially polymorphic repeats have focused on minisatellite repeats. For instance, Denoeud and colleagues [16] were more interested in highly polymorphic minisatellites and only used strict definitions of minisatellites (unit length greater than 17 nucleotides, for instance). Naslund and co-workers [17] used a logistic regression approach to predict potentially polymorphic repeats. However, they were specifically interested in minisatellites with a minimum repeat unit length of six nucleotides and not the full spectrum of repeat unit lengths. Denoeud and Vergnaud have carried out genomic comparisons of related bacteria to observe tandem repeat sequence length differences [18]. However, no such analysis has been carried out to detect human repeat polymorphism.
It has been hypothesized that an excess diversity of coding tandem repeats contributes to antigenic variation within the prokaryotic pathogen Neisseria [19]. Variations in the numbers of repeats within the collagen-like region in Bacillus anthracis correlated with variation of filament length on the spore surface and have been proposed to affect the properties of the spores in response to various environments [20]. Indeed, repeat-mediated variation may form an integral part of the ability of many pathogens to adapt and remain adapted to their hosts and environments [21] and has been proposed as a molecular basis for the rapid adaptation of both prokaryotes and eukaryotes to environmental changes [22]. Our investigations sought to find evidence of the existence of this in humans. We proposed that repeat polymorphism within host-defense proteins in the human population might be advantageous, as previously postulated [14], and thus we would expect higher levels of tandem repeat sequence length variation in these genes. Such proteins exhibit rapid rates of evolution in interspecies comparisons, consistent with positive selection for changes in response to pathogen selection pressures [23,24].
Here we report an investigation into the level of apparent polymorphism in human genes within the UniGene database, and examine whether such polymorphism is elevated in host-defense genes.
Results and discussion
Protein-coding repeat distribution
Of the 106,937 UniGene [15] sequence clusters, 14,953 (14%) contained coding sequence annotation. Of these, a total of 13,749 (13%) clusters had more than one sequence overlapping a repeat block, enabling a search for tandem repeat copy-number variants.
A total of 89,243 tandem repeats were detected in protein-coding regions of the 13,783 UniGene representative sequences. The detected repeats were predominantly those with short repeat unit lengths of two to six nucleotides (Figure 1a). The distribution showed a clear elevation of repeat units that were a multiple of three, which agrees with previous findings that protein-coding region repeats whose copy-number variation is likely to cause frameshift errors occur at a lower frequency in coding regions [25-27]. We noted a much greater excess of trimer repeats relative to dimers and tetramers in this dataset than in a previous genomic analysis of exonic sequences [27]. This is likely to largely reflect the exclusion of 5' and 3' untranslated regions (UTRs) from our dataset; intronic and intergenic regions in the genomic analysis had a still greater incidence of dimers and tetramers compared to trimers [27]. Thus, although there is an apparent mutational bias against trimer repeats in genomic sequences, within protein-coding regions they are the most frequent class of tandem repeats. Of the detected repeats 82% were 100% homogenous. Thus, 18% of the dataset included were inexact repeats, with a higher proportion of inexact repeats among the arrays composed of longer repeat units.
Range of tandem repeat copy-number variation
Detected variants were screened to ensure that they represented length variation arising as copy-number differences in genomic DNA, rather than intron retention or alternative splicing. Only length variations that corresponded to a length difference that was a multiple of the repeat unit were selected. This reduced the number of clusters with variation from 4,458 (16,483 query/hit pairs) to 623 (3,111 query/hit pairs). For this set, tandem repeats were detected in the variant sequence and checked to ensure that the observed copy-number was in agreement with the expected one, given the length of the hit block and the length of the repeat unit, further reducing the dataset to 218 clusters with observations of length variation (753 query/hit pairs).
In total, 249 unique repeat blocks (spanning 218 clusters) showed variation that was consistent with a change in repeat copy-number (Figure 1). We found 295 allelic variants that differed from the UniGene representative sequence (Additional data file 1) and 85.8% of these variants were a multiple of three nucleotides (253/295). Nearly 70% of variations that were a multiple of both three nucleotides and the repeat unit length arose within blocks of trinucleotide tandem repeats (Figure 1). Although some of the invariant repeats were imperfect, all the variant repeats were 100% homogenous (that is, every repeat unit was identical), and a large proportion were short (48% of variant repeat arrays were less than 20 nucleotides in length). The mean percentage match for repeats with array length less than 20 nucleotides was 98.52%. The mean percentage match for repeats with array length equal to or greater than 20 nucleotides was 90.50%.
Figure 2 illustrates the length differences observed between representative and other sequences. The majority of longer base differences were observed in repeats with a long repeat unit. Also, in most cases the majority of differences for a repeat of a given length are equal to one copy of that repeat, as indicated by the size of the circles in Figure 2. Among the longer repeat units, the variant alleles typically only differ by a single repeat unit (points along the diagonal). Allelic variants that differ by a larger number of repeat units are seen more often among the shorter repeats. The longest repeat units that exhibited polymorphism were 18 (3 representatives), 30 (2), 45 (1), 48 (2) and 57 (2) nucleotides in length (Figure 2, see also Additional data file 1). Of these large variants, the effects of the dopamine D4 receptor (DRD4) and SELPLG polymorphism have been well investigated [28,29], indicating probable effects on function and disease. The functional or clinical impact of the other variants remain to be evaluated, however. Clearly, the UniGene sampling approach is incomplete, and there are likely to be more large variant repeats in the human proteins; for example, the well known GP1BA polymorphism, with a unit size of 39 nucleotides, and the mucin 2 (MUC2) polymorphism [10,30]. These variants were not identified by this study, since the UniGene cluster sizes for these genes were too low to detect the common variants [31]. Three of the trimer repeats exhibited substantial length differences (39, 42 and 63 nucleotides, Figure 2), which are again likely to affect protein function. These were in the genes for the alpha 1A subunit of the voltage-dependent, P/Q type calcium channel (CACNA1A), the TATA-box binding protein (TBP) and the translocated promoter region to the activated MET oncogene (TPR) (Additional data file 1). While most of the CACNA1A allelic variants were in the 'normal' range of variation, the longest allele of 24 repeats was in the size range associated with the well studied trinucleotide-repeat expansion disease spinocerebellar ataxia 6 (SCA6) [32]. For TBP all eight allelic variants were below the length associated with a form of inherited ataxia [33,34]. TPR has not been associated with trinucleotide-repeat expansion diseases. A region of this oncogene has, however, been associated with nonrandom chromosomal deletions [35], and the role of this polymorphism in cancer may be of interest.
As an independent check for the completeness of our observations, the Human Gene Mutation Database (HGMD) [36] was queried with a set of all official HUGO gene symbols. A total of 18 contained coding-sequence repeat polymorphisms. Of these, eight (or 44%) were detected in our analysis - HD, ATXN1, ATXN2, AR, CACNA1A, TBP, SELPLG, and ATN1. Four of the remaining ten lacked coding-sequence annotation in the Hs.seq.uniq representative precluding the use of our method. One of the remaining six was a polymorphic mononucleotide repeat - these repeats were not included in our analysis. Two further genes contained cryptic GCN repeats. The last three had no variant hits in UniGene, either because of small cluster size (13, 170, 56), sequence error reducing the amount of hits (within-cluster alignments), or a lack of sufficient sequence coverage over the repeat region. Thus, in total, seven repeat variations were 'missed' either because of a lack of UniGene coding sequence annotation (4) or as a result of cluster size/sequence quality limitations (3), and three did not conform to the types of repeats considered in our analysis. Therefore, in relation to repeat variations previously associated with disease and considered in our analysis, we detected variations in 53% (8/15) of the associated genes.
This analysis highlights that fact that, while UniGene is a useful resource for looking at polymorphism, it has its limitations, specifically in relation to sample size, sequence quality and annotation. Of the 218 gene clusters with repeat variation, 34 had entries on the HGMD, eight of which - HD, ATXN1, ATXN2, AR, CACNA1A, TBP, SELPLG, and ATN1 - had coding-region repeat polymorphisms that were detected in our analysis. One further gene - VWF - was annotated as having a small deletion that corresponded to one of our repeat variants. Another gene - TWIST1 - was annotated as having a small deletion in the Saethre-Chotzen syndrome phenotype, which was detected in our analysis as a 12-nucleotide indel for a three-nucleotide repeat (GGC). While the variation observed in VWF may have arisen from a repeat slippage event, the variant for TWIST1 is unlikely to have done so. In addition to these variants, three genes - NUMBL, E2F4 and NOTCH4 - were annotated by Online Mendelian Inheritance in Man (OMIM) [37] as exhibiting trinucleotide repeat variation. Thus, 13 variants detected in our analysis were previously identified.
Frequency of repeat variants
Given the likely sampling errors and biases, we did not expect frequencies of repeat variants to closely reflect true population frequencies. However, for known repeat variations from the literature that were also detected in our analysis, we compared heterozygosities by querying the GDB database [38]. For a set of five genes that had heterozygosity information and existed in the GDB database (HD, AR, TBP, ATN1, HRC), the heterozygosity in GDB was broadly similar (values of 0.8, 0.63, 0.81, 0.79 and 0.55, respectively) to that estimated from this dataset (Additional data file 2).
Repeat copy-number and extent of variation
We compared the mean copy-number of the tandem repeats between clusters that have repeat variants and those without and found a significant difference (Mann-Whitney, p < 0.0001). As expected, the trend is for variant repeats to have a higher copy-number (Figure 3). This observation [39] has formed the basis of previous studies predicting repeat variation [13,14]. This difference in copy-number for the trimer repeats did not simply reflect a shift in the mean copy-number; there was a substantial upper tail in the distribution, indicating that the chance of a trimer being polymorphic increases as the copy-number increases. In contrast, there was no such marked tail of variants of relatively high copy-number for dimer repeats (Figure 3). This difference between dimer and trimer variation could represent a difference in mutational mechanisms, or, alternatively, the dimers may be subject to purifying selection against expansion, as most of the dimer variants are likely to cause frameshifts.
Origin of variation
Interestingly, the vast majority of dimer, tetramer and pentamer copy-number variants resulted in a length difference that was not divisible by three (Figure 1b). Given the preference for repeat variation that is a multiple of three nucleotides, we had anticipated that there would be a greater proportion of copy-number variants that expand or contract dimer and tetramer repeats by exactly three copies (for example, we expected to see a larger number of dimer tandem variants that differed in length by six nucleotides). The observation that such variants are very rare (Figure 1b), even though they do not disrupt the reading frame, strongly supports the stepwise mutation model for microsatellite repeats [40,41], and suggests that insertion/deletion mutations of more than one unit at a time are quite unusual. It is probable that the frameshifting copy-number variants are mainly recent mutations that are selectively deleterious, reducing the chance of gradual expansion of the tandem array variant over time; trimer repeat variants could typically be much older. Thus, the majority of copy-number mutations in tandem arrays with short unit sizes are likely to arise by slippage [42], which occurs most often in homogenous repeats [43,44]. Consistent with this, the majority of observed variants for these repeats differ by a single unit.
In contrast, for a number of the larger tandem repeats (unit size of 12 and above) the observed variants in some cases differ by more than one copy, with no sampling of an intermediate allele (Figure 2). Such longer repeat variants may potentially arise through recombination, rather than slippage mechanisms, giving the potential for the gain or loss of more than one unit at a time. It should also be mentioned that the use of UniGene to detect variation precludes the ability to determine if the variation exists at DNA or transcriptional level. Our requirement that observed length variations had to be consistent with a change in repeat copy-number minimized the likelihood of detecting variation resulting from an alternative splice site arising within a repeat block. This did not, however, rule out inclusion of alternative splices where the splice sites might coincide with boundaries of tandem repeat units. Inspection of the intron/exon structure of genes in our results using EnsEMBL [45] revealed no such examples (data not shown).
Frameshifting copy-number variation
This dataset is likely to underestimate the frequency of frameshifting repeat variants, as a large number of frameshifts stimulate nonsense-mediated RNA decay, biasing against their chance of being detected in UniGene. Messages carrying stop codons more than 50 nucleotides upstream of an intron are typically subject to rapid mRNA decay [46]. Secondly, nonsense polymorphisms typically occur at a low frequency in human proteins [47], reflecting selection against deleterious alleles, and it is possible that frameshifting tandem copy-number variants may similarly be at a lower frequency. Given the small sample size for many of the UniGene clusters, the incidence of frameshifting polymorphisms is probably strongly under-represented.
A few of the observed variants may not be true frameshifts, however, owing either to errors in coding-sequence annotation, sequencing errors, transcriptional errors or transcribed pseudogenes in the database. While we cannot definitively rule these out, the validation of repeat variants to ensure that they represent a change in repeat copy-number would reduce that possibility of some of these errors arising. Nevertheless, for the two reasons outlined above, we believe that the observation of one frameshifting tandem repeat polymorphism per 404 (34 out of 13,749) proteins surveyed (0.25%) represents a likely lower bound of the frequency. Wren et al. [14] predicted that 0.5% of proteins are likely to contain frameshifting tandem repeat polymorphisms.
It is of course possible that frameshifting tandem repeats can arise from sequencing errors, transcription errors or pseudogene transcripts. We inspected the 34 sequences containing frameshifting dinucleotide variants, and found that, in all but one sequence, the percentage of bases that were ambiguous (denoted by base 'N') was less than 1% (the outlier was 4%). We also searched the 51 frameshifting sequences and the representative allele against the human genome, and in each case both alleles hit the same sequence; that is, there was no evidence for the existence of a pseudogene with greater similarity to the frameshifted allele.
We cannot rule out the possibility of occasional transcriptional slippage giving rise to a small proportion of the observed variation: an experimental screen for such transcriptional errors estimated their frequency at approximately 1 in 5,000 transcripts in dinucleotide tandem repeats [48]: in our survey of 5,304 sequences containing 8,449 dinucleotide repeats, we found an incidence of 36 frameshifting dinucleotide mutations, compared with an expectation of less than two, arising from transcriptional errors. Secondly, two of the tetramer frameshifting repeats, and four of the dimer repeats, were observed in more than one sequence, which is a strong indication of a DNA, rather than a transcriptional, difference. None of the variants detected involved complete deletion of the repeat, with the lowest copy-number in the variant being 1.8 (see Additional data file 1).
Association of copy-number variation and host-defense functions
While previous work has shown clear ontological trends for repeats that exhibit variation, it was restricted to certain classes of repeats [49]. We tested whether there was an excess of tandem repeat polymorphic variation in host-defense proteins by comparing the frequency of polymorphic genes among those classified as being related to 'defense response' (GO:0006952) [50] or not. There were 484 UniGene clusters that mapped to defense-response proteins and 8,129 clusters that did not. The mean variation was marginally higher in the defense-response category but this was not significant (p = 0.982, Chi-squared test) (Table 1).
The ability to detect repeat variation within a given cluster is partially dependent on both the number of sequences in which we detected tandem repeats, and the number of repeat blocks in the sequence. These are highly correlated with the number of sequences in the cluster and sequence length, respectively (data not shown). It is possible that these two variables - cluster size and sequence length - might relate to protein groupings with certain functions. In addition, cluster size may be affected by ascertainment bias for certain genes highly expressed in well sampled tissues, and there may be an ascertainment bias towards variant sequences that have been preferentially selected for sequencing. Therefore, we performed a logistic regression where the dependent categorical variable described whether or not the cluster contained a variant repeat population, and tested this against the categorical 'defense response' variable (describing whether the cluster linked to the GO term). We considered as covariates the number of sequences within each cluster as well as the length of the protein. We found that variation was not dependent on the defense-response classification when both the number of sequences and the length of the protein were considered as covariates (p = 0.882) (Table 1).
Thus, we find no evidence that human host-defense proteins have an excess of tandem repeat variation. It is possible that the large size of human gene promoters and their innate variability (in SNPs, tandem repeats, indels and other polymorphisms) provides ample opportunity in response to pathogen challenges for rapid selection of variants modulating gene function. There may therefore be no strong long-term selection pressure to develop an innate reservoir of potential variation within the protein sequences themselves. We anticipate that it may be more likely that such advantageous tandem repeat polymorphisms would arise in host-defense proteins of organisms that lack the adaptive immune system and have much larger population sizes.
Association of tandem repeat copy-number variation and Gene Ontology (GO) terms
We investigated whether the occurrence of copy-number polymorphisms was associated with any other GO terms. Of the 362 level-4 terms in GO [50], 167 terms could be linked to our dataset and had at least one cluster linking to the term. We tested whether or not variation was significantly associated with any of these terms using a Fisher's exact test. This found 13 terms to be significant, of which only the term 'protein-binding' (GO:0005515) remained significant after Bonferroni correction for multiple testing. Again, we wished to ensure that the UniGene cluster size and the sequence length were not confounding the associations between variability and GO terms. Therefore, we performed the logistic regression described above, for which 67 of the 167 terms had a sufficiently large sample size to be tested. Twelve of these terms were significant, one of which remained significant after correcting for multiple testing. Again, this term was 'protein binding'. To ensure that the observed significance could not be largely attributed to differences in repeat copy-number between variants and non-variants (Figure 3) we performed the logistic regression with the mean repeat copy-number per cluster as an additional covariate. The significance remained the same under this model (p < 0.00001).
Length changes in repeats involved in protein-protein interactions may affect the evolution of cellular signaling pathways [51]. This process may be facilitated by an absence of selective constraint on the repeat if there are no deleterious effects on the phenotype. An elevation of sequence variability at the population level in these proteins is similarly consistent with lack of evolutionary constraint on the protein regions. Previous work has shown that for polyglutamine repeats between human and mouse, there is an association between new repeats and a high nonsynonymous sequence divergence rate, corresponding to regions of low purifying selection [52]. Further investigation of the classes of repeats that are polymorphic in different groups of genes is of interest [53] but sample sizes are too limited to draw strong inferences.
We investigated in more detail the 45 variant clusters linked to 'protein-binding'. Investigation of the daughter GO terms did not reveal any striking association with any subcategory (data not shown). A number of clusters corresponding to this category have previously been described to be associated with disease, particularly trinucleotide-repeat expansion diseases [54,55]. The existence of repeats in protein- and DNA-binding proteins has been linked to their functional roles [56-60]. The question is whether the polymorphisms in these repeats are likely to have a functional impact. There are two models that may explain the higher level of polymorphism. One is that these proteins are typically under low selective constraint, as repetitive regions in protein- and DNA-binding proteins are often substantially structurally disordered [60] and expansion is unlikely to destabilize the protein's overall folding. Supporting this is the observation that new repeats emerge in regions of proteins that are subject to lower-than-average levels of purifying selection [52]. The second model is that such polymorphisms are promoted by balancing selection or recent selection for adaptive change. In the dog, evidence has been found of repeat conservation across mammalian orders despite high mutation rates, suggesting strong stabilizing selection acting on these loci. In addition, it has been found that morphological differences between breeds of dog correlated with variations in repeat number [61]. Thus, in the presence of strong selection, significant repeat polymorphism can arise.
Overall incidence of tandem repeat polymorphism
We noted that our estimate of polymorphism was higher when only clusters with a larger sample size were used (for example, 3.06% among 3,331 tandem repeats for which the UniGene cluster size was at least 200 sequences), indicating that our overall estimate is a lower estimate of the true frequency. Wren et al. [14] predicted that around 92% of polymorphic repeats in protein-coding regions would be a multiple of three nucleotides, which is concordant with the observation seen in Figure 1b. They experimentally confirmed 40% (17/42) of their predicted polymorphic protein-coding repeats within a sample of at least 60 chromosomes. Of the 249 unique repeat polymorphisms detected in our analysis, 56% were below the minimum threshold used by Wren et al. to predict polymorphism. Thus, while the method of Wren et al. is a useful prediction algorithm, it fails to predict many observed polymorphisms in shorter tandem arrays. Predicted polymorphism reflects the consequences of mutation, while actual polymorphism reflects the combination of mutation and subsequent selection pressures, and therefore the two approaches may well lead to different conclusions.
It is not surprising that a purely computational prediction will have false negatives, as it must protect against the problem of predicting too many false positives. We make the following assumptions: first, the Wren et al. prediction method only provides coverage of 44% (standard error 0.03) of tandem repeat polymorphisms, given that 56% of our variants were below their thresholds for polymorphism prediction; second, only 40% (standard error 0.08) of predicted repeats are actually polymorphic; third, there is one computationally predicted polymorphic tandem repeat per 23,000 nucleotides of protein-coding DNA [14]; and fourth, the average length of protein-coding DNA is 1,666 nucleotides (based on the UniGene dataset analyzed here). This then implies a revised estimate of estimated polymorphic tandem repeat copy-number variation to 1 in 25,000 nucleotides (with a 95% confidence interval of 17,911-43,066) [62], and that the average frequency of polymorphic tandem repeats in human proteins is 6%. The existence of annotation and experimental error may bias this upwards, while the existence of nonsense-mediated RNA decay may bias the estimate downwards.
Since 14.24% (42/295) of the polymorphisms were not a multiple of three nucleotides, up to 1% of proteins may contain frameshifting tandem repeat polymorphisms. It is likely that a much greater number of genes contain rarer frameshifting copy-number variants below the 1% frequency threshold used to define polymorphisms [63].
Our analysis confirms that tandem repeat variation is an important source of variation in many proteins. Much of this variation is of potential relevance to protein function and disease. A more thorough evaluation of the frequency of coding-sequence tandem repeat polymorphism will be possible once the resequencing of human exons from a panel of individuals becomes available. This will allow an unbiased assessment of the extent of common frameshifting tandem repeat variants. However, characterization of the frequency of rarer frameshifting tandem repeats will require larger sample sizes than typical current resequencing projects, as many repeats with large biological effects, such as frameshifts, are likely to occur at low frequencies. Thus, extensive resequencing or genotyping through large cohorts of individuals will be required in order to define their true incidence and to provide a clearer picture of the balance of mutational and selection pressures acting on the generation, fixation and elimination of tandem repeat copy-number variants in human genes.
Materials and methods
Detection of tandem repeats
Two files, Hs.seq.uniq and Hs.seq.all, from the UniGene database [15] build 172 were downloaded. Hs.seq.uniq was used as the template for tandem repeat detection and consisted of one sequence per UniGene cluster that contained the longest region of high-quality sequence data. Hs.seq.all consisted of a redundant set of gene-orientated sequences - that is, multiple sequences can correspond to the same gene cluster identifier. Tandem repeats detected in Hs.seq.uniq were defined as the queries. Tandem repeat blocks detected in Hs.seq.all using the queries were defined as the hits.
To ensure that there was no significant bias arising from expressed sequence tags (ESTs) of cancerous origin, we eliminated these sequences from our results by using the TissueInfo [64] classification of EST libraries (December 2002).
Tandem repeats are often complex patterns and it was found that repeats were often detected as smaller sub-patterns when using a lower minimum score to report a repeat. This occurred for the 69-nucleotide repeat in MUC2 for instance, where the repeat unit was detected as a series of six- and three-nucleotide repeat units. As we wanted to detect the largest range of repeats possible while retaining repeat patterns that were correct, we decided to retain all repeats detected under default parameter settings and then to search for repeats using more sensitive parameters. Only repeats detected in the latter search that did not overlap with those in the former were included.
Tandem repeats were first detected in Hs.seq.uniq using the Tandem Repeats Finder (TRF) program version 3.21 [65] with default parameters for repeat detection. A minscore of 12 instead of 50 was used the second time round, which corresponds to a minimum of three copies of a 2-nucleotide repeat as an example. The TRF detection cutoff of 12 was deliberately chosen to be low: this was motivated by the desire to determine the level of repeat variation in all repeats, regardless of their mutational origin. Thus, of the repeats we investigated, 98% (87,787/89,243) had scores below the TRF default score of 50. Of the variants detected, 67% (167/249) had a TRF score below 50. Thus, searches for variant tandem repeats need to consider low copy-number repeats, as well as those high copy-number repeats which are more likely to be variant. For shorter arrays to be reported by TRF, they will need to be 100% homogeneous to be detectable. Clearly, there may be other insertions or deletions among short inexact repeat arrays that we have not detected. Sequences lacking 25 nucleotides of flanking sequence on both sides of the detected tandem repeat block were omitted from further analysis.
We restricted our analysis to variability among protein-coding repeat sequences. Definitions of coding sequence (CDS) start and stop points were taken from the sequence header of the Hs.seq.uniq sequences in UniGene. Sequences lacking CDS information and tandem repeat sequences that did not lie exclusively within coding regions were not included. Mononucleotide tandem repeats were excluded from the analysis, as we considered the probability of detecting sequence errors too great [66].
Detection of tandem repeat variation
Similarity of the tandem repeat region within the Hs.seq.uniq representative to the same region within other sequences within the cluster was assessed by matching up the corresponding sequences using their 25-nucleotide flanks. Length differences were detected by comparing the length of the representative tandem repeat block to that of the other sequences in the cluster.
Detected repeat blocks thus have the following properties: a 25-nucleotide flanking sequence on both sides (which is used to align repeat blocks from different sequences in the cluster), and they belong to a cluster containing more than one sequence overlapping the tandem repeat sequence block and its 25-nucleotide flanks.
Detected variants were screened to ensure that they represented length variation arising as copy-number differences in genomic DNA rather than intron retention or alternative splicing: Only length variations that corresponded to a length difference that was a multiple of the repeat unit were selected. For this set, tandem repeats were detected in the variant sequence and checked to ensure that the observed copy-number agreed with the expected one, given the length of the hit block and the length of the repeat unit.
We calculated the gene diversity (or heterozygosity) as
where Pi is the frequency of the ith of k repeat lengths at a locus ([67] and see Additional data file 2).
Gene Ontology (GO) data
To test the hypothesis that the number of genes with tandem repeat variation is elevated in genes involved in defense-related processes, the term 'defense response' (GO:0006952) was selected from GO. Human UniGene clusters linked to GO terms and their hierarchies were obtained by linking LocusLink to both UniGene and GO and also by linking UniGene to EMBL and then linking, via the EMBL accessions, to UniProt and thence to GO. Links were subsequently completed by adding links to all parent GO terms for each GO term using the GO_GRAPH_PATH and GO_TERM tables from the Gene Ontology database (dated 1 July 2004). By cross-referencing our GO term of interest with the file linking GO to UniGene, we were able to assign a binary classification (yes/no related to our GO term of interest) to each UniGene cluster. This allowed us to statistically assess the differences in the levels of variation between genes related and not related to the defense response. Significant terms were corrected for multiple testing using the Bonferroni method. Statistical analysis was carried out in STATA 8.
Additional data files
The following additional data is available with the online version of this paper. Additional data file 1 is a table listing the 295 repeat variants (spanning 218 UniGene clusters) detected in our analysis, with information on the repeats and a description of the cluster representative sequence. Additional data file 2 contains block lengths of repeats grouped into 249 unique repeat loci. For each locus, the heterozygosity of the repeat length allele frequencies has been calculated. Additional data file 3 contains data used for Figure 3. Counts of variant and invariant repeats of different unit lengths and copy-numbers are tabulated.
Supplementary Material
Additional data file 1
295 repeat variants (spanning 218 UniGene clusters) detected in our analysis, with information on the repeats and a description of the cluster representative sequence.
Click here for file
Additional data file 2
Block lengths of repeats grouped into 249 unique repeat loci. For each locus, the heterozygosity of the repeat length allele frequencies has been calculated.
Click here for file
Additional data file 3
Data used for Figure 3. Counts of variant and invariant repeats of different unit lengths and copy-numbers are tabulated.
Click here for file
Acknowledgements
We thank Philip Cotter and Kate Johnston for comments and suggestions during the preparation of this manuscript, and Patrick Dicker for advice on statistical tests used. The work was supported by the Higher Education Authority of Ireland.
Figures and Tables
Figure 1 Frequency of variant and invariant repeats. (a) Histogram of the frequencies of different length repeat units in the dataset. Repeats that are multiples of three occur at greater frequency across both variant and non-variant repeats. Mononucleotide repeats were not included in the analysis. Variants represent differences between the representative and the alleles that are a multiple of the unit length and consistent with a change in repeat copy-number. N, number of identified length variants (295 variants observed in 249 tandem repeats in 218 genes). For the non-variant repeats, N represents the number of unique invariant repeats. The x-axis is on a logarithmic scale. (b) Breakdown of repeat variants by the type of variant. Unit lengths 2 to 20 are shown here, encompassing 288 of the 295 variants. Areas in black above bars 2 and 4 represent variants of units this length that are also a multiple of three.
Figure 2 Weighted scatter-plot of the pattern of detected tandem repeat length variation. Length of repeat unit is plotted against the absolute difference between query and hit repeat block lengths. One variant corresponding to a length difference of 144 for a 48-nucleotide repeat has been omitted. Note that the length of repeat unit, rather than the tandem repeat array length, is plotted on the x-axis and most observed length differences are multiples of the corresponding unit length. The area of each circle is proportional to number of variants observed with a given unit length, and a given nucleotide difference between the representative and variant sequences.
Figure 3 Distribution of copy-numbers of tandem repeats. The x-axis indicates the number of tandem repeat loci of a given unit length (indicated by color key) and with a given copy-number (indicated on the x-axis, rounded to the nearest whole number). (a) Non-variants, N = 88,850; (b) variants, N = 249; copy-number for variants represents the average copy-number among variants.
Table 1 GO analysis of repeat variants
Term GO id Variants Non-variants Statistical tests
Linked Not linked Linked Not linked
(a) Primary hypothesis
Defense response GO:0006952 9 150 475 7,979 Chi-squared test 0.98 Logistic regression 0.88
(b) All level 4
Fishers exact Logistic regression
Fishers exact Bonferroni Logistic regression Bonferroni
Most significant terms
Protein binding GO:0005515 45 114 1,354 7,100 <0.00001 <0.0006 <0.00001 <0.0006
Morphogenesis GO:0009653 28 131 783 7,671 0.001 0.064 0.001 0.064
Intracellular GO:0005622 93 66 3,845 4,609 0.001 0.064 0.004 0.256
Transcription cofactor activity GO:0003712 10 149 166 8,288 0.002 0.128 <0.00001 <0.0006
RNA polymerase II transcription factor activity GO:0003702 9 150 150 8,304 0.003 0.192 0.001 0.064
Protein serine/threonine phosphatase complex GO:0008287 3 156 19 8,435 0.007 0.448 <0.00001 <0.0006
Helicase activity GO:0004386 6 153 93 8,361 0.01 0.64 0.007 0.448
Structural constituent of epidermis GO:0030280 2 157 7 8,447 0.011 0.704 0.001 0.064
Regulation of physiological process GO:0050791 37 122 1,339 7,115 0.016 1.024 0.014 0.896
Death GO:0016265 11 148 272 8,182 0.02 1.28 0.01 0.64
Pattern specification GO:0007389 2 157 17 8,437 0.047 3.008 0.013 0.832
Antigen binding GO:0003823 2 157 18 8,436 0.052 3.328 0.021 1.344
==== Refs
A novel gene containing a trinucleotide repeat that is expanded and unstable on Huntington's disease chromosomes. The Huntington's Disease Collaborative Research Group. Cell 1993 72 971 983 8458085 10.1016/0092-8674(93)90585-E
Verkerk AJ Pieretti M Sutcliffe JS Fu YH Kuhl DP Pizzuti A Reiner O Richards S Victoria MF Zhang FP Identification of a gene (FMR-1) containing a CGG repeat coincident with a breakpoint cluster region exhibiting length variation in fragile X syndrome. Cell 1991 65 905 914 1710175 10.1016/0092-8674(91)90397-H
Hui J Stangl K Lane WS Bindereif A HnRNP L stimulates splicing of the eNOS gene by binding to variable-length CA repeats. Nat Struct Biol 2003 10 33 37 12447348 10.1038/nsb875
Gebhardt F Zanker KS Brandt B Modulation of epidermal growth factor receptor gene transcription by a polymorphic dinucleotide repeat in intron 1. J Biol Chem 1999 274 13176 13180 10224073 10.1074/jbc.274.19.13176
Jeffreys AJ Royle NJ Wilson V Wong Z Spontaneous mutation rates to new length alleles at tandem-repetitive hypervariable loci in human DNA. Nature 1988 332 278 281 3347271 10.1038/332278a0
Jakupciak JP Wells RD Genetic instabilities in (CTG.CAG) repeats occur by recombination. J Biol Chem 1999 274 23468 23479 10438526 10.1074/jbc.274.33.23468
Richard GF Dujon B Haber JE Double-strand break repair can lead to high frequencies of deletions within short CAG/CTG trinucleotide repeats. Mol Gen Genet 1999 261 871 882 10394925 10.1007/s004380050031
La Spada AR Wilson EM Lubahn DB Harding AE Fischbeck KH Androgen receptor gene mutations in X-linked spinal and bulbar muscular atrophy. Nature 1991 352 77 79 2062380 10.1038/352077a0
Sutherland GR Richards RI Simple tandem DNA repeats and human genetic disease. Proc Natl Acad Sci USA 1995 92 3636 3641 7731957
Kenny D Muckian C Fitzgerald DJ Cannon CP Shields DC Platelet glycoprotein Ib alpha receptor polymorphisms and recurrent ischaemic events in acute coronary syndrome patients. J Thromb Thrombolysis 2002 13 13 19 11994555 10.1023/A:1015307823578
Holmer SR Hengstenberg C Kraft HG Mayer B Poll M Kurzinger S Fischer M Lowel H Klein G Riegger GA Schunkert H Association of polymorphisms of the apolipoprotein(a) gene with lipoprotein(a) levels and myocardial infarction. Circulation 2003 107 696 701 12578871 10.1161/01.CIR.0000048125.79640.77
Bugert P Hoffmann MM Winkelmann BR Vosberg M Jahn J Entelmann M Katus HA Marz W Mansmann U Boehm BO The variable number of tandem repeat polymorphism in the P-selectin glycoprotein ligand-1 gene is not associated with coronary heart disease. J Mol Med 2003 81 495 501 12879153 10.1007/s00109-003-0459-2
Fondon JW 3rdMele GM Brezinschek RI Cummings D Pande A Wren J O'Brien KM Kupfer KC Wei MH Lerman M Computerized polymorphic marker identification: experimental validation and a predicted human polymorphism catalog. Proc Natl Acad Sci USA 1998 95 7514 7519 9636181 10.1073/pnas.95.13.7514
Wren JD Forgacs E Fondon JW 3rdPertsemlidis A Cheng SY Gallardo T Williams RS Shohet RV Minna JD Garner HR Repeat polymorphisms within gene regions: phenotypic and evolutionary implications. Am J Hum Genet 2000 67 345 356 10889045 10.1086/303013
Schuler GD Boguski MS Stewart EA Stein LD Gyapay G Rice K White RE Rodriguez-Tome P Aggarwal A Bajorek E A gene map of the human genome. Science 1996 274 540 546 8849440 10.1126/science.274.5287.540
Denoeud F Vergnaud G Benson G Predicting human minisatellite polymorphism. Genome Res 2003 13 856 867 12695323 10.1101/gr.574403
Naslund K Saetre P von Salome J Bergstrom TF Jareborg N Jazin E Genome-wide prediction of human VNTRs. Genomics 2005 85 24 35 15607419 10.1016/j.ygeno.2004.10.009
Denoeud F Vergnaud G Identification of polymorphic tandem repeats by direct comparison of genome sequence from different bacterial strains: a web-based resource. BMC Bioinformatics 2004 5 4 14715089 10.1186/1471-2105-5-4
Jordon P Snyder LA Saunders NJ Diversity in coding tandem repeats in related Neisseria spp. BMC Microbiol 2003 3 23 14611665 10.1186/1471-2180-3-23
Sylvestre P Couture-Tosi E Mock M Polymorphism in the collagen-like region of the Bacillus anthracis BclA protein leads to variation in exosporium filament length. J Bacteriol 2003 185 1555 1563 12591872 10.1128/JB.185.5.1555-1563.2003
van Belkum A Scherer S van Alphen L Verbrugh H Short-sequence DNA repeats in prokaryotic genomes. Microbiol Mol Biol Rev 1998 62 275 293 9618442
Li YC Korol AB Fahima T Nevo E Microsatellites within genes: structure, function, and evolution. Mol Biol Evol 2004 21 991 1007 14963101 10.1093/molbev/msh073
Murphy PM Molecular mimicry and the generation of host defense protein diversity. Cell 1993 72 823 826 8458078 10.1016/0092-8674(93)90571-7
Shields DC Harmon DL Whitehead AS Evolution of hemopoietic ligands and their receptors. Influence of positive selection on correlated replacements throughout ligand and receptor proteins. J Immunol 1996 156 1062 1070 8557980
Metzgar D Bytof J Wills C Selection against frameshift mutations limits microsatellite expansion in coding DNA. Genome Res 2000 10 72 80 10645952
Dokholyan NV Buldyrev SV Havlin S Stanley HE Distributions of dimeric tandem repeats in noncoding and coding DNA sequences. J Theor Biol 2000 202 273 282 10666360 10.1006/jtbi.1999.1052
Subramanian S Mishra RK Singh L Genome-wide analysis of microsatellite repeats in humans: their abundance and density in specific genomic regions. Genome Biol 2003 4 R13 12620123 10.1186/gb-2003-4-2-r13
Chang FM Kidd JR Livak KJ Pakstis AJ Kidd KK The world-wide distribution of allele frequencies at the human dopamine D4 receptor locus. Hum Genet 1996 98 91 101 8682515 10.1007/s004390050166
Afshar-Kharghan V Diz-Kucukkaya R Ludwig EH Marian AJ Lopez JA Human polymorphism of P-selectin glycoprotein ligand 1 attributable to variable numbers of tandem decameric repeats in the mucinlike region. Blood 2001 97 3306 3307 11342464 10.1182/blood.V97.10.3306
Toribara NW Gum JR JrCulhane PJ Lagace RE Hicks JW Petersen GM Kim YS MUC-2 human small intestinal mucin gene structure. Repeated arrays and polymorphism. J Clin Invest 1991 88 1005 1013 1885763
Muckian C Hillmann A Kenny D Shields DC A novel variant of the platelet glycoprotein Ibalpha macroglycopeptide region lacks any copies of the 'perfect' 13 amino acid repeat. Thromb Haemost 2000 83 513 514 10744166
Matsuyama Z Kawakami H Maruyama H Izumi Y Komure O Udaka F Kameyama M Nishio T Kuroda Y Nishimura M Nakamura S Molecular features of the CAG repeats of spinocerebellar ataxia 6 (SCA6). Hum Mol Genet 1997 6 1283 1287 9259274 10.1093/hmg/6.8.1283
Koide R Kobayashi S Shimohata T Ikeuchi T Maruyama M Saito M Yamada M Takahashi H Tsuji S A neurological disease caused by an expanded CAG trinucleotide repeat in the TATA-binding protein gene: a new polyglutamine disease? Hum Mol Genet 1999 8 2047 2053 10484774 10.1093/hmg/8.11.2047
Zuhlke C Hellenbroich Y Dalski A Kononowa N Hagenah J Vieregge P Riess O Klein C Schwinger E Different types of repeat expansion in the TATA-binding protein gene are associated with a new form of inherited ataxia. Eur J Hum Genet 2001 9 160 164 11313753 10.1038/sj.ejhg.5200617
Dean M Park M Le Beau MM Robins TS Diaz MO Rowley JD Blair DG Vande Woude GF The human met oncogene is related to the tyrosine kinase oncogenes. Nature 1985 318 385 388 4069211 10.1038/318385a0
Stenson PD Ball EV Mort M Phillips AD Shiel JA Thomas NS Abeysinghe S Krawczak M Cooper DN Human Gene Mutation Database (HGMD): 2003 update. Hum Mutat 2003 21 577 581 12754702 10.1002/humu.10212
Hamosh A Scott AF Amberger JS Bocchini CA McKusick VA Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders. Nucleic Acids Res 2005 33 Database issue D514 D517 15608251 10.1093/nar/gki033
Letovsky SI Cottingham RW Porter CJ Li PW GDB: the Human Genome Database. Nucleic Acids Res 1998 26 94 99 9399808 10.1093/nar/26.1.94
Charmley P Concannon P Hood L Rowen L Frequency and polymorphism of simple sequence repeats in a contiguous 685-kb DNA sequence containing the human T-cell receptor beta-chain gene complex. Genomics 1995 29 760 765 8575771 10.1006/geno.1995.9940
Kimmel M Chakraborty R Stivers DN Deka R Dynamics of repeat polymorphisms under a forward-backward mutation model: within- and between-population variability at microsatellite loci. Genetics 1996 143 549 555 8722803
Ota T Kimura M A model of mutation appropriate to estimate the number of electrophoretically detectable alleles in a finite population. Genet Res 1973 22 201 204 4777279
Schlotterer C Tautz D Slippage synthesis of simple sequence DNA. Nucleic Acids Res 1992 20 211 215 1741246
Weber JL Informativeness of human (dC-dA)n.(dG-dT)n polymorphisms. Genomics 1990 7 524 530 1974878 10.1016/0888-7543(90)90195-Z
Kunst CB Leeflang EP Iber JC Arnheim N Warren ST The effect of FMR1 CGG repeat interruptions on mutation frequency as measured by sperm typing. J Med Genet 1997 34 627 631 9279752
Hubbard T Andrews D Caccamo M Cameron G Chen Y Clamp M Clarke L Coates G Cox T Cunningham F Ensembl 2005. Nucleic Acids Res 2005 33 Database issue D447 D453 15608235 10.1093/nar/gki138
Lykke-Andersen J Shu MD Steitz JA Human Upf proteins target an mRNA for nonsense-mediated decay when bound downstream of a termination codon. Cell 2000 103 1121 1131 11163187 10.1016/S0092-8674(00)00214-2
Hughes AL Packer B Welch R Bergen AW Chanock SJ Yeager M Widespread purifying selection at polymorphic sites in human protein-coding loci. Proc Natl Acad Sci USA 2003 100 15754 15757 14660790 10.1073/pnas.2536718100
van Den Hurk WH Willems HJ Bloemen M Martens GJ Novel frameshift mutations near short simple repeats. J Biol Chem 2001 276 11496 11498 11139590 10.1074/jbc.M011040200
Karlin S Burge C Trinucleotide repeats and long homopeptides in genes and proteins associated with nervous system disease and development. Proc Natl Acad Sci USA 1996 93 1560 1565 8643671 10.1073/pnas.93.4.1560
Ashburner M Ball CA Blake JA Botstein D Butler H Cherry JM Davis AP Dolinski K Dwight SS Eppig JT Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 2000 25 25 29 10802651 10.1038/75556
Hancock JM Simon M Simple sequence repeats in proteins and their significance for network evolution. Gene 2005 345 113 118 15716087 10.1016/j.gene.2004.11.023
Hancock JM Worthey EA Santibanez-Koref MF A role for selection in regulating the evolutionary emergence of disease-causing and other coding CAG repeats in humans and mice. Mol Biol Evol 2001 18 1014 1023 11371590
Alba MM Laskowski RA Hancock JM Detecting cryptically simple protein sequences using the SIMPLE algorithm. Bioinformatics 2002 18 672 678 12050063 10.1093/bioinformatics/18.5.672
Koide R Ikeuchi T Onodera O Tanaka H Igarashi S Endo K Takahashi H Kondo R Ishikawa A Hayashi T Unstable expansion of CAG repeat in hereditary dentatorubral-pallidoluysian atrophy (DRPLA). Nat Genet 1994 6 9 13 8136840 10.1038/ng0194-9
Kennedy WR Alter M Sung JH Progressive proximal spinal and bulbar muscular atrophy of late onset. A sex-linked recessive trait. Neurology 1968 18 671 680 4233749
Hamada H Seidman M Howard BH Gorman CM Enhanced gene expression by the poly(dT-dG).poly(dC-dA) sequence. Mol Cell Biol 1984 4 2622 2630 6098815
Lu Q Wallrath LL Granok H Elgin SC (CT)n (GA)n repeats and heat shock elements have distinct roles in chromatin structure and transcriptional activation of the Drosophila hsp26 gene. Mol Cell Biol 1993 13 2802 2814 8474442
Yee HA Wong AK van de Sande JH Rattner JB Identification of novel single-stranded d(TC)n binding proteins in several mammalian species. Nucleic Acids Res 1991 19 949 953 2017376
Richards RI Holman K Yu S Sutherland GR Fragile X syndrome unstable element, p(CCG)n, and other simple tandem repeat sequences are binding sites for specific nuclear proteins. Hum Mol Genet 1993 2 1429 1435 8242066
Colafranceschi M Colosimo A Zbilut JP Uversky VN Giuliani A Structure-related statistical singularities along protein sequences: a correlation study. J Chem Inf Model 2005 45 183 189 15667144
Fondon JW 3rdGarner HR Molecular origins of rapid and continuous morphological evolution. Proc Natl Acad Sci USA 2004 101 18058 18063 15596718 10.1073/pnas.0408118101
Armitage P Berry G Statistical Methods in Medical Research 1994 3 Oxford, UK: Blackwell Science
Day IN Alharbi KK Smith M Aldahmesh MA Chen X Lotery AJ Pante-de-Sousa G Hou G Ye S Eccles D Paucimorphic alleles versus polymorphic alleles and rare mutations in disease causation: theory, observation and detection. Curr Genomics 2004 5 431 438 10.2174/1389202043349156
Skrabanek L Campagne F TissueInfo: high-throughput identification of tissue expression profiles and specificity. Nucleic Acids Res 2001 29 E102 11691939 10.1093/nar/29.21.e102
Benson G Tandem repeats finder: a program to analyze DNA sequences. Nucleic Acids Res 1999 27 573 580 9862982 10.1093/nar/27.2.573
Weber JL David D Heil J Fan Y Zhao C Marth G Human diallelic insertion/deletion polymorphisms. Am J Hum Genet 2002 71 854 862 12205564 10.1086/342727
Weir BS Genetic Data Analysis II: Methods for Discrete Population Genetic Data 1996 2 Sunderland, MA: Sinauer
|
16086851
|
PMC1273636
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Jul 28; 6(8):R69
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r69
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r701608685210.1186/gb-2005-6-8-r70ResearchEvidence for a second class of S-adenosylmethionine riboswitches and other regulatory RNA motifs in alpha-proteobacteria Corbino Keith A [email protected] Jeffrey E [email protected] Jinsoo [email protected] Rüdiger [email protected] Brian J [email protected] Izabela [email protected] Maumita [email protected] Noam D [email protected] Ronald R [email protected] Department of Molecular, Cellular and Developmental Biology, Yale University, P.O. Box 208103, New Haven, CT 06520-8103, USA2 Department of Molecular Biophysics and Biochemistry, Yale University, P.O. Box 208103, New Haven, CT 06520-8103, USA3 Department of Chemistry, Yale University, P.O. Box 208103, New Haven, CT 06520-8103, USA4 Department of Physics, University of California, Berkeley, CA 94720-7200, USA2005 1 8 2005 6 8 R70 R70 28 4 2005 15 6 2005 1 7 2005 Copyright © 2005 Corbino et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Comparative sequence analysis and structural probing identified five RNA elements in the intergenic region of Agrobacterium tumefaciens and other α-proteobacteria. One of these RNA elements is probably a SAM-II, the only riboswitch class identified so far that is not found in Gram-positive bacteria.
Background
Riboswitches are RNA elements in the 5' untranslated leaders of bacterial mRNAs that directly sense the levels of specific metabolites with a structurally conserved aptamer domain to regulate expression of downstream genes. Riboswitches are most common in the genomes of low GC Gram-positive bacteria (for example, Bacillus subtilis contains examples of all known riboswitches), and some riboswitch classes seem to be restricted to this group.
Results
We used comparative sequence analysis and structural probing to identify five RNA elements (serC, speF, suhB, ybhL, and metA) that reside in the intergenic regions of Agrobacterium tumefaciens and many other α-proteobacteria. One of these, the metA motif, is found upstream of methionine biosynthesis genes and binds S-adenosylmethionine (SAM). This natural aptamer most likely functions as a SAM riboswitch (SAM-II) with a consensus sequence and structure that is distinct from the class of SAM riboswitches (SAM-I) predominantly found in Gram-positive bacteria. The minimal functional SAM-II aptamer consists of fewer than 70 nucleotides, which form a single stem and a pseudoknot. Despite its simple architecture and lower affinity for SAM, the SAM-II aptamer strongly discriminates against related compounds.
Conclusion
SAM-II is the only metabolite-binding riboswitch class identified so far that is not found in Gram-positive bacteria, and its existence demonstrates that biological systems can use multiple RNA structures to sense a single chemical compound. The two SAM riboswitches might be 'RNA World' relics that were selectively retained in certain bacterial lineages or new motifs that have emerged since the divergence of the major bacterial groups.
==== Body
Background
Riboswitches are structured RNA elements within the noncoding regions of some mRNAs that directly sense metabolites and regulate gene expression [1-4]. Riboswitches are known that respond to a wide range of metabolites including coenzymes [5-8], purines [9,10], amino acids [11,12], and a sugar-phosphate compound [13]. Most riboswitches are found within the 5' untranslated regions of bacterial mRNAs that encode biosynthetic enzymes or metabolite transporters. Ligand binding to the aptamer domain of a riboswitch stabilizes specific structural elements of an adjoining expression platform, which modulates the expression of downstream genes. The two most common types of expression platforms control either the formation of intrinsic transcription terminators that abort mRNA synthesis or the formation of alternate structures that mask ribosome-binding sites to prevent translation initiation.
Riboswitch aptamers have sequence and structural features that are typical of functional RNAs. Each riboswitch class is defined by a core of conserved base-paired elements and consensus nucleotides at specific positions interspersed with variable stems and loops. We have previously used comparative sequence analysis of intergenic regions (IGRs) from 94 microbial genomes to identify conserved RNA motifs residing upstream of functionally related genes in Bacillus subtilis that are candidates for new riboswitches [14]. Two of these RNA elements have subsequently proven to be novel riboswitch classes. Candidate RNAs termed glmS and gcvT function as glucosamine-6-phosphate dependent ribozymes [13] and cooperative glycine riboswitches [12], respectively.
Most riboswitches reported previously are found predominantly in Gram-positive bacteria, and representatives of all classes are present in B. subtilis. We speculated that other groups of bacteria might harbor different noncoding RNA domains, some of which could be novel riboswitches. We report here five novel structured RNA elements that were identified by focusing our comparative sequence analysis of IGRs on α-proteobacterial genomes. One of the five newfound motifs from Agrobacterium tumefaciens, termed metA, appears to function as a riboswitch that senses S-adenosylmethionine (SAM). This SAM-II riboswitch class has a consensus sequence and conserved structure that is distinct from the SAM-I riboswitch reported previously [15-18]. Compared with SAM-I aptamers, SAM-II aptamers are smaller and form a simpler secondary structure. However, the SAM-II aptamer exhibits a level of molecular discrimination that is similar to that observed for the SAM-I riboswitch. These findings demonstrate that biological systems use multiple RNA motifs to sense the same chemical compound.
Results and discussion
Identification of novel RNA motifs in α-proteobacteria
We searched α-proteobacterial genomes for new riboswitches and structured regulatory RNA elements by constructing a database of sequence comparisons between IGRs from 116 complete microbial genomes [19] (See also [14] and Materials and methods). We examined alignments and statistics from this database for examples where a conserved sequence motif occurred upstream of genes sharing a common function in different organisms. This initial screen encountered some α-proteobacterial sequence elements that had been previously described, including an ilvB leader peptide [20] and long repeat elements [21,22]. Other putative regulatory elements were further evaluated for their potential to form RNA structures by creating a secondary structure model and iteratively searching for additional matches. In the end, we identified five motifs specific to α-proteobacteria that are likely to be structured RNAs (Figure 1).
We experimentally corroborated our secondary structure models for these conserved RNA elements using in-line probing [23]. In this assay, the extent of spontaneous cleavage at each internucleotide linkage in an RNA molecule is determined by separating 5'-radiolabeled degradation products on a polyacrylamide gel. RNA cleavage occurs most rapidly at sites where nucleophilic attack by the 2' oxygen of a ribose approaches an 'in-line' geometry with respect to the phosphorus atom and adjoining 5' oxygen leaving group. Typically, linkages next to base-paired nucleotides in a structured RNA are rigidly held in a conformation that does not permit the formation of an in-line geometry, and therefore these sites cleave slowly. In contrast, internucleotide linkages that are in flexible regions of an RNA molecule occasionally sample an in-line geometry and are cleaved more rapidly. Therefore, regions with relatively low levels of degradation product in an in-line probing gel typically correspond to base-paired or other structured regions of an RNA.
Complete formatted sequence alignments, compilations of downstream genes, consensus structures, and in-line probing data for the five motifs are available (Additional data file 1). Sequence alignments of each RNA motif are also provided in Stockholm format (Additional data files 2, 3, 4, 5, 6) and have been deposited in the Rfam database [24].
The serC element
The short serC RNA element (Rfam: RF00517) consists of two conserved base-paired stems. Putative transcription start sites associated with near-consensus upstream promoter elements directly precede all examples of this motif, and the start codon for the serC gene is at most 11 nucleotides downstream of the final hairpin. This arrangement suggests that formation of the final hairpin would repress translation by sequestering the ribosome-binding site within the 3' side its base-paired stem and GNRA tetraloop. In-line probing of an RNA corresponding to nucleotides -46 to +11 relative to the serC start codon in A. tumefaciens (GenBank: NC_003305.1; nucleotides 788249 to 788193) supports this structure.
The serC motif is located upstream of an operon encoding serine transaminase (SerC) and phosphoglycerate dehydrogenase (SerA) in many α-proteobacteria. Together, these enzymes convert 3-phosphoglycerate into 3-phosphoserine during the first two steps of serine biosynthesis. SerC can also catalyze a related step in pyridoxal 5'-phosphate (PLP) biosynthesis involving a similar substrate. We have tested whether L-serine, L-threonine, PLP, pyridoxal, pyridoxine, pyridoxamine, or 4-pyridoxic acid are capable of directly binding to the A. tumefaciens RNA. None of these compounds have any effect on RNA structure as judged by in-line probing (data not shown). It is possible that an RNA-binding protein could be responsible for sensing a relevant metabolite, binding to the relatively small serC element, and derepressing translation. The PyrR protein performs a similar regulatory role for pyrimidine biosynthesis genes in B. subtilis [25].
The speF element
The extended speF element (Rfam: RF00518) is found upstream of proteins classified into COG0019 in several α-proteobacteria. Primary sequence conservation begins at the 5' end near a putative transcription start site and continues into a base-paired stem that is topped with a large insertion that can form a four-stem junction in some representatives. Following this stem, a stretch of around 80 conserved nucleotides appears to fold into a long bulged stem-loop. This model is tentatively supported by covariation at a few positions in the alignment, except for the outermost putative pairing elements where the sequence is absolutely conserved. The model is also supported by in-line probing patterns for the RNA corresponding to nucleotides -400 to +3 relative to the speF translation start site in A. tumefaciens (GenBank: NC_003305.1; nucleotides 205774 to 205372). There appear to be further conserved blocks of sequence within the more than 150 nucleotides remaining before the speF start codon, but we were unable to assign secondary structures there with much confidence.
Although COG0019 encodes diaminopimelate decarboxylases (lysA) in other groups of bacteria, a phylogenetic tree of protein sequences indicates that the genes downstream of this motif are orthologs of B. subtilis speF, an ornithine decarboxylase enzyme that catalyzes one of the first steps in polyamine biosynthesis. We have tested whether metabolites related to this pathway bind directly to the A. tumefaciens intergenic region and cause structural changes detectable by in-line probing. There is no measurable binding of L-ornithine, L-lysine, meso-diaminopimelate, putrescine, cadaverine, or spermidine to the speF RNA construct used in this study (data not shown).
The suhB element
The suhB element (Rfam: RF00519) was originally recognized upstream of one of nine A. tumefaciens ORFs, encoding proteins with similarity to archeal fructose-1,6-bisphosphatases (COG0483). After more matches were found, it became clear that this motif was most likely not a cis-acting regulatory element for the suhB gene but was more likely to be a small noncoding RNA that is transcribed from the opposite strand relative to the suhB gene. In this orientation, each representative carries a putative promoter and intrinsic terminator flanking the conserved sequence domain. Further searches for this motif revealed that multiple copies are present in many α-proteobacterial genomes (for example, five in Bradyrhizobium japonicum and four in Caulobacter crescentus) and that it is not associated with specific neighboring genes. The only evolutionarily conserved secondary structure in the suhB noncoding RNA, aside from the terminator stem, appears to be a short helix near its 5' end. In-line probing of an RNA corresponding to a portion of one A. tumefaciens intergenic region containing this motif (GenBank:NC_003305.1; nucleotides 979721 to 979594) also indicates that its characteristically conserved sequences reside in unstructured regions, suggesting that this family could be involved in some form of antisense gene regulation or other noncoding RNA function [26].
The ybhL element
The ybhL RNA motif (Rfam: RF00520) appears to be restricted to bacteria from the Rhizobiales. In-line probing data from an RNA corresponding to nucleotides -139 to +21 relative to the translation start site of the ybhL gene in A. tumefaciens (GenBank: NC_003304.1; nucleotides 2665399 to 2665558) indicate that this element folds into a doubly-bulged hairpin of around 60 nucleotides. Sequence covariation substantiates the formation of the outermost and innermost paired stems. A putative transcription start site is located close to the beginning of the hairpin within a region that appears highly conserved with our limited number of sequence examples. This RNA motif always occurs upstream of genes related to the Escherichia coli ybhL gene (COG0670), a putative integral membrane protein. Because the function of ybhL is not known, we were unable to formulate any hypotheses for the role of this RNA element.
The metA element
The metA RNA element (Rfam: RF00521) is found in a variety of α-proteobacteria, and there are even a few occurrences in other proteobacterial lineages and the Bacteroides group. This RNA was originally identified upstream of the metA gene in A. tumefaciens, but was subsequently found preceding other genes related to methionine and S-adenosylmethionine (SAM) biosynthesis. The RNA motif is compact with a single stem (P1) and pseudoknot (P2) that are both exceptionally well supported by covariation among more than 70 representatives (Figure 2a). Usually a possible transcription start site with near-consensus -35 and -10 promoter elements is located a few nucleotides before the first nucleotide of P1.
Many representatives also contain putative intrinsic terminators between P2 and the downstream ORF. This transcription terminator arrangement is characteristic of many known riboswitches, and suggests that the metA RNA is a regulatory element that functions as a genetic OFF switch [14]. In comparison, Gram-positive bacteria make extensive use of SAM-sensing riboswitches (Figure 2b) to repress a similar collection of methionine biosynthesis genes when SAM becomes abundant in the cell (Figure 2c), often with expression platforms that use transcription termination [15-18]. With consideration of these factors, we tested whether the simpler metA motif also functions as a natural aptamer for SAM.
The metA element binds SAM
RNA constructs corresponding to nucleotides -230 to -75 relative to the translation start site of the A. tumefaciens metA gene (GenBank: NC_003304.1; nucleotides 2703291 to 2703446) were prepared by in vitro transcription. The resulting 156-nucleotide RNA (termed 156 metA) contains the majority of the intergenic region but excludes the proposed terminator stem. In-line probing assays revealed that the 156 metA structure is greatly modulated in response to SAM concentrations ranging from 1 nM to 6 mM (Figure 3a). Mapping spontaneous cleavage patterns onto the secondary structure model for 156 metA (Figure 3b) reveals that all SAM-induced changes occur within the conserved metA sequence element. There are incidents of both increased and decreased rates of spontaneous RNA cleavage, indicating that SAM does not facilitate general RNA degradation. Rather, SAM associates with 156 metA to induce a precise structure that stabilizes certain RNA regions and destabilizes others, as has been seen for all riboswitches characterized previously. An apparent Kd value of around 1 μM (Figure 3c) for the RNA-SAM complex was determined by plotting the normalized fraction of RNA cleaved in several regions against the logarithm of the SAM concentration.
These results suggested that only the conserved core of this RNA is necessary for SAM recognition. Indeed, a smaller 68-nucleotide metA RNA (68 metA) encompassing only nucleotides -161 to -94 (GenBank: NC_003304.1; nucleotides 2703360 to 2703426) binds with an affinity of around 10 μM and displays a similar change in its spontaneous cleavage pattern (data not shown). Using 68 metA, we examined the importance of the formation of the pseudoknot stem (P2) for SAM binding by making two variants (Figure 3b). One variant carries disruptive mutations (M1: U132→C, C133→G) and the other carries these mutations and the corresponding compensatory mutations (M2: M1, G94→C, A95→G). These RNAs were subjected to in-line probing in the presence of 1 mM SAM (data not shown). Under these conditions, the spontaneous cleavage pattern of M1 did not change in response to SAM. In contrast, M2 exhibited wild-type levels of structural modulation. These results are consistent with covariation in the metA sequence alignment that suggests P2 stem formation is required for SAM binding.
We obtained further proof of direct binding between SAM and the A. tumefaciens metA RNA by equilibrium dialysis. Adding 10 μM of 156 metA to one side of an equilibrium dialysis chamber containing 100 nM S-adenosyl-L-methionine-(methyl-3H) ([3H]SAM), shifted the distribution of [3H]SAM to favor the RNA side of the membrane by 2.6-fold. A greater shift was not observed because our [3H]SAM sample contained an appreciable amount of radiolabeled breakdown products (see Materials and methods). If 125 μM of unlabeled SAM or the related compound S-adenosyl-L-homocysteine (SAH) are subsequently added to similarly prepared setups, only SAM is able to compete with the [3H]SAM and shift the ratio of tritium back to 1. This result demonstrates that 156 metA strongly discriminates against the demethylated form of SAM.
The genomic distribution of the metA element and its function as a receptor for SAM are consistent with its proposed function as a SAM riboswitch. SAM-II riboswitches found in α-proteobacteria have a consensus sequence and secondary structure that are distinct from SAM-I riboswitches found in the Gram-positive bacteria. A SAM-I riboswitch (the 124 yitJ aptamer from B. subtilis) has been shown to have a Kd for SAM of ~4 nM [17]. In contrast, the minimized aptamer from the A. tumefaciens SAM-II riboswitch upstream of metA has a much poorer affinity for SAM (68 metA, Kd around 10 μM). It has been shown that in vitro selected RNA aptamers that have greater information content generally exhibit greater ligand affinity [27]. The SAM-I and SAM-II aptamers follow this general trend, as low-affinity SAM-II aptamers carry two paired elements and only 24 nucleotides that are >80% conserved (Figure 2b). In comparison, SAM-I aptamers incorporate at least four paired stems and 54 conserved nucleotides.
The poorer affinity of the SAM-II aptamer does not necessarily mean that it would exhibit inferior in vivo genetic control as a riboswitch. The physiological environments for these riboswitches may be quite different since they operate in divergent groups of bacteria. Furthermore, the kinetics of transcription and ligand binding appear to be more important than equilibrium binding constants for determining whether a flavin mononucleotide (FMN) riboswitch triggers transcription termination [28]. The Kd for the truncated SAM-II aptamer examined in this study is roughly equal to the SAM concentrations needed to trigger transcription termination by SAM-I riboswitches in vitro [15,17]. Furthermore, the affinity of the SAM-II RNA is probably more than sufficient to sense SAM at biologically relevant concentrations. Endogenous SAM levels have been estimated to range from roughly 30 μM to 200 μM in E. coli cells grown in rich media [29]. Nevertheless, the ability of the SAM-II motif to function as an efficient riboswitch might be compromised if it were less capable of discriminating against metabolites with structures similar to SAM than the SAM-I aptamer. Therefore, we investigated the molecular specificity of the SAM-II riboswitch in more detail.
Molecular recognition characteristics of the SAM-II aptamer
We performed in-line probing assays with 156 metA in the presence of various SAM analogues to measure the discrimination of the SAM-II aptamer against related metabolites (Figure 4). No RNA structure modulation was seen in the presence of 1 mM SAH, S-adenosyl-L-cysteine (SAC), or methionine (Figure 4a). A more detailed molecular recognition study (Figure 4b,c) was conducted using a variety of chemically synthesized SAM derivatives (see Materials and methods) containing systematic single substitutions of functional groups that could potentially be recognized by the SAM-II aptamer (compounds a-f). It is important to note that the biologically active form of SAM used in our initial tests has the (-) sulfonium configuration [30], while the chemically synthesized compounds are racemic (±). Only two of these compounds modulated the riboswitch structure at a concentration of 1 mM. Full titrations indicated that racemic SAM (compound a) had a roughly twofold higher Kd than (-) SAM, and the 3-deaza SAM analogue (compound e) bound with a 50-fold higher Kd.
These analogue binding studies indicate that the SAM-II aptamer creates a binding compartment that recognizes functional groups on the entire surface of SAM. SAM-II discriminates more than 1,000-fold against binding SAM analogues lacking the ribose 2'- or 3'-hydroxyl groups and SAM analogues with single substitutions of the adenine 3-aza, 6-amino, or 7-aza groups. A majority of this affinity loss probably comes from disrupting hydrogen bonds or electrostatic interactions between the aptamer and metabolite, although secondary consequences of the chemical changes, such as altering the preferred ribose sugar pucker or purine ring electronic characteristics, may also contribute to the loss in affinity. Removal of either the carboxyl or amino group from the methionyl moiety is similarly detrimental and might disrupt hydrogen bonds or electrostatic interactions that the aptamer might form with the amino acid zwitterion. Not surpisingly, the aptamer also readily discriminates against the removal of the S-methyl group that is critical for the function of SAM as a coenzyme, probably due to the accompanying loss of positive charge on the sulfonium center. Finally, shortening the methionine side chain by one methylene group prevents SAM binding, most likely because it creates a distance constraint that prevents the simultaneous recognition of the methionyl and adenosyl moieties.
We have not investigated whether the 1-aza group of adenine is required for binding, but it is possible that the Watson-Crick face of the adenine base is recognized by a canonical base pair to an aptamer uridine, like that found in the adenine riboswitch [10,31,32]. There are six uracil residues that are absolutely conserved in putatively single-stranded regions of the SAM-II riboswitch and therefore candidates for this interaction (Figure 2b). The molecular recognition determinants for ligand binding by the SAM-II aptamer are depicted in Figure 4b.
The SAM-I riboswitch binds SAH and SAC around 100- and around 10,000-fold poorer than SAM, respectively [17]. The SAM-II aptamer discriminates greater than 1,000-fold against both these compounds, and therefore SAM-II appears to be at least as sensitive to the presence of the S-methyl group as SAM-I. Further binding studies with a panel of SAM analogues modified at the sulfonium center indicate that SAM-I tolerates these changes much better than SAM-II (Lim J, Winkler WC, Nakamura S, Scott V, Breaker RR, unpublished data). We are unable to quantitate discriminations of greater than 1,000-fold against analogues for SAM-II due to its poorer overall Kd. However, our findings indicate that the smaller size of the SAM-II aptamer does not prevent it from attaining the same exquisite discrimination required for efficient genetic control that is exhibited by SAM-I riboswitches.
Conclusion
Although multiple RNA solutions to small-molecule binding challenges are often found by in vitro selection (for example, ATP aptamers; [33-35]), it is now apparent that nature also exploits the structural diversity of RNA to employ multiple, unique mRNA motifs to sense a single metabolite. The SAM-II aptamer found primarily in α-proteobacteria has a much smaller conserved structure than the aptamer of the SAM-I riboswitch from Gram-positive bacteria. Despite having an overall lower affinity for SAM, the SAM-II aptamer appears to be adapted for precise genetic control and discriminates against closely related compounds at least as well as the SAM-I aptamer.
We see two main evolutionary scenarios that could explain the modern phylogenetic distribution of the SAM-I and SAM-II RNAs. SAM, a nucleotide-containing coenzyme, is thought to be a relic of an ancient 'RNA World' when all life processes were controlled primarily by RNA [36-40]. It is possible that RNA World organisms utilized multiple different SAM aptamers for regulatory purposes or as modules incorporated into extinct ribozymes that utilized SAM as a cofactor. According to this hypothesis, the current distribution of each riboswitch might reflect the selective retention of individual classes of SAM aptamers in the progenitors of different bacterial lineages. A second possibility is that the SAM riboswitches emerged more recently and that each aptamer developed independently sometime after the main bacterial lineages diverged billions of years ago [41]. Of course, a combination of ancient and more recent evolutionary events also could account for the distribution of these and other riboswitch classes.
SAM-II is the only known riboswitch that has not been found in the genome of the Gram-positive bacterium B. subtilis. We have also identified four other RNA motifs in A. tumefaciens that appear to be restricted to other α-proteobacterial genomes. Three of these are candidates for structured mRNA elements, and they join a growing list of putative 'orphan' RNA regulatory elements [14] that might respond to unknown cellular effectors in bacteria. Regardless of the true evolutionary provenance of riboswitches, it is likely that nature employs an even wider diversity of metabolite sensing mRNAs in modern organisms.
Materials and methods
Bioinformatics
An updated version of the BLISS database [14,19] containing the results of an all-versus-all BLAST comparison of IGRs from 116 microbial genomes was used to manually examine several α-proteobacterial genomes for conserved RNA elements. The BLISS website displays alignments of homology between bacterial IGRs along with compilations of sequence statistics, species distributions, and neighboring gene function assignments from the COG database [42] in a collaborative annotation environment. The updated version of BLISS is available on the web [19]. Further matches to the five motifs were found by iterative BLAST and filtered covariance model searches [43] of unfinished bacterial genomes and environmental sequences [44]. Phylogenetic trees were constructed with CLUSTALW [45] to clarify the specific functions of some genes assigned to ambiguous COGs.
In-line probing assays
RNA preparation, radiolabeling, and in-line probing assays were performed essentially as previously described [23]. DNA templates for in vitro transcription with T7 RNA polymerase promoters were prepared by whole-cell PCR from A. tumefaciens strain GV2260, except for 68 metA RNA mutants M1 and M2 where overlapping synthetic oligonucleotides were extended with reverse transcriptase. For each in-line probing reaction, around 1 nM 5' 32P-RNA was incubated for 40-48 h in a mixture of 50 mM Tris-HCl (pH 8.3 at 25°C), 20 mM MgCl2, 100 mM KCl, and various compounds as indicated. All compounds used for in-line probing were purchased from Sigma. SAM analogues were prepared as diastereomeric mixtures by the reaction of S-adenosylhomocysteine derivatives [46,47] and excess methyl iodide [48].
Equilibrium dialysis
Assays were performed by adding 100 nM S-adenosyl-L-methionine-(methyl-3H) to side 'a' and 10 μM metA RNA to side 'b' of a DispoEquilibrium Biodialyser with a 5 kDa MWCO (The Nest Group, Inc., Southboro, MA, USA) in 40 mM MgCl2, 200 mM KCl, 200 mM Tris-HCl (pH 8.5 at 23°C). The sample remaining on side 'a' of the chamber after 10 h of incubation at 23°C was replaced with fresh buffer to increase the final binding signal by preferentially removing non-interacting, radiolabeled metabolite breakdown products [5]. After a second 10 h incubation, the counts in each chamber were recorded. Unlabeled SAM or SAH was added to a concentration of 125 μM in side 'a' and the counts were measured again after a final 10 h incubation.
Additional data files
The following additional data are available with the online version of this article: A PDF file illustrating the formatted sequence alignments, compilations of downstream genes, consensus structures, and in-line probing data for all five RNA elements (Additional data file 1) and sequence alignments for each of the five RNA motifs in Stockholm format (Additional data files 2, 3, 4, 5 and 6).
Supplementary Material
Additional data file 1
A PDF file illustrating the formatted sequence alignments, compilations of downstream genes, consensus structures, and in-line probing data for all five RNA elements
Click here for file
Additional data file 2
serC RNA element sequence alignment in Stockholm format
Click here for file
Additional data file 3
speF RNA element sequence alignment in Stockholm format
Click here for file
Additional data file 4
suhB RNA element sequence alignment in Stockholm format
Click here for file
Additional data file 5
ybhL RNA element sequence alignment in Stockholm format
Click here for file
Additional data file 6
metA RNA element sequence alignment in Stockholm format
Click here for file
Acknowledgements
We thank S Dinesh-Kumar for his gift of A. tumefaciens strain GV2260 and JN Kim for work on the speF element. This research was funded by grants to R.R.B. from the NIH (GM 068819), NSF (EIA-0323510), and DARPA. J.E.B. is a Howard Hughes Medical Institute predoctoral fellow.
Figures and Tables
Figure 1 α-Proteobacterial RNA elements. (a) Consensus sequences and structures. Red and black positions for each RNA element indicate >95% and >80% conservation of a particular nucleotide, respectively. Purine (R) or pyrimidine (Y) designations are used when a single nucleotide is not >80% conserved. Solid black lines indicate variable regions, and solid grey lines are optional sequence insertions that are not present in all examples of an element. Circles represent single nucleotides whose presence (but not sequence) is conserved. Base pairs supported by strong (both bases in the pair vary) and weak (only one base in the pair varies) sequence covariation in a motif alignment have green and blue shaded backgrounds, respectively. (b) Phylogenetic distributions. Element names are chosen based on the proximity that representatives from A. tumefaciens have to genes.
Figure 2 The metA RNA element. (a) Sequence alignment of representative metA RNAs. Shaded nucleotides represent conserved base pairing regions. Lowercase and uppercase letters in the consensus line indicate 80% and 95% sequence conservation, respectively. A complete alignment is available in Additional data file 1. Organism abbreviations: Atu, Agrobacterium tumefaciens; Bja, Bradyrhizobium japonicum; Bme, Brucella melitensis; Mma, Magnetospirillum magnetotacticum; Mlo, Mesorhizobium loti; Rsp, Rhodobacter sphaeroides; Rpa, Rhodopseudomonas palustris; Sme, Sinorhizobium meliloti; Cbu, Coxiella burnetii; Bth, Bacteroides thetaiotaomicron; Bbr, Bordetella bronchiseptica. (b) Consensus sequence and structure of the SAM-I riboswitch aptamer found in Gram-positive bacteria. The consensus is updated from [17] and depicted using the same conventions as Figure 1a. The SAM-II aptamer structure is shown for comparison. (c) Comparison of genes in the methionine and SAM biosynthetic pathways found downstream of SAM-I and SAM-II riboswitches.
Figure 3 The metA element binds SAM. (a) In-line probing of 156 metA RNA from A. tumefaciens. 32P-labeled RNA (NR, no reaction) and products resulting from partial digestion with nuclease T1 (T1), partial digestion with alkali (-OH), and spontaneous cleavage during a 40 h incubation in the presence of varying of SAM concentrations (1 μM to 6 mM) were separated by polyacrylamide gel electrophoresis. Product bands corresponding to certain G residues (generated by T1 digestion) and full length 156 metA RNA (Pre) are labeled. (b) Sequence and secondary structure model for A. tumefaciens metA RNA. Sites of structural modulation for the 156 metA derived from in-line probing are circled with red, green and yellow representing reduced, increased, and constant scission in the presence of SAM, respectively. (c) Dependence of spontaneous cleavage in various regions of 156 metA on the concentration of SAM. Band intensities for the five regions (labeled 1-5) on the in-line probing gel in (a) were quantitated and normalized to the maximum modulation observed. Data from each of these sites corresponds to an apparent Kd of around 1 μM (producing half maximal modulation of cleavage) when plotted against the logarithm of the SAM concentration. Theoretical curves for single ligand binding at sites where cleavage increases (black) and decreases (gray) with a Kd of 1 μM are shown for comparison.
Figure 4 Molecular recognition characteristics of SAM-II aptamers. (a) In-line probing of A. tumefaciens 156 metA RNA in the presence of 1 mM SAM, SAH, SAC, and Met. See legend to Figure 3a for an explanation of the labels. (b) Chemical structures of SAM and a generalized SAM analogue. Arrows represent possible hydrogen bonds and electrostatic interactions that could serve as points of recognition by the aptamer. Circled interactions were determined to have strong (solid) or weak (dashed) contributions to binding affinity in singly substituted chemical analogues. Recognition of the N1 position of SAM was not tested. (c) Apparent Kd values of SAM analogues for binding to 156 metA. Columns (n, X, Y, Z, R1, R2, R3) correspond to groups on the core structure in (b). The S-methyl group (gray box) is not present for SAH and SAC.
==== Refs
Nudler E Mironov AS The riboswitch control of bacterial metabolism. Trends Biochem Sci 2004 29 11 17 14729327 10.1016/j.tibs.2003.11.004
Gelfand MS Mironov AA Jomantas J Kozlov YI Perumov DA A conserved RNA structure element involved in the regulation of bacterial riboflavin synthesis genes. Trends Genet 1999 15 439 442 10529804 10.1016/S0168-9525(99)01856-9
Henkin TM Transcription termination control in bacteria. Curr Opin Microbiol 2000 3 149 153 10745002 10.1016/S1369-5274(00)00067-9
Mandal M Breaker RR Gene regulation by riboswitches. Nat Rev Mol Cell Biol 2004 5 451 463 15173824 10.1038/nrm1403
Nahvi A Sudarsan N Ebert MS Zou X Brown KL Breaker RR Genetic control by a metabolite binding mRNA. Chem Biol 2002 9 1043 1049 12323379 10.1016/S1074-5521(02)00224-7
Nahvi A Barrick JE Breaker RR Coenzyme B12 riboswitches are widespread genetic control elements in prokaryotes. Nucleic Acids Res 2004 32 143 150 14704351 10.1093/nar/gkh167
Mironov AS Gusarov I Rafikov R Lopez LE Shatalin K Kreneva RA Perumov DA Nudler E Sensing small molecules by nascent RNA: a mechanism to control transcription in bacteria. Cell 2002 111 747 756 12464185 10.1016/S0092-8674(02)01134-0
Winkler WC Cohen-Chalamish S Breaker RR An mRNA structure that controls gene expression by binding FMN. Proc Natl Acad Sci USA 2002 99 15908 15913 12456892 10.1073/pnas.212628899
Mandal M Boese B Barrick JE Winkler WC Breaker RR Riboswitches control fundamental biochemical pathways in Bacillus subtilis and other bacteria. Cell 2003 113 577 586 12787499 10.1016/S0092-8674(03)00391-X
Mandal M Breaker RR Adenine riboswitches and gene activation by disruption of a transcription terminator. Nat Struct Mol Biol 2004 11 29 35 14718920 10.1038/nsmb710
Sudarsan N Wickiser JK Nakamura S Ebert MS Breaker RR An mRNA structure in bacteria that controls gene expression by binding lysine. Genes Dev 2003 17 2688 2697 14597663 10.1101/gad.1140003
Mandal M Lee M Barrick JE Weinberg Z Emilsson GM Ruzzo WL Breaker RR A glycine-dependent riboswitch that uses cooperative binding to control gene expression. Science 2004 306 275 279 15472076 10.1126/science.1100829
Winkler WC Nahvi A Roth A Collins JA Breaker RR Control of gene expression by a natural metabolite-responsive ribozyme. Nature 2004 428 281 286 15029187 10.1038/nature02362
Barrick JE Corbino KA Winkler WC Nahvi A Mandal M Collins J Lee M Roth A Sudarsan N Jona I New RNA motifs suggest an expanded scope for riboswitches in bacterial genetic control. Proc Natl Acad Sci USA 2004 101 6421 6426 15096624 10.1073/pnas.0308014101
McDaniel BA Grundy FJ Artsimovitch I Henkin TM Transcription termination control of the S box system: direct measurement of S-adenosylmethionine by the leader RNA. Proc Natl Acad Sci USA 2003 100 3083 3088 12626738 10.1073/pnas.0630422100
Epshtein V Mironov AS Nudler E The riboswitch-mediated control of sulfur metabolism in bacteria. Proc Natl Acad Sci USA 2003 100 5052 5056 12702767 10.1073/pnas.0531307100
Winkler WC Nahvi A Sudarsan N Barrick JE Breaker RR An mRNA structure that controls gene expression by binding S-adenosylmethionine. Nat Struct Biol 2003 10 701 707 12910260 10.1038/nsb967
Rodionov DA Vitreschak AG Mironov AA Gelfand MS Comparative genomics of the methionine metabolism in Gram-positive bacteria: a variety of regulatory systems. Nucleic Acids Res 2004 32 3340 3353 15215334 10.1093/nar/gkh659
The BLISS Database (Breaker Lab Intergenic Sequence Server)
Vitreschak AG Lyubetskaya EV Shirshin MA Gelfand MS Lyubetsky VA Attenuation regulation of amino acid biosynthetic operons in proteobacteria: comparative genomics analysis. FEMS Microbiol Lett 2004 234 357 370 15135544 10.1016/j.femsle.2004.04.005
Osteras M Stanley J Finan TM Identification of Rhizobium-specific intergenic mosaic elements within an essential two-component regulatory system of Rhizobium species. J Bacteriol 1995 177 5485 5494 7559334
Chen SL Shapiro L Identification of long intergenic repeat sequences associated with DNA methylation sites in Caulobacter crescentus and other alpha-proteobacteria. J Bacteriol 2003 185 4997 5002 12897020 10.1128/JB.185.16.4997-5002.2003
Soukup GA Breaker RR Relationship between internucleotide linkage geometry and the stability of RNA. RNA 1999 5 1308 1325 10573122 10.1017/S1355838299990891
Griffiths-Jones S Bateman A Marshall M Khanna A Eddy SR Rfam: an RNA family database. Nucleic Acids Res 2003 31 439 441 12520045 10.1093/nar/gkg006
Lu Y Turner RJ Switzer RL Function of RNA secondary structures in transcriptional attenuation of the Bacillus subtilis pyr operon. Proc Natl Acad Sci USA 1996 93 14462 14467 8962074 10.1073/pnas.93.25.14462
Wassarman KM Small RNAs in bacteria: diverse regulators of gene expression in response to environmental changes. Cell 2002 109 141 144 12007399 10.1016/S0092-8674(02)00717-1
Carothers JM Oestreich SC Davis JH Szostak JW Informational complexity and functional activity of RNA structures. J Am Chem Soc 2004 126 16 5130 5137 15099096 10.1021/ja031504a
Wickiser JK Winkler WC Breaker RR Crothers DM The speed of RNA transcription and metabolite binding kinetics operate an FMN riboswitch. Mol Cell 2005 18 49 60 15808508 10.1016/j.molcel.2005.02.032
Posnick LM Samson LD Influence of S-adenosylmethionine pool size on spontaneous mutation, dam methylation, and cell growth of Escherichia coli. J Bacteriol 1999 181 6756 6762 10542178
Haba G Jamieson GA Mudd AH Richard HH S-adenosylmethionine: The relation of configuration at the sulfonium center to enzymatic reactivity. J Am Chem Soc 1959 81 3975 3980 10.1021/ja01524a039
Noeske J Richter C Grundl MA Nasiri HR Schwalbe H Wohnert J An intermolecular base triple as the basis of ligand specificity and affinity in the guanine- and adenine-sensing riboswitch RNAs. Proc Natl Acad Sci USA 2005 102 1372 1377 15665103 10.1073/pnas.0406347102
Serganov A Yuan YR Pikovskaya O Polonskaia A Malinina L Phan AT Hobartner C Micura R Breaker RR Patel DJ Structural basis for discriminative regulation of gene expression by adenine- and guanine-sensing mRNAs. Chem Biol 2004 11 1729 1741 15610857 10.1016/j.chembiol.2004.11.018
Huang Z Szostak JW Evolution of aptamers with a new specificity and new secondary structures from an ATP aptamer. RNA 2003 9 1456 1463 14624002 10.1261/rna.5990203
Sassanfar M Szostak JW An RNA motif that binds ATP. Nature 1993 364 550 553 7687750 10.1038/364550a0
Sazani PL Larralde R Szostak JW A small aptamer with strong and specific recognition of the triphosphate of ATP. J Am Chem Soc 2004 126 8370 8371 15237981 10.1021/ja049171k
White HB 3rd Coenzymes as fossils of an earlier metabolic state. J Mol Evol 1976 7 101 104 1263263 10.1007/BF01732468
Benner SA Ellington AD Tauer A Modern metabolism as a palimpsest of the RNA world. Proc Natl Acad Sci USA 1989 86 7054 7058 2476811
Benner SA Ellington AD RNA world. Science 1991 252 1232 1718033
Jeffares DC Poole AM Penny D Relics from the RNA world. J Mol Evol 1998 46 18 36 9419222
Joyce GF The antiquity of RNA-based evolution. Nature 2002 418 214 221 12110897 10.1038/418214a
Gold L Brody E Heilig J Singer B One, two, infinity: genomes filled with aptamers. Chem Biol 2002 9 1259 1264 12498875 10.1016/S1074-5521(02)00286-7
Tatusov RL Fedorova ND Jackson JD Jacobs AR Kiryutin B Koonin EV Krylov DM Mazumder R Mekhedov SL Nikolskaya AN The COG database: an updated version includes eukaryotes. BMC Bioinformatics 2003 4 41 12969510 10.1186/1471-2105-4-41
Weinberg Z Ruzzo WL Exploiting conserved structure for faster annotation of non-coding RNAs without loss of accuracy. Bioinformatics 2004 20 Suppl 1 I334 I341 15262817 10.1093/bioinformatics/bth925
Venter JC Remington K Heidelberg JF Halpern AL Rusch D Eisen JA Wu DY Paulsen I Nelson KE Nelson W Environmental genome shotgun sequencing of the Sargasso Sea. Science 2004 304 66 74 15001713 10.1126/science.1093857
Thompson JD Higgins DG Gibson TJ CLUSTRAL W: Improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Nucleic Acids Res 1994 22 4673 4680 7984417
Borchardt RT Huber JA Wu YS Potential inhibitor of S-adenosylmethionine-dependent methyltransferases. 2. Modification of the base portion of S-adenosylhomocysteine. J Med Chem 1974 17 868 873 4845380 10.1021/jm00254a017
Borchardt RT Wu YS Potential inhibitors of S-adenosylmethionine-dependent methyltransferases. 3. Modifications of the sugar portion of S-adenosylhomocysteine. J Med Chem 1975 18 300 304 1133821 10.1021/jm00237a018
Borchardt RT Wu YS Potential inhibitors of S-adenosylmethionine-dependent methyltransferases. 5. Role of the asymmetric sulfonium pole in the enzymatic binding of S-adenosyl-L-methionine. J Med Chem 1976 19 1099 1103 978674 10.1021/jm00231a004
|
16086852
|
PMC1273637
|
CC BY
|
2021-01-04 16:35:47
|
no
|
Genome Biol. 2005 Aug 1; 6(8):R70
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r70
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r711608685310.1186/gb-2005-6-8-r71ResearchMicroRNA profiling of the murine hematopoietic system Monticelli Silvia [email protected] K Mark 1Xiao Changchun 1Socci Nicholas D 3Krichevsky Anna M 4Thai To-Ha 1Rajewsky Nikolaus 5Marks Debora S 6Sander Chris 3Rajewsky Klaus 1Rao Anjana 1Kosik Kenneth S 471 Department of Pathology, Harvard Medical School, and CBR Institute for Biomedical Research, Boston, MA 02115, USA2 Department of Biology and Genetics of Medical Sciences, Universitá degli Studi di Milano, 20133 Milan, Italy3 Computational Biology Center, Memorial Sloan-Kettering Cancer Center, New York, NY 10021, USA4 Department of Neurology, Brigham and Women's Hospital and Harvard Medical School, Boston, MA 02115, USA5 Center for Functional Comparative Genomics, Department of Biology, New York University, New York, NY 10003, USA6 Department of Systems Biology, Harvard Medical School, Boston, MA 02115, USA7 Neuroscience Research Institute, University of California Santa Barbara, Santa Barbara, CA 93106, USA2005 1 8 2005 6 8 R71 R71 23 2 2005 9 5 2005 1 7 2005 Copyright © 2005 Monticelli et al.; licensee BioMed Central Ltd.2005Monticelli et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The first report of systematic miRNA profiling in cells of the hematopoietic system suggests that, in addition to regulating commitment to particular cellular lineages, miRNAs might have a general role in cell differentiation and cell identity.
Background
MicroRNAs (miRNAs) are a class of recently discovered noncoding RNA genes that post-transcriptionally regulate gene expression. It is becoming clear that miRNAs play an important role in the regulation of gene expression during development. However, in mammals, expression data are principally based on whole tissue analysis and are still very incomplete.
Results
We used oligonucleotide arrays to analyze miRNA expression in the murine hematopoietic system. Complementary oligonucleotides capable of hybridizing to 181 miRNAs were immobilized on a membrane and probed with radiolabeled RNA derived from low molecular weight fractions of total RNA from several different hematopoietic and neuronal cells. This method allowed us to analyze cell type-specific patterns of miRNA expression and to identify miRNAs that might be important for cell lineage specification and/or cell effector functions.
Conclusion
This is the first report of systematic miRNA gene profiling in cells of the hematopoietic system. As expected, miRNA expression patterns were very different between hematopoietic and non-hematopoietic cells, with further subtle differences observed within the hematopoietic group. Interestingly, the most pronounced similarities were observed among fully differentiated effector cells (Th1 and Th2 lymphocytes and mast cells) and precursors at comparable stages of differentiation (double negative thymocytes and pro-B cells), suggesting that in addition to regulating the process of commitment to particular cellular lineages, miRNAs might have an important general role in the mechanism of cell differentiation and maintenance of cell identity.
==== Body
Background
MicroRNAs (miRNAs) represent a recently discovered class of small, noncoding RNAs, found in organisms ranging from nematodes to plants to humans. Many individual miRNAs are conserved across widely diverse phyla, indicating their physiological importance. The primary transcript (pri-miRNA) is generally transcribed by RNA polymerase II; it contains a typical stem-loop structure that is processed by a nuclear enzyme complex including Drosha and Pasha, and releases a 60- to 110-nucleotide pre-miRNA hairpin precursor [1]. The pre-miRNA is further processed by Dicer to yield the 19- to 22-nucleotide mature miRNA product, which is then incorporated into the RNA-induced silencing complex (RISC) [2-4]. RISC-bound miRNAs direct the cleavage and/or translational repression of messenger RNAs, thus providing post-transcriptional control of gene expression.
Like many transcription factors, miRNAs are important determinants of cellular fate specification. One of the most prominent and genetically best-studied examples is given by miRNAs involved in neuronal fate determination in Caenorhabditis elegans, where a cascade of several miRNAs and transcription factors regulate each other's activity to induce a different spectrum of putative chemoreceptors in the two main taste receptor neurons in C. elegans [5]. Furthermore, many miRNA genes are located at fragile sites, minimal loss of heterozygosity regions, minimal regions of amplification, or common breakpoints in human cancers, suggesting that miRNAs might play an important role in the pathogenesis of human cancer [6,7].
Hundreds of miRNAs have been identified in plants and animals, either through computational searches, RT-PCR-mediated cloning, or both. More than 200 human and rodent miRNAs have been reported and tabulated in the miRNA Registry [8], accounting for an estimated 1-2% of expressed human genes. Recent evidence suggests that the actual number of miRNAs is likely to be even larger [9,10]. MiRNAs have been implicated in biological processes ranging from cell proliferation and cell death during development to stress resistance, fat metabolism, insulin secretion and hematopoiesis [11]. However, for the most part, the regulation and function of most mammalian miRNAs are unknown. The bulk of the existing data on miRNA expression in mammalian cells has been derived from studies on whole tissues, which contain many heterogeneous cell types, or on transformed or established cell lines that may have diverged significantly from the primary cell types that they are assumed to represent [7,12-15]. To understand the role of miRNAs in mammalian development and differentiation, an important starting point is a systematic compilation of miRNAs expressed in individual cell types, especially those derived by differentiation from a common precursor.
The cells of the immune system originate from hematopoietic stem cells in the bone marrow, where many of them also mature. The hematopoietic stem cells give rise to both myeloid and lymphoid progenitors. The myeloid progenitor is the precursor of granulocytes, macrophages, dendritic cells, and mast cells of the innate immune system. Mast cells, whose blood-borne precursors are not well defined, terminate their differentiation in the body tissues, where they are widely distributed and where they orchestrate allergic responses and play a part in protecting mucosal surfaces against pathogens [16]. The common lymphoid progenitor gives rise to B and T lymphocytes and to natural killer cells. B lymphocytes differentiate in the bone marrow and T lymphocytes in the thymus; the stages of B and T cell development are defined by sequential rearrangement and expression of heavy- and light-chain immunoglobulin genes and TCR α and β chains, respectively. Mature B and T lymphocytes that have emigrated to the peripheral lymphoid organs, including the spleen and lymph nodes, but have not yet encountered their specific antigen are called 'naïve'. In the event of an infection, T lymphocytes that recognize the infectious agent are arrested in the lymphoid organs, where they proliferate and differentiate further into effector cells capable of combating the infection.
Because of the wealth of information available about the transcriptional and cellular networks involved in hematopoietic differentiation, the hematopoietic system is ideal for studying cell lineage specification. Many of the common progenitors of hematopoietic cells can be obtained as primary cells from humans and mice, and expanded and differentiated in vitro. Here we have performed a detailed analysis of miRNA expression in diverse hematopoietic cell types from the mouse, using a high-throughput system that allows analysis of many samples with minimal manipulation of the samples themselves. This has allowed us to identify miRNAs that are highly expressed in the hematopoietic system. Our results are consistent with a model of hematopoiesis in which transcriptional regulators act in concert with differentially expressed miRNAs to modulate the levels of mRNAs that control cell differentiation pathways.
Results
Microarray design
To probe the expression of miRNAs in a variety of different related and unrelated cell types, we chose to use miRNA arrays in preference to time-consuming Northern analysis that cannot be used efficiently with many different probes and samples. In the past year, several microarray methods have been developed [7,12,13,15,17-19]. Some of these groups [7,12,13,15,17] used cDNA or cRNA generated from total cellular RNA to apply to their microarrays. Other methods [18,19] rely heavily on several enzymatic steps, such as RNA ligation [18], or Klenow synthesis and exonuclease I degradation of ssDNA [19]. Instead, we chose a technique that does not involve reverse transcription of RNA and relies on only one enzymatic step ([20]; see Methods) thus reducing RNA manipulation to a minimum. In designing the arrays, we expanded the array dataset already developed by Krichevsky et al. [20]. The new generation of arrays contains 181 gene-specific oligonucleotide probes, corresponding to human, rat, and mouse miRNAs as reported in the miRNA Registry [8].
Data from the arrays
Figure 1a shows a typical array experiment comparing miRNA expression in bone marrow-derived mast cells (BMMC) and a hematopoietic progenitor cell line (Pu.1-/-) derived from mice lacking the Ets-family transcription factor PU.1 [21]. This cell line differentiates efficiently into mast cells when rescued with PU.1 under conditions where GATA2 expression is maintained; expression of PU.1 in the absence of GATA2 results in commitment to the macrophage lineage instead [22]. Visual comparison of three arrays performed with each RNA sample shows the high reproducibility of the arrays, and emphasizes the difference in miRNA expression relative to hippocampal RNA (Figure 1a). This high degree of reproducibility was maintained over a total of nine arrays, each performed in triplicate using three independent RNA samples (not shown). Statistical analysis confirmed the high level of reproducibility (Figure 1b): when the standard deviation over replicates was plotted versus the mean of each replicate, 86% of the spots were considered good (see legend to Figure 1b for details). RNA loading for all arrays and Northern blot experiments was evaluated by ethidium bromide staining of a denaturing acrylamide gel as shown in Figures 1a and 2d.
The arrays were repeated using cells from several stages of lymphocyte differentiation (Figure 2a). Among the cell types compared were pro-B cells, which are in the process of rearranging the heavy-chain immunoglobulin locus; mature splenic B cells, which express IgM and IgD B cell receptors and are competent to respond to antigen; double-negative thymocytes (DN T), which are just beginning to rearrange their T cell receptor chains and lack surface expression of the CD4 and CD8 co-receptors; naïve CD4 T 'helper' cells, which have exited the thymus, bear the CD4 co-receptor and a mature T cell receptor, and are fully capable of recognizing and responding to antigen; and Th1 and Th2 T helper subsets, which are derived by differentiation from a common precursor, the naïve CD4 T cell, and are characterized by selective expression of the cytokines IFNγ and IL4, respectively. Figure 2a shows representative array data for pro-B cells, mature splenic B cells, naïve T cells, and Th1 and Th2 clones. Note that miR-150 is highly expressed in B cells purified from mouse spleen, but not in pro-B cells isolated from bone marrow of Rag2-/- mice; it is also expressed in naïve T cells but is down-regulated in the Th1 and Th2 T cell clones (Figure 2a, arrow).
Validation of array data by Northern analysis
Before analyzing the entire dataset from the microarrays, we validated the array results by Northern blot analysis. Single-stranded DNA oligonucleotides complementary to over 40 different miRNAs were used as probes; they were chosen because they were expressed in at least one cell type in the hematopoietic lineages and/or were highly expressed in neurons. Several of these Northern blots were already published with the first description of the microarray methodology [20] to confirm the array specificity. We performed several other Northern blots that included hematopoietic cells and tissues. These data are shown in Figures 2, 3 and 4, and summarized in Tables 1 and 2. With minor exceptions as discussed below, the results of Northern analysis were consistent with the array data for most miRNAs, by simple visual inspection (Table 2), and when the hybridization signal intensity was quantified by phosphorimager (Figure 3 and Table 1). For example, Northern analysis of miR-150 expression confirmed its expression in spleen B but not pro-B cells (Figure 2b, lanes 6 and 7), and in naïve T cells but not the Th1 and Th2 clones, D5 and D10 (Figure 2c, lanes 5-7). Figure 2b also shows that miR-150 is expressed in thymocytes and splenic T cells (lanes 9 and 10), but not in ES cells, mouse embryo fibroblasts or hippocampus (lanes 11-13). The lack of expression in RAG2-/- spleen and thymus (lanes 5 and 8) confirms that expression in these organs is confined to T and B lymphoid-lineage cells, and that within these lineages, miR-150 expression is restricted to cells that have developed beyond the DN T and pro B stages of development. Figure 2c confirms that naïve T cells show high level expression of miR-150 (lane 7) whereas the precursor cell line Pu.1-/- and BMMC, which are of the myeloid lineage, do not (lanes 1-4). Equivalent RNA loading in all lanes was confirmed by ethidium bromide staining of a denaturing acrylamide gel as shown in Figure 2d.
Strikingly, miR-150 expression in naïve T cells is rapidly down-regulated upon TCR engagement, regardless of whether T cells are stimulated under Th1 or Th2 conditions (Figure 2c, lanes 8-12). The levels of expression of miR-150 were already reduced by ~50% after 12 h of stimulation with plate-bound αCD3 and αCD28 (lane 8), and by >90% after 25 h (lanes 9 and 11), indicating a rapid and highly inducible mechanism of down-regulation. Expression was barely detectable after 49 h (lanes 10 and 12) and remained undetectable 3 days after stimulation (data not shown). Furthermore, miRNA expression was extinguished in fully committed Th1 and Th2 T cell clones (lanes 5 and 6). Together, these results suggest a role for miR-150 either in maintaining the undifferentiated status of naïve T cells or in promoting early steps in T cell differentiation.
Figure 3 extends the concordance of the array data with the Northern analysis to five additional miRNAs, emphasizing the cell type-specific changes that take place during differentiation. In the left panels of Figure 3, miRNA expression in BMMC treated with cyclosporin A to prevent activation, or stimulated with PMA and ionomycin for the indicated times, is compared with expression in the Pu.1-/- 'precursor' cell line, which gives rise to mast cells when reconstituted with both PU.1 and GATA2 [22]. Three very different patterns are observed, exemplified by: miR-146 and 142s, which are expressed at essentially equivalent (low) levels in both the Pu.1-/- precursor cells and the fully-differentiated BMMC; miR-26a and 27a, which are expressed at low levels in the Pu.1-/- precursor cells and at three- to fourfold higher levels in fully differentiated BMMC; and miR-223, which is most highly expressed in the Pu.1-/- precursor cells and is barely detectable in the differentiated BMMC. The results of Northern analysis for these two cell types show an overall good concordance with the values obtained from the arrays (Figures 2, 3 and 4 and Table 1).
The right panels of Figure 3 compare expression of the same miRNAs in naïve T cells and fully-differentiated Th1 and Th2 cells (the D5 and D10 clones). Several expression patterns are evident: miR-146 is highest in Th1 cells and low in naïve T cells and Th2 cells; miRNAs 142s and 26a are expressed at higher levels in the precursor naïve T cells; miR-27a is equivalently expressed in both the precursor naïve T cells and the differentiated Th1 and Th2 cells; and miR-223 is very poorly expressed in all these T cell types. The relative expression of these miRNAs in naïve versus differentiated T cells was confirmed in primary cultures of Th1 and Th2 cells (see Table 1). There was full concordance of the Northern analysis with the miRNA array data for miRNAs 146, 142s, 26a and 223 (Figure 3, Table 1); however, as discussed further below, the signal for 27a and a handful of other miRNAs expressed at low levels in naïve T cells fell below the limit of detection on the microarrays.
Table 1 summarizes the results from Northern analysis of the miRNAs shown in Figures 2 and 3, as well as showing data for two additional miRNAs, let7d (let7 family) and miR-222. The miRNA expression pattern of D5 and D10 T cell clones was comparable with that of differentiated primary Th1 and Th2 cells respectively, validating the use of D5 and D10 cells as models for fully differentiated Th1 and Th2 cells. Like miR-150, the expression of miR-142s, miR-26a and let7d showed a rapid decline during differentiation of naïve T cells into Th1 or Th2 effectors. miR-27a was expressed at equivalent levels in naïve and differentiated T cells. miR-146 showed a Th1-skewed expression pattern: it's levels increased in Th1 cells and decreased in Th2 cells relative to it's expression in naïve T cells. We have not yet detected an miRNA with the converse expression pattern of high expression in Th2 cells relative to Th1. miR-222 was detectably expressed in BMMC, Pu.1-/- precursor cells and fully differentiated Th1 cells, but it's expression was not detectable in the other cell types tested (Table 1); in contrast, miR-16 was expressed in all cell types analyzed, but it's expression was relatively variable both in arrays and Northern blots, so quantification was not attempted (data not shown). Some of our data confirm published reports. For example, miR-223 is reported to be expressed in myeloid cells [7,23]; miR-125 and 128 are highly expressed in the brain [13,14]; and miR-16 is expressed in a wide variety of tissues [7,14,23] (see also heat map of expression in Figure 5a).
Figure 4 shows Northern blot data for additional miRNAs. Most of the data from Northern blots correlated at least qualitatively with the expression data from the arrays (Table 2; also compare data in Figures 3 and 4 to the heat map in Figure 5). Some exceptions were noted. For some of the miRNAs (miR-129, 151, 184, 185, 202, 212 and 351), we could not obtain any hybridization signal on Northern blots, so we were unable to compare Northern and array data. MiR-223 and miR-206 showed poor correlation between the arrays and the Northern blots: for miR-223, we detected a higher level of expression in pro-B in the arrays compared with what we detected on Northern blot, while for miR-206, the arrays showed high expression in pro-B and DN T that was undetectable by Northern blot. In a few other cases, the hybridization signal was lower in the arrays compared with Northern blots, but the relative expression levels between different cell types was similar. It is unclear at this point why the expression of some miRNAs appears different depending on the method used to detect them.
In a few cases, the probes used in Northern analysis hybridized to a cross-reacting band with a molecular weight higher than the mature or pre-miRNA molecules. In these cases the correlation between Northern blots (which use total RNA) and arrays (which use the low-molecular weight RNA) was only partial. Even though our system is designed to exclude RNA molecules bigger than ~300 nucleotides, we effectively obtained exclusion of molecules bigger than 60-80 nucleotides (as shown in [20]). Thus, the changes observed mainly reflect changes in mature miRNA levels, as also shown by the correlation with Northern blots. However, it remains possible that a strong expression of cross-reacting RNA close to this size might partially alter the array results; we observed such bands for miR-186, miR-188 and miR-321. Of note, miR-321 has been removed from the microRNA Registry because it was identified as a fragment of an Arg tRNA and not a miRNA.
Despite differences in methods as well as in the number of microRNAs analyzed, there is good agreement between our results and those of others, with regard to specificity and sensitivity, when array and Northern blot analyses are examined for similar cell types [13,14]. Similar to the findings and discussion of Miska et al. [13], we do not expect our microarray technique to provide sufficient specificity to distinguish reliably between hybridizing sequences that have only one or few nucleotide mismatches. Although hybridization signals from several control probes containing three staggered nucleotide mismatches were lower than that for the corresponding miRNA probes (see also Material and methods), our method cannot efficiently discriminate between close miRNA paralogs. This limitation is alleviated somewhat by the fact that for most miRNAs, the most closely related paralogs differ by five mismatches or more [13]. The sensitivity of the arrays is similar to that of Northern blots. Synthetic RNA oligonucleotides 'spiked' into cellular RNA samples prior to array hybridization were detected at a 2-20 fmole range. Northern blot allowed detection of as little as 1-10 fmole of synthetic oligonucleotides (data not shown). In summary, therefore, we saw substantial concordance between arrays and Northern blots, allowing us to identify cell type-specific differences in miRNA expression as well as differences between miRNAs expressed by precursor cells and their differentiated progeny. This led us to analyze the array data more extensively using computational methods.
Analysis of miRNA arrays
To identify patterns of miRNA expression among the cell types tested, we arranged the array data for miRNAs that were expressed at least three times over the background for at least one of the samples in a heat map (Figure 5a). Brown to white colors indicate increasing levels of miRNA expression in arrays. This analysis revealed a cluster of miRNAs that were preferentially expressed in the hippocampus compared with hematopoietic cells, as indicated by the blue bar in the left panel. MiRNAs expressed at higher levels in the hematopoietic system are indicated by the purple bar in the right panel.
To achieve a better understanding as to how miRNA expression patterns correlate with hematopoietic cell differentiation, we performed a hierarchical clustering of the normalized array data for hematopoietic cell types (Figure 5b). The subset of miRNAs detected in at least one hematopoietic cell sample was used to compute the distance function from the Pearson correlation between samples (Table 3). Standard hierarchical clustering with average linkage was used, and bootstrap resampling was employed to assess the robustness of the clustering results. This analysis showed that fully differentiated effector cells (Th1, Th2 and BMMC) are more closely related to each other in their miRNA expression pattern than to their respective precursor cells (DN T and Pu.1-/- precursor cells). The miRNA expression patterns of pro-B and DN T, precursor cells for the B and T lymphocyte lineages respectively, were also very closely related. Although the detected miRNA expression pattern of naïve T cells most closely resembled that of splenic B cells (Table 3), naïve T cells were excluded from the clustering analysis. This was because RNA isolated from naïve T cells yielded much lower overall array hybridization signals compared with RNA from the other cell types examined, causing the signal for a handful of expressed miRNAs to fall below the limit of detection for the microarrays (for example, miR-27a, see Figure 4 and Table 1), and making it impossible to accurately normalize the array data for naïve T cells relative to the signal obtained from other cell types.
Discussion
In summary, pairwise comparisons of the expression of 181 mature miRNAs in selected highly purified hematopoietic cell types at immature, mature, and effector stages revealed specific differences between related cell types (see also Additional Data Files 1, 2, 3). As described above, the differences were confirmed by Northern analyses (Figures 2, 3 and 4, and Tables 1 and 2) and revealed a subset of miRNAs expressed at higher levels in the hematopoietic system compared with neuronal tissue (Figure 5a). Figure 6 shows a schematic hematopoietic lineage tree, where lymphoid and myeloid cells derive from a common lymphoid and common myeloid progenitor, respectively. Both these cell types derive from a common precursor further upstream in the differentiation process and we used the Pu.1-/- cells as a model for such precursors. The figure also shows a summary of some confirmed changes in miRNA expression in the different stages of cell differentiation superimposed on the diagram, showing precursor-progeny relationships in lymphocyte and mast cell differentiation. Even though only selected hematopoietic cell lineages were analyzed, each differentiation step was characterized by changes in miRNA expression, with some miRNAs showing increased and some showing decreased expression. MiR-150 expression is of particular interest: this miRNA is up-regulated during the developmental stages of B and T cell maturation, but down-regulated again during the further differentiation of naïve T cells into effector Th1 and Th2 cells. miR-146 is also notable, since it is upregulated in Th1, but not Th2 cells. We therefore predict that these miRNAs probably play a role in establishing and/or maintaining cell identity in lymphocytes.
Differentiation of naïve T cells into Th1 and Th2 effector cells is a particularly tractable system for studying cell lineage specification. Northern blot analyses indicated that several miRNAs are rapidly down-regulated following activation of naïve T cells under both Th1 and Th2 differentiation conditions. miR-146 was a clear exception to this pattern, being up-regulated in Th1, but not Th2 cells. In this respect, miR-146 joins a group of Th1-associated genes that include cytokines (Ifnγ, Tnfα), chemokine receptors (Cxcr3, Ccr5), and transcription factors (Tbx21, Hlx) [6,24,25]. Our results suggest two hypotheses for further testing. Firstly, the transcription of pri-miRNAs may be controlled by the same transcription factors known to control specific cell differentiation events. For example, T-bet directs Ifnγ expression, and may also activate transcription of pri-miR-146. Conversely, miRNAs that are regulated during cell differentiation may target one or more of the mRNAs known to be differentially expressed between Th1 and Th2 cells. Although the predicted targets of conserved miRNAs represent a diverse array of gene products [26-30], targeting of key transcription factors would represent a particularly efficient means for miRNA participation in cell fate decisions.
Our results call to attention the 'common logic' and shared roles of transcriptional regulation by transcription factors and post-transcriptional regulation by miRNAs [5]. It is likely that both pathways of regulation are integral to successful regulation of hematopoietic cell differentiation. Decades of research have been devoted to the elucidation of the transcriptional networks involved in hematopoiesis. For example, the Ets family transcription factor, PU.1, is necessary for the generation of certain hematopoietic lineages but not others: it directs the differentiation of hematopoietic progenitors into macrophages, neutrophils, B lymphocytes and mast cells, but is not involved in erythroid or megakaryocytic differentiation [21,22]. Lineage specification is controlled by the level of PU.1 expression and by which partner transcription factors are co-expressed: for instance, low levels of PU.1 and co-expression of early B-cell factor (EBF) control differentiation to the B cell lineage, whereas high levels of PU.1 predispose to macrophage differentiation unless GATA2 is co-expressed, in which case differentiation is tilted to the mast cell lineage [22,31,32]. It will be informative to compare miRNA expression in Pu.1-/- progenitor cells that have been reconstituted to promote differentiation along these various lineages. Further study of the expression and function of miRNAs in hematopoiesis will probably uncover additional complexity and subtlety, as well as interconnections between miRNA and transcription factor networks [5].
Systematic analysis of miRNA expression patterns within our dataset indicates that besides influencing the process of commitment to a particular cellular lineage, miRNAs may play an important general role in the mechanism of cell differentiation and maintenance of cell identity. The highest degree of correlation in the expression pattern of miRNAs was observed 'horizontally' between hematopoietic cell types at a similar stage of differentiation. For instance, early B and T cell precursors are more closely related to each other in their miRNA expression patterns than to their more differentiated progeny, mature splenic B cells and naïve T cells. Most strikingly, fully differentiated effector cells, including the closely related Th1 and Th2 cells, but also the much more distantly related BMMC, are more closely related to each other in their miRNA expression pattern than to their respective precursor cells. The high degree of correlation in the miRNA expression patterns of these distantly related immune effectors suggests that a common set of miRNAs may be employed in both lineages to regulate similar effector functions, such as tissue homing and cytokine production. Alternatively, these findings may reflect a general role for some of these miRNAs in stabilizing gene expression and thereby lineage specification. This could be accomplished through the promiscuous targeting of many transcripts or by specific targeting of genes that regulate the plasticity of transcriptional states, such as chromatin-modifying proteins. Similarly, miRNAs shared among early precursors may regulate precursor cell self-renewal and maintenance of an undifferentiated state. The rapid loss of several miRNAs early in the process of differentiation of Th1 and Th2 effector cells from naïve T cell precursors is consistent with this concept.
Conclusion
We report miRNA expression patterns for diverse murine hematopoietic cells types, identify a subset of miRNAs preferentially expressed in the hematopoietic system compared with neuronal cells, and identify individual miRNA expression changes that occur during cell differentiation. Our data support the use of the miRNA microarray for detection of patterns of miRNA expression and for quantification of miRNA expression, with the obvious advantage that the expression of several hundred genes can be identified in the same sample at once, and with relatively small amounts of total RNA. Deciphering the miRNA expression status of cells under different conditions of development and activation and in different disease states will be useful to identify miRNA targets, and alterations in the pattern of miRNA expression may disclose new pathogenic pathways and new ways to target diseases.
Materials and methods
Tissue preparation, cells differentiation and RNA extraction
Hippocampi were dissected from 10-14 week old Balb/c mice. For pro-B cell preparation, bone marrow cells were isolated from femurs and tibias of 6-12 week old Rag2-/- C57BL/6 × 129 mice and pro-B cells were isolated using CD19 MACS beads. Spleen B cells were isolated using CD19 MACS beads from splenocytes obtained from 6-12 week old C57BL/6 mice. Double-negative thymocytes (DN T) were obtained from the thymi of 4-6 week old Rag2-/- C57BL/6 × 129 mice without purification; more than 90% of the cells were DN2 and DN3 (not shown). Due to the low amount of cells that can be obtained from one mouse, proB and DN T cells were purified from 20 Rag-/- mice, while the spleen B cells were obtained from 10 C57BL/6 mice, and the samples were pooled.
BMMC preparation and differentiation was as previously described [33,34]. Briefly, bone marrow cells were isolated from femurs and tibias of 6-12 week old BALB/c mice and maintained for 4-12 weeks in RPMI medium containing 50% WEHI-3 (American Type Culture Collection, VA, USA) conditioned supernatant as a source of IL-3.
Pu.1 knock-out cells were kindly provided by Dr Harinder Singh (University of Chicago), and were maintained in IMDM media supplemented with 10 ng/ml of recombinant IL-3 (Peprotech Inc., NJ, USA). Naïve CD4 cells were purified from spleen and lymph nodes of Tcrα-/- DO11 TCR transgenic mice by magnetic bead selection (Dynal, Oslo, Norway) as previously described [33]. Primary Th1 and Th2 cells were differentiated in culture for 7 days as previously described [35]. Murine Th1 (D5) and Th2 (D10) clones were maintained as previously described [36]. RNA was prepared using Ultraspec or Trizol reagents following manufacturer's instructions.
Oligonucleotide array for miRNA
This method has been previously described (see [20]). Briefly, trimer oligonucleotides (antisense to miRNAs) of 54-72 nucleotides at a final concentration of 7 μM were spotted on GeneScreen Plus (NEN) membranes with a 1536 pin plate replicator (V&P Scientific, CA, USA). Oligonucleotides were immobilized in 100 mM NaOH, after which the membranes were briefly neutralized in 5% SDS at room temperature and with 0.2% SDS at 72°C. Arrays were stored in 0.2% SDS at -4°C.
Total RNA (5-10 μg) from hippocampus tissue and various hematopoietic cell types was preheated at 80°C for 3 min, cooled on ice and filtered through Microcon YM-100 concentrators to obtain a low molecular weight (LMW) fraction of RNA enriched in molecules less than 60 nucleotides in size. The LMW RNA was end-labeled with 30 μCi of γ33P dATP (3000 Ci/mmole) with T4 polynucleotide kinase, and purified using the QIAgen Nucleotide Removal kit (QIAgen Inc., CA, USA).
For hybridization, membranes were first prehybridized in MicroHyb hybridization buffer (ResGen, AL, USA) at 37°C for at least 30 min, followed by an overnight hybridization in the same solution containing the RNA probe. Following hybridization, membranes were washed twice in 2 × SSC/0.5% SDS at 37°C and once in 1 × SSC/0.5% SDS at 37°C. Membranes were exposed to a phosphor storage screen, scanned using a Phosphor Imager, and signals were quantified using the ImageQuant software (Molecular Dynamics, CA, USA). For reuse, membranes were stripped with 0.2% SDS at 72°C, tested again by exposure to phosphorimager screen, and rehybridized three to five times. Each experiment included two to three independent RNA samples and to ensure accuracy of the hybridizations, each RNA sample was hybridized with three membranes.
To confirm specificity, a series of oligonucleotides with three mismatches (G>C or C>A) were included on the array. These mismatches resulted in a significant drop in signal intensity as compared with their cognates. The melting temperature of oligo probe:miRNA pairs could affect the sensitivity and specificity of the arrays for different miRNAs. An analysis of this correlation showed that hybridization signals significantly above background were obtained for probes in a wide range of melting temperatures (Additional Data File 4). Also, three synthetic 21-nt RNA oligonucleotides with sequences that do not correspond to any known miRNA, but that are exact complements to randomly spotted sequences, were added to the RNA samples at a known concentration as a reference for normalization.
Northern blot analysis
Total RNA (20 μg) was loaded and separated on a denaturing 12-15% polyacrylamide gel and transferred electrophoretically to a GeneScreen Plus or Nytran SuPerCharge membrane (Scheicher and Schuell, NH, USA). Membranes were UV-crosslinked. Probes were prepared by T4 polynucleotide kinase labeling of antisense oligonucleotides with γ32P dATP. Hybridization was performed with UltraHyb Hybridization buffer (Ambion, TX, USA) or Denhardt's solution at 37-42°C. Blots were washed at the same temperature with 2 × SSC/0.1% SDS with a brief final wash with 0.1 × SSC/0.1% SDS. Radiolabeled Decade RNA markers (Ambion) were loaded as size markers. tRNA and 5S RNA stained with ethidium bromide served as a sample loading control. For reuse, blots were stripped by boiling in 0.1 × SSC/0.1% SDS twice for 10 min and reprobed.
Data analysis
Before analysis, the raw data needed to be processed in order to handle overall scaling differences between the individual scans and negative values arising from the background subtraction. Although these are common issues in array-based experiments, it is not obvious what the optimal preprocessing algorithm should be. For the data presented here, the Variance Stabilization Normalization method [37] was used. The method has a considerable advantage in that it uses a generalized log transformation that can deal directly with negative values, eliminating the need to artificially shift or truncate these data points.
Standard hierarchical clustering with average linkage was used to cluster the hematopoietic samples. A subset of the miRNAs was used to compute the distance function. These miRNAs had to have a signal level of three times the background standard deviation in at least one of the hematopoietic samples. The Pearson correlation was used to compute the distance function with dist = (1-ρ)/2 where ρ is the correlation. To assess the robustness of the results, bootstrap resampling was carried out using a parametric method to add noise to the data. Gaussian noise with zero mean and a standard deviation equal to that for each spot's replicates was added to each point. One thousand resampled datasets were created and clustered with a consensus tree built from the results. The number at each node indicates how often that subtree appeared in the 1000 replica trees.
To identify miRNAs that were differentially expressed between the various sample subtypes, a variation of the standard t-test was used on the transformed expression values. To handle the low number of samples, a Bayesian correction method [38] was used to adjust the standard deviation. To account for the multiple testing problem, the False Discovery Rate (FDR) method was used and the lists were cut off at specific values of the FDR. Additionally, the results were filtered to include only those miRNAs that were expressed threefold above the background standard deviation in at least one sample.
Data availability
The primary microarray data is deposited in the ArrayExpress database with accession number E-MEXP-372.
Additional data files
The following additional data are available with the online version of this article: miRNAs differentially expressed between hippocampus and combined hematopoietic samples (Additional data file 1), miRNAs differentially expressed between BMMC and Th1, Th2 and Pu.1-/- cells (Additional data file 2), miRNAs differentially expressed between spleen B versus pro-B (Additional data file 3), and an analysis of the probe melting temperatures (Tm) versus the average signal obtained in the arrays (Additional data file 4).
Supplementary Material
Additional data file 1
For normalization the VSN method was used [37]. A standard t-test on the generalized log transformed expression levels was used to rank the miRNAs and compute p values. The results were filtered to discard any miRNA whose signal was not three times above the background in at least one of the two groups and whose absolute fold change was less than 3.
Click here for file
Additional data file 2
To handle the low sample numbers, the Bayesian corrected t-test [38] was used to rank and compute p values. The different comparisons had greatly varying significances, so different cut offs were used on each list. For the BMMC versus Th1 and Th2 comparisons, a FDR cutoff of 10-2 was used, with a filter of three times background levels and a fold change of three. For the BMMC versus Pu.1-/- comparison a FDR of 10-6 and four times background filter was used.
Click here for file
Additional data file 3
The FDR cut off was 10-6 and the background filter was four times.
Click here for file
Additional data file 4
There is a clear correlation between Tm and strength of signal for all the probes used (left panel). The same is true if only the probes for miRNAs that were identified as 'high in hematopoietic cells' (see Figure 5a) are considered (right panel). Also, the latter probes have Tm that span almost the entire range of Tm seen for all the probes, and have an average Tm (56°C) only slightly higher than the average of all the probes (54°C). This means that overall, the use of these different oligos as probes in the arrays is in fact valid, even though we may have lost some of the low Tm miRNAs as false negative signals, and some of the high Tm probes may have given a few false positive signals.
Click here for file
Acknowledgements
We thank Dr P Laslo and Dr H Singh for the Pu.1-/- cells. K.M.A. is a fellow of the Damon Runyon Cancer Research Fund (DRG-1682). C.X. is a Cancer Research Institute postdoctoral fellow. A special thanks to Prof A Siccardi for help and support. This work was supported by NIH grants to A.R., K.R., and K.S.K., and a grant from the Sandler Program for Asthma Research to A.R.
Figures and Tables
Figure 1 Design and reproducibility of microarrays. (a) Examples of microarrays: three membranes were used for each biological sample; arrays for Pu.1-/- cells, BMMC and hippocampus are shown. On the left, ethidium bromide staining of total RNA run on a denaturing gel for RNA quantification and quality control. (b) Plot of the standard deviation over replicates versus the mean of each replicate. The red line is a lowest fit to the distribution and the blue dotted line is twice the value of the red one. Points below the blue curve are considered good replicates; those above it are filtered out as too noisy. For this dataset, 86% of the spots were considered as good. BMMC, bone marrow-derived mast cells.
Figure 2 Comparison of miR-150 expression by arrays and Northern blotting. (a) Array data show that miR-150 (arrows) is highly expressed in spleen B and naïve T cells, but not in pro-B cells or fully differentiated Th1 and Th2 clones. (b) Northern blot analysis for miR-150 in various lymphoid and non-lymphoid tissues and cell types. U6 RNA levels are shown as loading control. (c) Northern blots of different cell types unstimulated or stimulated for the indicated amounts of time with either PMA and ionomycin (BMMC) or anti-CD3 and anti-CD28 (Th1 and Th2 primary cells). Preliminary data obtained both in Northern blot and arrays showed no difference in miRNA expression between BMMC left untreated or treated with cyclosporin A (CsA) (not shown). (d) Ethidium bromide staining of gel of total RNA from samples used in Figure 2c, showing equivalent RNA amounts. BMMC, bone marrow-derived mast cells; MEF, mouse embryo fibroblast.
Figure 3 Microarrays and Northern blots correlate qualitatively and quantitatively. Northern blots for miRNA expression in mast cells (left panels), or T cells (right panels). BMMC were treated with cyclosporin A or PMA and ionomycin for the indicated amounts of time. Loading control is the same as Figure 2d. First row underneath the panels: ratio between the intensities of the Northern blot bands as assessed by phosphorimager and quantified by ImageQuant; all the samples are compared with Pu.1-/- cells except for miR-223, where each sample is compared with BMMC. Second row: these are also ratios between the intensities of the Northern blot bands, but the T cell samples are compared directly to each other. This allows a better direct comparison with the numbers on the third row, which are the ratios of BMMC versus Pu.1-/- cells (left panels) and D5 versus D10 versus naïve T cells as obtained from the arrays (right panels). BMMC, bone marrow-derived mast cells; CsA, cyclosporin A; n.d., not detectable.
Figure 4 Additional Northern blots showing miRNA expression in various tissues and cell types. Shown are Northern blot data for the indicated miRNAs in different unstimulated cell types. Asterisks indicate bands of the size of pre-miRNA (60-70 nucleotides), which were detected in Northern blot for only a subset of the miRNAs analyzed. There was a good correlation overall between Northern blot data and expression data from the arrays (see heat map in Figure 5 and Table 2), with some exceptions, as discussed in the text. BMMC, bone marrow-derived mast cells; DN T, double-negative thymocyte.
Figure 5 Analysis of microarray data. (a) Heat map of miRNAs expressed at least three times over the background for at least one of the samples. (b) Hierarchical clustering of hematopoietic samples (see analysis in Table 3). DN T, double-negative thymocyte; MEF, mouse embryo fibroblast.
Figure 6 MiRNA expression and lineage commitment. Expression of miRNA in the hematopoietic system changes depending on the differentiation status. Part of the hematopoietic differentiation tree is represented: myeloid and lymphoid progenitors derive from a common progenitor, which is represented by the model Pu.1-/- cell line (see text). The common lymphoid progenitor gives rise to B and T lymphocytes and the common myeloid progenitor gives rise to mast cells and other cells types. Superimposed on this diagram are examples of miRNAs that were discovered to be differentially expressed between the indicated precursor/progeny pairs using array analysis with confirmation by Northern blot. DN, double-negative.
Table 1 Northern blot quantification of miRNA
miR BMMC Pu.1-/- Naïve Th1 (49 h) D5 Th1 Th2 (49 h) D10 Th2
150 n.d. n.d. 39.4 3.5 n.d. 1.0 n.d.
146 1.7 1.0 2.8 3.7 6.8 1.3 1.0
142s 2.0 1.0 32.0 7.5 3.8 4.7 3.9
26a 3.1 1.0 13.8 1.4 3.0 1.2 2.5
27a 3.9 1.0 1.3 0.9 1.4 0.8 1.5
223 1.0 45.0 n.d. n.d. n.d. n.d. n.d.
Let7d 2.3 1.0 4.7 2.1 2.5 1.7 1.7
222 1.0 1.2 n.d. n.d. 1.1 n.d. n.d.
Northern blots for miRNA expressed in mast cells, precursor cells, and T cells at various stages of differentiation were quantified by phosphorimager. BMMC, bone marrow-derived mast cell; DN T, double-negative thymocyte; n.d., not detectable.
Table 2 Correlation between arrays and Northern blot data
DNT Pro-B Spleen B BMMC Hippocampus
MiR 7 Arrays + ++ ++ +/- ++
Northerns + +++ ++ n.a. n.a.
MiR 24 Arrays +++ ++++ +++ ++ ++
Northerns ++ ++++ +++ n.a. n.a.
MiR 26a Arrays + + ++ +++ +++
Northerns + ++ +++ n.a. n.a.
MiR 29a Arrays + + ++ ++ +++
Northerns + + ++ nd +++
MiR 93 Arrays ++ ++ +/- ++ +/-
Northerns + ++ + n.a. n.a.
MiR 99a Arrays n.s.s. n.s.s. n.s.s. n.s.s. n.s.s.
Northerns +/- +/- - nd nd
MiR 101 Arrays ++ +++ + +++ ++
Northerns + + + nd ++
MiR-107 Arrays +++ +++ ++ ++ ++
Northerns + ++ + n.a. n.a.
MiR-127 Arrays +/- + + +/- +
Northerns - - - n.a. n.a.
MiR-142-3p Arrays +++ +++ ++++ +++ +
Northerns ++ +++ ++++ ++ n.a.
MiR-142-5p Arrays ++ + +++ ++ +
Northerns + ++ +++ + n.a.
MiR-144 Arrays n.s.s. n.s.s. n.s.s. n.s.s. n.s.s.
Northerns - - - n.a. n.a.
MiR-148 Arrays + + + + +/-
Northerns + + + n.a. n.a.
MiR-150 Arrays - + +++ + +/-
Northerns - - +++ n.a. -
MiR-181b Arrays + + + + ++
Northerns + + - n.a. +
MiR-191 Arrays +/- + + + +
Northerns + ++ ++ n.a. n.a.
MiR-199 Arrays + + + + +/-
Northerns - + - n.a. n.a.
MiR-206 Arrays ++ +++ + +/- +/-
Northerns - - - - n.a.
MiR-213 Arrays n.s.s. n.s.s. n.s.s. n.s.s. n.s.s.
Northerns - +/- - n.a. n.a.
MiR-223 Arrays + +++ + +/- +
Northerns +/- - + + n.a.
MiR-342 Arrays + + + + +
Northerns + +++ ++ n.a. n.a.
The table summarizes and compares the Northern blot data shown in Figure 4 and the array data shown in the heat map in Figure 5. Northern blot and array data were scored independently using an arbitrary scale from undetectable (-) to strongly detected (++++) to indicate relative signal intensity in each case. BMMC, bone marrow-derived mast cell; DN T, double-negative thymocyte; n.s.s., non statistically significant hybridization signal; n.a., not analyzed.
Table 3 Correlation coefficients (Pearson correlation)
Naïve DN T Pro-B Pu.1-/- Spleen B BMMC Th1 Th2
Naïve 1.00
DN T 0.50 1.00
Pro-B 0.47 0.94 1.00
Pu.1-/- 0.56 0.83 0.78 1.00
Spleen B 0.76 0.87 0.80 0.81 1.00
BMMC 0.63 0.74 0.68 0.78 0.79 1.00
Th1 0.68 0.85 0.79 0.84 0.88 0.92 1.00
Th2 0.67 0.83 0.80 0.84 0.87 0.90 0.96 1.00
The table represents a standard statistical correlation between the indicated samples, where 1 = perfectly correlated and 0 = uncorrelated. BMMC, bone marrow-derived mast cell; DN T, double-negative thymocyte.
==== Refs
Denli AM Tops BB Plasterk RH Ketting RF Hannon GJ Processing of primary microRNAs by the Microprocessor complex. Nature 2004 432 231 235 15531879 10.1038/nature03049
Bartel DP MicroRNAs: genomics, biogenesis, mechanism, and function. Cell 2004 116 281 297 14744438 10.1016/S0092-8674(04)00045-5
Ambros V The functions of animal microRNAs. Nature 2004 431 350 355 15372042 10.1038/nature02871
Meister G Tuschl T Mechanisms of gene silencing by double-stranded RNA. Nature 2004 431 343 349 15372041 10.1038/nature02873
Hobert O Common logic of transcription factor and microRNA action. Trends Biochem Sci 2004 29 462 468 15337119 10.1016/j.tibs.2004.07.001
Calin GA Dumitru CD Shimizu M Bichi R Zupo S Noch E Aldler H Rattan S Keating M Rai K Frequent deletions and down-regulation of micro-RNA genes miR15 and miR16 at 13q14 in chronic lymphocytic leukemia. Proc Natl Acad Sci USA 2002 99 15524 15529 12434020 10.1073/pnas.242606799
Calin GA Liu CG Sevignani C Ferracin M Felli N Dumitru CD Shimizu M Cimmino A Zupo S Dono M MicroRNA profiling reveals distinct signatures in B cell chronic lymphocytic leukemias. Proc Natl Acad Sci USA 2004 101 11755 117560 15284443 10.1073/pnas.0404432101
Griffiths-Jones S The microRNA Registry. Nucleic Acids Res 2004 32 D109 111 14681370 10.1093/nar/gkh023
Berezikov E Guryev V van de Belt J Wienholds E Plasterk RH Cuppen E Phylogenetic shadowing and computational identification of human microRNA genes. Cell 2005 120 21 24 15652478 10.1016/j.cell.2004.12.031
Xie X Lu J Kulbokas EJ Golub TR Mootha V Lindblad-Toh K Lander ES Kellis M Systematic discovery of regulatory motifs in human promoters and 3' UTRs by comparison of several mammals. Nature 2005 434 338 345 15735639 10.1038/nature03441
Ambros V MicroRNA pathways in flies and worms: growth, death, fat, stress, and timing. Cell 2003 113 673 676 12809598 10.1016/S0092-8674(03)00428-8
Lim LP Lau NC Garrett-Engele P Grimson A Schelter JM Castle J Bartel DP Linsley PS Johnson JM Microarray analysis shows that some microRNAs downregulate large numbers of target mRNAs. Nature 2005 433 769 773 15685193 10.1038/nature03315
Miska EA Alvarez-Saavedra E Townsend M Yoshii A Sestan N Rakic P Constantine-Paton M Horvitz HR Microarray analysis of microRNA expression in the developing mammalian brain. Genome Biol 2004 5 R68 15345052 10.1186/gb-2004-5-9-r68
Sempere LF Freemantle S Pitha-Rowe I Moss E Dmitrovsky E Ambros V Expression profiling of mammalian microRNAs uncovers a subset of brain-expressed microRNAs with possible roles in murine and human neuronal differentiation. Genome Biol 2004 5 R13 15003116 10.1186/gb-2004-5-3-r13
Baskerville S Bartel DP Microarray profiling of microRNAs reveals frequent coexpression with neighboring miRNAs and host genes. RNA 2005 11 241 247 15701730 10.1261/rna.7240905
Galli SJ Kalesnikoff J Grimbaldeston MA Piliponsky AM Williams CM Tsai M Mast cells as 'tunable' effector and immunoregulatory cells: recent advances. Annu Rev Immunol 2005 23 749 786 15771585 10.1146/annurev.immunol.21.120601.141025
Esau C Kang X Peralta E Hanson E Marcusson EG Ravichandran LV Sun Y Koo S Perera RJ Jain R MicroRNA-143 regulates adipocyte differentiation. J Biol Chem 2004 279 52361 52365 15504739 10.1074/jbc.C400438200
Thomson JM Parker J Perou CM Hammond SM A custom microarray platform for analysis of microRNA gene expression. Nat Methods 2004 1 47 53 15782152 10.1038/nmeth704
Nelson PT Baldwin DA Scearce LM Oberholtzer JC Tobias JW Mourelatos Z Microarray-based, high-throughput gene expression profiling of microRNAs. Nat Methods 2004 1 155 161 15782179 10.1038/nmeth717
Krichevsky AM King KS Donahue CP Khrapko K Kosik KS A microRNA array reveals extensive regulation of microRNAs during brain development. RNA 2003 9 1274 1281 13130141 10.1261/rna.5980303
DeKoter RP Walsh JC Singh H PU.1 regulates both cytokine-dependent proliferation and differentiation of granulocyte/macrophage progenitors. EMBO J 1998 17 4456 4468 9687512 10.1093/emboj/17.15.4456
Walsh JC DeKoter RP Lee H Smith ED Lancki DW Gurish MF Friend DS Stevens RL Anastasi J Singh H Cooperative and antagonistic interplay between PU.1 and GATA-2 in the specification of myeloid cell fates. Immunity 2002 17 665 676 12433372 10.1016/S1074-7613(02)00452-1
Chen CZ Li L Lodish HF Bartel DP MicroRNAs modulate hematopoietic lineage differentiation. Science 2004 303 83 86 14657504 10.1126/science.1091903
Ansel KM Lee DU Rao A An epigenetic view of helper T cell differentiation. Nat Immunol 2003 4 616 623 12830136 10.1038/ni0703-616
Murphy KM Reiner SL The lineage decisions of helper T cells. Nat Rev Immunol 2002 2 933 944 12461566 10.1038/nri954
Lewis BP Shih IH Jones-Rhoades MW Bartel DP Burge CB Prediction of mammalian microRNA targets. Cell 2003 115 787 798 14697198 10.1016/S0092-8674(03)01018-3
Lewis BP Burge CB Bartel DP Conserved seed pairing, often flanked by adenosines, indicates that thousands of human genes are microrna targets. Cell 2005 120 15 20 15652477 10.1016/j.cell.2004.12.035
Rajewsky N Socci ND Computational identification of microRNA targets. Dev Biol 2004 267 529 535 15013811 10.1016/j.ydbio.2003.12.003
John B Enright AJ Aravin A Tuschl T Sander C Marks DS Human microRNA targets. PLoS Biol 2004 2 e363 15502875 10.1371/journal.pbio.0020363
Krek A Grun D Poy MN Wolf R Rosenberg L Epstein EJ Macmenamin P da Piedade I Gunsalus KC Stoffel M Combinatorial microRNA target predictions. Nat Genet 2005 37 495 500 15806104 10.1038/ng1536
Akashi K Traver D Miyamoto T Weissman IL A clonogenic common myeloid progenitor that gives rise to all myeloid lineages. Nature 2000 404 193 197 10724173 10.1038/35004599
DeKoter RP Singh H Regulation of B lymphocyte and macrophage development by graded expression of PU.1. Science 2000 288 1439 1441 10827957 10.1126/science.288.5470.1439
Ansel KM Greenwald RJ Agarwal S Bassing CH Monticelli S Interlandi J Djuretic IM Lee DU Sharpe AH Alt FW Deletion of a conserved Il4 silencer impairs T helper type 1-mediated immunity. Nat Immunol 2004 5 1251 1259 15516924 10.1038/ni1135
Solymar DC Agarwal S Bassing CH Alt FW Rao A A 3' enhancer in the IL-4 gene regulates cytokine production by Th2 cells and mast cells. Immunity 2002 17 41 50 12150890 10.1016/S1074-7613(02)00334-5
Avni O Lee D Macian F Szabo SJ Glimcher LH Rao A T(H) cell differentiation is accompanied by dynamic changes in histone acetylation of cytokine genes. Nat Immunol 2002 3 643 651 12055628
Agarwal S Rao A Modulation of chromatin structure regulates cytokine gene expression during T cell differentiation. Immunity 1998 9 765 775 9881967 10.1016/S1074-7613(00)80642-1
Huber W von Heydebreck A Sultmann H Poustka A Vingron M Variance stabilization applied to microarray data calibration and to the quantification of differential expression. Bioinformatics 2002 18 S96 S104 12169536
Baldi P Long AD A Bayesian framework for the analysis of microarray expression data: regularized t-test and statistical inferences of gene changes. Bioinformatics 2001 17 509 519 11395427 10.1093/bioinformatics/17.6.509
|
16086853
|
PMC1273638
|
CC BY
|
2021-01-04 16:15:44
|
no
|
Genome Biol. 2005 Aug 1; 6(8):R71
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r71
|
oa_comm
|
==== Front
Genome BiolGenome Biology1465-69061465-6914BioMed Central London gb-2005-6-8-r721608685410.1186/gb-2005-6-8-r72MethodGenome-wide promoter extraction and analysis in human, mouse, and rat Xuan Zhenyu [email protected] Fang [email protected] Jinhua [email protected] Gengxin [email protected] Michael Q [email protected] Cold Spring Harbor Laboratory, 1 Bungtown Road, Cold Spring Harbor, NY 11724, USA2005 1 8 2005 6 8 R72 R72 29 3 2005 23 5 2005 11 7 2005 Copyright © 2005 Xuan et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
An investigation of how to improve mammalian promoter prediction by incorporating both transcript and conservation information leads to the creation of CSHLmpd, a mammalian promoter database.
Large-scale and high-throughput genomics research needs reliable and comprehensive genome-wide promoter annotation resources. We have conducted a systematic investigation on how to improve mammalian promoter prediction by incorporating both transcript and conservation information. This enabled us to build a better multispecies promoter annotation pipeline and hence to create CSHLmpd (Cold Spring Harbor Laboratory Mammalian Promoter Database) for the biomedical research community, which can act as a starting reference system for more refined functional annotations.
==== Body
Background
Gene transcription is regulated by transcription factors (TFs), binding mostly and specifically to the promoter regions. Recent developments of technologies for studying genome-wide transcriptional regulation include microarray expression and chromatin immunoprecipitation (ChIP). The analysis of data from such high-throughput technologies often requires a large set of promoter sequences. Some existing promoter databases for mammals, such as the Eukaryotic Promoter Database (EPD) [1] and the Database of Transcriptional Start Site (DBTSS) [2], were constructed by collecting experimentally identified promoter regions. The promoter data are, however, very limited in these databases. Computational methods have been developed to predict promoters in genomic sequences, but the performance is far from satisfactory, especially for non-CpG-island-related promoters [3,4]. Although known mRNAs have also been used to map the potential promoter regions [5-8] and genome-wide full-length cDNA sequencing projects have contributed lots of very valuable data [9-11], currently only 47-50% of human and mouse genes (or 21% of rat genes) have reference mRNAs (Table 1). It is therefore highly desirable to build a more comprehensive and accurate promoter dataset for the functional genomic community.
We have integrated sequence conservation with our promoter prediction program FirstEF [12] to improve the accuracy of prediction. FirstEF was developed as an ab initio human first-exon prediction program, which is capable of predicting noncoding first exons together with the corresponding promoters. It has been used in conjunction with mRNA/expressed sequence tags (EST) transcript information to produce an initial human promoter annotation pipeline (R. Davuluri and I. Gross, personal communication) because gene transcripts and models can be used to identify promoters with high confidence [13]. At the same time, TWINSCAN [14] and other studies [15] have shown that integrating genomic homology information can increase gene-prediction accuracy by about 10% compared with the use of ab initio methods alone, and conserved features in promoters have also been used to improve promoter identification in a small dataset [16]. Here, we set out to test if, and to what degree, integrating homology information from mouse and rat genomes can help to further improve human promoter prediction. We found that homologous sequence comparison can substantially increase the prediction accuracy. This enables us to build an improved multispecies promoter annotation pipeline by extracting known and predicted promoters, and to create a comprehensive mammalian promoter database (CSHLmpd) with on-the-fly analysis tools as a valuable public resource to facilitate future mammalian gene-regulatory network studies. As a convenient operational definition, we refer to 'promoter' in this paper as the genomic region (-700, +300) bp with respect to the transcription start site (TSS).
Results
We used orthologous genes to detect sequence conservation in promoter regions. To do this, we first identified all genic regions in the genomes on the basis of known and predicted transcripts, then collected all known promoters from present promoter annotations in the public databases and all predicted promoters produced by the original FirstEF. These promoters were then linked to downstream genes (see below). We took known promoters from the human-rodent orthologous genes and observed significant conservation in promoter sequences. We then used this conservation signal to improve de novo promoter prediction, and in the end constructed a reference promoter database for each of the three mammalian genomes.
Human, mouse and rat genes and orthologous gene sets
By aligning all known and predicted transcripts to the latest human, mouse and rat genomes we obtained 34,949, 35,073, 30,679 genes (see Materials and methods), which include 29,360, 25,571 and 22,643 canonical genes (based on RefSeq [17] mRNA and Ensembl [18] prediction) in these genomes, respectively. The orthologous relationship of these canonical genes is defined using EnsMart [19], which is based on similarity analysis of Ensembl transcripts and genes. We obtained 19,179 human-mouse-rat three-species orthologous gene triplets, and 1,967, 1,420 and 2,268 human-mouse, human-rat and mouse-rat two-species orthologous gene pairs respectively. Promoter conservation was studied in these orthologous genes.
Known promoter collection and promoter prediction in human, mouse and rat genomes
For each species we collected known promoters from EPD and DBTSS. We also collected known promoters from GenBank [20] by keyword search (see Materials and methods), and the promoter regions identified by luciferase assay and ChIP of TAF250 and RNA polymerase II in the Encyclopedia of DNA Elements (ENCODE) regions. These known promoter sequences were aligned with the genome by BLAT [21] to get the locations of TSSs. The total unique known TSSs in human, mouse and rat are 14,314, 8,141 and 943, respectively [21]. We also predicted 608,057, 449,132 and 427,130 promoters in these genomes separately using FirstEF with default parameter setting. Repeats in the genome were not masked. TSS locations of all known and predicted promoters were compared with the identified gene regions. A TSS is assigned to a gene when it is located in the genic region or upstream of the 5' end of the gene by no more than 5 kb (for RefSeq genes) or 20 kb (for other genes). By doing so, we obtained such 'gene-related' TSSs/promoters for further analysis. Predicted 'gene-related' promoters are also defined as 'transcript-supported promoters' if they overlap the 5' end of any transcript in a gene. Other predicted TSSs that were not gene-related were potential 'novel TSSs' and were not further analyzed. We used known promoters as training data to detect promoter conservation signal and then compared it with the signal in predicted promoters to reduce false-positive promoter predictions.
Statistical similarity among known promoters of orthologous genes
Pairwise comparison of known promoters
Of the orthologous gene pairs, 3,649 human-rodent and 214 mouse-rat pairs have known promoters in both species. We compared these known promoters by ClustalW [22] to measure the conservation in promoters. The conservation score is defined as the percentage of identical base-pairs in a 1 kb region. Using randomly selected known promoters of non-orthologous genes (see Materials and methods), we found that such conservation positively correlates with the GC content, especially when GC content is greater than 65%, and surprisingly, that the conservation distribution is independent of the species used for comparison (Figure 1a). We also measured the conservation for randomly selected 1 kb genomic DNA sequences, and found the same distribution of conservation score (Figure 1b, species-related data are not shown). Therefore, we chose the 99% quantile as the conservation cutoff for discriminating the pairwise 'high-scoring promoters' (that is, 1% error threshold or 1PET). We found that the conservation threshold is 48.8% for sequences of high GC content (greater than 65%), and 45.8% for the rest. The distribution of conservation score in known human-rodent promoter pairs is shown in Figure 1b, which consists of two mixed populations: one is similar to that of the sequence pairs in the two control sets, and the other is peaked much higher than 1PET.
We then defined a promoter pair as a homologous promoter pair, and the promoters as homologous promoters, if the conservation score is higher than 1PET (the pairwise cutoff rule). Using these cutoffs, we found 2,841 of 4,140 human known promoters in those 3,649 human-rodent orthologous gene pairs, and 152 of 229 mouse known promoters in those 214 mouse-rat orthologous gene pairs. In total, around 66-68% of known promoters can match highly conserved counterparts in the orthologous genes. The average conservation score is around 55% between human-rodent homologous promoter pairs, and 85% between mouse-rat homologous promoter pairs (Figure 1c).
Three-species promoter comparisons
We also analyzed known promoter conservation in 158 human-mouse-rat three-way orthologous gene triplets, which have 249 all-species promoter triplets. Using ClustalW to randomly align selected 1 kb sequences from human, mouse and rat genomes, we found that only 1% of the 1 kb triplets had conservation score higher than 21.8%. Here, the conservation score is defined as the percentage of identical base-pairs in the multiple alignments of 1 kb sequences. Using this cutoff, we identified 76 known promoter triplets, and the distribution of conservation score is shown in Figure 1d.
In the genome, functional regions (such as coding regions) are usually conserved under selection pressure during evolution. Hence the significantly higher conservation of homologous promoter pairs and triplets encouraged us to test whether it could be used to improve promoter prediction.
Improving promoter prediction by incorporating both mRNA annotation and promoter conservation information
We are able to combine the conservation signal in homologous promoters with promoter models used in FirstEF program to improve promoter prediction. We compared the performance of four methods. Method 0 is original FirstEF. Method 1 is a de novo FirstEF (with the post-clustering filter [23]) that only keeps the best-predicted promoters from the original FirstEF predictions within a 1,000 bp region. Method 2 uses transcript information to filter out the false positives of Method 0 predictions that are located within the gene region. Method 3 incorporates conservation signals into Method 2: first, predicted promoters are selected by using Method 2, and then for genes with homologous promoters, only the conserved predicted promoters will be reported (see Materials and methods and Figure 2). Here the conservation signal was measured between human and rodent promoters in the same orthologous gene pair, and the pairwise cutoff rule defined above was used to identify homologous promoters.
We collected 8,949 well annotated human genes, each of which has at least one known TSS and has at least one orthologous gene in mouse or rat, to do the test. There are in total 13,313 unique known TSSs for these human genes, with 9,806 being at least 500 bp apart (see Materials and methods). In both sets, we shortened each gene by 5 kb (or half of the gene length if the gene is shorter than 5 kb) from its 5' end to simulate 5' incomplete genes that are most common in the current gene annotations.
We found that by incorporating mRNA (Method 2) and promoter conservation information (Method 3), we could improve promoter prediction over the de novo FirstEF (Method 1) (Table 2). With conservation and mRNA information together, we achieved 66% in specificity and 69% in sensitivity on the 13,313 unique TSS set, corresponding to improvements of 20% and 2% respectively. Comparing this with the original FirstEF prediction (Method 0), we found that although sensitivity dropped 3%, an improvement of 20% in specificity is well worth the effort. Just using transcript information, Method 2 can improve on Method 1 by 11% in specificity and 3% in sensitivity (Table 2a). For those 9,806 known TSSs separated by at least 500 bp, we found that Method 3 still gives the largest improvement, with specificity (Sp) and sensitivity of prediction (Sn) reaching 60% and 66% (26% and 2% higher than those by Method 1), respectively (Table 2b). Of the 8,949 human genes, 5,893 (66%) have homologous promoters, and the specificity and sensitivity of promoter prediction for these genes by Method 3 are 69% and 82%, respectively (Table 2c). On the basis of the new definition of CpG-island [24], we found that the prediction of CpG-island related promoter has higher sensitivity and specificity (Figure 3a,b), consistent with the fact that FirstEF offers better prediction for CpG-related promoters than non-CpG-related ones. For CpG-island related promoters with homologous counterpart, the Sp and Sn of the prediction can reach 70% and 91% respectively. Very strikingly, the improvement for non-CpG related promoter prediction by homology information is much more dramatic (Figure 3). These results clearly show the considerable value of cross-species comparison in promoter prediction.
Incorporation of cross-species conservation in whole-genome promoter/TSS prediction
Encouraged by the enhancement in promoter prediction performance obtained by combining FirstEF promoter models with conservation signal and transcript information, we applied Method 3 to annotate human, mouse and rat genomes (Figure 2). In addition to the known and the original FirstEF-predicted TSSs, we defined two types of surrogate TSSs: bidirectional TSSs and RefSeq END TSSs. If the intergenic region between two adjacent 'head-to-head' (divergent) genes is shorter than 2 kb, their 5' ends are defined as bidirectional TSSs even if no promoter is predicted. For a gene with a RefSeq mRNA, the 5'-end location of the RefSeq mRNA is defined as RefSeq END TSS if there is no other known or predicted TSS linked to this gene. For each gene, we always keep its known promoters and assign these with the highest reliability. Method 3 was then used to select representative promoters from other predicted promoters of this gene, with homologous promoters having higher priority to be chosen (see Materials and methods for details) to reduce the false-positive rate. For simplicity, two TSSs of the same gene are regarded as alternative TSSs. By doing this, we obtained 55,513, 46,207 and 37,479 known and predicted promoters for 26,820, 22,228 and 21,125 genes in human, mouse and rat, respectively. With the current methods, we could not assign promoters for the remaining 8,129, 9,481 and 9,554 human, mouse and rat genes (most of them are predicted genes or only have single EST matches, see below). The detailed statistics are listed in Table 3. After comparing gene boundaries and TSSs to the CpG-islands (see Materials and methods), we found that most RefSeq genes are CpG-island related. In total, 68%, 54% and 56% promoters obtained above for human, mouse and rat are CpG-island related.
From the above promoter/TSS sets, we found 21,594, 21,501 and 17,257 homologous promoters for 13,432, 14,626 and 12,302 genes in human, mouse and rat. Of the mammalian canonical genes with orthologous genes, 60% to 70% have homologous promoters. However, our methods can assign promoters for only a small portion of the TWINSCAN and GenomeScan [25] predicted genes (42%), compared to 82% of the canonical genes (data not shown). This may be due either to the sensitivity of FirstEF, or to the fact that most predicted genes start from putative translational initiation sites (ATG) and the missing 5' exons and intron regions can span beyond our promoter search limit (20 kb upstream of the predicted gene boundary). The lack of complete 5' ends in non-RefSeq genes can also explain why we saw them to be less likely to be CpG-island related.
Cold Spring Harbor Laboratory Mammalian Promoter Database
To store the information about all the genes and promoters we annotated, we have constructed the Cold Spring Harbor Laboratory Mammalian Promoter Database (CSHLmpd [26]). It consists of three species-specific promoter sub-databases for human (HSPD), mouse (MMPD) and rat (RNPD). They are linked by homologous promoters wherever orthologous gene information is available. Each is currently equipped with two basic front-end components: a genome-wide browser, Gbrowse [27], to display information graphically; and a query-fetch system to query and extract promoters based on a gene identifier (such as GenBank accession number, UniGene [28] cluster ID, LocusLink [28] ID or gene name). In CSHLmpd, users can either search for promoters of their genes of interest in one species or get homologous promoters from other species. To make the database both a data resource and an analysis platform, we provide two sequence-alignment tools for homologous promoter analysis. ClusterW is for global multiple sequence alignment in the regions of user-selected promoters, and PromoterWise, a local alignment tool, is embedded to align each pair of promoter regions (E. Birney, unpublished data). We have also used MLAGAN [29] to do global multiple sequence alignment in the regions that include genes and their 5,000-bp upstream sequences to show the conservation at a larger scale. More promoter-analysis tools will be added in the future.
In addition, there is another related database, the Transcription Regulatory Element Database (TRED) [30]. It includes curated biological information, such as transcription factor binding sites (TFBSs) and regulation pathways/networks as well as cis-element analysis tools. Figure 4 shows some representative screen shots of the database user interface. For the user's convenience, we have classified the promoter quality in the following order (from the highest to the lowest): known promoters (EPD, DBTSS, GenBank annotation, promoters identified by luciferase assay or ChIP), RefSeq END promoters, transcript-supported promoters, bidirectional promoters, and other predicted promoters (see Materials and methods). If promoters with different qualities are linked to a gene, users can choose to retrieve only the most reliable one, any, or all of them. This promoter database is publicly available and all data are free for academic use.
Facilitating large-scale gene regulation studies and promoter array construction
Expression microarray and ChIP-chip (ChIP followed by microarray analysis of DNA) technologies have become important and widely used approaches to study gene expression and regulation at large scales. Being able to extract a large set of mammalian promoter sequences is a critical step for such studies.
To demonstrate the use of CSHLmpd, we have extracted a promoter sequence dataset for the Affymetrix human array HG-U133A. Out of the total of 22,283 probe sets for most known human genes [31] on this array, from the annotation we were able to obtain promoters from CSHLmpd for 20,903 of them. Because multiple probe sets can belong to the same gene, 13,014 promoters were retrieved. These include 6,052 known promoters and 4,550 predicted homologous promoters. No promoter could be assigned for only 1,380 probe sets. Among these, 448 were mapped to 353 genes without promoter information in our database, and 932 were created from poorly aligned mRNAs and ESTs, which were not used to construct the genes in the first place, or from other ESTs that do not overlap with any gene in our database (see Materials and methods). This HG-U133A Affymetrix promoter set can be freely downloaded from our FTP server [32], where one can also find separately prepared promoter sequence sets for all human, mouse and rat RefSeq genes. These RefSeq gene promoter sets include all DBTSS-defined promoters and RefSeq END TSS. Users can also create other customized promoter sequence sets for different arrays (or gene indices) using the CSHLmpd query tools. We also plan to provide more customized promoter sequence sets for making promoter chips that can be used for large-scale ChIP-chip studies or epigenetic mapping projects (such as for DNA methylation).
Discussion
Our method first collected known and predicted promoters in the whole genome. Then transcript and conservation information were used to filter the false positives from the predictions. Our test presented in this paper has proved that using both transcript and conservation information, together with FirstEF, will improve the accuracy of promoter prediction compared with the use of transcript information alone (for example, PromSer, Source). To our knowledge, this is the first attempt to integrate conservation information with de novo first-exon prediction on a genome-wide scale.
In collaboration with an experimental group (L. Stubbs, personal communication), we previously tested our FirstEF prediction on 48 human genes in chromosome 19 using reporter assays. Among these, 26 genes had promoters correctly predicted, and eight did not. This gave a sensitivity and specificity of 54% and 65%, respectively, at the gene level. However, there were a total of 105 predicted promoters around these genes, which led to a specificity of only 25% at the promoter level (data not shown). Therefore, while the experimental evaluation proves that de novo FirstEF performs well in predicting promoters for novel genes, it also shows its limitations on prediction specificity. A more systematic experimental test of 300 mouse promoters will be found in [33]. Our work presented here shows that both mRNA information and cross-species conservation can significantly improve the specificity of promoter prediction.
We have also demonstrated that conservation signal can be integrated with promoter models to improve the accuracy of promoter prediction. Our method uses conservation signal in the potential promoter regions, which can greatly reduce false positives when comparing using just mRNA or conservation information alone, especially when known mRNAs only have partial coding regions. Furthermore, without mRNA information, homologous information by itself cannot produce better overall prediction (data not shown), partly as because of a higher degree of conservation in exons. To decrease false predictions caused by exon conservation as much as possible, we not only used the information from known genes, but also predicted genes from some well known gene-finding methods. In this way, we can reduce the promoter search regions for known genes, and may obtain additional theoretical evidence for predicted genes when their promoters are predicted [4]. These potential novel genes with predicted promoters, especially when the promoters are evolutionarily conserved, could be valuable candidates for experimental validation. In our recent experiments, we have shown that about 25% of those novel genes have spliced transcripts [33].
To detect the conservation in promoter regions, we tested several different promoter definitions. They included upstream 200 bp of TSSs, -400 to +100 bp, -700 to +300 bp, and -1,500 to +500 bp around TSS. We found that the peak of the conservation score is closer to that of the control sequence set when promoter regions are too short or too long. Among these four promoter definitions, -700 to +300 bp around TSSs gave the best discrimination between the known promoter-training set and the control set. This indicated that many conserved TFBSs tend to cluster in the approximately 1 kb region near the TSS [34].
In our studies, we have observed that, if lower thresholds of the original FirstEF (such as Pexon = 0.3, Ppromoter = 0.25, Pdonor = 0.25) are used, the prediction sensitivity can be increased at the expense of specificity. In this case, however, even though mRNA and conservation information could help regain some specificity, the overall accuracy would actually be worse than that with default FirstEF thresholds (data not shown).
We cannot identify conservation signal for 27% of known human promoters and 17% of known rodent promoters (see our FTP site [32]). This may be due to the faster promoter divergence in the corresponding genes. The percentage of predicted promoters without homology that were detected was higher than that of known promoters because of the bias of existing known promoter data and false positives of promoter prediction. We hope to develop more sensitive methods for promoter-specific conservation detection in order to improve promoter prediction in the future.
Materials and methods
Human, mouse and rat genome releases
Human NCBI build 35 (May 2004), mouse mm5 (May 2004), and rat assembly rn3 (June 2003), were downloaded from the University of California at Santa Cruz (UCSC) website [35].
Genic region identification in the genomes
mRNAs from RefSeq and GenBank (mRNA), and transcripts predicted by Ensembl, TWINSCAN and GenomeScan (RefSeq XM) in the annotation of UCSC genome assemblies were obtained. They were aligned to the genomes by BLAT and Sim4 [36] programs. Transcripts with more than 10% nucleotides unaligned or with less than 95% identity in the aligned regions were excluded. Transcripts were regarded as overlapping if their exons shared at least 1 bp, and a genic region was defined as a continuous genomic DNA region that covers all overlapped transcripts. Gene type was based on the most reliable transcript for this gene, and the order of transcript reliability is: RefSeq > mRNA > Ensembl > RefSeq XM > TWINSCAN. All ESTs were also mapped to the genomes in the same way. ESTs that overlap an identified genic region were included as transcripts of this gene without changing the genic region boundary. The UniGene ID was linked to the gene on the basis of its transcripts. For genes with Ensembl transcript ID, using the information from Ensembl's EnsMart, we marked the orthologous gene sets in our identified genes.
Known promoter collection
All promoter sequences in EPD (release 74) and DBTSS (release 2.0) were extracted. Promoter information and sequences were also retrieved from GenBank (dated 21 February 2003) using 'exon number = 1', 'prim_transcript', 'precursor_mRNA', and 'promoter' as keywords. The promoter regions identified by luciferase assay and ChIP of TAF250 and RNA polymerase II in the ENCODE regions were obtained from the UCSC genome browser and included. All sequences were mapped to the genomes by BLAT to obtain their locations of TSSs. Two identical TSSs were regarded as one unique TSS.
Whole-genome promoter prediction
With default thresholds (Pexon = 0.5, Ppromoter = 0.4, Pdonor = 0.4), original FirstEF was run on each chromosome of the three genomes without repeat masking, and the output was filtered by different methods described below. Predicted and known TSSs were linked to the closest gene if they were located either in the gene region or in the 20 kb upstream of the gene (if the gene has RefSeq mRNA, the distance was limited to 5 kb), and these promoters/TSSs were collected as 'gene-related promoters/TSSs'. Predicted promoters overlapping the 5' end of any transcript in a gene are defined as 'transcript-supported promoters'.
Conservation in control sets
Regions of 1,000 bp were randomly extracted from the genome of each species to make sequence pairs or triplets. Control set I included 1 million such sequence pairs for every two species, and 1 million triplets for the three species. We also selected genes from different species that are not orthologs, and randomly picked promoters belonging to these genes to make 1 million promoter pairs and 1 million triplets for control set II. One million high-GC content (>65%) pseudo promoter pairs were also selected. ClustalW was used to carry out multiple sequence global alignment for each pair or triplet with the conservation score defined as the ratio of identical base-pairs divided by 1,000.
Calculation of conservation for known promoters in orthologous genes
For genes with known TSSs, we extracted (-700, +300) bp regions with respect to the TSSs from the genomes as promoter sequences. We aligned each promoter of a gene in one species with each of the known promoters of its orthologous genes by ClustalW and calculated the conservation scores. The maximum score of all these promoter pairs or triplets was used to describe the conservation of this promoter.
CpG island relationship
We used the new CpG-island definition [24] to search genomes of the three species to collect CpG islands. A gene is considered as CpG-island-related only if there is at least one CpG island overlapping the region of (-2,000 to around +500) bp at its 5' end. A TSS/promoter is considered as CpG-island-related if at least one CpG island can overlap the region of (-2,000, +500) bp with respect to the TSS.
Post-clustering script for selecting promoters at least 500 bp apart
For all the gene-related promoters, we first ordered the known ones on the basis of the distance between TSSs defined in the promoters to the gene 5' end defined by mapped transcripts. The promoters with shorter distances were then selected, and the rest were compared to the selected ones. Only those that were separated by at least 500 bp from any of the selected promoters were kept. The same selection procedure was used for homologous promoters, transcript-supported promoters and other promoters. As a result of such post-clustering, all the selected promoters of a gene were separated by at least 500 bp.
Evaluation of promoter prediction by simulation
The test set comprised 8,949 genes with 13,313 known TSSs. To simulate the 'partial genes' that often exist in the databases, we truncated each identified genic region by 5 kb (or half of the gene length if the gene is shorter than 5 kb) at the 5' end, including the parts of cDNAs that extend into this region. On the basis of such new gene boundaries, we reselected all gene-related promoters from the predictions by original FirstEF (Method 0). Each promoter was compared with promoters of the orthologous genes (if available) by ClustalW to calculate the conservation score, and they were defined as the homologous promoters if the conservation score obeyed the pairwise or three-way cutoff rules.
De novo FirstEF (Method 1) selected the best-predicted promoters (with the highest probability in the promoter region) from the original FirstEF predictions in a 1,000 bp region. Method 2 compared RNAs or predicted transcripts with original FirstEF predictions that were gene-related to filter out predicted promoters that were neither located in the upstream of the genic region nor transcript-supported, and Method 3 first used Method 2 to select promoters, and then for a gene with homologous promoters, only those homologous promoters were selected as output for the gene (see also Figure 2). Post-clustering was used in promoter selection from the output of Method 1, Method 2 and Method 3 for tests in the 9,806 known TSSs of 500 bp apart, and such combined methods were called Method 1s, Method 2s, and Method 3s respectively. A predicted TSS was regarded as a 'correct TSS' if its distance to a known TSS was shorter than 500 bp, and this known TSS was regarded as 'correctly predicted' simultaneously. The sensitivity of prediction (Sn) was defined as the ratio between the numbers of correctly predicted and known TSSs used in the validation. Specificity (Sp) was the number of correct TSSs divided by the total number of predicted promoters.
Cold Spring Harbor Laboratory mammalian promoter database construction
We first collected all gene-related TSSs in human, mouse and rat genomes. For genes with RefSeq mRNAs but no known or predicted promoters, the 5' ends of the RefSeq sequences were considered as their TSSs and called Refseq ND TSS. They were also defined as transcript-supported. For two adjacent divergent genes with their 5' ends less than 2 kb apart, we defined their 5' gene boundaries as 'bidirectional TSSs' if no other type of TSS could be found in the intergenic region between them. All promoters of the orthologous genes were aligned by ClustalW to find homologous promoters in the same way as done in the evaluation step. Method 3s was used to select the final promoter set. Known promoters filtered out by the post-clustering script were also included in the database after the selection to make the known promoter data as complete as possible. All these selected promoters were stored in a MySQL database. Gene features contained in the database include genome location, overlapping transcripts, UniGene ID, LocusLink ID, and gene name if available. Promoter features included TSS location, first donor and acceptor sites if available, corresponding gene, overlapped transcript for a transcript-supported promoter, and promoter type. Promoter type refers to the source type, which was also used to represent their reliability in the order of: known promoters (EPD, DBTSS, GenBank annotation, promoters identified by luciferase assay or ChIP), RefSeq END promoters, promoters of divergent genes (bidirectional TSS), transcript-supported promoters, as well as other gene-related promoters that were predicted. Homologous promoters were also marked. In addition to gene-related promoters, all other predicted promoters located in the intergenic regions were included in the database. They were regarded as predicted novel promoters and were of the lowest reliability.
Promoter set for the Affymetrix microarray
For each probe set in the gene chip, its gene index and/or chromosome location information were used to find the corresponding gene in our promoter database. The most reliable promoter of this gene was reported for this probe set. If no gene could be assigned to a probe set, the closest predicted novel promoter in its upstream region was taken if the distance between the promoter and probe set was less than 20 kb.
Data availability
All 8,949 human genes and 13,313 human known promoters used in the test can be downloaded from our FTP site at [37], the promoter set for Affymetrix array HG-U133A is in [38], the promoter set of all RefSeq genes is in [39], all known promoters in CSHLmpd can be downloaded from [40].
Acknowledgements
We thank Lisa Stubbs for providing experimental testing results before publication. We thank Ewan Birney for providing PromoterWise software, Lincoln Stein for providing Gbrowse. This work is supported by NIH grants HG01696, GM60513, CA88351, and HG002600.
Figures and Tables
Figure 1 Distribution of conservation scores in promoter alignments. (a) Pairwise promoter alignments of human-rodent and mouse-rat non-orthologous genes (control set II) with different promoter GC content. (b) Pairwise promoter alignments of most conserved promoter pairs and randomly selected 1 kb sequence pairs (control set I). (c) Alignments of mouse-rat and human-rodent homologous promoter pairs. (d) Three-way promoter alignments of homologous promoter triplets and sequence triplets from control set II.
Figure 2 Flowchart of the pipeline to construct the promoter database. Ovals indicate data and rectangles the method. The ovals shaded gray represent the data stored in CSHLmpd.
Figure 3 Sensitivity and specificity of promoter prediction for CpG-island related and non-CpG-island related promoters in different gene sets. (a) 5,893 human genes with homologous rodent promoters. (b) All 8,949 human genes in the test set. The definition of different methods is described in the text and in Materials and methods.
Figure 4 Screen shots of the CSHLmpd user interface. (a) Gbrowse for genome-wide gene and promoter display. (b) Homologous promoter search and analysis.
Table 1 Number of genes and transcripts of different types in the three mammalian genomes
Type HSPD MMPD RNPD
Gene* Transcript† Gene Transcript Gene Transcript
RefSeq 17,354 22,425 16,329 17,438 6,400 6,807
mRNA 8,846 106,279 2,641 40,552 1,967 11,116
Ensembl 3,160 33,653 6,601 31,022 14,276 27,989
RefSeq_XM 2,400 6,105 4,974 5,829 3,021 15,023
TwinScan 3,189 25,633 4,528 25,583 5,015 25,499
EST 0 4,488,530 0 3,254,853 0 477,321
Total 34,949 4,682,625 35,073 3,375,277 30,679 563,755
*Number of genes in non-overlapping gene types. †Number of all transcripts of this type.
Table 2 Sensitivity and specificity of promoter prediction with different methods
Sn Sp
(a) 13,313 unique TSSs in 8,949 human genes
Method 0* 72% 46%
Method 1† 67% 46%
Method 2‡ 70% 57%
Method 3§ 69% 66%
(b) 9,806 TSSs of 500 bp apart in 8,949 human genes
Method 1 + script¶ 64% 33%
Method 2 + script 67% 44%
Method 3 + script 66% 60%
(c) 6,356 TSSs of 500 bp apart in 5,893 human genes with homologous promoters
Method 1 + script 80% 37%
Method 2 + script 84% 46%
Method 3 + script 82% 69%
*Method 0 used original FirstEF alone to predict promoters in the upstream and genic regions of these genes. †Method 1 used de novo FirstEF to predict promoters in the upstream and genic regions of these genes. ‡Method 2 compared mRNAs or predicted transcripts with original FirstEF predictions to filter out promoters that were neither located in the upstream of the gene region nor overlapping with the 5'-end of any transcripts of this gene. §Method 3 tried to first find the promoters in one gene that have homologous rodent promoters. If no such promoters were found, it used Method 2 to select promoters for this gene. ¶script, a post-clustering script to select representative TSSs from the output of each method described above that were at least 500 bp apart (see Materials and methods for details).
Table 3 Statistics of promoters and genes in CSHLmpd
HSPD MMPD RNPD
Total genes 34,949 35,073 30,679
Known genes (RefSeq and mRNA) 26,200 18,970 8,367
Canonical genes (RefSeq, mRNA, and Ensembl) 29,360 25,571 22,643
Genes with promoters 26,820 25,592 21,125
Genes with homologous promoters 13,432 14,626 12,302
Predicted genes with promoters 4,340 7,343 13,230
Total promoters* 55,513 46,207 37,479
Known promoters 14,314 841 943
FirstEF predicted promoters 39,233 34,994 34,227
Transcript-supported FirstEF predicted promoters 19,331 16,913 11,798
RefSeq END promoters 1,828 2,988 2,270
Bidirectional gene promoters 138 84 39
Core promoters 26,820 25,592 21,125
Homologous promoters 21,594 21,501 17,257
Homologous known promoters 10,561 6,854 817
CpG-island related RefSeq genes 12,259 (71%) 9,831 (60%) 2,987 (47%)
CpG-island related other mRNA genes 2,679 (30%) 993 (38%) 907 (46%)
CpG-island related canonical genes 15,707 (54%) 12,293 (48%) 8,420 (37%)
CpG-island related promoters 37,572 (68%) 24,726 (54%) 20,826 (56%)
CpG-island related known promoters 10,332 (72%) 5,115 (63%) 444 (47%)
CpG-island related predicted promoters 26,936 (69%) 19,363 (55%) 20,207 (59%)
CpG-island related RefSeq END promoters 187 (10%) 201 (7%) 153 (7%)
CpG-island related bidirectional gene promoters 53 (38%) 47 (56%) 22 (56%)
CpG-island related homologous promoters 13,974 (82%) 11,867 (76%) 9,372 (80%)
*Predicted promoters were separated with other predicted or known promoters by at least 500 bp.
==== Refs
Cavin PR Junier T Bucher P The Eukaryotic Promoter Database EPD. Nucleic Acids Res 1998 26 353 357 9399872 10.1093/nar/26.1.353
Suzuki Y Yamashita R Nakai K Sugano S DBTSS: DataBase of human Transcriptional Start Sites and full-length cDNAs. Nucleic Acids Res 2002 30 328 331 11752328 10.1093/nar/30.1.328
Bajic VB Tan SL Suzuki Y Sugano S Promoter prediction analysis on the whole human genome. Nat Biotechnol 2004 22 1467 1473 15529174 10.1038/nbt1032
Scherf M Klingenhoff A Frech K Quandt K Schneider R Grote K Frisch M Gailus-Durner V Seidel A Brack-Werner R Werner T First pass annotation of promoters on human chromosome 22. Genome Res 2001 11 333 340 11230158 10.1101/gr.154601
Diehn M Sherlock G Binkley G Jin H Matese JC Hernandez-Boussard T Rees CA Cherry JM Botstein D Brown PO Alizadeh AA SOURCE: a unified genomic resource of functional annotations, ontologies, and gene expression data. Nucleic Acids Res 2003 31 219 223 12519986 10.1093/nar/gkg014
Halees AS Weng Z PromoSer: improvements to the algorithm, visualization and accessibility. Nucleic Acids Res 2004 32 W191 W194 15215378
Coleman SL Buckland PR Hoogendoorn B Guy C Smith K O'Donovan MC Experimental analysis of the annotation of promoters in the public database. Hum Mol Genet 2002 11 1817 1821 12140184 10.1093/hmg/11.16.1817
Trinklein ND Aldred SJ Saldanha AJ Myers RM Identification and functional analysis of human transcriptional promoters. Genome Res 2003 13 308 312 12566409 10.1101/gr.794803
Okazaki Y Furuno M Kasukawa T Adachi J Bono H Kondo S Nikaido I Osato N Saito R Suzuki H Analysis of the mouse transcriptome based on functional annotation of 60,770 full-length cDNAs. Nature 2002 420 563 573 12466851 10.1038/nature01266
Ota T Suzuki Y Nishikawa T Otsuki T Sugiyama T Irie R Wakamatsu A Hayashi K Sato H Nagai K Complete sequencing and characterization of 21,243 full-length human cDNAs. Nat Genet 2004 36 40 45 14702039 10.1038/ng1285
Gerhard DS Wagner L Feingold EA Shenmen CM Grouse LH Schuler G Klein SL Old S Rasooly R Good P The status, quality, and expansion of the NIH full-length cDNA project: the Mammalian Gene Collection (MGC). Genome Res 2004 14 2121 2127 15489334 10.1101/gr.2596504
Davuluri R Grosse I Zhang MQ Computational identification of promoters and first exons in the human genome. Nat Genet 2001 29 412 417 11726928 10.1038/ng780
Liu R States DJ Consensus promoter identification in the human genome utilizing expressed gene markers and gene modeling. Genome Res 2002 12 462 469 11875035 10.1101/gr.198002
Korf I Flicek P Duan D Brent MR Integrating genomic homology into gene structure prediction. Bioinformatics 2001 17 Suppl 1 S140 S148 11473003
Parra G Agarwal P Abril JF Wiehe T Fickett JW Guigó R Comparative gene prediction in human and mouse. Genome Res 2003 13 108 117 12529313 10.1101/gr.871403
Solovyev VV Shahmuradov IA PromH: Promoters identification using orthologous genomic sequences. Nucleic Acids Res 2003 31 3540 3545 12824362 10.1093/nar/gkg525
Pruitt KD Maglott DR RefSeq and LocusLink: NCBI gene-centered resources. Nucleic Acids Res 2001 29 137 140 11125071 10.1093/nar/29.1.137
Brooksbank C Camon E Harris MA Magrane M Martin MJ Mulder N O'Donovan C Parkinson H Tuli MA Apweiler R The European Bioinformatics Institute's data resources. Nucleic Acids Res 2003 31 43 50 12519944 10.1093/nar/gkg066
Kasprzyk A Keefe D Smedley D London D Spooner W Melsopp C Hammond M Rocca-Serra P Cox T Birney E EnsMart: a generic system for fast and flexible access to biological data. Genome Res 2004 14 160 169 14707178 10.1101/gr.1645104
Benson DA Karsch-Mizrachi I Lipman DJ Ostell J Wheeler DL GenBank. Nucleic Acids Res 2003 31 23 27 12519940 10.1093/nar/gkg057
Kent WJ BLAT - the BLAST-like alignment tool. Genome Res 2002 12 656 664 11932250 10.1101/gr.229202. Article published online before March 2002
Higgins DG Thompson JD Gibson TJ Using CLUSTAL for multiple sequence alignments. Methods Enzymol 1996 266 383 402 8743695
FirstEF
Takai D Jones PA Comprehensive analysis of CpG islands in human chromosomes 21 and 22. Proc Natl Acad Sci USA 2002 99 3740 3745 11891299 10.1073/pnas.052410099
Yeh RF Lim LP Burge CB Computational inference of homologous gene structures in the human genome. Genome Res 2001 11 803 816 11337476 10.1101/gr.175701
CSHL Mammalian Promoter Database (CSHLmpd)
Stein LD Mungall C Shu S Caudy M Mangone M Day A Nickerson E Stajich JE Harris TW Arva A Lewis S The generic genome browser: a building block for a model organism system database. Genome Res 2002 12 1599 1610 12368253 10.1101/gr.403602
Wheeler DL Church DM Federhen S Lash AE Madden TL Pontius JU Schuler GD Schriml LM Sequeira E Tatusova TA Wagner L Database resources of the National Center for Biotechnology. Nucleic Acids Res 2003 31 28 33 12519941 10.1093/nar/gkg033
Brudno M Do CB Cooper GM Kim MF Davydov E Green ED Sidow A Batzoglou S NISC Comparative Sequencing Program LAGAN and Multi-LAGAN: efficient tools for large-scale multiple alignment of genomic DNA. Genome Res 2003 13 721 731 12654723 10.1101/gr.926603
Zhao F Xuan Z Liu L Zhang MQ TRED: a Transcription Regulatory Element Database and a platform for in silico gene regulation studies. Nucleic Acid Res 2005 33 D103 D107 15608156 10.1093/nar/gki004
Liu G Loraine AE Shigeta R Cline M Cheng J Valmeekam V Sun S Kulp D Siani-Rose MA NetAffx: Affymetrix probesets and annotations. Nucleic Acids Res 2003 31 82 86 12519953 10.1093/nar/gkg121
Promoter sets
Dike S Balija VS Nascimento LU Xuan Z Ou J Zutavern T Palmer LE Hannon G Zhang MQ McCombie WR The mouse genome: experimental examination of gene predictions and transcriptional start sites. Genome Res 2004 14 2424 2429 15574821 10.1101/gr.3158304
Suzuki Y Yamashita R Shirota M Sakakibara Y Chiba J Mizushima-Sugano J Nakai K Sugano S Sequence comparison of human and mouse genes reveals a homologous block structure in the promoter regions. Genome Res 2004 14 1711 1718 15342556 10.1101/gr.2435604
UCSC Genome browser
Florea L Hartzell G Zhang Z Rubin GM Miller W A computer program for aligning a cDNA sequence with a genomic DNA sequence. Genome Res 1998 8 967 974 9750195
Human genes and promoters
Promoter set for Affymetrix array HG-U133A
Promoter set of all RefSeq genes
All known promoters in CSHLmpd
|
16086854
|
PMC1273639
|
CC BY
|
2021-01-04 16:05:41
|
no
|
Genome Biol. 2005 Aug 1; 6(8):R72
|
utf-8
|
Genome Biol
| 2,005 |
10.1186/gb-2005-6-8-r72
|
oa_comm
|
==== Front
BMC Health Serv ResBMC Health Services Research1472-6963BioMed Central London 1472-6963-5-361588514910.1186/1472-6963-5-36Research ArticleUsing autoregressive integrated moving average (ARIMA) models to predict and monitor the number of beds occupied during a SARS outbreak in a tertiary hospital in Singapore Earnest Arul [email protected] Mark I [email protected] Donald [email protected] Leo Yee [email protected] Department of Clinical Epidemiology, Tan Tock Seng Hospital, Singapore2 Communicable Disease Centre, Singapore2005 11 5 2005 5 36 36 15 10 2004 11 5 2005 Copyright © 2005 Earnest et al; licensee BioMed Central Ltd.2005Earnest et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The main objective of this study is to apply autoregressive integrated moving average (ARIMA) models to make real-time predictions on the number of beds occupied in Tan Tock Seng Hospital, during the recent SARS outbreak.
Methods
This is a retrospective study design. Hospital admission and occupancy data for isolation beds was collected from Tan Tock Seng hospital for the period 14th March 2003 to 31st May 2003. The main outcome measure was daily number of isolation beds occupied by SARS patients. Among the covariates considered were daily number of people screened, daily number of people admitted (including observation, suspect and probable cases) and days from the most recent significant event discovery. We utilized the following strategy for the analysis. Firstly, we split the outbreak data into two. Data from 14th March to 21st April 2003 was used for model development. We used structural ARIMA models in an attempt to model the number of beds occupied. Estimation is via the maximum likelihood method using the Kalman filter. For the ARIMA model parameters, we considered the simplest parsimonious lowest order model.
Results
We found that the ARIMA (1,0,3) model was able to describe and predict the number of beds occupied during the SARS outbreak well. The mean absolute percentage error (MAPE) for the training set and validation set were 5.7% and 8.6% respectively, which we found was reasonable for use in the hospital setting. Furthermore, the model also provided three-day forecasts of the number of beds required. Total number of admissions and probable cases admitted on the previous day were also found to be independent prognostic factors of bed occupancy.
Conclusion
ARIMA models provide useful tools for administrators and clinicians in planning for real-time bed capacity during an outbreak of an infectious disease such as SARS. The model could well be used in planning for bed-capacity during outbreaks of other infectious diseases as well.
==== Body
Background
Early isolation of infectious cases has been shown to be a key component for the successful management of SARS outbreaks[1]. Due to the potential for nosocomial transmission[2-6], the imperfect ability of clinical criteria to distinguish cases of SARS at presentation[7], the possibility of atypical presentations[8,9], and the lack of a sensitive diagnostic test in early disease[10], front-line clinicians need to err on the side of caution when admitting cases, and isolate patients until SARS can be clinically and virologically ruled out. As a result, the number of admissions and isolation beds required during management of SARS outbreaks can be expected to significantly exceed that used for actual SARS cases.
While hospitals are generally built with a fixed ratio of isolation to general ward beds, surge capacity for isolation beds can be met by conversion of single room wards, decanting of existing patients with lesser indications for isolation, and activation of isolation facilities at alternative sites and institutions. However, these processes require time, and the ability to forecast requirements is hence a critical component of efficient outbreak management.
During SARS outbreak in Singapore from 1 Mar to 31 May 2003, the Communicable Disease Centre (CDC) was the initial designated facility for the screening management of all SARS cases, beginning on 14 Mar 2003, two days after the WHO alert was sounded on 12 Mar[11]. To accommodate the surge in cases, the parent hospital of the CDC, Tan Tock Seng Hospital (TTSH) became designated as the central facility for management of all SARS cases in Singapore from 22 Mar 2003[12]. As a result of the above policy, 231 of 238 SARS cases diagnosed during the Singapore outbreak were admitted to TTSH. TTSH bed utilization patterns here hence reflect national level requirements for outbreak management. Various papers have described qualitative aspects of hospital management during SARS[13-15], but none have provided quantitative tools for predicting requirements for isolation beds. Yet such quantitative models would be of great utility to both hospital administrators and national level planners during outbreak management.
The main objective of this study is to apply autoregressive integrated moving average (ARIMA) models to make real-time predictions on the number of beds occupied in TTSH during the SARS outbreak, starting from 14 Mar 2003, when the CDC was activated, to 31 May 2003 when Singapore was declared SARS free.
Methods
This was a retrospective study design. Hospital admission and occupancy data for isolation beds was collected from Tan Tock Seng hospital for the period 14th March 2003 to 31st May 2003. The main outcome measure was daily number of isolation beds occupied by SARS patients, including those fulfilling WHO criteria for suspect and probable SARS[16], as well as those admitted not fulfilling WHO case definitions but admitted to isolation rooms for observation. Among the covariates considered were daily number of people screened, daily number of people admitted (including observation, suspect and probable cases) and days from the most recent significant event discovery. Key events considered were as follows:
1. 14th Mar: discovery of the TTSH outbreak
2. 22nd Mar: press release that TTSH would dedicated to SARS management
3. 4th Apr: discovery of an outbreak at Singapore General Hospital (SGH)
4. 11th Apr: discovery of an outbreak at National University Hospital (NUH)
5. 20th Apr: discovery and press release on an outbreak at Pasir Panjang Wholesale Market (PPWM)
6. 13th May: discovery of a cluster of febrile staff and patients at the Institute of Mental Health (IMH)
Details on the above can be found in the chronology of press releases on SARS events in Singapore[17]. Events 1–5 all involved probable SARS cases, whereas event 6 proved to be a false alarm[18].
We utilized the following strategy for the analysis. Firstly, we split the outbreak data into two. Data from 14th March to 21st April 2003 was used for model development. We used structural ARIMA models in an attempt to model the number of beds occupied[19]. Estimation is via the maximum likelihood method using the Kalman filter[20]. For the ARIMA model parameters, we considered the simplest parsimonious lowest order model.
We computed various permutations of the order of correlation (AR), order of integration (I) and order of moving average (MA), and chose the optimal combination of parameters using the mean square error. The correlogram and partial correlogram graphs were also used to help in deciding the order of moving average (MA) and auto-regressive (AR) terms to include in the model. To ensure the model was robust to symmetric nonnormality in the disturbances, including heteroskedasticity, we computed Huber/White/sandwich estimator of variance for the coefficient estimates[21]. Before modeling the bed occupancy, we examined whether the series was stationary. In the event of non-stationarity, we opted to set an a-prior value of 1 for starting the Kalman recursions[22].
We used the likelihood ratio test to determine if inclusion of other covariates helped improve the fit of the model. Based on the final model selected, we assessed the out-sample validity of the model, by applying the model to predict the number of beds occupied for the remaining period of the outbreak (i.e. 22nd April 2003 to 31st May 2003). In addition, we also made three-day forecasts for selected periods during the outbreak, starting from day 4 of the outbreak. We used the mean absolute percentage error (MAPE) to measure and quantify the quality of fit. A lower MAPE value will indicate a better fit of the data. All tests were conducted at the 5% level of significance, and data analysis was performed in Stata V7.0 (Stata Corporation, College Station, TX, USA).
Results
From 14th March 2003 to 31st May 2003, the median daily number of beds occupied was 134 (IQR: 105–193). The range was 15 to 238 beds. The number of beds occupied reached it's peak on the 24th and 28th of April 2003, with a total of 238 beds. For the final ARIMA model, we found that the ARIMA (1,0,3) model was the most suitable, with an auto-regression term of 1 and a moving average term of 3. The correlogram indicated that there was a significant autocorrelation out to about 3 lags, and this autocorrelation decayed slowly over time (figure 1). The partial auto correlation function (PACF) plot suggested that the only highly significant partial autocorrelation occurred at one lag (figure 2). We found that the AR(1) coefficient of 1.02 and MA(3) coefficient of -0.95 were significant (p < 0.05). The likelihood ratio test indicated that the total number of admissions on the previous day and number of probable cases admitted on the previous day were significant predictors, and these variables were thus included in the final model. Furthermore, the estimated variance of the white-noise disturbance was found to be 4.47 (see table 1). Days from most recent significant event discovery and number of patients screened were not found to be significant predictors of daily number of isolation beds occupied.
Figure 1 Correlogram of total beds occupied
Figure 2 Partial correlogram of total beds occupied
Table 1 Parameters for the final ARIMA model
Variable Coefficient 95 % CI p-value
Constant 17.03 6.43, 27.62 0.002
Previous day's total admissions -0.37 -0.63, -0.11 0.006
Previous day's probable case admissions 1.17 0.37, 1.97 0.004
ARMA Parameters
AR (1) 1.02 0.98, 1.06 <0.001
MA (1) -1.34 -2.94, 0.27 0.102
MA (2) -0.96 -1.80, -0.12 0.026
MA (3) -0.95 -1.52, -0.38 0.001
Sigma 4.47 1.40, 7.53 0.004
As we can see from figure 3, the predictions from the ARIMA model performed reasonably well, both for the training and validation data. The MAPE for the training set and validation set were 5.7% and 8.6% respectively. This translated to an error rate of ± 7 beds and ± 13 beds respectively. We have also provided the model parameters and their corresponding MAPE values for some of the alternative ARIMA models that we had considered (table 2).
Figure 3 Admissions, predicted and actual number of beds occupied
Table 2 Comparison of various selected ARIMA models
Model Training Set MAPE Validation Set MAPE
ARIMA (1,0,0) 6.0% 9.0%
ARIMA (1,0,1) 6.2% 9.2%
ARIMA (1,0,2) 5.9% 8.4%
ARIMA (1,0,3) 5.7% 8.6%
ARIMA (1,0,4) 5.3% 13.1%
ARIMA (1,1,3) 5.3% 18.7%
ARIMA (0,0,2) 9.8% 32.8%
ARIMA (0,0,3) 9.5% 16.4%
For three-day predictions, we found that the model fared reasonably well (see table 3). For day 4 to day 6 of the outbreak, the error rate was 6%. For day 7 to day 9, the rate was 10%, day 10 to day 12, 7% and finally, for day 13 to day 15, it was 9%. Although the MAPE values were within reasonable levels, we note that generally, the model under-predicts in the early stage of the outbreak, and over-predicts in the later stage of the outbreak (tables 3 and 4).
Table 3 Forecast of bed occupancy in the initial stage of the outbreak
Day of outbreak 3rd Day 6th day 9th Day 12th Day MAPE
Three-day forecast Actual number of beds
Day 4 36 37 6
Day 5 44 42
Day 6 51 46
Day 7 70 60 10
Day 8 81 72
Day 9 90 87
Day 10 92 99 7
Day 11 111 108
Day 12 125 111
Day 13 132 116 9
Day 14 131 122
Day 15 115 124
Table 4 Forecast of bed occupancy in the late stage of the outbreak
Day of outbreak 67th Day 70th day 73rd Day 76th Day MAPE
Three-day forecast Actual number of beds
Day 68 108 97 22
Day 69 132 106
Day 70 153 96
Day 71 134 129 16
Day 72 119 134
Day 73 105 136
Day 74 100 112 17
Day 75 101 119
Day 76 96 117
Day 77 79 103 35
Day 78 79 106
Day 79 75 105
Discussion
To the best of our knowledge, this is the first study to suggest the application of a known statistical method such as the ARIMA model, to predict and monitor the utilization of hospital isolation beds during the recent SARS outbreak in Singapore, for which Tan Tock Seng Hospital was the designated hospital for all patients presenting with SARS-like symptoms and exposures.
ARIMA models have traditionally found application in the financial sector. There has been limited literature on their use in healthcare; recent examples include their use in assessment of seasonal variation in selected medical conditions[23], and as a surveillance tool for outbreak detection[24]. There has been some research indicating that time series modeling may be more appropriate than the simple trend fitting approach, which suffers from model specification error[25]. ARIMA models have been used to forecast attendance at accident and emergency departments in the United Kingdom. Particularly, researchers have shown that the forecasting methodology can be improved by incorporating the ARIMA method[26]. Here, we show that the ARIMA model can be used over the much shorter time-frame of a single outbreak to forecast bed-utilization. The three-day forecasts from the model are fairly reasonable. The MAPE is low, allowing planners to confidently decide, with sufficient lead-time, on the need to open new isolation wards, each of which, in our setting, holds between 10 to 20 patients.
However, the model has its limitations. Firstly, unmeasured confounders could have affected the results of this study, although we have accounted for measured confounders by incorporating significant covariates into the final model. Secondly, the most significant covariate was the number of probable SARS cases admitted. This is not surprising, as probable SARS cases stayed longer in isolation facilities compared to cases which turned out not to be SARS (unpublished data). In this analysis, we used the final classification for each case after a variable period of observation and investigation, and not the admission classification.
This approach was chosen, as time required to confirm cases will be likely be shortened in any future outbreaks, in view of various advancements for SARS diagnostics[27,28]. However, it is still uncertain what proportion of SARS cases can be accurately classified on admission, and this may affect model performance. Another point to note is that the model generally under-predicts the number of beds occupied in the early stage of the outbreak and over-predicts at the later stage of the outbreak for three-day forecasts. One possible explanation could be that the model was derived from the first half of the data and applied to both the first and second halves of the outbreak. It is important for administrators to take this into account during bed planning, perhaps by allowing for appropriate buffer beds. Lastly, the ARIMA model parameters may differ under different practice protocols within different outbreak settings as well as between different SARS afflicted countries. It would be useful to calibrate the model using individual country level data. We would recommend that, in an actual outbreak, real-time calibration be performed, with additional data available on each day fed-back into the model to improve its predictive ability.
The application of ARIMA models in bed utilization is not only useful for an outbreak of SARS and emerging infectious diseases, but also for projecting resource requirements in bioterrorism events. It has been recognized by others that resource requirements will not include just isolation beds, but also outpatient resources[29], pharmaceuticals[30], as well as intensive care facilities[31]; few articles, however, have proposed any models for forecasting such requirements. The challenge, therefore, lies in the collection of timely surveillance and resource utilization data for this specific purpose, peace-time exploration of the most appropriate methods of analysis, and real-time validation and application in the event of an outbreak.
Conclusion
The ARIMA model that we developed for modeling the number of beds occupied during the SARS outbreak performed reasonably well, with a MAPE of 5.7% for the training set, and 8.6% for the validation set. In addition, we found that three-day forecasts provided a reasonable prediction of the number of beds required during the outbreak
ARIMA models provide useful tools for administrators and clinicians in planning the use of isolation beds during an outbreak of an infectious disease such as SARS. The model could be used in planning for bed-capacity during outbreaks of other infectious diseases, as well as predicting requirements for other critical resources.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
DN conceived the study and contributed to the study design, analysis and interpretation. AE contributed to the statistical analysis, interpretation and writing of the manuscript. MIC contributed to the statistical analysis, interpretation and writing of the manuscript. LYS contributed to the interpretation and writing of the manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
We would like to thank all the doctors, nurses, respiratory therapists, physiotherapists, radiographers, medical social workers, healthcare attendants and clerks who worked in Tan Tock Seng Hospital, Singapore during the SARS outbreak in 2003.
==== Refs
Lipsitch M Cohen T Cooper B Robins JM Ma S James L Gopalakrishna G Chew SK Tan CC Samore MH Fisman D Murray M Transmission dynamics and control of severe acute respiratory syndrome Science 2003 300 1966 70 12766207 10.1126/science.1086616
Lee N Hui D Wu A Chan P Cameron P Joynt GM Ahuja A Yung MY Leung CB To KF Lui SF Szeto CC Chung S Sung JJ A Major Outbreak of Severe Acute Respiratory Syndrome in Hong Kong N Engl J Med 2003 348 1986 94 12682352 10.1056/NEJMoa030685
Spurgeon D Canada reports more than 300 suspected cases of SARS BMJ 2003 326 897 12714456 10.1136/bmj.326.7395.897/a
Poutanen SM Low DE Henry B Finkelstein S Rose D Green K Tellier R Draker R Adachi D Ayers M Chan AK Skowronski DM Salit I Simor AE Slutsky AS Doyle PW Krajden M Petric M Brunham RC McGeer AJ National Microbiology Laboratory, Canada; Canadian Severe Acute Respiratory Syndrome Study Team Identification of Severe Acute Respiratory Syndrome in Canada N Engl J Med 2003 348 1995 2005 12671061 10.1056/NEJMoa030634
Dwosh HA Hong HH Austgarden D Herman S Schabas R Identification and containment of an outbreak of SARS in a community hospital CMAJ 2003 168 1415 20 12771070
Varia M Wilson S Sarwal S McGeer A Gournis E Galanis E Henry B Investigation of a nosocomial outbreak of severe acute respiratory syndrome (SARS) in Toronto, Canada CMAJ 2003 169 285 92 12925421
Rainer TH Cameron PA Smit D Ong KL Hung AN Nin DC Ahuja AT Si LC Sung JJ Evaluation of WHO criteria for identifying patients with severe acute respiratory syndrome out of hospital: prospective observational study BMJ 2003 326 1354 8 12816820 10.1136/bmj.326.7403.1354
Tan TT Tan BH Kurup A Oon LL Heng D Thoe SY Bai XL Chan KP Ling AE Atypical SARS and Escherichia coli bacteremia Emerg Infect Dis 2004 10 349 52 15030711
Tee AK Oh HM Lien CT Narendran K Heng BH Ling AE Atypical SARS in geriatric patient Emerg Infect Dis [serial online] 2004 Feb [cited 28 Feb 2004].
Cheng PK Wong DA Tong LK Ip SM Lo AC Lau CS Yeung EY Lim WW Viral shedding patterns of coronavirus in patients with probable severe acute respiratory syndrome Lancet 2004 363 1699 700 15158632 10.1016/S0140-6736(04)16255-7
World Health Organization, Communicable Disease Surveillance and Response Update 95 – SARS: Chronology of a serial killer 4 Jul 2003. Accessed 22 Oct 2003
Ministry of Health, Singapore MOH SARS Press Releases 22 Mar 2003: Enhanced Precautionary Measures To Break SARS Transmission Accessed on 11 Sep 2003.
McDonald LC Simor AE Su IJ Maloney S Ofner M Chen KT Lando JF McGeer A Lee ML Jernigan DB SARS in healthcare facilities, Toronto and Taiwan Emerg Infect Dis 2004 10 777 81 15200808
Loutfy MR Wallington T Rutledge T Mederski B Rose K Kwolek S McRitchie D Ali A Wolff B White D Glassman E Ofner M Low DE Berger L McGeer A Wong T Baron D Berall G Hospital preparedness and SARS Emerg Infect Dis 2004 10 771 6 15200807
Gopalakrishna G Choo P Leo YS Tay BK Lim YT Khan AS Tan CC SARS transmission and hospital containment Emerg Infect Dis 2004 10 395 400 15109403
World Health Organization, Communicable Disease Surveillance and Response Case definitions for surveillance of Severe Acute Respiratory Syndrome (SARS) Revised 1 May 2003. Accessed on 22 Oct 2003.
Ministry of Health, Singapore MOH SARS Press Releases Accessed on 25th Nov 2003.
Chong SA Subramaniam M Chua HC Lee CE SARS or not SARS: outbreak of fever in a state mental institute in Singapore Can J Psychiatry 2004 49 216 7 15101508
Box GEP Jenkins GM Reinsel GC Time Series Analysis: Forecasting and Control 1994 3 Englewood Cliffs, NJ: Prentice-Hall
Harvey AC Forecasting, structural time series models and the Kalman filter 1989 Cambridge: Cambridge University Press
Hamilton JD Time Series Analysis 1994 Princeton: Princeton University Press
Kalman RE A new approach to linear filtering and prediction problems Journal of Basic Engineering, Transactions of the ASME Series D 1960 82 35 45
Moineddin R Upshur RE Crighton E Mamdani M Autoregression as a means of assessing the strength of seasonality in a time series Popul Health Metr 2003 1 10 14675482 10.1186/1478-7954-1-10
Reis BY Mandl KD Time series modeling for syndromic surveillance BMC Med Inform Decis Mak 2003 3 2 12542838 10.1186/1472-6947-3-2
Farmer RD Emami J Models for forecasting hospital bed requirements in the acute sector J Epidemiol Community Health 1990 44 307 12 2277253
Milner PC Ten-year follow-up of ARIMA forecasts of attendances at accident and emergency departments in the Trent region Stat Med 1997 16 2117 25 9308136 10.1002/(SICI)1097-0258(19970930)16:18<2117::AID-SIM649>3.0.CO;2-E
Mahony JB Petrich A Louie L Song X Chong S Smieja M Chernesky M Loeb M Richardson S Ontario Laboratory Working Group for the Rapid Diagnosis of Emerging Infections Performance and Cost evaluation of one commercial and six in-house conventional and real-time reverse transcription-pcr assays for detection of severe acute respiratory syndrome coronavirus J Clin Microbiol 2004 42 1471 6 15070991 10.1128/JCM.42.4.1471-1476.2004
Guan M Chan KH Peiris JS Kwan SW Lam SY Pang CM Chu KW Chan KM Chen HY Phuah EB Wong CJ Evaluation and validation of an enzyme-linked immunosorbent assay and an immunochromatographic test for serological diagnosis of severe acute respiratory syndrome Clin Diagn Lab Immunol 2004 11 699 703 15242944 10.1128/CDLI.11.4.699-703.2004
Hupert N Mushlin AI Callahan MA Modeling the public health response to bioterrorism: using discrete event simulation to design antibiotic distribution centers Med Decis Making 2002 22 S17 25 12369227 10.1177/027298902237709
Cohen V Organization of a health-system pharmacy team to respond to episodes of terrorism Am J Health Syst Pharm 60 1257 63 2003 Jun 15 12845922
White SM Chemical and biological weapons. Implications for anaesthesia and intensive care Br J Anaesth 2002 89 306 24 12378672 10.1093/bja/aef168
|
15885149
|
PMC1274243
|
CC BY
|
2021-01-04 16:31:52
|
no
|
BMC Health Serv Res. 2005 May 11; 5:36
|
utf-8
|
BMC Health Serv Res
| 2,005 |
10.1186/1472-6963-5-36
|
oa_comm
|
==== Front
BMC Med EducBMC Medical Education1472-6920BioMed Central London 1472-6920-5-151588514110.1186/1472-6920-5-15Research ArticlePlay dough as an educational tool for visualization of complicated cerebral aneurysm anatomy Eftekhar Behzad [email protected] Mohammad [email protected] Ebrahim [email protected] Arman Rakan [email protected] Department of Neurosurgery, Sina Hospital, Tehran University, Iran2005 10 5 2005 5 15 15 19 12 2004 10 5 2005 Copyright © 2005 Eftekhar et al; licensee BioMed Central Ltd.2005Eftekhar et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Imagination of the three-dimensional (3D) structure of cerebral vascular lesions using two-dimensional (2D) angiograms is one of the skills that neurosurgical residents should achieve during their training. Although ongoing progress in computer software and digital imaging systems has facilitated viewing and interpretation of cerebral angiograms enormously, these facilities are not always available.
Methods
We have presented the use of play dough as an adjunct to the teaching armamentarium for training in visualization of cerebral aneurysms in some cases.
Results
The advantages of play dough are low cost, availability and simplicity of use, being more efficient and realistic in training the less experienced resident in comparison with the simple drawings and even angiographic views from different angles without the need for computers and similar equipment. The disadvantages include the psychological resistance of residents to the use of something in surgical training that usually is considered to be a toy, and not being as clean as drawings or computerized images.
Conclusion
Although technology and computerized software using the patients' own imaging data seems likely to become more advanced in the future, use of play dough in some complicated cerebral aneurysm cases may be helpful in 3D reconstruction of the real situation.
==== Body
Background
Imagination of the three-dimensional (3D) structure of cerebral vascular lesions using two-dimensional (2D) angiograms is one of the skills that neurosurgical residents should achieve during their training. Although ongoing progress in computer software and digital imaging systems has facilitated viewing and interpretation of cerebral angiograms significantly, the 2D nature of these images makes them still far from ideal. These facilities are not always readily available in many situations. Much of the actual mastering is achieved experientially through work on cadavers or in operating theatres.
Three-dimensional models made from materials such as wax, bronze and ivory have been used in the teaching of medicine for many centuries.
It is thought that the first three-dimensional model of the vascular tree was created by a follower of Mondino de'Luzzi in the 14th century. Molten wax was injected into the vascular system, forming a cast that was carefully dissected out from the surrounding tissue [1,2]. In the 17th and 18th centuries artists such as Ercole Lelli (1702–1766) turned to colored waxes to realistically recreate dissected figures and organs[3,4]. More recently, technology has started to displace this traditional way of teaching with the development of high quality visual and often interactive three-dimensional (3D) computer-generated images [5]. Computerized 3D models have not only been used for teaching anatomy and pathology, but also in different fields of neurosurgical training and interpretation of neuroradiological images [6-8].
Although there are some anatomic variations among different patients in all organs, the individual anatomy of brain vascular lesions in particular needs to be studied precisely before surgical intervention. This study is traditionally done through 2D angiograms. While modalities like Computed Tomography (CT) angiograms and Magnetic Resonance Angiography (MRA) gather data three dimensionally, this is only available through 2D monitors or printed materials. Current commercial computer software have made manipulation and viewing of the Magnetic Resonance Imaging (MRI) or CT images much easier, but their capabilities are far from ideal.
An evolving technology with potential application to medicine is three-dimensional printing, also known as rapid prototyping. Devices "print" three-dimensional data sets into solid models using various materials such as plastic, wax, and metal. Radiographic studies containing volumetric data can therefore be made into realistic three-dimensional physical anatomic models[2].
It seems that the trend is towards more complicated imaging technologies, both in training and surgical intervention.
Play dough is an old toy that has been used both in play and learning by our children for a long time. It is a time-honored educational tool on account of its simplicity of use, its pliability and relative inexpensiveness [9]. We could not find any reference in medical literature regarding similar usage as an educational tool for neurosurgical training in visualization of complicated aneurysm anatomy as an adjunct to cerebral angiography.
Methods
Where to use the play dough models?
We have used play dough as an adjunct to our traditional training tools and found some advantages for it. Its use is considered only in some visuospatially complicated aneurysms (figures 1 and 2) where imagination of the real anatomy based on 2D images or drawings may be difficult. It can be considered a good alternative for drawings sometimes done pre- and postoperatively by the consultants and residents in order to document what they are going to see or have seen during the operation. The preoperative model can be revisited after the operations to see where they went wrong.
Figure 1 Middle cerebral artery aneurysm angiogram, anteroposterior view. ACA Anterior Cerebral Artery, ICA Internal Carotid Artery, MCA Middle Cerebral Artery. Aneurysm is seen in the trifurcation of MCA
Figure 2 A Play Dough model based on the same patient's angiograms. ACA Anterior Cerebral Artery, ICA Internal Carotid Artery, MCA Middle Cerebral Artery. The model clarifies the location of the Aneurysm. Brain tissue shown in yellow.
How to make the models?
The play dough used was not different from what children use for playing. Only two colors, red for vessels and yellow or white for nervous tissue have been used. The trainee studies the patient's angiograms and other available images and based on her or his previous experiences and knowledge, imagines the 3D structure of the lesion. Then using the play dough the trainee materializes her or his concept of the structure (figures 1 and 2). Since the goal is only to clarify the vascular anatomy, building of the models takes a short time (less than 5 minutes on average). There is no need for previous experience with play dough.
Results
We have used this method for two years. In all cases, the preoperative models needed corrections and differed significantly from the true anatomy of the lesion. In those few cases where the models were made postoperatively, the results were much better, but interestingly the need for corrections in the minority of these cases could show us incorrect intraoperative 3D concept of the residents.
Discussion
Since the number of neurosurgical residents is limited, we could not conduct an acceptable quantitative study regarding the comparison of play dough models with other teaching methods and report the advantages and disadvantages quantitatively.
Advantages
Besides the low cost, availability and simplicity of use, play dough is much more efficient and realistic in training less experienced residents in comparison with the simple drawings and even angiographic views from different angles. It obviates the need for technological equipment. It helps discussion about the negative and positive points of different approaches with regard to the anatomy of surrounding vessels.
Disadvantages
One of the major disadvantages of play dough is the psychological resistance of residents to the use in surgical training of something that usually is considered to be a toy. With time, this seems to lessen, especially when the constructed models help them present their questions or comments. It may be more acceptable to those trainees who have an artistic flair. Play dough is not as clean as drawings or computerized images. Use of computer software is not only more in vogue and acceptable, but also strengthens the computing skills of trainees.
Design and conduct of a study to compare the application of this tool to other training methods may help better evaluation of play dough as an educational tool.
Conclusion
Although technologies like computerized software using the patients' own imaging data seems set to become more advanced in the future, use of play dough for some complicated cerebral aneurysm cases may be helpful in realistic 3D reconstruction of the lesions.
Competing interests
The author(s) declare that they have no competing interests.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
We thank Dr. Ali Saligheh Araghi Fellow in Intensive Care, University of Miami
Florida, for his comments and help. We would also like to thank Dr. Orla Dunne,
Mater Misericordiae University Hospital, Dublin for her editorial assistance.
==== Refs
Haviland TN Parrish LC A brief account of the use of wax models in the study of medicine J Hist Med Allied Sci 1970 25 52 75 4944593
Chen JC Amar AP Levy ML Apuzzo ML The development of anatomic art and sciences: the ceroplastica anatomic models of La Specola Neurosurgery 1999 45 883 91; discussion 891-2 10515484 10.1097/00006123-199910000-00031
Neave R Pictures in the round: moulage and models in medicine J Audiov Media Med 1989 12 80 84 2691552
Vernon T Peckham D The benefits of 3D modelling and animation in medical teaching J Audiov Media Med 2002 25 142 148 12554292 10.1080/0140511021000051117
Kling-Petersen T Rydmark M The BRAIN project: an interactive learning tool using desktop virtual reality on personal computers Stud Health Technol Inform 1997 39 529 538 10168945
Eftekhar B Ghodsi M Ketabchi E Rasaee S Surgical simulation software for insertion of pedicle screws Neurosurgery 2002 50 222 3; discussion 223-4 11844256 10.1097/00006123-200201000-00038
Phillips NI John NW Web-based surgical simulation for ventricular catheterization Neurosurgery 2000 46 933 6; discussion 936-7 10764268 10.1097/00006123-200004000-00031
Sharples M Jeffery NP du Boulay B Teather BA Teather D du Boulay GH Structured computer-based training in the interpretation of neuroradiological images Int J Med Inf 2000 60 263 280 10.1016/S1386-5056(00)00101-5
Girolametto L Weitzman E van Lieshout R Duff D Directiveness in teachers' language input to toddlers and preschoolers in day care J Speech Lang Hear Res 2000 43 1101 1114 11063233
|
15885141
|
PMC1274244
|
CC BY
|
2021-01-04 16:30:57
|
no
|
BMC Med Educ. 2005 May 10; 5:15
|
utf-8
|
BMC Med Educ
| 2,005 |
10.1186/1472-6920-5-15
|
oa_comm
|
==== Front
RetrovirologyRetrovirology1742-4690BioMed Central London 1742-4690-2-291587635810.1186/1742-4690-2-29ResearchElevated expression of CD30 in adult T-cell leukemia cell lines: possible role in constitutive NF-κB activation Higuchi Masaya [email protected] Takehiro [email protected] Naoki [email protected] Yasuaki [email protected] Ryouichi [email protected] Toshiki [email protected] Masahiko [email protected] Masayasu [email protected] Masahiro [email protected] Division of Virology, Niigata University Graduate School of Medical and Dental Sciences, Niigata 951-8510, Japan2 Division of Molecular Virology and Oncology, Faculty of Medicine, University of the Ryukyus, Nishihara, Okinawa 903-0215, Japan3 Department of Laboratory Medicine, Nagasaki University Graduate School of Biomedical Sciences, Nagasaki 825-8501, Japan4 Fourth Department of Internal Medicine, Faculty of Medicine, Kitasato University, Sagamihara, Kanagawa 228-8555, Japan5 Laboratory of Tumor Cell Biology, Department of Medical Genome Sciences, Graduate School of Frontier Sciences, The University of Tokyo, Minato-ku, Tokyo 108-109, Japan2005 6 5 2005 2 29 29 7 2 2005 6 5 2005 Copyright © 2005 Higuchi et al; licensee BioMed Central Ltd.2005Higuchi et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Human T-cell leukemia virus type 1 (HTLV-1) is associated with the development of adult T-cell leukemia (ATL). HTLV-1 encoded Tax1 oncoprotein activates the transcription of genes involved in cell growth and anti-apoptosis through the NF-κB pathway, and is thought to play a critical role in the pathogenesis of ATL. While Tax1 expression is usually lost or minimal in ATL cells, these cells still show high constitutive NF-κB activity, indicating that genetic or epigenetic changes in ATL cells induce activation independent of Tax1. The aim of this study was to identify the molecules responsible for the constitutive activation of NF-κB in ATL cells using a retroviral functional cloning strategy.
Results
Using enhanced green fluorescent protein (EGFP) expression and blasticidin-resistance as selection markers, several retroviral cDNA clones exhibiting constitutive NF-κB activity in Rat-1 cells, including full-length CD30, were obtained from an ATL cell line. Exogenous stable expression of CD30 in Rat-1 cells constitutively activated NF-κB. Elevated expression of CD30 was identified in all ATL lines examined, and primary ATL cells from a small number of patients (8 out of 66 cases).
Conclusion
Elevated CD30 expression is considered one of the causes of constitutive NF-κB activation in ATL cells, and may be involved in ATL development.
==== Body
Background
Adult T-cell leukemia (ATL) is an extremely aggressive human CD4+ T-cell leukemia (reviewed in [1]). ATL is resistant to chemotherapy and most patients die within one year of diagnosis. Human T-cell leukemia virus type 1 (HTLV-1) infection of CD4+ T-cells is the first step in ATL development. However, this alone is not sufficient for the development of leukemia because a minority of HTLV-1 infected subjects (approximately 5%) develop ATL on average 60–70 years after the infection (reviewed in [2,3]). In vitro, HTLV-1 transforms primary human CD4+ T-cells in an interleukin (IL)-2-dependent or an IL-2-independent manner. HTLV-1 encoded Tax1 protein is thought to play a critical role in T-cell transformation and leukemogenesis, as Tax1 itself immortalizes primary human CD4+ T-cells in vitro [4,5] and inhibits apoptosis induced by various stimuli in T-cell lines [6-9].
Tax1 is a multifunctional protein (reviewed in [2,3]). It activates the transcription of many cellular genes associated with cell growth, such as genes encoding cytokines [10-13], cytokine receptors [14-17], anti-apoptotic protein [8,18], cell cycle regulators [19-22], and proto-oncogenes [23]. Those proteins are thought to contribute to the deregulated proliferation of HTLV-1-infected cells. Accumulating evidence suggests that activation of cellular genes by Tax1, particularly through the nuclear factor-kappaB (NF-κB) pathway, is a critical process in transformation as well as the inhibition of apoptosis. For example, the transforming activity of Tax1 is abrogated by mutations that impair the ability of Tax1 to activate NF-κB [24-26]. Tax1 inhibits apoptosis of mouse T-cell lines by induction of the anti-apoptotic gene Bcl-xL through NF-κB activation [8,18].
In resting T-cells, NF-κB factors are sequestered in the cytoplasm, tightly associated with inhibitory proteins IκBs. Activation of NF-κB generally involves phosphorylation and degradation of IκBs, followed by nuclear translocation of NF-κB dimers and subsequent activation of the genes containing NF-κB binding sites (reviewed in [27]). Alternatively, NF-κB activation occurs by inducible processing of NFKB2/p100 with IκB-like inhibitory activity, into p52 with DNA binding activity, followed by nuclear translocation of p52 containing NF-κB dimers (reviewed in [28]). These two processes are largely dependent on an IκB kinase (IKK) complex comprised of two catalytic subunits, IKKα and IKKβ and a regulatory subunit IKKγ/NEMO. Tax1 interacts with the IKK complex through these three subunits and stimulates the catalytic activity [29-32].
In primary ATL cells as well as cell lines established from ATL patients, NF-κB is constitutively active as seen in HTLV-1 transformed cells [33]. It appears that this constitutive NF-κB activation contributes to the survival and chemotherapy resistance of ATL cells, since treatment of ATL cells with a NF-κB inhibitor, Bay 11-7082, induces apoptosis of these cells [34]. However, how NF-κB is constitutively activated in ATL cells is still largely unknown since the tax gene is mutated in some ATL cases [35,36] or the level of expression of Tax1 in these cells is extremely low, thereby being clearly insufficient to activate NF-κB [37,38]. There may be genetic or epigenetic changes that lead to tax-independent NF-κB activation, such as a gain of function of the NF-κB activating molecule(s) or a loss of function of the NF-κB regulator(s). The elucidation of the molecular mechanism of NF-κB activation in ATL cells is quite important in the light of prevention, diagnosis and treatment of ATL.
In order to identify the molecule(s) responsible for the constitutive NF-κB activation in ATL, we took a functional screening approach using a retroviral cDNA library from an ATL cell line and a reporter cell line that is easily distinguishable as a positive clone once NF-κB is activated. We obtained several cDNA clones that constitutively activate NF-κB. One of these, the full-length CD30, is a member of the TNF receptor superfamily and a marker of malignant Hodgkin and Reed-Sternberg (H-RS) cells in Hodgkin's lymphoma (HL) (reviewed in [39,40]). It is suggested that overexpression of CD30 in H-RS cells and HL cell lines contributes to CD30 ligand-independent constitutive NF-κB activation in these cells [41]. The results showed that CD30 is strongly expressed in all ATL cell lines examined, and that CD30 is expressed in primary ATL cells in a small number of ATL patients.
Results and Discussion
Screening of NF-κB activating molecules
In order to identify the molecule(s) responsible for the constitutive NF-κB activation in ATL cells, we employed a functional screening strategy using a retroviral cDNA library from an ATL cell line. In theory, if ATL cells express NF-κB activating molecules leading to the constitutive activation, it would be possible to obtain such clones using NF-κB activation as a positive selection marker (Figure 1A). We generated a retroviral cDNA library from ATL cell line TL-OmI, which had already been shown to have constitutive NF-κB activity in the absence of Tax1 [33]. As a reporter cell line, we generated a Rat-1 fibroblast cell line with a stably integrated blasticidin deaminase gene (bsr) fused to enhanced green fluorescent protein (EGFP) under five repeats of the NF-κB binding sequences from the IL-2 receptor α chain and the minimal HTLV-1 promoter [42]. The bsr and EGFP enabled us to easily identify NF-κB activated cells as surviving cells with green fluorescence in the presence of blasticidin. A pilot experiment, however, showed that the green fluorescent signal from this fusion protein in the cells after NF-κB activation stimuli (such as TNF-α treatment) was extremely low, probably due to the short half life of the fusion protein or a conformational change that interferes with EGFP activity (data not shown). Thus, we further stably transfected the EGFP gene regulated under the same NF-κB responsive promoter into the reporter cell line. This new reporter cell line, named Rat-1 κB-bsrEGFPx2, showed bright EGFP signals and blasticidin resistance after TNF-α treatment (Figure 1B and data not shown). This doubly transfected cell line has a critical advantage in this screening system. It is possible that retroviral cDNA is inserted near the bsrEGFP gene and the retroviral long terminal repeat (LTR) constitutively activates the expression of the bsrEGFP gene, resulting in a false positive clone. However, if it occurs in the new reporter cell line, these cells should have minimal EGFP signals because of the extremely low fluorescence intensity of the fusion protein and such cells could be easily eliminated during the screening process.
Figure 1 Strategy for cloning NF-κB activating molecules. A) A retroviral cDNA library from an ATL cell line is transduced to a reporter cell line expressing EGFP and bsr in response to NF-κB activation. Blasticidin-resistant and EGFP expressing cell clones are expanded and cDNA clones are obtained by PCR using the retrovirus vector specific primers. B) Visualization of NF-κB activation in reporter cells. Reporter cells were stimulated with TNF-α for 48 hours and tested for the expression of EGFP by FACS analysis.
After converting the plasmid library to the retroviral library by introduction into packaging cells, the resultant viruses were transduced into the Rat-1 κB-bsrEGFPx2 reporter cells. After selection in the presence of blasticidin, under an inverted fluorescence microscope, EGFP-positive cells were picked up and expanded, followed by genomic DNA extraction. PCR products amplified by the primers specific for the retroviral vector were cloned and the sequences were determined. Following three independent screenings, we obtained a total of 64 clones (Table 1).
Table 1 NF-κB activators isolated from the TL-OmI cDNA library.
cDNA No. of isolates Characteristics
NIK 58 N terminal deletion
CD30 3 Full length
LT-βR 2 Cytoplasmic region
RIP2 1 Full length
NF-κB inducing kinase (NIK) is a mitogen-activated protein kinase kinase kinase (MAP3K), which is involved in NFKB2/p100 processing and nuclear translocation of p52/RelB dimers, the so-called noncanonical pathway [43]. This pathway is activated by lymphotoxin-β (LT-β), CD40 ligand, and B cell activating factor (BAFF) and depends on IKKβ (reviewed in [28]). All the NIK clones we obtained possessed the intact kinase domain and the N-terminal amino acid deletion, starting at codon 417. It has been reported that the N-terminus of NIK contains a negative-regulatory domain and an N-terminal truncation mutant has higher NF-κB inducing activity than the wild type [44]. It is likely that this deletion was introduced by incomplete reverse transcription with oligo dT primer during the cDNA library construction process. It is interesting to note that none of the other MAP3Ks that can activate NF-κB, such as MEKK1 [45], were cloned. This selective isolation of NIK as well as its high frequency among the NF-κB-inducing clones indicates that NIK and/or the noncanonical pathway may play a central role in the constitutive NF-κB activation seen in various tumors.
The sequences of the two LT-β receptor (LT-βR) clones were identical and encoded a part of the cytoplasmic domain of the receptor (from codon 268 to 395). The retrovirus vector used in our experiments transcribes two mRNAs, one spliced and one unspliced. The unspliced mRNA may translate fusion genes of gag with inserted cDNA when they are in frame. The isolated LT-βR clone is in frame with gag and could be expressed as a fusion protein, which might induce constitutive NF-κB activation. This cloned LT-βR mutant is likely to be an artificial one generated during the library construction process as discussed above.
The receptor-interacting protein 2 (Rip2) is a serine/threonine kinase that contains a caspase-recruitment domain (CARD) at its carboxyl terminus and has been shown to induce NF-κB activation in an over-expression system [46]. Rip2 has been implicated in regulating both the innate and adaptive immune responses [47,48]. Recently, it has been reported that Rip2 participates in Bcl10-mediated NF-κB activation [49]. The Rip2 clone isolated in our study is full length and not in frame with gag. It is possible that Rip2 is over-expressed in ATL cells and this contributes to constitutive NF-κB activation. This hypothesis is currently under investigation.
Exogenous stable expression of CD30 induces constitutive NF-κB activation
CD30 is a member of the TNF receptor super family and is known as a marker of malignant Hodgkin and Reed-Sternberg (H-RS) cells in Hodgkin's lymphoma (HL). It has been suggested that overexpression of CD30 in H-RS cells and HL cell lines contributes to CD30 ligand (CD30L)-independent constitutive NF-κB activation in these cells [41]. The same possibility in ATL cells was further examined. One of the three CD30 clones (named kBL1) contains full-length CD30 in frame with gag (the other two clones were not completely sequenced). As described above, the retrovirus vector used in our experiments transcribes two mRNAs, one is a spliced one and the other is an unspliced one. The unspliced mRNA can translate fusion genes of gag with inserted cDNA when they are in frame. To determine that the fusion between CD30 and gag is responsible for its constitutive NF-κB inducing activity, we generated a retroviral vector that expresses only full-length CD30 by introducing a frame shift mutation upstream of the CD30 open reading frame of the cloned gene. We also constructed a retroviral vector for full-length CD30 cDNA (pMX CD30WT) out of frame with gag. Retroviral vectors for CD30 either in or out of frame with gag (pMX kBL1 or pMX kBL1ΔBglII respectively) and pMX CD30WT were introduced into packaging cells and the Rat-1-bsrEGFPx2 cells were infected with the resultant viruses. After 48 hours, EGFP signals were examined by fluorescence activated cell sorter (FACS) analysis. In all three cases, CD30 induced constitutive NF-κB activation, although CD30 in frame with gag had stronger NF-κB inducing activity, which means the fusion with gag indeed augments the activity (Figure 2). This result demonstrates that stably overexpressed CD30 can induce constitutive NF-κB activation in a ligand independent manner in Rat-1 cells, as described previously in human embryonic kidney cell line 293 [41].
Figure 2 Exogenous stable expression of CD30 induces constitutive NF-κB activation in Rat-1 cells Rat-1 κB-bsrEGFPx2 cells were infected with the pMX kBL1, pMX kBL1ΔBglII, or pMX CD30WT virus and tested for the expression of EGFP by FACS analysis. The cells infected with pMX virus were used as a negative control. CD30 expression was seen in cells infected with all three viruses containing CD30 gene (pMX kBL1, pMX kBL1ΔBglII, pMX CD30WT) but not pMX virus (data not shown).
Overexpression of CD30 in ATL cell lines
We next examined CD30 expression in ATL-derived T-cell lines, HTLV-1 transformed cell lines and HTLV-1 negative T-cell lines using FACS analysis (Figure 3). All ATL cell lines (TL-OmI, KOB, KK1 and ST1) showed strong CD30 expression whereas a B lymphoma cell line (BJAB) showed no staining (Figure 3A). HTLV-1 transformed cell lines (HUT-102, C5/MJ, MT-4 and SLB-1) also showed CD30 expression but the amount of the expression was various and lower than TL-OmI (Figure 3B). In HTLV-1 negative T-cell lines (Jurkat and MOLT-4), the expression of CD30 was significantly lower than TL-OmI (Figure 3C). Interestingly, NF-κB activity was much lower in Jurkat and MOLT-4 than ATL cell lines. Thus CD30 expression level is well correlated with the NF-κB activity, which suggests that overexpression of CD30 might be at least one of the factors that contributes to constitutive NF-κB activation in ATL cell lines. In HTLV-1 transformed cells, NF-κB activation is thought to be largely dependent on Tax1, however it is possible that relatively strong CD30 expression in HUT-102 and SLB-1 also contributes to constitutive NF-κB activation in these cells. In addition, CD30L expressed in ATL cell lines may possibly contribute to CD30 activation by a cell-cell contact mechanism. RT-PCR analysis for CD30 ligand showed that CD30L expression in TL-OmI cells was extremely weak compared with a Burkitt lymphoma cell line (EB-1), in which CD30L is weakly expressed (data not shown) [50]. This finding suggests that CD30L is not involved in the constitutive NF-κB activation in TL-OmI cells.
Figure 3 Elevated expression of CD30 in ATL cell lines. CD30 expression was examined in A) ATL, B) HTLV-1-transformed, and C) HTLV-1-negative cell lines by FACS analysis. A Burkitt lymphoma cell line (BJAB) was used as a negative control in A). TL-OmI was used as a standard for the CD30 expression level in B) and C).
Expression of CD30 in primary ATL cells
Next, we examined CD30 expression in primary ATL cells by FACS analysis. Peripheral blood lymphocytes (PBLs), lymph node cells, or ascitic fluid cells from ATL patients were stained with anti-CD30 antibody (Figure 4 and Table 2). ATL cases in which more than 30% of the cells expressed CD30 were classified as CD30-positive ones. CD30 expression was seen in 8 of 66 ATL cases (12.1%) and the CD30 expression was predominantly seen in the acute type (5 of 25 cases), representing the advanced stage of ATL (Figure 4B). Data of the FACS analysis (CD3, CD4, CD8, CD25, and CD30 expression) of the CD30-positive ATL cases are summarized in Table 2.
Table 2 Cell surface markers in CD30-positive ATL cases
% of Positive Cells
Case Sex Type Material CD3 CD4 CD8 CD25 CD30
1 M Acute PB 90.1 86.5 4.4 89.8 56.5
2 F Acute PB 18.9 78.3 3.0 81.0 48.7
3 F Acute PB 94.8 14.2 64.2 81.9 84.8
4 F Acute PB 89.3 96.3 2.6 93.3 35.5
5 M Acute LN 10.1 96.6 5.2 90.1 93.0
6 M Lymphoma LN 8.7 85.1 5.3 58.1 76.5
7 F Unknown LN 67.5 77.8 23.0 81.7 60.4
8 F Unknown Ascites 89.5 99.7 0.1 99.5 96.2
The percentage of positive cells was determined by immunofluorescence staining with respective antibodies and flow cytometric analysis.
Abbreviations: PB, peripheral blood; LN, lymph node.
Figure 4 CD30 expression in primary ATL cells. A) Primary ATL cells from a patient (case 8) were tested for the expression of CD3, CD4, CD8, CD25 and CD30 by FACS analysis. B) Summary of the number of CD30-positive ATL cases.
It has been reported that proteolytic cleavage of membrane-anchored CD30 releases a soluble fragment corresponding to the extracellular domain [51]. To examine the possibility that the low frequency of CD30 expression in primary ATL cells in the FACS analysis is due to this proteolytic processing, CD30 mRNA expression was examined in 8 ATL cases different from those used in the FACS analysis. Strong CD30 mRNA expression was seen in HUT-102 and PBLs activated by phytohemagglutinin (PHA), whereas the CD30 expression was seen in only one case (ATL8) diagnosed as the lymphoma type (Figure 5). The amount of CD30 mRNA expression in this case was lower than HUT-102 and it might not be sufficient to induce NF-κB activation by itself. However it is possible that weak CD30 expression still contributes to the constitutive NF-κB activation in cooperation with other signaling molecules in vivo. In summery, these FACS and RT-PCR data suggest that the expression of CD30 in ATL is not a common event and is limited to a small number of ATL cases. This is consistent with a previous report that CD30 expression was seen in 7 out of 36 cases (19.4%) when their lymph node biopsies were immunohistochemically stained with anti-CD30 antibody [52].
Figure 5 CD30 mRNA expression in primary ATL cells Primary ATL cells from ATL patients (lanes 5–12) and normal PBLs from healthy adult donors (lanes 1–3) were tested for CD30 (upper panel) and β-actin (lower panel) mRNA expression by RT-PCR analysis. The CD30 expression was seen in ATL8 (lane 12). PHA-stimulated PBLs (lane 4) and HUT-102 (lane 13) were used as a positive control.
The reason for the discrepancy between ATL cell lines and primary ATL cells in terms of CD30 expression is unknown at present. One possibility is that only CD30-positive primary ATL cells could be established as a cell line in vitro because of their stronger NF-κB activity or activation of other signaling pathways originating from CD30. In fact, CD30 activates not only NF-κB but also the mitogen activated protein kinase (MAPK) pathways, such as extracellular regulated kinase (ERK), Jun N-terminal kinase (JNK), and p38 MAPK pathways [53,54].
Recently, it has been reported that the noncanonical pathway is involved in constitutive NF-κB activation in ATL cells [55]. Although activation of the noncanonical pathway by CD30 has not yet been reported, it is likely that CD30 activates this pathway through association with TNF receptor associated factors (TRAFs) like LT-βR and CD40. In H-RS cells, which strongly express CD30, TRAF2 and TRAF5 make aggregates in the cytoplasm and co-localize with downstream signaling molecules, such as IKKα and IKKβ [56]. It would be interesting to see whether TRAF2 and TRAF5 also form aggregates in ATL cell lines and primary ATL cells expressing CD30.
In order to confirm that CD30 is involved in constitutive NF-κB activation and cell survival in ATL cell lines, we tried to knockdown CD30 expression in these cells by using short-hairpin RNAs. We generated 11 different short-hairpin RNAs for CD30 in total, but none of them showed any RNA interference effect. We also tried to introduce a decoy CD30 that lacks most of the cytoplasmic region and has been shown to induce apoptosis in H-RS cells [41], by using an adenovirus vector. However we were unable to obtain a sufficiently high titer adenovirus as a decoy CD30 mutant to carry out the experiment. Thus, whether elevated expression of CD30 actually contributes to constitutive NF-κB activation in ATL cell lines still remains unknown.
In this regard, the mechanism by which NF-κB is constitutively activated in ATL cells still remains a mystery. However, our data suggest that the elevated expression of CD30 plays a critical role in NF-κB activation in ATL cell lines and a small number of primary ATL cells. Other molecules belonging to the TNF receptor family, such as LT-βR, OX40, or downstream signaling molecules, could be involved in constitutive NF-κB activation in CD30-negative ATL cells, and the identification of such molecules would contribute to the prevention, diagnosis and treatment of ATL.
Conclusion
ATL cells have constitutive NF-κB activity which is important for the cells' survival. This NF-κB activation is independent of Tax protein expression. By screening a retroviral cDNA library from an ATL cell line to identify NF-κB activating molecules, we obtained several cDNA clones including full-length CD30. CD30 is strongly expressed in ATL cell lines and primary ATL cells from a small number of patients. Our results suggest that elevated expression of CD30 is one of the factors responsible for constitutive NF-κB activation in ATL cells.
Methods
Cell culture
Rat-1, a rat fibroblast cell line, was cultured in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal bovine serum (FBS). Human T-cell lines used in the present experiments have been characterized previously [33,57]. Jurkat and MOLT-4 are HTLV-1 negative human T-cell lines. HUT-102, C5/MJ, MT-4 and SLB-1 are HTLV-1-positive human T-cell lines. TL-OmI, KK1 [58], KOB [59], and ST1 [60] are HTLV-1-positive, ATL-derived cell lines. These cells were cultured in RPMI 10% FBS. Recombinant human IL-2 (Takeda Chemical Industries, Osaka, Japan) was added at 0.5 nM to the culture of KK1, KOB and ST1. A retrovirus packaging cell line Plat-E [61] was cultured in DMEM 10% FBS containing 1 μg/ml puromycin (Calbiochem, La Jolla, CA) and 10 μg/ml blasticidin (Invitrogen, San Diego, CA).
cDNA library construction
Poly (A)+ RNA was purified from TL-OmI using FastTrack 2.0 (Invitrogen). cDNA was synthesized by oligo(dT) primers using SuperScript Choice System (Invitrogen) according to the instructions provided by the manufacturer. The resulting cDNAs were size-fractionated through agarose gel electrophoresis, and cDNA fragments longer than 2.5 kb were extracted from the gel by using Qiaex II (Qiagen, Hilden, Germany). The cDNA fragments were then inserted into BstXI sites of the retroviral vector pMX [62] using BstXI adapters (Invitrogen). The ligated DNA was ethanol-precipitated and then electroporated into DH10B competent cells (Electromax DH10B; Invitrogen). About 1 × 106 independent clones were amplified on 150 mm LB/amp plates and plasmid DNA was purified by using Qiagen Plasmid Giga kit (Qiagen).
Generation of a reporter cell line
The NF-κB reporter plasmid κB-EGFP was constructed by replacing the luciferase gene (a BglII – BamHI fragment) of the κB-Luc plasmid [42] with EGFP (a HindIII – AflII fragment) from pEGFP-N3 (Clontech Laboratories, Palo Alto, CA) by blunt-end ligation. To construct the plasmid κB-bsrEGFP, which expresses bsrEGFP fusion protein, a PCR amplified bsr gene fragment was inserted in the ApaI and BamHI sites upstream of EGFP of the κB-EGFP plasmid. To prepare a NF-κB reporter cell line, Rat-1 cells (5 × 106) were transfected with 20 μg of κB-bsrEGFP and 1 μg of pcDNA3 (Invitrogen) by electroporation at 250 V and 975 μF. The transfected cells were cultured in 500 μg/ml G418 (Invitrogen), and resistant clones were screened for EGFP signals after being infected with retroviruses that express Epstein-Barr virus transforming protein LMP1. The selected cell clone (Rat-1 κB-bsrEGFP) was further transfected with κB-EGFP and pMik-HygB and cultured in 250 μg/ml hygromycin B (Wako Pure Chemical Industries, Osaka, Japan). Resistant clones were screened for EGFP signals after stimulation with 20 ng/ml TNF-α (Peprotech, London, UK).
Preparation of retroviruses and infection of reporter cells
Plat-E cells (2 × 106 cells) were seeded onto 60 mm dishes one day before transfection. The cDNA library (3 μg) was transfected using Fugene 6 (Roche Molecular Systems, Inc., NJ) according to the protocol provided by the manufacturer. Cells were cultured for 48 hours and the retroviral supernatant was harvested. For infection of reporter cells, 2.5 × 105 cells were seeded onto 100 mm dishes one day before infection and incubated with 10 ml DMEM 10% FBS containing 0.6 ml of the virus stock for 24 hours in the presence of polybrene (20 μg/ml). The medium was changed to fresh DMEM 10% FBS after 24 hours. After another 24 hours, the cells were incubated with medium containing 50 μg/ml blasticidin (Invitrogen).
Isolation of cDNA fragments from blasticidin-resistant clones
Genomic DNA was extracted from the blasticidin-resistant clones by DNeasy kit (Qiagen) and subjected to PCR to recover integrated cDNAs using pMX vector primers (5'-GGTGGACCATCCTCTAGACT-3' and 5'-CCCCTTTTTCTGGAGACTAAAT-3'). The PCR products were cloned into pGEM-T Easy vector (Promega, Madison, WI) and sequenced using BigDye Terminator v1.1 cycle sequencing kit (Applied Biosystems, Foster City, CA).
Expression plasmids
The retroviral vector pMX LMP1 was prepared by inserting an EcoRI – BamHI fragment of pSG5 F-LMP1 [63] in the EcoRI site of pMX by blunt-end ligation. pMX CD30WT was generated by inserting a MluI – NotI fragment of pCD30WT [64] in the EcoRI and NotI sites of pMX by blunt-end ligation. pMX kBL1 was generated by inserting a BamHI – SfiI fragment of pGEMT kBL1 in the BamHI – NotI sites of pMX. The BglII site of pMX kBL1 was destroyed by cutting by BglII, filling in by T4 polymerase, and self-ligation to make pMX kBL1ΔBglII.
Flow cytometric analysis
Heparinized peripheral blood, a piece of a lymph node, or ascites (in case no. 8) was collected from patients with ATL after obtaining informed consent in accordance with the Helsinki Declaration. Mononuclear cells were separated by Lymphoprep™ density gradient centrifugation (Axis-Shield PoC AS, Oslo, Norway). Morphological and surface marker analyses indicated that ATL cells in these samples always accounted for more than 80% of the total cell population in most cases. The study protocol was approved by the Human Ethics Review Committee of Nagasaki University Graduate School of Biomedical Sciences.
Primary ATL cells or T-cell lines were incubated for 30 min at 4°C with each PE-labeled or FITC-labeled monoclonal antibody (mAb). Cells were also incubated with isotype matched control antibodies. The following antibodies were used: PE-labeled mouse anti-human CD4 and CD25, FITC-labeled anti CD3 and CD8 (BD Biosciences Pharmingen, San Diego, CA); and PE-labeled mouse anti-human CD30 (Dako Corporation, Carpinteria, CA or Immunotech, Marseille, France). After washing with PBS, the cells were analyzed on FACScan flow cytometer using Cellquest software (Becton Dickinson, San Jose, CA).
Reverse transcription-polymerase chain reaction
Total cellular RNA was extracted with Trizol (Invitrogen) according to the protocol provided by the manufacturer. First-strand cDNA was synthesized from 1 μg total cellular RNA in a 20-μl reaction volume using an RNA PCR kit (Takara Shuzo, Kyoto, Japan) with random primers. Thereafter, cDNA was amplified for 35 cycles for CD30 and 28 cycles for β-actin. The oligonucleotide primers used were as follows: for CD30, sense, 5'-CTGTGTCCCCTACCCAATCT-3' and antisense, 5'-CTTCTTTCCCTTCCTCTTCCA-3'; [65] and for β-actin, sense, 5'-GTGGGGCGCCCCAGGCACCA-3' and antisense, 5'-CTCCTTAATGTCACGCACGATTTC-3'. Product sizes were 860-bp for CD30 and 548-bp for β-actin. Cycling conditions were as follows: denaturing at 94°C for 45 sec (for CD30) or for 30 sec (for β-actin), annealing at 62°C for 45 sec (for CD30) or 60°C for 30 sec (for β-actin) and extension at 72°C for 60 sec (for CD30) or for 90 sec (for β-actin). The PCR products were fractionated on 2% agarose gels and visualized by ethidium bromide staining.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
MH carried out the cDNA cloning and the functional analysis of CD30. TM and NM carried out the RT-PCR analysis. YY carried out the FACS analysis. MH, RH, TW, MT, MO and MF participated in the experimental design, data interpretation, and writing of the manuscript.
Acknowledgements
We are deeply indebted to the many patients with ATL and the control subjects who donated blood for these studies. We thank T. Kitamura for providing the retroviral vector pMX and the packaging cell line Plat-E. We also thank R. Fujita, S. Takizawa, and C. Yamamoto for the excellent technical assistance. This work was supported in part by a Grant-in-Aid for Scientific Research of Japan, Grant for Promotion of Niigata University Research Projects, and Tsukada Grant for Niigata University Medical Research.
==== Refs
Sugamura K Hinuma Y Levy JA Human retroviruses: HTLV-I and HTLV-II The Retrovirudae 1993 2 New York: Plenum Press 399 435
Yoshida M Multiple viral strategies of HTLV-1 for dysregulation of cell growth control Annu Rev Immunol 2001 19 475 496 11244044 10.1146/annurev.immunol.19.1.475
Matsuoka M Human T-cell leukemia virus type I and adult T-cell leukemia Oncogene 2003 22 5131 5140 12910250 10.1038/sj.onc.1206551
Grassmann R Berchtold S Radant I Alt M Fleckenstein B Sodroski JG Haseltine WA Ramstedt U Role of human T-cell leukemia virus type 1 X region proteins in immortalization of primary human lymphocytes in culture J Virol 1992 66 4570 4575 1351105
Akagi T Shimotohno K Proliferative response of Tax1-transduced primary human T cells to anti-CD3 antibody stimulation by an interleukin-2-independent pathway J Virol 1993 67 1211 1217 8437212
Brauweiler A Garrus JE Reed JC Nyborg JK Repression of bax gene expression by the HTLV-1 Tax protein: implications for suppression of apoptosis in virally infected cells Virology 1997 231 135 140 9143312 10.1006/viro.1997.8509
Mulloy JC Kislyakova T Cereseto A Casareto L LoMonico A Fullen J Lorenzi MV Cara A Nicot C Giam C Franchini G Human T-cell lymphotropic/leukemia virus type 1 Tax abrogates p53-induced cell cycle arrest and apoptosis through its CREB/ATF functional domain J Virol 1998 72 8852 8860 9765430
Tsukahara T Kannagi M Ohashi T Kato H Arai M Nunez G Iwanaga Y Yamamoto N Ohtani K Nakamura M Fujii M Induction of Bcl-x(L) expression by human T-cell leukemia virus type 1 Tax through NF-kappaB in apoptosis-resistant T-cell transfectants with Tax J Virol 1999 73 7981 7987 10482545
Kawakami A Nakashima T Sakai H Urayama S Yamasaki S Hida A Tsuboi M Nakamura H Ida H Migita K Kawabe Y Eguchi K Inhibition of caspase cascade by HTLV-I tax through induction of NF-kappaB nuclear translocation Blood 1999 94 3847 3854 10572100
Siekevitz M Feinberg MB Holbrook N Wong-Staal F Greene WC Activation of interleukin 2 and interleukin 2 receptor (Tac) promoter expression by the trans-activator (tat) gene product of human T-cell leukemia virus, type I Proc Natl Acad Sci U S A 1987 84 5389 5393 3037548
Himes SR Coles LS Katsikeros R Lang RK Shannon MF HTLV-1 tax activation of the GM-CSF and G-CSF promoters requires the interaction of NF-kB with other transcription factor families Oncogene 1993 8 3189 3197 7504230
Azimi N Brown K Bamford RN Tagaya Y Siebenlist U Waldmann TA Human T cell lymphotropic virus type I Tax protein trans-activates interleukin 15 gene transcription through an NF-kappaB site Proc Natl Acad Sci U S A 1998 95 2452 2457 9482906 10.1073/pnas.95.5.2452
Waldele K Schneider G Ruckes T Grassmann R Interleukin-13 overexpression by tax transactivation: a potential autocrine stimulus in human T-cell leukemia virus-infected lymphocytes J Virol 2004 78 6081 6090 15163701 10.1128/JVI.78.12.6081-6090.2004
Inoue J Seiki M Taniguchi T Tsuru S Yoshida M Induction of interleukin 2 receptor gene expression by p40x encoded by human T-cell leukemia virus type 1 EMBO J 1986 5 2883 2888 3024966
Maruyama M Shibuya H Harada H Hatakeyama M Seiki M Fujita T Inoue J Yoshida M Taniguchi T Evidence for aberrant activation of the interleukin-2 autocrine loop by HTLV-1-encoded p40x and T3/Ti complex triggering Cell 1987 48 343 350 3026643 10.1016/0092-8674(87)90437-5
Cross SL Feinberg MB Wolf JB Holbrook NJ Wong-Staal F Leonard WJ Regulation of the human interleukin-2 receptor alpha chain promoter: activation of a nonfunctional promoter by the transactivator gene of HTLV-I Cell 1987 49 47 56 3030566 10.1016/0092-8674(87)90754-9
Mariner JM Lantz V Waldmann TA Azimi N Human T cell lymphotropic virus type I Tax activates IL-15R alpha gene expression through an NF-kappa B site J Immunol 2001 166 2602 2609 11160322
Mori N Fujii M Cheng G Ikeda S Yamasaki Y Yamada Y Tomonaga M Yamamoto N Human T-cell leukemia virus type I tax protein induces the expression of anti-apoptotic gene Bcl-xL in human T-cells through nuclear factor-kappaB and c-AMP responsive element binding protein pathways Virus Genes 2001 22 279 287 11450946 10.1023/A:1011158021749
Akagi T Ono H Shimotohno K Expression of cell-cycle regulatory genes in HTLV-I infected T-cell lines: possible involvement of Tax1 in the altered expression of cyclin D2, p18Ink4 and p21Waf1/Cip1/Sdi1 Oncogene 1996 12 1645 1652 8622884
Santiago F Clark E Chong S Molina C Mozafari F Mahieux R Fujii M Azimi N Kashanchi F Transcriptional up-regulation of the cyclin D2 gene and acquisition of new cyclin-dependent kinase partners in human T-cell leukemia virus type 1-infected cells J Virol 1999 73 9917 9927 10559304
Iwanaga R Ohtani K Hayashi T Nakamura M Molecular mechanism of cell cycle progression induced by the oncogene product Tax of human T-cell leukemia virus type I Oncogene 2001 20 2055 2067 11360190 10.1038/sj.onc.1204304
Mori N Fujii M Hinz M Nakayama K Yamada Y Ikeda S Yamasaki Y Kashanchi F Tanaka Y Tomonaga M Yamamoto N Activation of cyclin D1 and D2 promoters by human T-cell leukemia virus type I tax protein is associated with IL-2-independent growth of T cells Int J Cancer 2002 99 378 385 11992406 10.1002/ijc.10388
Fujii M Niki T Mori T Matsuda T Matsui M Nomura N Seiki M HTLV-1 Tax induces expression of various immediate early serum responsive genes Oncogene 1991 6 1023 1029 1906155
Yamaoka S Inoue H Sakurai M Sugiyama T Hazama M Yamada T Hatanaka M Constitutive activation of NF-kappa B is essential for transformation of rat fibroblasts by the human T-cell leukemia virus type I Tax protein EMBO J 1996 15 873 887 8631308
Akagi T Ono H Nyunoya H Shimotohno K Characterization of peripheral blood T-lymphocytes transduced with HTLV-I Tax mutants with different trans-activating phenotypes Oncogene 1997 14 2071 2078 9160887 10.1038/sj.onc.1201045
Robek MD Ratner L Immortalization of CD4(+) and CD8(+) T lymphocytes by human T-cell leukemia virus type 1 Tax mutants expressed in a functional molecular clone J Virol 1999 73 4856 4865 10233947
Hayden MS Ghosh S Signaling to NF-kappaB Genes Dev 2004 18 2195 2224 15371334 10.1101/gad.1228704
Bonizzi G Karin M The two NF-kappaB activation pathways and their role in innate and adaptive immunity Trends Immunol 2004 25 280 288 15145317 10.1016/j.it.2004.03.008
Chu ZL DiDonato JA Hawiger J Ballard DW The tax oncoprotein of human T-cell leukemia virus type 1 associates with and persistently activates IkappaB kinases containing IKKalpha and IKKbeta J Biol Chem 1998 273 15891 15894 9632633 10.1074/jbc.273.26.15891
Harhaj EW Good L Xiao G Uhlik M Cvijic ME Rivera-Walsh I Sun SC Somatic mutagenesis studies of NF-kappa B signaling in human T cells: evidence for an essential role of IKK gamma in NF-kappa B activation by T-cell costimulatory signals and HTLV-I Tax protein Oncogene 2000 19 1448 1456 10723136 10.1038/sj.onc.1203445
Sun SC Harhaj EW Xiao G Good L Activation of I-kappaB kinase by the HTLV type 1 Tax protein: mechanistic insights into the adaptor function of IKKgamma AIDS Res Hum Retroviruses 2000 16 1591 1596 11080796 10.1089/08892220050193001
Xiao G Cvijic ME Fong A Harhaj EW Uhlik MT Waterfield M Sun SC Retroviral oncoprotein Tax induces processing of NF-kappaB2/p100 in T cells: evidence for the involvement of IKKalpha EMBO J 2001 20 6805 6815 11726516 10.1093/emboj/20.23.6805
Mori N Fujii M Ikeda S Yamada Y Tomonaga M Ballard DW Yamamoto N Constitutive activation of NF-kappaB in primary adult T-cell leukemia cells Blood 1999 93 2360 2368 10090947
Mori N Yamada Y Ikeda S Yamasaki Y Tsukasaki K Tanaka Y Tomonaga M Yamamoto N Fujii M Bay 11-7082 inhibits transcription factor NF-kappaB and induces apoptosis of HTLV-I-infected T-cell lines and primary adult T-cell leukemia cells Blood 2002 100 1828 1834 12176906 10.1182/blood-2002-01-0151
Furukawa Y Kubota R Tara M Izumo S Osame M Existence of escape mutant in HTLV-I tax during the development of adult T-cell leukaemia Blood 2001 97 987 993 11159527 10.1182/blood.V97.4.987
Okazaki S Moriuchi R Yosizuka N Sugahara K Maeda T Jinnai I Tomonaga M Kamihira S Katamine S HTLV-1 proviruses encoding non-functional TAX in adult T-cell leukemia Virus Genes 2001 23 123 135 11724264 10.1023/A:1011840918149
Kinoshita T Shimoyama M Tobinai K Ito M Ito S Ikeda S Tajima K Shimotohno K Sugimura T Detection of mRNA for the tax1/rex1 gene of human T-cell leukemia virus type I in fresh peripheral blood mononuclear cells of adult T-cell leukemia patients and viral carriers by using the polymerase chain reaction Proc Natl Acad Sci U S A 1989 86 5620 5624 2787512
Furukawa Y Osame M Kubota R Tara M Yoshida M Human T-cell leukemia virus type-1 (HTLV-1) Tax is expressed at the same level in infected cells of HTLV-1-associated myelopathy or tropical spastic paraparesis patients as in asymptomatic carriers but at a lower level in adult T-cell leukemia cells Blood 1995 85 1865 1870 7703492
Schneider C Hubinger G Pleiotropic signal transduction mediated by human CD30: a member of the tumor necrosis factor receptor (TNFR) family Leuk Lymphoma 2002 43 1355 1366 12389614 10.1080/10428190290033288
Al-Shamkhani A The role of CD30 in the pathogenesis of haematopoietic malignancies Curr Opin Pharmacol 2004 4 355 359 15251128 10.1016/j.coph.2004.02.007
Horie R Watanabe T Morishita Y Ito K Ishida T Kanegae Y Saito I Higashihara M Mori S Kadin ME Ligand-independent signaling by overexpressed CD30 drives NF-kappaB activation in Hodgkin-Reed-Sternberg cells Oncogene 2002 21 2493 2503 11971184 10.1038/sj.onc.1205337
Suzuki T Hirai H Murakami T Yoshida M Tax protein of HTLV-1 destabilizes the complexes of NF-kappa B and I kappa B-alpha and induces nuclear translocation of NF-kappa B for transcriptional activation Oncogene 1995 10 1199 1207 7700645
Xiao G Harhaj EW Sun SC NF-kappaB-inducing kinase regulates the processing of NF-kappaB2 p100 Mol Cell 2001 7 401 409 11239468 10.1016/S1097-2765(01)00187-3
Xiao G Sun SC Negative regulation of the nuclear factor kappa B-inducing kinase by a cis-acting domain J Biol Chem 2000 275 21081 21085 10887201 10.1074/jbc.M002552200
Lee FS Peters RT Dang LC Maniatis T MEKK1 activates both IkappaB kinase alpha and IkappaB kinase beta Proc Natl Acad Sci U S A 1998 95 9319 9324 9689078 10.1073/pnas.95.16.9319
McCarthy JV Ni J Dixit VM RIP2 is a novel NF-kappaB-activating and cell death-inducing kinase J Biol Chem 1998 273 16968 16975 9642260 10.1074/jbc.273.27.16968
Chin AI Dempsey PW Bruhn K Miller JF Xu Y Cheng G Involvement of receptor-interacting protein 2 in innate and adaptive immune responses Nature 2002 416 190 194 11894097 10.1038/416190a
Kobayashi K Inohara N Hernandez LD Galan JE Nunez G Janeway CA Medzhitov R Flavell RA RICK/Rip2/CARDIAK mediates signalling for receptors of the innate and adaptive immune systems Nature 2002 416 194 199 11894098 10.1038/416194a
Ruefli-Brasse AA Lee WP Hurst S Dixit VM Rip2 participates in Bcl10 signaling and T-cell receptor-mediated NF-kappaB activation J Biol Chem 2004 279 1570 1574 14638696 10.1074/jbc.C300460200
Gruss HJ DaSilva N Hu ZB Uphoff CC Goodwin RG Drexler HG Expression and regulation of CD30 ligand and CD30 in human leukemia-lymphoma cell lines Leukemia 1994 8 2083 2094 7528856
Pfreundschuh M Pohl C Berenbeck C Schroeder J Jung W Schmits R Tschiersch A Diehl V Gause A Detection of a soluble form of the CD30 antigen in sera of patients with lymphoma, adult T-cell leukemia and infectious mononucleosis Int J Cancer 1990 45 869 874 2159438
Ohtsuka E Kikuchi H Nasu M Takita-Sonoda Y Fujii H Yokoyama S Clinicopathological features of adult T-cell leukemia with CD30 antigen expression Leuk Lymphoma 1994 15 303 310 7866279
Harlin H Podack E Boothby M Alegre ML TCR-independent CD30 signaling selectively induces IL-13 production via a TNF receptor-associated factor/p38 mitogen-activated protein kinase-dependent mechanism J Immunol 2002 169 2451 2459 12193714
Zheng B Fiumara P Li YV Georgakis G Snell V Younes M Vauthey JN Carbone A Younes A MEK/ERK pathway is aberrantly active in Hodgkin disease: a signaling pathway shared by CD30, CD40, and RANK that regulates cell proliferation and survival Blood 2003 102 1019 1027 12689928 10.1182/blood-2002-11-3507
Hironaka N Mochida K Mori N Maeda M Yamamoto N Yamaoka S Tax-independent constitutive IkappaB kinase activation in adult T-cell leukemia cells Neoplasia 2004 6 266 278 15153339
Horie R Watanabe T Ito K Morisita Y Watanabe M Ishida T Higashihara M Kadin M Cytoplasmic aggregation of TRAF2 and TRAF5 proteins in the Hodgkin-Reed-Sternberg cells Am J Pathol 2002 160 1647 1654 12000717
Sugamura K Fujii M Kannagi M Sakitani M Takeuchi M Hinuma Y Cell surface phenotypes and expression of viral antigens of various human cell lines carrying human T-cell leukemia virus Int J Cancer 1984 34 221 228 6088403
Yamada Y Nagata Y Kamihira S Tagawa M Ichimaru M Tomonaga M Shiku H IL-2-dependent ATL cell lines with phenotypes differing from the original leukemia cells Leuk Res 1991 15 619 625 1861543 10.1016/0145-2126(91)90031-N
Maeda T Yamada Y Moriuchi R Sugahara K Tsuruda K Joh T Atogami S Tsukasaki K Tomonaga M Kamihira S Fas gene mutation in the progression of adult T cell leukemia J Exp Med 1999 189 1063 1071 10190897 10.1084/jem.189.7.1063
Yamada Y Ohmoto Y Hata T Yamamura M Murata K Tsukasaki K Kohno T Chen Y Kamihira S Tomonaga M Features of the cytokines secreted by adult T cell leukemia (ATL) cells Leuk Lymphoma 1996 21 443 447 9172809
Morita S Kojima T Kitamura T Plat-E: an efficient and stable system for transient packaging of retroviruses Gene Ther 2000 7 1063 1066 10871756 10.1038/sj.gt.3301206
Kitamura T Onishi M Kinoshita S Shibuya A Miyajima A Nolan GP Efficient screening of retroviral cDNA expression libraries Proc Natl Acad Sci U S A 1995 92 9146 9150 7568090
Izumi KM Kaye KM Kieff ED The Epstein-Barr virus LMP1 amino acid sequence that engages tumor necrosis factor receptor associated factors is critical for primary B lymphocyte growth transformation Proc Natl Acad Sci U S A 1997 94 1447 1452 9037073 10.1073/pnas.94.4.1447
Horie R Ito K Tatewaki M Nagai M Aizawa S Higashihara M Ishida T Inoue J Takizawa H Watanabe T A variant CD30 protein lacking extracellular and transmembrane domains is induced in HL-60 by tetradecanoylphorbol acetate and is expressed in alveolar macrophages Blood 1996 88 2422 2432 8839832
Gattei V Degan M Gloghini A De Iuliis A Improta S Rossi FM Aldinucci D Perin V Serraino D Babare R Zagonel V Gruss HJ Carbone A Pinto A CD30 ligand is frequently expressed in human hematopoietic malignancies of myeloid and lymphoid origin Blood 1997 89 2048 2059 9058727
|
15876358
|
PMC1274245
|
CC BY
|
2021-01-04 16:36:39
|
no
|
Retrovirology. 2005 May 6; 2:29
|
utf-8
|
Retrovirology
| 2,005 |
10.1186/1742-4690-2-29
|
oa_comm
|
==== Front
BMC Cell BiolBMC Cell Biology1471-2121BioMed Central London 1471-2121-6-241586971510.1186/1471-2121-6-24Research ArticleGuanylic nucleotide starvation affects Saccharomyces cerevisiae mother-daughter separation and may be a signal for entry into quiescence Sagot Isabelle [email protected] Jacques [email protected] Bertrand [email protected] Institut de Biochimie et Génétique Cellulaires, UMR CNRS 5095 – Université Victor Segalen / Bordeaux II 1, rue Camille Saint Saëns – F-33077 Bordeaux Cedex – France2005 4 5 2005 6 24 24 18 2 2005 4 5 2005 Copyright © 2005 Sagot et al; licensee BioMed Central Ltd.2005Sagot et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Guanylic nucleotides are both macromolecules constituents and crucial regulators for a variety of cellular processes. Therefore, their intracellular concentration must be strictly controlled. Consistently both yeast and mammalian cells tightly correlate the transcription of genes encoding enzymes critical for guanylic nucleotides biosynthesis with the proliferation state of the cell population.
Results
To gain insight into the molecular relationships connecting intracellular guanylic nucleotide levels and cellular proliferation, we have studied the consequences of guanylic nucleotide limitation on Saccharomyces cerevisiae cell cycle progression. We first utilized mycophenolic acid, an immunosuppressive drug that specifically inhibits inosine monophosphate dehydrogenase, the enzyme catalyzing the first committed step in de novo GMP biosynthesis. To approach this system physiologically, we next developed yeast mutants for which the intracellular guanylic nucleotide pools can be modulated through changes of growth conditions. In both the pharmacological and genetic approaches, we found that guanylic nucleotide limitation generated a mother-daughter separation defect, characterized by cells with two unseparated daughters. We then showed that this separation defect resulted from cell wall perturbations but not from impaired cytokinesis. Importantly, cells with similar separation defects were found in a wild type untreated yeast population entering quiescence upon nutrient limitation.
Conclusion
Our results demonstrate that guanylic nucleotide limitation slows budding yeast cell cycle progression, with a severe pause in telophase. At the cellular level, guanylic nucleotide limitation causes the emergence of cells with two unseparated daughters. By fluorescence and electron microscopy, we demonstrate that this phenotype arises from defects in cell wall partition between mother and daughter cells. Because cells with two unseparated daughters are also observed in a wild type population entering quiescence, our results reinforce the hypothesis that guanylic nucleotide intracellular pools contribute to a signal regulating both cell proliferation and entry into quiescence.
==== Body
Background
Guanylic nucleotides are critical for multiple crucial cellular processes such as replication, transcription, translation and signalization via small GTPases. Most cell types recycle GMP from guanosine or guanine, which are either taken up from the surrounding environment or synthesized by the intracellular metabolism. GMP can also be synthesized de novo from IMP via two consecutive enzymatic steps, both highly conserved through evolution. Initially, the inosine monophosphate dehydrogenase (IMPDH) catalyzes conversion of IMP into XMP, the first committed step in de novo GMP biosynthesis. Subsequently, GMP synthetase converts the newly produced XMP into GMP.
The expression of IMPDH encoding genes is tightly regulated. In both yeast and mammalian cells, high guanylic nucleotide levels repress the transcription of IMPDH encoding genes [1,2]. More strikingly, the transcription of IMPDH encoding genes in mammals is linked to cellular proliferation. In non-dividing cells, the expression level of IMPDHs is low, whereas it is highly increased in actively proliferating cells such as cancerous cells [3-6]. Further, IMPDH over-expression bypasses the anti-proliferative effect of p53, indicating that this p53 function requires proper control of IMPDH activity [7].
Saccharomyces cerevisiae cells have two major sources of guanylic nucleotides: intracellular IMP and extracellular guanine. Therefore GMP synthesis fully relies on IMPDH and GMP synthetase activities in the absence of guanine in the growth medium. Consistently, mutations in these cognate genes lead to guanine auxotrophy [8,9]. Similar to mammalian cells, S. cerevisiae genes involved in GMP biosynthesis are highly expressed during exponential growth, but are actively repressed through specific regulatory sequences when cells arrest proliferation upon nutrient limitation [10]. Moreover, the intracellular GTP/GDP ratio drastically decreases when yeast cells enter a quiescent state [11]. These experiments lead to the appealing hypothesis that intracellular guanylic nucleotides levels contribute to a signal regulating cell proliferation. However, the molecular pathways linking cell cycle progression to IMPDH activity, and thus to intracellular guanylic nucleotide pools remain unknown.
Mycophenolic acid (MPA) is a well-characterized non-competitive and reversible inhibitor of IMPDH that severely depletes intracellular GTP pool down to 10% of normal level [12,13]. MPA reduces or even abrogates proliferation of various cell types [14,15]. MPA particularly affects lymphocyte division and because it has few secondary effects, its pro-drug form, the mycophenolate mofetil (Cellcept, Roche), is in path to replace cyclosporine A as a commonly used immunosuppressive drug. At the cellular level, MPA inhibits lymphocyte cell cycle progression by arresting cells in G1. Although this effect correlates with the depletion of guanylic nucleotide pools [13,16], addition of guanosine and 8-aminoguanosine, which can partially replenish guanylic nucleotide pools, does not allow MPA treated cells to re-enter G2/M [17]. Thus, MPA treatment, although apparently blocking cells in G1, could also affect later steps of the cell cycle. In budding yeast, MPA treatment slows cell proliferation and causes various effects on gene expression and thus on the yeast proteome [18]. In a previous study, we have shown that MPA affects yeast cell size, DNA content, budding pattern and causes occasional perturbations of actin and microtubule cytoskeletons [18]. In addition, several mutants affected for various cellular functions are hypersensitive to MPA [19]. However, these data do not point at an obvious molecular process that would account for the effects of MPA on cell proliferation.
Here, to gain insight into the molecular relationships between intracellular guanylic nucleotide levels and cell cycle progression, we studied the effects of MPA treatment on Saccharomyces cerevisiae cell cycle progression. We first demonstrated that, although cells did not arrest in a particular cell cycle stage (confirming that there was no checkpoint for guanylic nucleotides in yeast), a large proportion of the population was slowed in telophase. We further observed that many MPA treated cells presented two unseparated daughter cells. We have shown that this specific morphology was due to a defect in mother-daughter separation and that it was probably a consequence of cell wall perturbations. To validate the results obtained with MPA, we developed yeast mutants in which guanylic nucleotide pools could be modulated by the composition of the growth medium. Using these genetic tools, we confirmed that cell separation was indeed the cell cycle step mostly perturbed by guanylic nucleotides starvation. Finally, the observation that cells entering quiescence also displayed the characteristic "two-daughter cells" morphology strongly suggests that a decrease in intracellular guanylic nucleotide levels may be part of a signal for yeast cells to enter stationary phase.
Results
Mycophenolic acid treatment particularly affects the last step of the yeast cell cycle
MPA treatment affects yeast growth in a concentration-dependent manner [18]. This growth defect can result from either a general slowing down of the entire cell cycle or a pause in a specific cell cycle step. To gain insight into this issue, wild type yeast cells were synchronized in G1 using alpha factor and released in either the absence or the presence of 100 μg/mL of MPA. At such a concentration, MPA does not affect cell viability nor totally arrest cell growth [18]. Higher MPA concentration gave similar effects, most probably because yeast cells detoxified the drug [19]. We then monitored cell cycle progression by FACS. As shown in figure 2A, although progression through initial cell cycle steps was slower for MPA treated cells, the population lagged most predominantly in a 2N DNA content stage. Fluorescence microscopy revealed that although for MPA-treated cells the anaphase onset was slightly delayed, its duration was almost similar to the anaphase of control cells (Fig. 2B). In fact, treated cells were mostly pausing in a stage where the DNA masses were totally separated (telophase, Fig. 2C). To our surprise, in the treated population we observed mother cells with two apparent daughter cells (Fig. 2D and 2E). The cells with two apparent daughters represented more than 50% of the population 300 minutes after the release from G1. To confirm this result, a non-synchronized population was treated with MPA. After 4 hours in the presence of MPA, more than 30% of the budding cells displayed two apparent daughters (Fig. 3), a phenotype we refer to as "bibudded", for simplicity.
Figure 2 Effects of MPA on the yeast cell cycle progression. A. Cell cycle progression of a yeast cell population synchronized in G1 with alpha factor and released in the absence (left) or in the presence of 100 μg/mL MPA (right) analyzed by FACS. B. Percentage of cells in anaphase in function of the progression through the cell cycle. Cells are the same as in A. More than 200 cells were counted for each time point. Cells were scored as being in anaphase when the mother cell DNA mass was clearly still connected to the DNA mass of the daughter cell, typically as the third cell shown in the – MPA panel of figure 2E. C. Percentage of cells in telophase in function of the progression through the cell cycle. Cells are the same as in A. More than 200 cells were counted for each time point. Cells were scored as being in telophase when displaying two clearly separated DNA masses, typically as the fourth cell shown in the – MPA panel of figure 2E. D. Percentage of cells with two or more daughter cells in function of the progression through the cell cycle. Cells are the same as in A. More than 200 cells were counted for each time point. E. Cells representative of each cell cycle stage (phase contrast and propidium iodide staining of the nucleus) for the untreated (top panel) or MPA treated (bottom panel) population. The unusual elongated cell morphology is due to the alpha factor treatment. Arrows indicate examples of second daughter cell appearing while the first daughter cell is not yet separated from the mother cell. Bar: 2 μm.
Figure 3 Effects of MPA treatment on an unsynchronized yeast cell population. Percentage of budding cells with two daughter cells after 4 hours growth in SD medium without (left) or with guanine (right) in the absence (grey bars) or in the presence (white bars) of 100 μg/mL MPA. More than 200 cells were counted for each condition.
A "bibudded" phenotype could result either from a single mother cell with two daughter cells, or from an unrelated G1 cell "sticking" to a normally budding cell. Since MPA treated cells were found to be sensitive to sonication, we used a fluorescence-based approach to distinguish between these two possibilities. Equal amounts of cells over-expressing a green variant of GFP were mixed with cells over-expressing a blue variant of GFP, grown to OD600 nm 0.2 and then treated with MPA.
After 4 hours of incubation, less than 1% (0.3% ± 0.3, N>100 for each of 4 independent experiments) of "bibudded" cells displayed one apparent daughter expressing a different GFP variant than its joined budding cell. Thus, more than 99% of the cells with two apparent daughters were monocolor, which is far more than the 50% expected for a random population of false "bibudded" cells. This experiment demonstrated that the population is indeed composed of a large proportion of cells with two attached daughters. In conclusion, MPA treatment particularly slowed down telophase and caused the appearance of cells with two unseparated daughters.
MPA effects are reversed by extracellular guanine and are not a consequence of translation inhibition
To generate guanylic nucleotides, the requirement for IMPDH activity can be bypassed by the addition of guanine into the growth medium (Fig. 1). To show that the effects of MPA on cell cycle progression specifically result from a decrease of intracellular guanylic nucleotide pools, we treated cells with MPA in guanine-supplemented growth medium. In these conditions, only 3% of the cells displayed two unseparated daughters (Fig. 3) and no growth defect was detected (data not shown and [18]).
Figure 1 Schematic representation of the purine nucleotide synthesis pathway in yeast. The solid thin lines represent the plasma membrane. ADP: adenosine-5'-diphosphate; AMP: adenosine-5'-monophosphate; ATP: adenosine-5'-triphosphate; GDP: guanosine-5'-diphosphate; GMP: guanosine-5'-monophosphate; GTP: guanosine-5'-triphosphate; IMP: inosine-5'-monophosphate; PRPP: 5-phosphoribosyl-1-pyrophosphate; SAMP: S-adenosine-5'-monophosphate; XMP: xanthosine-5'-monophosphate. Genes are shown in grey italic font and encode the following enzymatic activities: AAH1: adenine deaminase; ADE8: 5'-phosphoribosylglycinamide formyltransferase; ADE12: adenylosuccinate synthetase; ADE13: adenylosuccinate lyase; ADK1: AMP kinase; AMD1: AMP deaminase; APT1: adenine phosphoribosyltransferase; FCY2: purine cytosine permease; GUA1: GMP synthetase; GUK1: GMP kinase; HPT1: hypoxanthine-guanine phosphoribosyltransferase; IMD2, IMD3, and IMD4: IMP dehydrogenases (IMD1 is not indicated here because it is not expressed and is thought to be a pseudogene). Mycophenolic acid (MPA) inhibits IMP dehydrogenases.
As 100 μg/mL MPA causes a significant decrease of the translation efficiency [18], we considered the possibility that "bibudded" cells result from impaired translation. To this end, we examined cells treated for 4 hours with the general translational inhibitor cycloheximide at a concentration that abolished translation [20]. In contrast to MPA-treated cells, cycloheximide-treated cells did not exhibit more than one bud (0.8% ± 0.3% of the budding cells had two buds, N>200). Therefore, by decreasing intracellular guanylic nucleotide pools, MPA treatment caused a slowing down of the yeast cell cycle progression, the telophase being mostly affected. Consequently, "bibudded" cells accumulate and this, independently of MPA effects on translation efficiency.
Mutations affecting the guanylic nucleotide biosynthesis cause the emergence of cells with two unseparated daughters
Although addition of guanine to the growth medium fully reversed the effects of MPA, pharmacological studies have the caveat of possible secondary targets. Thus, we developed mutant yeast strains in which the guanylic nucleotide pools can be modulated by a simple change of the growth conditions. Because of genetic redundancy in IMPDH encoding genes, no single imd mutant is auxotroph for guanine, we therefore chose to use a mutant in the GUA1 gene which encodes GMP synthetase, the enzyme converting XMP into GMP (Fig. 1). A gua1Δ mutant is unable to synthesize guanylic nucleotides in the absence of guanine in the growth medium. To compare the effects of guanylic nucleotide starvation with the depletion of another purine, we combined the gua1Δ deletion with the ade8Δ deletion that leads to adenine auxotrophy. The double mutant ade8Δ gua1Δ is thus auxotroph for both guanine and adenine (Fig. 4A). When both guanine and adenine were provided in the growth medium, the double mutant ade8Δ gua1Δ grew like the isogenic ade8Δ control strain and, the ade8Δ gua1Δ population contained less than 3% of "bibudded" cells (Fig. 4A and 4B). After a 4-hour shift to a growth medium lacking guanine, 20% of the cells displayed two unseparated daughter cells. By contrast, shifting the cells to growth medium lacking adenine did not induce the emergence of "bibudded" cells (Fig. 4B). Therefore, "bibudded" cells appearance is specific to a guanylic nucleotide starvation.
Figure 4 Effects of guanylic nucleotide starvation using mutants of the purine nucleotide biosynthesis pathway. A. Growth of an ade8Δ gua1Δ double mutant and an isogenic ade8Δ single mutant on SD medium containing the indicated purines. The SD + ADE + GUA medium contains 25% Adenine / 75% Guanine ratio (total purine concentration of 0.3 mM). B. The ade8Δ gua1Δ strains was grown in SD medium containing 25% Adenine / 75% Guanine ratio to OD600 nm = 0.2; then shifted into the indicated medium. After 4 hours of incubation at 30°C, the percentage of cells with two daughters among budding cells was counted. More than 200 cells were counted for each condition. C. Growth of the WT strain, isogenic single amd1Δ or aah1Δ mutant strains and amd1Δ aah1Δ double mutant strain on SD medium, SD medium supplemented with adenine or SD medium supplemented with hypoxanthine. D. WT strain (left) and amd1Δ aah1Δ double mutant strain (right) were grown in SD medium to OD600 nm = 0.2 and then shifted into the indicated medium. After 4 hours of incubation at 30°C, the percentage of cells with two daughters among budding cells was counted. More than 200 cells were counted for each condition.
In order to diminish the intracellular guanylic nucleotide pools by another route, we constructed an aah1Δ amd1Δ double mutant. The AAH1 gene encodes adenine amino-hydrolase, an enzyme converting adenine into hypoxanthine. The AMD1 gene encodes the AMP deaminase enzyme, which converts AMP into IMP (Fig. 1). The de novo synthesis of IMP from PRPP is inhibited by the presence of adenine in the growth medium [21]. Consequently, in the presence of adenine, the aah1Δ amd1Δ double mutant has theoretically no path to synthesize GMP, neither from the de novo pathway nor from adenine. Accordingly, without extracellular guanine (or hypoxanthine), the guanylic nucleotide pools of the aah1Δ amd1Δ double mutant cannot be replenished and its growth is strongly affected (Fig. 4C). The aah1Δ amd1Δ double mutant was grown in SD medium without purine to OD600 nm = 0.2 and then shifted into a medium containing adenine as a sole source of purine. After 4 hours, the percentage of cells with two unseparated buds was counted. As shown in figure 4D, "bibudded" cells only arose when the aah1Δ amd1Δ double mutant cells were grown in a medium containing adenine. Thus, modifying the guanylic nucleotide pools through mutations in the purine nucleotide biosynthesis pathway provoked the same effect that MPA treatment on the yeast cell cycle: drastically impaired cell separation without cell cycle arrest, resulting in cells with two unseparated daughters.
Guanylic nucleotide starvation does not affect the formation of cellular structures required for completion of cytokinesis
The decrease of intracellular guanylic nucleotide pools caused the appearance of cells with two unseparated buds. We further characterized these abnormal cells to identify the cellular process(es) critically impaired by this starvation and thus likely responsible for this particular phenotype.
When synchronized cells were treated with MPA, the timing of emergence and the size of the second bud compared to the first one (see Fig. 2E) strongly suggested that MPA treated cells started a new cell cycle before the separation of the first daughter cell. Cells with two daughters follow a normal budding pattern (Fig. 2E and 5A), and are morphologically very different from the multibudded cells observed for polarity mutants impaired in bud emergence. Thus, the decrease of intracellular guanylic nucleotide pools did not likely affect cell polarity establishment. Further, most of the cells with two buds presented wild type, polarized actin patches and cables (Fig. 5A, middle lane), confirming previous studies showing that MPA treatment does not drastically affect the actin cytoskeleton [18]. In addition, nuclei of the "bibudded" cells are properly positioned (Fig. 5A, bottom lane) although abnormal mitosis occurs in less than 5% of the cells (Fig. 5A, + MPA right panel). Thus, polarization establishment and nuclear segregation were not drastically affected in cells with two unseparated daughters.
Figure 5 Localization of several cellular structures or fusion proteins in cells with two daughters. A. Cells were untreated (right panel) or treated (left panel) with 100 μg/mL MPA for 4 hours. DIC (top lane), actin Alexa-phalloidin staining (middle lane) and DAPI staining (bottom lane) are shown. B. Localization of endogenous Tem1p-GFP (left panel), endogenous Lte1p-GFP (middle panel) and GFP-Cdc12p expressed from a centromeric plasmid under the control of its own promoter (right panel) in cells untreated (right of each panel) or treated for 4 hours with 100 μg/mL MPA (left of each panel). Only cells displaying two buds are shown in the case of MPA treatment. Bar: 2 μm.
Microscopic observation of synchronized MPA-treated cells revealed that although anaphase on-set was delayed, its duration was almost similar to the anaphase in control cells (Fig. 2B). Therefore, MPA treatment did not critically affect the mitotic exit network (MEN). Further, MPA treatment did not disrupt proper localization of Tem1p, a RAS GTPase governing the MEN, neither the localization of its GTP exchange factor (GEF) Lte1p (Fig. 5B, left and middle panels). Taken together, these results suggest that MPA treatment does not significantly perturb anaphase progression, and consequently does not activate the anaphase checkpoint.
Because MPA severely affected progression through telophase, we examined the formation of cellular structures essential for the completion of cytokinesis. Figure 5A shows that both the first and the second daughter cell can build an actin cytokinetic ring. Further, more than 95% of the MPA treated cells (N>200) maintain proper localization of the septin Cdc12p (Fig. 5B, right panel), suggesting strongly that MPA-induced guanylic nucleotide pools depletion does not affect the septin ring formation and stability. Moreover, in "bibudded" cells, the septin ring can split to allow acto-myosin ring contraction for both the first and the second daughter cell (Fig. 5B, right panel). Therefore, guanylic nucleotide starved cells displayed all the structures required for cytokinesis completion, and the cellular processes leading to the mother-daughter separation defect must occur later in the cell cycle.
Guanylic nucleotide starvation affects cell wall separation
As cells with two unseparated daughters properly form both actin and septin rings, we examined whether the cytoplasm was still continuous between the first bud and its mother. After a 4 hour MPA treatment, we digested the cell wall with zymolyase and counted the number of remaining "bibudded" cells. Figure 6A shows that mild treatment with zymolyase cause a decrease of the "bibudded" cell number suggesting strongly that in those cells, the cytoplasm is no more continuous between the mother cell and at least one daughter cell. Thus, guanylic nucleotide depletion apparently did not affect cytoplasm constriction but rather a later step in the daughter cell separation process. To confirm this result, we observed MPA-treated cells by electron microscopy after Thiéry coloration, which reveals polysaccharides and therefore the cell wall. As shown in figure 6B, the cytoplasm between the mother and the first daughter cell is no longer continuous. Thus, guanylic nucleotide starvation did not affect cytoplasm closure but a later step in the daughter cell separation process. Further, the secondary septum of the first daughter cell appears normal by electron microscopy, although some small lacunae were occasionally observed (see inset of Fig. 6B). Therefore, we speculate that reduction of intracellular guanylic nucleotide pools by MPA treatment particularly impinges on the separation of the daughter cell by affecting the cell wall digestion.
Figure 6 Effects of MPA on the yeast cell wall. A. Percentage of unbudded (shadow bars), single budded (white bars) and of cells with two buds (grey bars) after 4 hours treatment with 100 μg/mL MPA followed by a mild digestion of the cell wall with zymolyase. More than 200 cells were counted for each condition. B. Electron microscopy pictures of cells with two buds obtained after 4 hours treatment with 100 μg/mL MPA. Untreated control cells are shown on the right panel each steps of the septum formation are illustrated (from top to bottom).
A large proportion of cells entering quiescence exhibit two daughters
In the course of our study, we noticed that a small but reproducible amount of untreated wild type cells displayed two unseparated buds (see for example Fig. 3), a morphology identical to MPA-treated "bibudded" cells. We then examined the frequency of "bibudded" cells during the growth of a wild type yeast population. To our surprise, whereas less than 2% of the cells with two unseparated buds could be observed during exponential phase, when yeast approached the diauxic shift, more than 30% of the budding population presented two apparent daughter cells (Fig. 7). This observation suggested a slowing of daughter cell separation upon the last divisions before stationary phase. Like previously, we verified the authenticity of the "bibudded" phenotype by mixing two populations of cells, each expressing one different GFP variant. When this mixed population reached the diauxic shift, we counted the number of cells with two apparent daughters and examined them by fluorescence microscopy. Similar to our results for MPA-treated cells, only 0.9% ± 0.7% of the apparent "bibudded" cells were bicolor (more than 200 "bibudded" cells were counted for each experiment). Thus, we concluded that those cells were indeed cells with two daughters. This was further supported by observation of those cells by electron microscopy (not shown). We obtained similar results with both the BY4742 and the FL100 genetic background (data not shown) demonstrating that the appearance of cells with two daughters is not specific to the BY4742 background. In conclusion, when a yeast culture approaches stationary phase, a significant proportion of cells behaves like cells in which the intracellular guanylic nucleotide pools have been depleted and give rise to cells with two unseparated daughters.
Figure 7 Number of cells with two daughters as a function of the age of the culture. WT cells were grown in SC medium and the percentage of cells with two daughters among budding cells as a function of time was counted (black circles). For each time point, more than 300 cells were counted. OD600 nm is indicated as a black bold line.
Discussion
Guanylic nucleotides are not only "building blocks" for nucleic acids but are also crucial for the regulation of many cellular processes such as G-proteins based signaling pathways. Therefore, cells must maintain their concentrations at a critical level. However, the molecular relationships between intracellular guanylic nucleotide levels and cell proliferation crucial events remain poorly understood.
Here, we have demonstrated that MPA treatment does not cause a firm cell cycle arrest in yeast. Treated cells continue to proliferate, although at a reduced rate (this study and [18]). Further, we have shown that conditional mutants unable to synthesize guanylic nucleotides do not arrest in a particular stage of the cell cycle. Therefore, our results establish that there is no guanylic nucleotide checkpoint is S. cerevisiae. By contrast, in mammalian cells, it was previously shown that MPA treatment cause an arrest of cellular proliferation but no guanylic nucleotide specific checkpoint has clearly been identified. Several reports have described that MPA treatment affects mammalian cells ability to commit into division by blocking the transition from G0 to the S phase of the cell cycle [13,17,22]. However, if MPA is added when cells have already entered the S phase, the cell cycle arrest occurs in G2/M [22]. Besides, the MPA-induced arrest is not fully reversed by the replenishment of guanylic nucleotide pools [13,17,22]. Here, we have shown that even if MPA treatment slows all stages of the yeast cell cycle progression, the most affected step is the mother-daughter cell wall separation, giving rise to "bibudded" cells. This result supports our previous observation that MPA treatment leads to the appearance of many cells with 3N DNA content [18]. Importantly, MPA treatment has the same effects on both non-synchronized and synchronized cells treated with the drug upon release from G1 (Fig. 2) or upon release from G2/M after nocodazole synchronization (I. S., B. D.-F. unpublished results). Therefore, unlike in mammals, in yeast, MPA treatment causes the same effects whatever the cell cycle stage of the cells at the time of drug addition. Further, our analysis of mutants demonstrates that the mother-daughter separation defect results solely from guanylic nucleotide pools depletion and is independent of potential MPA secondary targets.
In yeast, the fact that a guanylic nucleotides starvation causes a mother-daughter separation defect was unexpected. Indeed, one could have intuitively supposed that this depletion would rather provoke a drastic defect during DNA replication upon S phase or disturb cell cycle steps for which GTPase driven molecular processes are essential. In fact, upon MPA treatment, cells can still build a daughter cell and cell polarity was found even less affected in this study than in our earlier work [18]. Daughter cell appears to grow normally, mitosis proceeds unperturbed, and mother-daughter closure is properly achieved. Thus, guanylic nucleotides starvation does not critically affect the functions of key GTP binding proteins, such as Cdc42p, Tem1p, tubulin and septins. In this study we observed that in the "bibudded" cells, placement of the second daughter cell properly follows the axial budding pattern of haploid cells, suggesting that the bud-site selection machinery is properly located. In contrast, prolonged MPA treatment (48 hours) leads to a random budding pattern [18]. Thus, long-term guanylic nucleotides starvation may have more drastic effects. Most importantly, our experiments show that complete mother-daughter separation is not required for the mother cell to pass through START and to generate a second daughter cell (Fig. 2E). Therefore, our data confirm that no additional checkpoint blocks the cell cycle progression when anaphase is properly achieved.
What molecular targets trigger the mother-daughter separation defect upon guanylic nucleotide starvation? Observation of "bibudded" cells by electron microscopy revealed no obvious defect in the overall septum architecture of abnormally unseparated daughter cell. Nevertheless, MPA treatment increases cells sensitivity to SDS [19], zymolyase or sonication (I. S., B. D.-F. unpublished results), strongly suggesting cell wall defects. In fact, previous works have illustrated links between guanylic nucleotides metabolism and cell wall integrity, particularly through the synthesis of mannoproteins, essential components of the fungal cell wall. GDP-mannose is the common substrate for mannosyltransferases, enzymes catalyzing the addition of mannose residues on mannoproteins core oligosaccharides. In budding yeast, the GDP-mannose pyrophosphorylase Psa1p is an essential enzyme that synthesizes GDP-mannose, using GTP as a substrate. It was previously shown that MPA treatment affects Psa1p expression [18] and that Psa1p depletion leads to cell separation failure [23]. Further, Shimma et al have demonstrated that the major defect of a guk1 conditional mutant strain, that is impaired for GDP biosynthesis (Fig. 1), was a decrease in GDP-mannose level (to about 25% of the wild type levels) that leads to mannose outer chain elongation defects [24]. Accordingly, we have observed cell separation defect in guk1 cells (I. S., B. D.-F., unpublished results). In addition, yeast lacking mannosyltransferase encoding genes, such as OCH1, MNN10 or ANP1 display both hypersensitivity to MPA [19] and a cell separation defect similar to the one observed for guanylic nucleotide starved cells [25-27]. Therefore, one major consequence of guanylic nucleotide starvation could be a significant decrease in the GDP-mannose pool that in turn leads to a mother-daughter separation defect.
The most intriguing aspect of the regulation of intracellular guanylic nucleotide pools is the correlation between the IMPDH activity and cellular proliferation in mammalians models. Interestingly, the transcription of the IMD2 gene is actively shut off via regulatory sequences when yeast cells enter stationary phase upon nutrients limitation [2]. Transcriptome analyses have demonstrated that AAH1, HPT1 and GUA1 are among the most promptly down regulated genes when nutrients become limiting. Thus, it appears that an entire process is devoted to rapidly decrease the intracellular guanylic nucleotide pools when cells enter stationary phase. Here, we have demonstrated that the typical "bibudded" phenotype obtained during guanylic nucleotide starvation also occurs in untreated wild type cells achieving their last divisions upon nutrients limitation. Thus, an identical morphology is observed for both guanylic nucleotide starvation and entry into quiescence.
Finally, the GTP/GDP ratio is very sensitive to growth conditions, rapidly decreasing during the diauxic shift and drastically dropping upon nutrients starvation. Further, this ratio may regulate RAS GTPases activity by influencing its guanylic nucleotide loading equilibrium [11]. RAS and TOR pathways are key regulators that coordinate yeast proliferation with nutrients availability. Recent work has suggested that the TOR protein acts as ATP sensor in mammals [28]. Thus it is appealing to speculate that in parallel to the TOR pathway, intracellular guanylic nucleotides levels are part of a signal that regulate cell proliferation via the modulation of RAS GTPases activity.
Conclusion
Using either mycophenolic acid, a molecule that specifically inhibits the first committed step in de novo GMP biosynthesis or mutations in the guanylic nucleotide biosynthesis pathway, we have demonstrated that intracellular guanylic nucleotides limitation causes a mother-daughter cell wall separation defect in budding yeast. This defect leads to the emergence of cells with two unseparated daughters. These "bibudded" cells are also found in a population of cells entering quiescence upon nutrient limitation. These observations further suggest that guanylic nucleotide intracellular pools might contribute to a signal that regulates cell proliferation, particularly upon nutrients limitation and entry into stationary phase.
Methods
Strains, media and reagents
Yeast strains used were purchased from Euroscarf (Frankfurt, Germany) and are derivatives of BY4741 or BY4742 [29]. SD and SD casa media were described previously [2] and were supplemented when required with tryptophan (0.2 mM), uracil (1.8 mM), guanine (0.3 mM), adenine (0.3 mM), hypoxanthine (0.3 mM). Cycloheximide was purchased from Sigma and used at 50 μg/mL. Mycophenolic acid (MPA) was purchased from Amresco, (Ohio, USA) and was used at 100 μg/mL.
Plasmids
pB1594 integrates three tandem copies of GFP at the 3' end of the LTE1 coding region (Lte1p-3xGFP) and is a kind gift of D. Pellman [30]. pB1598 integrates three tandem copies of GFP at the 3' end of the TEM1 coding region (Tem1p-3xGFP) and is a kind gift of D. Pellman. pYB407 (GFP-Cdc12p – LEU2 – CEN) is a kind gift of Y. Barral [31]. pVTYS65T (URA3, 2μ) allows the expression of the GFP variant containing the S65T mutant (fluorescence maxima: Ex: 488 nm, Em: 511 nm) and has been previously described [32]. pVTYBFP2 allows the expression of a GFP variant containing mutations F64L, S65T, Y66H et Y145F which coding sequence has been optimized for expression in yeast (fluorescence maxima: Ex 380 nm, Em 440 nm see [33]). Details of the mutagenesis and constructions are available upon request.
Cell Biology techniques and Fluorescence Microscopy
Cells were synchronized by addition of alpha-factor at the final concentration of 10 μg/mL directly into the growth medium. After 3 hours of incubation at 30°C with agitation, more than 99% of the cells were unbudded. Cells were washed twice and released into fresh SD medium containing or not MPA. Propidium iodide (Sigma) staining was performed as described in [34] except that cells were fixed with ethanol 70%. Alexa-568-Phalloidin (Molecular Probes, Eugene, OR) was used to stain filamentous actin as described previously [35]. Zymolyase 20T (ICN Biomedicals, Costa Mesa, CA) was used at 0.2 mg/mL on formaldehyde fixed cells resuspended in PBS. Images were acquired with a Marianas system and analyzed with the Slidebook software (Intelligent Imaging Innovations, Inc. Denver, CO) except for GFP-Cdc12p expressing cells that were imaged with a previously described imaging system [32].
Electron Microscopy
Cells were grown in SD medium to OD600 nm = 0.2 at 30°C. MPA was added to the final concentration of 100 μg/mL and cells were grown for four more hours. Cells were then fixed with glutaraldehyde and osmic acid, and then stained by the method of Thiéry as previously described [36].
Authors' contributions
JS carried out the electron microscopy. IS did the other experimental work. BDF and IS conceived and coordinated the studies and drafted the manuscript. All authors read and approved the final manuscript.
Acknowledgements
The authors thank Dr. Y. Barral and Dr. D. Pellman for kindly providing us plasmids and other reagents. We thank C. Saint-Marc for technical assistance in constructing and characterizing the ade8Δ gua1Δ and amd1Δ aah1Δ double mutants. We specially want to express our gratitude to Dr. D. Pellman for having kindly allowed us to use his microscope and to Dr. Z. Storchova for her expertise and precious help for FACs experiments. We also acknowledge B. Pinson and A. Breton for helpful discussion and comments on the manuscript. We are very grateful to J. Moseley and E. Coic for their help in proofreading the manuscript. This work was supported by Grant # 4749 from The Association pour la Recherche Contre le Cancer.
==== Refs
Glesne DA Collart FR Huberman E Regulation of IMP dehydrogenase gene expression by its end products, guanine nucleotides Mol Cell Biol 1991 11 5417 25 1717828
Escobar-Henriques M Daignan-Fornier B Transcriptional regulation of the yeast gmp synthesis pathway by its end products J Biol Chem 2001 276 1523 30 11035032 10.1074/jbc.M007926200
Konno Y Natsumeda Y Nagai M Yamaji Y Ohno S Suzuki K Weber G Expression of human IMP dehydrogenase types I and II in Escherichia coli and distribution in human normal lymphocytes and leukemic cell lines J Biol Chem 1991 266 506 9 1670768
Nagai M Natsumeda Y Weber G Proliferation-linked regulation of type II IMP dehydrogenase gene in human normal lymphocytes and HL-60 leukemic cells Cancer Res 1992 52 258 61 1345808
Zimmermann AG Gu JJ Laliberte J Mitchell BS Inosine-5'-monophosphate dehydrogenase: regulation of expression and role in cellular proliferation and T lymphocyte activation Prog Nucleic Acid Res Mol Biol 1998 61 181 209 9752721
Collart FR Chubb CB Mirkin BL Huberman E Increased inosine-5'-phosphate dehydrogenase gene expression in solid tumor tissues and tumor cell lines Cancer Res 1992 52 5826 8 1356621
Liu Y Bohn SA Sherley JL Inosine-5'-monophosphate dehydrogenase is a rate-determining factor for p53-dependent growth regulation Mol Biol Cell 1998 9 15 28 9436988
Dujardin G Kermorgant M Slonimski PP Boucherie H Cloning and sequencing of the GMP synthetase-encoding gene of Saccharomyces cerevisiae Gene 1994 139 127 32 8112582 10.1016/0378-1119(94)90535-5
Hyle JW Shaw RJ Reines D Functional distinctions between IMP dehydrogenase genes in providing mycophenolate resistance and guanine prototrophy to yeast J Biol Chem 2003 278 28470 8 12746440 10.1074/jbc.M303736200
Escobar-Henriques M Collart MA Daignan-Fornier B Transcription initiation of the yeast IMD2 gene is abolished in response to nutrient limitation through a sequence in its coding region Mol Cell Biol 2003 23 6279 90 12917348 10.1128/MCB.23.17.6279-6290.2003
Rudoni S Colombo S Coccetti P Martegani E Role of guanine nucleotides in the regulation of the Ras/cAMP pathway in Saccharomycescerevisiae Biochim Biophys Acta 2001 1538 181 9 11336789 10.1016/S0167-4889(01)00067-2
Qiu Y Fairbanks LD Ruckermann K Hawrlowicz CM Richards DF Kirschbaum B Simmonds HA Mycophenolic acid-induced GTP depletion also affects ATP and pyrimidine synthesis in mitogen-stimulated primary human T-lymphocytes Transplantation 2000 69 890 7 10755546 10.1097/00007890-200003150-00038
Messina E Gazzaniga P Micheli V Barile L Lupi F Agliano AM Giacomello A Low levels of mycophenolic acid induce differentiation of human neuroblastoma cell lines Int J Cancer 2004 112 352 354 15352052 10.1002/ijc.20425
Yalowitz JA Jayaram HN Molecular targets of guanine nucleotides in differentiation, proliferation and apoptosis Anticancer Res 2000 20 2329 38 10953293
Morath C Zeier M Review of the antiproliferative properties of mycophenolate mofetil in non-immune cells Int J Clin Pharmacol Ther 2003 41 465 9 14703952
Allison AC Eugui EM Mycophenolate mofetil and its mechanisms of action Immunopharmacology 2000 47 85 118 10878285 10.1016/S0162-3109(00)00188-0
Laliberte J Yee A Xiong Y Mitchell BS Effects of guanine nucleotide depletion on cell cycle progression in human T lymphocytes Blood 1998 91 2896 904 9531600
Escobar-Henriques M Balguerie A Monribot C Boucherie H Daignan-Fornier B Proteome analysis and morphological studies reveal multiple effects of the immunosuppressive drug mycophenolic acid specifically resulting from guanylic nucleotide depletion J Biol Chem 2001 276 46237 42 11535588 10.1074/jbc.M103416200
Desmoucelles C Pinson B Saint-Marc C Daignan-Fornier B Screening the yeast "disruptome" for mutants affecting resistance to the immunosuppressive drug, mycophenolic acid J Biol Chem 2002 277 27036 44 12016207 10.1074/jbc.M111433200
Clark-Walker GD Linnane AW In vivo differentiation of yeast cytoplasmic and mitochondrial protein synthesis with antibiotics Biochem Biophys Res Commun 1966 25 8 13 5971759 10.1016/0006-291X(66)90631-0
Daignan-Fornier B Fink GR Coregulation of purine and histidine biosynthesis by the transcriptional activators BAS1 and BAS2 Proc Natl Acad Sci U S A 1992 89 6746 50 1495962
Quemeneur L Flacher M Gerland LM Ffrench M Revillard JP Bonnefoy-Berard N Mycophenolic acid inhibits IL-2-dependent T cell proliferation, but not IL-2-dependent survival and sensitization to apoptosis J Immunol 2002 169 2747 55 12193749
Warit S Zhang N Short A Walmsley RM Oliver SG Stateva LI Glycosylation deficiency phenotypes resulting from depletion of GDP- mannose pyrophosphorylase in two yeast species Mol Microbiol 2000 36 1156 66 10844699 10.1046/j.1365-2958.2000.01944.x
Shimma Y Nishikawa A bin Kassim B Eto A Jigami Y A defect in GTP synthesis affects mannose outer chain elongation in Saccharomyces cerevisiae Mol Gen Genet 1997 256 469 80 9413430 10.1007/s004380050591
Chepurnaya OV Kozhina TN Peshekhonov VT Korolev VG The REC41 gene of Saccharomyces cerevisiae: isolation and genetic analysis Mutat Res 2001 486 41 52 11356335
Lee BN Elion EA The MAPKKK Ste11 regulates vegetative growth through a kinase cascade of shared signaling components Proc Natl Acad Sci U S A 1999 96 12679 84 10535982 10.1073/pnas.96.22.12679
Mondesert G Clarke DJ Reed SI Identification of genes controlling growth polarity in the budding yeast Saccharomyces cerevisiae: a possible role of N-glycosylation and involvement of the exocyst complex Genetics 1997 147 421 34 9335583
Dennis PB Jaeschke A Saitoh M Fowler B Kozma SC Thomas G Mammalian TOR: a homeostatic ATP sensor Science 2001 294 1102 5 11691993 10.1126/science.1063518
Brachmann CB Davies A Cost GJ Caputo E Li J Hieter P Boeke JD Designer deletion strains derived from Saccharomyces cerevisiae S288C: a useful set of strains and plasmids for PCR-mediated gene disruption and other applications Yeast 1998 14 115 32 9483801 10.1002/(SICI)1097-0061(19980130)14:2<115::AID-YEA204>3.0.CO;2-2
Molk JN Schuyler SC Liu JY Evans JG Salmon ED Pellman D Bloom K The differential roles of budding yeast Tem1p, Cdc15p, and Bub2p protein dynamics in mitotic exit Mol Biol Cell 2004 15 1519 32 14718561 10.1091/mbc.E03-09-0708
Dobbelaere J Gentry MS Hallberg RL Barral Y Phosphorylation-dependent regulation of septin dynamics during the cell cycle Dev Cell 2003 4 345 57 12636916 10.1016/S1534-5807(03)00061-3
Sagot I Bonneu M Balguerie A Aigle M Imaging fluorescence resonance energy transfer between two green fluorescent proteins in living yeast FEBS Lett 1999 447 53 7 10218581 10.1016/S0014-5793(99)00258-6
Yang TT Sinai P Green G Kitts PA Chen YT Lybarger L Chervenak R Patterson GH Piston DW Kain SR Improved fluorescence and dual color detection with enhanced blue and green variants of the green fluorescent protein J Biol Chem 1998 273 8212 6 9525926 10.1074/jbc.273.14.8212
Pinson B Sagot I Borne F Gabrielsen OS Daignan-Fornier B Mutations in the yeast Myb-like protein Bas1p resulting in discrimination between promoters in vivo but notin vitro Nucleic Acids Res 1998 26 3977 85 9705508 10.1093/nar/26.17.3977
Sagot I Klee SK Pellman D Yeast formins regulate cell polarity by controlling the assembly of actin cables Nat Cell Biol 2002 4 42 50 11740491
Breton AM Schaeffer J Aigle M The yeast Rvs161 and Rvs167 proteins are involved in secretory vesicles targeting the plasma membrane and in cell integrity Yeast 2001 18 1053 68 11481676 10.1002/yea.755
|
15869715
|
PMC1274246
|
CC BY
|
2021-01-04 16:39:10
|
no
|
BMC Cell Biol. 2005 May 4; 6:24
|
utf-8
|
BMC Cell Biol
| 2,005 |
10.1186/1471-2121-6-24
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-381574061210.1186/1471-2105-6-38SoftwarehtSNPer1.0: software for haplotype block partition and htSNPs selection Ding Keyue [email protected] Jing [email protected] Kaixin [email protected] Yan [email protected] Xuegong [email protected] National Laboratory of Medical Molecular Biology; Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences (CAMS) and Peking Union Medical College (PUMC), Beijing 100005, China2 MOE Key Laboratory of Bioinformatics/Department of Automation, Tsinghua University, Beijing 100084, China3 Chinese National Human Genome Center, Beijing 100176, China2005 1 3 2005 6 38 38 25 6 2004 1 3 2005 Copyright © 2005 Ding et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
There is recently great interest in haplotype block structure and haplotype tagging SNPs (htSNPs) in the human genome for its implication on htSNPs-based association mapping strategy for complex disease. Different definitions have been used to characterize the haplotype block structure in the human genome, and several different performance criteria and algorithms have been suggested on htSNPs selection.
Results
A heuristic algorithm, generalized branch-and-bound algorithm, is applied to the searching of minimal set of haplotype tagging SNPs (htSNPs) according to different htSNPs performance criteria. We develop a software htSNPer1.0 to implement the algorithm, and integrate three htSNPs performance criteria and four haplotype block definitions for haplotype block partitioning. It is a software with powerful Graphical User Interface (GUI), which can be used to characterize the haplotype block structure and select htSNPs in the candidate gene or interested genomic regions. It can find the global optimization with only a fraction of the computing time consumed by exhaustive searching algorithm.
Conclusion
htSNPer1.0 allows molecular geneticists to perform haplotype block analysis and htSNPs selection using different definitions and performance criteria. The software is a powerful tool for those focusing on association mapping based on strategy of haplotype block and htSNPs.
==== Body
Background
Several recent genome-wide and experimental studies suggested that the genome consists of chromosome regions of strong inter-marker linkage disequilibrium (LD) (i.e., haplotype blocks) and has discrete boundaries defined by recombination hotspots [1-4]. There are a few common haplotypes of limited haplotype diversity within each haplotype block, which can be characterized by only a small number of haplotype tagging SNPs (htSNPs). Haplotype blocks and htSNPs have great implication for association-based mapping of disease genes, by significantly reducing the genotyping effort with only a modest loss of power [5]. A new genomic map (i.e., haplotype map) for characterizing the haplotype structure in human genome is now underway to speed up the searching for genes involved in complex diseases.
A range of operational definitions has been used to identify haplotype block structures [1,2,6,7], which can be roughly cataloged into three groups [19]. First, there are methods based on diversity in the sequence, such as those of Patil et al. [1] and Zhang et al. [7], which define blocks with low sequence diversity by some diversity measure. The second group is based on LD methods, such as that of Gabriel et al. [2], which defines blocks with generally high pairwise LD within blocks and low pairwise LD between blocks. Finally, there are methods that look for direct evidence of recombination, such as that of [6], using the four-gamete test developed by Hudson and Kaplan [8] and defining blocks as apparently recombination-free regions.
There is still no consensus on the performance criteria of htSNPs selection. Broadly, these criteria are categorized into two groups. One comprises the diversity criteria which evaluate the information captured from the original haplotype diversity [1,9], such as the α-percent coverage, requiring that the total frequencies of all haplotypes completely distinguished by the htSNPs set should be no less than α. The other group consists of the association-based criteria, concerned most directly with the issue of prediction – the ability of the reduced set H to detect unknown SNPs in the set A of all SNPs within the genome region of interest [10].
To provide molecular geneticists more convenience in analyzing haplotype block structures and in selecting htSNPs, we develop a computational tool, htSNPer1.0, with a graphical user interface (GUI) implementing the above algorithms for block structure partition and htSNPs selection.
Implementation
htSNPer1.0 is a computer program with a GUI for characterizing the haplotype block structure and selecting htSNPs. The core algorithm is implemented in C++ language, and the graphic interface is coded in Java. The software is platform-independent.
Here, we will be concerned with haplotype block partition and htSNPs selection of unphased autosomal SNPs genotype data. For the block definitions that can directly handle unphased genotype data such as Gabriel et al. [2] and those based on pairwise LD [11], the unphased data are first partitioned into blocks over which there is sufficient restriction of haplotype diversity. Then, haplotypes are estimated approximately within each block (by EM algorithm). Finally, based on these estimated haplotypes, htSNPs are selected according to certain htSNPs performance criterion [10]. For those block definitions that can only handle phased haplotype data [1,6], haplotypes are estimated first (by EM algorithm) from unphased genotype data. Then block partition and htSNPs selection are both based on these estimated haplotypes.
Haplotype estimation – EM algorithm
We apply the EM algorithm used by SNPHAP to estimate haplotypes from genotype data [18]. When the data consist of a large number of SNPs, the number of possible haplotype instances may become extremely large. In order to avoid this problem, the program starts from the first two SNPs and extends the solution by sequentially adding the rest SNPs. As each new SNP is added, the number of possible haplotypes is expanded considering all possible larger haplotypes. After EM algorithm estimating the posterior probabilities, the program deletes genotype assignments with posterior probability lower than 0.001. Then the posterior probabilities of the rest genotype assignments are recomputed.
We use the EM algorithm in SNPHAP because it is simple and fast, and can be easily integrated in our C++ code. There are other algorithms like HAPLOTYPER [15], PHASE [16] and PLEM [17] that are better studied and more widely used. However, one distinctive feature of htSNPer1.0 is to estimate haplotypes within each haplotype block. Within the blocks there is very limited haplotype diversity, so in such cases the algorithm in SNPHAP performs reasonably well. If one likes to do the haplotype phasing before block partition, he/she can use HAPLOTYPER [15], PHASE [16] or PLEM [17] to get more accurate estimation, and then input the estimated haplotypes to htSNPer1.0 to do the block partition and htSNP selection.
Definitions for haplotype blocks
htSNPer has integrated four haplotype block definitions: chromosome coverage [1], average pairwise LD |D'| [11], estimated pairwise LD confidence limits [2] with minor modifications by Wall and Prichard [14], and no historical recombination [6].
1. Chromosome coverage [1]. A block is defined as a region in which the sum frequencies of common haplotypes (whose frequency is over a threshold, e.g. 0.05) is no less than a threshold. For this definition of blocks we apply a dynamic programming for haplotype partitioning [7]. We define a Boolean function block (i, j) = 1 if the consecutive SNPs from SNPi to SNPj can be defined as a block according to the above definition, and block (i, j) = 0 otherwise. Let f (i, j) be the size of the minimal htSNP set found by GBB algorithm (see below) for α-percent coverage within the block from SNPi to SNPj. Given a block partition (1, i1), (i1 + 1, i2),..., (in-1 + 1, in), the total number of htSNPs for these n blocks is f (1, i1) + f (i1 + 1, i2) +...+ f (in-1 + 1, in). The optimal block partition is defined to be the one that minimizes the total number of htSNPs.
Denote Sj to be the total number of htSNPs for the optimal block partition of the first j SNPs, and set S0 = 0. According to dynamic programming theory, we have . Through this recursion the dynamic programming partitions the haplotypes for the optimal block partition.
2. Average pairwise LD |D'| [11]. Within a block the average pairwise |D'| is no less than a threshold.
3. Estimated pairwise LD confidence limits [2] with minor modifications by Wall and Prichard [14]. For details see Additional file 1.
4. No historical recombination [6]. A block is defined as a region without any historical recombination, which is examined by Four Gamete Test.
The above definitions of 2, 3 and 4 do not guarantee a unique solution for partition. In htSNPer1.0, blocks are searched from the start of the input data and expanded as long as possible by sequentially adding the next SNPs.
htSNPs selection criteria
htSNPer1.0 can find the minimal htSNP set of global optimum. Different definitions of optimum can be derived according to different htSNP performance criteria [10]. A generalized definition of "optimum" can be described as the minimal set of htSNPs that satisfies a given htSNP performance criterion. For example, weighted-average haplotype r2 is regarded as one of the most informative association-based htSNP performance measure by Weale et al. [10], which is defined as following:
Weighted-average haplotype
where
and we denote as the the frequency of haplotypes with allele 1 at SNP i, as the frequency of haplotypes in the gth htSNP-defined group (haplotypes within each group are identical at htSNP loci), and as the frequency of haplotypes both in the gth htSNP-defined group and with allele 1 at SNP i.
If the htSNP performance criterion is defined as the weighted-average haplotype r2 of the selected set of htSNPs should be at least 90% of the maximum possible value (which is the weighted-average haplotype r2 when all SNPs are selected as htSNPs), then the "optimum" according to this criterion can be described as the minimal set of htSNPs whose weighted-average haplotype r2 is at least 90% of the maximum possible value.
We have integrated the three htSNP performance criteria into our htSNPer software: α-percent coverage [1], explained proportion of Clayton's haplotype diversity [9], and weighted-average haplotype r2 [10].
α-percent coverage: the total frequencies of all haplotypes that are not completely distinguished by the htSNP set is less than 1 - α.
Explained proportion of Clayton's haplotype diversity: , where fi, fhaplo = g and fi,g are defined in the same way as above.
Weighted-average haplotype r2 : see above.
htSNPer1.0 takes advantages of a novel heuristic algorithm – Generalized Branch-and-Bound (GBB) algorithm, which is applicable for all kinds of htSNPs performance criteria, to search the minimal htSNPs set with both efficiency and global optimum, comparing to the exhaustive searching [7] which guarantees global optimum but runs very slowly, and to the greedy algorithm [1,13] which is faster but doesn't guarantee global optimum.
The GBB algorithm
Consider a block B containing N haplotypes and each haplotype has M bi-allelic SNPs markers. Each SNP marker can divide N haplotypes into two groups: one consists of all the haplotypes with its major allele, and the other with its minor allele. GBB algorithm is based on the following branching rule and Generalized Prune-rule, using the depth-first searching strategy (Figure 1).
1) Each node {T, R} in the searching tree consists of two parts: the test-set T and the discard-set R where T is the set of SNPs that have been selected, and R is the set of SNPs that should not be selected for the future. If the set of all SNPs is denoted as S, then the set of SNPs that can be used at the node is S\(T ∪ R). The search tree starts from the root node for which T = Φ and R = Φ.
2) A child node is generated by adding a SNP to T according to the branching rule. The node is pruned if it meets the Generalized Prune-rule.
Importance calculation
Given a certain node {T, R}, SNPs in T divide all the haplotypes into t non-overlapping groups called equivalence classes. Any haplotypes that belong to the same group are identical at all SNP sites in T. A biallelic SNP divides all the haplotypes into two groups: Gmajor and Gminor. To evaluate the competence of the SNP, the importance of a SNP is defined by
Branching rule
Given a node {T, R}, sort the SNPs in S\(T ∪ R) according to the importance calculation non-increasingly: I(SNP1 |T) ≥ I(SNP2 |T) ≥ … ≥ I(SNP|S|-|T|-|R| |T), create the children {T ∪ SNP1, R}, {T ∪ SNP2, R ∪ SNP1}, {T ∪ SNP3, R ∪ SNP1 ∪ SNP2}, ..., {T ∪ SNP|S|-|T|-|R|, SNPh}, and explore the children in this order.
Generalized prune-rule
Check whether the SNP subset T meets the htSNP performance criterion. If it does, prune the node when |T| ≥ U, or update U when |T| <U ; otherwise, prune the node when |T| ≥ U or |S| - |T| - |R| < 1 where U is the size of the best solution found so far.
The Importance Calculation in Branching rule is originally devised for the α-percent coverage criterion [1,7]. But it is also applicable for other criteria, although it may not be the best one. Actually, one can devise specific Branching-rule and Prune-rule according to specific htSNPs selection criterion in the GBB framework to achieve super efficiency and global optimization. The GBB framework and algorithm we proposed are applicable to all htSNP criteria, and are at least more efficient than enumeration.
Results and discussion
htSNPer1.0 takes advantages of a novel heuristic algorithm, Generalized Branch-and-Bound algorithm. It is applicable for all kinds of htSNPs performance criteria. The algorithm is of high computational efficiency and it can reach the global optimum.
htSNPer1.0 has integrated three htSNPs performance criteria and four haplotype block definitions. Besides genotype data, htSNPer1.0 can also handle haplotype data directly. It takes a simple flat-file as input. A dialogue box is used to set up parameters and for htSNPs selection algorithm (GBB algorithm and greedy algorithm). In the tabbed-output panel, htSNPer1.0 demonstrates the results both in the form of graphics and plain-texts. A graphical representation of haplotype block partition and htSNPs selection is provided in the graphic panel (Figure 2). In this example, there are 51 SNPs and 50 haplotypes in its input. The LD-based definition was used for haplotype partition, weighted-average haplotype r2 for htSNPs performance criteria, and branch-and-bound algorithm in htSNPs selection. In the text-output panel, there is more information about the analysis and results, such as the methods/criterion used on haplotype block definition and htSNPs selection. Users can also select different haplotype definitions and htSNPs performance criteria to compare the results from the result tree in the left panel.
Application example
Study sample
In order to compare the time used and the htSNPs numbers chosen with different softwares, we used the human chromosome 21 haplotype [1] as the test data. This dataset consists of 20 haplotype samples, and 24,047 common SNPs (minor allele frequency no less than 0.10). About 21% of the chromosome 21 data are missing data.
Results based on various haplotype block definitions and htSNP Selection Criteria
Results of the three different htSNPs performance criteria by htSNPer1.0:
Alpha-percent coverage: 3,953 blocks and 5,082 htSNPs.
Haplotype Diversity: 3,055 blocks and 4,619 htSNPs.
Weighted-average haplotype r2 : with this criterion and GBB algorithm, htSNPer can not run on our computer because of the large amount of required memory. Using greedy algorithm instead of GBB (see Additional file 1), we get 3,098 blocks and 6,962 htSNPs.
Using the same data set, the same htSNPs performance criterion of α-percent coverage but the different block searching algorithm, Patil et al. [1] reported 4,135 blocks and 4,563 htSNPs. Zhang et al. [7] reported 2,515 blocks and 3,582 htSNPs. About 21% of the chromosome 21 data are missing data and different programs use different strategy to handle missing data. All these contribute to the differences between the results of the above programs.
We also ran the different programs to compare the computational efficiency. The algorithms used for comparison were the GBB algorithm in htSNPer1.0, the greedy algorithm by Zhang et al. [13] and the enumeration algorithm in Zhang et al. [7]. For comparison, we used the diversity-based haplotype block definition and α-percent coverage criterion in htSNPs selection (α = 0.8). Running on our computer with 2.4 GHz AMD Athlon processor (1GB memory), with the block definition of Patil et al. [1], α-percent coverage htSNPs performance criterion and a dynamic programming for haplotype partition, htSNPer1.0 requires 3 hours 23 minutes to identify 3,953 blocks and 5,082 htSNPs. For comparison, the greedy algorithm in Zhang et al. [13] requires 20.74 seconds, identifying 3,766 blocks and 8,733 htSNPs, and the enumeration method in Zhang et al. [7] was too slow to apply on our computer on this human chromosome 21 dataset. The difference in efficiency may be partially due to the different block partition searching strategies applied by these three programs. Zhang et al. [13] uses greedy algorithm for block partition, while htSNPer1.0 and Zhang et al. [7] use dynamic programming algorithm.
Conclusion
In conclusion, htSNPer1.0 is a java-based program with Graphic User Interface. It allows molecular geneticists to perform haplotype block analysis and htSNPs selection using different definitions, different performance criteria, as well as different algorithms. The software is a powerful tool for those focusing on association mapping based on haplotype block and htSNPs strategy.
Availability and requirements
htSNPer1.0 is a graphic user interface and a platform-independent software. The software is available at . The source code is available on request to the authors.
htSNPer1.0 takes a plain text file as input, either unphased autosomal SNPs genotype data or phased haplotype data. It requires the Java Running Environment (Jre1.4 or later version) to run the program properly. Detailed tutorials, htSNPer1.0 help system and examples are distributed within htSNPer1.0 software. Please inform the corresponding author if you are a non-academic user.
Authors' contributions
K. Ding and J. Zhang provided software design, development and testing of the software, and J. Zhang wrote the source codes. K. Zhou participated in its design. Y. Shen and X. Zhang provided biological direction and validation of the tool. Y. Shen conceived of this project, and participated in its coordination. All authors have read and approved the final manuscript.
Supplementary Material
Additional File 1
It provides more details of the methodology that are not covered in the main text.
Click here for file
Acknowledgements
We thank Dr. Michael Q. Zhang for valuable discussion. This work is supported in part by China National Key Program on Basic Research (Grant G1998051003 to Y.S and 2004CB518605 to X.Z.), China National High-Tech R&D Program (Grant 863-102-10-03-05 to Y.S), and NSFC (Grants 39625007 and 39993420 to Y.S, 60275007 and 60234020 to X.Z).
Figures and Tables
Figure 1 An example for generialized branch-and-bounnd algorithm with 4 SNPs and 4 haplotypes. The htSNPs performance criterion is to distinguish all the different haplotypes. The depth-first searching starts from root, exploring nodes in the order N1, N2,..., N7. N2 is the globally optimal solution. N3,..., N7 are all pruned from further consideration.
Figure 2 Sample output from htSNPer1.0. The first line showed the SNPs index, and could be replaced by the SNPs coordinate in its input. The brown color blocks represent the haplotype block structure in this region. Three classes of dots represent the input SNPs, the SNPs over a threshold (e.g., 0.10; defined in the optional dialogue), and the htSNPs, respectively.
==== Refs
Patil N Berno AJ Hinds DA Barrett WA Doshi JM Hacker CR Kautzer CR Lee DH Marjoribanks C McDonough DP Nguyen BTN Norris MC Sheehan JB Shen N Stern D Stokowski RP Thomas DJ Trulson MO Vyas KR Frazer KA Fodor SPA Cox DR Blocks of Limited Haplotype Diversity Revealed by High-Resolution Scanning of Human Chromosome 21 Science 2001 294 1719 1723 11721056 10.1126/science.1065573
Gabriel SB Schaffner SF Nguyen H Moore JM Roy J Blumenstiel B Higgins J DeFelice M Lochner A Faggart M Liu-Cordero SN Rotimi C Adeyemo A Cooper R Ward R Lander ES Daly MJ Altshuler D The structure of haplotype blocks in the human genome Science 2002 296 2225 2229 12029063 10.1126/science.1069424
Jeffreys AJ Kauppi L Neumann R Intensely punctuate meiotic recombination in the class II region of the major histocompatibility complex Nat Genet 2001 29 217 222 11586303 10.1038/ng1001-217
May CA Shone AC Kalaydjieva L Sajantila A Je_reys AJ Crossover clustering and rapid decay of linkage disequilibrium in the Xp/Yp pseudoautosomal gene SHOX Nat Genet 2001 31 272 275 10.1038/ng918
Zhang K Calabrese P Nordborg M Sun FZ Haplotype block structure and its applications to association studies Am J Hum Genet 2002 71 1386 1394 12439824 10.1086/344780
Wang N Akey JM Zhang K Chakraborty R Jin L Distribution of recombination crossovers and the origin of haplotype blocks: the interplay of population history, recombination, and mutation Am J Hum Genet 2002 71 1227 1334 12384857 10.1086/344398
Zhang K Deng M Chen T Waterman MS Sun FZ A dynamic programming algorithm for haplotype block partition Proc Natl Acad Sci USA 2002 99 7335 7339 12032283 10.1073/pnas.102186799
Hudson R Kaplan N Statistical properties of the number of recombination events in the history of a sample of sequences Genetics 1985 111 147 164 4029609
Clayton D Choosing a set of haplotype tagging SNPs from a larger set of diallelic loci
Weale ME Depondt C Macdonald SJ Smith A Lai PS Shorvon SD Wood NW Goldstein DB Selection and Evaluation of Tagging SNPs in the Neuronal-Sodium-Channel Gene SCN1A: Implications for Linkage-Disequilibrium Gene Mapping Am J Hum Genet 2003 73 551 565 12900796 10.1086/378098
Reich DE Cargill M Bolk S Ireland J C SP Richter DJ Lavery T Kouyoumjian R Farhadian SF Ward R Lander ES Linkage disequilibrium in the human genome Nature 2001 411 199 204 11346797 10.1038/35075590
Anderson EC Novembre J Finding haplotype block boundaries by using the Minimum-Description-Length principle Am J Hum Genet 2003 73 336 354 12858289 10.1086/377106
Zhang K Jin L HaploBlockFinder: haplotype block analyses Bioinformatics 2003 19 1300 1301 12835279 10.1093/bioinformatics/btg142
Wall JD Pritchard JK Assessing the Performance of the Haplotype Block Model of Linkage Disequilibrium Am J Hum Genet 2003 73 502 515 12916017 10.1086/378099
Niu T Qin ZS Xu X Liu JS Bayesian Haplotype Inference for Multiple Linked Single-Nucleotide Polymorphisms Am J Hum Genet 2002 70 157 169 11741196 10.1086/338446
Stephens M Smith NJ Donnelly P A New Statistical Method for Haplotype Reconstruction from Population Data Am J Hum Genet 2001 68 978 989 11254454 10.1086/319501
Qin ZS Niu T Liu J Partition-Ligation Expectation-Maximization Algorithm for Haplotype Inference with Single-Nucleotide Polymorphisms Am J Hum Genet 2002 71 1242 1247 12452179 10.1086/344207
SNPHAP
Schwartz R Halldorsson BV Bafna V Clark AG Istrail S Robustness of Inference of Haplotype Block Structure J Comp Biol 2003 10 13 19 10.1089/106652703763255642
|
15740612
|
PMC1274247
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 1; 6:38
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-38
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-511576047810.1186/1471-2105-6-51SoftwareCoPub Mapper: mining MEDLINE based on search term co-publication Alako Blaise TF [email protected] Antoine [email protected] Baal Sjozef [email protected] Rob [email protected] Stefan [email protected] Ton [email protected] Jan [email protected] Guido [email protected] Department of Molecular Design & Informatics, Organon NV, P.O. Box 20, 5340 BH Oss, The Netherlands2 Department of Urology, Erasmus MC, P.O. Box 1738, 3000 DR Rotterdam, The Netherlands3 Department of Genetics, Erasmus MC, Rotterdam, The Netherlands4 Department of Medical Informatics, Erasmus MC, Rotterdam, The Netherlands2005 11 3 2005 6 51 51 21 12 2004 11 3 2005 Copyright © 2005 Alako et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
High throughput microarray analyses result in many differentially expressed genes that are potentially responsible for the biological process of interest. In order to identify biological similarities between genes, publications from MEDLINE were identified in which pairs of gene names and combinations of gene name with specific keywords were co-mentioned.
Results
MEDLINE search strings for 15,621 known genes and 3,731 keywords were generated and validated. PubMed IDs were retrieved from MEDLINE and relative probability of co-occurrences of all gene-gene and gene-keyword pairs determined. To assess gene clustering according to literature co-publication, 150 genes consisting of 8 sets with known connections (same pathway, same protein complex, or same cellular localization, etc.) were run through the program. Receiver operator characteristics (ROC) analyses showed that most gene sets were clustered much better than expected by random chance. To test grouping of genes from real microarray data, 221 differentially expressed genes from a microarray experiment were analyzed with CoPub Mapper, which resulted in several relevant clusters of genes with biological process and disease keywords. In addition, all genes versus keywords were hierarchical clustered to reveal a complete grouping of published genes based on co-occurrence.
Conclusion
The CoPub Mapper program allows for quick and versatile querying of co-published genes and keywords and can be successfully used to cluster predefined groups of genes and microarray data.
==== Body
Background
High throughput microarray analysis has made it possible to analyze the mRNA expression of most if not all human genes simultaneously [1,2]. The data generated from these analyses are overwhelming since hundreds of interesting differentially expressed genes can be identified in a single assay. Knowledge on expression levels of genes in different systems is useful, but does not directly answer biologically relevant questions, such as: What is the gene function? Where is the gene located within the genome? Where is the protein located within the cell? Most important is the answer to the question whether genes identified in microarray experiments have something in common, such as, are multiple genes part of a single biological pathway or proteins part of a protein complex? The public database which contains much of the relevant information to answer these questions is MEDLINE. Therefore, mining the MEDLINE database for all information on a set of genes of interest to extract and evaluate their co-occurrences with biological keywords and other genes, could reveal biologically relevant pathways [3-6].
The most widely used methodology to identify genes and proteins in text is by thesaurus-based concept extraction. Using a predefined gene name list, text phrases are compared to the thesaurus for matching. Complications for gene name thesauri are variations in full name spelling, use of abbreviations (gene symbols), the large number of synonyms (different name but same gene) and homonyms (same name but meaning different genes or unrelated concepts) [7,8]. Particularly homonyms in the form of abbreviations and acronyms create a serious problem of false positive assignment of a gene to a particular concept [9-13]. A complementary approach for gene/protein identification is "named entity recognition" in which a program learns to recognize concepts from text [14-16]. Due to the enormous synonym and homonym problems, named entity recognition encounters difficulties in achieving high performance gene name identification. A next step in text mining is linking of different concepts (such as gene names and keywords) that are identified. In the simplest method, co-occurrence of two concepts within the document can be used as an indication of linkage. Extensions of co-occurrence can include (i) the number of times a concept is found, (ii) how close concepts are to one another, such as, within a single sentence, and (iii) not just two, but the weighed combination of all concepts within a document. More sophisticated fact extraction methods can also retrieve information on the type of relationship between two concepts. Natural language processing (NLP) grammatically parses whole sentences to identify verbs and other connecting phrases that describe the correlation between concepts [3,4,6,17]. A third step in text mining takes linked concepts and groups them according to their co-occurrence and relationships. Again, this can be performed by simple clustering of the co-occurrence of pairs of concepts as well as complex multi-dimensional classification using weighed concept combinations [18,19]. This type of clustering of, for example, differentially expressed genes from a microarray experiment, can disclose, summarize, and visualize published knowledge, but can also be utilized for novel information discovery [5,20]. Although progress is being made in higher order literature processing, text mining applications in the field of genomics are mainly thesaurus and co-occurrence based. Such programs and methods to identify potential functional correlations between genes have been described [21-33]. Each of these applications has its unique advantages and limitations, showing the broad range of needs for text mining as well as the numerous extraction, linking, and discovery methods feasible.
We set out to create a well annotated and curated open source gene list including full names, symbols and aliases and a regular expression-based search method to identify genes in text databases such as MEDLINE. In addition to the gene thesaurus, specific keyword lists were generated for co-occurrence analyses. For each concept, PubMed identifiers (IDs) from MEDLINE documents containing the concept were extracted, all gene-gene and gene-keyword co-occurrence pairs identified and stored in a database for fast co-occurrence retrieval. This database can be mined using single or batches of concepts to retrieve co-occurrences that form the input in clustering programs to group genes and keywords according to their similarity in co-publications. The program, database and all thesauri are freely available and can be adapted to include updates, new thesauri, and search methods.
Implementation
Human gene thesaurus
A human gene thesaurus was compiled from the Affymetrix HG_U95 / HG_U133 and HUGO gene annotations (HG_U95 / HG_U133 annotation files from 2002) [34,8] (Table 1). In total, 15,621 annotated genes were included of which most gene descriptions consist of one or more full names, the gene symbol, and their aliases. The typical HUGO and Affymetrix full gene name descriptions contain commas, semicolons and often alternative names in parenthesis, which makes this description an inadequate direct search term. Full names were processed by replacing the commas and semicolons with the Boolean "AND" operator (Figure 1). All terms included in parentheses were deleted from "gene-level name" and placed in a separate field named "gene-level additional description". Both fields were semi-automatically curated to remove common words (such as protein, family, hypothetical, functional, human, tissue, yeast, etc), misspellings, and insert Boolean "OR" in case synonyms are described. From gene symbols and aliases fields, commas and semicolons separators were replaced by the Boolean "OR" operator. Two-letter symbols and aliases were removed from the thesaurus and all other abbreviations were compared to an English dictionary [35] to remove common English words (such as "AND", "CELL", etc.). The Microsoft Excel spreadsheet program was used for generating and curating gene thesaurus files and, as described by Zeeberg et al [36], conversion problems were encountered and when identified, manually corrected.
Semi-automatic stemming was performed on "gene-level name" and "gene-level additional description" fields by removing numbers, letters, and phrases like "alpha", "member", "type", "class", etc. This resulted in a stem-level gene name description. Although the current version of CoPub Mapper does not take this stem-level into account, these fields are part of the gene thesaurus and freely available.
Keyword thesauri
In total, five different keyword thesauri were compiled including the Gene Ontology "biological process", "cellular component", and "molecular function", as well as "diseases" and "tissues" (Table 1). In the disease thesaurus, commas were replaced with the Boolean "OR" operator. All keyword databases were manually curated to remove terms too specific or too common.
MEDLINE concept extraction and curation
The full MEDLINE baseline XML files (until January 2004) were obtained from the National Library of Medicine [37], extracted to small text files containing title, abstract and substances using BioPerl API. The title, substance and abstract fields from MEDLINE records from 1966 to January 2004 were searched for the presence of different case-insensitive gene and keyword concepts using Perl compatible regular expressions (PCRE). For the gene-level name descriptions the characters "] [.-)(,:;" and space were allowed preceding and following the gene-level name description and also an optional "s" was permitted to follow the name. Any space in the gene-level name description was allowed to be a space or a dash. The same regular expressions were applied to the gene name stem-level descriptions, except that, the description could also be followed by any single letter or a number between 0 and 99. Gene symbols and aliases could be preceded and followed by the characters "] [.-)(,:;" and space. After the first two characters, the presence of a dash was allowed in between the characters of the symbols and aliases (to take, for example, both "bcl2" and "bcl-2" into account). The concepts of the keyword files could be preceded and followed by the characters "][.-)(,:;" and space. In addition, "s" and "'s" were allowed to follow the disease concept. As for the gene-level name descriptions, a dash was allowed to be present between the words of a keyword concept. Per annotated gene or keyword, the PubMed IDs of MEDLINE records in which the concept was identified were stored in a MySQL database.
In order to identify potential problem concepts, 50 genes and 50 keywords with the highest number of PubMed IDs were manually inspected and curated if appropriate. In addition, a random selection of genes and all keywords that gave less than 2 MEDLINE hits were examined and this evaluation was used to optimise the thesauri and regular expressions search strategy described above.
To address the homonym issue, a correction was made for possible discrepancies between a parenthesised gene symbol and its expected name. All abbreviations in parenthesis in MEDLINE abstracts were retrieved in combination with 4 preceding words. In total, 1,105,669 MEDLINE records were identified where the abbreviation matched a gene symbol or alias. For all these records, 4 words preceding the abbreviation were compared to the gene-level name description of that particular gene. If none of the words resembled partly the gene name, the PubMed ID was removed from that particular gene's PubMed ID list. Using this method, 603,580 records were deleted from the gene hit database resolving part of the gene-unrelated concept homonym problems. Manual inspection of 173 random records revealed that, extrapolated, 79 % of the 603,580 records was correctly removed, while 7 % of the 502,089 non-removed records should have been deleted.
In our examination of genes with the highest number of PubMed IDs and our first CoPub Mapper analyses, we noticed a distinct contamination of records identifying gene symbols and aliases by abbreviation used for cell lines (such as PC3 which is an alias for 3 genes as well as a prostate cancer cell line). Since full names of cell line abbreviations are rarely put in writing, the homonym correction did not eliminate these discrepancies. A list of cell line names was retrieved [38] and gene symbols and aliases that fitted a cell line name were further processed. From 106 genes that included one of the cell line homonym names, all MEDLINE records were deleted in which the cell line name was mentioned without the presence of the stem-level gene name. In total, 100,213 PubMed IDs were eliminated. A manual inspection of 78 randomly chosen records showed that 87 % were correctly removed.
Database set-up and CoPub Mapper program
A file was generated that contains a unique query ID and the probeset IDs, UniGene (combination of Aug 2002 and Oct 2003 builds) and RefSeq identifiers for each of the individual 15,621 entries in the gene thesaurus (alias_affygene). In addition, a file with the gene name, symbol and aliases and unique query ID was created (query_affygene).
The retrieved PubMed IDs from each field (gene names, symbols and aliases) of the 15,621 unique gene thesaurus query IDs were non-redundantly combined into a MySQL database (lit_affygene) and a separate data-file (litstat_affygene) in which the number of PubMed IDs per query was counted. Furthermore, the PubMed IDs from the keyword thesauri were per concept stored (query_keyword, lit-keyword and litstat_keyword). Per gene-gene pair and gene-keyword pair, overlaps in PubMed IDs were identified and separately stored in the database (pair_keyword_affygene). From these paired files, a pairstat file was generated containing the number of PubMed IDs of each concept, the number of overlapping PubMed IDs between the two concepts and a relative score. The relative score is based on the mutual information measure and was calculated as
S = PAB/PA * PB
in which PA is the number of hits for concept A divided by the total number of PubMed IDs, PB is the number of hits for concept B divided by the total number of PubMed IDs, and PAB is the number of co-occurrences between concepts A and B divided by the total number of PubMed IDs. The relative score is produced as a log10 conversion and in the batch search option in a 1–100 scaled log10 conversion:
R = 10log S
and the scaled log transformed relative score:
R' = 1 + 99 * (R - Rmin) / (Rmax - Rmin)
where Rmin and Rmax are the lowest and highest R values in each pairstat file, respectively.
The CoPub program was generated in Python and runs as a web-based application (CGI script). The text output of a batch search can be saved and imported into a clustering program such as Cluster [39] and SpotFire (Spotfire, Göteborg, Sweden). The HTML output of "number of hits", "relative score", and batch search results are hyperlinked to the MEDLINE database at the European Bioinformatics Institute [40] for direct manuscript retrieval.
Performance evaluation using ROC (receiver operating characteristics) curves
In order to investigate whether the CoPub Mapper output could group genes according to their MEDLINE co-occurrence profile, 8 different groups of genes were defined based on common gene ontology (GO) terms [41], the BRCA1 BioCarta pathway [42], or a microarray experiment (Table 2). In the UniGEM V microarray experiment, the gene expression profile of prostate stroma cells was compared to prostate epithelial cells [43]. A set of 28 annotated genes, higher expressed in epithelial cells as compared to stromal cells (more than 2-fold) were randomly selected.
The 150 genes from the eight selected gene groups are pooled into one set. The selected genes were entered into CoPub Mapper to generate the co-occurrence matrix of relative scores of genes versus genes and genes versus the 5 different keyword thesauri. Relative scores were only generated in case more than 2 co-publications occurred per concept-concept pair. The genes versus genes matrix was hierarchical clustered and visualised using Cluster and TreeView [39] (Figure 2).
For a systematic evaluation of performance we applied Receiver Operating Characteristics (ROC) graphs and the area under the ROC curve (AUC) as an outcome measure. To use this method all genes from the 8 subgroups are pooled into one set. To calculate an AUC for every gene we used the following procedure. A gene from the pooled set is selected as a seed. The seed is paired with all other genes in the set and non-centered Pearson correlation coefficients are calculated based on their co-occurrence profiles. The co-occurrence profile is one row of the co-occurrence matrix under investigation. The genes are ordered by their correlation coefficients, with the highest value at the first rank. To generate a ROC curve, the obtained ranking of the genes is viewed as the outcome of a classifier. For a seed, genes from the same subgroup are called positives and all other genes are called negatives. ROC curves are two-dimensional graphs in which the true-positive (TP) rate is plotted against the false-positive (FP) rate. The TP rate is defined as correctly classified positives divided by all positives. The FP rate is defined as incorrectly classified negatives divided by all negatives. While running down the list, for every rank the true and false positive rate are calculated, by taking all encountered genes to be classified as positive and all not yet encountered genes as negative. The AUC of the ROC curve is calculated. The procedure is repeated until an AUC has been calculated for every gene in the pooled set. An average AUC is calculated per subgroup. The AUC measure varies between 0 and 1. Random ordering gives an AUC of 0.5 and an AUC of 1 represents perfect ordering, i.e. all positives are at the top of the list with no negatives in between, indicating perfect co-occurrence clustering of the genes in the subgroup [44].
Results
Validation of CoPub Mapper co-occurrence profiling
To validate the usefulness of the CoPub Mapper output, we evaluated how well genes with known relations could be grouped according to their MEDLINE co-occurrence profile. As shown in Figure 2, partial clustering of the initial 8 groups occurred upon their gene-gene co-occurrence profile evaluation. To quantify this grouping, ROC (receiver operating characteristics) curves were generated and the AUCs (Area Under Curve) for each gene calculated. In Figure 3, the median AUCs ± SD of the genes per group are depicted. Most of the 8 groups and in particular the BRCA1-associated genes clustered well together in the gene-keyword comparisons (median AUC of 0.93 ± 0.07). The ubiquitin-associated genes performed worst (median AUC of 0.6 ± 0.11). With respect to the thesaurus selection, the overall clustering of the 8 groups using the "genes versus genes self" comparison, performed best with an average AUC of 0.76 ± 0.13. The "genes versus diseases" and "genes versus tissues" comparisons were for many of the 8 groups not resulting in clustering higher than expected by random chance. In other words, from co-publication analysis of genes with disease or tissue keywords, the commonality between the genes, as defined by the 8 groups, could rarely be traced (Figure 3). As shown in Table 2, six groups of genes were selected based on gene ontology keywords, using two from each of the annotation trees (biological process, molecular function, and cellular component). As expected and without exception, the AUC of the 6 groups of genes was higher using their corresponding GO-derived thesaurus compared to using the other two GO-derived thesauri. For example, the molecular function annotated group of "acetyltransferases" was clustered best using the "genes versus molecular function" co-publication comparison (AUC of 0.81 as compared to 0.65 using the biological process thesaurus and 0.59 using the cellular component thesaurus). This shows that the selection of keywords for co-occurrence analysis is an important determinant in optimal text-based grouping of genes.
Microarray analysis using CoPub Mapper
In order to validate the CoPub Mapper program with real microarray data, a set of differentially expressed genes was selected from a comparison between ovaries of healthy women and women suffering from Poly Cystic Ovary Syndrome (PCOS) [45]. PCOS is characterized by a combination of chronic anovulation, hyperandrogenism and cysts in ovaries and is the most common cause of anovulatory infertility. Also hyperinsulinemia and obesity can be observed in many PCOS patients [46,47].
A set of 230 dysregulated DNA fragments representing 189 genes were used as input for CoPub Mapper (see Table 1 in [45]). Gene-keyword pairs were obtained from biological processes and diseases. Relative scores were only generated in case 3 or more co-publications occurred per gene-keyword pair. From these 189 genes, 104 were annotated and had at least 3 co-publications with one of the keywords. Resulting matrices were exported as text files and opened and merged in Spotfire. Hierarchical clustering was used to group genes and keywords. Figure 4 shows that subsets of genes form clusters with subsets of biological processes and diseases. Zooming in on these clusters confirms the relation of certain genes with e.g. PCOS, diabetes, obesity, gametogenesis, immune response. Characterization of all clusters revealed known and unknown relations of these PCOS dysregulated genes with biological processes and diseases.
Single Gene-Keyword extraction
The CoPub Mapper includes an option to query the database for all genes and keywords co-published with a single gene of interest. In addition, a keyword of interest can be selected and all genes with 2 or more co-occurrences can be extracted. As examples, the top ten genes (Table 3) and top ten diseases (Table 4) co-published with the androgen receptor are shown. An assessment of the 2 lists identified the puromycin-sensitive aminopeptidase gene (NPEPPS) as an example of a homonym (Table 3, fourth gene). The PSA alias of NPEPPS is mainly used to specify prostate specific antigen. The prostate specific antigen gene (KLK3) is regulated by the androgen receptor and correctly found many times to be co-published with the androgen receptor (Table 3, second gene). Due to the homonym curation described in the Systems and Methods section, the number of co-occurrences of the androgen receptor with NPEPPS (246) is lower than with KLK3 (414). Before homonym curation, NPEPPS and KLK3 had 634 and 635 co-publications with the androgen receptor, respectively. The top ten list of diseases co-published with the androgen receptor (Table 4) is a near perfect reflection of the known diseases associated with androgen receptor activity and aberrations.
In Table 5, the top ten genes are listed that are most often co-published with the keyword "prostate cancer". Again, the incorrect identification of NPEPPS in 4507 MEDLINE entries is due to the PSA homonym.
Meta-analysis: all genes versus keywords
In order to provide a summary of all gene-keyword co-occurrences, CoPub Mapping was performed using all 15,621 annotated genes as input in the different gene-keyword thesauri co-occurrence comparisons. Relative scores were only computed if in at least two articles a co-occurrence was observed. Elimination of single gene-keyword co-publications was carried out to eradicate non-reproduced findings and to make the large matrices manageable. A second selection was made to eliminate genes which included only low relative scores. Many genes have multiple co-publications with very common keywords such as "cancer" (disease thesaurus), "cytoplasm" (cellular component thesaurus), etc. If not functionally relevant, these co-occurrences have typically a low relevance score. Genes with only low relevance scores were eliminated by removing those genes that did not have 1 or more scaled relevance scores of more than a threshold (between 39 and 52) in which 20 % of genes were eliminated. The hierarchical clustered genes-diseases co-publication matrix is displayed in Figure 5. 5626 genes (rows) versus 1275 diseases (columns) were grouped according to their co-publication profiles. The enlarged section shows the amount of detail present in the matrix (Figure 5B). The vertical lines in the matrix are caused by co-publication of almost all genes with very common disease keywords such as "cancer", "neoplasm", and "carcinoma". Horizontal lines are genes co-published with many diseases, such as "insulin", "interleukin 6", and "keratin 3A". If low relevance scores are masked by hiding values below 30 in TreeView or SpotFire, these streaks become less prominent.
Clustering and visualisation of only highly significant co-occurrences will result in discrete groups of genes and keywords as shown in Figure 6. Stringent selection criteria were implemented including: (i) each gene had to be co-published with at least two different keywords with a relevance score of more than 50, and (ii) a co-occurrence must have been described in at least 3 publications per gene-keyword combination. From the 10,203 genes co-occurring with cellular component keywords, 1135 genes were retrieved using the stringent selection criteria mentioned above. As expected, these genes were clustered according to well-known cellular components of which some examples are depicted (Figure 6).
Discussion
With the implementation of high-throughput technologies in many fields of research, problems have shifted from data gathering to data comprehension. Linking data from different sources, such as microarray expression data to biomedical text corpora, can assist in the disclosure, summary, and visualisation of knowledge. This is particularly valuable when from high throughput data, only a few items can be selected for further detailed low-throughput examination. Co-occurrence analysis of concepts using the MEDLINE literature database, is an effective tool to extract and categorize published knowledge. CoPub Mapper output was successfully used to cluster predefined groups of genes and resulted in a commonsensical clustering of PCOS microarray data. In addition, CoPub Mapper uncovered relationships between genes using single concept searches and provided an overall gene-keyword clustered summary of the literature. One obvious limitation of gene-driven text mining is the incomplete study and publication of all human genes. Out of approximately 30,000 human genes, we included 15,621 annotated genes of which 10,700 were mentioned at least once and 9,769 at least twice in MEDLINE. The use of human gene names, symbols and aliases does not necessarily mean a human-specific literature search. Many gene names and symbols are shared by other species as well.
The main advantages of CoPub Mapper above most other co-publication programs, are its modularity of keyword databases and the pre-calculated co-occurrences. Based on the results from the predefined groups of genes, the choice of keyword database made a substantial difference in clustering efficiency as determined by AUC calculations. Utilisation of a single joint thesaurus could counteract clustering due to inclusion of irrelevant non-discriminating keywords. Another illustration that keyword selection is an important issue, is the prevalence of common keywords such as "cancer" (disease), "membrane" (cellular component), "metabolism" (biological process), "receptor" (molecular function), and "blood" (tissue). These keywords are co-published with nearly any gene of interest and were identified using CoPub Mapper. Although the relative score is generally low, these co-occurrences will influence the clustering process. Manual removal or stringent selection criteria before clustering can largely eliminate this potential bias. Addition of new keyword thesauri such as species, technologies, drugs, toxicology, pathology, etc. is feasible. Pre-calculation of co-publication of all possible gene-gene and gene-keyword pairs and storage in the pairstat data file, makes querying the database extremely efficient. Although the data are present, CoPub Mapper is not programmed for co-occurrence querying of more than 2 concepts. We are currently integrating CoPub Mapper into the Sequence Retrieval System (SRS) for multi-concept interrogation and direct linkage to other databases (such as microarray data, Gene Ontology, OMIM, SwissProt, LocusLink, UniGene, Ensembl, etc.) [48].
Comparing the gene expression profiles of normal versus PCOS ovaries has identified a large number of genes representing networks and pathways that are deregulated in PCOS. However, the gene names and symbols hardly ever point to specific signal transduction pathways. The relation of genes with their function, localization and context has been described in literature. Here we show that within the list of differentially expressed genes some are linked to PCOS, obesity, diabetes and gametogenesis. This is without surprise and easily explained [46,47]. Other genes are linked to cell proliferation, differentiation and cancer. Most of them were downregulated which correlates with the observed arrest in growth and differentiation of follicles. Other clusters with no obvious link to PCOS may shed new light on the genes and pathways involved in the disease.
One of the major challenges associated with compiled heterogeneous text records such as MEDLINE, is correct gene recognition and assignment. The lack of consistent gene naming has resulted in a flood of synonyms and homonyms [7]. Although the synonym issue can be resolved by accumulating all different gene names and symbols, the correction for homonyms is still a daunting task. In order to include different spelling forms and the word context, we performed the text searches case insensitive and with predefined rules of regular expression.
The homonym problem consists of (i) different genes with identical gene name, symbol, or alias, and (ii), more frequently, a gene name, symbol or alias used for other terms than genes [9]. In the curated CoPub Mapper gene thesaurus, 1,286 of the 15,621 annotated genes (8.2 %) share a symbol or alias. In order to limit both aspects of the homonym problem, we (i) eliminated 2 letter symbols and aliases, (ii) deleted all symbols and aliases present in the English dictionary, (iii) manually curated terms with exceptionally high number of hits, (iv) corrected for cell line names, and (v) deleted records in which the preceding description of parenthesised symbols or aliases did not match the corresponding gene name. This last method has been used before to make an inventory of the homonym problem and provide strategies for correction, such as the one used here [9-13]. Although these measures effectively reduced the homonym problem, one will regularly encounter incorrect record assignment and invalid co-occurrence quotation using CoPub Mapper. Additional optimisation of the gene thesaurus might further reduce this problem to some extent, but other correction approaches should be considered. One of the most promising strategies to achieve disambiguation is based on the preferential co-occurrence of other concepts [9,10]. For example, concepts generally co-published with PSA meaning Poultry Science Association, will be very different from concepts co-published with PSA representing prostate specific antigen. Based on these preferential co-occurring concepts, one can assign the correct meaning to an ambiguous term.
Besides disclosure, summary, and visualisation of known facts using co-publication, one could also discover novel linkages among genes and between genes and other concepts. One possibility to identify unpublished, but plausible links, is to screen for black squares surrounded by red ones in a clustered co-occurrence heat map as shown in Figure 5. The fact that a particular gene-disease combination was not found in MEDLINE (black square), but clustered together with other co-published gene-disease pairs (red squares), could indicate an unpublished association. This approach shows analogies with the Swanson discovery framework in which concept A is known to relate to B and B is associated with C [49,50]. Combining all data, the deduction that A relates to C can be hypothesised and tested [49,51-53].
Conclusion
CoPub Mapper is a program that identifies and rates co-published genes and keywords starting from a single concept search or batch-wise from a set of genes. Its modularity and pre-calculated co-occurrences allow for quick and versatile querying. The regular-expression search strategy and homonym correction makes the keyword database comprehensive and less contaminated with false positive classifications. CoPub Mapper can be used to summarize, evaluate and categorise annotated genes from microarray analyses based on co-occurrences with biological keywords and other published genes.
Availability and requirements
The CoPub Mapper program is available for free use at this URL: or
Authors' contributions
GJ, SvB, and JP conceived the approach and participated in the early design. BTFA and AV developed and optimised the software. TR developed and performed the homonym correction algorithm. RJ performed and interpreted the AUC ROC analyses and SV performed the MEDLINE gene and keyword searches. The project was supervised by GJ, TR and JP. All authors read and approved the final manuscript.
Acknowledgements
We thank Edwin van den Heuvel, Victor de Jager, Rene van Schaik, Jacob de Vlieg, and BioASP for their support, NLM (National Library of Medicine) for licensing of MEDLINE and Jan Kors and Jeannette Kluess for careful reading of the manuscript.
Figures and Tables
Figure 1 Flow diagram of the processing and curation of the gene names, symbols and aliases. Gene names, symbols and aliases were retrieved from Affymetrix HG_U95 / HG_U133 and the HUGO databases.
Figure 2 Clustered view of gene co-occurrences among a collection of 8 groups of selected genes. Of the 150 genes, the relative scores of co-occurrences were calculated and clustered using hierarchical clustering. A co-occurrence was only taken into account when at least two articles mention the gene-gene pair. Using this criterion, 45 genes did not co-publish with any of the other 149 genes. To which group (Table 2) a gene belongs to is indicated in the right part of the figure. Image contrast in TreeView was set at 50. Scaled (1–100) relative scores are represented in a red spectrum with bright red being the highest score. A relative score of zero or no score are in black.
Figure 3 Receiver operating characteristics (ROC) of the 8 selected groups of genes to quantify their coherence upon clustering of literature co-occurrences. Co-occurrences of the 150 genes were determined with the genes themselves, or the 5 different keyword thesauri. A co-occurrence was only taken into account when at least two articles mention the gene-gene or gene-keyword pair. The co-occurrence matrixes were Pearson correlation clustered and the distances between genes determined. For each gene, it was determined whether the next closest clustered gene was a group member. Genes from the same group were scored as true positive and any other gene as false positive to generate a ROC curve. For each gene, the area under the ROC curve (AUC) was determined and the median of all the group members per group ± SD depicted. Scaling is from an AUC of 0.3 to 1. An AUC of 0.5, representing a random ordering is highlighted with a thick line.
Figure 4 Hierarchical clustering of literature co-occurences of 104 genes (rows) versus 761 biological processes and diseases (columns). A co-occurrence was only taken into account when at least three articles mention the gene-keyword pair. Hierarchical clustering of CoPub Mapper results using genes differentially expressed in PCOS ovaries. From 221 regulated genes 104 genes contain a gene name, symbol or alias and produce a gene-keyword pair with biological processes or diseases. 104 modulated genes returned 761 keywords denoting biological processes or diseases. Hierarchical clustering was performed using Spotfire using the Complete Linkage method and Correlation as Similarity Measure. Several subclusters were identified shown here with blue boxes; between parenthesis the number of genes in a cluster. A: PCOS, Obesity, Insulin Resistance (4); B & D: Gametogenesis (5&8); C: Cell adhesion, Angiogenesis (19); E & H: Immune response, Inflammation (14&11); F: Cancer, Cell growth, Differentiation (32); G: Inflammatory diseases (6).
Figure 5 Hierarchical clustering of literature co-occurrences of 5626 genes (rows) versus 1275 diseases (columns). A co-occurrence was only taken into account when at least two articles mention the gene-disease pair. Each gene had to have at least once a high (1–100 scaled) relevance score of >46. A: Overview of all 5626 genes and 1275 diseases. B: Enlargement of a small subsection of genes showing the amount of detail present in the CoPub Mapper analysis.
Figure 6 Hierarchical clustering of literature co-occurrences of 1135 genes (rows) versus 177 cellular components (columns). A co-occurrence was only taken into account when at least three articles mention the gene-cellular component pair. Each gene had to have at least twice a high (1–100 scaled) relevance score of >50. Relative scores of less then 50 were masked in the TreeView program. Some of the cellular component concepts responsible for clustering of genes are indicated.
Table 1 CoPub Mapper gene and keyword database information. Gene names, symbols and aliases were retrieved from Affymetrix HG_U95 / HG_U133 [54] and the HUGO databases [55]. The keyword thesauri include the three Gene Ontology subsections [41], diseases [56] and tissues/organs [57].
Thesaurus Data Source Number of terms Number of terms with MEDLINE hits Total number of MEDLINE citations
Gene Affymetrix HG_U95-133 HUGO 15,621 10,700 5,932,448
Molecular Function Gene Ontology 962 851 6,616,546
Cellular Component Gene Ontology 218 196 1,890,561
Biological Process Gene Ontology 767 621 3,455,950
Diseases Karolinska Institute 1475 1444 6,099,280
Tissues National Library of Medicine 309 307 9,083,831
Table 2 CoPub Mapper test groups. Eight groups of genes with a common function, process, cellular location, or microarray expression profile, were defined from gene ontology (GO), BioCarta, or a microarray experiment. The genes used for CoPub Mapper analysis were randomly selected from larger sets of genes part of the 8 different groups.
Test groups # Genes Source
smooth muscle contraction 12 GO (Biological Process)
acetyltransferase 18 GO (Molecular Function)
nuclear pore 15 GO (Cellular Component)
nucleosome 17 GO (Cellular Component)
ubiquitin 24 GO (Molecular Function)
hypoxia 26 GO (Biological Process)
BRCA1 11 BioCarta
Epithelial-specific genes 27 UniGEM V microarray: stroma vs epithelial cells
Table 3 CoPub Mapper single gene pair output. Output of the "Single Gene Pair Mapper" in which the top ten genes co-published with the androgen receptor are listed according to number of co-publications (Pmid hits).
Gene Name Gene Symbols Gene Alias Pmid Hits
progesterone receptor PGR NR3C3 605
kallikrein 3, prostate specific antigen KLK3 PSA 414
nuclear receptor subfamily 3, group C, member 1; glucocorticoid receptor NR3C1 GCR, GRL 389
aminopeptidase puromycin sensitive NPEPPS MP100, PSA 246
sex hormone-binding globulin SHBG ABP 179
gonadotropin-releasing hormone 1, leutinizing-releasing hormone GNRH1 GNRH, GRH, LHRH, LNRH 157
prolactin PRL 131
insulin INS 125
epidermal growth factor, beta-urogastrone EGF URG 123
tumor protein p53 TP53 P53 94
Table 4 CoPub Mapper single gene biological concept output. Output of the "Single Gene Biological Term Mapper" in which the top ten diseases co-published with the androgen receptor are listed according to their relevance score.
Keywords Number of hits log10 Relative Score
Androgen-Insensitivity Syndrome 229 3.07
Kennedy Disease 21 2.56
Muscular Atrophy Spinal 133 2.12
Prostate Cancer 932 1.93
Gynecomastia 59 1.88
Hypospadia 81 1.79
Sex Chromosome Aberrations 2 1.78
Hirsutism 76 1.78
Robinow Syndrome 2 1.71
X-Linked Myotubular Myopathy 2 1.65
Table 5 CoPub Mapper single gene biological concept output. Output of the "Single Gene Biological Term Mapper" in which the top ten genes co-published with the prostate cancer disease-keyword are listed according to number of co-publications.
Gene name Gene Symbols Gene Aliases Number of hits log10 Relative Score
kallikrein 3, prostate specific antigen KLK3 PSA 6628 2.55
aminopeptidase puromycin sensitive NPEPPS MP100, PSA 4507 2.57
androgen receptor, dihydrotestosterone receptor DHTR, NR3C4 932 1.93
acid phosphatase, prostate ACPP 546 2.22
gonadotropin-releasing hormone GNRH1 GNRH, GRH, LHRH, LNRH 522 1.24
1, leutinizing-releasing hormone tumor protein p53 TP53 P53 431 0.96
B-cell CLL/lymphoma 2 BCL2 346 1.17
insulin INS 318 0.05
epidermal growth factor, beta- urogastrone EGF URG 251 0.72
cyclin-dependent kinase inhibitor 1A CDKN1A CAP20, CDKN1, CIP1, MDA-6, P21, SDI1, WAF1 190 0.98
==== Refs
Brown PO Botstein D Exploring the new world of the genome with DNA microarrays Nat Genet 1999 21 33 37 9915498 10.1038/4462
Duggan DJ Bittner M Chen Y Meltzer P Trent JM Expression profiling using cDNA microarrays Nat Genet 1999 21 10 14 9915494 10.1038/4434
de Bruijn B Martin J Getting to the (c)ore of knowledge: mining biomedical literature Int J Med Inf 2002 67 7 18 10.1016/S1386-5056(02)00050-3
Hirschman L Park JC Tsujii J Wong L Wu CH Accomplishments and challenges in literature data mining for biology Bioinformatics 2002 18 1553 1561 12490438 10.1093/bioinformatics/18.12.1553
Mack R Hehenberger M Text-based knowledge discovery: search and mining of life-sciences documents Drug Discov Today 2002 7 S89 S98 12047886 10.1016/S1359-6446(02)02286-9
Shatkay H Feldman R Mining the biomedical literature in the genomic era: an overview J Comput Biol 2003 10 821 855 14980013 10.1089/106652703322756104
Pearson H Biology's name game Nature 2001 411 631 632 11395736 10.1038/35079694
Wain HM Lush MJ Ducluzeau F Khodiyar VK Povey S Genew: the Human Gene Nomenclature Database, 2004 updates Nucleic Acids Res 2004 32 D255 D257 14681406 10.1093/nar/gkh072
Weeber M Schijvenaars BJ Van Mulligen EM Mons B Jelier R Van Der Eijk CC Kors JA Ambiguity of Human Gene Symbols in LocusLink and MEDLINE: Creating an Inventory and a Disambiguation Test Collection Proc AMIA Symp 2003 704 708 14728264
Liu H Johnson SB Friedman C Automatic resolution of ambiguous terms based on machine learning and conceptual relations in the UMLS J Am Med Inform Assoc 2002 9 621 636 12386113 10.1197/jamia.M1101
Chang JT Schutze H Altman RB Creating an online dictionary of abbreviations from MEDLINE J Am Med Inform Assoc 2002 9 612 620 12386112 10.1197/jamia.M1139
Pustejovsky J Castano J Cochran B Kotecki M Morrell M Automatic extraction of acronym-meaning pairs from MEDLINE databases Medinfo 2001 10 371 375 11604766
Wren JD Garner HR Heuristics for identification of acronym-definition patterns within text: towards an automated construction of comprehensive acronym-definition dictionaries Methods Inf Med 2002 41 426 434 12501816
Tanabe L Wilbur WJ Generation of a large gene/protein lexicon by morphological pattern analysis J Bioinform Comput Biol 2004 1 611 626 15290756 10.1142/S0219720004000399
Yeganova L Smith L Wilbur WJ Identification of related gene/protein names based on an HMM of name variations Comput Biol Chem 2004 28 97 107 15130538 10.1016/j.compbiolchem.2003.12.003
Zhou G Zhang J Su J Shen D Tan C Recognizing names in biomedical texts: a machine learning approach Bioinformatics 2004 20 1178 1190 14871877 10.1093/bioinformatics/bth060
Yandell MD Majoros WH Genomics and natural language processing Nat Rev Genet 2002 3 601 610 12154383
Van Der Eijk CC Van Mulligen EM Kors JA Mons B Van Den Berg J Constructing an associative concept space for literature-based discovery J Am Soc Inf Sci Technol 2004 55 436 444 10.1002/asi.10392
Jelier R Jenster G Dorssers LC Van Der Eijk CC Van Mulligen EM Mons B Kors JA Co-occurrence based meta-analysis of scientific texts: retrieving biological relationships between genes Bioinformatics 2005
Swanson DR Medical literature as a potential source of new knowledge Bull Med Libr Assoc 1990 78 29 37 2403828
Chaussabel D Sher A Mining microarray expression data by literature profiling Genome Biol 2002 3 RESEARCH 0055 12372143 10.1186/gb-2002-3-10-research0055
Becker KG Hosack DA Dennis G JrLempicki RA Bright TJ Cheadle C Engel J PubMatrix: a tool for multiplex literature mining BMC Bioinformatics 2003 4 61 14667255 10.1186/1471-2105-4-61
Jenssen TK Laegreid A Komorowski J Hovig E A literature network of human genes for high-throughput analysis of gene expression Nat Genet 2001 28 21 28 11326270 10.1038/88213
Masys DR Welsh JB Lynn FJ Gribskov M Klacansky I Corbeil J Use of keyword hierarchies to interpret gene expression patterns Bioinformatics 2001 17 319 326 11301300 10.1093/bioinformatics/17.4.319
Raychaudhuri S Chang JT Imam F Altman RB The computational analysis of scientific literature to define and recognize gene expression clusters Nucleic Acids Res 2003 31 4553 4560 12888516 10.1093/nar/gkg636
Hu Y Hines LM Weng H Zuo D Rivera M Richardson A LaBaer J Analysis of genomic and proteomic data using advanced literature mining J Proteome Res 2003 2 405 412 12938930 10.1021/pr0340227
Glenisson P Coessens B Van Vooren S Mathys J Moreau Y De Moor B TXTGate: profiling gene groups with text-based information Genome Biol 2004 5 R43 15186494 10.1186/gb-2004-5-6-r43
Tanabe L Scherf U Smith LH Lee JK Hunter L Weinstein JN MedMiner: an Internet text-mining tool for biomedical information, with application to gene expression profiling Biotechniques 1999 27 1210 1217 10631500
Chiang JH Yu HC Hsu HJ GIS: a biomedical text-mining system for gene information discovery Bioinformatics 2004 20 120 121 14693818 10.1093/bioinformatics/btg369
Lin SM McConnell P Johnson KF Shoemaker J MedlineR: an open source library in R for Medline literature data mining Bioinformatics 2004 20 3659 3661 15284107 10.1093/bioinformatics/bth069
Stapley BJ Benoit G Biobibliometrics: information retrieval and visualization from co-occurrences of gene names in Medline abstracts Pac Symp Biocomput 2000 529 540 10902200
Iliopoulos I Enright AJ Ouzounis CA Textquest: document clustering of Medline abstracts for concept discovery in molecular biology Pac Symp Biocomput 2001 384 395 11262957
Raychaudhuri S Schutze H Altman RB Using text analysis to identify functionally coherent gene groups Genome Res 2002 12 1582 1590 12368251 10.1101/gr.116402
Liu G Loraine AE Shigeta R Cline M Cheng J Valmeekam V Sun S Kulp D Siani-Rose MA NetAffx: Affymetrix probesets and annotations Nucleic Acids Res 2003 31 82 86 12519953 10.1093/nar/gkg121
The Online Plain Text English Dictionary
Zeeberg BR Riss J Kane DW Bussey KJ Uchio E Linehan WM Barrett JC Weinstein JN Mistaken Identifiers: Gene name errors can be introduced inadvertently when using Excel in bioinformatics BMC Bioinformatics 2004 5 80 15214961 10.1186/1471-2105-5-80
National Library of Medicine
Human and Animal Cell Line Names
Eisen MB Spellman PT Brown PO Botstein D Cluster analysis and display of genome-wide expression patterns Proc Natl Acad Sci U S A 1998 95 14863 14868 9843981 10.1073/pnas.95.25.14863
European Bioinformatics Institute
Gene Ontology
BioCarta
Smid M Dorssers LC Jenster G Venn Mapping: clustering of heterologous microarray data based on the number of co-occurring differentially expressed genes Bioinformatics 2003 19 2065 2071 14594711 10.1093/bioinformatics/btg282
Hanley JA McNeil BJ The meaning and use of the area under a receiver operating characteristic (ROC) curve Radiology 1982 143 29 36 7063747
Jansen E Laven JS Dommerholt HB Polman J Van Rijt C Van Den HC Westland J Mosselman S Fauser BC Abnormal gene expression profiles in human ovaries from polycystic ovary syndrome patients Mol Endocrinol 2004 18 3050 3063 15308691 10.1210/me.2004-0074
Guzick DS Polycystic ovary syndrome Obstet Gynecol 2004 103 181 193 14704263
Solomon CG The epidemiology of polycystic ovary syndrome. Prevalence and associated disease risks Endocrinol Metab Clin North Am 1999 28 247 263 10352918
Zdobnov EM Lopez R Apweiler R Etzold T The EBI SRS server – recent developments Bioinformatics 2002 18 368 373 11847095 10.1093/bioinformatics/18.2.368
Smalheiser NR Swanson DR Using ARROWSMITH: a computer-assisted approach to formulating and assessing scientific hypotheses Comput Methods Programs Biomed 1998 57 149 153 9822851 10.1016/S0169-2607(98)00033-9
Swanson DR Fish oil, Raynaud's syndrome, and undiscovered public knowledge Perspect Biol Med 1986 30 7 18 3797213
Srinivasan P Libbus B Mining MEDLINE for implicit links between dietary substances and diseases Bioinformatics 2004 20 I290 I296 15262811 10.1093/bioinformatics/bth914
Weeber M Vos R Klein H De Jong-Van Den Berg LT Aronson AR Molema G Generating hypotheses by discovering implicit associations in the literature: a case report of a search for new potential therapeutic uses for thalidomide J Am Med Inform Assoc 2003 10 252 259 12626374 10.1197/jamia.M1158
Wren JD Bekeredjian R Stewart JA Shohet RV Garner HR Knowledge discovery by automated identification and ranking of implicit relationships Bioinformatics 2004 20 389 398 14960466 10.1093/bioinformatics/btg421
Affymetrix
HUGO Gene Nomenclature Committee
Karolinska Institiute Alphabetic List of Specific Diseases/Disorders
Medical Subject Headings
|
15760478
|
PMC1274248
|
CC BY
|
2021-01-04 16:02:48
|
no
|
BMC Bioinformatics. 2005 Mar 11; 6:51
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-51
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-521576299310.1186/1471-2105-6-52SoftwareWEBnm@: a web application for normal mode analyses of proteins Hollup Siv Midtun [email protected] Gisle [email protected] Nathalie [email protected] Computational Biology Unit, Bergen Center for Computational Science, University of Bergen, Thormøhlensgt.55, N-5008 Bergen, Norway2005 11 3 2005 6 52 52 21 12 2004 11 3 2005 Copyright © 2005 Hollup et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Normal mode analysis (NMA) has become the method of choice to investigate the slowest motions in macromolecular systems. NMA is especially useful for large biomolecular assemblies, such as transmembrane channels or virus capsids. NMA relies on the hypothesis that the vibrational normal modes having the lowest frequencies (also named soft modes) describe the largest movements in a protein and are the ones that are functionally relevant.
Results
We developed a web-based server to perform normal modes calculations and different types of analyses. Starting from a structure file provided by the user in the PDB format, the server calculates the normal modes and subsequently offers the user a series of automated calculations; normalized squared atomic displacements, vector field representation and animation of the first six vibrational modes. Each analysis is performed independently from the others and results can be visualized using only a web browser. No additional plug-in or software is required. For users who would like to analyze the results with their favorite software, raw results can also be downloaded. The application is available on . We present here the underlying theory, the application architecture and an illustration of its features using a large transmembrane protein as an example.
Conclusion
We built an efficient and modular web application for normal mode analysis of proteins. Non specialists can easily and rapidly evaluate the degree of flexibility of multi-domain protein assemblies and characterize the large amplitude movements of their domains.
==== Body
Background
Molecular modeling provides several powerful tools for computing the dynamics of proteins. Normal Mode Analysis (NMA) is a well suited approach to study dynamics of proteins, especially when the protein is relatively big (several thousand amino acids) and the time scale of the dynamical events of interest are longer than what molecular dynamics (MD) simulations can reach, typically a few nanoseconds. These methods are based on the hypothesis that the vibrational normal modes exhibiting the lowest frequencies (also named soft modes) describe the largest movements in a protein and are the ones functionally relevant.
Several tools based on NMA have been developed [1-16] and successfully applied to predict the collective, large amplitude motions of several macromolecules of different sizes, e.g. the F(1)-APTase[17], RNA polymerases[18] or bigger systems such as virus capsids[19]. Lately, web tools have appeared making this technique accessible to a larger number of users. The elNémo[20], web interface to the Elastic Network Model, offers normal modes calculations and a fairly large number of analyses for each calculated mode; degree of collectivity, animation (PDB downloadable files or animated GIF images) for each mode using three different views for the protein, comparison between experimental and predicted B-factors, maximum distance fluctuation between all pairs of Cα atoms and normalized mean squared atomic displacements. If two structures are uploaded, the cumulative overlap between the modes and the conformational difference is calculated. Delarue et al. [21] have developed another application based on the Elastic Network Model. The application offers calculations of normal modes on all atoms (the users can also choose to use only Cα) and provides an animation for each calculated mode (PDBmovies) that can be visualized with e.g. PyMol. The same group has developed a server performing normal modes calculations using a more general molecular mechanics force field, Gromacs, and which also provides animation of the vibrations corresponding to each calculated mode. The use of such a force field increases the computational cost of the computation and the system size is therefore limited to 5000 atoms. The NMA movie generator, available from the web pages of the database of macromolecular movements (MolMovDB[22]), calculates the five lowest frequency normal modes for a PDB structure file which can be either uploaded to the server or chosen by its PDB or SCOP identifiers. Animated GIF images of the vibrations are generated and compared with the pre-calculated flexibility regions based on supplied B-factors or multiple structural alignments for the corresponding fold family for one-domain fold proteins.
The Molecular Vibrations Evaluation Server (MoVies[23]) provides vibrational study of proteins and nucleic acids, using modified AMBER force field[24] and a self-consistent harmonic approximation method. Starting from a structure file in the PDB format, the application performs normal modes calculations and several analyses, and on completion the results are sent to the user by email. Of special interest is the evaluation of hydrogen bond disruption probability.
The ProMode database [25] is a database of normal mode analysis of proteins. Results of normal mode analysis for a large number of proteins are made accessible via a web interface. For each mode, an animation and the axes of the movement (as calculated by DynDom[26]) can be viewed using the Chime plugin. Fluctuations of atom positions and torsion angles, correlation between Cα atom displacements are plotted for each mode; the averages of these values over all modes are also stored in the database. Dynamical domains for each mode, characterized using DynDom, are given. Although NMA results for a large number of proteins can be very quickly retrieved from ProMode, not all proteins available in the Protein Data Bank are present and users cannot submit their own structure file.
We developed a web application for calculation of normal mode analysis which offers fast calculation of the 200 lowest frequency modes and different types of analyses: deformation energy, animation of the vibration, atomic squared displacements and vector field analysis. Results of each analysis can be visualized using only a web browser, without any additional plug-in or program. Alternatively, the users can download raw data and visualize them using their favorite software. We have carefully designed our web application into independent modules so that the users can perform only the analyses they are interested in, and in this way avoid spending time waiting for results of analysis irrelevant to their particular question. The modular structure will, in the future, allow us to easily add new functionality. The core of the application is written in the Python programming language, using the Molecular Modeling ToolKit [27] (MMTK). It contains an implementation of the approximate normal analysis method developed by Hinsen[10] which calculates low-frequency domain motions at negligible computational cost. Zope[28] is used for the web interface, which communicates with the core through an application server. Details of the implementation are given below, followed by an example calculation on a large transmembrane protein.
Implementation
1. Web-interface
The first step for the user is to upload a pdb file containing the structure. Pressing the submit button starts the normal mode calculation, which runs to completion without doing any further analysis. No limit is set for the system size (i.e. number of residues). When the calculation is finished, the user is directed to a page which displays the result of the energy deformation analysis. Low average deformation energy indicates a mode with large rigid regions, i.e. a mode with a large degree of collectivity, which has a good chance of describing domain motions. This page is meant to help users judge for which mode(s), if any, the analysis will be significant in terms of large collective movements. They can then decide to perform further analysis of the calculated modes and are given the possibility to choose among three different analyses (see description below). Results of each analysis are stored and can at any time be viewed either in a separate window, or downloaded as a ZIP archive together with results of all other analyses performed up to that moment.
Normalized squared atomic displacements can be retrieved in two different formats. Users can download text files containing two columns, the first one corresponding to the amino acid numbers of the sequence in the structure file (PDB) submitted and the second one containing the normalized displacement corresponding to each amino acid. Alternatively, the user can retrieve PDF plots representing the variation of normalized atomic displacements vs. amino acid number. These plots are generated using the R programming language[29] and RPy [30], a Python interface to R. Thus, we provide the users with the possibility to see the results directly from their web browser without any additional plugins or program, but we also, for users who want to have more flexibility, provide the raw data.
Mode animations are provided for the six first significant modes (i.e. modes 7 to 12, see Methods section), as animated gif images or as DCD trajectory files. The DCD file format is a binary format for trajectories from MD simulations that is common to the CHARMm[1], XPlor[31] and NAMD[32] programs. DCD files can be read by VMD[33]. Unlike with animated gifs, visualizing DCD files with VMD allows the users to manipulate the protein themselves (rotate, zoom, highlight specific regions, etc..) which might offer a better insight in the calculated domain movements. On the other hand, this requires that the user has VMD installed on his computer and is sufficiently used to it. Therefore, we have decided to offer the possibility to choose the orientation of the protein before the animated gif images are generated. Rasmol[34,35] is used to generate image files of the different conformations along the mode vector (see Methods section). The images are then concatenated to produce an animation (animated GIF file) using Image Magick [36]. The resulting animation is a sequence of five conformations, with a delay of 1/25 second between them.
Vector field representations help characterize the domain displacements with vectors representing the direction and the relative displacements of the different regions of the protein. Using VMD, the web application generates a picture of the protein and the vectors for modes 7 to 12. Using the same setup as for the mode animations, the user can choose the orientation of his system. Additionally, VMD 'state' files are generated and available for download, allowing a more interactive inspection of the vector fields.
2. Application server
The web interface of WEBnm@ is written using the DTML language of the Zope[28] webserver. The analysis core, written in Python, runs under the BIAZ application server[37]. BIAZ is connected to Zope using a pipe (see Figure 1). The purpose of the BIAZ application server is to simplify the development of web interfaces for computationally demanding applications; it has been developed and is used to run the computational services of the Norwegian Bioinformatics Platform . BIAZ itself is written in Common Lisp(CL), and applications in CL or Python are currently supported. The application server fetches the results after completion of the computation and sends them to the web interface (Zope). The division between core application and web interface also makes the code more readable, and thus maintainable. The core application code becomes thereby usable in other contexts as well.
Results: example calculation on SERCA1 Ca-ATPase
The calcium ATPase from the sarcoplasmic reticulum, is constituted of 3 cytoplasmic domains, named Actuator (A, amino acids 1 to 40 (NTer) and 124 to 243), Nucleotidic (N, 360 to 604) and Phosphorylation (P, 330 to 359 and 605 to 737), and 10 transmembrane helices hosting the calcium binding sites. It is known that the cytoplasmic domains undergo large amplitude movements during the active transport of calcium ions. We recently reported a NMA study of the E1Ca form of the Ca-ATPase, starting from its x-ray structure (PDB ref 1EUL) [38]. Using MMTK, we could show that the N and A domains undergo the largest amplitude movements, as revealed by the lowest frequency modes. We highlighted a large amplitude movement of the transmembrane helices, which "twist-opens" the lumenal side of the protein.
In what follows, we explain how to use WEBnm@ to perform the same type of analysis (we use here the PDB ID 1SU4, instead of 1EUL) and especially how to interpret the results given by our application. We show that we obtain the same results with WEBnm@ as we obtained using a non automated procedure [38]. After the uploading of the structure file (PDB format) on the main page (Figure 1), normal modes are calculated. The server is directed to an html page with a table containing deformation energies for modes 7 through 20. The deformation energy of a mode is a measure of the collectivity of the movements associated to this mode. The lower the deformation energy, the higher the degree of collectivity. A high degree of collectivity means that large regions of the protein, possibly domains, are displaced. Although the deformation energies have no quantitative physical meaning (and therefore no unit), values obtained on different proteins can be compared. In our example (Cf. Figure 2a), the value of the deformation energy for the first mode is extremely low (135.2). In comparison, the deformation energy of the first mode for lysozyme is 2378.5 (pdb id: 153l), 795.0 for the MscL (pdb id: 1msl) and 5881.7 for trypsin (pdb id: 1anb), which is not known to undergo large amplitude domains movements.
The user can then choose to proceed to further analyis (Cf. Figure 2b), for example generate an animation for each of the 6 first modes (7 through 12). The next page (Figure 2c) offers the users the possibility to orient the system properly to ensure the best view of the movements by choosing a rotation angle over the x, y and z axes. A preview will be generated for each chosen set of angles. Once the user has decided upon a set of angles, he can check the 'I'm done' radio button, and then press the 'Perform' button and animations will be generated. The user is then brought back to the 'Analysis' page (Figure 2d) where a logo has now appeared next to 'Mode Animation'. By clicking on this icon, a new window containing the animated images (gif format) will be opened (Figure 2e). This goes for all additional analyses. A click on an icon opens a new window with the results of the corresponding analysis. At any moment, one can download the analyses performed up to that point as a ZIP archive that contains all result files.
Figure 3 displays the plot obtained by calculating the normalized atomic squared displacements. For example, one can see that the displacements associated with modes 7 (top left plot) concern mostly the domain N (aminoacids number 360 to 604) and the L1–2 (aa 78 to 89), L7–8(aa 852 to 896) and L9–10(aa 949 to 965) loops.
Conclusion
WEBnm@ allows efficient calculation of normal modes for proteins and is available to everyone from . Calculation of the modes for the Ca-ATPase, which contains 994 residues, takes about 4 minutes. Our web application has several other advantages; a user can choose which analyses to perform so that no time is wasted on analysis he/she is not interested in. Result pages for each analysis are independent and open in separate windows. All results are presented on the web pages, no additional programs or plugins are needed for visualization. However, results are also provided in other formats (x, y format for normalized squared atomic displacements, PDB for structure and DCD for trajectories) in case users want to use their favorite program to visualize and analyze their results. This allows anyone to calculate normal modes for relatively large systems, without having the required resources (i.e. memory) to do it in-house. At any time, result files of the calculation performed up to that moment can be downloaded in a ZIP file. Although WEBnm@ is not the first tool of his kind, it is probably the fastest and provides functionalities that are not found elsewhere.
The architecture of WEBnm@ is totally modular. It is meant to welcome an increasing number of functionalities (structure comparison between different conformations of a protein, domain determination, etc...). Decision on future developments will also be based on users' requests.
Methods
Normal modes calculations
A normal mode analysis (NMA) consists of the diagonalization of the matrix of the second derivatives of the energy with respect to the displacements of the atoms, in mass-weighted coordinates (Hessian matrix). The eigenvectors of the Hessian matrix are the normal modes, and its eigenvalues are the squares of the associated frequencies. We use the approximate normal modes calculation method developed by Hinsen [10] and implemented in the MMTK package[27]. This method represents the low-frequency domain motions very well at negligible computational cost. The force field used is slightly different from the one used in the original publication and has been described in reference [13]. It uses only the Cα atoms of the protein, which are assigned the masses of the whole residues they represent.
Briefly, the functional form of the force field is
V(r) is the harmonic pair potential describing the interaction between the Cα atoms:
where is the pair distance vector (Ri - Rj) in the input configuration and k is the pair force constant:
Two hundred modes are calculated for proteins containing less than 1200 residues. For proteins containing more than 1200 residues, N/6 modes are calculated (N being the number of residues). The first six modes (zero-frequency modes) correspond to global rotation and translation of the system and are ignored in the analyses. Thus, the lowest frequency mode of interest is mode 7. Deformation energy and normalized atomic displacements analyses are performed for modes 7 through 20 while mode animations and vector fields are calculated for modes 7 through 12.
Deformation energy
As in DomainFinder[10,11], a deformation energy is calculated for each atom. Deformation energy depends on the changes in the distance between the atom in question and each of its close neighbors. Low deformation energies indicate relatively rigid regions, whereas high deformation energies indicate flexible regions. The application returns the average deformation energy for each mode. Low average deformation energy indicates a mode with large rigid regions, which has a good chance of describing domain motions.
Normalized squared atomic displacements
Normalized squared atomic displacements (Di) for each amino acid (resid) or Cα atom (i = 1 to n) are calculated as follows:
where di is the component of the eigenvector corresponding to the ith residue.
Normal mode animations
Subsequent structures of a given animation are generated by applying eigenvectors of the corresponding mode to the Cα coordinates of the structure submitted to the server. Two structures of the protein are generated in each direction (i.e. +a*mode, +2*a*mode, -a*mode, -2*a*mode). The 'a' factor is arbitrary; we choose to set it equal to 10 as a default value since this gives the best visual insight on the movements.
Vector fields
A vector field representation is calculated as described by Thomas et al. [39]. The vector field is calculated over cubic regions with an edge length of 3 Å, containing on average 1.3 Cα atoms. The vector field defined on a regular lattice at the center of each cube is the mass-weighted average of the displacements of the atoms in the cube.
Authors' contributions
SMH designed the modular architecture, developed the graphical presentation of results for mode animations and vector field representations, and served as the main driving force in building the latest version of WEBnm@. GS is the developer of the BIAZ application server; several features of BIAZ were developed especially for WEBnm@. NR wrote the core analysis code, designed and supervised the project, and edited the manuscript. This work is a truly collaborative effort of all three authors. All authors read and approved the final manuscript.
Acknowledgements
Funding for this work was provided by FUGE (Norwegian functional genomics program) through the technology platform for bioinformatics. Inge Jonassen and Konrad Hinsen are thankfully acknowledged for their pertinent advices and careful reading of our manuscript.
Figures and Tables
Figure 1 WEBnm@ architecture. a. The main page of WEBnm@ is a form where users can input a structure file in the PDB format. b. The server consists of two parts, the graphical web interface and the core of the program, written in Python, which performs the actual computation. The two parts communicate via a web application server, BIAZ.
Figure 2 Snapshots of an example calculation of the SERCA1 Ca-ATPase. a. Presentation of the 10 lowest frequency modes with their average deformation energy. b. Page presenting the available analyses. c. Page where user can choose the orientation of the system for the animations. d. Apparition of an icon on the page presenting the list of analyses. e. Page displaying animated gif image of the first 6 modes (7 to 12). f. List of analyses page after that both mode animations and atomic displacements have been calculated, two icons are present.
Figure 3 Normalized atomic displacements plots. Plots for modes 7 to 12 are generated on the same page and converted to a PDF file.
==== Refs
Brooks BR Bruccoleri RE Olafson BD States DJ Swaminathan S Karplus M CHARMM: A Program for Macromolecular Energy, Minimization, and Dynamics Calculations J Comput Chem 1983 4 187 217 10.1002/jcc.540040211
Go N Noguti T Nishikawa T Dynamics of a small globular proteins in terms of low frequency normal modes Proc Natl Acad Sci USA 1983 80 3696 3700 6574507
Levitt M Sander C Stern PS Protein normal-mode dynamics: trypsin inhibitor, crambin, ribonuclease and lysozyme J Mol Biol 1985 181 423 447 2580101 10.1016/0022-2836(85)90230-X
Schulz GE Domain motions in proteins Curr Opin Struct Biol 1991 1 883 888 10.1016/0959-440X(91)90082-5
Mouawad L Perahia D Diagonalization in a mixed basis: a method to compute normal-modes for large macromolecules Biopolymers 1993 33 599 611 10.1002/bip.360330409
Marques O Sanejouand YH Hinge-bending motion in citrate synthase arising from normal mode calculations Proteins 1995 23 557 560 8749851
Tirion MM Large amplitude elastic motions in proteins from a single-parameter, atomic analysis Phys Rev Lett 1996 77 1905 1908 10063201 10.1103/PhysRevLett.77.1905
Bahar I Atilgan AR Erman B Direct evaluation of thermal fluctuations in proteins using a single-parameter harmonic potential Folding Des 1997 2 173 181
Cornell WD Louise-May S P. Schleyer NATCJGPKHSPS Normal Mode Analysis Encyclopedia of Computational Chemistry 1998 Chichester, UK, John Wiley & Sons. 1904 1913
Hinsen K Analysis of domain motions by approximate normal mode calculations Proteins 1998 33 417 429 9829700 10.1002/(SICI)1097-0134(19981115)33:3<417::AID-PROT10>3.0.CO;2-8
Hinsen K Thomas A Field MJ Analysis of domain motions in large proteins Proteins 1999 34 369 382 10024023 10.1002/(SICI)1097-0134(19990215)34:3<369::AID-PROT9>3.0.CO;2-F
Berendsen HJ Hayward S Collective protein dynamics in relation to function Curr Opin Struct Biol 2000 10 165 169 10753809 10.1016/S0959-440X(00)00061-0
Hinsen K Petrescu AJ Dellerue S Bellissent-Funel MC Kneller GR Harmonicity in slow protein dynamics Chem Phys 2000 261 25 37 10.1016/S0301-0104(00)00222-6
Tama F Gadea FX Marques O Sanejouand YH Building-block approach for determining low-frequency normal modes of macromolecules Proteins 2000 41 1 7 10944387 10.1002/1097-0134(20001001)41:1<1::AID-PROT10>3.0.CO;2-P
Hayward S Becker OM, MacKerell AD, Roux B and Watanabe M Normal mode analysis of biological molecules Computational biochemistry and biophysics 2001 New-York, Marcel Dekker, Inc. 153 168
Li G Cui Q A coarsed-grained normal mode approach for macromolecules: an efficient implementation and application to Ca++-ATPase Biophys J 2002 83 2457 2474 12414680
Cui Q Li G Ma J Karplus M A normal mode analysis of structural plasticity in the biomolecular motor F(1)-ATPase J Mol Biol 2004 340 345 372 15201057 10.1016/j.jmb.2004.04.044
Yildirim Y Doruker P Collective motions of RNA polymerases. Analysis of core enzyme, elongation complex and holoenzyme J Biomol Struct Dyn 2004 22 267 280 15473702
Tama F Brooks CL Diversity and identity of mechanical properties of icosahedral viral capsids studied with elastic network normal mode analysis J Mol Biol 2005 345 299 314 15571723 10.1016/j.jmb.2004.10.054
Suhre K Sanejouand YH ElNemo: a normal mode web server for protein movement analysis and the generation of templates for molecular replacement Nucleic Acids Res 2004 32 W610 4 15215461 10.1093/nar/gkh053
Delarue M Lindahl E Normal mode calculation and visualisation using Pymol 2004
Echols N Milburn D Gerstein M MolMovDB: analysis and visualization of conformational change and structural flexibility Nucleic Acids Res 2003 31 478 482 12520056 10.1093/nar/gkg104
Cao ZW Xue Y Han LY Xie B Zhou H Zheng CJ Lin HH Chen YZ MoViES: molecular vibrations evaluation server for analysis of fluctuational dynamics of proteins and nucleic acids Nucleic Acids Research 2004 32 W679 W685 15215475
Cornell W Cieplak P Bayly C Gould I Merz KM Ferguson D Spellmeyer D Fox T Caldwell J Kollman P A second generation force field for the simulation of proteins and nucleic acids J Am Chem Soc 1995 117 5179 5197 10.1021/ja00124a002
Wako H Kato M Endo S ProMode: a database of normal mode analyses on protein molecules with a full-atom model Bioinformatics 2004 20 2035 2043 15059828 10.1093/bioinformatics/bth197
Hayward S Berendsen HJ Systematic analysis of domain motions in proteins from conformational change: new results on citrate synthase and T4 lysozyme Proteins 1998 30 144 154 9489922 10.1002/(SICI)1097-0134(19980201)30:2<144::AID-PROT4>3.0.CO;2-N
Hinsen K The Molecular Modeling Toolkit : a new approach to molecular simulations J Comput Chem 2000 21 79 85 10.1002/(SICI)1096-987X(20000130)21:2<79::AID-JCC1>3.0.CO;2-B
Zope Open Source web application server.
R Copyright (C) 1989, 1991 Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
Moreira W Warnes GR RPy 2003
Brünger AT "XPLOR Manual Version 3.1" Yale UNiversity Press; New Haven 1993
Kalé L Skeel R Bhandarkar M Brunner R Gursoy A Krawetz N Phillips J Shinozaki A Varadarajan K Schulten K NAMD2: Greater scalability for parallel molecular dynamics. Journal of Computational Physics 1999 151 283 312 10.1006/jcph.1999.6201
Humphrey W Dalke A Schulten K VMD - Visual Molecular Dynamics J Molec Graphics 1996 14 33 38 10.1016/0263-7855(96)00018-5
Sayle R Milner-White EJ RasMol: Biomolecular graphics for all TIBS 1995 20 374 7482707
Bernstein HJ Recent changes to RasMol, recombining the variants TIBS 2000 9 453 455
Cristy J Randers-Pehrson G Image Magick 2003
Saelensminde G The Biaz application server 2003
Reuter N Hinsen K Lacapere JJ Transconformations of the SERCA1 Ca-ATPase: a normal mode study Biophys J 2003 85 2186 2197 14507684
Thomas A Hinsen K Field MJ Perahia D Tertiary and quaternary conformational changes in aspartate transcarbamylase : a normal mode study. Proteins 1999 34 96 112 10336386 10.1002/(SICI)1097-0134(19990101)34:1<96::AID-PROT8>3.0.CO;2-0
|
15762993
|
PMC1274249
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 11; 6:52
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-52
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-531576298510.1186/1471-2105-6-53Research ArticleEvolutionary sequence analysis of complete eukaryote genomes Blair Jaime E [email protected] Prachi [email protected] S Blair [email protected] NASA Astrobiology Institute and Department of Biology, The Pennsylvania State University, 208 Mueller Laboratory, University Park, Pennsylvania 16802-5301, USA2005 11 3 2005 6 53 53 25 10 2004 11 3 2005 Copyright © 2005 Blair et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Gene duplication and gene loss during the evolution of eukaryotes have hindered attempts to estimate phylogenies and divergence times of species. Although current methods that identify clusters of orthologous genes in complete genomes have helped to investigate gene function and gene content, they have not been optimized for evolutionary sequence analyses requiring strict orthology and complete gene matrices. Here we adopt a relatively simple and fast genome comparison approach designed to assemble orthologs for evolutionary analysis. Our approach identifies single-copy genes representing only species divergences (panorthologs) in order to minimize potential errors caused by gene duplication. We apply this approach to complete sets of proteins from published eukaryote genomes specifically for phylogeny and time estimation.
Results
Despite the conservative criterion used, 753 panorthologs (proteins) were identified for evolutionary analysis with four genomes, resulting in a single alignment of 287,000 amino acids. With this data set, we estimate that the divergence between deuterostomes and arthropods took place in the Precambrian, approximately 400 million years before the first appearance of animals in the fossil record. Additional analyses were performed with seven, 12, and 15 eukaryote genomes resulting in similar divergence time estimates and phylogenies.
Conclusion
Our results with available eukaryote genomes agree with previous results using conventional methods of sequence data assembly from genomes. They show that large sequence data sets can be generated relatively quickly and efficiently for evolutionary analyses of complete genomes.
==== Body
Background
The use of complete genomes for phylogenetic analysis has greatly improved our understanding of prokaryote evolution [1-3]. However, until recently, relatively few complete genome sequences were available for such analyses in eukaryotes. As this improves, there will be a greater demand on methodology for evolutionary analysis of complete genomes. Previous whole-genome studies of eukaryotes have focused on gene and gene family presence-absence [4-7], lineage-specific gene loss [8,9], insertion-deletion markers and introns [6,10,11], and other non-sequence based information. While these approaches have their advantages, previous studies have not used complete genome sequences (nucleotides and/or amino acids) for reconstructing evolutionary relationships. At the same time, the complexity of eukaryote genomes, with numerous gene duplications and losses in different lineages, has created a challenge for sequence-based phylogeny estimation. Here, we outline a conservative approach designed to utilize the wealth of evolutionary information present in complete genome sequences by identifying orthologs in multiple eukaryotes for the purpose of evolutionary analysis.
Methods for the identification of clusters of orthologs and lineage-specific paralogs have proven useful for classifying gene function and identifying cases where genes have been differentially lost or duplicated in different lineages [12-14]. However, such assemblages of data contain a mixture of orthologs, paralogs, and missing data as a result of gene loss, and are not generally suitable for large-scale phylogenetic sequence analysis of organismal evolution. Our approach for comparing multiple genome sequences involves the identification of single-copy orthologs across a number of genomes for evolutionary analysis (Figure 1). We refer to such strict (1:1) orthologs as panorthologs, in reference to their presumed "complete" orthology, in contrast to synorthologs, which contain a mixture of species divergences and gene duplication events. In other words, panorthologs are those genes (or clusters of sequences) that contain only species divergences and do not contain in-paralogs, out-paralogs, or co-orthologs [15]. On the other hand, synorthologs are those genes (or clusters of sequences) that contain species divergences and any combination of paralogy (in-paralogs and out-paralogs). While the use of panorthologs is conservative and reduces the number of usable genes or proteins, it also lowers the probability that errors will be made in confusing a species divergence with a gene duplication event. Because the ability to identify orthologs is diminished in analyses of small to moderate numbers of species or genomes, such a conservative method is appropriate in those cases. This conservative approach has been used to identify the number of shared, unduplicated proteins in Homo sapiens, Drosophila melanogaster, Caenorhabditis elegans, and Saccharomyces cerevisiae, where it was determined that such proteins perform primarily anabolic rather than catabolic functions [16].
We compare our phylogenetic results and divergence time estimates for an analysis of seventeen published eukaryote genomes to a previous study that assembled nuclear protein sequence data in a more conventional manner from public databases [17]. While the phylogenetic relationships between the organisms included in this study are not controversial, with the exception of the position of nematodes [18], this general approach will prove useful as more genomes, including those with questionable phylogenetic affinity, are sequenced. In addition, this approach facilitates the estimation of divergence times between organisms with numerous molecular clock methods.
Results
The number of orthologous clusters per pairwise comparison and the percentage of those clusters showing panorthology are presented in Table 1. On average, pairwise orthologous clusters contained approximately 60.3% panorthologs; exceptions include comparisons between fungi, including Encephalitozoon (average 89% panorthology), and all comparisons with Arabidopsis (average 34.6% panorthology). Comparisons within metazoans averaged 54.7% panorthology, with Mus and Rattus showing the highest number of shared transcripts (16,413 orthologous clusters; 79.2% panorthology) as expected due to their recent evolutionary divergence. Previous analyses showed approximately 12,400 panorthologs between Mus and Rattus [19]. Caenorhabditis elegans and C. briggsae, who diverged roughly 100 Ma [20], also shared a large number of transcripts (12,416 orthologous clusters; 84.4% panorthology), which is similar to a previous estimate of 12,155 panorthologs [21]. The number of orthologous clusters between Drosophila and Anopheles found here (7072, 61.3% panorthology) is also similar to a previous estimate of approximately 6130 panorthologs [22]. Pairwise comparisons with the smallest genome, the Guillardia nucleomorph, averaged ~176 orthologous clusters, but the percentage of panorthologs varied greatly, from a low of 25.1% with Arabidopsis to as high as 97.2% with Encephalitozoon.
The intersection of nine metazoan genomes resulted in a large number of shared genes. Among the nine genomes, 285 panorthologs were found, totaling 97,581 amino acids. The neighbor-joining tree of that concatenation is shown in Figure 2; all nodes in this tree received 100% bootstrap support. The intersection of all seventeen eukaryote genomes included in this study resulted in three shared genes (t-complex protein delta subunit, proteasome beta type-1 subunit, and Nip7p biogenesis factor) and orthology was confirmed manually. The reconstructed trees for the three genes showed long-branch attraction errors associated with the intracellular parasite Encephalitozoon and the Guillardia nucleomorph (data not shown). This was expected because both have highly reduced genomes and high rates of substitution across many genes as a result of their current or ancestral parasitic and symbiotic lifestyles [23,24]. For this reason, the intersection of the remaining fifteen genomes was determined, resulting in ten panorthologs. The intersection of genomes from twelve multicellular eukaryotes resulted in 63 panorthologs. The functional classifications of the panorthologs found here are similar to those identified in previous studies [14], with the most frequently represented functions being transcription, translation, replication and repair, and RNA processing. The phylogenetic trees reconstructed from the concatenated datasets both showed the expected relationships (Figure 3a and 3b) [17]. All nodes in these trees received very high bootstrap support, with only one node showing less than 95% bootstrap support (animals + fungi in Figure 3a). The long branch observed in Plasmodium (Figure 3a) may be the result of both the long evolutionary separation from the other eukaryotes included in this study, and the high (A-T) composition of the genome [25] leading to biased amino acid compositions among proteins [26].
Phylogenetic trees were also reconstructed for each panortholog to test for congruence with well-supported phylogenies from the concatenated data (see Additional file 1). We found that in most cases, the consensus values calculated from individual trees agree with the high bootstrap support of the concatenated analysis. Two exceptions were the low consensus values for the accepted close relationship between animals and fungi and the contested position of nematodes. Both taxa showed slightly longer branch lengths, and long-branch attraction artifacts [27] may be affecting the individual datasets, causing low consensus values. Also, recent empirical [28] and simulation [29] studies suggest that results from multigene analyses are more accurate when a tree is derived from a concatenated dataset of individual genes rather than a consensus of trees from multiple analyses.
Divergence times were estimated for both the 15-genome and 12-genome datasets (Table 2). Results were consistent with previous studies [17,30-33], showing an early divergence between plants, animals, and fungi (animals/fungi vs. plants ~1670 Ma, animals vs. fungi ~1400 Ma), and a Precambrian origin for animals (~900–1100 Ma). To specifically address the deuterostome-arthropod divergence within animals, two additional datasets were assembled to maximize the number of proteins analyzed: the intersection of seven genomes (Homo, Mus, Rattus, Takifugu, Drosophila, and Anopheles; Arabidopsis as outgroup) and the intersection of four genomes (Homo, Takifugu, and Drosophila; Arabidopsis as outgroup). The seven genome intersection contained 380 panorthologs (132,190 amino acids; Figure 4a), and yielded a vertebrate-arthropod divergence time of 955 Ma. The four genome intersection contained 753 panorthologs (287,000 amino acids; Figure 4b) and yielded a vertebrate-arthropod divergence time of ~1100 Ma. Although this last estimate was derived from more than five times the data previously used, the divergence time is remarkably consistent with previous large-scale studies [17], and suggests that bilaterian animals originated hundreds of millions of years before the first fossil evidence of their existence in the Cambrian (~520 Ma). With the exception of the maximum fossil-based time estimate used in the tetrapod-actinopterygian fish calibration, the other fossil constraints used here are minimums, and therefore the resulting time estimates are minimums [34]. The agreement between our results and those of previous studies using different methods of data assembly suggests that our genome intersection approach is correctly assembling orthologs. Younger time estimates of the vertebrate-arthropod divergence have been obtained in some studies [35-37]. However, those results are problematic because they also contain estimates which are inconsistent (too young) with undisputed aspects of the fossil record, such as the oldest red algae (1200 Ma), green algae (1000 Ma), and stramenopiles (1000 Ma) [38-40]. Possible reasons for their inconsistency are discussed elsewhere [41].
Discussion
The complete genome sequence of an organism is essentially the maximum amount of discrete, genetically encoded information available for evolutionary analyses. However, orthology determination has been a major obstacle in the analysis of complete genomes, especially in eukaryotes where considerable gene duplication and loss has created additional complexity. Our approach for evolutionary analysis of complete eukaryote genome sequences is both simple and fast compared with the conventional method of gene-by-gene orthology determination using similarity searches in the public databases. The results of this approach applied to a subset of the available eukaryote genomes show agreement with previous results using conventional (non-genomic) approaches. In addition, the relatively high consensus values for most nodes indicate general agreement in tree topology among individual panorthologs.
The relatively low number of common genes in our intersections of 12–17 genomes is a combination of using panorthologs and including distantly related species. Genes are more likely to duplicate over long periods of evolutionary time, as in the time elapsed since plants separated from animals (~1600 Ma) [17,32]. Therefore, a better approach with such distantly related species (e.g., all eukaryotes), and those groups with high levels of gene duplication and gene loss (e.g., nematodes), would be to relax the orthology criterion and include synorthologs. In that case, a representative or consensus sequence may be chosen from among in-paralogs. On the other hand, the implementation used here (panorthologs) should yield many genes in analyses of genomes from closely related species (e.g., within mammals), even if large numbers of species are used.
The use of sequence data for comparative genomics and phylogenetics has several advantages over the use of datasets based on the presence and absence or position of genes, introns and insertions. Sequence data can provide a larger number of characters for analysis, yielding hundreds of thousands of amino acid sites and more than a million nucleotide sites in some cases. Also, statistical models of sequence change are better known than those for non-sequence based data. Finally, the assembly of sequence data from complete genomes of multiple organisms not only facilitates phylogenetic and divergence time analyses, but a diversity of other comparative evolutionary analyses requiring sequence information [42,43].
Conclusion
Unlike previous studies of complete eukaryote genomes, here we have used a fast, conservative approach to assemble orthologous clusters of proteins for phylogenetic analysis and divergence time estimation. Our results are similar to previous studies that used conventional (slower) gene-by-gene data mining. We find that complete genome sequences support the close evolutionary relationship between animals and fungi, and also that molecular divergences between animals occurred approximately 400 million years before the Cambrian Explosion of fossils. Our approach will be further tested as more eukaryote genomes are sequenced.
Methods
Multigenome intersection approach for evolutionary analysis (MIA): Reciprocal BLAST [44] searches of genomes versus themselves and versus all other genomes included in the analysis were used to generate lists of pairwise similarity scores for each transcript. These scores were then used to generate orthologous clusters of proteins by first determining the "primary" ortholog pair through reciprocal best hits, then adding lineage-specific paralogs (in-paralogs) as implemented in INPARANOID [45]. The settings used here (sequence overlap cut-off 50%, group overlap cut-off 50%, in-paralog confidence cut-off 5%) were considered optimal in the sense that all closely related lineage-specific paralogs (and alternative transcripts) will be placed in the same pairwise cluster, minimizing the probability that the same gene will be represented in more than one cluster. In-paralogs are added to clusters if they are more similar to one member of the primary ortholog pair than the two primary orthologs are to each other [45].
Only pairwise ortholog clusters can be obtained using INPARANOID. For phylogenetic analysis, ortholog sets for a larger number of genomes (at least four) must be constructed. Therefore, we combined the pairwise ortholog clusters from groups of species using a relational database. The intersection between ortholog clusters was determined by iteratively comparing sequence identification tags and reducing the intersected clusters at each round to exclude clusters that represent relationships other than panorthology. For example, consider genomes A, B and C. First, ortholog clusters are determined for each pairwise genome comparison, which results in clusters: A-B, B-C and A-C. Intersection of sets A-B and B-C is obtained by searching for common sequences of genome B in the two sets and merging the two sets accordingly into an ortholog cluster set of A-B-C. This combined set is reduced to include only clusters with panorthology relationships, i.e. only clusters with one sequence for each of the genomes A, B and C are retained. Further, the combined cluster A-B-C is compared with the pairwise cluster A-C based on common sequence tags for genome C. Any cluster from the combined A-B-C set that does not agree with the sequence grouping of genomes A and C is removed. This last step serves as an important check for orthology in each iteration of the intersection procedure, and is similar to the construction of three-member COGs (clusters of orthologous groups) [46]. The steps described above were performed iteratively in order to add more species to the ortholog clusters. Any number of genomes can be intersected (tested up to seventeen here), and an outgroup can be treated as part of the intersection or added separately by using a pairwise comparison to one of the in-group taxa. All programming was written in Perl.
Analysis of Eukaryote Genomes: Complete protein transcripts were obtained for the following eukaryote genomes [three letter abbreviation]: Homo sapiens [Hsa] (version 34b.2) [16,47,48]; Mus musculus [Mmu] (version 32.2) [48,49]; Rattus norvegicus [Rno] (version 3b.1) [19,48]; Takifugu rubripes [Tru] (version 3.0) [50,51]; Ciona intestionalis [Cin] (version 1.0) [51,52]; Drosophila melanogaster [Dme] (version 3.1) [53,54]; Anopheles gambiae [Aga] (version 2a.2) [48,55]; Caenorhabditis elegans [Cel] (version 120) [56,57]; Caenorhabditis briggsae [Cbr] (version 25) [21,57]; Saccharomyces cerevisiae [Sce] [58,59]; Neurospora crassa [Ncr] (version 3) [60,61]; Ashbya gossypii [Ago] (version 1.0) [62,63]; Encephalitozoon cuniculi [Ecu] [24,64]; Arabidopsis thaliana [Ath] (version 5.0) [65,66]; Cyanidoschyzon merolae [Cme] [67,68]; Guillardia theta nucleomorph [Gtn] [23,64]; and Plasmodium falciparum 3D7 [Pfa] (version 4.1) [25,69]. Some genome transcripts were given unique sequence identifiers to avoid redundancy when sequence tags are truncated. The intersection of these seventeen genomes was determined as described above.
Each panortholog was aligned [70] and individual datasets were concatenated. Both individual panorthologs and concatenations were analyzed using maximum likelihood [71] to determine alpha parameters for the gamma rate-heterogeneity correction. Phylogenetic trees of concatenated datasets were reconstructed using neighbor-joining (Poisson + gamma correction model) with bootstrapping (2000 replicates) [72] and maximum likelihood (JTT + gamma correction model) with 1000 puzzling steps [73]. Phylogenetic trees of individual datasets were reconstructed using maximum likelihood (Poisson + gamma correction model) [71] and a consensus tree was derived [74]. Consensus values (i.e. the proportion of trees recovering a specific node) were calculated for each dataset.
Divergence times were estimated for concatenated datasets using Bayesian inference (JTT model) [75] as described previously [17]. The following fossil dates were used as minimum constraints: Mus-Rattus 12 Ma [76], primate-rodent 65 Ma [77], tetrapod-actinopterygian fish 425 Ma (lower bound) and 495 Ma (upper bound) [78], vertebrate-urochordate 520 Ma [79], Drosophila-Anopheles 250 Ma [80], chordate-arthropod 545 Ma [77], green algae/plants-red algae 1200 Ma [40].
Authors' contributions
JEB and PS designed and developed the methodology. PS programmed the genome intersection software. JEB carried out the evolutionary analyses and drafted the manuscript. SBH coordinated the research and assisted with drafting the manuscript.
Supplementary Material
Additional File 1
Consensus trees of individual gene trees. Consensus trees of individual gene (panortholog) trees, showing percentage of individual gene trees supporting each node.
Click here for file
Acknowledgements
The authors would like to thank Uthra Ramaswamy for additional computational support and Sankar Subramanian for helpful comments. This work was supported by grants to SBH from the National Science Foundation and the NASA Astrobiology Institute.
Figures and Tables
Figure 1 Flowchart of multigenome intersection approach (MIA). 1) Complete genomes are reciprocally compared against themselves and all other genomes with BLAST. 2) Pairwise ortholog clusters are identified using similarity scores and imported into a local database. 3) The intersection between genomes is determined by iteratively comparing sequence identification tags and retaining those clusters showing panorthology. 4) Additional genomes are added and checked as in the previous step. 5) Sequence data files are generated for evolutionary analysis.
Figure 2 Neighbor-joining tree of nine metazoan genomes, 285 panorthologs (97,581 amino acid positions, alpha = 1.28). All nodes are supported significantly (>95%) in bootstrap analyses of neighbor-joining and maximum likelihood. The arrow indicates an alternative root [6, 18].
Figure 3 Neighbor-joining trees of complete eukaryotic genome sequence analyses. (A) The intersection of fifteen eukaryotic genomes, 10 panorthologs (5094 amino acid positions, alpha = 1.01). (B) The intersection of genomes from twelve multicellular eukaryotes, 63 panorthologs (23,571 amino acid positions, alpha = 1.15). All nodes are supported significantly (>95%) in bootstrap analyses of neighbor-joining and maximum likelihood, with the exception of node indicated by an asterisk (94% with maximum likelihood) in (A).
Figure 4 Neighbor-joining trees of genomes used to address deuterostome-arthropod divergence time. (A) The intersection of seven eukaryotic genomes, 380 panorthologs (132,190 amino acid positions, alpha = 1.38). (B) The intersection of four eukaryotic genomes, 753 panorthologs (287,000 amino acid positions, alpha = 1.46). All nodes are supported significantly (>95%) in bootstrap analyses of neighbor-joining and maximum likelihood.
Table 1 Number of orthologous clusters (upper-right) and percentage panorthologs (lower-left) per pairwise comparisona.
Hsa Mmu Rno Tru Cin Dme Aga Cel Cbr Sce Ncr Ago Ecu Ath Cme Gtn Pfa
Hsa 14571 14201 9881 6100 5009 5081 3794 4114 1898 2197 1856 759 2892 1628 177 1198
Mmu 68.5 16413 9885 6115 4933 5112 4194 4182 1924 2240 1887 755 2894 1523 183 1241
Rno 73.2 79.2 9708 6016 5001 5023 4127 4079 1871 2207 1825 754 2836 1630 177 1235
Tru 61.6 61 67.5 4872 4970 4974 4109 4090 1464 2224 1808 725 2814 1591 177 1260
Cin 49.3 50.3 55.6 58.2 4520 4554 2980 3848 1823 2090 1740 700 2669 1494 170 1199
Dme 36.7 37.5 41.2 43.8 60.6 7072 3904 3822 1753 1967 1738 713 2476 1460 178 1141
Aga 41.3 41.7 47.4 51.5 72.4 61.3 3973 3926 1833 2107 1793 717 2641 1597 177 1214
Cel 34.2 37.4 42.2 45.7 65.1 49.9 58.8 12416 1549 1702 1593 697 2235 1368 171 990
Cbr 42 42.9 47.5 54.5 79.7 56.7 68.9 84.4 1611 1836 1561 691 2205 1348 168 1083
Sce 42.9 43.7 48.8 49.5 72.4 54.8 63.9 60.3 69 2604 4036 683 1867 1321 177 958
Ncr 48.3 47.3 52.8 59.2 83.1 59.2 71.9 67.9 77.9 84.3 2560 648 2182 1388 168 880
Ago 48.6 47.4 52.9 59.3 82.9 59.9 72.5 69.5 78.3 93.4 96.4 686 1818 1282 172 893
Ecu 42.6 39.5 43.5 55.7 77 55.4 66.8 68.4 76.4 77 91.8 91.3 741 567 143 525
Ath 29 28.5 32.2 32.4 40.9 33.3 37 32.9 36.4 36.6 39.3 38.6 33.1 2126 199 1295
Cme 48.2 45.4 52.1 59.6 83.2 57.6 71.3 68.5 77.7 78.7 90.4 92 90.8 40.9 198 849
Gtn 43.5 30.6 33.3 64.9 84.7 54.5 66.7 71.9 77.4 62.7 92.9 88.4 97.2 25.1 91.9 190
Pfa 46.2 42.1 46.9 58.7 80.9 57.9 71.6 71.4 78 79.6 91.6 91.3 91.8 36.6 91.2 94.7
aThree-letter abbreviations listed in Methods.
Table 2 Bayesian divergence time estimates (± one standard deviation) for eukaryote genome datasets.
Divergence 15 Genomes (5094 aa)a 12 Genomes (23,571 aa) 7 Genomes (132,190 aa) 4 Genomes (287,000 aa)
Mus – Rattus 37 ± 5 50 ± 8 40 ± 9 n/ac
Primate – Rodent 93 ± 10 117 ± 15 120 ± 20 n/a
Tetrapod – Fishb 459 ± 20 460 ± 20 460 ± 20 458 ± 20
Vertebrate – Ciona 771 ± 47 756 ± 58 n/a n/a
Drosophila – Anopheles 445 ± 38 500 ± 48 477 ± 51 n/a
Chordate – Arthropod 949 ± 66 912 ± 89 955 ± 92 1109 ± 103
C. elegans – C. briggsae 89 ± 13 114 ± 22 n/a n/a
Coelomata – Nematoda 1166 ± 89 1074 ± 116 n/a n/a
Ashbya – Saccharomyces 311 ± 39 n/a n/a n/a
Saccharomycetes – Neurospora 900 ± 80 851 ± 96 n/a n/a
Animal – Fungi 1493 ± 125 1303 ± 155 n/a n/a
Arabidopsis – C. merolae 1414 ± 121 n/a n/a n/a
Animal/Fungi – Plant 1671 ± 145 n/a n/a n/a
aNumber of amino acids (aa). bTetrapod – actinopterygian fish divergence constrained between 425 and 495 Ma [78]. cDivergence not available (n/a) for timing.
==== Refs
Wolf YI Rogozin IB Grishin NV Tatusov RL Koonin EV Genome trees constructed using five different approaches suggest new major bacterial clades BMC Evolutionary Biology 2001 1 8 11734060 10.1186/1471-2148-1-8
Raymond J Zhaxybayeva O Gogarten JP Gerdes SY Blankenship RE Whole-genome analysis of photosynthetic prokaryotes Science 2002 298 1616 1620 12446909 10.1126/science.1075558
Brochier C Forterre P Gribaldo S Archaeal phylogeny based on proteins of the transcription and translation machineries: tackling the Methanopyrus kandleri paradox Genome Biology 2004 5 R17 15003120 10.1186/gb-2004-5-3-r17
Korbel JO Snel B Huynen MA Bork P SHOT: a web server for the construction of genome phylogenies Trends in Genetics 2002 18 158 162 11858840 10.1016/S0168-9525(01)02597-5
House CH Runnegar B Fitz-Gibbon ST Geobiological analysis using whole genome-based tree building applied to the Bacteria, Archaea, and Eukarya Geobiology 2003 1 15 26 10.1046/j.1472-4669.2003.00004.x
Wolf YI Rogozin IB Koonin EV Coelomata and not Ecdysozoa: evidence from genome-wide phylogenetic analysis Genome Research 2004 14 29 36 14707168 10.1101/gr.1347404
Copley RR Aloy P Russell RB Telford MJ Systematic searches for molecular synapomorphies in model metazoan genomes give some support for Ecdysozoa after accounting for the idiosyncrasies of Caenorhabditis elegans Evolution and Development 2004 6 164 169 15099303 10.1111/j.1525-142X.2004.04021.x
Krylov DM Wolf YI Rogozin IB Koonin EV Gene loss, protein sequence divergence, gene dispensability, expression level, and interactivity are correlated in eukaryotic evolution Genome Research 2003 13 2229 2235 14525925 10.1101/gr.1589103
Hughes AL Friedman R Differential loss of ancestral gene families as a source of genomic divergence in animals Proceedings of the Royal Society of London B: Biological Sciences 2004 271 S107 109 15101434 10.1098/rspb.2003.2556
Rogozin IB Wolf YI Sorokin AV Mirkin BG Koonin EV Remarkable interkingdom conservation of intron positions and massive, lineage-specific intron loss and gain in eukaryotic evolution Current Biology 2003 13 1512 1517 12956953 10.1016/S0960-9822(03)00558-X
Coghlan A Wolfe KH Origins of recently gained introns in Caenorhabditis PNAS 2004 101 11362 11367 15243155 10.1073/pnas.0308192101
Lee Y Sultana R Pertea G Cho J Karamycheva S Tsai J Parvizi B Cheung F Antonescu V White J Holt I Liang F Quackenbush J Cross-referencing eukaryotic genomes: TIGR Orthologous Gene Alignments (TOGA) Genome Research 2002 12 493 502 11875039 10.1101/gr.212002
Li L Stoeckert CJJ Roos DS OrthoMCL: identification of ortholog groups for eukaryotic genomes Genome Research 2003 13 2178 2189 12952885 10.1101/gr.1224503
Koonin EV Fedorova ND Jackson JD Jacobs AR Krylov DM Makarova KS Mazumder R Mekhedov SL Nikolskaya AN Rao BS Rogozin IB Smirnov S Sorokin AV Sverdlov AV Vasudevan S Wolf YI Yin JJ Natale DA A comprehensive evolutionary classification of proteins encoded in complete eukaryotic genomes Genome Biology 2004 5 R7 14759257 10.1186/gb-2004-5-2-r7
Sonnhammer ELL Koonin EV Orthology, paralogy and proposed classification for paralog subtypes Trends in Genetics 2002 18 619 620 12446146 10.1016/S0168-9525(02)02793-2
Lander ES Linton LM Birren B Nusbaum C Zody MC Baldwin J Devon K Dewar K Doyle M FitzHugh W Funke R Gage D Harris K Heaford A Howland J Kann L Lehoczky J LeVine R McEwan P McKernan K Meldrim J Mesirov JP Miranda C Morris W Naylor J Raymond C Rosetti M Santos R Sheridan A Sougnez C Stange-Thomann N Stojanovic N Subramanian A Wyman D Rogers J Sulston J Ainscough R Beck S Bentley D Burton J Clee C Carter N Coulson A Deadman R Deloukas P Dunham A Dunham I Durbin R French L Grafham D Gregory S Hubbard T Humphray S Hunt A Jones M Lloyd C McMurray A Matthews L Mercer S Milne S Mullikin JC Mungall A Plumb R Ross M Shownkeen R Sims S Waterston RH Wilson RK Hillier LW McPherson JD Marra MA Mardis ER Fulton LA Chinwalla AT Pepin KH Gish WR Chissoe SL Wendl MC Delehaunty KD Miner TL Delehaunty A Kramer JB Cook LL Fulton RS Johnson DL Minx PJ Clifton SW Hawkins T Branscomb E Predki P Richardson P Wenning S Slezak T Doggett N Cheng JF Olsen A Lucas S Elkin C Uberbacher E Frazier M Gibbs RA Muzny DM Scherer SE Bouck JB Sodergren EJ Worley KC Rives CM Gorrell JH Metzker ML Naylor SL Kucherlapati RS Nelson DL Weinstock GM Sakaki Y Fujiyama A Hattori M Yada T Toyoda A Itoh T Kawagoe C Watanabe H Totoki Y Taylor T Weissenbach J Heilig R Saurin W Artiguenave F Brottier P Bruls T Pelletier E Robert C Wincker P Smith DR Doucette-Stamm L Rubenfield M Weinstock K Lee HM Dubois J Rosenthal A Platzer M Nyakatura G Taudien S Rump A Yang H Yu J Wang J Huang G Gu J Hood L Rowen L Madan A Qin S Davis RW Federspiel NA Abola AP Proctor MJ Myers RM Schmutz J Dickson M Grimwood J Cox DR Olson MV Kaul R Shimizu N Kawasaki K Minoshima S Evans GA Athanasiou M Schultz R Roe BA Chen F Pan H Ramser J Lehrach H Reinhardt R McCombie WR de la Bastide M Dedhia N Blocker H Hornischer K Nordsiek G Agarwala R Aravind L Bailey JA Bateman A Batzoglou S Birney E Bork P Brown DG Burge CB Cerutti L Chen HC Church D Clamp M Copley RR Doerks T Eddy SR Eichler EE Furey TS Galagan J Gilbert JG Harmon C Hayashizaki Y Haussler D Hermjakob H Hokamp K Jang W Johnson LS Jones TA Kasif S Kaspryzk A Kennedy S Kent WJ Kitts P Koonin EV Korf I Kulp D Lancet D Lowe TM McLysaght A Mikkelsen T Moran JV Mulder N Pollara VJ Ponting CP Schuler G Schultz J Slater G Smit AF Stupka E Szustakowski J Thierry-Mieg D Thierry-Mieg J Wagner L Wallis J Wheeler R Williams A Wolf YI Wolfe KH Yang SP Yeh RF Collins F Guyer MS Peterson J Felsenfeld A Wetterstrand KA Patrinos A Morgan MJ Szustakowki J de Jong P Catanese JJ Osoegawa K Shizuya H Choi S Chen YJ Initial sequencing and analysis of the human genome Nature 2001 409 860 921 11237011 10.1038/35057062
Hedges SB Blair JE Venturi ML Shoe JL A molecular timescale of eukaryote evolution and the rise of complex multicellular life BMC Evolutionary Biology 2004 4 2 15005799 10.1186/1471-2148-4-2
Blair JE Ikeo K Gojobori T Hedges SB The evolutionary position of nematodes BMC Evolutionary Biology 2002 2 7 11985779 10.1186/1471-2148-2-7
Gibbs RA Weinstock GM Metzker ML Muzny DM Sodergren EJ Scherer S Scott G Steffen D Worley KC Burch PE Okwuonu G Hines S Lewis L DeRamo C Delgado O Dugan-Rocha S Miner G Morgan M Hawes A Gill R Celera Holt RA Adams MD Amanatides PG Baden-Tillson H Barnstead M Chin S Evans CA Ferriera S Fosler C Glodek A Gu Z Jennings D Kraft CL Nguyen T Pfannkoch CM Sitter C Sutton GG Venter JC Woodage T Smith D Lee HM Gustafson E Cahill P Kana A Doucette-Stamm L Weinstock K Fechtel K Weiss RB Dunn DM Green ED Blakesley RW Bouffard GG De Jong PJ Osoegawa K Zhu B Marra M Schein J Bosdet I Fjell C Jones S Krzywinski M Mathewson C Siddiqui A Wye N McPherson J Zhao S Fraser CM Shetty J Shatsman S Geer K Chen Y Abramzon S Nierman WC Havlak PH Chen R Durbin KJ Egan A Ren Y Song XZ Li B Liu Y Qin X Cawley S Cooney AJ D'Souza LM Martin K Wu JQ Gonzalez-Garay ML Jackson AR Kalafus KJ McLeod MP Milosavljevic A Virk D Volkov A Wheeler DA Zhang Z Bailey JA Eichler EE Tuzun E Birney E Mongin E Ureta-Vidal A Woodwark C Zdobnov E Bork P Suyama M Torrents D Alexandersson M Trask BJ Young JM Huang H Wang H Xing H Daniels S Gietzen D Schmidt J Stevens K Vitt U Wingrove J Camara F Mar Alba M Abril JF Guigo R Smit A Dubchak I Rubin EM Couronne O Poliakov A Hubner N Ganten D Goesele C Hummel O Kreitler T Lee YA Monti J Schulz H Zimdahl H Himmelbauer H Lehrach H Jacob HJ Bromberg S Gullings-Handley J Jensen-Seaman MI Kwitek AE Lazar J Pasko D Tonellato PJ Twigger S Ponting CP Duarte JM Rice S Goodstadt L Beatson SA Emes RD Winter EE Webber C Brandt P Nyakatura G Adetobi M Chiaromonte F Elnitski L Eswara P Hardison RC Hou M Kolbe D Makova K Miller W Nekrutenko A Riemer C Schwartz S Taylor J Yang S Zhang Y Lindpaintner K Andrews TD Caccamo M Clamp M Clarke L Curwen V Durbin R Eyras E Searle SM Cooper GM Batzoglou S Brudno M Sidow A Stone EA Payseur BA Bourque G Lopez-Otin C Puente XS Chakrabarti K Chatterji S Dewey C Pachter L Bray N Yap VB Caspi A Tesler G Pevzner PA Haussler D Roskin KM Baertsch R Clawson H Furey TS Hinrichs AS Karolchik D Kent WJ Rosenbloom KR Trumbower H Weirauch M Cooper DN Stenson PD Ma B Brent M Arumugam M Shteynberg D Copley RR Taylor MS Riethman H Mudunuri U Peterson J Guyer M Felsenfeld A Old S Mockrin S Collins F Genome sequence of the Brown Norway rat yields insights into mammalian evolution Nature 2004 428 493 521 15057822 10.1038/nature02426
Coghlan A Wolfe KH Fourfold Faster Rate of Genome Rearrangement in Nematodes Than in Drosophila Genome Res 2002 12 857 867 12045140 10.1101/gr.172702
Stein LD Bao Z Blasiar D Blumenthal T Brent MR Chen N Chinwalla A Clarke L Clee C Coghlan A Coulson A D'Eustachio P Fitch DH Fulton LA Fulton RE Griffiths-Jones S Harris TW Hillier LW Kamath R Kuwabara PE Mardis ER Marra MA Miner TL Minx P Mullikin JC Plumb RW Rogers J Schein JE Sohrmann M Spieth J Stajich JE Wei C Willey D Wilson RK Durbin R Waterston RH The Genome Sequence of Caenorhabditis briggsae: A Platform for Comparative Genomics PLoS Biology 2003 1 E45 14624247 10.1371/journal.pbio.0000045
Zdobnov EM von Mering C Letunic I Torrents D Suyama M Copley RR Christophides GK Thomasova D Holt RA Subramanian GM Mueller HM Dimopoulos G Law JH Wells MA Birney E Charlab R Halpern AL Kokoza E Kraft CL Lai Z Lewis S Louis C Barillas-Mury C Nusskern D Rubin GM Salzberg SL Sutton GG Topalis P Wides R Wincker P Yandell M Collins FH Ribeiro J Gelbart WM Kafatos FC Bork P Comparative genome and proteome analysis of Anopheles gambiae and Drosophila melanogaster Science 2002 298 149 159 12364792 10.1126/science.1077061
Douglas S Zauner S Fraunholz M Beaton M Penny S Deng LT Wu X Reith M Cavalier-Smith T Maier UG The highly reduced genome of an enslaved algal nucleus Nature 2001 410 1091 1096 11323671 10.1038/35074092
Katinka MD Duprat S Cornillot E Metenier G Thomarat F Prensier G Barbe V Peyretaillade E Brottier P Wincker P Delbac F El Alaoui H Peyret P Saurin W Gouy M Weissenbach J Vivares CP Genome sequence and gene compaction of the eukaryote parasite Encephalitozoon cuniculi Nature 2001 414 450 453 11719806 10.1038/35106579
Gardner MJ Hall N Fung E White O Berriman M Hyman RW Carlton JM Pain A Nelson KE Bowman S Paulsen IT James K Eisen JA Rutherford K Salzberg SL Craig A Kyes S Chan MS Nene V Shallom SJ Suh B Peterson J Angiuoli S Pertea M Allen J Selengut J Haft D Mather MW Vaidya AB Martin DMA Fairlamb AH Fraunholz MJ Roos DS Ralph SA McFadden GI Cummings LM Subramanian GM Mungall C Venter JC Carucci DJ Hoffman SL Newbold C Davis RW Fraser CM Barrell B Genome sequence of the human malaria parasite Plasmodium falciparum Nature 2002 419 498 511 12368864 10.1038/nature01097
Pizzi E Frontali C Low-Complexity Regions in Plasmodium falciparum Proteins Genome Res 2001 11 218 229 11157785 10.1101/gr.GR-1522R
Felsenstein J Cases in which parsimony or compatibility methods will be positively misleading Syst Zool 1978 27 401 410
Rokas A Williams BL King N Carroll SB Genome-scale approaches to resolving incongruence in molecular phylogenies Nature 2003 425 798 804 14574403 10.1038/nature02053
Gadagkar SR Rosenberg MS Kumar S Inferring species phylogenies from multiple genes: Concatenated seqeunce tree versus consensus gene tree Journal of Experimental Zoology Part B: Molecular and Developmental Evolution 2005 304B 64 74 10.1002/jez.b.21026
Wray GA Levinton JS Shapiro LH Molecular evidence for deep Precambrian divergences among metazoan phyla Science 1996 274 568 573 10.1126/science.274.5287.568
Feng DF Cho G Doolittle RF Determining divergence times with a protein clock: update and reevaluation Proc Nat Acad Sci U S A 1997 94 13028 13033 10.1073/pnas.94.24.13028
Wang DY Kumar S Hedges SB Divergence time estimates for the early history of animal phyla and the origin of plants, animals and fungi Proceedings of the Royal Society of London B: Biological Sciences 1999 266 163 171 10097391 10.1098/rspb.1999.0617
Nei M Xu P Glazko G Estimation of divergence times from multiprotein sequences for a few mammalian species and several distantly related organisms Proc Natl Acad Sci U S A 2001 98 2497 2502 11226267 10.1073/pnas.051611498
Hedges SB Kumar S Precision of molecular time estimates Trends in Genetics 2004 20 242 247 15109778 10.1016/j.tig.2004.03.004
Peterson KJ Lyons JB Nowak KS Takacs CM Wargo MJ McPeek MA Estimating metazoan divergence times with a molecular clock Proc Natl Acad Sci U S A 2004 101 6536 6541 15084738 10.1073/pnas.0401670101
Douzery EJP Snell EA Bapteste E Delsuc F Philippe H The timing of eukaryote evolution: Does a relaxed molecular clock reconcile proteins and fossils? Proc Natl Acad Sci U S A 2004 101 15386 15391 15494441 10.1073/pnas.0403984101
Aris-Brosou S Yang Z Bayesian models of episodic evolution support a late precambrian explosive diversification of the Metazoa Molecular Biology and Evolution 2003 20 1947 1954 12949122 10.1093/molbev/msg226
Kumar S Mesoproterozoic megafossil Chuaria-Tawuia association may represent parts of a multicellular plant, Vindhyan Supergroup, Central India Precambrian Res 2001 106 187 211 10.1016/S0301-9268(00)00093-0
Woods KN Knoll AH German TN Xanthophyte Algae from the Mesoproterozoic/Neoproterozoic Transition: Confirmation and Evolutionary Implications GSA Abstracts with Programs 1998 30 A232
Butterfield NJ Bangiomorpha pubescens n. gen., n. sp.: Implications for the Evolution of Sex, Multicellularity, and the Mesoproterozoic/Neoproterozoic Radiation of Eukaryotes Paleobiology 2000 26 386 404
Blair JE Hedges SB Molecular Clocks Do Not Support the Cambrian Explosion Mol Biol Evol 2005 22 387 390 15537810 10.1093/molbev/msi039
Clark AG Glanowski S Nielsen R Thomas PD Kejariwal A Todd MA Tanenbaum DM Civello D Lu F Murphy B Ferriera S Wang G Zheng X White TJ Sninsky JJ Adams MD Cargill M Inferring nonneutral evolution from human-chimp-mouse orthologous gene trios Science 2003 302 1960 1963 14671302 10.1126/science.1088821
Kumar S Subramanian S Mutation rates in mammalian genomes Proc Natl Acad Sci U S A 2002 99 803 808 11792858 10.1073/pnas.022629899
Altschul SF Madden TL Schaffer AA Zhang J Zhang Z Miller W Lipman DJ Gapped BLAST and PSI-BLAST: a new generation of protein database search programs Nucleic Acids Res 1997 25 3389 3402 9254694 10.1093/nar/25.17.3389
Remm M Storm CE Sonnhammer EL Automatic clustering of orthologs and in-paralogs from pairwise species comparisons Journal of Molecular Biology 2001 314 1041 1052 11743721 10.1006/jmbi.2000.5197
Tatusov RL Galperin MY Natale DA Koonin EV The COG database: a tool for genome-scale analysis of protein functions and evolution Nucleic Acids Research 2000 28 33 36 10592175 10.1093/nar/28.1.33
Venter JC Adams MD Myers EW Li PW Mural RJ Sutton GG Smith HO Yandell M Evans CA Holt RA Gocayne JD Amanatides P Ballew RM Huson DH Wortman JR Zhang Q Kodira CD Zheng XH Chen L Skupski M Subramanian G Thomas PD Zhang J Gabor Miklos GL Nelson C Broder S Clark AG Nadeau J McKusick VA Zinder N Levine AJ Roberts RJ Simon M Slayman C Hunkapiller M Bolanos R Delcher A Dew I Fasulo D Flanigan M Florea L Halpern A Hannenhalli S Kravitz S Levy S Mobarry C Reinert K Remington K Abu-Threideh J Beasley E Biddick K Bonazzi V Brandon R Cargill M Chandramouliswaran I Charlab R Chaturvedi K Deng Z Di Francesco V Dunn P Eilbeck K Evangelista C Gabrielian AE Gan W Ge W Gong F Gu Z Guan P Heiman TJ Higgins ME Ji RR Ke Z Ketchum KA Lai Z Lei Y Li Z Li J Liang Y Lin X Lu F Merkulov GV Milshina N Moore HM Naik AK Narayan VA Neelam B Nusskern D Rusch DB Salzberg S Shao W Shue B Sun J Wang Z Wang A Wang X Wang J Wei M Wides R Xiao C Yan C Yao A Ye J Zhan M Zhang W Zhang H Zhao Q Zheng L Zhong F Zhong W Zhu S Zhao S Gilbert D Baumhueter S Spier G Carter C Cravchik A Woodage T Ali F An H Awe A Baldwin D Baden H Barnstead M Barrow I Beeson K Busam D Carver A Center A Cheng ML Curry L Danaher S Davenport L Desilets R Dietz S Dodson K Doup L Ferriera S Garg N Gluecksmann A Hart B Haynes J Haynes C Heiner C Hladun S Hostin D Houck J Howland T Ibegwam C Johnson J Kalush F Kline L Koduru S Love A Mann F May D McCawley S McIntosh T McMullen I Moy M Moy L Murphy B Nelson K Pfannkoch C Pratts E Puri V Qureshi H Reardon M Rodriguez R Rogers YH Romblad D Ruhfel B Scott R Sitter C Smallwood M Stewart E Strong R Suh E Thomas R Tint NN Tse S Vech C Wang G Wetter J Williams S Williams M Windsor S Winn-Deen E Wolfe K Zaveri J Zaveri K Abril JF Guigo R Campbell MJ Sjolander KV Karlak B Kejariwal A Mi H Lazareva B Hatton T Narechania A Diemer K Muruganujan A Guo N Sato S Bafna V Istrail S Lippert R Schwartz R Walenz B Yooseph S Allen D Basu A Baxendale J Blick L Caminha M Carnes-Stine J Caulk P Chiang YH Coyne M Dahlke C Mays A Dombroski M Donnelly M Ely D Esparham S Fosler C Gire H Glanowski S Glasser K Glodek A Gorokhov M Graham K Gropman B Harris M Heil J Henderson S Hoover J Jennings D Jordan C Jordan J Kasha J Kagan L Kraft C Levitsky A Lewis M Liu X Lopez J Ma D Majoros W McDaniel J Murphy S Newman M Nguyen T Nguyen N Nodell M Pan S Peck J Peterson M Rowe W Sanders R Scott J Simpson M Smith T Sprague A Stockwell T Turner R Venter E Wang M Wen M Wu D Wu M Xia A Zandieh A Zhu X The sequence of the human genome Science 2001 291 1304 1351 11181995 10.1126/science.1058040
Ensembl Genome Browser http://www.ensembl.org/
Waterston RH Lindblad-Toh K Birney E Rogers J Abril JF Agarwal P Agarwala R Ainscough R Alexandersson M An P Antonarakis SE Attwood J Baertsch R Bailey J Barlow K Beck S Berry E Birren B Bloom T Bork P Botcherby M Bray N Brent MR Brown DG Brown SD Bult C Burton J Butler J Campbell RD Carninci P Cawley S Chiaromonte F Chinwalla AT Church DM Clamp M Clee C Collins FS Cook LL Copley RR Coulson A Couronne O Cuff J Curwen V Cutts T Daly M David R Davies J Delehaunty KD Deri J Dermitzakis ET Dewey C Dickens NJ Diekhans M Dodge S Dubchak I Dunn DM Eddy SR Elnitski L Emes RD Eswara P Eyras E Felsenfeld A Fewell GA Flicek P Foley K Frankel WN Fulton LA Fulton RS Furey TS Gage D Gibbs RA Glusman G Gnerre S Goldman N Goodstadt L Grafham D Graves TA Green ED Gregory S Guigo R Guyer M Hardison RC Haussler D Hayashizaki Y Hillier LW Hinrichs A Hlavina W Holzer T Hsu F Hua A Hubbard T Hunt A Jackson I Jaffe DB Johnson LS Jones M Jones TA Joy A Kamal M Karlsson EK Karolchik D Kasprzyk A Kawai J Keibler E Kells C Kent WJ Kirby A Kolbe DL Korf I Kucherlapati RS Kulbokas EJ Kulp D Landers T Leger JP Leonard S Letunic I Levine R Li J Li M Lloyd C Lucas S Ma B Maglott DR Mardis ER Matthews L Mauceli E Mayer JH McCarthy M McCombie WR McLaren S McLay K McPherson JD Meldrim J Meredith B Mesirov JP Miller W Miner TL Mongin E Montgomery KT Morgan M Mott R Mullikin JC Muzny DM Nash WE Nelson JO Nhan MN Nicol R Ning Z Nusbaum C O'Connor MJ Okazaki Y Oliver K Overton-Larty E Pachter L Parra G Pepin KH Peterson J Pevzner P Plumb R Pohl CS Poliakov A Ponce TC Ponting CP Potter S Quail M Reymond A Roe BA Roskin KM Rubin EM Rust AG Santos R Sapojnikov V Schultz B Schultz J Schwartz MS Schwartz S Scott C Seaman S Searle S Sharpe T Sheridan A Shownkeen R Sims S Singer JB Slater G Smit A Smith DR Spencer B Stabenau A Stange-Thomann N Sugnet C Suyama M Tesler G Thompson J Torrents D Trevaskis E Tromp J Ucla C Ureta-Vidal A Vinson JP Von Niederhausern AC Wade CM Wall M Weber RJ Weiss RB Wendl MC West AP Wetterstrand K Wheeler R Whelan S Wierzbowski J Willey D Williams S Wilson RK Winter E Worley KC Wyman D Yang S Yang SP Zdobnov EM Zody MC Lander ES Initial sequencing and comparative analysis of the mouse genome Nature 2002 420 520 562 12466850 10.1038/nature01262
Aparicio S Chapman J Stupka E Putnam N Chia J Dehal P Christoffels A Rash S Hoon S Smit A Gelpke MDS Roach J Oh T Ho IY Wong M Detter C Verhoef F Predki P Tay A Lucas S Richardson P Smith SF Clark MS Edwards YJK Doggett N Zharkikh A Tavtigian SV Pruss D Barnstead M Evans C Baden H Powell J Glusman G Rowen L Hood L Tan YH Elgar G Hawkins T Venkatesh B Rokhsar D Brenner S Whole-Genome Shotgun Assembly and Analysis of the Genome of Fugu rubripes Science 2002 297 1301 1310 12142439 10.1126/science.1072104
DOE Joint Genome Institute http://www.jgi.doe.gov/
Dehal P Satou Y Campbell RK Chapman J Degnan B De Tomaso A Davidson B Di Gregorio A Gelpke M Goodstein DM Harafuji N Hastings KE Ho I Hotta K Huang W Kawashima T Lemaire P Martinez D Meinertzhagen IA Necula S Nonaka M Putnam N Rash S Saiga H Satake M Terry A Yamada L Wang HG Awazu S Azumi K Boore J Branno M Chin-Bow S DeSantis R Doyle S Francino P Keys DN Haga S Hayashi H Hino K Imai KS Inaba K Kano S Kobayashi K Kobayashi M Lee BI Makabe KW Manohar C Matassi G Medina M Mochizuki Y Mount S Morishita T Miura S Nakayama A Nishizaka S Nomoto H Ohta F Oishi K Rigoutsos I Sano M Sasaki A Sasakura Y Shoguchi E Shin-i T Spagnuolo A Stainier D Suzuki MM Tassy O Takatori N Tokuoka M Yagi K Yoshizaki F Wada S Zhang C Hyatt PD Larimer F Detter C Doggett N Glavina T Hawkins T Richardson P Lucas S Kohara Y Levine M Satoh N Rokhsar DS The draft genome of Ciona intestinalis: insights into chordate and vertebrate origins Science 2002 298 2157 2167 12481130 10.1126/science.1080049
Adams MD Celniker SE Holt RA Evans CA Gocayne JD Amanatides PG Scherer SE Li PW Hoskins RA Galle RF George RA Lewis SE Richards S Ashburner M Henderson SN Sutton GG Wortman JR Yandell MD Zhang Q Chen LX Brandon RC Rogers YH Blazej RG Champe M Pfeiffer BD Wan KH Doyle C Baxter EG Helt G Nelson CR Gabor GL Abril JF Agbayani A An HJ Andrews-Pfannkoch C Baldwin D Ballew RM Basu A Baxendale J Bayraktaroglu L Beasley EM Beeson KY Benos PV Berman BP Bhandari D Bolshakov S Borkova D Botchan MR Bouck J Brokstein P Brottier P Burtis KC Busam DA Butler H Cadieu E Center A Chandra I Cherry JM Cawley S Dahlke C Davenport LB Davies P de Pablos B Delcher A Deng Z Mays AD Dew I Dietz SM Dodson K Doup LE Downes M Dugan-Rocha S Dunkov BC Dunn P Durbin KJ Evangelista CC Ferraz C Ferriera S Fleischmann W Fosler C Gabrielian AE Garg NS Gelbart WM Glasser K Glodek A Gong F Gorrell JH Gu Z Guan P Harris M Harris NL Harvey D Heiman TJ Hernandez JR Houck J Hostin D Houston KA Howland TJ Wei MH Ibegwam C Jalali M Kalush F Karpen GH Ke Z Kennison JA Ketchum KA Kimmel BE Kodira CD Kraft C Kravitz S Kulp D Lai Z Lasko P Lei Y Levitsky AA Li J Li Z Liang Y Lin X Liu X Mattei B McIntosh TC McLeod MP McPherson D Merkulov G Milshina NV Mobarry C Morris J Moshrefi A Mount SM Moy M Murphy B Murphy L Muzny DM Nelson DL Nelson DR Nelson KA Nixon K Nusskern DR Pacleb JM Palazzolo M Pittman GS Pan S Pollard J Puri V Reese MG Reinert K Remington K Saunders RD Scheeler F Shen H Shue BC Siden-Kiamos I Simpson M Skupski MP Smith T Spier E Spradling AC Stapleton M Strong R Sun E Svirskas R Tector C Turner R Venter E Wang AH Wang X Wang ZY Wassarman DA Weinstock GM Weissenbach J Williams SM WoodageT Worley KC Wu D Yang S Yao QA Ye J Yeh RF Zaveri JS Zhan M Zhang G Zhao Q Zheng L Zheng XH Zhong FN Zhong W Zhou X Zhu S Zhu X Smith HO Gibbs RA Myers EW Rubin GM Venter JC The genome sequence of Drosophila melanogaster Science 2000 287 2185 2195 10731132 10.1126/science.287.5461.2185
Berkeley Drosophilia Genome Project
Holt RA Subramanian GM Halpern A Sutton GG Charlab R Nusskern DR Wincker P Clark AG Ribeiro JM Wides R Salzberg SL Loftus B Yandell M Majoros WH Rusch DB Lai Z Kraft CL Abril JF Anthouard V Arensburger P Atkinson PW Baden H de Berardinis V Baldwin D Benes V Biedler J Blass C Bolanos R Boscus D Barnstead M Cai S Center A Chaturverdi K Christophides GK Chrystal MA Clamp M Cravchik A Curwen V Dana A Delcher A Dew I Evans CA Flanigan M Grundschober-Freimoser A Friedli L Gu Z Guan P Guigo R Hillenmeyer ME Hladun SL Hogan JR Hong YS Hoover J Jaillon O Ke Z Kodira C Kokoza E Koutsos A Letunic I Levitsky A Liang Y Lin JJ Lobo NF Lopez JR Malek JA McIntosh TC Meister S Miller J Mobarry C Mongin E Murphy SD O'Brochta DA Pfannkoch C Qi R Regier MA Remington K Shao H Sharakhova MV Sitter CD Shetty J Smith TJ Strong R Sun J Thomasova D Ton LQ Topalis P Tu Z Unger MF Walenz B Wang A Wang J Wang M Wang X Woodford KJ Wortman JR Wu M Yao A Zdobnov EM Zhang H Zhao Q Zhao S Zhu SC Zhimulev I Coluzzi M della Torre A Roth CW Louis C Kalush F Mural RJ Myers EW Adams MD Smith HO Broder S Gardner MJ Fraser CM Birney E Bork P Brey PT Venter JC Weissenbach J Kafatos FC Collins FH Hoffman SL The genome sequence of the malaria mosquito Anopheles gambiae Science 2002 298 129 149 12364791 10.1126/science.1076181
The C. elegans Sequencing Consortium Genome sequence of the nematode C. elegans: a platform for investigating biology. Science 1998 282 2012 2018 9851916 10.1126/science.282.5396.2012
Wormbase
Goffeau A Barrell BG Bussey H Davis RW Dujon B Feldmann H Galibert F Hoheisel JD Jacq C Johnston M Louis EJ Mewes HW Murakami Y Philippsen P Tettelin H Oliver SG Life with 6000 genes Science 1996 274 546, 563 7 8849441 10.1126/science.274.5287.546
Saccharomyces Genome Database
Galagan JE Calvo SE Borkovich KA Selker EU Read ND Jaffe D FitzHugh W Ma LJ Smirnov S Purcell S Rehman B Elkins T Engels R Wang S Nielsen CB Butler J Endrizzi M Qui D Ianakiev P Bell-Pedersen D Nelson MA Werner-Washburne M Selitrennikoff CP Kinsey JA Braun EL Zelter A Schulte U Kothe GO Jedd G Mewes W Staben C Marcotte E Greenberg D Roy A Foley K Naylor J Stange-Thomann N Barrett R Gnerre S Kamal M Kamvysselis M Mauceli E Bielke C Rudd S Frishman D Krystofova S Rasmussen C Metzenberg RL Perkins DD Kroken S Cogoni C Macino G Catcheside D Li W Pratt RJ Osmani SA DeSouza CP Glass L Orbach MJ Berglund JA Voelker R Yarden O Plamann M Seiler S Dunlap J Radford A Aramayo R Natvig DO Alex LA Mannhaupt G Ebbole DJ Freitag M Paulsen I Sachs MS Lander ES Nusbaum C Birren B The genome sequence of the filamentous fungus Neurospora crassa Nature 2003 422 859 868 12712197 10.1038/nature01554
Broad Institute
Dietrich FS Voegeli S Brachat S Lerch A Gates K Steiner S Mohr C Pohlmann R Luedi P Choi S Wing RA Flavier A Gaffney TD Philippsen P The Ashbya gossypii genome as a tool for mapping the ancient Saccharomyces cerevisiae genome Science 2004 304 304 307 15001715 10.1126/science.1095781
Ashbya Genome Database
European Bioinformatics Institute
The Arabidopsis Genome Initiative Analysis of the genome sequence of the flowering plant Arabidopsis thaliana Nature 2000 408 796 815 11130711 10.1038/35048692
The Institute for Genomic Research
Matsuzaki M Misumi O Shin IT Maruyama S Takahara M Miyagishima SY Mori T Nishida K Yagisawa F Yoshida Y Nishimura Y Nakao S Kobayashi T Momoyama Y Higashiyama T Minoda A Sano M Nomoto H Oishi K Hayashi H Ohta F Nishizaka S Haga S Miura S Morishita T Kabeya Y Terasawa K Suzuki Y Ishii Y Asakawa S Takano H Ohta N Kuroiwa H Tanaka K Shimizu N Sugano S Sato N Nozaki H Ogasawara N Kohara Y Kuroiwa T Genome sequence of the ultrasmall unicellular red alga Cyanidioschyzon merolae 10D Nature 2004 428 653 657 15071595 10.1038/nature02398
Cyanidioschyzon merolae Genome Project
PlasmoDB
Thompson JD Gibson TJ Plewniak F Jeanmougin F Higgins DG The CLUSTALX windows interface: flexible strategies for multiple sequence alignment aided by quality analysis tools Nucleic Acids Research 1997 25 4876 4882 9396791 10.1093/nar/25.24.4876
Yang Z PAML: a program package for phylogenetic analysis by maximum likelihood Bioinformatics 1997 13 555 556
Kumar S Tamura K Jakobsen I Nei M MEGA2: molecular evolutionary genetics analysis software Bioinformatics 2001 17 1244 1245 11751241 10.1093/bioinformatics/17.12.1244
Strimmer K von Haeseler A Quartet Puzzling: A Quartet Maximum-Likelihood Method for Reconstructing Tree Topologies Mol Biol Evol 1996 13 964 969
Felsenstein J Phylip: Phylogeny Inference Package 2002 3.6(a3) Seattle, Department of Genome Sciences, University of Washington
Kishino H Thorne JL Bruno WJ Performance of a divergence time estimation method under a probabilistic model of rate evolution Molecular Biology and Evolution 2001 18 352 361 11230536
Jacobs LL Downs WR Tomida Y, Li CK and Setoguchi T The evolution of murine rodents in Asia Rodent and Lagomorph Families of Asian Origins and Diversification 1994 8 Tokyo, National Science Museum Monographs 149 156
Benton MJ The Fossil Record 2 1993 New York, Chapman & Hall
Donoghue PCJ Smith MP Sansom IJ Donoghue PCJ and Smith MP The origin and early evolution of chordates: molecular clocks and the fossil record Telling the Evolutionary Time: Molecular Clocks and the Fossil Record 2004 New York, CRC Press 190 223
Shu DG Chen L Han J Zhang XL An Early Cambrian tunicate from China Nature 2001 411 472 473 11373678 10.1038/35078069
Gaunt MW Miles MA An insect molecular clock dates the origin of the insects and accords with palaeontological and biogeographic landmarks Molecular Biology and Evolution 2002 19 748 761 11961108
|
15762985
|
PMC1274250
|
CC BY
|
2021-01-04 16:02:48
|
no
|
BMC Bioinformatics. 2005 Mar 11; 6:53
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-53
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-541576638410.1186/1471-2105-6-54Methodology ArticleIn silico microdissection of microarray data from heterogeneous cell populations Lähdesmäki Harri [email protected] llya [email protected] Valerie [email protected] Olli [email protected] Wei [email protected] Institute of Signal Processing, Tampere University of Technology, P.O.Box 553, 33101 Tampere, Finland2 Cancer Genomics Laboratory, University of Texas M. D. Anderson Cancer Center, 1515 Holcombe Blvd., Box 85, Houston, TX 77030, USA2005 14 3 2005 6 54 54 28 10 2004 14 3 2005 Copyright © 2005 Lähdesmäki et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Very few analytical approaches have been reported to resolve the variability in microarray measurements stemming from sample heterogeneity. For example, tissue samples used in cancer studies are usually contaminated with the surrounding or infiltrating cell types. This heterogeneity in the sample preparation hinders further statistical analysis, significantly so if different samples contain different proportions of these cell types. Thus, sample heterogeneity can result in the identification of differentially expressed genes that may be unrelated to the biological question being studied. Similarly, irrelevant gene combinations can be discovered in the case of gene expression based classification.
Results
We propose a computational framework for removing the effects of sample heterogeneity by "microdissecting" microarray data in silico. The computational method provides estimates of the expression values of the pure (non-heterogeneous) cell samples. The inversion of the sample heterogeneity can be facilitated by providing accurate estimates of the mixing percentages of different cell types in each measurement. For those cases where no such information is available, we develop an optimization-based method for joint estimation of the mixing percentages and the expression values of the pure cell samples. We also consider the problem of selecting the correct number of cell types.
Conclusion
The efficiency of the proposed methods is illustrated by applying them to a carefully controlled cDNA microarray data obtained from heterogeneous samples. The results demonstrate that the methods are capable of reconstructing both the sample and cell type specific expression values from heterogeneous mixtures and that the mixing percentages of different cell types can also be estimated. Furthermore, a general purpose model selection method can be used to select the correct number of cell types.
==== Body
Background
Recent developments in high-throughput genomic technologies have revolutionized the approaches aimed at understanding biological systems and emphasized the need for computational and systems biology research. Microarray analysis, for instance, can provide massive amounts of information about a biological sample by simultaneously measuring thousands of transcript levels. Application of such methodologies has already yielded important molecular insight into cellular phenotypes under various experimental conditions [1] and provided new knowledge about the development and treatment of human diseases, such as cancers [2-4]. During the last several years, microarray technology has undergone continued improvement with better quality control in the overall measurement process, ranging from hybridization conditions to image processing techniques [5]. Nevertheless, to fully harness the power of the microarray technology to study biological materials such as cancer tissues, one has to deal with a source of measurement variability that comes from the biological materials themselves, which rarely consist of homogeneous cell populations. For example, except for a few types of immune-privileged tissues such as the brain, most solid tumor tissues contain infiltrating lymphocytes as a result of the immune response. Most tumor tissues also contain endothelial cells as part of the necessary vasculature systems that provide nutrients for the tumor cells. The complexity of this problem is that different tumor tissues contain different proportions of these non-tumor cells. Therefore, if tumor tissues are used without consideration of such a mixing phenomenon, measurement of differential gene expression will certainly be confounded by the heterogeneous cell populations. In some studies [6], pathologists carefully evaluated the tissues and only selected tissues with more than a certain percentage of tumor cells. This prescreening step, however, results in the exclusion of many tumor tissues for the study and contributes to the small sample size problem in some of the studies. Alternatively, laser capture microdissection (LCM) technology can be used to purify the tumor cells from mixed populations [7]. This approach has been very successful in DNA-based studies because of the relatively high stability of DNA. However, for microarray studies, which require less stable RNA, LCM has seen limited success because it is much more challenging to maintain RNA stability during the microdissection process. Other drawbacks of LCM are that such procedures are time-consuming and yield insufficient quantities of RNA, thus requiring multiple amplification steps that may confound quantitative inferences from gene expression data.
A recent paper by Ghosh [8] introduced a mixture model based framework for determining differential expression in the presence of mixed cell populations. In this study, we aim at reconstructing the actual expression values of the pure cell types from the heterogeneous mixtures. That is, we develop a computational method for removing the effect of mixing from heterogeneous samples and to microdissect microarray data in silico. Similar analytical approaches have been previously proposed by Lu et al. [9], Stuart et al. [10] and Venet et al. [11]. Lu et al. focused on estimating the fraction of cells in different phases of the cell cycle whereas Stuart et al. considered the problem of estimating the cell type specific expression patterns over all samples. Here we focus on estimating both the sample and cell type specific expression values using carefully controlled microarray experiments. The inversion of the 'cell mixing effect' can be made appreciably easier by providing estimates of the mixing percentages of different cell types in each measurement, which can be measured by an experienced pathologist. The entire process does not hinge upon such measurements, however, as the mixing percentages can be estimated within the modeling framework. Venet et al. [11] introduced some preliminary methods and results for tackling the same problem as we consider here. In particular, they used a similar regression based framework as in [10] and as we do. We also consider the problem of selecting the correct number of cell types using the cross-validation model selection framework.
Results
The microarray data to which we apply our computational methods consists of five different heterogeneous mixtures of lymph node and colon cancer samples which are hereafter abbreviated as normal and RKO, respectively. For more details, see Materials and methods Section. Each heterogeneous mixture consists of different fractions of different cell samples, see Table 3.
Inversion of sample heterogeneity
The first goal is to invert the mixing effect caused by sample heterogeneity. We apply the linear model developed in Materials and methods Section to the heterogeneous microarray data. The obtained results are presented below.
Because of the inherent variability of individual gene expression values, the performance of the inversion method cannot be estimated based on results for individual genes. (For illustration purposes, the results of inversion of the mixing effect for individual genes are discussed and shown later on in connection with Figures 6 and 7.) Thus, we examine the performance of our method globally, by comparing the measured and estimated expression values of all the genes simultaneously. For performance evaluation and visualization purposes, the dimensionality of the 4704-dimensional expression profiles is reduced using the standard principal component analysis (PCA). The effect of the sample heterogeneity is the same for all the genes within one array. Therefore, for each array, it is useful to combine the results over all the genes. In other words, instead of looking at individual genes, we combine the expression values of all the genes and visualize the results using the most significant principal components. For comparison purposes, we also show the samples used as a reference in the conducted microarray experiments. Since the number of measurements is far smaller than the number of genes, we use a standard approach when solving the PCA eigenvector-eigenvalue problem. Let and , i = 1, ..., K, denote the measured mixture and reference samples, respectively, and let and denote the estimated RKO and normal expression profiles. Let , where . Instead of finding the eigenvalues of the original sample covariance matrix DTD, we compute them for the matrix DDT. The eigenvalues of DTD and DDT are the same and the eigenvectors of DTD can be obtained from the eigenvectors of DDT by multiplying them by DT. Results of the inversion of the sample heterogeneity are shown in Figures 1 and 2. In Figure 1, all five heterogeneous samples are used to estimate the expression values of the pure colon cancer and lymphocyte samples. The two most significant PCA components of all the heterogeneous samples, reference samples, and the estimated expression profile of the pure colon cancer cells and lymphocytes are shown. Figure 1 clearly shows that the heterogeneous samples ('m1' through 'm5') are located almost on a straight line in the 2-dimensional PCA space. Furthermore, the line on which the heterogeneous samples are lying is parallel to the first principal component, suggesting that the most significant variation in the data is due to the linear mixing effect. The estimated expression profile of the pure colon cancer cells and lymphocytes are close to samples number #1 and #5, respectively, indicating that the inversion of the mixing phenomenon produces reasonable results.
The results are more easily appreciated when only the most significant PCA component is shown. As discussed above, the variation in the most significant PCA component is due to the mixing effect. The results in Figure 2 (a) are as in Figure 1, but now shown in 1-dimension in order to facilitate the interpretation. Results in Figure 2 (b), in turn, are as in Figure 2 (a) except that the inversion was done using only the samples #2, #3, and #4. This represents a more difficult and realistic case, since fewer mixtures are available.
When comparing Figure 2 (a) with Figure 2 (b), one can conclude that the method performs slightly better when more samples are used to estimate the true expression profiles – a result that was expected. Overall performance, however, is good in both cases. The estimated expression values for the pure colon cancer (RKO) are close to the mixture #1, as it should be since the mixture #1 corresponds to a measurement of the pure colon cancer. Similarly, the estimated expression values of the pure lymphocytes are close to the mixture #5 as well as to all of the reference samples (note that samples used in the reference channel (Cy5) are the same lymphocytes as the ones used in the mixtures). In Figure 1 and 2 (a), the most significant PCA component and the two most significant PCA components explain about 70.0% and 81.9% of the total variation in the data, respectively. For the reduced data, for which the results are shown in Figure 2 (b), a slightly smaller fraction of the variance is explained, namely about 67.3% and 81.2%. The results obviously depend on the optimality criterion for which we used the standard least squares. Less outlier sensitive results can be obtained with robust regression methods, such as the Huber estimator with the iteratively reweighted least squares implementation [12,13] or median based regression methods [12,14]. The robust methods provided similar global results, but improved results for some individual genes that contained one or more outliers.
Optimization of mixing percentages
In practice, the true mixing percentages are not known but must be measured by some means. Therefore, they are also likely to contain some error. So, in addition to inverting the mixing effect, it is also useful to simultaneously estimate the most likely value of the mixing percentages. This problem can be formulated as type of optimization problem, the details of which are shown in the Materials and methods Section. The proposed optimization scheme was applied to the heterogeneous microarray data. Since the heterogeneous samples #1 and #5 correspond to the cases where only colon cancer cells and lymph node cells are used, respectively, we may assume that α1 = 1 and α5 = 0. Thus, we only estimate the value of the three remaining mixing parameters. However, practically the same results are obtained when all the five mixing parameters are estimated. We found that the convergence of the above method is practically independent of the initialization in step 1. The convergence of the optimization method is illustrated in Figure 3 by showing the evolution of the value of the objective function. Parameters in Â(1) are initialized using the measured values shown in Table 3.
The found optimal values of the mixing percentages are shown in Table 1. The values of the estimated mixing parameters are in a good agreement with the results shown in Figure 2. That is, for instance, the heterogeneous sample #2 is quite close to the heterogeneous sample #1 (α2 ≈ 0.9296) and the heterogeneous sample #4 is fairly far away from the heterogeneous sample #5 (α4 ≈ 0.3796). Note that estimation of the mixing parameters may also compensate for some other errors/biases in the data than just the mixing percentages.
The obtained expression estimates for the pure colon cancer and lymph node samples, when all five heterogeneous samples are used in estimation, are shown in Figure 4. Again, the two most significant PCA components of all the heterogeneous samples, reference samples, and the estimated expression profiles of the pure colon cancer cells and lymphocytes are shown. It is instructive to compare these results with the ones shown in Figure 1. Because the heterogeneous samples are again located almost on a straight line, we use 1-dimensional visualization for the results. Figure 5 shows the obtained expression estimates in 1-dimensional PCA space.
Again, the estimated expression values for the pure colon cancer cells (RKO) are close to those from mixture #1, as it should be, since mixture #1 corresponds to a measurement of the pure colon cancer cells. Similarly, the estimated values from the lymph node sample are close to those from mixture #5 as well as to all of the reference samples. In Figures 4 and 5 (a), the first PCA component and the first two PCA components explain about 69.2% and 81.7% of the total variation in the data. For the reduced data, for which the results are shown in Figure 5 (b), the fractions of variance explained are about 65.1% and 80.4%. Although the fraction of variance explained is slightly smaller than without the optimization of the mixing parameters, the optimized mixing parameters provide a better fit to the data.
Confidence intervals
Above we were only interested in estimating the expression values of the pure cell types. Often it is also useful to assess the confidence intervals of the obtained expression estimates. For that purpose, we consider two methods: one based on Gaussian approximation and the other using bootstrap. (For more details, see Materials and methods Section.)
For illustration purposes, Figure 6 shows the 90% estimated confidence intervals for a set of genes by pooling each of them with the 50 closest genes. The horizontal and vertical axes correspond to the fraction of lymph node cells and the normalized expression value, respectively. In other words, the different heterogeneous mixtures are placed on the x-axis according to the corresponding mixing fractions. The vertical lines at x = 0 and x = 1 expand over the maximum of the two confidence intervals. In most of the cases the two confidence intervals are in good agreement. The confidence intervals can be tightened by measuring more heterogeneous mixtures.
The proposed inversion methods for the sample heterogeneity were also tested on standard non-replicated microarray data by treating the replicated measurements for each gene as individual "genes." The obtained results were qualitatively similar with the ones shown above and only slightly more variable. In a similar fashion, we examined the effect of low quality replicates on the heterogeneity inversion. Slightly less variable results were obtained with a method [15] that detects and removes unreliable replicates prior to the averaging. A drawback of such unreliable spot detection is that, without any missing value estimation method, some of the genes will be excluded from further analysis.
Selection of the number of cell types
It is known that the heterogeneous mixtures used in our experiments consist of only two cell types. However, in general case, heterogeneous mixtures may contain an unknown number of cell types. In those cases, it is useful to assess the validity of the model (i.e., the number of cell types) as well. As introduced in Materials and methods Section, the linear mixing model can be extended to incorporate more than just two cell types. We use a general purpose cross-validation for model selection. In particular, we apply the so-called leave-one-out cross-validation (LOOCV) and test the one, two, and three cell type models. (For more computational details, see Materials and methods Section.)
For the three cell type model, the number of samples does not permit us to optimize the mixing percentages for each cross-validation training data set separately. Therefore, within the cross-validation loop, we use fixed mixing percentages and only estimate the expression values. For the two and three cell type models we use the estimated mixing percentages shown in Tables 1 and 2, respectively. The relative LOOCV errors for the one, two, and three cell type models are 1.79, 1.00, and 2.28, respectively. The results suggest that the two cell type model is indeed the correct one.
Discussion
This paper presents an inversion method for the effects of sample heterogeneity. The proposed method is successfully applied to a carefully controlled microarray data consisting of five different heterogeneous mixtures of lymph node and colon cancer samples. The results demonstrate that both the sample and cell type specific expression values can be reconstructed from heterogeneous mixtures. In some situations, such as cancer metastases in the lymph node, lymphocytes constitute a major cell type beside tumors. Hence, with careful sample preparation, the two cell type model can directly be applied to such cases. For unknown heterogeneous mixtures obtained from more complex cancer samples, the analysis may be a bit more difficult. For example, contaminating cells may include several cell types, such as fibroblasts, endothelial cells, macrophages and lymphocytes. As the proposed method can be applied to any cell types and to any number of cell types, the method works in principle in more complex cases as well. Requirement for the number of measurements necessary for reliable inversion, however, increases together with the number of cell types present in the sample.
We have emphasized that proper inversion of the mixing effect results in more accurate expression values of the pure cell types. While this is true, it must be noted that clinically relevant information may also be incorporated into other populations than the pure (cancer) cells. For example, the degree of lymphocyte infiltration may be clinically important and could be used to complement microarray analysis. However, for comparative microarray analysis, it is important to make comparisons between homogeneous samples so as to minimize the confounding influence of different proportions of contaminating cell types.
Application of 'in silico microdissection' to detection of differentially expressed genes
In order to illustrate the above 'in silico microdissection' in practice, consider the following (hypothetical) experimental setting. Given the three middle mixture measurements (#2, #3, and #4), a goal is to identify a set of genes which are differentially expressed between the colon cancer and the lymph node samples. In a simple approach, often used in practice, the most heterogeneous sample would be discarded since it is measured to contain about 56% (resp., 44%) of colon cancer cells (resp., lymphocytes), thus giving no direct discriminative information about the underlying two samples. For illustration purposes, let us measure the expression difference of a given gene between these two samples using the fold-change, i.e., the expression value of the ith gene in the colon cancer sample, , is regarded as being differentially over-expressed (resp., under-expressed) if the ratio of to the expression value of the same gene in the lymph node sample, , is at least 2 (resp., smaller than 1/2). Of course, in practice, more sophisticated methods for detecting differential expression, including correction for multiple testing, should be used. However, for illustrative purposes, this example will suffice. Since only the heterogeneous samples are available, without any inversion of the mixing effect, one must compare the mixture measurement and . Figure 7 shows some example genes whose expression difference (i.e., the fold-change) between the two heterogeneous samples is within the given threshold (above 1/2 and below 2), but after the 'in silico microdissection,' the expression difference exceeds even a more stringent criterion (approximately 4-fold-change). The measured mixing percentages are used in the estimation (see Table 3). It is clear from this example that the proposed method is able to correctly detect differential expression even from heterogeneous samples, especially when the direct use of such samples may fail to find differential expression. Indeed, the conclusions we can draw based on the red stars are consistent with those that are based on the true homogeneous samples represented by blue squares in Figure 7.
As is evident from the example above, heterogeneity in the biological sample preparation can hinder further statistical analysis steps. Not only can the heterogeneity blur the identification of differentially expressed genes, it can also cause contrary effects. Presence of a considerable percentage of additional cell types can result in the identification of differentially expressed genes that may be unrelated to the biological question being studied. Similarly, irrelevant gene combinations can be discovered in the case of gene expression based classification. For an illustration, see [16] where the authors analyzed a colon cancer data set contaminated with muscle cells.
Although the microarray technology has been improved during the recent years, the measurements are still moderately noisy. The easiest and the most widely used approach for improving the measurement quality is to capture replicated measurements. This may become costly because each additional measurement requires an extra spot on the array, or an extra array. An alternative approach based on so-called composite microarrays was introduced in [17], where several different oligos representing different genes are printed on each spot. The multiplexing results in a mixing effect similar to the one introduced in this manuscript, and the phenomenon can be inverted to get the reconstructed expression values for single genes. The benefit is to obtain more replicated measurements without proportionately increasing the number of printed spots. Closely related ideas have also been introduced from an error-correcting microarray design point of view in [18]. The standard non-repeated microarray method does not tolerate "drop-outs": if a spot is badly corrupted and its intensity cannot be read, the expression value of the corresponding gene will be missed. Khan et al. showed that a certain amount of "drop-outs" can be recovered from the multiplexed samples, thus providing more error-resilient measurements. Following the methods developed in [17,18], instead of multiplexing individual genes on spots, one may wish to multiplex different samples on arrays, thus allowing a fault-tolerant recovery of expression values in the case of corrupted array(s). As a future extension, one can also consider multiplexing both the genes on spots and the samples on arrays. Similar methods for inverting the sample heterogeneity have also been studied in the context of time-series gene expression measurements in [19,20], where the fundamental mixing effect is not due to the different tissue types present in the sample, but due to the loss of synchrony of the cell population. It would be worthwhile to simultaneously study the sample heterogeneity and the loss of synchrony in the future.
Conclusion
In this paper, we proposed a computational framework for removing the effects of sample heterogeneity. In addition to providing estimates of the expression values of the pure (non-heterogeneous) cell samples, the proposed computational methods can also be used to estimate the mixing percentages of different cell types. Furthermore, we also proposed a way of applying general-purpose model selection method for the selection of the correct number of cell types. Application of the proposed methods to a carefully controlled cDNA microarray data obtained from heterogeneous samples shows that the computational methods can invert the effect of sample heterogeneity and, at the same time, estimate the mixing percentages of the different cell types. Furthermore, a general purpose model selection method can be used to select the correct number of cell types.
Materials and methods
Microarray production
RNA isolation, microarray production, and microarray hybridization were carried out as described previously in [21]. RNA from normal human lymph node was purchased from a commercial source (Stratagene, La Jolla, CA). Five μg aliquots of total RNA from normal lymph node and RKO colon cancer cell line were reverse transcribed using Superscript II RT (Invitrogen, Carlsbad, CA) in conjunction with oligodT-T7 primers according to the manufacturer's suggested protocol. The second strand was synthesized using 10U E. coli DNA ligase (vendor), 40U E. coli DNA polymerase I (vendor), and 2U E. coli Rnase H (vendor). This reaction was stopped with EDTA and then cleaned with Qiagen's PCR Purification kit (Qiagen, Valencia, CA). The double stranded cDNA was then amplified by an in vitro transcription reaction (Ambion, Austin, TX) and cleaned with Qiagen's Rneasy kit (Qiagen, Valencia, CA). Each amplified cRNA sample was then quantitated using a Beckman DU640 spectrophotometer (Beckman, Fullerton, CA). Five μg amplified cRNA from Stratagene's normal lymph node was labeled with Cy5 for each microarray hybridization. Mixtures of appropriate volumes of cRNA from normal lymph node and RKO were labeled with Cy3 in a reverse transcription reaction using Superscript II RT (see Table 3). Labeled samples were co-hybridized overnight at 60°C in a humidified incubator on a cDNA microarray containing 4704 human genes in duplicate produced in-house. The 4704 genes represent most of the known genes in the cDNA library we used to generate the microarrays. For the purpose of this study, the identity of the genes is not very important since we only study the general effect of sample heterogeneity. As the mixing effect is the same for all the genes, we expect to have similar results when the whole genome arrays are used. Slides were scanned with an LS-IV laser scanner (Genomic Solutions, Ann Arbor, MI). In total, five different heterogeneous mixtures were measured. The measured mixing percentages are shown in Table 3.
Preprocessing
The microarray data consists of five different heterogeneous mixtures of lymph node and colon cancer samples which are hereafter abbreviated as normal and RKO, respectively. (For more details, see Microarray production Section above and Table 3.) The gene expression data set was preprocessed as follows. The replicated background-subtracted signal intensities were averaged and log2-transformed, and the dye-bias effect was corrected in the log2-domain using the standard lowess smoothing-based normalization (see e.g. [22]) with smoothing parameter f = 0.7. Because the averaging effect (source of heterogeneity) takes place on the molecular level, the phenomenon must be modeled using the absolute expression values. Therefore, after the correction of the dye-bias, the data were transformed back to the original domain using the inverse of the log2-transformation. Correspondingly, single-channel data were used for further analysis. In order to mitigate the between array variability, the data were further standardized for each array and the two channels separately.
Modeling sample heterogeneity
The two samples, RKO colon cancer cells and normal lymphocytes, are mixed at the extracted RNA level. Therefore, without any further verification, the model can be assumed to be linear. Lymphocytes were used because tumor tissues often contain infiltrating lymphocytes, especially in tumor metastases in the lymph nodes. Let and denote the expression level of the ith gene in the colon cancer (RKO) and in the lymph node (normal) samples, respectively. Assuming only two different cell types are mixed, the sample heterogeneity is modeled by a simple linear model
where denotes the expression value of the ith gene in the kth heterogeneous sample, and 0 ≤ αk ≤ 1 denotes the fraction of the colon cancer cells in the kth mixture. It is worth noting that we use the same mathematical model for the sample heterogeneity as in [9-11]. Also note that in Equation (1) it is assumed that the expression level in RKO () and normal () is "fixed" and does not change between heterogeneous measurements. In other words, the measurements come from the same heterogeneous sample with different mixing fractions. In order to allow variation in the expression values between different samples/treatments/time points, the same model can be applied separately to each set of measurements from the other samples/treatments/time points. The same model can also be extended to more than two cell types (for more details, see Selection of the number of cell types Section below).
Inversion of sample heterogeneity
The first objective is to invert the mixing effect shown in Equation (1), that is, to obtain estimates for the expression values of the pure colon cancer cells and the pure lymphocytes. In practice, however, the measured expression values, yi, include one or more sources of noise. By making some distributional assumptions, one could use standard model-based estimation methods. However, in order to avoid making additional modeling assumptions, we prefer to use a general purpose least squares method to estimate the gene expression levels corresponding to the pure samples.
Let the number of genes be n and assume that one has measured the expression values for K different heterogeneous mixtures. Thus, one has measurements , 1 ≤ i ≤ n, 1 ≤ k ≤ K. Let us also assume for now that the mixing percentages are known or have been measured. For the ith gene the sample heterogeneity can be expressed as (excluding all noise terms)
When including all n genes, the above model can be rewritten as
where 0 denotes the K-by-2 zero matrix. Let the block matrix in Equation (2) above be denoted as à Assuming the column rank of A is full, the well-known least squares solution is given by
where . Due to the structure of the matrix Ã, the least squares solution can be obtained gene-wise as .
The Gauss-Markov theorem says that the standard least squares solution is indeed the best linear unbiased estimate if the noise in the measurements is additive and i.i.d. with constant variance. However, a common observation is that the homoscedasticity does not always hold for microarray data, but instead, the noise variance depends on the underlying signal intensity [23,24]. Such heteroscedasticity may decrease the power of the inversion method shown in Equation (3). Fortunately, the structure of the matrix à ensures that the inversion can also be performed for each gene separately. Consequently, it is not necessary for the homoscedasticity to hold globally. Indeed, all we need to assume is that the noise variance is approximately constant for each gene separately.
Also note that, in this two cell type model, no prior knowledge about the expression values of either of the two cell types is needed since the method estimates the expression values for both of the two cell types. The same is also true for more general models including more cell types, assuming the model is sufficiently over-determined (see also Selection of the number of cell types Section below).
Optimization of mixing percentages
In practice, the mixing percentages must be measured by some means. Therefore, they are also likely to contain some error. So, overall, one would like to estimate not only the expression values for the pure cell types but also the most likely value of the mixing percentages.
Assuming the model in Equation (2) is sufficiently over-determined, the mixing parameters can be adjusted computationally, too. Let us again consider the case where only two different cell types are mixed. Note that K denotes the number of different heterogeneous mixtures measured. Therefore, the regression matrix à in Equation (2) has only K free parameters. Since the number of expression values to be estimated is 2n, the total number of free parameters in Equation (2) is 2n + K. The number of equations in Equation (2) is Kn. Hence, the model is over-determined if Kn > 2n + K, which, for a fixed n ≥ 3, holds if K > 2. (Note that in our case we have measured five different heterogeneous mixtures, i.e., K = 5.) As above, no assumptions on the noise distributions are being made and we use the least squares method. This results in the following optimization problem
A similar optimization problem was also introduced in [11].
Because the objective function in Equation (4) above is minimized over both à and x, the objective function is not linear in the parameters anymore and, therefore, cannot be solved as in Equation (3). In general, any iterative optimization method can be used to get a solution. Iterative methods usually become inefficient/unstable as the number of parameters to be optimized increases. In this case, the number of free parameters in à and x is 2n + K. Therefore, we use a two-step approach in the optimization. In the first step, given proper initial value for Ã, the least squares solution for x is found using Equation (3). In the second step, the mixing percentages are optimized in the least squares sense (subject to the constraints 0 ≤ αk ≤ 1 for all 1 ≤ k ≤ K) using the previously found value for x. These two steps are then repeated, essentially resulting in a type of expectation-maximization (EM) approach. A similar iterative procedure was also proposed in [11], except with different constraints. Note that when Equation (4) is minimized over Ã, given the value of x, the optimization problem is again linear in its parameters. Assuming the constraints are not violated, the standard equation (similar to the one in Equation (3)) can be applied. If that is not the case, then any general-purpose constrained optimization method may be applied. Let (resp. Â(j)) denote the value of x (resp. Ã) after the jth iteration. Details of the algorithm are shown in Figure 8. Clearly, at each iteration of steps 2 and 3, the value of the objective function is decreased. Thus, a minimum will be found.
Confidence intervals
It is useful to assess the confidence intervals of the obtained expression estimates. As explained above, the Gauss-Markov theorem is applied gene-wise that greatly alleviates the issue of heteroscedasticity. Should the noise variance σ2 be constant, then the variance of the estimated expression values would be . Due to the special structure of the matrix à (i.e., the gene-wise inversion of the mixing effect), the variance of the estimated expression values for the ith gene can be expressed as
where is the noise variance for the ith gene. A straightforward way of obtaining an estimate of the variance is to compute the sample noise variance for each gene and then apply Equation (5) to get . That would result in somewhat sensitive variance estimates since there are only K = 5 error residuals associated with each gene. A better alternative is to pool genes which have approximately the same average expression value and then compute the sample noise variance from the error residuals of the pooled genes. Although we do not assume Gaussian noise distribution, we can resort to the Gaussian approximation when computing the confidence intervals. For example, using the Gaussian approximation, the 1 - 2α confidence interval for estimated expression value of the ith gene in the colon cancer cells is , where Φ-1(·) is the inverse of the standard normal cumulative distribution function and denotes the (1,1) element of the estimated variance matrix (similarly for the lymph node sample: ). Alternatively, the confidence intervals can be obtained using the non-parametric bootstrap framework [25]. Here we consider the method in which one re-samples the error residuals with replacement (within the set of pooled genes) and computes the confidence intervals directly from the α and 1 - α percentiles of the bootstrap distribution of the expression estimates.
Selection of the number of cell types
Although it is known that only two cell types are mixed in our experiments there may be other experimental settings where the number of cell types may be unknown. Then it is useful to assess the validity of the model as well. As was mentioned above, the linear mixing model can be extended to incorporate more than just two cell types using a straightforward extension: , where denotes the expression value of the ith gene in the jth cell type, and 0 ≤ ≤ 1 denotes the fraction of the jth cell type in the kth mixture. The mixing percentages must also satisfy for all k. The significance of different 'regression coefficients' could be tested using standard regression-based statistical tests. Since those tests apply only to Gaussian noise we recommend using a general purpose cross-validation for model selection (see e.g. [26]). Here we consider the leave-one-out cross-validation (LOOCV) and test the one, two, and three cell type models. Thus, each heterogeneous sample is left out from the training data at a time, the regression coefficients are estimated based on the remaining four samples, and the model is then tested on the sample which was left out from the training data set.
Authors' contributions
VD and WZ conducted the experiments and HL, IS, OY-H and WZ developed the computational methods. HL, IS and WZ prepared the manuscript.
Acknowledgements
This study was partially supported by Tampere Graduate School in Information Science and Engineering (TISE), Academy of Finland, the Tobacco Settlement Fund to M. D. Anderson Cancer Center as appropriated by the Texas Legislature, a generous donation from Kaddorie Foundation, a grant from the Goodwin Fund, and the Cancer Center Support Grant from NCI/NIH.
Figures and Tables
Figure 1 Results of the sample heterogeneity inversion in the 2-dimensional PCA space. All five heterogeneous samples are used to estimate the expression profiles of the pure colon cancer cells and lymphocytes. Symbols: estimated expression profiles of the pure colon cancer cells and lymphocytes (gray stars), mixture samples (green triangles), and reference samples (red circles). The labels next to each green triangle (resp. red circle) denote the number of the heterogeneous (resp. reference) sample, e.g., 'm1' = mixture sample #1 and 'r1' = reference sample #1, etc. (see also Table 3). The estimated expression profile of the pure colon cancer cells and lymphocytes have labels 'e1' and 'e5', respectively. See text for further details.
Figure 2 Results of the sample heterogeneity inversion in the 1-dimensional PCA space. (a) All five heterogeneous samples, and (b) only the heterogeneous samples #2, #3, and #4 are used to estimate the expression profiles of the pure colon cancer cells and lymphocytes. The height of each bar corresponds to the value of the most significant PCA component. Each bar corresponds to a heterogeneous sample, reference sample, or estimated expression profile and is labelled with the corresponding text.
Figure 3 Evolution of the value of the objective function. The red (resp. blue) graph corresponds to the value of the objective function after step 2 (resp. step 3).
Figure 4 Results of the combined sample heterogeneity inversion and the estimation of the most likely values of the mixing parameters in the 2-dimensional PCA space. All five heterogeneous samples are used to estimate the expression profiles of the pure colon cancer and lymphocyte. Symbols: estimated expression profiles (gray stars), mixture samples (green triangles), and reference samples (red circles). See text for further details.
Figure 5 Results of the combined sample heterogeneity inversion and the estimation of the most likely values of the mixing parameters in the 1-dimensional PCA space. (a) All five heterogeneous samples, and (b) only the heterogeneous samples #2, #3, and #4 are used to estimate the expression profiles of the pure colon cancer cells and lymphocytes. Each bar corresponds to a heterogeneous sample, reference sample, or estimated expression profile and is labelled with the corresponding text. The height of each bar corresponds to the value of the most significant PCA component.
Figure 6 Estimated 90 % confidence intervals for the estimated expression values of the pure cell types. The horizontal and vertical axes correspond to the fraction of lymph node cells and the normalized expression value, respectively. Symbols: the measured expression values (blue circles), the estimated expression values of the pure cell types (red stars), regression-based confidence intervals (red points), and bootstrap-based confidence intervals (red x-marks).
Figure 7 Detecting differentially expressed genes. A set of genes which are not found to be significantly differentially expressed based on the heterogeneous measurements (samples #2 and #4, blue circles). After the inversion of the mixing effect, however, the expression difference between the estimated pure colon cancer cells and lymphocytes (red stars) meet even a more stringent criterion of differential expression. The horizontal and vertical axes correspond to the fraction of lymph node cells and the normalized expression value, respectively. Symbols: the heterogeneous samples (blue circles), the estimated expression values (red stars), and the measured expression values of the pure colon cancer cells (blue squares). See text for more details.
Figure 8 The two-step optimization algorithm. Details of the two-step algorithm used for the optimization problem shown in Equation (4).
Table 1 The estimated mixing percentages. The found optimal values of the mixing percentages.
sample #2 sample #3 sample #4
RKO 92.96 65.06 37.96
normal 7.04 34.94 62.04
Table 2 The estimated mixing percentages for the three cell type model. The found optimal values of the mixing percentages for the three cell type model.
cell type sample #1 sample #2 sample #3 sample #4 sample #5
RKO 98.15 67.94 58.70 30.74 0
normal 1.22 16.39 36.53 62.97 96.95
the 3rd cell type 0.63 15.66 4.77 6.29 3.05
Table 3 The measured mixing percentages. The measured mixing percentages (RKO/normal) in the five heterogeneous samples.
sample #1 sample #2 sample #3 sample #4 sample #5
RKO 100 80 56 30 0
normal 0 20 44 70 100
==== Refs
Spellman PT Sherlock G Zhang MQ Iyer VR Anders K Eisen MB Brown PO Botstein D Futcher B Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization Mol Biol Cell 1998 9 3273 3297 9843569
Golub T Slonim D Tamayo P Huard C Gaasenbeek M Mesirov J Coller H Loh M Downing J Caligiuri M Bloomfield C Lander E Molecular classification of cancer: class discovery and class prediction by gene expression monitoring Science 1999 286 531 537 10521349 10.1126/science.286.5439.531
van't Veer LJ Dai H van de Vijver MJ He YD Hart AA Mao M Peterse HL van der Kooy K Marton MJ Witteveen AT Schreiber GJ Kerkhoven RM Roberts C Linsley PS Bernards R Friend SH Gene expression profiling predicts clinical outcome of breast cancer Nature 2002 415 530 536 11823860 10.1038/415530a
Zhang W Ramdas L Shen WP Song WS Hu L Hamilton SR Apoptotic response to 5-fluorouracil treatment is mediated by reduced polyamines, non-autocrine fas ligand and induced tumor necrosis factor receptor 2 Cancer Biol Ther 2003 2 572 578 14614330
Zhang W Shmulevich I Astola J Microarray Quality Control 2004 John Wiley and Sons
Fuller GN Rhee CH Hess K Caskey L Wang R Bruner JM Yung WKA Zhang W Reactivation of insulin-like growth factor binding protein 2 expression in glioblastoma multiforme: a revelation by parallel gene expression profiling Cancer Res 1999 59 4228 4332 10485462
Emmert-Buck MR Bonner RF Smith PD Chuaqui RF Zhuang Z Goldstein SR Weiss RA Liotta LA Laser capture microdissection Science 1996 274 998 1001 8875945 10.1126/science.274.5289.998
Ghosh D Mixture models for assessing differential expression in complex tissues using microarray data Bioinformatics 2004 20 1663 1669 14988124 10.1093/bioinformatics/bth139
Lu P Nakorchevskiy A Marcotte EM Expression deconvolution: a reinterpretation of DNA microarray data reveals dynamic changes in cell populations Proc Natl Acad Sci USA 2003 100 10370 10375 12934019 10.1073/pnas.1832361100
Stuart RO Wachsman W Berry CC Wang-Rodriguez J Wasserman L Klacansky I Masys D Arden K Goodison S McClelland M Wang Y Sawyers A Kalcheva I Tarin D Mercola D In silico dissection of cell-type-associated patterns of gene expression in prostate cancer Proc Natl Acad Sci U S A 2004 101 615 620 14722351 10.1073/pnas.2536479100
Venet D Pecasse F Maenhaut C Bersini H Separation of samples into their constituents using gene expression data Bioinformatics 2001 17 S279 287 11473019
Rousseeuw PJ Leroy AM Robust Regression and Outlier Detection 1987 John Wiley
Holland PW Welsch RE Robust regression using iteratively reweighted least-squares Commun Stat Theory Methods 1977 A6 813 827
Rousseeuw PJ Least median of squares regression J A Stat Assoc 1984 79 871 881
Hao X Sun B Hu L Lähdesmäki H Dunmire V Feng Y Zhang S-W Wang H Wu C Wang H Fuller GN Symmans WF Shmulevich I Zhang W Differential gene and protein expression in primary breast malignancies and their lymph node metastases as revealed by combined cDNA microarray and tissue microarray analysis Cancer 2004 100 1110 1122 15022276 10.1002/cncr.20095
Ben-Dor A Bruhn L Friedman N Nachman I Schummer M Yakhini Z Tissue classification with gene expression profiles J Comput Biol 2000 7 559 584 11108479 10.1089/106652700750050943
Shmulevich I Astola J Cogdell D Hamilton SR Zhang W Data extraction from composite oligonucleotide microarrays Nucleic Acids Res 2003 31 e36 12655024 10.1093/nar/gng036
Khan AH Ossadtchi A Leahy RM Smith DJ Error-correcting microarray design Genomics 2003 81 157 165 12620393 10.1016/S0888-7543(02)00032-0
Lähdesmäki H Huttunen H Aho T Linne M-L Niemi J Kesseli J Pearson R Yli-Harja O Estimation and inversion of the effects of cell population asynchrony in gene expression time-series Signal Process 2003 83 835 858 10.1016/S0165-1684(02)00471-1
Bar-Joseph Z Farkash S Gifford DK Simon I Rosenfeld R Deconvolving cell cycle expression data with complementary information Bioinformatics 2004 20 I23 I30 15262777 10.1093/bioinformatics/bth915
Shmulevich I Hunt K El-Naggar A Taylor E Ramdas L Laborde P Hess KR Pollock R Zhang W Tumor specific gene expression profiles in human leiomyosarcoma: an evaluation of intratumor heterogeneity Cancer 2002 94 2069 2075 11932911 10.1002/cncr.10425
Quackenbush J Microarray data normalization and transformation Nat Genet 2002 496 501 12454644 10.1038/ng1032
Huber W von Heydebreck A Sultmann H Poustka A Vingron M Variance stabilization applied to microarray data calibration and to the quantification of differential expression Bioinformatics 2002 18 S96 104 12169536
Durbin BP Rocke DM Variance-stabilizing transformations for two-color microarrays Bioinformatics 2004 20 660 677 15033873 10.1093/bioinformatics/btg464
Efron B Tibshirani RJ An introduction to the bootstrap 1993 New York: Chapman & Hall
Hastie T Tibshirani R Friedman J The Elements of Statistical Learning: Data Mining, Inference, and Prediction 2001 Springer-Verlag
|
15766384
|
PMC1274251
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 14; 6:54
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-54
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-551576638310.1186/1471-2105-6-55Research ArticleSpeeding disease gene discovery by sequence based candidate prioritization Adie Euan A [email protected] Richard R [email protected] Kathryn L [email protected] David J [email protected] Ben S [email protected] Medical Genetics Section, Department of Medical Sciences, The University of Edinburgh, Edinburgh, UK2005 14 3 2005 6 55 55 22 10 2004 14 3 2005 Copyright © 2005 Adie et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Regions of interest identified through genetic linkage studies regularly exceed 30 centimorgans in size and can contain hundreds of genes. Traditionally this number is reduced by matching functional annotation to knowledge of the disease or phenotype in question. However, here we show that disease genes share patterns of sequence-based features that can provide a good basis for automatic prioritization of candidates by machine learning.
Results
We examined a variety of sequence-based features and found that for many of them there are significant differences between the sets of genes known to be involved in human hereditary disease and those not known to be involved in disease. We have created an automatic classifier called PROSPECTR based on those features using the alternating decision tree algorithm which ranks genes in the order of likelihood of involvement in disease. On average, PROSPECTR enriches lists for disease genes two-fold 77% of the time, five-fold 37% of the time and twenty-fold 11% of the time.
Conclusion
PROSPECTR is a simple and effective way to identify genes involved in Mendelian and oligogenic disorders. It performs markedly better than the single existing sequence-based classifier on novel data. PROSPECTR could save investigators looking at large regions of interest time and effort by prioritizing positional candidate genes for mutation detection and case-control association studies.
==== Body
Background
Over the last twenty years the genes underlying more than a thousand classically Mendelian disorders have been successfully identified. By contrast, only a relatively small number of genetic components of complex traits have been characterized [1].
Regions of interest identified through complex-trait linkage studies regularly exceed 30 centimorgans in size and can contain hundreds of genes. The traditional candidate-gene approach to reducing this number of genes to a manageable level involves attempting to match functional annotation to knowledge of the disease or phenotype under investigation. Unfortunately this approach has been characterized by unsubstantiated and unreplicated claims [2].
Problems arise firstly because the link between genotype and phenotype in complex disorders tends to be weak; matching a single gene's functional annotation to a phenotype is unlikely to be successful unless the gene in question is clearly related to some known pathogenesis of the disease. Secondly, functional annotation of the human genome is incomplete and biased towards better studied genes which have higher levels of annotation. Furthermore, assigning functional annotation is a time-consuming process which is unavoidably error-prone [3,4] and, if taken at face value, misannotated genes can mislead or delay researchers [5].
Van Driel et al. developed a web-based system for automating the annotation based candidate-gene approach [6] that collates expression and phenotypic data from nine different databases and returns genes that conform to investigator-defined criteria. Recently several other candidate-gene identification systems that rely on grouping Gene Ontology (GO) terms have been described [7,8], notably POCUS [9], which finds genes across multiple susceptibility loci that share Interpro [10] domains and GO terms. These systems all rely on functional annotation to make correct predictions, but given that such annotation is incomplete and inherently biased towards a particular subset of genes, a more robust option might be to use sequence-based features instead.
It has been suggested that the genes underlying human hereditary disease share certain distinctive, sequence-based features such as larger gene size [11]. By using machine learning algorithms we aimed to discover such common patterns that could be applied to create an automatic classification scheme capable of identifying genes more likely than not to be involved in disease.
Machine learning has moved rapidly from the field of experimental artificial intelligence to that of applied science. Bioinformatics researchers have been quick to adopt machine learning algorithms in a variety of different situations and their use is now widespread [12]. Lopez-Bigas et al. recently presented a relatively successful decision tree created using such techniques [13] which used amino-acid length and a measure of sequence conservation across species of genes as features to predict genes likely to be involved in hereditary disease. Our approach was related but examined a broader set of features and algorithms, producing a significantly more successful classifier that is able to predict genes involved in both Mendelian and more complex traits. We have also created a web interface to allow researchers to easily classify individual genes or whole regions of the genome and made it freely accessible at .
Results
Defining features and building the training set
A set of features was chosen based on a comparative study of ~ 18,000 known genes from Ensembl [14] which are not known to be involved in human disease and the 1,084 Ensembl genes also listed in Online Mendelian Inheritance in Man (OMIM) [15]. The feature set (described in Table 1) reflects the structure, content and phylogenetic extent (the extent to which a gene is conserved back through evolution based on homologs in other species) of each gene examined. We included signal peptide and transmembrane domain predictions; though these are strictly speaking functional attributes they can be calculated with a high degree of accuracy directly from sequence.
Table 2 lists the features we found to be different between the Ensembl genes in OMIM and those not in OMIM. Using the Mann-Whitney U test we found highly significant differences between the gene, cDNA and protein sizes of the two sets (P < 0.001). The genes listed in OMIM were significantly larger and encoded larger proteins; this confirms previous findings [11,13] which noted that the genes and proteins involved in human disease tend to be larger than average. Similarly, we found that the genes listed in OMIM were far more likely to have well conserved best reciprocal hit (BRH) homologs with other species and in particular with mice; this also concurs with previous studies [13,16]. The percentage of gene products that are secreted was much higher in the set of genes listed in OMIM than on average (P < 0.0001) and perhaps unsurprisingly given the larger sizes of genes involved in disease and the correlation between gene size and exon number we found a highly significant difference in the number of exons per gene (P < 0.001 using the Mann-Whitney U test). Genes listed in OMIM had a median of 10 exons while genes not known to be involved in disease had a median number of 8. Genes listed in OMIM were more frequently expressed in specific tissues (P < 0.001) and again this confirms previous findings [17] – however, it was decided to exclude tissue specificity from our feature set in order to avoid potential bias (see Methods).
We also found novel differences between the two sets of genes. There was a small difference (P < 0.028) in the number of CpG islands at the 5' end of genes listed in OMIM and those not, with slightly more genes listed in OMIM having 5' CpG islands, which are associated with both housekeeping genes and to a lesser extent tissue specific genes [18]. There was also a significant difference (P < 0.01) in the length of the 3' UTR between genes listed in OMIM (median 488 bp) and those not involved in disease (median 446 bp). There was also a significant disparity (P < 0.01) in the distance to the nearest neighbouring gene – genes listed in OMIM had a median distance of 52 kb to their neighbours while genes not known to be involved in disease had a median distance of 46 kb. To our knowledge these features have not been previously reported.
Graphs showing the different distributions of selected features in the two sets are shown in Figure 1. Though some of the differences we found have previously been described in literature, the discrepancy in 3' UTR length has to our knowledge not been examined before and cannot be easily explained in terms of correlation to other, known feature differences. Two other novel features are the distance to the nearest neighbouring gene and the number of exons; both of these are quite strongly correlated to gene size (with Spearman correlation coefficients of 0.69 and 0.71, respectively).
We also studied the number of Interpro domains in each set of genes and found significant differences but concluded that a bias existed towards better studied genes. Therefore we excluded this feature from our study (see Methods).
Automatic classifiers are created by being trained on a set of genes that has already been classified manually. Our training set of genes was made up of the 1,084 genes found in both OMIM and Ensembl (the "disease genes") and 1,084 Ensembl genes not known to be involved in disease (the "control genes") which were selected at random from the larger set of ~ 18,000 as a representative sample.
Choosing an algorithm
We used Weka [19] as the platform for our machine learning experiments. A variety of different machine learning methods were examined but the alternating decision tree algorithm was chosen as the basis of our classification scheme as it couples high accuracy with a relatively small set of rules [20]. The advantage of decision tree based schemes over other popular algorithms such as k-Nearest Neighbour, Support Vector Machines and Bayesian Networks is that the rules that are produced for classifying instances can be interpreted more easily by non-expert users. This is particularly true for the alternating decision tree algorithm, which typically produces trees that are just as predictive as those created by more traditional decision tree algorithms but that are far more concise and thus easier to understand. Alternating decision trees also allowed us to measure the contribution of each feature to the final classification of a gene, which might provide insight into the essential differences between those genes more and less likely to be involved in disease.
Alternating decision trees are created by adding rules to the tree in an iterative fashion in the order of their predictive power, with the more effective rules being added first. These rules are automatically derived from the differences between the disease and control genes in the training set provided. A new rule is added to the tree either as a new "node" or as a child of an existing node. With Weka, the number of nodes to add to the tree is specified by the user before training begins. Too few nodes and the tree will be sparse, without enough cumulative discriminatory power to make confident classifications. Too many nodes, on the other hand, will result in an overly-complex tree where later nodes with weak predictive power can distort the effects of earlier, more predictive nodes.
On the basis of past experience we chose to limit the size of our alternating decision tree to fifteen nodes, which is a good balance of predictive power and complexity. As each node represents a rule that tests a single sequence feature, this meant that of the two dozen sequence features available a maximum of fifteen would be used in the final tree.
An alternating decision tree with fifteen nodes was produced by training on the training set of genes and is shown in Figure 2. We also produced trees with ten and twenty nodes for comparison and discovered using the measurements described below that classifier performance was indeed poorer than with the fifteen node version (details not shown).
The alternating decision tree
A gene is classified with the tree in Figure 2 by beginning at the node marked "Start" and then following each branch in turn. Upon reaching a node that contains an assumption – for example, that the gene length is larger than a given number – the "yes" or "no" branch is followed as appropriate. If the relevant feature – the paralog percentage identity, for example – is "unknown", neither branch is followed. Adding up each of the numbers in rectangles that are encountered along the way results in a final score that reflects the relative confidence of the classification. The classification itself is based on the sign of the score – if negative the gene is generally more likely to be involved in hereditary disease, if positive the gene is generally less likely to be involved in hereditary disease.
We tested the classifier on our training set of genes. 77% of the disease genes were correctly identified (that is to say had a negative score). In contrast, 42% of the 1,084 control genes were classified as disease genes (were false positives). As this is a predictive approach – we cannot say a priori how much of the genome and thus the representative sample in the training set is made up of genes that are involved in disease but are not yet characterized – at least some of these apparently incorrect classifications are likely to be correct.
We ran a tenfold cross-validation test to get a conservative estimate of how our classifier might perform on unseen data. Tenfold cross-validation is a widely used technique in machine learning and involves partitioning the whole training set into ten independent "folds" each with the same balance of disease genes and control genes. The classifier is trained on nine of the partitions and tested on the remaining partition. This is repeated until each partition has been tested on a new classifier built with the remainder of the training set and simulates the performance of the chosen algorithm and feature set on unseen data. On average, 70% of the disease genes were correctly identified during cross-validation with 43% of control genes classified as false positives. This is comparable to the results obtained by Lopez-Bigas et al. [13] during cross-validation. Table 3 contains more detailed statistics relating to classifier performance.
As the alternating decision tree outputs a score that can be thresholded, it is a relatively simple matter to increase specificity (precision) at the expense of sensitivity (recall). Receiver Operating Characteristic (ROC) curves can be used to visualise classifier performance with different combinations of specificity and sensitivity. The x-axis of a ROC curve represents the fraction of false positives and the y-axis the fraction of true positives in the classifier results. As the number of true positives (sensitivity) increases, so too does the number of false positives (decreasing specificity). Figure 3 shows the ROC curves for the classifier on the training set and the two test sets which are described below.
Table 4 shows the relative importance of the eleven different sequence features used by the classifier. We calculated these values by testing the classifier on our training set and, for each gene, keeping track of the percentage contribution of each feature to the final score. These percentages were then averaged out over all genes predicted as likely to be involved in disease. It should be noted that while the percentages given accurately reflect the relative contribution of each feature to our classifier they are meaningless when taken out of context; by themselves, for example, GC content and the % identity of a worm homolog are not necessarily equally predictive features for distinguishing between genes that are more likely and those less likely to be involved in disease.
We implemented our classifier as a standalone script in Perl and designed an associated web interface to aid in the interpretation of the results produced. The resulting software is named PROSPECTR (for PRiOrization by Sequence &PhylogEnetic features of CandidaTe Regions) and is freely accessible together with training and test sets of genes at . The web interface allows researchers to quickly obtain scores for regions of the genome or individual genes of interest.
Further testing
Evaluating classifier performance on the training set alone is potentially misleading as over-fitting may have occurred. Over-fitting happens when a classifier generalises only to the extent necessary to work well on the training data, resulting in poor performance on data that was not seen during the training process. Cross-validation provides a measure of the performance of our approach in general, but doesn't reflect actual PROSPECTR performance accurately as the alternating decision trees created for each fold are different. We therefore created two test sets independent of the training set.
The first independent test set (the "HGMD set") contained 675 genes associated with disease listed in the Human Gene Mutation Database [21] and 675 genes not known to be involved in disease that were picked at random from Ensembl. The second (the "oligogenic set") contained 54 genes not known to be involved in disease and picked at random from Ensembl and 54 genes not listed in OMIM but associated with different oligogenic disorders including inflammatory bowel disease, Parkinsons, Retinitis Pigmentosa and autosomal recessive limb-girdle muscular dystrophy.
We were unable to obtain a sizeable, reliable set of genes involved in complex traits; this meant that classifier performance could not be tested on the components of complex disease. This may change in the future as resources such as the Genetic Association Database [22] develop further and more association data becomes accessible.
71% (478) of the disease genes from the HGMD set and 72% (39) of the genes from the oligogenic set were correctly identified by the classifier, with 42% (282) and 41% (22) of control genes misclassified respectively. These results are similar to those obtained on the training set, suggesting that over-fitting did not occur. They also suggest that our sequence-based approach works equally well for finding genes involved in both oligogenic and monogenic disorders.
Lopez-Bigas et al [13] used a larger set of disease genes during training. Only 260 of the genes from the HGMD set were independent of the training sets of both PROSPECTR and the Lopez-Bigas classifier. As a comparative measure, these 260 genes were scored using both classifiers. PROSPECTR correctly identified 72% (189) of the disease genes while the Lopez-Bigas classifier identified 47% (123).
PROSPECTR, however, had a higher false positive rate, categorising ~ 44% of the whole human genome as likely to be involved in disease while the Lopez-Bigas classifier categorised ~ 31% of the whole human genome as likely to be involved in disease. To see how this might have affected recall we calculated the Kappa statistic [23] for the results from both classifiers. The Kappa statistic is a measurement of agreement between predicted and actual classifications and takes false positive rates into account. It is a number between 1 (symbolising perfect agreement between predicted and actual classifications) and 0 (symbolising no agreement). On the independent HGMD set of 260 genes and assuming a false positive rate of 31%, the Lopez-Bigas classifier had a Kappa statistic of 0.158 while PROSPECTR assuming a false positive rate of 44% had a Kappa statistic of 0.282, a factor of almost twofold. This suggests that PROSPECTR is substantially more adept than the Lopez-Bigas classifier at correctly classifying unseen data.
By ranking genes by score in descending order, it is possible for PROSPECTR to create a list of genes for any given locus the top of which is enriched for genes that have a higher probability of being involved in disease.
To test this we took the HGMD set and for each gene created an artificial locus 30 Mb in size consisting of the gene from the HGMD set and all known genes within 15 Mb on either side on the same chromosome. The gene taken from the HGMD set was in each case designated the "target gene" and by scoring each gene in the artificial locus and then ranking them we were able to see where the target gene appeared in the ordered list that was created.
For the 675 genes from the HGMD set the average number of genes per list was 202. Target genes were in the top 5% of the ordered list 68 times (10% of the time), top 10% 125 times (18%), top 50% 510 times (75%) and the top 75% 639 times (94%).
We repeated the procedure for the 1,084 genes listed in OMIM from our training set. The average number of genes per list was 198 and target genes were in the top 5% 171 times out of 1084 (15%) and the top 50% 873 times (80%).
The genes from the training and HGMD sets are mostly Mendelian monogenic disorders; to see if the classifier was equally successful at enriching loci involved in more complex diseases we took the list of 219 genes likely to be involved in oligogenic disorders used as a test set by POCUS [9].
For these 219 genes involved in oligogenic disorders the average number of genes per list was 209 and the target gene was in the top 5% 29 times (13.4%) and the top 50% 172 times (79%). Figure 4 shows a graphical representation of these results.
Performance on different types of mutation
Gene records from the HGMD contain information about the number of different mutations associated with any phenotypes linked to that gene, split into three types: nucleotide substitutions, micro-lesions and gross lesions (including repeat variations and complex rearrangements). For example, the Huntington gene (HD) is recorded as being implicated in Huntington disease, which is associated with a gross lesion. The Haemoglobin beta gene (HBB) is recorded as being implicated in sickle cell anaemia, associated with nucleotide substitutions.
Of the HGMD set we used to test performance, 297 genes were associated with nucleotide substitutions only, 55 with gross lesions only and 27 with micro-lesions only. We tested each subset separately to determine if the underlying cause of disease influenced PROSPECTR's performance.
We found that 75% and 77% of the genes involved in disease and associated only with nucleotide substitutions or only with micro-lesions, respectively, were correctly identified by PROSPECTR. However, only 54% of the genes involved in disease and associated only with gross lesions were identified. This suggests that the decision tree used by PROSPECTR is better at identifying genes likely to be involved in disease because of small or point mutations than genes involved in disease because of more drastic events like gross deletions, insertions and chromosomal aberrations.
Whole genome analysis
PROSPECTR was used to score every known gene in the Ensembl database on the likelihood that it is involved in human hereditary disease. We normalised the score α given to each gene with the equation
where gamma (γ) represents Euler's constant so that it fell between 0 and 1 with higher scores suggesting a higher likelihood of involvement in disease.
97 genes had a score over 0.75, of which 36 (~ 33%) are listed in either the HGMD or OMIM and are thus already known to be involved in disease (this represents a more than threefold enrichment; Ensembl contained ~ 19,500 known genes of which ~ 9% were known disease genes). By contrast, in the set of 4,357 genes which scored less than 0.3 only ~ 0.8% (35 genes) are already known to be involved in disease.
A list of the 61 genes that scored higher than 0.75 but are not already known to be involved in disease is included as supplementary material (see Additional File 1). By searching for references in PubMed we discovered that 9 of these genes (~ 15% of the total) are already candidates for involvement in diseases including Alzheimers (ABCA2), osteoporosis (COL4A1 and COL4A2) and schizophrenia (SLIT3).
Discussion
Relative performance
PROSPECTR has a number of advantages over existing classification schemes designed to differentiate between genes more and less likely to be involved in disease.
PROSPECTR appears to perform significantly better on unseen data than the decision tree classifier presented by Lopez-Bigas et al. The Lopez-Bigas classifier is less likely to be useful as a predictive tool for two, related reasons: firstly, it achieves perfect accuracy when tested on the training set, even though the training set is known to be inconsistent. The Lopez-Bigas classifier suggests that ~ 31% of the genome is made up of predicted disease genes. Thus if genes were picked at random to make up the control set during training it should be assumed that ~ 31% of them are actually disease genes which have not yet been characterized. Perfect accuracy on the training set is therefore undesirable – by ignoring the possibility that the set of control genes might contain disease genes the classifier loses flexibility and predictive power. Secondly, the fact that all uncharacterized disease genes were predicted as being control despite their presumed strong similarity (in terms of sequence features) to other disease genes suggests that it is highly likely that at least some degree of overfitting occurred, which would further impair performance on novel data.
PROSPECTR's use of a spread of sequence-based features representing the structure, content and phylogenetic extent of candidate genes allows investigators to see exactly which features are contributing the most towards a particular classification. In addition, it requires no detailed phenotypic knowledge of the disease in question and can score whole chromosomes in minutes. The use of sequence-based features avoids the bias inherent in current functional annotation, where better studied genes are far more likely to have better and more extensive annotation. Furthermore, by relying less on phylogenetic conservation we reduce the amount of potential bias from imperfect homology prediction (see Eliminating Bias in Methods).
Classifier mechanics
Other researchers have examined some of the sequence-based differences between genes listed in OMIM and genes not known to be involved in disease but there have been no comprehensive studies. The data we present here collates all of the known sequence-based differences and introduces some new ones – for example, the differences in 3' UTR length between the two sets of genes are statistically significant and, though a correlation with gene size exists, it is relatively weak (a Spearman correlation coefficient of 0.35). Further research is needed to suggest how all of these differences might relate to, for example, a gene's relative importance or position in a protein-protein interaction map or biological pathway.
The length of the 3' UTR is thought to be related to translational efficiency and mRNA stability [24], which in turn affects the level of expression of the gene. Two other novel sequence-based features where we found significant differences between disease and non-disease genes – the distance to the nearest neighbouring gene and the number of exons – might also be directly related to expression levels [25]. In this work we were able to confirm the suggestions from previous studies that there is a significant difference in tissue specificity between disease and non-disease genes [17] – perhaps similar differences exist between the two sets in patterns of overall gene expression levels.
It seems remarkable that disease genes would share sequence features to such an extent. In particular gene length and protein length when taken together as features for an alternating decision tree classifier with fifteen nodes can be reasonably predictive (69% of disease genes correctly classified from the training set with 51% misclassification, details not shown).
A complex web of correlations exists between gene and protein size, levels of expression and rates of evolution, which perhaps explains why predictive power remains relatively high after removing features other than gene and protein size. Additionally, larger genes might simply be bigger targets for mutation [13], or be more likely to have sequence features like overlapping gene groups, multiple amino acid runs [26] and motifs associated with mutational hotspots which might increase the chance of them succumbing to some disease causing mutation.
An alternative hypothesis is that PROSPECTR does not predict genes likely to be involved in disease at all, but the opposite: it derives its predictive power from discounting those genes which are unlikely to be involved in disease as mutations usually result in a phenotype which is either lethal (in which case we wouldn't class it as a disease gene), undetectable (in which case we couldn't class it as a disease gene) or very weak (in which case the classification of the gene would be debateable).
We have shown that PROSPECTR performs well on an oligogenic test set. However, one might expect the biological mechanisms of cause and effect to differ between simple Mendelian and more complex traits and therefore the classifiers dealing with either type may also have to differ. Currently no sizeable dataset of genes involved in complex disease exists; until one is created and examined we cannot tell how PROSPECTR will perform when used to find the genes underlying complex disease. We would thus advise caution when using PROSPECTR to search for genes involved in complex traits.
Future directions
PROSPECTR can create lists of genes the tops of which are enriched for those genes that are likely to be involved in human disease. Substantial enrichment is highly likely with this sequence-based approach, although investigators still need to carry out functional comparisons and fine scale mapping to reduce lists to one or two candidates for each region of interest. By contrast, functional classifiers might present only a handful of high quality suggestions for each of the regions studied but equally might not return the target gene at all as their threshold for successful detection is too high.
One way of speeding the candidate gene discovery process further without sacrificing accuracy might be to combine existing techniques that use functional annotation with a sequence-based approach similar to the one we describe here. It may be possible to create a combined classifier greater than the sum of its parts by lowering the threshold of a successful functional annotation based classifier and then dismissing false positives using a sequence-based approach.
The alternating decision tree used in PROSPECTR was trained using all genes from OMIM and as such is suited for general use. However, there might well be some value in creating custom classifiers targeted to a particular area of interest; for example, genes involved in neurological disorders. If the training set was still sufficiently large enough to be representative then one might expect more precision when scoring candidate genes in similar disorders. As a first step towards this we have made instructions for creating custom classifiers available on the PROSPECTR website.
Conclusion
On average, PROSPECTR successfully enriches lists of candidate genes 2-fold ~ 77% of the time, 5-fold ~ 37% of the time and 25-fold ~ 11% of the time. It does so for both monogenic and oligogenic disorders and on the basis of a compact set of rules which look at sequence-based features. These features reflect the structure and content of the genes in question as well as the phylogenetic extent and are much less likely to be biased towards better studied genes than manual annotation.
The rules involved are easily interpretable which gives some insight into how the classifier works and the importance of various features relative to each other, signposting new avenues of investigation into the differences between the types of disease and non-disease genes.
We predict that the growing availability of relevant protein-protein interaction data and better functional annotation will greatly improve candidate identification techniques for oligogenic and complex disorders, as shared or compensated pathways become clearer. However, robust genome-wide functional annotation is still some way off. In the meantime, using PROSPECTR as a quick, unbiased method to rank genes in order of their likelihood of involvement in disease could save investigators much time and effort when examining larger regions of interest, prioritizing candidates for more in-depth functional characterization, mutation detection and case control studies.
Our implementation of PROSPECTR is readily available on the web at .
Methods
We used Online Mendelian Inheritance in Man (OMIM) and the Human Gene Mutation Database (HGMD) to obtain lists of disease genes and MySQL client access to Ensembl to retrieve the sequences for those genes. We also used Ensembl to provide genes as a control set by selecting reasonably sized representative sets at random. The genes in the control set were not listed in either OMIM or the HGMD.
To create our initial feature set we collated information from Ensembl, NCBI's Homologene, Interpro [10], SwissPROT and the Novartis Gene Expression Atlas [27]. This initial feature set included all of the features listed in Table 1 as well as information relating to tissue expression and protein domain distribution. The relevant information for all known genes in Ensembl was stored in a local MySQL database. We then compared features from a set of 1,084 genes listed in OMIM with a representative sample from the control set made up of ~ 18,000 genes from Ensembl not listed in OMIM. Features that were considered to have a reasonable degree of predictive power were selected to create the feature set to be made available to the alternating decision tree algorithm. To calculate the tissue specificity of each gene we used the same method as Winter et. al [17].
We used Weka as the platform for building our classifier. Weka is free, open source Java application and is readily available on the internet [19]. Experiments were carried out using the Explorer interface to Weka using the ADTree classifier. The accuracy, precision, recall, AUC (the area under the ROC curve; used as a performance metric) and Kappa statistics in Table 3 were obtained directly from Weka. We used custom Perl scripts to create artificial loci and then rank the scores of the genes they contained. We wrote the PROSPECTR software using Perl and the Apache web server.
Eliminating bias and sources of error
We studied the feature set for potential bias. In particular, the number of Interpro domains described on each gene appeared to have more to do with a bias towards better studied genes than with disease gene association. When we compared the number of Interpro domains between the OMIM genes and a group of genes not known to be involved in disease but with at least one reference in literature (according to their SwissPROT record) no significant differences were found. We therefore eliminated any features based on Interpro domains as potentially biased.
Though highly significant differences in tissue expression patterns were detected, it was decided to exclude the tissue specificity feature from the training set as it introduced a bias towards disease genes; reliable, normalised tissue expression data was available for ~ 95% of genes implicated in disease but only two thirds of control genes. Genes without the relevant data could have been ignored or had a best guess value assigned to them, but this would have undermined the classifier's usefulness for detecting novel disease gene candidates and introduced new sources of bias.
Determining phylogenetic extent by looking at homologs is also a potential source of bias as disease genes are better characterized and more transcript evidence is available. Imperfection in gene prediction is a major hindrance to accurate orthology prediction [16]. On average, around a third (~ 34%) of the predictive power of our classifier comes from features related to phylogenetic extent.
There exists a possibility that the set of OMIM genes that make up the training set is itself biased towards genes containing features which somehow make linking a disease to an allele of that gene easier. However, we believe this to be unlikely.
Firstly, it is important to remember that, at least with Mendelian disorders, it has been the goal of identifying the gene behind a particular common disease that has driven research, not matching diseases to genes that are easier to find, clone or characterize. Secondly, although the OMIM database has been collecting information about Mendelian disorders for many years the majority of confirmed disease genes have been added more recently after having been mapped and characterized with the help of publicly available sequence data and modern molecular biology techniques – none of which are obviously biased towards particular sequence features. Finally, given the combined size (~ 1,700 genes) of the training and test datasets it seems reasonable to assume that we are working with a representative sample of disease genes.
Authors' contributions
BP, RA and EA conceived of the study, which was coordinated by BP. EA carried out the work with Perl and Weka with help from RA who also participated in testing. KE and DP participated in the analysis of the results and stimulated discussion. All authors helped to draft the manuscript.
Supplementary Material
Additional File 1
The 61 top scoring genes not known to be implicated in disease.
Click here for file
Acknowledgements
The authors wish to thank Colin Semple (MRC Human Genetics Unit, Edinburgh) for discussion and his comments on the manuscript.
Figures and Tables
Figure 1 Histograms of selected features. Histograms showing distributions of selected features in both "disease genes" (those listed in OMIM) and control genes (those not). Data was binned for graphing purposes. Distributions are shown for (A) gene length in kilobases; (B) protein length in amino acids; (C) % identity of the best reciprocal hit (BRH) homolog in mouse; (D) Ka (a measure of non-synonymous change between species) of the BRH homolog in mouse; (E) number of exons and (F) 3' UTR length in basepairs.
Figure 2 The alternating decision tree. The alternating decision tree used to classify instances. A gene is classified with the tree by beginning at the node marked "Start" and then following each branch in turn. Upon reaching a node which contains an assumption the "yes" or "no" branch is followed as appropriate. If the relevant feature is "unknown", neither branch is followed. Adding up each of the numbers in rectangles that are encountered along the way results in a final score which reflects the relative confidence of the classification. The classification itself is based on the sign of the score.
Figure 3 Receiver Operating Characteristic (ROC) curves. Receiver Operating Characteristic (ROC) curves for the training set (A) and the two test sets (B and C). The true positive rate is measured along the y-axis and the false positive along the x-axis. The area under the resulting curve is a measure of classifier performance.
Figure 4 Performance over artificial loci. Relative performance on the sets of artificial loci created from the training set (yellow line), HGMD test set (the blue line) and oligogenic test set (the green line). The gray line represents the value expected if there had been no enrichment. The x axis represents the % of the ranked list in which the target gene was found; the y axis represents how frequent that occurrence was. For example, in the training set (the yellow line) the target gene was in the top 30% of the ranked list around 56% of the time.
Table 1 The feature set. The list of features which were made available to the machine learning application (Weka) to build the alternating decision tree.
Feature Source Description
Gene length EnsemblMart 22.1 Length of gene in bp.
CDS length EnsemblMart 22.1 Length of coding sequence in bp.
cDNA length EnsemblMart 22.1 Length of complementary DNA in bp.
Protein length EnsemblMart 22.1 Length of protein in aa.
Length of 3' UTR EnsemblMart 22.1 The length of the 3' untranslated region (UTR) in bp
Length of 5' UTR EnsemblMart 22.1 The length of the 5' untranslated region (UTR) in bp
Distance to nearest neighbouring gene EnsemblMart 22.1 Distance to the next known gene on the same chromosome on either strand in bp.
Number of exons EnsemblMart 22.1 Number of exons in the gene.
GC EnsemblMart 22.1 GC content (as a %) of gene
Transmembrane EnsemblMart 22.1 Prediction of transmembrane domains (1 for yes or 0 for no)
Signal peptide EnsemblMart 22.1 Prediction of signal peptide (1 for yes or 0 for no)
Paralog EnsemblMart 22.1 If the gene has a paralog in the human genome (1 for yes or 0 for no)
Paralog % identity EnsemblMart 22.1 % protein identity of best paralog in the human genome. Genes without paralogs have "unknown" entered here.
Mouse homolog % identity Homologene % protein identity of mouse homolog. Genes without a mouse homolog have "0" entered here.
Rat homolog % identity Homologene % protein identity of rat homolog. Genes without a rat homolog have "0" entered here.
Worm homolog % identity Homologene % protein identity of worm homolog (potentially 0, see above)
Fly homolog % identity Homologene % protein identity of fly homolog (potentially 0, see above)
Yeast homolog % identity Homologene % protein identity of yeast homolog (potentially 0, see above)
Arabidopsis homolog % identity Homologene % protein identity of Arabidopsis homolog (potentially 0, see above)
Mouse homolog Ka Homologene Measure of non-synonymous changes between human and mouse homolog.
Mouse homolog Ks Homologene Measure of synonymous changes between human and mouse homolog.
Mouse homolog Ka / Ks Homologene Ratio of above two fields.
CpG island at 3' end of gene EnsemblMart 22.1 If a CpG island exists at the 3' end of the gene (1 or 0)
CpG island at 5' end of gene EnsemblMart 22.1 If a CpG island exists at the 5' end of the gene (1 or 0)
Table 2 Significant differences between the control set and disease set of genes. The features found to be significantly different between Ensembl genes found in OMIM and those not in OMIM. Significance was calculated using the Mann-Whitney U test unless otherwise noted.
Feature Median in control set Median in disease set Significance
Gene length 19 k 27 k P < 0.001
cDNA length 2,126 bp 2,442 bp P < 0.001
Protein length 383 aa 494 aa P < 0.001
3' UTR length 446 bp 488 bp P < 0.01
Exon number 8 10 P < 0.001
Distance to neighbouring gene 46 kb 52 kb P < 0.01
Protein identity with BRH in mouse 80% 87% P < 0.001
Gene encodes signal peptide 17% 35% P < 0.0001 (calculated using the chi squared test)
5' CpG islands 12% 16% P < 0.028 (calculated using the chi squared test)
Table 3 More detailed classifier performance statistics. For each set of genes tested, five statistics that reflected performance were calculated. Accuracy is the overall accuracy of the classifier; precision reflects the classifier's specificity and recall reflects classifier sensitivity. The area under curve (AUC) is the area underneath the ROC curve drawn for each set of genes (see Figure 3) and represents classifier performance across all combinations of sensitivity and specificity. It ranges from 0 to 1, where 1 represents 100% accuracy, 0.5 represents performance no better than random and 0 represents 0% accuracy. The Kappa statistic is a measurement of agreement between predicted and actual classifications and takes false positive rates into account. It is a number between 1 (symbolising perfect agreement between predicted and actual classifications) and 0 (symbolising no agreement).
Test Set Nodes in tree Accuracy Precision Recall AUC Kappa
Training (OMIM) set 15 67% 65% 77% 0.75 0.35
10 × cross validation 15 63% 62% 70% 0.70 0.27
HGMD set 15 64.5% 63% 71% 0.69 0.29
Oligogenic set 15 65% 63% 72% 0.76 0.31
Table 4 Relative contribution of each feature to classification as disease gene. An estimate of the relative contribution of each sequence feature in the final score used by the alternating decision tree for classifying genes as being involved in disease. The percentages are based on the average absolute contribution to the cumulative absolute score of each disease gene in the training set.
Feature % Contribution to final score
Signal peptide 23%
Mouse homolog % identity 21%
Length of 3' UTR 12%
Number of exons 7%
Rat homolog % identity 7%
Worm homolog % identity 6%
GC 6%
CDS length 5%
Gene length 4%
Mouse homolog Ka 3%
Paralog % identity 2%
==== Refs
Glazier AM Nadeau JH Aitman TJ Finding Genes That Underlie Complex Traits Science 2002 298 2345 2349 12493905 10.1126/science.1076641
McCarthy M Smedley D Hide W New methods for finding disease-susceptibility genes: impact and potential Genome Biology 2003 4 119 14519189 10.1186/gb-2003-4-10-119
Devos D Valencia A Intrinsic errors in genome annotation Trends in Genetics 2001 17 429 431 11485799 10.1016/S0168-9525(01)02348-4
Gilks WR Audit B De Angelis D Tsoka S Ouzounis CA Modeling the percolation of annotation errors in a database of protein sequences Bioinformatics 2002 18 1641 1649 12490449 10.1093/bioinformatics/18.12.1641
Pallen M Wren B Parkhill J 'Going wrong with confidence': misleading sequence analyses of CiaB and ClpX Molecular Microbiology 1999 34 195 10540297 10.1046/j.1365-2958.1999.01561.x
Van Driel MA Brunner HG Leunissen JAM Kemmeren PPCW Cuelenaere K A new web-based data mining tool for the identification of candidate genes for human genetic disorders European Journal of Human Genetics 2003 11 57 63 12529706 10.1038/sj.ejhg.5200918
Freudenberg J Propping P A similarity-based method for genome-wide prediction of disease-relevant human genes Bioinformatics 2002 18 110S 1115
Perez-Iratxeta C Bork P Andrade MA Association of genes to genetically inherited diseases using data mining Nature Genetics 2002 31 316 319 12006977
Turner FS Clutterbuck DR Semple CAM POCUS: mining genomic sequence annotation to predict disease genes Genome Biology 2003 4
Mulder NJ Apweiler R Attwood TK Bairoch A Barrell D Bateman A Binns D Biswas M Bradley P Bork P The InterPro Database, 2003 brings increased coverage and new features Nucl Acids Res 2003 31 315 318 12520011 10.1093/nar/gkg046
Smith NGC Eyre-Walker A Human disease genes: patterns and predictions Gene 2003 318 169 175 14585509 10.1016/S0378-1119(03)00772-8
Kapetanovic IM Rosenfeld S Izmirilan G Overview of Commonly Used Bioinformatics Methods and Their Applications Ann NY Acad Sci 2004 1020 10 21 15208179 10.1196/annals.1310.003
Lopez-Bigas N Ouzounis CA Genome-wide identification of genes likely to be involved in human genetic disease Nucl Acids Res 2004 32 3108 3114 15181176 10.1093/nar/gkh605
Hammond MP Birney E Genome information resources – developments at Ensembl Trends in Genetics 2004 20 268 272 15145580 10.1016/j.tig.2004.04.002
Hamosh A Scott AF Amberger J Bocchini C Valle D McKusick VA Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders Nucl Acids Res 2002 30 52 55 11752252 10.1093/nar/30.1.52
Huang H Winter E Wang H Weinstock K Xing H Goodstadt L Stenson P Cooper D Smith D Alba MM Evolutionary conservation and selection of human disease gene orthologs in the rat and mouse genomes Genome Biology 2004 5 R47 15239832 10.1186/gb-2004-5-7-r47
Winter EE Goodstadt L Ponting CP Elevated Rates of Protein Secretion, Evolution, and Disease Among Tissue-Specific Genes Genome Res 2004 14 54 61 14707169 10.1101/gr.1924004
Gardiner-Garden M Frommer M CpG islands in vertebrate genomes Journal of Molecular Biology 1987 196 261 282 3656447 10.1016/0022-2836(87)90689-9
Frank E Hall M Trigg L Holmes G Witten IH Data mining in bioinformatics using Weka Bioinformatics 2004 261
Freund Y Mason L The Alternating Decision Tree Learning Algorithm Proceedings of the Sixteenth International Conference on Machine Learning 124 133
Stenson PD Ball EV Mort M Philips AD Shiel JA Thomas NST Abeysinghe S Krawczak M Cooper DN Human Gene Mutation Database (HGMD®): 2003 update Human Mutation 2004 21 577 581 10.1002/humu.10212
Becker KG Barnes KC Bright TJ Wang SA The Genetic Association Database Nature Genetics 2004 36 431 432 15118671 10.1038/ng0504-431
Forbes AD Classification algorithm evaluation: five performance measures based on confusion matrices Journal of Clinical Monitoring 1995 11 189 206 7623060
Tanguay RL Gallie DR Translational efficiency is regulated by the length of the 3' untranslated region Molecular Cellular Biology 1996 16 146 156
Chiaromonte F Miller W Eric E Gene Length and Proximity to Neighbors Affect Genome-Wide Expression Levels Genome Res 2003 13 2602 2608 14613975 10.1101/gr.1169203
Karlin S Chen C Gentles AJ Cleary M Associations between human disease genes and overlapping gene groups and multiple amino acid runs PNAS 2002 99 17008 17013 12473749 10.1073/pnas.262658799
Su AI Cooke MP Ching KA Hakak Y Walker JR Wiltshire T Orth AP Vega RG Sapinoso LM Moqrich A Large-scale analysis of the human and mouse transcriptomes PNAS 2002 99 4465 4470 11904358 10.1073/pnas.012025199
|
15766383
|
PMC1274252
|
CC BY
|
2021-01-04 16:02:50
|
no
|
BMC Bioinformatics. 2005 Mar 14; 6:55
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-55
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-561576929010.1186/1471-2105-6-56SoftwareTMB-Hunt: An amino acid composition based method to screen proteomes for beta-barrel transmembrane proteins Garrow Andrew G [email protected] Alison [email protected] David R [email protected] School of Biochemistry and Microbiology, University of Leeds, Leeds, LS2 9JT, UK2005 15 3 2005 6 56 56 1 11 2004 15 3 2005 Copyright © 2005 Garrow et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Beta-barrel transmembrane (bbtm) proteins are a functionally important and diverse group of proteins expressed in the outer membranes of bacteria (both gram negative and acid fast gram positive), mitochondria and chloroplasts. Despite recent publications describing reasonable levels of accuracy for discriminating between bbtm proteins and other proteins, screening of entire genomes remains troublesome as these molecules only constitute a small fraction of the sequences screened. Therefore, novel methods are still required capable of detecting new families of bbtm protein in diverse genomes.
Results
We present TMB-Hunt, a program that uses a k-Nearest Neighbour (k-NN) algorithm to discriminate between bbtm and non-bbtm proteins on the basis of their amino acid composition. By including differentially weighted amino acids, evolutionary information and by calibrating the scoring, an accuracy of 92.5% was achieved, with 91% sensitivity and 93.8% positive predictive value (PPV), using a rigorous cross-validation procedure.
A major advantage of this approach is that because it does not rely on beta-strand detection, it does not require resolved structures and thus larger, more representative, training sets could be used. It is therefore believed that this approach will be invaluable in complementing other, physicochemical and homology based methods. This was demonstrated by the correct reassignment of a number of proteins which other predictors failed to classify. We have used the algorithm to screen several genomes and have discussed our findings.
Conclusion
TMB-Hunt achieves a prediction accuracy level better than other approaches published to date. Results were significantly enhanced by use of evolutionary information and a system for calibrating k-NN scoring. Because the program uses a distinct approach to that of other discriminators and thus suffers different liabilities, we believe it will make a significant contribution to the development of a consensus approach for bbtm protein detection.
==== Body
Background
Beta-barrel transmembrane proteins
The beta-barrel is one of only two membrane spanning structural motifs currently identified [1]. It is proven with high resolution structures for many proteins expressed within the outer membranes of gram negative bacteria and is also widely expected for several proteins expressed in the outer membranes of mitochondria [2] and chloroplasts [3]. In addition, the structure of a protein found spanning the outer membrane of Mycobacteria (an acid fast gram positive bacterium) was recently resolved revealing two consecutive membrane spanning beta-barrels [4]. As with alpha-helical transmembrane (ahtm) proteins, beta-barrel transmembrane (bbtm) proteins play both functionally important and diverse roles [1].
Currently, over 92 bbtm protein structures are present in the protein databank [5], including 23 families as defined in PDB_TM [6]. They are classified in the SCOP hierarchy, in 3 different folds [7], the transmembrane beta-barrels (described as not a true fold, but a gathering of beta-barrel membrane proteins), the integral outer membrane protein TolC fold and the Leukocidin (pore forming toxins) fold. The transmembrane beta-barrels consist of four SCOP superfamilies; OmpA-like, OmpT-like, OmpLA and the Porins; and include channels, enzymes and receptors. These superfamilies vary in numbers of subunits, where each subunit contributes a single barrel. The TolC fold, consists of one SCOP superfamily and includes proteins involved in secretion and expression of outer membrane proteins (OMPs) [8]. These proteins are trimeric with each subunit contributing four strands to a single barrel, and contain large stretches of alpha-helix, which stretch across the periplasm. Finally, the Leukocidin fold consists of heptameric pore forming toxins with each subunit contributing 2 strands to the barrel. TolC, Leukocidin and the Mycobacterial porin MspA (which is not yet classified within SCOP) can thus be considered "non-typical" bbtm proteins. From the diversity of bbtm proteins in different SCOP folds, it seems likely that these proteins have multiple evolutionary origins.
These structures have helped reveal a number of features concerning transmembrane (TM) beta-strands and their organisation [9]. TM beta-strands show an inside-outside dyad repeat motif of alternating residues facing the lipid bilayer and the inside of the barrel. Outside (lipid bilayer facing) residues are typically hydrophobic whilst inside (facing inside of barrel) residues are of intermediate polarity. TM beta-strands are often flanked by a layer of aromatic residues, believed to be involved in maintaining the protein's stability within the membrane [10]. Structures have also revealed an even number of strands, with N and C termini on the same side of the membrane. Strands form an antiparallel beta-meander topology with alternating long and short loops. The number of TM beta-strands in a barrel has been shown to range from 8–22 strands, with a range of 6–22 (most frequently 12) residues.
In contrast to ahtm proteins, which are easy to identify through TM alpha-helices composed of 20 or more hydrophobic residues [11], the short and cryptic nature of TM beta-strands makes their discrimination difficult. Prediction is complicated further with beta-strands of some globular proteins superficially resembling those of bbtm proteins.
BBTM protein discriminators
Despite these difficulties, numerous methods have recently been published for the identification of these proteins, most commonly focusing on identification of TM beta-strands. Methods include rule based approaches [12], an architecture based approach [13], Hidden Markov Models (HMMs) [14-18], a neural network based method [19], a combined neural network and support vector machine [20], composition of transmembrane beta strands combined with secondary structure prediction [21] and an approach based on architecture [13] combined with isoleucine and asparagine abundance [22]. Of these, the first two give no indication of discriminatory accuracy, but the others range from 80 to 90%.
Whilst this level of accuracy may seem acceptable if analysing a particular sequence of interest, problems will occur when screening an entire genome for potential bbtm proteins, owing to the fact that a large number of sequences are being tested of which these molecules only constitute a small fraction. There is therefore a need for programs with higher accuracy and in particular higher specificity, in order to minimise the false discovery rate.
Amino acid composition based protein classification
This paper describes TMB-Hunt, an amino acid composition based program for the identification of bbtm proteins. Amino acid composition has been analysed for bbtm proteins [13], however whole sequence composition has not previously been used for discrimination. Many previous studies have shown how amino acid composition can be successfully applied to protein sequence analysis, including prediction of structural class [23], discrimination of intra- and extra cellular proteins [24] and distinguishing between membrane protein type [25]. Amino acid composition is often used for prediction of subcellular location, as an alternative to signal detection based methods [26-29] which are prone to errors in automated gene prediction at the 5' end [30]. The limitation of this technique, however, is that the correlation of cell location with amino acid composition is not absolute. It was suggested that composition differences are a consequence of different requirements for protein folding, stability and transportation [24,26]. Subsequently it has been shown that amino acid composition differences correlate most strongly with surface residues [27]. Thus, composition has been particularly useful in discriminating between ntm and ahtm proteins, which consist of large numbers of hydrophobic amino acids in contact with the lipid bilayer. This feature has enabled algorithms to be developed capable of distinguishing between the two classes with >97% accuracy [31], based on identification of the TM alpha-helices.
Because TMB-Hunt puts no emphasis on identification of TM beta-strands, we were not dependent on sequences with resolved structures and training sets could be much larger than those used for other predictors [12-22]. As a result, bbtm proteins with structures more diverse than those used by other predictors were included, resulting in a greater degree of sensitivity. TMB-Hunt is at least as accurate as other predictors, but its major advantage is that it adopts a completely different approach to other methods and is likely therefore to be valuable in consensus approaches, which should be much more successful at hunting for new families of candidate bbtm proteins in diverse proteomes.
Implementation
Training sets
Training sets for bbtm, ahtm and non-TM (ntm) proteins were gathered from a number of manually curated and published sources. The PDB accessions of 3159 ntm proteins were acquired from PDB-REPRDB via the Papia database [32], and respective sequences were extracted.
Sequences of ahtm proteins were downloaded from a test set available at the Sanger centre [33]. Four datasets were available of varying quality. Dataset A comprised 37 sequences where structural information was available. Dataset B contained 23 sequences with very good biochemical characterisation from at least two complementary methods. Dataset C contained 129 sequences with some biochemical characterisation and where annotation was only reliable for part of the sequence. Dataset D contained sequences with no biochemical characterisation and only hydrophobicity or an alignment as a basis for their characterisation. Datasets A, B and C were used.
Beta-barrel transmembrane protein sequences were downloaded from a number of resources including:
957 from UniProt [34] using a keyword search for 'Transmembrane' and 'Outer Membrane' and taxonomy filter for only bacteria
134 from the transporter classification (TC) database [35]
35 extracted from the PDB files of beta-barrel outer membrane proteins in SCOP [7].
All these datasets were manually created and rechecked to ensure no obvious spurious sequences were present. Sequences of less than 120 residues were removed from the training set. Sequences were next grouped into clusters using BLASTclust and a sequence similarity threshold of 23%. Amino acid composition profiles were produced for each group using evolutionary information, as described below. Dataset details are summarised in Table 1.
The final dataset included numerous types of bbtm protein not included in the training sets of other predictors. Inclusion of such a diverse range of proteins was important as it covers a wide range of evolutionary origins and physicochemical adaptations. TolC, Alpha-hemolysin and the Mycobacterial Porin Family are bbtm proteins with resolved structures, not used by other predictors, either because of their unusual structure or because their structure was resolved after the predictor had been completed. Fimbrial, pili and flagellar associated proteins were also included, as were non-bacterial proteins e.g. the mitochondrial porin (VDAC), plastid bbtm proteins (e.g. OEP24) and chloroplast porins (Toc75).
Sequences used for proteome screening were downloaded from the NCBI FTP site [36]. Sequences used for annotation comparison were downloaded via SRS [37,38] from Uniprot [34].
k-nearest neighbour algorithm
The k-nearest neighbour algorithm is a simple instance-based learning method for performing general, non-parametric classification [39,40]. Each object or instance (a protein in this case) is associated with a class which can be unknown (class 0), bbtm (1), ahtm (2) or ntm (3). For query proteins of unknown class, predictions are made by using information from a training set of proteins where the class is known. The prediction is made on the basis of a set of k objects from the training set which are most similar (in the sense described below) to the query protein. This technique is thus a local approximation, focusing on the neighbourhood of the query instance. A major advantage of this algorithm is that it is robust to noisy data (given a large dataset), as taking the weighted average of the nearest neighbours smoothes out isolated training instances.
Proteins are represented by x = (fa (x), a ∈ A; c(x)), where c(x) represents the class c ∈ {0,1,2,3} as defined above, A is the set of naturally occurring amino acids and fa(x) denotes the relative frequency of the amino acid a. The distance between two proteins xi and xj in this representation is measured by the standard Euclidean metric.
Given a query protein xq, the algorithm first finds the k closest instances in the training set according to this metric, and then assigns a score S(xq, c) for each possible class c,
where δ(c1, c2) = 1 if the classes c1 and c2 are equal and zero otherwise. Thus the score for each class is a sum of positive contributions from each of the nearest neighbours from that class, where the contribution is weighted according to the reciprocal square distance between query instance and neighbour with closer neighbours contributing more strongly.
Since we are very often concerned with binary classification problems (e.g. distinguishing bbtm proteins from proteins in any other class), it is also useful to define a discrimination score,
which is the score from one class (e.g. bbtm proteins) minus the scores from other classes.
Calibration and scoring
In making predictions a standard nearest neighbour algorithm would simply predict the class of xq to be the class c with the highest score S(xq, c). However, this procedure is problematical in cases such as this where the training set is unbalanced, containing many more ntm proteins than either of the other two classes. Statistical chance means that the k-nearest neighbour sets tend to contain more proteins from the dominant class, leading to this class as the dominant prediction even in the presence of substantial evidence for membership of one the other classes in the nearest neighbour set. One approach to this problem would be to reduce representation of the dominant class to produce a balanced training set, but this procedure involves wasting useful information. It would also be possible to down-weight information from the dominant class, but we found that a more effective approach was to use the distributions of D(x,c) scores in the training set proteins, divided between proteins in class c, and proteins in other classes from which they are to be distinguished. For clarity, in the remainder of this section we will consider c = 1, where the classification problem is to distinguish bbtm proteins from any others, and D will denote the discrimination score D(x,c = 1) for an arbitrary protein x.
Empirical cumulative probability distributions for D in the case above are shown in Figure 1. As expected, plots showed a higher mean discrimination score for bbtm (mean = 0.078, standard deviation = 0.115) than other proteins (mean = -0.206, standard deviation = 0.171). These distributions do not deviate significantly from the normal distribution. Using these distributions it is possible to convert discrimination scores into a convenient log likelihood ratio (beta-barrel score),
R(D) = log(p(bbtm|D)/p(other|D)),
where p(bbtm|D) denotes the probability of a bbtm protein obtaining a score of at least D, and p(other|D) denotes the probability of a protein from the other class obtaining a score of D or greater. Negative values of R indicate a query protein more likely to come from the other class, and positive values indicate a protein more likely to come from the bbtm class.
An alternative probabilistic interpretation of the D score is the expected number of proteins from the other class scoring D or greater, E(D) = Np(other|D), where N indicates the number of query sequences tested. This measure takes account of the multiple testing involved in screening large numbers of sequences in a genome, and is related to the standard Bonferroni correction. It is directly analogous to the E-values reported by the popular sequence search programs FASTA [41] and BLAST [42].
Differential dimension weightings
To account for some dimensions contributing information more valuable to classification than others, weights were applied to each of the dimensions used in calculating Euclidean distances. The modified Euclidean distance calculation was:
where ga is the weight applied to amino acid a.
A genetic algorithm was employed to calculate the optimal weightings for each dimension. Genetic algorithms are an optimisation approach, based on Darwinian principles, which assume that given a population of individuals, environmental pressures cause natural selection thus increasing the overall fitness of the population [43]. Application of a genetic algorithm requires a population of solutions, termed chromosomes, whose fitness can be measured using an objective function. Based on fitness, the better candidates are chosen to seed the next generation through a combination of crossover and/or mutation. This will result in the evolution of successively better solutions. The process is carried out until an optimal solution or time limit is reached.
The algorithm initiates by constructing a random population of chromosomes (i.e. potential solutions), represented as vectors, with each element of the vector termed a gene, representing a weight for a particular dimension of the Euclidean space. Fitness for chromosomes was measured by the Matthews Correlation Coefficient (MCC) value returned from a 'leave homologues out' cross-validation analysis (see below) using a fixed set of 100 bbtm proteins and 100 ntm proteins. Once fitness for each of the chromosomes within a generation was determined, the fittest were used to create offspring through a process of crossover and mutation. Crossovers involve the construction of a new vector, using random genes taken from two or more parents. Mutations involved randomly mutating 1 in 8 genes.
Inclusion of evolutionary information
Random noise in amino acid composition was reduced by inclusion of evolutionary information. Evolutionary information was included by building a feature vector using both the query sequence, as well as a number of close homologues (as determined by a BLAST query against Uniprot/SwissProt with an E-value threshold of 0.0001, and a maximum of 25 homologues) to calculate an average amino acid composition vector for the sequence and its close evolutionary relatives. A weighted average composition was used, with more distant homologues contributing more to the average (since the more distant sequences contain more new information). Weights were assigned by first carrying out all-against-all alignments within the set using BLAST, then weighting sequences according to their average distance to other sequences. The weights were calculated as
where Wk denotes the weight applied to sequence k, and pk the average percentage difference (100 minus the percentage identity) from sequence k to other sequences.
Performance
Cross-validation studies were used to assess performance. Two approaches were used, 'leave-one out' cross-validations and 'leave-homologues out' cross-validations. The first of these methods involved removing in turn profiles from the training set and seeing if the algorithm could correctly reassign one of the sequences used to build the profile. Removal of profiles and their construction using sequences in clusters of >23% identity meant that sequences should not then be correctly reassigned due to 'self-detection' by a close homolog. However, even sequences of <23% identity can be homologues and show significant similarity e.g. over shorter fragments of the sequence, therefore a 'leave homologues out' cross-validation was used as a stricter alternative. This meant pre-computing sequences similar (with a BLAST E-value threshold <1) to each query sequence, and leaving these out of the training set when testing. This procedure eliminates any homolog whose sequence is sufficiently similar to be detected with BLAST.
Performance was measured using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy and MCC, which are defined in terms of true positives (TP), false positives (FP), true negatives (TN) and false negatives (FN).
Sensitivity is a measure of the percentage of bbtm proteins correctly classified and is calculated with, 100*TP/(TP+FN). Specificity is the percentage of non-bbtm correctly classified as is calculated as 100*TN/(TN+FP). The PPV is the percentage of predicted bbtm proteins that are correct and is calculated by, 100*TP/(TP+FP). The NPV is the percentage of predicted non-bbtm proteins that are correct and is calculated using 100*TN/(TN+FN). Accuracy is a measure of the total number of correctly assigned proteins and is measured by, 100*(TP+TN)/t, where t is the total number of sequences queried. However this statistic can be misleading in circumstances with bias in the test set composition. Therefore, the Matthews Coefficient Correlation (MCC) is an alternative measure that accounts for both under and over predictions.
This returns a value between -1 and 1, with 1 meaning everything is correctly assigned and -1 meaning everything is incorrectly assigned. Given two prediction classes (e.g. bbtm and ntm) and a random probability of assigning queries to either, a score of 0 would be expected by random classification.
Results
TMB-Hunt uses a k-Nearest Neighbour (k-NN) algorithm to classify query instances, using the class (bbtm, ahtm or ntm) of their nearest neighbours, as defined by differences in amino acid composition. A number of steps were involved in optimisation, including selection of the numbers of neighbours used (k), amino acid weightings and scoring statistics. Once optimised, performance of the program was assessed and it was applied to the screening of several genomes.
K-values
An optimal k-value was chosen using a series of cross-validation tests. These were computed with a range of parameters and, consistently, the program found that accuracy showed a weak peak at k = 5 and gradually declined thereafter. However performance was generally insensitive to the precise value of k, with similar performance shown for moderate values ≥ 5.
Differential amino acid weightings
A genetic algorithm was used to calculate optimal amino acid weightings for differentiating between bbtm and ntm proteins. The results are shown in Figure 2, alongside weights derived from average compositional differences between the classes. Amino acids contributing the most to classification include Cys, Phe, His, Met, Asn, Gln and Thr. Those contributing the least include Glu, Pro and Tyr. The greatest contributing amino acid, Phe contributed 3.76 times more than the lowest, Pro.
Interestingly, these weights did not completely correlate with compositional differences (Figure 2). Phe had the greatest GA weighting, with 0.077, but had a relatively small composition difference between training sets, with corresponding weight 0.042 (ranked 15th of 20) and Glu had a fairly large composition difference (ranked 7th) but lower GA weighting (ranked 16th). However, there were some correlations, with Asn, His, Cys and Met ranked 2nd, 3rd, 4th and 5th in the GA weightings and 4th, 2nd, 1st and 6th respectively in the composition difference rankings.
Weights significantly differed from those used by Liu [21] who found, using a Fisher's Discrimination Ratio, that the amino acids most useful for distinguishing between beta-strands of globular and membrane proteins were Gly, Val, Ile, Asn, Leu and Cys. These differences can be attributed to the fact that Liu tried to identify differences in strand residues, whereas our method identifies differences in the composition of entire sequences.
Performance
The ability of the program to discriminate between different classes was tested using a 'leave homologues out' cross-validation (see methods) and was defined in terms of PPV, sensitivity and accuracy. Figure 3 shows how PPV, sensitivity and accuracy vary over a range of discrimination scores. Performance results are summarised in Tables 2,3, with the optimal cut-off point (discrimination score giving the highest accuracy) used. Table 2 summarises the performance difference between the program with various features, i.e. weighted amino acids and query sequence evolutionary information. Table 3 describes the ability of the program to discriminate between various protein classes with two different settings. Without inclusion of query sequence evolutionary information, the program was better at discriminating between bbtm and ntm proteins than bbtm and ahtm, with accuracies of 85% and 77.5% respectively. This difference was reduced with the inclusion of query sequence evolutionary information and weighted amino acids, with a prediction accuracy of 92.5% for discrimination between both bbtm and ntm proteins and bbtm and ahtm proteins.
Results reported so far have used cross-validations based on removing detectable homologues (BLAST E-value<1) from the training set. The results have shown high accuracy discriminations. This indicates that amino acid composition can be used to identify bbtm proteins. It is not possible to know the extent of very distant homology in the training set, since this is often only apparent when 3D structures are determined. It is not clear therefore whether the good performance we observe results from the detection of distant homologues, or whether the composition signal is a characteristic of many evolutionary unrelated families of bbtm protein. It seems likely that both explanations contribute to the results, which indicate at the very least that composition is an important feature of these proteins that is preserved over long evolutionary distances and may be shared by unrelated bbtm proteins.
The program was extremely fast, able to query 400 sequences in <1 minute on a 2 Ghz Pentium processor. When using evolutionary information, speed was limited by a BLAST query against Uniprot/Swissprot, and 'all against all' BLAST runs to identify the similarities of homologues. However, even with evolutionary information TMB-Hunt is still faster than Prof-TMB, of a similar speed to Pred-TMBB and only marginally slower than BOMP.
Specific examples
Cross-validation results were reviewed specifically for a number of bbtm proteins that are non-typical, controversial, expressed in membranes other than the outer membrane of gram negative bacteria or for bbtm proteins of gram negative bacteria that have recently been structurally resolved. The aim of TMB-Hunt is identification of novel families of bbtm protein. Unfortunately a fair comparison of the abilities of various predictors to detect novel families is difficult owing to unavoidable uncertainties about training set contents and in some cases (e.g. BOMP) a lack of user control in specificity thresholds. In an attempt to make this comparison we chose examples that for the reasons given should not be well represented in the training sets of other predictors. The ability of TMB-Hunt to identify novel families is given with results coming from cross-validation tests. Table 4 gives details of prediction results using TMB-Hunt and compares them with three other web-based bbtm protein predictors; BOMP, Prof-TMB, Pred-TMBB.
Pred-TMBB and TMB-Hunt both correctly classified non-typical bbtm proteins TolC [8] (P02930), Alpha-hemolysin [44] (P09616) and the Mycobacterial Porin [4] (Q9RLP7), whilst these were classified as non-bbtm by BOMP and Prof-TMB. The secreted pore-forming toxin, Alpha-hemolysin is difficult to classify because the majority of its beta-strands are non-membrane. Alpha-hemolysin is homoheptameric, with each subunit contributing 2 strands to a 14 strand TM barrel. In addition to the 2 TM strands, each subunit consists of 14 soluble strands which make up a cap and rim domain. The Mycobacterial Porin, has not been included in the training sets of any currently published predictors, because its structure has only recently been resolved [4] and because, at 10 nm width, the outer membrane of gram positive Mycobacteria is unlike that of gram negative bacteria at 4 nm width [45]. TolC has been a problem in classification because each of the three subunits contributes just 4 strands to the beta-barrel and contains large stretches of alpha-helix.
To confirm that the predictor was not just selecting proteins destined for the outer membranes of gram negative bacteria, we also tested with a number of mitochondrial and chloroplast bbtm proteins. All the predictors tested were able to correctly classify the mitochondrial porin VDAC (Q9RLP7), but only BOMP and Pred-TMBB classified Tom40 (Q18090) as a bbtm protein. Only Prof-TMB and TMB-Hunt (using the 'leave-one out' cross-validation) classified Toc75 (Q43715) as a bbtm protein and only Pred-TMBB and TMB-Hunt identified OEP24 (O49929).
All four predictors tested were able to correctly identify proteins with recently resolved structures i.e. Tsx [46] (P22786), FadL [47] (P10384), BtuB [48] (P06129) except BOMP which misclassified NalP [49] (Q8GKS5). BOMP was the only predictor tested which did not classify Secretin [50] (P31700) as a bbtm protein but all four classified the Usher protein [51] (P30130) as bbtm. A 60 kDa cysteine rich outer-membrane protein [52] (P26758), was the only example that was not classified as a bbtm protein by any of the predictors. However the experimental evidence that this is a genuine bbtm protein is weak and it has been suggested that it is falsely annotated [21]. It should be noted that PSORT-B 2.0 [53] identified all of these examples as outer membrane proteins, including the 60 kDa rich outer membrane protein. However it classified these using strong homology to sequences within its training set and thus did not give a representation of its ability to predict novel families of bbtm proteins.
Differences in the prediction results of these algorithms with these examples suggests that combined approaches could result in a higher overall accuracy.
Genome screening
Figure 4 demonstrates typical results seen when screening a genome. It demonstrates that due to the large number of sequences queried, a number of sequences get scores with an E-value >1 but a beta barrel score indicative of a bbtm protein (i.e. >0). These sequences are said to be in the 'twilight zone' because it is impossible to classify them as either bbtm or not. To reduce the number of sequences within this zone, sequences without signal peptides were removed. Sequences were accepted if a signal peptide was predicted using SignalP 3.0 with either the Neural Network [54] or HMM [55] modes, so as to minimise the number of potential candidates removed. Similar filtering systems have been applied in previous bbtm protein screening attempts [3,16,56]. Signal peptide filtering poses certain risks owing to errors in the prediction of the 5' ends of genes [30] and imperfections in signal peptide prediction algorithms, but these risks are outweighed by the reduction of FP sequences within the twilight zone.
A range of organisms with completed genomes were screened for bbtm proteins, including several bacteria, a protozoan, a fungus, a nematode and an angiosperm. Table 5 shows the results of proteomes screened. Plasmodium falciparum, Saccharomyces cerevisiae, Caenorhabditis elegans and Arabidopsis thaliana were screened as eukaryotic tests. To date, the only predicted eukaryotic bbtm proteins are those of the mitochondrial and chloroplast outer membranes, however the possibility of other eukaryotic bbtm protein families should not be ignored. Three examples of where they could exist are i) organelles of endosymbiotic bacterial origin other than the mitochondria and chloroplasts e.g. the apicoplast of apicomplexan parasites including the malaria parasite Plasmodium [57] or ii) novel double membrane systems e.g. the outer membranes of the parasitic worm schistosomes, which contains two overlaid phospholipid bilayers [58] and iii) toxins e.g. TT95 which is a pore forming molecule produced by the parasitic nematode Trichuris [59] but which does not contain any predicted TM helices.
Screening eukarotic genomes for bbtm proteins is a more complex process than with prokaryotes owing to larger numbers of sequences queried and a wider range of targeting signals. TMB-Hunt is able to identify mitochondrial and chloroplast outer membrane bbtm proteins (Table 4), but these were missed during eukaryotic genome screening due to prior removal of sequences without signal peptides. Owing to the wide range of eukaryotic protein targeting pathways, eukaryotic sequences should ideally be screened without prior filtering, however this would result in much larger numbers of sequences within the twilight zone. Another alternative would be an addition to the score whenever targeting signals are detected.
TMB-Hunt did not predict many bbtm proteins in eukaryotes; 3 with an E-value <1 in P. falciparum (0.03% of all proteins screened), 4 in S. cerevisiae (0.07%), 23 in Arabidopsis thaliana (0.07%) and 26 in C. elegans (0.1%), with the majority of selected sequences in A. thaliana and C. elegans being closely related and described as hypothetical or putative proteins. Only 1 eukaryotic protein got an E-value <0.1, a P. falciparum gene annotated as a serine protease with an E-value of 0.032.
The mean percentage of proteins in Gram negative bacterial proteomes, with an E-value <1, was 1.37%, with a range of 0.65–2.46%. The figure was highest in proteobacteria, possibly reflecting biases in the training set, with homologies to training instances enabling statistically significant scores (E-values) for many sequences. However given that the numbers of bbtm proteins in various bacterial phyla is not known, it may be that these results reflect true figures. Previous results [17] identified smaller numbers of bbtm proteins in some genomes e.g. Aquifex aeolicus, Thermatoga maritima and Trepanoma palidium although the numbers of sequences screened were not given.
Escherichia coli O157:H7 proteins downloaded from Uniprot were screened in order to compare results with high quality annotation (Figure 5). In total, 249 sequences got a positive beta barrel score when, given the number of sequences queried, 133 would be expected. Thus assuming the remaining 116 sequences are genuine bbtm proteins, the proteome contains (116/4005) × 100 = 2.896% bbtm proteins (a number consistent with other predictions). Of these 249 sequences, 69 had an E-value<1, that is 1.72% of all proteins queried. These 69 included 15 proteins described as outer membrane and TM, 40 hypothetical or putative bbtm proteins described as probable OMPs or with homology to OMPs, 6 hypothetical proteins without homology to well annotated proteins, 4 flagellar proteins, 3 lipoproteins and 1 well known ahtm protein. The 15 proteins described as outer membrane and TM should be bbtm proteins and the 40 with homology to OMPs are probably bbtm proteins. The flagellar are possible bbtm proteins as several flagellar proteins are known bbtm proteins. The 6 hypothetical proteins without homology to well annotated proteins possibly represent novel families of bbtm protein. The 3 lipoproteins are non-bbtm proteins and the 1 ahtm protein could be easily filtered using a ahtm protein predictor.
TMB-Hunt proved successful in that Uniprot annotation suggests that the vast majority of bbtm proteins (65 of the 69 (>95%)) it predicted were probably bbtm proteins. However, several more probable bbtm proteins were found in the twilight zone, suggesting that this algorithm alone does not infallibly detect all bbtm proteins, even in organisms well represented in the training set. In comparing results with BOMP, we found it rejected the lipoproteins that TMB-Hunt incorrectly classified as bbtm (Q8XBQ1, Q7ABP6, Q7ABA4), whilst correctly classifying a number of proteins annotated as bbtm proteins which were within the TMB-Hunt twilight zone (e.g. Q7AGG6, Q7AY93). However we found that BOMP also incorrectly rejected a large number of annotated bbtm proteins that we classified with an E-value <1 (e.g. Q7AAR4, Q7A9N7). Similar patterns were found with Pred-TMBB and Prof-TMB. These differences are further evidence suggesting that combining algorithms could lead to a higher overall accuracy.
Because composition is correlated with physicochemical environment [26], TMB-Hunt struggles with differentiation between bbtm proteins and proteins occupying similar environments i.e. lipoproteins and periplasmic proteins. However TMB-Hunt gets a stronger signal from bbtm proteins as they effectively occupy 3 environments, the transmembrane (where there is a preference for amino acids which form TM beta-strands) and either side of it, whereas lipoproteins and periplasmic proteins will occupy only one side of the membrane. The liability of TMB-Hunt is thus different to that of topology based predictors which typically report difficulties in discriminating between beta-strands of bbtm proteins and some globular proteins.
Conclusion
A program called TMB-Hunt has been described which identifies bbtm proteins using the amino acid composition of entire sequences. TMB-Hunt uses a novel method for calibration of results from the k-NN algorithm and uses evolutionary information from close homologues to build composition profiles. We suggest that these methods can be used to boost the accuracy of other k-NN and composition based classifiers.
TMB-Hunt was found to have several advantages over existing methods. Firstly, a cross-validation analysis showed performance to be superior to that of other bbtm protein predictors. Secondly, unlike previous predictors which are dependent on TM beta-strand detection, this method does not require resolved structures and thus larger more representative training sets could be used. Thirdly, by adopting a novel approach, we believe that the major benefit of this program is that it has different liabilities to others. This was demonstrated by its ability to correctly classify several proteins with which previous predictors struggled. Finally, it is extremely quick, capable of screening >400 sequences per minute. TMB-Hunt has been successfully applied to the screening of several genomes, however, numerous proteins fell into the twilight zone, where it was impossible to statistically categorise them as either bbtm or not. It is therefore intended that it will be included as part of a consensus approach, which can be used to hunt for novel families of bbtm protein.
Availability and requirements
Project name: TMB-Hunt
Project home page: A web server is available at .
Operating system: LINUX
Programming languages: ANSI C and Perl
Other requirements: None
Licence: GPL
Any restrictions to non-academics: None
Abbreviations
AA – Amino acid
ahtm – Alpha-helical transmembrane
bbtm – Beta-barrel transmembrane
BLAST – Basic Local Alignment Search Tool
GA – Genetic Algorithm
HMM – Hidden Markov Models
k-NN – k-Nearest Neighbour
MCC – Matthews Correlation Coefficient
ntm – Non Transmembrane
OMP – Outer membrane protein
PDB – Protein DataBank
PPV – Positive predictive value
TM – Transmembrane
Authors' contributions
AGG constructed the datasets, wrote and tested the programs, screened genomes and built the website. AA suggested the project and analyzed the genome screening results. DRW oversaw the construction of the programs and helped develop the methods. All authors have read and approved the final manuscript.
Supplementary Material
Additional File 1
TMB-Hunt source code and training sets
Click here for file
Additional File 2
E. coli O157:H7 (Uniprot sequence) query results. Various proteomes screened, examples of results and queries, help files.
Click here for file
Acknowledgements
The authors would like to thank the MRC for funding and three anonymous reviewers for their constructive criticism.
Figures and Tables
Figure 1 Probabilities used for development of a calibrated score. Probability (y-axis), p(D'≥D), for observing a score D' greater than or equal to D (x-axis) for either bbtm (■) or ntm (▲) proteins. Plots were made by calculating the frequencies of bbtm and ntm proteins identified above certain discrimination scores (using weighted amino acids, no evolutionary information and a 'leave homologues out' cross-validation).
Figure 2 Comparison between GA weightings and difference ratios. Relationship between GA derived weights for amino acids and weights based simply on average compositional distances between classes.
Figure 3 TMB-Hunt performance over a range of discrimination scores. Accuracy (x), sensitivity (▲) and PPV(■) of the predictor at range of discrimination score thresholds. The above results were taken for the predictor discriminating between bbtm and non-bbtm proteins, using the 'leave homologues out' cross-validation, with weighted amino acids and evolutionary information for the query sequence. Similar patterns were found with all settings i.e. using weighted amino acids, no evolutionary information, 'leave homologues out' cross-validation and discriminating between bbtm and ntm proteins.
Figure 4 Range of E-values and BB-scores from E. coli screening. Sequences with a predicted signal peptide from the proteome of E. coli, were screened using the algorithm described. Sequences were then sorted by their E-values and plotted graphically. The graph demonstrates that in proteome screening with this tool there a number of sequences will be identified with positive bb scores, but E-values >1. Sequences with these scores are described as being in the twilight zone.
Figure 5 Uniprot annotation of predicted E. coli bbtm proteins. Numbers of E. coli O157:H7 sequences with a TMB-Hunt E value <= 1 with different categories of annotation in Uniprot.
Table 1 Sequence datasets used to generate training sets.
Training dataset Sources Initial number sequences Sequences >120 AA Size after redundancy removal
ntm PDB-REPRDB [32] 3159 2290 1763
ahtm Sanger all-alpha membrane datasets A, B and C [33] 189 166 132
bbtm TC-DB [35], Uniprot [34] and PDB [5] 1126 1107 196
Three training datasets were generated using sequences from various sources. Datasets were filtered for sequences of <120 AA and clustered to remove redundancy.
Table 2 Program performance using different settings.
BBTM vs NTM
% Sensitivity % Specificity % PPV % NPV % Accuracy
Plain 83 87 86.5 83.7 85
Weighted AAs 84 91 90.3 85 87.5
Evolutionary information 89 94 93.7 89.5 91.5
Evolutionary information + weighted AAs 91 94 93.8 91.3 92.5
Ability of the program to discriminate between bbtm and ntm proteins, using the 'leave homologues out' cross-validation method and with a range of different features. The plain mode indicates neither evolutionary information or weighted amino acids were included.
Table 3 Ability of program to differentiate between various protein classes.
A. Plain
% Sensitivity % Specificity % PPV % NPV % Accuracy
bbtm vs ntm
83 87 86.5 83.7 85
bbtm vs ahtm 83 72 74.8 80.1 77.5
B. Evolutionary Information plus weighted AAs
% Sensitivity % Specificity % PPV %NPV % Accuracy
bbtm vs ntm 91 94 93.8 91.3 92.5
bbtm vs ahtm 88 97 96.7 88.9 92.5
A shows the ability of the program to differentiate between various protein classes without inclusion of evolutionary information or differential amino acid weightings. B shows the improvements given the inclusion of these features. Performance was assessed using the 'leave homologues out' cross-validation.
Table 4 Comparison of various predictors with specific examples.
BOMP Prof-TMB Pred-TMBB TMB-Hunt: Leave Homs Out TMB-Hunt: Leave One Out
NalP – Q8GKS5 0Γ 12.32 2.92 10.73 10.73
TSX – P22786 1 10.92 2.94 4.47 4.47
FadL – P10384 1 9.47 2.88 0.8 0.8
BtuB – P06129 1 10.39 2.91 10.82 10.82
Secretin – P31700 0Γ 3.73Γ 2.90 5.48 5.48
Usher – P30130 1 10.46 2.95 10.79 10.79
60 kDA cysteine rich OMP – P26758 0Γ 2.42Γ 3.03Γ -1.70Γ -1.70Γ
Mycobacterial Porin – Q9RLP7 0Γ 5.65Γ 2.84 7.74 7.74
TolC – P02930 0Γ 1.85Γ 2.90 6.76 10.64
Alpha hemolysin – O68404 0Γ 0.83Γ 2.88 9.89 9.89
VDAC – Q60931 1 6.55 2.88 5.24 5.24
Tom40 – Q18090 1 4.79Γ 2.92 -1.04Γ -1.04Γ
Toc75 – Q43715 0Γ 6.50 2.99Γ -1.41Γ 1.24
OEP24 – O49929 0Γ 3.11Γ 2.87 1.55 1.55
All programs were run via their web interfaces, using default settings. Sequences classified as non-bbtm are marked using Γ. BOMP [22] values indicate the number bbtm proteins predicted given the number of sequences queried. Prof-TMB [17] returns a z-score statistic for which 50% of bbtm proteins get a z-score of >= 10 at an accuracy of 80% and 35% bbtm proteins get a z-score >= 6 at an accuracy of 35%. Pred-TMBB [16] returns a threshold score, for which sequences with threshold scores >2.965 are assumed not to be bbtm proteins. Beta-barrel scores, were given for TMB-Hunt. These were calculated without inclusion of evolutionary information, using 'leave homologues out' and 'leave one out' cross-validations. Beta-barrel scores >0 indicate that there is a greater probability that the sequence is from a bbtm protein.
Table 5 Proteomes screened.
Organism Proteins No. signal peptide % proteins with signal peptide No. bbtm protein <E = 1 % of proteins with signal peptide bbtm E<= 1 % bbtm proteins <E = 1
Escherichia coli 5341 1032 19.32 87 8.43 1.63
E. coli Ш 4005 782 19.52 69 8.82 1.72
Pseudomonas aeruginosa 5567 1142 20.51 137 12 2.46
P. aeruginosa Ш 5567 1412 25.36 137 9.7 2.46
Staphylococcus aureus 2632 409 15.54 18 4.4 0.68
Aquifex aeolicus 1560 187 11.98 16 8.55 1.02
Chlamydia trachomatis 895 145 16.20 17 11.7 1.89
Thermatoga maritima 1858 265 14.26 12 4.53 0.65
Trepanoma pallidum 1036 203 19.59 12 5.91 1.16
Bacteroides thetaiotaomicron 4778 1614 33.78 131 8.12 2.74
Deinococcus radiodurans 3182 689 21.65 25 3.62 0.76
Rhodopirellula baltica 7325 1584 20.66 49 3.09 0.67
Plasmodium falciparum Ш 9178 1613 17.57 3 0.18 0.03
Arabidopsis thaliana 28860 5569 19.30 23 0.41 0.07
Caenorhabditis elegans Ш 22561 5778 22.60 26 0.45 0.12
Saccharomyces cerevisiae 5866 651 11.09 4 0.61 0.07
Several proteomes were screened, representing the major kingdoms of life. Proteomes were first filtered for sequences with signal peptides. Remaining sequences were then each queried, returning bb scores and E-value statistics. All proteomes were downloaded from the NCBI FTP site except those denoted Ш, downloaded from Uniprot/SwissProt for superior annotation.
==== Refs
Wimley WC The versatile beta-barrel membrane protein. Curr Opin Struct Biol 2003 13 404 411 12948769 10.1016/S0959-440X(03)00099-X
Casadio R Jacoboni I Messina A De Pinto V A 3D model of the voltage-dependent anion channel (VDAC) FEBS Lett 2002 520 1 7 12044860 10.1016/S0014-5793(02)02758-8
Schleiff E Eichacker LA Eckart K Becker T Mirus O Stahl T Soll J Prediction of the plant beta-barrel proteome: a case study of the chloroplast outer envelope Protein Sci 2003 12 748 759 12649433 10.1110/ps.0237503
Faller M Niederweis M Schulz GE The structure of a mycobacterial outer-membrane channel. Science 2004 303 1189 1192 14976314 10.1126/science.1094114
Bernstein FC Koetzle TF Williams GJ Meyer EFJ Brice MD Rodgers JR Kennard O Shimanouchi T Tasumi M The Protein Data Bank: a computer-based archival file for macromolecular structures. J Mol Biol 1977 112 535 542 875032
Tusnady GE Dosztanyi Z Simon I Transmembrane proteins in the Protein Data Bank: identification and classification. Bioinformatics 2004 20 2964 2972 15180935 10.1093/bioinformatics/bth340
Murzin AG Brenner SE Hubbard T Chothia C SCOP: a structural classification of proteins database for the investigation of sequences and structures. J Mol Biol 1995 247 536 540 7723011 10.1006/jmbi.1995.0159
Postle K Vakharia H TolC, a macromolecular periplasmic 'chunnel' Nat Struct Biol 2000 7 527 530 10876231 10.1038/76726
Schulz GE beta-Barrel membrane proteins. Curr Opin Struct Biol 2000 10 443 447 10981633 10.1016/S0959-440X(00)00120-2
Yau WM Wimley WC Gawrisch K White SH The preference of tryptophan for membrane interfaces Biochemistry 1998 37 14713 14718 9778346 10.1021/bi980809c
Viklund H Elofsson A Best alpha-helical transmembrane protein topology predictions are achieved using hidden Markov models and evolutionary information. Protein Sci 2004 13 1908 1917 15215532 10.1110/ps.04625404
Zhai Y Saier MHJ The beta-barrel finder (BBF) program, allowing identification of outer membrane beta-barrel proteins encoded within prokaryotic genomes. Protein Sci 2002 11 2196 2207 12192075 10.1110/ps.0209002
Wimley WC Toward genomic identification of beta-barrel membrane proteins: composition and architecture of known structures. Protein sci 2002 11 301 312 11790840 10.1110/ps.29402
Martelli PL Fariselli P Krogh A Casadio R A sequence-profile-based HMM for predicting and discriminating beta barrel membrane proteins. Bioinformatics 2002 18 S46 53 12169530
Liu Q Zhu YS Wang BH Li YX A HMM-based method to predict the transmembrane regions of beta-barrel membrane proteins. Comput Biol Chem 2003 27 69 76 12798041 10.1016/S0097-8485(02)00051-7
Bagos PG Liakopoulos TD Spyropoulos IC Hamodrakas SJ A Hidden Markov Model method, capable of predicting and discriminating beta-barrel outer membrane proteins. BMC Bioinformatics 2004 5 29 15070403 10.1186/1471-2105-5-29
Bigelow HR Petrey DS Liu J Przybylski D Rost B Predicting transmembrane beta-barrels in proteomes. Nucleic Acids Res 2004 32 2566 2577 15141026 10.1093/nar/gkh580
Bagos PG Liakopoulos TD Spyropoulos IC Hamodrakas SJ PRED-TMBB: a web server for predicting the topology of beta-barrel outer membrane proteins Nucleic Acids Res 2004 32 W400 4 15215419
Gromiha MM Ahmad S Suwa M Neural network-based prediction of transmembrane beta-strand segments in outer membrane proteins. J Comput Chem 2004 25 762 767 14978719 10.1002/jcc.10386
Natt NK Kaur H Raghava GP Prediction of transmembrane regions of beta-barrel proteins using ANN- and SVM-based methods. Proteins 2004 56 11 18 15162482 10.1002/prot.20092
Liu Q Zhu Y Wang B Li Y Identification of beta-barrel membrane proteins based on amino acid composition properties and predicted secondary structure. Comput Biol Chem 2003 27 355 361 12927109 10.1016/S1476-9271(02)00085-3
Berven FS Flikka K Jensen HB Eidhammer I BOMP: a program to predict integral beta-barrel outer membrane proteins encoded within genomes of Gram-negative bacteria. Nucleic Acids Res 2004 32 W394 9 15215418
Zhang CT Chou KC An optimization approach to predicting protein structural class from amino acid composition Protein Sci 1992 1 401 408 1304347
Nakashima H Nishikawa K Discrimination of intracellular and extracellular proteins using amino acid composition and residue-pair frequencies J Mol Biol 1994 238 54 61 8145256 10.1006/jmbi.1994.1267
Chou KC Elrod DW Prediction of membrane protein types and subcellular locations Proteins 1999 34 137 153 10336379 10.1002/(SICI)1097-0134(19990101)34:1<137::AID-PROT11>3.0.CO;2-O
Cedano J Aloy P Perez-Pons JA Querol E Relation between amino acid composition and cellular location of proteins. J Mol Biol 1997 266 594 600 9067612 10.1006/jmbi.1996.0804
Andrade MA O'Donoghue SI Rost B Adaptation of protein surfaces to subcellular location J Mol Biol 1998 276 517 525 9512720 10.1006/jmbi.1997.1498
Park KJ Kanehisa M Prediction of protein subcellular locations by support vector machines using compositions of amino acids and amino acid pairs. Bioinformatics 2003 19 1656 1663 12967962 10.1093/bioinformatics/btg222
Bendtsen JD Nielsen H von Heijne G Brunak S Improved prediction of signal peptides: SignalP 3.0 J Mol Biol 2004 340 783 795 15223320 10.1016/j.jmb.2004.05.028
Casadei R Strippoli P D'Addabbo P Canaider S Lenzi L Vitale L Giannone S Frabetti F Facchin F Carinci P Zannotti M mRNA 5' region sequence incompleteness: a potential source of systematic errors in translation initiation codon assignment in human mRNAs. Gene 2003 4 185 193 10.1016/S0378-1119(03)00835-7
Rost B Fariselli P Casadio R Topology prediction for helical transmembrane proteins at 86% accuracy Protein sci 1996 5 1704 1718 8844859
Noguchi T Matsuda H Akiyama Y PDB-REPRDB: a database of representative protein chains from the Protein Data Bank (PDB). Nucleic Acids Res 2001 1 219 220 10.1093/nar/29.1.219
Moller S Kriventseva EV Apweiler R A collection of well characterised integral membrane proteins. Bioinformatics 2000 16 1159 1160 11159338 10.1093/bioinformatics/16.12.1159
Bairoch A Apweiler R Wu CH Barker WC Boeckmann B Ferro S Gasteiger E Huang H Lopez R Magrane M Martin MJ Natale DA O'Donovan C Redaschi N Yeh LS The Universal Protein Resource (UniProt) Nucleic Acids Res 2005 33 Database Issue D154 9 15608167
Busch W Saier MH The transporter classification (TC) system, 2002. Crit Rev Biochem Mol Biol 2002 37 287 337 12449427 10.1080/10409230290771528
The NCBI FTP site,
The sequence retrieval system:
Etzold T Argos P SRS--an indexing and retrieval tool for flat file data libraries. Comput Appl Biosci 1993 9 49 57 8435768
Cover T Hart P Nearest neighbour pattern classification IEEE Trans Inform theory 1967 IT-13 21 27 10.1109/TIT.1967.1053964
Friedman JH Baskett F Shustek LJ An algorithm for finding nearest neighbors IEEE Trans Inform Theory 1975 C-24 1000 1006
Pearson WR Lipman DJ Improved tools for biological sequence comparison Proc Natl Acad Sci U S A 1988 85 2444 2448 3162770
Altschul SF Gish W Miller W Myers EW Lipman DJ Basic local alignment search tool. J Mol Biol 1990 215 403 410 2231712 10.1006/jmbi.1990.9999
Eiben AE Schoenauer M Evolutionary computing Information Processing Letters 2002 82 1 6 10.1016/S0020-0190(02)00204-1
Song L Hobaugh MR Shustak C Cheley S Bayley H Gouaux JE Structure of staphylococcal alpha-hemolysin, a heptameric transmembrane pore Science 1996 274 1859 1866 8943190 10.1126/science.274.5294.1859
Brennan PJ Nikaido H The envelope of mycobacteria. Annu Rev Biochem 1995 64 29 63 7574484 10.1146/annurev.bi.64.070195.000333
Ye J Van Den Berg B Crystal structure of the bacterial nucleoside transporter Tsx Embo J 2004 23 3187 3195 15272310 10.1038/sj.emboj.7600330
van den Berg B Black PN Clemons WMJ Rapoport TA Crystal structure of the long-chain fatty acid transporter FadL Science 2004 304 1506 1509 15178802 10.1126/science.1097524
Chimento DP Mohanty AK Kadner RJ Wiener MC Substrate-induced transmembrane signaling in the cobalamin transporter BtuB Nat Struct Biol 2003 10 394 401 12652322 10.1038/nsb914
Oomen CJ Van Ulsen P Van Gelder P Feijen M Tommassen J Gros P Structure of the translocator domain of a bacterial autotransporter Embo J 2004 23 1257 1266 15014442 10.1038/sj.emboj.7600148
Bitter W Secretins of Pseudomonas aeruginosa: large holes in the outer membrane Arch Microbiol 2003 179 307 314 12664194
Thanassi DG Stathopoulos C Dodson K Geiger D Hultgren SJ Bacterial outer membrane ushers contain distinct targeting and assembly domains for pilus biogenesis J Bacteriol 2002 184 6260 6269 12399496 10.1128/JB.184.22.6260-6269.2002
Everett KD Hatch TP Architecture of the cell envelope of Chlamydia psittaci 6BC J Bacteriol 1995 177 877 882 7532170
Gardy JL Laird MR Chen F Rey S Walsh CJ Ester M Brinkman FS PSORTb v.2.0: expanded prediction of bacterial protein subcellular localization and insights gained from comparative proteome analysis Bioinformatics 2005 21 617 23 15501914 10.1093/bioinformatics/bti057
Nielsen H Engelbrecht J Brunak S von Heijne G Identification of prokaryotic and eukaryotic signal peptides and prediction of their cleavage sites Protein Eng 1997 10 1 6 9051728 10.1093/protein/10.1.1
Nielsen H Krogh A Prediction of signal peptides and signal anchors by a hidden Markov model Proc Int Conf Intell Syst Mol Biol 1998 6 122 130 9783217
Casadio R Fariselli P Finocchiaro G Martelli PL Fishing new proteins in the twilight zone of genomes: the test case of outer membrane proteins in Escherichia coli K12, Escherichia coli O157:H7, and other Gram-negative bacteria Protein Sci 2003 12 1158 1168 12761386 10.1110/ps.0223603
Fichera ME Roos DS A plastid organelle as a drug target in apicomplexan parasites Nature 1997 390 407 409 9389481 10.1038/37132
Gobert GN Stenzel DJ McManus DP Jones MK The ultrastructural architecture of the adult Schistosoma japonicum tegument Int J Parasitol 2003 33 1561 1575 14636672 10.1016/S0020-7519(03)00255-8
Barker GC Bundy DA Isolation of a gene family that encodes the porin-like proteins from the human parasitic nematode Trichuris trichiura Gene 1999 229 131 136 10095112 10.1016/S0378-1119(99)00039-6
|
15769290
|
PMC1274253
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 15; 6:56
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-56
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-571577400810.1186/1471-2105-6-57Methodology ArticleCombining Affymetrix microarray results Stevens John R [email protected] RW [email protected] Department of Statistics, Purdue University, 150 N. University Street, West Lafayette, Indiana 47907-2067, USA2 Department of Agronomy, Purdue University, Lilly Hall of Sciences, 915 W. State Street, West Lafayette, Indiana 47907-2054, USA2005 17 3 2005 6 57 57 28 10 2004 17 3 2005 Copyright © 2005 Stevens and Doerge; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
As the use of microarray technology becomes more prevalent it is not unusual to find several laboratories employing the same microarray technology to identify genes related to the same condition in the same species. Although the experimental specifics are similar, typically a different list of statistically significant genes result from each data analysis.
Results
We propose a statistically-based meta-analytic approach to microarray analysis for the purpose of systematically combining results from the different laboratories. This approach provides a more precise view of genes that are significantly related to the condition of interest while simultaneously allowing for differences between laboratories. Of particular interest is the widely used Affymetrix oligonucleotide array, the results of which are naturally suited to a meta-analysis. A simulation model based on the Affymetrix platform is developed to examine the adaptive nature of the meta-analytic approach and to illustrate the usefulness of such an approach in combining microarray results across laboratories. The approach is then applied to real data involving a mouse model for multiple sclerosis.
Conclusion
The quantitative estimates from the meta-analysis model tend to be closer to the "true" degree of differential expression than any single lab. Meta-analytic methods can systematically combine Affymetrix results from different laboratories to gain a clearer understanding of genes' relationships to specific conditions of interest.
==== Body
Background
Microarray technology allows simultaneous assessment of transcript abundance for thousands of genes. This exciting research tool permits the identification of genes which are significantly differentially expressed between conditions. With the use of microarrays becoming more commonplace, it is not unusual for several different laboratories to investigate the genetic implications of the same condition(s). Each lab may produce its own list of candidate genes which they believe to be related to the condition of interest. As a result of sound statistical approaches, each lab will also have for each candidate gene some quantitative measure that serves as the basis for the claim of statistical significance.
Of interest in this paper are the methods by which these quantitative measures may be combined across labs to arrive at a more comprehensive understanding of the effects of the different candidate genes. Where the term "analysis" is used to describe the quantitative approaches to draw useful information from raw data, the term "meta-analysis" [1] refers to the approaches used to draw useful information from the results of previous analyses. Meta-analysis has been predominantly used in the medical and social sciences, in situations where several studies may have been conducted to investigate the effect of the same treatment, and the researcher seeks to combine the results of the different studies in a meaningful way in order to arrive at a single estimate of the true effect of the treatment. For the current application, meta-analytic approaches can be employed to combine the results from several different labs without having access to the original raw data that yielded the initial results. Such approaches have particular utility with the results of Affymetrix GeneChip® microarrays and other fabricated arrays, where results are given in a uniform format that readily lends itself to comparison between labs and combination across labs.
A measure of the degree or magnitude of differential expression provides more information regarding a gene's relation to a disease or condition of interest than does a statement regarding its significance or nonsignificance. This information is useful because it allows for greater precision of estimation of the gene's effect with respect to the condition of interest. That is, to arrive at a clearer understanding of a gene's true effect relating to the condition of interest, it is most helpful to have a quantitative measure of the magnitude of differential expression rather than a simple declaration of significance.
Prior applications of meta-analysis to microarray data have either sought to combine P-values or to combine results across platforms (i.e., combining Affymetrix and cDNA array results) [2-6]. Combining only P-values, while useful in obtaining more precise estimates of significance, does not provide information that is easily interpretable by a biologist, may not indicate the direction of significance (e.g., up- or down-regulation), and most importantly, gives no information regarding the magnitude of the estimated expression change. Similarly, while a "vote-counting" approach based on P-values [6] addresses differences in lists of significant genes from separate experiments, it gives no information regarding the magnitude of the estimated expression change. While an "integrative correlation" approach [5] will help identify genes with reproducible expression patterns, it also does not provide any information regarding the magnitude of the estimated expression change
Previous attempts to combine results across microarray platforms (i.e., technologies) assume that spot intensities or signal values for a given gene can be directly compared even though they represent different segments of the gene. That is, a spot for a given gene on a cDNA array represents the entire gene, while each spot for the same gene on an Affymetrix array represents a specific small section of the gene. Thus, combining results across technologies using only spot intensities is problematic from a biological perspective because the measurements represent different physical quantities. Even if the average spot intensity on an Affymetrix array is used, it is not certain that this average spot intensity value is at all comparable to the spot intensity value of the gene on a cDNA array.
Moreau et al. [4] report that 'after appropriate filtering, ratio and intensity data from different platforms can be compared and are amenable to' be used in a meta-analysis. However, "filtering", or "averaging out" outliers or non-reproducible spots, requires some subjectivity in the method of choice and may force agreement between platforms where no agreement should exist due to fundamental technological differences. Parmigiani et al. [5] attempt to address this problem of cross-platform consistency by identifying a set of genes whose expression patterns are essentially reproducible across platforms. However, even for these "reproducible" genes, there remains the question of how to systematically combine their corresponding results from the several laboratories to arrive at a single quantitative measure of differential expression. At the very least, if results are to be combined across platforms in a meta-analysis, the use of covariates [7] should be employed to account for the underlying differences between oligonucleotide (e.g., Affymetrix) and cDNA platforms. The focus of the current application is restricted to standard Affymetrix microarray results, and a method to combine results across laboratories is proposed and evaluated.
Results
Affymetrix technology
The Affymetrix GeneChip® microarray [8] represents individual genes by 25-mer segments (probes) fixed to the chip, and also makes use of mismatch probes differing at position 13. Each gene on the chip is typically represented by the same number of probe pairs on the chip (usually 14–20), although exceptions exist. It is now possible for some organisms' entire genomes to be represented on a single microarray (e.g., Arabidopsis). Appropriately prepared tissue sample is hybridized to the array and the array is scanned, producing raw data consisting of the intensities of the individual spots on the array. These intensities come in pairs, with PM denoting the intensity of a perfect-match probe and MM denoting the intensity of the corresponding mismatch probe.
Affymetrix algorithms
Affymetrix has developed statistical algorithms [9] that employ these individual spot intensities for the purpose of estimating the true expression levels of individual genes in single samples. Furthermore, the Affymetrix approach compares gene expression levels in two different tissues (samples or treatment conditions) and reports a "signal log ratio" (SLR) with 95 percent confidence bounds. The signal log ratio is the signed Iog2 of the signed fold change (FC) familiar to biologists [9]. That is, FC = 2SLR if SLR ≥ 0 and FC = (-1)2-SLR if SLR < 0. The algorithm used by Affymetrix to compute the SLR is based on Tukey's biweight algorithm [10] and for each gene takes a weighted average of the log2 of the ratio of PM - MM between treatments or conditions, with weights related to the deviations from the median log2 ratio for the gene, and with adjustments made when PM <MM. The resulting weighted average is the SLR.
Between the two conditions of interest, each gene either changes its level of expression or the level remains the same. A declaration of significant differential expression results from sufficient evidence that the gene is not expressed the same in the two conditions (i.e., that the SLR differs significantly from zero). Tukey's biweight algorithm provides an estimate for the variability of the SLR and an approximate distribution for the SLR estimate. The Affymetrix software (Microarray Suite Version 5.0, or MAS 5.0) reports a 95 percent confidence interval for the SLR [11], from which the estimated standard error can be computed. Individual laboratories can use this information to make a declaration of significant differential expression.
It should be noted that this SLR estimate represents a measure of differential expression between two chips (generically referred to as a base sample chip and an experimental sample chip). In practice, of course, it is recommended that experiments involve more than two chips, but the current MAS 5.0 algorithm is designed to represent differential expression only between two chips at a time. Other approaches exist to measure differential expression (dChip [12] and RMA [13], for example), and future work will evaluate their performance in meta-analyses. However, for the purposes of this current work, the focus of differential expression will rely on the SLR estimate of differential expression between two chips, because these are the estimates provided automatically by the commercial MAS 5.0 software.
If we let denote the estimate of the SLR θi,k for gene k in lab i, and be the upper bound for the 95 percent confidence interval for the same, then both and are reported by the Affymetrix software. The Affymetrix documentation [9] gives the estimated standard error of as si,k = ( - ) / ti,k(.975), where ti,k(.975) is the upper .025 critical value of the t distribution with dfi,k degrees of freedom, where dfi,k = max(0.7(ni,k - 1), 1), with ni,k representing the number of probe pairs representing gene k on each array in study i. The estimated variance of the SLR estimate is vi,k = .
Accordingly, lab i could then test for significant differential expression of gene k (i.e., test the hypothesis : θi,k = 0) by use of the test statistic Ai,k = /si,k. Under , Ai,k approximately follows the t distribution with dfi,k degrees of freedom. The significance P-value (Pi,k) for gene k in lab i is the value such that |Ai,k| is the upper Pi,k/2 critical value of the t distribution with dfi,k degrees of freedom, and is rejected at the αi,k level if Pi,k <αi,k. That is, if Pi,k is sufficiently small, then lab i would declare gene k significantly differentially expressed.
Meta-analysis
The general meta-analytic framework [7] assumes that a measurable relationship exists between certain quantities of interest, and n independent studies have been conducted to examine this relationship. In turn, this relationship can be quantified so that each study produces an estimate of the relationship. If the estimates are appropriately standardized, then each study's estimate can be termed an "effect size" estimate. An effect size is essentially a standardized quantitative expression of the relationship of interest. For example, several different laboratories may investigate which of two drugs are better at treating a particular disease. In this case, the relationship of interest is the difference between the drugs' effects. If each laboratory produces an estimate standardized such that estimates from all laboratories address the same quantity and are on the same scale, then these estimates are effect size estimates.
There are three main classes of effect size estimates [14]. The first and perhaps most common is the standardized difference estimate, such as Hedges's g, similar to the t-statistic in a two sample study: . The second is the standardized relation estimate, such as the sample correlation coefficient r. The third is the measure of significance, such as the P-value from a particular hypothesis test. (Although not an effect size in the traditional sense, the measure of significance approach is mentioned here for the sake of completeness.)
In order to be combined across studies, effect size estimates must address the same measure or quantity, be standardized, and (with the exception of P-values, which are combined differently [15]) include some measure of variability of the effect size estimate [16]. Once each study i has provided its effect size estimate and its measure of variability vi, a meta-analysis can be performed. Three main meta-analytic approaches exist: fixed effects, random effects, and hierarchical Bayes. The first two approaches are summarized here in order of increasing complexity, and the third is the subject of Choi et al. [3] and a future research interest. The three approaches are discussed more fully in Cooper and Hedges [14] and DuMouchel and Normand [17].
Fixed effects meta-analysis model
Assume that n independent studies have provided effect size estimates and measures of variability vi, i = 1, ..., n. The most general meta-analytic approach assumes that
with sampling error εi ~ N(0, ). That is, each is an estimate of a true fixed underlying effect size θi, and it is assumed that θ1 = ... = θn with the common value θ. This is referred to as the homogeneity assumption and can be interpreted as assuming that all studies examined and provided estimates of the same parameter θ, and any differences between the estimates are attributable to sampling error alone. This common value parameter θ is estimated as a weighted average of the effect size estimates:
The weights wi are chosen to minimize the variance of , and this is achieved by , where vi is the estimated variance of . The variance of is .
The underlying assumption of homogeneity : θ1 = ... = θn can be tested by use of the test statistic
Under , Q is approximately distributed as . Then this χ2 distribution can serve as the basis for an approximate test of homogeneity. If Q is larger than the upper αQ critical value of the distribution, is rejected at the αQ level. Alternatively, the homogeneity P-value PQ is the value such that Q is the upper PQ critical value of the distribution, and is rejected at the αQ level if PQ <αQ.
The test of significance : θ = 0 can be considered by use of the test statistic . Under , Z is distributed as N(0, 1), and if |Z| is larger than the upper αZ/2 critical value of the N(0,1) distribution, is rejected at the αZ level. Alternatively, the significance P-value PZ is the value such that |Z| is the upper PZ/2 critical value of the N(0, 1) distribution, and is rejected at the αZ level if PZ <αZ.
Meta-analysis in the context of a microarray experiment assumes that several laboratories have provided quantitative measurements of differential expression (the effect size) for a number of genes along with variability estimates. For the fixed effects model, the homogeneity assumption () provides that for each laboratory the gene is expressed the same, and differences between laboratories are due to sampling error only. On the other hand, the hypothesis has the biological interpretation that there is no change in gene expression between the conditions of interest. This test of significance identifies genes that are significantly differentially expressed between the two conditions, using information from multiple laboratories.
Random effects meta-analysis model
In practice, the homogeneity assumption (and the resulting fixed effects model) tends to be overly simplistic but is presented in this paper for the sake of completeness. This assumption can be relaxed to make the meta-analysis model more appropriate. The basic random effects model [18] assumes n independent studies have provided effect size estimates and measures of variability vi, i = 1, ..., n. In addition, the model assumes that
In this framework, θ is the population mean effect size, and there are two error components, δ and ε, corresponding to between-study and within-study variability, respectively. Each study seeks to make statements regarding this quantity θ, and so takes a sample of individuals from a certain population in order to study the underlying effect size θ. However, due to differences between studies such as time, location, equipment, and other uncontrollable (and possibly unknown) factors, each study will in fact be estimating a slightly different quantity. That is, due to differences between studies, study i is estimating θi, a random effect size from the population of all possible effect sizes. The error component δi ~ N(0, Δ2) is the random deviation of θi from θ (representing variability between studies). In this basic model, Δ2 represents the random variation between studies. Within study i, the actual estimate will vary from the "true" effect size θi based on which random sample is selected. That is, replicates within a study will result in slightly different estimates of the effect size due to sampling error. Here, εi ~ N(0, ) is sampling error (representing variability within study i).
Q is calculated as in the fixed effects model. (Note that the fixed effects model assumes that Δ2 = 0.) The random effects model uses this Q value to calculate new weights , where
Then the meta-analysis estimate for the population mean effect size θ is
The variance of is . The test of significance : θ = 0 can be considered by use of the test statistic . Under , Zw is distributed as N(0, 1), and the significance P-value is calculated in the same manner as in the fixed effects model.
When the random effects meta-analysis is applied in the context of a microarray experiment, again it is assumed that several laboratories have provided quantitative measurements of differential expression (the effect size) for a given gene along with variability estimates. The random effects model assumes that there is some true degree of differential expression for the gene, and each lab is actually estimating a slightly different true degree of differential expression. That is, each laboratory has a slightly different "true" degree of differential expression. In addition, the estimate from each laboratory varies randomly about its true degree of differential expression due to sampling error. Then Δ2 is a measure of the amount of variation between the laboratories' true degrees of differential expression, and the test of significance is used to identify differentially expressed genes by using information across multiple laboratories.
Meta-analysis with Affymetrix data
Our motivation for applying meta-analytic techniques to microarray data is threefold. First, standard platforms (e.g., Affymetrix) make combining results across labs straightforward and eliminate the usual criticism of meta-analyses that "apples and oranges" are being mixed [16] because the estimates being combined across labs have each been standardized by the same algorithms [9] in such a way that they are in fact estimates of the same underlying effect. Furthermore, any known differences between laboratories such as sample tissue type can be incorporated into the meta-analysis by use of covariates [7]. Second, combining raw data may provide more information than combining results, but raw data are not always easy to obtain, and it is conceivable that raw data may become unavailable while published (or unpublished) results are available. Third, if it can be shown that meta-analysis produces similar results to the pooling of raw data, then it can be argued that meta-analytic approaches are more efficient in the sense that they only require easily obtainable results rather than the raw data.
The uniformity of chip design and data acquisition from Affymetrix oligonucleotide microarray experiments readily lends itself to a meta-analysis. Given n studies examining the differences in gene expression between two treatments (e.g., healthy vs. diseased), a meta-analysis can combine each study's signal log ratio (SLR) estimates in a meaningful way by taking the SLR as the effect size estimate. The SLR satisfies the criteria for an effect size (i.e., comparability of estimates, standardization to the same scale, and availability of a variance estimate). The SLR for a given gene represents the degree of differential expression between two conditions, and is directly comparable between labs since it estimates the same physical quantity. The SLR from Affymetrix is standardized in the sense that a SLR of zero means no differential expression is observed, and the algorithms used to produce the SLR place all SLR estimates on the same scale. Finally, a variance for the SLR estimate is provided by the Affymetrix algorithms [9,10].
A general fixed effects model can be employed to perform a meta-analysis to estimate the true effect size (signal log ratio, SLR) θk of gene k. In addition, the test of homogeneity can be evaluated to determine whether the n studies are in fact estimating the same true underlying value of θk, i.e., whether θ1,k = ... = θn,k. If this homogeneity assumption is found to be reasonable, then a test of significance can be considered to determine whether the true signal log ratio θk is significantly different from zero (i.e., whether gene k is significantly differentially expressed between the two conditions). If the homogeneity assumption is deemed unreasonable, then the random effects model can be employed to account for inter-study variability.
Simulation example
In order to evaluate the usefulness of this meta-analytic approach, a simulation study was conducted. The purpose of this simulation study was to illustrate how the results of the meta-analysis compare with the actual ("truth") simulation setting. A simple simulation model was developed with the sole purpose of generating "raw" probe-level data with certain genes "known" to be differentially expressed. While this model may not account for all sources of possible variability, it is nonetheless adequate for the purposes of the current work.
Simulation model
"Raw" probe-level data were generated from a model assuming that mismatch intensities (MM) are random background noise, which is an underlying assumption of the Affymetrix approach [9]. Our investigation of real data indicated that mismatch intensities appear to follow a long-tailed Gamma distribution. Based on this, a random mismatch intensity is simulated for each probe l of each gene k such that MMkl ~ Gamma(α, β), with mean and variance [19].
In this simulation, larger values of the shape parameter α indicate more signal being detected by mismatch probes, with the peak of the distribution of MM intensities being moved away from zero. Larger values of the scale parameter β make high MM intensities less likely by pulling in the tail of the distribution. For the purposes of this simulation, it was assumed that mismatch intensities did not vary across labs or treatments.
Once the background mismatch intensities were obtained, the perfect match (PM) intensities were generated via the model
Yijkl = μ + Li + Gk + P(G)(k)l + LGik + ρk(Tj + LTij + TGjk + LTGijk + TP(G)j(k)l) + ε(ijk)l (7)
where Yijkl is the log2 of the PM - MM difference for probe l of gene k under treatment j in lab i. N labs were considered with each lab using the same two treatments. The term ρk ~ Bernoulli(p) is 1 if gene k is differentially expressed between conditions j = 1 and j = 2, and is 0 otherwise. The parameter p corresponds to the percentage of genes that are differentially expressed, with higher values resulting in more differentially expressed genes. In this model, Li is the effect of lab i, Tj is the effect of treatment j, Gk is the effect of gene k, P(G)(k)l is the effect of probe l of gene k, ε(ijk)l is a random error term, and the other terms are interaction effects. To introduce more between-lab variability, the error variance was allowed to be different in each lab. That is, ε(ijk)l ~ N(0, ) for the error terms in lab i. Each term (X) in the model is assumed to be a random effect from a N(0, ) distribution, except for the constant μ, the fixed effect Tj, and the ρk term. The parameters p, μ, Tj, σ1, ..., σN, and σX for X = L, G, P(G), LG, LT, TG, LTG, and TP(G) can be adjusted to introduce various sources of variability in the "observed" simulated data.
These simulated data can be used to generate "observed" SLR estimates for each gene in each lab. These "observed" SLR estimates can then be combined systematically in a meta-analysis. Note that the "true" SLR value for each gene can be obtained by using the same parameter values as in the simulation model but dropping all lab and error terms. Then the adaptive nature of the meta-analytic approach can be illustrated by comparing the "true" SLR values with the estimates from each lab and from the meta-analysis models.
Simulated data
The simulation was conducted in the R environment [20] with code requiring the use of the affy package [21] from the Bioconductor project [22,23]. While not the purpose of this investigation, the simulation was performed based upon the Affymetrix rat neuro chip RN_U34 with model parameter settings N = 6, α = 0.1, β = 0.0003, p = 0.05, μ = 2.5, σL = 0, σG = 0.5, σP(G) = 0.3, σLG = 0.1, T1 = -0.2, T2 = 0.2, σLT = 0.1, σTG = 1.0, σLTG = 0.13, σTP(G) = 0.5, σ1 = .48, σ2 = .60, σ3 = .72, σ4 = .84, σ5 = .96, and σ6 = 1.08. These parameter settings were selected to produce a distribution of MM intensities similar to that observed in real data (Figure 1a,b) and to force the distribution of signal log ratio (SLR) estimates to fall within a reasonable range with some variation between laboratories (Figure 1c,d).
Most SLR estimates were near zero (Figure 1c), indicating nondifferential expression, while some genes had larger absolute SLR's with smaller standard errors, an indication of significant differential expression. The data were simulated such that there were similar patterns between labs while allowing for lab differences, as evidenced by a comparison of the SLR's from two simulated labs (Figure 1d). While the estimates from the two simulated labs were clearly similar, there were obvious differences between the labs, although not as different as could be observed in real data. As a result, these two labs might produce slightly different lists of significantly differentially expressed genes. The simulation parameters can be adjusted to introduce varying degrees of difference between experiments, and this will affect the final claim made by the meta-analysis regarding statistical significance of differential expression.
Fixed effects meta-analysis results
The results from the six simulated labs were combined using the fixed effects meta-analysis model (Figure 2). The test of homogeneity : θ1,k = ... = θ6,k was performed for each gene k, k = 1, ..., 1322 (the RN_U34 chip has 1322 features or reference sequences), and the P-values for each were summarized in a histogram of homogeneity P-values (Figure 2a). Clearly there was widespread violation of the homogeneity assumption, as evidenced by the abundance of smaller homogeneity P-values.
When the False Discovery Rate (FDR) [24] was controlled at 0.05, 88 of the 1322 genes failed the homogeneity test. That is, there appeared to be significant interlaboratory differences, such that the laboratories did not appear to provide estimates of the same true degree of differential expression for all genes. This appeared to be true for genes across a wide range of fixed effects meta-analysis SLR estimates, as evidenced by the lack of a clear relationship between fixed-effects SLR estimates and homogeneity P-values (Figure 2b). As a result, the random effects meta-analysis model was deemed more appropriate to adjust for the lack of homogeneity.
Random effects meta-analysis results
The same data from the six simulated labs were used in a random effects meta-analysis, and the resulting SLR estimates were similar to those from the fixed effects meta-analysis (Figure 3a). The test of significance : θk = 0 was performed for k = 1, ..., 1322 (i.e., for all 1322 genes), and the P-values for each value of k were summarized in a histogram of significance P-values (Figure 3b). As a result of the parameter selections for the simulation, an abundance of small P-values was observed, indicating a large number of significantly differentially expressed genes. A comparison of the meta-analysis SLR estimates with the significance P-values (Figure 3c) showed a trend of smaller P-values for larger absolute SLR. Similar to the results from a single lab (Figure 1c), most meta-analysis SLR estimates were close to zero (Figure 3d), but the standard errors were slightly lower overall for the meta-analysis estimates, after combining the SLR estimates across labs.
The differences between the fixed effects and random effects models can be summarized by considering bubble plots for a single gene (Figure 4a,b), with bubble area proportional to weights used in the meta-analysis. For this particular gene, laboratory 2 estimated a SLR considerably smaller than the SLR's from the other labs with very small variance and hence very large weight in the fixed effects meta-analysis. As a result, the fixed effects meta-analysis estimated the true SLR for this gene to be closest to the SLR from lab 2. Such a result would call into question the SLR estimate from lab 2 for this gene. The random effects model took this lack of homogeneity into account and appropriately down-weighted the SLR estimates for this gene from all six labs. In particular, the weight for this gene in lab 2 was reduced from 88.5 in the fixed effects model to 1.8 in the random effects model. For comparison, the weights for this gene in the other five labs ranged from 3.0 to 11.9 in the fixed effects model and from 1.2 to 1.6 in the random effects model. Whereas the fixed effects model declared this gene significantly differentially expressed, the random effects model did not (controlling the FDR at 0.05 in both models). Thus, the random effects model was not overly influenced by any single lab's SLR estimate.
Comparing simulated results and "truth"
When the FDR was controlled at 0.05, 72 of the 1322 genes were declared by the random effects meta-analysis to be significantly differentially expressed based on the results of the test of significance . Individually, the six labs identified between 44 and 58 significantly differentially expressed genes (controlling the FDR at 0.05 for each lab) (Table 1). For each lab, most of its significant genes were declared significant by both the fixed effects and random effects meta-analyses.
These results demonstrate how a meta-analysis handles discrepancies between labs. A meta-analysis can be useful in finding genes that are statistically significantly differentially expressed and not just declared significant by one or more labs due to random variation between labs. For example, lab 1 declared 46 genes significant and lab 2 declared 49 genes significant, but these two labs declared only 33 of the same genes significant (Table 1). These 33 are not necessarily the most significant in either lab. That is, the 33 are not necessarily the genes with the smallest lab 1 P-values or smallest lab 2 P-values, but are those genes with the smallest P-values from both labs. Alternatively, rather than considering all genes declared significant by any of the labs, the random effects meta-analysis combines information across all six labs in a well-structured manner and declares 72 genes significantly differentially expressed.
While the numbers of correctly identified differentially expressed genes do not vary drastically between the individual labs (Table 1), the meta-analyses tend to correctly identify a higher number of differentially expressed genes. A comparison of the results from this SLR-based meta-analysis with the results from a previously-proposed meta-analysis approach based on combining P-values [2] is also summarized in Table 1. A slight modification to this P-value approach was necessary to account for differences in experimental design. Where the previous approach implicitly required multiple control and experimental sample arrays from each lab, this simulation data (as well as the real data presented subsequently in this work) did not satisfy this requirement in all labs. To modify the previous approach for the current data, probe-specific perfect match (PM) intensity differences between the experimental and control conditions were used to obtain a paired t statistic. Then the same permutation approach [2] was used to obtain a significance P-value for each gene in each laboratory based on this paired t statistic, and these P-values were combined across laboratories as previously proposed. In general, the SLR-based approach presented here tended to result in more genes found significantly differentially expressed by the meta-analysis than this previously proposed P-value approach (Table 1). In addition, the P-value approach does not provide a final quantitative estimate of the degree of differential expression for each gene, as does the currently proposed SLR-based approach. The meta-analysis SLR estimates tend to be much closer to the true SLR values than do the estimates from individual labs (Figure 5).
Integration-Driven Discovery (IDD)
One of the benefits of a meta-analysis is also one of the benefits of pooling raw data, that is the increased power to detect significant differences. It is possible that while a given gene is not declared significantly differentially expressed by any one lab, the combination of results across labs in a meta-analysis provides sufficient evidence to declare significant differential expression. Choi et al. [3] use the term "Integration-Driven Discovery" (IDD) to refer to a gene identified as differentially expressed by the results of a meta-analysis, but not identified as differentially expressed by any of the individual studies or labs. In this case, the term "integration" is used in the unification sense rather than the mathematical, since the results of several different studies are being integrated into a single meta-analysis.
As shown in Table 2, our particular simulation study produced 21 IDD's (i.e., 21 of the 72 genes declared significant by the random effects meta-analysis were not declared significant by any of the six labs). Of these 21 IDD's, 6 were truly differentially expressed; that is, our simulation study produced 6 true IDD's and 15 false IDD's. An examination of the SLR estimates for these IDD genes (Figure 4c) indicated that IDD's will tend to occur when 'small but consistent' [3] effect size estimates are combined. In addition, high variability of each lab's estimate may cause individual labs to not declare a gene significant while the meta-analysis estimate will have a lower variance, making a declaration of significance more likely.
Integration-Driven Revision (IDR)
In the simulation results presented here, there were 4 genes declared significant by at least two of the simulated labs that were not declared significant by the random effects meta-analysis (Table 2). A closer examination of the SLR estimates for these particular genes (Figure 4d) revealed that while at least two of the labs individually declared the gene significant, the SLR estimates between the six labs differed sufficiently to make the variance of the meta-analysis SLR estimate large. This increased variance of the meta-analysis SLR estimate caused the meta-analysis to declare these genes not significant. In addition, some labs' variability estimates may be artificially low due to chance, thus forcing a false declaration of differential expression at the individual lab level. The random effects meta-analysis is able to account for this possibility by down-weighting overly influential results.
We introduce the term "Integration-Driven Revision" (IDR) to describe a gene identified as differentially expressed by multiple studies or labs, but determined by the results of a meta-analysis to be not differentially expressed. While multiple laboratories might promote such a gene for further study because of its large and significant effect size, the meta-analysis would conclude that, due to the inconsistencies in effect sizes across labs, the gene is not significantly differentially expressed. Whereas Integration-Driven Discoveries (IDD's) will tend to occur when 'small but consistent' [3] effect size estimates are combined, Integration-Driven Revisions (IDR's) will tend to occur when large but inconsistent effect size estimates are combined. Of the 4 IDR's made in this simulated study, 3 were not truly differentially expressed; that is, our simulation study made 3 true IDR's and 1 false IDR's.
As noted previously, the simulation parameters can be adjusted to introduce varying degrees of difference between experiments. Increased inter-laboratory variability, or greater inconsistency among effect size estimates, will tend to affect the numbers of IDD's and IDR's made by the meta-analysis. Because IDD's occur when effect size estimates are small but consistent and IDR's occur when effect size estimates are large but inconsistent, greater inter-laboratory variability will tend to result in fewer IDD's and more IDR's being made.
Real data example
Several laboratories have investigated the genetic basis for EAE (experimental autoimmune encephalomyelitis, the mouse model for human multiple sclerosis) by use of Affymetrix technology, and have reported their findings in published papers [25-29]. In each of these published papers, mention is made of appropriate care of the mice following the ethics guidelines at the respective institutions. The three laboratories providing data are Offner [26,29], Carmody [28], and Ibrahim [25,27]. Each laboratory measured gene expression in a base (naive or control) sample and in an experimental (EAE-induced) sample, with some laboratories using multiple experiments. For the current purposes, an "experiment" is a single array-to-array comparison to study differential expression. Seven total experiments from the three laboratories are summarized in Table 3. While not all labs used the same measure of differential expression in their publications, here the Affymetrix MAS 5.0 algorithm [11] provides the SLR estimates for each lab from the respective raw data sets.
The use of different Affymetrix chip versions presents a non-trivial challenge in comparing and combining results across laboratories. The same gene may be represented on two different chip versions, and yet the names reported by the two chips may differ. Also, different sets of probes may represent the same gene on different chip versions, resulting in different probe set names on different chip versions. For example, the gene 1200011I18Rik on chip Mu11KsubA is identified by Probe Set ID AA000151_at, while on chip MG_U74Av2 it is Probe Set ID 104759_at. Furthermore, different chip versions may have different sets of genes represented on them. In order to combine the results across labs (and consequently, across chip versions), each gene must have a "name" recognized by all chips in the meta-analysis.
Previous meta-analyses of microarray results ([2,6], for example) have relied on Unigene cluster numbers to essentially achieve a uniform gene naming scheme across chip versions and platform types. Other recent work [30] proposed combining raw data from common probes into new probesets based on Unigene clusters. Because the focus of the current work is on combining the results of the Affymetrix algorithms, SLR estimates corresponding to the same Unigene cluster numbers are combined across all experiments. This approach will allow a gene to have multiple SLR estimates (corresponding to different original probe set names) from the same experiment. The Unigene number corresponding to each probe set on an array is available through the NetAffx feature [31] of the Affymetrix website [8].
There were considerable differences in the SLR estimates from the different labs, as represented in Figure 6. Some of these differences may be due to the use of different mouse strains, tissue types, and chip versions in the different laboratories. However, even experiments from the same laboratory tended to show disagreement, highlighting the need for biological replicates to provide more precise estimates of the degree of differential expression for each gene. This inter-laboratory variability also illustrates the need for methods to systematically combine results across laboratories. The fixed effects and random effects meta-analysis models were employed to combine these estimates across all laboratories for each Unigene number.
Similar to Table 1 for the simulated example, Table 4 summarizes the overlap in the numbers of genes declared significantly differentially expressed by each experiment in this observed data example. Based on the results of the random effects model, 3,671 genes are identified as statistically significantly differentially expressed. There were 12,775 unique Unigene numbers represented by the genes across all arrays in this meta-analysis, so approximately 28.7% of the genes represented were declared significantly differentially expressed. This may support the prediction made by Carmody et al. [28] that about 28.9% of the genes in the mouse genome 'may be relevant for autoimmune inflammation', or affected by EAE. Similar to the simulated data (Table 1), a comparison of the results from this SLR-based meta-analysis with the results from an implementation of the previously proposed P-value approach [2] indicates that even with real data, the SLR-based approach tends to identify more genes as significantly differentially expressed (Table 4).
As in the simulated data, the meta-analysis of these observed data produced Integration-Driven Discoveries (IDD's) and Integration-Driven Revision (IDR's). Similar to Table 2 for the simulated data, Table 5 summarizes these findings for the observed data. There were 65 IDD's and 5518 IDR's made. Of the IDR's, 32 were made for genes declared significantly differentially expressed by all seven experiments but not by the random effects meta-analysis. Figure 7 presents bubble plots for representative IDD and IDR genes from the observed data, similar to Figure 4 for the simulated data. As in the simulated data, IDD's tend to occur when small but consistent effect sizes are combined, and IDR's tend to occur when large but inconsistent effect sizes are combined. (See Additional file 1 : EAE.Random.Effects.Results.csv for the final estimates for all 12,775 genes.)
Discussion
Before any clinical decision is made based on the results of a meta-analysis, a biological validation of the results should be performed. Microarray technology is well-suited for hypothesis generation, and a meta-analysis can be used to effectively combine results across multiple laboratories to refine the list of candidate genes deserving biological validation. This approach will tend to yield more informative results when each lab has used biological and technical replicates in their experimental design [32]. The use of replicates at the laboratory level provides both added power to detect differential expression and more precise estimation of the true degree of differential expression for each gene under consideration.
The model used to generate data for the simulation example can be adjusted to account for various sources of variation and relationships between genes. It is of great interest to investigate how such relationships affect the outcome of a meta-analytic approach. An extension of the fixed effects and random effects models to the hierarchical Bayes approach is also being investigated with the hope of improving the meta-analysis approach as applied to microarrays and to incorporate prior knowledge. Included in this extension is the use of covariates in the meta-analysis framework to account for known differences between labs and the appropriate modeling of possible dependence among effect size estimates. We feel the use of covariates will provide insight into the effects of different labs, tissues, and microarray platforms on the observed differential expressions of genes. Separating out the effects of these covariates will facilitate the identification of those genes which are differentially expressed between two conditions rather than appearing differentially expressed due to external influences such as lab, tissue, or platform. For example, the differences observed between experiments from the same laboratory (Table 4) may be explained by differences in mouse strain or tissue sample, and the inclusion of covariate information in the model would adjust for this.
In the examples presented here, all studies involved used the Affymetrix platform and the data were summarized using the same normalization strategy with the MAS 5.0 algorithm [11]. When multiple studies have employed different platforms (such as cDNA and other oligonucleotide arrays) or normalization strategies, then some adjustments to the approach presented here will be necessary. In particular, a readily-available quantitative measure of differential expression common to all platforms involved is needed. In addition, it will be of great interest to consider the effect of a platform covariate in the extended meta-analysis model.
Although we have demonstrated our approach using both simulated rat data and real observed data from essentially genetically homogeneous mice, its utility with human data is of great interest. Along with the increased variability in human data comes an increase in the information about each individual subject and subpopulation. Therefore the incorporation of such covariate information is an important subject of our future work. We anticipate that the use of covariate information with human data will be particularly informative in identifying biologically significant subpopulations - for example, in identifying genes that are related to a disease in one subpopulation but not in another.
Conclusion
The signal log ratio (SLR), automatically reported by MAS 5.0 [11], is naturally suited to serve as an effect size estimate in a meta-analysis of results from multiple laboratories. In order to perform a meta-analysis of microarray results as presented here, the following components are needed for each probe set from each experiment: the corresponding Unigene ID, the SLR estimate, and the estimated variance of the SLR estimate. The random effects meta-analysis model is better suited than the fixed effects model for the analysis of microarray results because of the lack of homogeneity of effects from different laboratories. Genes not declared significantly differentially expressed by any single lab but then declared significantly differentially expressed by the meta-analysis are referred to as Integration-Driven Discoveries, or IDD's [3]. In addition to the identification of IDD's, our meta-analysis method identified genes declared significantly differentially expressed by multiple (and possibly all) laboratories but not significantly differentially expressed by the meta-analysis. These genes are referred to as Integration-Driven Revisions, or IDR's. The simulation example demonstrated how the final SLR estimates from the meta-analysis models tend to be much closer to the "true" SLR values than do the SLR estimates from any single lab. These meta-analytic approaches to microarray results provide a systematic method to combine results from different laboratories with the purpose of gaining clearer insight into the true degree of differential expression for each gene.
Authors' contributions
For this research RWD initiated the underlying concept of meta-analysis as applied to microarray technology, and coordinated the focus. RWD is the Ph.D. advisor of JRS. JRS developed and evaluated the simulation model, wrote the R code for the respective analyses and graphical displays, and drafted the manuscript. Both authors read and approved the final manuscript.
Supplementary Material
Additional File 1
EAE.Random.Effects.Results.csv Summary of random effects meta-analysis model results for all 12,775 genes, including SLR estimates, declaration of statistically significant differential expression, and status as an Integration-Driven Discovery (IDD) or Integration-Driven Revision (IDR). Note that the Gene column refers to Unigene cluster number unless the entry has an extension such as _at, in which case it refers to an Affymetrix probe set name for which no Unigene number was available. This file is comma-delimited and can be opened in Microsoft Excel.
Click here for file
Acknowledgements
We thank Drs. Robert Meisel (Purdue University) and Paul Mermelstein (University of Minnesota) for use of their RN_U34 Affymetrix data that provided our simulation parameters. We also thank Drs. Halina Offner (Oregon Health Sciences University), Ruaidhrí J. Carmody (University of Pennsylvania School of Medicine), and Saleh M. Ibrahim (University of Rostock), as well as their colleagues, for providing access to their raw Affymetrix data. We also thank two anonymous reviewers for their helpful suggestions to improve this work.
Figures and Tables
Figure 1 Summary of real and simulated RN_U34 Affymetrix chip data. (a) The real mismatch (MM) intensities are from a RN_U34 Affymetrix chip, and their histogram closely resembles a Gamma distribution with a long tail. (b) The simulated MM intensities are drawn from a Gamma distribution to resemble the real MM intensities. (c) The relationship between the SLR from a simulated lab and the standard error of the SLR for each gene on the RN_U34 Affymetrix chip. Large absolute SLR's with small standard errors indicate significant differential expression. (d) A comparison of SLR estimates from two simulated labs shows general agreement, with some differences between labs.
Figure 2 Fixed effects meta-analysis of simulated results. (a) The abundance of smaller homogeneity P-values indicates widespread violation of the homogeneity assumption. (b) The plot of homogeneity P-values versus fixed effects SLR estimates shows that this lack of homogeneity exists across a large range of SLR estimates.
Figure 3 Random effects meta-analysis of simulated results. (a) A comparison of the fixed effects and random effects meta-analysis estimates of SLR shows general agreement between the models. (b) The histogram of significance P-values shows an abundance of significantly differentially expressed genes, as evidenced by the large number of smaller significance P-values. (c) The smallest significance P-values tended to occur for genes whose random effects meta-analysis SLR estimates were large in absolute value. The reference line is the P-value cut-off used to control the False Discovery Rate at 0.05. Using this cut-off, the random effects meta-analysis declared 72 of the 1322 genes significantly differentially expressed. (d) Demonstration of how the SLR's estimated from the meta-analysis relate to their standard errors.
Figure 4 Bubble plots from the simulation example. Bubble area is proportional to weights used in the meta-analysis. Dashed lines represent the 95 percent confidence interval for the true value of the SLR, adjusted to control the FDR for all 1322 genes at 0.05. The red dotted line represents the true SLR value. The green bubbles represent labs which claimed significant differential expression for the gene. When zero lies outside the confidence interval, the meta-analysis declares the gene significantly differentially expressed. (a) SLR estimates from the six labs for a particular gene, with fixed effects weights. (b) Plot for the same gene as in (a), but with random effects weights. (c) Plot for one of the twenty-one Integration-Driven Discovery (IDD) genes declared significant by none of the six simulated labs but significant by the random effects meta-analysis. (d) Plot for one of the four Integration-Driven Revision (IDR) genes declared significant by multiple labs but not significant by the random effects meta-analysis.
Figure 5 Comparison of SLR estimates with true values from simulation example. Green squares represent type I errors, genes incorrectly claimed as differentially expressed. Red triangles represent type II errors, genes incorrectly claimed to be not significantly differentially expressed. The SLR estimates from the random effects meta-analysis tend to approximate the true values much better than does any single lab.
Figure 6 Comparison of SLR estimates from two experiments in the EAE example. This plot illustrates the variation between experiments and the consequent need for a method of systematically combining results from different experiments.
Figure 7 Bubble plots from the EAE example. Bubble area is proportional to weights used in the meta-analysis. Dashed lines represent the 95 percent confidence interval for the true value of the SLR, adjusted to control the FDR for all 1322 genes at 0.05. The green bubbles represent experiments which claimed significant differential expression for the gene. When zero lies outside the confidence interval, the meta-analysis declares the gene significantly differentially expressed. (a) Bubble plot for one of the sixty-five Integration-Driven Discovery (IDD) genes declared significant by none of the seven observed experiments but significant by the random effects meta-analysis. (b) Bubble plot for one of the thirty-two Integration-Driven Revision (IDR) genes declared significant by all seven observed experiments but not significant by the random effects meta-analysis.
Table 1 Comparison of results from the simulated data. Comparison of numbers of genes in common declared significant (i.e., significantly differentially expressed) by simulated labs 1 through 6, the SLR-based fixed effexts and random effects meta-analyses, and the previously proposed P-value-based meta-analysis [2]. The (i, j)th element of this table is the number of genes declared significant by both lab i and lab j, with F here representing the fixed effects meta-analysis, R the random effects meta-analysis, and T the "truth" behind the simulation model. P represents the meta-analysis based on P-values [2]. Each lab (and meta-analysis) had the False Discovery Rate (FDR) controlled at 0.05.
Simulated Lab Meta-Analysis
1 2 3 4 5 6 Fixed Random P-value Truth
1 46 33 34 34 31 31 43 35 33 35
2 49 34 37 31 34 47 39 35 38
3 54 34 32 36 44 41 38 41
4 51 30 35 48 38 35 39
5 44 32 39 37 34 37
6 58 48 40 37 41
F 137 72 45 58
R 72 45 56
P 45 45
T 70
Table 2 Summary of results for the simulated data. Numbers of genes declared significant (i.e., significantly differentially expressed) by different numbers of labs and the fixed and random effects meta-analyses in the simulation example. The False Discovery Rate (FDR) for the meta-analyses and each lab separately was controlled at 0.05. There were 21 Integration-Driven Discoveries (IDD's) and 4 Integration-Driven Revisions (IDR's).
Fixed Effects Model Random Effects Model
Num. of Labs Declaring Number of Genes Declared Number of Genes Declared
Significance Not Significant Significant Not Significant Significant
0 1152 51 1182 21
1 33 38 64 7
2 3 8 4 4
3 0 4 0 4
4 0 3 0 3
5 0 7 0 7
6 0 26 0 26
1188 137 1250 72
Table 3 Summary of observed data for the EAE example. A summary of the seven observed experiments involving EAE data to be combined in a meta-analysis.
Lab Experiment ID Base Sample Experimental Sample Chip Version
Offner Of.1 Of 1. Naive Of1.EAE1 MG_U74Av2
Offner Of.2 Of2.Naivel Of2.EAEl MG_U74Av2
Offner Of.3 Of2.Naive2 Of2.EAE2 MG_U74Av2
Carmody Ca.1 Ca.Naive1 Ca.Acute1 MG_U74A
Carmody Ca.2 Ca.Naive2 Ca.Acute2 MG_U74A
Ibrahim Ib.A Ib.ControlA Ib.PeakA Mu11KsubA
Ibrahim Ib.B Ib.ControlB Ib.PeakB Mu11KsubB
Table 4 Comparison of results from the EAE example. Comparison of numbers of genes in common declared significant (i.e., significantly differentially expressed) by the observed experiments, the SLR-based fixed effects and random effects meta-analyses, and the previously proposed P-value-based meta-analysis [2]. Each experiment (and meta-analysis) had the False Discovery Rate (FDR) controlled at 0.05.
Observed Experiment ID Meta-Analysis
Ib.A Ib.B Ca.1 Ca.2 Of.1 Of.2 Of.3 Fixed Random P-value
Ib.A 2952 402 1327 1253 1474 1354 1349 2336 1216 265
Ib.B 2902 996 950 1093 1067 1065 2456 1546 236
Ca.1 4471 2834 2797 2476 2461 3548 1555 402
Ca.2 4165 2646 2301 2324 3243 1464 333
Of.1 5001 2763 2807 3834 1578 355
Of.2 4911 3035 3669 1344 335
Of.3 5041 3728 1289 305
F 8263 3623 388
R 3671 205
P 453
Table 5 Summary of results from the EAE example. Numbers of genes declared significantly differentially expressed by different numbers of experiments and the fixed and random effects meta-analyses in the observed data example. The False Discovery Rate (FDR) for the meta-analyses and each experiment separately was controlled at 0.05. There were 65 Integration-Driven Discoveries (IDD's) and 5518 Integration-Driven Revisions (IDR's). There were 32 IDR's that had been declared significantly differentially expressed by all seven experiments.
Fixed Effects Model Random Effects Model
Number of Experiments Declaring Number of Genes Declared Number of Genes Declared
Significance Not Significant Significant Not Significant Significant
0 1792 301 2028 65
1 804 2265 1558 1511
2 749 1409 1869 289
3 625 1512 1662 475
4 328 1319 1084 563
5 152 908 617 443
6 54 464 254 264
7 8 85 32 61
4512 8263 9104 3671
==== Refs
Glass GV Primary, Secondary, and Meta-Analysis of Research Educational Research 1976 5 3 8
Rhodes DR Barrette TR Rubin MA Ghosh D Chinnaiyan AM Meta-Analysis of Microarrays: Interstudy Validation of Gene Expression Profiles Reveals Pathway Dysregulation in Prostate Cancer Cancer Research 2002 62 4427 4433 12154050
Choi JK Yu U Kim S Yoo OJ Combining Multiple Microarray Studies and Modeling Interstudy Variation Bioinformatics 2003 19 i84 i90 12855442 10.1093/bioinformatics/btg1010
Moreau Y Aerts S Moor BD Strooper BD Dabrowski M Comparison and Meta-Analysis of Microarray Data: From the Bench to the Computer Desk Trends in Genetics 2003 19 570 577 14550631 10.1016/j.tig.2003.08.006
Parmigiani G Garrett-Mayer ES Anbazhagan R Gabrielson E A Cross-Study Comparison of Gene Expression Studies for the Molecular Classification of Lung Cancer Clinical Cancer Research 2004 10 2922 2927 15131026
Rhodes DR Yu J Shanker K Deshpande N Varambally R Ghosh D Barrette T Pandey A Chinnaiyan AM Large-scale meta-analysis of cancer microarray data identifies common transcriptional profiles of neoplastic transformation and progression Proceedings of the National Academy of Sciences 2004 101 9309 9314 10.1073/pnas.0401994101
Hedges LV Olkin I Statistical Methods for Meta-Analysis 1985 Academic Press, San Diego, CA
Affymetrix
Affymetrix Statistical Algorithms Description Document 2002 Affymetrix, Santa Clara, CA
Hoaglin DC Mosteller F Tukey J Understanding Robust and Exploratory Data Analysis 1983 John Wiley and Sons, New York
Affymetrix Affymetrix Microarray Suite User's Guide Version 50 2001 Affymetrix, Santa Clara, CA
Li C Wong WH Parmigiani G, Garrett ES, Irizarry RA, Zeger SL DNA-Chip Analyzer (dChip) The Analysis of Gene Expression Data: Methods and Software 2003 Springer, NY
Irizarry RA Bolstad BM Collin F Cope LM Hobbs B Speed TP Summaries of Affymetrix GeneChip probe level data Nucleic Acids Research 2003 31 el5 10.1093/nar/gng015
Cooper H Hedges LV The Handbook of Research Synthesis 1994 Russell Sage Foundation
Fisher R Statistical Methods for Research Workers 1941 8 Oliver and Boyd, Edinburgh, UK
Glass GV Integrating Findings: The Meta-Analysis of Research Review of Research in Education 1978 5 351 379
DuMouchel W Normand SL Stangl DK, Berry DA, Marcel Dekker Computer-modeling and Graphical Strategies for Meta-analysis Meta-Analysis in Medicine and Health Policy 2000 127 178
DerSimonian R Laird N Meta-Analysis in Clinical Trials Controlled Clinical Trials 1986 7 177 188 3802833 10.1016/0197-2456(86)90046-2
Casella G Berger RL Statistical Inference 1990 Duxbury Press, Belmont, CA
The Comprehensive R Archive Network
Gautier L Cope L Bolstad BM Irizarry RA affy – analysis of Affymetrix GeneChip data at the probe level Bioinformatics 2004 20 307 315 14960456 10.1093/bioinformatics/btg405
BioConductor: open source software for bioinformatics
Ihaka R Gentleman R A Language for Data Analysis and Graphics Journal of Computational and Graphical Statistics 1996 5 299 314
Benjamini Y Hochberg Y Controlling the False Discovery Rate: a Practical and Powerful Approach to Multiple Testing Journal of the Royal Statistical Society B 1995 57 289 300
Ibrahim SM Mix E Bottcher T Koczan D Gold R Rolfs A Thiesen HJ Gene expression profiling of the nervous system in murine experimental autoimmune encephalomyelitis Brain 2001 124 1927 1938 11571211 10.1093/brain/124.10.1927
Matejuk A Dwyer J Zamora A Vandenbark AA Offner H Evaluation of the Effects of 17β-Estradiol (17β-E2) on Gene Expression in Experimental Autoimmune Encephalomyelitis Using DNA Microarray Endocrinology 2002 143 313 319 11751623 10.1210/en.143.1.313
Mix E Pahnke J Ibrahim SM Gene-Expression Profiling of Experimental Autoimmune Enchephalomyelitis Neurochemical Research 2002 27 1157 1163 12462414 10.1023/A:1020925425780
Carmody RJ Hilliard B Maguschak K Chodosh LA Chen YH Genomic scale profiling of autoimmune inflammation in the central nervous system: the nervous response to inflammation Journal of Neuroimmunology 2002 133 95 107 12446012 10.1016/S0165-5728(02)00366-1
Matejuk A Dwyer J Hopke C Vandenbark AA Offner H 17β-Estradiol Treatment Profoundly Down-Regulates Gene Expression in Spinal Cord Tissue in Mice Protected from Experimental Autoimmune Encephalomyelitis Archivum Immunologiae et Therapiae Experimentalis 2003 51 185 193 12894873
Morris JS Yin G Baggerly K Wu C Zhang L Identification of Prognostic Genes, Combining Information Across Different Institutions and Oligonucleotide Arrays Critical Assessment of Microarray Data Analysis (CAMDA) 2003 Conference Paper 2003
Liu G Loraine AE Shigeta R Cline M Cheng J Valmeekam V Sun S Kulp D Siani-Rose MA NetAffx: Affymetrix probesets and annotations Nucleic Acids Research 2003 31 82 86 12519953 10.1093/nar/gkg121
Parmigiani G Garrett ES Irizarry RA Zeger SL Parmigiani G, Garrett ES, Irizarry RA, Zeger SL The Analysis of Gene Expression Data: An Overview of Methods and Software The Analysis of Gene Expression Data: Methods and Software 2003 Springer, NY
|
15774008
|
PMC1274254
|
CC BY
|
2021-01-04 16:02:51
|
no
|
BMC Bioinformatics. 2005 Mar 17; 6:57
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-57
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-581577400210.1186/1471-2105-6-58Methodology ArticleTowards precise classification of cancers based on robust gene functional expression profiles Guo Zheng [email protected] Tianwen [email protected] Xia [email protected] Qi [email protected] Jianzhen [email protected] Hui [email protected] Jing [email protected] Haiyun [email protected] Chenguang [email protected] Eric J [email protected] Qing [email protected] Shaoqi [email protected] Department of Computer Science, Harbin Institute of Technology, Harbin 150001, China2 Department of Bioinformatics, Harbin Medical University, Harbin 150086, China3 School of Biological Science and Technology, Tongji University, Shanghai, 200092, China4 Department of Molecular Cardiology and Department of Cardiovascular Medicine, the Cleveland Clinic Foundation, 9500 Euclid Avenue, Cleveland, Ohio 44195, USA2005 17 3 2005 6 58 58 4 10 2004 17 3 2005 Copyright © 2005 Guo et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Development of robust and efficient methods for analyzing and interpreting high dimension gene expression profiles continues to be a focus in computational biology. The accumulated experiment evidence supports the assumption that genes express and perform their functions in modular fashions in cells. Therefore, there is an open space for development of the timely and relevant computational algorithms that use robust functional expression profiles towards precise classification of complex human diseases at the modular level.
Results
Inspired by the insight that genes act as a module to carry out a highly integrated cellular function, we thus define a low dimension functional expression profile for data reduction. After annotating each individual gene to functional categories defined in a proper gene function classification system such as Gene Ontology applied in this study, we identify those functional categories enriched with differentially expressed genes. For each functional category or functional module, we compute a summary measure (s) for the raw expression values of the annotated genes to capture the overall activity level of the module. In this way, we can treat the gene expressions within a functional module as an integrative data point to replace the multiple values of individual genes. We compare the classification performance of decision trees based on functional expression profiles with the conventional gene expression profiles using four publicly available datasets, which indicates that precise classification of tumour types and improved interpretation can be achieved with the reduced functional expression profiles.
Conclusion
This modular approach is demonstrated to be a powerful alternative approach to analyzing high dimension microarray data and is robust to high measurement noise and intrinsic biological variance inherent in microarray data. Furthermore, efficient integration with current biological knowledge has facilitated the interpretation of the underlying molecular mechanisms for complex human diseases at the modular level.
==== Body
Background
Gene expression profile (GEP) has been widely used to address the relationship between disease phenotypes and the cellular expression patterns. Numerous data mining methods have been proposed for precise classification of disease phenotypes (subtypes) using high dimension GEPs [1-5]. Although much progress in applying microarray technology to versatile biological kingdoms has been witnessed in recent time, further advancing its efficiency and power in elucidating complex biological mechanisms would very likely rely on our ability to handle the high dimension genetic information mixed with measurement noises [6,7], intrinsic biological variance [8,9], and a large number of irrelevant genes [10,11]. However, lack of coherence in biological interpretations often occurring in analysis of gene expression profiling can be remedied partially by integrating with a knowledge-mining tool such as Onto-Express developed by Draghici et al. [12,13].
Cellular biology is essentially to study an interacting network of various functional gene modules that coordinately carry out highly integrated cellular functions in somewhat isolated fashions [14-16]. The assumption that genes express and perform their functions in modular fashions in cells has been supported by accumulated multiple lines of evidence from, among others, gene expression and protein-protein interaction studies [17-19]. Inspired by the insight that genes often interplay as a module to realize a highly integrated cellular function, we propose an alternative approach to analyzing the high dimension microarray data by formulating the disease classification problem from a perspective of modularity. In this study, we map genes to their categories in Gene Ontology (GO) [20,21], which provides a unified gene function classification system across genomes. After annotating each individual gene to a GO functional category, we identify gene functional categories enriched with differentially expressed genes. These categories, defined as differentially expressed functional modules, are very likely to be relevant with experimental conditions, or specifically, with the disease type discrimination. For each functional module, we construct a representative functional feature, and then employ a traditional data mining toolbox to train the rule(s) for classifying disease types based on the newly built functional expression profiles (FEPs). Instead of analyzing raw expressions of single genes, we consider the gene expressions within a functional module as an integrative data point to shrink the feature dimension. This modular approach is flexible and also statistically robust to high measurement noise and intrinsic biological variance inherent in microarray data. Furthermore, efficient integration with current biological knowledge support provided in the GO database has facilitated the interpretation of the underlying molecular mechanisms at the modular level. We apply the alternative approach to analyze four publicly available datasets to demonstrate its performance and statistical properties. The results from analysis of two datasets (NCI60 dataset [1] and the lymphoma dataset [2]) are described in the main text. To obtain a robust and convincing comparison of FEP and GEP, we have undertaken analysis of two additional large-scale datasets and have described the detailed results in the supplement [see Additional file 1].
Results
Description of the two datasets
NCI60 dataset consists of 9,703 cDNAs whose expression levels were measured in 60 cancer cell lines derived from a variety of tissues and organs. A subset of NCI60 (31 samples of four cancer types) is used in this study, including 8 samples from renal cancer (RE), 7 samples from colon cancer (CO), 8 samples from leukaemia (LE), 8 samples from melanoma (ME), respectively. The same criterion as in [1] is used to identify the differentially expressed genes (i.e., log2 (ratio) > 2.8 or log2 (ratio) < -2.8 in at least four cell lines). A total of 1160 genes are filtered. The lymphoma dataset contains the expression profiles of 18,000 cDNAs for 42 samples of the diffuse large B-cell lymphoma (DLBCL), 9 samples from follicular lymphoma (FL), 11 samples from chronic lymphocyte leukaemia (CLL) and 24 samples from the healthy sources (NORMAL) prepared from activated human blood B cells and resting/activated blood T cells, respectively. The 4,026 genes, originally filtered by Alizadeh et al. [2], is used in this study. As suggested by Alizadeh et al. [2], we exclude 9 samples (eight NORMAL samples and one DLBCL sample). Using the criterion "log2 (ratio) > 1 or log2 (ratio) < -1 in at least eight cell lines" [2], we identify a total of 705 differentially expressed genes.
FEP based analysis of NCI60 dataset
Based on NCI60 dataset, 114 differentially expressed GO modules are identified according to a statistical test described in the Methods section and their functional expression profiles, a 114 × 31 matrix, are denoted with FEP114A or FEP114M when arithmetic mean or median is used to summarize the overall activity of a module, respectively. The 114 differentially expressed GO categories are annotated with a total of 617 differentially expressed genes. For comparison, we also perform classification analysis using the expression profiles of the 617 differentially expressed and annotated genes (GEP617) or the 1160 differentially expressed genes (GEP1160).
Recursive partition analysis of the 114 functional features using median as the summary measure identify three significant functional signatures for multiple cancer subtypes. Figure 1A depicts the decision tree trained on the FEP114M. The internal nodes of the tree are denoted with the functional modules from Gene Ontology. The leaf nodes give the classification results for different cancer types: the total number of samples contained over the number of the incorrectly predicted samples. Figure 1B depicts the functional expression profiles of the three modules (GO:0005923, GO:0007345 and GO:0009887). The three modules are annotated with 9, 41 and 148 genes (figure 1C), respectively. We identify 4, 11 and 35 genes that are differentially expressed between four cancer types, respectively. Hypergeometric tests indicate that all the three modules are significantly (or highly significantly, p-value < 0.01) enriched with differentially expressed genes, with the probability of observing a more extreme of 0.0150, 0.0322 and 0.0079, respectively.
One advantage for FEP based analysis is to allow us to interpret our findings at the modular level. Based on the trained tree, we observe that RE is distinct from the remaining cancer types and is characterized with the up-regulation of GO:0007345 (termed with embryogenesis and morphogenesis), suggesting that the abnormal up-regulation (possibly over-expression) of the genes that determine embryogenesis and morphogenesis may contribute to development of RE, too. To look for knowledge support, we search G2D database [22,23]. Interestingly, significant association of GO:0007345 with RE has been documented previously. PUBMED searching provides further evidence to support the trained hypothesis. Gene F2R (thrombin receptor), which is differentially expressed between the investigated cancers and is also annotated in the module, is pivotal in proliferation and motility of prostate cancer cells [24], colon cancer cells [25] and breast carcinoma cells [26]. We thus propose that GO:0007345 may be an important functional target for the molecular pathogenesis of RE. Further distinction between the remaining three cancers can be made by looking at the module GO:0009887, which acts for organogenesis and is down-regulated in ME, but is up-regulated in both LE and CO. G2D database indicates that GO:0009887 is indeed significantly associated with both LE and CO. By searching PUBMED, we find that CYP1B1 (a member of cytochrome P450 enzyme), a differentially expressed gene annotated in this module, was reported to be associated with high risk for developing several forms of cancers [27], which is again consistent with our finding. The third module, GO:0005923, contains a cluster of genes for tight junction and is identified for distinguishing between caners LE and CO. Its association with LE has been documented in G2D. In addition, experiment studies agree with our finding that three of the four differentially expressed genes (CLDN1, CLDN4 and CLDN5) annotated in the module are members of the claudin family, which were demonstrated to be related to the invasiveness and metastatic phenotype of pancreatic and colorectal cancers [28,29]. In short, the above biological knowledge mining supports our analysis. Intuitively, the functional expression patterns, as demonstrated in figure 1B, are clearly distinguishable between the four cancers. RE samples have the highest expressions in all three modules and ME samples have the lowest. Nevertheless, two outliers (RE:SN12C and ME:L0XIMVT) have marked deviations from their respective groups and thus not surprisingly they have been misclassified.
To provide an unbiased evaluation on the utility of the trained three modules for multi-class cancer diagnosis, we perform a five-fold cross validation procedure, as described in the Methods section, to verify the trained classifier in terms of accuracy, precision and recall. As shown in figure 2A, the classification accuracies for four gene expression measures (FEP114A, FEP114M, GEP617 and GEP1160) are 51.6%, 67.7%, 71.0% and 64.5%, respectively. Comparing the two summary measures, median (FEP114M) generally perform better than mean (FEP114A), evaluated in terms of the three criteria. Using the low dimension function profile and median measure, we have achieved comparable results to those using the high dimension gene expression profiles (GEP617 and GEP1160). However, this application implicates that there is a space for further improvement in multi-class cancer diagnosis using tumour gene expression signatures or functional signatures, perhaps by combining with the other contributed clinical risk factors and histopathological information, to some extent which has reflected the complex nature of cancers.
FEP based analysis of the lymphoma dataset
For the lymphoma dataset, 44 differentially expressed GO modules are identified and their functional expression profiles make up a 44 × 77 matrix, called FEP44A or FEP44M when arithmetic mean or median is used to be the summary measure, respectively. The 44 differentially expressed GO modules are annotated with a total of 383 differentially expressed genes. Again for comparison, we also perform classification analysis using the raw expression profiles of the 383 differentially expressed and annotated genes (GEP383) or the 705 differentially expressed genes (GEP705).
By a coincidence, we also identify three functional modules for distinguishing lymphoma subtypes. Figure 3A displays the decision tree trained on FEP44M of the lymphoma dataset. Figure 3B gives the expression patterns of the three functional modules (GO:0006875, GO:0009611 and GO:0019865) for 77 tissue samples. Over half of annotated genes in all the modules are differentially expressed between the tissue samples (figure 3C), i.e., 5, 28 and 4 out of 8, 49 and 5 genes measured in the three modules, respectively. Hypergeometric tests indicate that all the three modules are significantly (or highly significantly, p-value < 0.01) enriched with differentially expressed genes, with the probability of observing a more extreme of 0.0396, 0.0008 and 0.0205, respectively.
Because G2D database lacks the data for DLBCL and FL, we resort to PUBMED for knowledge support. The first identified module, GO:0006875 (metal ion homeostasis) is a parental category of GO:0006874 (calcium ion homeostasis). Three genes (Hs.241392, Hs.73817 and Hs.237356) in the small size module are differentially expressed between the lymphoma tissue types. Anghileri et al. [30] showed that calcium-overload can lead to proliferation and neoplastic transformation of lymphocytes in mice and suggested the involvement of the calcium homeostasis change in lymphoma induction. At the second layer of the trained tree, GO:0009611 (response to wounding) distinguishes DLBCL (up-regulated, clearly visible in figure 3B) from normal samples. One differentially expressed gene annotated in this GO module, BCL6 (zinc finger protein 51), was found to be frequently translocated and hypermutated in diffuse large-cell lymphoma (DLBCL), and it thus may be involved in the pathogenesis of DLBCL [31]. The functional module labelled immunoglobulin binding (GO:0019865) may be an important modular marker for separating the two lymphoma subtypes (FL and CLL). One differentially expressed gene annotated in this GO module, CD23 (Fc fragment of IgE, low affinity II), was identified to be associated with chronic lymphocytic leukaemia (CLL) [32], which is again consistent with our data. Median measure FEP44M achieves the highest accuracy (88.31%) for classification of lymphoma tissue types (figure 4A). Again, as we expected, median perform better than mean (FEP44A) in terms of accuracy, precision and recall. Of special note, FEP44M attains a nearly perfect precision or recall rate (98%) to distinguish DLBCL from others (figures 4B and 4C), implicating its utility to clinically isolate this particular lymphoma using the identified modular signatures.
We present in Additional file 1 the detailed numeric results for further comparison of different gene expression measures using four datasets (plus two additional large-scale datasets). In all the four datasets, FEPs have achieved comparable or better classification performance than those GEPs do. In short, we have provided convincing evidence to support FEP as a robust gene expression measure, as a useful summary index for efficient data reduction and as a way towards precise classification of biological phenotypes at the modular levels.
Discussion
With the rapid accumulation of gene functional knowledge, GO functional modules have been widely applied in inferring the unknown functions of genes based on their expression profiles (e.g. [33-35]), but there is an open space for development of the timely and relevant computational algorithms that use robust functional expression profiles towards precise classification of complex human diseases at the modular level. In this paper, we have presented an alternative approach to analyzing microarray data. The central idea is to transform the gene expressions to modular level to achieve both robustness and precise classification with better biological interpretation. We first map genes onto their functional modules according to GO hierarchy, and then consider the newly built modules as the features for learning. Because the modules are evaluated with a summary measure(s), its variance is considerably reduced. For this reason, function expression profiling is robust to measurement or biological noises, outliers and distribution assumptions underlying some approaches.
Recent time has witnessed the attempts to study human diseases at the modular levels. Hanczar et al. [36] grouped the whole set of genes with k-means clustering of the averaged expression values in each cluster and then trained a SVM classifier based on these integrated values. Huang et al. [37,38] chose a Bayesian formalism of singular value decompositions (SVD) coupled with binary regression and stochastic regularization. Our approach differs from these methods in at least three aspects. First, we construct a module based on the well established GO categories in order to achieve better biological interpretation. Second, we identify statistically significant modules enriched with differentially expressed genes to avoid inclusion of some noise modules. Third, we can easily procure biological knowledge (e.g., GO in this study) because of the very nature of the proposed methodology.
Traditional methods for reducing dimension of gene expression profiles are feature selection [39], for examples, wrappers, filters and embedded algorithms. However, if only an optimal gene subset is extracted, many genes of the same (or similar) function(s) would be excluded due to redundancy. We have thus proposed an ensemble approach to mining disease-relevant genes by constructing a gene forest [40,41]. Alternatively, one may consider analyzing the gene expression profiles at the modular level to avoid unnecessary loss of important information. The proposed modular approach can achieve both goals simultaneously: reducing the dimension of microarray data by transforming the single gene expressions into modular expressions and improving the interpretability on the data mining results. The trained functional modules can be presented graphically and are easily understood by biologists. In fact, a trained tree implicates a decision rule(s) that determine the interactions of modules and can be used to elucidate the complex cellular processes that lead to distinct biological types. Our approach could be regarded as a way of identifying disease-relevant functional modules (selected by decision trees) guided by precise classification of cancers. Compared with many tools (e.g. Onto-Express, FatiGO, GoMiner [42] and GOAL [43]) developed mainly for gene function annotation using the data acquired from microarrays or other high-throughput techniques, our approach focuses on identification of the modules of high disease discriminating power, thus implicating stronger evidence of their relevancy with the studied disease. However, caution should be taken in interpretation of the module selection for refinement of the biological phenotypes investigated, especially when normal controls are not included. In this case, the modules relevant to disease subtypes should be considered as important molecular signatures which may also be the disease-causing modules.
In the study, genes are annotated to the modular terms in GO as granted. Nevertheless, the classification system with modules hierarchically structured is neither the most efficient nor the optimal for pursuing specific biological tasks, for example, classifying cancer types using modular signatures. In the context of microarray experiments, a large number of cDNA sequences often remain not being annotated by GO because of either their unknown functions or ambiguous annotations. To extract maximal information from microarray data, one may consider performing computational function assignments of gene products using the strategy proposed by Vinayagam et al. [44]. Therefore, further investigations on an alternative classification system(s) or an extension of the GO system and choices of more efficient indices for functional expression profiling are warranted.
Conclusion
In summary, we have proposed an alternative approach to analyzing gene expression profiles at the modular levels, where the functional expression profiles replace the traditional gene expression profiles. We have applied the alternative approach to four large-scale microarray datasets, and have achieved comparable or better classification performance by using the functional expression profiles. It should be recognized that median or other modular measures are generally robust to noises because they are less sensitive to any single individual gene expression value. However, for the same reason, they are conservative in using full information of microarray experiments, so it cannot be vouched that FEP always has better performance than GEP does. Despite this fact, the improved biological interpretability and the advantages of robustness to measurement noise and intrinsic biological variance of gene expression data would promote its application in biomedical research.
Methods
Gene annotation and definition of the differentially expressed functional modules
In GO database, gene functional categories are tagged with functional terms and organized in three directed acyclic graphs where the root nodes are tagged with "biological process" (BP), "molecular function" (MF) and "cellular component" (CC), respectively. There are two kinds of relationships between a child category and its parent categories in GO: 'Is-a', where the child category is an instance of its parents, and 'Part-of', where the child category is a part of its parents. Up to the present, GO contains a total of over 17,000 categories (or called modules), with 8625 categories in the BP ontology, 1407 categories in the CC ontology, and 7336 categories in the MF ontology. All the information in GO can be downloaded in a relational database file format to local computers. With the existing gene function knowledge, known genes can be annotated to certain GO categories corresponding to their most specific function(s). As implied by the ontology structure, one gene annotated to a category is also within the ancestor categories on the same path.
During the annotation step, a gene can be annotated with multiple GO categories. Not all of these categories, however, are to be used in this study. Only the categories that contain significantly larger number of differentially expressed genes than by random are kept for the following analysis. As Khatri et al. [45] and Al-Shahrour at al. [46,47] did, we perform a statistical test to decide whether a GO category is significantly enriched with differentially expressed genes that are aroused (induced or repressed) by the experiment conditions. Suppose that a total of N genes (set A) for the analyzed data are annotated in GO in which a set of C genes (set B) are differentially expressed. For a given GO category, a gene is either in the category or not in the category. Suppose further that n genes out of set A and k genes out of set B are in the category. If the k differentially expressed genes are effectively a random sample uniformly selected from set B, the expected value of k is (n/N) C. As a gene can be selected only once, this is the sampling without replacement and can therefore be appropriately modelled by a hypergeometric distribution [45]. The probability of observing at least k differentially expressed genes in the GO category of n genes can be computed as follows:
The p-value calculated above corresponds to a one-sided test and a smaller p-value relates to a higher likelihood of a GO category's enrichment with differentially expressed genes. Only the categories that contain significantly (p ≤ 0.05) larger number of differentially expressed genes than that by random are kept for the following analysis. In this study, to avoid the possible loss of the true positives, we do not perform a multiple-tests correction for multiple GO categories evaluated. Therefore, the p-value quoted should be considered as a heuristic measure, useful as an indicator of roughly rating of the relative enrichment of differentially expressed genes for each GO category. We remove a redundant category if all the genes annotated to a category are also annotated to one of its child categories. In this case, we retain one of the child categories because its function(s) is more specifically defined. In the following text, we refer to a GO category significantly enriched with differentially expressed genes as a 'gene functional module', or a 'module' for short.
Construction of the functional expression profiles
After extracting the differentially expressed functional modules, we compute two summary measures: arithmetic mean and median (the 50% quantile) of all the gene expression values in each module to capture the activity of the module. The modular measure(s) can have multiple sources of variations including systematic experiment variation, treatment effects, chip variation and biological variation [6-9]. The distributions for individual (raw) gene expression are usually not known in prior and could be contaminated with outliers and possibly distorted due to heteroscedasticity [48]. Therefore, mean or median measure for the modular activity can be good remedy statistics for the location parameter. When the data are Gaussian or symmetrically distributed, sample mean has a higher statistical efficiency compared to sample median. If there are outliers, however, sample median is a robust measure for the modular activity [49].
Evaluation of the functional profiles based on a decision tree
Based on the functional expression profiles computed with the two summary measures, we can now apply a proper classification algorithm as do traditionally for the individual measures of gene expressions. In this study, we chose a decision tree model (e.g., C4.5 [50,51]) to train the classification rule. Since there often are only limited numbers of instances in microarray experiments, we adopt a k-fold (k = 5 in this study) cross-validation procedure. In the k-fold cross-validation, we divide the data into k subsets of approximately equal size. We train the classifier on the k-1 subset and use the remaining subset to test the performance of the classifiers. The performance for each classifier is evaluated in terms of three measures: accuracy, precision and recall rate, which are defined as follows:
where TP, TN, FP and FN denote true positive, true negative, false positive and false negative, respectively. Each sample in the test set can be categorized in one of the four outcomes. True positives are class members according to both the classifier and sample label (disease type). True negatives are non-members according to both. False positives are samples that the classifier places within the given class, but sample labels are non-members. False negatives are samples that the classifier places outside the class, but sample labels are members. Accuracy is a percentage quantity for the number of times that the classifier is correct in its classification and it conveys the right intuition when the positive and negative populations are roughly equal in size. Precision is the percentage of times that the classifier is correct in its classification of positive samples. Recall is the percentage of known positive samples that the classifier would classify as being positive.
Biological knowledge support
We apply G2D (Candidate Genes to Inherited Diseases) database [22,23] to associate a phenotype (disease) to a GO module trained using a decision tree. G2D database is built by text-mining approach. First, the number of papers in MEDLINE that contain a MeSH-C term (describing a phenotypic feature) and a MeSH-D term (describing a chemical object) are counted, and then the corresponding phenotypic term and the chemical term are judged as associated if they occur together in many abstracts. Second, a chemical term is judged as associated to a GO term corresponding to a functional module appearing on the decision tree if they appear to be related by many sequences from RefSeq. Third, if a phenotypic term is associated to a chemical term that has a functional counterpart, then the phenotypic term is associated to the functional term. We search PUBMED manually to get further supporting evidence. If one or several differentially expressed genes, which are annotated to one functional module in the decision tree, are suggested by existing literature to be functionally related to one disease type, the investigated functional module is then deemed to be functionally relevant to the disease type.
List of abbreviations used
Gene expression profile (GEP), function expression profile (FEP), Gene Ontology (GO), renal cancer (RE), colon cancer (CO), leukaemia (LE), melanoma (ME), diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL), chronic lymphocyte leukaemia (CLL), healthy sources (NORMAL), arithmetic mean (A), median (M), Candidate Genes to Inherited Diseases (G2D), biological process (BP), molecular function (MF), cellular component (CC).
Authors' contributions
This study was undertaken by a collaborative team of four institutes as indicated, led by ZG, TZ, XL and SR, who also conceived of the proposal of the study and drafted the manuscript. JX, QW (Qi Wang) and HY participated in writing the computing codes and applied the data mining strategy to the field datasets. JZ, HW, CW, EJT and QW (Qing Wang) implemented the search for biological knowledge support and provided constructive advice for the biological interpretation of the results. All authors participated in reading, approving and revising the manuscript.
Supplementary Material
Additional File 1
Further comparison of different gene expression measures for classification of biological phenotypes using four large-scale datasets
Click here for file
Acknowledgements
We would like to thank the reviewer and the editors for helps and suggestions. This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 30170515, 30370388 and 30370798), the Chinese 863 Program (Grant Nos. 2003AA2Z2051 and 2002AA2Z2052), the 211 Project, the Tenth 'Five-year' Plan, Harbin Medical University and the Heilongjiang Province Department of Education Outstanding Overseas Scientist grant (to SR).
Figures and Tables
Figure 1 Training classification rules for four cancer types based on functional expression profiles of 114 modules. A – Decision tree trained with the NCI60 FEP median measure. The internal nodes of the tree are denoted with the functional modules from Gene Ontology. The leaf nodes give the classification results for the cancer types. The numbers in the leaf nodes are the total number of samples contained over the number of the incorrectly predicted samples. B – Functional expression profiles of the three identified modules. For the identified GO modules from decision analysis, their functional expression profiles are demonstrated with a colouring spectrum of their medians. Each GO module corresponds to a row, and the column denotes the functional expression for each cell line. At the top are names of cell lines (renal cancer (RE), colon cancer (CO), leukaemia (LE), melanoma (ME)). Samples with a missing value or the null value are coded with black colour, a positive with red colour and a negative with green colour. C – numbers of genes annotated and differentially expressed in the three identified modules.
Figure 2 Comparison of different gene expression measures for classification of cancer types in terms of accuracy (A), precision (B) and recall (C).
Figure 3 Training classification rules for lymphoma subtypes based on functional expression profiles of 44 GO modules. A – Decision tree trained with the lymphoma FEP median measure. The internal nodes of the tree are denoted with the functional modules from Gene Ontology. The leaf nodes give the classification results for the lymphoma subtypes. The numbers in the leaf nodes are the total number of samples contained over the number of the incorrectly predicted samples. B – Functional expression profiles of the three identified modules. For the identified GO modules from decision analysis, their functional expression profiles are demonstrated with a colouring spectrum of their medians. Each GO module corresponds to a row, and the column denotes the functional expression for each cell line. At the top are names of cell lines (diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL), chronic lymphocyte leukaemia (CLL), and the healthy sources (NORMAL)). Samples with a missing value or the null value are coded with black colour, a positive with red colour and a negative with green colour. C – Numbers of genes annotated and differentially expressed in the three identified modules.
Figure 4 Comparison of different gene expression measures for classification of lymphoma tissues in terms of accuracy (A), precision (B) and recall (C).
==== Refs
Ross DT Scherf U Eisen MB Perou CM Rees C Spellman P Iyer V Jeffrey SS Van de Rijn M Waltham M Pergamenschikov A Lee JC Lashkari D Shalon D Myers TG Weinstein JN Botstein D Brown PO Systematic variation in gene expression patterns in human cancer cell lines Nat Genet 2000 24 227 235 10700174 10.1038/73432
Alizadeh AA Eisen MB Davis RE Ma C Lossos IS Rosenwald A Boldrick JC Sabet H Tran T Yu X Powell JI Yang L Marti GE Moore T Hudson JJ Lu L Lewis DB Tibshirani R Sherlock G Chan WC Greiner TC Weisenburger DD Armitage JO Warnke R Levy R Wilson W Grever MR Byrd JC Botstein D Brown PO Staudt LM Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling Nature 2000 403 503 511 10676951 10.1038/35000501
Antoniadis A Lambert-Lacroix S Leblanc F Effective dimension reduction methods for tumor classification using gene expression data Bioinformatics 2003 19 563 570 12651713 10.1093/bioinformatics/btg062
Zhang H Yu CY Singer B Xiong M Recursive partitioning for tumor classification with gene expression microarray data Proc Natl Acad Sci U S A 2001 98 6730 6735 11381113 10.1073/pnas.111153698
Li L Jiang W Li X Moser KL Guo Z Du L Wang Q Topol EJ Rao S A robust hybrid between genetic algorithm and support vector machine for extracting an optimal feature gene subset Genomics 2005 85 16 23 15607418 10.1016/j.ygeno.2004.09.007
Novak JP Sladek R Hudson TJ Characterization of variability in large-scale gene expression data: implications for study design Genomics 2002 79 104 113 11827463 10.1006/geno.2001.6675
Tu Y Stolovitzky G Klein U Quantitative noise analysis for gene expression microarray experiments Proc Natl Acad Sci U S A 2002 99 14031 14036 12388780 10.1073/pnas.222164199
Elowitz MB Levine AJ Siggia ED Swain PS Stochastic gene expression in a single cell Science 2002 297 1183 1186 12183631 10.1126/science.1070919
Swain PS Elowitz MB Siggia ED Intrinsic and extrinsic contributions to stochasticity in gene expression Proc Natl Acad Sci U S A 2002 99 12795 12800 12237400 10.1073/pnas.162041399
Herzel H Beule D Kielbasa S Korbel J Sers C Malik A Eickhoff H Lehrach H Schuchhardt J Extracting information from cDNA arrays Chaos 2001 11 98 107 12779445 10.1063/1.1336843
Xing EP Jordan MI Karp RM Feature selection for high-dimensional genomic microarray data 2001 (ICML2001). 601 608
Draghici S Khatri P Martins RP Ostermeier GC Krawetz SA Global functional profiling of gene expression Genomics 2003 81 98 104 12620386 10.1016/S0888-7543(02)00021-6
Draghici S Onto-Expression
Hartwell LH Hopfield JJ Leibler S Murray AW From molecular to modular cell biology Nature 1999 402 C47 52 10591225 10.1038/35011540
Barabasi AL Oltvai ZN Network biology: understanding the cell's functional organization Nat Rev Genet 2004 5 101 113 14735121 10.1038/nrg1272
Rives AW Galitski T Modular organization of cellular networks Proc Natl Acad Sci U S A 2003 100 1128 1133 12538875 10.1073/pnas.0237338100
D'Haeseleer P Liang S Somogyi R Genetic network inference: from co-expression clustering to reverse engineering Bioinformatics 2000 16 707 726 11099257 10.1093/bioinformatics/16.8.707
Azuaje FJ Bodenreider O Incorporating ontology-driven similarity knowledge into functional genomics: An exploratory study: In IEEE Forth Symposium on Bioformatics and Bioengineering (BIBE2004)May 19-21; Taichung, Taiwan. 2004
Bader GD Hogue CW Analyzing yeast protein-protein interaction data obtained from different sources Nat Biotechnol 2002 20 991 997 12355115 10.1038/nbt1002-991
Ashburner M Ball CA Blake JA Botstein D Butler H Cherry JM Davis AP Dolinski K Dwight SS Eppig JT Harris MA Hill DP Issel-Tarver L Kasarskis A Lewis S Matese JC Richardson JE Ringwald M Rubin GM Sherlock G Gene ontology: tool for the unification of biology. The Gene Ontology Consortium Nat Genet 2000 25 25 29 10802651 10.1038/75556
GO Gene Ontology Consortium
Perez-Iratxeta C Bork P Andrade MA Association of genes to genetically inherited diseases using data mining Nat Genet 2002 31 316 319 12006977
Perez-Iratxeta C Bork P Andrade MA G2D: Candidate Genes to Inherited Diseases
Liu J Bastian M Kohlschein P Schuff-Werner P Steiner M Expression of functional protease-activated receptor 1 in human prostate cancer cell lines Urol Res 2003 31 163 168 12883880 10.1007/s00240-003-0309-2
Darmoul D Gratio V Devaud H Lehy T Laburthe M Aberrant expression and activation of the thrombin receptor protease-activated receptor-1 induces cell proliferation and motility in human colon cancer cells Am J Pathol 2003 162 1503 1513 12707033
Yin YJ Salah Z Grisaru-Granovsky S Cohen I Even-Ram SC Maoz M Uziely B Peretz T Bar-Shavit R Human protease-activated receptor 1 expression in malignant epithelia: a role in invasiveness Arterioscler Thromb Vasc Biol 2003 23 940 944 12637343 10.1161/01.ATV.0000066878.27340.22
Agundez JA Cytochrome p450 gene polymorphism and cancer Curr Drug Metab 2004 5 211 224 15180491 10.2174/1389200043335621
Michl P Barth C Buchholz M Lerch MM Rolke M Holzmann KH Menke A Fensterer H Giehl K Lohr M Leder G Iwamura T Adler G Gress TM Claudin-4 expression decreases invasiveness and metastatic potential of pancreatic cancer Cancer Res 2003 63 6265 6271 14559813
Miwa N Furuse M Tsukita S Niikawa N Nakamura Y Furukawa Y Involvement of claudin-1 in the beta-catenin/Tcf signaling pathway and its frequent upregulation in human colorectal cancers Oncol Res 2000 12 469 476 11939410
Anghileri LJ Mayayo E Domingo JL Thouvenot P Cellular calcium homeostasis changes in lymphoma-induction by ATP iron complex Oncol Rep 2002 9 61 64 11748456
Pasqualucci L Migliazza A Basso K Houldsworth J Chaganti RS Dalla-Favera R Mutations of the BCL6 proto-oncogene disrupt its negative autoregulation in diffuse large B-cell lymphoma Blood 2003 101 2914 2923 12515714 10.1182/blood-2002-11-3387
Schwarzmeier JD Shehata M Hilgarth M Marschitz I Louda N Hubmann R Greil R The role of soluble CD23 in distinguishing stable and progressive forms of B-chronic lymphocytic leukemia Leuk Lymphoma 2002 43 549 554 12002758 10.1080/10428190210323
Brown MP Grundy WN Lin D Cristianini N Sugnet CW Furey TS Ares MJ Haussler D Knowledge-based analysis of microarray gene expression data by using support vector machines Proc Natl Acad Sci U S A 2000 97 262 267 10618406 10.1073/pnas.97.1.262
Lagreid A Hvidsten TR Midelfart H Komorowski J Sandvik AK Predicting gene ontology biological process from temporal gene expression patterns Genome Res 2003 13 965 979 12695321 10.1101/gr.1144503
Mateos A Dopazo J Jansen R Tu Y Gerstein M Stolovitzky G Systematic learning of gene functional classes from DNA array expression data by using multilayer perceptrons Genome Res 2002 12 1703 1715 12421757 10.1101/gr.192502
Hanczar B Courtine M Benis A Hennegar C Clement K Zucker JD Improving classification of microarray data using prototype-based feature selection ACM SIGKDD Explorations Newsletter 2003 5 23 230
Huang E Cheng SH Dressman H Pittman J Tsou MH Horng CF Bild A Iversen ES Liao M Chen CM West M Nevins JR Huang AT Gene expression predictors of breast cancer outcomes Lancet 2003 361 1590 1596 12747878 10.1016/S0140-6736(03)13308-9
Huang E Ishida S Pittman J Dressman H Bild A Kloos M D'Amico M Pestell RG West M Nevins JR Gene expression phenotypic models that predict the activity of oncogenic pathways Nat Genet 2003 34 226 230 12754511 10.1038/ng1167
Kohavi R John GH Wrappers for feature subset selection Artificial Intelligence 1997 97 273 324 10.1016/S0004-3702(97)00043-X
Li X Rao S Wang Y Gong B Gene mining: a novel and powerful ensemble decision approach to hunting for disease genes using microarray expression profiling Nucl Acids Res 2004 32 2685 2694 15148356 10.1093/nar/gkh563
Li X Rao S Zhang T Guo Z Moser KL Topol EJ An ensemble method for gene discovery based on DNA microarray data Sci China C Life Sci 2004 47 396 405 15623151
Zeeberg BR Feng W Wang G Wang MD Fojo AT Sunshine M Narasimhan S Kane DW Reinhold WC Lababidi S Bussey KJ Riss J Barrett JC Weinstein JN GoMiner: a resource for biological interpretation of genomic and proteomic data Genome Biol 2003 4 R28 12702209 10.1186/gb-2003-4-4-r28
Volinia S Evangelisti R Francioso F Arcelli D Carella M Gasparini P GOAL: automated Gene Ontology analysis of expression profiles Nucleic Acids Res 2004 32 W492 9 15215435
Vinayagam A Konig R Moormann J Schubert F Eils R Glatting KH Suhai S Applying Support Vector Machines for Gene Ontology based gene function prediction BMC Bioinformatics 2004 5 116 15333146 10.1186/1471-2105-5-116
Khatri P Draghici S Ostermeier GC Krawetz SA Profiling gene expression using onto-express Genomics 2002 79 266 270 11829497 10.1006/geno.2002.6698
Al-Shahrour F Diaz-Uriarte R Dopazo J FatiGO: a web tool for finding significant associations of Gene Ontology terms with groups of genes Bioinformatics 2004 20 578 580 14990455 10.1093/bioinformatics/btg455
Al-Shahrour F Diaz-Uriarte R Dopazo J FatiGO: Data mining with Gene Ontology
Loguinov AV Mian IS Vulpe CD Exploratory differential gene expression analysis in microarray experiments with no or limited replication Genome Biol 2004 5 R18 15003121 10.1186/gb-2004-5-3-r18
Goodall C Mosteller F and Tukey JW M-estimators of location: An outline of the theory Understanding Robust and Exploratory Data Analysis 1983 New York, 339 403
Quinlan JR C4.5: Programs for Machine Learning 1993 San Francisco, Morgan Kaufmann
Quinlan JR C4.5 Release 8
|
15774002
|
PMC1274255
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 17; 6:58
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-58
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-601577402210.1186/1471-2105-6-60Methodology ArticleSNPHunter: a bioinformatic software for single nucleotide polymorphism data acquisition and management Wang Lin [email protected] Simin [email protected] Tianhua [email protected] Xin [email protected] Program for Population Genetics, Harvard School of Public Health, Boston, MA 02115, USA2 Department of Epidemiology, Harvard School of Public Health, Boston, MA 02115, USA3 Division of Preventive Medicine, Department of Medicine, Brigham and Women's Hospital, Harvard Medical School, Boston, MA 02215, USA2005 18 3 2005 6 60 60 12 1 2005 18 3 2005 Copyright © 2005 Wang et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Single nucleotide polymorphisms (SNPs) provide an important tool in pinpointing susceptibility genes for complex diseases and in unveiling human molecular evolution. Selection and retrieval of an optimal SNP set from publicly available databases have emerged as the foremost bottlenecks in designing large-scale linkage disequilibrium studies, particularly in case-control settings.
Results
We describe the architectural structure and implementations of a novel software program, SNPHunter, which allows for both ad hoc-mode and batch-mode SNP search, automatic SNP filtering, and retrieval of SNP data, including physical position, function class, flanking sequences at user-defined lengths, and heterozygosity from NCBI dbSNP. The SNP data extracted from dbSNP via SNPHunter can be exported and saved in plain text format for further down-stream analyses. As an illustration, we applied SNPHunter for selecting SNPs for 10 major candidate genes for type 2 diabetes, including CAPN10, FABP4, IL6, NOS3, PPARG, TNF, UCP2, CRP, ESR1, and AR.
Conclusion
SNPHunter constitutes an efficient and user-friendly tool for SNP screening, selection, and acquisition. The executable and user's manual are available at .
==== Body
Background
With the ever-increasing volume of single nucleotide polymorphisms (SNPs) deposited in publicly available databases such as National Center for Biotechnology Information (NCBI) dbSNP, laboratory geneticists are faced with the routine need of selecting an appropriate set of SNPs in both gene mapping and molecular evolution studies. The major bottleneck in the workflow for SNP-based studies has shifted away from SNP discovery toward SNP selection. Although it is beyond dispute that several web-based applications and stand-alone software packages are available for handling SNP data, including viewGene [1], Genotools SNP manager [2], SNPbox [3], SNPicker [4], and SNPper [5], these applications go off on a tangent when it comes to selecting the best SNP set because their applications focus on primer design (e.g. SNPbox and SNPicker), SNP visualization (e.g. viewGene), specific platform applications such as MassARRAY technology (e.g. Genotools SNP manager), and SNP search (e.g. SNPper). In light of the surging interest in haplotype inference [6,7] and haplotype-based association studies, the power of a linkage disequlibrium (LD) study is determined not only by the number of SNPs used, but also by the quality. Contemporary geneticists aim to maximize the statistical power in detecting a disease-susceptible locus by selecting a "best set" of closely linked SNPs given a limited (and often fixed) genotyping budget [8].
In the case when a large number of SNPs are available for a susceptibility gene of interest, genotyping all SNPs on all samples is an inefficient utilization of resources. Recently, a cost-effective two-stage method has been proposed to identify disease-susceptibility markers [9]. In stage I, a set of SNPs (S1), spaced in a predefined interval, is selected (e.g., evenly spaced every 3 to 5 Kb in and surrounding a candidate gene [10]). The genotypes of markers in S1 are then used to define LD blocks and to reconstruct haplotypes within blocks across the candidate gene locus in a representative random sample, C1, of the original source population (e.g., a multiethnic cohort of men and women [11]). In stage II, a representative set (S2; S2 ⊆ S1) of htSNPs is selected on the basis of the LD characterization in the random sample of C1, and S2 is then genotyped in a much larger case-control set C2 (C1 ⊂ C2), and haplotype-based association tests are performed in C2, nested in the original source population.
Both stages are critical to the success of an association study. However, it is not a trivial task to select S1 (i.e. a set of evenly spaced SNPs in and surrounding a candidate gene) because the number of available SNPs for each human gene varies dramatically (from <10 to >200) because of varying gene sizes and SNP densities. Furthermore, we certainly do not simply keep common SNPs in S1 [i.e., minor allele frequency (MAF) ≥ 5%]; missense and regulatory SNPs should still be considered to be included in S1 even if their MAFs fall below 5% [11]. Hand-picking S1 by "eyeballing" is extremely labor-intensive, time-consuming, and error-prone for candidate genomic regions with hundreds and thousands of SNPs. Furthermore, obtaining a SNP flanking sequence long enough (~200 bp), and with annotation of nearby potential SNPs, is essential for the successful design of a SNP genotyping assay. Unfortunately, the flanking sequences of many SNPs recorded in dbSNP are short (<100 bp) and without any annotation of nearby SNPs.
NCBI's dbSNP offers a comprehensive SNP searching tool [12]. However, tools are still needed to easily and efficiently locate the desired SNPs, to evaluate their annotations, and to export them in suitable formats for downstream analyses. To meet such needs, we have developed the program SNPHunter, a tool with a friendly graphical user interface (GUI) that works as a portal between the user and NCBI dbSNP [13]. The program can extract and export SNP data retrieved from dbSNP, import saved SNP data, and offers a very flexible SNP selection function with graphic illustration of SNP position, function and heterozygosity. Furthermore, it retrieves any arbitrarily-defined, user-specified length of SNP flanking sequence with annotation of all nearby SNPs.
Architectural structure
SNPHunter was written using Microsoft Visual Basic .NET. A schematic diagram of the architectural framework for SNPHunter is shown in Figure 1. This tool relies on an HTTP parser that delegates the user's query to databases including dbSNP [13], MapViewer, LocusLink [14], and AceView at NCBI (Figure 1, left), and parses the retrieved data. It consists of three modular components, SNP Search, SNP Management, and LocusLink SNP. In the SNP Search module, the user inputs the gene symbol of interest and chooses SNPs based on heterozygosity (HET), chromosomal position, and functional class (Figure 1, upper right). The user can also specify whether upstream/downstream sequences of the gene should also be included for search. In the SNP Management module, the user fetches and manages detailed information for SNPs retrieved in the SNP Search module or the user's own SNP list. In the LocusLink module, SNPHunter reads in a list of LocusLink gene IDs (i.e. Entrez Gene IDs) and performs a batch-mode SNP search via LocusLink (Although NCBI LocusLink was superseded by the NCBI Entrez Gene, this SNP search mode is still fully functional).This module is very useful for obtaining SNP data with a large number of genes. SNPHunter creates a SNP summary and pops up a new "Filter SNP" panel (Figure 1, lower right). SNP filtering can be performed on all or selected genes according to user-specified filtering criteria. One advantage of the SNPHunter's local filter is that it does not rely on any Web server to perform the filtering. Once the user has downloaded all the SNP information and exported it to a local file, SNPHunter will perform filtering either automatically or manually, which means that the user can further modify the selection after automatic filtering. The selected SNP list can be exported to local directories for storage or further analyses. This batch-mode search operation is impressively fast. In the example shown in Figure 1, all the SNPs on the six genes were retrieved and downloaded in 10 sec, and automatic filtering on a regular personal computer with one Intel Pentium 4 2.8 GHz processor took another 7 sec.
Implementation
A detailed description of the implementation of the three modules has been presented in the User's Manual [15]. In brief, since retrieval of the flanking sequences of a desired SNP relies on knowledge of its genomic coordinate, in an ad hoc mode, SNPHunter first pinpoints the SNP's genomic coordinate from dbSNP's reference SNP (refSNP) record, strand orientation, and the SNP's corresponding contig number. Moreover, SNPHunter communicates with the NCBI MapViewer database and retrieves the corresponding sequence centering at the desired SNP, with the sequence lengths specified by users. SNPHunter will detect all neighboring SNPs located within a user-defined radius around the SNP of interest. Once the SNP's genomic coordinate and contig data are retreived, SNPHunter also obtains nearby SNP data on all neighboring SNPs by querying dbSNP for all available SNPs that lie within the user-defined radius. Once the starting and ending coordinates of a particular gene are determined by SNPHunter through NCBI's AceView, the 5' upstream and 3' downstream regions of the gene can be retrieved according to user-defined lengths. In a batch-mode, SNPHunter communicates with NCBI's LocusLink to fetch the SNPs that reside within each LocusLink gene. Since LocusLink has a curated SNP list for each gene included in the LocusLink database, this batch-mode search offers a reliable, efficient way to conduct a systematic SNP search for a large set of candidate genes (e.g. belonging to the same biological pathway/network). Furthermore, SNP data can be stored in the user's local directories, and SNP filtering can be performed automatically according to user-defined criteria.
Application example
To demonstrate the SNP selection process from dbSNP using SNPHunter, we applied SNPHunter for S1 selection for 10 biological candidate genes (Table 1) for a type 2 diabetes mellitus (DM) case-control study. These 10 candidate genes were chosen on the basis of their biochemical and physiological functions.
We used the following four SNP selection criteria:
(1) Genome coverage: SNPs should cover the gene region as well as its 30 Kb 5' upstream and 30 Kb 3' downstream regions (the gene sizes are shown in Table 1).
(2) Functionality priority: coding SNPs (cSNPs; including both synonymous and nonsynonymous SNPs) and splice site SNPs (ssSNPs) must be kept; for SNPs located in the 5' upstream region and 3' downstream regions, the function is defined according to existing in vivo/in vitro experimental data. The priority of SNP selection is nonsynonymous SNPs > synonymous SNPs > ssSNPs > 5' upstream SNPs > 3' downstream SNPs > intronic SNPs.
(3) Priority based on HET: For cSNPs and ssSNPs, no HET threshold is set (HET can be calculated using the POLYMORPHISM software [16]); for intronic and 5' upstream or 3' downstream region SNPs, those SNPs with HET values going above the threshold of 0.095 (which correspond to MAF ≥ 5%) have higher priorities.
(4) SNP density: The SNPs should be relatively evenly distributed across the gene region (as well as the 30 Kb 5' upstream and 30 Kb 3' downstream regions) with a density of 5–50 SNPs/Kb depending on the gene sizes (see Table 1). The goal is that for gene sizes < 10 Kb, we use a density of 50 SNPs/Kb; for gene sizes 10–100 Kb, we use a density of 10 SNPs/Kb; for gene sizes > 100 Kb, we use a density of 5 SNPs/Kb.
To date, there are no turn-key solutions that can select the best SNP set automatically. Our SNP selection procedure is an iterative process consisting of the following four major steps:
(a) Retrieve all SNPs regardless of HET values according to SNP selection criterion (1).
(b) Select all cSNPs and ssSNPs; in addition, 5' upstream, 3' downstream and intronic SNPs with HET ≥ 0.095 will also be selected according to SNP selection criterion (2).
(c) Enforce a relatively even SNP density according to SNP selection criterion (4). We implement this by setting the maximum inter-marker distance d (i.e., for a given set of selected SNPs S, if there exists a pair of neighboring SNPs (SNPi, SNPj), where the physical distance between SNPi and SNPj is <d, the program recursively picks a random SNP, say SNPk, between SNPi and SNPj and inserts SNPk in the middle of SNPi and SNPj; by mathematical induction, this process will guarantee that S will eventually be a saturated set, S', at a resolution level of d). Re-adjust the marker density by iteratively adding available SNPs in the priority order set by SNP selection criterion (2) and (3) until we come to a target number of SNPs with desired density, according to SNP selection criterion (4).
(d) Include any non-redundant SNPs from sources other than dbSNP, such as from literature review.
Using these criteria and selection procedures, we selected a total of 670 SNPs for the 10 genes listed in Table 1.
Besides SNP selection, SNPHunter allows the retrieval of genomic coordinates and flanking sequences for specific SNPs and gives graphic illustration of all the SNPs within the gene of interest as well. Figure 2 gives an illustration of the 28 SNPs found in a 2.7 Kb region spanning the tumor necrosis factor (TNF) gene from NCBI dbSNP.
Discussion
The motivation for developing SNPHunter is to allow the efficient and accurate selection of S1 (see Background) because of its intrinsic value in LD studies, particularly in a case-control setting. A few Web resources, such as NCBI's Entrez, Ensembl's EnsMart [17] and SNPper [5] provide SNP database searching and SNP information downloading according to user-specified criteria. These tools, each with its own unique capabilities and focuses, have benefited the work of geneticists. However, few of them are dedicated solely for SNP search purposes and for the management of SNP data. Although SNPper [5] offers a very helpful function of filtering SNP sets, it is a locally stored SNP-centric database resource maintained by the Children's Hospital Informatics Program, Harvard Medical School, and requires regular data downloads from NCBI dbSNP. By contrast, SNPHunter is designed to work as a stand-alone application that retrieves the most-updated SNP and sequence data without the need for complicated local database support. Thus, the user is relieved from maintaining a local database and updating the data frequently. The ability to export and to save every dataset locally in plain text format provides the user with the freedom for later reuse or any other customized analysis without any website support. In addition, SNPHunter offers a very friendly GUI, allowing researchers without much computer background to perform SNP searches easily and efficiently. Moreover, its batch search and automatic SNP selection proved very efficient in large-scale candidate genes study. Table 2 lists features comparisons between SNPHunter and other major SNP related software/web tools.
It is worth noting that SNPHunter relies on dbSNP for data retrieval, and thus is deprived of the independence whereas other application with local database support usually has. What's more, SNP selection should not be limited to NCBI dbSNP, although dbSNP represents the largest publicly available SNP database that can be accessed via the Internet worldwide. Some SNPs reported in the earlier literature have not yet been incorporated into dbSNP. Furthermore, there are several on-going gene re-sequencing projects for selected human genes, such as SeattleSNPs or SNP500Cancer [18]. Therefore, SNPs from these other sources, if not yet included in dbSNP, should also be considered in SNP selection. Nevertheless, NCBI dbSNP has been steadily updated and has gradually emerged as one of the most comprehensive SNP depositories.
Conclusion
In summary, SNPHunter allows for customized SNP searches (both ad hoc-mode and batch-mode) by directly retrieving and managing SNP information from the NCBI dbSNP database, eliminating tedious and costly local database maintenance on the user's side. To date, SNPHunter has received more than 1000 downloads worldwide. We hope this simple program can serve as an efficient and reliable tool for researchers everywhere to facilitate their genetic studies.
Availability and requirements
Project name: SNPHunter
Project home page:
Operating system(s): Microsoft Windows
Programming language: Visual Basic .NET
Other requirements: Microsoft .NET Framework 1.0 or above.
License: None
Any restrictions to use by non-academics: Contact authors
Authors' contributions
Simin Liu, Tianhua Niu, and Lin Wang identified the need to develop such a program, initiated the project, and designed the basic functions. Lin Wang wrote the source code for the software and interface design. Tianhua Niu and Xin Xu contributed with ideas on overall design, feature requirements, and implementation. All authors participated in the drafting of the manuscript and approved the final version.
Acknowledgements
We wish to thank the many SNPHunter users for their constructive comments, especially Illumina, Inc. We thank Melissa Veno for editorial assistance. This work was supported in part by National Institutes of Health grants R01 HG002518, R01 DK062290, R01 DK066401, and R01 HL073882.
Figures and Tables
Figure 1 A depiction of the architecture structure of SNPHunter. SNPHunter allows the user to perform (1) an ad hoc search by gene symbol [an example of "Tumor Necrosis Factor" (Gene Symbol, TNF) is shown]; and (2) batch-mode search. It can be seen that a total of 722 SNPs were found on a total of 6 user-specified genes (with LocusLink ID listed on the right). With automatic filtering, 405 SNPs were picked and the gaps between them were also calculated.
Figure 2 The Gene View panel which gives a graphic illustration of all the SNPs within the gene of TNF (LocusLink ID: 7124). The height of a SNP bar indicates the heterozygosity of that SNP, and the four dotted horizontal lines means heterozygosity of 0, 0.25, 0.5, and 0.75 respectively. Out of 28 available SNPs, two have been dropped and plotted as "gray". For those 26 selected SNPs, "green" means "coding: synonymy unknown" or "synonymous"; "red" means "non synonymous"; "orange" means all the others. And for the convenience of SNP selection, there is a red triangle indicating the current focused SNP, which is SNP rs3093663.
Table 1 Size, location, and the estimated number of SNPs for each of the 10 candidate genes for type 2 diabetes mellitus.
Gene Symbol Locus-Link Location Size (bp) (# exons)a Total # SNPs (#/Kb)b # SNPs MAF ≥ 5% (density)c Func SNPsd # SNPs selected for S1
5' 30 Kb 3' 30 Kb Gene region Total
CAPN10 11132 2q37.3 48411(32) 308 (6) 27 (0.6) 1 3 11 25 39
FABP4 2167 8q21 5616 (6) 119 (21) 4 (0.7) 2 1 1 20 22
IL6 3569 7p21 6130 (10) 291 (47) 49 (8.0) 8 3 8 40 51
NOS3 4846 7q36 24637(29) 299 (12) 67 (2.7) 28 7 6 62 75
PPARG 5468 3p25 146974(16) 627 (4) 124 (0.8) 5 0 7 142 149
TNF 7124 6p21.3 2775 (5) 381 (137) 42 (15.1) 6 17 14 20 51
UCP2 7351 11q13 34786 (26) 135 (4) 13 (0.4) 2 4 10 20 34
CRP 1401 1q21-q23 2306 (18) 171 (74) 33 (14.3) 1 5 18 21 44
ESR1 2099 6q25.1 472913 (23) 1403 (3) 207 (0.4) 9 6 16 85* 107
AR 367 Xq11.2 180270 (13) 241 (1) 15 (0.1) 8 1 1 96 98
a Gene statistics from NCBI AceView; number of exons was calculated as number of intron plus 1.
b Including SNPs belonging to all 6 NCBI dbSNP's function classes (i.e. Coding Nonsynonymous, Coding Synonymous, mRNA UTR, Splice site, Intron, and Others) that cover the gene locus as well as the 30-Kb 5' upstream and 30-Kb 3' downstream flanking sequences of the locus.
c Note that a subset of dbSNP entries do not include information on allele frequency or heterozygosity. The numbers of SNPs reported here indicate those SNPs with MAF information and MAF ≥ 5%, including 30-Kb 5' upstream and 30-Kb 3' downstream flanking regions.
d Defined as Coding Nonsynonymous, Coding Synonymous, or Splice-site SNPs.
* SNP distance (maximum inter-marker distance between selected SNPs) was controlled at 10 Kb for the ESR1 gene.
Table 2 Comparisons between SNPHunter and other publicly available software/tool.
Software/Web tool Type Maintenance requirement SNP search related functions Primer design
SNP search Arbitrary flanking sequence length Nearby SNP annotation SNP selection interface Graphic SNP illustration
NCBI dbSNPa Web-Server None for user; server contains latest updated data Yes No No No Yes No
SNPper Web-Server None for user; server needs to update periodically Yesb Yes Yes No No Yes
SNPHunter Web-Client (depends on dbSNP for new SNP retrieval) None for user; rely on dbSNP for data update Yesb Yes Yes Yesc Yes No
SNPicker Web-Client (depends on NEB restriction-enzyme data) None for user; rely on NEB for data update No No No No No Yes
SNPBox Web-Server None No No No No No Yes
viewGene Software depends on annotated record Needs to download annotated data Yes Yes Yes No Yes No
a dbSNP's Entrez query.
b Both SNPper and SNPHunter can do batch search for a list of genes instead of single gene at a time.
c SNPHunter offers automatic SNP selection as well as manual selection.
==== Refs
Kashuk C SenGupta S Eichler E Chakravarti A ViewGene: a graphical tool for polymorphism visualization and characterization Genome Res 2002 12 333 338 11827953 10.1101/gr.211202
Pusch W Kraeuter KO Froehlich T Stalgies Y Kostrzewa M Genotools SNP manager: a new software for automated high-throughput MALDI-TOF mass spectrometry SNP genotyping Biotechniques 2001 30 210 215 11196313
Weckx S De Rijk P Van Broeckhoven C Del-Favero J SNPbox: web-based high-throughput primer design from gene to genome Nucleic Acids Res 2004 32 W170 2 15215373 10.1093/nar/gnh168
Niu T Hu Z SNPicker: a graphical tool for primer picking in designing mutagenic endonuclease restriction assays Bioinformatics 2004 20 3263 3265 15201186 10.1093/bioinformatics/bth360
Riva A Kohane IS SNPper: retrieval and analysis of human SNPs Bioinformatics 2002 18 1681 1685 12490454 10.1093/bioinformatics/18.12.1681
Niu T Qin ZS Xu X Liu JS Bayesian haplotype inference for multiple linked single-nucleotide polymorphisms Am J Hum Genet 2002 70 157 169 11741196 10.1086/338446
Niu T Algorithms for inferring haplotypes Genet Epidemiol 2004 27 334 347 15368348 10.1002/gepi.20024
Hoh J Wille A Zee R Cheng S Reynolds R Lindpaintner K Ott J Selecting SNPs in two-stage analysis of disease association data: a model-free approach Ann Hum Genet 2000 64 413 417 11281279 10.1046/j.1469-1809.2000.6450413.x
Thompson D Stram D Goldgar D Witte JS Haplotype tagging single nucleotide polymorphisms and association studies Hum Hered 2003 56 48 55 14614238 10.1159/000073732
Stram DO Haiman CA Hirschhorn JN Altshuler D Kolonel LN Henderson BE Pike MC Choosing haplotype-tagging SNPS based on unphased genotype data using a preliminary sample of unrelated subjects with an example from the Multiethnic Cohort Study Hum Hered 2003 55 27 36 12890923 10.1159/000071807
Haiman CA Stram DO Pike MC Kolonel LN Burtt NP Altshuler D Hirschhorn J Henderson BE A comprehensive haplotype analysis of CYP19 and breast cancer risk: the Multiethnic Cohort Hum Mol Genet 2003 12 2679 2692 12944421 10.1093/hmg/ddg294
dbSNP [http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=Snp&cmd=Limits]
Sherry ST Ward MH Kholodov M Baker J Phan L Smigielski EM Sirotkin K dbSNP: the NCBI database of genetic variation Nucleic Acids Res 2001 29 308 311 11125122 10.1093/nar/29.1.308
Pruitt KD Maglott DR RefSeq and LocusLink: NCBI gene-centered resources Nucleic Acids Res 2001 29 137 140 11125071 10.1093/nar/29.1.137
PPGWebsite [http://www.hsph.harvard.edu/ppg/software.htm]
Niu T Struk B Lindpaintner K Statistical considerations for genome-wide scans: design and application of a novel software package POLYMORPHISM Hum Hered 2001 52 102 109 11474211 10.1159/000053361
Kasprzyk A Keefe D Smedley D London D Spooner W Melsopp C Hammond M Rocca-Serra P Cox T Birney E EnsMart: a generic system for fast and flexible access to biological data Genome Res 2004 14 160 169 14707178 10.1101/gr.1645104
Packer BR Yeager M Staats B Welch R Crenshaw A Kiley M Eckert A Beerman M Miller E Bergen A Rothman N Strausberg R Chanock SJ SNP500Cancer: a public resource for sequence validation and assay development for genetic variation in candidate genes Nucleic Acids Res 2004 32 D528 D532 14681474 10.1093/nar/gkh005
|
15774022
|
PMC1274256
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 18; 6:60
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-60
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-611577748210.1186/1471-2105-6-61SoftwareGraphical representation of ribosomal RNA probe accessibility data using ARB software package Kumar Yadhu [email protected] Ralf [email protected] Sebastian [email protected] Bernhard [email protected]öckner Frank Oliver [email protected] Rudolf [email protected] Harald [email protected] Wolfgang [email protected] Lehrstuhl für Mikrobiologie, Technische Universität München, D-85350 Freising Germany2 Max Plank Institute for Marine Microbiology, D-28359 Bremen, Germany3 International University Bremen, D-28759 Bremen, Germany4 Lehrstuhl für Rechnertechnik und Rechnerorganisation, Technische Universität München, D-85748 Garching, Germany2005 21 3 2005 6 61 61 12 10 2004 21 3 2005 Copyright © 2005 Kumar et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Taxon specific hybridization probes in combination with a variety of commonly used hybridization formats nowadays are standard tools in microbial identification. A frequently applied technology, fluorescence in situ hybridization (FISH), besides single cell identification, allows the localization and functional studies of the microbial community composition. Careful in silico design and evaluation of potential oligonucleotide probe targets is therefore crucial for performing successful hybridization experiments.
Results
The PROBE Design tools of the ARB software package take into consideration several criteria such as number, position and quality of diagnostic sequence differences while designing oligonucleotide probes. Additionally, new visualization tools were developed to enable the user to easily examine further sequence associated criteria such as higher order structure, conservation, G+C content, transition-transversion profiles and in situ target accessibility patterns. The different types of sequence associated information (SAI) can be visualized by user defined background colors within the ARB primary and secondary structure editors as well as in the PROBE Match tool.
Conclusion
Using this tool, in silico probe design and evaluation can be performed with respect to in situ probe accessibility data. The evaluation of proposed probe targets with respect to higher-order rRNA structure is of importance for successful design and performance of in situ hybridization experiments. The entire ARB software package along with the probe accessibility data is available from the ARB home page .
==== Body
Background
The introduction and use of comparative sequence analysis of appropriate marker genes as a powerful tool in taxonomy has substantially contributed to the rapid growth of molecular sequence databases such as EMBL [1], GenBank [2], and ribosomal RNA (rRNA) databases [3-5]. Evidently, molecular phylogenetic analyses have greatly influenced the restructuring of systematics especially in the case of prokaryotes. Nowadays, identification and classification at the species and higher taxonomic levels mainly relies on a genotypic approach, typically involving an analysis of small, and to a lesser extent, large ribosomal RNA gene (rRNA) structures. The backbone of the current taxonomy of the prokaryotes is almost exclusively based upon a phylogenetic network derived from comparative sequence analysis of the small subunit rRNAs and respective phylogenetic marker genes [6]. As 'living fossils', these molecules at least roughly reflect the evolutionary history of the respective organisms. The mosaic-like primary structures comprising highly variable to highly conserved or invariant regions provide diagnostic information for different levels of phylogenetic relationship. Consequently, this information can be used to identify oligonucleotide target regions unique to phylogenetic entities, for use as taxon-specific hybridization probes or PCR primers. Depending on the target site such oligonucleotide probes or probe combinations can be designed for phylogenetic groupings as diverse as bacterial species or an entire phylum.
Ever since the fluorescence in situ hybridization (FISH) technique became an integral part of the rRNA approach to microbial ecology and evolution [7], rRNA-targeted oligonucleotide probes have evolved into a widely used tool for the direct, cultivation-independent identification and enumeration of individual microbial cells or specific groups of bacteria in simple to complex natural environments. In this regard, a good probe design and careful further evaluation in silico plays a crucial role to ensure sensitivity and specificity of a potential probe in its practical application. Besides uniqueness of the target sequence, number, character and position of diagnostic residues, comprehensiveness with respect to the inclusion of members of the desired target group (taxon) and exclusion of non-members along with a target molecule or region accessibility in the real hybridization experiment, have to be taken into consideration. Recently, data on in situ accessibility of rRNA targets in several microorganisms have become available [8-11].
Since biology is a highly visual science, there is a general demand for tools to visualise the variety of biological knowledge as diagrams, illustrations, two-dimensional and three-dimensional reconstructions, and other types of graphical formats. Hence, the visualization of molecular data in an interactive and intuitive graphical user interface ideally will serve as third eye for a molecular biologist. In this paper, we describe how the ARB software package [3] provides a workbench for designing, evaluation and visualization of oligonucleotide probes in more intuitive way, using interactive graphical user interface to visually examine characteristics and criteria of target regions.
Implementation
Sequence data
Periodically retrieved raw gene data comprising small subunit rRNA from public databases such as EBI [1], Genbank [2], the RDP[4], and the sequence data determined in our laboratory and other partner groups are imported into the ARB database, processed according to a variety of criteria and finally provided as curated databases at the ARB projects web-site [13]. The current public release of small subunit rRNA database [3] containing only complete sequences was taken for designing, evaluation and visualization of probes and targets, respectively. Partial sequences are avoided as they greatly limit the probe design by reducing the number of potential target regions and also give no hint about the specificity of existing probes that target to non-sequenced regions of the respective rRNAs.
The positional tree (PT) server
The PT-Server [3] is a suffix tree server implemented in the ARB software which is used for indexing all sequence data represented in the underlying ARB sequence database. Once established, the particular PT-Server allows rapid and exact searching for target regions with respect to sequence identity or uniqueness.
Probe design and probe match
Probe design is carried out using the PROBE Design tool (PDT) of ARB software involving following steps:
1. The user selects the target group or a species of interest.
2. The parameters such as size of the probe and the probable physico-chemical characteristics like %GC content, melting temperature (Tm) according to the 4°C GC, 2°C AT rule [14], and self-complementarity (hair-pin bonds) are specified. Optionally, a range of allowed target positions within the sequence alignment of the respective database can be defined.
3. Potential probe candidates are searched involving the respective PT-Server. Both, target and probe sequence are displayed in a result list. Ranking within this list follows estimated probe quality according to criteria defined for probe design such as number, character and position of diagnostic residues, coverage of the target group, physicochemical demands, which are displayed in separate probe results window along with relevant information.
4. Once the user selects the desired probe in the result list, it can be evaluated against the entire database by using the PROBE Match tool (PMT) of ARB. PMT, by default evaluates the targets for the sequence (strand) stored in the database. Optionally, the complementary sequence (opposite strand) can be evaluated as well. Members of the target group are displayed in a separate PROBE Match window along with other information such as number of mismatches, weighted mismatches, E. coli positions, reverse complementarity and local alignment of probe targets (Figure 1).
Results and discussion
As the demand for oligonucleotide probes that can identify and quantify bacteria by nucleic acid hybridization is permanently increasing, in silico evaluation and visualization of such probes and targets are necessary, particularly, when used for FISH experiments. Target accessibility is among the crucial criteria to be evaluated with respect to experimental success of the respective probe based identification and detection system [7-12]. To facilitate this evaluation procedure, new functionalities were added to the ARB software package providing a more intuitive graphical environment. As an example, oligonucleotide probes were designed for the enterobacteria group represented by 947 database entries. The 5'-UGGAGGGGGAUAACUACU-3' probe was selected from the list of potential probes and evaluated against the background of the full dataset of complete and partial small subunit rRNA sequences. The selected probe perfectly matches the respective target of 497 members of the enterobacteria group (Figure 1). The same probe has been visualized in all the screenshots presented in the paper.
Although a phylogenetic probe is primarily judged in terms of its taxonomic range to identify the members of its intended target taxon to the exclusion of non-target bacteria, for a practical consideration it must also fulfil certain other criteria with respect to its applicability depending on the particular hybridization format. In case of the fluorescence in situ hybridization approach the results of the accessibility studies conducted by Fuchs and co-workers on the 16S and 23S rRNA of Escherichia coli and other organisms are among such criteria. They showed that some regions of E. coli ribosome are virtually inaccessible for oligonucleotide probes when FISH is performed [8,9]. They proposed a color code assigned to six intensity classes of in situ hybridization signals. Within the ARB program, these classes are coded in respective SAIs (so called Sequence Associated Information) and optionally visualized as background colors of the sequences in primary structure (ARB_Edit4), secondary structure (SEC_Edit), and probe visualization windows (PROBE Match) of ARB. All the displays produced by the ARB software are interconnected and any changes in one window are automatically updated in other windows as well. Simultaneous visualization and evaluation of oligonucleotide probes in different levels allows the user to look carefully and closely into the proposed probe candidates in silico, before carrying out further in situ or in vivo studies. More importantly, the user can perform a variety of sequence related operations such as importing sequence data, aligning, treeing, designing, evaluation and visualization of probes, performing statistical calculations and many other functions using interoperating and user friendly tools controlled from a common graphical platform within the ARB software package.
Visualization of potential probe candidates and the sequence associated information (SAI) such as higher order structure, conservation, G+C content, transition-transversion profiles and in situ target accessibility patterns, is possible at three different levels: the local alignment (PROBE Match tool), global alignment (ARB Primary Structure Editor) and secondary structure levels (Secondary Structure Editor).
Visualization of SAI in probe match window
Visualization of probe candidates in a local alignment along with additional sequence associated information (SAI) can be managed with the PROBE Match SAI window. The neighboring region up to nine nucleotides on either terminus of the potential probe target is retrieved from the database. A local alignment of the extracted rRNA sequence is established and displayed along with the respective unique identifier such as ARB short_name, accession number, or any other underlying database fields (eg., Full Name, Group) (Figure 2, 3, 4). The user can select any information that is associated with the sequences (SAI) such as secondary structure masks (Figure 2) or any statistical calculations performed on the sequence level like sequence consensus, positional variability using parsimony method (Figure 3) or any other user defined models, filters or statistics as well as in situ accessibility maps for visualization (Figure 4). Different background colors can be assigned to characters and values or character groups and ranges of values of the particular SAIs, respectively. Optionally, the real characters or values contained in such SAIs can directly be visualized below the individual sequences. This offers a researcher a deeper insight in to the proposed oligonucleotide probe targets for careful examination of probe candidates in silico before making any decision on the selection of probe.
Visualization of SAI in ARB primary structure editor
On the global alignment level, the user selected oligonucleotide probe is visualized in different background colors in the primary structure editor window of ARB [3]. The primary structure editor (Figure 5) of ARB displays multiple sequence alignments generated by the respective ARB software tools [3] of the selected sequences from the underlying database in the user-defined colors and symbols.
As already described for the local alignment level, any type of SAI can be visualized by the user defined background colors for the individual alignment columns. Customized color selections can be assigned to the different types of SAIs mentioned before. By scrolling the mouse or the use of ARB search tools, the user gets an easy access to the information for any range or the selection of sequences. In the context of probe evaluation for in situ hybridization experiments, mapping of experimentally derived in situ accessibility patterns onto the primary structures of interest certainly provides valuable support to the users for probe evaluation. Part of a multiple 16S rRNA sequence alignment is shown in the figure 5. The brightness classes defined for 16S rRNA structural model of Methanosaeta sedula [11] are mapped on the aligned sequences and indicated by background colors according to Behrens et al [11].
Visualisation of SAI in ARB secondary structure editor
Theoretically as well as experimentally derived secondary structure information of SSU rRNA [15-17] is used more profoundly in sequence alignment refinement and probe design and evaluation. The tertiary structure of the SSU rRNA of the bacterium Thermus thermophilus which had been elucidated with atomic resolution by X-ray diffraction crystallography of ribosomal subunit [17] allows evaluating the exactness of the secondary structure model. The secondary structure of SSU rRNA has a crucial role in evaluating the proposed probe candidates prior to the actual experimentation. The ARB Secondary structure editor (Figure 5) provides the user with more intuitive graphical display of the secondary structure model of SSU rRNA. The user can visualize the entire SSU rRNA sequence of any organism in the respective database which fits into the common consensus model. The localization of proposed oligonucleotide probe targets can be visualized in customizable background colors.
Conclusion
The evaluation of proposed probe target position with respect to higher-order rRNA structure is of more importance especially when probes are intended to be used for in situ hybridizations [7-12]. Albeit there have been several software programs developed for the design of rRNA targeted oligonucleotide probes [18,19], the criteria taken to design the probes are generally restricted to certain parameters such as size, nucleotide composition, specificity definition, and the general hybridisation behavior. None of the software described [18,19] takes into account the special requirements of rRNA targeted probes that are destined for FISH applications which is, the structure dependant probe accessibility of the ribosomal RNA. This feature has been developed and implemented in ARB. Using this tool, in silico probe design and evaluation can be performed with respect to in situ probe accessibility data. By identifying and excluding the probes targeting sites with a poor accessibility the number of time consuming empirical tests can be reduced.
Availability and requirements
The entire ARB software and the periodic updates of well aligned and annotated ribosomal RNA databases are made freely available for the scientific community via World Wide Web [13]. Currently, the ARB Software is available for PCs running LINUX operating systems and SUN SOLARIS systems.
Authors' contributions
YK developed and implemented the tool and drafted the manuscript. RW participated in design and implementation. SB and BF provided the accessibility data and revised the manuscript. FOG, RA and HM critically revised the manuscript. WL initiated the development of the tool and supervised the ARB project.
Acknowledgements
This work was partially supported by grants to WL of the Bavarian Research Foundation (BSF) and the German Ministry of Education and Research (bmb+f).
Figures and Tables
Figure 1 Probe match window. Probe Match Window showing results of in silico evaluation of the probe candidate (5'-UGGAGGGGGAUAACUACU-3') designed as part of a probe combination for the enterobacteria group. It hits 497 members of the group containing 947 species. Additional information such as number of mismatches, E. coli position, region up- and down-stream of the probe target in the actual nucleotide sequence along with the local alignment are also shown.
Figure 2 Screenshot displaying 16S rRNA secondary structure model where helix region is colored in blue, starting and ending positions of helix halves are in red and bases without background represent commonly non base paired positions.
Figure 3 Screenshot showing positional variability filter generated by the respective ARB tool in various background colors (light purple – 7; red – 8; dark blue – 9; green – AB; blue – CD; yellow – EF; light grey – GH; grey – IJ; increasing numbers followed by the alphabetical order of letters indicate increasing degree of sequence conservation).
Figure 4 Probe visualization window. Probe visualization window displaying local alignments of probe target regions for members of the target group. Column statistics performed on the sequence alignments and structure masks are visualized in different background colors in the probe visualization window. This screenshot shows 16S rRNA accessibility map, where the experimentally determined relative fluorescence intensities are visualized in different colors (orange (0.8 – 0.61); green (0.6 – 0.41) [11]).
Figure 5 ARB Primary and secondary structure windows. ARB primary and secondary structure windows showing the distribution of relative fluorescence intensities of oligonucleotide probes targeting 16S rRNA structure of M. sedula [11]. The different background colors indicate brightness range of different classes (classes I through VI) with respect to the observed fluorescence intensities. Numbers displayed in black and lower case denotes the respective nucleotide positions and the numbers showing in red represent helix numbers in the model.
==== Refs
Kulikova T Aldebert P Althorpe N Baker W Bates K Browne P van den Broek A Cochrane G Duggan K Eberhardt R Faruque N Garcia-Pastor M Harte N Kanz C Leinonen R Lin Q Lombard V Lopez R Mancuso R McHale M Nardone F Silventoinen V Stoehr P Stoesser G Tuli A Tzouvara K Vaughan R Wu D Zhu W Apweiler R The EMBL nucleotide sequence database Nucleic Acids Res 2004 32 D27 30 14681351 10.1093/nar/gkh120
Benson DA Karsch-Mizrachi I Lipman DJ Ostell J Wheeler DL Genbank: update Nucleic Acids Res 2004 32 D23 26 14681350 10.1093/nar/gkh045
Ludwig W Strunk O Westram R Richter L Meier H Yadhukumar Buchner A Lai T Steppi S Jobb G Förster W Brettske I Gerber S Ginhart AW Gross O Grumann S Hermann S Jost R König A Liss T Lüßmann R May M Nonhoff B Reichel B Strehlow R Stamatakis A Stuckmann N Vilbig A Lenke M Ludwig T Bode A Schleifer KH ARB: a software environment for sequence data Nucleic Acids Res 2004 32 1363 1371 14985472 10.1093/nar/gkh293
Maidak BL Cole JR Lilburn TG Parker CT Saxman PR JrFarris RJ Garrity GM Olsen GJ Schmidt TM Tiedje JM The RDP-II (Ribosomal Database Project) Nucleic Acids Res 2001 29 173 174 11125082 10.1093/nar/29.1.173
Wuyts J PerrieÁre G Van de Peer Y The European ribosomal RNA database Nucleic Acids Res 2004 32 D101 103 14681368 10.1093/nar/gkh065
Ludwig W Klenk HP Garrity G Overview: a phylogenetic backbone and taxonomic framework for prokaryotic systematics Bergey's Manual of Systematic Bacteriology 2001 2 New York: Springer 49 65
Amann R Ludwig W Schleifer KH Phylogenetic identification and in situ detection of individual microbial cells without cultivation Microbiol Rev 1995 59 143 169 7535888
Fuchs BM Wallner G Beisker W Schwippl I Ludwig W Amann R Flow cytometric analysis of the in situ accessibility of Escherichia coli 16S rRNA for fluorescently labeled oligonucleotide probes Appl Environ Microbiol 1998 64 4973 4982 9835591
Fuchs BM Syutsubo K Ludwig W Amann R In situ accessibility of the Escherichia coli 23S rRNA for fluorescently labeled oligonucleotide probes Appl Environ Microbiol 2001 67 961 968 11157269 10.1128/AEM.67.2.961-968.2001
Inàcio J Behrens S Fuchs BM Fonseca I Spencer-Martins I Amann R In situ Accessibility of Saccharomyces cerevisiae 26S rRNA to Cy3-Labeled Oligonucleotide Probes Comprising the D1 and D2 Domains Appl Environ Microbiol 2003 69 2899 2905 12732564 10.1128/AEM.69.5.2899-2905.2003
Behrens S Ruehland C Inàcio J Huber H Fonseca A Spencer-Martins S Fuchs BM Amann R In Situ accessibility of small-subunit rRNA of members of the domains Bacteria, Archaea and Eucarya to Cy3-Labeled oligonucleotide probes Appl Environ Microbiol 2003 69 1748 1758 12620867 10.1128/AEM.69.3.1748-1758.2003
Behrens S Fuchs BM Mueller F Amann R Is the In situ Accessibility of the 16S rRNA of Escherichia coli for Cy3-Labeled Oligonucleotide Probes Predicted by a Three-Dimensional Structure Model of the 30S Ribosomal Subunit? Appl Environ Microbiol 2003 69 4935 4941 12902289 10.1128/AEM.69.8.4935-4941.2003
The ARB project
Suggs SV Hirose T Miyake T Kawashima EH Johnson MJ Itakura K Wallace RB Brown D, Fox CF Use of synthetic oligodeoxyribonucleotides for the isolation of specific cloned DNA sequences Developmental biology using purified genes 1981 New York: Academic Press Inc 683 693
Cannone JJ Subramanian S Schnare MN Collett JR D'Souza LM Du Y Feng B Lin N Madabusi LV MuÈller KM Pande N Shang Z Yu N Gutell RR The comparative RNA Web (CRW) site: an online database of comparative sequence and structure information for ribosomal, intron and other RNAs BMC Bioinformatics 2002 3 15 10.1186/1471-2105-3-15
Gutell RR Collection of small subunit (16S- and 16S-like) ribosomal RNA structures Nucleic Acids Res 1993 21 3051 3054 8332526
Wimberly BT Brodersen DE Clemons WM Morgan-Warren RJ Carter AP Vonrhein C Hartsch T Ramakrishnan V Structure of the 30S ribosomal subunit Nature 2000 407 327 339 11014182 10.1038/35030006
Ashelford KE Weightman AJ Fry JC PRIMROSE: a computer program for generating and estimating the phylogenetic range of 16S rRNA oligonucleotide probes and primers in conjunction with the RDP-II database Nucleic Acids Res 2002 30 3481 3489 12140334 10.1093/nar/gkf450
Pozhitkov AE Tautz D An algorithm and program for finding sequence specific oligo-nucleotide probes for species identification BMC Bioinformatics 2002 3 9 11882251 10.1186/1471-2105-3-9
|
15777482
|
PMC1274257
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 21; 6:61
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-61
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-621578013410.1186/1471-2105-6-62Methodology ArticleA standard curve based method for relative real time PCR data processing Larionov Alexey [email protected] Andreas [email protected] William [email protected] Breast Unit, Western general Hospital, Edinburgh, UK2 Novartis Pharmaceuticals, Biostatistics, CH – 4002 Basel, Switzerland3 Breast Unit, Edinburgh University, Edinburgh, UK2005 21 3 2005 6 62 62 11 11 2004 21 3 2005 Copyright © 2005 Larionov et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Currently real time PCR is the most precise method by which to measure gene expression. The method generates a large amount of raw numerical data and processing may notably influence final results. The data processing is based either on standard curves or on PCR efficiency assessment. At the moment, the PCR efficiency approach is preferred in relative PCR whilst the standard curve is often used for absolute PCR. However, there are no barriers to employ standard curves for relative PCR. This article provides an implementation of the standard curve method and discusses its advantages and limitations in relative real time PCR.
Results
We designed a procedure for data processing in relative real time PCR. The procedure completely avoids PCR efficiency assessment, minimizes operator involvement and provides a statistical assessment of intra-assay variation.
The procedure includes the following steps. (I) Noise is filtered from raw fluorescence readings by smoothing, baseline subtraction and amplitude normalization. (II) The optimal threshold is selected automatically from regression parameters of the standard curve. (III) Crossing points (CPs) are derived directly from coordinates of points where the threshold line crosses fluorescence plots obtained after the noise filtering. (IV) The means and their variances are calculated for CPs in PCR replicas. (V) The final results are derived from the CPs' means. The CPs' variances are traced to results by the law of error propagation.
A detailed description and analysis of this data processing is provided. The limitations associated with the use of parametric statistical methods and amplitude normalization are specifically analyzed and found fit to the routine laboratory practice. Different options are discussed for aggregation of data obtained from multiple reference genes.
Conclusion
A standard curve based procedure for PCR data processing has been compiled and validated. It illustrates that standard curve design remains a reliable and simple alternative to the PCR-efficiency based calculations in relative real time PCR.
==== Body
Background
Data processing can seriously affect interpretation of real time PCR results. In the absence of commonly accepted reference procedures the choice of data processing is currently at the researcher's discretion. Many different options for data processing are available in software supplied with different cyclers and in different publications [1-7]. However, the basic choice in relative real time PCR calculations is between standard curve and PCR-efficiency based methods. Compared to the growing number of studies addressing PCR efficiency calculations [3,5,8-10] there is a shortage of publications discussing practical details of the standard curve method [11]. As a result, the PCR efficiency approach appears as the method of choice in data processing for relative PCR [12]. However, when reliability of results prevails over costs and labor load, the standard curve approach may have advantages.
The standard curve method simplifies calculations and avoids practical and theoretical problems currently associated with PCR efficiency assessment. Widely used in many laboratory techniques this approach is simple and reliable. Moreover, at the price of a standard curve on each PCR plate it also provides the routine validation for methodology. To benefit from the advantages of the standard curve approach and to evaluate its practical limitations we designed a data processing procedure implementing this approach and validated it for relative real time PCR.
Results
Description of the data processing procedure
Source data
Raw fluorescence readings were exported from Opticon Monitor software and processed in MS Excel using a VBA script (the mathematical formulae, script and samples of source data are attached to the electronic version of publication, see Additional files 1 and 2).
Noise filtering
The random cycle-to-cycle noise was reduced by smoothing with a 3 point moving average (two-point average in the first and the last data points). Background subtraction was performed using minimal value through the run. If significant scattering in plateau positions was observed it was removed by amplitude normalization (normalizing by maximal value in the cell over the whole PCR run). The noise filtering is illustrated in the Figure 1.
Crossing points calculation
The crossing points (CPs) were calculated directly as the coordinates of points in which the threshold line actually crossed the broken lines representing fluorescence plots obtained after the noise filtering (Figure 2). If several intersections were observed the last one was used as the crossing point.
Standard curve calculation
A standard curve was derived from the serial dilutions by a customary way. Relative concentrations were expressed in arbitrary units. Logarithms (base 10) of concentrations were plotted against crossing points. Least square fit was used as the standard curve.
Threshold selection
The optimal threshold was chosen automatically. The VBA script examined different threshold positions calculating coefficient of determination (r2) for each resulting standard curve. The maximum coefficient of determination pointed to the optimal threshold (typically the maximum r2 was larger than 99%).
Calculating means and variances of means for crossing points in PCR replicas
The optimal threshold was used to calculate CPs for unknown samples. Means and variances of means were then calculated for CPs in PCR replicas.
Derivation of non-normalized values from crossing points
The non-normalized values were calculated from the CPs' means by the standard curve equation followed by exponent (base 10). The variances were traced by the law of error propagation.
Summarizing data from several reference genes to a single normalizing factor
Two options are available in the VBA script to summarize data from multiple reference genes:
- Arithmetic mean (deprecated),
- Geometric mean (recommended).
Calculation of normalized results for target genes
The final results representing relative expression of target genes were calculated by dividing the non-normalized values by the above normalization factor. The normalized results' variances were derived by the law of error propagation.
When confidence intervals or coefficients of variation were needed they have been calculated from the corresponding variances (see Additional file 1 with formulae for details).
Procedure testing and validation
We tested this procedure on the measurement of expression of 6 genes in 42 breast cancer biopsies (Figure 3, Table 1).
To validate the assumption of a Normal distribution for the initial data (i.e. CPs) we studied distributions of crossing points in four plates, each of which represented a 96× PCR replica. The observed distributions were symmetric, bell-shaped and close to a Normal distribution (Figure 4, Table 2).
Transformation of the Normal distribution through PCR data processing was analyzed by a computer simulation. It showed that the shape of resulting distributions significantly depends on the initial data dispersion. At low variation in crossing points (SD < 0.2 or CV < 1%) the distributions remain close to Normal through all steps of data processing (Figure 5-A). In contrast, at higher initial dispersion (crossing points' SD > 0.2 or CV > 1%) the PCR data processing transformed the Normal distribution such that the resulting distributions became asymmetric and far from normal (Figure 5-C).
Addressing the use of amplitude normalization we studied several factors potentially affecting PCR plateau level. On the gels run immediately after PCR the weak bands initially visible without staining because of SYBR Green originated from PCR mixes were remarkably increased after additional staining with SYBR Green (Figure 6). When PCRs were run with different concentrations of primers, enzyme, and using different caps for PCR plate, neither increase of primers nor addition of enzyme influenced the plateau level and scattering. However, the caps design did affect the plateau position (Figure 7).
Discussion
PCR data processing is a complex procedure that includes a number of steps complementing each other. Many different options have been suggested by different authors at each step of PCR data processing. In the discussion below we go through our procedure on a step-to-step basis shortly discussing the available options and explaining our choices. In general, we preferred the simplest functioning solutions. In statistical treatment we looked for valid practical estimations rather than for mathematically exact solutions. Because of lack of relevant theoretical data we paid especial attention to the amplitude normalisation and to statistical processing of intra-assay PCR replicas. To validate these sections of our procedure we had to address some basic theoretical issues.
PCR data processing may need to be optimized for specific PCR machines and chemistry. The discussed processing was optimized for data obtained on an Opticon Monitor 2 machine (MJ Research) using the QuantiTect SYBR Green PCR kit (Qiagen).
Smoothing
Smoothing is necessary if noticeable non-specific scattering from cycle to cycle is observed on the raw fluorescence plots. Apart from moving averages there are other more sophisticated mathematical approaches to filter this kind of noise e.g. sigmoidal fitting [13]. However, this fit is no more than a mathematical abstraction fitting PCR plot. Until the development of a genuine mathematical model of real time PCR, all other fits will not be related to PCR per se. Therefore, since simple 3 point moving average produced acceptable results there was no obvious need for more complex methods.
Background subtraction
Background subtraction is a common step in PCR data processing. Often it requires operator's involvement to choose between several available options (e.g. subtraction of minimal value through the run, subtraction of average over a certain cycle ranges, different kinds of "trends", etc). To avoid the operator involvement we always subtract the minimal value observed in the run. This option has a clear interpretation and works well. It is important that the baseline subtraction is performed after smoothing. So the noise potentially affecting minimal values has already been reduced before baseline subtraction.
Amplitude normalization
Amplitude normalization unifies plateau positions in different samples. Although amplitude normalization was available in some versions of Light-Cycler software and has been used by some researchers [14] this step still is not common in PCR data processing. The caution with regard to the amplitude normalization is probably caused by current lack of understanding of the plateau phase in PCR.
Amplitude normalization is based on the suggestion that in ideal PCR, output is determined by the initially available PCR resources. In this case PCRs prepared from the same master mix will run out of the same limiting resource in different samples. The resource can run out sooner (abundant template) or later (rare template) but finally the same amount of PCR products will be produced in all samples. This assumption is valid for ideal PCR but in practice it may not always hold (for example, non-specific PCR products may also consume PCR resources). The factors potentially leading PCR to the plateau include utilization of primers or nucleotides, thermal inactivation of DNA polymerase, competition between primers and PCR products for annealing, enzyme inactivation by PCR products and accumulation of inhibitors [15]. The plateau may also be affected by factors influencing the detection of PCR products: e.g. by PCR volume and by concentration of probe or SYBR-Green in PCR mix [14,16,17]. In practice the plateau phase is probably caused by different factors depending on the particular PCR design and PCR mix composition.
In this work we used QuantiTect SYBR Green PCR kit (Qiagen). With this kit neither increase of primers nor addition of enzyme notably affected the plateau positions (Figure 7). The fact that bands on PCR gels were remarkably enlarged by additional staining with SYBR Green (Figure 6) suggests that the plateaus observed in PCRs could had been caused simply by limited SYBR Green concentration. Therefore, in samples prepared with the same master mix, the plateau scattering could be considered as a non-specific noise and should be removed.
What may cause the plateau scattering in fluorescence plots? In certain cases, it may be optical factors. Freshwater et al [18] showed that refraction and reflection notably affects the plateau scattering in different types of tubes (Figure 8). This is in agreement with our observations in which (i) we failed to observe positive correlation between plateau positions and the volumes of bands on PCR gels and (ii) plateau scattering may be reduced by passive dye normalization (data not shown). Potentially, other factors may also play a role in plateau scattering: e.g. non-uniform evaporation across PCR plates[18].
So far, lack of understanding of the PCR plateau nature makes the amplitude normalization an optional step. When used, amplitude normalization should be empirically validated in each individual plate. Linearity of the standard curve may act as an empirical test for amplitude normalization, i.e. if the standard curve is good so the amplitude normalization does not alter the results and the procedure may be employed. Our experience is that amplitude normalization usually improves the standard curve (Figure 9).
Finally, a "PCR-specific" explanation of plateau scattering can not explain the scattering observed in PCR replicas (Figure 10A). After amplitude normalization the fluorescence plots in replicas often converge toward a single line (Figure 10B). In our experiments this reduced CV in replicas by a factor of 2 to 7. Therefore, when a marked plateau scattering is observed at a particular PCR, amplitude normalization should be considered.
Threshold selection
As long as the standard curve provides both basis and empirical validation for PCR results the threshold may be put at any level where it produces a satisfactory standard curve. At the same time, the linearity of standard curve is theoretically explained at exponential phase of PCR only. Therefore, the common practice is to put the threshold as low as possible to cross the fluorescence plots in the exponential phase. For this reason we usually restrict the search of the optimal threshold position to the lower half of the fluorescence plot.
Crossing point calculation
Currently the most established methods of crossing point calculations are the fit point method and the second derivative maximum method [4]. The fit point method reliably allocates the threshold level in the exponential phase and reduces minor inaccuracies by aggregating data from several points. The second derivative maximum method eliminates interactivity during threshold selection and baseline subtraction. These are robust and reliable methods.
Our calculation method also produces good results. In addition, it is simple and does not alter the initial mathematical definition of crossing points.
Statistical treatment of PCR replicas
The next step in the data processing is derivation of results from crossing points. Two separate issues need to be addressed during this step: (i) best-fit values and (ii) errors in replicates. Calculation of best-fit values is simple with standard curve methodology (see formulae in Additional file 1) but statistical assessment of errors in replicates requires detailed consideration.
Description and interpretation of intra-assay PCR variation
PCR uncertainty is usually characterized by coefficient of variation. This reflects the fact that the errors propagated to non-normalized values and to final results are higher at higher best-fit values. This is not always the case with the crossing points. However, coefficients of variation still may be used for rough comparison of CPs' dispersions because the CPs' absolute values vary in quite a limited range (typically between 20 and 30 cycles).
Importantly, that during PCR interpretation the statistical significance of differences between samples should not be based on intra-assay variation. Intra-PCR replicates account only for errors originated from PCR. At the same time the uncertainty in final results is usually more affected by pre-PCR steps [1]. In this case the replicates of the whole experiment (including sampling, RNA extraction and reverse transcription) are needed to derive statistical differences between samples. If the amount of starting material is limited or replicates are unavailable (for example when studying tumor biopsies) the preliminary assessment of replicates in an experimental set of similar samples is required to base statistical comparison between samples (type B evaluation of uncertainty according to Taylor and Kuyatt [19]). This type of statistical treatment is not included in the described data processing. Even though in our experiments the intra-assay PCR variation can not be directly used for statistical inferences, we routinely use it as an internal quality check for PCR.
Starting point for statistical assessment
Two different approaches may be utilized for initial statistical handling of intra-assay PCR replicates. Either CP values are first averaged and then transformed to non-normalized values or vice versa. Both approaches may yield similar results, as long as the arithmetic mean is used for the CP values and geometric mean for the non-normalized quantities. We prefer to start statistical assessment using unmodified source data i.e. we average crossing points before transformation to the non-normalized values.
Crossing point distribution in PCR replicas
To choose appropriate statistical methods to deal with crossing points, we started from the assessment of crossing points' distributions in PCR replicates. Distributions of crossing points were studied in four PCR plates each of those represented a 96× replicate. The distributions were close to the Normal (Table 2, Figure 4). Combined analysis of a number of PCR reactions, made in triplicates or quadruplicates, confirmed this result (data not shown). Therefore, Normal distribution satisfactorily reflects the distribution of crossing points in PCR replicates. This allowed us to use arithmetic mean and mean's variances to estimate best-fit values and their uncertainty in crossing points.
Error propagation
The CPs' variances were traced to final results by the law of error propagation. This assumed the normality of distributions not only in crossing points but also at the later steps of data processing. Strictly speaking, this assumption is not completely true: the data processing deforms normal distribution. Three functions are used to calculate results from crossing points: linear function (linear standard curve), exponent (calculation of non-normalized values) and ratio (normalizing by reference genes). Among them only linear function keeps normality of distribution. Exponent and ratio distort it. At the same time, the degree of the introduced distortion depends on particular numeric parameters. Analyzing the deformation of normal distribution at the parameters typical for real time PCR we found that at low initial dispersions the resulting distributions remain close to normal (Figure 5A). Therefore, the convenient parametric methods can be used in PCR data processing if crossing points' CV in replicas does not exceed 1% (for a typical PCR it roughly corresponds to crossing points' SD ≤ 0.2 and to CV in non-normalized values ≤ 14%, see Table 3). At higher initial dispersions the resulting distributions become asymmetric and require special statistical treatment (Figure 5C). Actually observed in our experiments crossing points' CVs usually were less than 0.5% (Table 2).
Additionally the analysis confirmed the remarkable increase of relative variation at each step of data processing. E.g. 2% CV at crossing points resulted to 28% CV in the non-normalized values and to 40% CV in the final results (Table 3). This also complicates interpretation of results with high dispersion in crossing points.
Standard curves
In line with the common practice, we interpreted the standard curve as an ordinary linear function ignoring its statistical nature and uncertainty because the uncertainty was usually quite small (typical coefficient of determination above 99%). With sufficient number and range of standard dilutions and proper laboratory practice it is always should be possible to produce the standard curve of sufficient quality.
Specific design of standard curves may differ for different genes depending on the variability of their expression. For relatively stabile genes (e.g. Actin beta or GAPD) we usually were able to obtain good standard curves using 5–6 two-fold dilutions. To cover the dynamic range for genes with less stable expression (e.g. Mammaglobin 1 in breast cancers) more dilutions (up to 8) and/or higher factor at each dilution (3–5 fold) were needed. We usually run standards in triplicates (as well as the target specimens).
Even though the standard curves could be quite reproducible [12] we consider the presence of standard curves on each plate to be a good laboratory practice. Additionally, there is no great economy in sharing standard curves between PCR plates, when the plates are filled up with samples. For example, 6-point standard curve in triplicates takes just 18 cells: this is less than 20% of 96-plate and less than 5% of 386-plate. Therefore sharing of standard curves reduces costs and labour only in pilot experiments with small number of samples. However, even in pilot experiments the repeatability of shared standard curves should be validated on a regular basis.
Summarizing data from several reference genes
Several reference genes are required for accurate relative quantification [1,20]. Different ways may be used to derive a single normalizing factor out of several genes. To explore this in the attached version of VBA script we made available two options: arithmetic and geometric mean.
Arithmetic mean is the most "intuitive" way. However, it has a major disadvantage: it depends on arbitrary choice of the absolute values for reference genes. For example, the normalizing factor will differ, if a reference gene is described either as a fraction of 1 (absolute values from 0 to 1) or in percents (values 0% to 100%). Importantly, this can change the relative values of the normalizing factor in different samples. In contrast, if geometric mean is used, the arbitrary choice of units for any reference gene will not affect the relative values of normalizing factor in different samples. Neither arithmetic nor geometric mean accounts for differences in uncertainties of different reference genes. In practice this implies similar variances in all reference genes. This assumption seems reasonable in most of the cases. However, if this assumption does not hold the weights reciprocal to variances could be introduced.
Obviously, the different ways of summarizing data from reference genes will produce different results. At the same time, at truly stable expression of reference genes the general tendencies in results should be similar. Currently we calculate the single normalizing factor by geometric mean, because it better fits to the relative nature of measurements as well as to the logarithmic scale of gene expression changes [20,21].
Unfortunately common practice tends to ignore the uncertainty of normalizing factor. Our procedure estimates this uncertainty using the law of error propagation (see formulae in Additional file 1).
Methods based on PCR efficiency and individual shapes of fluorescent plots
Standard curve approach was chosen for our procedure because currently PCR efficiency assessment may complicate data processing. The main complication is that actual efficiency of replication is not constant through the PCR run being high at exponential phase and gradually declining toward the plateau phase. However, most current methods of PCR efficiency assessment report "overall" efficiency as a single value. Additionally, PCR efficiency may be calculated in different ways that can "overestimate" or "underestimate" the "true" PCR efficiency [12]. In contrast, the standard curve method is based on a simple approximation of data obtained in standard dilutions to unknown samples.
At present the most popular method of PCR efficiency assessment is based on the slope of standard curve. This method does not account for PCR efficiencies in individual target samples. In contrast, recent publications on PCR efficiency assessment were concentrated on the analysis of individual shapes of fluorescence plots [8-10]. Potentially this may lead to better mathematical understanding of PCR dynamic and to new practical solutions in PCR quantification [13].
Limitations of our data processing
This section summarizes conditions that must be adhered to in order to obtain valid results with our data processing:
• all PCRs must achieve doubtless plateau and no non-specific PCR products should be observed to use amplitude normalization;
• standard curves with coefficient of determination above 99% are required to ignore uncertainty of regression and to validate the use of amplitude normalization;
• low dispersion in PCR replicates (crossing points' CV < 1% or SD < 0.2) is required to use the conventional statistical methods.
These limitations are linked: amplitude normalization provides the low dispersion in replicas needed for statistical treatment.
Conclusion
In this article we described a procedure for relative real time PCR data processing. The procedure is based on the standard curve approach, does not require PCR efficiency assessment, can be performed in fully automatic mode and provides statistical assessment of intra-assay PCR variation. The procedure has been carefully analyzed and tested. The standard curve approach was found a reliable and simple alternative to the PCR-efficiency based calculations in relative real time PCR.
Methods
Tissue samples, RNA extraction, reverse transcription
Breast cancer biopsies were taken from 21 patients before and after treatment with an aromatase inhibitor. Samples were obtained in the Edinburgh Breast Unit (Western General Hospital, Edinburgh) with patients' informed consent and ethical committee approval. Biopsies were snap frozen and stored in liquid nitrogen until RNA extraction. Before RNA extraction the frozen tissue was defrosted and stabilized in RNA-later-ICE reagent (Ambion). Total RNA was extracted with RNeasy-mini columns (Qiagen). Amount and purity of RNA were evaluated by spectrophotometer. RNA integrity was confirmed by agarose gel electrophoresis.
cDNA was synthesised with SuperScript III reverse transcriptase (Invitrogen) in accordance with the manufacturer's recommendations. Briefly:
1) oligo(dT)20 primers and dNTPs were added to total RNA,
2) the mix was heated to 65°C for 5 min and then chilled on ice,
3) first-Strand buffer, DDT, RNase inhibitor (RNaseOUT, Invitrogen) and Reverse transcriptase were added to specimens,
4) reverse transcription was carried out for 60 minutes at 50°C.
PCR
Calibrator preparation, cDNA dilution and PCR plate set up were performed as illustrated in Figure 11. Briefly:
1. Aliquots of cDNA samples running on the same plate were pooled and the pool was used as calibrator.
2. cDNAs were diluted with water prior PCR.
3. The set of samples consisting of the diluted cDNAs and the dilutions of the calibrator were used for several PCR plates: one plate for each gene.
4. For each sample the whole PCR mix including primers and cDNA was prepared before dispensing into the plate.
5. Samples were loaded to 96× PCR plates by 15 μl per cell in triplicates or quadruplicates.
Primer's sequences are given in Table 1. Primers were designed basing on the sequences published in GenBank and using Primer-3 software [22]. To avoid genomic DNA amplification the primers were either located in different exons or across exon-exon boundaries. Primers were synthesized in Sigma Genosys or in Cancer Research UK. PCR was performed using QuantiTect SYBR Green PCR kit (Qiagen), Opticon-2 PCR machine (MJ Research), white 96× PCR plates and plain PCR caps (MJ Research). The cycling parameters for all genes were the following: hot-start 95°C 15 min, 45 cycles of (denaturation 94°C 15 sec, annealing 56°C 30 sec, elongation 72°C 30 sec, plate read), final elongation 72°C 5 min, melting curve 65–95°C. Gradient PCRs confirmed 56°C as appropriate annealing temperature for all primers.
Several additional PCRs were run with different amount of primers (0.1 μM, 0.3 μM, 0.9 μM), different amount of enzyme (0.8U, 1.5U and 3.1U of HotStarTaq, Qiagen were added to 15 μl PCRs made with QuantiTect SYBR Green PCR mix, Qiagen) and different caps (domed and plain caps, MJ Research).
PCR product electrophoresis
Electrophoreses were run immediately after PCRs. 10 μl of PCR products were mixed with 2 μl of loading buffer. 6 μl of the mix per well was loaded into 10% PAAG (TBE Ready Gel, Biorad). Electrophoresis was run at 100 V for ~1 hr using MiniProtean-II cell (Biorad).
Prior electrophoresis 1 μl of 1:100 Sybr-Green-1 (Molecular Probes) was added into molecular weight marker (PCR Low Ladder Set, Sigma) but not into the PCR samples. After electrophoresis the gels were stained for 10 min in fresh prepared 1:10000 SybrGreen-1 (Molecular Probes). Photos were taken before and after staining using the GelDocMega4 gel documentation system (Uvitec).
96× PCR replicas
To study distributions of crossing points in PCR replicas four PCR plates have been run with a 96× replica on each. The distributions were evaluated using histograms, skewness and kurtosis measures, and the Kolmogorov-Smirnov test for Normality (see Table 2 and Figure 4).
Normal distribution transformation through the data processing
The transformation of Normal distribution through data processing was studied by computer simulation (Figure 12).
Basing on the above empirical observations (Table 2, Figure 4) the crossing points were simulated by sampling from the Normal. Samples of 1,000 random normal numbers were obtained using standard Excel data analysis tool. A pair of such samples was used to simulate CPs for one target and one reference genes. Then the simulated CPs were processed in the same way as real PCR data. The distributions obtained at each step of data processing were evaluated for normality by histograms, skewness and kurtosis measures, and the Kolmogorov-Smirnov test.
Parameters used in calculations were close to actual parameters typically observed in our PCRs (MeanCP = 20, Slope = -0.3, Intercept = 7). The resulted true values for non-normalized and normalized results were 10 and 1 correspondingly.
To study error propagation at different initial dispersions we performed simulations using the Normal distributions with different variances (CV 0.5%, 1%, 1.5%, 2%, 3%, 4% and 5%; the means were always 20). Detailed illustration for CV 1% is presented in Figure 12. The summary of simulation results is presented in Figure 5 and Table 3.
Excel VBA macros
The calculations where performed using MS Excel VBA script included to the electronic version of publication (see Additional file 2).
List of abbreviations
GAPD – glyceraldehyde-3-phosphate dehydrogenase
CP (CPs) – crossing point (crossing points)
SD – standard deviation
CV – coefficient of variation
r2 – coefficient of determination in linear regression
Authors' contributions
AL carried out the main body of the project including PCR, statistics and programming.
WM conceived of the study and participated in its design and co-ordination.
AK verified statistical methods and mathematical calculations.
All co-authors contributed to the manuscript preparation.
Supplementary Material
Additional File 1
Pdf file with formulae.
Click here for file
Additional File 2
ZIP file containing VBA macros (PCR1.xls), test data for the above macros
(Target1.csv, Target2.csv, Target3.csv, Target4.csv, Target5.csv,
Reference1.csv, Reference2.csv) and instruction to the above macros
(Instructions.pdf). Unzip file into a separate folder on your PC and follow
the instructions.
Click here for file
Acknowledgements
The study was supported by an educational grant from Novartis. Preliminary results were presented at 1St International qPCR Symposium (3–6 March, 2004, Freising-Weihenstephan, Germany,[23,24]). We thank Mr. Tzachi Bar for the valuable discussion during this conference.
Figures and Tables
Figure 1 Noise filtering. Axes: vertical – fluorescence, horizontal – cycle number, A Source data, B Smoothing, C Baseline subtraction, D Amplitude normalization
Figure 2 Direct calculation of crossing points.
Figure 3 Expression of Cyclin B1 mRNA in breast cancer biopsies. The observed decrease of Cyclin B1 expression after treatment was expected in most but not all cases. Bars show actual 95% confidence intervals estimated by the described statistical procedure in a set of real clinical specimens (NB – these are confidence intervals for intra-assay PCR variation only).
Figure 4 Distribution of crossing points in PCR replicas. Axes: vertical – relative frequency (%), horizontal – crossing points. Histogram represents a typical crossing points' distribution in 96× replica (Plate 1 from Table 2). The Kolmogorov-Smirnov test has not revealed significant deviations from the Normal distribution. The red line shows a Normal fit.
Figure 5 Transformation of normal distribution through data processing. Axes: vertical – relative frequency (%), horizontal – results. Red lines show Normal fits. A: At CPs' CV 0.5% the deviations from normality were not detectable using the Kolmogorov-Smirnov test. B: At CPs' CV 1% the deviations from normality were not detectable in non-normalized values though moderate deviations were detectable in final results. C: At CPs' CV 2% deviations from normality were detectable in both non-normalized values and in final results.
Figure 6 Effect of staining with SYBR Green 1 on PCR gel. A: Before staining. B: After staining. Before electrophoresis SYBR Green1 was added to marker but not to samples.
Figure 7 Effect of different factors on plateau position. A: More enzyme in blue than in red samples B: More primers in blue than in red samples C: Domed and plain caps
Figure 8 Optical factors affect the plateau scattering. SYBR Green real time PCR in frosted plates (green) and white plates (blue). Frosted plates cause increased plateau scattering because of inconsistent reflection and refraction (Reproduced from [18], with ABgene® permission).
Figure 9 Effect of amplitude normalization on standard curve.
Figure 10 Effect of amplitude normalization on plateau scattering in 96× replica. Axes: vertical – Fluorescence, horizontal – Cycle. Data for plate 3 from Table 2.
Figure 11 PCR set up.
Figure 12 Computer simulation of PCR data processing. Computer simulation of PCR data processing at 1% CV in crossing points (see Methods for details).
Table 1 Primers' sequences
Short name Full name GenBank number Primers
SCGB2A2 Mammaglobin 1 (Secretoglobin, family 2A, member 2) NM_002411 TCC AAG ACA ATC AAT CCA CAA G
AAA ATA AAT CAC AAA GAC TGC TG
SCGB2A1 Mammaglobin 2 (Secretoglobin, family 2A, member 1) NM_002407 AAG ACC ATC AAT TCC GAC ATA
CAC CAA ATG CTG TCG TAC ACT
CCNB1 Cyclin B1 NM_031966 CAT GGT GCA CTT TCC TCC TT
CAG GTG CTG CAT AAC TGG AA
CKS2 CDC28 protein kinase regulatory subunit 2 NM_001827 TTC ATG AGC CAG AAC CAC AT
CTC GTG CAC AGG TAT GGA TG
PTN Pleiotrophin (heparin binding growth factor 8, neurite growth-promoting factor 1) NM_002825 GTG CAA GCA AAC CAT GAA GA
GCT CGC TTC AGA CTT CCA GT
LPIN2 Lipin 2 NM_014646 TTG TTG CTG CAG ATT GAT CC
CCA AAT GGC AAT GGA TTT TC
ACTB Actin, beta NM_001101 GGA GCA ATG ATC TTG ATC TT
CCT TCC TGG GCA TGG AGT CCT
GAPD glyceraldehyde-3-phosphate dehydrogenase NM_002046 TGC ACC ACC AAC TGC TTA GC
GGC ATG GAC TGT GGT CAT GAG
Primers for GAPD were taken from Vandesompele et al [20]
Table 2 Crossing points' distributions observed in PCR replicas
Plate Number of replicates Mean CP SD CV Skewness Kurtosis Kolmogorov-Smirnov test
1 96 21.48 0.06 0.3% 0.1 -0.1 Normal
2 94 18.09 0.07 0.4% 1.5 5.7 Sharper than normal
3 96 20.09 0.04 0.2% 0.1 -0.3 Normal
4 96 18.13 0.10 0.5% 0.5 1.0 Normal
Table 3 Magnitude of propagated error at different steps of data processing
SD in crossing points CV in crossing points CV in non-normalized values CV in normalized results
0.1 0.5% 7% 10%
0.2 1.0% 14% 20%
0.3 1.5% 22% 31%
0.4 2.0% 28% 40%
0.6 3.0% 45% 66%
In all instances mean values are 20 in crossing points, 10 in non-normalized values and 1 in final results. See Figures 5 and 13 for more details.
==== Refs
Bustin SA Quantification of mRNA using real-time reverse transcription PCR (RT-PCR): trends and problems J Mol Endocrinol 2002 29 23 39 12200227 10.1677/jme.0.0290023
Muller PY Janovjak H Miserez AR Dobbie Z Processing of gene expression data generated by quantitative real-time RT-PCR Biotechniques 2002 32 1372 4, 1376, 1378-9 12074169
Pfaffl MW A new mathematical model for relative quantification in real-time RT-PCR Nucleic Acids Res 2001 29 e45 11328886 10.1093/nar/29.9.e45
Pfaffl MW Horgan GW Dempfle L Relative expression software tool (REST) for group-wise comparison and statistical analysis of relative expression results in real-time PCR Nucleic Acids Res 2002 30 e36 11972351 10.1093/nar/30.9.e36
Livak KJ Schmittgen TD Analysis of relative gene expression data using real-time quantitative PCR and the 2(-Delta Delta C(T)) Method Methods 2001 25 402 408 11846609 10.1006/meth.2001.1262
Roshe Applied Science Overview of LightCycler Quantification Methods Technical Note No LC 10 2003
Applied Biosystems Guide to Performing Relative Quantitation of Gene Expression Using Real-Time Quantitative PCR 2004
Tichopad A Dilger M Schwarz G Pfaffl MW Standardized determination of real-time PCR efficiency from a single reaction set-up Nucleic Acids Res 2003 31 e122 14530455 10.1093/nar/gng122
Liu W Saint DA A new quantitative method of real time reverse transcription polymerase chain reaction assay based on simulation of polymerase chain reaction kinetics Anal Biochem 2002 302 52 59 11846375 10.1006/abio.2001.5530
Bar T Stahlberg A Muszta A Kubista M Kinetic Outlier Detection (KOD) in real-time PCR Nucleic Acids Res 2003 31 e105 12930979 10.1093/nar/gng106
Rutledge RG Cote C Mathematics of quantitative kinetic PCR and the application of standard curves Nucleic Acids Res 2003 31 e93 12907745 10.1093/nar/gng093
Pfaffl MW Bustin SA Quantification strategies in real time PCR A-Z of quantitative PCR IUL biotechnology series ; 5 2004 La Jolla, CA, International University Line
Rutledge RG Sigmoidal curve-fitting redefines quantitative real-time PCR with the prospective of developing automated high-throughput applications Nucleic Acids Res 2004 32 e178 15601990 10.1093/nar/gnh177
Wittwer CT Herrmann MG Moss AA Rasmussen RP Continuous fluorescence monitoring of rapid cycle DNA amplification Biotechniques 1997 22 130 1, 134-8 8994660
Kainz P The PCR plateau phase - towards an understanding of its limitations Biochim Biophys Acta 2000 1494 23 27 11072065
Zipper H Lämmle K Buta C Brunner H Bernhagen J Vitzthum F Investigations on the binding of SYBR Green I to double-stranded (ds)DNA: In Proceedings of the joint annual fall meeting , German Society for Biochemistry and Molecular Biology (GBM) & German Society for Expermental and Clinical Pharmacology and Toxicology (DGPT) September 7-10 2002; Halle (Saale), Germany. 2002 177
Vitzthum F Geiger G Bisswanger H Brunner H Bernhagen J A quantitative fluorescence-based microplate assay for the determination of double-stranded DNA using SYBR Green I and a standard ultraviolet transilluminator gel imaging system Anal Biochem 1999 276 59 64 10585744 10.1006/abio.1999.4298
Freshwater S van der Valk A O'Shaughnessy M Ng S Baker S Pfaffl MW The effect of consumable type on the sensitivity and reproducibility of qPCR: In Proceedings of the 1st International qPCR Symposium and Application Workshop 3rt - 6th March 2004; Freising-Weihenstephan, Germany. 2004 88
Taylor BN Kuyatt CE Guidelines for evaluating and expressing the uncertainty of NIST measurement results NIST technical note ; 1297 1994 1994 Gaithersburg, MD, U.S. Department of Commerce, Technology Administration, National Institute of Standards and Technology 20 p.
Vandesompele J De Preter K Pattyn F Poppe B Van Roy N De Paepe A Speleman F Accurate normalization of real-time quantitative RT-PCR data by geometric averaging of multiple internal control genes Genome Biol 2002 3 RESEARCH0034 12184808 10.1186/gb-2002-3-7-research0034
Szabo A Perou CM Karaca M Perreard L Quackenbush JF Bernard PS Statistical modeling for selecting housekeeper genes Genome Biol 2004 5 R59 15287981 10.1186/gb-2004-5-8-r59
Rozen S Skaletsky HJ Krawetz S and Misener S Primer3 on the WWW for general users and for biologist programmers Bioinformatics Methods and Protocols: Methods in Molecular Biology 2000 Totowa, NJ,, Humana Press 365 386
Larionov AA Hulme MJ Miller WR Pfaffl MW Amplitude normalization in real time PCR data processing: 3rt - 6th March 2004; Freising-Weihenstephan, Germany. 2004 56 57
Larionov AA Miller WR Pfaffl MW Data processing in real time PCR: In Proceedings of the 1st International qPCR Symposium and Application workshop Freising-Weihenstephan, Germany. 2004 28 29
|
15780134
|
PMC1274258
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 21; 6:62
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-62
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-641578014610.1186/1471-2105-6-64SoftwareMBEToolbox: a Matlab toolbox for sequence data analysis in molecular biology and evolution Cai James J [email protected] David K [email protected] Xuhua [email protected] Kwok-yung [email protected] Department of Microbiology, University of Hong Kong, Pokfulam, Hong Kong, China2 Department of Biochemistry, University of Hong Kong, Pokfulam, Hong Kong, China3 Department of Biology, University of Ottawa, Canada2005 22 3 2005 6 64 64 3 12 2004 22 3 2005 Copyright © 2005 Cai et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
MATLAB is a high-performance language for technical computing, integrating computation, visualization, and programming in an easy-to-use environment. It has been widely used in many areas, such as mathematics and computation, algorithm development, data acquisition, modeling, simulation, and scientific and engineering graphics. However, few functions are freely available in MATLAB to perform the sequence data analyses specifically required for molecular biology and evolution.
Results
We have developed a MATLAB toolbox, called MBEToolbox, aimed at filling this gap by offering efficient implementations of the most needed functions in molecular biology and evolution. It can be used to manipulate aligned sequences, calculate evolutionary distances, estimate synonymous and nonsynonymous substitution rates, and infer phylogenetic trees. Moreover, it provides an extensible, functional framework for users with more specialized requirements to explore and analyze aligned nucleotide or protein sequences from an evolutionary perspective. The full functions in the toolbox are accessible through the command-line for seasoned MATLAB users. A graphical user interface, that may be especially useful for non-specialist end users, is also provided.
Conclusion
MBEToolbox is a useful tool that can aid in the exploration, interpretation and visualization of data in molecular biology and evolution. The software is publicly available at and .
==== Body
Background
MATLAB integrates programming, visualization and computation in an easy-to-use environment and is widely used in scientific and engineering studies. One of the most attractive features of MATLAB is that the basic data element of the system is a matrix that does not require dimensioning. This allows users to solve many technical computing problems, especially those with matrix and vector formulations, in a very effective way. The MATLAB environment itself offers a comprehensive set of built-in functions and many toolboxes have been developed, and are often freely available, for more specialized needs.
However, to our knowledge, these advantages in the MATLAB environment have not been fully utilized in the area of molecular biology and evolution. Only a few MATLAB toolboxes or functions are freely available for data analysis, exploration, and visualization of nucleotide and protein sequences. MATHWORKS has recently provided a bioinformatics toolbox, however this toolbox has relatively limited functions for molecular evolutionary studies. MBEToolbox, is presented here to fulfil the most obvious needs in sequence manipulation, genetic distance estimation and phylogeny inference under the MATLAB environment. Moreover, this toolbox provides an extensible, functional framework to formulate and solve problems in evolutionary data analysis. It facilitates the rapid construction of both general applications, as well as special-purpose tools for evolutionary biologists, in a fraction of the time it would take to write a program in a scalar, noninteractive language such as C or FORTRAN.
Implementation
MBEToolbox is written in the MATLAB language and has been tested on the WINDOWS platform with MATLAB version 6.1.0. The main functions implemented are: sequence manipulation, computation of evolutionary distances derived from nucleotide-, amino acid- or codon-based substitution models, phylogenetic tree construction, sequence statistics and graphics functions to visualize the results of analyses. Although it implements only a small fraction of the multiplicity of existing methods used in molecular evolutionary analyses, interested users can easily extend the toolbox.
Input data and formats
MBEToolbox requires a single ASCII file containing the nucleotide or amino acid sequence alignment in either PHYLIP [1], CLUSTALW [2] or fasta format. The toolbox does provide a built-in CLUSTALW [2] interface if an unaligned sequence file is provided. Protein-coding DNA sequences can be automatically aligned based on the corresponding protein alignment with the command alignseqfile.
After input, in common with the MATHWORKS bioinformatics toolbox, MBEToolbox represents the alignment as a numeric matrix with every element standing for a nucleic or amino acid character. Nucleotides A, C, G and T are converted to integers 1 to 4, and the 20 amino acids are converted to integers 1 to 20. A header, containing information about the names and type of the sequences as well as the relevant genetic code for protein-coding nucleotides, is attached to the alignment matrix to form a MATLAB structure. An example alignment structure, aln, in MATLAB code follows:
aln =
seqtype: 2
geneticcode: 1
seqnames: {1 × n cell}
seq: [n × m double]
where n is the number of sequences and m is the length of the aligned sequences. The type of sequence is denoted by 1, 2 or 3 for sequences of non-coding nucleotides, protein coding nucleotides and amino acids, respectively.
Sequence manipulation and statistics
The alignment structure, aln, can be manipulated using the MATLAB language. For example, aln.seq(x,:) will extract the xth sequence from the alignment, while aln.seq(:, [i: j]) will extract columns i to j from the alignment. Users may easily extract more specific positions by using functions developed in the toolbox, such as extractpos(aln, 3) or extractdegeneratesites to obtain the third codon positions or fourfold degenerate sites, respectively. For each sequence, some basic statistics such as the nucleotide composition (ntcomposition) and GC content, can be reported. Other functions include the calculation of the relative synonymous codon usage (RSCU) and the codon adaptation index (CAI), counts of segregating sites, taking the reverse complement or translating a sequence, and determining the sequence complexity.
Evolutionary distances
The evolutionary distance is one of the important measures in molecular evolutionary studies. It is required to measure the diversity among sequences and to infer distance-based phylogenies. MBEToolbox contains a number of functions to calculate evolutionary distances based on the observed number of differences. The formulae used in these functions are analytical solutions of a variety of Markov substitution models, such as JC69 [3], K2P [4], F84 [1], HKY [5] (see [6] for detail). Given the stationarity condition, the most general form of Markov substitution models is the General Time Reversible (GTR or REV) model [7-10]. There is no analytical formula to calculate the GTR distance directly. A general method, described by Rodriguez et al. [9], has been implemented here. In this method a matrix F, where Fij denotes the proportion of sites for which sequence 1 (s1) has an i and sequence 2 (s2) has a j, is formed. The GTR distance between s1 and s2 is then given by
where ∏ denotes the diagonal matrix with values of nucleotide equilibrium frequencies on the diagonal, and tr(A)denotes the trace of matrix A. The above formula can be expressed in MATLAB syntax directly as:
>> d = -trace(PI*logm(inv(PI)*F))
MBEToolbox also calculates the gamma distribution distance and the LogDet distance [11] (i.e., Lake's paralinear distance [12]).
For alignments of codons, the toolbox provides calculation or estimation of the synonymous (Ks) and non-synonymous (Ka) substitution rates by the counting method of Nei and Gojobori [13], the degenerate methods of Li, Wu and Luo [14] and the method of Li or Pamilo and Bianchi [15,16], as well as the maximum likelihood method through PAML [17]. All these methods for calculating Ks and Ka require that the input sequences are aligned in the appropriate reading frame, which can be performed by the function alignseqfile. Unresolved codon sites will be removed automatically. In addition, several quantities, including the number of substitutions per site at only synonymous sites, at only non-synonymous sites, at only four-fold-degenerate sites, or at only zero-fold-degenerate sites can be calculated. The output from these calculations are distance matrices which can be exported into text or excel files, or used directly in further operations.
Phylogeny inference
Two distance-based tree creation algorithms, Unweighted Pair Group Method with Arithmetic mean (UPGMA) and neighbor-joining (NJ) [18] are provided and trees from these methods can be displayed or exported. Maximum parsimony and maximum likelihood algorithms can be applied to nucleotide or amino acid alignments through an interface to the phylip package [1]. As properly implemented maximum likelihood methods are the best vehicles for statistical inference of evolutionary relationships among species from sequence data, several maximum likelihood functions have been explicitly implemented in MBEToolbox. These functions allow users to incorporate various evolutionary models, estimate parameters and compare different evolutionary trees.
The simplest case of estimation of the evolutionary distance between two sequences, s1 and s2, can be considered as the estimation of the branch length (the number of substitutions along a branch) separating ancestor and descendent nodes. Branch lengths, relative to a calibrated molecular clock, can reveal the time interval for this separation. A continuous time Markov process is generally used to model evolution along the branch from s1 to s2. A transition rate matrix, Q, is used to indicate the rate of changing from one state to another. For a specified time interval or distance, t, the transition probability matrix is calculated from P(t) = eQt. If there are N sites, the full likelihood is
In this equation, and are the ith bases of sequences 1 and 2 respectively; is the expected frequency of base .
In MBEToolbox, to calculate the likelihood, L, at a given time interval (or distance) t, we have to specify a substitution model by using an appropriate model defining function, such as modeljc, modelk2p or modelgtr for non-coding nucleotides, modeljtt or modeldayhoff for amino acids, or modelgy94 for codons. These functions return a model structure composed of an instantaneous rate matrix, R, and an equilibrium frequency vector, pi which give Q, (Q = R*diag(pi)). Once the model is specified, the function likelidist(t, model, s1, s2) can calculate the log likelihood of the alignment of the two sequences, s1 and s2, with respect to the time or distance, t, under the substitution model, model.
In most cases we wish to estimate t instead of calculating L as a function of t, so the function optimlikelidist (model, s1, s2) will search for the t that maximises the likelihood by using the Nelder-Mead simplex (direct search) method, while holding the other parameters in the model at fixed values. This constraint can be relaxed by allowing every parameter in the model to be estimated by functions, such as optimlikelidistk2p, that can estimate both t and the model's parameters. Figure (1a and 1b) illustrates the estimation of the evolutionary distance between two ribonuclease genes through the fixed- and free-parameter K2P models, respectively. When the K2P model's parameter, kappa, is fixed, the result and trace of the optimisation process is illustrated by the graph of L and t (Fig. 1a). When kappa is a free parameter, a surface shows the result and trace of the optimisation process (Fig. 1b).
When calculating the likelihood of a phylogenetic tree, where s1 and s2 are two (descendant) nodes in a tree joined to an internal (ancestor) node, sa, we must sum over all possible assignments of nucleotides to sa to get the likelihood of the distance between s1 and s2. Consequently, the number of possible combinations of nucleotides becomes too large to be enumerated for even moderately sized trees. The pruning algorithm introduced by Felsenstein [19] takes advantage of the tree topology to evaluate the summation in a computationally efficient (but mathematically equivalent) manner. This and a simple and elegant mapping from a 'parentheses' encoding of a tree to the matrix equation for calculating the likelihood of a tree, developed in the MATLAB software, PHYLLAB [20], have been adopted in likelitree.
Combination of functions
Basic operations can be combined to give more complicated functions. A simple combination of the function to extract the fourfold degenerate sites with the function to calculate GC content produces a new function (countgc4) that determines the GC content at 4-fold degenerate sites (GC4). A subfunction for calculating synonymous and nonsynonymous differences between two codons, getsynnonsyndiff, can be converted into a program for calculating codon volatility [21] with trivial effort. Similarly, karlinsig which returns Karlin's genomic signature (the dinucleotide relative abundance or bias) for a given sequence can be easily re-formulated to estimate relative di-codon frequencies, which may be a new index of biological signals in a coding sequence. In addition, the menu-driven user interface, MBEGUI, is also a good example illustrating the power of combination of basic MBEToolbox functions.
Graphics and GUI
Good visualisation is essential for successful numerical model building. Leveraging the rich graphics functionality of MATLAB, MBEToolbox provides a number of functions that can be used to create graphic output, such as scatterplots of Ks vs Ka, plots of the number of transitions and transversions against genetic distance, sliding window analyses on a nucleotide sequence and the Z-curve (a 3-dimensional curve representation of a DNA sequence [22]). A simple menu-driven graphical user interface (GUI) has been developed by using GUIDE in MATLAB. The top menu includes File, Sequences, Distances, Phylogeny, Graph, Polymorphism and Help submenus (Fig. 2). It aids the usage of the most frequently required functions so that users do not have to run any scripts or functions from the MATLAB command line in most cases.
Results and discussion
Vectorization simplifies programming
MATLAB is a matrix language, which means it is designed for vector and matrix operations. Programming can be simplified and made more efficient by using algorithms that take advantage of vectorization (converting for and while loops to the equivalent vector or matrix operations). The MATLAB compiler in version 7.0 will automatically recognize and vectorize loops without recursion. An example of vectorization is the calculation of Z-scores [23] for Smith-Waterman alignments [24] to give a measure of the significance of an alignment score against a background of scores from randomly generated sequences with the same composition and length. Hence, Z-scores are designed to overcome the bias due to the composition of the alignment and are usually calculated by comparing an actual alignment score with the scores obtained on a set of random sequences generated by a Monte-Carlo process. The Z-score is defined as:
Z(A, B) = (S(A, B) - mean)/standard deviation
where S(A, B) is the Smith-Waterman (S-W) score between two sequences A and B. The mean and standard deviation are taken from realignments of the permuted sequences. The algorithm is implemented as follows in MATLAB with as few as 15 lines of code:
function [z,z_raw] = zscores(s1,s2,nboot)
ml = length(s1);
m2 = length(s2);
% Initialise two vectors holding Z-score of
% s1_rep and s2_rep, i.e., replicate samples
% of sequences s1 and s2.
v_z1 = zeros(1,nboot);
v_z2 = zeros(1,nboot);
z_raw = smithwaterman(s1,s2);
for (k = 1:nboot),
s1_rep = s1(:,randperm(m1));
v_z1(1,k) = smithwaterman(s1_rep, s2) ;
s2_rep = s2(:,randperm(m2));
v_z2(1,k) = smithwaterman(s1, s2_rep);
end
z1 = (z_raw-mean(v_z1))./std(v_z1);
z2 = (z_raw-mean(v_z2))./std(v_z2);
z = min(z1,z2);
where randperm(n) is a vector function returning a random permutation of the integers from 1 to n and smithwaterman performs local alignment by the standard dynamic programming technique.
Extensibility
An important distinction between compiled languages with subroutine libraries and interactive environments like MATLAB is the ease with which problems can be specified and solved in the latter. Moreover, MATLAB toolboxes are traditionally organised in a less object-oriented mode and, consequently, functions are more independent of each other and easier to combine and extend. Several examples were given in the Implementation section.
Comparison with other toolboxes
Some other toolboxes have been developed in MATLAB for bioinformatics related analyses. These include PHYLLAB [20] and MATARRAY [25] as well as the bioinformatics toolbox developed by MATHWORKS. Other examples can be found at the link and file exchange maintained at MATLAB CENTRAL [26]. PHYLLAB is a molecular phylogeny toolbox which also provides some functions for sequence and tree input and manipulation. Its main focus is on creating a maximum likelihood tree based on Bayesian principles using a Markov chain Monte Carlo method to compute posterior parameter distributions. MATARRAY is focussed on the analysis of gene expression data from microarrays and provides normalization and clustering functions but does not address molecular evolution. The bioinformatics toolbox from MATHWORKS provides a range of bioinformatics functions, including some related to molecular evolution.
MBEToolbox provides a much broader range of molecular evolution related functions and phylogenetic methods than either the more specialized PHYLLAB project or the more general bioinformatics toolbox from MATHWORKS. These extra functions include IO in PHYLIP format, statistical and sequence manipulation functions relevant to molecular evolution (e.g. count segregating sites), evolutionary distance calculation for nucleic and amino acid sequences, phylogeny inference functions and graphic plots relevant to molecular evolution (e.g. Ka vs Ks). As such it makes an important contribution to the bioinformatics analyses that can be performed in the MATLAB environment.
A novel enhanced window analysis
To test for the selective pressures in the different lineages of a phylogenetic tree, the nonsynonymous to synonymous rate ratio (Ka/Ks) is normally estimated [27-29]. Values of Ka/Ks = 1, > 1, or < 1 indicate neutrality, positive selection, or purifying selection, respectively. However, Ks and Ka are measurements of average synonymous and nonsynonymous substitutions per site along the whole length of the sequences. Average Ks and Kavalues give neither the pattern of intragenic fluctuation of selective constraints, nor region- or site-specific information. A sliding window method is usually adopted to examine the intragenic pattern of the substitution rates and to test for the occurrence of significant clusters of variant regions [30-33]. Significant heterogeneity in Ks would indicate that the neutral substitution rate varies across the gene, whereas heterogeneity in Kamay indicate that selective constraints vary along the gene. The results and accuracy of sliding window methods, either overlapping or non-overlapping, depend on both the size of the window and the moving distance adopted. Large window lengths may obliterate the details of patterns in Ks or Ka, whereas small window lengths usually result in larger statistical fluctuations. Hence, the resolution of a sliding window is usually limited.
A mathematical formalism, similar to the Z'-curve [34], is introduced here to solve this problem. Consider a subsequence based analysis of Ks or Ka. In the n-th step, count the cumulative numbers of Ks or Ka occurring from the first to the n-th nucleotide position in the gene sequences being inspected. Let denote either Ks or Ka and denote the cumulative at the n-th sequence position. is usually an approximately mono-increasing linear function of n. The points (, n), n = 1, 2, ..., N are fit by a least square method to a linear function, f() = βn, to give a straight line with β being its slope. We define
The two-dimensional curve of ( ~ n) gives an alternative representation of the normal sliding window curve.
To compare these two curve representations, the example dataset of Suzuki and Gojobori [35], which contains the coding regions of two hepatitis C virus strains (HCV-JS – Genbank Acc.: D85516 and HCV-JT – Genbank Acc.: D11168), was used. The entire coding sequence is divided into eight regions (C, El, E2, NS2, NS3, NS4, NS5A, NS5B). Some of the coding regions have been combined as these short ORFs are unlikely to yield meaningful Ks and Ka values. The reduction of Ks in the C, El and NS5B regions, as well as its elevation in NS3, which have been shown in previous studies [35], are not clear in a standard sliding window representation (Fig. 3a). In contrast a sharp increase in the ( ~ n) curve (Fig. 3b), indicates an increase in , while a drop in the curve indicates a decrease in . This new method has been implemented in the function plotSlidingKaKs. Since it is derived from the sliding window method, it is called the enhanced sliding window method.
Limitations
The current version of this toolbox lacks novel algorithms yet it implements a variety of existing algorithms. There are some limitations in the practical use of MBEToolbox. First, though the toolbox provides many methods to infer and handle sequence and evolutionary analyses, the full range of these features can only be accessed through the MATLAB command line interface, as in the majority of MATLAB packages. Second, some of the functions cannot handle ambiguous nucleotide or amino acid codes in the sequences. The future development of MBEToolbox will overcome these present limitations.
Conclusion
The MBEToolbox project is an ongoing effort to provide an easy-to-use, yet powerful, analysis environment for molecular biology and evolution. Currently, it offers a substantial set of frequently used functions to manipulate sequences, to calculate genetic distances, to infer phylogenetic trees and related analyses. MBEToolbox is a useful tool which should inspire evolutionary biologists to take advantage of the MATLAB environment.
Availability and requirements
Project name: MBEToolbox
Project web page:
Operating system: WINDOWS 95/98/2000/XP
Programming language: MATLAB 6.0 or higher
Other requirements: Statistics Toolbox
License: GPL
Any restrictions on use by non-academics: License needed
Authors' contributions
JJC designed and implemented the software and wrote the draft of the manuscript. DKS participated in the design and revised the manuscript. XX participated in the design and provided suggestions for future development. KYY supervised and participated in the design of the study. All authors read and approved the final version of the manuscript.
Acknowledgements
This work was supported by the AIDS Trust Fund (MSS 083), Research Grant Council Grant (HKU 7363/03M), and University Development Fund, University of Hong Kong. JJC would like to thank Dr. Nam-Kiu Tsing (Department of Mathematics, University of Hong Kong) for valuable discussions, and Dr. Richard E. Strauss (Department of Biological Sciences, Texas Tech University) for allowing his NJ routine to be adapted for MBEToolbox and for releasing his MATLAB library to the public.
Figures and Tables
Figure 1 Log-likelihood of evolutionary distance. (a) Likelihood as function of K2P distance. The distance is estimated by maximising the likelihood of the alignment with the bias of transitions to transversions, kappa, held fixed. (b) Likelihood as a function of distance and kappa. Both the distance and kappa are optimised simultaneously. The maximum likelihood peaks are marked with *. The two sequences used are the coding regions of Tamarin eosinophil-derived neurotoxin (Acc. No.: U24099) and human eosinophil cationic gene (Acc. No: NM_002935).
Figure 2 MBEToolbox GUI. (a) Distances submenu; (b) Phylogeny submenu; (c) Graph submenu; and (d) Polymorphism submenu.
Figure 3 A comparison between sliding window and enhanced sliding window methods. Sliding window analysis of Ks and Ka for the concatenated coding regions of two hepatitis C virus strains, HCV-JS and HCV-JT. The number of codons for the C, El, E2, NS2, NS3, NS4, NS5A, and NS5B genes are 191, 192, 426, 217, 631, 315, 447, and 591, respectively. The different coding regions are separated by vertical lines. (a) illustrates the result of a normal sliding window analysis; (b) illustrates the result of the enhanced sliding window analysis. Beginnings and ends of regions poor in synonymous substitutions (slope < 0) are indicated by the arrows a and b (genes C and El) and e and f (gene NS5B). A region rich in synonymous substitutions (slope > 0) in gene NS3 is indicated by arrows c and d.
==== Refs
Felsenstein J PHYLIP – Phylogeny Inference Package (Version 3.2) Cladistics 1989 5 164 166
Thompson J Higgins D Gibson T CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice Nucl Acids Res 1994 22 4673 4680 7984417
Jukes TH Cantor C Munro HN Evolution of protein molecules Mammalian Protein Metabolism 1969 New York: Academic Press 21 132
Kimura M A simple method for estimating evolutionary rates of base substitutions through comparative studies of nucleotide sequences J Mol Evol 1980 16 111 120 7463489
Hasegawa M Kishino H Yano T Dating of the human-ape splitting by a molecular clock of mitochondrial DNA J Mol Evol 1985 22 160 174 3934395
Nei M Kumar S Molecular evolution and phylogenetics 2000 Oxford, UK: Oxford University Press
Lanave C Preparata G Saccone C Serio G A new method for calculating evolutionary substitution rates J Mol Evol 1984 20 86 93 6429346
Tavare S Some probabilistic and statistical problems in the analysis of DNA sequences Lectures on Mathematics in the Life Sciences 1986 17 57 86
Rodriguez F Oliver JL Marin A Medina JR The general stochastic model of nucleotide substitution J Theor Biol 1990 142 485 501 2338834
Yang Z Estimating the pattern of nucleotide substitution J Mol Evol 1994 39 105 111 8064867
Steel MA Recovering a tree from the leaf colourations it generates under a Markov model Appl Math Lett 1994 7 19 32 10.1016/0893-9659(94)90024-8
Lake JA Reconstructing evolutionary trees from DNA and protein sequences: paralinear distances Proc Natl Acad Sci USA 1994 91 1455 1459 8108430
Nei M Gojobori T Simple methods for estimating the numbers of synonymous and nonsynonymous nucleotide substitutions Mol Biol Evol 1986 3 418 426 3444411
Li WH Wu CI Luo CC A new method for estimating synonymous and nonsynonymous rates of nucleotide substitution considering the relative likelihood of nucleotide and codon changes Mol Biol Evol 1985 2 150 174 3916709
Li WH Unbiased estimation of the rates of synonymous and nonsynonymous substitution J Mol Evol 1993 36 96 99 8433381
Pamilo P Bianchi NO Evolution of the Zfx and Zfy genes: rates and interdependence between the genes Mol Biol Evol 1993 10 271 281 8487630
Yang Z Phylogenetic Analysis by Maximum Likelihood (PAML) Version 30 2000 London: University College
Saitou N Nei M The neighbor-joining method: a new method for reconstructing phylogenetic trees Mol Biol Evol 1987 4 406 425 3447015
Felsenstein J Evolutionary trees from DNA sequences: a maximum likelihood approach J Mol Evol 1981 17 368 376 7288891
Rzhetsky A Morozov P Markov chain Monte Carlo computation of confidence intervals for substitution-rate variation in proteins Pac Symp Biocomput 2001 6 203 214 11262941
Plotkin JB Dushoff J Fraser HB Detecting selection using a single genome sequence of M. tuberculosis and P. falciparum Nature 2004 428 942 945 15118727 10.1038/nature02458
Zhang R Zhang CT Z curves, an intutive tool for visualizing and analyzing the DNA sequences J Biomol Struct Dyn 1994 11 767 782 8204213
Pearson WR Lipman DJ Improved tools for biological sequence comparison Proc Natil Acad Sci U S A 1988 85 2444 2448
Smith TF Waterman MS Identification of common molecular subsequences J Mol Biol 1981 147 195 197 7265238 10.1016/0022-2836(81)90087-5
Venet D MatArray: a Matlab toolbox for microarray data Bioinformatics 2003 19 659 660 12651728 10.1093/bioinformatics/btg046
MATLAB Central
Sharp PM In search of molecular darwinism Nature 1997 385 111 112 8990106 10.1038/385111a0
Akashi H Within- and between-species DNA sequence variation and the 'footprint' of natural selection Gene 1999 238 39 51 10570982 10.1016/S0378-1119(99)00294-2
Crandall K Kelsey C Imamichi H Lane H Salzman N Parallel evolution of drug resistance in HIV: failure of nonsynonymous/synonymous substitution rate ratio to detect selection Mol Biol Evol 1999 16 372 382 10331263
Clark AG Kao T Nonsynonymous Substitution at Shared Polymorphic Sites Among Self-Incompatibility Alleles of Solanaceae Proc Natl Acad Sci USA 1991 88 9823 9827 1946408
Ina Y ODEN: a program package for molecular evolutionary analysis and database search of DNA and amino acid sequences Comput Appl Biosci 1994 10 11 12 8193950
Endo T Ikeo K Gojobori T Large-scale search for genes on which positive selection may operate Mol Biol Evol 1996 13 685 690 8676743
Choi SS Lahn BT Adaptive evolution of MRG, a neuron-specific gene family implicated in nociception Genome Res 2003 13 2252 2259 14525927 10.1101/gr.1431603
Zhang CT Wang J Zhang R A novel method to calculate the G+C content of genomic DNA sequences J Biomol Struct Dyn 2001 19 333 341 11697737
Suzuki Y Gojobori T Salemi M, Vandamme A Analysis of coding sequences The phylogenetic handbook: a practical approach to DNA and protein phylogeny 2003 Cambridge, UK: Cambridge University Press 283 311
|
15780146
|
PMC1274259
|
CC BY
|
2021-01-04 16:02:51
|
no
|
BMC Bioinformatics. 2005 Mar 22; 6:64
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-64
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-651578415210.1186/1471-2105-6-65Methodology Article"Harshlighting" small blemishes on microarrays Suárez-Fariñas Mayte [email protected] Asifa [email protected] Knut M [email protected] Center for Studies in Physics and Biology, The Rockefeller University, 1230 York Ave, Box 212, New York, NY 10021, USA2 Laboratory of Investigative Dermatology, The Rockefeller University, 1230 York Ave, Box 178, New York, NY 10021, USA3 General Clinical Research Center, The Rockefeller University, 1230 York Ave, Box 322, New York, NY 10021, USA2005 22 3 2005 6 65 65 26 10 2004 22 3 2005 Copyright © 2005 Suárez-Fariñnas et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Microscopists are familiar with many blemishes that fluorescence images can have due to dust and debris, glass flaws, uneven distribution of fluids or surface coatings, etc. Microarray scans show similar artefacts, which affect the analysis, particularly when one tries to detect subtle changes. However, most blemishes are hard to find by the unaided eye, particularly in high-density oligonucleotide arrays (HDONAs).
Results
We present a method that harnesses the statistical power provided by having several HDONAs available, which are obtained under similar conditions except for the experimental factor. This method "harshlights" blemishes and renders them evident. We find empirically that about 25% of our chips are blemished, and we analyze the impact of masking them on screening for differentially expressed genes.
Conclusion
Experiments attempting to assess subtle expression changes should be carefully screened for blemishes on the chips. The proposed method provides investigators with a novel robust approach to improve the sensitivity of microarray analyses. By utilizing topological information to identify and mask blemishes prior to model based analyses, the method prevents artefacts from confounding the process of background correction, normalization, and summarization.
==== Body
Background
Analysis of hybridized microarrays starts with scanning the fluorescent image. For high-density oligonucleotide arrays (HDONAs) such as Affymetrix GeneChip® oligonucleotide (Affy) arrays, the focus of this paper, each scanned image is stored pixel-by-pixel in a 'DAT' file. As the first step in measuring intensity of the hybridization signal, a grid is overlaid, the image is segmented into spots or features, and the pixel intensities within each of these are summarized as a probe intensity estimate (See reviews [1] and [2] for cDNA chips). The probe-level intensity estimates are stored in a 'CEL' file. Each gene is represented by pairs of probes, each representing another characteristic sequences and a 'mismatch', which is identical, except for the Watson-Crick complement in the center. Expression of a gene is estimated from such a probe set by applying algorithms for background correction, normalization, and summarization.
The quality of data scanned from a microarray is affected by a plethora of potential confounders, which may act during printing/manufacturing, hybridization, washing, and reading. Each chip contains a number of probes specifically designed to assess the overall quality of the biochemistry, such as 'checkerboards' in the corners and borders, whose purpose is, e.g., to indicate problems with the biotinylated B2 hybridization. Affymetrix software provides for a number of criteria to assess the overall quality of a chip, such as percent present calls, scaling factor, background intensity, and overall pixel-to-pixel variation (raw Q). Software packages such as Bioconductor for R [3] have implemented biochemical quality control tools such as RNA degradation plots. If a quality problem is found, however, these criteria and tools do not easily suggest a remedy and they have little sensitivity to detect localized artefacts, like a speck of dust or a localized hybridization problem. Although such physical blemishes obviously affect the expression estimates, they have hitherto been only narrowly addressed in the literature. Thus, there are currently no safeguards to signal potential physical blemishes. Instead, researchers are merely advised to carefully inspect the chip images visually [4,5]. Given the high variance among the hundreds of thousands of probes and their random allocation on the chip, it is impossible to visually detect any but the starkest artefacts. For two-colour cDNA arrays, a Bayesian network approach has been proposed [6], based on the 'features' of the pixel distribution within each probe, yet, due to the standardized manufacturing process, the probes on an oligonucleotide array have too few 'features' for such an approach to be effective.
As the price of microarrays continues to drop, a typical microarray experiment now contains several chips, each representing a sample obtained under conditions that were similar except for the experimental factor under investigation. Having collections of chips available offers new strategies not only for analyzing the effect of the experimental factor, but also for identifying blemishes. The power of having several chips available was first harnessed for estimating mRNA expression levels by the 'robust multichip average' (RMA) method [7]. One of the assumptions underlying the RMA model is that probes across chips are highly correlated, due to differences in their affinity [8,9] and because only a small proportion of genes are differentially expressed in any experimental setting. This correlation should be even higher for the mismatches, because they are less likely to be affected by the specific changes in gene expression induced by the experimental factor. Given the volume of pixel level data, (>50 megapixels per image) it is desirable to devise algorithms that work from the 100 times smaller probe level files, the same information used in traditional signal value estimation approaches.
Figure 1 shows how the large probe-to-probe variance can obscure the subtle changes caused by all but the starkest blemishes. Since probes vary in intensity by four orders of magnitude, a biologically relevant change of 30% in brightness in a small region can easily go undetected. In fact one of the chips shown in Figure 1 is affected by several blemishes. However, Figure 1 shows that not only do the internal standards have a very consistent pattern, as one would expect, but also that high expression values are correlated across chips. As we will show, drawing on these correlations allows for a simple and efficient method to identify areas on individual chips where the assumption of spatially uncorrelated errors is clearly violated. We shall use the chip-to-chip correlation to devise a 'harshlight' that makes the blemishes stand out.
Results
Data
Psoriasis is thought to be due to an overly active immune system [10,11]. To study how the immune response of leukocytes isolated from blood can be affected by drugs that may serve to control autoimmune diseases like psoriasis, blood was drawn from five volunteers under a protocol that had been approved by The Rockefeller University Hospital Institutional Review Board [12].
For each subject, peripheral blood mononuclear cells (PBMCs) were isolated and cultured in six Petri dishes. Four cultures were activated with an anti CD3/CD28 antibody, two of which were pre-treated with a repressor drug. Two cultures served as control without drug or activation. One of the two sets of control, activated, and pre-treated cultures (subject 1 and 2) was analyzed after 6, the other after 24 hrs. (For subjects 3, 4, and 5, only one time point is available.) All samples were hybridized to Affymetrix HuU95av2 chips.
Artefacts identified on probe-level (CEL) files
Figure 1 displays the six chips obtained from one subject's PBMC sample. This subject was chosen because one of the chips (upper row, centre) exhibits a variety of blemishes, which are discussed below, see Figure 2b: a 'bright spot' in the upper-right corner, a 'dark spot' in the upper centre, 'dark clouds' in the upper and lower right centre, and two 'shadowy circles' reaching beyond the left border. Part of the upper circle is included in the chip portion depicted in Figure 1.
Similar results were obtained for all subjects (data not shown). None of the artefacts would have been detected by visual inspection of the pseudo image (Figure 2a). Even after having seen the filtered image, most blemishes are difficult to identify at best. Interestingly, some chips appear to have a preponderance of specific artefacts, suggesting that at least some of the blemishes are caused by specific environmental factors during hybridization, and providing the first indication for the validity of the proposed method. The chip used as the background in Figure 3 has 'dark clouds' in the upper left corner and, albeit to a lesser degree, in both lower corners. Of the two chips with several smaller artefacts, one had three spots that resemble the 'dark spot' in Figure 2. Only the bright scratch at the bottom of one of the chips could have been detected by mere visual inspection of the chip, although even this chip passed the Rockefeller University's Gene Array Resource Center's quality control.
Average vs. median in the filtering procedure
The proposed filtering process relies on identifying deviations of a probe on one chip from a measure of central tendency for this probe across chips. Thus, if few chips have high intensity 'outliers' for one probe, the chips with normal intensities may appear to be negative 'outliers'. One would expect that the six-chip filter is less likely to generate such 'ghosting' artefacts than the three chip filter. We compared the use of medians vs. arithmetic means as the reference. As we had predicted based on the understanding that errors are more likely to be outliers than white noise, using medians not only resulted in less 'ghosting', but also in fewer isolated cells being considered artefacts and, thereby, better contrast (Figure 4).
Validation of probe-level artefacts by going back to the pixel-level image
Our method allows us to identify spatially correlated regions that are unlikely to originate from random fluctuations. To demonstrate that the statistical anomalies detected in the pseudo images at the probe level (Figure 2 and Figure 3) are, in fact, physical blemishes, we inspected the corresponding raw image at the pixel level. The regular artefacts seen (shadow, circle, cloud, etc.) are clearly blemishes, even if the precise nature of the physical blemish may not be known. Still, the difference in features between blemishes suggests different causes.
A number of factors are known to cause bright or dark spots in fluorescence micrographs. Dust on the front cover slip will cause a dark, out-of-focus shadow. Common white paper is bleached with strongly fluorescent dyes, so fibres from tissue paper ordinarily used for cleaning cause intense glare. Many organic solvents, detergents, and other chemicals will fluoresce when concentrated, so leftover droplets or condensates will appear as bright regions, regardless of whether they are in front or behind the focal plane. A crack in the glass would ordinarily be invisible to fluorescence microscopy – except for its ability to accumulate such substances. Glass will normally be coated with substances to prevent the direct binding of fluorophores to it; however, any damage to the fragile coating will cause fluorescent streaks. Illumination with a coherent source such as a laser, as opposed to a broadband source such as a xenon lamp, has specific artefacts such as speckle. In addition, the arrays themselves are manufactured through photolithographic techniques and may contain occasional damage.
Dirt
The visible bright artefact at the bottom-left of Figure 3 is the only blemish in our dataset that did not require 'harshlighting' to be visible. The magnification in Figure 5a shows a structure in an area of 25 × 25 probes. Figure 5b shows the corresponding area in the raw image, clearly exhibiting this artefact to be a piece of debris lying in front of the active array surface in the optical path. While the exact physical nature of this debris is unclear, there can be no doubt that probes highlighted at the bottom of Figure 3b are, in fact, a blemish.
Dark and bright spots
A very 'dark spot' was seen in the lower left corner of Figure 3b. The probe level pseudo image (Figure 6a) shows a dark region, but only the raw image reveals the characteristic of this blemish: an elliptical spot with sharp boundaries which pass through the inside of probes. Still, the grid is visible underneath, as in one of the examples given by Simon, Korn, et al. [13] for cDNA arrays. The dark probes in Figure 6a are therefore likely to be caused by a physical blemish that has 'stained' the image with a dark oval, a mechanical/optical artefact that invalidates the measured intensities of the probes in the region, so all affected probes in the region should be excluded from further analysis. The 'dark spot' in Figure 2 (upper centre) also had a well defined border, although with less contrast (not shown). Three similar artefacts were seen in yet another chip, as shown in the composite picture (Figure 3).
The bright spot on the upper right corner Figure 2 clearly is of different nature. The zoomed area of the DAT file of the second chip (activated) of subject 2 shown in Figure 7b reveals a diffuse area of brightness that covers around 20 probes. Because this bright cloud is out of focus, it is difficult to assess whether its physical location was in front of or behind the focal plane; it could be a leftover detergent condensate in the plastic back panel of the chip. The artefact is less visible in the pseudo image than in the raw image, because the low granularity of the pseudo image enforces an artificial grid structure. Moreover, the Affymetrix image analysis algorithm, taking the 75 percentile of the pixels as an estimate of the probe, may make it more difficult to detect these artefacts through visual inspection because the brightness in areas with low pixel-to-pixel variation is lowered for all percentiles above the median. Although they were easily seen in the filtered pseudo image, neither the 'bright spot' nor the 'dark spot' could have been identified by visual inspection of the original pseudo image. Even on the raw image, only an extremely thorough search for areas of low pixel-to-pixel contrast or boundaries with high contrast across probes could have detected these artefacts based on a single chip alone. Thus, blemishes involving only 9 to 25 probes would often be overlooked in a visual inspection of both the raw and the pseudo image. Given the high variance across pixels, any image processing algorithm aiming at detecting such blemishes at high sensitivity would also create many false positive results.
Dark clouds
For the 'dark clouds', the raw image at first did not show any recognizable feature. Upon closer inspection, however, we noted that the 'dark cloud' in subject 1 had higher pixel-to-pixel variance (Figure 8). The noise does not seem to have a physical origin, as the fluctuations appear to be single-pixel in extent, giving the raw image a 'grainy' appearance.
The areas outside the dark clouds do not appear to be any grainier, so it does not seem to be a change of exposure setting or other simple global change. The image analysis software reports a single, global pixel-to-pixel variation Qraw; it would be useful to have a local quality measure as well, in a fashion similar to the reported background estimate for probe intensities. All dark clouds we found impinge on the array borders. We have no conjecture as to the physical origin of this problem.
Shadowy circles
The two artefacts crossing the left border of Figure 2 suggest yet another reason for blemishes on microarrays. Only one of our chips displayed this artefact, but it did so twice on the same border. Neither the raw image nor physical examination of the chip in a dissection microscope provided any hints to the possible cause (data not shown).
There are myriad possible explanations for what caused this striking artefact. A perfectly round structure with outliers concentrated near its perimeter, evocative of the 'coffee stain rings' phenomenon [14], suggests that a bubble (or a drop) may have formed, during the microfluidic stage, condensation after the washing stage, or as a manufacturing defect.
Thus, to further elucidate the potential cause of this artefact, we plotted the observed vs. the expected intensity (median across the other five chips) for each probe in the area depicted above (Figure 9). We then marked the points below the .10 percentile of all deviations (3) in this area, which formed the 'shadowy circle'. These points were seen over a wide range in expected intensity (7 to 14 in log2 units), although their density is higher for lower intensities. Notably, their intensity was consistently lower than the expected intensity, as though something had only partially interfered with hybridization – or partially stripped the fluorophores prior to readout, or affected probe sensitivity.
Relevance
To determine the extent to which such artefacts may affect standard analyses, we compared the activated vs. the repressed samples (two each) for patient 2, and studied whether masking the blemishes affects the list of differentially expressed genes.
We searched for blemishes all four chips; after manually circling each affected area, we masked (declared missing) all points in the upper or lower 10th percentile within that area, respectively. We used either the lower or upper 10% since one of our findings is that all artefacts seem to have the common characteristic shown, for instance, in Figure 9, that outliers within an artefact are either (almost) exclusively brighter ('bright spot') or darker (all other blemishes) than expected We conducted separate analyses for the original and the masked data. We estimated the signal value for each probe using the Bioconductor implementation (affy package 1.3.28, R.1.8) of the MAS5 algorithm with default parameters, after modifying the summarization and normalization steps to allow for missing data. The overall effect is shown in Figure 10a, with a maximum difference of 4.6 log2.
Genes whose expression estimates changed by more than 0.1 log2 through filtering were considered as 'altered' by filtering. The 'bright spot', where about 39 probes were affected, altered the expression of 16 genes by up to 1.37 log2. The 'shadowy circle' altered the expression of about 380 genes; more than 50 of them by more than 0.5 log2. The 'dark spot' affected 47 probes, altering expression of 103 genes by up to 1.6 log2. The 'cloud' altered the expression of 700 genes, 83 of them by more than 0.5 log2. The dirt covering around 25 × 25 probes, affected around 376 probes, altering 148 genes, 16 of them by more than 0.5 to a maximum of 1.26 log2.
Finally, we compared the two conditions (absence vs. presence of a repressor), mirroring masked probes on both on the affected and the corresponding chip. As an exploratory criterion, we used the modified (paired) t-test suggested in Smyth [15] from the limma package of the Bioconductor project [16]. As shown in Figure 10b, the effects of identifying genes as differentially expressed can be dramatic, demonstrating the potential value of detecting blemishes and masking affected areas on microarrays.
Validity
We validate the proposed method using data from the Spike-in HUG133 experiments [17]. This data set consists of 3 technical replicates of 14 separate hybridizations of 42 spiked transcripts at concentrations from 0.125 pM to 512 pM arrayed as a Latin Square. Our interest is to assess whether masking the blemishes improves the ability to detect differentially expressed genes. We used the Affycomp package of the Bioconductor project, which encompass a series of tools developed by [18] to compare the performance of expression measures for Affymetrix GeneChips. Figure 11 shows that masking blemishes has little effect for large fold changes, as one would expect, while the ROC curve (sensitivity) vs. (1-specificity) shows a substantial improvement for small (2 fold) changes. Other statistics are also improved in this case: the average false positive decreases (from 2818 to2763) while the true positives increases (from 14.33 to 14.57). Comparing by range of intensities, the area under the curve (AUC) is bigger for the masked data in the lower intensities (0.003 vs. 0.010) while keep similar performance in the medium and low range (data not shown), resulting in a bigger average weighted AUC for the filtered data (0.002 vs. 0.007) (a detailed description of these statistics can be found in [18] and in the affycomp vignette). Thus, our masking procedure improves the sensitivity/specificity to detect small differential expression, especially in the range of low intensities.
Discussion
As an alternative approach to identify blemishes, one might try to look at the residuals from parametric estimations in the background subtraction or summarization stages; e.g., looking at the residuals of the PM-MM difference model [19] or the RMA model over the PM values [20] to identify possible aberrations. Unfortunately, the variety of models currently being discussed attests to the fact that each model has its drawbacks. While random variation can typically be handled by statistical methods, systematic errors in the choice of the model assumptions may have a drastic impact on these processes. The proposed method is robust in the sense that only few assumptions are made. Another advantage of our approach is that we can include mismatch probes which are especially suitable to identify aberrations, because they are less sensitive to gene expression variations.
Moreover, in any such model of expression estimation the residuals of the entire probe set containing a faulty probe is likely to be affected, so that errors are spread across the probe set and hence over the image; if one probe in a probe set is an outlier, e.g., very bright, all other probes would be slightly dark ghost images, similar to the 'ghosting' seen in Figure 4. Utilizing topological information for identification and elimination of blemishes has the advantage that suspect probes are identified before background correction, normalization and summarization take place. Thus, faulty data will not confound the preprocessing steps and further statistical analysis.
With the next generation of Affymetrix chips, the relevance of correcting for blemishes will even increase. Here, we analyzed U95 chips with 16 probe pairs per probe set. To make room for more probe sets, the number of pairs per set has been reduced to (as few as) 11 on the U133 chips. This, however, not only increases the standard error by 20%, and, thus the effect of any artefacts on the results, but also reduces the ability of model based methods to draw on probe set information. The number of neighbouring cells on a microarray, in contrast, is not adversely affected by reducing the size of the probe sets. In fact, smaller probe sets make it less likely that probe pairs from the same set are in close vicinity.
Conclusion
We have presented an extremely simple method for finding blemishes on microarrays. The method's simplicity makes it robust and it does not rely on estimating model parameters. It sensitively tagged blemishes on chips that had passed our Gene Array Resource Center's quality control mechanism. Only one blemish (Figure 5) could have been readily seen in the raw images. That we found clear evidence of physical blemishes in the raw images for most of the artefacts identified on the pseudo images attests to the validity of the findings.
We have applied our method to an experimental dataset and were able to identify anomalies of different type. Approximately 25% of our chips are blemished, often more than once, and blemishes can cover areas from a few dozen to hundreds of probes. We examined the potential impact these blemishes have on the experiments. Failure to remove the blemishes from further analysis can materially affect the detection of subtle changes in experiments testing similar conditions. When applied to the Spike-in data set, the proposed method had an overall better sensitivity/(1-specificity) ratio.
For the future we propose to develop pattern recognition algorithms to automatically find and mask out suspected blemishes, and to modify the extant background correction and summarization algorithms to be able to properly handle missing data from blemish removal.
Methods
Let X(i), i = 1, ..., n, represent the intensity values of the i-th of n chips, each consisting of m × m (e.g., 650 × 650) cells . Assuming that biological systems respond to relative, rather than absolute differences in gene expression, for each pair of chips a matrix of pointwise (log) ratios is defined as
Given that the intensity at each cell is highly determined by the sequence of the probe [8], the spatial distribution of differences in log-intensities should have no identifiable features, except for probes belonging to probe sets related to the genes that are differentially expressed under the conditions the samples were taken. Here, we assume that the proportion of differentially expressed genes is small. Thus, since probes belonging to a probe set are (more or less) randomly distributed across the chip, cells of related genes are rarely located next to each other, so that no obvious pattern should be discernable. If, however, chip X(i) has a localized 'defect', this should result in a similar pattern across all R(i,i'≠i) in the region of the defect. To allow for visual inspection of such pattern, we draw on the fact that the distribution of differences in log-intensities should be (more or less) symmetrical, except for outliers caused by rare events affecting small areas in particular chips. Probe-wise outliers (due to both differential expression and defects) can be identified by comparing each chip to a measure of central tendency derived from all other chips. Although other measures of central tendency will be discussed below, we start our discussion with the special case of the arithmetic mean, which is known to be optimal in the classical linear model ([21])
Let R(i,i') = Δ(i,i') + D(i) - D(i') + ε where Δ(i,i') indicates the random contribution from the differentially expressed genes, D(i) describes the defects of the i-th chip, and ε other random errors. Then, D(i) contributes not only to (bars indicating the average over the index replaced by dot), but also, albeit with only 1/n of the intensity, to each of the other as a 'negative shadow' or 'ghost' image. As the number of chips n increases, however, the law of large numbers allows for approximating the linear equation system (1), with hats indicating estimators, as
From (2), we get the linear equation system:
where I = (δj = j')j,j' = 1...n and J = (1)j,j' = 1...n. A system has the trivial solution Y = D whenever column sums are zero (JY = 0). As (2) guarantees that , setting yields the solution
as the linear model estimate for the deviation of the i-th chip from the other chips. As the number of chips increases, ghosting reduces, so that any discernable pattern in in the limit would suggest a defect.
The above justification for obtaining residuals within the linear model by subtracting the average is well known. Still, spelling out and justifying the individual steps above helps in two ways. First, we can fine tune the method for the particular situation we are faced with and, second, we can provide numerical examples comparing the proposed non-parametric with the traditional parametric approach. The justification for the choice of the arithmetic mean (average) as the measure of central tendency in linear models relies either on the law of large numbers and the central limit theorem or on the assumption that the distribution of errors is symmetrical, in general, and Gaussian, in particular. Neither assumption is easily justified for the errors caused by defects on a chip.
The arithmetic mean is known to be relatively sensitive to outliers. Thus, to discriminate outliers from observations close to the centre of the non-outliers, one would need either a very large number of chips or a measure of central tendency that is less likely to be affected by the outliers themselves. While microarray 'experiments' now typically consist of more than a single chip, the number of chips analyzed under comparable conditions is still too small to rely on the central limit theorem for outlier detection. With the number of chips in the single digits, even 'Winsorization' may not be feasible. Moreover, the need for choosing some Winsorization cut-off points adds an undesirable level of arbitrariness to the results. The median, as the most robust form of Winsorization, provides for a simple alternative measure of central tendency:
Acknowledgements
The authors wish to thank Marcelo O. Magnasco for helpful discussions and support. M.S.F. acknowledges a Woman in Science fellowship from RU. K.M.W. was supported in part by GCRC grant M01-RR00102 from the National Center for Research Resources at the National Institutes of Health. This paper is, in part, based on a presentation given at the 2004 Joint Statistical Meetings in Toronto, Canada.
Figures and Tables
Figure 1 Detail of six chips for the second patient. Upper left corners of six chips (250 × 250 cells) with samples from a cell culture evaluated at different time points (rows: 6 hrs, 24 hrs) under different experimental conditions (columns: none, activated, activated in the presence of a repressor). Boxed areas indicate internal standards.
Figure 2 Blemishes of one chip (Activated 6 h) for the second patient. Left (a), whole chip pseudo-image. The box indicates the portion of this chip depicted in the center image of the first row of Figure 1. Right (b), filtered image based on the set of the three chips in the first row (with median adjustment).
Figure 3 Collage of Artefacts. b) Areas with artefacts obtained from seven chips. a) Composition of the raw areas corresponding to the areas denoted in (b). Filtered image based on the set of three chips with median adjustment.
Figure 4 Median vs. Average filter. The "bright spot" artefact (of Figure 2). Top row: raw image from the same are of three chips showing gene expression from the same sample under three experimental conditions. 3 chip filtering relies on information from the three presented chips measured at 6 hrs only, while 6 chip filtering also draws on the chips observed at 24 hrs.
Figure 5 "Dirt". Detail of the artefact seen at the bottom of Figure 3. a) CEL file b) DAT file. The size of this artefact is approximately 25 × 25 probes in the CEL file and 0.5 × 0.5 mm on the chip).
Figure 6 "Dark Spot". Detail of the 'dark spot' artefact seen in the lower left corner of Figure 3. a) CEL file b) DAT file.
Figure 7 "Bright Spot". Detail of the 'bright spot' in the upper right corner of Figure 3. a) CEL file b) DAT file.
Figure 8 "Dark Clouds". Detail of the region containing the 'dark cloud' at the right border of Figure 1. Top row: sample area around the 'dark cloud', Bottom row: corresponding area from a control chip. 1st column: CEL file, second column: DAT file; third column: detail from the indicated area in the DAT file.
Figure 9 "Shadowy circles". Analysis of the two 'shadowy circles' at the left border of Figure 2. Observed intensities vs. expected intensities.
Figure 10 Influence on expression values. a) Expression values for four chips of subject 2, original data vs. filtered data. b) T-statistics for the comparison of activated vs. activated in the presence of the repressor; raw vs. filtered data. Open circles: p < 0.01 with the original data only, solid black dots: p < 0.01 with the filtered set only; large grey solid dots: p < 0.01 with both sets.
Figure 11 ROC curves for Spike-in 133 data. Receiver Operator curves. a) using only arrays which nominal fold changes are equal to 2. b) as a but fold change equal to 8. c) as a but fold change equal to 1024.
==== Refs
Brown CS Goodwin PC Sorger PK Image metrics in the statistical analysis of DNA microarray data PNAS 2001 98 8944 8949 11481466 10.1073/pnas.161242998
Jain AN Tokuyasu TA Snijders AM Segraves R Albertson DG Pinkel D Fully Automatic Quantification of Microarray Image Data Genome Res 2002 12 325 332 11827952 10.1101/gr.210902
Ihaka R Gentleman R A language for data analysis and graphics Journal of Computational and Graphical Statistics 1996 5 299 314
Parmigiani G Garrett ES Irizarry RA Zeger SL The analysis of gene expression data: methods and software 2003 New York: Springer
Affymetrix I GeneChip Expression Analysis: Data Analysis Fundamentals 2004
Hautaniemi S Edgren H Vesanen P Wolf M Jarvinen A-K Yli-Harja O Astola J Kallioniemi O Monni O A novel strategy for microarray quality control using Bayesian networks Bioinformatics 2003 19 2031 2038 14594707 10.1093/bioinformatics/btg275
Irizarry RA Hobbs B Collin F Beazer-Barclay YD Antonellis KJ Scherf U Speed TP Exploration, normalization, and summaries of high density oligonucleotide array probe level data Biostatistics 2003 4 249 264 12925520 10.1093/biostatistics/4.2.249
Naef F Magnasco MO Solving the riddle of the bright mismatches: labeling and effective binding in oligonucleotide arrays Phys Rev E Stat Nonlin Soft Matter Phys 2003 68 011906 Epub 012003 Jul 011916. 12935175
Wu Z Irizarry RA Gentleman R Martinez Murillo F Spencer F A model based background adjustment for oligonucleotide expression arrays Journal of the American Statistical Association 2004 99 909 917 10.1198/016214504000000683
Lew W Bowcock AM Krueger JG Psoriasis vulgaris: cutaneous lymphoid tissue supports T-cell activation and 'Type 1' inflammatory gene expression Trends in Immunology 2004 25 295 305 15145319 10.1016/j.it.2004.03.006
Zhou X Krueger JG Kao M-CJ Lee E Du F Menter A Wong WH Bowcock AM Novel mechanisms of T-cell and dendritic cell activation revealed by profiling of psoriasis on the 63,100-element oligonucleotide array Physiol Genomics 2003 13 69 78 12644634
Bayliffe AI Haider A Haws TF Kaplow Y Krueger JG Liy AG Thompson PW Wang X-J Lamb JR PPAR-g mechanisms in reducing cutaneus inflammation European Academy of Dermatology and Venereology; Florence, IT 2004
Simon RM Korn EL McShane LM Radmacher MD Wright GW Zhao Y Design and Analysis of DNA Microarray Investigations 2003 New York: Springer
Deegan RD Pattern formation in drying drops Physical Review E 2000 61 475 485 10.1103/PhysRevE.61.475
Smyth GK Linear models and empirical Bayes methods for assessing differential expression in microarray experiments Statistical Applications in Genetics and Molecular Biology 2004 3 3
Website title
Website title
Cope L Irizarry R Jaffee H Wu Z Speed T A benchmark for affymetrix GeneChip expression measures BIOINFORMATICS 2004 20 323 331 14960458 10.1093/bioinformatics/btg410
Li C Wong WH Model-based analysis of oligonucleotide arrays: Expression index computation and outlier detection PNAS 2001 98 31 36 11134512 10.1073/pnas.011404098
Collin F Brettschneider J Bolstad B Speed T Quality Assessment of Gene Expression Data for Affymetrix Genechips Affymetrix GeneChip Microarray Low-Level Workshop; Berkeley, UC 2003
Searle SR Linear Models 1971 New York: Wiley
|
15784152
|
PMC1274260
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 22; 6:65
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-65
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-671578414010.1186/1471-2105-6-67Methodology ArticleEvaluation of gene importance in microarray data based upon probability of selection Fu Li M [email protected] Casey S [email protected] Pacific Tuberculosis and Cancer Research Organization, Pasadena, California, USA2 University of Florida, Gainesville, Florida, USA2005 22 3 2005 6 67 67 19 11 2004 22 3 2005 Copyright © 2005 Fu and Fu-Liu; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Microarray devices permit a genome-scale evaluation of gene function. This technology has catalyzed biomedical research and development in recent years. As many important diseases can be traced down to the gene level, a long-standing research problem is to identify specific gene expression patterns linking to metabolic characteristics that contribute to disease development and progression. The microarray approach offers an expedited solution to this problem. However, it has posed a challenging issue to recognize disease-related genes expression patterns embedded in the microarray data. In selecting a small set of biologically significant genes for classifier design, the nature of high data dimensionality inherent in this problem creates substantial amount of uncertainty.
Results
Here we present a model for probability analysis of selected genes in order to determine their importance. Our contribution is that we show how to derive the P value of each selected gene in multiple gene selection trials based on different combinations of data samples and how to conduct a reliability analysis accordingly. The importance of a gene is indicated by its associated P value in that a smaller value implies higher information content from information theory. On the microarray data concerning the subtype classification of small round blue cell tumors, we demonstrate that the method is capable of finding the smallest set of genes (19 genes) with optimal classification performance, compared with results reported in the literature.
Conclusion
In classifier design based on microarray data, the probability value derived from gene selection based on multiple combinations of data samples enables an effective mechanism for reducing the tendency of fitting local data particularities.
==== Body
Background
Based on the concept of simultaneously studying the expression of a large number of genes, a DNA microarray is a chip on which numerous probes are placed for hybridization with a tissue sample. The DNA microarray has recently emerged as a powerful tool in molecular biology research, offering high throughput analysis of gene expression on a genomic scale. However, biological complexity encoded by a deluge of microarray data is being translated into all sorts of computational, statistical or mathematical problems.
Driven by the growing genomic technology, molecular medicine has become a rapidly advancing field. An important research topic is to identify disease-related gene expression patterns based on microarray analysis. In one approach, genes are selected for constructing a clinically useful classifier for disease diagnosis. The genes thus selected often shed light on the fundamental molecular mechanisms of the disease [1]. As addressed in several research works [1-5], the problem of gene selection considered in this context is a difficult one because there are thousands of genes at hand but only a very limited number of samples are available. Mathematically, this problem is characterized by high data dimensionality. To develop a classifier, dimensionality reduction by gene selection is essential. Genes selected for constructing a classifier are believed to be important. Typically, only a small fraction of genes differentially expressed in the diseased tissue will be selected.
There exist two related but different objectives for gene selection. As mentioned above, one objective is to construct a classifier or predictor for classifying, diagnosing, or predicting the type of cancer tissue according to the expression pattern of selected genes in the tissue [6]. The other objective is to determine whether the changes in gene expression across two conditions are significant (e.g., SAM) [7]. The present work is developed in the first context.
Here, we report new theoretical developments and research results as an extension of our earlier work [4,8], presenting a new probabilistic analysis of gene selection from microarray data, which distinguishes our work from other related work.
Results
Probability analysis of selected genes
Under very high data dimensionality, questions can be raised of whether genes could have been selected by chance and whether selected genes are sufficiently significant beyond any doubt due to inherent uncertainty or data particularity. Quite often, not identical sets of genes are selected from different subsets of the data. At the fundamental level, it would be important to distinguish between the case of diverse patterns and the case of false patterns. To address the problem, we take the approach that takes into account both statistical significance and performance issues. The bootstrapping technique lends itself well as far as the first issue is concerned.
Suppose we randomly draw samples from a given domain and conduct a gene selection experiment. Assume that we select one gene out of a total of p genes. The probability of the event that a particular gene is selected in a single trial is 1/p. According to the information theory, the smaller the probability is, the more informative the event is. Given a large p, it seems that the event is significant, and this would be true only if we have a particular gene in mind before gene selection; otherwise, the probability should be adjusted for the presence of p genes, and then it becomes clear that any gene selected in a single trial is non-informative. Now suppose we conduct multiple trials and ask the question of whether any gene repeatedly selected across trials is significant. Here we devise an analysis for the question.
Theorem
In r multiple independent trials conducted for gene selection, select one gene out of a total of p genes in each trial. Given the level of significance α, a gene is considered significant if it is selected r times in r trials and
Proof
The probability of the event that the same gene is selected r times in r trials is (1/p)r. Since there are p genes, the adjusted probability (analogous to Bonferroni's correction) is p(1/p)r. Therefore,
Equivalently,
Thus,
Note that the value of is negative. The result follows. €
Corollary 1
The minimum threshold value of r for reaching the given level of significance is
where ⌈⌉ is the ceiling operator. This is because r must be an integer greater than or equal to the real threshold.
For example, consider the leukemia data [1]. There are 7129 genes. Assume α = 0.05. From Eq. (1), rθ = 2.
Consider a more general case: what is the probability of the event that a gene is selected r times in m trials? The adjusted probability becomes
where is the combinatorial function that returns the number of possibilities for choosing r from m objects. Assume a large p so that. Then, we have
The level of significance (α in Eq. (1) and (2)) is set to 0.05 by convention in the present work.
Reliability analysis of gene selection
The innovative feature of our method is to conduct reliability analysis for arriving at the gene expression signature. The analysis assesses the repeatability of genes selected and determines the repeatability for gene selection using M-fold cross-validation.
In the 10-fold cross-validation approach, the data set is divided into 10 disjoint subsets of about equal size. Genes are selected on the basis of nine of these subsets, and then the remaining subset is used to estimate the predictive error of the trained classifier using only the selected genes. This process is repeated 10 times, each time leaving one set out for testing and the others for training. The cross-validation error rate is given by the average of the 10 estimates of the error rate thus obtained.
In each cross-validation cycle, we conduct SVM-RFE gene ranking and selection operations, as described in the Methods section. We select a minimal set of genes by collecting from the top rank one by one and picking the set associated with minimum error in each cross-validation cycle. There is no guarantee that the same subset of genes will be selected in each of the 10 cycles in 10-fold cross-validation. However, vital genes tend to be selected more consistently than others across cycles. The significance of a gene is correlated with the repeatability of selection according to the probabilistic analysis given earlier. We associate each selected gene with a repeatability value indicating how many times it is selected in the cross-validation experiment. The biological or clinical interpretation of "repeatability" would depend on the objective and design of the microarray experiment. We may consider the validity of a selected gene by its reliability in the sense that the more often a gene is selected, the less likely chance is a factor.
To select the final set of genes, we need to determine the repeatability threshold. A gene is in the final set if its repeatability reaches (i.e., no less than) the threshold. To this end, second 10-fold cross-validation is performed. Then we choose the repeatability threshold that is associated with the minimal cross-validation error under the given level of significance (α = 0.05). Recall that a gene with a higher repeatability is associated with a small P value, as shown earlier.
To extend the method from two-class to multi-class classification, we adopt the one-against-all others strategy under which genes are selected for each class one at a time and then combined. For each class, all the other classes are grouped as a single class. In this way, a multi-class gene selection problem is converted into a series of two-class problems. The program was written in Matlab [9]. An SVM Matlab toolbox as well as Mathlab is required for the program use.
Case analyses
In cancer research, our current goal is to develop a molecular classifier based on tissue gene expression patterns for diagnosis and subtype classification. With this in mind, we evaluate our method using well-known benchmark microarray data sets including those concerning small round blue cell tumors, colon cancer, leukemia as well as perturbed data sets.
The small round blue cell tumors (SRBCTs) data set includes 63 training samples and 25 test samples derived from both tumor biopsy and cell lines [10]. In consistency with other reports in the literature, we used the test set of 20 samples after 5 non-SRBCT samples were removed. The data set consists of four types of tumor in childhood, including Ewing's sarcoma (EWS), rhabdomyosarcoma (RMS), neuroblastoma (NB), and Burkitt lymphoma (BL). After initial screening, the data set in the public domain contains 2308 genes.
The colon cancer data set contains 62 tissue samples, each with 2000 gene expression values [11]. The tissue samples include 22 normal and 40 colon cancer cases. In this study, we used all the 62 samples in the original data.
The leukemia data consist of 72 tissue samples, each with 7129 genes expression values [1]. The samples include 47 ALL (acute lymphoblastic leukemia) and 25 AML (acute myeloid leukemia). The original data have been divided into a training set of 38 samples and a test set of 34 samples.
The reference method with which we compared our method applied a technique referred to as SVM-RFE [3] to select genes from the training data without reliability assessment. The reference method [12] is a multi-class extension of the SVM-RFE method used for two-class classification. The SVM-RFE method (two-class or multi-class) has not been applied to the SRBCT data before. We implemented the computer algorithm of the reference method for comparison with ours. The same experimental conditions were applied to both methods.
Small round blue cell tumor classification
On the SRBCT data, our method selected 19 genes (Table 1) from the microarray gene expression data of the 63 training samples. The SVM classifier trained on the 63 training samples using the 19 selected genes was tested on the 20 different test samples. Both the training and test predictive accuracies were 100%. That is, the trained SVM classifier can accurately predict the tumor class using the 19 gene expression data for both seen and unseen samples. Since the classifier may tend to fit the training data, the generalization performance of the classifier is indicated by the test accuracy.
The reference method selected 8 genes with 100% training accuracy but with only 90% test accuracy. It seemed that the reference method did not select enough genes even though the selected genes could correctly classify all the training samples – an example of over-generalization, whereas our bootstrap-like strategy adequately dealt with this problem by taking into account of both reliability and diversity in gene selection.
We examined the consensus of genes selected by our method and by two other best-known methods: the method of Khan et al. [10] based on artificial neural networks and the method of Tibshirani et al. [13] based on shrunken centroids, and we found that there was high consensus between our and their results. Out of the 19 genes selected by our method, 18 genes were also selected by Khan's method and 16 genes by Tibshirani's method (Table 1). While agreement among results produced by different methods may imply similarities in the inductive biases, these two other methods use fundamentally different representational biases. Thus, such agreement should not be taken for granted and would instead serve as substantial evidence indicative of the validity and significance of our method.
Whether the selected genes served as meaningful markers for cancer classification was further confirmed by cluster analysis and visualization. To this end, we applied a hierarchical clustering program developed by Eisen [14] to the gene expression data of the selected genes. By visual inspection of the gene expression map, four clearly separated clusters (Figure 1) were identified. Upon verification, each cluster corresponded exactly to a distinct tumor group with 100% accuracy. Thus, a diagnostic chip can be designed based on the selected genes. This result also provides additional evidence to support our method.
Colon cancer diagnosis
In performance analysis, we conducted multiple experiments with random data partitions. In each experiment, the data were randomly and equally split into training and test sets. The training set was used for gene selection and classifier training, and the test set for determining the predictive performance of the classifier based on the genes selected by the given algorithm. Our method outperformed the reference method by a small margin. This result reflects the underlying fact that there are multiple possible ways of selecting genes for constructing a classifier with comparable performance using different methods.
Our program selected 15 genes from the colon cancer data (Table 2). The selected genes allow the separation of cancer from normal samples in the gene expression map (Figure 2, Table 3). Some genes were selected because their activities resulted in the difference in the tissue composition between normal and cancer tissue. Other genes were selected because they played a role in cancer formation or cell proliferation. It was not surprise that some genes implicated in other types of cancer such as breast and prostate cancers were identified in the context of colon cancer because these tissue types shared similarity.
Our method is supported by the meaningful biological interpretation of selected genes, as discussed below. New biological hypotheses can be formulated to further investigate the relationship of a particular gene with colon cancer. For example, what is the role of profilin 1 protein in colon cancer? Some discovered genes could potentially serve as novel targets for drugs, vaccines, or gene therapy.
Leukemia classification
On the leukemia data, our method selected four genes (Table 4) from the microarray gene expression data of 38 training samples. The SVM classifier trained on the 38 training samples using the selected genes was tested on the 34 different test samples. The training and test accuracies were 100% and 97.06%, respectively. In addition, the AML and ALL samples formed separate clusters in the gene expression map of the selected genes.
The reference method also selected four genes and achieved the same level of test accuracy as our method. The original algorithm of SVM-RFE [3] selected 8 or 16 genes on this data set. The method based on shrunken centroids [13] selected 21 genes on this data set. A recent study indicated that the unbiased error estimate of the classifier using a small number of selected genes was virtually non-zero on the leukemia data set [6]. Taken together, the evidence showed that our method produced optimum results in terms of both predictive performance and the number of selected genes.
Perturbed data
In practical circumstances, noise may arise during sample collection and handling, slide preparation, hybridization, or image analysis, as reflected by variations in microarray results generated from different laboratories. To address this issue, we also conducted performance evaluation of our gene selection method based on perturbed data. 20 data sets were produced by randomly perturbing 5% (rounded up to the nearest integer) of the training cases, reversing their class labels and leaving the test cases intact, in the domains of colon cancer diagnosis and leukemia classification (ten in each domain). The average test predictive accuracies with our method in the two domains were 85.49% and 88.61%, respectively, compared with 80.65% and 86.11% with the reference method. The result suggests the potential advantage with our method in smoothing out data variations due to various sources in practice.
Discussion
Both cross-validation and bootstrapping are standard statistical methods for arriving at an unbiased estimate of the true error rate associated with a classifying or predicting system. Bootstrapping has also been used for assessing the reliability or stability of phylogenetic trees [15] or cluster analysis [16]. Bootstrapping is a method for random re-sampling with replacement for a number of times and estimates the error rate by the average error rate over the number of iterations. Cross-validation is a method of assessing the reliability of error; however, its application to learning the pattern in the data is novel. As discussed later, stability emerges as an important issue in gene selection. Here we propose to use bootstrapping or cross-validation for analyzing the issue. Our experience showed that cross-validation was more efficient than bootstrapping. For instance, genes selected based on a single 10-fold cross-validation were more accurate in prediction than those selected using bootstrapping with 10 re-sampling iterations. Since the SVM-based gene selection algorithm is time-consuming, we consider only cross-validation for assessment of error and stability in this study.
In the original SVM-RFE algorithm [3], error estimation and gene selection are not independent processes because both are based on the same training set. However, it is important to correct for the selection bias by performing a cross-validation or applying a bootstrap external to the selection process [6,17]. Our implementation of SVM-RFE is based on this idea.
Genes selected for cancer diagnosis or classification can be validated by their biological significance since these genes are expected to show differential expression between normal and cancer tissue or among subtypes of cancer, and as such, they are implicated in cancer-related mechanisms or pathways. Genes with unknown roles may be discovered through gene selection and later verified by biological studies.
From the SRBCT data set, genes selected by our method for a particular type of cancer/tumor against other types are generally consistent with its tissue of origin. For example, genes selected for neuroblastoma (NB) are characteristic for nerve cells, such as neuronal N-cadherin, and meningioma 1; genes selected for rhabdomyosarcoma (RMS) are characteristic for muscle cells, such as alpha sarcoglycan, and slow skeletal troponin T1; genes selected for Burkitt lymphoma (BL) are characteristic for lymphocytes or blood cells, such as major histocompatibility complex (class II, DM alpha). Some genes discovered by means of microarray analysis have been reported in the biological literature, e.g., over-expression of MIC2 in Ewing's sarcoma (EWS) [18]. Some genes are over-expressed in a certain type of tumor but lack specificity. For instance, FGFR4 (fibroblast growth factor receptor 4) was noted to be highly expressed only in RMS and not in normal muscle, but it is also expressed in some other cancers and normal tissues [10]. A gene that is under-expressed in a particular type of tumor compared with other types can also be selected as a diagnostic marker. For instance, cold shock domain protein A selected for NB was under-expressed in this tumor, consistent with the fact that this gene is expressed in B cells and skeletal muscle but not in the brain [13].
With our method, four muscle-related genes (H20709, T57882, T92451, and J02854) were selected from the colon cancer data, reflecting the fact that normal colon tissue had higher muscle content, whereas colon cancer tissue had lower muscle content (biased toward epithelial cells) [11]. The selection of 60s ribosomal protein L30E agreed with an observation that ribosomal protein genes had lower expression in normal than in cancer colon tissue [11]. The selected interferon inducible protein 1-8D genes were found to be expressed in adenocarcinoma cell lines [19]. There was a potential connection of another selected gene, human chitotriosidase, to cancer [3]. The implications of cancer among other selected genes are explained as follows. S-100 protein can stimulate cellular proliferation and may function as a tumor growth factor [20]. Profilin 1 protein can suppress tumorigenicity in breast cancer cells. A study showed consistently lower profilin 1 levels in tumor cells [21]. The reduced expression of P27 protein was linked to the possibility of colon carcinoma [22]. The A20 protein can inhibit a specific apoptotic pathway [23]. Recall that apoptosis is a major mechanism for tumor suppression. The guanine nucleotide-binding protein is involved in signal transduction and its abnormality may contribute to cancer development [24]. A thyroid receptor interactor could be a target gene of a certain oncogene. The alpha trans-inducing protein (bovine herpesvirus type 1) may be linked to oncogenic activity.
In the related work [3], 7 genes were selected from the colon cancer data: H08393, M59040, T94579, H81558, R88740, T62947, and H64807. For all of them, a possible link to cancer was found in the biological literature. These 7 genes, however, do not include any muscle-specific gene, despite that muscle content offered a discriminating index for colon cancer [11].
In a typical microarray data analysis problem, the data dimensionality is high and the sample size is relatively small. Under this condition, the problem of finding a classification model is under-constrained, and the model found tends to fit the training data so closely that it fails to generalize to unseen data. To address the issue of data overfitting, the SVM has the capability of controlling the model complexity to the point where a satisfactory solution can be produced. On the other hand, the ability of causal discovery based on the SVM-RFE approach or an alternative approach is discounted by the finding that most genes selected are selected only once from one data split to another in M-fold cross-validation [25]. This means that the SVM is not free of the data-overfitting problem at least in the context of gene selection from microarray data, and it raises the question about stability or reliability of gene selection, as we address here.
The research finding that the SVM may assign zero weights to strongly relevant variables and non-weights to weakly relevant (red-herring) features [26] implies the disadvantage with this approach for discovery of causal variables associated with the target variable concerned. This however can be understood since the SVM-RFE is aimed to identify the best features for maximum margin of separation between different classes of samples, regardless of causal implications. In reality, causal variables are not necessarily most discriminant, as the target variable is not always categorized according to its causal factors. The issue of causality becomes even more complicated because of confounding variables leading to so-called spurious causation. The method presented here is developed in the context of cancer subtype classification and evaluated in terms of predictive performance rather than the capability of causal inference. However, some methods are both predictive and causal [26,27].
We emphasize the importance of holding back some data to improve generalization and diversity of the learning outcome. In application of M-fold cross-validation to n samples, M can assume a value ranging from 2 to n. A small M is not sufficient to assess the repeatability of selected genes while a large M (e.g., M = n in the leave-one-out experiment) is associated with high degree of redundancy on data for training and low diversity of genes selected. This argument suggests that there exists an optimum M value. So we conducted experiments to compare predictive accuracies for three cases: M = 5, 10, and 15. Among the three cases, 10-fold cross-validation achieved the best results. It is thus consistent with our intuitive analysis. However, there is no proof that 10-fold cross-validation is always the best choice. In practice, the optimum M value should be determined by the value associated with the best cross-validation accuracy.
This study highlights the importance of reliability assessment of genes selected from a large-scale microarray data. We show how to derive the P value of each selected gene in multiple gene selection trials based on different data partitions. The importance of a gene is indicated by its associated P value. The distinctive feature of our method is that gene selection is determined by both ranking and reliability analyses. Reliability analysis is conducted using M-fold cross-validation. Some gene selection methods [3,28] use cross-validation to determine the number of selected genes by minimum cross-validation error but not by optimum repeatability as in our method. Thus, reliability analysis comprising repeatability measurement and optimum repeatability determination defines the novelty of our method, which has enabled a more accurate and cost-effective cancer classifier to be constructed, compared with other methods. Notice, however, the argument about reliability or stability must rest on the assumption of sound performance, as will be clear from the apparent stability with some trivial approaches to gene selection such as the one based on lexicographic ordering of gene names. In fact, the theory behind the analytical scheme we developed is a general one and can therefore be extended to other performance-based gene selection methods.
Conclusion
The DNA microarray technology has become a standard tool for gathering genome-wide gene expression information. Molecular classification based on gene expression information has emerged as an important approach to cancer diagnosis. A cost-effective approach is to select a small set of genes for classifier design. Moreover, it may be ineffective to use whole microarray data for classification purposes because the data dimensionality (i.e., the number of variables/genes) is often several orders of magnitude greater than the available sample size.
Experience shows that different sets of genes can be selected from different combinations of microarray data instances with the same gene selection algorithm. At the same time, it is noticed that a biologically significant gene tends to be selected repeatedly across different combinations of data instances. We have developed a method for analyzing this situation. In the domain of small round blue cell tumor subtype classification, we have demonstrated that the method we developed selected only 19 genes that provided 100% accuracy on both training and test data sets. In comparison, the approach based on artificial neural networks [10] selected 96 genes, and the shrunken centroid method [13] selected 43 genes. Thus, our method suggests a mechanism for effectively reducing the tendency of fitting local data particularities in the process of gene selection for classifier design based on microarray data.
Methods
This section provides the details of the methods, but the novelty aspects are described in the "Results" section.
Classification based on support vector machines
We use the method of support vector machines (SVM) [29,30] for classification. The SVM has been demonstrated as a useful tool for analyzing microarray data [31]. Consider n training samples {(, yi) | 1 ≤ i ≤ n }, where , is the input feature vector for the ith sample and yi is the corresponding target class (output). The basic problem for training an SVM can be reformulated as: given a set of n training instances, each represented as (, yi), maximize
subject to
The optimal hyperplane that separates different classes of objects can be constructed from the solutions αI's to this maximization problem. The SVM can perform a nonlinear transformation via the inner-product kernel to map the input space into a new high-order feature space where the patterns are linearly separable with high probability. The use of such a kernel function can lead to a decision function that is non-linear in the input space but its image is linear in the transformed space. When the samples are not linearly separable, whether in the input or transformed space, a soft-margin algorithm as an extension of the basic algorithm is available [32].
The SVM used in this study employed the linear kernel since we found that it yielded a better result than a non-linear kernel for the data under investigation, and this observation is also consistent with the literature [3]. All SVM parameters were set to the standard values in accordance with the convention: s = 0 (C-SVM), t = 0, c = 100, v = 10.
Data normalization in the case of cDNA arrays proceeded as follows: the local background intensity is subtracted from the value of each spot on the array; the two channels are normalized against the median values on that array; the Cy5/Cy3 fluorescence ratios and log10-transformed ratios are calculated from the normalized values. In addition, genes that do not change significantly can be removed through a filter in a process called data filtration.
Gene selection
An SVM-based gene selection algorithm has two main components: gene ranking and gene selection. Gene ranking results in a sorted list of genes in decreasing order of importance for classification. This issue is complicated since some genes become important only if combined with other genes. After genes are ranked, genes are selected according to their ranks.
When there are a large number of features, a conservative strategy is to determine the least important feature one at a time recursively. In this work, we adopted the SVM-RFE (recursive feature elimination) algorithm [3] where the least important feature is identified and removed in each iteration, remaining features are re-evaluated, and the process repeats until no more features are left for consideration. For the linear kernel, the importance of a feature is determined by the associated weight magnitude, and the least important feature refers to the one with the smallest weight value. SVM-RFE essentially implements the strategy of backward feature elimination. In principle, feature ranking becomes more accurate as less important features are removed successively. To improve the speed, a chunk of least important features was eliminated per step until there were 256 genes remained, from which point, one gene was remove per step. The RFE ranking criterion is given by
Rank(gi) <Rank(gj) ⇔ Order-of-Elimination(gi) >Order-of-Elimination(gj)
That is, the later a gene is eliminated, the higher (smaller) rank it has. So, the first-rank gene is last removed.
Authors' contributions
L. Fu developed the method and conducted the experiments. C. Fu-Liu interpreted the data. Both authors drafted, read, and approved the manuscript.
Acknowledgements
This work is supported by National Institutes of Health and National Science Foundation under grants HL-080311 and IIS-0221954. E. S. Youn assisted in coding the algorithm.
Figures and Tables
Figure 1 The gene expression map of the 19 genes selected by our method in the domain concerning classification of SRBCTs. The map was generated by Eisen's hierarchical clustering program called CLUSTER and viewed by the TREEVIEW program. Four sample clusters are visually recognizable, corresponding exactly to the four predefined tumor classes (NB, EWS, BL, and RMS) with 100% accuracy.
Figure 2 The gene expression map of the 15 genes selected from the colon cancer microarray data set using our method. Two major sample clusters can be recognized by visual inspection, corresponding to normal and cancer tissue samples, respectively.
Table 1 Genes selected by our method on the microarray dataset of small round blue-cells tumors. Those genes also selected using the methods of Tibshirani et al. [13] and Khan et al. [10] are respectively marked by the symbol •.
Image ID P Value Gene Description Tibshirani et al. Khan et al.
21652 2.3 × 10-5 catenin (cadherin-associated protein), alpha 1 • •
878280 2.3 × 10-5 collapsin response mediator protein 1 •
377461 < 0.000001 caveolin 1, caveolae protein • •
325182 2.3 × 10-5 cadherin 2, N-cadherin (neuronal) • •
1435862 0.02 MIC2 surface antigen (CD99) • •
42558 0.02 L-arginine:glycine amidinotransferase • •
812105 < 0.000001 transmembrane protein • •
41591 < 0.000001 meningioma 1 • •
810057 < 0.000001 cold shock domain protein A •
183337 0.02 major histocompatibility complex, class II, DM alpha • •
796258 < 0.000001 sarcoglycan, alpha • •
1409509 0.02 troponin T1, skeletal, slow • •
788107 < 0.000001 amphiphysin-like •
770394 < 0.000001 Fc fragment of IgG, receptor, transporter, alpha • •
82225 0.02 secreted frizzled-related protein 1 •
814260 < 0.000001 follicular lymphoma variant translocation 1 • •
784224 < 0.000001 fibroblast growth factor receptor 4 • •
308163 2.3 × 10-5 ESTs • •
212542 < 0.000001 cDNA DKFZp586J2118 • •
Table 2 15 genes selected from the colon cancer microarray data set (62 samples) using our method.
Gene Accession # P value Definition
H20709 < 0.000001 myosin light chain alkali, smooth-muscle isoform
X57351 < 0.000001 interferon-inducible protein 1-8D
T94579 < 0.000001 human chitotriosidase precursor
T47377 < 0.000001 S-100P protein (human)
T98835 < 0.000001 alpha trans-inducing protein (bovine herpesvirus type 1)
T61661 3.0 × 10-5 profilin I (human)
X67325 3.0 × 10-5 H. sapiens p27
T58861 0.02 60s ribosomal protein L30E
T61446 0.02 putative DNA binding protein A20
H88360 0.02 guanine nucleotide-binding protein G(OLF), alpha subunit
L38810 0.02 Homo sapiens thyroid receptor interactor (TRIP1)
T57882 0.02 myosin heavy chain, nonmuscle type A
T92451 0.02 tropomyosin, fibroblast and epithelial muscle-type
J02854 0.02 myosin regulatory light chain 2, smooth muscle isoform
K03474 0.02 human mullerian inhibiting substance gene
Table 3 Diagnosis results of the colon cancer data samples based on 15 selected genes, in correspondence with the gene expression map.
Normal Tissue Cancer Tissue
Sample
Diagnosis
Sample
Diagnosis
Normal-01 normal Cancer-01 cancer
Normal-02 normal Cancer-02 normal
Normal-03 normal Cancer-03 cancer
Normal-04 normal Cancer-04 cancer
Normal-05 normal Cancer-05 cancer
Normal-06 normal Cancer-06 cancer
Normal-07 normal Cancer-07 cancer
Normal-08 cancer Cancer-08 cancer
Normal-09 normal Cancer-09 cancer
Normal-10 normal Cancer-10 cancer
Normal-11 normal Cancer-11 cancer
Normal-12 normal Cancer-12 cancer
Normal-13 normal Cancer-13 cancer
Normal-14 normal Cancer-14 cancer
Normal-15 normal Cancer-15 cancer
Normal-16 normal Cancer-16 cancer
Normal-17 normal Cancer-17 cancer
Normal-18 normal Cancer-18 cancer
Normal-19 normal Cancer-19 cancer
Normal-20 cancer Cancer-20 cancer
Normal-21 normal Cancer-21 cancer
Normal-22 normal Cancer-22 cancer
Cancer-23 cancer
Cancer-24 cancer
Cancer-25 cancer
Cancer-26 cancer
Cancer-27 cancer
Cancer-28 normal
Cancer-29 cancer
Cancer-30 normal
Cancer-31 cancer
Cancer-32 cancer
Cancer-33 cancer
Cancer-34 cancer
Cancer-35 cancer
Cancer-36 normal
Cancer-37 cancer
Cancer-38 cancer
Cancer-39 cancer
Cancer-40 cancer
Table 4 Genes selected by our method on the leukemia microarray dataset. Those genes also selected using the methods of Golub et al.[1] and SVM-RFE (the reference algorithm) are respectively marked by the symbol •.
Access Number P Value Gene Description Golub et al. SVM-RFE
M27891 < 0.000001 CST3 Cystatin C • •
Y00787 < 0.000001 INTERLEUKIN-8 PRECURSOR • •
M19507 0.006 MPO Myeloperoxidase •
L20688 0.006 Ly-GDI
==== Refs
Golub TR Slonim DK Tamayo P Huard C Gaasenbeek M Mesirov JP Coller H Loh ML Downing JR Caligiuri MA Bloomfield CD Lander ES Molecular classification of cancer: class discovery and class prediction by gene expression monitoring Science 1999 286 531 537 10521349 10.1126/science.286.5439.531
Xiong M Li W Zhao J Jin L Boerwinkle E Feature (gene) selection in gene expression-based tumor classification Mol Genet Metab 2001 73 239 247 11461191 10.1006/mgme.2001.3193
Guyon I Weston J Barnhill S Vapnik V Gene selection for cancer classification using support vector machines machine learning 2002 46 389 422 10.1023/A:1012487302797
Fu LM Youn ES Improving reliability of gene selection from microarray functional-genomics data IEEE Transactions on Information Technology in Biomedicine 2003 7 191 196 14518732 10.1109/TITB.2003.816558
Lee KE Sha N Dougherty ER Vannucci M Mallick BK Gene selection: a Bayesian variable selection approach Bioinformatics 2003 19 90 97 12499298 10.1093/bioinformatics/19.1.90
Ambroise C McLachlan GJ Selection bias in gene extraction on the basis of microarray gene-expression data Proc Natl Acad Sci U S A 2002 99 6562 6566 11983868 10.1073/pnas.102102699
Tusher VG Tibshirani R Chu G Significance analysis of microarrays applied to the ionizing radiation response Proc Natl Acad Sci U S A 2001 98 5116 5121 11309499 10.1073/pnas.091062498
Fu LM Fu-Liu CS Multi-class cancer subtype classification based on gene expression signatures with reliability analysis FEBS Lett 2004 561 186 190 15013775 10.1016/S0014-5793(04)00175-9
Fu LM Cancer Subtype Classification Based on Gene Expression Signatures
Khan J Wei JS Ringner M Saal LH Ladanyi M Westermann F Berthold F Schwab M Antonescu CR Peterson C Meltzer PS Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks Nat Med 2001 7 673 679 11385503 10.1038/89044
Alon U Barkai N Notterman DA Gish K Ybarra S Mack D Levine AJ Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon tissues probed by oligonucleotide arrays Proc Natl Acad Sci U S A 1999 96 6745 6750 10359783 10.1073/pnas.96.12.6745
Ramaswamy S Tamayo P Rifkin R Mukherjee S Yeang CH Angelo M Ladd C Reich M Latulippe E Mesirov JP Poggio T Gerald W Loda M Lander ES Golub TR Multiclass cancer diagnosis using tumor gene expression signatures Proc Natl Acad Sci U S A 2001 98 15149 15154 11742071 10.1073/pnas.211566398
Tibshirani R Hastie T Narasimhan B Chu G Diagnosis of multiple cancer types by shrunken centroids of gene expression Proc Natl Acad Sci U S A 2002 99 6567 6572 12011421 10.1073/pnas.082099299
Eisen MB Spellman PT Brown PO Botstein D Cluster analysis and display of genome-wide expression patterns Proc Natl Acad Sci U S A 1998 95 14863 14868 9843981 10.1073/pnas.95.25.14863
Baxevanis AD Ouellette BFF Bioinformatics 2001 New York, NY, John Wiley & Sons
Kerr MK Churchill GA Bootstrapping cluster analysis: assessing the reliability of conclusions from microarray experiments Proc Natl Acad Sci U S A 2001 98 8961 8965 11470909 10.1073/pnas.161273698
Statnikov A Aliferis CF Tsamardinos I Hardin D Levy S A comprehensive evaluation of multicategory classification methods for microarray gene expression cancer diagnosis Bioinformatics 2004
Kovar H Dworzak M Strehl S Schnell E Ambros IM Ambros PF Gadner H Overexpression of the pseudoautosomal gene MIC2 in Ewing's sarcoma and peripheral primitive neuroectodermal tumor Oncogene 1990 5 1067 1070 1695726
Fujimoto T Nishikawa A Iwasaki M Akutagawa N Teramoto M Kudo R Gene expression profiling in two morphologically different uterine cervical carcinoma cell lines derived from a single donor using a human cancer cDNA array Gynecol Oncol 2004 93 446 453 15099960 10.1016/j.ygyno.2004.02.012
Klein JR Hoon DS Nangauyan J Okun E Cochran AJ S-100 protein stimulates cellular proliferation Cancer Immunol Immunother 1989 29 133 138 2720706 10.1007/BF00199288
Janke J Schluter K Jandrig B Theile M Kolble K Arnold W Grinstein E Schwartz A Estevez-Schwarz L Schlag PM Jockusch BM Scherneck S Suppression of tumorigenicity in breast cancer cells by the microfilament protein profilin 1 J Exp Med 2000 191 1675 1686 10811861 10.1084/jem.191.10.1675
Dai JY Liang XP Wen JL Li CY Deng CZ Zhang ZH [Expression of P27 protein and cyclin E in colon cancer] Ai Zheng 2003 22 1093 1095 14558959
Beyaert R Heyninck K Van Huffel S A20 and A20-binding proteins as cellular inhibitors of nuclear factor-kappa B-dependent gene expression and apoptosis Biochem Pharmacol 2000 60 1143 1151 11007952 10.1016/S0006-2952(00)00404-4
Daaka Y G proteins in cancer: the prostate cancer paradigm Sci STKE 2004 2004 re2 14734786
Aliferis CF Tsamardinos I Massion P Statnikov A Fananapazir N Hardin D Machine Learning Models For Classification Of Lung Cancer and Selection of Genomic Markers Using Array Gene Expression Data 2003
Hardin D Tsamardinos I Aliferis CF A theoretical characterization of linear SVM-based feature selection: ; Banff, Alberta, Canada. 2004 ACM Press, New York, NY
Tsamardinos I Constantin F. Aliferis CF Alexander Statnikov A Time and sample efficient discovery of Markov blankets and direct causal relations: ; Washington, D.C.. 2003
Cho JH Lee D Park JH Lee IB New gene selection method for classification of cancer subtypes considering within-class variation FEBS Lett 2003 551 3 7 12965195 10.1016/S0014-5793(03)00819-6
Haykin S Neural Networks: A Comprehensive Foundation 1999 Second Upper Saddle River, NJ, Prentice Hall
Cristianini N Shawe-Taylor J Support Vector Machines 2000 Cambridge, UK, University Press
Brown MP Grundy WN Lin D Cristianini N Sugnet CW Furey TS Ares MJ Haussler D Knowledge-based analysis of microarray gene expression data by using support vector machines Proc Natl Acad Sci U S A 2000 97 262 267 10618406 10.1073/pnas.97.1.262
Cortes C Vapnik V Support vector networks Machine Learning 1995 20 273 297
|
15784140
|
PMC1274261
|
CC BY
|
2021-01-04 16:02:50
|
no
|
BMC Bioinformatics. 2005 Mar 22; 6:67
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-67
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-681578809510.1186/1471-2105-6-68Research ArticleFeature selection and nearest centroid classification for protein mass spectrometry Levner Ilya [email protected] Department of Computing Science, University of Alberta, Canada2005 23 3 2005 6 68 68 9 12 2004 23 3 2005 Copyright © 2005 Levner; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The use of mass spectrometry as a proteomics tool is poised to revolutionize early disease diagnosis and biomarker identification. Unfortunately, before standard supervised classification algorithms can be employed, the "curse of dimensionality" needs to be solved. Due to the sheer amount of information contained within the mass spectra, most standard machine learning techniques cannot be directly applied. Instead, feature selection techniques are used to first reduce the dimensionality of the input space and thus enable the subsequent use of classification algorithms. This paper examines feature selection techniques for proteomic mass spectrometry.
Results
This study examines the performance of the nearest centroid classifier coupled with the following feature selection algorithms. Student-t test, Kolmogorov-Smirnov test, and the P-test are univariate statistics used for filter-based feature ranking. From the wrapper approaches we tested sequential forward selection and a modified version of sequential backward selection. Embedded approaches included shrunken nearest centroid and a novel version of boosting based feature selection we developed. In addition, we tested several dimensionality reduction approaches, namely principal component analysis and principal component analysis coupled with linear discriminant analysis. To fairly assess each algorithm, evaluation was done using stratified cross validation with an internal leave-one-out cross-validation loop for automated feature selection. Comprehensive experiments, conducted on five popular cancer data sets, revealed that the less advocated sequential forward selection and boosted feature selection algorithms produce the most consistent results across all data sets. In contrast, the state-of-the-art performance reported on isolated data sets for several of the studied algorithms, does not hold across all data sets.
Conclusion
This study tested a number of popular feature selection methods using the nearest centroid classifier and found that several reportedly state-of-the-art algorithms in fact perform rather poorly when tested via stratified cross-validation. The revealed inconsistencies provide clear evidence that algorithm evaluation should be performed on several data sets using a consistent (i.e., non-randomized, stratified) cross-validation procedure in order for the conclusions to be statistically sound.
==== Body
Background
Advances in protein mass spectrometry have have recently shown great potential for high-throughput disease classification and biomarker identification. In turn, fast and accurate detection of diseases, such as early cancer detection, can revolutionize the field of medical diagnosis. Typically, serum samples are analyzed by a mass spectrometer, producing a high dimensional abundance histogram. Next, informative features are extracted from the high dimensional data and are presented to a classifier. In turn, the classifier outputs a decision about the status of the patient with respect to a particular disease (e.g., healthy or diseased). Recently, numerous feature selection and classification techniques have been shown to perform well on several isolated data sets. However, current literature does not contain rigorous comparative studies analyzing the merits of individual feature selection and classification algorithms across several data sets. This paper analyzes several state-of-the-art feature selection methods coupled with a very fast nearest centroid classifier. In addition, we present a novel combination of boosted feature extraction coupled with the nearest centroid classifier, which consistently outperforms all other algorithms tested in terms of classification accuracy.
Mass spectrometry analysis
Discovered by Sir J.J. Thomson in the early part of the 20th century Mass spectrometry (MS) is a technique for 'weighting' individual molecules, fragments of molecules or individual atoms that have been ionized. In a vacuum environment an ion source vaporizes and charges the sample matter, which is then deflected into a magnetic or electric field. The mass spectrometer then measures the molecular masses along with abundances and masses of fragments that are produced as a result of molecular breakdown. The fundamental measurement unit of the MS is the mass-to-change ratio (M/Z). For proteomic applications, Daltons (Da) are used to measure mass, while the electric potential of a single electron is the measurement unit for charge (z). The spectrum is a graph of ion intensity as a function of mass-to-charge ratio and is often depicted as a histogram.
Time-of-Flight (TOF)
In time-of-flight (TOF) instruments, positive ions are produced by periodic bombardment of the sample with brief pulses of either electrons, secondary ions, or laser-generated photons. The ions produced by the laser are then accelerated by an electric field pulse and passed into a field-free drift tube. Ideally, all ions entering the tube will have the same kinetic energies, and their velocities must therefore vary inversely with their masses, with lighter particles arriving at the detector earlier than the heavier ones. The ions therefore drift through a field-free path and are separated in space and time-of-flight [5].
Matrix assisted laser desorption/ionization (MALDI)
By incorporating the (bio)molecules in a large excess of matrix molecules, strong intermolecular forces are reduced. The matrix molecules absorb the energy from the laser light and transfer it into excitation energy of the solid system. The effect is an instantaneous phase transition of small molecular layers of the sample into a gaseous state. Thus solid (and liquid) material can be easily analyzed by TOF MS.
Surface-enhanced laser desorption/ionization (SELDI)
This method uses protein chip arrays with different selective surfaces such as cation or anion exchange surfaces, hydrophobic surfaces and metal binding surfaces. Biofluids such as cell lysate, plasma or urine are applied onto the selective surface and, after washing, a subset of proteins is specifically bound. The chip is then analyzed in a (MALDI) TOF-MS which generates a protein spectrum of the different molecular masses present on the protein chip. This technology is therefore highly suited for research into molecular mechanisms of disease and biomarker identification.
Related research
Mass Spectrometry (MS) based pattern recognition is rapidly becoming a broad and fruitful research field. This section, provides details on current state of research within the field of early cancer diagnosis based on proteomic pattern recognition.
Ovarian cancer studies
In [19], genetic algorithms together with self-organizing maps were used to distinguish between healthy women and those afflicted with ovarian cancer. Although cross-validation studies were not conducted, the approach was able to correctly classify all cancer stricken patients and 95% of healthy women, on a single test set.
Using the same data sets in [17], the researchers employed Principle Component Analysis (PCA) [14] for dimensionality reduction and Linear Discriminant Analysis (LDA) [8] coupled with a nearest centroid classifier [18] for classification. For each of the train/test data splits, 1000 cross-validation runs with re-sampling were conducted. When training sets were larger than 75% of the total sample size, perfect (100%) accuracy was achieved on the OC-WCX2b data set. Using only 50% of data for training, the performance dropped by 0.01%. Unfortunately, the probabilistic approach used in this study can leave some samples unclassified. For the OC-H4 data set, the system had a 92.45% sensitivity and 91.95% specificity when 75% of the data was used for training. However, only 98.60% of the data samples were classified. Similarly, for the OC-WCX2a data set 97.34% sensitivity and 96.99% specificity was attained on 99.92% of the test data, when 75/25 train/test split was used.
In [30], the researchers compared two feature extraction algorithms together with several classification approaches on a MALDI TOF acquired data. The T-statistic, also known as the student-t test [21], was used to rank features in terms of their relevance. Then two feature subsets were greedily selected (respectively having 15 and 25 features each). Support vector machines (SVM), random forests, linear/quadratic discriminant analysis (LDA/QDA), k-nearest neighbors, and bagged/boosted decision trees were subsequently used to classify the data. In addition, random forests were also used to select relevant features with previously mentioned algorithms used for classification. Again 15 and 25 feature sets were selected and classification algorithms applied. When the T-statistic was used as a feature extraction technique, SVM, LDA and random forests classifiers obtained the top three results (with accuracy in the vicinity of 85%). On the other hand, classification improved to approximately 92% when random forests were used as both feature extractors and classifiers. Similar performance was also achieved using the the nearest-neighbor algorithm, a close relative of the nearest centroid algorithm [28] we will be using in this study. While the results appear promising, the authors provide little motivation as to why 15 and 25 feature sets were selected. Other that the fact that LDA and QDA need the number of features to be less than the number of samples, the actual size of the selected feature set seems to be an arbitrary choice. In practice, determining the size of the feature set is an added burden, placed on the software developer and, ideally, should be eliminated. Furthermore, testing several feature sets of various sizes and selecting the set with the best performance can lead to overfitting. With that in mind we propose to automatically select features and the size of the feature set using an internal leave-one-out cross-validation procedure (LOOCV) discussed in the following sections.
Using the same MALDI TOF data set as in [30], researchers in [26] applied the nearest shrunken centroid approach to classify the MS samples. Using only seven features their method achieved a classification error rate of approximately 23%. More recently, in [12], both the GA approach and the nearest shrunken centroid approach have been found inferior to the boosting based feature selection approach. Further investigation, in [16], confirmed the poor performance of the nearest shrunken centroid on the ovarian cancer (OC-H4) and the prostate cancer (PC-H4) data sets.
Prostate cancer studies
In [1], the researchers used a decision tree algorithm to differentiate between healthy individuals and those with prostate cancer. This study used the SELDI TOF MS to acquire the mass spectra which corresponds to our PC-IMAC-Cu data set. In order to select relevant features, the area under the Receiver Operating Characteristics (ROC) curves was used to identify informative peaks which were subsequently used by the decision tree classification algorithm. The researchers did not perform cross-validation, but on a single test set the classifier achieved an 81% sensitivity and a 97% specificity, yielding a balanced accuracy (BACC) of 89%.
In [22], the performance was improved on the PC-IMAC-Cu data set by the use of boosting. As is [1], the area under the curve (AUC) criteria was used to identify relevant features. For subsequent feature selection and classification, the researchers used decision stumps together with AdaBoost and its variant, Boosted Decision Stump Feature Selection (BDSFS) method. A key difference between the two methods is that BDSFS selects features without replacement, whereas boosted decision stumps (BDS) allows for selection of the same feature multiple times. The BDS algorithm achieved perfect accuracy on the single test set for the prostate cancer data set. However, a randomized 10-fold cross-validation procedure yielded an average sensitivity of 98.5% and an average specificity of 97.9%, for an overall BACC of 98%. For the BDSFS, the results were considerably worse, with a sensitivity of 91.1% and a specificity of 94.3%. The BDS algorithm used all 124 features selected by the AUC, and required 500 rounds of boosting. On the other hand, the BDSFS algorithm used just 21 features which were easily interpretable. The researchers informally report that other classifiers had similar classification accuracies but were more difficult to interpret. Although, this is the highest reported accuracy on this data set, the BDS algorithm [9] required over 500 rounds of boosting which complicates the identification of key relevant features necessary to differentiate heathy individuals from those afflicted with prostate cancer.
In [29], the same PC-IMAC-Cu data set was analyzed using several classifiers. Using a filter-based ANOVA F-statistic to rank the preselected peaks, relevant features were selected in sets of increasing size. Classification was performed with k-nearest-neighbors (kNN), linear/quadratic discriminants (LDA/QDA), and suport vector machines (SVM) using 100-fold randomized cross-validation strategy. Linear SVM achieved the best accuracy of 91% using just eight peaks.
In [17], the researchers again used PCA for dimensionality reduction and LDA for classification. The PC-IMAC-Cu data set was obtained from the authors of [1] and, in the same fashion as with the ovarian cancer set, the researchers conducted a detailed study using various train/test set sizes. For each train/test data split, 1000 cross-validation runs (with re-sampling) were conducted. When training sets were larger than 75% of the total sample size, average accuracy of 88% was achieved (88.46% sensitivity and 88.98% specificity). Using only 50% of data for training, the performance dropped to 86%. In comparison to ovarian cancer sets, the lower accuracy suggests that this data set is much more difficult to classify correctly using the PCA/LDA algorithm.
In [20,31], researchers used Genetic Algorithms (GA's) for feature selection and Self Organizing Maps (SOM's) for classification of prostate cancer (data set PC-H4 in our study). This approach achieved a specificity of 95% and a sensitivity of 71%, for an average accuracy of 83%. Although cross-validation was carried out, the results were not presented.
In [6], the aforementioned studies on prostate cancer raised the following question: Why do the features and classification performance vary so drastically across studies? The results indicate that different SELDI-TOF approaches combined with different machine learning techniques for pattern recognition produce highly variable results in terms of relevant features and classification accuracy. Furthermore, such results also indicate that the MS spectra contains a large number of features relevant to the task of discriminating heathy individuals from those afflicted with cancer. An alternative explanation, found in [2], seems to suggest chemical/electronic noise and/or bias introduced during the acquisition of the MS spectra. This further motivates the need for comparative studies done on a regular basis using several mass spectrometry techniques in conjunction with a number of machine learning approaches done on several data sets.
Data sets
For this study five data sets were acquired. Each sample in each data set is represented as a vector of real valued features forming the spectra. Each feature in turn represents the quantity (parts per million) of ions with a specific m/z ratio. In essence, each sample spectrum is a histogram describing the composition of the sample bio-fluid or tissue sample. Each data set is named based on the type of disease tested, OC for Ovarian Cancer and PC for Prostate Cancer, as well as the type of SELDI affinity chip used to produce the mass spectra. This naming scheme was adopted from [17]. The following data sets were used for this study:
OC-H4
This ovarian cancer set was obtained using the H4 protein chip from Ciphergen. It contains 100 diseased and 100 healthy samples which were manually prepared. Each spectra contains 15,156 features (M/Z values) in this data set.
OC-WCX2a
This ovarian cancer set obtained using the WCX2 protein chip. It contains the same 100 diseased and 100 healthy samples as the OC-H4 data set which were re-precessed using the WCX2 protein chip. For this data set the samples were also processed by hand. Each spectra contains 15,156 features (M/Z values) in this data set.
OC-WCX2b
This ovarian cancer set was also obtained using the WCX2 protein chip. However, a robotic instrument replaced the manual chip preparation for this data set. This data set contains 92 healthy and 162 diseased samples, all different from the two previous data sets. Each spectra contains 15,156 features (M/Z values) in this data set.
PC-H4
The spectra were collected using the H4 protein chip, which was prepared by hand. There are 322 total samples: 190 samples with benign prostate hyperplasia with PSA levels greater than 4, 63 samples with no evidence of disease and PSA level less than 1, 26 samples with prostate cancer with PSA levels 4 through 10, and 43 samples with prostate cancer with PSA levels greater than 10. Each sample is again a histogram composed of 15,156 features. For this set we combined samples with benign prostate hyperplasia and those with no evidence of disease into the healthy class. The rest of the samples formed the diseased class.
PC-IMAC-Cu
The spectra were collected using the IMAC-Cu metal binding chip, and were prepared by hand. There are 324 total samples: 167 samples with prostate cancer, 77 with benign prostate hyperplasia and 82 samples with no evidence of disease. Each sample is composed of 16,382 features. For this set we also combined samples with benign prostate hyperplasia and those with no evidence of disease into the healthy class. The rest of the samples formed the diseased class.
Results
This section presents the evaluated feature selection algorithms in conjunction with the base classification technique. In addition the empirical evaluation results are presented.
Centroid classification method
A fast and simple algorithm for classification is the centroid method [10,18]. This algorithm assumes that the target classes correspond to individual (single) clusters and uses the cluster means (or centroids) to determine the class of a new sample point. A prototype pattern for class Cj is defined as the arithmetic mean:
where xi's are the training samples labeled as class Cj. Recall that the training sample is a MS spectra represented as a multi-dimensional vector (denoted in bold). In a similar fashion, we can obtain a prototypical vector for all the other classes. During classification, the class label of an unknown sample x is determined as:
where d(x, y) is a distance function or:
where s(x, y) is a similarity metric. This simple classifier will form the basis of our studies. It works with any number of features and its run-time complexity is proportional to the number of features and the complexity of the distance or similarity metric used. Preliminary experiments in [15], were conducted to establish which similarity/distance metric is most appropriate for the centroid classification algorithm, and the L1 distance metric was selected. Defined by:
L1 (x, μ) = || x - μ||1 (1)
with ||y||1 = |y(i)|, and y(i) being the value of the ith feature. The value L1(x, μ) has a linear cost in the number of features. In this study, data sets contain two classes and hence the number of calls to the distance metric is also two. Therefore, the centroid classifier, at run-time, is linear in the number of features. During training, two prototypes are computed and the cost of computing each prototype is O(mN), where N is the number of features and m is the number of training samples which belong to a given class. Note that m only varies between data sets and not during training or feature selection processes. Thus, we can view m as a constant and conclude that the centroid classifier has O(N)cost in the training phase.
Nearest shrunken centroid
A special purpose feature selection algorithm for the nearest centroid algorithm was developed by Tibshirani et al. and presented in [10,25,26]. The algorithm, related to the lasso method, tries to shrink the class prototypes () towards the overall mean:
Briefly, the algorithm calculates:
where , s is a vector of pooled within class variances for each feature and division is done component wise. We can now view the class centroid as:
where denotes component wise multiplication. By decreasing dj we can move the class centroid towards the overall centroid. When a component of the class centroid is equal to the corresponding component of the overall mean for all classes, the feature no longer plays a part in classification and is effectively removed. Hence, as dj shrinks progressively more features are removed. To decrease dj soft thresholding is used to produce with:
Where dj(i) is the ithcomponent of the vector dj. The shrunken centroid is then computed by replacing dj with in equation 4. In our experiments we used 20 different values for δ, NSC(20), {0.5, 1, 1.5, ..., 10}.
We also tried 200 different values for δ also in the range (0, 10] in increments of 0.05, but attained the same BACC score while incurring ten times the computational cost (results not shown).
Filter-based feature selection
Filter methods attempt to select features based on simple auxiliary criteria, such as feature correlation, to remove redundant features. In order to be tractable, such approaches decouple the feature selection process from the performance component, but may ultimately select irrelevant features as a result. In general, filter-based methods are designed for a specific type of feature. Since the mass spectra is composed of continuous features, we use univariate statistical tests. Instead of selecting features by invoking a classifier as in wrapper-based approaches, univariate statistics simply rank individual features. The student-t test (T-test), the Kolmogorov-Smirnov test (KS-test) [21] and the P-test [11] algorithms are the commonly used statistics. These 'goodness-of-fit' tests compare feature values of samples belonging to class 1 to feature values of samples belonging to class 2. The goal is to determine if the feature values for class 1 come from a different distribution than those for class 2. The key difference between these tests are the assumptions they make. The T-test assumes that both distributions have identical variance, and makes no assumptions as to whether the two distributions are discrete or continuous. On the other hand, the KS-test assumes that the two distributions are continuous, but makes no other assumptions.
In the case of the T-test, the null hypothesis is μ1 = μ2, indicating that the mean of feature values for class 1 is the same as the mean of the feature values for class 2. In the case of the KS-test, the null hypothesis is cdf(1) = cdf(2), meaning that feature values from both classes have an identical cumulative distribution. Both tests determine if the observed differences are statistically significant and return a score representing the probability that the null hypothesis is true. Thus, features can be ranked using either of these statistics according to the significance score of each feature. In addition to the T-test and KS-test, we also use a simpler feature ranking criteria called the P-test and denoted as:
where σi is the standard deviation for class i. This can be seen as a simplified version of the student-t score that ignores sample size and ranks features solely on the basis of their mean and standard deviation. Both the benefits and drawbacks of these statistical tests stem from the assumption that the features are independent. On one hand, the independence assumption makes these algorithms computationally efficient. On the other hand, the independence assumption clearly may not hold for all data sets, thereby producing suboptimal feature rankings.
In [30], the researchers used the T-test to rank each feature but chose to test classification algorithms with 15 and 25 top-ranked features, without any apparent justification. The apparent focus of their research is on comparing classifiers rather than the two feature extraction methods (T-test and random forests). In contrast, we show that feature ranking coupled with greedy forward selection using internal leave-one-out cross-validation (LOOCV) can automatically find a feature subset of an arbitrary size that improves performance with respect to using the centroid algorithm without any feature selection.
Wrapper-based feature selection
Wrapper Methods attempt to evaluate feature relevance within the context of a given task and avoid intractability by using greedy/heuristic search methods. In other words, the number of possible subsets is greatly restricted by the greedy selection procedure, and each candidate feature subset is evaluated using the actual performance element (i.e., training a classifier/regressor using a subset of features). Thus far, a variety of greedy algorithms have been proposed to select feature sets sequentially. Sequential Forward (respectively Backward) selection (SFS and SBS) methods start from an empty (respectively full) set of features and at each step add (respectively remove) a single feature which produces the greatest increase in performance. The SFS technique, as described, is easily applicable to the MS data. On the other hand, the SBS algorithm, much like a full search over all subsets, is still computationally intractable. Our informal estimates revealed that a naive application of the SBS algorithm to all five data sets, used in this study, would take approximately 100 years to complete on the hardware platform available to us. Thus, in order to make SBS tractable, we implemented several heuristics. First, rather than searching through all features within the active set, and removing a feature that produces the greatest improvement in performance, we stop at the first feature whose removal does not degrade the overall performance as determined by the internal LOOCV approach. Now that each loop of SBS terminates at the first candidate feature, we can re-order the features based on the probability of each feature being irrelevant and/or redundant. To do so we use the KS-test to rank and re-order all features. Thus, the SBS search starts by first testing a feature deemed most likely to be irrelevant by the KS-test. The second heuristic added to the SBS algorithm involves recording the stoping position of the last iteration. In the standard SBS, each iteration of the algorithm tests all features in the active set. However, since the previously added heuristic lets SBS terminate the innermost loop at the first feature deemed unnecessary, re-testing previously examined features has less utility than looking at the uninspected features. Hence, rather than re-starting the search from the beginning, each iteration of the modified SBS starts the feature search from the previous stopping position. Upon reaching the end of the feature index array, the search is restarted from the beginning.
Boosting
In addition to SFS and the modified SBS, we also use boosting which has been shown to perform very well on the PC-IMAC-Cu data set in [22]. To determine the merit of this embedded feature selection approach, we created two versions of the boosting algorithm. The first version is a standard boosting algorithm [23] that uses a weighted nearest centroid method as the base learner. As in the standard nearest centroid, the first round of boosting assigns equal weights to each sample and calculates the nearest centroid for each of the two classes. Each training sample is then classified and re-weighted based on the outcome of classification. If a sample is misclassified, it receives a higher weight (for the next boosting round), whereas if the sample was correctly classified its weight is decreased. The next round of boosting creates new centroids based on the adjusted sample weights and the process repeats itself until training error becomes zero or a predefined number of boosting rounds is reached. This version of the algorithm does not perform feature selection and is used to assess the performance of the second version of boosted nearest centroid algorithm.
The second version of the algorithm extends the boosting algorithm by enabling feature selection. This version, called boosted feature extraction (boostedFE), is similar to sequential forward selection (SFS) in that during each round of boosting the algorithm searches over all features and selects a single best feature upon which to build the weighted nearest centroid classifier. Although variants of this approach have been used in [22] and [27], to the best of our knowledge this is the first time the boostedFE algorithm has been coupled with the (weighted) nearest centroid classifier. The finer aspects of this algorithm are presented in the discussion section of this paper.
Dimensionality reduction
Feature selection algorithms attempt to select relevant features with respect to the performance task, or conversely remove redundant or irrelevant ones. In contrast, the goal of dimensionality reduction techniques is to literally transform the raw input features while preserving the global information content. In essence, the dimensionality reduction algorithms attempt to extract features capable of reconstructing the original high dimensional data, irrespective to the classification label assigned to each data point. For example, principle components analysis (PCA) [14] attempts to find a linear combination of principal components that preserves the variance of the data. In order to test dimensionality reduction algorithms, we have procured the Q5 code used in [17], which uses PCA in conjunction with linear discriminant analysis (LDA) to classify the sample mass spectra. Briefly, PCA projects the MS spectra onto a low dimensional linear manifold required by the LDA algorithm, which cannot use more features than training instances. In turn the LDA algorithm attempts to project the data onto a hyperplane which minimizes within-class scatter, while maximizing between-class distance. Once the data has been projected into the LDA subspace, the nearest centroid approach is used to classify new instances. In our experiments, we test both PCA/LDA + nearest centroid as well as PCA + nearest centroid approaches. This design is meant to assess the merit of individual components, namely PCA and LDA.
Empirical evaluation
We conducted experiments on three ovarian and two prostate data sets, previously used in [1,2,4,17,19,20,22]. Sets OC-H4, OC-WCX2a, OC-WCX2b, and PC-H4 contain 15,156 features (i.e., m/z values), while the last data set PC-IMAC-Cu contains 16,382 features.
We used a stratified three-fold cross-validation procedure, for all experiments, whereby each data set was split into three subsets of equal size. Each test fold used one of the three subsets with the remaining two subsets used for training. Within the training phase an internal leave-one-out cross-validation (LOOCV) loop was used for for all feature selection methods (with the exception of dimensionality reduction approaches). In this manner, test set performance remains unbiased by the feature selection process. For PCA and PCA/LDA algorithms, the maximal number of principal components usable by the LDA algorithm was selected and is further described in [17]. The results presented in Figure 1 and Tables 1, 2, 3, 4, 5 express performance statistics averaged over the three test folds. Balanced accuracy (BACC) is taken as the arithmetic mean of sensitivity and specificity and is formally defined in the List of Abbreviations section along with the rest of the performance measures. The BACC measure is related to the standard BER (Balanced Error Rate), where BER = 1 - BACC is commonly used for evaluation of feature selection algorithms [16].
Classification accuracy
Figure 1 and Table 3 present balanced accuracy (BACC) results across the five data sets. The results indicate that boosting based feature extraction (boostedFE) produces the most significant improvement in classification accuracy with respect to performance of the nearest centroid algorithm without feature selection (NoFE). On four of the five data sets boostedFE attained equal or better BACC than any other algorithm tested. On the OC-H4 data set boosting without feature selection slightly outperformed boostedFE algorithm by approximately 3%. However, the difference is not statistically significant as indicated by the paired student-t test at 95% significance level (the probability of the null hypothesis being true is 73.7%). In all other cases boostedFE outperformed the other algorithms including the boosted nearest centroid algorithm.
To make our results comparable with those of Qu et al. in [22], we reran the boosted feature extraction algorithm using ten fold cross-validation scheme on the PC-IMAC-Cu data set and obtained BACC of 98.1%. More specifically our algorithm attained 100% specificity and 96.25% sensitivity. Qu et al. achieved a 98.5% sensitivity and 97.9% specificity averaged over ten 90/10% randomized train/test splits. However, their boosted decision stumps algorithm required 500 rounds of boosting to achieve such a high performance level. As a result, identification of relevant features and their significance is made difficult if not impossible. To find at least some of the relevant features within the PC-IMAC-Cu data set, in [22] the researchers employed the BDSFS algorithm which found 21 relevant features but had a significantly lower accuracy. In contrast, our boostedFE nearest centroid algorithm only required, on average, 5(± 2.8) boosting rounds to achieve comparable classification accuracy. To be fair, we note that the the BDS and BDSFS algorithms used in [22] were ran on pre-processed data, whereby 124 peaks were extracted by the AUC procedure. Hence the performance of our boostedFE algorithm is only comparable in terms of classification accuracy and the number of features selected to the BDS + AUC preprocessing. The quality of features in terms of biological relevance cannot be assessed using this or any of the other tested datasets due to i) biologically confounding factors introduced during sample acquisition and ii) ill-defined data preprocessing steps (the next section discusses these topics in more detail).
The rest of the tested algorithms did not produce consistent results. Some algorithms performed well on one or two of the data sets, but not on all of them as shown in Table 3 (and Figure 1, bottom graph). In contrast, boostedFE consistently produced high quality results on all the tested data sets. In addition, boostedFE produced results with the lowest variance across the cross-validation folds as shown in Tables 1 and 2 by low standard deviation scores. Again, the OH-H4 data set is the exception, where boostedFE has a high standard deviation for the BACC score. A closer look at Table 1 shows that the boostedFE algorithm had 100% (± 0.0) sensitivity but only 70.7% (± 0.26) specificity.
In terms of merely increasing the classification accuracy without performing feature selection, the standard boosting algorithm improved average performance by over 11% as seen in Table 3. Analysis of the training data revealed that in most cases boosting terminated in less than 21 rounds, indicating that for the five datasets used in this study, very few prototypes were needed for accurate sample classification. To see this, recall that in each round of boosting two centroids are produced, one for each class but the size of the training set ranges from 130 samples to 212 samples. Hence, boosting effectively abstracted the training samples into prototypes, producing about 21 class prototypes for each class. Unfortunately, this approach is unlikely to provide insight into the underlying biological factors, provided they exist, due to its use of the full mass spectra.
Surprisingly, the sequential backward selection (SBS) performed rather poorly across all relevant aspects, such as accuracy, running times and size of selected feature subsets. Even more surprising was the poor classification accuracy of T-test, NSC(20), and PCA/LDA algorithms, which appear highly accurate in publications [26] and [17]. Again, the effects of pre-processing steps need to be factored in when comparing our results and those of other studies. Detailed experimental results of this study are presented in Tables 1 and 2 in order to show additional performance statistics such as sensitivity, specificity and positive predictive value obtained under our experimental conditions.
For the OC-H4 data set, it appears that the filter-based methods, SFS, and NSC(20) improve specificity at the cost of decreased sensitivity. In contrast, SBS and boosting based methods improve both with respect to the basic nearest centroid algorithm. This trend resurfaces again for the PC-H4 dataset. This time all algorithms increase specificity at the cost of decreased sensitivity. It is interesting to note that both methods were created via the Ciphergen H4 ProteinChip array and both datasets had their baseline subtracted.
Feature sets
Table 4 presents statistics on sizes of selected feature subsets. Clearly, the SBS algorithm produces the largest subsets, while the rest of the algorithms produce feature subsets of significantly smaller size. In contrast, the SFS and T-test consistently select very small sets of features across data sets. The boostedFE algorithm also performs quite consistently in terms of the number of features selected. For PCA and PCA/LDA there is no clear way to identify relevant features. However, the number of principal components can be viewed as the degree of compression for a given data set. The number of components for PCA and PCA/LDA is the same and, furthermore, it is constant for a given data set since we select the maximal number of principal components (as in [17]) usable by LDA. In turn, the number of usable dimensions for LDA is given by: min(#samples, #features) – #classes and ranges between 130 and 212 dimensions for our data sets. Hence the number of dimensions used by PCA and PCA/LDA algorithms is comparable to the size of the feature sets selected by the SBS algorithm.
Computational cost comparison
Table 5 shows the computational costs of each feature selection algorithm run on each of the data sets. All experiments were conducted on a dual CPU Athlon 1400+, running Linux. The algorithms were implemented in Matlab 6.5. As expected, filters and dimensionality reduction algorithms have low computational costs. This is due to the fact that the computational complexity of these algorithms is largely governed by sample size. Hence, the run-time performance reflects the small sample size, as compared to the number of features, within the tested data sets. On the other hand, wrapper-based and boosting approaches are computationally much more expensive, in some cases by several orders of magnitude. This is due to fact that feature subsets are evaluated by training a classifier and evaluating performance on a validation set. In our case we use LOOCV, an even costlier but more accurate approach for evaluating the quality of a set of features. However, since the LOOCV approach was also used for the filter methods, the added computational costs can be directly attributed to repeated classifier training. The nearest shrunken centroid method has an additional factor influencing computational cost, namely the number of △ values examined. The NSC(20) used only 20. We also tested NSC(200) which attained very similar classification results at the cost exceeding that of SBS.
Discussion
While it was expected that SBS would be the most costly algorithm, and that it would produce the largest feature subsets, what is surprising is the noticeably poor overall performance as seen from Table 3. It appears that the additional heuristics we have added to make the algorithm tractable, had a negative impact on the performance of SBS, or that it is simply a poor choice for feature selection in the presence of so many features. On the other hand, SFS is computationally nearly an order of magnitude cheaper than SBS, produces compact feature sets, and has the second best balanced accuracy after boostedFE. From the filter-based approaches, both the KS-test and P-test outperform the T-test in terms of both classification accuracy and running times. T-test, on the other hand, consistently produces very stable features sets as seen from Table 4. Out of the three filter approaches tested, only the T-test appears in the surveyed literature. The P-test, has been used in [11] for gene selection in DNA microarrays. To the best of our knowledge we are the first to use the Kolmogorov-Smirnov test for feature selection in proteomics.
Unexpectedly, the subspace projection methods, namely PCA and PCA/LDA do not perform well under the outlined experimental conditions. This is clearly in contradiction to the results presented in [17]. In fact, Table 3 shows that the nearest centroid classifier without feature selection outperforms PCA on all but one data set. Intuitively, the poor performance of PCA, causes the PCA/LDA combination to also perform rather poorly on three of the five data sets. We should note that the randomized re-sampling testing strategy as used in [17] and [22] along with a number of other papers has been shown to be overly optimistic due to the correlations between test and train sets (see [7] and references within for a detailed explanation). Hence, we believe that this testing methodology has a significant impact on performance. On the other hand, stratified cross-validation approaches, such as the one we have adopted in this paper, remove correlations between test sets, giving more accurate performance estimates. As a consequence, all performance statistics appear 'deflated' in comparison to results reported in previous studies. However, we believe that these, 3-fold cross-validation results, provide more realistic performance estimates and can be used to make statistically sound inferences.
Nearest centroid, SFS, and boosting
The choice of nearest centroid classifier to study feature selection was not an arbitrary one. Although the nearest centroid is one of the simplest classifiers found in the literature, nevertheless it is capable of classifying raw mass spectra without any feature selection. In addition, it is extremely fast and therefore allows the use of costly wrapper methods, such as SFS, SBS, and boostedFE, which may otherwise be intractable. Hence, not only does the nearest centroid classifier able to provide a base-line for evaluation of feature selection algorithms, it also allows us to test a number of algorithms previously inapplicable in the domain of proteomic mass spectrometry. For the two class problems considered, the nearest centroid algorithm is linear and implicitly encodes a thresholding hyperplane separating the two classes. However, when combined with boosting the algorithm becomes capable of encoding non-linear boundaries. As mentioned previously, the use of boosting effectively abstracts the training samples into prototypes. Integration of sequential forward selection (SFS) yields a further improvement. By merging weighted nearest centroid with boosting and SFS, the new algorithm is able to simultaneously select relevant features and learn a highly accurate classifier. Thus boostedFE, fulfills both rolls as a feature selection and classification algorithm. By testing the nearest centroid without feature selection, SFS, boosting, and boostedFE, we can easily gauge the effect each component has on the performance of boostedFE. In fact this piece-wise analysis can easily explain why boosting outperformed boosting FE on the OC-H4 data set.
From Table 1, we can see that SFS performed worse than NoFE (meaning nearest centroid without feature selection), hence when boosting and SFS were used together the net effect actually lowered performance in comparison to boosting without feature selection. More specifically, we can see from Table 1 that the specificity of SFS for the OC-H4 data set was extremely low (51.5%) and was accompanied with a very high standard deviation of (± 42%).
Feature analysis
The aim of this paper was to profile a number of feature selection algorithms coupled with the nearest centroid classifier. Our goal was to examine performance in terms of computational time, feature set sizes and, most importantly, classification accuracy. However, due to the concerns raised in [2,24] regarding the quality of ovarian and prostate cancer data, we make no attempt to interpret the results of feature selection from a biological standpoint. Furthermore, data preprocessing strategies, themselves being actively studied [3], should also be examined in future investigations due to their influence on feature selection and classification results. In order to truly assess biological underpinnings of discriminative m/z values, it is imperative that datasets free from flaws, which confound biology with instrument noise, collection bias, and/or other "artifacts of sample effects" [2], are used in further studies. In addition, the effectiveness of preprocessing methods can only be assessed with respect their ability to improve identification of relevant biological factors governing class discrimination.
Conclusion
Mass spectrometry based disease diagnosis is an emerging field, with the potential to revolutionize early medical diagnosis. However, due to the vast amount of information captured by the high-resolution mass spectrometry techniques, the supervised training of classifiers is problematic. Specifically, the many thousands of raw attributes forming the mass spectra frequently contain a large amount of redundancy, information irrelevant to a particular disease, and measurement noise. Therefore, aggressive feature selection techniques are crucial for learning high-accuracy classifiers and realizing the full potential of mass spectrometry based disease diagnosis. This paper analyzed dimensionality reduction, filter, wrapper, and boosting based approaches to feature selection and compared the results to previously published state-of-the-art performance. In addition, a novel combination of nearest centroid classifier coupled with boosting based feature selection (boostedFE) was presented and evaluated. Experimental results indicate that sequential forward selection, P-test, and KS-test perform reasonably well across the proteomic data sets we acquired. However, the aforementioned algorithms lack consistency. On the other hand, the proposed boostedFE algorithm greatly reduces the dimensionality of the data and significantly improves classification accuracy. In contrast to all other algorithms, its performance is much more consistent across all five data sets used in the experiments.
Future research will investigate the extent to which the features selected by the boostedFE approach can be used in conjunction with more sophisticated classifiers, such as artificial neural networks and support vector machines. In addition, future studies should investigate whether the boostedFE + nearest centroid combination can serve as a meta-wrapper for more sophisticated classification algorithms. From a biological perspective, the significance of the selected features and their value in identifying potential biomarkers should be investigated. A prerequisite for this task is the production of datasets where biological factors are not confounded by instrumentation noise, sample acquisition bias and/or other experimental design flaws. The production of these datasets would also enable future studies to accurately assess the effectiveness of preprocessing techniques, critical for producing diagnostic tools which indeed base classification on underlying biological factors encoded within the mass spectra.
List of abbreviations
In this section we define the various measures used. Respectively, TP, TN, FP, FN, stand for the number of true positive, true negative, false positive, false negative samples at classification time.
Sensitivity is defined as and is also known as Recall.
Specificity is defined as .
PPV (Positive Predictive Value) is defined as and is also known as Precision.
NPV (Negative Predictive Value) is defined as .
BACC (Balanced Accuracy) is defined as This measure defines the average of sensitivity and specificity.
% correct is defined as and measures the overall percentage of samples correctly classified.
Acknowledgements
Deepest thanks to anonymous reviewers and Dr. Vadim Bulitko who provided comments on the initial draft. Ovarian and prostate cancer data sets: OC-H4, OC-WCX2a, OC-WCX2b, and PC-H4 were provided by the National Cancer Institute, Clinical Proteomics Program Databank [13]. The PC-IMAC-Cu, prostate cancer set, was provided by the authors of [1]. We are grateful to the authors of [17] for making the PCA/LDA code publicly available. Funding for this research was provided by University of Alberta, NSERC and AICML.
Figures and Tables
Figure 1 Performance of Feature Extraction Algorithms on five cancer data sets. Both graphs show balanced accuracy (BACC) score. Top: Results grouped by data set. Bottom: Results grouped by feature extraction algorithm.
Table 1 Detailed performance statistics for ovarian cancer data sets Bold columns represent the mean of the respective performance measure, while columns labeled as (std) correspond to the standard deviation across the three cross-validation folds.
OC-H4
Corr Corr(std) BACC 3ACC(std Spec Spec(std) Sens Sens(std) PPV PPV(std)
No FE 0.763 0.05 0.763 0.05 0.848 0.16 0.677 0.11 0.841 0.13
PCA 0.712 0.07 0.712 0.07 0.727 0.25 0.697 0.12 0.768 0.20
PCA/LDA 0.727 0.07 0.727 0.07 0.636 0.31 0.818 0.18 0.744 0.19
SFS 0.747 0.22 0.747 0.22 0.980 0.02 0.515 0.42 0.931 0.06
SBS 0.823 0.08 0.823 0.08 0.899 0.13 0.747 0.08 0.891 0.12
P-test 0.763 0.20 0.763 0.20 0.929 0.05 0.596 0.38 0.863 0.09
T-test 0.747 0.19 0.747 0.19 0.929 0.02 0.566 0.38 0.856 0.08
KS-test 0.702 0.22 0.702 0.22 0.909 0.09 0.495 0.35 0.766 0.28
NSC(20) 0.621 0.19 0.621 0.19 0.949 0.06 0.293 0.32 0.743 0.29
Boosted 0.884 0.06 0.884 0.06 0.990 0.02 0.778 0.11 0.986 0.03
Boosted FE 0.854 0.13 0.854 0.13 1.000 0.00 0.707 0.26 1.000 0.00
OC-WCX2a
Corr Corr(std) BACC 3ACC(std Spec Spec(std) Sens Sens(std) PPV PPV(std)
No FE 0.773 0.09 0.773 0.09 0.828 0.02 0.717 0.18 0.800 0.05
PCA 0.682 0.18 0.682 0.18 0.687 0.14 0.677 0.25 0.671 0.18
PCA/LDA 0.899 0.02 0.899 0.02 0.889 0.10 0.909 0.06 0.900 0.09
SFS 0.949 0.03 0.949 0.03 0.980 0.03 0.919 0.05 0.979 0.04
SBS 0.854 0.15 0.854 0.15 0.929 0.08 0.778 0.23 0.903 0.12
P-test 0.944 0.03 0.944 0.03 0.970 0.03 0.919 0.06 0.969 0.03
T-test 0.965 0.02 0.965 0.02 0.949 0.05 0.980 0.02 0.953 0.04
KS-test 0.929 0.02 0.929 0.02 0.970 0.03 0.889 0.05 0.968 0.03
NSC(20) 0.944 0.04 0.944 0.04 0.990 0.02 0.899 0.08 0.989 0.02
Boosted 0.914 0.06 0.914 0.06 1.000 0.00 0.828 0.12 1.000 0.00
Boosted FE 0.965 0.01 0.965 0.01 1.000 0.00 0.929 0.02 1.000 0.00
OC-WCX2b
Corr Corr(std) BACC 3ACC(std Spec Spec(std) Sens Sens(std) PPV PPV(std)
No FE 0.837 0.14 0.834 0.12 0.822 0.07 0.846 0.20 0.891 0.05
PCA 0.901 0.05 0.893 0.03 0.867 0.03 0.920 0.07 0.926 0.02
PCA/LDA 1.000 0.00 1.000 0.00 1.000 0.00 1.000 0.00 1.000 0.00
SFS 0.992 0.01 0.991 0.01 0.989 0.02 0.994 0.01 0.994 0.01
SBS 0.901 0.14 0.903 0.13 0.911 0.10 0.895 0.17 0.942 0.07
P-test 0.980 0.02 0.975 0.03 0.956 0.05 0.994 0.01 0.976 0.03
T-test 0.837 0.07 0.834 0.04 0.822 0.05 0.846 0.13 0.897 0.01
KS-test 0.984 0.02 0.983 0.02 0.978 0.04 0.988 0.01 0.988 0.02
NSC(20) 0.972 0.02 0.973 0.03 0.978 0.04 0.969 0.03 0.988 0.02
Boosted 0.980 0.01 0.982 0.00 0.989 0.02 0.975 0.02 0.994 0.01
Boosted FE 1.000 0.00 1.000 0.00 1.000 0.00 1.000 0.00 1.000 0.00
Table 2 Detailed performance statistics for prostate cancer data sets Bold columns represent the mean of the respective performance measure, while columns labeled as (std) correspond to the standard deviation across the three cross-validation folds.
PC-H4
Corr Corr(std) BACC 3ACC(std Spec Spec(std) Sens Sens(std) PPV PPV(std)
No FE 0.732 0.05 0.777 0.06 0.698 0.05 0.855 0.09 0.439 0.06
PCA 0.530 0.20 0.516 0.18 0.540 0.24 0.493 0.21 0.248 0.11
PCA/LDA 0.692 0.15 0.667 0.14 0.710 0.22 0.623 0.33 0.431 0.17
SFS 0.885 0.05 0.827 0.17 0.929 0.03 0.725 0.36 0.728 0.03
SBS 0.773 0.03 0.729 0.09 0.806 0.11 0.652 0.27 0.498 0.07
P-test 0.813 0.02 0.728 0.11 0.877 0.08 0.580 0.28 0.572 0.07
T-test 0.816 0.04 0.709 0.14 0.897 0.05 0.522 0.31 0.575 0.07
KS-test 0.826 0.04 0.784 0.14 0.857 0.08 0.710 0.35 0.579 0.05
NSC(20) 0.791 0.04 0.736 0.10 0.833 0.12 0.638 0.31 0.529 0.07
Boosted 0.850 0.06 0.810 0.11 0.881 0.04 0.739 0.22 0.627 0.10
Boosted FE 0.960 0.01 0.906 0.03 1.000 0.00 0.812 0.07 1.000 0.00
PC-IMAC-Cu
Corr Corr(std) BACC 3ACC(std Spec Spec(std) Sens Sens(std) PPV PPV(std)
No FE 0.709 0.13 0.711 0.13 0.767 0.14 0.655 0.12 0.750 0.15
PCA 0.618 0.07 0.619 0.07 0.654 0.21 0.583 0.20 0.652 0.08
PCA/LDA 0.746 0.03 0.748 0.03 0.818 0.07 0.679 0.04 0.800 0.06
SFS 0.795 0.03 0.798 0.03 0.912 0.14 0.685 0.07 0.914 0.13
SBS 0.758 0.15 0.760 0.15 0.818 0.15 0.702 0.15 0.802 0.17
P-test 0.771 0.06 0.773 0.06 0.843 0.15 0.702 0.05 0.840 0.12
T-test 0.765 0.06 0.766 0.06 0.805 0.10 0.726 0.09 0.803 0.08
KS-test 0.789 0.03 0.791 0.03 0.862 0.09 0.720 0.05 0.854 0.07
NSC(20) 0.761 0.09 0.764 0.09 0.868 0.12 0.661 0.17 0.849 0.10
Boosted 0.823 0.05 0.826 0.05 0.950 0.09 0.702 0.11 0.949 0.09
Boosted FE 0.908 0.02 0.911 0.02 1.000 0.00 0.821 0.03 1.000 0.00
Table 3 Overall performance comparison Performance of each feature extraction algorithms averaged across data sets. Balanced accuracy (BACC) reported in increasing order.
Average BACC (+/-) OC-H4 OC-WCX2a OC-WCX2b PC-H4 PC-IMAC-Cu
PCA 0.684 0.139 0.712 0.682 0.893 0.516 0.619
No FE 0.771 0.044 0.763 0.773 0.834 0.777 0.711
T-test 0.804 0.100 0.747 0.965 0.834 0.709 0.766
NSC(20) 0.808 0.148 0.621 0.944 0.973 0.736 0.764
PCA/LDA 0.808 0.137 0.727 0.899 1.000 0.667 0.748
SBS 0.814 0.070 0.823 0.854 0.903 0.729 0.760
P-test 0.837 0.114 0.763 0.944 0.975 0.728 0.773
KS-test 0.838 0.115 0.702 0.929 0.983 0.784 0.791
SFS 0.863 0.103 0.747 0.949 0.991 0.827 0.798
Boosted 0.883 0.070 0.884 0.914 0.982 0.810 0.826
Boosted FE 0.927 0.057 0.854 0.965 1.000 0.906 0.911
Table 4 Feature set size comparison
OC-H4 (+/-) OC-WCX2a (+/-) OC-WCX2b (+/-) PC-H4 (+/-) PC-IMAC (+/-)
SFS 3 1.15 3 1.15 3 1.15 4 1.00 5 1.53
SBS 139 195.19 94 120.75 454 769.62 205 162.69 136 144.24
P-test 2 1.53 6 3.06 41 66.97 3 2.08 1 0.58
T-test 5 2.08 3 0.58 2 0.58 2 0.58 3 0.58
KS-test 2 1.00 7 4.36 63 106.52 2 1.53 2 0.58
Boosted FE 7 3.51 3 0.58 3 0.58 8 1.15 10 5.03
Table 5 Computational cost comparison Results presented in CPU seconds and in increasing order. All experiments were conducted using Matlab code on a dual CPU Athlon 1400+ running Linux.
Ave. CPU Time (+/-) OC-H4 OC-WCX2a OC-WCX2b PC-H4 PC-IMAC-Cu
No FE 1.19 0.40 0.84 0.87 1.11 1.31 1.82
P-test 2.42 1.11 1.41 1.58 4.12 2.08 2.90
PCA 19.39 7.70 12.69 12.08 17.41 25.57 29.21
PCA/LDA 20.58 8.08 13.53 12.95 18.52 26.88 31.03
KS-test 27.36 3.07 25.56 24.55 29.96 25.40 31.34
Boosted 543.42 332.13 371.73 134.56 507.62 688.22 1014.97
T-test 649.84 39.39 622.97 623.37 645.66 639.14 718.04
SFS 3164.70 1477.13 2178.24 2175.75 2516.42 3269.37 5683.70
BoostedFE 3356.89 2236.51 2679.97 1336.25 1841.57 3997.65 6928.99
NSC(20) 5434.20 3321.34 3717.30 1345.60 5076.20 6882.20 10149.70
SBS 23934.82 6655.33 17032.94 17244.80 29913.61 24574.69 30908.07
==== Refs
Adam B Qu Y Davis JW Ward MD Clements MA Cazares LH Semmes OJ Schellhammer PF Yasui Y Feng Z Wright GL Jr Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men Cancer Research 2002 62 3609 3614 12097261
Baggerly KA Morris JS Coombes KR Reproducibility of seldi-tof protein patterns in serum: Comparing data sets from different experiments Bio Informatics 2004 4
Keith BaggerlyA Jeffrey MorrisS Jing Wang David Gold Lian-Chun Xiao Kevin CoombesR A comprehensive approach to the analysis of matrix-assisted laser desorption/ionization-time of flight proteomics spectra from serum samples Proteomics 2003 3 1667 1672 12973722 10.1002/pmic.200300522
Conrads TP Zhou M Petricoin EF IIILiotta L Veenstra TD Cancer diagnosis using proteomic patterns Expert Reviews in Molecular Diagnostics 2003 3 411 420 10.1586/14737159.3.4.411
Cotter RJ Time-of-Flight Mass Spectrometry 1994 American Chemical Society, Washington, DC
Diamandis E Proteomic patterns in biological fluinds: Do they represent the future of cancer diagnostics Clinical Chemistry (Point/CounterPoint) 2003 48 1272 1278 10.1373/49.8.1272
Thomas G Dietterich. Approximate statistical test for comparing supervised classification learning algorithms Neural Computation 1998 10 1895 1923 9744903 10.1162/089976698300017197
Duda R Hart P Pattern Classification and Scene Analysis 1973 John Wiley & Sons, New York
Freund Y Schapire R A decision-theoretical generalization of on-line learning and an application to boosting Computer System Science 1997 55 119 139 10.1006/jcss.1997.1504
Hastie T Tibshirani R Friedman J The Elements of Statistical Learning Springer Series in Statistics 2001 Springer Verlag, New York
Inza I Larranaga P Blanco R Cerrolaza AJ Filter versus wrapper gene selection approaches in dna microarray domains Artificial Intelligence in Medicine 2004 31 91 103 15219288 10.1016/j.artmed.2004.01.007
Jeffries NO Performance of a genetic algorithm for mass spectrometry proteomics BMC Bioinformatics 2004 5
Johann D Clinical proteomics program databank. Technical report, National Cancer Institute, Center for Cancer Research, NCI-FDA Clinical Proteomics Program 2003
Michael Kirby Geometric Data Analysis: An Empirical Approach to Dimensionality Reduction and the Study of Patterns 2001 John Wiley & Sons, New York
Levner I Proteomic pattern recognition Technical report, University of Alberta, No: TR04-10 2004
Levner I Bulitko V Lin G Isabelle Guyon, Steve Gunn, Masoud Nikravesh, Lofti Zadeh Feature extraction for classification of proteomic mass spectra: A comparative study Feature Extraction, Foundations and Applications 2005 Springer
Lilien RH Farid H Donald BR Probabilistic disease classification of expression-dependent proteomic data from mass spectrometry of human serum Computational Biology 2003 10
Park H Jeon M Rosen JB Lower dimensional representation of text data based on centroids and least squares BIT 2003 43 1 22 10.1023/A:1026039313770
Petricoin EF Ardekani AM Hitt BA Levine PJ Fusaro VA Steinberg SM Mills GB Simone C Fishman DA Kohn EC Liotta LA Use of proteomic patterns in serum to identify ovarian cancer The Lancet 2002 359 572 577 11867112 10.1016/S0140-6736(02)07746-2
Petricoin EF Ornstein DK Paweletz CP Ardekani A Hackett PS Hitt BA Velassco A Trucco C Wiegand L Wood K Simone C Levine PJ Linehan WM Emmert-Buck MR Steinberg SM Kohn EC Liotta LA Serum preteomic patterns for detection of prostate cancer Journal of the National Cancer Institute 2002 94 1576 1578 12381711
Press WH Teukolsky SA Vetterling WT Flannery BP Numerical Recipes in C: The Art of Scientifi Computing 2002 Second Cambridge University Press
Qu Y Adam B Yasui Y Ward MD Cazares LH Schellhammer PF Feng Z Semmes OJ Wright GL Jr Boosted decision tree analysis of surface-enhanced laser desorption/ionization mass spectral serum profiles discriminates prostate cancer from noncancer patients Clinical Chemistry 2002 48 1835 1843 12324514
Robert SchapireE A brief introduction to boosting IJCAI 1999 1401 1406
Sorace J Zhan M A data review and re-assessment of ovarian cancer serum proteomic profiling BioInformatics 2003 4
Tibshirani R Hastie T Narasimhan B Chu G Class prediction by nearest shrunken centroids, with applications to dna microarrays Statistical Science 2003 18 104 117 10.1214/ss/1056397488
Tibshirani R Hastiey T Narasimhanz B Soltys S Shi G Koong A Le Q Sample classifcation from protein mass spectrometry by 'peak probability contrasts' BioInformatics 2004
Paul Viola Michael Jones Robust real-time object detection International Journal of Computer Vision 2003
Keung CK Lam W Ling CX Learning good prototypes for classification using filtering and abstraction of instances Pattern Recognition 2002 35 1491 1506 10.1016/S0031-3203(01)00131-5
Michael Wagner Naik DN Kasukurti S Pothen A Devineni RR Adam BL Semmes OJ Wright GL Jr Computational protein biomarker prediction: a case study for prostate cancer BMC Bioinformatics 2004 5
Wu B Abbott T Fishman D McMurray W Mor G Stone K Ward D Williams K Zhao H Comparison of statistical methods for classifcation of ovarian cancer using mass spectrometry data BioInformatics 2003 19
Wulfkuhle JD Liotta LA Petricoin EF Proteomic applications for the early detection of cancer Nature Reviews 2003 3 267 275 12671665 10.1038/nrc1043
|
15788095
|
PMC1274262
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 23; 6:68
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-68
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-691578810610.1186/1471-2105-6-69SoftwareWildfire: distributed, Grid-enabled workflow construction and execution Tang Francis [email protected] Ching Lian [email protected] Liang-Yoong [email protected] Yun Ping [email protected] Praveen [email protected] Arun [email protected] Information Science Research, Bioinformatics Institute, 30 Biopolis Street, #07-01, Matrix, 138671 Singapore2 Operations Research Lab, DSO National Laboratories, 20 Science Park Drive, 118230 Singapore3 Singapore Biomedical Computing Resource, Bioinformatics Institute, 30 Biopolis Street, #07-01, Matrix, 138671 Singapore4 Global Software Group, Motorola Electronics Pte Ltd, 12 Ang Mo Kio St. 64, Ang Mo Kio Industrial Park 3, 569088 Singapore2005 24 3 2005 6 69 69 22 11 2004 24 3 2005 Copyright © 2005 Tang et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
We observe two trends in bioinformatics: (i) analyses are increasing in complexity, often requiring several applications to be run as a workflow; and (ii) multiple CPU clusters and Grids are available to more scientists. The traditional solution to the problem of running workflows across multiple CPUs required programming, often in a scripting language such as perl. Programming places such solutions beyond the reach of many bioinformatics consumers.
Results
We present Wildfire, a graphical user interface for constructing and running workflows. Wildfire borrows user interface features from Jemboss and adds a drag-and-drop interface allowing the user to compose EMBOSS (and other) programs into workflows. For execution, Wildfire uses GEL, the underlying workflow execution engine, which can exploit available parallelism on multiple CPU machines including Beowulf-class clusters and Grids.
Conclusion
Wildfire simplifies the tasks of constructing and executing bioinformatics workflows.
==== Body
Background
Seemingly small steps in usability of bioinformatics applications have, perhaps, been the most important to the bioinformatics consumer. Suites such as Accelrys SeqWeb and EMBOSS/Jemboss [1,2], through consistent user interface elements, have narrowed the usability gap and made individual applications accessible to the non-specialist bioinformatician.
Bioinformatics analysis is becoming more complex, often requiring several applications to be run in combination in a workflow. Beowulf-class clusters have made supercomputing affordable, allowing us to execute workflows faster, including some which would previously have been infeasible. Traditionally, building such workflows required programming expertise, often in scripting languages such as perl. The usability gap in bioinformatics has now moved from individual applications to both construction and execution of workflows.
Existing efforts in narrowing this gap include Jemboss [2], Taverna/Freefluo [3], ICENI [4] and Biopipe [5]. Jemboss, though it does not support workflows directly, addresses usability of bioinformatics applications by providing a graphical user interface to EMBOSS. The user interface replaces the command-line options of the EMBOSS applications with interface elements such as check boxes, drop-down lists and text boxes. This simplifies the applications for users unfamiliar with command-line interfaces; even for command-line enthusiasts, it simplifies learning of new applications, which might only be used occasionally, since the interface is consistent across applications. Jemboss can run the EMBOSS application on the same machine ("Stand-alone mode") or remotely using a SOAP protocol.
Taverna, by default, constructs workflows which use Web-Services for components. The interface requires the user to connect together output and input ports of components to build a workflow. Taverna relies on Soaplab [6] to convert the EMBOSS command-line applications into Web-Services. However, Soaplab appears to have lost the help text annotations of the different input fields which is characteristic of other EMBOSS interfaces. ICENI is also a service-oriented workflow framework and has a Netbeans-based user interface.
Biopipe is a workflow framework which also allows for execution of workflows across clusters. However, Biopipe only allows for pipelines, not more general workflows with iterative loops; in particular, the particle swarm optimisation example presented later cannot be implemented in Biopipe. Also, Biopipe currently does not have a user-friendly interface for building pipelines.
In the next section, we introduce Wildfire which provides an integrated environment for construction and execution of workflows. It provides an intuitive interface based on a drawing analogy and, like Jemboss, presents program options using graphical user interface elements; thus Wildfire hides the precise syntax of scripting languages and command-line options from the user. Jemboss can run several independent processes in the background, but it has no dependency handling facility, whereas Wildfire allows the user to compose applications into a workflow. We illustrate by presenting some examples in the Results section. In contrast to Taverna and ICENI, Wildfire works directly with program executables, rather than Web- or Grid-Services. For execution, it uses GEL (Grid Execution Language [7]) which can run the workflow over the compute nodes of a cluster, similar to Biopipe. However, GEL can also run executables directly, or on the Grid. Thus, Wildfire and GEL bring supercomputing power to the bioinformatician.
Implementation
Wildfire allows the user to visually construct workflows. For execution, Wildfire exports the workflow as a GEL script, and then calls a GEL interpretor to execute it. The GEL interpretor can either run on the same machine as Wildfire, or on a remote compute server. Figure 1 summarises the interaction between Wildfire and GEL.
Wildfire is implemented in Java, and has been tested on Windows and Linux platforms. On a Linux platform, the user can run workflows directly on the same machine: ideal for developing and testing small examples on a laptop, while reserving the multi-processor servers and clusters for running the workflow on real data.
We next describe the two main activities enabled by Wildfire: construction and execution of workflows.
Workflow construction
When constructing workflows, the user does not need to work directly with the syntax of scripting languages such as GEL or perl. Rather, the user is presented with a graphical workflow canvas. On the canvas, a workflow component can be (i) an atomic component, (ii) a subworkflow or (iii) a loop (both parallel and sequential). An atomic component approximately corresponds to an EMBOSS application; in particular, each atomic component has an ACD (Ajax Command Definition [8]) description of its parameters and options. The user can select the atomic components from a customisable list of templates, which by default includes all the EMBOSS 2.8.0 applications (see Availability and requirements section). Components are visually rendered on the canvas as yellow rectangles labelled with the component name (e.g. EMBOSS program name), and a unique numerical identifier which can be used distinguish instances from the same component template. Sequential dependencies between components are created by drawing an arrow between them. By default, components not linked by arrows are assumed to be independent (and so can be executed in parallel).
Double clicking on an atomic component in the workflow will bring up a properties window resembling that of Jemboss (see Fig. 2). Wildfire uses Jemboss code to parse the ACD [8] description of the application to construct the form and provides default values where defined. These forms simplify configuration options by replacing the command-line flags and switches with graphical user interface elements such as drop-down menus. Help text annotations for the input fields save the user the effort of looking up UNIX man- or EMBOSS tfm-pages.
Wildfire extends the Jemboss interface by allowing the user to use expressions (similar to spreadsheet formulae) in the text fields. For example, in Fig. 2, the query file for blastall is = $flie. The first letter is an equals symbol (=) and indicates that this is not a literal string, but an expression. The remainder is the expression meaning "the value of variable $flie". Here the value of $flie is determined by the pforeach container, as shown in the background window, and denotes a parallel composition of blastall instances with $file set to the different files matching *_dice*.fna. The output file is
= $file . ".out"
which is an expression meaning "the value of variable $flie with .out appended". Another example of an expression is
= $f % ".fasta"
which means "the value of variable $f without the .fasta extension". The % and . operators can be mixed, for example
= $f % ".fasta" . ".pep"
which replaces the .fasta extension with .pep.
In addition, the user can add his/her own command-line programs to the list of atomic components by providing a description of its command-line options using an extended ACD syntax. The Wildfire user interface has a facility to help the user write ACD files for new atomic components. The interface shields the user from the complex ACD syntax.
Other than defining the dependencies between components and the invocation arguments, the user can place input files required by the workflow in subdirectories within the workflow directory. Wildfire can instruct GEL to copy files from these subdirectories into the working directory before a component is executed. Any instance can specify input files, thus allowing for files to be staged-in in a just-in-time manner. However, a common workflow pattern is one which specifies all input files to copied only by the first instance in the workflow.
Workflow execution
For execution, Wildfire exports a programmatic description of the workflow, in a scripting language called GEL [7], which is passed to a GEL interpretor for execution. GEL is a scripting language with parallel constructs characterising common parallel workflow execution patterns. It is designed to be a generic parallel scripting language which can be executed on different types of homogeneous and heterogeneous parallel hardware such as shared-memory SMP servers, clusters with a shared disk image, and Grids without a shared disk image. There currently exist interpretors that can run GEL scripts on SMP servers, clusters with Platform LSF, PBS or Sun GridEngine (SGE), and on Condor Grids [9,10]. GEL is similar to APST [11], NIMROD [12] and DAGMan (part of Condor) but also allows for cyclic dependencies between jobs. The reader is referred to [7] for a more thorough description of GEL.
When developing small workflows, the user can run the workflow on the same machine (see Availability and requirements section). In this way, Wildfire can be used as a stand-alone application without access to the network.
Alternatively, the user can choose to send the workflow to a remote server and run it there. In this case, Wildfire uses the secure shell (SSH) protocol to send the necessary files over to, and then run the GEL interpretor on the remote server (see Fig. 3). The GEL interpretor can execute the atomic components directly if the server has multiple processors. If the server is a cluster, then GEL can submit the atomic components as jobs to the queue manager. In either case, the GEL interpretor will try to use multiple processors where possible. Remote server execution is useful for workflows with large data sets since GEL will make use of multiple processors. It is also useful if the atomic components are not installed on the local machine.
Wildfire and GEL do not require super-user privileges to install: they can be installed in the "home" directory. For the client-server mode of operation, only an SSH service on the server is required; there is no need to configure other services such as SOAP over HTTP or Web-/Grid-Services, and the firewall is only required to allow incoming SSH connections. Most modern UNIX-style configurations already provide an SSH service.
Wildfire can also use GEL to break up the workflow and run parts of it concurrently on different supercomputers using Condor. (Note: GEL 1.0 uses the Globus [13] protocols to provide Grid support. GEL 2.0 uses Condor for Grid execution and future support for Globus Grids will be via Condor-G.) This is useful for very large workflows which require as many compute resources as possible. In practice, it is more useful when not all components are available on any one machine, for example, because of licence availability.
Wildfire monitors execution of the individual atomic components and feedback is provided via annotations on the canvas which are updated in real-time.
The exported GEL script can also be run directly using an interpretor via the command line. This allows a workflow to be run in batch mode independently of Wildfire, and is useful for very long-running workflows or those that have to be run repeatedly.
Results
We describe three applications of Wildfire. In the first application, we construct a workflow for analysis of human tissue-specific transcripts by comparing them against the known exons. This example shows how Wildfire can make use of the parallel capabilities of supercomputers. The second example considers a particle swarm optimisation algorithm implemented as a workflow, and shows that Wildfire can express workflows requiring iteration. The last example cross validates an allergenicity prediction algorithm. The number of parallel processes in this workflow can only be determined at run-time.
Tissue-specific Gene Expression Analysis
To study tissue-specific gene expression in humans, we compare the known exons against a database of 16,385 transcripts obtained from the Mammalian Gene Collection. Since the human genome contains many exons, the extraction process is time consuming, but it is easily parallelised. The standard organisation of the 24 chromosomes into separate files provides a natural partitioning of the exon extraction problem: we extract the exons from each chromosome in parallel. To further increase the granularity of the problem and so exploit more parallelism, we break up each of the 24 files of exons into five smaller files, resulting in a total of 120 files. We blast each of these smaller files against the database of transcripts.
The workflow as constructed in Wildfire is shown in Fig. 4, and its implementation without Wildfire is described in detail in [14]. The atomic component exonx is a program developed in-house to extract and store exons from a genbank file in fasta format; dice is a perl script used to break up a fasta file into smaller pieces. The noop.sh component is required to instruct GEL to copy the input files into the working directory. (The initial copying of input files will likely be an implicit feature of workflows in future versions of Wildfire, and so the explicit noop. sh component will no longer be necessary.) The remaining components (GNU gunzip, NCBI BLAST formatdb and blastall) are standard applications which we have incorporated as atomic components using our ACD editor.
The whole workflow takes less than 6000 seconds to run on a 128 CPU Pentium III cluster, whereas a sequential version of the same workflow required almost nine times longer. Profiling of the workflow shows that breaking up the 24 files of exons more evenly would significantly improve performance. Since all programs are run via the scheduler on the cluster, the workflow follows whatever scheduling policies are configured at the component-level. Hence, the workflow in its current form is already efficient from the point of view of resource use.
Swarm optimisation
Real-life optimisation problems are often intractable and heuristics are the only choice for finding near optimal solutions. Particle Swarm Optimisation [15] is such a heuristic based on simulation of information exchange between leaders and followers observed in, for example, bird flocking.
The algorithm simulates individuals flying through the search space. On each iteration, the individuals are separated into a set of leaders and a set of followers, based on their fitness. The followers use the locations of the leaders to change their flying direction, i.e. search velocity. The location of each individual is computed based on its current location and flying direction. The new location is used to rank the fitness of individuals and subsequently the leader and follower sets.
Note that efficient Swarm Optimisation implementations exhibit both (1) iteration and (2) parallelism: successive generations must be simulated until a termination condition is met, and simulation of each generation entails simulation of many independent individuals. Therefore, a workflow tool suitable for implementing such algorithms must support iteration and parallelism.
The workflow in Fig. 5 is a simplified implementation of a swarm algorithm by Ray et al. [16] implemented as a workflow. The algorithm is applied to a parameter estimation problem for a biochemical pathway model consisting of 36 unknowns and eight ordinary differential equations [17]. Components init1, eval1 and init2 are used to initialise and rank the individuals. Component test determines whether the workflow should terminate and extract collects together the results on termination of the simulation. Component eva12 is used to evaluate the fitness of an individual; note the outer parallel loop evaluates the fitness of each follower. The remaining components are used to select the leaders and followers.
Component test depends on both init2 and reassign; on the first iteration, test can start executing only after init2 has terminated, and on subsequent iterations, only after reassign has terminated. Since reassign itself hereditarily depends on test (i.e. reassign depends on eval2 which depends on classify which itself depends on test) we see there is a cyclic dependency. The while loop in GEL allows such dependencies and so is crucial for this workflow.
Allergenicity prediction
Allergenicity prediction is the process of determining whether a new protein sequence is an allergen or not. Proteins known to induce allergic responses have been documented in allergen databases. One approach to allergenicity prediction is to determine, automatically, motifs from sequences in such a database, and then search for these motifs in the query sequences.
The objective of the workflow in Fig. 6 is to test the accuracy of the approach described above, where protein sequence motifs are identified using an algorithm [18] based on wavelet analysis. From a group of 817 sequences, known to be allergens, we take a learning set consisting of a randomly selected subset covering 90% to be used for identification of motifs. The remaining 10% are used as query sequences for allegenicity prediction. We use the predictions to assess the accuracy of this approach.
Initially, we use ClustalW to generate the pairwise global alignment distances between the protein sequences in the learning set. We then use these distances to cluster the protein sequences by partitioning around medoids using the R project [19]. We use ClustalW again to align each cluster of protein sequences and we use the wavelet analysis algorithm on each aligned cluster to identify motifs in the protein sequences. For each identified motif, we build an HMM profile [20,21] in parallel. Note that the number of motifs is not know a priori; it can only be known at run-time. The HMM profiles are then used to search for the motifs in each query sequence, and thus predict whether it is an allergen or not. The accuracy of the predictions is computed to assess the effectiveness of this approach.
Discussion
Bioinformaticians can use Wildfire as an integrated environment to construct and to execute workflows. The graphical user interface elements ease workflow construction by hiding the syntax of scripting languages. The constructed workflows can be executed across multiple processors (i) in the same server, (ii) in a cluster, or (iii) across several supercomputers across the Grid.
Wildfire is preconfigured to allow applications from EMBOSS to be used as workflow components. The user can add his own applications to be used as components by creating the necessary ACD files. Wildfire also hides this syntax by providing a wizard for creating and editing ACD files.
In a typical scenario, the user develops a workflow by visually constructing and executing small examples on his desktop or laptop. When the workflow is ready, the user can run it on real data on a shared resource such as a compute cluster running a scheduler. GEL workflows, such as those constructed in Wildfire, run well on shared resources since the components of the workflow are run through the scheduler. This gives the scheduler more opportunities to schedule the components with respect to whatever policies are configured: for example, a fair share policy would allow jobs from other users to run even when the workflow would otherwise monopolise the whole cluster.
We have described three applications of Wildfire in three different fields, and welcome readers to try Wildfire for themselves and solicit recommendations for improvements.
Availability and requirements
Wildfire
Wildfire is run on the client computer and allows the user to visually construct workflows. Wildfire invokes GEL (see below) to execute workflows. EMBOSS 2.8.0 [22] is required (2.9.0 is not yet supported) to run EMBOSS workflows.
Availability:
Operating systems: Platform Independent (tested on Windows, i386 Linux)
Programming Language: Java
Other requirements: Java 1.4.2, GEL
Licence: GPL
Currently, stand-alone mode is not available for Windows.
GEL
GEL is an interpretor which executes GEL scripts generated by Wildfire. Currently only UNIX platforms are supported. GEL supports LSF, SGE and PBS clusters, SMP servers and Condor-based Grids.
Availability:
Operating systems: GNU-style UNIX (tested on i386 Linux, ia64 Linux, Spare Solaris, Alpha Tru64)
Programming Language: Java
Other requirements: Java 1.4.2, bash
Licence: free for non commercial use, see
Authors' contributions
FT lead and coordinated the software engineering aspects of the project, and drafted this manuscript. FT and CCL co-designed Wildfire and GEL. CCL and HLY programmed, tested and debugged the software. LYP designed the tissue-specific transcript analysis workflow. PI and AK designed the allergen prediction workflow. AK participated in the concept, design and testing of the software and contributed to successive revisions of this manuscript. All authors read and approved the manuscript.
Acknowledgements
The authors would like to thank Tan Chee Meng and Prof. Tapabrata Ray for their contribution to the Particle Swarm Example, and various anonymous referees for improvements in this manuscript.
Figures and Tables
Figure 1 Relationship between Wildfire and GEL. Relationship between Wildfire and GEL. Wildfire is an interactive application which allows users to construct workflows using a drawing analogy. Wildfire executes the workflow by exporting it as a GEL script which is executed using a suitable GEL interpretor. There are GEL interpretors for execution on (i) the Grid, using Condor, (ii) a cluster, using LSF, PBS or SGE, and (iii) the same machine, which could be a laptop, desktop or multi-processor server.
Figure 2 Elements of the Wildfire interface. Elements of the Wildfire interface. The main window in the background shows the workflow canvas. Its left panel lists the atomic workflow components; this list is preconfigured with all EMBOSS applications and can be customised. The foreground window shows the properties form for the blastall program.
Figure 3 Remote execution of workflows. Remote execution of workflows. In the case of large workflows or when the applications are not available on the client machine, it is possible to execute the workflow remotely. In this case, Wildfire uses the secure shell (SSH) protocol to send the files to the remote machine and start execution. The remote machine can be a cluster or multiprocessor server.
Figure 4 Tissue-specific gene expression analysis. Tissue-specific gene expression analysis. We initially start with 24 compressed genbank files, one for each human chromosome. We decompress (gunzip) the genbank file, extract (exonx) from it all the exons into one fasta file and then break up (dice) this fasta file into 5 smaller fasta files; we do this for each chromosome in parallel. At the same time, we decompress the file of transcripts and use formatdb to format it for use as a BLAST database. Finally, we blast the exons against the transcript database in parallel. Note that components are rendered as rectangles and the bottom half shows a unique numerical identifier; this can be used to distinguish components derived from the same template, e.g. gunzip:2 and gunzip:6.
Figure 5 Swarm optimisation example. Swarm optimisation example. Component eval1 is executed four times in parallel within a parallel for loop. The circle denotes a while loop, with test as loop guard: if test returns false, then we follow the bottom branch to extract, otherwise we follow the right branch and test test again after reassign. Component eval2 is executed in parallel once for each file matching follower_sol*.
Figure 6 An allergenicity prediction workflow. The allergenicity prediction workflow from [23] constructed in Wildfire. The components in this workflow are all custom applications, or custom scripts calling standard applications. Components format, group, join_scr and process_scr are administrative programs which translate and convert files from one format to another. Component rscript uses the R-project to cluster the amino acid sequences from a database of known allergens. For each cluster, we align its sequences and use a wavelet algorithm to predict motifs. The resulting motifs are used to construct HMM profiles using hmmbuild. Finally, we use these profiles with hmmpfam to predict allergenic sequences. Components join_scr and process_scr collate and summarise the results.
==== Refs
Rice P Longden I Bleasby A EMBOSS: The European Molecular Biology Open Software Suite Trends in Genetics 2000 16 276 277 10827456
Carver T Bleasby A The design of Jemboss: a graphical user interface to EMBOSS Bioinformatics 2003 19 1837 1843 14512356
Oinn T Addis M Ferris J Marvin D Greenwood M Carver T Pocock MR Wipat A Li P Taverna: A tool for the composition and enactment of bioinformatics workflows Bioinformatics 2004 20 3045 3054 15201187
Furmento N Lee W Mayer A Newhouse S Darlington J ICENI: An Open Grid Service Architecture Implemented with Jini SuperComputing 2002 2002
Hoon S Ratnapu KK Chia JM Kumarasamy B Juguang X Clamp M Stabenau A Potter S Clarke L Stupka E Biopipe: a flexible framework for protocol-based bioinformatics analysis Genome Res 2003 13 1904 1915 12869579
Senger M Rice P Oinn T Cox SJ Soaplab – a unified Sesame door to analysis tools Proceedings, UK e-Science, All Hands Meeting 2003 509 513
Chua Ching Lian Tang F Issac P Krishnan A GEL: Grid Execution Language J Parallel and Distributed Computing
de Boer T AJAX Command Definition (ACD files)
Litzkow MJ Livny M Mutka MW Condor: A Hunter of Idle Workstations Proceedings of 8th International Conference on Distributed Computing Systems 1988 104 111
Condor Team Condor Home Page 2002
Berman FD Wolski R Figueira S Schopf J Shao G Application-Level Scheduling on Distributed Heterogeneous Networks Proceedings of Supercomputing 1996 1996
Abramson D Sosic R Giddy J Hall B Nimrod: A Tool for Performing Parameterised Simulations Using Distributed Workstations HPDC 1995 112 121
Foster I Kesselman C Globus: A Metacomputing Infrastructure Toolkit The International Journal of Supercomputer Applications and High Performance Computing 1997 11 115 128
Chua CL Tang F Lim YP Ho LY Krishnan A Implementing a Bioinformatics Workflow in a Parallel and Distributed Environment Parallel and Distributed Computing: Applications and Technologies, of LNCS, Springer 2005 3320 1 4
Eberhart RC Kennedy J A New Optmizer using Particle Swarm Theory Sixth International Symposium on Micro Machine and Human Science, IEEE Service Center 1995 39 43
Ray T Tai K Seow K An Evolutionary Algorithm for Constrained Optimization Proceedings of the Genetic and Evolutionary Computation Conference, Morgan Kaufmann 2000 771 777
Moles CG Mendes P Banga JR Parameter Estimation in Biochemical Pathways: A Comparison of Global Optimization Methods Genome Research 2003 13
Krishnan A Li KB Issac P Rapid detection of conserved regions in protein sequences using wavelets In Silico Biology 2004 4 133 48 15107019
R Language Definition
Durbin R Eddy S Krogh A Mitchison G Biological sequence analysis CUP 1998
Eddy S HMMER User's Guide
EMBOSS
Li KB Issac P Krishnan A Predicting allergenic proteins using wavelet transform Bioinformatics 2004 20 2572 8 15117757
|
15788106
|
PMC1274263
|
CC BY
|
2021-01-04 16:02:48
|
no
|
BMC Bioinformatics. 2005 Mar 24; 6:69
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-69
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-711579041810.1186/1471-2105-6-71Research ArticlePrediction of a common structural scaffold for proteasome lid, COP9-signalosome and eIF3 complexes Scheel Hartmut [email protected] Kay [email protected] Bioinformatics Group, Memorec Biotec GmbH, Stöckheimer Weg 1, D-50829 Köln, Germany2005 24 3 2005 6 71 71 22 11 2004 24 3 2005 Copyright © 2005 Scheel and Hofmann; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The 'lid' subcomplex of the 26S proteasome and the COP9 signalosome (CSN complex) share a common architecture consisting of six subunits harbouring a so-called PCI domain (proteasome, CSN, eIF3) at their C-terminus, plus two subunits containing MPN domains (Mpr1/Pad1 N-terminal). The translation initiation complex eIF3 also contains PCI- and MPN-domain proteins, but seems to deviate from the 6+2 stoichiometry. Initially, the PCI domain was defined as the region of detectable sequence similarity between the components mentioned above.
Results
During an exhaustive bioinformatical analysis of proteasome components, we detected multiple instances of tetratrico-peptide repeats (TPR) in the N-terminal region of most PCI proteins, suggesting that their homology is not restricted to the PCI domain. We also detected a previously unrecognized PCI domain in the eIF3 component eIF3k, a protein whose 3D-structure has been determined recently. By using profile-guided alignment techniques, we show that the structural elements found in eIF3k are most likely conserved in all PCI proteins, resulting in a structural model for the canonical PCI domain.
Conclusion
Our model predicts that the homology domain PCI is not a true domain in the structural sense but rather consists of two subdomains: a C-terminal 'winged helix' domain with a key role in PCI:PCI interaction, preceded by a helical repeat region. The TPR-like repeats detected in the N-terminal region of PCI proteins most likely form an uninterrupted extension of the repeats found within the PCI domain boundaries. This model allows an interpretation of several puzzling experimental results.
==== Body
Background
In eukaryotic organisms, there exist at least three distinct multi-protein assemblies that are jointly referred to as 'PCI complexes' [1] and have a similar subunit architecture despite their fundamentally different function: i) the proteasome lid, a subcomplex of the 19S proteasome regulator and the 26S proteasome, ii) the COP9 signalosome or CSN complex, and iii) the eukaryotic translation initiation factor eIF3. As a common feature, these complexes are composed of multiple subunits harbouring the PCI domain, named after the three participating complexes [2], sometimes also referred to as the PINT domain [3]. Other subunits of these complexes are characterized by a second shared homology domain called MPN (Mpr1-Pad1 N-terminal) [2,3].
Among these complexes, the proteasome lid and the CSN share a particular degree of analogy. Both complexes consist of eight core subunits, six of them of the PCI class and two of the MPN class. As described previously [1] and summarized in table 1, there is a clear 1:1 correspondence between the paralogous sets of PCI and MPN subunits. In addition, a similar ring-shaped structure was observed for the two complexes [4] and there is evidence that in those rings paralogous subunits occupy equivalent positions [5]. By contrast, the eIF3 complex has a smaller number of PCI subunits (table 1) and its two MPN subunits are absent in several unicellular eukaryotes. Unlike the proteasome lid and the CSN, the eIF3 complex contains a number of non-PCI/non-MPN subunits, which are required for its function in translation.
Despite the common homology domains and a similar structure, the functions of the three PCI complexes are very different. The proteasome lid, in combination with the 'base' complex containing a hexameric ring of AAA-ATPases, forms the 19S regulatory particle, which in turn constitutes an essential subcomplex of the 26S proteasome [6]. The lid complex contains an intrinsic deubiquitinating activity, which is encoded by the MPN subunit Rpn11 that has the hallmarks of a metalloprotease [7-9]. No specific function has been described for the PCI subunits of the lid. The CSN complex has been first described as a regulator of photomorphogenesis in plants, but seems to regulate diverse cellular processes like signal transduction, regulation of transcription or cell proliferation [10,11]. Csn5, an MPN-bearing subunit of the signalosome, which is analogous to Rpn11, also encodes a metalloprotease that is essential for the removal of the ubiquitin-like protein Nedd8 from cullins [12] The third PCI complex, the translation initiation factor eIF3, promotes the formation of preinitiation complexes and works as a scaffold by binding to other initiation factors, to ribosomes and to mRNA [13,14]. Both MPN subunits of eIF3 lack the residues necessary for metal binding [8,15] and are most likely catalytically inactive.
So far, the metal-containing MPN subunits and the non-PCI/non-MPN portion of the complex constitute the only known carriers of functionality. The PCI proteins themselves seem to be the main building blocks of the complexes, a fact already suggested by their high abundance. There are several hints that the PCI subunits are crucial for proper complex assembly [16-19]. The MPN subunits of the three complexes are rather well conserved and the detection of MPN domains and their boundaries is relatively straightforward. By contrast, the degree of conservation between PCI subunits is highly variable. Sequence similarity between the corresponding subunits of proteasome lid and CSN is generally easy to spot, while the detection of similarity between other paralogous PCI subunits typically requires sophisticated sequence comparison approaches, such as the generalized profile method [2,20]. A particular challenge is the detection of the highly divergent PCI domains in the budding yeast CSN-like complex [21] and those of the eIF3 complex, where only three PCI subunits could be detected in the initial survey [2]. Due to this difficulties, it is to be expected that there are still a number of highly divergent PCI domain proteins in eukaryotic genomes, which have eluded detection so far. A second issue in the bioinformatical definition of the PCI domain concerns the position of its N-terminal boundary. In general, homology domains are thought to correspond to structural domains in the sense of autonomous folding units; they are typically characterized by a pronounced loss of sequence similarity at the domain boundaries. While this is true for the PCI domain C-terminus, the N-terminal domain boundary is blurred through a gradual decay in sequence similarity instead of a sharp drop. As a consequence, different PCI domain boundaries have been used in the literature [2,3] and in various domain databases like PROSITE [22], Pfam [23] and SMART [24]. The corresponding accession numbers are PS50250, PF01399 and SM00088, respectively.
During an exhaustive bioinformatical analysis of proteasome subunits and other components of the ubiquitin/proteasome system, we obtained two independent results jointly suggesting that a structure-based redefinition of the PCI domain is appropriate: on one hand, we detected multiple instances of TPR-like repeats in the N-terminus of many PCI proteins, which suggests that the homology between the proteasome and CSN components is not restricted to the PCI domain itself. On the other hand, we detected a previously overlooked PCI domain in the novel eIF3 subunit eIF3k [25]. Most interestingly, an X-ray structure of eIF3k has been published recently [26]. Based on this structure and on our alignment data, we suggest a bipartite consensus model for the canonical PCI proteins, consisting of a C-terminal 'winged helix' domain preceded by an extended helical repeat region. We use this model to re-evaluate some bioinformatical and experimental findings that have been enigmatic so far.
Results
TPR-like helical repeats in PCI proteins
In most PCI proteins, the canonical PCI domain occupies a region of approximately 190 residues close to the carboxy-terminus of the sequence. The N-terminal non-PCI portion of the proteins is moderately conserved between species and only poorly conserved between different PCI subunits – even between the analogous subunits of the lid and the CSN. Upon submitting those PCI proteins to profile- or HMM-based domain detection services, no significant matches were obtained for the N-termini of the proteins. However, the PROSITE [27] profile for the tetratrico-peptide repeat (TPR) yielded a number of closely sub-significant matches in multiple PCI proteins, e.g. Rpn7 from S. bayanus (P value = 0.01, Ref [28]) and Csn1 from E. histolytica (P value = 0.06, Uniprot: Q8WQ58). The TPR repeat family [29] is very heterogeneous, and TPR motif descriptors such as the PROSITE profile are known to miss several instances of bona fide TPR repeats. Upon closer inspection, most PCI proteins exhibit multiple regions of similarity to profiles derived from established TPR repeats (matches schematically shown in figure 1), although the similarity scores for each of the single regions do not reach statistical significance. No relevant similarity scores were obtained for other helical repeat motifs, such as HEAT or Armadillo repeats.
To further investigate if a TPR-like structure should be assumed for the N-terminal portions of all PCI proteins, we performed a secondary structure prediction for each of the protein families individually. To that aim, we constructed multiple alignments for representative members of each subunit family and submitted the alignment to PHD and JPred prediction servers [30,31]. As a result, all PCI subunits of lid and CSN are predicted to adopt an all-helical secondary structure upstream of the PCI domain. Interestingly, these helical regions merge seamlessly into the PCI domain, at least if the longer PCI versions of PROSITE and Pfam are used. This finding is in agreement with the observation of several regions with weak TPR-similarity within the N-terminal part of the PCI domain itself (see figure 1). Further support for a TPR-like structure comes from a sequence-based fold recognition for lid and CSN subunits using the Superfamily-service [32]. Several subunits like Rpn7 from budding yeast and human Csn1 were found to have significant scores for the TPR fold upstream of the PCI domain (data not shown).
The predicted all-helical secondary structure of the non-PCI portion of lid and CSN subunits consists of several short helices that appear to occur in pairs. To test whether those bi-helical segments correspond to the structural elements of a TPR-like repeat, we selected several examples starting immediately upstream of the predicted PCI domains. When multiple alignments of those bi-helical segments were used for profile construction and in subsequent database searches several bona fide TPR proteins were found to match within the TPR region, with the bi-helices being in the correct TPR register, these segments were also classified as TPR-like. No matches to established HEAT- or Armadillo-repeat proteins were found, demonstrating that the scores are not just caused by an arbitrary helical repeat arrangement.
It should be pointed out that none of the singular observations described above is able to prove a statistically significant sequence relationship between the N-terminal portions of PCI proteins and true TPR-repeats. Taken as a whole, the results strongly suggest that there is a general tendency of PCI domains to be preceded by an α-helical repeat structure that has at least some specific relationship to the tetratrico-peptide repeat.
A previously unrecognized PCI domain in eIF3k
In the first surveys of recognizable PCI domains, only three PCI subunits of the eIF3 complex had been detected [2]. More recently, a number of novel eIF3 components have been identified: eIF3j [33], eIF3k [25] and eIF3l [34]. Among these novel subunits, only eIF3l has been reported to harbour a PCI domain [34], interestingly also preceded by a TPR-region. In order to find further indications of divergent PCI domains, we performed a thorough profile analysis of all uncharacterized eIF3 subunits.
A generalized profile was constructed from the conserved portion of representative eIF3k orthologs from vertebrates, invertebrates, plants and fungi. After a scaling step, the resulting profile was run against a nonredundant protein database. Apart from the eIF3k proteins already used for profile construction, the only other sequences matching with significance were selected PCI subunits of the proteasome and the CSN, among them rice Csn8 (p = 0.01) and the drosophila Rpn12 homologue (p = 0.05). All of the twenty top-scoring sequences could be identified as either Csn8- or Rpn12-homologs. As shown in table 1, Csn8 and Rpn12 are the corresponding PCI subunits in the CSN and the lid, respectively. Csn8 and Rpn1 are the most divergent PCI subunits of the proteasome and the signalosome, respectively, and their PCI domains appear to be shorter than that of the more typical family members. Our observations provide good bioinformatical evidence that eIF3k is the fifth PCI-containing subunit of the eIF3 complex and most likely a direct analogue of Csn8 and Rpn12 (figure 2, table 1).
A structural model for the canonical PCI domain
The discovery of a PCI domain in eIF3k is of particular importance, as a three-dimensional structure of eIF3 has been solved recently [26]. So far, no structural information on the PCI domain has been available, and a structural model for the canonical PCI domain based on the alignment shown in figure 2 should allow interesting insights into the architecture of the PCI complexes.
A detailed analysis of the eIF3k structure [26] reveals a bipartite structure of two subdomains that are in close contact through a large inter-domain surface patch (figure 3a). The C-terminal half-domain is a globular α/β structure with an "αβααββ" arrangement. The three β-strands are very short and form an antiparallel sheet. The whole C-terminal part can be classified as a "winged helix" fold and thus is referred to as "WH-domain" [26]. By contrast, the N-terminal half-domain is entirely helical with a core of six regularly-spaced helices that form three antiparallel helical hairpin elements. The resulting superhelix is reminiscent of the solenoids found in helical repeats such as HEAT, Armadillo and TPR. Somewhat unusual are the short 3–10 helices that connect the consecutive α-hairpins. According to Wei et al. [26] the N-terminal half-domain resembles structurally mainly HEAT and Armadillo repeats, and thus the name "HAM-domain" was proposed. The bipartite structure of eIF3k is in good overall agreement with the secondary structure predictions for the single PCI domain families and also with our result of TPR-like helical repeats partially overlapping the PCI domain. It was therefore of special interest to make a detailed comparison of the eIF3k structure and the profile-guided alignment of the canonical PCI superfamily shown in figure 2.
Within the N-terminal subdomain, the sequence conservation between the different PCI domain families is relatively poor and some aspects of the alignment shown in figure 2 are not very reliable. Nevertheless, there is a good correspondence between the helices that build the α-hairpins of eIF3k and the uninterrupted sequence blocks in the PCI alignment. The gap-regions in the PCI alignment are typically caused by insertion events in selected PCI subfamilies. In no case, a deletion of one or more of the hairpin helices is observed. This finding suggests that the helical hairpin structure is conserved in most or all PCI domains. Our alignment and the derived secondary structure predictions suggest that the short 310 helices that connect the helical hairpins in eIF3k are absent in most other PCI proteins. As mentioned in the previous paragraphs, there are several instances of subsignificant sequence similarity to TPR repeats found also within the N-terminal subdomain of the PCI domain. By contrast, no similarity to HEAT or Armadillo-repeats has been observed. Thus, we prefer to interpret the helical hairpin structure of the N-terminal subdomain as atypical TPR-like repeats rather than as the HEAT/Armadillo repeats suggested by Wei et al. [26].
The globular C-terminal subdomain (WH) is generally better conserved than the helical N-terminal domain and as a consequence, the part of the alignment covering this structural subdomain shown in figure 2 is more reliable. The "αβααββ" arrangement of α- and β-regions is distributed over two large sequence blocks with a single major gap region between "αβα" and "αββ". As can be seen in figure 2, no important secondary structure element is interrupted by a gap found in the PCI alignment. Like in the N-terminal subdomain, the WH portion shows a good concordance between the secondary structure predicted from the canonical PCI families and the structural elements of the eIF3k structure, apart from minor problems in predicting one of the very short β-strands.
Taken together, the comparison of the PCI alignment with the eIF3k structure (figure 3) shows that the two structures are clearly compatible and suggests that the canonical PCI domains will have an analogous bipartite fold similar to that shown in figure 3. The prediction of TPR-like helical repeats N-terminal of the proper PCI domain suggests that they form an extension of the helical repeat region of the first PCI subdomain. The implications of this model for the overall PCI structure will be discussed below.
Discussion
Revision of the eIF3 complex stoichiometry
While the proteasome lid and the CSN complex have an analogous architecture of six PCI-subunits and two MPN-subunits, the more distantly related eIF3 complex has so far only three readily detectable PCI proteins: eIF3a (EIF3S10), eIF3c (EIF3S8) and eIF3e (EIF3S6) [2]. Recent work by Morris-Desbois et al. [34] has grouped eIF3l (EIFS6IP) with the PCI components of eIF3, and our work described above adds eIF3k (EIF3S11) to the ranks of PCI proteins. Besides the PCI subunits, vertebrate eIF3 complexes also contain two MPN proteins: eIF3f (EIF3S1) and eIF3h (EIF3S3). Unlike the situation in the lid and CSN complexes, both MPN subunits of eIF3 have lost their metal-coordinating residues and are most likely catalytically inactive. In addition, several unicellular eukaryotes, including budding yeasts, do not seem to have any eIF3-associated MPN proteins.
Comparing the stoichiometry of eIF3 with the two better-conserved PCI complexes, only one PCI subunit seems to be missing. Our sequence analysis efforts have also included other known eIF3 subunits, but no indications for further PCI domains could be obtained (data not shown). Given the high degree of PCI sequence divergence, it cannot be fully excluded that one of the non-PCI/non-MPN subunits (eIF3b, eIF3d, eIF3g, eIF3i, eIF3j) harbours a cryptic PCI domain that has eluded our detection. On the other hand, it is well conceivable that eIF3 has a deviating subunit composition. In yeast and several other organisms, not only the MPN proteins are missing but also the number of PCI components is reduced, as eIF3e and eIF3l are absent. At present, it is not clear whether the corresponding positions in the complex are left empty or are filled by additional copies of the remaining PCI components. In evolutionary terms, it appears likely that the eIF3 complex is a 'degraded' copy of an ancient lid-like complex, which has lost its MPN+/JAMM mediated catalytic activity and potentially some of its PCI subunits. In turn, by acquiring a group of novel non-PCI/non-MPN subunits, the eIF3 complex has gained a functionality that is different (and potentially even completely unrelated) to the proteasome lid and its cousin, the CSN complex.
The bipartite structure of the PCI domain
In our original discovery note [2], we had defined the PCI domain as a homology domain, i.e. as a region of localized similarity found within multiple proteins that are otherwise unrelated. The results presented here suggest that this view should be revised. The sequence regions detected as PCI domains by bioinformatical methods seems to consist of two structurally distinct domains. The C-terminal portion, which in eIF3k is referred to as the WH-domain, is much better conserved in sequence than the N-terminal portion, and the C-terminal boundary of the PCI homology domain is relatively well defined by a notable loss of sequence conservation. By contrast, the N-terminal boundary of the homology domain has always been ill-defined, as the overall sequence conservation in this region is low and different families of PCI proteins appear to lose their similarity at different positions. As a consequence, different domain databases and their associated web-servers detect PCI-domains (or the synonymous 'PINT' domains) of varying length in the order PROSITE > Pfam > SMART.
Using the eIF3k-derived structural model, most of these observations can be readily explained. The C-terminal PCI/PINT boundary, which is agreed on by all domain databases, corresponds to the C-terminal boundary of the structural WH-like domain. The N-terminal boundary of the PINT domain, as described in the SMART database, essentially corresponds to the N-terminus of the WH-like domain. The PCI domain of the Pfam database corresponds to the WH-portion plus a single α-helical hairpin repeat. Finally, the PCI domain as described in the PROSITE database covers the WH-portion and all three helical hairpin repeats found in the eIF3k structure. Of the three representations, the PINT domain of the SMART database is structurally most correct, as it describes a true autonomously folding domain. The observation that some PCI families lose their sequence conservation at different N-terminal positions can be explained by assuming a variable number of helical repeat motifs for those proteins. As an extreme example, only the WH-like region could be detected in eIF3e by our profile searches, and the secondary structure prediction for the eIF3e family suggests a β-structure instead of the usual helical-hairpin repeats upstream of the WH region. This finding can be taken as a further hint for structural and functional independence of the N- and C-terminal sub-regions of the PCI homology domain.
The nature of the N-terminal helical repeat extension
Our finding of TPR-like repeats preceding many PCI domains, combined with the helical repeat structure of the N-terminal portion of the PCI domain itself, leads to the interesting question if these repeats are of the same type and may form a continuous solenoid structure. The authors of the eIF3k crystal structure propose a structural relationship between the eIF3k N-terminus and the HEAT motif based on superposition calculations with DALI [35]. By contrast, our sequence-based analysis methods rather point to an evolutionary relationship to the TPR motif, both for the region preceding the PCI domain and for the first helical hairpin of the PCI domain itself. A related finding was reported for Rpn3 and Csn12 elsewhere, where a homology domain termed "PAM" (PCI-associated module) with TPR-like properties has been proposed [36]. Our findings suggest that the PAM-domain is a special case of a more widespread preference of the WH-portion of the PCI domain to be preceded by TPR-like repeats. In addition, both our results and those of Ciccarelli et al. argue in favour of a continuity between the N-terminal repeats and those found within the PCI domain.
Due to the borderline sequence similarity between the classical TPR motif and the distinct helical hairpins of the PCI proteins, a completely novel type of bi-helical repeats distinct from TPR and HEAT/Armadillo or some kind of intermediary form cannot be ruled out. Structurally, HEAT and TPR repeats are relatively similar and both tend to form superhelical solenoid structures [37]. Without assuming a particular repeat family, we have attempted a rough estimation of what a typical PCI component of the lid or the CSN complex might look like. Figure 3b shows schematically a PCI protein with a WH-like domain at the C-terminus (green), preceded by three helical-repeats assumed to lie within the PCI boundaries according to PROSITE (dark blue), which are in turn preceded by three additional helical repeats that represent the TPR-related N-terminal extension (light blue). As we do not assume a particular repeat family with a well-known radius of solenoid curvature, we use the values derived from the first two helical hairpins of the eIF3 structure instead. It should be stressed that the model of figure 3b with its 'boomerang'-shaped architecture can only give a very coarse approximation of the real situation. Both the solenoid curvature and the exact number of N-terminal repeat extensions are rough estimations. Nevertheless, the model appears to be roughly compatible with the electron density maps of the lid and CSN complexes [4].
A structural scaffold for three multi-protein complexes
PCI proteins constitute the main components of the proteasome lid and the CSN complex and also form the structural core of the translation initiation factor eIF3. So far, no catalytic activity has been described for PCI proteins. Given the lack of invariant polar residues, such a role appears unlikely. The role of the PCI domains is most likely that of a scaffold for the other complex subunits and other binding partners. There are at least three distinct structural roles that PCI proteins have to fulfil: i) maintaining the integrity of the complex by binding to other PCI proteins, ii) attaching the MPN-subunits to the complex, and iii) binding to other partners such as the base-complex in the case of the proteasome lid or the RNA-binding subunits of the eIF3 complex.
The assignment of these functionalities to the different regions of the PCI proteins, and equally important, the source for the subunit interaction specificity or promiscuity have been subject to several experimental studies, some of them published while others have been presented at a recent meeting on PCI complexes [38]. The PCI model presented here will be certainly useful, both for the interpretation of the experimental results, and for the design of new experiments e.g. those based on domain truncations or domain swaps. According to our analysis, in some proteins the PCI domain is restricted to a C-terminal WH-like part. As these proteins are also components of PCI-complexes, a role of the WH domain in PCI:PCI domain interaction is very likely. On the other hand, TPR-repeats in general form versatile protein-interaction surfaces [29] and we expect the same to be true for the TPR-like repeats found in the PCI proteins.
Tsuge et al. analysed truncated forms of human Csn1 for its interaction with other PCI subunits of the CSN complex [16]. A construct containing residues 197–500 (corresponding to the entire PCI region and some C-terminal material) was able to bind to Csn2, Csn3 and Csn4. Another construct starting at position 340, and thus lacking the helical-repeat region, no longer bound to Csn2 and Csn4 but maintained binding to Csn3. By contrast, a construct 197–307 that lacks the WH-like region was only able to bind to Csn4. These experiments suggest that both the WH portion and the helical-repeat part of the PCI proteins have a role in PCI:PCI interactions, although they seem to interact with different subunits of the complex. The importance of both subdomains is confirmed by a recent study of Isono et al., who analyse multiple point mutations in the lid subunit Rpn7 [39]. In their hands, both mutations in the N-terminal helical repeat part of Rpn7 and mutations within the WH-like region are able to abrogate binding to Rpn3, another PCI subunit of the proteasome lid. So far, no information is available on the PCI regions involved in binding to the MPN subunits.
Conclusion
In summary, we believe the PCI domain could play a role as a universal binding domain supporting intra-complex interactions as well as recruitment of additional ligands. The model presented here is a first step to the understanding of the supramolecular architecture of three important complexes and certainly will facilitate the interpretation of further experimental results. Nevertheless, a full understanding of the interaction mode between PCI- and MPN-domain proteins will certainly require experimentally determined high-resolution structures of the components – or ideally, that of an intact complex.
Methods
Database searches
Sequence database searches were performed with a nonredundant data set constructed from current releases of SwissProt, TrEMBL, and GenPept [40,41]. Generalized profile construction [20] and searches were run locally using the pftools package, version 2.1. (program available from the URL ). Generalized profiles were constructed using the BLOSUM45 substitution matrix [42] and default penalties of 2.1 for gap opening and 0.2 for gap extension. Statistical significance of profile matches was derived from the analysis of the score distribution of a randomized database [43]. Database randomization was performed by individually inverting each protein sequence, using SwissProt 34 as the data source. Only sequence matches found with a probability of p < 0.01 were included into subsequent rounds of iterative profile refinement.
Multiple alignments
For sufficiently related proteins, multiple alignments were calculated by T-coffee [44], using excised domains instead of the entire sequences. For alignments of highly divergent sequences, such as the whole PCI family, the overall alignment was generated by profile-guided assembly of family-specific subalignments. If necessary, manual adjustments were introduced in the final alignment step. Whenever possible, positions for insertion- and deletion events were placed according to predicted secondary structure elements derived from the subfamilies involved.
Structure prediction
For each subfamily-specific multiple alignment, secondary structure elements were predicted using web services of PHD [30] and JPred [31]. JPred uses a set of different algorithms for secondary structure prediction and calculates a consensus prediction. By combining the JPred and PHD derived secondary structure predictions, every position within a given alignment was assigned to "helical", "sheet" or "none". For attempting a sequence-based fold recognition, representative sequences were submitted to the 'Superfamily' web service [32].
Authors' contributions
HS performed the sequence analysis and participated in writing the manuscript. KH designed the study, participated in the sequence analysis and in writing the manuscript. All authors read and approved the final manuscript.
Figures and Tables
Figure 1 TPR-like motifs upstream and inside the PCI domain. The proposed domain topologies of selected human PCI proteins were investigated by profile techniques. Besides the common PCI domain (blue) short stretches of ~35 aa each are depicted in orange and light blue. These stretches show weak to medium similarity to TPR segments in established TPR proteins and merge seamlessly into the PCI domain in several PCI subunits. Corresponding accession numbers are listed in figure 2.
Figure 2 Multiple sequence alignment of human PCI subunits from proteasome lid, CSN and eIF3. Shown are only the segments matched by the PROSITE PCI domain. Conserved residues printed on black background were found in at last 50 % of ~60 PCI proteins from selected species, from which only human representatives are shown. Grey background was assigned to positions occupied by residues with similar physicochemical properties in at least 50 % of the sequences. The alignment was shaded using BOXSHADE . Above the PCI alignment secondary structure prediction as calculated from JPred [31] is presented. In these calculations sequences of eIF3k homologues were not included. Secondary structure elements of eIF3k as derived from PDB structure 1RZ4 are shown in a separate row. The abbreviations denote the following secondary structure types: E extended (sheet) and H helix. In addition, structural subdomain classification ('HAM', 'WH') as described in Wei et al. [26] and domain boundaries according to PCI profiles from PROSITE and Pfam are provided. Sequence names correspond to the following SwissProt database entries: eIF3k (Q9UBQ5), PSMD3 (O43242), PSMD6 (Q15008), PSMD8 (P48556), PSMD11 (O00231), PSMD12 (O00232), PSMD13 (Q9UNM6), CSN1 (Q13098), CSN2 (P61201), CSN3 (Q9UNS2), CSN4 (Q9BT78), CSN8 (Q99627).
Figure 3 (A) shows the overall structure of eIF3k from the PDB-entry 1RZ4 [26] with β-strands and α-helices represented as ribbons and cylinders, respectively. Regions of the structure with sequence similarity to canonical PCI domain are rendered in colour. Regions belonging to the WH subdomain are shown in green, while conserved structure elements of the helical hairpin regions are shown in dark blue. The connection between β-strand 2 and 3 is not resolved and thus missing in 1RZ4. Other regions (extreme N- and C- termini, connecting helices between hairpins, unstructured regions) are shown in grey. (B) Model of a PCI protein with three additional helical hairpins upstream of the PCI domain. Within the PCI domains, only regions that can be modelled on the eIF3k template are shown. The N-terminal extension is shown in light blue, the other colours are as in figure 3a.
Table 1 PCI complexes and their subunit correspondence
Domain proteasome lid CSN eIF3
PCI Rpn7 / PSMD6 Csn1 eIF3a, eIF3c, eIF3e, eIF3l
PCI Rpn6 / PSMD11 Csn2 eIF3a, eIF3c, eIF3e, eIF3l
PCI Rpn3 / PSMD3 Csn3 eIF3a, eIF3c, eIF3e, eIF3l
PCI Rpn5 / PSMD12 Csn4 eIF3a, eIF3c, eIF3e, eIF3l
PCI Rpn9 / PSMD13 Csn7a,b eIF3a, eIF3c, eIF3e, eIF3l
PCI Rpn12 / PSMD8 Csn8 eIF3k
MPN+ Rpn11 / PSMD14 Csn5 -
MPN Rpn8 / PSMD7 Csn6 eIF3f, eIF3h
==== Refs
Kim T Hofmann K von Arnim AG Chamovitz DA PCI complexes: pretty complex interactions in diverse signaling pathways Trends Plant Sci 2001 6 379 386 11495792 10.1016/S1360-1385(01)02015-5
Hofmann K Bucher P The PCI domain: a common theme in three multiprotein complexes Trends Biochem Sci 1998 23 204 205 9644972 10.1016/S0968-0004(98)01217-1
Aravind L Ponting CP Homologues of 26S proteasome subunits are regulators of transcription and translation Protein Sci 1998 7 1250 1254 9605331
Kapelari B Bech-Otschir D Hegerl R Schade R Dumdey R Dubiel W Electron microscopy and subunit-subunit interaction studies reveal a first architecture of COP9 signalosome J Mol Biol 2000 300 1169 1178 10903862 10.1006/jmbi.2000.3912
Fu H Reis N Lee Y Glickman MH Vierstra RD Subunit interaction maps for the regulatory particle of the 26S proteasome and the COP9 signalosome Embo J 2001 20 7096 7107 11742986 10.1093/emboj/20.24.7096
Glickman MH Rubin DM Coux O Wefes I Pfeifer G Cjeka Z Baumeister W Fried VA Finley D A subcomplex of the proteasome regulatory particle required for ubiquitin-conjugate degradation and related to the COP9-signalosome and eIF3 Cell 1998 94 615 623 9741626 10.1016/S0092-8674(00)81603-7
Verma R Aravind L Oania R McDonald WH Yates JR Koonin EV Deshaies RJ Role of Rpn11 metalloprotease in deubiquitination and degradation by the 26S proteasome Science 2002 298 611 615 12183636 10.1126/science.1075898
Maytal-Kivity V Reis N Hofmann K Glickman MH MPN+, a putative catalytic motif found in a subset of MPN domain proteins from eukaryotes and prokaryotes, is critical for Rpn11 function BMC Biochem 2002 3 28 12370088 10.1186/1471-2091-3-28
Yao T Cohen RE A cryptic protease couples deubiquitination and degradation by the proteasome Nature 2002 419 403 407 12353037 10.1038/nature01071
Serino G Deng XW The COP9 signalosome: regulating plant development through the control of proteolysis Annu Rev Plant Biol 2003 54 165 182 14502989 10.1146/annurev.arplant.54.031902.134847
Schwechheimer C Deng XW COP9 signalosome revisited: a novel mediator of protein degradation Trends Cell Biol 2001 11 420 426 11567875 10.1016/S0962-8924(01)02091-8
Cope GA Suh GS Aravind L Schwarz SE Zipursky SL Koonin EV Deshaies RJ Role of predicted metalloprotease motif of Jab1/Csn5 in cleavage of Nedd8 from Cul1 Science 2002 298 608 611 12183637 10.1126/science.1075901
Asano K Kinzy TG Merrick WC Hershey JW Conservation and diversity of eukaryotic translation initiation factor eIF3 J Biol Chem 1997 272 1101 1109 8995409 10.1074/jbc.272.28.17668
Dever TE Translation initiation: adept at adapting Trends Biochem Sci 1999 24 398 403 10500305 10.1016/S0968-0004(99)01457-7
Tran HJ Allen MD Lowe J Bycroft M Structure of the Jab1/MPN domain and its implications for proteasome function Biochemistry 2003 42 11460 11465 14516197 10.1021/bi035033g
Tsuge T Matsui M Wei N The subunit 1 of the COP9 signalosome suppresses gene expression through its N-terminal domain and incorporates into the complex through the PCI domain J Mol Biol 2001 305 1 9 11114242 10.1006/jmbi.2000.4288
Valasek L Phan L Schoenfeld LW Valaskova V Hinnebusch AG Related eIF3 subunits TIF32 and HCR1 interact with an RNA recognition motif in PRT1 required for eIF3 integrity and ribosome binding Embo J 2001 20 891 904 11179233 10.1093/emboj/20.4.891
Lier S Paululat A The proteasome regulatory particle subunit Rpn6 is required for Drosophila development and interacts physically with signalosome subunit Alien/CSN2 Gene 2002 298 109 119 12426099 10.1016/S0378-1119(02)00930-7
Freilich S Oron E Kapp Y Nevo-Caspi Y Orgad S Segal D Chamovitz DA The COP9 signalosome is essential for development of Drosophila melanogaster Curr Biol 1999 9 1187 1190 10531038 10.1016/S0960-9822(00)80023-8
Bucher P Karplus K Moeri N Hofmann K A flexible motif search technique based on generalized profiles Comput Chem 1996 20 3 23 8867839 10.1016/S0097-8485(96)80003-9
Maytal-Kivity V Pick E Piran R Hofmann K Glickman MH The COP9 signalosome-like complex in S. cerevisiae and links to other PCI complexes Int J Biochem Cell Biol 2003 35 706 715 12672462 10.1016/S1357-2725(02)00378-3
Hulo N Sigrist CJ Le Saux V Langendijk-Genevaux PS Bordoli L Gattiker A De Castro E Bucher P Bairoch A Recent improvements to the PROSITE database Nucleic Acids Res 2004 32 Database issue D134 7 14681377 10.1093/nar/gkh044
Bateman A Coin L Durbin R Finn RD Hollich V Griffiths-Jones S Khanna A Marshall M Moxon S Sonnhammer EL Studholme DJ Yeats C Eddy SR The Pfam protein families database Nucleic Acids Res 2004 32 Database issue D138 41 14681378 10.1093/nar/gkh121
Letunic I Copley RR Schmidt S Ciccarelli FD Doerks T Schultz J Ponting CP Bork P SMART 4.0: towards genomic data integration Nucleic Acids Res 2004 32 Database issue D142 4 14681379 10.1093/nar/gkh088
Mayeur GL Fraser CS Peiretti F Block KL Hershey JW Characterization of eIF3k: a newly discovered subunit of mammalian translation initiation factor elF3 Eur J Biochem 2003 270 4133 4139 14519125 10.1046/j.1432-1033.2003.03807.x
Wei Z Zhang P Zhou Z Cheng Z Wan M Gong W Crystal Structure of Human eIF3k, the First Structure of eIF3 Subunits J Biol Chem 2004 279 34983 34990 15180986 10.1074/jbc.M405158200
Hofmann K Bucher P Falquet L Bairoch A The PROSITE database, its status in 1999 Nucleic Acids Res 1999 27 215 219 9847184 10.1093/nar/27.1.215
Kellis M Patterson N Endrizzi M Birren B Lander ES Sequencing and comparison of yeast species to identify genes and regulatory elements Nature 2003 423 241 254 12748633 10.1038/nature01644
D'Andrea LD Regan L TPR proteins: the versatile helix Trends Biochem Sci 2003 28 655 662 14659697 10.1016/j.tibs.2003.10.007
Rost B Liu J The PredictProtein server Nucleic Acids Res 2003 31 3300 3304 12824312 10.1093/nar/gkg508
Cuff JA Clamp ME Siddiqui AS Finlay M Barton GJ JPred: a consensus secondary structure prediction server Bioinformatics 1998 14 892 893 9927721 10.1093/bioinformatics/14.10.892
Gough J Karplus K Hughey R Chothia C Assignment of homology to genome sequences using a library of hidden Markov models that represent all proteins of known structure J Mol Biol 2001 313 903 919 11697912 10.1006/jmbi.2001.5080
Valasek L Hasek J Nielsen KH Hinnebusch AG Dual function of eIF3j/Hcr1p in processing 20 S pre-rRNA and translation initiation J Biol Chem 2001 276 43351 43360 11560931 10.1074/jbc.M106887200
Morris-Desbois C Rety S Ferro M Garin J Jalinot P The human protein HSPC021 interacts with Int-6 and is associated with eukaryotic translation initiation factor 3 J Biol Chem 2001 276 45988 45995 11590142 10.1074/jbc.M104966200
Holm L Sander C Dali/FSSP classification of three-dimensional protein folds Nucleic Acids Res 1997 25 231 234 9016542 10.1093/nar/25.1.231
Ciccarelli FD Izaurralde E Bork P The PAM domain, a multi-protein complex-associated module with an all-alpha-helix fold BMC Bioinformatics 2003 4 64 14687415 10.1186/1471-2105-4-64
Kajava AV What curves alpha-solenoids? Evidence for an alpha-helical toroid structure of Rpn1 and Rpn2 proteins of the 26 S proteasome J Biol Chem 2002 277 49791 49798 12270919 10.1074/jbc.M204982200
Chang EC Schwechheimer C ZOMES III: the interface between signalling and proteolysis EMBO Rep 2004 5 1041 1045 15514681 10.1038/sj.embor.7400275
Isono E Saeki Y Yokosawa H Toh-e A Rpn7 Is required for the structural integrity of the 26 S proteasome of Saccharomyces cerevisiae J Biol Chem 2004 279 27168 27176 15102831 10.1074/jbc.M314231200
Bairoch A Apweiler R The Swiss-Prot Protein Sequence Data Bank and Its Supplement Trembl Nucleic Acids Research 1997 25 31 36 9016499 10.1093/nar/25.1.31
Benton D Recent changes in the GenBank On-line Service Nucleic Acids Research 1990 18 1517 1520 2326192
Henikoff S Henikoff JG Amino acid substitution matrices from protein blocks Proceedings of the National Academy of Sciences of the United States of America 1992 89 10915 10919 1438297
Hofmann K Sensitive protein comparisons with profiles and hidden Markov models Brief Bioinform 2000 1 167 178 11465028
Notredame C Higgins DG Heringa J T-Coffee: A novel method for fast and accurate multiple sequence alignment J Mol Biol 2000 302 205 217 10964570 10.1006/jmbi.2000.4042
|
15790418
|
PMC1274264
|
CC BY
|
2021-01-04 16:02:48
|
no
|
BMC Bioinformatics. 2005 Mar 24; 6:71
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-71
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-721579040210.1186/1471-2105-6-72DatabaseGeneKeyDB: A lightweight, gene-centric, relational database to support data mining environments Kirov SA [email protected] X [email protected] E [email protected] D [email protected] B [email protected] J [email protected] Graduate School for Genome Science and Technology, Oak Ridge National Laboratory-University of Tennessee, Oak Ridge, USA2 Life Sciences Division, Oak Ridge National Laboratory, Oak Ridge, USA3 Department of Engineering and Computer Science, Baylor University, Waco, USA2005 24 3 2005 6 72 72 2 11 2004 24 3 2005 Copyright © 2005 Kirov et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The analysis of biological data is greatly enhanced by existing or emerging databases. Most existing databases, with few exceptions are not designed to easily support large scale computational analysis, but rather offer exclusively a web interface to the resource. We have recognized the growing need for a database which can be used successfully as a backend to computational analysis tools and pipelines. Such database should be sufficiently versatile to allow easy system integration.
Results
GeneKeyDB is a gene-centered relational database developed to enhance data mining in biological data sets. The system provides an underlying data layer for computational analysis tools and visualization tools. GeneKeyDB relies primarily on existing database identifiers derived from community databases (NCBI, GO, Ensembl, et al.) as well as the known relationships among those identifiers. It is a lightweight, portable, and extensible platform for integration with computational tools and analysis environments.
Conclusion
GeneKeyDB can enable analysis tools and users to manipulate the intersections, unions, and differences among different data sets.
==== Body
Background
As we move toward large-scale research into complex molecular and cellular networks, the research community will need to develop new interfaces to complex data sets. Existing databases and interfaces, such as those at EBI and NCBI[1,2], often use sequence records as the central organizing unit. A database organization around a genome sequence record, for example, might be ideal for the purpose of a genome analysis, while the analysis of biological networks would be better organized around genes and gene products. LocusLink [3] (soon to be replaced by Entrez Gene [4]) is an example of a resource that adapts the more suitable gene-centric view. While having excellent user interfaces (UIs), LocusLink does not provide robust application programming interfaces (APIs). Even though an API could use web interface or a flat file database, this would make the analysis tool unacceptably slow. In particular, APIs are needed for computers to process the sets of the genes and gene products that are found in these biological networks. Both computational tools and advanced data mining environments need to use these APIs to access and manipulate large, diverse, and intersecting sets of data. EBI's EnsMART [5] is a resource that permits a comparable manipulation of data about sets of genes and provides an API along with the UI. The database, however, is somewhat difficult to store locally due to its large size and complexity. Another database that is to some extent similar with respect to the design is the DRAGON database[6]. We have developed GeneKeyDB, a relational database, in an attempt to address these issues. A schematic comparison of these databases and GeneKeyDB can be seen in Table 1.
An interesting alternative to the above mentioned databases is BioMART[7]. This is not a database in the conventional sense (though the underlying data can also be downloaded). It extracts and integrates data from several sources, creating customized database. While this approach is very powerful, not all data present in GeneKeyDB is available from BioMART sources (for example CGAP expression data or Homologene). Still BioMART requires a human intervention to retrieve the customized database through a web interface, where GeneKeyDB can be updated entirely through scripts.
The development of GeneKeyDB is motivated by a desire to have a smaller-sized database that could tightly interoperate with different local computational tools and local data sets. While providing support for different data mining tools, the database may remain lightweight, by just storing the keys (database identifiers) of objects, some important attributes, and the relationships among the objects. The database does not actually need to store large objects, like sequence records, as long any tool built on top of GeneKeyDB can retrieve the subset of the objects of interest on demand from other sources – EBI, NCBI or local storage. The system allows us to create centrally shared functions to manipulate biological data sets. Each analysis tool, analysis pipeline, or computer can rely on these core functions and data found in GeneKeyDB.
Construction and content
GeneKeyDB consists of several sub-modules (Figure 1), corresponding to the represented data sources or supported services- LocusLink, EnsEMBL, HomoloGene, GoTreeMachine (GOTM), MGI comparative map and CGAP tables [1,2,8-10]. Additional modules can be easily added through the central key of the database. For example we are developing a Cis-Regulatory Elements database called PSITE (manuscript in preparation) which is already integrated to GeneKeyDB through LocusLink ids (Entrez Gene identifiers). In a similar way a supporting GOTM database module has been created previously [10]. Consistency between the database sub-modules is achieved through the creation of novel joining relations. The overall effect is a schema where no data connection is more than two relational tables apart. Summary of the data, provided by each source can be seen in Table 2.
The database is created using Oracle and the parsers are written in Perl. Bioperl and ENSEMBL API are also used to create some of the tables. A brief overview of the database creation process is shown in Figure 2. Though we do not cleanse the source data, there are natural restrictions that should be enforced. For example no RefSeq accession number can be mapped to two LocusLink identifiers, no ensembl gene stable identifier can be mapped to more than one LocusLink identifiers, and the combination of the gene symbol and organism name should be unique.
LocusLink tables are derived through parsing LocusLink flat files and inferring additional data such as absolute chromosome coordinates, calculated by joining NCBI contig builds (see supplemental data at GeneKeyDB website [11]). These tables describe fundamental information about a particular gene: name, description, associated accession numbers, chromosome location, suggested function, comparative map information among other variables. ENSEMBL originating information also occupies a significant part of the database. It holds the relationship between LocusLink identifiers and ENSEMBL, the original tables used to create this correlation, and additional inferred data such as absolute chromosome coordinates. The CGAP data is parsed into several tables, two of which hold the expression and cDNA library data, while the rest hold functional annotation. These tables are created and updated automatically twice per week, except Affymetrix data, which must be downloaded manually from a password protected area. The update takes approximately 12 hours, including download, parsing and loading. Each sub module can be processed in a distributed fashion (after the central module, Locuslink is parsed), thus significantly reducing the time needed for the update. This allows new modules to be added without noticeable impact over the time that the update process will take This allows us to user more than one processor, for example At the conclusion of the process, new data is exported both as comma delimited and tab delimited files along with the create statements for Oracle, mysql and PostgreSQL RDBMS.
It is important to note that no extensive cleansing of the data is performed during the database creation and update process. This allows automatic updates and eliminates some well known problems created by data cleansing [12]. At the same time, the database automates the creation of views that display the best solution to possible conflicting entries, usually derived from different sources. An example of such conflicts are chromosome coordinates (UCSC, NCBI, Ensembl) and groups of orthologues (Homologene, MGI).
Currently we are migrating from LocusLink data to Entrez Gene data as it becomes available from NCBI.
Utility and discussion
GeneKeyDB is primarily used for automating the integration of large data sets with a variety of computational tools that examine secondary gene relationships, such as the analysis of gene regulatory networks and their evolution. Current tools heavily relying on GeneKeyDB include the BSA pipeline (manuscript in preparation), GoTree machine and other utilities available on our web site. GoTree and Webgestalt [13] for example, rely on GeneKeyDB to generate gene ontology relationships.
The database was constructed, following some of the guidelines and logic described by Waterman et al. [14]. GeneKeyDB can be used to answer similar to or even more complex question than the ones described in the same work as shown by the following set of examples (more details available from GeneKeyDB website):
1. Define a set of candidate genes, based on the genome localization and tissue expression pattern[15]
2. Get all genes expressed in a specific tissue that share a conserved protein domain[16].
3. Get a set of orthologues to a group of genes (defined by the input) and obtain their genome coordinates [17]
These examples are designed to help formulate sets of genes, sharing some functional, sequence or expression similarity. These results are not permanent as the information differs between releases. The times for executing the queries was less than 30 s. per each one and have been produced under Oracle 9.0i database on Sparc Ultra-4, SunOS 5.8.
While such queries are complex, GeneKeyDB's lightweight nature allows its conversion to popular databases such as Microsoft Access and FileMaker Pro, where graphical user interfaces allow for the creation of queries with only a rudimentary knowledge of SQL. Even though Microsoft Access and FileMaker have limited capabilities compared to Oracle, MySQL or PostgreSQL, they succeeded in handling GeneKeyDB.
Conclusion
The development of more exhaustive high-throughput experimental procedures has led to the accumulation of abundant biological data resources. As a consequence, the LocusLink approach of one-gene-at-a-time is an insufficient method to properly mine the data and therefore, analysis of sets of genes is more productive. Additionally, converting between multiple database identifiers is still a challenge as there is no current method to uniquely identify the same gene across multiple databases. Therefore a data mining environment that can synchronize multiple sources and provide general annotation information is going to be beneficial for comparing and using results originating from different experimental groups. One excellent existing resource is EnsMART[5], but its complexity and size can be overwhelming, especially if system integration is needed. GeneKeyDB can be used to address the same issues, but may be run locally and is roughly 20 times smaller. At approximately 1 GB, it is easily integrated with new and existing computational tools. For the same reason, GeneKeyDB can be used to support distributed analysis for purposes of genome annotation, regulatory network predictions, and other analysis pipelines. Another advantage is that any tool based on GeneKeyDB will rely on standard SQL to manipulate data instead of a proprietary code. On the other hand, GeneKeyDB could easily interact with BioSQL database[18], with GeneKeyDB providing the central keys and relations between them, while BioSQL can supply sequence data and additional annotation also through standard SQL. This creates a system that is both very flexible and powerful.
Compared to the most similar database, DRAGON, GeneKeyDB has significantly more keys and relationships (CGAP, LocusLink, MGI, Homologene, Affymetrix identifiers, etc), where DRAGON has only some PFAM, Unigene and Incyte data as TREMBL, Transfac and Interpro data is not implemented yet in this database (see DRAGON database website [19]).
Through the GeneKeyDB database, it is possible to bring together different data such as evolutionary (HomoloGene, MGI comparative map), expression (CGAP), physical location and functional annotation (GO) in a high throughput fashion, making this resource valuable both to experimental and bioinformatics groups. Though the database is most useful when installed locally, some web based functionality also exists. The most significant function that this system provides is a small lightweight database that can enable analysis tools and local database to have the flexible functions of database queries about genes, gene products, and sets of genes in the course of their large-scale analysis.
Current and future development includes improvement to the MySQL/PostgreSQL export, migration toward Entrez Gene (done gradually as new data becomes available at the FTP site), test integration with BioSQL, etc. We intend to convert GeneKeyDB to an open source project and share our parsers with other relevant open source project such as BioPerl wherever appropriate. We expect this step to help GeneKeyDB maintenance.
Availability and requirements
SQL access (through web), Oracle, PostgreSQL and MySQL downloads are available from .
Project Name: GeneKeyDB
Project Homepage:
Operating System: Platform independent, tested under Linux and Unix
Other Requirements: RDBMS, preferably Oracle, MySQL or PostgreSQL
License: GNU GPL
Any Restrictions to use by non-academics: License needed
List of abbreviations
RDBMS- relational database management system
SQL- simple query language
UIs- user interfaces
APIs- application programming interfaces
GOTM- GO Tree Machine
GRIF- Gene Reference Into Function
Authors' contributions
SK drafted the manuscript and is currently implementing the newest features in GeneKey. SK, EB and BZ are using aspects of GeneKeyDB to extend the data mining of gene sets. DS, SK, BZ, XP and EB developed the GeneKeyDB database. JS guided and coordinated execution of the project. All authors read and approved the final manuscript.
Acknowledgements
We thank Sean Davis at NHGRI for critically reading the manuscript, Wes Hickey, Aaron Douthit and Travis Taylor for their technical help in developing the web user interface for the LocusLink data and Suzanne Baktash for technical help preparing the manuscript. This work was supported by the INIA project (NIH/NIAAA, U01-AA013532), the BISTI project (NIH/NIDA, P01-DA015027) and the ORNL LDRD project (DOE, AC05-00OR22725).
Figures and Tables
Figure 1 GeneKeyDB sub-modules, external database identifiers and connecting tables. The connecting tables may convert between the central key and another unique key used throughout the sub-module and are shown next to the connector lines.
Figure 2 A workflow schema of GeneKeyDB creation and export to other RDBMS. *LocusLink is parsed first as other sub-modules depend on it with respect to the central key of the database. PROD refers to the current production stage database.
Table 1 Comparison of different databases, which could be used to annotate and analyze large-scale biological data. Simplified joins refers to the availability of a central key and the ability to join tables through simple queries. Remote access column refers only to a machine access to a database server.
Database SQL Web interface Simplified joins 1Remote access Automatic updates provided Type Structure available as Modular design Download
GeneKeyDB Yes 2Indirect Yes No Yes Oracle Oracle, mysql, postgresql Yes Yes
LocusLink (Entrez Gene) No Yes Na Na No Flat file(s) Flat file only Na Yes
RefSeq No Yes Na Na No Flat file Flat file only Na Yes
Ensembl (core databases) Yes Yes No Yes No Mysql Mysql No Yes
EnsMART Yes Yes Yes Yes No Mysql Mysql No Yes
Dragon Yes Yes No No No Mysql Flat file only No Data files only
HomoloGene No Yes Na Na No Flat files, XML XML No Yes
1Refers only to machine access to the relational databases.
2GeneKeyDB serves as a data mining environment to different tools, therefore these tools are could also be considered a part of the interface layer.
Table 2 Summary of the attributes provided by each source
Sub module Source Attributes
CGAP CGAP Expression data, LocusLink ID to Unigene and Genbank Accession; Unigene to KEGG/GO/Biocarta
UCSC UCSC RefGene LocusLink ID to chromosome coordinates and exon structure
LocusLink LocusLink (Entrez Gene) Locuslink to Genbank accesion number, RefSeq Accession numbers, Gene descriptions (symbol, name, etc.), GRIF, Pubmed, OMIM, CDD, map location,
MGI MGI comparative map Homology data
Homologene Homologene Homology data
Ensembl Ensembl LocusLink ID to Ensebml Gene and Transcript Stable IDs, Contig data
==== Refs
Brooksbank C Camon E Harris MA Magrane M Martin MJ Mulder N O'Donovan C Parkinson H Tuli MA Apweiler R Birney E Brazma A Henrick K Lopez R Stoesser G Stoehr P Cameron G The European Bioinformatics Institute's data resources Nucleic Acids Res 2003 31 43 50 12519944 10.1093/nar/gkg066
Wheeler DL Church DM Federhen S Lash AE Madden TL Pontius JU Schuler GD Schriml LM Sequeira E Tatusova TA Wagner L Database resources of the National Center for Biotechnology Nucleic Acids Res 2003 31 28 33 12519941 10.1093/nar/gkg033
Pruitt KD Maglott DR RefSeq and LocusLink: NCBI gene-centered resources Nucleic Acids Res 2001 29 137 140 11125071 10.1093/nar/29.1.137
Entrez Gene
Kasprzyk A Keefe D Smedley D London D Spooner W Melsopp C Hammond M Rocca-Serra P Cox T Birney E EnsMart: a generic system for fast and flexible access to biological data Genome Res 2004 14 160 169 14707178 10.1101/gr.1645104
Bouton CM Pevsner J DRAGON View: information visualization for annotated microarray data Bioinformatics 2002 18 323 324 11847082 10.1093/bioinformatics/18.2.323
BioMart homepage
Riggins GJ Strausberg RL Genome and genetic resources from the Cancer Genome Anatomy Project Hum Mol Genet 2001 10 663 667 11257097 10.1093/hmg/10.7.663
Harris MA Clark J Ireland A Lomax J Ashburner M Foulger R Eilbeck K Lewis S Marshall B Mungall C Richter J Rubin GM Blake JA Bult C Dolan M Drabkin H Eppig JT Hill DP Ni L Ringwald M Balakrishnan R Cherry JM Christie KR Costanzo MC Dwight SS Engel S Fisk DG Hirschman JE Hong EL Nash RS Sethuraman A Theesfeld CL Botstein D Dolinski K Feierbach B Berardini T Mundodi S Rhee SY Apweiler R Barrell D Camon E Dimmer E Lee V Chisholm R Gaudet P Kibbe W Kishore R Schwarz EM Sternberg P Gwinn M Hannick L Wortman J Berriman M Wood V de la Cruz N Tonellato P Jaiswal P Seigfried T White R The Gene Ontology (GO) database and informatics resource Nucleic Acids Res 2004 32 Database issue D258 61 14681407
Zhang B Schmoyer D Kirov S Snoddy J GOTree Machine (GOTM): a web-based platform for interpreting sets of interesting genes using Gene Ontology hierarchies BMC Bioinformatics 2004 5 16 14975175 10.1186/1471-2105-5-16
GeneKeyDB website
Heiko Muller JCF Problems, Methods, and Challenges in Comprehensive Data Cleansing Technical Report HUB-IB-164, Humboldt University Berlin 2003
WebGestalt
Waterman M Uberbacher E Spengler S Smith FR Slezak T Robbins RJ Marr T Kingsbury DT Gilna P Fields C Genome informatics I: community databases J Comput Biol 1994 1 173 190 8790463
GeneKeyDB example 1
GeneKeyDB example 2
GeneKeyDB example 3
Open Bioinformatics Foundation
DRAGON database
|
15790402
|
PMC1274265
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 24; 6:72
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-72
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-751579042110.1186/1471-2105-6-75Methodology ArticleRanking the whole MEDLINE database according to a large training set using text indexing Suomela Brian P [email protected] Miguel A [email protected] Ontario Genomics Innovation Centre, Ottawa Health Research Institute, 501 Smyth Rd, Ottawa, Ontario K1H 8L6, Canada2005 24 3 2005 6 75 75 8 9 2004 24 3 2005 Copyright © 2005 Suomela and Andrade; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The MEDLINE database contains over 12 million references to scientific literature, with about 3/4 of recent articles including an abstract of the publication. Retrieval of entries using queries with keywords is useful for human users that need to obtain small selections. However, particular analyses of the literature or database developments may need the complete ranking of all the references in the MEDLINE database as to their relevance to a topic of interest. This report describes a method that does this ranking using the differences in word content between MEDLINE entries related to a topic and the whole of MEDLINE, in a computational time appropriate for an article search query engine.
Results
We tested the capabilities of our system to retrieve MEDLINE references which are relevant to the subject of stem cells. We took advantage of the existing annotation of references with terms from the MeSH hierarchical vocabulary (Medical Subject Headings, developed at the National Library of Medicine). A training set of 81,416 references was constructed by selecting entries annotated with the MeSH term stem cells or some child in its sub tree. Frequencies of all nouns, verbs, and adjectives in the training set were computed and the ratios of word frequencies in the training set to those in the entire MEDLINE were used to score references. Self-consistency of the algorithm, benchmarked with a test set containing the training set and an equal number of references randomly selected from MEDLINE was better using nouns (79%) than adjectives (73%) or verbs (70%). The evaluation of the system with 6,923 references not used for training, containing 204 articles relevant to stem cells according to a human expert, indicated a recall of 65% for a precision of 65%.
Conclusion
This strategy appears to be useful for predicting the relevance of MEDLINE references to a given concept. The method is simple and can be used with any user-defined training set. Choice of the part of speech of the words used for classification has important effects on performance. Lists of words, scripts, and additional information are available from the web address .
==== Body
Background
As the amount of textual information generated by scientific research expands, there is an increasing need for effective literature mining that can help scientists gather relevant knowledge encoded in text documents. The challenge is to develop methods of automated information extraction to support building logical databases and discover new knowledge from online journal collections. A large amount of information for biological research is available in the form of free text such as MEDLINE abstracts. Abstracts are collected and maintained in the MEDLINE database which currently contains references to over 12 million articles dating back to the mid 1960's in domains of molecular biology, biomedicine and medicine, and currently growing by almost half a million articles per year.
MEDLINE articles of interest can be searched for through the PubMed server [1] with queries using a Boolean combination of free text or controlled vocabulary keywords. The usefulness of free text keyword searching will depend on the word content in the title and/or abstract of references of interest. Some interfaces map free text terms to a corresponding Medical Subject Heading (MeSH) [2]. Subject heading (thesaurus, controlled vocabulary) searching can also be a powerful strategy for finding information. Subheadings can help to focus the scope of the search space. This strategy is appropriate for researchers interested in a narrow concept to retrieve a small slice of references for visual inspection. However, there are certain computational analyses of the literature or database developments that would require the ranking of the complete MEDLINE database of references as to their relation to a topic of interest. For example, given any two articles it would be useful to decide which one relates more to a topic.
Many machine learning methods have been applied to the problem of document classification [3,4]. Typically such algorithms learn from a set of text that has already been classified (training set) how to classify another set of documents (test set). Naïve Bayes, k-nearest neighbors, decision trees, neural networks, and support vector machines are a few of the most common machine learning algorithms [4]. A key difference between these methods is the way that the documents are represented by the features selected (most often words or phrases from the text) [5]. This results in differences not only in performance but also in the time that is needed to train the method of choice on the test set. Those differences become very significant when the training and the test sets become on the order of thousands. As a result, only naïve Bayesian learning has been applied to the ranking of MEDLINE abstracts (where the test set is on the order of millions of abstracts) and only using training sets of one hundred examples [6]. Just to give an example, a recent survey of these machine learning algorithms on comparatively small sets of text documents required more than five years of CPU time [7]. Applying these methods to classifying thirteen million vectors which are each the width of the number of words used in all the articles in MEDLINE (several hundred thousand even after removal of rare terms and stop words) would certainly be an impossible computational task.
An alternative is given by text indexing based on word frequencies [8]. The titles and abstracts of MEDLINE references contain words that are indicative of specific topics which can be detected by examining how a given word is used more often in references dealing with the topic than in unrelated references. We have previously used this to find keywords in MEDLINE abstracts describing protein families [9] or genetic disease [10], by using the ratio of word usage in a group of pre-selected abstracts with respect to word usage in MEDLINE.
Here we propose to use this approach not just to extract keywords but also to evaluate the entire MEDLINE database with respect to a topic of interest, in a reasonable amount of time such that it can be used in an article search query. The idea is that the learning procedure does not rely on discriminating whole MEDLINE abstracts, but on the words inside, which is much less computationally expensive. This is translated into a dictionary of scored words that can be used later to score any abstract according to the words it contains.
Because the approach is relatively inexpensive, we can evaluate different scoring schemes. We will discuss those and comment on how the performance of the approach is affected by the part of speech (e.g., noun, verb, or adjective) used for the analysis.
Results
The training set
The starting point of our algorithm is a set of articles associated (or believed to be associated) with a topic of interest. The system is trained with this set and therefore we define it as the training set. To ease evaluation of the method, we chose a subject for which the fraction of articles in the database would be neither too small nor too large of a subset of MEDLINE. In this work we used the topic stem cells and we took advantage of the annotation of MEDLINE entries with terms of the MeSH keyword hierarchy to select the training set. For this we obtained by license the complete MEDLINE database (November 2003 release, National Library of Medicine).
The MeSH vocabulary contains 22,568 descriptors, and 139,000 headings called Supplementary Concept Records. An average of 10 MeSH indexing terms are applied to each MEDLINE citation by NLM indexers, who after reading the full text of the article will choose the most specific MeSH heading(s) that describe the concepts discussed. The MeSH indexing terms are organized into concept hierarchies (directed acyclic graphs) that represent is-a and part-whole relationships [2]. Indexers can also assign subheadings to further describe a particular aspect of a MeSH concept. In addition to assigning MeSH terms that describe the topic of the article, the indexer provides terms that reflect the age group of the population studied, the nature of the studies (e.g. human vs. animal, male vs. female), and the material represented (Publication Types such as Clinical Trials, Editorial, Review). Thus MeSH headings serve as a telegraphic surrogate of the concepts contained in a journal article.
We selected all MEDLINE entries annotated with either the MeSH term stem cells or any of its 15 children terms in the MeSH keyword hierarchy (see the list in Table 1). The resulting set contained 81,416 articles with abstracts in MEDLINE (See additional file 1). Entries without abstracts were discarded because our training is based upon the words present in both the abstract and title of the reference. This training set of stem cell references represents ~0.5% of MEDLINE.
Keyword scoring
The property to be analysed is the frequency of certain words in abstracts. Based on our previous experience in the classification of abstracts, we registered the presence of words in the abstract (and title), ignoring the cardinality of words within a single abstract as done in [11], such that a word appearing many times within one abstract would not carry additional weight over a word appearing only once [12]. Additionally, we restricted the analysis to words which commonly convey meaning, that is, nouns, verbs, and adjectives, and not adverbs or conjunctions which would be more appropriate for style studies than for information extraction purposes [13].
Accordingly, we registered the frequencies of 100,196 unique nouns, 20,243 adjectives, and 7,970 verbs in all MEDLINE entries with abstracts (6,803,293 out of a total of 12,330,355 references from the year 1965 until November 2003) that appeared in at least 100 abstracts. Frequencies of 19,117 unique nouns, 6,452 adjectives, and 3,174 verbs were counted in the training set of stem cell references. To reduce noise we further filtered the list by considering only words that occurred in more than 100 of the 81,416 entries in the training set (2,256 nouns, 1,193 adjectives, and 748 verbs). Words were always distinguished by their part of speech. For example, of the combined set of 3,449 nouns and adjectives, 104 occurred in the literature both as a noun and as an adjective. Each noun was treated as a separate keyword from its adjective counterpart.
Each word (nouns, adjectives or verbs) occurring in 100 or more of the entries in the training set was scored by the ratio of frequency of occurrence in the training set divided by its frequency of occurrence in all of MEDLINE (see Methods). The top 30 scoring nouns, adjectives, and verbs are listed in Table 2, Table 3, and Table 4, respectively. (The complete lists are given in additional file 2). The set of nouns was much larger than the sets of adjectives or verbs. The top keyword scores were also higher for nouns than for adjectives and verbs.
Reference scoring
We studied two variables when scoring MEDLINE references on the basis of their word scores: one was the part(s) of speech used, and the other was the number of words used for the score. To make this analysis feasible in terms of computing time, we constructed a set of MEDLINE references with the training set and an additional equal number of references with abstracts chosen at random from the rest of MEDLINE, which we will call the random set, ideally not related to stem cells. The self-consistency of a given scoring scheme was measured by the fraction of references from the training set that ranked in the top half when the whole set of 162,832 references was scored. The scores for training and random sets are given as supplementary material (additional files 3 and 4, respectively).
We first analyzed the effect of scoring MEDLINE references using the average of all keywords in the abstract and title and compared the results to the average of only the top 5, and the top 10 words with the highest scores, which gave worse results and a small improvement, respectively (data not shown). For the rest of experiments we used the average of all words.
We then studied the influence of the part of speech used to score the references. Figure 1 shows the fraction of articles from the training set that was retrieved when selecting a variable range of top-scoring articles. Nouns were better keywords than were adjectives, or verbs. Using both nouns and adjectives as keywords slightly improved retrieval for the top-ranking articles, but weakened prediction of middle and low-ranking stem cell articles. Accordingly, we adopted as a scoring scheme the average over all nouns with scores.
We computed the scores for all MEDLINE references with abstracts. As expected, the MEDLINE score distribution agreed with the score distribution of the random set and was well below the score distribution of the training set (see Figure 2). However, the considerable overlap between the background and the training set was indicative that neither all references in the training set were dealing with stem cells in a strict sense nor all references in the random set were unrelated to stem cells.
Close inspection of the top ranking references from the random set revealed that they were also likely to be of interest to anybody wanting to read about stem cells (see Table 5 and Discussion). For these reasons, we measured performance of the method by observing recall and precision in a set evaluated by a human expert and not used to train the algorithm, a typical way to evaluate literature mining algorithms [14].
Recall and precision of the algorithm
We collected a test set of 6,923 MEDLINE entries randomly chosen from articles published during January 2004, and therefore not included in our training set. Their score distribution was in agreement with the MEDLINE background (Figure 2). According to a human evaluator with expertise in the field of stem-cell biology (MAA) there were 204 articles relevant to the topic of stem cells in the set, all with scores clearly above the background.
Figure 3 displays the recall and precision of the algorithm according to expanding thresholds in the scoring. The selection of the set of the 204 articles with best scores (roughly above a score of 2.14) retrieved 132 true positives and missed 72. For that set, the precision (fraction of selected articles that are true positives) and the recall (fraction of true positives selected) of the algorithm were of 65%.
The first false positive, PMID: 14707522 ranked at position 2, mentions the highly scoring keyword 'fibroblast' (see Table 2) but in the context of an inherited disease (glutaric aciduria type I). Similarly, other articles with high scores that were not considered to be relevant to stem cells by the human expert were usually talking of cells, genes, and proteins relevant to stem cells, but in a context not directly related to stem cell biology, such as cancer or metabolic disease. The worst scoring positive, PMID: 14702195 ranked at position 1910, was a review dealing with the use of neural stem cells for therapy of neurodegenerative diseases. The score was very low because its abstract does not contain any mention to relevant facts about stem cells.
This type of analysis is subjective because it reflects the prejudices of a particular human expert; however, it is indicative of the general agreement between human selection and automated ranking. The complete list of scored abstracts and the results of the human evaluation are available in Additional file 5.
Discussion
We have introduced a simple strategy to judge the relevance of a text according to a topic of interest based on a training set of text. The method relies on different frequencies of discriminating words between the training set and other non-relevant articles. This algorithm is appropriate for information extraction of molecular biology data from the MEDLINE database of scientific references.
Our analysis of more than six million MEDLINE entries with abstracts indicated that there were 128,409 unique keywords (100,196 nouns, 20,243 adjectives, 7,970 verbs) appearing in at least 100 abstracts. For comparison, the OED, the largest English-language dictionary, contains 290,000 entries with about 616,500 word forms [15]. OED omits many slang words, proper names, scientific and technical terms, and jargon (there are over a million named species of insects). Most estimates of the total vocabulary of English are well over three million words, but only ~200,000 words are still commonly used. An educated person has a vocabulary of ~20,000 words and uses ~2,000 per week through conversation.
To test the system, we constructed a set of references related to the topic of stem cells taking those annotated with the corresponding keywords of the MeSH hierarchy (see Methods). This set contained 81,416 MEDLINE references. There were 28,743 unique keywords (19,117 nouns, 6,452 adjectives, 3,174 verbs) extracted from the training set of 81,416 stem cell references.
We then focused on words that were used more often in the training set of stem cell references than elsewhere. Regarding those words, it was not surprising that a high proportion of the keywords extracted were proper names and scientific jargon. In order to be sure of choosing relevant words (and not those that could be present in the training set by pure chance) we took only those used in more than 100 references in the training set: only 2,256 (12%) of the nouns, 1,193 (18%) of the adjectives, and 748 (24%) of the verbs.
The words were scored by their different usage in stem cell references compared to MEDLINE, and all MEDLINE references with abstracts were ranked by the average of scores of their keywords (see Methods). The best keywords (mesoderm, fibroblast, foreskin, stem, mesenchyme) were mostly related to sources of stem cells and therefore were identifying relevant references. The worst keywords (hospital, care, health, practice, management) were totally off-topic and abstracts with many of these generic words would often rank poorly with respect to their relevance to stem cells.
The self-consistency analysis of the algorithm with a set combining the training set with an equally large set of randomly selected references was used to compare the performance of the algorithm for different parts of speech and simple scoring mechanisms. Nouns were found to be superior to verbs and adjectives and the average score of all nouns in the abstract and title was found to be most appropriate. We observed truly stem cell related articles in the random set of articles that were not annotated with stem cells MeSH terms, and also articles in the abstract set which were not relevant to the subject.
In order to further evaluate the capabilities of the method, we compared the results obtained with those returned by a human expert from a set of 6,923 articles not used for training. A precision of 65% was found for a recall level of 65% in the retrieval of the 204 articles deemed by the human to be relevant.
It would be interesting to see how the algorithm presented here performs when searching for different concepts such as stem cells. Evaluation of the self-consistency of the algorithm is relatively simple, so any user can have a good idea of whether there is enough information in the training set to allow distinction from the rest of the database and see how the part of the speech chosen affects performance. However, the least we can do here is to note that the part of speech that gave better performance were nouns. We propose two predictions. Firstly, the optimal part of speech could be related to the part of speech of the topic under consideration; in this case nouns are the best keywords because the topic is an object, stem cells; if the topic was a verb, such as interact or phosphorylate, we expect that a small number of verbs will work better. Our second prediction is that our algorithm will often work better using names as keywords, as it will be easier to discriminate topics composed of nouns or nouns and adjectives than bare adjectives or verbs. This is for the reason that nouns are used to name a person, place, thing, act, or concept, whereas adjectives indicate qualities of the nouns, and verbs tell of doing or being something. Therefore, context is often needed to determine the meaning of adjectives and verbs whereas nouns are relatively context-insensitive, especially in science. Most keywords used in molecular biology [2,16-18], are nouns which are sometimes complemented with an adjective, such as mitochondrial membrane.
Ideally biomedical texts should have a lower degree of linguistic variation than other genres [17]. However the naming conventions in biology and biomedicine are highly non-standardized even when it comes to the fundamental concepts. In theory, terms should be mono-referential (one-to-one correspondence between terms and concepts), but in practice we have to deal with ambiguities (i.e. homography – the same term corresponds to many concepts) and variants (i.e. synonymy – many terms lead to the same concept). One approach to solve the ambiguities of the natural text used in abstracts has been the indexing of the literature in the MEDLINE database by keywords drawn from the MeSH controlled terminology that was originally developed to categorize the citations contained in Index Medicus. The annotation of MEDLINE with MeSH terms at the National Library of Medicine helps users to link their search terms to abstracts containing different terms with the same meaning [18].
Annotation of articles with MeSH headings are optionally flagged with subheadings and importance markers (Major / Minor). However some applications might require a fuzzy association to subjects, for example, one reference can be more strongly relevant to stem cells than another. This could be important for example when setting up priorities between references. Another reference could be possibly relevant to stem cells with a low likelihood. This could matter if a researcher wanted to find out any possible relation of a gene to stem cells, even if it is a remote association. The approach presented in this work allows the ranking of any MEDLINE reference with respect to its relevance to a topic.
A different problem with MeSH terms particular to the subject of stem cells is that many references were annotated with stem cells MeSH terms because of the usage of stem cells as a technique. For example PMID: 15105256 is annotated with the MeSH term stem cells because mouse embryonic stem cells were used to raise chimeric mice using a method previously described, yet the major finding of the publication really has nothing to do with stem cells. Such an article would likely not be interesting to a researcher working on the biology of stem cells. As discussed previously [19], such information will be contained in the Methods section of the corresponding article and would often be omitted from the abstract. Thus our algorithm defeats this problem by using a different focus to avoid the imprecision caused by trusting MeSH annotations alone.
Regarding the computational time needed by our method, the extraction of specific parts of speech from MEDLINE requires several hours on a reasonably fast machine, but this only has to be done once. Newer entries added to the MEDLINE database can be parsed monthly or more frequently if desired. The main bottleneck is the production of a ranked list of MEDLINE references, which is not a problem if one is interested in only one concept such as stem cells. The real limitation arises if one considers using this strategy to the mining of ranked reference lists relating to many concepts. For each concept query, a set of training references must be collected, keyword scoring tables constructed, and all abstracts in MEDLINE must be scored. Providing a real-time interface for arbitrary concept queries is possible but would require some combination of the large storage requirements of pre-processed tables and cluster (or distributed, Beowulf) computing. A more realistic approach would be a local implementation of our approach according to the interests and requirements of individual researchers.
In our implementation, the training set was collected using a selection of MEDLINE references annotated with a subset of terms from the MeSH hierarchy that we considered to be relevant to the subject of stem cells. However, there are many other ways of selecting sets of MEDLINE references relevant to topics, such as links from databases like OMIM [10], HSSP [9], or by manual selection. The garbage in garbage out principle applies here as in many other applications where the quality of the training set matters, so if the selection is too messy the algorithm might not pick any relevant discriminating keywords.
To make our analysis as impartial and simple as possible, only MeSH terms in the sub tree of stem cell were considered, but there are other terms elsewhere in the MeSH hierarchy (e.g. Embryo Research) that would also be good indicators that a given article is talking about stem cells. It would be feasible to determine the nearest neighbours of an arbitrary MeSH term, and by setting a threshold similarity factor one could include all the MeSH terms within a certain semantic distance of one another in a clustering manner. Surely this would improve the performance of the relevance prediction algorithm. However, considering that we are in a stage of testing and illustrating the method, we employed a simple approach of using a MeSH term and its children in the MeSH hierarchy.
The cosine distance between vectors of word usage can be used to measure distance between MEDLINE abstracts [20]. However, this measure takes into account all the parts of speech, as well as the number of times each word is used in a body of text. The purpose of the cosine measure is to offer an objective distance between entries independent of the user's interest in a particular topic. Therefore our scoring is more appropriate, which is not a distance but rather an absolute value used to derive a ranking upon learning from a training set, a typical strategy in information retrieval [14]. Eventually, the cosine distance might be refined to use only certain parts of speech (such as nouns). We can assume this would give better results when searching MEDLINE neighbours of a given entry in MEDLINE, provided that the user is interested in topics similar to those contained in biological keyword systems.
Conclusion
This report describes an approach to compute a ranked list of publications according to relevance to a topic of interest, given a training set of MEDLINE references. It is evident that the analysis of the word usage in the abstracts of publications associated with a given concept can be used for literature mining. The strong dependency of the quality of the results with the part of speech used must be taken into consideration. Even if the procedure applied in this work may seem to be too simplistic given the existence of sophisticated methods such as naïve Bayesian classifiers, support vector machines, and neural networks, one should not forget that we are dealing with test sets of millions of abstracts, and training sets of tens of thousands, and that the variation of each single item to be classified is very large because they are composed of some hundred words. In situations like this, sophistication leads very quickly to impossibility of computation and pragmatic approaches are needed. We have produced a method that works and the conclusions obtained regarding the part of speech used may be useful for others working in information extraction from natural language.
Methods
The databases used were the December 2003 MEDLINE [1] and the 2004 MeSH keyword hierarchy [2]. The stem cell training set was selected by taking all references annotated with any MeSH term with a "TreeNumber" identifier of the type A11.872.x.y (for any x and y values).
All titles and abstracts in MEDLINE were processed using the Tree-Tagger part of speech parser [21] to extract separate lists of nouns, adjectives, or verbs, along with their frequency of occurrence. For each keyword found in some training set reference we computed the fraction of references in the training set using the keyword, and the fraction of references in the whole of MEDLINE using the keyword. Each keyword received a score which is the ratio of the frequency of usage in training set over the fraction of usage in the whole of MEDLINE. A score above one indicates that the word was used more often in the training set than in the rest of MEDLINE.
In order to remove irrelevant words associated to the training set by chance one can require that the words appear with a minimum frequency. We chose an absolute number of 100 times in our training set of 81,416 references (~0.1%).
We scored MEDLINE references based on the average score of all keywords in their abstract and title (this is the method used in XplorMed [22]). Words without a score (because they were present less than 100 times in the training set) were not taken into account. For comparison, scores of the top five, or the top ten keywords were also tested. The scoring was performed using nouns, adjectives, verbs, and nouns plus adjectives, as keywords.
Authors' contributions
BPS designed and tested the scripts that generated and benchmarked the lists of MEDLINE references ranked by their "stemness", and generated the figures of the manuscript. MAA provided advice and guidance, and evaluated the method's precision and recall. Both authors collaborated in the writing of the manuscript.
Supplementary Material
Additional File 1
Stem Cell referencesList in plain text format (stemcellpapers.txt) of 81,416 PubMed Identifiers (PMIDs) linked to abstracts in MEDLINE that have one or more MeSH terms which are members of the set of terms related to stem cell.
Click here for file
Additional File 2
Nouns, adjectives, and verb scores. A zip compressed file (file2.zip) containing three lists in plain text format (sortednounscores.txt, sortedadjectivescores.txt, sortedverbscores.txt) of the computed scores for 2,256 nouns, 1,193 adjectives, and 748 verbs.
Click here for file
Additional File 3
Stem Cell MEDLINE reference scoresA zip compressed file (file3.zip) containing lists in plain text format (stemcellpaperscores-adjectives.txt, stemcellpaperscores-nouns.txt, stemcellpaperscores-nounsadjectives.txt, stemcellpaperscores-verbs.txt) of 81,416 PMIDs of stem cell references and their scores according to nouns, adjectives, verbs, and combined nouns/adjectives.
Click here for file
Additional File 4
Subset of MEDLINE reference scoresA zip compressed file (file4.zip) containing lists in plain text format (paperscores-adjectives.txt, paperscores-nouns.txt, paperscores-nounsadjectives.txt, paperscores-verbs.txt) of 81,416 PMIDs of references randomly selected from MEDLINE and their scores according to nouns, adjectives, verbs, and combined nouns/adjectives.
Click here for file
Additional File 5
List of 6,923 scored abstractsA zip compressed file (file5.zip) containing a table in plain text format with tab separated columns (paperscores-nouns-recent.txt) of 6,923 PMIDs of references not included in the training set with their scores, and a human evaluation of their relevance to the topic of stem cells. Scripts are available on request. TreeTagger is available from [21].
Click here for file
Acknowledgements
Thanks to all the members of the Bioinformatics Group at the Ottawa Health Research Institute for countless helpful discussions, to the National Library of Medicine for the licensing and distributing MEDLINE, and to H. Schmidt (University of Stuttgart) for distributing TreeTagger.
This work is part of projects funded by the Ontario Genomics Institute, the Ontario Research and Development Challenge Fund, the Canadian Foundation for Innovation, the Ontario Innovation Trust, and the Canadian Stem Cell Network. MAA is the recipient of a Canada Research Chair.
Figures and Tables
Figure 1 Self-consistency test of the algorithm. Fraction of references from the stem cell training set (F) retrieved when selecting a number (N) of top-scoring references in a mixed set combining the training set and the random set. Nouns are better discriminators with F = 0.87 for the top half of the list. F was 0.79 for adjectives, 0.73 for verbs, and 0.70 for nouns plus adjectives. Performance could not be theoretically perfect because there were articles in the training set which were not relevant to stem cells, and there were articles in the random set which were relevant to stem cells.
Figure 2 Distribution of scores in MEDLINE sets. For each of the sets of MEDLINE references analyzed in this work we plot the distribution of score values (using the average over all nouns). The complete MEDLINE (black line with X's) has a maximum around 0.65. The training set composed of 81,416 references annotated with MeSH terms related to stem cells (magenta with diamonds) has a maximum at 2.75 and a "hump" at 1.5. This type of distribution is due to the fact that this set includes both references truly related to stem cells and others that are not and agree more with the general MEDLINE background distribution of scores. The random set of 81,416 references (red with triangles) has, logically, an identical distribution to the whole of MEDLINE. The 6,923 randomly selected MEDLINE references (green with squares) used for the recall and precision test also follow the background distribution. Of those, the 204 references evaluated as stem cell related by a human expert (blue bars) had significantly higher scores than the background distribution of MEDLINE.
Figure 3 Recall and precision of the algorithm. The recall and the precision of the algorithm were checked in a set of 6,923 references not included in the training set. Manual examination of the set resulted in the identification of 204 references (positives) relevant to stem cells. Recall was measured as TP/(TP+FN) and precision as TP/(TP+FP), where TP is true positives, FP is false positives, and FN is false negatives.
Table 1 MeSH keywords that are children of Stem Cell in the MeSH hierarchy
Fibroblasts
Colony-Forming Units Assay
Stem Cell Transplantation
Tumor Stem Cells
Erythroid Progenitor Cells
Myeloid Progenitor Cells
Myocytes, Cardiac
Myocytes, Smooth Muscle
Muscle Cells
Muscle Fibers
Satellite Cells, Skeletal Muscle
Totipotent Stem Cells
Multipotent Stem Cells
Pluripotent Stem Cells
Mesoderm
Table 2 Nouns that occur >100 times in stem cell MEDLINE references
Ranking Score Noun
1 43.8 mesoderm
2 37.7 fibroblast
3 33.6 Stem
4 29.6 foreskin
5 28.0 mesenchyme
6 27.0 progenitor
7 26.3 noggin
8 25.4 epiblast
9 24.8 endoderm
10 23.7 tenon
11 23.5 somite
12 23.3 Zellweger
13 23.0 gastrulation
14 22.7 notochord
15 22.2 ectoderm
16 22.0 XP
17 21.5 ES
18 20.6 xeroderma
19 19.5 Cockayne
20 18.9 myotome
21 18.7 gastrula
22 18.1 CFC
23 17.8 keloid
24 16.0 granulocyte-macrophage
25 15.7 haematopoiesis
26 15.5 blastula
27 15.3 Rous
28 15.0 Werner
29 14.7 BMP
30 14.4 stem
Table 3 Adjectives that occur >100 times in stem cell MEDLINE references
Ranking Score Adjective
1 51.3 embryoid
2 35.7 somital
3 32.4 mesodermal
4 30.5 totipotent
5 27.8 fibroblastic
6 19.4 mesenchymal
7 18.6 semisolid
8 18.5 immortal
9 17.1 haemopoietic
10 15.4 committed
11 14.4 ectodermal
12 14.3 vegetal
13 13.3 hematopoietic
14 13.2 senescent
15 12.2 diploid
16 12.1 dermal
17 12.0 endodermal
18 11.8 skinned
19 11.7 confluent
20 11.3 sonic
21 11.2 haematopoietic
22 10.6 myogenic
23 10.5 erythropoietic
24 10.5 embryonic
25 10.4 quiescent
26 10.4 morphogenetic
27 9.2 inductive
28 9.1 primitive
29 9.1 dorsoventral
30 8.7 unscheduled
Table 4 Verbs that occur >100 times in stem cell MEDLINE references
Ranking Score Verb
1 11.6 immortalize
2 10.9 plait
3 10.6 engraft
4 10.5 skin
5 9.7 purge
6 9.2 seed
7 9.2 subculture
8 8.5 reprogram
9 8.3 rejoin
10 8.2 passage
11 7.6 stem
12 7.4 recapitulate
13 7.2 nucleate
14 7.0 proliferate
15 6.8 culture
16 6.6 condition
17 6.4 transform
18 6.2 ruffle
19 6.1 deregulate
20 6.1 explant
21 5.4 populate
22 5.4 rescue
23 5.3 wound
24 5.2 round
25 5.0 migrate
26 4.9 cultivate
27 4.7 sort
28 4.3 transplant
29 4.5 differentiate
30 4.3 internalize
Table 5 High scoring references not annotated with stem cell MeSH terms.
Ranking1 PMID Title
139 9811585 Hematopoietic induction and respecification of A-P identity by visceral endoderm signaling in the mouse embryo.
160 8714368 The role of fibroblast growth factor-2 (FGF-2) in hematopoiesis.
174 11672504 Molecular regulation of embryonic hematopoiesis and vascular development: a novel pathway.
177 8805699 Positional cloning of a global regulator of anterior-posterior patterning in mice
404 2910353 Dual role of fibronectin in hematopoietic differentiation.
426 10441547 Regulative development of the sea urchin embryo: signalling cascades and morphogen gradients.
638 11730936 The dynamics of bone marrow stromal cells in the proliferation of multipotent hematopoietic progenitors by substance P: an understanding of the effects of a neurotransmitter on the differentiating hematopoietic stem cell.
720 3659868 Early cardiogenesis in the newt embryo.
750 79573 Calcium-binding protein of the chick chorioallantoic membrane. II. Vitamin K-dependent expression.
801 7538068 A conserved enhancer of the human and murine Hoxa-7 gene specifies the anterior boundary of expression during embryonal development.
1 Rank was assigned by computing the average score of all the nouns present in the abstract and title of an article, and comparing this score with that of other articles in the merged list of 162,832 references. The merged set was constructed from 81,416 references randomly selected from MEDLINE combined with 81,416 references that are annotated with the MeSH term stem cells or one of its children in the MeSH hierarchy.
==== Refs
NLM MEDLINE 2004
NLM Medical Subject Headings (MeSH) 2004
Mitchell TM Machine Learning 1997 Boston, WCB/McGraw-Hill
Yang Y Liu X A re-examination of text categorization methods Annual ACM Conference on Research and Development in Information Retrieval 1999 Berkeley, CA, ACM Press 42 49
Kim W Aronson AR Wilbur WJ Automatic MeSH term assignment and quality assessment Proc AMIA Symp 2001 319 323 11825203
Wilbur WJ Boosting naive Bayesian learning on a large subset of MEDLINE Proc AMIA Symp 2000 918 922 11080018
Forman G Cohen I Learning from Little: Comparisons of Classifiers Given Little Training: 4 AD/9/20; Pisa, Italy. 2004
Salton G Automatic Text Processing -- The Transformation, Analysis, and Retrieval of Information by Computer 1989 Addison-Wesley
Andrade MA Valencia A Automatic extraction of keywords from scientific text: application to the knowledge domain of protein families Bioinformatics 1998 14 600 607 9730925 10.1093/bioinformatics/14.7.600
Andrade MA Bork P Automated extraction of information in molecular biology FEBS Lett 2000 476 12 17 10878241 10.1016/S0014-5793(00)01661-6
Perez-Iratxeta C Keer HS Bork P Andrade MA Computing fuzzy associations for the analysis of biological literature Biotechniques 2002 32 1380 1385 12074170
Dumais ST Platt J Heckerman D Sahami M Inductive learning algorithms and representations for text categorization. In CIKM-98: Proceedings of the Seventh International Conference on Information and Knowledge Management 1998
Netzel R Perez-Iratxeta C Bork P Andrade MA The way we write EMBO Rep 2003 4 446 451 12728240 10.1038/sj.embor.embor833
Manning C Schütze H Foundations of Statistical Natural Language Processing 1999 Cambridge, MA, MIT Press
Simpson JA and Weiner ESC Oxford English Dictionary 1989 2 Oxford University Press
Harris MA Clark J Ireland A Lomax J Ashburner M Foulger R Eilbeck K Lewis S Marshall B Mungall C Richter J Rubin GM Blake JA Bult C Dolan M Drabkin H Eppig JT Hill DP Ni L Ringwald M Balakrishnan R Cherry JM Christie KR Costanzo MC Dwight SS Engel S Fisk DG Hirschman JE Hong EL Nash RS Sethuraman A Theesfeld CL Botstein D Dolinski K Feierbach B Berardini T Mundodi S Rhee SY Apweiler R Barrell D Camon E Dimmer E Lee V Chisholm R Gaudet P Kibbe W Kishore R Schwarz EM Sternberg P Gwinn M Hannick L Wortman J Berriman M Wood V de la CN Tonellato P Jaiswal P Seigfried T White R The Gene Ontology (GO) database and informatics resource Nucleic Acids Res 2004 32 Database issue D258 D261 14681407
Hahn U Romacker M Schulz S Creating knowledge repositories from biomedical reports: the MEDSYNDIKATE text mining system Pac Symp Biocomput 2002 338 349 11928488
Barnes JC Conceptual biology: a semantic issue and more Nature 2002 417 587 588 12050632 10.1038/417587b
Shah PK Perez-Iratxeta C Bork P Andrade MA Information extraction from full text scientific articles: where are the keywords? BMC Bioinformatics 2003 4 20 12775220 10.1186/1471-2105-4-20
Wilbur WJ Yang Y An analysis of statistical term strength and its use in the indexing and retrieval of molecular biology texts Comput Biol Med 1996 26 209 222 8725772 10.1016/0010-4825(95)00055-0
Institut fuer maschinelle Sprachverarbeitung US Tree-Tagger 2004
Perez-Iratxeta C Bork P Andrade MA XplorMed: a tool for exploring MEDLINE abstracts Trends Biochem Sci 2001 26 573 575 11551795 10.1016/S0968-0004(01)01926-0
|
15790421
|
PMC1274266
|
CC BY
|
2021-01-04 16:02:48
|
no
|
BMC Bioinformatics. 2005 Mar 24; 6:75
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-75
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-781579677710.1186/1471-2105-6-78SoftwareCGMIM: Automated text-mining of Online Mendelian Inheritance in Man (OMIM) to identify genetically-associated cancers and candidate genes Bajdik Chris D [email protected] Byron [email protected] Shawn [email protected] Steven [email protected] Angela [email protected] Cancer Control Research Program, BC Cancer Agency, 600 West 10th Avenue, Vancouver BC, V5Z 4E6, Canada2 Genome Sciences Centre, BC Cancer Agency, 600 West 10th Avenue, Vancouver BC, V5Z 4E6, Canada2005 29 3 2005 6 78 78 29 10 2004 29 3 2005 Copyright © 2005 Bajdik et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Online Mendelian Inheritance in Man (OMIM) is a computerized database of information about genes and heritable traits in human populations, based on information reported in the scientific literature. Our objective was to establish an automated text-mining system for OMIM that will identify genetically-related cancers and cancer-related genes. We developed the computer program CGMIM to search for entries in OMIM that are related to one or more cancer types. We performed manual searches of OMIM to verify the program results.
Results
In the OMIM database on September 30, 2004, CGMIM identified 1943 genes related to cancer. BRCA2 (OMIM *164757), BRAF (OMIM *164757) and CDKN2A (OMIM *600160) were each related to 14 types of cancer. There were 45 genes related to cancer of the esophagus, 121 genes related to cancer of the stomach, and 21 genes related to both. Analysis of CGMIM results indicate that fewer than three gene entries in OMIM should mention both, and the more than seven-fold discrepancy suggests cancers of the esophagus and stomach are more genetically related than current literature suggests.
Conclusion
CGMIM identifies genetically-related cancers and cancer-related genes. In several ways, cancers with shared genetic etiology are anticipated to lead to further etiologic hypotheses and advances regarding environmental agents. CGMIM results are posted monthly and the source code can be obtained free of charge from the BC Cancer Research Centre website .
==== Body
Background
Cancers are complex diseases with multiple genetic and environmental factors contributing to their development. The most prominent success stories in cancer genetics to date have involved genes that produce a recognizable pattern of disease within certain rare families. Most cancers, however, are sporadic and appear in people who do not have a clear family history of the disease. These cancers are currently being studied in epidemiological investigations that examine genetics, environmental exposures or both. The studies often compare "cases" or affected individuals to "controls" or unaffected individuals, to determine which group has a higher frequency of a particular gene variant or a greater level of exposure to an environmental agent. The studies require logical hypotheses regarding the genes to be tested and clear criteria for case definition. Cases may be defined as people who have any of several types of cancer, if those types are related. For example, epidemiologic studies of BRCA1 mutation carriers might benefit from information collected about both breast and ovarian cancer cases. But what genes are associated with a group of cancers, and what cancers are associated with a particular gene? The answers can be found in literature regarding cancer genetics, microbiology, clinical medicine, epidemiology and other sciences. More than 1% of all human genes are associated with cancer [1] and information about the association between genes and cancer changes constantly.
Online Mendelian Inheritance in Man (OMIM; ) is a computerized database of information about genes and heritable traits in human populations. The database was created by Victor McKusick at Johns Hopkins University and is now edited by him and colleagues around the world.[2] We consider it a particularly high-quality data source because it is curated by a knowledgeable team, based on information reported in the scientific literature, and continuously updated. OMIM is maintained on the Internet by the National Center for Biotechnology Information at the US National Institutes of Health.[3] Data mining aims to discover unexpected trends and patterns from large sets of data [4], and the rapid growth of biomedical literature underscores the value of text-mining in particular. Text-mining has been described as a modular process involving document categorization, named entity tagging, fact and information extraction, and collection-wide analysis.[5] In document categorization, a subset of potentially relevant documents is retrieved to increase the efficiency of subsequent steps. Named entity tagging identifies the important entities or objects mentioned in the article, often using a list of synonyms. Fact and information extraction identifies the relationships between entities. Finally, in collection-wide analysis, information extracted from different documents is integrated.
Many research studies aim to explore the association between genes and cancer. The design of these studies requires the identification of appropriate patient groups and candidate genes, and both steps can benefit from effective text-mining of public data sources. OMIM is a high-quality information source and considered a key reference database by the genetics community. Our objective was to establish an automated text-mining system for OMIM that will identify genetically-related cancers and cancer-related genes.
Implementation
We developed the computer program CGMIM to text-mine OMIM. The software considers 21 major cancer types identified by the National Cancer Institute of Canada [ref [6], page 18, Table 1]. CGMIM recognizes genetically-related cancers by identifying cancer types mentioned in association with a specific gene. For pairs of cancer types, CGMIM generates a table with rows and columns for each cancer type, and cells containing the number of OMIM gene entries that mention an association with those cancers. We refer to this table as the siteXsite matrix. If several OMIM entries mention one type of cancer, and several entries mention another type of cancer, then some entries will mention both types of cancer by chance alone. If the mention of different cancers occurred at random, the expected number of genes (E) in OMIM that mention two specific types of cancer can be estimated as the total number of genes related to cancer, multiplied by the probabilities that an entry mentions each individual cancer type. The latter probabilities are estimated as the proportion of genes in OMIM that are related to each cancer type. Explicitly, if there are N genes related to cancer, GA genes related to cancer type A and GB genes related to cancer type B, then
EAB = (GA/N) × (GB/N) × N (1)
where EAB is the expected number of genes related to both cancer types A and B. The observed number of genes (O) is the number of OMIM entries that mention both cancer types, and O/E indicates whether the number of genes associated with a pair of cancer sites is different than chance alone would predict. An O/E value of 1.0 indicates the number of entries observed is the number expected by chance. An approximate 95% confidence interval (95% CI) is O/E ± (1.96/√E).
Our text-mining algorithm begins by separating paragraphs of an OMIM entry into constituent sentences, and assumes sentences end with a period followed by a space. There are many words and phrases that refer to cancer. A breast cancer might be described as a breast tumor, breast carcinoma or mammary gland neoplasm. A list of synonyms for each cancer type was developed using the International Classification of Disease for Oncology (ICD-O) [7] and augmented by familiar lay terminology. Other variation occurs as the result of English grammar. Breast cancer might be referred to as cancer of the breast, and several cancers might be referred to in a list (e.g., "cancer of the ovary, breast, and skin"). The algorithm identified OMIM entries for each type of cancer by finding sentences that included both a site synonym and a cancer synonym. For phrases in the synonym list, CGMIM searched for sentences containing all of the individual words.
"Stemming" was used to remove capitalization and common suffixes from words, and thereby changes similar words to identical word fragments. The process is best demonstrated with an example.
Unstemmed
Large-cell lymphomas comprise approximately 25% of all non-Hodgkin lymphomas in children and young adults, and approximately one-third of these tumors have a t(2;5)(p23;q35) translocation.
Stemmed
larg-cell lymphoma compris approxim 25% of all non-hodgkin lymphoma in children and young adult, and approxim on-third of these tumor have a t(2;5)(p23;q35) transloc
We used an established algorithm ("Porter's algorithm") to perform the stemming.[8]
Our list of synonyms was stemmed and then compared to the stemmed sentences in OMIM. An OMIM entry may contain alternative entry names, mapping information, a text summary, references to key publications, examples of known allelic variants, and a clinical synopsis of the corresponding phenotype. Some of these fields are subjective, such as the examples of allelic variants, and we restricted our search to the text summary.
Finally, not all OMIM entries refer to specific genes. Some entries refer to heritable traits for which no gene has been identified. In addition, more than one OMIM entry can refer to the same gene. This typically occurs when the entry for a trait is linked to a gene that was previously identified and described in a separate OMIM entry. Because OMIM is dynamically organized and updated, this type of multiple referencing is unavoidable. To restrict searches to only the OMIM entries for genes, CGMIM compares each entry name and alternative names with a list of gene names assigned by the Human Genome Organization (HUGO; ).
We performed manual searches of the OMIM database to identify the strengths and weakness of the computerized search method, and to iteratively modify the software. This involved selecting a sample of OMIM entries and reading through the text to determine whether the entries referred to a cancer, or if entries were identified by CGMIM where, in reality, there was no true cancer reference. We also reviewed the entries to identify sentences that referred to cancer, but for which evidence indicated there was no association. (E.g., "An early study showed the gene was not related to breast cancer.") While an OMIM entry might include a sentence of that sort, another sentence in the entry might cite evidence supporting the association. (E.g., "A subsequent study showed the gene was related to breast cancer.") Despite the negative statement, this example OMIM entry mentions evidence supporting the association and hence would be included when tallying entries associated with the cancer.
CGMIM was written in the Perl computer language and implemented on a Linux workstation. OMIM is updated daily and we created static copies of the database to provide a stable reference for search evaluation. The copies of OMIM used to develop CGMIM were downloaded between March and October of 2003, and each copy contained more than 14,000 entries.
Results and discussion
In the OMIM database on September 30, 2004, CGMIM identified 1943 genes related to cancer. BRCA2 (OMIM *164757), BRAF (OMIM *164757) and CDKN2A (OMIM *600160) were each related to 14 types of cancer. The OMIM entries for all three genes mention leukemia, melanoma, breast cancer, colorectal cancer, pancreatic cancer, stomach cancer, ovarian cancer and prostate cancer. The entry for BRCA2 also mentions cancer of the brain, larynx, cervix, uterus, thyroid and kidney. The entry for BRAF also mentions lymphoma and cancer of the lung, bladder, testes, cervix and uterus. The entry for CDKN2A also mentions lymphoma and cancer of the lung, bladder, brain, esophagus and kidney. Each gene defines a large group of related cancers.
The numbers of genes associated with each pair of cancer types are summarized in the siteXsite matrix (Figure 1). Diagonal cells in the matrix contain the total numbers of genes identified for each cancer type; off-diagonal cells are the numbers of genes identified by both the row and the column titles. For example, there were 45 genes related to cancer of the esophagus, 121 genes related to cancer of the stomach, and 21 genes related to both. The cancer mentioned by the greatest number of OMIM entries was leukemia, and the greatest number of OMIM gene entries that mention a combination of two cancers was 143 for lymphoma and leukemia. For some pairs of cancer sites, no genes were identified.
The numbers in the off-diagonal cells depend on the number of genes related to the individual cancers. Based on the number of OMIM entries that mention leukemia and lymphoma individually, the number expected to mention both is 98.3 and the ratio of the observed and expected values is 1.5 (95% CI 1.3–1.7). (In equation (1), GLEUKEMIA = 643, GLYMPHOMA = 297 and N = 1943.) This indicates there are 50% more genes related to both cancers than would be expected by chance. Table 1 provides a list of 20 pairs of cancer types where the ratio of the observed and expected number of genes in the siteXsite matrix is greatest. The table indicates that fewer than three genes in OMIM should mention both cancer of the esophagus and cancer of the stomach by chance, but 21 entries mention both cancers. This more than seven-fold discrepancy suggests that cancers of the esophagus and stomach might be more related than current literature suggests. Similar conclusions might be made for the other pairs of cancer types in Table 1.
We randomly selected 25 genes related to cancer and manually reviewed text of the corresponding OMIM entries. All of the entries correctly mention one or more types of cancer, but for 20% of those entries, one of the cancers was only mentioned in the context of evidence suggesting no association.
CGMIM can assist in designing effective studies of genetically-related cancers. CGMIM uses a high-quality database of genetic information to produce a summary of gene and cancer associations. A group of cancer types might be related by physical proximity in the body (e.g., prostate and bladder cancer), a shared physiologic function (e.g., cancers involving the digestive tract), a common exposure (e.g., cancers caused by air pollution) or a common genetic characteristic (e.g., cancers in tissues that express BRCA1). The identification of such groups becomes more difficult and time-consuming as the literature about genes and cancer expands, and efficient text-mining tools have increasing value.
In several ways, groups of cancers that have shared genetic factors are anticipated to lead to further etiologic hypotheses and advances regarding environmental agents. First, grouping cancers will be especially useful if a group combines several cancers that are rare and difficult to study individually. Second, knowledge of genetic pathways might suggest an environmental factor associated with all of the cancers. For example, a grouping defined by a vitamin receptor gene would suggest vitamin intake as a possible environmental agent in the etiology of all of the cancers. Third, CGMIM will allow us to design studies that might extend gene-cancer associations to include cancers at other sites. The groups can also be used to identify cancers that should be considered together in a definition of family history, and in selection of genetic tests that might be adopted for high-risk families. During development of CGMIM, we observed changes in OMIM and the cancer groups that it produced from one week to another. This illustrates the need for a tool that can routinely perform the analysis, as opposed to a set of results based on the OMIM contents from a particular day.
OMIM is based on published material from the scientific literature. The number of genes identified by our program does not necessarily indicate the relatedness of two or more cancer types, but rather what is known about those cancers. This reflects what research has been funded, performed and published. There is more funding for certain types of cancer, there are more journals that address certain types of cancer, and there are more people studying certain types of cancer. Published information reflects our knowledge base and the scientific literature is hence a valid basis for identifying cancer groups and genes for further study. In some cases, evidence about an association was based on studies of cell lines or non-human organisms. In other cases, evidence was based on anecdotal observations in a small number of people. Some associations were based on several independent studies that each involved hundreds of patients.
There are sentences in OMIM that contain phrases such as "is not related to breast cancer". We could not create an algorithm that recognized all negative references without overlooking positive valid ones. Some OMIM entries report both negative and positive evidence of an association. These "mixed" entries are tallied as positive reports by CGMIM, consistent with our interest in positive associations. Other sentences in OMIM describe evidence of gene expression in both cancerous and normal tissue. E.g., "... has been shown to be expressed in breast cancer cells and prostate cells". The sentences are incorrectly interpreted as mentions of prostate cancer. Manual review of OMIM indicated that a minority of apparent associations (about 20%) between a gene and specific type of cancer were the result of negative evidence and are thus "false-positive" text-mining associations. We suggest that a manual review of OMIM associations always precede subsequent study design and analysis. We assume the excess 20% is included in every cell of the siteXsite matrix. Thus expected values also include the 20% excess, and the O/E ratios are not affected.
Other databases might be used as the basis for assessing scientific knowledge regarding genetic cancer groupings, but OMIM offers several advantages. OMIM is based on all publications in the PubMed database that are related to a specific human gene or trait. Results based on mining all of PubMed would be of interest, but would involve a much larger volume of literature and lack the expert review that is characteristic of OMIM. More specialized cancer groupings also might be created using computerized conference proceedings or journal contents. Likewise, a list of synonyms might be determined from other sources such as the UMLS (Unified Medical Language System) Specialist Lexicon of the National Cancer Institute. We used ICD-O terminology because it is the basis for most scientific writing on cancer.
This project used resources that have been developed by the US National Institutes of Health and Human Genome Project.[3] Our approach is exhaustive of the information reported in OMIM, will produce a computer algorithm for near-automatic updating of the review, and has the potential to be extended to other computerized databases. We will use CGMIM along with other criteria to guide the design of studies of genes and environment in cancer etiology.
Conclusion
CGMIM uses an expert database of genetic information to determine a summary of gene and cancer associations. The software identifies genes that are associated with a particular type of cancer, groups of cancers that share a common genetic association, and pairs of cancer types where there are more related genes than expected by chance.
Availability and requirements
• Project name: CGMIM
• Project home page:
• Operating system: The source code for CGMIM can be downloaded from the CGMIM homepage and run under Linux.
• Programming language: The source code for CGMIM can be downloaded from the CGMIM homepage and is written in Perl.
Abbreviations
OMIM is Online Mendelian Inheritance in Man; HUGO is the Human Genome Organisation; ICD-O is the International Classification of Disease for Oncology
Authors' contributions
The software was developed by BK and SR under the direction of SJ and CDB. The website and manuscript were created by CDB and AB. Funding for the project was obtained by CDB, AB and SJ.
Acknowledgements
Chris Bajdik and Steven Jones are scholars of the Michael Smith Foundation for Health Research. This work was supported by a research grant from the Canadian Cancer Etiology Research Network. Steve Sung assisted with design of the CGMIM website and Chris Young assisted with manual searches of the OMIM database.
Figures and Tables
Figure 1 A siteXsite matrix for 21 major cancer types as reported by the National Cancer Institute of Canada. Matrix cells indicate the number of genes related to cancers named in the row and column labels. Cell entries are based on cancers mentioned in Online Mendelian Inheritance in Man (OMIM; ) searched September 30, 2004.
Table 1 The twenty pairs of cancer types with the highest ratio of observed (O) to expected (E) number of associated genes. The O/E ratio and a 95% confidence interval (95%CI) are provided. Results are based on cancers mentioned in Online Mendelian Inheritance in Man (OMIM; ) searched September 30, 2004.
Pair of Cancer Types Number of Genes Related to Both O/E Ratio and 95%CI
Observed (O) Expected (E)
cervix – larynx 2 0.16 12.6 ± 4.9
larynx – mouth 2 0.17 11.6 ± 4.8
larynx – uterus 2 0.25 7.9 ± 3.9
esophagus – stomach 21 2.80 7.5 ± 1.2
larynx – stomach 3 0.44 6.9 ± 3.0
bladder – esophagus 10 1.51 6.6 ± 1.6
larynx – myeloma 1 0.15 6.6 ± 5.1
larynx – esophagus 1 0.16 6.2 ± 4.9
cervix – esophagus 6 1.02 5.9 ± 1.9
bladder – cervix 8 1.47 5.4 ± 1.6
cervix – uterus 8 1.59 5.1 ± 1.6
larynx – bladder 1 0.23 4.3 ± 4.1
pancreas – stomach 23 6.10 3.8 ± 0.8
cervix – stomach 10 2.74 3.7 ± 1.2
esophagus – mouth 4 1.11 3.6 ± 1.9
brain – kidney 19 5.39 3.5 ± 0.8
bladder – testis 18 5.12 3.5 ± 0.9
bladder – prostate 17 5.05 3.4 ± 0.9
bladder – pancreas 11 3.28 3.4 ± 1.1
cervix – lung 20 6.14 3.3 ± 0.8
==== Refs
Futreal PA Coin L Marshall M Down T Hubbard T Wooster R Rahman N Stratton MR A census of human cancer genes Nat Rev Cancer 2004 4 177 183 14993899 10.1038/nrc1299
Hamosh A Scott AF Amberger J Bocchini C Valle D McKusick VA Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders Nucleic Acids Research 2002 30 52 55 11752252 10.1093/nar/30.1.52
Wheeler DL Church DM Edgar R Federhen S Helmberg W Madden TL Pontius JU Schuler GD Schrimi LM Sequeira E Suzek TO Tatusova TA Wagner L Database resources of the National Center for Biotechnology Information: update Nucleic Acids Research 2004 32 D35 40 14681353 10.1093/nar/gkh073
Han J Kamber M Data Mining: Concepts and Techniques 2001 First Morgan Kaufmann Publishers
de Bruin B Martin J Getting to the (c)ore of knowledge: mining biomedical literature Int J Medical Informatics 2002 67 7 18 10.1016/S1386-5056(02)00050-3
National Cancer Institute of Canada Canadian Cancer Statistics Toronto 2004
Fritz A Percy C Jack A Shanmugaratnam K Sobin L Parkin DM Whelan S International Classification of Diseases for Oncology 2000 Third World Health Organization
Porter MF An algorithm for suffix stripping Program 1980 14 130 137 It has since been reprinted in Sparck Jones, Karen, and Peter Willet (1997) Readings in Information Retrieval San Francisco. Morgan Kaufmann
|
15796777
|
PMC1274267
|
CC BY
|
2021-01-04 16:02:49
|
no
|
BMC Bioinformatics. 2005 Mar 29; 6:78
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-78
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-851580790410.1186/1471-2105-6-85SoftwareCGHPRO – A comprehensive data analysis tool for array CGH Chen Wei [email protected] Fikret [email protected] H-Hilger [email protected] Steffen [email protected] Reinhard [email protected] Max-Planck Institute for Molecular Genetics, Ihnestrasse 73, 14195 Berlin, Germany2005 5 4 2005 6 85 85 18 11 2004 5 4 2005 Copyright © 2005 Chen et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Array CGH (Comparative Genomic Hybridisation) is a molecular cytogenetic technique for the genome wide detection of chromosomal imbalances. It is based on the co-hybridisation of differentially labelled test and reference DNA onto arrays of genomic BAC clones, cDNAs or oligonucleotides, and after correction for various intervening variables, loss or gain in the test DNA can be indicated from spots showing aberrant signal intensity ratios.
Now that this technique is no longer confined to highly specialized laboratories and is entering the realm of clinical application, there is a need for a user-friendly software package that facilitates estimates of DNA dosage from raw signal intensities obtained by array CGH experiments, and which does not depend on a sophisticated computational environment.
Results
We have developed a user-friendly and versatile tool for the normalization, visualization, breakpoint detection and comparative analysis of array-CGH data. CGHPRO is a stand-alone JAVA application that guides the user through the whole process of data analysis. The import option for image analysis data covers several data formats, but users can also customize their own data formats. Several graphical representation tools assist in the selection of the appropriate normalization method. Intensity ratios of each clone can be plotted in a size-dependent manner along the chromosome ideograms. The interactive graphical interface offers the chance to explore the characteristics of each clone, such as the involvement of the clones sequence in segmental duplications. Circular Binary Segmentation and unsupervised Hidden Markov Model algorithms facilitate objective detection of chromosomal breakpoints. The storage of all essential data in a back-end database allows the simultaneously comparative analysis of different cases. The various display options facilitate also the definition of shortest regions of overlap and simplify the identification of odd clones.
Conclusion
CGHPRO is a comprehensive and easy-to-use data analysis tool for array CGH. Since all of its features are available offline, CGHPRO may be especially suitable in situations where protection of sensitive patient data is an issue. It is distributed under GNU GPL licence and runs on Linux and Windows.
==== Body
Background
DNA sequence copy number changes have shown to play an important role in the aetiology of cancer and congenital disorders. Comparative Genomic Hybridization (CGH) is a molecular cytogenetic method for the detection of chromosomal imbalances [1], which does not depend on the availability of chromosome spreads and is not confined to the analysis of growing cells. Unfortunately, conventional chromosomal CGH has a low resolution. Recently, this drawback has been overcome by the introduction of array CGH. Here differentially labelled test and reference DNA are co-hybridized onto microarrays of several thousand evenly spaced DNA clones or oligonucleotides representing specific regions of the human genome[2,3]. The resolution of this technique depends on the number of different DNA spots printed on the glass slides and on the size of the DNA clones used. Currently most available CGH arrays allow the reliable detection of deletions and duplications if they are larger than 1 Mb. However, with arrays containing more densely spaced BAC (Bacterial Artificial Chromosomes) clones containing on average 150 kb of human DNA, much higher resolutions can be achieved [4]. Up to now, array CGH has been predominately used in highly specialized laboratories, and most of the data analysis programs currently available are not able to process the output of array CGH experiments in an easy and comprehensive way. For example, the two R packages, aCGH and DNAcopy, can identify copy number transitions on chromosomes by Unsupervised Hidden Markov Model [5] and Circular Binary Segmentation [6], but the application of these tools requires basic programming skills in R language. CGH-Plotter is a MATLAB toolbox with a graphic user interface. It detects the regions of amplifications and deletions using k-means clustering and dynamic programming. However, like aCGH and DNAcopy, CGH-Plotter can only be used to analyse already normalized array data in a specific format [7]. In addition, these programs output display the results in a non-interactive plot. As a visualization tool for array CGH, SeeGH can display the data in a user friendly interface[8]. It allows users to explore the results in a conventional karyotype diagram with annotation. However, without the essential statistical methods for characterizing the genomic profile, seeGH is rather a visualization than an analysis tool for array CGH.
Here we present a comprehensive data analysis software for array CGH. The program combines analysis and visualization of array CGH data. Furthermore it supports comparative analysis of complete study groups based on absolute and relative frequencies of aberrations.
Implementation
Software design and information sources
CGHPRO was programmed in Java and MySQL was used as the back-enddatabase. The decision to use Java and MySQL was based on their public availability, their platform independence and the fact that MySQL can handle large data files with high throughput. Two "R" packages from Bioconductor[9], DNAcopy and aCGH, were implemented in our software, which enable a platform-independent characterization of genomic profiles. Up to now, CGHPRO has been tested in a Linux and a Windows 2000 environment.
To date, CGHPRO allows the import of result files from GenepixPro5.0, Agilent and Imagene, but users can also customize the program to support their own data format. All essential data required to meet future standards of "minimal information about an array CGH experiment" can be stored in the database. Such standards have already been defined for gene expression analysis [10], but are still lacking for array CGH.
The genome annotation that has been integrated into the current version of CGHPRO is based on the UCSC Genome Browser [11], but users can choose different versions to meet their own needs.
The example chosen for demonstration of this software is a male versus female co-hybridization onto a 14000 BAC array, comprising a genome wide 1 Mb resolution BAC array (clones kindly provided by Nigel Carter, Wellcome Trust Sanger Centre, [12]) and the tiling path of nine chromosomes from the Human "32 k" BAC Re-Array set, a series of overlapping BAC clones obtained from BACPAC Resources Centre at Children's Hospital Oakland Research Institute [13,14]. Detailed protocols for the generation of the arrays and the hybridization are available at our website [15].
The dataset for the comparative analysis was generated artificially.
Results and discussion
Database design
In CGHPRO a back-end database using MySQL has been implemented. The database stores the description of each analyzed chip (glass slide) as an entry in the table 'analyzedChips'. This description includes all essential information about the experimental and data analysis procedure, e.g. the number of spots that have been excluded and the normalization algorithm applied. A separate table named according to the Chip ID saves the original data from the image analysis software as well as the results from data analysis. A table called 'clonePosition' is used to store the mapping information for each clone. The information comprises data that might influence the reliability of the clone's hybridization characteristics, such as content of repetitive sequences and most importantly, its involvement in segmental duplications, which can be visualized by a colour code, as discussed below. In order to be able to use this feature properly, users with cDNA- or oligo-arrays have to adjust the chromosome positions of their probes. Instead of the actual chromosome position, one should define the chromosomal region, which the probe represents.
Data input
CGHPRO allows the import of output files from several image analysis software packages, as listed above. Essential data are extracted and spots flagged as "poor" by the image analysis software are excluded automatically. Mapping information and related annotations for each clone are fetched from the back-end database.
Mapping information for each clone, based on a specific version of UCSC Genome Browser, has to be provided by the user. For this purpose, a tab-delimited file has to be loaded into the back-end database, which must include six fields for each clone, the unique identifier, the respective chromosome, the positions of the first and last base pair, the source of the clone, and the user-specified comments of the clone. For the complete tiling path, as distributed by the BACPAC Resources Centre, the mapping information based on the April 2003 assembly of UCSC Genome Browser comes along with the software.
The way this information is acquired differs from other recently published programs like ArrayCGHbase [16] or CAP [17], both of which provide these data by directly accessing the respective genome browser. This may be an advantage when looking for the most recent update, but it may pose problems for diagnostic and related applications, where patient confidentiality is important and precludes online data analysis. Offline analysis also speeds up the process, as it is not dependent on server capacities or transfer rates.
Graphical analysis of hybridization characteristics
CGHPRO provides a variety of graphical data representation tools to visualize the data before and after normalization. Scatterplots allow for estimates of the noise within a given data set. Spatial dependency of log ratios can be detected with Boxplots for all different subgrids. MAplots are implemented to detect intensity-dependent effects of the log ratios distribution [18], and the distributions of signal intensity ratios are displayed as histograms. This feature is supplemented by QQ plots, which demonstrate to what extent the ratio follows the normal distribution.
Visual inspection enables identification of clones that should be excluded from further analysis, and at the same time it provides a basis for choosing the appropriate method of normalization.
Data normalization
The goal of normalization is to remove any systematic bias in the measured fluorescence intensities. Such systematic bias can originate from different labelling efficiencies of the used fluorochromes, different scanning parameters, spatial effects and/or other effects. In CGHPRO there are four options to remove these biases: Global Median, LOWESS (Locally Weighted Regression and Smoothing Scatterplots) [19], Subgrid Median and Subgrid LOWESS. Global methods assume that the bias is constant across the whole chip. Using Global Median, the Median value of log2 ratios is chosen as the correcting value. LOWESS function is especially useful for removing intensity-dependent biases. In order to reduce the influence of spatial effects, as may be detected by Boxplots, the above two methods can be applied to every subgrid separately. According to our experience, at least in our settings, the median subgrid works best for the normalization of array CGH data.
If one clone is spotted more than once on the chip, the replicate spots are automatically identified by their common ID. After normalization, the normalized ratios for the replicates are averaged and the standard deviations are calculated. This average ratio will later be used to represent the ratio of each clone. In subsequent analysis, users can set a threshold based on the number of replicas and standard deviation, such that clones exhibiting inconsistent results can be excluded.
Characterization of genomic profiles
The eventual goal of array CGH is the characterization of the individual genomic profile. Up to now, the common method is to use fixed thresholds, which should be dependent on the variability of the data. CGHPRO allows users to set the threshold either directly, or smooth the data first and then set a threshold based on the smoothed results. For smoothing, CGHPRO provides two options. When using moving average, which is applied to each chromosome separately, a window of adjustable size moves along the clones, which are ordered according to their base pair positions on the chromosome. The smoothed ratio of the clone at a window's centre will be the average ratio of the clones within the window.
The second smoothing strategy is to segment the clones, which are ordered along the chromosome, into sets with equal copy numbers. Then the data can be smoothed via averaging within the sets.
CGHPRO includes two methods for the segmentation of chromosomes into regions with identical copy numbers, namely 'Unsupervised Hidden Markov Partition' created by Jane Fridlyand [5] and 'Circular Binary Segmentation' first published by Adam Olshen [6]. We have implemented the two methods by linking the two R packages, aCGH and DNAcopy, to our program. Based on the smoothed ratios generated by one of these two algorithms, we have introduced the Median Absolute Deviation (MAD) as an objective measurement of data scattering.
Data display
Genomic display
The graphical interface of CGHPRO allows to explore the results in an interactive interface (Figure 1). In the Genome Display, the window consists of 24 sub-panels, each containing one chromosome. In each sub-panel, the ratios of clones are plotted in a size-dependent manner along the ideogram. As described below, several display parameters can be modified.
In each sub-panel, there are three lines along each chromosome. The yellow line represents a log ratio of zero; the individually adaptable green and red lines mark the negative and positive log ratios, respectively. The smoothed log ratios calculated by moving average, DNAcopy or aCGH can be chosen to be displayed as a black line called Smooth Line. Optionally, the original data can be blanked out.
Each clone is colour-coded according to its involvement in segmental duplications, as defined by the following formula: (Σ Length of Duplication * Copy Number)/ Length of Clone. Based on the factors determined this way, the clones are grouped into seven classes that can be looked up separately by clicking on the button with the corresponding colour in the top right corner.
Segmental duplications, which comprise ~5% of the human genome, are copies of genomic DNA with >90% sequence identity that range in size from 1 to >200 kb and are present in at least two locations in the human genome [20]. Highlighting segmental duplications is useful for the recognition of clones that may show misleading ratio scores [21]. Moreover, this feature also allows to relate chromosomal rearrangements to duplicated genomic regions. It has already been shown that segmental duplications increases the chances of non-allelic homologous recombination and that genomic regions flanked by these duplications are particularly prone to rearrangements [22]. Clicking on each sub panel will open a separate window and allow zooming on a specific chromosome.
Chromosome display
Chromosome Display provides a detailed view of the selected chromosome (Figure 2). In addition to the features provided by the Genome Display, the Chromosome Display supports the search for clones, zooming in or out, as well as the export of images. Upon clicking on a clone, information on its exact localization, contents of simple repeats, its involvement in segmental duplications, as well as information on number, position and ratio of the present replicas will be displayed in a text box. A key feature added to the Chromosome Display is a right-click mouse event, which will open a pop-up menu, offering several zoom options. Finally, Chromosome Display can be exported as an image file in Portable Network Graphics (png) format.
Comparative analysis of different chips
Once stored in the database, all entries can be used for comparative analysis at the genomic, chromosomal and clone-by-clone level. Genomic View is especially suitable for the summarizing display of chromosomal aberrations in a series of cases. In this mode, the absolute frequencies of aberrations within a study group are displayed alongside the chromosome ideograms ordered in a 6 × 4 grid. Upon clicking on the chromosomes of interest in the list located at the left side of the screen, the program switches to the Chromosome View and zooms in the respective chromosome. In addition to the absolute frequencies of aberrations, the relative frequencies can also be shown, which makes it easier to compare study groups of different size (Figure 3). For detailed analysis, the clone-by-clone view can be used. This mode supports "mouse over functionality", which displays further clone information in the bottom text field. Additionally the clone-by-clone matrix paves the way for the implementation of further algorithms such as hierarchical clustering.
As in all other views, balanced regions are indicated in yellow, while deleted and gained regions are shown in red and green, respectively. The simultaneous display of results from several experiments can assist in the definition of shortest regions of overlap, can help to reveal patterns of chromosomal aberrations, and can facilitate the recognition of odd clones.
Conclusion
CGHPRO represents a comprehensive and easy-to-use data analysis tool for array CGH. The software features the test of hybridization quality, normalization and visualization, as well as interactive data exploration and the comparative analysis of complete study groups. By providing all features offline, CGHPRO can be especially suitable in situations where protection of sensitive patient data is an issue. CGHPRO is written in Java and requires MySQL and R. The program runs on Linux and Windows operating systems.
It is freely available for use under the terms of the GNU General Public Licences (GPL) at the project's homepage. The open design of CGHPRO allows the easy adaptation to specific needs and the future incorporation of new features.
Availability and requirements
Project name: CGHPRO
Project home page:
Operating system: Linux and Windows 2000
Programming language: Java, SQL, R
Other requirements: Java 1.4 or higher, MySQL database and R
License: GNU General Public License
Any restriction to use by non-academics: Contact authors
Authors' contributions
WC was the principal programmer of the CGHPRO software. FE has performed the hybridization experiment and tested the program, HHR and SL contributed ideas for display and analyzing features and RU was responsible for defining the practical requirement catalog and contributed to manuscript preparation.
Acknowledgements
This work was supported by the Nationales Genomforschungsnetzwerk, grant 01GR0203. Additionally we would like to thank BACPAC Resources Center. and the Mapping Core and Map Finishing groups of the Wellcome Trust Sanger Institute for initial clone supply and verification, Sarah Shoichet for reading the manuscript, Martin Vingron for fruitful discussions and the anonymous reviewers for their helpful comments.
Figures and Tables
Figure 1 Genome Display exemplified by a male versus female hybridization on a 14000 BAC DNA array: Circles 1–4: (1) Colour coding table indicating the involvement of clones in segmental duplication (2) Black line representing the smoothed ratios calculated by DNAcopy (3) and (4) red and green bars to the left and right side of the ideogram highlighting regions of losses and gains, respectively.
Figure 2 Chromosome Display: (A) Detection of a copy number polymorphism, which encompass about 300 kb. (B) Zoom-in view of the relevant region (red rectangle in (A)).
Figure 3 Comparative Analysis: CGHPRO supports the visualization of absolute (A) and relative (B) frequencies of chromosomal aberrations in a series of cases. Results can be displayed simultaneously for all or for single chromosomes, as shown here for chromosome 11.
==== Refs
Kallioniemi A Kallioniemi OP Sudar D Rutovitz D Gray JW Waldman F Pinkel D Comparative genomic hybridization for molecular cytogenetic analysis of solid tumors Science 1992 258 818 821 1359641
Solinas-Toldo S Lampel S Stilgenbauer S Nickolenko J Benner A Dohner H Cremer T Lichter P Matrix-based comparative genomic hybridization: biochips to screen for genomic imbalances Genes Chromosomes Cancer 1997 20 399 407 9408757 10.1002/(SICI)1098-2264(199712)20:4<399::AID-GCC12>3.0.CO;2-I
Pinkel D Segraves R Sudar D Clark S Poole I Kowbel D Collins C Kuo WL Chen C Zhai Y Dairkee SH Ljung BM Gray JW Albertson DG High resolution analysis of DNA copy number variation using comparative genomic hybridization to microarrays Nat Genet 1998 20 207 211 9771718 10.1038/2524
Ishkanian AS Malloff CA Watson SK DeLeeuw RJ Chi B Coe BP Snijders A Albertson DG Pinkel D Marra MA Ling V MacAulay C Lam WL A tiling resolution DNA microarray with complete coverage of the human genome Nat Genet 2004 36 299 303 14981516 10.1038/ng1307
Fridlyand J Snijders AM Pinkel D Albertson DG Jain A Hidden Markov models approach to the analysis of array CGH data Journal of Multivariate Analysis 2004 90 132 153 10.1016/j.jmva.2004.02.008
Olshen AB Venkatraman ES Lucito R Wigler M Circular binary segmentation for the analysis of array-based DNA copy number data Biostatistics 2004 5 557 572 15475419 10.1093/biostatistics/kxh008
Autio R Hautaniemi S Kauraniemi P Yli-Harja O Astola J Wolf M Kallioniemi A CGH-Plotter: MATLAB toolbox for CGH-data analysis Bioinformatics 2003 19 1714 1715 15593402 10.1093/bioinformatics/btg230
Chi B DeLeeuw RJ Coe BP MacAulay C Lam WL SeeGH – a software tool for visualization of whole genome array comparative genomic hybridization data BMC Bioinformatics 2004 5 13 15040819 10.1186/1471-2105-5-13
BioConductor
Brazma A Hingamp P Quackenbush J Sherlock G Spellman P Stoeckert C Aach J Ansorge W Ball CA Causton HC Gaasterland T Glenisson P Holstege FC Kim IF Markowitz V Matese JC Parkinson H Robinson A Sarkans U Schulze-Kremer S Stewart J Taylor R Vilo J Vingron M Minimum information about a microarray experiment (MIAME)-toward standards for microarray data Nat Genet 2001 29 365 371 11726920 10.1038/ng1201-365
Human Genome Browser Gateway
Fiegler H Carr P Douglas EJ Burford DC Hunt S Scott CE Smith J Vetrie D Gorman P Tomlinson IP Carter NP DNA microarrays for comparative genomic hybridization based on DOP-PCR amplification of BAC and PAC clones Genes Chromosomes Cancer 2003 36 361 374 12619160 10.1002/gcc.10155
Osoegawa K Mammoser AG Wu C Frengen E Zeng C Catanese JJ de Jong PJ A bacterial artificial chromosome library for sequencing the complete human genome Genome Res 2001 11 483 496 11230172 10.1101/gr.169601
Krzywinski M Bosdet I Smailus D Chiu R Mathewson C Wye N Barber S Brown-John M Chan S Chand S Cloutier A Girn N Lee D Masson A Mayo M Olson T Pandoh P Prabhu AL Schoenmakers E Tsai M Albertson D Lam W Choy CO Osoegawa K Zhao S de Jong PJ Schein J Jones S Marra MA A set of BAC clones spanning the human genome Nucleic Acids Res 2004 32 3651 3660 15247347 10.1093/nar/gkh700
MPI array CGH protocols
arrayCGHbase- Matrix CGH Reloaded
L'Institut Curie: Bioinformatics
Yang YH Dudoit S Luu P Lin DM Peng V Ngai J Speed TP Normalization for cDNA microarray data: a robust composite method addressing single and multiple slide systematic variation Nucleic Acids Res 2002 30 e15 11842121 10.1093/nar/30.4.e15
Cleveland WS Robust Locally Weighted Regression and Smoothing Scatterplots Journal of the American Statistical Association 1979 74 829 836
Bailey JA Gu Z Clark RA Reinert K Samonte RV Schwartz S Adams MD Myers EW Li PW Eichler EE Recent segmental duplications in the human genome Science 2002 297 1003 1007 12169732 10.1126/science.1072047
Locke DP Segraves R Nicholls RD Schwartz S Pinkel D Albertson DG Eichler EE BAC microarray analysis of 15q11–q13 rearrangements and the impact of segmental duplications J Med Genet 2004 41 175 182 14985376 10.1136/jmg.2003.013813
Stankiewicz P Lupski JR Genome architecture, rearrangements and genomic disorders Trends Genet 2002 18 74 82 11818139 10.1016/S0168-9525(02)02592-1
|
15807904
|
PMC1274268
|
CC BY
|
2021-01-04 16:02:51
|
no
|
BMC Bioinformatics. 2005 Apr 5; 6:85
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-85
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-151578014410.1186/1471-2334-5-15Technical AdvanceHuman immunodeficiency virus type 1 (HIV-1) proviral DNA load in purified CD4+ cells by LightCycler® Real-time PCR Kabamba-Mukadi Benoît [email protected] Philippe [email protected] Jean [email protected]ère Nicole [email protected]éus Monique [email protected] Patrick [email protected] Laboratoire de référence SIDA, Université Catholique de Louvain, Avenue Hippocrate 54, B1200 Bruxelles, Belgium2 Service de Médecine interne, Clinique Saint Joseph, Liège, Belgium2005 21 3 2005 5 15 15 3 12 2004 21 3 2005 Copyright © 2005 Kabamba-Mukadi et al; licensee BioMed Central Ltd.2005Kabamba-Mukadi et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The human immunodeficiency virus type 1 (HIV-1) proviral DNA persists in infected cells, even after prolonged successful HAART. In the present study, a relative quantification assay of HIV-1 proviral DNA by LightCycler® real-time PCR based on SYBR Green I detection was developed in comparison to the number of purified CD4+ cells as estimated by the quantification of the β-globin gene.
Methods
The ability of the designed gag primers to quantify HIV-1 Group M and the PCR efficiency were assessed on HIV-1 reference isolate subtypes A, B, C and D. The 8E5 cell line containing a single defective copy of HIV-1 proviral DNA was used as a standard for both the HIV-1 target gene and the β-globin reference gene. The assay was applied on thirty consecutive patient samples received for RNA viral load determinations and on retrospective samples from fifteen patients undergoing 2 years of structured treatment interruption (STI).
Results
The lower limit of quantification was 50 HIV-1 DNA proviral copies per CD4+ cell sample. The dynamic range was from 50 to 106 HIV-1 DNA copies per CD4+ cell sample with intra- and inter-assay coefficients of variability ranging from 3.1% to 37.1%. The β-globin reference gene was quantified down to a limit of 1.5 pg of DNA/μl (approximately 5 cells) with intra- and interassay coefficients of variability ranging from 1.8% to 21%. DNA proviral load varies widely among HIV-1 infected patients. Proviral load and plasma viral load rebound were high in STI patients who took longer to achieve an undetectable plasma viral load under therapy. A statistically significant correlation was observed between DNA proviral load and RNA steady state viral load in STI patients (p-value = 0,012).
Conclusion
We have developed a fast, sensitive and specific relative quantification assay of HIV-1 proviral DNA in purified CD4+ cells. The assay enables the monitoring of HIV-1 proviral load, which may be useful to monitor therapy efficacy especially in patients with undetectable plasma RNA viral load, and allows the exploration of viral reservoirs.
==== Body
Background
At present, HIV-1 infected patients are followed by monitoring plasma HIV-1 RNA viral load, allowing follow up of the immediate effects of treatment, and CD4+ T cell count, allowing initiation or reinitiation of the therapy as immune function decreases [1]. The HIV-1 proviral DNA load could be an alternative viral marker, as it is known that proviral DNA persists in infected cells, even after prolonged successful HAART as evidenced by undetectable plasma RNA viral load. A decline in DNA might indicate a long-term impact of the HAART on the reservoirs and the long-term effectiveness of the treatment [2-11]. But data regarding the decline in DNA are sometimes conflicting. Some authors noted decreased levels after one year of antiretroviral triple combination therapy [12] and others reported stable HIV-1 DNA levels over several years in PI-ART naive patients [13,14]. Several assays for the quantification of HIV type 1 proviral DNA in peripheral blood mononuclear cells (PBMC) have previously been reported, and many of them are based either on the principle of conventional PCR requiring post PCR analysis or on real-time PCR on total PBMC, using specific probe detection [9,10,15-23]. Recently, a multiplex real-time PCR for quantification of HIV-1 DNA and the human albumin gene in CD4+ cells has been published [23] using TaqMan probes and an Epstein-Barr virus (EBV) standard curve. Sequence-specific hybridisation probes provide the most specific real-time analysis of amplified target sequences but the very high variability of HIV sequences within subtypes, even in HIV variants in a given patient, led us to choose a sequence-independent detection with SYBR Green I. One of the experimental prerequisites for relative quantification consists in having identical PCR efficiencies for both target gene and reference gene in sample, standard, and calibrator. PCR efficiency may vary between EBV standard PCR and HIV-1 PCR. We therefore used the 8E5 cell line containing a single defective copy of HIV-1 proviral DNA as a standard for both the HIV-1 target gene and the β-globin reference gene in order to carry out a relative quantification. Another recent publication described a real-time PCR based on LightCycler® technology revealed through SYBR green fluorochrome to quantify the HIV-1 proviral DNA load[24]. The article describes quantification in PBMC of HIV-1 seropositive patients. As 95 to 99% of infected cells are CD4+ cells[25], we developed a relative quantification assay of HIV-1 proviral DNA in purified CD4+ cells on the LightCycler® by real-time PCR compared to cell quantification by β-globin PCR.
Methods
Primer design and selection
Several primers were designed in the gag-pol junction region with low variability on the basis of a consensus sequence obtained by aligning complete sequences of HIV-1 subtypes A to J from the Entrez nucleotides database [26]. The Vector NTI® Suite (InforMax, Bethesda, USA) software was used for the design of primers. Selection of primer sets was first based on a specific signal without primer dimer formation when testing DNA from 8E5 cell dilution series and negative controls. Primer and MgCl2 concentrations were optimised by combining primer concentrations ranging from 0.3 mM to 1.2 mM and MgCl2 concentrations from 2 mM to 5 mM. The selected primers were tested on DNA samples obtained from HIV-1 subtypes A, B, C, and D reference strains and from clinical samples including subtypes B and non-B. We also tested the absence of signal with HIV-2 DNA. The final primer set selection was based on the efficiency of the PCR reaction. Differences in efficiency may have a major impact on the calculation of the initial amount. In theory, the optimum efficiency in PCR is two, meaning that every PCR product is replicated once every cycle. In reality, however, many PCR parameters influence PCR efficiency which is calculated according to the formula E = 10-1/slope. The LightCycler® Software calculates the slope of the standard curve by plotting crossing points against the logarithm of concentration for each standard point of a 10-fold dilution series. A slope value of -3.33 indicates the maximal PCR efficiency.
Reference strains and patients
Four different reference strains of HIV-1 and 2 of HIV-2 from the National Institute for Biological Standards and Control (NIBSC, UK) were used in this project: HIV-1 Primary isolates-Uganda (HIV-1 subtype A, NIBSC repository reference ARP177.5) [27], 8E5/LAV (HIV-1 subtype B, NIBSC repository reference ARP110) [28], HIV-1 SE12808/ SE14784 (Subtype C, NIBSC repository reference ARP197.1, ARP197.2) and HIV1-ELI (HIV-1 subtype D, NIBSC repository reference EVA117) [29], HIV-2 ROD reference strain (subtype A, NIBSC repository reference EVA121.1) [30,31] and HIV-2 EHO reference strain (subtype B, NIBSC repository reference EVA132) [32].
Samples from 15 patients undergoing 2 years (start 2001) of structured treatment interruption (STI) and 30 consecutive clinical samples received for RNA plasma viral load determinations were included in the study. Informed consent following the Helsinki declaration was obtained from each patient. The HIV-1 seropositive status was confirmed according to the accepted methods. HIV-1 RNA plasma viral load was performed using the HIV-1 Amplicor™ Monitor 1.5 (Roche, Branchburg, NJ, USA). The viral RNA and the proviral DNA sequences were determined by using an in house RT-PCR method applied on reverse transcriptase and protease genes [adapted from [33]]. HIV RNA was extracted from plasma (QIAmp Viral Mini Kit™, QIAGEN, Leiden, The Netherlands) and a direct cycle sequencing was used with BigDye terminator chemistry on an ABI Prism 310 (Applied Biosystems, Foster City, USA).
CD4+ cells were isolated from 10 ml of patient EDTA whole blood samples by an immunomagnetic method using anti-CD4 coated magnetic beads (Dynabeads M450 CD4, Dynal A S, Oslo, Norway) according to the manufacturer's protocol and were stored at -80°C until use. The complete process takes approximately 3 hours and the purity of the CD4+ cell preparation was about 99% as estimated by Becton Dickinson Fascan Flow Cytometer technology (data not shown). HIV-2 clinical samples were obtained from patients who were diagnosed with HIV-2 infection at our AIDS reference laboratory. The 8E5 cell line, the CD4+ blood donor cells infected with HIV-1 reference strains, and the H9 cell line infected with HIV-2 reference strains were cultured in RPMI 1640 medium supplemented with 10% FetalClone 1® (HyClone, Logan, Utah, USA), 1% Glutamine 200 mM and 0.1% Gentamycin 50 mg/ml. 8E5 cells were counted by the Coulter automated haematology analyser and diluted to 106 cells per aliquot stored at -80°C.
DNA purification
DNA was extracted from purified patient CD4+ cells diluted in 200 μl PBS, using the High Pure® PCR Template Preparation Kit (Roche Diagnostics GmbH, Mannheim, Germany) according to the manufacturer's recommendations. To concentrate the HIV-1 target gene in patient samples, DNA was eluted in a volume of 50μl. DNA from the T-lymphoblastoid 8E5 cells, and CD4+ blood donor cells infected with HIV-1 reference strains were eluted in 200μl. The 8E5 cell DNA was purified from 106 cells. A negative control (template replaced by Nuclease-free water) was included in each PCR run. Particular attention was paid to using DNAse and RNAse free materials. Depending on the number of samples, the whole DNA purification process required approximately 1 hour.
HIV-1 DNA real time PCR assay
A series of ten-fold dilutions of 8E5 DNA corresponding to 2.5 × 103 to 2.5 HIV-1 DNA copies per μl (2.5 × 103 to 2.5 cells/μl) was included in each experiment in order to generate an external standard curve. The PCR mixture (total volume 20 μl) in nuclease free water contained 2 μl of LightCycler® FastStart DNA Master SYBR Green 1, a ready-to-use "Hot Start" reaction mix (Roche Diagnostics GmbH, Mannheim, Germany), final concentrations of 4 mM MgCl2 and 0.5μM of each primer and 5μl of purified DNA or negative control. All samples were analysed in duplicate. The amplification protocol for HIV-1 on the LightCycler® was as follows: a 10 min denaturation step at 95°C for polymerase activation, a "touch down" PCR step of 10 cycles consisting of 10 seconds (s) at 95°C, 10 s at 65°C, and 30 s at 72°C, followed by 40 cycles consisting of 10 s at 95°C, 10 s at 55°C, and 30 s at 72°C. The fluorescence was measured at the end of each elongation step. The next step was a slow heating (0.1°C per s) of the amplified product from 65°C to 95°C in order to generate a melting temperature curve. This curve served as a specificity control. The entire cycling process including data analysis took less than one hour and was monitored using the LightCycler® software program (version 3.5). Second derivative maximum mode was chosen with baseline adjustment set in the arithmetic mode.
A fragment from the human β-globin gene was amplified in parallel with the HIV-1 gag gene to quantify the total number of investigated cellular genomes. The β-globin real-time quantitative PCR was performed using the LightCycler® Control Kit DNA (Roche Diagnostics GmbH, Mannheim, Germany) according to the manufacturer's recommendations. The PCR mixture (total volume 20μl) in nuclease free water contained 2μl of LightCycler® FastStart DNA Master SYBR Green 1 and 2μl of LightCycler® Control Kit DNA primer mix, 2μl of 10-fold dilution of purified DNA or negative control, and a final concentration of 4 mM MgCl2. The DNA control from the kit corresponds to 30 ng of human genomic DNA. The same 8E5 DNA standard samples corresponding to 2.5 × 103 to 2.5 cells/μl were included in each experiment in order to generate an external standard curve as for the HIV-1 experiment. All samples were run in duplicate. A melting temperature step of β-globin amplicons was performed.
To validate the use of the 8E5 cells as standards, we first performed a real-time PCR quantification of the human β-globin gene using the DNA control of the kit to generate a standard curve. The values of the 8E5 cell count as performed by the Coulter automated haematology analyser were compared with the values obtained by extrapolating the cell count from the β-globin real-time PCR quantification. In the second step, 8E5 cells were used as standards and we compared the kit DNA control value obtained from the LightCycler® quantification with the theoretical value.
Relative quantification was carried out using the LightCycler® Relative quantification Software (version 1.0). The calculation of data is based on the crossing point (Cp) values obtained by the LightCycler® Software. Results are calculated as the target/reference ratio of the sample divided by the target/reference ratio of the calibrator. This corrects for sample inhomogeneity and variability of detection. The result of the HIV-1 proviral quantification was expressed as log10 number of DNA copies per 106 CD4+ cells.
Statistical analysis
Statistical analysis was performed using the parametric Pearson correlation test or the nonparametric Spearman correlation method. The critical p-value required to reject the null hypothesis (there is no proof of significant correlation between the variables) and accept the alternative hypothesis is equal to or less than 0,05.
Results
Primer selection
Out of several primers designed in the HIV-1 gag-pol junction region, the sequences of the selected HIV-1 gag forward primer BK1F and the reverse primer BK1R were respectively: 5'-GTA ATA CCC ATG TTT TCA GCA TTA TC-3' and 5'-TCT GGC CTG GTG CAA TAG G-3'. These amplify a 181-bp amplicon in the gag region of low variability among HIV-1 subtypes.
Standard validation
The use of 8E5 cells as standards was validated. As shown in Figure 1, the HIV-1 and the β-globin curves generated by using 8E5 cells had comparable slopes, indicating equal PCR efficiency. The comparison of the 8E5 cell count as performed by the Coulter automated haematology analyser with the values obtained by real-time PCR quantification of β-globin genes showed similar results, as did the comparison of the DNA control sample value obtained from LightCycler® quantification with the theoretical value (Figure 2).
Figure 1 the HIV-1 and the human globin standard curves created by a ten-fold dilution series of 8E5 cell DNA show comparable slopes, indicating equal PCR efficiency. The LightCycler Software calculates the slope of the standard curve by plotting crossing points against the logarithm of concentration for each standard point. Each standard point was run in five replicates. (_____) HIV-1 DNA standard curve, (_ _ _) Human globin standard curve.
Figure 2 Validation of the use of 8E5 cells as standards for human globin PCR. The figure shows similar results of the 8E5 cell count as performed by the Coulter automated hematology analyzer compared with the values obtained by real-time PCR quantification of the human globin gene. The comparison of the theoretical and calculated values of the DNA control sample from LightCycler Kit shows equal results. Each standard point was run in five replicates.
Specificity of the assay
As described in the methods section, the specificity of the PCR reaction was tested with DNA purified from the 8E5 cell line carrying a single copy of subtype B proviral DNA per cell, with CD4+ blood donor cells infected with HIV-1 subtype A, C, and D reference strains and with H9 cells infected with HIV-2 ROD (subtype A) and EHO (subtype B) reference strains. Clinical samples from HIV-1 or HIV-2 infected patients were included. The subtyping results based on proviral protease sequences showed 65% of subtype B viruses and 35% of non-subtype B viruses, which included subtypes A, C, D, F1 and CRF AE [34].
HIV-1 amplicons produced in the PCR reaction had the same melting temperature of 85.30°C, with a slight variation depending on the HIV-1 subtypes and on the type of initial sample, indicating that the signal was specific. The melting temperature of β-globin amplicons was about 86°C. Furthermore, the specificity of the HIV-1 amplification was confirmed by the presence of the expected band size on the electrophoresis gel (agarose 2%) carried out with the PCR products obtained from standards, reference strains, and patient samples. Direct sequencing of amplified products using the PCR primers showed the expected HIV-1 sequences (data not shown). No amplification was obtained with HIV-2 DNA purified from patient samples and from reference strains belonging to HIV-2 subtypes A and B. The four HIV-1 reference strains obtained from NIBSC were amplified. The results showed that the assay amplified DNA purified from HIV-1 reference isolate subtypes A, B, C, and D with similar PCR efficiency (Table 1).
Table 1 PCR efficiencies of HIV-1 group M amplification
Sample concentration: Subtype A Subtype B Subtype C Subtype D
Log dilution Mean Cpa Mean Cp Mean Cp Mean Cp
-2 26,16 22,90 24,61 26,13
-3 29,50 26,02 28,10 29,27
-4 32,82 29,75 31,79 32,86
-5 36,72 32,98 34,83 36,72
Slopeb -3,50 -3,37 -3,43 -3,54
R2c 0,99 0,99 0,99 0,99
(a) Mean of crossing point values calculated using two replicates of each point in 3 runs.
(b) The LightCycler Software calculates the slope of the curve by plotting crossing points against the logarithm of concentration for each point of a 10-fold dilution series. A slope value of -3.33 indicates the maximal PCR efficiency.
(c) The LightCycler Software also calculates the coefficient of correlation R2.
Sensitivity and reproducibility of the assay
Ten replicates of 1 copy, 5 copies, and 10 copies of DNA purified from 8E5 cells were tested in the PCR reaction. The detection of one HIV-1 DNA copy per PCR failed in most cases while five and ten copies were always detected. The lower limit of quantification was set at 5 HIV-1 DNA proviral copies per PCR or 50 HIV-1 DNA proviral copies per CD4+ cell sample. A HIV-1 DNA proviral load below 50 copies per CD4+ cell sample was reported as undetectable. The dynamic range was from 50 to 106 HIV-1 DNA proviral copies per CD4 lymphocyte sample. We tested the intra and inter-assay reproducibility of our technique with a series of ten-fold dilutions of 8E5 DNA from 12.5 × 103 to 12.5 HIV-1 DNA copies per PCR. The intra-assay reproducibility was evaluated using five replicates of each point and the inter-assay reproducibility was calculated on 10 runs. The intra- and inter-assay coefficients of variability of the HIV-1 DNA copy number ranged from 3.1% for high provirus concentrations to 37.1% for low concentrations. The quantification of the human β-globin reference gene by using the "LightCycler® Control Kit DNA" enabled a lower limit of detection of 1.5 pg/μl (approximately 5 cells) with intra- and interassay coefficients of variability ranging from 1.8% for high DNA concentrations to 21% for low DNA concentrations. We observed an inhibition of the human β-globin reference gene amplification at high amounts of the cellular genome. The inhibition was shown by a maximum fluorescence of the amplification curve, which was lower than the maximum fluorescence of the standard points. In our assay, this problem was solved by diluting (10-fold) DNA from patient samples. The influence of the type of sample on the PCR reaction was tested by diluting the standard 8E5 cells in HIV seronegative plasma from a blood donor before DNA extraction. A slight difference (about 1 Cp) was observed within the range of 12.5 to 12.5 × 103HIV-1 DNA copies per PCR between the Cp values obtained for 8E5 cells diluted in PBS buffer as compared to 8E5 cells diluted in the plasma, in both HIV-1 and β-globin assays. We also successfully tested the quantification of both HIV-1 and β-globin genes without the need to include a standard curve in every run by loading an external standard curve generated in a different run.
HIV-1 DNA viral load in clinical specimens
In order to test our quantitative real-time PCR assay on DNA from patient CD4+ cell samples, the assay was first applied on thirty consecutive patient samples received for RNA viral load determinations regardless of whether or not they were receiving antiretroviral therapy. Of the 30 patients, 10 were treatment naive while 20 were antiretroviral-experienced adults (Table 2). If we put all patients (N = 30) together, no significant correlation was found between DNA relative proviral load and either plasma RNA viral load or CD4+ cell count (Figure 3 and Figure 4). The only observed correlative tendency was between DNA relative proviral load and plasma RNA viral load in the naive patients (Figure 5), but this was not statistically significant (p-value >0,05 by Pearson test). DNA proviral load varies widely among HIV-1 infected patients in the same or at different disease stages.
Table 2 Patient characteristics
Naive group Treated patients STI patients
Number 10 20 15
Genders
Males 5 8 8
Females 5 12 6
Transsexuals 1
Mean years old 42 42 43
Geographic origin
Europeans 5 12 12
Other/Unknown 5 8 3
Mean range follow up
Yearsa 4 (range 1 – 11) a 7.5 (range 1 – 15)a 6 (range 2 – 7) a
CD4+ count
cells × 106/l Median 391 (range 109–805) Median 338 (range 1 – 790) Median 717 (range 216 – 1356)
<200 cells × 106/l (Number) b 1 6 0
Plasma HIV-RNA
Log10 copies/ml Median 4.8 (range 2.7 – 5.4) Median 4.4 (range 1.7 – 5.9) Median 3.2 (range 1.7 – 4.7) c
<50 copies/ml (Number)b 0 5 3
HIV1-DNA
Log10 copies/106 CD4 cells Median 3.0 (range 2.6 – 3.7) Median 3.3 (range 2.2 – 4.6) Median 2.9 (range 2.1 – 4.0)
<5 copies/PCR (Number) b 6 12 5
(a) Number of years between HIV-1 infection diagnosis and DNA proviral load testing.
(b) Number of patients with viral or proviral load under limit of quantification or with CD4+ count <200 cells 106/l.
(c) Steady state viral load after STI.
Figure 3 No correlation was found between DNA proviral load and plasma RNA viral load in thirty consecutive unselected patient samples received for RNA viral load determinations.
Figure 4 No correlation was found between DNA proviral load and CD4+ cell count in thirty consecutive unselected patient samples received for RNA viral load determinations.
Figure 5 Correlative tendency between DNA relative proviral load and plasma RNA viral load in the naive patients but not statistically significant (p-value >0,05 by Pearson test)
The assay was also applied on samples from fifteen patients undergoing 2 years of STI (Table 2). At the start of the STI, all patients were receiving at least a triple-therapy regimen including two nucleoside reverse transcriptase inhibitors (NRTIs) plus 1 or 2 PIs. All STI patients had received previous monotherapy or dual therapy before HAART with a mean total treatment time of five years. At the initiation of STI, 11 patients out of the 15 had an HIV-1 RNA viral load of less than 50 copies/ml; four had a baseline HIV-1 RNA viral load of less than 572 copies/ml (median of 2 log copies/ml) and the mean CD4+ cell count was 748 × 106 cells/l. Out of the 15 STI patients, seven had a steady state viral load less than 3.3 log copies/ml, including three patients with undetectable HIV-1 RNA viral load (<50 copies/ml), while eight patients had a high HIV-1 RNA viral load ranging from 3.9 to 4.7 log copies/ml. All seven patients with a low steady state plasma viral load showed a DNA proviral load under 2.5 log copies/106 CD4+ cells, including four patients with a proviral load under the limit of quantification (<5 DNA copies/PCR). Of eight patients with a high steady state plasma viral load (>3.9 log copies/ml), three showed a DNA proviral load under 2.5 log copies/106 CD4+ cells while five had a DNA proviral load above 2.5 log copies/106 CD4+ cells. A statistically significant correlation was observed between DNA proviral load and RNA steady state viral load in STI patients (Figure 6). The Pearson's coefficient of correlation was about 0,578 with an associated p-value of 0,012. When applying the nonparametric Spearman correlation method, the associated p-value was about 0,018. Proviral load and plasma viral load rebound were high in STI patients who took longer to achieve an undetectable plasma viral load under therapy.
Figure 6 The assay was also applied on samples from 15 patients undergoing 2 years of STI. A statistically significant correlation was observed between DNA proviral load and RNA steady state viral load in STI patients. (p-value = 0,012, Pearson correlation test)
Discussion
A relative quantification assay of HIV-1 proviral DNA in purified CD4+ cells was developed on the LightCycler® real-time PCR instrument expressed per 106 CD4+ cells as derived from estimated by the quantification of the β-globin gene. In healthy individuals, CD4+ cells represent approximately 30 to 50% of the PBMC and their fraction may be considerably lower and variable over time in HIV-infected patients. Theoretically, to improve the sensitivity and accuracy of an assay for low-copy-number samples, a larger amount of DNA should be added to the PCR reaction mix when purified from PBMCs than directly from CD4+ cells. High amounts of the cellular genome may lead to PCR inhibition. In addition, published data show low proviral loads, ranging from 1 to < 5 log10 HIV-1 DNA copies (median 2–3 log10) in 106 PBMC equivalents [14-17,19,20,22-24]. Methods quantifying the HIV-1 DNA proviral load in PBMC may therefore lead to undetectable proviral loads. However, in any case a direct comparison between DNA load in CD4+ cells and PBMCs could be useful as the CD4 isolation adds extra work (i.e. costs), which should be weighed against the clinical benefit. Sequence-specific hybridisation probes provide the most specific real-time analysis of amplified target sequences but the very high variability of HIV sequences within subtypes, even in HIV variants in a given patient led us to choose a sequence-independent detection with SYBR Green I. The ability of SYBR Green I to bind double-stranded DNA molecules with emission of a fluorescent signal allows the PCR reaction to be monitored. As SYBR Green I binds to all double-stranded DNA molecules, a melting temperature curve needs to be constructed to ensure the specificity of the signal. Thus specific primer selection and PCR optimisation is crucial to avoid primer dimer formation. We tested the influence of possible inhibitors in the DNA sample. A slight variation of the melting temperature (about 1°C) and the Cp values (about 1Cp) was observed, depending on the HIV-1 subtypes and on the type of initial sample (EDTA whole blood or culture supernatant). The specificity of the HIV-1 amplification was confirmed by the presence of the expected band size on the electrophoresis gel carried out with the PCR products from standards, reference strains, and patient samples. Direct sequencing of amplified products showed the expected HIV-1 sequences. No amplification was obtained with HIV-2 DNA. The ability of the assay to quantify HIV-1 Group M and the PCR efficiency were assessed on HIV-1 reference isolate subtypes A, B, C, and D. The results showed similar PCR efficiencies. Particular attention should be given to the quantification of the β-globin gene. We observed an inhibition with a high amount of the reference gene. Diluting (10-fold) DNA from patient samples resolves this problem. The lower limit of quantification is five HIV-1 DNA proviral copies per PCR and 1.5 pg of cellular DNA/μl (approximately five cells) for the human β-globin reference gene quantification. The dynamic range of HIV-1 DNA quantification is between 50 and 106 proviral copies per CD4 with coefficients of variability of the HIV-1 DNA copy number ranging from 3.1% for high provirus concentrations to 37.1% for low concentrations. Our method shows acceptable technical sensitivity and specificity. Quantification could be performed without the need to include a standard curve in every run by loading an external standard curve generated in a different run. Relative quantification was carried out by using the LightCycler® Relative quantification Software (version 1.0). The result of HIV-1 proviral quantification was expressed as log10 number of DNA copies per 106 CD4+ cells. The crossing points that are calculated by the LightCycler® Software are a function of the amplification efficiency. Efficiency differences have a major impact on the accuracy of initial amount calculation. The assay developed met one of the experimental prerequisites for PCRs, HIV-1 target gene and β-globin reference gene, i.e., having identical PCR efficiencies in sample, standard, and calibrator.
The assay was successfully tested on 30 consecutive unselected patient samples. No significant correlation was found between DNA relative proviral load and either plasma RNA viral load or CD4+ cell in this group, overall. Similar observations have been reported in the literature [9,24]. DNA proviral load varies widely among HIV-1 infected patients in the same or at different disease stages.
The assay was also applied on samples from 15 patients undergoing 2 years of STI. In fact, of the 15 STI patients, all seven patients with a low steady state plasma viral load showed a DNA proviral load under 2.5 log copies/106 CD4+ cells, including four patients with a proviral load under the limit of quantification (<5 DNA copies/PCR). Of the remaining eight patients presenting a high steady state plasma viral load (>3.9 log copies/ml), three showed a DNA proviral load under 2.5 log copies/106 CD4+ cells while five had a DNA proviral load above 2.5 log copies/106 CD4+ cells. We observed that in patients who needed a change in antiretroviral treatment to achieve an undetectable plasma viral load or who took longer to achieve an undetectable viral load under treatment, the proviral load was >2.5 log copies/PCR and the plasma viral load rebound was high. However, a large and prospective study is needed to confirm the value of this observation and particularly to assess the predictive value of proviral load in the setting of STI.
Conclusion
We have developed a fast, sensitive and specific assay, which enables the monitoring of HIV-1 proviral load in CD4+ cells by LightCycler® real-time PCR based on SYBR Green I quantification. This should enable us to evaluate prospectively the proviral load as a prognostic marker in therapy and in the evaluation of treatment interruptions.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
BKM conceived of the study and designed it together with PH and PG. BKM developed the HIV-1 DNA real-time PCR and performed the assay and sequencing reactions. PH assembled the clinical samples. ND did the CD4+ cell preparation and the all the viral culture work. BK and PG drafted the manuscript. JR and MB reviewed the manuscript. All authors contributed to the final version of the manuscript, read and approved it.
Pre-publication history
The pre-publication history for this paper can be accessed here:
==== Refs
Guidelines for the Use of Antiretroviral Agents in HIV-1-Infected Adults and Adolescents October 29, 2004
Bennett JM Kaye S Berry N Tedder RS A quantitative PCR method for the assay of HIV-1 provirus load in peripheral blood mononuclear cells J Virol Methods 1999 83 11 20 10598078 10.1016/S0166-0934(99)00096-8
Chun TW Stuyver L Mizell SB Ehler LA Mican JA Baseler M Lloyd AL Nowak MA Fauci AS Presence of an inducible HIV-1 latent reservoir during highly active antiretroviral therapy Proc Natl Acad Sci USA 1997 94 13193 13197 9371822 10.1073/pnas.94.24.13193
Burgard M Izopet J Dumon B Tamalet C Descamps D Ruffault A Vabret A Bargues G Mouroux M Pellegrin I Ivanoff S Guisthau O Calvez V Seigneurin JM Rouzioux C HIV RNA and HIV DNA in peripheral blood mononuclear cells are consistent markers for estimating viral load in patients undergoing long-term potent treatment AIDS Res Hum Retroviruses 2000 10 1939 1947 11153076 10.1089/088922200750054666
De Milito A Titanji K Zazzi M Surrogate markers as a guide to evaluate response to antiretroviral therapy Curr Med Chem 2003 10 349 365 12570696
Finzi D Hermankova M Pierson T Carruth LM Buck C Chaisson RE Quinn TC Chadwick K Margolick J Brookmeyer R Gallant J Markowitz M Ho DD Richman DD Siliciano RF Identification of a reservoir for HIV-1 in patients on highly active antiretroviral therapy Science 1997 278 1295 1300 9360927 10.1126/science.278.5341.1295
Garrigue I Pellegrin I Hoen B Dumon B Harzic M Schrive MH Sereni D Fleury H Cell-associated HIV-1 DNA quantification after HAART-treated primary infection in patients with persistently undetectable plasma HIV-1 RNA AIDS 2000 14 2851 2855 11153666 10.1097/00002030-200012220-00006
Ngo-Giang-Huong N Deveau C Da Silva I Pellegrin I Venet A Harzic M Sinet M Delfraissy JF Meyer L Goujard C Rouzioux C Frnech PRIMO Cohort Study Group Proviral HIV-1 DNA in subjects followed since primary HIV-1 infection who suppress plasma viral load after year of highly active antiretroviral therapy. French PRIMO Cohort Study Group AIDS 2001 15 665 673 11371680 10.1097/00002030-200104130-00001
Pellegrin I Caumont A Garrigue I Merel P Schrive MH Fleury H Dupon M Pellegrin JL Ragnaud JM Predictive Value of Provirus Load and DNA Human Immunodeficiency Virus Genotype for Successful Abacavir-Based Simplified Therapy J Infect Dis 2003 187 38 46 12508144 10.1086/345860
Saitoh A Hsia K Fenton T Powell CA Christopherson C Fletcher CV Starr SE Spector SA Persistence of human immunodeficiency virus (HIV) type 1 DNA in peripheral blood despite prolonged suppression of plasma HIV-1 RNA in children J Infect Dis 2002 15 1409 1416 11992275 10.1086/340614
Wong JK Hezareh M Gunthard HF Havlir DV Ignacio CC Spina CA Richman DD Recovery of replication-competent HIV despite prolonged suppression of plasma viremia Science 1997 278 1291 1295 9360926 10.1126/science.278.5341.1291
Aleman S Visco-Comandini U Lore K Sönnerborg A Long-term effects of antiretroviral combination therapy on HIV type 1 DNA levels AIDS Res Hum Retroviruses 1999 15 1249 54 10505673 10.1089/088922299310142
Hockett RD JrSaag MS Kilby JM Sfakianos G Wakefield TB Bucy RP Stability in the HIV vDNA pool in peripheral CD4+ T cells of untreated patients by single tube quantitative PCR J Virol Methods 2000 87 1 12 10856747 10.1016/S0166-0934(00)00139-7
Kostrikis LG Touloumi G Karanicolas R Pantazis N Anastassopoulou C Karafoulidou A Goedert JJ Hatzakis A Multicenter Hemophilia Cohort Study Group Quantitation of human immunodeficiency virus type 1 DNA forms with the second template switch in peripheral blood cells predicts disease progression independently of plasma RNA load J Virol 2002 76 10099 108 12239284 10.1128/JVI.76.20.10099-10108.2002
Christopherson C Kidane Y Conway B Krowka J Sheppard H Kwok S PCR-based assay to quantify human immunodeficiency virus type 1 DNA in peripheral blood mononuclear cells J Clin Microbiol 2000 38 630 634 10655358
Comandini UV Sönnerborg A Vahlne A Yun Z Quantification of HIV-1 proviral DNA from peripheral blood mononuclear cells using a high throughput four-competitor competitive PCR J Virol Methods 1997 69 171 180 9504762 10.1016/S0166-0934(97)00153-5
Gratzl S Moroni C Hirsh HH Quantification of HIV-1 viral RNA and proviral DNA by isotopic competitive PCR J Virol Methods 1997 66 269 282 9255738 10.1016/S0166-0934(97)00064-5
Guenthner PC Hart CE Quantitative, competitive PCR assay for HIV-1 using a microplate-based detection system BioTechniques 1998 24 810 816 9591131
Izopet J Tamalet C Pasquier C Sandres K Marchou B Massip P Puel J Quantification of HIV-1 proviral DNA by a standardized colorimetric PCR-based assay J Med Virol 1998 54 54 59 9443109 10.1002/(SICI)1096-9071(199801)54:1<54::AID-JMV8>3.0.CO;2-O
Jurriaans S Dekker JT de Ronde A HIV-1 viral DNA load in peripheral blood mononuclear cells from seroconverters and long-term infected individuals AIDS 1992 6 635 641 1503683
Stieger M Démollière C Ahlborn-Laake L Mous J Competitive polymerase chain reaction assay for quantification of HIV-1 DNA and RNA J Virol Methods 1991 34 149 160 1804850 10.1016/0166-0934(91)90095-H
Zhao Y Yu M Miller JW Chen M Bremer EG Kabat W Yogev R Quantification of Human Immunodeficiency Virus Type 1 Proviral DNA by Using TaqMan Technology J Clin Microbiol 2002 40 675 678 11825994 10.1128/JCM.40.2.675-678.2002
Eriksson LE Leitner T Wahren B Bostrom AC Falk KI A multiplex real-time PCR for quantification of HIV-1 DNA and the human albumin gene in CD4+ cells APMIS 2003 111 625 33 12969018 10.1034/j.1600-0463.2003.1110605.x
Gibellini D Vitone F Schiavone P Ponti C Placa M Carla Re M Quantitative detection of human immunodeficiency virus type 1 (HIV-1) proviral DNA in peripheral blood mononuclear cells by SYBR green real-time PCR technique JCV 2004 29 282 289 15018857 10.1016/S1386-6532(03)00169-0
Girard PM Katlama Ch Pialoux G VIH Paris: Doin 2001
The Entrez Nucleotides database
Cheingsong-Popov R Osmanov S Pau CP Schochetman G Barin F Holmes H Francis G Ruppach H Dietrich U Lister S Weber J Serotyping of HIV type 1 infections: definition, relationship to viral genetic subtypes, and assay evaluation. UNAIDS Network for HIV-1 Isolation and Characterization AIDS Res Hum Retroviruses 1998 14 311 318 9519892
Folks TM Powell D Lightfoote M Koenig S Fauci AS Benn S Rabson A Daugherty D Gendelman HE Hoggan MD Biological and biochemical characterization of a cloned Leu-3-cell surviving infection with the acquired immune deficiency syndrome retrovirus J Exp Med 1986 280 90 3014036 10.1084/jem.164.1.280
Alizon M Wain-Hobson S Montagnier L Sonigo P Genetic variability of the AIDS virus: nucleotide sequence analysis of two isolates from African patients Cell 1986 46 63 74 2424612 10.1016/0092-8674(86)90860-3
Clavel F Guetard D Brun-Vezinet F Chamaret S Rey MA Santos-Ferreira MO Laurent AG Dauguet C Katlama C Rouzioux C Isolation of a new human retrovirus from West African patients with AIDS Science 1986 233 343 346 2425430
Guyader M Emerman M Sonigo P Clavel F Montagnier L Alizon M Genome organization and transactivation of the human immunodeficiency virus type 2 Nature 1987 326 662 669 3031510 10.1038/326662a0
Rey MA Krust B Laurent AG Guetard D Montagnier L Hovanessian AG Characterization of an HIV-2-related virus with a smaller sized extracellular envelope glycoprotein Virology 1989 173 258 267 2683362 10.1016/0042-6822(89)90242-0
Schmit JC Cogniaux J Hermans P Van Vaeck C Sprecher S Van Remoortel B Witvrouw M Balzarini J Desmyter J De Clercq E Vandamme AM Multiple drug resistance to nucleoside analogues and nonnucleoside reverse transcriptase inhibitors in an efficiently replicating human immunodeficiency virus type 1 patient strain J Infect Dis 1996 174 962 968 8896496
Advanced Biological Laboratories S.A
|
15780144
|
PMC1274269
|
CC BY
|
2021-01-04 16:28:14
|
no
|
BMC Infect Dis. 2005 Mar 21; 5:15
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-15
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-101575242310.1186/1471-2334-5-10Research ArticleEmergency vaccination of rabies under limited resources – combating or containing? Eisinger Dirk [email protected] Hans-Hermann [email protected] Thomas [email protected]üller Thomas [email protected] Department of Ecological Modelling, UFZ-Centre for Environmental Research Leipzig/Halle, Leipzig, Germany2 Friedrich-Loeffler-Institut, Federal Research Institute for Animal Health, Wusterhausen, Germany2005 7 3 2005 5 10 10 21 10 2004 7 3 2005 Copyright © 2005 Eisinger et al; licensee BioMed Central Ltd.2005Eisinger et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Rabies is the most important viral zoonosis from a global perspective. Worldwide efforts to combat the disease by oral vaccination of reservoirs have managed to eradicate wildlife rabies in large areas of central Europe and North-America. Thus, repeated vaccination has been discontinued recently on a geographical scale. However, as rabies has not yet been eradicated globally, a serious risk of re-introduction remains. What is the best spatial design for an emergency vaccination program – particularly if resources are limited? Either, we treat a circular area around the detected case and run the risk of infected hosts leaving the limited control area, because a sufficient immunisation level has not yet been built up. Or, initially concentrate the SAME resources in order to establish a protective ring which is more distant from the infected local area, and which then holds out against the challenge of the approaching epidemic.
Methods
We developed a simulation model to contrast the two strategies for emergency vaccination. The spatial-explicit model is based on fox group home-ranges, which facilitates the simulation of rabies spread to larger areas relevant to management. We used individual-based fox groups to follow up the effects of vaccination in a detailed manner. Thus, regionally – bait distribution orientates itself to standard schemes of oral immunisation programs and locally – baits are assigned to individual foxes.
Results
Surprisingly, putting the controlled area ring-like around the outbreak does not outperform the circular area of the same size centred on the outbreak. Only during the very first baitings, does the ring area result in fewer breakouts. But then as rabies is eliminated within the circle area, the respective ring area fails, due to the non-controlled inner part.
We attempt to take advantage of the initially fewer breakouts beyond the ring when applying a mixed strategy. Therefore, after a certain number of baitings, the area under control was increased for both strategies towards the same larger circular area. The circle-circle strategy still outperforms the ring-circle strategy and analysis of the spatial-temporal disease spread reveals why: improving control efficacy by means of a mixed strategy is impossible in the field, due to the build-up time of population immunity.
Conclusion
For practical emergency management of a new outbreak of rabies, the ring-like application of oral vaccination is not a favourable strategy at all. Even if initial resources are substantially low and there is a serious risk of rabies cases outside the limited control area, our results suggest circular application instead of ring vaccination.
==== Body
Background
Rabies is life-threatening for humans [1] and the most important viral zoonosis from a global perspective [2]. In Europe and North-America, wildlife is the main reserve (i.e. foxes or raccoons). Aerial distribution of vaccine filled baits proved to be a method which can be used for controlling rabies in these species, as they are attainable via baits, and an efficient oral rabies virus vaccine is available [3,4]. Therefore, disease managers have been making huge efforts in rabies control over the last 25 years [2,5-9]. Long-term and large scale oral vaccination of wildlife eradicated rabies at the regional scale in central Europe and the Americas [10-16]. Consequently, repeated vaccination in these regions has now ended [8,16,17] and, eventually, its host populations will be completely susceptible to new rabies infection. Therefore, we must be aware of a reintroduction as long as rabies persists anywhere in the world, [18] and we have to develop emergency measures designed for a local outbreak in non-immunised wildlife populations. Thus recent contingency planning appears comparable to the situation in the UK at the end of the last century when an introduction from continental Europe was expected [19-23]. A lot of literature is available from that period concerning how a newly introduced rabies epidemic potentially spreads or will be controlled [21,22,24,25]. However, empirical knowledge has been accumulated in the mean time regarding large-scale field application of oral vaccination, recognition of successful strategies or operating population immunity levels, and termination of repeated baiting. It appears worthwhile to exploit these sources, in order to adjust contingency plans for future rabies control in general and the event of rabies re-introduction in particular.
How should disease management react to re-introduction, i.e. detection of an infection within a rabies-freed area? Revitalising country-wide vaccination campaigns appears to be not very well-adapted to detection of a local rabies outbreak [26]. A WHO [27] recommendation suggests 5,000 sq. km of compact vaccination area as the minimum sustainable strategy, but there are no details regarding the plausible spatial configuration. Field practice demonstrates that modern aerial distribution of vaccine-filled baits performs precisely, even on complex spatial distribution patterns [28,29].
Thus, alternative control application schemes can be considered as emergency strategies, which 1) are able to restrict the spatial extent of the control area, 2) are able to eradicate the disease and finally, 3) are logistically practicable.
Disease managers usually think of combating an outbreak by immediately controlling all areas at risk [17]. But in an emergency, when the outbreak is very local at first, to what spatial extent must a control area be designed to cover "all areas at risk", or in practice, to what distance might the disease spread until a protective immunisation level has been built up [24,26,30]? How can we exclude potential breakouts of infected hosts just before the control measures succeed in the controlled area? The most appealing counter-measure would fence in the epidemic first and eradicate afterwards. The "fence" could be realised by a ring-shaped area of competently vaccinated hosts at an adequate distance from the detected outbreak (see for example [31]). The host population of such a ring area is already well immunised before first infections will reach the inner border of the ring – hence the outbreak is actually contained. But, the ring approach promises another advantage: Although we BAIT an equally sized area compared to circle application, the larger spatial extent of the ring allows for an increased control area because the outer border of the ring is beyond the circle of respective size. Indeed, the inner part of the ring must not be treated in the beginning which could be important if we have to cope with logistic and/or resource limitations immediately after an outbreak (Vos, pers. comm). While aiming at a serious contingency plan, we still have to analyse the comparability of the two different approaches from the epidemiological discussion: Centred on the detected outbreak, we treat either a circular compact area or the equal-sized area, but arrange it in a ring around an omitted inner part (i.e. equal size of baited area, equal baiting program, equal number of baits and same bait density).
We have identified two strategic alternatives: (i) Combating – refers to vaccination applied in a circle which (a) aims at immediate treatment of the surrounding area of the outbreak to keep the number of rabies cases low, but (b) accepts the risk of early breakout due to an unfinished build up of immunisation level. (ii) Containing – refers to ring vaccinations applied at a distance from the outbreak which (a) aims to prevent a breakout of rabid animals through a readymade immunisation level within the ring, before the epidemic reaches it, but (b) accepts higher numbers of cases in the inner part (Fig. 1).
Figure 1 Spatial design of an emergency vaccination. Schematic design of the vaccinated area in an emergency situation (hashed – vaccinated area, blank – not vaccinated, stars – detected cases of rabies). (a) Circle design: The surrounding area of the first detected case of rabies is vaccinated. (b) Ring design: The immediate surroundings are not vaccinated, but a ring-shaped surface around the detection area is vaccinated.
We use an explicit simulation model of the fox-rabies system to compare the different spatial designs of vaccination. We analyse how long the circle or the ring design can keep the rabies epidemic inside the control area. We compare the two spatial designs for the application of mixed strategies, i.e. the definition of most-rewarding-point-in- time, in order to change from ring to circle, as compared to a pure circle strategy. In case of a rabies outbreak in a previously rabies-free region, the results determine which of the strategies should be applied and how to benefit the most from limited resources.
Methods
Model background
We evaluated the management strategies with a model of the rabies-fox system which is tailored to emergency control planning. We applied a spatially-explicit, individual-based, time-discrete modelling approach [32-34]. This approach had already proved practical in studying the spread and control of rabies in foxes [26,35-37] and to provide useful management support [17]. Thus, previous models [26,35,36] were enhanced in order to cope with the new question. The rules of rabies dynamics between the fox family groups, as well as individual dispersal of juveniles after maturity, were adopted from the basic model [35,36]. The spatial unit of a fox group home-range was also maintained because it had proven suitable for studying the disease spread on the regional scale [20,38,39]. Group home-ranges are implemented within regular grid cells, and the obtained results would not change if the grid cells were replaced by irregular shaped home-ranges of a given mean-size. This is because the dispersal movements are modelled relative to group home-range size [35,40,41] and not in metric measures [36,42-45]. This approach incorporates an implicit adjustment of the fox density effect [40]. A metric reference to fox densities within central Europe is realised by means of the mean-fox-group home-ranges (i.e. cells) corresponding to 1 sq. km[46]. Compared to previous rabies models which had addressed an introduction of rabies [20,25,31,47,48], it was necessary to extend the simulation area (i.e. 256 times 256 cells corresponding to ~65.536 sq. km in a central European scenario) in order to allow relevant dimensions of the control area. The representation of the control area by a regular ring or circle within the model is an abstraction. When applied to real landscapes the vaccination areas are non-regular, as they are usually determined by administrative borders, hence certain excess areas must be baited additionally. Thus the geometric simplification in the model represents the required core area, which must at least be covered by the vaccination area defined along administrative units. But, the aim of our study necessitates a further step in scaling down the basic model [49]. The temporal resolution was refined to a weekly time step, since the success of an emergency vaccination depends on the time of introduction of rabies into the fox population, the time until detection of the epidemic, and the timing of the initial vaccination campaign [7,17,20]. The model rules were complemented with ecological characteristics of, and disease transmission between, individual foxes of a group. The individual-based representation enables a locally varying immunisation level due to non-homogeneous bait uptake [50,51] or individual foxes moving across the border of the vaccination area. The effect of these issues might be negligible for vaccination success on a geographical scale, [36] but it becomes serious for the few rabid animals after an outbreak or a spatially limited vaccination area WITHIN a landscape.
Basic fox population model
Each cell comprises a family group [38,52] of age-classified individual foxes (juvenile, adult). Fox groups in the field contain on average 2–3 adults (i.e. 1 male and 1–2 females) before reproduction [52-55]. The pattern is realised in the model by assuming a maximum group size of 5 adults [46,56] together with the mortality and dispersal process. For parameterisation see Table 1.
Table 1 Parameters of the model, default values and reference.
Parameter Value Reference
Population Ecology
PMaxAdultsPerGroup 5 Adjusted to [46]
PMonthlyMortalityAdults 6.1 % [58]
PMonthlyMortalityJuveniles 12.0 % [58]
PLitterMeanSize 5.5 [44,64,117]
PLitterStdDev 1.5 [44]
Juveniles' Dispersal
PDispersalProbabilityNotToLeave 15.0 % [43,44,70]
PDispersalIntrinsicMaxDistance 60 steps [35]
PDispersalMaximumDistance 100 steps [43,70]
PDispersalMortalityPerStep 2.0 % Adjusted to [69]
PDispersalLengthOfDispersalPeriod 8 weeks [67]
Rabies Epidemiology
PIncubationPeriodMean 3.5 weeks [74]
PIncubationPeriodMinimum 2 weeks [74]
PTransmissionProbabilityPerNeighbourGroup 14.0 % Following [36,76]
PTransmissionBasicProbabilityMaiting 14.0 % Following [36,76]
Management Strategy
PManagementDetectionProbablity 2.0 % [24,93]
PManagementBaitDensity 20 bpkm2 [7,10,11,13,17,79]
PVacArea 6,400
10,800
16,000 cells Variable in accordance with simulation experiments
Mortality
Without rabies, adult foxes have a monthly mortality of 6.1% [57,58]. Juveniles are subjected to a monthly mortality of 12% until dispersal [58]. After dispersal they are treated as adults [59-61].
Reproduction
Reproduction is scheduled in the first week of April. All non-empty cells produce a litter of a normally distributed number of cubs with mean of 5.5 and a standard deviation of 1.5 [44,57,58,62-64]. Fox groups which consist of exactly one individual reproduce with 50% probability. This rule accounts for floaters and multiple mating males as well as for non-reproducing males [53,65,66].
Dispersal
With these population dynamics, on average 3.5 juveniles per group emerge in the maturity dispersal (Goretzki, pers. comm.). The dispersal occurs for 8 weeks from October to November [64,67]. Thus, during that phase per time-step one eighth of all cells are selected randomly. Out of each selected cell, all juveniles move consecutively according to the following dispersal algorithm. The dispersing individual is randomly assigned with a main direction from 8 cells of 360 degrees, which is maintained in each step with 50% probability [60,67-69]. In the remaining steps the individual deviates to the left or to the right by one cell with equal probability (i.e. 25%; see [35]). The probability to settle in a cell (PSettle) increases with the distance travelled (PSettleDistance; [40]), but decreases with the number of adult foxes already in there (CrowdingFactor):
PSettle = PSettleDistance(Step) * CrowdingFactor(NumAdultFoxes)
PSettleDistance(Step) = (15% + (1-0.15) * Step/60)
The dispersal of one individual is limited to 100 steps, i.e. a maximum of 100 fox group home-ranges will be passed [43,70]. During each step we assume a mortality of 2% (adjusted to 22% dispersed foxes found dead by [69]). The emerging frequency distribution of dispersal distances is shown in Figure 2.
Figure 2 Dispersal distances. The cumulative distribution of dispersal distances as a result of the model algorithm. 51% move at most 10 cells, whilst only 3.5% disperse farther than 40 cells (indicated by vertical lines). The insert shows the dispersal kernel of the model together with field data observed by Jensen [43]. For this graph, the metrics of the cells are scaled as 0.8*0.8 sq. km.
Rabies transmission
Each fox has a disease state (susceptible, infected, infectious, or immune). The state is updated according to weekly time-steps. If infection was introduced in a cell by neighbourhood contact one adult fox is randomly selected. If this fox is not susceptible but "immune", nothing happens, otherwise its state changes from "susceptible" to "infected". The "infected" fox gets infectious after a negative exponential distribution with a minimum of 2 weeks and an effective mean of 3.5 weeks [71-74]. During the following infectious period of 1 week, a fox can transmit the disease [41,75]. It is assumed that infected cubs will die of rabies, but can only transmit the infection if their incubation period ends after the dispersal [15].
Local Contacts
An infectious fox passes the infection on to all other susceptible foxes within the cell [15,21,41].
Neighbourhood Contacts
If there is at least one infected fox in a cell, then the 8 neighbouring cells have a probability of 14% of getting infected [36,76], i.e. approach of Infection Communities [39] or 'group infection rate' in [41].
Mating Contacts (additionally in January and February)
If there is an infected fox in a cell, any neighbouring cell within a distance of up to 3 cells will be infected with a probability of 0.141, 0.142 and 0.143 respectively [36,41,76,77].
Dispersal Contacts
There are hardly any infections during dispersal [15,53,67]. But juvenile foxes dispersing in their incubation period will cause standard transmission after settlement [15,78].
Distribution of baits
Regional
Standard vaccination protocol on the regional scale comprises biannual campaigns with 18–20 baits distributed per sq km [7,10,11,13,17,79]. Accordingly, two vaccination events are performed in the model: one in the first week of April and one in the second week of September.
Local
Grid cells represent the spatial equivalent of home-ranges of fox families, [38] which do not have equal area size [80] and hence will not receive an equal number of baits [81,82]. We approximate this non-equal assignment of bait pieces to spatial fox group home-ranges by simulating the distribution of effective bait numbers on the ground found for standard aerial delivery (Fig. 3) [82]. The baits randomly drawn to fall into a fox group home-range are assumed to be lost with 80% probability according to empirical findings, i.e. baits lost to competitors [51,83-88], baits unfound or only partly consumed [17,88-90]. The baits remaining in a particular cell are distributed randomly to the respective individuals, independent of their state. The "susceptible" foxes permanently turn towards "immune" two weeks after receiving at least one piece of bait [73,90,91]. With these rules an immunisation level of 70–80% emerges after 2 campaigns (Fig. 4) as empirically documented by vaccination campaigns in the field with 18–20 baits per sq. km [16,88,92,93].
Figure 3 Distribution of baits. The frequency distribution of the number of baits received per fox group according to [82]. The grid model draws from this distribution and accounts for each bait a probability of 80% of being lost (e.g. to competitor animals) before assigning explicit baits randomly to individuals of a group.
Figure 4 Immunisation. Immunisation level found in the fox population of the model (Circle – circle strategy; Ring – ring strategy [PVacArea = 10,800]; Large-scale – vaccination of the whole region). Biannual vaccination is always performed with 50% starting in autumn and 50% in spring. (a) Development of the immunisation level in the vaccinated area over time (100 repetitions): For the large-scale vaccination, both immunisation rate per campaign and final level of population immunity correspond to field data estimated during past control programs [16,87,91-93,104]. Dispersing non-immunised foxes lower the average immunisation level in the circular and ring-shaped vaccination areas. (b) The immunisation level after 3 vaccination campaigns by distance to the centre of the control area (100 repetitions): The immunisation level is lowered at the borders of the vaccinated area.
Emergency vaccination
Rabies detection
The mid cell of the grid receives an external infection during a randomly chosen week of the year. Subsequently, any rabies case will be detected with a probability of 2% [16,24,93,94]. Rabid juvenile foxes will be detected only from August onwards [15].
First vaccination campaign
If one infected fox is detected, we assume a preparation time of 2 months until the first vaccination campaign is scheduled (Vos, pers. comm.). Further campaigns are performed according to the standard protocol: autumn and spring [7,17,79] – with the only exception being that the second campaign will not be performed less than two months after the initial baiting.
Spatial design – ring vs. circle
The relative assessment of the competing spatial designs is based on regular edges. Using the Moore neighbourhood, adjacent and diagonal cells are assumed to have equally scaled distances. The modelled emergency area is always centred on the first detected rabies case ignoring other "infected" cells on the grid.
The parameter vaccination area (PVacArea) corresponds to the maximum amount of cells that could be treated immediately after detection. PVacArea is set to be 6,400, 10,800 or 16,000 cells. The values are selected to provide useful width of the ring area (i.e. 20 km, 30 km or 40 km wide ring areas respectively). The area could be calculated into a necessary amount of baits after scaling the mean area of fox group home-ranges. For instance, in rural Europe fox group density of ~1 per sq. km is agreed [20,51] which fixes the mean area of the cells in the model at 1 sq. km Thus, the amount of baits per campaign used in the three scenarios is roughly: 128,000, 216,000 or 320,000 respectively when applying 20 baits per sq. km.
Circle strategy
The vaccinated area is compact around the detected outbreak and implemented in the model as a solid square. According to PVacArea, the region is 80*80, 104*104 or 126*126 cells respectively.
Ring strategy
60*60 cells remain without baits. This inner part should compensate for an annual spread of rabies of 30 km [27,41]. Around the interior, a ring of cells is assumed to be treated with baits. The treated area is determined by PVacArea and corresponds to a width of 20, 30 or 40 cells respectively [27]. The surrounded area (i.e. not baited + baited) thus covers: 10,000, 14,400 and 19,600 cells respectively. Figure 5(a) and 5(b) show screen shots of the simulations with these strategies treating an area of equal size.
Figure 5 Examples of simulation runs. Example simulation run visualised after an infected fox was detected and the first vaccination was applied (PVacArea = 10,800; (1) – first detection of rabies set as centre of the control area, (2) – first infection of rabies). (a) Experiment 1: circle strategy. (b) Experiment 1: ring strategy. (c) Experiment 2: Ring strategy when PVacArea has been doubled. White – empty group, green – group of "susceptible" foxes, light blue – group with at least one "immune" fox, red – group with at least one "infected" fox, black – group with at least one "infectious" fox. If foxes at different states are within one cell, only the last of the list is shown.
Simulation experiments
Simulation experiments are performed on a grid of 256 × 256 which totals 65,536 cells. We ran each simulation scenario 10,000 times to cover stochastic effects.
Experiment 1 – Containment of the epidemic with fixed resources
We assess which strategy performs better at confining the epidemic inside the control area over the short and long term. The frequency of infected foxes outside the control area provides the quantitative measure. The 3 sizes of vaccinated areas (PVacArea) remain constant throughout a simulation run.
Experiment 2 – Search for the optimal switch point from ring to circle strategy with increasing resources
The aim is to identify the strategy or a mixture of strategies which performs best in final eradication of the epidemic. The inner part of the ring has to be baited in the end to achieve eradication. Thus resource limitation is assumed to be eased at some point in time, and the baited area (PVacArea) is doubled afterwards. The resource extension is assumed with a lag of either 1 or up to 5 vaccination campaigns. Again, initial vaccination areas (PVacArea) will have 3 different sizes but they are doubled after the time lag. Technically, the following spatial configuration is applied in this experiment: Either we already start baiting the circle whose surface gets doubled after the time lag. Or, we start baiting the ring of the same size and after the time lag we continue baiting the circle of doubled surface, which of course contains the ring. Hence, the final treated area is always a solid square of 112*112, 146*146 or 178*178 cells respectively. Figure 5(c) shows an example of the configuration.
Model 'robustness'
We followed the pattern-orientated approach [49,95-98] for validation of the experimental results and qualitative debugging of the model logic [99,100]. Hence, we compared population parameters as re-read from the model to empirical data. The model successfully reproduces the fox ecology (e.g. fox densities during the year from around 1.5 to 3 foxes per sq. km [44,68,77,101], the dispersal distances (Fig. 2), the spread of rabies (Fig. 5b) [102,103], the development of immunisation level (Fig. 4a) [16,87,91-93,104] and the time period up to local eradication of an epidemic (Fig. 6c) [16,91,105]).
Figure 6 Emergency vaccination with fixed resources. Emergency vaccination with fixed resources (R – Ring strategy, C – Circle strategy; PVacArea = 6,400, 10,800, and 16,000; 10,000 repetitions). (a) Risk of a rabies breakout of the control area with respect to the number of vaccination campaigns performed: Initially there are fewer breakouts for the ring strategy, but in the long run, the circle strategy always performs better. (b) Average number of "infected" foxes between consecutive vaccination campaigns. Only simulation runs with rabies inside the control area are considered: As the inner part of the ring is not vaccinated, the epidemic can develop inside. (c) Chance of eradication with respect to the number of vaccination campaigns performed. With circle strategy rabies was eradicated in 80% of the repetitions after three vaccination campaigns (vertical line) but never with ring strategy.
When parameters of the model were altered, only the quantitative results changed, but neither the qualitative results nor the conclusions did. But there is one noteworthy difference between large-scale and local vaccination concerning immunisation level. In emergency control the relatively small baited areas are surrounded by a susceptible neighbourhood and thus non-immunized foxes will regularly disperse into the baited region and vice versa. Indeed, the immunisation level maintained by the circle or ring strategy was measured lower than for the large-scale application (Fig. 4a), in particular at the edges of the control areas (Fig. 4b). Nevertheless, the resulting immunisation in the model was still sufficient to eradicate rabies locally, which is in agreement with recent findings about potential over-baiting during the past control programs in Europe [15,50,82,106,107].
Results
Experiment 1: Containing with fixed resources
In all scenarios we found noteworthy frequency of rabies infections spreading beyond the vaccinated area (Fig. 6a). For the medium amount of resources, about 1% of all simulation runs ended up with breakouts after 2 years. Independent of the amount of applied resources in the long run, the ring strategy performs worse than the circle strategy. In the ring strategy the number of rabies cases rises quickly (Fig. 6b) and the epidemic is not eradicated. On the other hand, by distributing the same resources in the circle strategy, the epidemic often gets eradicated (Fig. 6c). Only for the first two vaccination campaigns the ring strategy performs better in containing rabies.
We detail the spatio-temporal dynamics of the simulated epidemic (Fig. 7 &8) to understand why the sole ring strategy performs badly. The strategy is characterised by an increasing risk of breakouts over time. Early breakouts are seldom due to the distant outer border of the control area (Fig. 7b &8b; black line). But only the ring itself is treated with baits and the epidemic can spread out within the non-vaccinated inner part (Fig. 7b). The growing number of infections close to the baited area challenges the ring (Fig. 7b; olive line) and, eventually, infections beyond the outer border of the ring rise with time (Fig. 8b; compare black and olive line). By contrast, the risk of breakouts associated with the circle strategy decreases with time. Soon after the outbreak, infections occur outside the circular control area, which would still be inside the control area under the respective ring strategy (comp. Fig. 8a &8b). However, the probability of eradication increases with time i.e. the number of vaccination campaigns (Fig. 6c), and thus the risk of still having an epidemic which could breakout diminishes (Fig. 8a; compare black and olive line).
Figure 7 Spatio-temporal spread of rabies. The series show the frequency of infected fox groups at increasing distances from the centre of the vaccination area (average of 40,000 repetitions; solid line- circle strategy, hashed line – ring strategy; PVacArea = 10,800). The diagrams depict the frequency distribution after consecutive vaccination campaigns: black – after one; olive – after 3; green – after 5 campaigns respectively. The shaded areas indicate the extent of the vaccinated areas. (a) Circle strategy: The outbreak is soon suppressed. (b) Ring strategy: Rabies can devolve inside the non-vaccinated part. (c) Mixed strategies – Comparison of the epidemic situation just before control area is doubled (i.e. time lag = 3; hashed line = former ring strategy; solid line former circle strategy): There are more cases of rabies inside the final control area (i.e. shaded) when starting from the ring strategy, as compared to the former circle.
Figure 8 Infections beyond control area. Infections found beyond the control area's outer border (legend see Fig. 7, but notice that the x-axis was cut below 50 cells and the y-axis zoomed in because of the small numbers of recorded outbreaks). Only infections caused by foxes out of the control area are considered. (a) Circle strategy: Some cases might escape the smaller control area at the beginning. (b) Ring strategy: Fewer cases can escape initially, but the number of breakouts rises with time.
Figure 9 illustrates, qualitatively, the risk of breakouts over time. From this risk analysis we expect a crossover point before which the ring strategy has a lower risk and after which the circle strategy has the lower risk of breakouts. To check the prediction, we re-analyse data of Figure 8. We directly equate the risk of breakouts to the number of infections beyond the control area that are actually caused by foxes leaving the control area, and secondary infections are ignored (Fig. 10). Indeed, initially fewer infections are found beyond the outer border of the control area of the ring and later beyond that of the circle. Therefore, we attempt to profit from the initial advantage of ring vaccination by mixing strategies, i.e. starting control with ring vaccination (left down in Fig. 9) and later switching to the circle strategy (right down branch in Fig. 9).
Figure 9 Qualitative evaluation of risk of breakout. The conceptual scheme depicts the risk of rabies breakout over time. Initially, the risk of rabies breakout is higher for the circle design compared to the ring design as the outer border of the vaccinated area is closer to the location of the detected rabies cases. The risk decreases as rabies ceases with repeated control. With the ring design, rabies can develop freely inside and the risk of breakout increases with time. Ring vaccination has to be stopped and eradication of the epidemic started no later than the switch point.
Figure 10 Temporal risk analysis. Frequency of infected fox groups beyond the border of the control area after repeated vaccination campaigns calculated from Fig. 8: Whereas the risk of infections decreases with the circle strategy, the risk increases with the ring strategy. The cross point is around the third vaccination campaign.
Experiment 2: Eradication with increasing resources
According to Figure 10 we expect at least one mixed strategy (switching from ring to circle after k baitings) to perform better than the continuous circle approach. Following this idea, we conducted experiment 2: The initial strategy is changed after k vaccination campaigns towards a circle application. In practice, the change could commence when resource limitations are overridden. Thus resources are doubled after the switch and the final circular control areas are IDENTICAL for all mixed strategies, i.e. ring-circle and circle-circle.
Surprisingly, the strategy which immediately starts with a circle is still favourable (Fig. 11a). Indeed, the mixed strategy results in more breakouts and less eradication. Eradication also takes longer when using the mixed strategy as compared to the circle strategy (Fig. 11b), because the time lag before vaccination starts in the inner part of the ring is simply added to the time until eradication.
Figure 11 Emergency vaccination with increasing resources. The vaccinated area (PVacArea) is doubled with a time lag of 1 up to 5 vaccination campaigns (R – Ring strategy, C – Circle strategy; PVacArea = 6,400, 10,800, and 16,000; 10,000 repetitions). (a) Risk of rabies breakout from the control area with respect to the time lag when the vaccinated area was doubled: We cannot find the predicted switch point (see text); the circle strategy still performs better. (b) Chance of eradication with respect to the time lag when vaccinated area was doubled: In contrast to experiment 1 there is eradication now in the ring strategy; however the control success is shifted by the time lag.
Discussion
Expert knowledge and biological data about the host, the epidemiology of the infection or even the efficacy of management measures are quite often vague in the sense that they are never measured, they are examined with highly differing results or they are even difficult to sample precisely [52,108]. Nevertheless, it is necessary to overcome these uncertainties because management decisions have to be made [7]. Usually, modelling studies tend to select one particular configuration and substantiate the parameter choice with the help of logical arguments. We suggest an approach that is oriented to robust conclusions for practical management in terms of different or even antagonistic model assumptions. There are initial studies in the literature [109,110] which opt for the development of a more general methodology. The approach parallels standard techniques of model validation or sensitivity analysis. But we are no longer troubleshooting at the level of particular values which the model acquires in relation to slight changes in assumed parameters. We are only interested in changes at the level of conclusions made for management application, i.e. whether there are hypothetical scenarios which could falsify the management decision just derived from the model results. That is how practical management often performs [15]. Thus, while targeting useful support for these decisions, we have already covered the need for 'robustness' during the shaping of our management proposal.
We compared two spatial strategies of local emergency vaccination for controlling a rabies outbreak. One refers to the immediate control of infection within a smaller treated area. The alternative was theorised to overcome the drawback of a spatially limited strategy by providing equal resources in a ring around the affected area which contains the infection at the cost of more cases in the centre.
Simulation of the two spatial strategies revealed the true dynamics of the models. The ring strategy in general does not outperform the circle strategy. The predicted advantage of the ring strategy can only be found in the short term (Fig. 6a). Therefore, we determined mixed strategies and searched for the most rewarding point in time for switching between the ring strategy, which was better in the short term, and the circle strategy, which was better in the long term (Fig. 10).
Why aren't we able to identify the switch point as suggested by the risk analysis? Figure 10 actually proposes a switch around the third campaign. However, no matter which point we chose for switching from ring to circle in experiment 2, we did not find a clear advantage at all for the mixed strategy. There is only one plausible reason for this disagreement between the mind and the simulation model. The switch between the two strategies comprises a time lag until the protection level is reached inside the ring, i.e. until the inner part of the former ring actually acts like a circle. Our qualitative risk assessment (Fig. 9) assumes a perfect change between strategies. In practice however, we have to consider the temporal delay of building up population immunity which was shown to be at least 2 vaccination campaigns [16,87,91-93,104] in noteworthy agreement with our simulations (Fig. 4a). Thus the change in strategy has to take place two baitings in advance of the theoretical suggestion in order to profit with the mixed strategy. But the ideal switch point found in the simulation is around the 3rd baiting campaign (Fig. 10). After subtracting the building up time of 2 vaccination campaigns in practice we need to change at baiting 1, i.e. we must apply the circle strategy from the very beginning.
Consequently, in the field we cannot benefit from the alternative baiting scheme and hence are forced to focus contingency planning on a compact control area around the detected outbreak. After testing the respective models, we can reject a-priori any pure vaccination field trial that attempts to distribute vaccine baits with a ring-like strategy.
Our findings are not contradicted by the successful application of cordon-sanitaire vaccination at borders of large-scale vaccination areas in the field [17,27,111-115]. In fact, the basic difference between a ring-like vaccination around a new outbreak and the cordon-sanitaire is the aim of control: In the outbreak situation we do not need to accept the rabies inside the ring, but in the second, the border situation, we have to. Rabies persists "on the other side" of the cordon, i.e. if neighbouring countries have not (successfully) combated the disease. Although the ring-like emergency vaccination does provide some protection for the surrounding area in the same manner in which the cordon-sanitaire vaccination does, in the emergency situation we aim for ultimate eradication, and in order to achieve this, our results clearly require the treatment of a compact circle-like control area.
The only threat for success of control is the migration of infected foxes from the limited area under treatment into the non-vaccinated surroundings. We cannot limit the distance infected foxes disperse, but we can reduce the number of them by means of the control itself. It is the circle strategy in which the number of rabies infections is lowered right from the beginning.
We present only 3 widths of the ring (Table 2). This is because ring width below 20 km cannot be expected to be protective [17]. Even though we analysed ring dimensions of 50 km and more in accordance with EU recommendations [17], there is no need to present these results. The difference between ring and circle strategy is less pronounced compared to that of the 40 km ring (Fig. 11a). Indeed, wider rings do provide decreasing gain in the outer radius due to the non-linear relationship between radius and area (Table 2, line 4). When the inner non-vaccinated part is reduced, there is no useful gain left, thus circle and ring become equivalent (Table 2, line 5). On the other hand, an extended inner part reduces the protective ability of the ring and simultaneously triples the non-vaccinated area (Table 2, line 6).
Table 2 Geometry of circle vs. ring. Gain in the maximum distance from the point of outbreak by using a ring instead of a circle. Presented scenarios with input parameters framed. Bold – respective calculations for scenarios not presented. Smaller inner non-vaccinated area reduces the gain, whereas a large inner radius reduces the thickness (i.e. protective ability) of the ring.
Vaccinated area [km2] ring thickness [km] ring – inner radius [km] ring – outer radius [km] circle – outer radius [km] gain [km]
6400 20 30 50 40 10
10800 30 30 60 52 8
16000 40 30 70 63 7
22000 50 30 80 74 6
10800 43 10 53 52 1
10800 22 50 72 52 20
If the ring design strategically loses, i.e. in eradicating rabies, are there other benefits which outweigh the disadvantage? There are two other potential benefits to consider: economy and public health. With modern aerial bait disposal controlled by the GPS, logistic costs do not differ substantially between the two spatial designs (Mürke, pers. comm.). The cost is mainly linked to the total number of baits needed for the program. However, we demonstrated that in the ring design the time lag until the spared inner part is vaccinated directly adds to the time until eradication (Fig. 11b). This increases the cost of the ring design as compared to the earlier eradication in the circle design.
The remaining public health issue is the principal objective [7,71,116]. Eradication of the disease is the only method for achieving this objective [93]. However, within the inner part of the ring the epidemic roams freely and thus imposes a risk to humans and livestock which makes it less competitive than the compact circle approach. Additionally, eradication takes longer and thus the threat to public health is prolonged. Consequently, we can rule out the ring as a non-viable approach in terms of eradication, economy and public health.
In all respects we concluded that the circle performs better. But we still have to deal with possible early breakouts. Whilst zero risk strategies perhaps represent political demand, they are probably neither possible nor the most cost-effective approach. However, additional measures could be applied for an improvement of the performance of the circle strategy and will be considered in the ongoing analysis: Firstly, better monitoring programs could lower the time until detection of an outbreak, which consequently leads to an earlier eradication. Secondly, it is not clear whether immediate vaccination with a risk of imperfect placement performs better than waiting with the first vaccination until a monitoring program has provided a better understanding of the spatial extent of the outbreak. Thirdly, circles baited with spatially varying density of baits could provide both, the required fast suppression of the epidemic and the largest possible control area (see [31] for a combined simulation of poisoning and ring vaccination). And finally, follow-up programs can be designed. Indeed, the circle strategy with a control area of 10.800 sq. km has already provided a very low likelihood of breakouts. Thus, the strategic approach could extend the initial circular area to a non-circular control area (but still vaccinated on a regular basis), according to detected breakouts. Raised awareness after the reintroduction of rabies and particular border surveillance around the baited area, would guarantee fast detection of breakouts. Thus, we recognise the need for a more detailed cost-benefit, which explores the cost of extensions of the control area, versus the benefit of reducing the amount of resources applied to the initial hazard area.
Conclusion
If vaccination is the only approved measure for fighting a rabies outbreak within a completely susceptible fox population, then the only feasible contingency plan is to vaccinate a compact area centred on the epidemic. The ring strategy which leaves an inner part non-vaccinated must be ruled out in all concerns: strategically, since it under-performs in eradication levels, economically, since eradication takes longer and public health, since it allows more cases of rabies.
Yet even in the circle strategy, there remains some risk of early breakouts of rabies from the control area. Thus, further studies should concentrate on optimizing the emergency strategy concerning timing and benefits of additional monitoring programs. Furthermore, a detailed cost-benefit analysis of potential strategic alternatives is needed in order to improve the outcome of a contingency plan.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
DE and HHT developed the model, and drafted the manuscript. DE implemented the code, and performed the simulation experiments. HHT, TS and TM developed the alternative strategy. TM provided the background of practical rabies management. All read and approved the final manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
DE was partly funded by Impfstoffwerke Dessau-Tornau GmbH. We gratefully acknowledge improvements according to referees' comments.
==== Refs
Rupprecht CE Smith JS Fekadu M Childs JE The ascension of wildlife rabies: a cause for public health concern or intervention? Emerg Infect Dis 1995 1 107 114 8903179
Hanlon CA Childs JE Nettles VF Recommendations of a national working group on prevention and control of rabies in the United States. III: Rabies in wildlife J Am Vet Med Assoc 1999 215 1612 1619 14575027
Johnston DH Voigt DR MacInnes CD Bachmann P Lawson KF Rupprecht CE An aerial baiting system for the distribution of attenuated or recombinant rabies vaccines for foxes, racoons and skunks Rev Infect Dis 1988 10 660 664
Johnston DH Bachmann P Lawson KF MacInnes CD Voigt DR Pond BA Nunan CP Ayers NR Bögel K, Meslin FX and Kaplan M Design considerations for aerial bait distribution of rabies vaccines Wildlife Rabies Control 1992 Kent, Wells Medical Ltd. 160 167
Meslin FX Zoonoses in the world - Current and future trends Schweiz Med Wochenschr 1995 125 875 878 7770747
Stöhr K Karge E Gädt H Kokles R Ehrentraut W Witt W Fink HG Orale Immunisierung freilebender Füchse gegen Tollwut - Vorbereitung und Durchführung der ersten Feldversuche in den ostdeutschen Bundesländern. Mh Vet Med 1990 45 782 786
Stöhr K Meslin FX Progress and setbacks in the oral immunisation of foxes against rabies in Europe Vet Rec 1996 139 32 35 8839488
Müller WW Review of reported rabies cases data in Europe to the WHO Collaborating Centre in Tübingen from 1997 to 2000 Rabies Bulletin Europe 2000 24 11 19
Zanoni RG Kappeler A Müller UM Müller C Wandeler A Breitenmoser U Tollwutfreiheit der Schweiz nach 30 Jahren Fuchstollwut / Rabies free status of Switzerland after 30 years of fox rabies Schweiz Arch Tierheilkd 2000 142 423 429 11004890
Breitenmoser U Müller U Kappeler A Zanoni RG The final stage of rabies in Switzerland [German] Schweiz Arch Tierheilkd 2000 142 447 454 11004893
Brochier B Deschamps P Costy F Hallet L Leuris J Villers M Péharpré D Mosselmans F Beier R Lecomte L Mullier P Roland H Bauduin B Kervyn T Renders C Escutenaire S Pastoret PP Elimination de la rage en Belgique par la vaccination du renard roux (Vulpes vulpes) Ann Med Vet 2001 145 293 305
Müller WW Where do we stand with oral vaccination of foxes against rabies in Europe Arch Virol Suppl 1997 13 83 94 9413528
Bruyere V Janot C La France bientôt déclarée officiellement indemne de rage Bulletin épidémiologique mensuel de la rage animale en France 2000 30 1 2
Müller T Schlüter H Oral immunization of red foxes (Vulpes vulpes L.) in Europe – a review J Etlik Vet Microbiol 1998 9 35 39
Vos A Oral vaccination against rabies and the behavioural ecology of the red fox (Vulpes vulpes) J Vet Med B 2003 50 477 483 10.1046/j.1439-0450.2003.00712.x
MacInnes CD Smith SM Tinline RR Ayers NR Bachmann P Ball DGA Calder LA Crosgrey SJ Fielding C Hauschildt P Honig JM Johnston DH Lawson KF Nunan CP Pedde MA Pond B Stewart RB Voigt DR Elimination of rabies from red foxes in eastern Ontario J Wildlife Dis 2001 37 119 132
European Commission: Report of the Scientific Committee on Animal Health and Animal Welfare: The oral vaccination of foxes against rabies [http://www.europa.eu.int/comm/food/fs/sc/scah/out80_en.pdf] 2002 Brussels
MacKenzie D Will rabies bite back? New Sci 1997 24 25
Harris S Smith GC If rabies comes to Britain New Sci 1990 128 20 21
Smith GC Harris S Rabies in urban foxes (Vulpes vulpes) in Britain: the use of a spatial stochastic simulation model to examine the pattern of spread and evaluate the efficacy of different control regimes Philos Trans R Soc Lond Biol Sci 1991 334 459 479 1686115
White PCL Harris S Smith GC Fox contact behaviour and rabies spread: a model for the estimation of contact probabilities between urban foxes at different population densities and its implications for rabies control in Britain J Appl Ecol 1995 32 693 706
Evans ND Pritchard AJ A control theoretic approach to containing the spread of rabies IMA J Math Appl Med Biol 2001 18 1 23 11339335
MAFF Report of the committee of inquiry into rabies Final report 1971 London, HMSO, Ministry of agriculture fisheries and food
Bacon PJ The consequences of unreported fox rabies J Environ Manage 1981 13 195 200
Smith GC Modelling rabies control in the UK: the inclusion of vaccination Mammalia 1995 59 629 637
Thulke HH Tischendorf L Staubach C Selhorst T Jeltsch F Schlüter H Wissel C The spatio-temporal dynamics of a post vaccination resurgence of rabies in foxes and emergency control planning Prev Vet Med 2000 47 1 21 11018731 10.1016/S0167-5877(00)00167-7
WHO Report of WHO Seminar on wildlife rabies control, Geneva 2-5 July 1990 1992 Geneva, WHO
Vos A Mührke HH Holzhofer E Gschwender P Schuster P A satellite navigated and computer supported fully automatic system for distributing oral vaccine-baits against rabies: SURVIS: 2001. 2001 Peterborough, Canada, Proceedings of the XIIth International Meeting on Advances in Rabies Research and Control in the Americas, Nov. 12-16 28 11344588
Müller T Stöhr K Teuffert J Stöhr P Erfahrungen mit der Flugzeugbeköderung von Ködern zur oralen Immunisierung der Füchse gegen Tollwut in Ostdeutschland Dtsch Tierärztl Wochenschr 1993 100 203 207
van den Bosch F Metz JAJ Diekman O The velocity of spatial population expansion J Math Biol 1990 28 529 565 10.1007/BF00164162
Smith GC Wilkinson D Modeling control of rabies outbreaks in red fox populations to evaluate culling, vaccination, and vaccination combined with fertility control J Wildl Dis 2003 39 278 286 12910754
Mollison D Kuulasmaa K Bacon PJ Spatial Epidemic Models: Theory and Simulations Population Dynamics of Rabies in Wildlife 1985 London, Academic Press 291 309
Durrett R Mollison D Spatial Epidemic Models Epidemic Models - Their Structure and Relation to Data 1995 Cambridge, Cambridge University Press 187 201
DeAngelis DL Gross LJ Individual-based models and approaches in ecology: populations, communities and ecosystems 1992 London, Chapman Hall
Jeltsch F Müller MS Grimm V Wissel C Brandl R Pattern formation triggered by rare events: lessons from the spread of rabies Proc R Soc Lond B 1997 264 495 503 9149424 10.1098/rspb.1997.0071
Tischendorf L Thulke HH Staubach C Müller MS Jeltsch F Goretzki J Selhorst T Müller T Schlüter H Wissel C Chance and risk of controlling rabies in large-scale and long-term immunized fox populations Proc R Soc Lond B 1998 265 839 846 9633109 10.1098/rspb.1998.0368
Suppo C Naulin JM Langlais M Artois M A modelling approach to vaccination and contraception programmes for rabies control in fox populations Proc R Soc Lond B 2000 267 1575 1582 11007334 10.1098/rspb.2000.1180
Macdonald DW The ecology of carnivore social behaviour Nature 1983 301 379 384 10.1038/301379a0
Thulke HH Tischendorf L Staubach C Müller MS Schlüter H Neue Antworten zur Frage der weiteren Tollwutbekämpfung in Deutschland Dtsch Tierärztl Wochenschr 1997 104 492 495
Trewhella WJ Harris S McAllister FE Dispersal distance, home range size and population density in the red fox (Vulpes vulpes): a quantitative analysis J Appl Ecol 1988 25 423 434
Macdonald DW Bacon PJ Fox society, contact rate and rabies epizootiology Comp Immunol Microbiol Inf Dis 1982 5 247 256 10.1016/0147-9571(82)90045-5
Garnerin P Hazout S Valleron AJ Estimation of two epidemiological parameters of fox rabies: the length of incubation period and the dispersion distance of cubs Ecol Modell 1986 33 123 135 10.1016/0304-3800(86)90036-0
Jensen B Movements of red fox (Vulpes vulpes L.) in Denmark investigated by marking and recovery Danish Review of Game Biology 1973 8 3 20
Goretzki J Ahrens M Stubbe C Tottewitz F Sparing H Gleich E Zur Ökologie des Rotfuchses (Vulpes vulpes L.,1758) auf der Insel Rügen: Ergebnisse des Jungfuchsfanges und der Jungfuchsmarkierung Beitr Jagd- u Wildforsch 1997 22 187 199
Englund J Zimen E Yearly variations of recovery and dispersal rates of fox cubs tagged in swedish coniferous forest Biogeographica Vol18 - The Red Fox 1980 The Hague, Dr.W.Junk B.V. Publishers 195 205
Goszczynski J Home ranges in red fox: territoriality diminishes with increasing area Acta Theriol 2002 47 103 114
David JM Andral L Artois M Computer simulation model of the epi-enzootic disease of vulpine rabies Ecol Modell 1982 15 107 125 10.1016/0304-3800(82)90056-4
Ball FG Bacon PJ Spatial models for the spread and control of rabies incorporating group size Population Dynamics of Rabies in Wildlife 1985 London, Academic Press 197 222
Thulke HH Grimm V Müller MS Staubach C Tischendorf L Wissel C Jeltsch F From pattern to practice: a scaling-down strategy for spatially explicit modelling illustrated by the spread and control of rabies Ecol Modell 1999 117 179 202 10.1016/S0304-3800(98)00198-7
Johnston DH Improving efficiency and reducing costs in oral rabies vaccination programs: 2001. 2001 Peterborough, Canada, Proceedings of the XIIth International Meeting on Advances in Rabies Research and Control in the Americas, Nov. 12-16 28 28
Trewhella WJ Harris S Smith GC Nadian AK A field trial evaluating bait uptake by an urban fox (Vulpes vulpes) population J Appl Ecol 1991 28 454 466
Cavallini P Variation in the social system of the red fox Ethol Ecol Evol 1996 8 323 342
Niewold FJJ Zimen E Aspects of the social structure of red fox populations: a summary Biogeographica Vol18 - The Red Fox 1980 The Hague, Dr.W.Junk B.V. Publishers 185 193
Macdonald DW Zimen E Social factors affecting reproduction amongst red foxes (Vulpes vulpes L., 1758) Biogeographica Vol18 - The Red Fox 1980 The Hague, Dr.W.Junk B.V. Publishers 123 175
Baker PJ Robertson CPJ Funk SM Harris S Potential fitness benefits of group living in the red fox, Vulpes vulpes Anim Behav 1998 56 1411 1424 9933538 10.1006/anbe.1998.0950
Von Schantz T Female cooperation, male competition, and dispersal in red fox, Vulpes vulpes OIKOS 1981 37 63 68
Ansorge H Daten zur Fortpflanzungsbiologie und zur Reproduktionsstrategie des Rotfuchses, Vulpes vulpes, in der Oberlausitz Säugetierkd Inf 1990 3 185 199
Stiebling U Untersuchungen zur Habitatnutzung des Rotfuches, (Vulpes vulpes L., 1758), in der Agrarlandschaft als Grundlage für die Entwicklung von Strategien des Natur- und Artenschutzes sowie der Tierseuchenbekämpfung 2000 PhD thesis. HU Berlin
Tackmann K Löschner U Mix H Staubach C Thulke HH Ziller M Conraths FJ A field trial to control Echinococcus multilocularis-infections of the red fox (Vulpes vulpes) in an endemic focus in Brandenburg, Germany Epidemiol Infect 2001 127 577 587 11811893
Harris S Trewhella WJ An analysis of some of the factors affecting dispersal in an urban fox (Vulpes vulpes) population. J Appl Ecol 1988 25 409 422
Allen SH Sargeant AB Dispersal patterns of red foxes relative to population density J Wildl Manage 1993 57 526 533
Stubbe M Stubbe W Zur Populationsbiologie des Rotfuchses Vulpes vulpes (L.) Hercynia N F , Leipzig 1977 14 160 177
Vos AC Aspekte der Dynamik einer Fuchspopulation nach dem Verschwinden der Tollwut 1993 PhD thesis. Ludwig-Maximillians-Universität München, Forstwirtschaftliche Fakultät
Lloyd HG The red fox 1980 London, B.T.Batsford Ltd.
Macdonald DW Bunce RGH Fox populations, habitat characterization and rabies control J Biogeogr 1981 8 145 151
Zimen E Long range movements of the red fox, Vulpes vulpes Acta Zool Fennica 1984 171 267 270
Storm GL Montgomery GG Fox MW Dispersal and Social Contact among Red Foxes: Results from Telemetry and Computer Simulation The Wild Canids 1975 New York, Van Nostrand Reinhold Co. 237 246
Storm GL Andrews RD Phillips RL Bishop RA Siniff DB Tester JR Morphology, reproduction, dispersal, and mortality of midwestern red fox populations Wildl Monogr 1976 49 1 82
Woollard T Harris S A behavioural comparison of dispersing and non-dispersing foxes (Vulpes vulpes) and an evaluation of some dispersal hypotheses J Anim Ecol 1990 59 709 722
Steck F Wandeler A The epidemiology of fox rabies in Europe Epidemiol Rev 1980 2 71 96 7000539
Aubert MFA Bögel K, Meslin FX and Kaplan M Epidemiology of fox rabies Wildlife rabies control 1992 Kent, Wells Medical Ltd. 9 18
Bacon PJ Bacon PJ A Systems Analysis of Wildlife Rabies Epizootics Population Dynamics of Rabies in Wildlife 1985 London, Academic Press 109 130
Barrat J Aubert MF Current status of fox rabies in Europe Onderstepoort J Vet Res 1993 60 357 363 7777320
Reichert HU Simulationsstudien zur Ausbreitung und Bekämpfung der Tollwut bei Füchsen mit einem stochastischen, räumlichen Modell 1989 PhD thesis. University Frankfurt a. Main
Charlton KM Campball JB The pathogenesis of rabies Rabies 1988 Bosten, Cluever Acad. Publ. 101 150
Müller MS Ein gitterbasiertes Modell zur Tollwutausbreitung bei Füchsen (Vulpes vulpes) 1995 Diploma thesis. University Marburg/Lahn
Toma B Andral L Epidemiology of fox rabies Adv Virus Res 1977 21 1 36 324250
Kappeler A Untersuchungen zur Altersbestimmung und zur Altersstruktur verschiedener Stichproben aus Rotfuchspopulationen (Vulpes vulpes) in der Schweiz 1985 Universität Bern
Wachendörfer G Frost JW Gutmann B Hofmann J Schneider LG Eskens U Dingeldein W Experiences with oral immunization of foxes against rabies in Hesse [German] Tierärztl Praxis 1986 14 185 196
Dekker JJA Stein A Heitkonig IMA A spatial analysis of a population of red fox (Vulpes vulpes) in the Dutch coastal dune area J Zool 2001 255 505 510
Breitenmoser U Müller U How to do the wrong thing with the highest possible precision - a reflection on the use of GPS in rabies vaccination campaigns Rabies Bulletin Europe 1997 21 11 13
Thulke HH Selhorst T Müller T Wyszomirski T Müller U Breitenmoser U Assessing anti-rabies baiting - what happens on the ground? BMC Infect Dis 2004 4 9 15113448 10.1186/1471-2334-4-9
Linhart SB Some factors affecting the oral rabies vaccination of free-ranging carnivores Rev Sci Tech 1993 12 109 113 8518438
Selhorst T Thulke HH Müller T Cost-efficient vaccination of foxes (Vulpes vulpes) against rabies and the need for a new baiting strategy Prev Vet Med 2001 51 95 109 11530197 10.1016/S0167-5877(01)00209-4
Bachmann P Bramwell RN Fraser SJ Gilmore DA Johnston DH Lawson KF MacInnes CD Matejka FO Miles HE Pedde MA Wild carnivore acceptance of baits for delivery of liquid rabies vaccine J Wildl Dis 1990 26 486 501 2250325
Marks CA Bloomfield TE Bait uptake by foxes (Vulpes vulpes) in urban Melbourne: the potential of oral vaccination for rabies control Wildl Res 1999 26 777 787
Brochier B Thomas I Iokem A Ginter A Kalpers J Paquot A Costy F Pastoret PP A field trial in Belgium to control fox rabies by oral immunisation Vet Rec 1988 123 618 621 3218039
Vos A Müller T Schuster P Schlüter H Neubert A Oral vaccination of foxes against rabies with SAD B19 in Europe, 1983-1998: a review Veterinary Bulletin 2000 70 1 6
Kappeler A Die orale Immunisierung von Füchsen gegen Tollwut in der Schweiz 1991 PhD thesis. Universität Bern
Müller WW Review of rabies in Europe Med Pregl 1998 1 9 74 9769649
Masson E Aubert MFA Barrat J Vuillaume P Comparison of the efficacy of the antirabies vaccines used for foxes in France Veterinary Research 1996 27 255 266 8767887
Stöhr K Stöhr P Müller T Orale Fuchsimpfung gegen Tollwut - Ergebnisse und Erfahrungen aus den ostdeutschen Bundesländern Tierärztl Umschau 1994 49 203 211
Schlüter H Müller T Tollwutbekämpfung in Deutschland. Ergebnisse und Schlubfolgerungen aus über 10-jähriger Bekämpfung Tierärztl Umschau 1995 50 748 758
Braunschweig AV Zimen E Ein Modell für die Fuchspopulation in der Bundesrepublik Deutschland Biogeographica Vol18 - The Red Fox 1980 The Hague, Dr.W.Junk B.V. Publishers 97 106
Grimm V Mathematical models and understanding in ecology Ecol Modell 1994 75/76 641 651 10.1016/0304-3800(94)90056-6
Grimm V Frank K Jeltsch F Brandl R Uchmánski J Wissel C Pattern-oriented modelling in population ecology Sci Total Environ 1996 183 151 166 10.1016/0048-9697(95)04966-5
Wiegand T Jeltsch F Hanski I Grimm V Using pattern-oriented modeling for revealing hidden information: a key for reconciling ecological theory and application OIKOS 2003 100 209 222 10.1034/j.1600-0706.2003.12027.x
Railsback SF Concepts from complex adaptive systems as a framework for individual-based modelling Ecol Modell 2001 139 47 62 10.1016/S0304-3800(01)00228-9
Grimm V Visual debugging: a way of analyzing, understanding, and communicating bottom-up simulation models in ecology Nat Resourc Model 2002 15 23 38 10.1216/nrm/1030539095
Hansen F Tackmann K Jeltsch F Wissel C Thulke HH Controlling Echinococcus multilocularis - ecological implications of field trials Prev Vet Med 2003 60 91 105 12900151 10.1016/S0167-5877(03)00084-9
Wandeler A Wachendörfer G Förster U Krekel H Schale W Müller J Steck F Rabies in wild carnivores in central Europe: I. Epidemiological studies Zentralblatt für Veterinärmedizin B 1974 21 735 756
Hengeveld R Hengeveld R The stochastic structure of the wave front of rabies in central Europe Dynamics of biological invasions 1989 London, Chapman & Hall 116 125
Sayers BMA Ross JA Saengcharoenrat P Mansourian BG Bacon PJ Pattern analysis of the case occurrences of fox rabies in Europe Population dynamics of rabies in wildlife 1985 London, Academic Press 235 254
Brochier B Costy F Pastoret PP Elimination of fox rabies from belgium using a recombinant vaccinia-rabies vaccine - an update Vet Microbiol 1995 46 269 279 8545965 10.1016/0378-1135(95)00091-N
Masson E Bruyere V Vuillaume P Lemoyne S Aubert M Rabies oral vaccination of foxes during the summer with the VRG vaccine bait Vet Rec 1999 30 595 605 1059640
Farry SC Henke SE Beasom SL Fearneyhough MG Efficacy of bait distributional strategies to deliver canine rabies vaccines to coyotes in southern Texas J Wildl Dis 1998 34 23 32 9476222
Thomson PC Algar D The uptake of dried meat baits by foxes and investigations of baiting rates in Western Australia Wildl Res 2000 27 451 456
Zimen E Zimen E Fox social ecology and rabies control Biogeographica Vol18 - The Red Fox 1980 The Hague, Dr.W.Junk B.V. Publishers 277 285
Selhorst T Thulke HH Müller T Threshold analysis of cost-efficient oral vaccination strategies against rabies in fox (vulpes vulpes) populations: 2000. 2000 Edinborough, Society for Veterinary Epidemiology and Preventive Medicine 71 84
Hansen F Jeltsch F Tackmann K Staubach C Thulke HH Processes leading to a spatial aggregation of Echinococcus multilocularis in its natural intermediate host Microtus arvalis Int J Parasit 2004 34 37 44 10.1016/j.ijpara.2003.10.003
Ulbrich F Ergebnisse der oralen Fuchsimmunisierung gegen Tollwut im Freistaat Sachsen im Zusammenhang mit der grenzüberschreitenden Tollwutgefährdung Tierärztl Umschau 1999 54 219 223
Murray JD Stanley EA Brown DL On the spatial spread of rabies among foxes Proc R Soc Lond B 1986 229 111 150 2880348
Murray JD Seward WL On the spatial spread of rabies among foxes with immunity J Theor Biol 1992 156 327 348
Brandl R Jeltsch F Grimm V Müller MS Kummer G Modelle zu lokalen und regionalen Aspekten der Tollwutausbreitung Z Ökol Nat schutz 1994 3 207 216
Wandeler A Capt S Gerber H Kappeler A Kipfer R Rabies epidemiology, natural barriers and fox vaccination Parassitologia 1988 30 53 57 3268773
Hanlon CA Niezgoda M Morrill PA Rupprecht CE The incurable wound revisited: progress in human rabies prevention? Vaccine 2001 19 2273 2279 11257347 10.1016/S0264-410X(00)00516-8
Gortazar C Ferreras P Villafuerte R Martin M Blanco JC Habitat related differences in age structure and reproductive parameters of red foxes Acta Theriol 2003 48 93 100
|
15752423
|
PMC1274270
|
CC BY
|
2021-01-04 16:28:15
|
no
|
BMC Infect Dis. 2005 Mar 7; 5:10
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-10
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-131576928910.1186/1471-2334-5-13Technical AdvanceA dual fluorescent multiprobe assay for prion protein genotyping in sheep Van Poucke Mario [email protected] Jo [email protected] Marc [email protected] Zeveren Alex [email protected] Luc J [email protected] Department of Animal Genetics and Breeding, Faculty of Veterinary Medicine, Ghent University, Heidestraat 19, B-9820 Merelbeke, Belgium2 Center for Medical Genetics, Faculty of Medicine and Health Sciences, Ghent University Hospital, De Pintelaan 185, B-9000 Ghent, Belgium2005 15 3 2005 5 13 13 22 9 2004 15 3 2005 Copyright © 2005 Van Poucke et al; licensee BioMed Central Ltd.2005Van Poucke et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Scrapie and BSE belong to a group of fatal, transmissible, neurodegenerative diseases called TSE. In order to minimize the risk of natural scrapie and presumed natural BSE in sheep, breeding programmes towards TSE resistance are conducted in many countries based on resistance rendering PRNP polymorphisms at codons 136 (A/V), 154 (R/H) and 171 (R/H/Q). Therefore, a reliable, fast and cost-effective method for routine PRNP genotyping in sheep, applicable in standard equipped molecular genetic laboratories, will be a vital instrument to fulfill the need of genotyping hundreds or thousands of sheep.
Methods
A dual fluorescent multiprobe assay consisting of 2 closed tube PCR reactions containing respectively 4 and 3 dual-labelled fluorescent ASO probes for the detection in real-time of the 7 allelic variants of sheep PRNP mentioned above.
Results
The assay is succesfully performed using unpurified DNA as a template for PCR, without any post-PCR manipulations and with semi-automatic determination of the PRNP genotypes. The performance of the assay was confirmed via PCR-RFLP and sequencing in a cross-validation study with 50 sheep.
Conclusions
We report the development and validation of a robust, reliable and reproducible method for PRNP genotyping of a few to many sheep samples in a fast, simple and cost-effective way, applicable in standard equipped molecular genetic laboratories. The described primer/probe design strategy can also be applied for the detection of other polymorphisms or disease causing mutations.
==== Body
Background
TSE or prion diseases are a group of fatal, transmissible, neurodegenerative diseases occurring in human (e.g. CJD and vCJD) and animals (e.g. scrapie in sheep and BSE in cattle). These complex diseases (for reviews see [1]) are characterized by an accumulation in the central nervous system of PrPSc, a disease-causing isoform of the host encoded cellular prion protein, termed PrPC [2]. In sheep, it is known that polymorphisms at codons 136 (A/V), 154 (R/H) and 171 (R/H/Q) of PRNP, the gene encoding PrPC, are associated with TSE resistance/susceptibility [3].
Based on epidemiological evidences, it is almost certain that scrapie is not transmissible to human, in contrast to BSE, where there is strong evidence that it causes vCJD in human [4]. Moreover, also sheep can be infected by BSE via experimental transmission, although the ARR/ARR PRNP genotype seems to render the highest degree of resistance to both classical scrapie and BSE [5]. Because there is great concern about the possibility that sheep are also infected by BSE in nature, the EU forces all EU countries to conduct breeding programmes in order to select for sheep with the most TSE resistant genotypes to minimize the risk of TSE, or eventually eradicate the disease from sheep stocks [6].
In order to accomplish this goal, a reliable, fast and cost-effective method for PRNP genotyping in sheep is indispensable. To date, a number of methods are used, including sequencing [7], PCR-RFLP [8], DGGE analysis [9], primer extension assay [10], ARMS [11], TaqMan-MGB assay [12] and reverse hybridization (Stefan Roels, personal communication). Here we report the development and validation of a new, dual fluorescent multiprobe assay, consisting of 2 closed tube PCR reactions containing respectively 4 and 3 dual-labelled fluorescent ASO probes for the detection in real-time of the 7 allelic variants of sheep PRNP mentioned above.
It is important to note that the last few years there are more and more reports of sheep that are infected with atypical scrapie, including sheep with the ARR/ARR genotype. New insights in this new form of scrapie, that seems to be associated with PRNP codons 141 and 154 [13], will probably have a major impact on the ongoing breeding programmes for reduced TSE susceptibility and will probably result in the development of adapted genotyping methods.
Methods
DNA preparation from blood samples
Whole blood samples from 58 Belgian sheep (Texel, Suffolk, Swifter, Bleu du Maine, Ardense Voskop, Vlaams Schaap, Lakens Schaap, Belgisch Melkschaap) were collected in tubes containing anti-coagulant and stored at -20°C. Two hundred μl blood was washed 3 times with 500 μl TE (10 mM Tris-HCl pH 8; 1 mM EDTA pH 8) to roughly clean-up the white blood cells. The pellet was resuspended in 100 μl Lysis Buffer K (10 mM Tris-HCl pH 8; 50 mM KCl; 0.5% Tween 20) supplemented with 100 μg/ml proteinase K (Roche Diagnostics, Belgium) and incubated for 45' at 56°C to release the DNA. The lysate was incubated for 10' at 95°C to inactivate the proteinase K and centrifugated at 16.100 × g for 1' to pellet down the cell debris. This unpurified DNA solution has an average concentration of 50 ng/μl.
PCR primer and probe design
Primers and dual-labelled ASO probes were designed with Primer Express software version 2.0 (Applied Biosystems, USA) and synthesized by Sigma-Genosys (UK). Their sequences and Tm (calculated via Tm Utility [14]) are listed in Table 1.
Table 1 An overview of the primers and probes used in the dual fluorescent multiprobe assay
Primer/Probe Name Sequence (5' → 3') Length (bp) Tm (°C) avTm (°C)
Forward primer GCCTTGGTGGCTACATG 17 59.35 -
Reverse primer CTGTGATGTTGACACAGTCAT 21 59.60 -
A136-probe FAM-TGCTCATGGCACTTCCCA-BHQ1 18 62.98 56.27
V136-probe HEX-CTGCTCATGACACTTCCCAG-BHQ1 20 61.86 55.33
R154-probe TexasRed-CCGTTACTATCGTGAAAACATGTAC-BHQ2 25 61.08 56.59
H154-probe Cy5-CCGTTACTATCATGAAAACATGTACC-BHQ2 26 61.06 56.67
R171-probe FAM-CCAGTGGATCGGTATAGTAACCA-BHQ1 23 62.27 57.19
H171-probe HEX-AGACCAGTGGATCATTATAGTAACCA-BHQ1 26 61.96 57.66
Q171-probe TexasRed-CCAGTGGATCAGTATAGTAACCAGA-BHQ2 25 62.07 58.46
Fluorescent labels and quenchers are in italic, SNPs are underlined. (av)Tm was calculated using Tm Utility [14].
The primers generate a 180-bp amplicon of PRNP exon 3 containing codons 136, 154 and 171. The primer sequences contain no described SNPs, maximizing the range of animals that can be tested, and contain no more than 2 Gs or Cs in the last 5 bp and no more than 3 consecutive Gs, minimizing self-complementarity and the formation of primer-dimers and hairpin structures.
A dual-labelled ASO probe was designed for each of the 7 allelic variants, in which the SNP is localized approximately in the middle of the sequence. The probe sequence does not start with a G, preventing quenching of the fluorophore, and contains no more than 3 consecutive Gs. For all possible candidate probes of each SNP, the melting temperature was calculated for the perfect match (Tm) and for the single mismatch (avTm) with the other allelic variant. Those probes were selected for which the Tm with the perfect match was higher than the Tm of the primers and for which the avTm with the single mismatch was lower than the Tm of the primers (see Table 1).
In PCR 1, the 4 allelic variants of codons 136 (A/V) and 154 (R/H) are determined simultaneously with 4 allele-specific probes, each containing a different fluorophore. A combination of the fluorophores FAM/HEX/TexasRed/Cy5 was chosen as described by Ugozzoli et al. [15]. In PCR 2, the 3 allelic variants of codon 171 (R/H/Q) are determined simultaneously with 3 allele-specific probes, containing resp. FAM/HEX/TexasRed as fluorophore. The fluorophores FAM/HEX are quenched with BHQ1, and TexasRed/Cy5 with BHQ2.
Quality test for dual-labelled ASO probes
In order to check the quality of the probes, 1 μl 10 μM probe was digested with 1 U of RQ1 DNase (Promega, USA) for 30' at 37°C in a final volume of 10 μl. After adding 40 μl of 10 mM Tris-HCl pH 8, the fluorescence was measured by using the Imaging Services of the iCycler iQ Real-Time PCR Detection System (Bio-Rad Laboratories, USA) with 50 μl of the Bio-Rad calibration dye as a reference in a separate tube.
Dual fluorescent multiprobe assay
Both PCRs of the assay were performed in the iCycler iQ Real-Time PCR Detection System (Bio-Rad Laboratories, USA) using normal PCR tubes (ABgene, UK), in a total volume of 15 μl containing iQ Supermix (50 mM KCl, 20 mM Tris-HCl pH 8.4, 0.8 mM dNTPs, 0.375 U iTaq DNA polymerase, 3 mM MgCl2 and stabilizers; Bio-Rad Laboratories, USA), 400 nM of each primer and probe, and ~150 ng DNA. The real-time PCR program for both reactions consists of an iTaq DNA polymerase activation and DNA denaturation step (3' at 95°C), followed by 40 amplification cycles (denaturation for 20" at 95°C and annealing-elongation for 40" at 62°C). The fluorescent signals, generated by the cleavage of the dual-labelled ASO probes, were detected in real-time during the annealing-elongation step.
Data was analysed by the iCycler iQ Real-Time PCR Detection System Software version 3.0a (Bio-Rad Laboratories, USA). For every probe, RFU-values were measured every cycle and after background normalization plotted against the cycle number. Based on these real-time amplification plots, Ct-values were calculated in the PCR Quantification tab of the Data Analysis module by user defined assignation of the Baseline Cycles and the Threshold position. The obtained data, Ct-values for detection and 'N/A' for non-detection, was then copied to a Microsoft Excel-spreadsheet [see Additional file 1] to semi-automatically determine the PRNP genotype. As positive controls, 8 samples of known genotypes representing all combinations of tested variants of each polymorphic locus and 1 NTC, should be included in every assay. Only if all of those samples are genotyped correctly, the genotypes of the unknown samples should be considered as reliable.
Gel electrophoresis and elution
The 180-bp amplification product from sheep with known genotypes was visualized by 2% agarose multi-purpose molecular grade (Bioline, UK) gel electrophoresis, eluted in 20 μl with the Geneclean II kit (Bio101, USA), 1000 times diluted with 10 mM Tris-HCl pH 8 and used as a template during the optimalization of the genotyping assay. These purified PCR products were also used as positive controls in later assays. To avoid possible contamination of samples with these PCR controls, the setup of the control PCRs was performed after the setup of the sample PCRs, with the same PCR mix, while following standard procedures for good laboratory practise.
PCR-RFLP
The PCR-RFLP assay is described by Peelman & Van Poucke [16], validated via a ring-test organized by the International Society of Animal Genetics in 2003–2004, and already used to genotype more than 3000 Belgian sheep [17].
Direct sequencing
A PRNP amplicon of 315 bp, containing codons 136, 154 and 171, was amplified with PCR primers OariPRNPseqF 5'-GGAGGCTGGGGTCAAGGT-3' and OariPRNPseqR 5'-GGTGGTGGTGACTGTGTGTTG-3'. PCR was performed in the T3 Thermocycler (Biometra, Germany) in a total volume of 10 μl containing ~150 ng DNA, 500 nM of each primer, 0.8 mM dNTPs, 2 mM MgCl2, BioTaq buffer and 0.5 U BioTaq polymerase (Bioline, UK). The PCR program consists of a DNA denaturation step (5' at 95°C), followed by 30 amplification cycles (denaturation for 30" at 95°C, annealing for 30" at 63°C and elongation for 1' at 72°C), and a final elongation step for 10' at 72°C. Amplification products were visualized by 2% agarose multi-purpose molecular grade (Bioline, UK) gel electrophoresis and eluted in 20 μl with the Geneclean II kit (Bio101, USA). Approximately 200 ng of this purified PCR product was used in combination with 2 pmol OariPRNPseqR primer for a direct sequencing reaction with the Thermo Sequenase Cy5 Dye Terminator Sequencing Kit according to the manufacturers' instructions (Amersham Biosciences, Denmark). The reaction, consisting of 30 cycles (30" at 95°C, 30" at 58°C and 1'20" at 72°C), was performed in the T3 Thermocycler (Biometra, Germany) and analysed on the ALFexpress Sequencing system (Amersham Biosciences, Denmark).
Results
Dual fluorescent multiprobe assay optimalization
DNA was released from 8 sheep blood samples with known PRNP genotypes (VRQ/VRQ, VRQ/ARR, ARR/ARR, ARR/ARH, ARH/ARH, ARH/ARQ, AHQ/AHQ and ARR/AHQ), representing all combinations of tested variants of each polymorphic locus.
With those DNA samples as a template, the annealing temperature range of the primers was determined in which only the correct PCR fragment was amplified by a temperature gradient experiment in the iCycler iQ Real-Time PCR Detection System (data not shown). This experiment was performed with iQ Supermix, ~150 ng unpurified DNA and 200 nM primers in a total volume of 15 μl.
After probe quality testing, the hybridization temperature range was determined for each probe separately, in which highly specific hybridization occured without mismatch hybridization, within the optimal annealing temperature range of the primers. This was performed with a real-time temperature gradient experiment in the iCycler iQ Real-Time PCR Detection System (data not shown). The experiment was performed in a total volume of 15 μl with iQ Supermix, 200 nM primers and probe, and ~150 ng purified PCR product for 3 different samples (+/+, +/- and -/-) and H2O as NTC.
Finally, considering the optimal hybridization temperature range of all probes, the optimal hybridization temperature and the primer and probe concentrations were determined for the 2 fluorescent multiprobe PCRs. For both closed tube PCRs an annealing/hybridization temperature of 62°C in combination with 400 nM primers and probes resulted in an accurate genotyping assay, even with unpurified DNA as a template. The obtained amplification curves from unpurified DNA samples with all possible outcomes for each probe are shown in Figure 1.
Figure 1 (A-G) – Amplification plots obtained with the dual fluorescent multiprobe assay for sheep PRNP genotyping. Amplification plots are shown for a homozygous positive (+/+), a heterozygote (+/-), a homozygous negative (-/-) and a no template control (NTC) with (A) A136-probe (FAM-labelled) in PCR 1, (B) V136-probe (HEX-labelled) in PCR 1, (C) R154-probe (TexasRed-labelled) in PCR 1, (D) H154-probe (Cy5-labelled) in PCR 1, (E) R171-probe (FAM-labelled) in PCR 2, (F) H171-probe (HEX-labelled) in PCR 2 and (G) Q171-probe (TexasRed-labelled) in PCR 2. (H) An amplification plot for 41 test samples and 9 control samples for the latter probe. For all amplification plots, unpurified DNA was used as a template.
Semi-automatic PRNP genotype determination
In order to speed up data analyses a Microsoft Excel-spreadsheet was developed for semi-automatic PRNP genotype determination [see Additional file 1]. Ct-values (in case of detection) or 'N/A' (in case of non-detection), obtained with the iCycler iQ Real-Time PCR Detection System Software version 3.0a (Bio-Rad Laboratories, USA), have to be entered (copy-paste) in the 'Ct-values Samples' tab of the spreadsheet for all probes of all test samples, together with their 'Run ID' and 'Sample ID', in order to generate the corresponding genotype. An impossible combination will result in a blank cell. The same thing should be done for the control samples in the 'Ct-values Controls' tab. The program automatically determines if the control samples are genotyped correctly. From the 'Results' tab, 'Sample ID' and genotype for every sample can be copy-pasted to other software programs, with a reminder if all control samples were genotyped correctly or not.
Cross-validation study
To evaluate the performance of the dual fluorescent multiprobe assay for PRNP genotyping in sheep, a total of 50 DNA samples that had been previously genotyped via PCR-RFLP [16,17], were tested in a blind manner. For every allelic variant homozygous and heterozygous genotypes were included at least 3 times. Concordant results were obtained for all 50 samples. Scatter plots for the allelic discrimination of both variants of codons 136 and 154, obtained via the Allelic Discrimination feature of the Data Analysis module in the Threshold Cycle Display Mode, are shown in Figure 2. Since the software is not able to distinguish more than 2 allelic variants, no plots are shown for codon 171. These results were also confirmed by direct sequencing. In addition, the assay was used to genotype more than 600 Belgian sheep within the TSE resistant breeding programme framework. No discrepancies with the rules of Mendelian inheritance were observed.
Figure 2 (A-B) – Scatter plots based on Ct-values obtained with the dual fluorescent multiprobe assay for sheep PRNP genotyping. In these scatter plots, Ct-values for 2 alleles of a specific codon are plotted on the xy axes from 50 sheep samples. (A) A136-probe (FAM-labelled) vs V136-probe (HEX-labelled) showing homozygous A136-genotypes (red closed dots), heterozygous genotypes (dark green closed triangles), homozygous V136-genotypes (light green closed squares), and a NTC (light blue closed diamond). (B) R154-probe (Texas-Red) vs H154-probe (Cy5-labelled) showing homozygous R154-genotypes (red closed dots), heterozygous genotypes (dark green closed triangles), homozygous H154-genotypes (light green closed squares), and a NTC (light blue closed diamond).
Discussion
To minimize the risk ofTSE, breeding programmes of sheep to select resistant genotypes will be implemented in the near future in all EU countries [6]. An assay as described above will therefore be a vital instrument for a lot of molecular genetic laboratories to fulfill the need of genotyping hundreds or thousands of sheep. However, the most resistant ARR/ARR sheep are not 100% resistant to classical scrapie and BSE, and are furthermore susceptible to atypical scrapie [3,5,13]. Because scientists are searching for new resistance markers, it is likely that, based on new findings, future breeding programmes will be adapted. As a consequence, also the genotyping methods will have to be adapted, which can easily be done with this assay.
The assay can be performed on the iCycler iQ Real-Time PCR detection system (Bio-Rad) or on equally performing machines. It consists of 2 closed tube PCR reactions with the same PCR primers, spanning the codons 136, 154 and 171 of sheep PRNP. In PCR 1, polymorphisms at codons 136 (A/V) and 154 (R/H) are simultaneously detected in real time with 4 different dual-labelled fluorescent ASO probes. In PCR 2, polymorphisms at codon 171 (R/H/Q) are simultaneously detected in real time with 3 different dual-labelled fluorescent ASO probes. Although it was not the purpose of this assay, it is possible to distinguish homozygotes from heterozygotes, based on the shape of the amplification plots (Figure 1). These can not only serve as internal controls, but can also identify 'complex' PRNP genotypes [18].
A probe for the detection of the K171 variant was not included in the assay, because little is known about its association with TSE resistance/susceptibility [19]. If recommended, the probe can be included in PCR 2. The assay was not tested on sheep containing the K171 allele, but because of the low Tm of the R171, H171 and Q171 probes with the K171 sequence, a K171 allele won't be detected with the current assay. In case of a homozygote no allele will be detected in PCR 2 and in case of a heterozygote only the other allele will be detected in PCR 2. However, the shape of the amplification plot will reveal that only 1 allele was detected. This will also be the case for every other mutation in the probe sequence, depending on the Tm of the probes with the mutated sequence, which is inherent in a primer/probe based technique. Until now, no such problem was observed during routine genotyping of more than 600 Belgian sheep.
Although the sequencing method is still the golden standard for detecting all possible polymorphisms, it is a very time consuming and expensive technique not suitable for routine typing of large samples numbers in smaller service laboratories. The assay described here is an alternative method for routine typing of a defined number of polymorphisms. The capability of the iCycler to simultaneously excite and detect 4 different fluorophores, enables the use of 4 different probes in a single PCR. Only 2 PCR reactions, each with a total volume of 15 μl, are required for accurate genotyping calls for all 15 possible genotypes. Since the allelic variants are detected in real-time, no post-PCR manipulations are required, such as restriction digests and/or electrophoresis, reducing time, costs and possible carry-over contamination. Data can be copied to a Microsoft Excel-spreadsheet for semi-automatic determination of the genotype [see Additional file 1]. The control samples, which serve as external controls, are successfully analysed for many times during routine genotyping, proving the reproducibility of the assay.
Taking into account the 9 control samples, both PCR 1 and PCR 2 can be conducted in 1 single run using a well factor plate for calibration for less than 40 samples. For 40 samples or more (with a maximum of 87 samples per run), the assay should be performed in 2 runs (PCR 1 and PCR 2) using the experimental plate for calibration. Every run takes 1h25. When using robotic workstations and multiple real-time PCR detection systems this assay can easily be used in laboratories dealing with a large number of samples.
The assay is robust since it can be performed with unpurified DNA as template for PCR, as shown in Figure 1 and 2, although the amplification plots usually have lower Ct-values and higher RFU-values with purified PCR products (data not shown). Since the discrimination factor in the assay is whether or not a Ct-value is generated (the value itself is not important), the DNA concentration and quality don't influence the genotyping result, as long as a PCR product is generated. So, an extra DNA purification step could be included for bad preserved blood samples.
When applying the proposed primer/probe design strategy, the Tm of the primers can be chosen randomly, as long as the Tm of each probe with its perfect match target sequence is higher and its avTm with the other allelic variant (mismatch) is lower than the Tm of the primers (see Table 1). A ΔTm of 2°C in both directions is sufficient to obtain the desired probe specificity for allelic discrimination (see Figure 1 and 2), without the need for expensive MGB- or LNA-probes [20,21]. We recommend to check every probe for fluorophore respons before starting any PCR optimization.
The performance of the assay was demonstrated via a cross-validation study. The correlation between the genotypes of 50 sheep generated with the dual fluorescent multiprobe assay and with PCR-RFLP and direct sequencing as reference methods was 100%. In addition, no discrepancies against the rules of Mendelian inheritance were observed during routine genotyping of more than 600 Belgian sheep.
Conclusions
We have developed and validated a dual fluorescent multiprobe assay for robust, reliable and reproducible genotyping of few to many sheep samples in a fast, simple and cost-effective way, practicable in most standard equipped molecular genetic laboratories. The outlined primer/probe design strategy can also be applied for the detection of other polymorphisms or disease causing mutations.
List of abbreviations
+/+ Homozygous positive
+/- Heterozygous
-/- Homozygous negative
A Alanine
ARMS Amplification Refractory Mutation System
ASO Allele-Specific Oligonucleotide
avTm Melting Temperature when annealed to Allelic Variant
BHQ Black Hole Quencher
bp basepairs
BSE Bovine Spongiform Encephalopathy
CJD Creutzfeldt-Jacob Disease
Ct Threshold Cycle
DGGE Denaturing Gradient Gel Electrophoresis
DNA DeoxyriboNucleic Acid
EU European Union
H Histidine
LNA Locked Nucleic Acid
MGB Minor Groove Binding
N/A Not Available
NTC No Template Control
PCR Polymerase Chain Reaction
PRNP gene encoding PrPC
PrPC host encoded cellular prion protein
PrPSc disease-causing isoform of PrPC
Q Glutamine
R Arginine
RFLP Restriction Fragment Length Polymorphism
RFU Relative Fluorescence Units
SNP Single Nucleotide Polymorphism
Tm Melting Temperature
TSE Transmissible Spongiform Encephalopathy
U Unit(s)
V Valine
vCJD variant Creutzfeldt-Jacob Disease
vs versus
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
MVP designed the project and protocols involved, carried out the assays and drafted this manuscript. JV participated in primer/probe design and provided real-time PCR support. MM carried out the DNA preparations. AVZ and LJP participated in the design of the project.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Supplementary Material
Additional File 1
A Microsoft Excel-spreadsheet for semi-automatic determination of the sheep PRNP genotype (based on codons 136, 154 and 171).
Click here for file
Acknowledgements
The authors wish to thank Dominique Vander Donckt and Linda Impe for excellent technical assistance. This work was supported by FPS Health, Food Chain Safety and Environment (Project S-6151). Jo Vandesompele is supported by a grant from the Flemish Institute for the Promotion of Innovation by Science and Technology in Flanders (IWT).
==== Refs
Prusiner SB Prion Biology and Diseases 2004 New York: Cold Spring Harbor Laboratory Press
Prusiner SB The prion diseases Brain Pathol 1998 8 499 513 9669700
Hunter N Scrapie and experimental BSE in sheep Br Med Bull 2003 66 171 183 14522858 10.1093/bmb/66.1.171
Bosque PJ Bovine spongiform encephalopathy, chronic wasting disease, scrapie, and the threat to humans from prion disease epizootics Curr Neurol Neurosci Rep 2002 2 488 495 12359101
Houston F Goldmann W Chong A Jeffrey M Gonzalez L Foster J Parnham D Hunter N Prion diseases: BSE in sheep bred for resistance to infection Nature 2003 423 498 12774113 10.1038/423498a
European Commission Decision 2003/100/EC of 13 February laying down minimum requirements for the establishment of breeding programmes for resistance to transmissible spongiform encephalopathies in sheep (Text with EEA relevance) (notified under document number C(2003) 498) Official Journal 2003 L041 41 45
Junghans F Teufel B Buschmann A Steng G Groschup MH Genotyping of German sheep with respect to scrapie susceptibility Vet Rec 1998 143 340 341 9795405
Yuzbasiyan-Gurkan V Krehbiel JD Cao Y Venta PJ Development and usefulness of new polymerase chain reaction-based tests for detection of different alleles at codons 136 and 171 of the ovine prion protein gene Am J Vet Res 1999 60 884 887 10407484
Belt PB Muileman IH Schreuder BE Bos-de Ruijter J Gielkens AL Smits MA Identification of five allelic variants of the sheep PrP gene and their association with natural scrapie J Gen Virol 1995 76 509 517 7897344
Zsolnai A Anton I Kuhn C Fesus L Detection of single-nucleotide polymorphisms coding for three ovine prion protein variants by primer extension assay and capillary electrophoresis Electrophoresis 2003 24 634 638 12601731 10.1002/elps.200390074
Buitkamp J Semmer J A robust, low- to medium-throughput PRNP genotyping system BMC Infect Dis 2004 4 30 15345029 10.1186/1471-2334-4-30
Garcia-Crespo D Oporto B Gomez N Nagore D Benedicto L Juste RA Hurtado A PrP polymorphisms in Basque sheep breeds determined by PCR-restriction fragment length polymorphism and real-time PCR Vet Rec 2004 154 717 722 15214515
Moum T Olsaker I Hopp P Moldal T Valheim M Moum T Benestad SL Polymorphisms at codons 141 and 154 in the ovine prion protein gene are associated with scrapie Nor98 cases J Gen Virol 2005 86 231 235 15604451 10.1099/vir.0.80437-0
Idaho Technology Inc. Downloads & Upgrades site
Ugozzoli LA Chinn D Hamby K Fluorescent multicolor multiplex homogeneous assay for the simultaneous analysis of the two most common hemochromatosis mutations Anal Biochem 2002 307 47 53 12137778 10.1016/S0003-2697(02)00016-7
Peelman LJ Van Poucke M PRNP genotype frequency sampling of the most important sheep breeds in Belgium Vl Diergeneeskd Tijdschr 2003 72 20 26
Roels S Renard C De Bosschere H Geeroms R Van Poucke M Peelman L Vanopdenbosch E Detection of polymorphisms in the prion protein gene in the Belgian sheep population: some preliminary data Vet Q 2004 26 3 11 15072136
Dawson M Warner R Nolan A McKeown B Thomson J 'Complex' PrP genotypes identified by the National Scrapie Plan Vet Rec 2003 152 754 755 12833938
Gombojav A Ishiguro N Horiuchi M Serjmyadag D Byambaa B Shinagawa M Amino acid polymorphisms of PrP gene in Mongolian sheep J Vet Med Sci 2003 65 75 81 12576708 10.1292/jvms.65.75
Kutyavin IV Afonina IA Mills A Gorn VV Lukhtanov EA Belousov ES Singer MJ Walburger DK Lokhov SG Gall AA Dempcy R Reed MW Meyer RB Hedgpeth J 3'-minor groove binder-DNA probes increase sequence specificity at PCR extension temperatures Nucleic Acids Res 2000 28 655 661 10606668 10.1093/nar/28.2.655
Ugozzoli LA Latorra D Pucket R Arar K Hamby K Real-time genotyping with oligonucleotide probes containing locked nucleic acids Anal Biochem 2004 324 143 152 14654057 10.1016/j.ab.2003.09.003
|
15769289
|
PMC1274271
|
CC BY
|
2021-01-04 16:28:14
|
no
|
BMC Infect Dis. 2005 Mar 15; 5:13
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-13
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-161578013610.1186/1471-2334-5-16Research ArticleWorldwide trends in quantity and quality of published articles in the field of infectious diseases Bliziotis Ioannis A [email protected] Konstantinos [email protected] Paschalis I [email protected] Antonia I [email protected] Matthew E [email protected] Alfa Institute of Biomedical Sciences(AIBS), Athens, Greece2 Alfa HealthCare, Athens, Greece3 Department of Medicine, Tufts University School of Medicine, Boston, Massachusetts, USA2005 21 3 2005 5 16 16 26 11 2004 21 3 2005 Copyright © 2005 Bliziotis et al; licensee BioMed Central Ltd.2005Bliziotis et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Trying to confront with the widespread burden of infectious diseases, the society worldwide invests considerably on research. We evaluated the contribution of different world regions in research production in Infectious Diseases.
Methods
Using the online Pubmed database we retrieved articles from 38 journals included in the "Infectious Diseases" category of the "Journal Citation Reports" database of the Institute for Scientific Information for the period 1995–2002. The world was divided into 9 regions based on geographic, economic and scientific criteria. Using an elaborate retrieval system we obtained data on published articles from different world regions. In our evaluation we introduced an estimate of both quantity and quality of research produced from each world region per year using: (1) the total number of publications, (2) the mean impact factor of publications, and (3) the product of the above two parameters.
Results
Data on the country of origin of the research was available for 45,232 out of 45,922 retrieved articles (98.5 %). USA and Western Europe are by far the most productive regions concerning publications of research articles. However, the rate of increase in the production of articles was higher in Eastern Europe, Africa, Latin America and the Caribbean, and Asia during the study period. The mean impact factor is highest for articles originating in the USA (3.42), while it was 2.82 for Western Europe and 2.73 for the rest of the world (7 regions combined).
Conclusion
USA and Western Europe make up a striking 80% of the world's research production in Infectious Diseases in terms of both quantity and quality. However, all world regions achieved a gradual increase in the production of Infectious Diseases articles, with the regions ranking lower at present displaying the highest rate of increase.
==== Body
Background
Infectious diseases constitute a major health problem both in developed and developing countries. Old and emerging infectious diseases contribute substantially to morbidity and mortality worldwide. For this reason, the society invests considerably on infectious diseases research, in order to achieve scientific progress and develop new therapeutic interventions.
The research productivity by various world regions has been studied for several biomedical fields. In general, the USA and Western Europe are the leaders of global biomedical research, although their relative contribution varies for different fields of research [1-4]. Several studies have focused on the scientific production of European Union's countries, in various biomedical fields, including Infectious Diseases [5-7]. However, the literature lacks studies estimating the quantity and quality of worldwide research production in Infectious Diseases. The purpose of our study was to evaluate the contribution of different world regions in scientific research in the field of Infectious Diseases. We also evaluated the trends in quality and quantity of published articles from different world regions.
Methods
We used the electronic PubMed database [8] and data from the Journal Citation Reports (JCR) database of the Institute for Scientific Information (ISI) [9]. We searched for articles included both in the "Infectious Diseases" category of the JCR and in PubMed database. Articles published prior to 1995 were not included in the analysis, because the full address of the authors of the papers was frequently not registered in PubMed prior to this year. Because JCR had available data up to the year 2002 at the time of our analysis, our data collection and evaluation refers to the period 1995–2002. A total of 38 journals were included. Two independent investigators conducted the data collection (IAB, PIV).
For the purpose of this study, the world was divided into 9 regions based on a combination of geographic, economic and scientific criteria [10]. The 9 regions are Western Europe, Eastern Europe, United States of America (USA), Canada, Latin America and the Caribbean, Africa, Japan, Asia (excluding Japan), and Oceania.
In our search of different fields in the Pubmed database we used a phrase consisting of four parts joined together by the so-called Boolean operators, i.e. AND, OR, and NOT. Each search was limited to a specific year using the "Limits" function, which is incorporated in the search engine. We only analyzed data on original articles and reviews, excluding publication types, such as letters, editorials, and news reports. For example in order to search for articles published in the "Journal of Infectious Diseases" and whose first author's address was in Europe, we used the following text: Journal of Infectious Diseases [journal] AND journal article [pt] AND (Andorra [AD] OR Austria [AD] OR... Wales [AD]) NOT (Australia [AD] OR Canada [AD] OR...). In the first parenthesis of the search phrase, the countries of the implicated region are included. In the second parenthesis, after the word NOT, certain addresses are excluded in order to avoid double counting.
Subsequently, the results of these searches (the number of articles produced by each world region in a specific journal within a year) were summed up. For confirmation, the sum of articles produced by all different world regions in a journal, was compared to the actual total number of articles published in that journal for a specific year. This number was obtained from PubMed without using any address limits. Using this methodology we were able to cross-examine missed or unretrieved addresses. This occurred occasionally, in cases of articles with no address registered, and in cases of articles where only the affiliated institution or the city (not the country) was recorded. If less than 5% of the total articles of a specific journal during a year had missing addresses, we did not include these articles in our calculations, assuming that the numerical error was not significant. On the other hand, if more than 5% of the total articles of a specific journal during a year, had missing addresses, we performed searches for the author's address by checking other articles of the same author within the same year.
The number of published articles was considered as an index of quantity of research productivity. The mean impact factor of the published articles was considered as an index of quality of research productivity. Finally, the product of the number of articles published in a journal multiplied by the impact factor of the journal, for the year studied, was considered as an index evaluating combined the quantity and quality of research productivity. The sum of these products from all journals, for each world region within a year, was named "total product" for each region within the studied year. The impact factor for each journal was obtained from the JCR database of the ISI.
To further evaluate factors associated with the research published in Infectious Diseases journals we used relevant "World Development Indicators" [11] from the online databases of the World Bank. The research productivity of different world regions (estimated by the "total product") was evaluated in relation to total population, gross domestic product (GDP) in standard 1995 US dollars, and gross national income (GNI) per capita (Atlas method).
We used the absolute figures and the average annual rates of increase of scientific output (research productivity) of different world regions to calculate future performance using a projection model. Also, we performed correlation statistical analysis of the absolute numbers of published articles between the different world regions during the years of the study period (1995–2002) using Pearson correlation testing. In addition, we performed correlation statistical analysis to examine the research productivity of the specified world regions compared with the total world production.
Results
The journals that were included in our analysis are shown in Table 1. Using the methodology described above, we managed to retrieve and categorize 45,232 out of 45,922 articles, (98.5%) from the implicated journals indexed in Pubmed during the study period. The total production of articles in each defined world region, as well as the relative contribution of each region to the total production in the field of Infectious Diseases, is displayed in Table 2. As shown in this table, USA and Western Europe are by far the most productive regions (79.8% of the articles published worldwide, throughout the whole period studied, came from these two regions). As expected, the difference between the USA and Western Europe increases when both the number of articles and impact factor are taken into account, due to the higher impact factor that USA had throughout the study period. In the years 2000–2002 Western Europe's production exceeded that of the USA, although USA researchers produced more articles in all previous years. The last column shows the "total product" of research published in the field of Infectious Diseases for each region for the whole study period.
Table 1 Title of journals included in the field of Infectious Diseases of the Institute for Scientific Information (ISI) indexed both by ISI and PubMed.
Title of journal Study period
AIDS 1995 – 2002
AIDS Patient Care STDS 2002
AIDS Research and Human Retroviruses 1995 – 2002
American Journal of Infection Control 1995 – 2002
Antiviral Therapy 2000 – 2002
BMC Infectious Diseases 2002
Clinical and Diagnostic Laboratory Immunology 1995 – 2002
Clinical Infectious Diseases 1995 – 2002
Clinical Microbiology and Infection 2002
Current Opinion in Infectious Diseases 2000 – 2002
Diagnostic Microbiology and Infectious Disease 1995 – 2002
Emerging Infectious Diseases 1995 – 2002
Epidemiology and Infection 1995 – 2002
European Journal of Clinical Microbiology and Infectious Diseases 1995 – 2002
Infection 1995 – 2002
Infection and Immunity 1995 – 2002
Infection Control and Hospital Epidemiology 1995 – 2002
Infectious Agents and Disease 1995 – 1996
Infectious Disease Clinics of North America 1995 – 2002
International Journal of Antimicrobial Agents 2000 – 2002
International Journal of Hygiene and Environmental Health (prior to 1999: Zentralblatt fur hygiene und umweltmedizin) 1997 – 2002
International Journal of STD & AIDS 1995 – 2002
International Journal of Tuberculosis and Lung Disease 1998 – 2002
JAIDS-Journal of Acquired Immune Deficiency Syndromes (prior to 1998: Journal of Acquired Immune Deficiency Syndromes and Human Retrovirology) 1995 – 2002
Japanese Journal of Infectious Diseases 2000 – 2002
Journal of Antimicrobial Chemotherapy 1995 – 2002
Journal of Hospital Infection 1995 – 2002
Journal of Human Virology 2001 – 2002
Journal of Infection 1995 – 2002
Journal of Viral Hepatitis 1997 – 2002
Leprosy Review 1997 – 2002
Microbial Drug Resistance 1997 – 2002
Pediatric AIDS and HIV Infection 1997
Pediatric Infectious Disease Journal 1995 – 2002
Scandinavian Journal of Infectious Diseases 1995 – 2002
Sexually Transmitted Diseases 1995 – 2002
Sexually Transmitted Infections (prior to 1997: Genitourinary Medicine) 1999 – 2002
The Journal of Infectious Diseases 1995 – 2002
Table 2 Number of articles published in journals included in the "Infectious Diseases" category of "Journal Citation Report" database and indexed by PubMed, from different world regions, for the period 1995–2002.
Number of articles (% percentage within a calendar year)
WORLD AREAS 1995 1996 1997 1998 1999 2000 2001 2002 1995–2002 1995–2002*
USA 2118 (46.98) 2078 (45.09) 2162 (43.66) 2425 (42.86) 2346 (40.27) 2530 (38.92) 2543 (38.48) 2481 (37.76) 18683 (41.30) 63804
Western Europe 1673 (37.11) 1804 (39.14) 1863 (37.62) 2136 (37.75) 2248 (38.59) 2577 (39.65) 2579 (39.03) 2539 (38.64) 17419 (38.51) 49033
Asia (excluding Japan) 175 (3.88) 160 (3.47) 235 (4.75) 317 (5.60) 316 (5.42) 376 (5.78) 413 (6.25) 437 (6.65) 2429 (5.37) 5927
Japan 106 (2.35) 141 (3.06) 177 (3.57) 135 (2.39) 226 (3.88) 258 (3.97) 273 (4.13) 260 (3.96) 1576 (3.48) 4113
Canada 155 (3.44) 126 (2.73) 151 (3.05) 179 (3.16) 181 (3.11) 194 (2.98) 183 (2.77) 207 (3.15) 1376 (3.04) 4510
Latin America and Caribbean 76 (1.69) 82 (1.78) 101 (2.04) 127 (2.24) 130 (2.23) 154 (2.37) 173 (2.62) 189 (2.88) 1032 (2.28) 2978
Oceania 89 (1.97) 108 (2.34) 111 (2.24) 138 (2.44) 142 (2.44) 155 (3.97) 144 (2.18) 145 (2.21) 1032 (2.28) 3153
Africa 70 (1.55) 66 (1.43) 97 (1.96) 137 (2.42) 159 (2.73) 151 (2.32) 177 (2.68) 153 (2.33) 1010 (2.23) 2913
Eastern Europe 46 (1.02) 44 (0.95) 55 (1.11) 64 (1.13) 78 (1.34) 105 (1.62) 123 (1.86) 160 (2.43) 675 (1.49) 1325
Total 4508 (100) 4609 (100) 4952 (100) 5658 (100) 5826 (100) 6500 (100) 6608 (100) 6571 (100) 45232 (100) 137756
*Number of articles published multiplied by their impact factor
We observed a continuous increase in the production of research articles from all world regions during the period 1995–2002 (Table 2). There was a strong and statistically significant correlation between the absolute numbers of published articles between the different world regions during the years of the study period (1995–2002). The median (range) of the Pearson correlation test values between comparisons of 36 possible couples of the specified 9 world regions was 0.88 (0.48 – 0.99). Thirty of 36 comparisons had statistical significance at levels < 0.05 (24 of them had statistical significance at levels < 0.01). The comparisons that did not have statistical significance (p > 0.05) were between world regions with relative small numbers of published articles in the field of Infectious Diseases.
In addition, a strong and statistically significant correlation was noted between the annual research production of the specified world regions with that of the total world production; specifically the median (range) of Pearson correlation test results of these analyses were 0.94 (0.70–0.99). However, the rate of increase of research productivity in the field of Infectious Diseases was higher in Eastern Europe, Africa, Latin America and the Caribbean, and Asia. Using a projection model we estimated that these regions would reach USA's production level in 23 years and Western Europe's production level in 29 years, provided that each region maintains the average rate of increase of research production achieved in the 8-year-period studied.
Table 3 presents the mean impact factor of published articles in the field of Infectious Diseases for each region in the studied years. A mean value of the impact factor is also presented for the whole 8-year-period. The mean impact factor, for the whole period, is highest for articles originating in the USA. Interestingly, Canada ranks second and Western Europe ranks sixth regarding the mean impact factor of published articles. Eastern Europe has the lowest mean impact factor among all world regions.
Table 3 Mean impact factor of articles published in journals included in the "Infectious Diseases" category of "Journal Citation Report" database and indexed by Pubmed, from different world regions, for the period 1995–2002.
Mean impact factor
WORLD AREAS 1995 1996 1997 1998 1999 2000 2001 2002 1995–2002 (25thpercentile, median, 75thpercentile)
USA 3.08 3.44 3.12 3.26 3.47 3.60 3.71 3.54 3.42 (2.29 3.51 4.21)
Canada 2.88 3.71 2.91 3.00 3.29 3.48 3.49 3.43 3.28 (2.08 3.20 4.18)
Oceania 2.69 3.17 3.01 2.84 2.75 3.53 3.28 2.99 3.05 (1.79 2.80 4.18)
Africa 2.83 2.87 3.12 2.46 3.00 2.87 3.01 2.90 2.89 (1.63 2.20 4.18)
Latin America and Caribbean 2.62 3.13 2.69 2.84 2.87 2.97 2.70 3.14 2.89 (1.77 2.63 4.04)
Western Europe 2.46 2.91 2.56 2.68 2.85 2.97 3.02 2.89 2.82 (1.41 2.36 3.93)
Japan 2.77 3.23 2.80 2.85 2.28 2.57 2.32 2.59 2.61 (1.35 2.52 4.03)
Asia (excluding Japan) 2.28 2.61 2.37 2.47 2.46 2.44 2.41 2.47 2.44 (1.34 2.01 3.24)
Eastern Europe 1.69 2.12 2.12 1.83 2.15 2.26 1.89 1.77 1.96 (1.20 1.58 2.29)
Mean (for all regions) 2.77 3.17 2.83 2.93 3.06 3.18 3.21 3.10
Figure 1 depicts the worldwide trends of research productivity in the period 1995–2002. USA ranks first among all studied world regions, even during the period 2000–2002 in which investigators from Western Europe published a greater number of articles than investigators from USA. Eastern Europe had the most significant relative growth in the "total product" of research between 1995 and 2002.
Figure 1 Graph displaying the worldwide trends of "total product" of research productivity (number of articles published multiplied by their impact factor) in Infectious Diseases, for different world regions, in the period 1995–2002.
Table 4 presents the quality and quantity of published research adjusted for the regional population and to the gross national income per capita (GNIPC). Specifically, it presents the ratio of scientific "total product" per population divided by the gross national income per capita for each region annually and the respective mean ratio for the whole period. USA and Canada are on the top of this list regarding the cumulative production of research in Infectious Diseases during the period 1995–2002. Interestingly, with the aforementioned adjustments Oceania ranks third on this list.
Table 4 Research output of different world areas, published in journals included in the category of "Infectious Diseases" of the Institute for Scientific Information (ISI), adjusted for population and gross national income per capita (GNIPC).
Number of publications multiplied by the impact factor per million of population divided by the GNIPC (in 10,000 1995 US dollars per capita)
WORLD AREAS 1995 1996 1997 1998 1999 2000 2001 2002 Average
USA 8.8 9.4 8.4 9.5 9.4 10.1 10.5 9.6 9.5
Canada 7.7 7.9 7.1 8.4 8.8 9.5 8.9 9.6 8.5
Oceania 5.3 7.4 6.9 7.8 7.4 10.2 8.5 7.6 7.6
Africa 4.2 3.8 6.0 6.4 8.8 7.7 9.1 7.4 6.7
Western Europe 4.5 5.7 5.0 5.9 6.4 7.4 7.4 6.9 6.2
Asia (excluding Japan) 1.4 1.3 1.7 2.3 2.2 2.5 2.6 2.7 2.1
Latin America & the Caribbean 1.2 1.5 1.5 1.9 2.0 2.4 2.4 3.1 2.0
Eastern Europe 0.8 0.9 1.2 1.2 1.6 2.1 2.0 2.4 1.5
Japan 0.6 0.8 0.9 0.7 0.9 1.2 1.1 1.2 0.9
Discussion
Our study shows that USA and Western Europe make up a striking 80% of the world's research production in terms of both quantity and quality of articles published in Infectious Diseases journals. In addition, our study shows that scientific publications in Infectious Diseases journals increased from 1995 through 2002. The product of the number of published articles multiplied by the impact factor of the journals ("total product"), an index that estimates combined the quantity and quality of produced publications, also increased during the study period. The increased number of published articles in the Infectious Diseases journals, during the study period, is mainly attributed to the introduction of new titles of journals as well as an increase of the number of articles published in some of the journals in the field; both of these trends are mainly the result of increased demand for publishing due to increased production of research data [12,13].
Another interesting finding in our study was the relative reduction in research productivity of the USA compared to the rest of the world. This finding was also observed in the past by other investigators including the U.S. share of research articles in the leading basic and clinical research articles [14,15]. These observations may reflect mainly the improvement of the scientific output, including biomedical research, by several world areas as a result of the general improvement of their economic indices rather than absolute worsening of these factors in the USA.
We provide some data about the relative contribution in research productivity of different world regions in the field of Infectious Diseases. This quantitative data may be used in comparing the productivity of areas of the world with diverse economic status and priorities for funding of different social needs. In addition, our data may be useful as baseline information in evaluating the return of investment on research in Infectious Diseases in areas of the world where this is needed most, i.e. in the developing countries. Specifically, our data show that USA and Canada are the most productive regions when population and GNIPC are taken into account. However, it is interesting that Africa ranks fourth in research productivity when adjustments for these two factors are made. When interpreting this result, one should take into account that a large part of the research originating from this region is the result of multinational/multiregional collaborations, a fact that was not evaluated in this study. Nevertheless, our analysis shows that articles produced by investigators in Africa represent an important scientific contribution to the field of Infectious diseases, due to the very low GNIPC of the area, as well as a satisfactory return of the resources invested for Infectious Diseases research in the area.
Our study has several limitations in both the collection and interpretation of data. First, we used JCR criteria for including medical journals in the study. Articles published in non JCR-cited journals were not included, although they contribute to scientific production [16]. Moreover, we used the JCR impact factor. Although the impact factor has often been criticized as a tool for measuring scientific research quality [17-19], thus far it has not been replaced by any other worldwide-accepted method. JCR uses several criteria in order to include a journal in its databases, and up today the impact factor represents the best method of biomedical journal categorization [20-22]. Also, we used the PubMed, which is an easily accessible and widely used database. Nevertheless, some scientific articles are not included in this database and consequently were not analyzed in our study. In addition, in PubMed only the address of the first author is registered; thus studies that were created by multinational/multi-regional cooperation were counted as originating from only one region of the world. Another problem with the collection of data was associated with the fact that the search system we created was not able to retrieve the addresses of all articles. However, we managed to retrieve 98.5% of all published articles by performing meticulous searches for the address of the first author. Therefore, we assumed that the number of missed articles did not significantly affect our study results.
Another limitation is associated with the division of the world into different regions. Our categorization takes into account geographic, economic, and, most importantly, scientific criteria but despite that, alternative approaches would also be appropriate. For example, Canada could be grouped together with USA, and Japan could be studied together with the other Asian countries. Nevertheless, Canada and Japan represent powerful autonomous scientific world regions and thus we examined them as separate regions. In addition, when interpreting the results, one should take into account that many articles regarding infectious diseases are published in journals of other JCR categories such as "Medicine, General and Internal", "Medicine, Research and Experimental", "Virology", and "Parasitology" and not in the "Infectious Diseases" category. However, we believe that this fact adds no systematic bias in the analysis of our data.
Conclusion
In summary, we evaluated the worldwide trends of research productivity in the field of Infectious Diseases during an 8-year recent period. The results of this study showed a reassuring trend; the fact that developing world regions achieved a higher rate of increase of research productivity than the developed world regions. This is probably the result of increased awareness about the significance of infectious diseases in the developing world regions as well as improved infrastructure supporting research and development in these areas.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
MEF conceived the idea for the study; IAB and PIV collected the data; KP and AIK did the statistical analysis; IAB drafted the manuscript; all authors contributed in the writing and preparation of the manuscript. All authors read and approved the final manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
==== Refs
Keiser J Utzinger J Tanner M Singer BH Representation of authors and editors from countries with different human development indexes in the leading literature on tropical medicine: survey of current evidence BMJ 2004 328 1229 32 15059851 10.1136/bmj.38069.518137.F6
Vergidis PI Karavasiou AI Paraschakis K Bliziotis I Papastamataki PA Falagas ME A bibliometric analysis of worldwide trends in research productivity in Microbiology Eur J Clin Microbiol Infect Dis
Rosmarakis ES Vergidis PI Soteriades ES Paraschakis K Papastamataki PA Falagas ME Estimates of global production in cardiovascular diseases research Int J Cardiol
Rahman M Fukui T Biomedical publication-global profile and trend Public Health 2003 117 274 80 12966750 10.1016/S0033-3506(03)00068-4
Ugolini D Mela GS Oncological research overview in the European Union. A 5-year survey Eur J Cancer 2003 39 1888 94 12932667 10.1016/S0959-8049(03)00431-3
Mela GS Martinoli C Poggi E Derchi LE Radiological research in Europe: a bibliometric study Eur Radiol 2003 13 657 62 12664100
Ramos JM Gutierrez F Masia M Martin-Hidalgo A Publication of European Union research on infectious diseases (1991–2001): a bibliometric evaluation Eur J Clin Microbiol Infect Dis 2004 23 180 4 14986155 10.1007/s10096-003-1074-4
National Library of Medicine Index Medicus database (PubMed), Bethesda, Maryland 2004
Institute for Scientific Information. SCI Science Citation Index-Journal Citation Reports, Philadelphia, The Institute for Scientific Information 2004
United Nations Statistical Yearbook 42nd issue, CD-Rom Edition, United Nations, New York 2004
World Development Indicators The World Bank, Washington 2004
Boldt J Haisch G Maleck WH Changes in the impact factor of anesthesia/critical care journals within the past 10 years Acta Anaesthesiol Scand 2000 44 842 9 10939697 10.1034/j.1399-6576.2000.440710.x
Jemec GB Impact factors of dermatological journals for 1991–2000 BMC Dermatol 2001 1 7 11710969 10.1186/1471-5945-1-7
Stossel TP Stossel SC Declining American representation in leading clinical-research journals N Engl J Med 1990 322 739 742 2308603
Rahman M Fukui T A decline in the U.S. share of research articles N Engl J Med 2002 347 1211 2 12374892 10.1056/NEJM200210103471523
Winkmann G Schweim HG [Medical-bioscientific databanks and the Impact Factor] Dtsch Med Wochenschr 2000 125 1133 41 11147369 10.1055/s-2000-7581
Whitehouse GH Impact factors: facts and myths Eur Radiol 2002 12 715 7 11960216 10.1007/s00330-001-1212-2
Seglen PO Why the impact factor of journals should not be used for evaluating research BMJ 1997 314 498 502 9056804
Barnaby DP Gallagher EJ Alternative to the Science Citation Index impact factor as an assessment of emergency medicine's scientific contributions Ann Emerg Med 1998 31 78 82 9437346
Gensini GF Conti AA [The impact factor: a factor of impact or the impact of a (sole) factor? The limits of a bibliometric indicator as a candidate for an instrument to evaluate scientific production] Ann Ital Med Int 1999 14 130 3 10399377
Garfield E Citation indexes for science; a new dimension in documentation through association of ideas Science 1955 122 108 11 14385826
Luukkonen T Bibliometrics and evaluation of research performance Ann Med 1990 22 145 50 2393549
|
15780136
|
PMC1274272
|
CC BY
|
2021-01-04 16:28:15
|
no
|
BMC Infect Dis. 2005 Mar 21; 5:16
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-16
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-171578414810.1186/1471-2334-5-17Research ArticleEfficacy of two distinct ethanol-based hand rubs for surgical hand disinfection – a controlled trial according to prEN 12791 Kampf Günter [email protected] Christiane [email protected] Bode Chemie GmbH & Co., Scientific Affairs, Melanchthonstr. 27, 22525 Hamburg, Germany2 Institut für Hygiene und Umweltmedizin, Ernst-Moritz-Arndt Universität Greifswald, Walther-Rathenau-Str. 49a, 17489 Greifswald, Germany3 Bode Chemie GmbH & Co., Microbiology, Melanchthonstr. 27, 22525 Hamburg, Germany2005 22 3 2005 5 17 17 30 12 2004 22 3 2005 Copyright © 2005 Kampf and Ostermeyer; licensee BioMed Central Ltd.2005Kampf and Ostermeyer; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Aim of the study was to determine the efficacy of two distinct ethanol-based hand rubs for surgical hand disinfection in a controlled cross-over trial according to prEN 12791.
Methods
20 subjects were included. Hands were washed for 1 min with soap. The bacterial prevalue was obtained by rubbing finger tips in TSB for 1 min. Then, each subject treated the hands with the reference procedure (n-propanol, 60% v/v) or the product (Sterillium® Rub, based on 80% ethanol; Avagard, based on 61% ethanol and 1% chlorhexidine gluconate) which were all applied in 3 to 4 portions each of 3 ml for a total of 3 min. Bacterial postvalues (immediate effect) were taken from one hand, the other hand was gloved for 3 h. After gloves were taken off the second postvalue was taken for the assessment of a sustained effect.
Results
Bacterial pre-values were between 4.38 ± 0.66 and 4.46 ± 0.71. Sterillium® Rub achieved the required immediate (mean log10-reduction of 2.59 ± 1.19) and sustained effect (1.73 ± 1.08) compared with the reference treatment (immediate effect: 2.58 ± 1.16; sustained effect: 1.67 ± 0.96). Avagard, however, did not achieve the required immediate (1.82 ± 1.40) and sustained effect (1.41 ± 1.08) in comparison to the reference disinfection (immediate effect: 2.98 ± 0.90; sustained effect: 2.56 ± 1.17; p < 0.01; Wilcoxon test).
Conclusion
Based on our data, Sterillium® Rub can be regarded to be effective for surgical hand disinfection, but Avagard can not. The addition of 1% chlorhexidine gluconate to 61% ethanol (w/w) did not outweigh an ethanol concentration of 80% (w/w).
==== Body
Background
The new CDC guideline on hand hygiene has indicated that the efficacy of alcohols is superior to many other active agents such as chlorhexidine gluconate or povidone iodine, also on the resident hand flora [1]. Alcohol-based hand rubs are commonly used for surgical hand disinfection in Europe [2]. Their in vivo efficacy is usually tested according to prEN 12791 under practical conditions against a reference treatment [3]. This means that a product shall not be significantly less effective compared to a reference alcohol after 0 and 3 h (gloved hand). This test method is well suitable to discriminate the efficacy of various types of preparations based on different active agents [4]. To our knowledge, the efficacy of preparations for surgical hand disinfection based on different concentrations of ethanol has never been compared according to prEN 12791. The aim of this study was to evaluate the efficacy of two ethanol-based hand rubs for surgical hand disinfection, Sterillium Rub (80% ethanol, w/w) and Avagard (61% ethanol w/w, and 1% chlorhexidine gluconate) according to prEN 12791.
Methods
Twenty subjects were included for each of two experiments. Hands were pre-washed with soap for 1 min. The bacterial prevalue was obtained by rubbing finger tips in tryptic soy broth (TSB) for 1 min. Afterwards, each subject treated the hands with the reference alcohol (n-propanol, 60% v/v) or the product. For the reference disinfection, n-propanol was applied in 3 to 4 portions each of 3 ml in order to keep the skin moist for a total of 3 min. Sterillium Rub™ based on 80% w/w ethanol and Avagard based on 61% w/w ethanol and 1% chlorhexidine gluconate were also applied in 3 to 4 portions in order to keep the skin moist for a total of 3 min. Bacterial postvalues (immediate effect) were taken from one hand by rubbing finger tips in TSB containing neutralizers (3% Tween 80, 3% lecithin, 0.1% histidine, and 0.1% cysteine) for 1 min, the other hand was gloved for 3 h. After gloves were taken off the second postvalue was taken by rubbing finger tips in TSB for 1 min for the assessment of a sustained effect. The bacterial concentration in the sampling fluid was determined by serial dilution and surface culture. The differences between the log10 pre- and postvalues were calculated individually for each subject [5]. Means of these differences were analyzed with the Wilcoxon matched-pairs signed-ranks test [6].
Results
Sterillium Rub™ was found to be equally effective as the reference alcohol both in the immediate effect and after 3 h. The difference of the mean bacterial reduction at 0 h and 3 h between the reference treatment and Sterillium Rub™ was not significant (p > 0.1; Wilcoxon matched-pairs rank test; Table 1).
Table 1 Mean log10-reduction (RF) ± s.d. of Sterillium Rub™ (3 min) and Avagard in comparison to the reference alcohol (60% v/v n-propanol; 3 min) for surgical hand disinfection according to prEN 12791.
Product Pre-value 0 h 3 h
mean RF p-value mean RF p-value
Sterillium Rub 4.39 ± 0.83 2.59 ± 1.19 > 0.1 1.73 ± 1.08 > 0.1
Reference treatment 4.44 ± 0.90 2.58 ± 1.16 1.67 ± 0.96
Avagard 4.46 ± 0.71 1.82 ± 1.40 0.009 1.41 ± 1.08 0.008
Reference treatment 4.38 ± 0.66 2.98 ± 0.90 2.56 ± 1.17
Avagard was found to be less effective than the reference alcohol in both the immediate effect (0 h) and after 3 h. The difference of the mean bacterial reduction between the reference treatment and Avagard was significant at 0 h (p = 0.009) and 3 h (p = 0.008; Table 1).
Discussion
Sterillium Rub™ was found to meet the requirements of prEN 12791 (version 1997) on the bactericidal efficacy for a surgical hand rub, but Avagard did not. The reason is probably a too low concentration of ethanol (61% w/w) in Avagard [7]. It has been shown earlier that ethanol at a concentration of 60% is far less effective against the resident hand flora than ethanol at 80% or more [8,9]. In addition, chlorhexidine gluconate at 1% in Avagard did not compensate for the low efficacy of 61% w/w ethanol. Even after 3 hours, there was no sustained effect under the gloved hand which raises doubts on the justification of this agent in the formulation.
This finding is in line with previously reported data. The efficacy of ethanol-based hand rubs on the resident hand flora varies considerably depending mainly on the concentration of the active agent. An immediate effect of 1.0 to 1.32 log10-reduction has been described with 70% w/w ethanol, a better effect of the resident skin flora can be found with ethanol at 85% w/w (mean reduction: 2.1 to 2.5 log10-steps) or 95% w/w (mean reduction: 2.1 log10-steps) [7]. The combination of 61% ethanol with 1% chlorhexidine gluconate has been described earlier to have superior bactericidal efficacy compared with an antimicrobial liquid soap based on 4% chlorhexidine, especially after 5 and 21 days use [10]. In another report the combination of 61% ethanol with 1% chlorhexidine was significantly more effective than 4% chlorhexidine soap on day 1 and 2 but not on day 5 [11] which is to some extent controversial to the data derived from the other study. In general, a better efficacy should be expected with an alcohol-based leave-on preparation containing chlorhexidine gluconate compared with a chlorhexidine-containing rinse-off preparation.
Alcohols without the addition of non-volatile agents such as quaternary ammonium compounds or chlorhexidine gluconate are regarded to have no sustained efficacy [1]. It is quite difficult to clearly define a sustained activity in surgical hand antisepsis. In the new CDC guideline, "persistent" activity is defined as the prolonged or extended antimicrobial activity that prevents or inhibits the proliferation or survival of microorganisms after application of the product [1]. But it remains unclear how such a persistent effect can be determined. According to the European norm prEN 12791 (version 1997) on products for surgical hand disinfection, a preparation has sustained efficacy if the mean RF is not significantly lower after 3 h, compared with the reference treatment. This reference treatment itself leads to a mean bacterial density on the hands which is usually significantly lower after 3 h compared to baseline [12] which can be described as a sustained efficacy. A persistent efficacy was defined as an efficacy after 3 hours which is significantly superior compared with the reference treatment regardless of the presence of a non-volatile active agents [5]. This definition was based on the knowledge that the reference alcohol does not contain any non-volatile active agent. A preparation with a non-volatile active agent such as chlorhexidine gluconate, however, may have an additional effect in comparison to the reference treatment. But following this definition, a preparation based on 85% ethanol (w/w) without any non-volatile active agent such as chlorhexidine gluconate has been described to have persistent activity [9]. In the new version of prEN 12791 (2003), the terminology has been changed. Now the term "sustained effect" describes the formerly "persistent effect" which may lead to some confusion. It would certainly be helpful to clearly define scientific terms and appropriate requirements in surgical hand disinfection which, ideally, are accepted worldwide [13].
The potential benefit of chlorhexidine gluconate is thought to be a prolonged effect. Repeated application is thought to increase the antimicrobial activity on the resident hand flora. If chlorhexidine gluconate is used in a "leave-on" preparation like Avagard it can be expected that the non-volatile active agent chlorhexidine gluconate remains on the skin and will continue to have antimicrobial activity. In our study we were not able to show that such an effect can be measured after a single application. In addition, permanent exposure to chlorhexidine salts has been shown to lead to adaptation or even resistance. Exposure of Pseudomonas aeruginosa to 5 mg/L chlorhexidine diacetate over a period of 12 days was able to increase the minimal inhibitory concentration (MIC) from ≤ 10 mg/L to 70 mg/L [14]. A similar observation was made after exposure of six strains of Pseudomonas stutzeri to gradually increasing concentrations of chlorhexidine diacetate which led to an increase of the MIC from 2.5 to 50 mg/L after 12 days [15]. Adaptation was also found with Streptococcus sanguis strains which were exposed to variable concentrations of chlorhexidine over a period of 10 weeks resulting in an increase of the MIC from 16 mg/L to up to 128 mg/L [16]. The resistance which has been developed on permanent exposure to chlorhexidine has been described to be stable and to include cross-resistance to other antiseptic agents (like triclosan or benzalkonium chloride) and antibiotics (like gentamicin, ampicillin, and erythromycin) [15]. Although this effect has to our knowledge not been reported with resident skin bacteria, it nevertheless underlines the potential of chlorhexidine gluconate once bacteria are permanently exposed to sub-lethal concentrations of the agent. If even antibiotic resistance can emerge by permanent exposure to chlorhexidine gluconate the potential benefit should be substantial to justify the addition of chlorhexidine gluconate in a surgical hand rub preparation from our point of view.
It is known that chlorhexidine salts are difficult to neutralize in experimental settings which may lead to false favorable results [17-20]. In one study, there was no effect at all against enterococci including VRE if neutralization of remaining chlorhexidine was ensured after the exposure time [21]. In the present study, neutralization of residual chlorhexidine was achieved after the exposure time which may be the explanation for the lacking effect after 3 hours.
In Europe, the efficacy of alcohol-based hand rubs for surgical hand disinfection is assessed using prEN 12791 [3]. The test principle is the cross-over evaluation with a reference alcohol (n-propanol 60%, v/v) which has been shown to have the best efficacy on the resident hand flora together with a "within-subject-comparison" of the bacterial reductions [8]. In addition, the test method has been described to yield reproducible results [22]. In the US, hand antiseptics are usually evaluated according to the test method published in the tentative final monograph for healthcare antiseptic products [23]. This test method is designed for rinse-off preparations. It does not include a reference treatment in the test on volunteers, but a preparation has to fulfill certain minimum requirements at various test days with higher requirements after 5 days [23]. This test philosophy is hard to understand and to justify since a patient who is treated on a Monday should have the same level of safety compared with the patient who is treated on a Friday. Inclusion of a reference treatment can be regarded to be superior since tests are done on the resident hand flora which may vary considerably in number and composition of bacterial species. The efficacy of a test preparation has to be equal to a suitable reference procedure at any time point providing the same level of safety for any patient. Only with inclusion of a reference treatment a true comparison of the efficacy between preparations can be achieved [4].
Alcohol-based hand rubs have been shown to have a better antimicrobial efficacy on both the transient and resident hand flora [2,7]. That is why is has been recommended in the new CDC guideline on hand hygiene that they may well be used for surgical hand disinfection [1] although it remains unclear if the use of preparations with a higher effect on the hand flora has an additional impact on the incidence of surgical site infections [24]. But another benefit has also been described: The use of a well formulated alcohol-based hand rub can improve the skin conditions of the surgeons resulting in significantly less skin dryness and significantly less skin irritation once they have changed from an antimicrobial soap to a well formulated alcohol-based hand rub [24]. Apart from the efficacy of a preparation, the dermal tolerance should also be considered [25].
Conclusion
A high concentration of ethanol (80% w/w) was found to be effective on the resident hand flora after 0 and 3 hours, a lower concentration of ethanol (61% w/w), however, was not sufficiently effective if tested according to prEN 12791. The addition of 1% chlorhexidine gluconate to the 61% ethanol did not provide a substantial improvement of the bactericidal efficacy after 3 hours.
Competing interests
Both authors are paid employees of Bode Chemie GmbH & Co., Hamburg, Germany.
Authors' contributions
GK designed the study, analysed and interpreted the data. CO coordinated the study and acquired the data. Both authors drafted and revised the article.
Pre-publication history
The pre-publication history for this paper can be accessed here:
==== Refs
Boyce JM Pittet D Guideline for hand hygiene in health-care settings. Recommendations of the healthcare infection control practices advisory committee and the HICPAC/SHEA/APIC/IDSA hand hygiene task force MMWR - Morbidity & Mortality Weekly Report 2002 51 1 45
Rotter ML Arguments for the alcoholic hand disinfection Journal of Hospital Infection 2001 48 S4 S8 11759024 10.1016/S0195-6701(01)90004-0
Labadie JC Kampf G Lejeune B Exner M Cottron O Girard R Orlick M Goetz ML Darbord JC Kramer A Recommendation for surgical hand disinfection - requirements, implementation and need for research. A proposal by representatives of the SFHH, DGHM and DGKH for a European discussion. Journal of Hospital Infection 2002 51 312 315 12183149 10.1053/jhin.2002.1243
Marchetti MG Kampf G Finzi G Salvatorelli G Evaluation of the bactericidal effect of five products for surgical hand disinfection according to prEN 12054 and prEN 12791 Journal of Hospital Infection 2003 54 63 67 12767849 10.1016/S0195-6701(03)00039-2
prEN 12791 Chemical disinfectants and antiseptics. Surgical hand disinfection. Test method and requirement (phase 2, step 2) 1997 Brussels, CEN - Comité Européen de Normalisation
Altman DG Practical statistics for medical research 1991 1 London, Chapman & Hall
Kampf G Kramer A Epidemiologic background of hand hygiene and evaluation of the most important agents for scrubs and rubs Clinical Microbiology Reviews 2004 17 863 893 15489352 10.1128/CMR.17.4.863-893.2004
Rotter ML Mayhall CG Hand washing and hand disinfection Hospital epidemiology and infection control 1999 2nd Philadelphia, Lippincott Williams & Wilkins 1339 1355
Kampf G Kapella M Suitability of Sterillium Gel for surgical hand disinfection Journal of Hospital Infection 2003 54 222 225 12855239 10.1016/S0195-6701(03)00087-2
Larson EL Aiello AE Heilman JM Lyle CT Cronquist A Stahl JB Comparison of different regimes for surgical hand preparation AORN Journal 2001 73 412 420 11218929
Mulberry G Snyder AT Heilman J Pyrek J Stahl J Evaluation of a waterless, scrubless chlorhexidine gluconate / ethanol surgical scrub for antimicrobial efficacy American Journal of Infection Control 2001 29 377 382 11743484 10.1067/mic.2001.118842
Kampf G Ostermeyer C Influence of applied volume on efficacy of 3-minute surgical reference disinfection method prEN 12791 Applied and Environmental Microbiology 2004 70 7066 7069 15574901 10.1128/AEM.70.12.7066-7069.2004
Kampf G Goroncy-Bermes P Fraise A Rotter M Terminology in surgical hand disinfection - a new tower of Babel in infection control J Hosp Infect 2005 58 269 271 15694992 10.1016/j.jhin.2004.09.020
Thomas L Maillard JY Lambert RJ Russell AD Development of resistance to chlorhexidine diacetate in Pseudomonas aeruginosa and the effect of a "residual" concentration Journal of Hospital Infection 2000 46 297 303 11170761 10.1053/jhin.2000.0851
Tattawasart U Maillard JY Furr JR Russell AD Development of resistance to chlorhexidine diacetate and cetylpyridinium chloride in Pseudomonas stutzeri and changes in antibiotic susceptibility Journal of Hospital Infection 1999 42 219 229 10439995 10.1053/jhin.1999.0591
Westergren G Emilson CG In vitro development of chlorhexidine resistance in Streptococcus sanguis and its transmissibility by genetic transformation Scandinavian Journal of Dental Research 1980 88 236 243 6932090
Kampf G Höfer M Rüden H Inaktivierung von Chlorhexidin bei der in vitro Desinfektionsmitteltestung Zentralblatt für Hygiene und Umweltmedizin 1998 200 457 464
Sheikh W Development and validation of a neutralizer system for in vitro evaluation of some antiseptics Antimicrobial Agents and Chemotherapy 1981 19 429 434 7247368
Shimizu M Okuzumi K Yoneyama A Kunisada T Araake M Ogawa H Kimura S In vitro antiseptic susceptibility of clinical isolates from nosocomial infections Dermatology 2002 204 21 27 12011516 10.1159/000057720
Werner HP Engelhardt C Problematik der Inaktivierung am Beispiel des in vitro-Tests Hygiene + Medizin 1978 3 326 330
Kampf G Höfer M Wendt C Efficacy of hand disinfectants against vancomycin-resistant enterococci in vitro Journal of Hospital Infection 1999 42 143 150 10389064 10.1053/jhin.1998.0559
Rotter ML Simpson RA Koller W Surgical hand disinfection with alcohols at various concentrations: parallel experiments using the new proposed European standards methods Infection Control and Hospital Epidemiology 1998 19 778 781 9801287
Anonymous Tentative final monograph for health care antiseptic products; proposed rule Federal Register 1994 59 31401 31452
Parienti JJ Thibon P Heller R Le Roux Y von Theobald P Bensadoun H Bouvet A Lemarchand F Le Coutour X Hand-rubbing with an aqueous alcoholic solution vs traditional surgical hand-scrubbing and 30-day surgical site infection rates - a randomized equivalence study JAMA 2002 288 722 727 12169076 10.1001/jama.288.6.722
Kampf G Rudolf M Shaffer M Dermal tolerance of Sterillium Rub in the repeated insult patch test Infect Control Hosp Epidemiol 2000 21 104
|
15784148
|
PMC1274273
|
CC BY
|
2021-01-04 16:28:16
|
no
|
BMC Infect Dis. 2005 Mar 22; 5:17
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-17
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-181579481710.1186/1471-2334-5-18Research ArticleA2 gene of Old World cutaneous Leishmania is a single highly conserved functional gene Garin Yves JF [email protected] Pascale [email protected] Francine [email protected] Jean-Pierre [email protected] Francis [email protected] Frédéric [email protected] Laboratoire de Parasitologie-Mycologie, Hôpital Saint-Louis, Assistance Publique Hôpitaux de Paris, U.F.R. Lariboisière, Université Paris VII, France2 Laboratoire de Parasitologie et Centre National de Référence des Leishmania, C.H.U. de Montpellier, Montpellier, France2005 28 3 2005 5 18 18 27 10 2004 28 3 2005 Copyright © 2005 Garin et al; licensee BioMed Central Ltd.2005Garin et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Leishmaniases are among the most proteiform parasitic infections in humans ranging from unapparent to cutaneous, mucocutaneous or visceral diseases. The various clinical issues depend on complex and still poorly understood mechanisms where both host and parasite factors are interacting. Among the candidate factors of parasite virulence are the A2 genes, a family of multiple genes that are developmentally expressed in species of the Leishmania donovani group responsible for visceral diseases (VL). By contrast, in L. major determining cutaneous infections (CL) we showed that A2 genes are present in a truncated form only. Furthermore, the A2 genomic sequences of L. major were considered subsequently to represent non-expressed pseudogenes [1]. Consequently, it was suggested that the structural and functional properties of A2 genes could play a role in the differential tropism of CL and VL leishmanias. On this basis, it was of importance to determine whether the observed structural/functional particularities of the L. major A2 genes were shared by other CL Leishmania, therefore representing a proper characteristic of CL A2 genes as opposed to those of VL isolates.
Methods
In the present study we amplified by PCR and sequenced the A2 genes from genomic DNA and from clonal libraries of the four Old World CL species comparatively to a clonal population of L. infantum VL parasites. Using RT-PCR we also amplified and sequenced A2 mRNA transcripts from L. major.
Results
A unique A2 sequence was identified in Old World cutaneous Leishmania by sequencing. The shared sequence was highly conserved among the various CL strains and species analysed, showing a single polymorphism C/G at position 58. The CL A2 gene was found to be functionally transcribed at both parasite stages.
Conclusion
The present study shows that cutaneous strains of leishmania share a conserved functional A2 gene. As opposed to the multiple A2 genes described in VL isolates, the CL A2 gene is unique, lacking most of the nucleotide repeats that constitute the variable region at the 5'end of the VL A2 sequences. As the variable region of the VL A2 gene has been shown to correspond to a portion of the protein which is highly immunogenic, the present results support the hypothesis of a possible role of the A2 gene in the differential tropism of CL and VL leishmania parasites.
==== Body
Background
Leishmaniases are among the most important protozoan infections that affect humans in the world. The disease is widespread in 88 endemic countries with 350 million people at risk, 12 million people permanently affected, and an estimated annual incidence of 1.5–2 million cases [2]. This results in a global morbidity of 2,357 thousands DALYs (Disability Adjusted Life Years: number of healthy years of life lost due to premature death and disability) and a mortality rate of 59,000/year [3].
A surprisingly broad spectrum of clinical expressions is observed in humans ranging from asymptomatic to cutaneous (CL), diffuse cutaneous, mucocutaneous and visceral (VL) diseases, and an intermediary form known as post-Kala-Azar dermal leishmaniasis.
The various clinical issues of Leishmania infection depend on a complex host-parasite relationships where both the genetic or immunological status of the host [4-6] and the proper parasite biodiversity in terms of tropism and virulence [7,8] appear as determinant factors. A number of parasitic factors have been identified as susceptible to play a role in virulence/protection mechanisms in leishmaniases [9]. Among these, since its first identification in Leishmania infantum [10] several lines of evidence indicate that the A2 gene/protein family could be one of the most eligible candidate factor of virulence in VL infections: i) A2-proteins and mRNA transcripts are developementally expressed at the amastigote intracellular stage while undetectable in the promastigote [11], ii) Inhibition of A2-expression in Leishmania donovani using anti-sense RNA or by generation of partial knock-out mutants results in reduced virulence in vivo, iii) by contrast, increased parasite levels are observed in spleen of mice infected with A2-expressing transfected L. major [11,12] iv) A protective immunity can be achieved experimentally in mice by immunization with recombinant A2 protein or DNA vaccination showing that A2 from L. donovani is highly immunogenic and represents a potential antigen for protection in VL [13,14] and more recently in L. amazomensis infections [15].
A2 genes were detected by karyotype analysis in L. donovani, L. infantum and L. chagasi (Old World and New World VL) and in L. mexicana and L. amazonensis (New World DCL and MCL, respectively) but not in cutaneous species from the Old World (L. tropica, L. aethiopica and L. major) and the New World (L. brasiliensis, L. guyanensis and L. panamensis) [16]. Accordingly, A2-antibodies were found in sera from human and dogs naturally infected with L. chagasi (VL) [17] and in patients with VL in Sudan and India and CL due to L. mexicana, while they were not detected in L. tropica and L. brasiliensis infections (CL) [16].
While long considered absent in the L. tropica group, we identified by sequencing an A2 gene from crude PCR products of two strains of L. major (AF532102, AF532103) showing that the gene is present in L. major in a truncated form lacking most of the repeated motives that are present at the 3'end variable region of the VL A2 genes. Moreover, the L. major A2 gene was found subsequently to be non expressed and was considered to represent a pseudogene [1].
These observations raised the question of a possible role of the structure/functionality of A2 genes in the cutaneous or visceral tropism of leishmania parasites. As no data were available on the A2 gene of CL Leishmania except for L. major, our objective was to investigate this gene in Old World CL species. We amplified and sequenced A2-genes of additional strains of Old World CL species and in a clonal lineage of a L. infantum mediterranean strain isolated from a VL patient. Our results show that: i) The A2 sequence is extremely conserved both among strains and species of Old World CL Leishmania, ii) The CL A2 gene is a single copy gene of only 153 base pairs (bp) encoding for a protein of 51 amino acids, as opposed to A2 of VL species that are multicopy genes of varying length, ii) The CL A2 gene is functionally transcribed at the promastigote and amastigote stages.
Methods
Parasites
Strains. Six strains of one visceral and four cutaneous Old World Leishmania species were used for sequencing: L. infantum and L. major, L. tropica, L. killicki and L. aethiopica, respectively (Table I). Four of the five cutaneous strains were reference strains recommended by the W.H.O. [18]. In addition two L. donovani VL strains LEM3467 and LEM3566 were used for PCR analysis. All strains originated from the International Leishmania Cryobank and Identification Center, Montpellier, France.
Table 1 Tropism Species Strain Zymodeme Allele type GeneBank accession number
gDNA2 mRNA Protein3
Visceral (LV) L. infantum MHOM/FR/92/ LEM2385 Cl 1 MON-29 II AY255807 AAP21103
III AY255808 AAP21104
IV AY255809 AAP21105
Cutaneous L. major IPAP/MA/86/ LEM898 MON-25 I.1 AF5321022 AY25581 AAM95954
MHOM/SU/73/ 5 ASKH1 MON-4 I.2 AY185122 AAP21106
AAO27297
(LC) L. aethiopica MHOM/ET/72/ L1001 MON-14 I.1 AY255804 AAP21100
L. killicki MHOM/TN/86/ LEM904- CL1 MON-8 I.1 AY255805 AAP21101
L. tropica MHOM/SU/74/ K271 MON-60 I.1 AY255806 AAP21102
1WHO recommended reference strains. 2 Direct genomic DNA sequencing.
3Predicted protein.
Strains originated from the International Leishmania Cryobank and Identification Center, Montpellier, France.
Parasite clones. Parasite clonal lineages were obtained from L. infantum MHOM/FR/92/LEM2385 and from L. major MHOM/SU/73/5 ASKH strains using a microplate technique as previously described [19].
Culture and isolation. Promastigotes were cultivated at 27°C in HOSMEM liquid medium [20] supplemented with hemin 10 μM (Sigma, Saint Quentin Fallavier, France) and 10% fetal calf serum (Gibco, Cergy-Pontoise, France). Parasites were inoculated into 25 ml culture flasks at day 0 (d0) at a final concentration of 105 ml-1. Amastigote organisms were isolated from foot-pad (L. major) or spleen (L. infantum) of Balb/c mice inoculated subcutaneously or intraveinously with 107 log-phase promastigotes, respectively. Parasites were washed twice in PBS and counted in Malassez chambers.
DNA extraction
Washed parasites (100 μl PBS / ≈109 parasites) were lysed by thermal shock in Eppendorf tubes, 1 mn in boiling water – 1 mn in melting ice, three times. DNA extraction was performed using classical phenol/chloroform/isoamylic alcool protocol and precipitation was made using NaCl/ethanol procedure. The DNA was dissolved in 40 μl of sterile water.
PCR and sequencing
Amplification of the parasite DNA matrix (50 ng) was made using L2/R3 primers (5'-TTGGCAATGCGAGCGTCACAGTC / 5'- CAACGCGTACGATAATGCCACA). The L2/R3 primers correspond to the 5' end position 16301 and 16603 of the inverse-complementary strand of the AC010851 sequence, respectively. The PCR was performed in a reaction mixture of 50 μl containing either 1 or 3 mM MgCl2, 200 μM each dNTP, 25 pmol of each primer (Proligo, Paris, France), 1 U Taq polymerase (Eurogentec, Seraing, Belgium). L2/R3-PCR conditions consisted to denaturation for 3 mn at 94°C, followed by 35 amplification cycles at 94°C for 1 mn, 1 mn at 58°C, 1 mn at 72°C, then one cycle at 72°C for 5 mn. Amplification of cDNA from bacterial culture medium (0.5 μl) was made using M13 forward-20 / M13 reverse (Qiagen PCR cloning kit, Qiagen, Courtaboeuf, France) with 1 mM MgCl2. M13 forward-20 / M13 reverse -PCR conditions consisted in a hot-start denaturation for 10 mn 95°C, followed by addition of 1 U Taq polymerase, 30 amplification cycles at 94°C for 30 sec, 30 sec at 48°C, 1 mn at 72°C, then one cycle at 72°C for 5 mn. Five microliters of PCR product was electrophoresed in 2% agarose gel in the presence of ethidium bromide, and visualized under UV light. A 50-bp ladder (Sigma) was used as MW marker.
For sequencing, the two strands of PCR-amplified DNA were purified with QIAquick PCR Purification Kit (Qiagen) and sequenced with the corresponding PCR primer set using the BigDye Terminator Sequencing Kit V3.1 (Applied Biosystems, Courtaboeuf, France) on an automated sequencer 3100 Genetic analyser (Applied Biosystems).
RNA extraction and reverse transcription-PCR (RT-PCR)
L. major IPAP/MA/86/LEM898 total RNA was extracted from promastigote cultures or infected organs (≈106-107 parasites) using Rneasy Plant Mini Kit (Qiagen). To eliminate any remaining DNA the RNA extract mixture (5 μl) was additionnaly treated by Dnase Rnase-free (Eurogentec) for 30 min at 37°C in a final volume of 30 μl. As a negative control, an aliquote of the sample (5 μl) was subsequently digested with Rnase A (Qiagen) 700 mg 1 h at 25°C in a final volume of 16 μl. RT was performed in a total volume of 20 μl by 50-min incubation at 42°C followed by 15 min at 70°C to inactivate the reverse transcriptase. The reaction mixture contained a sixth of the initial volume of the RNA extraction products, and the following final reagent concentrations: 1X hexanucleotide mix (Roche, Meylan France), 500 μM dNTP mix, 40U Rnase inhibitor, 1X first strand buffer, 100 mM dithiothreitol, and 200 U Super Script II (Invitrogen, Cergy Pontoise, France). Two microliters of RT products were PCR-amplified with L2/R3 primer set.
DNA libraries
L2/R3-PCR products from genomic DNA matrix were synthetized as described above except for the MgCl2 concentration (3 mM). PCR products were purified on QIAquick column (Qiagen). Poly-A treatment, insertion of the PCR products into pDrive vector, transformation of E. coli EZ competent cells (Qiagen) and cloning were performed using a PCR Cloning Kit (Qiagen) as described by the supplier. A2-containing genomic clones were screened by digestion of M13-PCR products with Sau3AI endonuclease (BioLabs, Saint Quentin en Yvelines, France) which cuts off A2-gene nucleotidic sequences at position 33–34 (reference : L. major MHOM/IR/-/173; AF532103).
GeneBank accession numbers
Accession numbers for genomic DNA, mRNA and putative protein sequences are given in Table I.
Results
PCR
L2/R3 PCR products obtained from parasite crude genomic DNA resolved in different electrophoretic patterns, according to the species. For all CL isolates (Fig. 1-A) one single band of about 260 bp was evidenced when PCR was performed using stringent or not stringent conditions (1 or 3 mM Mgcl2), thus these products were available for direct sequencing. By contrast PCR amplification products could be obtained only under non stringent conditions (3 mM MgCl2) for VL species, resolving in a complex electrophoresis pattern (Fig. 1-B). These products were shown by direct sequencing to be a mixture of A2 sequences and of non specific products resulting probably from a certain degree of mispriming due to non stringent conditions. As a consequence direct sequencing of L2R3 PCR products from crude genomic DNA of VL species could not be performed, thus A2 sequences were obtained from clone libraries of L. infantum MHOM/FR/92/LEM2385-clone 1.
Figure 1 PCR electrophoresis patterns. Electrophoretic patterns of PCR products obtained from crude parasite genomic DNAs using 3 mM MgCl2. A. CL isolates: 1, L. major LEM898; 2, L. major LEM134; 3, L. aethiopica LEM144; 4, L. tropica LEM419; 5, L. killicki LEM904; Light arrow: A2 gene. B. VL isolates: 1, L. infantum LEM2385-cl1; 2, L. donovani LEM3467-cl3 3; L. donovani LEM3566; square bracket: A2-gene area.
The PCR performed with M13 primers on clone libraries corroborated the above results. All genomic-DNA clones originating from L. major, L. tropica, L. aethiopica and L. killicki resolved in a unique band of 300 bp. By contrast, three different bands of 370, 410 and 460 bp were identified from the L. infantum clone library (data not shown). Only 1/4 of the library clones (20 clones were sequenced) corresponded to cloned A2-gene. The other clones corresponded to unknown or microsatellite structures (3 clones) in BLAST analysis study.
Sequencing (Fig. 2)
Figure 2 A2 nucleotide sequences. Leishmania major IPAP/MA/86/LEM 898 and MHOM/SU/73/5 ASKH LEM134 (allele types I.1 and I.2) and Leishmania infantum MHOM/FR/92/LEM2385 clone 1 (allele types II to IV). Internal nucleotide repeats are shown in blue. The polymorphism C/G at position 58 is highlighted in yellow.
Old-World CL strains
A single sequence of 258 nucleotides was evidenced by directly sequencing the crude L2/R3-PCR products of the genomic-DNA of the three L. major strains. The three sequences were found identical with the exception of a single polymorphism C/G at position 58 for the L. major 134 and 173 as compared with L. major 898. Additional nucleotide sequences from 32 genomic clones of the above mentioned L. major strains were all identical to the corresponding crude sequences. Moreover, the sequences obtained from crude L2/R3-PCR products and from 15 genomic clones of L. aethiopica, L. tropica and L. killicki were totally identical to the L. major-898 sequence. These sequences were referred to as A2-gene allele typeI.1 and I.2, respectively.
L. infantum VL strain
Three A2-gene sequences of 371, 413 and 464 nucleotides were isolated from the genomic library of L. infantum MHOM/FR/92/LEM2385 clone 1 (L. infantum-2385.1). These sequences were referred to as A2 type II-, III- and IV-alleles.
Comparative analysis of A2-gene alleles from CL (type I) and VL (Type II, III and IV) strains showed that the genes are composed of a common nucleotide sequence at the 5' end followed by a region of varying length inserted from the position 97 of the ORF to the 3' end which is a stretch of a number of more or less conserved repeated nucleotide patterns.
mRNA expression
Results of RT-PCR on Dnase-treated L. major IPAP/MA/86/LEM898 mRNA extracts are presented in Fig. 3. A single band was evidenced in RT-PCR products from both promastigotes and foot pads of infected mice. A faint band was also observed in popliteal lymph node extracts (data not shown). The sequencing of the RT-PCR cDNA product resulted in a 258-nucleotide sequence (AY255810) identical to the sequence obtained from the crude genomic DNA, and corresponding to a putative A2 protein of 51 AA.
Figure 3 RT-PCR on IPAP/MA/86/LEM898 strain Dnase-treated mRNA extracts. Additional Rnase-digestion (1–3). Uninfected spleen (1,4); Cultured promastigotes (2,5); Foot-pad (3,6). Controls: Genomic DNA (7); H2O (8). Arrow : A2 mRNA transcripts.
Discussion
Leishmania A2-genes were first identified in two strains of the L. donovani complex, L. infantum LV9 and L. donovani 1S2D determining visceral infections [10,21]. In these VL strains A2 genes were shown to be organized in several clusters each comprising multiple A2 genes of varying length that are tandemly associated with related sequences (A2rel) [10,11,22]. However, these results were obtained from leishmania strains isolated from naturally infected hosts which are known to be most likely composed of multiple parasite populations [23,24]. Therefore, the existence of multiple A2 genes remained to be confirmed using a genetically pure parasite clonal lineage. As in the present study three different A2-alleles type II, III and IV were sequenced from the L. infantum MHOM/FR/92/LEM2385 Clone-1 genomic library, our results provide additional evidence that A2 of VL species is a multiple gene family. Alleles type I, II and III differ only in the number and arangement of the repeated motives at the 3'end variable region of the gene as previously described in VL strains [10]. However, in the present study we identified A2 sequences showing a limited number of repeats and consequently a length of only 371 to 464 bp contrasting with the previously published A2 genes of about 700–800 bp. The inability to evidence A2 sequences > 1 kb in the present study is most probably due to the limits of the PCR technique performed on crude genomic parasite DNA or to the absence of long A2 sequences in this strain. Thus these results are not contradictory to the previously published data but bring additional information on the variability of the L. infantum A2 genes.
We previously identified A2 sequences in two strains of L. major IPAP/MA/86/LEM898 and MHOM/IR/00/173 (AF532102 and AF532103, respectively). These sequences were 95% (88 nt/93) identical to the S69693 stage-specific S antigen homolog (A2) of L. infantum VL [10] at the 5'-end of the ORF (nucleotides 74 to 167). By contrast, a major deletion of the 3'end variable region of repeated nucleotide motives was observed in L. major A2 genes.
In the present study, the amplification of Old World Leishmania genomic DNA with L2/R3 A2-gene primer set resolved in a single amplification product on gel electrophoresis, as opposed to the complex pattern observed with L. infantum. Accordingly, a single sequence of 258 nucleotides was obtained by direct sequencing crude DNA products from all CL Leishmania, whatever the strain or species. A common sequence was also isolated from 32 genomic clones of L. major LEM898 and 15 clones of L. tropica, L. aethiopica and L. killicki. The A2 sequence shared by all strains and species of Old World leishmania presented a single polymorphism C/G at position 58 of the ORF. These results show for the first time that the A2 gene of Old World cutaneous Leishmania is unique and highly conserved, contrasting strongly with the multiple A2 sequences of varying length observed in VL isolates.
RT-PCR on mRNA extracts from strain IPAP/MA/86/LEM898 followed by sequencing evidenced that the CL A2 gene is functional. It is noteworthy that this finding does not presume of the expression of the protein at the post-transcriptional level, however it contrasts with the previous suggestion based on the failure to demonstrate A2 gene transcripts in L. major that the A2-gene of L. major is a non-expressed pseudogene [1]. RT-PCR amplification of both promastigote and amastigote mRNAs resulted in a similar signal on gel electrophoresis in the present study, showing that L. major A2-gene is transcribed at both amastigote and promastigote stages. However, these results do not signify that L. major A2 are not developmentally expressed since RT-PCR is not quantitative. Actually, L. infantum promastigotes were reported to express very low levels of A2 [10,22].
Conclusion
The present study evidence that Old World cutaneous species of Leishmania share a common, highly conserved and functional A2 gene. In CL Leishmania the A2 gene is a single gene in which the 3'end variable region is almost entirely deleted, contrasting with VL A2 genes that are a family of multiple genes where the 3'end portion is a variable stretch of nucleotide repeats. Further investigations are needed for exploring the potential role of theses structural differences between CL and VL A2 genes in governing both the proper parasite virulence/tropism and host susceptibility/protection response conditioning the multiple clinical issues observed in human Leishmania infections.
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
YJFG conceived and conducted the study and drafted the manuscript. PM carried out the work at the technical level. FP and JPD supplied and characterized the parasite strains. FD is the Chief Manager of the Laboratory and revised the manuscript and FL directed the biomolecular and bioinformatic analyses and participated to the writing of the manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
We thank H. Bui, Jean Dausset Foundation, Centre d'Etude du Polymorphisme Humain (C.E.P.H.), Paris, France, for the sequencing.
==== Refs
Zhang WW Mendez S Ghosh A Myler P Ivens A Clos J Sacks DL Matlashewski G Comparison of the A2 gene locus in L. donovani and L. major and its control over cutaneous infection J Biol Chem 2003 278 35508 35515 12829719 10.1074/jbc.M305030200
WHO The leishmaniases and Leishmania/HIV co-infections. WHO Fact Sheet No 116 2000
WHO Reducing risks, promoting healthy life. The World Health Report 2002, WHO, Geneva, Switzerland The World Health Report 2002 2002
Blackwell JM Genetic susceptibility to leishmanial infections: studies in mice and man Parasitology 1996 112 Suppl S67 S74 8684837
Gradoni L Gramiccia M Leishmania infantum tropism: strain genotype or host immune status ? Parasitol Today 1994 10 264 267 15275441 10.1016/0169-4758(94)90142-2
Kubar J Marty P Lelievre A Quaranta JF Staccini P Caroli-Bosc C Le Fichoux Y Visceral leishmaniosis in HIV-positive patients: primary infection, reactivation and latent infection. Impact of the CD4+ T-lymphocyte counts AIDS 1998 12 2147 2153 9833855 10.1097/00002030-199816000-00009
Garin YJ Sulahian A Pratlong F Meneceur P Gangneux JP Prina E Dedet JP Derouin F Virulence of Leishmania infantum is expressed as a clonal and dominant phenotype in experimental infections Infect Immun 2001 69 7365 7373 11705909 10.1128/IAI.69.12.7365-7373.2001
Honore S Garin YJ Sulahian A Gangneux JP Derouin F Influence of the host and parasite strain in a mouse model of visceral Leishmania infantum infection FEMS Immunol Med Microbiol 1998 21 231 239 9718213 10.1016/S0928-8244(98)00079-0
Matlashewski G Leishmania infection and virulence Med Microbiol Immunol (Berl) 2001 190 37 42 11770107
Charest H Matlashewski G Developmental gene expression in Leishmania donovani: differential cloning and analysis of an amastigote-stage-specific gene Mol Cell Biol 1994 14 2975 2984 7545921
Zhang WW Matlashewski G Characterization of the A2-A2rel gene cluster in Leishmania donovani: involvement of A2 in visceralization during infection Mol Microbiol 2001 39 935 948 11251814 10.1046/j.1365-2958.2001.02286.x
Zhang WW Matlashewski G Loss of virulence in Leishmania donovani deficient in an amastigote-specific protein, A2 Proc Natl Acad Sci U S A 1997 94 8807 8811 9238059 10.1073/pnas.94.16.8807
Ghosh A Zhang WW Matlashewski G Immunization with A2 protein results in a mixed Th1/Th2 and a humoral response which protects mice against Leishmania donovani infections Vaccine 2001 20 59 66 11567746 10.1016/S0264-410X(01)00322-X
Ghosh A Labrecque S Matlashewski G Protection against Leishmania donovani infection by DNA vaccination: increased DNA vaccination efficiency through inhibiting the cellular p53 response Vaccine 2001 19 3169 3178 11312013 10.1016/S0264-410X(01)00023-8
Coelho EA Tavares CA Carvalho FA Chaves KF Teixeira KN Rodrigues RC Charest H Matlashewski G Gazzinelli RT Fernandes AP Immune responses induced by the Leishmania (Leishmania) donovani A2 antigen, but not by the LACK antigen, are protective against experimental Leishmania (Leishmania) amazonensis infection Infect Immun 2003 71 3988 3994 12819086 10.1128/IAI.71.7.3988-3994.2003
Ghedin E Zhang WW Charest H Sundar S Kenney RT Matlashewski G Antibody response against a Leishmania donovani amastigote-stage-specific protein in patients with visceral leishmaniasis Clin Diagn Lab Immunol 1997 4 530 535 9302200
Carvalho FA Charest H Tavares CA Matlashewski G Valente EP Rabello A Gazzinelli RT Fernandes AP Diagnosis of American visceral leishmaniasis in humans and dogs using the recombinant Leishmania donovani A2 antigen Diagn Microbiol Infect Dis 2002 43 289 295 12151189 10.1016/S0732-8893(02)00410-8
WHO Control of the leishmaniases. Report of a WHO Expert Committee World Health Organ Tech Rep Ser 1990 793 1 158 2124015
Garin YJ Meneceur P Sulahian A Derouin F Microplate method for obtaining Leishmania clonal populations J Parasitol 2002 88 803 804 12197138
Berens RL Marr JJ An easily prepared defined medium for cultivation of Leishmania donovani promastigotes J Parasitol 1978 64 160 627959
Zhang WW Charest H Ghedin E Matlashewski G Identification and overexpression of the A2 amastigote-specific protein in Leishmania donovani Mol Biochem Parasitol 1996 78 79 90 8813679 10.1016/S0166-6851(96)02612-6
Ghedin E Charest H Matlashewski G A2rel: a constitutively expressed Leishmania gene linked to an amastigote-stage-specific gene Mol Biochem Parasitol 1998 93 23 29 9662025 10.1016/S0166-6851(98)00027-9
Pacheco RS Grimaldi GJ Momen H Morel CM Population heterogeneity among clones of New World Leishmania species Parasitology 1990 100 Pt 3 393 398 2163502
Cupolillo E Grimaldi GJ Momen H Genetic diversity among Leishmania (Viannia) parasites Ann Trop Med Parasitol 1997 91 617 626 9425364 10.1080/00034989760716
|
15794817
|
PMC1274274
|
CC BY
|
2021-01-04 16:28:15
|
no
|
BMC Infect Dis. 2005 Mar 28; 5:18
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-18
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-191579677910.1186/1471-2334-5-19Research ArticleBCG skin reaction in Mantoux-negative healthy children Singla Mohit [email protected] Vaibhav [email protected] Sukhbir [email protected] Rakesh Pal [email protected] Section of Immunobiology, Yale University School of Medicine, 300 Cedar Street, TAC S 630, New Haven, CT 06510, USA2 Medical Internist, Government Medical College, Amritsar, India3 Department of Pharmacology, Himachal Dental College, Sundernagar, India4 Department of Pediatrics, Singla Hospital, Hoshiarpur, India2005 29 3 2005 5 19 19 20 3 2004 29 3 2005 Copyright © 2005 Singla et al; licensee BioMed Central Ltd.2005Singla et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Tuberculosis poses a great challenge, especially in children. The response of BCG Test may be different in previously vaccinated children and needs to be considered before interpreting positivity for TB. This study has been carried out to determine the pattern of BCG reaction comparing previously vaccinated with non-vaccinated children.
Methods
The study was conducted in the healthy school children aged 4–6 years. The BCG skin reaction in Mantoux-negative children was compared between children with and without previous BCG scar. After the Mantoux and BCG Test, the analysis of variance was done as per protocol.
Results
Out of 50 children previously BCG vaccinated, 39(78%) showed exaggerated BCG test responses while out of another 50 children who were not vaccinated for TB, only 9(18%) showed exaggerated BCG Test response (p-value < 0.00001). Average induration obtained in children who were immunized with BCG at birth was much greater than those who were not immunized. 80% and 76% males and females respectively in Group I showed exaggerated BCG response while 16% and 20% males and females respectively of Group II showed exaggerated BCG response.
Conclusion
The present study indicates that normal healthy children may have a mild exaggerated BCG Test response i.e. induration up to 8 mm because of prior BCG vaccination. Therefore, BCG Test, though important should not be the only criteria for start of chemotherapy for TB in children as the side effects of drugs may cause much morbidity. An induration up to 8 mm after the BCG Test can be normal in Indian settings due to exposure to Mycobacterium in environment and/or BCG vaccine.
==== Body
Background
Tuberculosis remains a leading cause of death from any single infectious disease, accounting for over a quarter of all avoidable deaths. It infects about 1 billion people and causing an estimated 1–2 million deaths annually [1]. Nearly 3–4 million children in India have tuberculosis while 94 million are at risk to infection. 40% of these children by the age of 6 yrs and 80% by the age of 16 years are labeled as infected [2].
Bacillus Calmette-Guerin (BCG) vaccine, which consists of live attenuated Mycobacterium tubercle bovine Danish 1331 strain, has been extensively used as a protective measure against tuberculosis for the last half century.
Under Universal immunization Programme (UIP), primary BCG vaccination is given at birth or within the first month. The attenuated bovine non-pathogenic Mycobacterium induces tuberculin sensitivity and potentates the defense mechanism enabling the recipient to combat re-infection when exposed to pathological strains of mycobacterium later. This primary infection enables the vaccinated person to mobilize immune processes more rapidly when challenged by further natural infections.
A definite scar evidences an effective BCG vaccination and absence of this denotes that no immunity is attributable to the vaccination [14]. However, even with proper vaccination the immunity has been observed to decline within a few years, on the basis on which revaccination at five years has been suggested as an optional dose by the Indian Academy of Pediatrics [6]. Vijayalakshmi et al[11] suggested that revaccination may be contemplated after 8 years.
BCG given as a diagnostic test is based on Koch's phenomenon. When BCG is given to a child with tuberculosis, the reaction occurs at the site of vaccination within 48–72 hrs as against the usual late reaction i.e. after 3–6 weeks in child without tuberculosis. The reaction is measured with a standard plastic scale and the maximum size of the reaction is noted. The appearance of papule or induration more than 5 mm in size at the test site is considered as positive BCG Test [7-11].
BCG Test response is graded as mild (5–10 mm), moderate (10–20 mm) or severe (more than 20 mm) [6]. The reaction is complete if all the stages i.e. papule, pustule, ulcer and scab are seen, but is incomplete when only a papule or nodule is seen and the reaction does not progress to a stage of pustule or scab formation.
BCG response is more sensitive than the Mantoux test (Tuberculin test) and hence has been considered a better tool for epidemiological investigations. However, the effect of previous BCG vaccine on the subsequent revaccination is not well established and only a few elaborate studies have been carried out in this regard. It has been seen that the Mantoux test is not reliable as a post-vaccination check, because in vitro, cell-mediated immune responses may be demonstrable even when Mantoux test becomes negative [11,12].
The response of BCG may be different in previously vaccinated children and needs to be considered before interpreting positivity. The exact type of reaction and range of induration is not clear cut or well established.
There is possibility of over diagnosis of tuberculosis on the basis of BCG Test results, if previous BCG vaccination status is not taken into consideration. It is to be noted that the BCG Test may be positive in previously vaccinated children even when Mantoux test is negative.
This study has been carried out to determine the pattern of BCG reaction comparing previously vaccinated with non-vaccinated children.
Aims and objectives
1. To assess the BCG skin reaction in Mantoux-negative healthy children at 4–6 years of age previously vaccinated with BCG, as evidenced by the presence of scar [14]
2. To compare the reaction between children with and without previous BCG scar.
Methods
The study was conducted to determine the BCG response in 100 healthy, Mantoux-negative children in the age group of 4–6 years in relation to previous vaccination status. The informed consent of the parents was taken. Basic information, including the immunization status, was obtained on a written proforma from the parents of 193 children of 4–6 yrs age group belonging to kindergarten schools of urban and rural areas of Ludhiana. The immunization status of the children was confirmed by the presence of BCG scar[14]. Children with measles or other serious infections in previous 6 months were excluded. Also, only those with body weight above 80%of reference weight (50th percentile of the Harvard standard for the chronological age) were selected. Children presenting with clinical features of any disease, including tuberculosis, were excluded.
Mantoux test was given by injecting 0.1 ml (5 TU) of PPD intradermally on volar surface of left forearm. The test was read after 48–72 hrs and induration of 5 mm or less was taken as negative. Those with induration above 5 mm were excluded. The children were divided into two groups on the basis of previous BCG vaccination.
Group I: BCG given at birth.
Group II: BCG not given at birth.
The first 25 boys and 25 girls in Group I and Group II each who fulfilled the above criteria were included in the study. The children in both groups were given the BCG Test and the reaction was observed for erythema and induration after 1 day, 3 days, 7 days and 6 weeks as detailed below. Children in the Group I with induration 10 mm or more and in Group II with 6 mm or more were further investigated with x-ray chest to rule out any chance of tuberculosis and thus eliminate false positives in the study.
BCG test
Freshly diluted freeze-dried 3CC vaccine were mixed. Dry vaccine and solvent was carried from the refrigerator in an icebox to the place of vaccination. 0.1 ml of the solution containing 0.1 mg BCG vaccine was administered intradermally in the deltoid region using sterile disposable needles.
Observation of Reaction
The reaction was observed after 24 hrs, 72 hrs, one week and 6 weeks. The horizontal and vertical diameters of induration were measured. All the measurements were made by non-stretchable white fiberglass tape measuring to the nearest of 1 mm. The reactions in the two groups were recorded.
Results
Out of 50 children in Group I (children who were given prior BCG vaccination), 39 (i.e. 78 per cent) had exaggerated BCG Test responses while in Group II (children who were not given prior BCG vaccine) only 9 (18 per cent) out of 50 cases had exaggerated BCG Test response (Refer to Table 1). Critical difference at 5% of 0.64 and F-ratio of 60.4 indicates high statistically significant difference between the two groups.
Table 1 Comparison of Exaggerated BCG response in two groups
Number of cases Exaggerated response cases
Study Group Male Female Total Male Female Total
Total % Total % Total %
Group I 25 25 50 20 80 19 75 39 78
Group II 25 25 50 4 16 5 20 9 18
Total 50 50 100 24 48 24 48 48 48
When the exaggerated BCG Test response cases in Group I were further evaluated for size of the induration, it was seen that 72% cases had induration between 6 to 8 mm (Refer to Table 2). Only 2 cases had induration of greater or equal to 10 mm. These two cases were further investigated with x-ray chest to rule out the possibility of tuberculosis. Similarly in Group II, 12% had exaggerated BCG test response of 6 mm whereas rest 6% had 7 mm induration.
Table 2 Exaggerated BCG response in relation to scar size
Exaggerated response
Scar size Total cases Total %
No scar 50 9 18
Faint scar 27 21 77
Good scar 11 8 72
Large scar 12 10 83
Almost same number of males and females in both groups had exaggerated BCG response (Refer to Table 1). No statistically significant difference was observed between males and females in either of the two groups.
Highly significant statistical difference of exaggerated BCG response between Group I (BCG given at birth) and Group II (BCG not given at birth) were obtained.
To test the significance of this difference of the exaggerated response to BCG Test between Group I and Group II, a Standardized Normal Test for Proportions (Z-Test) was used. For the calculated value of Z = 7.518, the p-value is less than 0.00001; thus the exaggerated response of Group I was found to be highly significant in comparison to that of Group II.
Discussion
BCG has been extensively used as a protective measure against tuberculosis for the last half a century [5]. In the previous two decades encouraging results have been reported regarding BCG vaccination as a diagnostic test for tuberculosis and it has been termed BCG Test [3,4].
The effect of previous BCG on subsequent revaccination response is not well established and only a few elaborate studies have been carried out in this regard. It has been seen that Mantoux test is not reliable as a post-vaccination check; because in-vitro cell mediated immune response may be demonstrable even when Mantoux test becomes negative [11,12].
The response of BCG Test may be different in previously vaccinated children. It needs to be reconsidered before interpreting postivity. The present study was carried out to determine the pattern of BCG reaction comparing previously vaccinated with non-vaccinated children.
As already mentioned very few studies regarding BCG Test in prior BCG vaccinated children have been carried out. One such study is by P.M.Udani[7]: BCG Test in the diagnosis of tuberculosis in children – Indian Pediatric, (1982) addendum regarding "BCG Test in children who have received prior BCG is available." In this study he arrived at three sets of conclusions.
First
Negative BCG Test: the infant or child may not get any reaction because of BCG vaccination and body behaves as if no prior immunization was given. The prior BCG vaccine given at birth did not contribute to immune response of the child.
Second
Strongly positive BCG Test, a classical accelerated reaction. This reaction is similar to reaction seen in a patient having tuberculous disease.
Third
Mildly positive reaction. The child may get induration between 6–9 mm in this reaction. He attributed the third type of reaction because of prior BCG vaccination.
In the present study, most (about 84%) of the exaggerated BCG Test response in Group I had reaction between 6–8 mm similar to the Third type of reaction evidenced by Udani[7]. In the study conducted by Udani[7], no inclusion and exclusion criteria for study have been mentioned, and probably all children who were given prior BCG vaccination were included in the study. As mentioned earlier the present study compares the results of the BCG Test in BCG vaccinated children with BCG non-vaccinated children.
Although the 50 boys and 50 girls who were included in this study were not exactly selected on the basis of formal random sampling procedures, there is little possibility of bias. Even though there was no entry criterion, except for the body weight above 80% and exclusion of any case with any disease process, we understand that the absence of formal random allocation is a limitation to the study. However, the highly significant results (p-value < 0.00001) considerably overshadow the limitation.
Diagnosis of TB in children is not easy. Like in adults one should not always think of Therapeutic Trial in children as the drugs are very toxic and the side effects like Optic Neuritis cannot be easily diagnosed in children. So Diagnosing and treating TB in Children is a double-edged sword. These days PCR and other new investigations have become useful, but they just add to the old regime of investigation and are not fully predictive of disease on their own. So coming to a diagnosis of TB in children need a lot of clinical skills and interpreting the investigations in the right way.
There is no doubt that in some children it may be difficult to decide whether BCG Test is positive because of prior BCG vaccination or due to natural tubercular infection. Over diagnosis of tuberculosis is usually made by using exaggerated BCG Test response as a parameter to decide anti-tubercular therapy in patients in whom otherwise history, clinical features, tuberculin test sensitivity, x-ray chest, etc. are not fully suggestive of the Koch's disease. The present study indicates that normal healthy children may have a mild exaggerated BCG Test response i.e. induration up to 8 mm because of prior BCG vaccination. Therefore, the patients who were immunized with BCG at birth and tested positive for BCG Test, and in whom other parameters for diagnosis of tuberculosis are not suggestive, are perhaps receiving unnecessary anti-tubercular therapy. A positive test may be of value after correlating it with the age of the child, nutrition, size of tuberculin test induration, history of contact and symptoms to diagnose and institute chemotherapy [1].
Conclusion
78% (39 out of 50) of children in Group I (prior BCG vaccination given) and 18% (9 out of 50) in Group II (prior BCG vaccination not given) had exaggerated BCG response; defined as induration greater than or equal to 6 mm. Statistical difference (p-value<0.00001) between Group I and Group II was highly significant. 72% of children in Group I, who showed exaggerated BCG response, had induration in range of 6–8 mm. Only 6% of children showed induration of 9, 10 or more than 10 mm. 22% of children in Group I had induration less than 6 mm. Children in Group II showing exaggerated BCG response had induration of either 6 or 7 mm. Differences between male and female readings in each of two groups were statistically insignificant. Day 1 and 3 readings were statistically more significant as compared to day 7 and 6-week readings. All children in Group II showing exaggerated BCG response and in Group I showing induration of 10 mm or more were investigated with x-ray chest to rule out tuberculosis. Even in absence of prior BCG vaccination, some normal healthy children of Group II exhibited mildly exaggerated BCG response. This can probably be explained on the basis of tubercular infection not warranting treatment. Significant number (78%) of normal children with previous BCG vaccination, who were Mantoux negative and received BCG as a 'diagnostic test' or as revaccination, showed, exaggerated BCG response. Since these children are healthy, an exaggerated BCG response in such a situation must not be construed as an evidence of tubercular disease.
Abbreviations
BCG – Bacillus Calmette and Guerrin
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
MS carried out designed the study, conducted all the fieldwork and drafted the manuscript. VS has participated in the statistical analysis and editing of the script. SSS helped in co-ordination and draft of the manuscript. RPG conceived the study and provided guidance in carrying out the work.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
We thank Dr. Paula Kavathas from Yale University School of Medicine, who gave us the precious guidance in the review of the manuscript.
==== Refs
International Union against Tuberculosis World Health Organization Study Group Tuberculosis Control Report of a Joint IUAT-WHO Study Group Geneva, WHO 1982
Udani PM Tuberculosis in children in India – A major health hazard Pediatr Clin India 1983 18 14 42
Fridman E Silverman F Use of BCG vaccine as a new diagnostic test for tuberculosis Pediatr 1952 9 280 285
Frappier A Guy R A new practical BCG skin test for detection of total tuberculin allergy Can J Public Health 1958 41 72 83 15406448
WHO Expanded Programme on Immunization Report and working paper 31st Session of WHO Reg Committee Mongolia SEARO 1978 21 28
Park JE Park K Textbook of Preventive and Social Medicine 10 290 292
Udani PM BCG test The diagnosis of tuberculosis in children Indian Pediatr 1982 19 563 581 7174084
Udani PM Protective efficacy of BCG vaccine Indian Pediatr 1982 19 739 752 6984429
Choudhari VP Sinah MM Verma IC BCG and Mantoux intradermal testsin diagnosis of tuberculosis Indian Pediatr 1974 11 535 538 4548414
Bhandari B Mandowara SL Chliaparwal Rajesh BCG Dilemma, Indian Pediatr 1982 19 165 173 7118240
Vijayalakshrni NV Rao DV Murthy KJR Jam SN A study of the tuberculin test and its correlation with in vitro responses Lung India 1989 7 63 66
Seth V Malaviya AN Sahai V Arora N Sundaram KR Cell mediatedImmuno response in childhood, Tuberculosis Indian J Med Res 1981 73 68 73 6972352
Udani PM Scoring technique in the diagnosis of tuberculosis in children (score to settle with a killer) Souvenir XVIII National Conference of Indian Academy of Paediatrics 1980 30 37
Pereira SM Bierrenbach AL Dourado I Barreto ML Ichihara MY Hijjar MA Rodrigues LC Sensitivity and specificity of the BCG scar reading Rev Saude Publica 2003 37 254 9 [Article in Portuguese] 12700850
Immunization schedule in India Need for a Change Indian Academy of Pediatrics 1992 14 137 140
|
15796779
|
PMC1274275
|
CC BY
|
2021-01-04 16:28:16
|
no
|
BMC Infect Dis. 2005 Mar 29; 5:19
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-19
|
oa_comm
|
==== Front
BMC Infect DisBMC Infectious Diseases1471-2334BioMed Central London 1471-2334-5-211580790310.1186/1471-2334-5-21Research ArticleBartonella seropositivity in children with Henoch-Schonlein purpura Robinson Joan L [email protected] Donald W [email protected] Errol [email protected] Dorothy [email protected] Harvey [email protected] Department of Pediatrics and Stollery Children's Hospital, University of Alberta, Edmonton AB, Canada2 Dynacare Kasper Medical Laboratories, Edmonton, AB, Canada3 Bacterial and Rickettsial Zoonosis Section, National Microbiology Laboratory, Canadian Science Centre for Human and Animal Health, Winnipeg, MB, Canada2005 5 4 2005 5 21 21 18 10 2004 5 4 2005 Copyright © 2005 Robinson et al; licensee BioMed Central Ltd.2005Robinson et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
An association between Henoch-Schonlein purpura (HSP) and seropositivity for Bartonella henselae (BH) has been described. The objective of this study was to see if such an association exists in northern Alberta.
Methods
Immunofluorescent antibody testing utilizing an antigen prepared from B. henselae was undertaken on sera from six children with current HSP, 22 children with remote HSP, and 28 controls that were matched for age. Blood from the six children with current HSP was analysed by polymerase chain reaction (PCR) assay with primers derived from the citrate synthase (gltA) gene for the detection of Bartonella DNA.
Results
The seropositivity rate for BH was 61% in cases versus 21% in controls (p < 0.03). The PCR assay was negative in all six current cases.
Conclusion
There is an increased seropositivity rate for BH in children with HSP. However, it is not clear if infection with B. henselae or a related Bartonella species can result in HSP, or if the increased seropositivity is from non-specific or cross-reacting antibodies.
==== Body
Background
Henoch Schonlein purpura (HSP) is an idiopathic form of vasculitis, which manifests as a characteristic painless palpable purpuric rash most pronounced on the buttocks and the extensor surfaces of the lower extremities. The vasculitis can also involve the bowel, resulting in abdominal pain. In severe cases, there can be melena, malabsorption, pancreatitis or intussussception [1]. Joint involvement occurs in the majority of cases. Renal involvement occurs in about half of cases, and usually results in a reversible, asymptomatic IgA-mediated nephritis, but about 1% of patients progress to chronic renal failure [1]. Impressive testicular swelling can occur. About 10–20% of patients have recurrences of HSP – typically within a few weeks of the disease appearing to resolve. Evidence of recent infection with group A streptococcus, Epstein-Barr virus (EBV), varicella, parvovirus B19, Campylobacter, or Mycoplasma have all been found in patients with HSP [2,3], but these organisms do not appear to be etiologic agents.
Bartonella henselae is a fastidious gram-negative organism, and is the etiologic agent for cat-scratch disease (CSD) [4]. Less commonly, infection with this organism results in encephalitis, splenic or hepatic abscesses, or osteomyelitis [4]. The organism is presumed to be carried by fleas, which then transmit it to cats, resulting in feline bacteremia. A cat bite or scratch then transmits the organism to humans. A 2002 study from Florida demonstrated that 67% of patients with a recent diagnosis of HSP had serologic evidence of infection with B. henselae (versus 14% of a control group) [5]. It is uncertain if this means that B. henselae causes HSP or if there is a non-etiologic association between HSP and B. henselae.
The objective of this study was to determine if children in northern Alberta with a current or remote diagnosis of HSP have evidence of infection with B. henselae or a related Bartonella species using both serology and nucleic acid amplification.
Methods
Study population
This study was approved by the Health Ethics Review Board of the University of Alberta. Pediatricians were asked to notify us of children with a current or remote diagnosis of HSP, and health records from the Stollery Children's Hospital for 1997–2001 were searched to identify children with this diagnosis. After informed consent was obtained, data were collected from the parents, the patient, and the medical record on the symptoms the child had at the time of diagnosis, the number of recurrences that had occurred to date, the level of exposure to cats, and the results of any biopsies that were done. The diagnosis of HSP was based on either i) the presence of a classic rash with palpable purpuric lesions mainly on lower limbs and buttocks, or ii) an atypical rash and either abdominal pain, joint pain, lower gastrointestinal bleeding, or laboratory evidence of nephritis. Patients were considered to have current HSP if onset of initial or recurrent symptoms was less than 42 days prior to enrollment, recent HSP if symptoms started 42 or more days prior to enrollment but had not yet resolved, and remote HSP if symptoms started 42 or more days prior to enrollment and had resolved.
Paired sera were collected for B. henselae serology from test subjects, with the convalescent sera being collected approximately two weeks after the acute sera. Blood was drawn for amplification of Bartonella-specific genomic sequences by PCR assay from patients that were considered to have current HSP. Bartonella henselae serology was also run on controls that had been matched for age (< 3 yr, 4–7 yr, 8–12 yr, or > 12 yr). Control sera were originally collected for other diagnostic purposes, and no clinical information was available on these children. The technicians were blinded as to the source of the specimens (cases versus controls) and all specimens were run in a single batch.
Sample size
The assumption was made that if B. henselae infection were the sole causative organism of HSP, patients with a current or remote diagnosis of HSP would be sero-positive half the time, as waning antibody titers occur [6]. Assuming that the seropositivity rate in the control group could be as high as 15%, and that the diagnosis of HSP is accurate, enrolling 20 patients would yield the power to demonstrate that over half of cases are seropositive at a confidence interval of 95%.
Serology
Analysis of sera for immunoglobulin (Ig) G antibodies to B. henselae antigen was performed using an indirect immunofluorescence assay (IFA) method with positive and negative controls. A suspension of B. henselae (ATCC 49882) in 0.1% formal saline was prepared with bacteria grown in-house on brain heart infusion agar supplemented with 5% sheep blood. The suspension was spotted onto 12-well slides (#ER-202W, Erie Scientific, USA) and air-dried for 1 hour. Slides were fixed in cold acetone for 15 minutes, air-dried, and stored at -80 degrees Celsius. For initial screening, sera from controls and test subjects were diluted 1:32 in FTA Hemagglutination Buffer (#211248, Becton Dickinson, USA). A 1:32 dilution of goat anti-human IgG (whole molecule) FITC conjugated antiserum was used to detect IgG antibodies. Sera reactive at 1:32 were serially titrated two-fold to endpoint. A titer of 1:64 or higher was interpreted as evidence for infection at an undetermined time [6]. A titer of 1:256 or higher was interpreted as evidence for recent infection.
Polymerase chain reaction
DNA was extracted from six blood clots using a QIAamp DNA Mini Kit (Qiagen, Chatsworth, CA). Polymerase chain reaction amplification to detect Bartonella-specific sequences of the citrate synthase gene (gltA) [7] was attempted and found to be negative. Simultaneous amplification with porphobilinogen deaminase gene primers [8] verified that the extracted DNA was of sufficient quality for amplification.
Statistics
The proportion of seropositivity to B. henselae was compared in the cases and controls using a chi-square test. A comparison was done of seropositivity in current cases versus remote cases, and in cases with initial onset of symptoms ≤ 12 months before the serology was drawn, versus those who had onset of symptoms > 12 months earlier. A p value of <0.05 was considered to be significant.
Results
Serology
There were 28 patients (21 males, 7 females) enrolled in the study. In addition to meeting the clinical criteria for HSP, two of these patients had HSP nephritis on renal biopsy, and two others had leukocytoclastic vasculitis on skin biopsy. Six patients had current HSP (five were initial episodes, and one was a recurrence), none had recent HSP, and 22 had remote disease. Seventeen of the 28 cases had contact with cats (nine had contact but no bite or scratch, five had a bite or scratch from an adult cat, and three others had a bite or scratch from a kitten). Results of serology are shown in Table 1. Sixty-one percent of cases were seroreactive to B. henselae antigen versus 21% of controls (p < 0.003). Paired sera were drawn 17 and 18 days apart from two patients with current HSP, and had static titers of 1:32 and 1:128, respectively. Figure 1 shows that B. henselae titers did not appear to be related to the time that had elapsed since the diagnosis of HSP and Table 2 shows that seropositivity did not seem to be related to the season of the year in which the initial episode of HSP occurred.
Table 1 Results of serology for B. henselae in patients with Henoch-Schonlein purpura, and in a control group
n Titer < 1:64 Titer 1:64 – 1: 128 Titer ≥ 1: 256 Total seropositives
Current HSP 6 3 (50%) 2 (33%) 1 (17%) 3 (50%)
Remote HSP 22 8 (36%) 10 (45%) 4 (18%) 14 (63%)
HSP ≤ 12 months earlier 11 6 (54%) 2 (18%) 3 (27%) 5 (45%)
HSP > 12 months earlier 17 5 (29%) 10 (59%) 2 (12%) 12 (71%)
All HSP 28 11 (39%) 12 (43%) 5 (18%) 17 (61%)
Controls 28 22 (79%) 6 (21%) 0 6 (21%)
Figure 1 Relationship between results of serology for B. henselae and time since onset of Henoch-Schonlein purpura
Table 2 Seasonality of seropositivity for Bartonella henselae in patients with Henoch-Schonlein purpura
Month of initial onset of Henoch-Schonlein purpura Seropositives/Total number tested
January-March 3/8
April-June 1/1
July-September 6/8
October-December 7/11
Polymerase chain reaction
From each of the six patients diagnosed with current HSP, a single blood sample was collected on one of the following days post onset of symptoms: 7, 18, 19, 20, 24 and 34. All blood samples were negative by PCR assay for Bartonella-specific sequences of the citrate synthase gene.
Discussion
A possible association between HSP and B. henselae infection has been noted in one previous study, with an increased seroprevalence in children with recent onset of HSP [5]. In this confirmatory study, there was a significantly increased seroreactivity rate to B. henselae antigen in 28 children with a current or remote history of HSP than in controls (p < 0.03). This is the first study to use PCR to attempt to identify B. henselae in children with HSP, but the organism was not detected in any of the six cases tested. In the absence of demonstration of genomic sequences specific to B. henselae, it is not clear if B. henselae causes HSP. Many studies have demonstrated that serological differentiation between B. henselae and B. quintana infections is impossible, with a cross-reactivity rate between these species of up to 95% [6]. It is possible that the antibodies to B. henselae in HSP patients observed here are actually cross-reactive to B. quintana or to other bacteria, or are a non-specific reaction to inflammation.
If B. henselae or a related Bartonella species is the sole causative agent of HSP, one might expect an even higher seroprevalence than the 61% in this study, and 67% in the previous study [5]. Perhaps HSP is a "final common pathway" of infection with multiple organisms including Bartonella. Another explanation would be that for the patients with current HSP, serology was run too early in the course of their disease. A study of CSD found that titers peaked between 2 and 16 weeks after onset of symptoms [9]. Serology was drawn 7 to 39 days after the onset of symptoms of HSP in the six patients with current HSP in this study, and mainly within 28 days in the previous study [5]. Obtaining more follow-up titers in patients with HSP might increase the seropositivity rate or demonstrate a rise in titers. Measuring IgM titers to B. henselae in patients with acute HSP could also be useful, but potential problems include the lack of sensitivity of the IgM assay, and potential cross-reactivity with IgM to EBV [6].
Another possibility is that the clinical manifestations of HSP constitute an immunologic reaction to an infection that has resolved, such that the titers are already waning when the clinical features of HSP become evident. Therefore, it is possible that some seronegative patients with remote HSP had seroreverted by the time serology was drawn. In the initial study of B. henselae serology in CSD, a rapid decline in titers as measured by IFA was noted [6]. A recent study that used an enzyme immunoassay found only 25% of patients maintained IgG to B. henselae for more than 12 months [10]. However, in the current study, there was no apparent relationship between the time since diagnosis of HSP and the B. henselae titers. Other reasons for false-negative serologic results could be that there are differences in serologic response to different strains of B. henselae [6], or that the clinical diagnosis of HSP was incorrect in some cases.
The B. henselae seroprevalence rate in controls of 21% and 14% in the current study and the Florida study [5] respectively are higher than in a summary of North American studies where the rate ranged from 2% [11] to 6% [6]. However, the seroprevalence was 37% in adult blood donors and 18% in children with respiratory illnesses in a study done in British Columbia, Canada [12]. It is well recognized that interpretation of the IFA is subjective, and it is probably not valid to compare titers obtained in different laboratories. Because there have been no previous seroprevalence studies in Alberta, it is not clear if our seroprevalence rate is higher than predicted or if the specificity of the IFA is low. Evidence for the latter is that during the finalization of this manuscript, case and control specimens after being stored for 2 years were re-analyzed by an in-house IFA assay in a different laboratory, and all had titres of <1:50. However, a previous study demonstrated that in-house assay to be less sensitive than a commercial one [13]. A lack of specificity of our IFA would not account for the higher seroprevalence rate in cases than in controls.
Bartonella was not detected by PCR in the blood of six acute cases of HSP in the current study, but it is possible that the PCR assay that was used is not sufficiently sensitive, or it may have detected the organisms if done earlier in the course of HSP. In a study of cat-scratch disease, the PCR was positive on the lymph nodes in 10/21 cases, with 9/10 specimens obtained in the first 6 weeks of illness being positive but only 1/11 obtained after 6 weeks of illness being positive [9]. Another explanation for the failure to detect Bartonella in the blood could be that the organism is present in the liver, spleen, or lymph nodes in HSP rather than in the blood. If the clinical manifestations of HSP constitute an immunologic reaction to an infection that has resolved, it may be too late to detect the organism by PCR by the time the diagnosis of HSP is evident.
Biopsy of the rash of HSP shows leukoclastic vasculitis. This is a hypersensitivity reaction that can be caused by a wide variety of infections, drugs, insect bites, cold exposure, and malignancies [14]. Although leukoclastic vasculitis has been described in two cases of CSD [5,14], one might expect many more cases if CSD and HSP are caused by the same organism. If B. henselae is the cause of HSP, it is a bit surprising that 39 % of our cases had no history of contact with cats. However, in a previous study of serologically proven CSD, only 57% of patients could recall contact with cats [9]. It is possible that humans can acquire infection directly from fleas [4]. In addition, there is some evidence that dogs can transmit B. henselae [15].
In summary, patients with a current or remote diagnosis of HSP have increased seropositivity to B. henselae. However, no association was found between the antibody titre and the time since the onset of HSP, and blood samples in six patients with acute HSP were negative by PCR assay for Bartonella-specific sequences. To prove if B. henselae is related to HSP, future studies should perform more extensive serologic follow-up, confirm serologic results with a commercial assay, and perform PCR and cultures on blood and all available tissues as early in the course of the disease as possible. Furthermore, B. henselae serology should be done in children with other inflammatory conditions to determine if the seroreactivity is a non-specific inflammatory response.
Abbreviations
BH – Bartonella henselae
CSD – cat-scratch disease
EBV – Epstein Barr virus
HSP – Henoch-Schonlein Purpura
IFA – immunofluorescence assay
PCR – polymerase chain reaction
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
JR wrote the protocol and the manuscript. DWS did the statistical analysis and reviewed the manuscript. EP, DM, and HA ran the laboratory tests and reviewed the manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
The authors would like to thank Bonita E. Lee MD for assistance with the figure.
==== Refs
Tizard EJ Henoch-Schonlein Purpura Arch Dis Child 1999 80 380 383 10086951
Szer IS Henoch Schonlein purpura: when and how to treat J Rheumatol 1996 23 1661 1665 8877944
Kraft DM McKee D Scott C Henoch Schonlein purpura: a review Am Fam Physician 1998 58 405 408 9713395
Bass JW Vincent JM Person DA The expanding spectrum of Bartonella infections: II. Cat-scratch disease Pediatr Infect Dis J 1997 16 163 179 9041596 10.1097/00006454-199707000-00013
Ayoub EM McBride J Schmiedereer J Anderson B Role of Bartonella henselae in the etiology of Henoch-Schonlein purpura Pediatr Infect Dis J 2002 21 28 31 11791094
Sander A Berner R Ruess M Serodiagnosis of cat scratch disease: response to Bartonella henselae in children and a review of diagnostic methods Eur J Clin Microbiol Infect Dis 2001 20 392 401 11476439 10.1007/s100960100520
Norman AF Regnery R Jameson P Greene C Krause DC Differentiation of Bartonella-like isolates at the species level by PCR-restriction fragment length polymorphism in the citrate synthase gene J Clin Micro 1995 33 1797 1803
Shimizu H Shimizu C Bruns JC Detection of novel RNA viruses: morbilliviruses as a model system Mol Cell Probes 1994 8 209 214 7969194 10.1006/mcpr.1994.1029
Ridder GJ Boedeker CC Technau-Ihling K Grunow R Sander A Role of cat-scratch disease in lymphadenopathy in the head and neck Clin Infect Dis 2002 35 643 649 12203159 10.1086/342058
Metzkor-Cotter E Kletter Y Avidor B Varon M Golan Y Ephros M Giladi M Long-term serological analysis and clinical follow-up of patients with cat scratch disease Clin Infect Dis 2003 37 1149 54 14557957 10.1086/378738
Demers DM Bass JW Vincent JM Person DA Noyes DK Staege CM Samlaska CP Lockwood NH Regnery RL Anderson BE Cat-scratch disease in Hawaii: etiology and seroepidemiology J Pediatr 1995 127 23 26 7608806
Cimolai N Benoit L Hill A Lyons C Bartonella henselae infection in British Columbia: evidence for an endemic disease among humans Can J Microbiol 2000 46 908 912 11068677 10.1139/cjm-46-10-908
Maurin M Rolain JM Raoult Comparison of in-house and commercial slides for detection by immunofluoresence of immunoglobulins G and M against Bartonella henselae and Bartonella Quintana Clin Diagn Lab Immunol 2002 9 1004 9 12204950 10.1128/CDLI.9.5.1004-1009.2002
Hashkes PJ Trabulsi A Passo MH Systemic cat-scratch disease presenting as leukoclastic vasculitis Pediatr Infect Dis J 1996 15 93 95 8684888 10.1097/00006454-199601000-00023
Gundi VA Bourry O Davoust B Raoult D La Scola B Bartonella clarridgeaiae and B. henselae in dogs, Gabon Emerg Infect Dis 2004 10 2261 2 15672535
|
15807903
|
PMC1274276
|
CC BY
|
2021-01-04 16:28:16
|
no
|
BMC Infect Dis. 2005 Apr 5; 5:21
|
utf-8
|
BMC Infect Dis
| 2,005 |
10.1186/1471-2334-5-21
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 1625067010.1371/journal.pmed.0020333Research ArticleGenetics/Genomics/Gene TherapyOpthalmologyOphthalmologyGene TherapyPharmacological and rAAV Gene Therapy Rescue of Visual Functions in a Blind Mouse Model of Leber Congenital Amaurosis Recovery of Vision in Blind MiceBatten Matthew L
1
Imanishi Yoshikazu
1
2
Tu Daniel C
3
Doan Thuy
4
Zhu Li
1
5
Pang Jijing
6
Glushakova Lyudmila
6
Moise Alexander R
1
2
Baehr Wolfgang
7
8
9
Van Gelder Russell N.
3
10
11
Hauswirth William W
6
Rieke Fred
4
Palczewski Krzysztof
1
2
5
12
*1Department of Ophthalmology, University of Washington, Seattle, Washington, United States of America,2Department of Pharmacology, Case School of Medicine, Case Western Reserve University, Cleveland, Ohio, United States of America,3Department of Ophthalmology and Visual Sciences, Washington University School of Medicine, St. Louis, Missouri, United States of America,4Department of Physiology and Biophysics, University of Washington, Seattle, Washington, United States of America,5Department of Chemistry, University of Washington, Seattle, Washington, United States of America,6Department of Ophthalmology, and Powell Gene Therapy Center, University of Florida, Gainesville, Florida, United States of America,7Department of Ophthalmology, University of Utah, Salt Lake City, Utah, United States of America,8Department of Biology, University of Utah, Salt Lake City, Utah, United States of America,9Department of Neurobiology and Anatomy, University of Utah, Salt Lake City, Utah, United States of America,10Department of Molecular Biology, Washington University School of Medicine, St. Louis, Missouri, United States of America,11Department of Pharmacology, Washington University School of Medicine, St. Louis, Missouri, United States of America,12Department of Pharmacology, University of Washington, Seattle, Washington, United States of AmericaLightman Susan Academic EditorMoorfields Eye HospitalUnited Kingdom*To whom correspondence should be addressed. E-mail: [email protected]
Competing Interests: WWH and the University of Florida own equity in a company, Applied Genetic Technologies Corporation, that may commercialize some of the technology described in this work. KP owns shares of the company Retinagenix. The University of Washington and Retinagenix may commercialize some of the technology described in this work.
Author Contributions: MLB, YI, DCT, TD, LZ, JP, LG, ARM, WB, RNVG, WWH, FR, and KP contributed to conception and design, or acquisition of data, or analysis and interpretation of data. MLB, YI, DCT, TD, LZ, ARM, WB, RNVG, WWH, FR, and KP contributed to writing the paper.
11 2005 1 11 2005 2 11 e33319 5 2005 12 8 2005 Copyright: © 2005 Batten et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.
Restoring Retinal Function in a Mouse Model of Hereditary Blindness
Tackling Inherited Blindness
Background
Leber congenital amaurosis (LCA), a heterogeneous early-onset retinal dystrophy, accounts for ~15% of inherited congenital blindness. One cause of LCA is loss of the enzyme lecithin:retinol acyl transferase (LRAT), which is required for regeneration of the visual photopigment in the retina.
Methods and Findings
An animal model of LCA, the Lrat
−/− mouse, recapitulates clinical features of the human disease. Here, we report that two interventions—intraocular gene therapy and oral pharmacologic treatment with novel retinoid compounds—each restore retinal function to Lrat
−/− mice. Gene therapy using intraocular injection of recombinant adeno-associated virus carrying the Lrat gene successfully restored electroretinographic responses to ~50% of wild-type levels (p < 0.05 versus wild-type and knockout controls), and pupillary light responses (PLRs) of Lrat
−/− mice increased ~2.5 log units (p < 0.05). Pharmacological intervention with orally administered pro-drugs 9-cis-retinyl acetate and 9-cis-retinyl succinate (which chemically bypass the LRAT-catalyzed step in chromophore regeneration) also caused long-lasting restoration of retinal function in LRAT-deficient mice and increased ERG response from ~5% of wild-type levels in Lrat
−/− mice to ~50% of wild-type levels in treated Lrat
−/− mice (p < 0.05 versus wild-type and knockout controls). The interventions produced markedly increased levels of visual pigment from undetectable levels to 600 pmoles per eye in retinoid treated mice, and ~1,000-fold improvements in PLR and electroretinogram sensitivity. The techniques were complementary when combined.
Conclusion
Intraocular gene therapy and pharmacologic bypass provide highly effective and complementary means for restoring retinal function in this animal model of human hereditary blindness. These complementary methods offer hope of developing treatment to restore vision in humans with certain forms of hereditary congenital blindness.
A combination of intraocular gene therapy and pharmacologic bypass provides a complementary way to restore retinal function in an animal model of human hereditary blindness.
==== Body
Introduction
Development of successful treatments for inherited and acquired retinal disease caused by gene mutations represents a major challenge [1]. Recessive congenital defects arising from gene inactivation and subsequent disruption of a metabolic pathway are particularly amenable to pharmacological treatment or somatic gene therapy. Oral administration of appropriate compounds can correct visual deficits in humans and other animals by bypassing a block in the retinoid cycle [2]. Sorsby's fundus dystrophy and Leber congenital amaurosis (LCA) are two examples. Sorsby's fundus dystrophy, an autosomal-dominant retinal degeneration caused by mutations in the tissue inhibitor of the metalloproteinases-3 gene, leads to night blindness [1]. Vitamin A (retinol [ROL]) administered orally has been shown to significantly restore photoreceptor function in affected individuals [3]. LCA is an early-onset recessive human retinal degeneration that can be caused by mutations in the gene encoding retinal pigment epithelium 65 (RPE65), a key protein involved in the production and recycling of 11-cis-retinal (11-cis-RAL) in the eye. Approximately 15% of patients with LCA have been found to have mutations in Rpe65 [4,5]. Humans with this form of LCA and Rpe65
−/− mice both have severely impaired rod and cone function [6]. The biochemical block caused by the absence of Rpe65
−/− can be bypassed with synthetic cis-retinoids administered orally, resulting in a dramatic improvement in photoreceptor physiology [7].
Somatic gene therapy has also been very successful in many animal models of retinal degeneration [8]. Most notably, a canine model with a naturally occurring Rpe65 deficiency, the Rpe65
−/− dog, bears a phenotype similar to that of human LCA patients and Rpe65
−/− mice. A recombinant adeno-associated virus (rAAV) carrying wild-type (WT) Rpe65 (rAAV-Rpe65) restored visual function in this model of childhood blindness [9].
The lecithin:retinol acyl transferase (LRAT)-deficient mouse is an animal model of LCA [10]. Mutations in the gene encoding LRAT are associated with early-onset severe retinal dystrophy, or LCA [11]. The prevalence of mutations at this locus among patients with LCA is unknown but it is likely uncommon, as three of 267 probands with severe early-onset retinal dystrophy carried mutations in Lrat [11]. LRAT is a key enzyme involved in storage of ROL in the form of retinyl esters (REs) in specific structures known as retinosomes [12]. Without LRAT, no 11-cis-RAL chromophore is produced, and visual function is severely impaired [10].
Here, we aimed to investigate whether we could rescue visual function by pharmacological intervention and gene transfer therapy in LRAT-deficient mice, and to assess the advantages and disadvantages of oral administration of retinoids and gene therapy. Comparison of these two approaches is an important prelude for treatment in humans.
Methods
Animals
All animal experiments employed procedures approved by the University of Washington and conformed to recommendations of the American Veterinary Medical Association Panel on Euthanasia and recommendations of the Association of Research for Vision and Ophthalmology. Animals were maintained in complete darkness, and all manipulations were done under dim red light employing a Kodak No. 1 safelight filter (transmittance > 560 nm). Lrat
−/− mice were generated and genotyped as described previously [10]. Typically, 6- to 12-wk-old mice were used in all experiments. In the case of rAAV-Lrat treatment, 2- to 3-wk-old mice were used.
rAAV1-VMD2-mLrat Vector
The pTR-UF5 backbone [13] was used for generation of a pTR-VMD2-mLRAT plasmid construct. An EcoR1 fragment containing the full-length mouse Lrat cDNA was excised from pCR-TOPO II Blunt-Lrat and blunt ligated into the Not1 site of the UF5 cassette, replacing the Gfp gene. Orientation of the cDNA was confirmed by restriction analysis and by sequencing. The Kpn1-Xba1 fragment of the placF-VMD2 plasmid (from D. Zack, Johns Hopkins University) [14] includes the −585/+38 upstream region of the human VMD2 gene (Chromosome 11q13) and was subcloned into the Kpn1-Xba1 sites of the pTR-UF5 cassette (replacing the CMV promoter) upstream of mouse Lrat cDNA. Sequence analysis confirmed the orientation and reading frame of the Lrat cDNA. A serotype 1 AAV vector was produced in the presence of a mini-Ad helper plasmid pDG38 by double transfection of HEK293 cells, followed first by purification over an iodixanol gradient and then by high-Q FPLC column chromatography (Pharmacia, Uppsala, Sweden). Vector particle titers were determined by quantitative PCR. The rAAV1-VMD2-mLrat vector was prepared at 4 × 1013 physical particles/ml. Exclusive RPE expression is seen in mice using a Gfp reporter gene in an analogous rAAV1-VMD2 vector (Glushakova and Hauswirth, unpublished data).
Preparation of Retinoids and Oral Gavage
All-trans-retinyl acetate (R-Ac), all-trans-R-Palm, all-trans-RAL, all-trans-ROL, and 9-cis-RAL were purchased from Sigma-Aldrich (St. Louis, Missouri, United States). 9-cis-ROL, 9-cis-R-Palm, 9-cis-R-Ac, and 9-cis-retinyl succinate (9-cis-R-Su) were prepared from 9-cis-RAL. To prepare 9-cis-R-Ac, 100 mg of 9-cis-RAL was reduced with 50 mg of sodium borohydride in 0.7 ml of ethanol at 0 °C for 30 min, and 9-cis-ROL was purified by organic extraction and dried under argon. Solid 9-cis-ROL and 80 mg of 4-dimethylaminopyridine were dissolved in 0.4 ml of dry CH2Cl2, and 0.1 ml of acetic acid anhydride was added. After 6 h at 10 °C, the reaction was quenched with 0.1 ml of ethanol, CH2Cl2 was removed by flowing argon at 20 °C, and 9-cis-R-Ac was purified by organic extraction and dried under argon. To prepare all-trans-R-Su or 9-cis-R-Su, solid all-trans-ROL or 9-cis-ROL was dissolved in 0.2 ml of pyridine, with 100 mg of succinic acid anhydride added and reacted overnight at 10 °C. CH2Cl2 was removed by flowing argon at 20 °C, and all-trans-R-Su or 9-cis-R-Su was then purified by organic extraction and dried down under argon. Retinoids were dissolved in pure canola oil (Western Family Foods, Tigard, Oregon, United States), and concentrations were measured spectrophotometrically. Retinoids at a final concentration of 40 mg/ml in canola oil were administered to Lrat
−/− mice using a 1-ml syringe and a 20-gauge, 3.5-cm long gavage needle. Mice were allowed to rest for 3 d following gavage. A single 5-mg dose (125 μl) of retinoid was used to compare the different retinoids. Doses of gavaged 9-cis-R-Ac were 1, 2, 5, 10, 20, and 40 μmol to determine dose effect. Multiple gavages of 9-cis-R-Ac were performed with doses of 1, 5, and 10 μmol for up to ten consecutive treatments.
Photobleaching of 9-cis-R-Ac Gavaged Mice
Lrat
−/− mice were gavaged four times with 10 μmol doses of 9-cis-R-Ac. Caged mice were placed on a bench under fluorescent light of average luminosity 600 cd × m−2 and allowed to photobleach for 1, 3, and 10 d. Mice were then placed in the dark for 1 d, sacrificed, and the eyes collected for retinoid analysis. A subset of mice were gavaged with 10 μmol 9-cis-R-Ac following photobleach, placed in the dark for 1 d, sacrificed, and the eyes collected for retinoid analysis.
Pupillary Light Responses
Pupillary light responses (PLRs) were recorded from dark-adapted mice under infrared conditions using a CCD video camera fitted with close-up lens and an IR filter. Data analysis was performed by video pupillometry. Light stimuli were provided by a halogen source; wavelength and intensity were manipulated with neutral density and narrow bandwidth (10 nm) interference filters (Oriel, Stratford, Connecticut, United States). Irradiance measurements (W/m2) were made using a radiometer (Advanced Photonics International, White Plains, New York, United States).
Analyses of Retinoids and Visual Pigments
All procedures were performed under dim red light as described previously [7,15,16]. Retinoid analysis was performed on an Agilent 1100 series HPLC equipped with a diode array detector and Agilent Chemstation A.10.01 software (Agilent, Palo Alto, California, United States). A normal phase column (Beckman Ultrasphere Si 5μ, 4.6 × 250 mm [Beckman Instruments, Fullerton, California, United States]) and an isocratic solvent system of 0.5% ethyl acetate in hexane (v/v) for 15 min followed by 4% ethyl acetate in hexane for 65 min at a flow rate of 1.4 ml/min at 20 °C (total 80 min) with detection at 325 nm were used.
All of the experimental procedures related to the analysis of dissected mouse eyes, derivatization, and separation of retinoids have been described previously [7]. Rhodopsin and isorhodopsin measurements were performed as described previously [17]. Typically, two mouse eyes were used per assay, and the assays were repeated three to six times. Whole livers were homogenized for 30 s in 4 ml of retinoid derivatization buffer (50 mM MOPS, 10 mM NH2OH, and 50% ethanol in H2O [pH 7.0]) using a Polytron PT1200 motorized homogenizer (Polytron, Bad Wildbad, Germany), and allowed to sit at room temperature for 30 min. Retinoid analysis was performed on 1 ml of liver homogenate following the same extraction procedure used for eyes. Mouse blood was collected from the eye socket using heparinized microhematocrit capillary tubes (Fisher Scientific) immediately following removal of the eyes. Blood was transferred to a tared 1.5-ml Eppendorf tube and weighed (Eppendorf, Hamburg, Germany). Next, 1 ml of retinoid derivatization buffer was added and vortexed (Eppendorf Mixer 5432) for 30 min high speed at room temperature. Blood was then used for retinoid analysis following the same extraction procedure used for eyes. Rhodopsin from Lrat
−/− mice treated with rAAV-Lrat virus was isolated by immunoaffinity chromatography as described previously [18].
Electroretinography
Mice were anesthetized by intraperitoneal injection using 20 μl/g body weight of 6 mg/ml ketamine and 0.44 mg/ml xylazine diluted with 10 mM sodium phosphate (pH 7.2), containing 100 mM NaCl. The pupils were dilated with 1% tropicamide. A contact lens electrode was placed on the eye, and ground electrodes were placed on the scalp and tail. Electroretinograms (ERGs) were recorded with the universal testing and electrophysiologic system UTAS E-3000 (LKC Technologies, Gaithersburg, Maryland, United States). The light intensity was calibrated and computer-controlled. The mice were placed in a Ganzfeld chamber, and scotopic responses to flash stimuli were obtained from both eyes simultaneously. Flash stimuli had a range of intensities (−3.7 to 2.8 log cd·s·m−2), and white light flash duration was adjusted according to intensity (from 20 μs to 1 ms). Three to five recordings were made with intervals of 10 s or longer, and for higher intensity trials, intervals were 10 min or as indicated. Five animals were typically used for recording of each point in gavage conditions. ERGs were performed on all mice treated with rAAV-Lrat. The results were examined using the one-way ANOVA test.
Recordings from Rods
Suction electrode recordings from rod photoreceptors followed published procedures [19]. In brief, a small piece of retina was shredded with fine needles, and the resulting suspension was placed in a chamber on the microscope stage. Single outer segments were drawn by suction into a tightly fitting glass electrode, and changes in outer segment current in response to brief light flashes were measured. All procedures were carried out using infrared illumination (>950 nm). Mice for these experiments were dark adapted for at least 12 h. C57Bl/6J mice were used as controls. All experiments were carried out at 35–37 °C.
Liver RA Analysis
The method of Kane et al. was used with slight variation [20]. Whole mouse livers were removed from Lrat
−/− mice after eye removal during the above experiments. Livers were weighed and then frozen in liquid N2. Frozen livers were transferred to a 15-ml glass centrifuge tube (Corex #8441 [Corning Life Sciences, Acton, Massachusetts, United States]) containing ice-cold phosphate buffered saline (PBS) at a 1:4 ratio of liver to PBS (w/v) to make a 25% homogenate and were homogenized 30 s using a Polytron PT1200 motorized homogenizer. Next, 500 μl of homogenate was transferred to an 8-ml glass tube on ice, and 1 ml of ice-cold ethanol and 5 μl of 5 M NaOH were added and vortexed. Finally, 4 ml of ice-cold hexane was added, and the mixture was vortexed and centrifuged for 5 min using a Beckman J2-HS centrifuge with a JS13.1 swinging bucket rotor for 5 min at 4,000 rpm, 4 °C. The hexane layer was removed and discarded. The hexane extraction was repeated one more time. To extract RA, 20 μl of 12 M HCl was added to the remaining aqueous solution and vortexed. Hexane extraction was performed as above, but hexane was retained from both extractions and dried down under blowing argon at 20 °C. Residue was dissolved in 300 μl of 1,000:4.3:0.675 hexane:2-propanol:acetic acid (v/v) and transferred to an amber glass HPLC vial with glass insert.
For the first separation, 100 μl of sample was injected into the Agilent 1100 HPLC described above. Two tandem normal phase columns were used. The first column was a Varian Microsorb Silica 3μ, 4.6 × 100 mm column (Varian, Palo Alto, California, United States), and the second was the Beckman column described above. An isocratic solvent system of 1,000:4.3:0.675 hexane:2-propanol:glacial acetic acid (v/v) at a flow rate of 1 ml/min at 20 °C with detection at 355 nm was used [21]. The system was calibrated using standards of all-trans-RA and 9-cis-RA purchased from Sigma-Aldrich.
Immunocytochemistry and Histology
Procedures have been described previously [22]. Anti-LRAT monoclonal antibody [10] was directly coupled with Alexa488, or detected by anti-mouse IgG labeled with Cy3. Sections were analyzed under an epifluorescence microscope (Nikon, Tokyo, Japan). Low magnification images were captured with a digital camera (ORCA-ER, Hamamatsu Photonics, Bridgewater, New Jersey, United States) or a Zeiss LSM 510 NLO confocal microscope (Zeiss, Oberkochen, Germany).
Retinas of 17-mo-old Lrat
−/− and WT mice were marked, enucleated, and immersed immediately in a fixative of 4% paraformaldehyde in 0.1 M phosphate buffer (pH 7.4). Following fixation for 4–5 h at 4 °C, the eyecup containing the optic nerve was postfixed in 1% osmium tetroxide in phosphate buffer, dehydrated through a series of graded ethanol, and embedded in Spurr's resin. Sections 0.5–1 μm thick were imaged using a Leica (Wetzlar, Germany) DM-R microscope with Prior stage, using Syncroscan RT software from Syncroscopy (Frederick, Maryland, United States). The scaling for measurement is 182 nm/pixel and 5.5 pixels/micrometer at 40×. The thicknesses of rod outer segments (ROSs) in micrometers as a function of distance from the ONH were measured after import into Deneba Canvas software (ACD Systems, Saanichton, British Columbia, Canada).
Transmission EM
For transmission electron microscopy (EM), mouse eyecups were fixed primarily by immersion in 2.5% glutaraldehyde and 1.6% paraformaldehyde in 0.08 M PIPES (pH 7.4) containing 2% sucrose, initially at room temperature for ~1 h, then at 4 °C for the remainder of 24 h. The eyecups were then washed with 0.13 M sodium phosphate (pH 7.3), and secondarily fixed with 1% OsO4 in 0.1 M sodium phosphate (pH 7.4), for 1 h at room temperature. The eyecups were dehydrated through a CH3OH series and transitioned to the epoxy embedding medium with propylene oxide. The eyecups were embedded for sectioning in Eponate 812. Ultrathin sections (60–70 nm) were stained with aqueous saturated uranium acetate and Reynold's formula lead citrate prior to survey and micrography with a Philips CM10 EM (Philips Electron Optics, Eindhoven, The Netherlands).
Statistical Analyses
Data were expressed as mean ± standard error of the mean (SEM) Liver data was presented as mean ± SEM of one-quarter of whole livers. Blood data was presented as mean ± SEM per 100 mg of blood.
Results
Lrat
−/− Mice Are a Model of Vitamin A Deficiency
We analyzed the retinoid content of Lrat
−/− and Lrat
+/+ mice to determine how the absence of LRAT influenced the levels of the 11-cis-RAL chromophore and its derivatives. Lrat
−/− mice had diminished levels of all-trans-retinoids in the eye [10] and 100- to 1,000-fold lower levels of cis-retinoids compared with WT mice (Figure S1). Cis-retinoids were also observed in Rpe65-deficient mice, but at 5% of the level of WT mice [23]. Formation of cis-retinoids is likely due to the propensity of retinoids to spontaneously isomerize and become trapped by opsin in the photoreceptors. Lrat is expressed in the liver as well as in the eye; the amount of REs in the liver of Lrat
−/− mice was more than 20,000-fold lower than WT mice (Table 1). Circulating all-trans-ROL from the diet was also reduced, producing lower retinoid concentrations in the blood (Table 1). No visual pigments were measurable in the Lrat
−/− mice by direct spectrophotometric analysis.
Table 1 Retinoid Analysis of Tissues from Lrat
−/− Mice Gavaged with 20 μmol 9-cis-R-Ac as a Function of Post-Treatment Time
Visual Pigment in Lrat-Deficient Mice Is Restored by Oral Gavage with 9-cis-Retinoid
We synthesized a series of 9-cis-retinoids and ester derivatives to determine whether the phenotype of Lrat
−/− mice could be reversed by chemically bypassing the LRAT-catalyzed metabolic step (Figure 1A). Oral gavage of these exogenous retinoids led to the generation of visual pigment in Lrat
−/− mice. Visual pigments were measured chromatographically by the retention of 9-cis-RAL oximes (Figure 1B) or spectrophotometrically (unpublished data) by the level of isorhodopsin (i.e., opsin + 9-cis-RAL chromophore). Both methods were highly reproducible and yielded similar results. Visual pigment was rescued in Lrat
−/− mice by oral gavage with 9-cis-RE, 9-cis-RAL, or 9-cis-ROL, whereas all-trans isomers were ineffective. Among REs, 9-cis-R-Ac and 9-cis-R-Su were most efficient (by weight per dose) in restoring pigment (Figure 1B). Esters are readily metabolized in the small intestine and are more inert than RALs and ROLs (Figure 1C). For these reasons 9-cis-R-Ac was chosen for the remaining experiments.
Figure 1 Retinoid Structures, Specificity of Retinoids in Regeneration of Visual Pigment, and Model of Absorption of 9-cis-R-Ac in Mammals
(A) Structures of retinoids used for gavage studies.
(B) Levels of all-trans-RAL oximes, 9-cis-RAL oximes (corresponding to formation of visual pigment isorhodopsin), and all-trans-REs in Lrat
−/− mouse eyes gavaged with 5 mg of each retinoid before 48–72 h dark adaptation (n ≥ 3, data shown with standard deviation [SD]).
(C) Model of absorption of 9-cis-R-Ac in mammals.
Kinetics of Visual Pigment Rescue and Retinoid Clearance by Oral 9-cis-R-Ac in Lrat-Deficient Mice
Gavage with 9-cis-R-Ac produced a transient increase in retinoid levels in the liver and a more sustained increase in levels in the eye (Figure 2). The visual pigment was restored in Lrat
−/− mice 4–5 h after gavage with 9-cis-R-Ac. Visual pigment levels measured by HPLC remained nearly constant for 96 h (Figure 2G). By 120 d following a single oral gavage with 9-cis-R-Ac, dark-reared Lrat
−/− mice retained more than 50% of the pigment (Figure 2H). Considering that~10% of ROSs are phagocytosed daily and replaced by newly formed discs at their base [24], these data suggest that even in the absence of LRAT, the chromophore may be efficiently recycled via hydrolysis of isorhodopsin and directional transport of 9-cis-RAL from the RPE to the photoreceptors.
Figure 2 Retinoids in the Liver and Eyes of Lrat
−/− Mice after 9-cis-R-Ac Treatment
(A–F) Normal-phase HPLC analysis of nonpolar retinoids extracted from the tissues of dark-adapted Lrat
+/+ or Lrat
−/− mice following gavage with retinoids. Peaks marked * represent the solvent change artifact, and peaks labeled (1) indicate an unidentified non-specific compound with λmax= 270 nm. In all experiments, mice were dark-adapted for 48 h after gavage. Shown are results from eyes of dark-adapted control Lrat
+/+ (gray) and Lrat
−/− mice (black) (A); eyes of dark-adapted Lrat
−/− mouse gavaged with all-trans-R-Ac (B); eyes of dark-adapted Lrat
−/− mouse gavaged with 9-cis-R-Ac (C); liver tissue from dark-adapted control Lrat
+/+ (gray) and Lrat
−/− (black) mice (D); liver tissue from dark-adapted Lrat
−/− mouse gavaged with all-trans-R-Ac (E); and liver tissue from dark-adapted Lrat
−/− mouse gavaged with 9-cis-R-Ac (F).
(G–J) Time course of the levels of nonpolar and polar retinoids in the tissues of Lrat
−/− mice following gavage with 9-cis-R-Ac measured by HPLC. After gavage, mice were dark adapted for indicated time before HPLC analysis (n ≥ 3, data shown with SD). The graphs depict: a short time course of 9-cis-RAL oxime levels detected in Lrat
−/− mice eyes following a 20 μmol gavage of 9-cis-R-Ac (G); a longer time course of 9-cis-RAL oxime levels in Lrat
−/− mice eyes following a 20 μmol gavage of 9-cis-R-Ac (H); time course of 9-cis-ROL blood levels in Lrat
−/− mice following gavage with 20 μmol 9-cis-R-Ac (I); and time course of RE and RA levels in liver of Lrat
−/− mice following a 20 μmol gavage with 9-cis-R-Ac (J).
Despite the absence of LRAT, REs were formed transiently in the liver and blood after gavage, suggesting that another enzyme is capable of the esterification reaction, such as acyl coenzyme A:diacylglycerol acyltransferase [25]. In less than 10 h, more than 90% of retinoids, including nuclear receptor activators all-trans-RA and 9-cis-RA, were cleared (Figure 2I and 2J; Table 1). Thus within 4–5 h after gavage, peripheral hydrolysis of REs is faster than synthesis, and retinoids are quickly metabolized or secreted. These data suggest efficient uptake of 9-cis-retinoids by the visual pigment and rapid clearance of excess exogenous retinoids from key metabolizing and transporting tissues, lowering the potential for toxicity.
Regeneration of Visual Pigments Improves the Morphology of Rod Photoreceptors
Maximal visual pigment formation required ~4 μmol of 9-cis-R-Ac (Figure 3A). The level of isorhodopsin in the eye of Lrat
−/− mice after a single gavage was ~60%–70% of the WT level of rhodopsin (Figure 3). When total opsin was isolated by immunoaffinity chromatography, the UV-vis spectrum showed that all opsin present in the retina was regenerated (unpublished data), suggesting that all opsin was properly folded but is present in ROS at lower amounts than in WT ROS. Multiple 5- to 10-μmol doses spread over weeks produced almost full recovery of WT levels of visual pigment (Figure 3B).
Figure 3 Levels of 9-cis-RAL Oximes in the Eyes of Lrat
−/− Mice after a Single or Multiple Dose of 9-cis-R-Ac
(A) The level of 9-cis-RAL in Lrat
−/− mouse eyes after a varying dose of 9-cis-R-Ac.
(B) The level of 9-cis-RAL in Lrat
−/− mouse eyes after a varying size and number of doses of 9-cis-R-Ac. The solid gray line represents a maximal level of isorhodopsin as measured by the level of 9-cis-retinal oximes in Lrat
−/− mouse eyes after ten gavages; dashed gray line indicates the SD. The maximal level of isorhodopsin is comparable to the level of rhodopsin in WT mice (blue dotted line, shown as pmol of 11-cis-retinal/eye).
(C) The level of 9-cis-RAL in Lrat
−/− mouse eyes after 9-cis-R-Ac treatment and light exposure or after exposure to light and re-gavage (n ≥ 3, data shown with SD).
(D and E) Changes in the RPE-ROS interface in control Lrat
−/− mice and Lrat
−/− mice treated with 9-cis-R-Ac. Treated Lrat
−/− mice were gavaged with 9-cis-R-Ac (10 μmol per gavage) six times, 3 d apart, and analyzed (D). Control retina from age-matched (8 wk old) untreated Lrat
−/− mice (E). Considerably improved RPE-ROS processes were observed in all treated mice. RPE, retinal pigment epithelium; ROS, rod outer segments; IS, inner segments. Scale bar, 10 μm.
Exposure to light released the isomerized chromophore as all-trans-ROL in treated Lrat-deficient mice (Figure 3C, black) with kinetics proportional to the intensity of the bleaching light. Subsequent gavages restored the chromophore efficiently (Figure 3C, red). Multiple gavages were effective in mice up to 12 mo old (Figure S2A). Histological analysis showed that the retina of untreated Lrat
−/− mice degenerated slowly (Figure S2B), leaving functionally intact photoreceptors available for treatment in the older mice.
Since ROS structural morphology is thought to be dependent on functional rhodopsin, this result (in conjunction with recycling of the chromophore as discussed above) suggested that ROS morphology might be improved with treatment. We used EM to evaluate the RPE-ROS interface. Control and gavaged Lrat-deficient mice were analyzed. The thickness of the ROS layer was substantially improved, from 10.7 ± 0.16 μm to 14.2 ± 2.2 μm in treated mice, and the RPE-ROS interface showed closer apposition in comparable areas of the retina (Figure 3D and 3E; n = 5, p < 0.002).
Improvement of Rod Responses in Lrat
−/− Mice Treated with 9-cis-R-Ac
Outer segment membrane currents of Lrat
+/+ and Lrat
−/− rods were directly measured with suction electrodes (Figure 4). Rods of untreated Lrat
−/− mice were ~2,000-fold less sensitive than Lrat
+/+ rods. Gavage of Lrat
−/− mice with 9-cis-R-Ac restored near-WT sensitivity, although several treatments were required.
Figure 4 Rescue of Visual Responses Measured by Single-Cell Recording and ERG Responses of single Lrat
+/+ and Lrat
−/− Rods
(A–D) Flash families measured for a Lrat
+/+ rod (A), a control Lrat
−/− rod (B), a Lrat
−/− rod after a single gavage with 9-cis-R-Ac (C), and a Lrat
−/− rod after three gavages with 9-cis-R-Ac (D). Each panel superimposes average responses to five to 20 repeats of a flash, with the flash strength increasing by a factor of two for each successively larger response.
(E) Stimulus-response relations for the same cells in (A–D). Maximal response amplitudes are plotted against the flash strength. Fits are saturating exponential functions, used to estimate the strength of the flash producing a half-maximal response (Lrat
+/+, 26 photons/μm2; Lrat
−/−, 43,000 photons/μm2; singly treated Lrat
−/−, 590 photons/μm2; and multiply treated Lrat
−/−, 69 photons/μm2).
(F and G) Comparison of WT mice scotopic single flash ERG to Lrat
−/− 9-cis-R-Ac gavaged mice and Lrat
−/− and Lrat
+/+ control mice. Lrat
−/− mice were gavaged nine times with 5 μmol 9-cis-R-Ac over a 1-mo time period (n ≥ 3, data shown with SD).
Figure 4A and 4B show flash families recorded from Lrat
+/+ and Lrat
−/− rods. Lrat
−/− rods had a smaller dark current than Lrat
+/+ rods (mean ± SEM: Lrat
−/−, 4.8 ± 0.4 pA, n = 21; Lrat
+/+, 15 ± 1 pA, n = 22). In addition, the dim flash response reached a peak more quickly in Lrat
−/− rods (Lrat
−/−, 108 ± 4 msec; Lrat
+/+, 228 ± 9 msec). Sensitivity was estimated by plotting the amplitude of the response versus flash strength (Figure 4E). The flash necessary to produce a half-maximal response was ~2,000 times brighter in Lrat
−/− rods (Lrat
−/−, 33,000 ± 2,000 photons/μm2; Lrat
+/+, 18 ± 1 photons/μm2). The decreased dark current, decreased sensitivity, and faster response kinetics of Lrat
−/− rods were all consistent with bleaching adaptation produced by residual opsin activity in the absence of a chromophore [26–28]. This phenotype is more severe than that of Rpe65
−/− rods, which are ~150-fold less sensitive than normal [16]. This may be explained by the lower levels of 9-cis-RAL (see Figure S1) [23] seen in Lrat
−/− mice compared with Rpe65
−/− mice. The sensitivity of Lrat
−/− rods was markedly enhanced when the mice were gavaged with 9-cis-R-Ac. Figure 4C and 4D show flash families recorded from Lrat
−/− mice rods after a single gavage (Figure 4C) and multiple gavages (Figure 4D). Figure 4E shows the corresponding stimulus-response relations. The dark current was restored to near-normal levels after a single treatment with 9-cis-R-Ac (15 ± 1 pA, n = 10), and the half-maximal flash strength was reduced to 390 ± 20 photons/μm2 (n = 10). Correcting for the ~3-fold reduced quantum efficiency from rhodopsin to isorhodopsin, the sensitivity of singly-treated rods was ~8-fold less than normal. Treatment with multiple doses of 9-cis-R-Ac increased the sensitivity. The half-maximal flash strength after three treatments was 87 ± 6 photons/μm2 (n = 26). Correcting for the lowered 9-cis-RAL quantum efficiency, this is within a factor of two of WT mouse rod responses.
Restoration of ERG in Lrat
−/− Mice Treated with 9-cis-R-Ac
ERG responses in treated and control Lrat
−/− mice confirmed the improvements described above for individual rod signaling. The a-wave (generated by photoreceptors) and b-wave (generated by bipolar cells) of dark-adapted Lrat
−/− mice were about 5% of the amplitude of those observed in WT mice (Figure 4F). The Lrat
−/− b-wave was also reduced substantially, in proportion to the decreased a-wave. Following multiple 9-cis-R-Ac gavage treatment, ERG a- and b-wave amplitudes increased to about half of WT levels (Figure 4F and 4G).
AAV-Lrat Rescue of Visual Function in Lrat-Deficient Mice
As an alternative to oral gavage, we tested the ability of treatment with rAAV carrying the Lrat gene to restore function in Lrat-deficient mice. Immunolocalization of LRAT in rAAV-Lrat-treated Lrat
−/− mouse eyes showed expression in the vicinity of the injection site (Figure 5A) and was not uniform (Figure 5B). Higher resolution flat-mounted immunocytochemistry showed that Lrat was expressed specifically in the RPE of treated mice, but was not observed in control mice (Figure 5C and 5D). Treatment led to production of rhodopsin as determined by retinoid analysis and isolation of rhodopsin (Figure S3). The measurements of the level of rhodopsin as measured both directly by spectroscopy and indirectly by cis-retinoids (11- or 9-) are in good agreement (see also [29]).
Figure 5 Immunocytochemistry and ERG of rAAV-Lrat-Treated Lrat
−/− Mice
(A) Immunolocalization of LRAT (green) in rAAV-Lrat treated Lrat
−/− mouse eye. Anti-LRAT antibody was directly labeled by Alexa 488. Expression of LRAT is locally restricted (arrowheads) in the eye. Nuclei are stained by Hoechst 33342 (blue).
(B) Higher-power magnification image of (A). LRAT (green) is observed specifically in the RPE cell layer.
(C) Subcellular localization of LRAT in rAAV-Lrat treated Lrat
−/− mouse eye. RPE cells were labeled by anti-LRAT and detected by Cy3-labeled secondary antibody (red). Nuclei are stained by Hoechst 33342 (blue). RPE cell layer was mounted flat on a coverslip and imaged. LRAT is localized in the internal membrane of the RPE cells. Similar localization was observed for WT mouse RPE cells [12].
(D) Control flat-mounted RPE cell layer of untreated Lrat
−/− mouse. Anti-LRAT antibody does not show any non-specific labeling. Bars indicate 100 μm.
(E) Comparison of scotopic single flash ERG of rAAV-Lrat treated, Lrat
+/+, and Lrat
−/− control mice as measured by a-wave amplitudes, (n ≥ 16, data shown with SD).
(F) A plot of a-wave amplitudes at 2.8 cd·s·m−2 intensity as a function of post-treatment time for rAAV-Lrat treated mice.
Treatment with rAAV-Lrat increased the sensitivity of the a- and b-waves in Lrat
−/− mice (Figure 5E and 5F). The rescue was comparable to 9-cis-R-Ac-treated mice. The largest improvement in sensitivity was seen 6 wk after injection of the virus, while the response gradually declined from 6–31 wk post-treatment (Figure 5F). The improvement in a- and b-wave sensitivity following rAAV-Lrat treatment showed considerable variability, most likely due to the inherent variability in the amount of vector surgically delivered to these small eyes (Figure S4).
The visual pigment of rAAV-Lrat-treated Lrat-deficient mice was augmented by oral gavage with 9-cis-R-Ac as described above. This augmentation was observed by HPLC separation and quantification of 11-cis-RAL oximes (a result of rAAV rescue) and 9-cis-RAL oximes (Figure S3A–S3C). ERG responses in Lrat
−/− mice treated with rAAV-Lrat and 9-cis-R-Ac further improved to reach the levels of 9-cis-R-Ac treated mice (Figure S3D and S3E).
Both rAAV-Lrat and 9-cis-R-Ac Treatment Restore PLRs in Lrat-Deficient Mice
PLRs were used to assay rescue of retinal signaling to the brain (Figure 6). Restoration of PLR implies successful transmission of photic information to the olivary pretectum of the brain. PLR of control Lrat
−/− mice were ~3.5 log units less sensitive than heterozygous or WT animals (Figure 6I). Irradiance-response relations for PLR showed that both 9-cis-R-Ac and rAAV-Lrat treatments each increased Lrat
−/− PLR sensitivity by ~2.5 log units (Figure 6I). Figure 6A–6D and Video S1 compare the PLRs of the same Lrat
−/− mouse before and after 9-cis-R-Ac treatment, illustrating the significant gain in photosensitivity observed following treatment. Similarly, Figure 6E–6H and Video S2 show that rAAV-Lrat treatment conferred substantially increased sensitivity to light stimuli as compared with the control contralateral eye of the same animal. Similar improved PLR were recently obtained for Rpe65
−/− mice treated with 9-cis-RAL [30].
Figure 6 Light-Induced Pupillary Constriction of Lrat
−/− Mice Before and After Treatment with 9-cis-R-Ac or rAAV-Lrat
(A–H) 470 nm light (4.79 × 1013 photons·cm−2·sec−1) was used to stimulate pupillary constriction. Untreated Lrat
−/− pupil before (A) and during (B) light exposure. Same mouse as in (A and B) subsequent to treatment with 9-cis-R-Ac, before (C) and during (D) light exposure. Control, untreated pupil of Lrat
−/− mouse before (E) and during (F) light exposure. Contralateral eye of mouse shown in (E and F) treated with rAAV-Lrat, before (G) and during (H) light exposure.
(I) Irradiance-response relations for PLR to 470 nm light.
Complementarity of Viral and Pharmacological Rescue of Lrat
−/− Mice
To determine if viral and pharmacologic treatments could be combined, we performed oral gavage on Lrat
−/− mice previously rescued with rAAV-Lrat. Biochemical augmentation of visual pigment was observed directly, as the elution of 11-cis-RAL oximes (a result of AAV rescue) and 9-cis-RAL oximes (produced by oral gavage) were well separated on a HPLC column (Figure S3A–S3C). While ERG amplitudes of virally-rescued animals were below those seen in mice rescued by oral gavage, ERG responses in Lrat
−/− mice treated with rAAV-Lrat and 9-cis-R-Ac further improved to reach the levels of 9-cis-R-Ac treated mice (Figure S3D and S3E).
Discussion
Here, we demonstrate that inborn errors of metabolism leading to blinding diseases can be successfully treated by gene therapy or pharmacological intervention. Visual pigment formation, tissue morphology, and visual function as measured by single-cell recordings, ERG, and PLR were significantly improved after treating a mouse model of LCA due to deficiency in LRAT with intraocular injections of rAAV-Lrat or by oral gavage with 9-cis-R-Ac. Pharmacological interventions to treat some forms of retinal disease have been attempted in animal models of human blinding conditions, among them treatment of rd1 mice with the Ca2+ channel blocker D-cis-diltiazem and other blockers that prevent rise of intracellular Ca2+ to toxic levels [31], and treatment of rd1 mice with leukemia inhibitory factor, which is involved in the down-regulation of genes involved in synthesis and degradation of cGMP [32]. Other studies have used inhibitors of apaptosis, an approach applicable to a wide range of retinal disorders. Several studies suggest that intraocular injection of neurotrophic factors (e.g., brain-derived neurotrophic factor and ciliary neurotropic factor ) can protect the murine retina from light damage, or delay photoreceptor degeneration in animal models (for example [33]). In most cases, the beneficial effects of treatment last less than a month and require repeated administrations.
Dryja and colleagues showed that ROL supplementation slows the rate of photoreceptor degeneration caused by a T4M rhodopsin mutation in mice [34]. More recently, 9-cis-RAL application to Rpe65
−/− mice resulted in formation of an active iso-rhodopsin and an improvement in the ERG of these animals for a time period of up to 6 mo after treatment [7,16]. The Rpe65 mutation blocks the retinoid cycle by preventing the generation of 11-cis-RAL, the chromophore of visual pigments. Oral application of 9-cis-RAL circumvents this blockage. The blockade has also been rescued by rAVV-Rpe65 gene transfer method in naturally Rpe65-deficient dogs [9]. Rpe
−/− mice display enhanced esterification properties that could lead to trapping 9-cis-retinoids in the form of 9-cis-REs. Thus it was very important to examine if this pharmacological approach would be successful in mice lacking the key esterifying enzyme. Combining oral retinoid treatment (using next-generation retinoid pro-drugs) and viral rAAV-Lrat somatic gene therapy in Lrat
−/− mice led to remarkable rescue of visual functions in these mice. We found that both methods increased ERG responses from ~5% of wild-type levels in Lrat
−/− mice to ~50% of wild-type levels in treated Lrat
−/− mice. Retinoid treatment led to increased levels of visual chromophore from undetectable levels to 600 pmoles per eye. The ROS dark current was restored to near-normal levels of 15 ± 1 pA after a single treatment with 9-cis-R-Ac, and the half-maximal flash strength after three treatments was 87 ± 6 photons/μm2. Correcting for the lowered 9-cis-RAL quantum efficiency, this is within a factor of 2 of WT mouse rod responses. Irradiance-response relations for PLR showed that both 9-cis-R-Ac and rAAV-Lrat treatments each increased Lrat
−/− PLR sensitivity by ~2.5 log units.
Rescue of Visual Functions by Pharmacological Treatment and Gene Transfer
Our previous findings described the Lrat
−/− mouse as a model of LCA with pathological characteristics similar to those found in patients affected by mutations in the LRAT gene [11]. In the current study, we demonstrate that visual functions can be rescued in Lrat
−/− mice, as measured by recovery of visual chromophore and pigment, and by single-cell and ERG responses. Successful restoration of retinotectal signaling, as measured by pupillary responses, was also achieved. Pharmacological treatment was successful in every experimental trial, with restoration of about half-maximal ERG responses compared with the WT mice. While full restitution of the ERG could not be obtained in Lrat
−/− mice in the tested experimental conditions, our finding of nearly complete restitution of single cell responses suggests that remodeling of the neuronal retina in Lrat
−/− may limit functional rescue [35]. Possibly, the ROS of Lrat
−/− mice are shorter as a result of damage caused by phototransduction triggered by free opsin [27,28], a phenomenon reminiscent of ROS shrinkage under continuous stimulation by light (photostasis) or as a result of vitamin A deprivation [36–38]. However, the rod photoreceptors resist complete degeneration in Lrat
−/− mice (even in 17-mo-old mice raised in a normal light cycle), while the cones undergo almost complete degeneration. In fact, the ROS in 17-mo-old mice are nearly as long as in 2-mo-old Lrat
−/− mice (see Figure S2B). A similar observation has been made for Rpe65-deficent mice [39]. Multiple gavages during a 2-wk period led to morphological restoration once the ROS completed the 12-d cycle of phagocytosis and renewal [24]. The improved rescue after multiple gavages was confirmed not only by retinoid analysis by HPLC, but also by morphological analysis and electrophysiological tests. In single-cell recordings, the ~3-fold difference in sensitivity between WT ROS and ROS from multiply treated Lrat
−/− mice could be accounted for by the discrepancy in the quantum yield of opsin loaded with 11-cis-RAL and 9-cis-RAL chromophores. These differences are small compared to the range of intensities over which vision operates [40].
Oral synthetic retinoid treatment was remarkably successful in rescuing the Lrat
−/− phenotype because of the highly specific mechanism of ROL transport and retention [41]. ROL absorption is followed by entrapment of REs in the lipid droplets of hepatic stellate cells and the retinosome structures found in the RPE [12,42]. ROL is mobilized from the intestine or liver as ROL ester components of chylomicrons or as ROL complexed with ROL-binding protein and transthyretin or albumin [12,42–44]. Cis-retinoids employed in this study used the existing retinoid transport system and were efficiently delivered to the eye (see Figure 2C).
We have chosen 9-cis-retinoids in our studies because they display higher stability than 11-cis-retinoids in storage and handling and at the low pH of the stomach, and they require simpler chemical synthesis, yet they share many advantages of 11-cis-retinoids, including the efficient formation of highly sensitive pigments and a similar metabolic pathway of degradation and transport in mammals. Based on their chemical properties, pro-drug cis-REs appear to be a better alternative to RAL or ROL because of their stability and efficient metabolic transformation to active 9-cis-RAL in the eye. Interestingly, other pro-drugs, such as vitamin E, are also delivered as more hydrophilic succinate esters [45]. Such REs lower the hydrophobicity and non-specific diffusion of these lipid-soluble compounds within the membranes of the digestive tract. In the RPE, 9-cis-ROL is transiently trapped as REs (Table 1), which then undergo hydrolysis and oxidation to regenerate the active chromophore 9-cis-RAL (see Figure 1C). The absorption and transformation of cis-retinoids might be similar to that of all-trans-retinoids because of the multiple overlapping activities of the enzymes involved and their low specificity toward different retinoid isomers [41].
Viral treatment was also effective in restoring the normal retinoid cycle in the eye and rescuing visual function. The rescue peaked at 6 wk and then slowly decayed. The expression of Lrat was localized to the injection site, but significant rescue of visual pigment rhodopsin was observed (~50%). More eye-to-eye variability was observed with viral rescue than with pharmacologic rescue. Because of their small size and relatively large lenses, which occupy about 70% of the eye, subretinal injection surgery is more complicated for mouse eyes than it would be for larger eyes.
While direct comparison of rescue methods is difficult, it appears that both therapeutic methods provide efficient recovery of higher order visual responses, as tested by PLRs. The clear advantage of oral retinoid treatment is its ease of administration compared with the subretinal injections required for viral vectors. Its primary disadvantage is its potential for long-term systemic toxicity compared to vector targeting of LRAT to the RPE. The usual advantage of rAAV-delivered gene therapy over systemic drugs is the persistence of passenger gene expression after a single administration; however, in this case there was a gradual but distinct loss of delivered LRAT function after several months. This has not been observed in other systems (RPE65 dog and mouse studies) and may reflect an Lrat
−/−-specific effect on RPE viability. The slow and gradual decline in ERG amplitude in treated mice is also observed with age in Lrat
+/+ mice. Although the decline rate during the first 7 mo of life for Lrat
+/+ mice is slight (T. Maeda and KP, unpublished data) and is slower than in rAAV-treated mice, the direct comparison is not precise, as untreated retinas of Lrat
−/− degenerate slowly, and the level of opsin expression appears to be slightly lower (unpublished data). Thus, the observed decline in ERG amplitudes could be a result of several different factors, including the stability of rAAV infection, retina remodeling in Lrat
−/− mice [35], and shorter ROS structure in Lrat
−/− mice (this study) and degeneration of cone photoreceptors (unpublished data). Importantly, 9-cis-R-Ac treatment appeared to improve rescue in virally treated animals, and resulted in reconstitution of both native 11-cis-retinoids and administered 9-cis-retinoids simultaneously.
Insight into Retinoid Metabolism
The level of 9-cis-RAL is maintained in the eyes of 9-cis-R-Ac-treated Lrat
−/− dark-adapted mice for over 4 mo. This finding suggests that the rod-RPE system efficiently recycles the chromophore of phagocytosed visual pigment. This phenomenon can be readily observed, because retinoids are only transiently stored in organs other than the eye in Lrat
−/− mice. Approximately 10% of the ROS is renewed each day [46]. The entire length of the ROS is thus completely renewed in less than 2 wk. However, the 9-cis-RAL based pigment persists in the eye for 4 mo post-gavage at a ~50% level. This can be explained only by the recovery of 9-cis-chromophore from isorhodopsin phagocytosed by the RPE. Second, even though REs are transiently formed in the eye (Table 1), they do not support the retinoid cycle, suggesting that LRAT is a key enzyme in this process and that formation of esters is insufficient in the completion of the cycle. This observation suggests that localization of these esters is crucial for the proper function of the retinoid cycle. Third, we confirmed reports obtained by Fan et al. [23] that mice with an impaired retinoid cycle appear to spontaneously generate 9-cis-RAL that is then trapped by ROS containing almost exclusively free opsin.
Leber Congenital Amaurosis and Potential Treatment
The major advantage of retinoid-based treatments of LCA resulting from the deficiency of LRAT is that these compounds are not stored in the liver for any prolonged time and are quickly oxidized and secreted. Retinoid absorption in mammals is an active process driven by esterification/hydrolysis cycles in the intestine and liver and in the RPE. Esterification is carried out mainly by LRAT [47], as evidenced by the fact that the absence of LRAT makes retinoids vulnerable to quick elimination from the body. In the absence of LRAT, the equilibrium between ROLs and REs is clearly shifted in favor of free ROL. Low levels of REs are present in mice deficient in LRAT expression (Lrat
−/−), but these esters are formed transiently as a possible consequence of acyl-CoA:retinol acyltransferase activity [25, 48].
An important finding is that the pharmacological treatment can be sustained multiple times, and that several low-dose treatments show cumulative effects. This treatment is possible at any age, as the morphological structure of the retina is to some degree preserved in both young and old mice. These data are reminiscent of results obtained following gene therapy of RPE65-deficient mice and dogs [7,9,12,16,49,50].
Cis- and all-trans-RALs are irreversibly oxidized to RAs by RAL dehydrogenase types 1, 2, 3, and 4 [51]. All-trans-RA and its 9-cis-isomer are important regulators of gene expression via nuclear receptors [52]. For any retinoid-based therapy there can be negative side effects resulting from potential production of RAs and their teratogenic effects. To prevent their accumulation, RAs are oxidized by CYP26A1, CYP26B1, and CYP26C1 to 4-hydroxy-RA, 4-oxo-RA, and 18-hydroxy-RA [53]. As demonstrated here, only a low and transient level of RAs is observed in the Lrat
−/− mice gavaged with 9-cis-R-Ac, arguing that the potential side effects due to RA production are highly attenuated. More toxicological studies are needed before pharmacological treatment can be proposed for treatment of similar human conditions. It is possible that a combination of viral and pharmacological treatment could be the most efficient means of restoring vision in patients afflicted with LCA, for whom no treatment is currently available.
The rescue of PLRs in Lrat
−/− demonstrates that treatment of the retinal defects results in ultimate improvement in signaling from eye to brain. PLRs are mediated by both inner retinal, non-visual photoreceptors [54] and rod and cone input. While measurement of the PLR in this study could not readily distinguish between rescue of inner and outer retinal photoreceptors, given the observed dramatic improvement of the ERG, these results suggest that rescued outer retinal function may be at least partly responsible for rescue of the PLR. Determination of the extent to which pharmacologic or somatic gene therapy rescue of Lrat
−/− mice would allow restoration of actual visual function will require additional study. It is unknown at present to what extent mice lacking Lrat develop normal intraretinal circuitry, and it is similarly unknown to what extent these mice develop amblyopia from visual deprivation. The apparent visual improvement associated with rAAV rescue in dogs mutant in the Rpe65 gene [9], however, does suggest that form vision may be restored in at least some forms of LCA.
Our findings establish that both chromophore supplementation and somatic gene therapy are effective in improving visual functions in the Lrat
−/− mouse, an animal model of LCA that accurately replicates the pathology of the disease. The treatments are optimally effective in combination and if the chromophore supplementation is continued at low doses for a longer period of time. Applying 9-cis-R-Ac treatment to virally treated animals appeared to improve rescue, and resulted in reconstitution of both native 11-cis- and administered 9-cis- retinoids simultaneously. There is a great value in combined approaches, for several reasons. First, either one may prove to be more suitable for a specific age group of patients, thus offering effective treatment for a wider age range; second, a partial and regional rescue by rAVV, as observed in the Rpe65-deficient dog (unpublished data), could be augmented by retinoid treatment, reaching the whole retina; third, the rescue in rAVV-treated animals allows storage of 9-cis-REs that might be mobilized when needed (for example, high bleaching levels). Clinically, it is likely that pharmacologic and somatic gene therapeutic approaches, if successful, could be used in complementary fashion; for example, treatment of appropriate patients with oral retinoids could begin in infancy to avoid amblyopia, avoiding the difficulties associated with surgery in very young patients, while at older ages long-lasting drug-free treatment might be achieved by surgical introduction of viral vectors.
This work on the rescue of vision extends previous studies by Van Hooser et al. [7,16] in part by employing novel compounds with potentially better drug-like properties, and to the second LCA mouse model. In contrast to Rpe65
−/− mice, Lrat
−/− mice do not have significant capacity to store retinoids outside of the visual pigment; however, the pharmacological approach was highly successful. There have been reports documenting the potential adverse effects of high doses of retinoids [55]. During these and related studies of both acute and prolonged (up to 1 y) multiple dose treatments, no adverse effects were observed for mice of several genetic backgrounds. Number and size of litters, coat grooming, survival after gavage, post-natal development, and growth/weight curves were unaffected by treatment. Lrat
−/− mice may be particularly resistant to potential toxicity of RE compounds, as lack of esterification of the retinoid agents actively leads to rapid removal through secretion and oxidation to oxo/hydroxo RA compounds. This reaction occurs not only in the eye but also in the liver, where LRAT is also normally expressed. The transient formation and accumulation of retinoids observed in these mice soon after treatment disappeared within a day in most cases. Only visual pigments had the capability to retain these retinoids for a long period of time, as shown in Figure 2H. These observations raise the hope that, after formal toxicological studies, these RE pro-drugs have the potential to be extended effectively to humans.
Supporting Information
Figure S1 Chromatographic Separation of Retinoids Extracted from 16 Lrat
−/− Mouse Eyes
(A) Peak 1, syn-11-cis-RAL oxime; 2, undetermined compound with λmax at 315 nm; 3, syn-all-trans-RAL oxime; 4, syn-9-cis-RAL oxime; 5, 13-cis-ROL; and 6, all-trans-ROL.
(B) Retinoids bound to immunoaffinity purified opsin from 40 untreated Lrat
−/− mouse eyes (one-third of the sample was loaded on the HPLC column).
(125 KB PDF).
Click here for additional data file.
Figure S2 Levels of 9-cis-RAL in Mice of Different Ages Treated with 9-cis-R-Ac
(A) 9-cis-RAL oxime levels in Lrat
−/− mouse eyes following gavage with 20 μmol 9-cis-R-Ac at differing ages. The analysis is performed after 72 h of post-gavage dark adaptation (n ≥ 3).
(B) ROS thickness as a function of the retinal location from the optic nerve head (in mm). Four age-matched retinas of WT and Lrat
−/− were analyzed both in inferior/superior and nasal/temporal orientations, and the data plotted with Microsoft Excel. The dotted lines represent data from 2-mo-old Lrat
+/− and Lrat
−/− mice as published earlier [10].
(60 KB PDF).
Click here for additional data file.
Figure S3 Retinoid Analysis and ERG of rAAV-Lrat Treated Lrat
−/− Mice Augmented with 9-cis-R-Ac
(A) Lrat
−/− mouse treated with rAAV-Lrat. Inset, immunoaffinity-purified rhodopsin from the retina of Lrat
−/− mice treated with rAAV-Lrat virus. E, corresponding fraction in elution.
(B) Lrat
−/− mouse treated with 9-cis-R-Ac.
(C) Lrat
−/− mouse treated with rAAV-Lrat and 9-cis-R-Ac. Peaks marked * represent the solvent change artifact. Inset, the chromatogram of retinoids from WT mouse.
(D) Scotopic single-flash ERG a-waves.
(E) Scotopic single-flash ERG b-waves (n ≥ 10, data shown with SEM).
(175 KB PDF).
Click here for additional data file.
Figure S4 Scatter Plot of a- and b-waves of ERGs Obtained from rAAV-Lrat-Treated and Control Lrat
−/− Mice
(A) ERG a-waves of 6- to 7-wk-old Lrat
−/− control mice.
(B) ERG b-waves of 6- to 7- wk-old Lrat
−/− control mice.
(C) ERG a-waves of 6- to 7- wk-old rAAV-Lrat-treated mice.
(D) ERG b-waves of 6- to 7- wk-old rAAV-Lrat-treated mice (n ≥ 39).
(129 KB PDF).
Click here for additional data file.
Video S1 Effect of 9-cis-R-Ac Treatment on the PLR of Lrat
−/− Mice
PLRs of an individual Lrat
−/− mouse (left eye) recorded prior (control) and 72 h subsequent (9-cis-R-Ac) to three 5-μmol doses of 9-cis-R-Ac. Presence of light stimulus (30-s pulse of narrow bandpass 470 nm light, 1.38 × 1014 photons·cm−2·sec−1) is represented in the movie by a light bulb symbol.
(3.0 MB MOV).
Click here for additional data file.
Video S2 Effect of Intraocular Injection of rAAV-Lrat on the PLR of Lrat
−/− Mice
PLRs of an individual Lrat
−/− mouse recorded from rAAV-Lrat-treated left eye (rAAV-Lrat) and non-treated right eye (control). Presence of light stimulus (30-s pulse of narrow bandpass 470 nm light, 4.79 × 1013 photons·cm−2·sec−1) is represented in the movie by a light bulb symbol.
(3.0 MB MOV).
Click here for additional data file.
Accession Number
The GenBank (http://www.ncbi.nlm.nih.gov/) accession number of Chromosome 11q13 is AF139813.
Patient Summary
Background
Some causes of blindness are inherited. Leber congenital amaurosis is one inherited disease that causes degeneration and loss of activity of the retina—the tissue at the back of the eye. Hence, babies have a severe loss of vision at birth as well as roving eye movements (nystagmus), deep-set eyes, and sensitivity to bright light. One cause of Leber congenital amaurosis is loss of an enzyme called lecithin:retinol acyl transferase (LRAT), which is required for regeneration of a pigment necessary for the eye to detect light. Currently there is no treatment for this condition.
Why Was This Study Done?
There are animal models of this disease, and previous work has suggested that there are two possible ways of treating this condition. One is by giving by mouth synthetic pigments similar to those in the eye. Another is by gene therapy with viruses that replace the abnormal gene with normal copies. The authors wanted to test these methods in a mouse model, and see if the two approaches worked well together.
What Did the Researchers Do and Find?
They treated mice that had a genetic defect mimicking the human condition by placing a virus carrying the normal gene directly into their eyes. They also gave the mice pro-drugs (compounds that are turned into the active drug inside the body) by mouth, which bypassed the missing step in the regeneration of the retinal compound. The authors found that they got the best results either using individual methods or when they tried both approaches together. They could show that the mice had the electrical impulses that are a sign that the eye is working correctly, and in addition, their pupils responded to light.
What Do These Findings Mean?
These results are an early step to making these treatments available to patients. Obviously, they would only help patients who had this particular genetic defect, and possible defects in genes from this same metabolic pathway. Before these treatments were to be used, many other questions would need to be answered, including whether these oral compounds might be toxic if given repeatedly—as they would need to be. Also, acceptable methods of placing the gene therapy into human eyes would need to be found.
Where Can I Get More Information Online?
The Foundation Fighting Blindness funds research on retinal degenerative diseases, and has a number of pages of information for patients and families:
http://www.blindness.org/
Contact a Family is a UK charity that has information on specific conditions and rare disorders:
http://www.cafamily.org.uk
MedlinePlus also has a series of pages on retinal disorders:
http://www.nlm.nih.gov/medlineplus/retinaldisorders.html
We thank Dan Possin for EM analysis and Rebecca Birdsong for comments on the manuscript. This research was supported by National Institutes of Health grants EY09339 to KP, EY08123 to WB, EY14988 to RNVG, and EY11123 and EY13729 to WWH; grants from Research to Prevent Blindness to the Department of Ophthalmology at the University of Utah and Washington University; the Culpepper Medical Scholar grant from Rockefeller Brothers Foundation to RNVG; a Center Grant from the Foundation Fighting Blindness to the University of Utah; and a grant from the EK Bishop Foundation. Dan Possin was supported by a Vision CORE Grant EY01730. Retinagenix had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Citation: Batten ML, Imanishi Y, Tu DC, Doan T, Zhu L, et al. (2005) Pharmacological and rAAV gene therapy rescue of visual functions in a blind mouse model of Leber congenital amaurosis. PLoS Med 2(11): e333.
Abbreviations
AAVadeno-associated virus
EMelectron microscopy
ERGelectroretinogram
LCALeber congenital amaurosis
LRATlecithin:retinol acyl transferase
PLRpupillary light response
RAretinoic acid
R-Acretinyl acetate
RALretinal
REretinyl ester
ROLretinol
ROSrod outer segment
RPEretinal pigment epithelium
R-Suretinyl succinate
SDstandard deviation
SEMstandard error of the mean
WTwild-type
==== Refs
References
Rattner A Sun H Nathans J Molecular genetics of human retinal disease Annu Rev Genet 1999 33 89 131 10690405
Baehr W Wu SM Bird AC Palczewski K The retinoid cycle and retina disease Vision Res 2003 43 2957 2958 14611932
Jacobson SG Cideciyan AV Regunath G Rodriguez FJ Vandenburgh K Night blindness in Sorsby's fundus dystrophy reversed by vitamin A Nat Genet 1995 11 27 32 7550309
Lotery AJ Namperumalsamy P Jacobson SG Weleber RG Fishman GA Mutation analysis of 3 genes in patients with Leber congenital amaurosis Arch Ophthalmol 2000 118 538 543 10766140
Cremers FP van den Hurk JA den Hollander AI Molecular genetics of Leber congenital amaurosis Hum Mol Genet 2002 11 1169 1176 12015276
Redmond TM Yu S Lee E Bok D Hamasaki D Rpe65 is necessary for production of 11-cis -vitamin A in the retinal visual cycle Nat Genet 1998 20 344 351 9843205
Van Hooser JP Aleman TS He YG Cideciyan AV Kuksa V Rapid restoration of visual pigment and function with oral retinoid in a mouse model of childhood blindness Proc Natl Acad Sci U S A 2000 97 8623 8628 10869443
Campochiaro PA Gene therapy for retinal and choroidal diseases Expert Opin Biol Ther 2002 2 537 544 12079489
Acland GM Aguirre GD Ray J Zhang Q Aleman TS Gene therapy restores vision in a canine model of childhood blindness Nat Genet 2001 28 92 95 11326284
Batten ML Imanishi Y Maeda T Tu DC Moise AR Lecithin-retinol acyltransferase is essential for accumulation of all-trans-retinyl esters in the eye and in the liver J Biol Chem 2004 279 10422 10432 14684738
Thompson DA Li Y McHenry CL Carlson TJ Ding X Mutations in the gene encoding lecithin retinol acyltransferase are associated with early-onset severe retinal dystrophy Nat Genet 2001 28 123 124 11381255
Imanishi Y Batten ML Piston DW Baehr W Palczewski K Noninvasive two-photon imaging reveals retinyl ester storage structures in the eye J Cell Biol 2004 164 373 383 14745001
Zolotukhin S Potter M Hauswirth WW Guy J Muzyczka N A “humanized” green fluorescent protein cDNA adapted for high-level expression in mammalian cells J Virol 1996 70 4646 4654 8676491
Esumi N Oshima Y Li Y Campochiaro PA Zack DJ Analysis of the VMD2 promoter and implication of E-box binding factors in its regulation J Biol Chem 2004 279 19064 19073 14982938
Maeda T Van Hooser JP Driessen CA Filipek S Janssen JJ Evaluation of the role of the retinal G protein-coupled receptor (RGR) in the vertebrate retina in vivo J Neurochem 2003 85 944 956 12716426
Van Hooser JP Liang Y Maeda T Kuksa V Jang GF Recovery of visual functions in a mouse model of Leber congenital amaurosis J Biol Chem 2002 277 19173 19182 11897783
Palczewski K Van Hooser JP Garwin GG Chen J Liou GI Kinetics of visual pigment regeneration in excised mouse eyes and in mice with a targeted disruption of the gene encoding interphotoreceptor retinoid-binding protein or arrestin Biochemistry 1999 38 12012 12019 10508404
Zhu L Jang GF Jastrzebska B Filipek S Pearce-Kelling SE A naturally occurring mutation of the opsin gene (T4R) in dogs affects glycosylation and stability of the G protein-coupled receptor J Biol Chem 2004 279 53828 53839 15459196
Sampath AP Rieke F Selective transmission of single photon responses by saturation at the rod-to-rod bipolar synapse Neuron 2004 41 431 443 14766181
Kane MA Chen N Sparks S Napoli JL Quantification of endogenous retinoic acid in limited biological samples by LC/MS/MS Biochem J 2005 388 363 369 15628969
Klvanova J Brtko J Selected retinoids: Determination by isocratic normal-phase HPLC Endocr Regul 2002 36 133 141 12463969
Haeseleer F Jang GF Imanishi Y Driessen CA Matsumura M Dual-substrate specificity short chain retinol dehydrogenases from the vertebrate retina J Biol Chem 2002 277 45537 45546 12226107
Fan J Rohrer B Moiseyev G Ma JX Crouch RK Isorhodopsin rather than rhodopsin mediates rod function in RPE65 knock-out mice Proc Natl Acad Sci U S A 2003 100 13662 13667 14578454
Bok D Retinal photoreceptor-pigment epithelium interactions. Friedenwald lecture Invest Ophthalmol Vis Sci 1985 26 1659 1694 2933359
Yen CL Monetti M Burri BJ Farese RV The triacylglycerol synthesis enzyme DGAT1 also catalyzes the synthesis of diacylglycerols, waxes, and retinyl esters J Lipid Res 2005 46 1502 1511 15834126
Dowling JE Chemistry of visual adaptation in the rat Nature 1960 188 114 118 13724150
Palczewski K Saari JC Activation and inactivation steps in the visual transduction pathway Curr Opin Neurobiol 1997 7 500 504 9287193
Woodruff ML Wang Z Chung HY Redmond TM Fain GL Spontaneous activity of opsin apoprotein is a cause of Leber congenital amaurosis Nat Genet 2003 35 158 164 14517541
Saari JC Garwin GG Van Hooser JP Palczewski K Reduction of all-trans-retinal limits regeneration of visual pigment in mice Vision Res 1998 38 1325 1333 9667000
Fu Y Zhong H Wang M-HH Luo D-G Liao HW Intrinsically photosensitive retinal ganglion cells detect light with a vitamin A-based photopigment, melanopsin Proc Natl Acad Sci U S A 2005 102 10339 10344 16014418
Frasson M Sahel JA Fabre M Simonutti M Dreyfus H Retinitis pigmentosa: Rod photoreceptor rescue by a calcium-channel blocker in the rd mouse Nat Med 1999 5 1183 1187 10502823
LaVail MM Yasumura D Matthes MT Lau-Villacorta C Unoki K Protection of mouse photoreceptors by survival factors in retinal degenerations Invest Ophthalmol Vis Sci 1998 39 592 602 9501871
Okoye G Zimmer J Sung J Gehlbach P Deering T Increased expression of brain-derived neurotrophic factor preserves retinal function and slows cell death from rhodopsin mutation or oxidative damage J Neurosci 2003 23 4164 4172 12764104
Li T Sandberg MA Pawlyk BS Rosner B Hayes KC Effect of vitamin A supplementation on rhodopsin mutants threonine-17 –> methionine and proline-347 –> serine in transgenic mice and in cell cultures Proc Natl Acad Sci U S A 1998 95 11933 11938 9751768
Jones BW Watt CB Frederick JM Baehr W Chen CK Retinal remodeling triggered by photoreceptor degenerations J Comp Neurol 2003 464 1 16 12866125
Penn JS Williams TP Photostasis: Regulation of daily photon-catch by rat retinas in response to various cyclic illuminances Exp Eye Res 1986 43 915 928 3817032
Hofmann KP Schleicher A Emeis D Reichert J Light-induced axial and radial shrinkage effects and changes of the refractive index in isolated bovine rod outer segments and disc vesicles: Physical analysis of near-infrared scattering changes Biophys Struct Mech 1981 8 67 93 7326356
Naash MI LaVail MM Anderson RE Factors affecting the susceptibility of the retina to light damage Prog Clin Biol Res 1989 314 513 522 2608676
Znoiko SL Rohrer B Lu K Lohr HR Crouch RK Downregulation of cone-specific gene expression and degeneration of cone photoreceptors in the rpe65
−/− mouse at early ages Invest Ophthalmol Vis Sci 2005 46 1473 1479 15790918
Baylor DA Photoreceptor signals and vision. Proctor lecture Invest Ophthalmol Vis Sci 1987 28 34 49 3026986
McBee JK Palczewski K Baehr W Pepperberg DR Confronting complexity: The interlink of phototransduction and retinoid metabolism in the vertebrate retina Prog Retin Eye Res 2001 20 469 529 11390257
Imanishi Y Gerke V Palczewski K Retinosomes: New insights into intracellular managing of hydrophobic substances in lipid bodies J Cell Biol 2004 166 447 453 15314061
Vogel S Piantedosi R O'Byrne SM Kako Y Quadro L Retinol-binding protein-deficient mice: Biochemical basis for impaired vision Biochemistry 2002 41 15360 15368 12484775
Harrison EH Mechanisms of digestion and absorption of dietary vitamin A Annu Rev Nutr 2005 25 87 103 16011460
Kline K Yu W Sanders BG Vitamin E and breast cancer J Nutr 2004 134 3458S 3462S 15570054
LaVail MM Rod outer segment disk shedding in rat retina: Relationship to cyclic lighting Science 1976 194 1071 1074 982063
Ruiz A Winston A Lim YH Gilbert BA Rando RR Molecular and biochemical characterization of lecithin retinol acyltransferase J Biol Chem 1999 274 3834 3841 9920938
O'Byrne SM Wongsiriroj N Libien JM Vogel S Goldberg IJ Retinoid absorption and storage is impaired in mice lacking lecithin: Retinol acyltransferase (LRAT) J Biol Chem 2005 E-pub ahead of print
Liang Y Fotiadis D Filipek S Saperstein DA Palczewski K Organization of the G protein-coupled receptors rhodopsin and opsin in native membranes J Biol Chem 2003 278 21655 21662 12663652
Aleman TS Jacobson SG Chico JD Scott ML Cheung AY Impairment of the transient pupillary light reflex in Rpe65
−/− mice and humans with Leber congenital amaurosis Invest Ophthalmol Vis Sci 2004 45 1259 1271 15037595
Zhao D McCaffery P Ivins KJ Neve RL Hogan P Molecular identification of a major retinoic-acid-synthesizing enzyme, a retinaldehyde-specific dehydrogenase Eur J Biochem 1996 240 15 22 8797830
Chambon P A decade of molecular biology of retinoic acid receptors Faseb J 1996 10 940 954 8801176
Fujii H Sato T Kaneko S Gotoh O Fujii-Kuriyama Y Metabolic inactivation of retinoic acid by a novel P450 differentially expressed in developing mouse embryos EMBO J 1997 16 4163 4173 9250660
Berson DM Strange vision: Ganglion cells as circadian photoreceptors Trends Neurosci 2003 26 314 320 12798601
Collins M.D Mao G.E Teratology of retinoids Annu Rev Pharmacol Toxicol 1999 39 399 430 10331090
|
16250670
|
PMC1274279
|
CC BY
|
2021-01-05 10:40:29
|
no
|
PLoS Med. 2005 Nov 1; 2(11):e333
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020333
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 1625301210.1371/journal.pmed.0020343Research ArticleImmunologyInfectious DiseasesMicrobiologyVirologyHematologyPathologyVaccinesInfectious DiseasesProlonged Activation of Virus-Specific CD8+T Cells after Acute B19 Infection Prolonged Activated CD8 Response to B19Isa Adiba
1
*Kasprowicz Victoria
2
3
Norbeck Oscar
1
Loughry Andrew
3
Jeffery Katie
4
Broliden Kristina
1
Klenerman Paul
3
Tolfvenstam Thomas
1
5
Bowness Paul
2
6
1Institution for Medicine, Infectious Disease Unit, Center for Molecular Medicine, Karolinska Institutet, Karolinska University Hospital, Stockholm, Sweden,2MRC Human Immunology Unit, Weatherall Institute of Molecular Medicine, Oxford, United Kingdom,3Peter Medawar Building for Pathogen Research, Nuffield Department of Medicine, Oxford University, Oxford, United Kingdom,4Department of Virology, John Radcliffe Hospital, Oxford, United Kingdom,5Division of Clinical Virology, Karolinska Institutet, Karolinska University Hospital, Stockholm, Sweden,6Nuffield Orthopaedic Centre NHS Trust, Oxford, United KingdomMoss Paul Academic EditorUniversity of BirminghamUnited Kingdom*To whom correspondence should be addressed. E-mail: [email protected]
Competing Interests: PK is on the Editorial Board of PLoS Medicine.
Author Contributions: PK, TT, and PB designed the study. AI, VK, ON, and AL performed the experimental work. AI and VK analyzed the data. KJ and KB enrolled patients. AI, VK, PK, TT, and PB contributed to writing the paper.
12 2005 1 11 2005 2 12 e34318 4 2005 17 8 2005 Copyright: © 2005 Isa et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.
A Persistent Immune Response to an Acute Virus
Background
Human parvovirus B19 (B19) is a ubiquitous and clinically significant pathogen, causing erythema infectiosum, arthropathy, transient aplastic crisis, and intrauterine fetal death. The phenotype of CD8+ T cells in acute B19 infection has not been studied previously.
Methods and Findings
The number and phenotype of B19-specific CD8+ T cell responses during and after acute adult infection was studied using HLA–peptide multimeric complexes. Surprisingly, these responses increased in magnitude over the first year post-infection despite resolution of clinical symptoms and control of viraemia, with T cell populations specific for individual epitopes comprising up to 4% of CD8+ T cells. B19-specific T cells developed and maintained an activated CD38+ phenotype, with strong expression of perforin and CD57 and downregulation of CD28 and CD27. These cells possessed strong effector function and intact proliferative capacity. Individuals tested many years after infection exhibited lower frequencies of B19-specific cytotoxic T lymphocytes, typically 0.05%–0.5% of CD8+ T cells, which were perforin, CD38, and CCR7 low.
Conclusion
This is the first example to our knowledge of an “acute” human viral infection inducing a persistent activated CD8+ T cell response. The likely explanation—analogous to that for cytomegalovirus infection—is that this persistent response is due to low-level antigen exposure. CD8+ T cells may contribute to the long-term control of this significant pathogen and should be considered during vaccine development.
Although parvovirus B19 apparently causes an acute viral infection, it appears to induce a persistent activated CD8 + T cell response.
==== Body
Introduction
Human parvovirus B19 (B19) is a ubiquitous, single-stranded DNA virus. The 5.6-kb genome codes for only three major proteins, the two overlapping capsid proteins VP1 and VP2 and the non-structural protein NS1. B19 targets immature erythroid cells in the bone marrow after respiratory transmission. Common manifestations of the infection are the benign febrile illness erythema infectiosum followed by an acute arthropathy (in approximately 5% of children and up to 50% of adult infections) that spontaneously resolves within 3 wk. Some adult patients who develop chronic arthritis can fulfil the diagnostic criteria of rheumatoid arthritis [1,2]. Other severe manifestations include transient aplastic crisis in individuals with increased red cell turnover, and chronic anaemia in immunocompromised patients. Furthermore, infection during pregnancy is a major cause of fetal death. Clinical resolution of acute infection is associated with the emergence of antiviral IgG, which is maintained lifelong [3]. However, although antibodies are of importance, evidence for an important role for cellular immune responses in the control of B19 infection is also emerging. In some individuals with apparently intact antibody responses, virus replication continues long term [4]. Readily detectable CD8+ T cell responses in three asymptomatic seropositive individuals have also been observed [5]. CD8+ T cells play an essential role in the control of viral infections by direct killing of virus-infected cells and through cytokine secretion. After a primary viral infection, naïve antigen-specific T cells expand clonally and undergo several differentiation stages, followed by a contraction phase mediated by apoptosis. Expression of the chemokine receptor CCR7 on memory T cells has been used to divide this population into central and effector memory subsets. CCR7+ “central” memory lymphocytes mostly home to secondary lymphoid organs and have a high proliferative capacity in response to antigen re-encounter. CCR7− “effector” memory cells home to non-lymphoid organs and are capable of rapid effector function [6]. Peptide–major histocompatibility complex multimers can be used to further quantify and phenotype T cells bearing T cell receptors of appropriate peptide/human leukocyte antigen (HLA) specificity [7,8]. Amongst “effector” memory populations, a spectrum of phenotypes exists, as assessed by surface expression of markers such as CD27, CD28, and CD57, and intracellular expression of perforin. Amongst the viral infections studied, cytomegalovirus (CMV) is associated with the emergence of the largest long-term memory populations with the most “mature” phenotype, CD27/CD28 low, CD57 high, and often perforin positive [9]. The factors that ultimately determine the phenotype and function of these populations of lymphocytes are not fully understood, but continued exposure to antigen is thought to be important. Persistent infections such as hepatitis C virus (HCV), Epstein–Barr virus (EBV), and HIV also show antiviral populations that are “effector” (CD62L or CCR7 low), but show lower levels of maturation than CMV-specific responses [7].
Its small stable genome and clear seroconversion illness make B19 an ideal viral infection in which to study the evolution of antigen-specific T cells longitudinally. Most other significant viruses in humans are either highly variable (HCV and HIV) or have extremely large genomes (EBV and CMV). In our previous analyses, we observed a surprisingly robust response to B19 in three individuals with an asymptomatic seropositive state [5]. We aimed in this study to define the origins of such responses and track a range of antiviral populations during and after acute disease.
Methods
Study Participants
Eleven previously healthy immunocompetent adults presenting to their general practitioner with symptoms of fever, arthralgia, fatigue, and rash were prospectively identified (B19 IgM-positive) at the Departments of Clinical Virology at the Oxford Radcliffe Hospitals, Oxford, United Kingdom, and Karolinska University Hospital, Stockholm, Sweden. The two cohorts were ascertained and studied independently. Clinical details of the patients are shown in Table 1. The timing of blood samples is given from onset of symptoms for both cohorts. B19 DNA in serum was detected by nested PCR amplifying 284 bp in the NS1 gene with a sensitivity of 103 DNA copies/ml (Table 1) [5]. In addition, five healthy B19 IgG-positive, B19 IgM/DNA-negative healthy laboratory volunteers were studied as a “remotely infected” cohort. One gave a history of acute rash and arthritis 10 y previously; the remainder had no history suggestive of B19 infection and were likely infected in childhood. No interferon-γ (IFNγ)– or B19-specific T cells were detected in B19 IgG/IgM/DNA-negative healthy controls (data not shown). Peripheral blood mononuclear cells (PBMCs) were separated from heparinized blood samples within 8 h of sampling, by gradient centrifugation on Ficoll-Paque (Amersham Biosciences, Uppsala, Sweden) or Lymphoprep (Fresenius Kabi Norge, Halden, Norway). DNA was extracted from PBMCs using a QIAamp DNA Mini kit (VWR International, Stockholm, Sweden) or PURGENE DNA purification kit (Gentra Systems, Minneapolis, Minnesota, United States). B19 IgM and IgG were tested using a commercial EIA (Biotrin International, Dublin, Ireland).
Table 1 Data on Individuals Acutely and Remotely Infected with B19
Ethical approval for the study was obtained from the local ethical committee at Huddinge University Hospital, Karolinska Institutet, Stockholm, Sweden, and from the Oxfordshire Clinical Research Ethics Committee (CO2.113), Oxford, United Kingdom, and informed patient consent was obtained.
HLA Tissue Typing
HLA class I genotyping was performed using multiplex PCR on DNA extracted from PBMCs (ABC SSP Unitray, Dynal Biotech, Oslo, Norway).
Generation of HLA–Peptide Multimeric Complexes
HLA–B19 peptide multimeric complexes were constructed as previously described [10] and are listed in Table 2. Briefly, recombinant β2 microglobulin and HLA heavy chain (modified by deletion of the transmembrane–cytosolic tail and addition of a BirA enzymatic biotinylation site) were expressed in BL21pLysS Escherichia coli. The resulting inclusion bodies were purified and solubilized in urea, and refolded by limiting dilution. Peptides were purchased (Biopeptide, San Diego, California, United States) or synthesized by F-moc chemistry and were of greater than 95% purity. The refolded complex was then concentrated, buffer exchanged, biotinylated with BirA enzyme, and purified by fast protein liquid chromatography. HLA class I peptide complexes were stored at −80 °C. Before staining, the monomers were tetramerized with phycoerythrin or Alexa 647–labeled streptavidin (Molecular Probes, Eugene, Oregon, United States) at a molar ratio of 1:4. In addition, a Pro5 Pentamer containing the HLA-B*40 TEADVQQWL peptide conjugated to allophycocyanin was purchased (ProImmune, Oxford, United Kingdom). Specificity of the MHC monomers was confirmed using T cell lines established in vitro. When tested against non-HLA-matched cytotoxic T lymphocyte (CTL) lines, no population was detected (<0.02%; data not shown).
Table 2 HLA Restriction and Peptide Sequence of B19 Epitope Multimers Used
FACS Analysis
Cryopreserved PBMCs were thawed and washed twice with RPMI-1640 supplemented with 10% fetal calf serum, L-glutamine, penicillin, streptomycin, and Hepes buffer at pH 7.5. In preliminary experiments, freshly isolated PBMCs were stained in parallel with the same results. PBMCs (2.5 × 105 cells) were stained with the respective major histocompatibility complex multimer and incubated for 20–30 min at 37 °C. After two washes with PBE (2 mM EDTA and 0.05% BSA, in PBS [pH 7.4]), cells were co-stained with the appropriate monoclonal antibodies for 15 min on ice, and fixed in 1%–2% formaldehyde. Monoclonal antibodies used were directly conjugated and purchased ( Becton-Dickinson, Stockholm, Sweden). Four-colour FACS was performed using fluorochrome-coupled anti-human CD3-, CD8-, CD27-, CD28-, CD38-, CD45RA-, CD45RO-, CD57-, CD62L-, CCR7-, and perforin-specific antibodies. For perforin staining, the cells were permeabilized for 15 min using permeabilizing solution (Becton Dickinson, Palo Alto, California, United States) before staining with perforin monoclonal antibody. Cell acquisition was performed with a four-colour FACS by using a FACSCalibur with CellQuest software (Becton Dickinson).
T Cell Lines and Functional Assays
PBMCs were pulsed with 50 μM of the respective epitope and cultured at 2 × 106 cells/ml in 24-well plates for 12–18 d. At day three, 10 units/ml of IL-2 was added. Half of the medium was replaced each third day with fresh medium containing 10 units/ml of IL-2.
Ex Vivo IFNγ ELISpots
IFNγ ELISpots were performed as described previously [11]. Briefly 2.5 × 105 PBMCs were stimulated in triplicates with peptide pools/PHA (Sigma, St. Louis, Missouri, United States). Synthetic peptides, 15-mer, overlapping by ten amino acids spanning the entire B19 protein sequence were used in pools of ten at a final concentration of 10 μM. IFNγ responses were confirmed using individual peptides, with optimal epitopes and HLA restriction identified by synthesis of truncated peptides, prepulsing, and extensive washing of HLA-matched and -mismatched target cells, and synthesis of HLA–peptide multimeric complexes [12]. Using these techniques we confirmed previous epitopes in two independent cohorts, and identified two new epitopes shown in Table 2, the nonamer FYT restricted by HLA-A*24 and FPG restricted by HLA-B*35. The latter is in addition to the HLA-B*35 epitope described previously [5].
Intracellular Cytokine Staining
Short-term stimulation of PBMCs was carried out using either peptide pools or individual peptides as previously described [13]. Epitope HLA restriction required the prepulsing of matched and mismatched PBMCs (from B19 seronegative individuals) with peptide at 20 μM for 1 h at 37 °C. Cells were washed and added to CTLs at a 10:1 ratio for 1 h before the addition of Brefeldin A.
Chromium Release Assay
CTLs were set up as described above. The cytolysis was performed by killing of chromium-labelled targets as described previously by Nixon et. al. [14] using LBL721.220 transfected with HLA-A*0201 as target cells.
Results
B19-Specific CD8+ T cells Expand and Persist at High Frequency Following Resolution of Acute Symptomatic Infection
Eleven adults with acute B19 infection (five in Stockholm cohort and six in Oxford cohort) and five remotely infected seropositive individuals were studied. The clinical details, symptom duration, and HLA types are shown in Table 1. Surprisingly, in both cohorts (studied independently) the frequency of B19-specific (multimer-positive) CD8+ T cells in peripheral blood samples continued to increase for many months following symptom resolution. Figure 1A shows representative staining from patient O3. The frequency of A24 FYT tetramer-staining CD8+ T cells increases up to 15 mo after symptom presentation. Patient O3 presented with rash and arthritis; however, symptoms resolved within 5 wk and clinical examination was entirely normal at the time of the second and subsequent venesections. Figure 1B and 1C show the levels of multimer-positive CD8+ T cells for the Oxford and Stockholm cohorts of acutely infected individuals over time. Only one sample was obtained for patient O6; all other individuals were studied 2–7 times. All acutely infected individuals showed B19-specific CD8+ T cell percentages ranging from 0.09% to 4.5% total CD8+ T cells. The levels rose for at least the first 4 mo and frequently persisted for 12–32 mo after symptom onset. The frequencies at 22 mo (or nearest time point) were significantly greater than at first sampling (Willcoxon rank sum test, p = 0.0020), and than those observed for remotely infected seropositive individuals (Figure 1B; Mann–WhitneyU test, p = 0.0022).
Figure 1 B19-Specific CD8+ T Cells Persist at High Levels for Many Months after Acute Infection
(A) Representative A24 FYT tetramer staining of individual O3\′s PBMCs. Plots are gated on live CD8+ lymphocytes stained directly ex vivo. Percentages shown are those of tetramer-positive CD8+ T cells. Time points indicated refer to the number of months after first symptoms reported. Symptoms in this individual lasted 5 wk.
(B) Frequency of B19-specific responses over time for six acutely infected individuals in the Oxford cohort (O1–O5) and five remotely infected individuals (OR1–OR5). In one case, two epitopes were studied.
(C) Frequency of B19-specific responses over time for five acutely infected individuals in the Stockholm cohort (S1–S5). In two cases, two epitopes were studied.
B19 DNA could be detected by nested PCR in serum for different time periods (see Table 1). No B19 DNA could be amplified from individuals in the remotely infected group. IgG levels rose quickly and were maintained at a stable level in patients studied (data not shown).
B19-Specific CD8+ T Cells Remain Activated Following Resolution of Acute Symptomatic Infection
In order to understand the origins and potential pathogenic role of virus-specific CD8+ T cells during and after acute infection, we studied the activation phenotype of these cells ex vivo. Figure 2 shows examples of T cell phenotypic marker expression at early and late sample time points for three different patients using three different B19 peptide–HLA multimers. Figure 2A shows that HLA-B*40 TEA pentamer-positive CD8+ T cells from individual S2 are largely perforin positive and CD62L negative, and that these changes become more pronounced from the early (4 mo) to late (21 mo) time points. Figure 3 shows sequential phenotypic data for all of the acutely and remotely infected Oxford cohort. The upper left panel shows that all acutely infected patients showed increases in perforin expression on their B19-specific T cells over the first year following symptom development. By contrast, the seropositive remotely infected individuals showed low levels of perforin expression on tetramer-positive cells. These differences were statistically significant (Mann–Whitney U test, p = 0.0079). Figure 4 shows sequential data from two individuals, S1 and S2, of the Stockholm cohort, each studied with two different HLA–peptide multimers at five or six time points after acute presentation.
Figure 2 Representative Ex Vivo Phenotyping of B19-Specific CD8+ T Cell Populations over Time
Percentages shown are the frequency of marker-positive cells amongst multimer-positive cells (plots gated on live CD8+ lymphocytes ex vivo).
(A) Patient S2 B40 TEA pentamer staining at 4 mo (left) and 21 mo (right), showing perforin and CD62L levels.
(B) Patient O3 A24 FYT tetramer staining at 2 mo (left) and 20 mo (right), showing CD27 and CD28 staining.
(C) Patient O5 A2 GLC tetramer staining at 4 mo (left) and 18 mo (right), showing CD38 and CD57 staining.
Figure 3 Acutely infected individuals Maintain Activated Mature Effector Cells after Resolution of Acute Infection
Remotely infected individuals show a less mature/activated phenotype but express only low levels of CCR7 and CD62L (Oxford cohort). The y-axis show the frequency of marker-positive cells amongst tetramer-positive CD8+ cells, while the x-axis show the number of months after symptom onset. Data are shown in perforin, CD38, CD57, CD27, CD28, and CCR7: data are derived from six acutely infected individuals in the Oxford cohort (O1–O3, O5, and O6; insufficient cells available for O4) and five remotely infected individuals (OR1–OR5).
Figure 4 Longitudinal Phenotype Analysis of B19-Specific CD8+ T Cell Responses in Acutely Infected Individuals in Stockholm Cohort
(A) Blood from patient S1 was stained with the two tetramers A2 LLH and A2 GLC.
(B) Blood from patient S2 was stained with tetramers A2 GLC and the Pro5 B40 TEA pentamer. The top panels show the frequency of tetramer-positive cells over time for the two different responses in each individual (data equivalent to those in Figure 1C). The subsequent panels show the frequency of B19-specific CD8+ cells positive for perforin, CD38, CD57, and CD62L in both patients.
CD62L expression was found to be low at all time points in the acutely infected cohort as well as in the remotely infected cohort (see Figure 2A, lower panels; Figure 4, bottom panels; data not shown). Figure 3 also shows that, for the Oxford acutely infected cohort, CCR7 expression was low and fell over the study period. B19-specific CD8+ T cells from four of the five remotely infected individuals had low CCR7 expression (see Figure 3, bottom right panel).
Figure 2B (upper panels) shows the frequency of CD27 expression on A24 FYT tetramer-positive T cells for patient O3. Figure 3 (upper right panel) shows the data from all of the Oxford cohort. CD27 was downregulated over time in acutely infected individuals. Three of the remotely infected individuals had high levels of CD27 expression on their tetramer-positive populations, while two had slightly lower levels, with only approximately 50% of tetramer-positive cells expressing CD27. B19-specific T cells from the acutely infected individuals also showed downregulation of the CD28 marker over time. This is shown for individual O3 in Figure 2B (bottom panels). Two months after symptom onset, 86% of multimer-positive cells expressed CD28, while at 20 mo only 4% of multimer-positive cells expressd CD28. Figure 3 (middle right panel) shows a similar pattern of CD28 downregulation over time for multimer-positive cells from the Oxford acutely infected cohort. Similar results were seen for the Stockholm cohort (data not shown). By contrast, B19-specific CD8+ T cells from remotely infected individuals were predominately CD28+. The T cells 20 mo post-infection in acutely infected individuals expressed significantly lower CD28 than in the remotely infected cohort (Mann–Whitney U test, p = 0.015).
Persistent CD38 Expression Following B19 Infection Suggests Ongoing Low-Level Antigenic Stimulation
All acutely infected individuals maintained high levels of CD38 expression for more than 10 mo after symptom onset, as shown for the Oxford cohort in Figures 2C and 3 and for two individuals of the Stockholm cohort in Figure 3. Strikingly, for one patient, O3, there was a dramatic fall in CD38 expression at 27 mo compared to the 20-mo time point. Figure 4 shows the sequential CD38 expression of multimer-positive T cells of two specificities for Stockholm patients S1 and S2. S1 shows fluctuations over a period of over 32 mo; S2 shows a gradual decline in CD38 expression from high levels over a comparable period.
CD57 expression levels increased over time in almost all acutely infected individuals. This is shown for one Oxford individual in Figure 2C, for the Oxford acutely infected cohort in Figure 3 (bottom left panel), and for two of the Stockholm cohort in Figure 4. Thus, for individual S2 (Figure 4) a high level of CD57 expression was maintained at 30 mo, the last time point tested. By contrast, Figure 3 (bottom left panel) shows that remotely infected individuals have low levels of CD57 expression on their B19-specific CD8+ T cells. T cells 20 mo post-infection in acutely infected individuals expressed significantly higher CD57 than in the remotely infected cohort (Mann–Whitney U test, p = 0.0043).
Differing Patterns of CD45 Isoform Expression Follow Acute B19 Infection
CD45 isoform expression showed variation between individuals. Most acutely infected patients from the Stockholm and Oxford cohorts expressed high levels of CD45RO within a few months of symptom onset, with variable downregulation over time (data not shown). In the remotely infected cohort, three individuals expressed high levels of CD45RO on tetramer-positive cells. Two of these individuals, however, had only just over 50% of their tetramer-positive cells expressing CD45O (data not shown). CD45RA was downregulated in some patients in the Oxford cohort, whereas in the Stockholm cohort the B19-specific CD8+ cells during the entire study period generally expressed high frequency of CD45RA (data not shown).
B19-Specific CD8+ T Cells Have Efficient Effector Function
Figure 5A shows that PBMCs from patient O1 at the 2-mo time point are capable of immediate IFNγ release on stimulation with the FYTPLADQF peptide. The number of IFNγ-releasing T cells corresponds to approximately 0.3% of CD8 T cells (since at this time point, 40% of PBMCs were CD8+). At this time point, 1.6% of CD8+ cells were stained with the A24 FYT tetramer (Figure 5A, right panel). Thus, the numbers of cells capable of immediate IFNγ secretion were approximately 5-fold lower than the numbers of tetramer-staining cells. Figure 5B shows that B40-restricted TEA-specific T cells from patient S2 (at 30 mo) were able to proliferate rapidly in vitro on stimulation with peptide. After 12 d, 90% of surviving CD8+ T cells were B40 TEA-specific. B19-specific T cell lines were capable of rapid IFNγ production, as assessed by intracellular cytokine release assay (data not shown).
Figure 5 B19-Specific CD8+ T Cells Secrete IFNγ Ex Vivo, Proliferate, and Show Cytolytic Function In Vitro
(A) Left panel shows that PBMCs from acutely infected patient O1 secrete IFNγ ex vivo after 18 h of FYTPLADQF peptide stimulation. Negative control (zero spots) and two peptide-stimulated wells from an ELISpot plate are shown. Numbers represent IFNγ-secreting cells per 250,000 PBMCs. Right panel shows A24 FYT tetramer staining of PBMCs at same time point, displaying the number of tetramer-positive cells expressed as a percentage of CD8+ T cells.
(B) Tetramer staining of patient S2\′s PBMCs ex vivo (left) and after short-term TEADVQQWL peptide stimulation in vitro (right).
(C) Ex vivo IFNγ ELISpot results for remotely infected individual OR3. Mean and standard deviations of triplicates are shown. Cells were stimulated for 18 h with no peptide, GLCPHCINV, or TEADVQQWL.
(D) 51Cr release assay using HLA-A2-restricted GLCPHCINV-specific CTLs from individual OR1. PBMCs were stimulated for 14 d with GLCPHCINV peptide and cytolysis was tested against HLA-A*0201-transfected LBL.721.220 target cells at various effector-to-target-cell (E:T) ratios.
Lastly, we studied the function of B19-specific CD8+ T cells from remotely infected individuals (who in all cases had no symptoms attributable to possible B19 infection for at least 5 y). Figure 5C shows that these cells were capable of direct ex vivo IFNγ release on 18 h culture with GLCPHCINV or TEADVQQWL peptides. A T cell line derived by culture of OR1 PBMCs with peptide in vitro was capable of HLA-A*0201-restricted B19-peptide-specific cytolysis (Figure 5D).
Discussion
In this study we observed, to our knowledge for the first time, a striking pattern of evolving, long-lived CD8 immune responses against B19 in 11 adults with primary B19 infection. Although the symptoms of this virus are generally short-lived and the virus is not classically regarded as persistent, the immune responses showed a sustained activated state many months after initial infection. This pattern was observed in a range of patients in two different clinical centres and appears to represent a new and distinct style of host–virus relationship. This is the first example to our knowledge of an “acute” human viral infection inducing persistent activated CD8 T cell responses.
The CD8+ T cell responses tracked here were mapped using a comprehensive screening system facilitated by the compact and stable viral genome of B19. The virus has only one NS gene, which appears to be the major target of CD8+ T cell responses during acute disease, with a range of epitopes identified [5,12]. HLA-A2-restricted epitopes were commonly targeted, but, interestingly, no clear-cut dominance of one over the other was consistently seen, in contrast to infections such as CMV and HIV [15,16]. CTL responses to HLA-A2 and non-HLA-A2 epitopes showed similar kinetics, frequencies, and phenotypes. Thus, future studies might reasonably track defined epitopes, rather than requiring individual mapping, as is the case in, for example, HCV [17–19].
We used, to our knowledge for the first time, HLA–peptide multimeric complexes to detect CD8+ T cell responses during acute B19 infection. Surprisingly, these continued to increase in magnitude at later time points, long after resolution of acute symptoms. In some cases, these responses reached high levels in blood, and were sustained over many months. Even responses of lower frequency appeared to show this delayed expansion. This differs to responses to almost all other human viruses studied in such detail. Responses to HIV and HCV are strong in acute infection, but typically decline as virus is controlled [20]. EBV-specific responses to latent antigens may increase over time [21]. Few data exist on acute CMV infection in immunocompetent humans, but in murine infection, a phenomenon of “memory inflation” is seen for some but not all epitopes [22–24]. Here, responses showed gradual accumulation over time after a short acute response and a lag period of about 8–10 wk after acute infection. The current study did not have the resolution to determine whether B19-specific responses were biphasic in this manner, although in the mouse such phenomena may differ substantially between different virus preparations, doses, and experimental settings.
The B19-specific T cell populations underwent contraction 1.5–2 y after acute infection. The kinetics of this contraction were not defined in this study, but “remotely infected” individuals, who had no recent history of infection, and in some cases may have been primarily infected as long ago as 30 y previously, showed smaller populations of B19-specific T cell populations . However, as we have noted previously, B19-specific CTLs are readily detectable—often much larger than equivalent responses to viruses such as influenza and comparable to some EBV-specific responses [5]. A single individual who had a documented B19 infection 10 y previously showed populations of a size and phenotype (see below) similar to the other remotely infected persons.
In addition to a sustained and prolonged expansion of antiviral responses, we also observed continued maturation of B19-specific CD8+ T cells in the cohorts of acutely infected individuals studied. This expansion was consistent across a range of markers, all of which have been linked to the evolution of antiviral “effector” memory cells against persistent virus infections. Consistent with these data is the finding of sustained CD8 effector function of the Swedish cohort over time, as evidenced by IFNγ production in response to viral peptides [12]. Antiviral T cell responses to CMV have been extensively studied and are typically regarded as exhibiting a “mature” phenotype associated with loss of expression of co-stimulatory molecules CD27 and CD28, sustained loss of lymph node homing markers CD62L and CCR7, and generally positive expression of intracellular perforin [7,25–27]. A particular feature of these cells is sustained expression of CD57, which is considered to be a marker of terminally differentiated cells [20]. Although there are substantial differences between individuals, many groups report re-expression of CD45RA on such highly differentiated cells [28–31]. All of these features were clearly reproduced in the B19-specific responses tracked in the months following acute infection. Interestingly, the gradual evolution of responses from a CD28+ to CD28− status, in concert with changes in other markers, could be clearly tracked over time. These phenotypic changes have not been extensively investigated in human CMV, but in murine CMV, CD28 loss does appear to occur relatively early, and to be subsequently maintained in the immunodominant populations [32].
It is generally considered likely, although not proven, that such marker evolution represents a maturation pathway, driven by restimulation in vivo with antigen. The nature or duration of the encounter, coupled with the survival of antigen-specific cells, may lead to the typical appearances of T cells specific for different persistent virus infections [7,25]. The findings that not only are the cells phenotypically mature but also strongly activated in vivo (CD38+), is consistent with continuous encounter with antigen over the post-infection period. This period of restimulation appears to be sustained, but, unlike in CMV, appears to wane over time, perhaps after 1.5–2 y. Eventual disappearance of antigen would be consistent with the relatively less differentiated phenotype seen in remotely infected individuals. Nevertheless, although such populations are smaller in size, less activated, and less mature, they retain a CD62L-low phenotype. Murine T cell populations that exist in the absence of antigen typically show slow reversion to a CD62L-high state [33], even in the case of CMV [32]. Thus antigenic drive may be reduced but would appear to be still sufficient to maintain an “effector” memory T cell population.
The striking features of the T cell responses to B19 infection indicate persistence of antigen long after the resolution of acute infection. The status of the virus in the post-acute period is not fully understood. Direct nucleic acid analysis has revealed loss of detection in blood after 3 wk in infected volunteers [3], although this may be prolonged in some cases [4]. Nested PCR analysis in our study revealed the presence of B19 DNA in blood at early time points during acute infection, but such assays were negative at time points 6–12 mo after infection, when T cell populations remained activated. It is possible that B19 persists, in the blood and other sites including bone marrow, joints, or skin. Detection of B19 DNA in the bone marrow of asymptomatic remotely infected volunteers has been reported [3]. The finding of parvovirus in skin remains restricted to a single study of a specific genotype distinct from the genotype 1 strains found in our study participants [34]. Low-level, contained replication at a tissue site for weeks or months after infection does seem like the most likely explanation for the immune responses seen, and more sensitive quantitative PCR assays are being validated to address this question. It is also possible that viral antigen is retained, for example, on follicular dendritic cells, following the extremely high burden seen acutely. Alternatively subgenomic particles may be generated in the post-acute period, in the absence of full viral replication. Parvovirus is not thought to establish true latency or integrate into the host genome, so the mechanism behind this low-level persistence remains to be explored.
The relationship between joint or bone marrow pathology and the T cell responses seen is not clear. Indeed, since the most active CD8+ T cell responses were seen at stages where joint symptoms had resolved, these T cell responses are unlikely to be directly involved. It remains an open and interesting question, however, whether these or perhaps CD4+ T cells are involved in the prolonged inflammatory arthritis seen in a proportion of cases. A previous study did identify an HLA association with B19-induced arthritis syndrome, suggesting a significant role for cellular immune responses [1]. All our patients presented—as is common in adults—with joint symptoms, and in principle an acutely infected but asymptomatic group would provide an ideal comparison to address such a question. It is also possible that the CTL responses seen might be involved in immune-mediated pathology in the bone marrow. Although the virus itself may lead to direct death of the critical progenitor cells, in other settings vigorous T cell responses can also contribute importantly to marrow suppression through lytic and non-lytic pathways [35].
In summary, this is the first demonstration to our knowledge of a virus not thought to cause true or classical persistent infection leading to a persistent activated CD8+ T cell response. Responses of this quantity and quality (i.e., CD27 and CD28 low and CD57 and perforin high) have only previously been seen for infection with CMV, a virus known to establish persistent infection. Our data suggest that B19 persists in some form after acute infection, and provokes sustained activated CD8+ T cell responses, which might ultimately play a role in viral clearance. Defining the role of CTLs in this setting will be of value not only in expanding further our understanding of the role of T cells in acute and persistent viral infections but also in vaccine design and in immunotherapy, as has been applied to treatment of EBV and CMV infections in immunosuppressed individuals. B19 is a small virus and attracts relatively little attention from clinicians and immunologists [36], but clearly attracts a great deal of attention from the immune system. Now that progress has been made in the definition of the kinetics and antigenic targets of these immune responses, further studies in specific clinical settings where the virus remains a significant problem could readily redress this balance.
Patient Summary
Background
Parvovirus B19 is a small virus that is very common. Usually it causes mild symptoms of fever and a typical “slapped cheek” rash on the face, which may then spread; in around one in 20 children, and one in two adults it can also cause joint pain. Other, more serious complications occur in people whose red cells do not last as long as usual, for example, people with sickle cell disease; these people can get a severe aplastic anemia—the bone marrow stops making blood completely for a time. The virus can also cause fetal death if a woman contracts it while pregnant.
Why Was This Study Done?
It is not clear how the body's defenses—immune system—work to clear this virus. There is some evidence that in some people the virus might continue to divide in the body for a long time after the first infection. The authors wanted to study one part of the immune system—T cells—in people who had just recently been infected with the virus and compare the findings with people who had had the infection a long time previously.
What Did the Researchers Do and Find?
They compared findings from two groups of people: 11 who had recently had the infection (six from Oxford and five from Stockholm) and five who had had the virus many years previously. They found that over the year following the infection, one particular type of T cell continued to increase in numbers and responsiveness to the B19 virus, despite the fact that the patients' clinical symptoms had gotten better.
What Do These Findings Mean?
It seems the virus remained in the patients' bodies for a considerable time after they appeared to have recovered, and the virus continued to stimulate T cells to respond to it. These results may be useful in designing a strategy to develop a vaccine for this virus.
Where Can I Get More Information Online?
The Health Protection Agency in the United Kingdom has a Web page of information on parvovirus:
http://www.hpa.org.uk/infections/topics_az/parvovirus/gen_info.htm
MedlinePlus also has links to further information:
http://www.nlm.nih.gov/medlineplus/ency/article/000977.htm
This study was financially supported by the Medical Research Council UK, the Tobias Foundation, the Swedish Cancer Foundation, and the specific Programme for Research and Technological Development “Quality of Life and Management of Living Resources, Human Parvovirus Infection: Towards Improved Understanding Diagnosis and Therapy” (QLK2-CT-2001–00877) of the Swedish Medical Research Council, the Wellcome Trust, and the Commission of the European Communities. However, the study does not necessarily reflect the views of these funders and in no way anticipates the European Commission's future policy in this area. We are also grateful to Mr. Tim Rostron for assistance with tissue typing. We would also like to thank the patients and volunteers who donated blood for the study. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Citation: Isa A, Kasprowicz V, Norbeck O, Loughry A, Jeffery K, et al (2005) Prolonged activation of virus-specific CD8+ T cells after acute B19 infection. PLoS Med 2(12): e343.
Abbreviations
B19human parvovirus B19
CMVcytomegalovirus
CTLcytotoxic T lymphocyte
EBVEpstein–Barr virus
HCVhepatitis C virus
HLAhuman leukocyte antigen
IFNγinterferon-γ
PBMCperipheral blood mononuclear cell
==== Refs
References
Gendi NS Gibson K Wordsworth BP Effect of HLA type and hypocomplementaemia on the expression of parvovirus arthritis: one year follow up of an outbreak Ann Rheum Dis 1996 55 63 65 8572737
van Elsacker-Niele AM Kroes AC Human parvovirus B19: Relevance in internal medicine Neth J Med 1999 54 221 230 10399450
Heegaard ED Brown KE Human parvovirus B19 Clin Microbiol Rev 2002 15 485 505 12097253
Lundqvist A Tolfvenstam T Bostic J Soderlund M Broliden K Clinical and laboratory findings in immunocompetent patients with persistent parvovirus B19 DNA in bone marrow Scand J Infect Dis 1999 31 11 16 10381211
Tolfvenstam T Oxenius A Price DA Shacklett BL Spiegel HM Direct ex vivo measurement of CD8(+) T-lymphocyte responses to human parvovirus B19 J Virol 2001 75 540 543 11119624
Sallusto F Lenig D Forster R Lipp M Lanzavecchia A Two subsets of memory T lymphocytes with distinct homing potentials and effector functions Nature 1999 401 708 712 10537110
Appay V Dunbar PR Callan M Klenerman P Gillespie GM Memory CD8+ T cells vary in differentiation phenotype in different persistent virus infections Nat Med 2002 8 379 385 11927944
Altman JD Moss PA Goulder PJ Barouch DH McHeyzer-Williams MG Phenotypic analysis of antigen-specific T lymphocytes Science 1996 274 94 96 8810254
Kern F Khatamzas E Surel I Frommel C Reinke P Distribution of human CMV-specific memory T cells among the CD8pos. subsets defined by CD57, CD27, and CD45 isoforms Eur J Immunol 1999 29 2908 2915 10508265
Ogg GS McMichael AJ HLA-peptide tetrameric complexes Curr Opin Immunol 1998 10 393 396 9722914
Lalvani A Brookes R Hambleton S Britton WJ Hill AV Rapid effector function in CD8+ memory T cells J Exp Med 1997 186 859 865 9294140
Norbeck O Isa A Pohlmann C Broliden K Kasprowicz V Sustained CD8+ T-cell responses induced after acute parvovirus B19 infection in humans J Virol 2005 79 12117 12121 16140790
Klenerman P Phillips RE Rinaldo CR Wahl LM Ogg G Cytotoxic T lymphocytes and viral turnover in HIV type 1 infection Proc Natl Acad Sci U S A 1996 93 15323 15328 8986810
Nixon DF Townsend AR Elvin JG Rizza CR Gallwey J HIV-1 gag-specific cytotoxic T lymphocytes defined with recombinant vaccinia virus and synthetic peptides Nature 1988 336 484 487 2461519
Goulder PJ Sewell AK Lalloo DG Price DA Whelan JA Patterns of immunodominance in HIV-1-specific cytotoxic T lymphocyte responses in two human histocompatibility leukocyte antigens (HLA)-identical siblings with HLA-A*0201 are influenced by epitope mutation J Exp Med 1997 185 1423 1433 9126923
Jin X Demoitie MA Donahoe SM Ogg GS Bonhoeffer S High frequency of cytomegalovirus-specific cytotoxic T-effector cells in HLA-A*0201-positive subjects during multiple viral coinfections J Infect Dis 2000 181 165 175 10608763
Lauer GM Barnes E Lucas M Timm J Ouchi K High resolution analysis of cellular immune responses in resolved and persistent hepatitis C virus infection Gastroenterology 2004 127 924 936 15362047
Lauer GM Nguyen TN Day CL Robbins GK Flynn T Human immunodeficiency virus type 1-hepatitis C virus coinfection: Intraindividual comparison of cellular immune responses against two persistent viruses J Virol 2002 76 2817 2826 11861849
Day CL Seth NP Lucas M Appel H Gauthier L Ex vivo analysis of human memory CD4 T cells specific for hepatitis C virus using MHC class II tetramers J Clin Invest 2003 112 831 842 12975468
Lechner F Wong DK Dunbar PR Chapman R Chung RT Analysis of successful immune responses in persons infected with hepatitis C virus J Exp Med 2000 191 1499 1512 10790425
Callan MF Tan L Annels N Ogg GS Wilson JD Direct visualization of antigen-specific CD8+ T cells during the primary immune response to Epstein-Barr virus In vivo J Exp Med 1998 187 1395 1402 9565632
Karrer U Wagner M Sierro S Oxenius A Hengel H Expansion of protective CD8+ T-cell responses driven by recombinant cytomegaloviruses J Virol 2004 78 2255 2264 14963122
Karrer U Sierro S Wagner M Oxenius A Hengel H Memory inflation: Continuous accumulation of antiviral CD8+ T cells over time J Immunol 2003 170 2022 2029 12574372
Holtappels R Pahl-Seibert MF Thomas D Reddehase MJ Enrichment of immediate-early 1 (m123/pp89) peptide-specific CD8 T cells in a pulmonary CD62L(lo) memory-effector cell pool during latent murine cytomegalovirus infection of the lungs J Virol 2000 74 11495 11503 11090146
van Lier RA ten Berge IJ Gamadia LE Human CD8(+) T-cell differentiation in response to viruses Nat Rev Immunol 2003 3 931 939 14647475
Zhang D Shankar P Xu Z Harnisch B Chen G Most antiviral CD8 T cells during chronic viral infection do not express high levels of perforin and are not directly cytotoxic Blood 2003 101 226 235 12393740
Moss P Khan N CD8(+) T-cell immunity to cytomegalovirus Hum Immunol 2004 65 456 464 15172445
Gillespie GM Wills MR Appay V O'Callaghan C Murphy M Functional heterogeneity and high frequencies of cytomegalovirus-specific CD8(+) T lymphocytes in healthy seropositive donors J Virol 2000 74 8140 8150 10933725
Geginat J Lanzavecchia A Sallusto F Proliferation and differentiation potential of human CD8+ memory T-cell subsets in response to antigen or homeostatic cytokines Blood 2003 101 4260 4266 12576317
Marchant A Appay V Van Der Sande M Dulphy N Liesnard C Mature CD8(+) T lymphocyte response to viral infection during fetal life J Clin Invest 2003 111 1747 1755 12782677
Mizobuchi T Yasufuku K Zheng Y Haque MA Heidler KM Differential expression of Smad7 transcripts identifies the CD4+CD45RChigh regulatory T cells that mediate type V collagen-induced tolerance to lung allografts J Immunol 2003 171 1140 1147 12874199
Sierro S Rothkopf R Klenerman P Evolution of diverse antiviral CD8(+) T cell populations after murine cytomegalovirus infection Eur J Immunol 2005 35 1113 1123 15756645
Barber DL Wherry EJ Ahmed R Cutting edge: Rapid in vivo killing by memory CD8 T cells J Immunol 2003 171 27 31 12816979
Hokynar K Soderlund-Venermo M Pesonen M Ranki A Kiviluoto O A new parvovirus genotype persistent in human skin Virology 2002 302 224 228 12441066
Binder D Fehr J Hengartner H Zinkernagel RM Virus-induced transient bone marrow aplasia: Major role of interferon-alpha/beta during acute infection with the noncytopathic lymphocytic choriomeningitis virus J Exp Med 1997 185 517 530 9053452
Riddell SR Greenberg PD T cell therapy of human CMV and EBV infection in immunocompromised hosts Rev Med Virol 1997 7 181 192 10398482
Shade RO Blundell MC Cotmore SF Tattersall P Astell CR Nucleotide sequence and genome organization of human parvovirus B19 isolated from the serum of a child during aplastic crisis J Virol 1986 58 921 936 3701931
|
16253012
|
PMC1274280
|
CC BY
|
2021-01-05 10:39:22
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e343
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020343
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 1757074910.1371/journal.pmed.0020345Research ArticleGenetics/Genomics/Gene TherapyDiabetes/Endocrinology/MetabolismEndocrinologyDiabetesGeneticsGenetic Prediction of Future Type 2 Diabetes Prediction of Type 2 DiabetesLyssenko Valeriya
1
*Almgren Peter
1
Anevski Dragi
1
2
Orho-Melander Marju
1
Sjögren Marketa
1
Saloranta Carola
3
4
Tuomi Tiinamaija
3
4
Groop Leif
1
the Botnia Study Group 1Department of Clinical Sciences, Diabetes and Endocrinology, Lund University, University Hospital Malmö, Malmö, Sweden,2School of Mathematical Sciences, Chalmers University of Technology, Gothenburg, Sweden,3Department of Medicine, Division of Diabetology, Helsinki University Hospital, Helsinki, Finland,4Folkhälsan Research Center, Institute of Genetics, Helsinki, FinlandHattersley Andrew Academic EditorPeninsular Medical School, ExeterUnited Kingdom*To whom correspondence should be addressed. E-mail: [email protected]
Competing Interests: LG is a member of the editorial board of PLoS Medicine.
Author Contributions: VL extracted, genotyped, and analyzed the data, and drafted the report. PA and DA were responsible for the statistical analyses, MOM and MS for genotyping, and CS and TT for the phenotype data. LG designed the study and supervised all parts of the work including drafting the final report. All researchers took part in the revision of the report and approved the final version.
12 2005 1 11 2005 2 12 e34522 11 2004 23 8 2005 Copyright: © 2005 Lyssenko et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.
Predicting the Development of Type 2 Diabetes
Background
Type 2 diabetes (T2D) is a multifactorial disease in which environmental triggers interact with genetic variants in the predisposition to the disease. A number of common variants have been associated with T2D but our knowledge of their ability to predict T2D prospectively is limited.
Methods and Findings
By using a Cox proportional hazard model, common variants in the PPARG (P12A), CAPN10 (SNP43 and 44), KCNJ11 (E23K), UCP2 (−866G>A), and IRS1 (G972R) genes were studied for their ability to predict T2D in 2,293 individuals participating in the Botnia study in Finland. After a median follow-up of 6 y, 132 (6%) persons developed T2D. The hazard ratio for risk of developing T2D was 1.7 (95% confidence interval [CI] 1.1–2.7) for the PPARG PP genotype, 1.5 (95% CI 1.0–2.2) for the CAPN10 SNP44 TT genotype, and 2.6 (95% CI 1.5–4.5) for the combination of PPARG and CAPN10 risk genotypes. In individuals with fasting plasma glucose ≥ 5.6 mmol/l and body mass index ≥ 30 kg/m2, the hazard ratio increased to 21.2 (95% CI 8.7–51.4) for the combination of the PPARG PP and CAPN10 SNP43/44 GG/TT genotypes as compared to those with the low-risk genotypes with normal fasting plasma glucose and body mass index < 30 kg/m2.
Conclusion
We demonstrate in a large prospective study that variants in the PPARG and CAPN10 genes predict future T2D. Genetic testing might become a future approach to identify individuals at risk of developing T2D.
In a large prospective study, Lyssenko and colleagues show that variants in the PPARG and CAPN10 genes can help predict whether a person will develop Type 2 diabetes.
==== Body
Introduction
Type 2 diabetes (T2D) is a multifactorial disease in which environmental triggers interact with genetic variants in the predisposition to the disease [1]. T2D is characterized by impaired insulin secretion and insulin action in target tissues such as muscle and liver [2]. Many patients with a genetic predisposition to T2D also have a predisposition to weight gain, and obesity is a strong risk factor for T2D [3].
Although several candidate genes have been associated with T2D [4–7], many findings have been difficult to replicate. The list of genes with support in extensive meta-analyses is relatively short, including genes encoding for PPARG, calpain 10, Kir 6.2, and insulin receptor substrate 1 (IRS1) [8]. The PPARG P12A polymorphism is associated with enhanced insulin sensitivity and protects against T2D [4,9–11]. Although the individual risk reduction for carriers of the rare A allele is only 15%, the population attributable risk of the common allele is about 25%. Two intronic single nucleotide polymorphisms (SNPs) (43 and 44) in the gene encoding for the cystein protease calpain 10 (CAPN10) confer increased susceptibility to insulin resistance and T2D [12–15]. The ATP-sensitive potassium channel Kir 6.2 (KCNJ11) forms together with the sulfonylurea receptor SUR1 (ABCC8), an octamer protein that regulates transmembrane potential and thereby glucose-stimulated insulin secretion in pancreatic β-cells. A E23K polymorphism in KCNJ11 has been associated with T2D [16–18]. Carriers of a G972R polymorphism in the IRS1 gene (IRS1) have been shown to have reduced insulin content in pancreatic islets [19]. Although the meta-analyses suggested a role for the G972R polymorphism in T2D [8,20,21], two recent large case-control studies failed to replicate this association [22,23].
In addition to the genes listed above, we considered it worthwhile to also include the uncoupling protein 2 gene (UCP2) in the analysis because some studies have associated a polymorphism in the promoter of the gene (UCP2 −866G>A) with increased risk of T2D and impaired insulin secretion [24–27], whereas other studies have reported reduced risk of T2D [28]. Increased expression of UCP2 in pancreatic islets is associated with increased uncoupling and thereby decreased ATP production required for insulin secretion [29].
In this study, we tested variants in a number of candidate genes for T2D for their ability to predict diabetes in 2,293 individuals without diabetes participating in the Botnia prospective study in western Finland.
Methods
Study Participants
The Botnia study is a family-based study aiming to identify genes increasing susceptibility to T2D [30,31]. Details of the study cohort and sampling strategy have been presented earlier [31]. In brief, individuals with T2D from the area of five health-care centers in western Finland were invited to participate, together with their family members [31]. An oral glucose tolerance test (OGTT) was performed for all participants aged 18–70 y who had fasting plasma glucose concentration (FPG) lower than 11 mmol/l. Participants without diabetes, either family members of T2D patients or control participants (spouses without first or second degree family history of diabetes), between 18–70 y were offered prospective visits every 2–3 y. During the study period (which started in 1990 and was closed for this analysis in 2002), 1,869 relatives of T2D patients from 577 extended pedigrees (approximately three persons per pedigree) and 424 controls without family history of diabetes participated in at least OGTTs with a median follow-up of 6 y (range 2–12 y). Of the participants in both these groups, 1,569 had normal glucose tolerance and 724 had impaired fasting glucose and/or impaired glucose tolerance at baseline. Carriers of mutations causing maturity onset diabetes of the young (n = 20) were excluded from the present study. Glucose tolerance was defined according to the current World Health Organization criteria [32]. All participants gave informed consent, and the local ethics committee approved the study.
Anthropometric Measurements and Assays
The participants' weight, height, waist and hip circumference, and blood pressure were measured as previously reported [30,31]. Body mass index (BMI) was calculated as weight (in kilograms) divided by height (in meters) squared. All participants participated in a 75-g OGTT after a 12-h overnight fast. Fasting blood samples were drawn for the measurement of high density lipoprotein cholesterol, triglyceride, and free fatty acid concentrations, and at −10, 0, 30, 60, and 120 min for the measurement of plasma glucose and serum insulin. Insulin resistance was estimated as homeostasis model assessment index (HOMAIR) using a computer-based model [33] and β-cell function as the ratio of incremental insulin to glucose responses during the first 30 min of the OGTT (ΔI/ΔG = ΔI30 min fasting/ΔG30 min fasting); this index is also called the insulinogenic index. The disposition index was used to adjust insulin secretion for the degree of insulin resistance (insulinogenic index/HOMAIR).
Genotyping
Genotyping of SNPs was performed with a polymerase chain reaction–restriction fragment length polymorphism (PCR-RFLP) method and agarose gel electrophoresis for IRS1 G972R, or using the Multiplex SNaPshot kit (Applied Biosystems, Stockholm, Sweden) for single base pair extension on ABI 3100 (Applied Biosystems) for CAPN10 SNP43 and 44, or with an allelic discrimination assay-by-design method on ABI 7900 (Applied Biosystems) for KCNJ11 E23K, IRS1 G972R, and UCP2 −866G>A (Table S1). By randomly regenotyping 10%–20% of the samples, we achieved concordance rates of 99% for PPARG P12A, CAPN10 SNP44, KCNJ11 E23K, IRS1 G972R, and UCP2 −866G>A, and 98% for CAPN10 SNP43. All genotypes were in Hardy–Weinberg equilibrium except CAPN10 SNP44, which showed a moderate deviation from equilibrium (p = 0.035). Genotyping errors are an unlikely explanation for this deviation from equilibrium, because in genotyping 1,880 samples of CAPN10 SNP44 using two different methods (allelic discrimination and single base extension) the concordance rate was 99%.
Statistical Analyses
Variables are presented as median (interquartile range). A χ2 test was used for comparison of group frequencies. Survival analysis was used to estimate the effect of genetic variants (risk and non-risk genotypes defined from previous studies) on the risk of developing T2D and shown by Kaplan–Meier survival curves as the distribution of age at onset (the proportion of individuals developing T2D at certain age). The risk of developing T2D was expressed as a hazard function (the negative slope divided by the survival curve) using an age-adjusted Cox proportional hazard regression model [34]. The hazard function is the (conditional) probability for the development of diabetes during a time interval divided by the length of that time interval, for an individual that is diabetes-free at the start of the time interval. The relative effect is presented as the ratio between the hazard functions (hazard ratio [HR]) of the two groups. HRs quantify the effect size of both discrete variables (carriers versus non-carriers) and continuous variables (used in the interaction analysis below where HR measures the effect of an increase in one unit of the continuous variable). All survival analyses were stratified for gender and adjusted for family history of diabetes and BMI (when appropriate). The information that an individual did not or did belong to a nuclear family with at least one other affected member was coded as zero or one, respectively, and used as a covariate in the Cox regression analyses. All survival analyses were performed with a robust variance estimate to adjust for within family dependence extended to the large pedigrees. In using a robust variance estimate we treated each pedigree (instead of each individual) as an independent entity for calculating the variance of the estimates.
Expected risk genotypes were defined according to earlier reports (PP genotype of PPARG, GG genotype of CAPN10 SNP43, TT genotype of CAPN10 SNP44, EK/KK genotypes of KCNJ11, GR/RR genotypes of IRS1, GG genotype of UCP2). However, the risk TT genotype of CAPN10 SNP44 was in opposite direction compared to other studies [13,14] and selected based upon a previous report from the Botnia study [15] showing that the combination of the TT genotype of SNP44 and the GG genotype of SNP43 in CAPN10 was significantly more frequent in patients with T2D than in control individuals. Therefore, in the present study we refer to the TT genotype of CAPN10 SNP44 as an at-risk genotype. Individuals with missing data were excluded from the analyses.
Analyses of interaction between effect of phenotype (P) defined as insulin secretion (disposition index) and insulin action (HOMAIR) and genotype (G) (1 = risk and 0 = non-risk) on age at onset of T2D were performed using the following Cox proportional hazards model: h(t) = h0(t)exp(β1
P + β2
G + β3
P*G), in which h(t) is the hazard function and h0(t) is the baseline hazard function, with β1 and β2 measuring the univariate effects and β3 measuring the interaction. If there is an interaction (β3 ≠ 0), the HR for carriers and non-carriers of the risk genotype will not be the same. Thus, in different genotype carriers the HR of T2D associated with x units increase/decrease in the phenotypic value (P + x) is HR = exp(β1 + β3)*x for the risk genotype carriers and HR = exp(β1*x) for the non-risk genotype carriers. A logistic regression analysis was applied to explore the relationship between FPG and BMI, with genetic factor (defined as 1 = risk and 0 = non-risk) as dependent variable and FPG, BMI, and an interaction term as covariates. All statistical analyses were performed using Number Crunching Statistical Systems version 2004 (NCSS, Kaysville, Utah, United States), R (www.r-project.org, and Stata (StataCorp, College Station, Texas, United States). Two-sided p-values of less than 0.05 were considered statistically significant.
Results
In total, 2,293 persons (1,051 men and 1,242 women) were included in the study (Table 1). Of them, 1,078 (47%) had non-normal FPG (≥ 5.6 mmol/l), 280 (12.3%) had BMI ≥ 30 kg/m2, and 160 (7%) had both elevated FPG and BMI ≥ 30 kg/m2. Of the 2,293 persons included, 132 (6%) (67 men and 65 women; 40 with normal and 92 with abnormal glucose tolerance) developed diabetes during the follow-up period of 6 y (converters).
Table 1 Clinical Characteristics of the Participants at Baseline
PPARG
The allele and genotype frequencies of the PPARG P12A polymorphism were similar to those previously reported in Caucasians [4], with 73.3% of participants carrying the risk PP genotype (Table 2). Of all individuals who developed T2D, 109 (82.6%) had the PP genotype, which also significantly increased the risk of subsequent T2D (HR 1.7, p = 0.016) (Figure 1; Table 3). Because we have previously shown that a family history of diabetes, non-normal FPG (≥ 5.6 mmol/l), and BMI ≥ 30kg/m2 identify individuals at high risk of T2D [31], we now tested whether the PPARG risk genotype could replace family history in this prediction. In fact, the incidence of T2D was increased in carriers of the PP genotype with elevated FPG and high BMI as compared with the PA/AA genotype carriers without any other risk factors (22.9% versus 1.5%, p < 0.001) (Figure 2). This corresponded to a HR of 13.5 (95% confidence interval [CI] 4.5–40.7, p < 0.001) estimated by the Cox model (Table 3). The PPARG genotype also influenced the relationship between BMI and FPG; there was a stronger correlation between BMI and FPG in carriers of the PPARG PP as compared to the PA/AA genotypes (0.23 versus 0.15, p = 0.041), suggesting a steeper increase in FPG for any increase in BMI in carriers of the risk genotype. Furthermore, we observed a significant interaction between the PPARG P12A polymorphism and HOMAIR (p = 0.004), indicating that with increasing insulin resistance [31] carriers of the PP genotype had a greater risk of developing T2D than carriers of the PA/AA genotypes (Figure 3).
Figure 1 Unadjusted Kaplan–Meier Diabetes-Free Survival Probability Curves
Curves for different carriers of PPARG P12A (PP versus PA/AA), CAPN10 SNP44 (TT versus TC/CC), UCP2 −866 G/A (GG versus GA/AA), and the combination of PPARG and CAPN10 SNP43/44 (PP/GG/TT versus other). y-Axis shows probability of diabetes-free survival time. x-Axis shows follow-up time in years. The HR of developing T2D in different genotype carriers obtained from Cox proportional hazards regression stratified on sex and adjusted for age, BMI, and family history of diabetes with robust variance estimate is shown (see also Table 3).
Figure 2 The Effects of Risk Genotypes of the PPARG P12A Polymorphism (PP), the Combination of CAPN10 SNP43/44 (GG/TT), and the Combination of PPARG and CAPN10 SNP43/44 (PP/GG/TT) Together with FPG and BMI for the Risk of Developing T2D
y-Axis denotes incident diabetes estimated as the proportion (percent) of participants who developed diabetes during the follow-up period in the groups with each risk factor defined as risk genotype, elevated FPG (≥5.6 mmol/l), and high BMI (≥30 kg/m2). The absolute number of individuals who developed diabetes in the groups with each risk factor is given within the bars (in parentheses) and in Table S2. The incidence of T2D was significantly increased in carriers of the risk PP genotype, GG/TT genotypes, and PP/GG/TT genotypes with elevated FPG and high BMI as compared with individuals carrying low risk genotypes without risk factors (χ2 test, p < 0.001).
Figure 3 The Effect of Insulin Resistance Together with the Risk Genotype of the PPARG P12A Polymorphism on Risk of Developing T2D
y-Axis denotes HR and its 95% CI. x-Axis denotes increase in insulin resistance estimated as HOMAIR.
Table 2 Allele and Genotype Frequencies of the Studied Polymorphisms
Table 3 Risk of Developing T2D in Different Genotype Carriers of the Studied Polymorphisms
The G allele of SNP43 and the C allele of SNP44 were in strong linkage disequilibrium (D′ = 0.99, p < 0.001). Fifty percent of the participants had the risk genotype (GG) of SNP43 and 62.1% had the risk genotype (TT) of SNP44. A total of 534 (24%) individuals carried both the GG (SNP43) and the TT (SNP44) genotypes. Seventy (54.3%) of the converters had the GG (SNP43) genotype, while 91 (70.0%) had the TT (SNP44) genotype. While SNP43 had no effect on its own, the SNP44 TT genotype was associated with a moderately increased risk of T2D (HR 1.5, p = 0.035) (see Figure 1; Table 3). The incidence of T2D was highest in individuals carrying both SNP43 (GG) and SNP44 (TT) genotypes and other risk factors, particularly those with high BMI, as compared with individuals with low risk genotypes without any other risk factors (36.7% versus 3.0%, p < 0.001) (see Figure 2). In the age-adjusted Cox analysis, the corresponding HR was 13.2 (95% CI 6.0–28.7, p < 0.001) (Table 3). Also, there was a significant interaction between the GG genotype of SNP43 and HOMAIR, and between the combination of both risk genotypes (GG/TT) and HOMAIR (p = 0.037 and p = 0.028, respectively), showing that the risk conferred by worsening of insulin sensitivity [31] increased more in carriers of these genotypes than in carriers of the low risk genotypes.
UCP2
Fifty-eight (44.3%) of the converters were homozygous for the risk genotype (GG) of the UCP2 −866G>A variant. The GG genotype was associated with a modestly increased risk of T2D (HR 1.4, p = 0.049) (see Figure 1; Table 3). This risk was not influenced by BMI and FPG at baseline. However, the GG genotype was also associated with increased risk of developing T2D in patients with earlier onset of diabetes (HR 2.0, 95% CI 1.2–3.3, p = 0.0057) (Table 4). Furthermore, the GG genotype was also more frequent among patients with earlier than late onset of T2D (60.3% versus 39.7%, p = 0.042; χ2 test and odds ratio 2.5, 95% CI 1.2–5.3, p = 0.016; logistic regression analyses adjusted for gender, BMI, and family history of diabetes). None of the other tested genotypes predicted significantly earlier onset of T2D.
Table 4 Risk of Developing Earlier Onset T2D in Different Genotype Carriers of the Studied Polymorphisms
IRS1
Twenty-three (17.6%) converters carried the RR/RG genotypes of the IRS1 G972R polymorphism. Whereas the R allele had no independent effect on T2D risk, it increased the risk of T2D in a dominant fashion (RR or RG versus GG) in participants with elevated FPG and BMI ≥ 30 kg/m2 to 9.3 (95% CI 3.6–23.9, p < 0.001) (see Table 3).
KCNJ11
Ninety-six converters (73.8%) had the risk EK/KK genotypes of the KCNJ11 E23K polymorphism, but these genotypes did not increase risk of future T2D, neither alone nor in combination with elevated FPG or high BMI. In line with previous findings of an effect of this variant on insulin secretion [16–18], there was a significant interaction between the KCNJ11 E23K polymorphism and the disposition index (p < 0.001), suggesting that the risk of T2D associated with a low disposition index [31] is increased by the EK/KK genotypes.
Combined Genetic Effects
In total, 1,028 (45.9%) individuals carried risk genotypes in both PPARG (PP) and CAPN10 SNP44 (TT), whereas 397 (18.2%) individuals had three risk genotypes: PPARG (PP), CAPN10 (TT), and UCP2 (GG). The effect of both the PPARG PP and CAPN10 SNP44 TT genotypes on the risk of subsequent T2D when present in the same individual was greater (HR 2.6, 95% CI 1.5–4.5) than the individual risks (Table 3). The effect was even stronger when the combination of at-risk GG and TT genotypes of both SNP43 and 44 of the CAPN10 gene (HR 3.3, 95% CI 1.7–6.8) (Table 3; see Figure 1) was included in the analysis. Again, the incidence of T2D in participants with the combination of SNP43 (GG) and SNP44 (TT) in CAPN10, the PP genotype in the PPARG gene, elevated FPG, and high BMI was markedly higher than in those with low risk genotypes and no other risk factors (44.7% versus 3.0%, p < 0.001) (see Figure 2), with a HR of 21.2 (95% CI 8.7–51.4, p < 0.001) (Table 3). Also, these combinations influenced the correlation between BMI and FPG, yielding a steeper increase in FPG for any increase in BMI in carriers of the risk genotype combinations than in carriers of the non-risk genotypes (0.28 versus 0.19, p = 0.041).
Discussion
The key finding of the present study was that variants in the PPARG and CAPN10 genes increased the future risk for T2D, particularly in individuals with other risk factors. A question often raised about genetic association studies of polygenic diseases is whether the information can be used to predict the disease, since the risk conferred by the variant is usually rather modest (odds ratio < 1.2–1.3) [6]. In T2D association studies, cases are usually ascertained through diabetes clinics and thereby possibly enriched by carriers of more severe genetic variants. It was therefore encouraging to see that polymorphisms in some genes previously shown to be associated with T2D in case-control studies (particularly P12A in PPARG and SNP44 in CAPN10) [4,15] could predict T2D in high risk individuals from families with T2D.
The relative risk (λs) of developing T2D for members of the Botnia families is about three. However, this risk is greatest in obese (BMI ≥ 30 kg/m2) individuals with FPG above normal (≥5.6 mmol/l) and a family history of T2D [31]. Replacing the family history with the PPARG and CAPN10 variants, and particularly with their combination, gave almost the same strong prediction of subsequent T2D. These genotypes also influenced the relationship between BMI and FPG, i.e., in carriers of the risk genotypes there was a steeper increase in FPG for any given increase in BMI.
Several papers have examined the effect of single gene variants on the risk of conversion to T2D in interventional trials like the Finnish Diabetes Prevention Study [9,17] and the STOP-NIDDM trial [11]. However, it is important to know the effect of these genetic variants on risk of future T2D in an observational study before conclusions can be drawn on their putative additive or synergistic effects, together with the effects of specific factors such life style changes [9,17] or acarbose use [11]. Many of these studies have provided conflicting results in different subgroups; this is a natural corollary of their design, which breaks the initial cohorts down in relatively small subgroups with limited power. Although the present study also has limited power, it is to our knowledge the largest of its kind, and it also provides information on key T2D variants in the same paper.
PPARG
The P12A polymorphism in PPARG is to date the best replicated genetic variant for T2D, with a cumulative odds ratio from published studies of about 1.25 and overall p < 0.001 [4,8]. The P12A variant is located in an extra exon B in the 5′-end of the adipose-specific PPARG2 isoform and shows reduced transcription of target genes. The A allele has been associated with increased insulin sensitivity [35], particularly, enhanced suppression of lipolysis [36]. In support of this, there was a significant interaction between the P12A polymorphism and HOMAIR (which measures insulin resistance). The Nurse's Health Study also reported reduced risk for developing diabetes in carriers of the A allele [10]. Most recently, the STOP-NIDDM trial also reported that the PP genotype predicted conversion to diabetes in women in the acarbose intervention group [11]. These combined data, however, contrast with findings in the Finnish Diabetes Prevention Study, which reported an increased risk of developing diabetes in carriers of the A allele compared to individuals with the P12P genotype [9]. This effect was restricted to the control group, whereas the few A12A carriers in the intervention group lost more weight than the P12P carriers. As enhanced insulin sensitivity is a risk factor for weight gain, the A allele has also been associated with more rapid weight regain after weight reduction [37]. It is therefore possible that the protective effect of the A allele is attenuated in very obese individuals. Differences in BMI cannot fully explain the different results, since the risk conferred by the P12P genotype was maintained after adjusting for BMI in the present study. The effect of the P12A variant on BMI and lipolysis is also dependent upon intake of saturated fat [38]. Taken together, the data suggest a complex interaction between the P12A polymorphism in the PPARG gene and diet, body weight, and insulin sensitivity in predicting risk of future T2D.
Notably, the HR in the present prospective study was higher than the odds ratios in previous case-control studies [4]. The same also applies to the odds ratios obtained in the intervention studies discussed [10,11]. Although there could be several possible explanations for this discrepancy, a likely explanation is the accuracy by which the cases (converters) and controls (non-converters) were defined in the prospective study as compared with a case-control association study. In the prospective study all individuals underwent repeated OGTTs to define glucose tolerance status, while in case-control studies the definition of normal glucose tolerance is often based upon one OGTT. Therefore, we assume that we have a certain proportion of controls misclassified. We simulated (1,000 times) this situation by introducing 5%, 10%, and 20% misclassification of non-converters regarding the HR of 1.72 for the P12A polymorphism in the PPARG gene to predict future diabetes. A 5% misclassification would result in a decrease in HR to 1.31 (minimum 0.85; maximum 1.97), 10% to 1.18 (minimum 0.84; maximum 1.68), and 20% to 1.08 (minimum 0.82; maximum 1.38). Of course, there also could be other factors contributing, such as change in diabetes prevalence (which almost doubled) between the time when the cases in the case-control study and the converters in the prospective study were diagnosed. Finally, our study was carried out in a high risk population of first degree relatives of patients with T2D.
CAPN10
The discovery that intronic SNPs in the CAPN10 gene explained the linkage to Chromosome 2q in Mexican-Americans represented the first successful positional cloning of a T2D gene [12]. It also raised a number of questions, e.g., how could these intronic SNPs in a gene encoding a cystein protease confer increased risk of T2D? Several recent meta-analyses have demonstrated a consistent but modest risk of SNP43 and 44 (odds ratio 1.15–1.20) for the association with T2D [13,14]. In opposite direction to other studies [13,14], in the Botnia study the TT genotype of SNP44 has been associated with increased risk of T2D [15]. However, a stronger association was seen for the combination of both the GG genotype of SNP43 and the TT genotype of SNP44 [15]. Carriers of the GG genotype of SNP43 have decreased CAPN10 mRNA levels in skeletal muscle, which correlates with more severe insulin resistance [39]. Our data of an interaction between HOMAIR and the CAPN10 SNP43 supports this notion. The question arises whether the combined effect of variants in the PPPARG and CAPN10 genes on risk for future T2D can be solely explained by the variants' effect on insulin sensitivity. A combined effect of the two genes on both insulin secretion and action would be more plausible. However, the knowledge of molecular mechanisms by which calpain 10 would increase susceptibility to T2D is limited.
UCP2
Impaired insulin secretion has been shown to predominate over insulin resistance in individuals with early onset T2D [40]. In line with this view, the UCP2 variant was a strong predictor of T2D with earlier onset. The promoter variant in the UCP2 gene has been associated with increased expression of the gene in adipose tissue [41]. If this variant is associated with increased UCP2 mRNA levels in human pancreatic β-cells (which is not known), this could result in increased uncoupling and, in turn, in decreased formation of ATP and impaired insulin secretion.
KCNJ11
Genetic variants in the KCNJ11 gene have not only been associated with T2D [16,18], but also with a severe form of neonatal diabetes [42]. Whereas these neonatal mutations result in a 10-fold activation of the ATP-dependent potassium channel, the E23K variant results in only a 2-fold increase in activity [43]. The KCNJ11 E23K variant did not significantly increase the risk for T2D in our study's participants. We have no explanation for this finding other than lack of power (the study had only 52% power to detect an effect of KCNJ11 E23K on risk of developing T2D) or the presence of other unidentified risk factors in the patients with manifest T2D. We did, however, observe an interaction between the EK and/or KK genotypes and impaired β-cell function, supporting a role in insulin secretion.
Conclusion
In conclusion, we show in a large observational prospective study that genetic variants in candidate genes can predict future T2D, particularly in association with conventional risk factors such as obesity and abnormal glucose tolerance. With accumulating data from prospective studies, it should be possible to define whether there will be a future role for these variants in genetic prediction of T2D and whether these variants will influence response to prevention or treatment.
Although this study is, to our knowledge, the largest of its kind thus far, it still has limited power to detect an effect of low-frequency alleles. It is therefore obvious that larger studies with longer follow-up are needed to replicate the findings. One such resource will be the Malmö Diabetes Prevention cohort, in which 22,000 individuals have been followed for more than 20 y. They are presently being restudied to obtain DNA and information on whether they have developed diabetes or not. While waiting for these results, it will be important to create consortia to merge data from available prospective studies.
Supporting Information
Table S1 Primers and Probes Used in the Study
(92 KB PDF).
Click here for additional data file.
Table S2 The Number of Individuals Who Developed T2D Carrying Different Risk Factors
(93 KB PDF).
Click here for additional data file.
Accession Numbers
The NCBI Entrez (http://www.ncbi.nlm.nih.gov/gquery/gquery.fcgi) accession numbers for the polymorphisms discussed in this paper are CAPN10 SNP43 (rs3792267), CAPN10 SNP44 (rs2975760), IRS1 G972R (rs1801278), KCNJ11 E23K (rs5219), PPARG P12A (rs1801282), and UCP2 −866G>A (rs659366).
Patient Summary
Background
Type 2 diabetes, also known as adult onset or non-insulin-dependent diabetes, is increasing in frequency around the world. Many different factors work together to make someone more likely to develop diabetes, including factors in their environment—for example, the food they eat—and in their family background—the genes they inherited from their parents. Many studies have been done looking at which genes are associated with diabetes, but few have tried to see whether it is possible to predict who will get diabetes in future from looking at a person's genes before any symptoms develop.
Why Was This Study Done?
These authors wanted to look at changes in five genes previously shown to be associated with diabetes in a group of people who were to be followed prospectively—that is, from before they developed diabetes—and see if it was possible to predict who would get diabetes.
What Did the Researchers Do and Find?
They studied 2,293 people in Finland who were family members or spouses of people with type 2 diabetes, but who themselves did not have diabetes. They followed these people for up to 12 years, starting in 1990. In total, 132 of these individuals (6%) developed diabetes during this time. They found that changes in two of the genes, PPARG (which is involved in how the body regulates fat tissue) and CAPN10 (which is involved in modifying certain proteins), were associated with people having a higher chance of getting type 2 diabetes. This chance was increased substantially when the participants already had slightly raised blood glucose, and a high body mass index.
What Do These Findings Mean?
In some people, it does seem possible to use certain genes to predict whether a person will develop type 2 diabetes. However, environmental factors are also very important, and any risk is much increased in people who are already overweight.
Where Can I Get More Information Online?
MedlinePlus has many links to pages of information on diabetes:
http://www.nlm.nih.gov/medlineplus/ency/article/001214.htm
The Finnish Diabetes Association has information on diabetes in general and more specifically for Finland:
http://www.diabetes.fi/english/
This work was supported by grants from the Sigrid Juselius Foundation, European Community (Genomic Integrated Force for Type 2 Diabetes, grant QLG2-Ct-1999–0546), Folkhälsan Research Foundation, the Swedish Research Council, Academy of Finland, the Lundberg Foundation, the Novo Nordic Foundation, the European Federation for the Study of Diabetes (Sankyo Pharma), the Diabetes Research Foundation, the Albert Påhlssons Foundation, Crafoord Foundation, and the Anna-Lisa and Sven-Eric Lundgren Foundation. We thank the patients for their participation and the Botnia research team, as well as Malin Svensson and Barbro Gustavsson for excellent technical assistance. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
The Botnia Study Group
Investigator from Vasa, Finland: Mikael Nissén.
Investigators from Jakobstads Health Center, Jakobstads Hospital, and Folkhalsan Östanlid, Jakobstad, Finland: Bo Isomaa, Leena Sarelin, and Carola Svenfelt.
Investigators from Korsholms Health Center, Korsholm, Finland: Ulla-Britt Björk, Nils Holmström, and Jessica Strand.
Investigators from Malax Health Center, Malax, Finland: Lisbeth Åkerman and Inga-Britt Stenback.
Investigators from Närpes Health Center, Närpes, Finland: Björn Forsén, Monika Gullström, Maja Häggblom, and Susann Söderback.
Investigators from Vasa Health Center, Vasa, Finland: Kaj Lahti, Marianne Nyman, and Sonja Paulaharju.
Investigators from Department of Medicine, Helsinki University Central Hospital; Folkhalsan Research Center, Department of Genetics, and Research Program för Molecular Medicine, University of Helsinki, Helsinki, Finland: Seija Heikkinen, Paula Kokko, Merja Lahtinen, Mikko Lehtovirta, and Virve Lundgren.
Investigators from Department of Medicine, Helsinki University Central Hospital; Research Program for Cardiovascular Diseases, University of Helsinki, Helsinki, Finland: Hannele Hildén and Marja-Riitta Taskinen.
Investigators from Department of Clinical Sciences, Diabetes, and Endocrinology, Lund University, Malmö University Hospital, Malmö, Sweden: Esa Laurila and Margareta Svensson.
Citation: Lyssenko V, Almgren P, Anevski D, Orho-Melander M, Sjögren M, et al. (2005) Genetic prediction of future type 2 diabetes. PLoS Med 2(12): e345.
Abbreviations
BMIbody mass index
CIconfidence interval
FPGfasting plasma glucose concentration
HOMAIRhomeostasis model assessment index
HRhazard ratio
OGTToral glucose tolerance test
T2Dtype 2 diabetes
SNPsingle nucleotide polymorphism
==== Refs
References
Zimmet P Alberti KG Shaw J Global and societal implications of the diabetes epidemic Nature 2001 414 782 787 11742409
Bonadonna RC Alterations of glucose metabolism in type 2 diabetes mellitus. An overview Rev Endocr Metab Disord 2004 5 89 97 15041783
Chan JM Rimm EB Colditz GA Stampfer MJ Willett WC Obesity, fat distribution, and weight gain as risk factors for clinical diabetes in men Diabetes Care 1994 17 961 969 7988316
Altshuler D Hirschhorn JN Klannemark M Lindgren CM Vohl MC The common PPARgamma Pro12Ala polymorphism is associated with decreased risk of type 2 diabetes Nat Genet 2000 26 76 80 10973253
Barroso I Luan J Middelberg RP Harding AH Franks PW Candidate gene association study in type 2 diabetes indicates a role for genes involved in β-cell function as well as insulin action PLoS Biol 2003 1 e20 10.1371/journal.pbio.0000020 14551916
Florez JC Hirschhorn J Altshuler D The inherited basis of diabetes mellitus: Implications for the genetic analysis of complex traits Annu Rev Genomics Hum Genet 2003 4 257 291 14527304
Laukkanen O Pihlajamaki J Lindstrom J Eriksson J Valle TT Common polymorphisms in the genes regulating the early insulin signalling pathway: Effects on weight change and the conversion from impaired glucose tolerance to type 2 diabetes. The Finnish Diabetes Prevention Study Diabetologia 2004 47 871 877 15127203
Parikh H Groop L Candidate genes for type 2 diabetes Rev Endocr Metab Disord 2004 5 151 176 15041791
Lindi VI Uusitupa MI Lindstrom J Louheranta A Eriksson JG Association of the Pro12Ala polymorphism in the PPAR-gamma2 gene with 3-year incidence of type 2 diabetes and body weight change in the Finnish Diabetes Prevention Study Diabetes 2002 51 2581 2586 12145174
Memisoglu A Hu FB Hankinson SE Liu S Meigs JB Prospective study of the association between the proline to alanine codon 12 polymorphism in the PPARgamma gene and type 2 diabetes Diabetes Care 2003 16 2915 2917
Andrulionyte L Zacharova J Chiasson JL Laakso M Common polymorphisms of the PPAR-gamma2 (Pro12Ala) and PGC-1alpha (Gly482Ser) genes are associated with the conversion from impaired glucose tolerance to type 2 diabetes in the STOP-NIDDM trial Diabetologia 2004 47 2176 2184 15592662
Horikawa Y Oda N Cox NJ Li X Orho-Melander M Genetic variation in the gene encoding calpain-10 is associated with type 2 diabetes mellitus Nat Genet 2000 26 163 175 11017071
Weedon MN Schwarz PE Horikawa Y Iwasaki N Illig T Meta-analysis and a large association study confirm a role for calpain-10 variation in type 2 diabetes susceptibility Am J Hum Genet 2003 73 1208 1212 14574648
MSong Y Niu T Manson JE Kwiatkowski DJ Liu S Are variants in the CAPN10 gene related to risk of type 2 diabetes? A quantitative assessment of population and family-based association studies Am J Hum Genet 2004 74 208 222 14730479
Orho-Melander M Klannemark M Svensson MK Ridderstrale M Lindgren CM Variants in the calpain-10 gene predispose to insulin resistance and elevated free fatty acid levels Diabetes 2002 51 2658 2664 12145185
Gloyn AL Weedon MN Owen KR Turner MJ Knight BA Large-scale association studies of variants in genes encoding the pancreatic beta-cell KATP channel subunits Kir6.2 (KCNJ11) and SUR1 (ABCC8) confirm that the KCNJ11 E23K variant is associated with type 2 diabetes Diabetes 2003 52 568 572 12540637
Laukkanen O Pihlajamaki J Lindstrom J Eriksson J Valle TT Polymorphisms of the SUR1 (ABCC8) and Kir6.2 (KCNJ11) genes predict the conversion from impaired glucose tolerance to type 2 diabetes. The Finnish Diabetes Prevention Study J Clin Endocrinol Metab 2004 89 6286 6290 15579791
Florez JC Burtt N De Bakker PI Almgren P Tuomi T Haplotype structure and genotype-phenotype correlations of the sulfonylurea receptor and the islet ATP-sensitive potassium channel gene region Diabetes 2004 53 1360 1368 15111507
Marchetti P Lupi R Federici M Marselli L Masini M Insulin secretory function is impaired in isolated human islets carrying the Gly(972)→Arg IRS-1 polymorphism Diabetes 2002 51 1419 1424 11978638
Jellema A Zeegers MP Feskens EJ Dagnelie PC Mensink RP Gly972Arg variant in the insulin receptor substrate-1 gene and association with type 2 diabetes: A meta-analysis of 27 studies Diabetologia 2003 46 990 995 12819898
Zeggini E Parkinson J Halford S Owen KR Frayling TM Association studies of insulin receptor substrate 1 gene (IRS1) variants in type 2 diabetes samples enriched for family history and early age of onset Diabetes 2004 53 3319 3322 15561966
Florez JC Sjogren M Burtt N Orho-Melander M Schayer S Association testing in 9,000 people fails to confirm the association of the insulin receptor substrate-1 G972R polymorphism with type 2 diabetes Diabetes 2004 53 3313 3318 15561965
van Dam RM Hoebee B Seidell JC Schaap MM Blaak EE The insulin receptor substrate-1 Gly972Arg polymorphism is not associated with type 2 diabetes mellitus in two population-based studies Diabet Med 2004 21 752 758 15209769
Wang H Chu WS Lu T Hasstedt SJ Kern PA Uncoupling protein-2 polymorphisms in type 2 diabetes, obesity, and insulin secretion Am J Physiol Endocrinol Metab 2004 286 E1 E7 12915397
Sasahara M Nishi M Kawashima H Ueda K Sakagashira S Uncoupling protein 2 promoter polymorphism −866G/A affects its expression in beta-cells and modulates clinical profiles of Japanese type 2 diabetic patients Diabetes 2004 53 482 485 14747301
D'Adamo M Perego L Cardellini M Marini MA Frontoni S The −866A/A genotype in the promoter of the human uncoupling protein 2 gene is associated with insulin resistance and increased risk of type 2 diabetes Diabetes 2004 53 1905 1910 15220218
Bulotta A Ludovico O Coco A Di Paola R Quattrone A The common −866G/A polymorphism in the promoter region of the UCP2 gene is associated with reduced risk of type 2 diabetes in Caucasians from Italy J Clin Endocrinol Metab 2005 90 1176 1180 15562023
Krempler F Esterbauer H Weitgasser R Ebenbichler C Patsch JR A functional polymorphism in the promoter of UCP2 enhances obesity risk but reduces type 2 diabetes risk in obese middle-aged humans Diabetes 2002 51 3331 3335 12401727
Chan CB De Leo D Joseph JW McQuaid TS Ha XF Increased uncoupling protein-2 levels in beta-cells are associated with impaired glucose-stimulated insulin secretion: Mechanism of action Diabetes 2001 50 1302 1310 11375330
Groop L Forsblom C Lehtovirta M Tuomi T Karanko S Metabolic consequences of a family history of NIDDM (the Botnia study): Evidence for sex-specific parental effects Diabetes 1996 45 1585 1593 8866565
Lyssenko V Almgren P Anevski D Perfekt R Lahti K Predictors of and longitudinal changes in insulin sensitivity and secretion preceding onset of type 2 diabetes Diabetes 2005 54 166 174 15616025
World Health Organization Definition, diagnosis, and classification of diabetes mellitus and its complications. Report of a WHO consultation. Part 1: Diagnosis and classification of diabetes mellitus 1999 Geneva World Health Organization Available: http://whqlibdoc.who.int/hq/1999/WHO_NCD_NCS_99.2.pdf . Accessed 14 September 2005
Levy JC Matthews DR Hermans MP Correct homeostasis model assessment (HOMA) evaluation uses the computer program Diabetes Care 1998 21 2191 2192 9839117
Klein JP Moeschberger ML Survival analysis: Techniques for censored and truncated data, 2nd ed 2003 New York Springer 536
Deeb SS Fajas L Nemoto M Pihlajamaki J Mykkanen L A Pro12Ala substitution in PPARgamma2 associated with decreased receptor activity, lower body mass index and improved insulin sensitivity Nat Genet 1998 20 284 287 9806549
Stumvoll M Haring H Reduced lipolysis as possible cause for greater weight gain in subjects with the Pro12Ala polymorphism in PPARgamma2? Diabetologia 2002 45 152 153 11845236
Nicklas BJ van Rossum EF Berman DM Ryan AS Dennis KE Genetic variation in the peroxisome proliferator-activated receptor-gamma2 gene (Pro12Ala) affects metabolic responses to weight loss and subsequent weight regain Diabetes 2001 50 2172 2176 11522688
Luan J Browne PO Harding AH Halsall DJ O'Rahilly S Evidence for gene-nutrient interaction at the PPARgamma locus Diabetes 2001 50 686 689 11246892
Baier LJ Permana PA Yang X Pratley RE Hanson RL A calpain-10 gene polymorphism is associated with reduced muscle mRNA levels and insulin resistance J Clin Invest 2000 106 R69 R73 11018080
O'Rahilly S Spivey RS Holman RR Nugent Z Clark A Type II diabetes of early onset: A distinct clinical and genetic syndrome? Br Med J 1987 294 923 928 3107658
Esterbauer H Schneitler C Oberkofler H Ebenbichler C Paulweber B A common polymorphism in the promoter of UCP2 is associated with decreased risk of obesity in middle-aged humans Nat Genet 2001 28 178 183 11381268
Gloyn AL Pearson ER Antcliff JF Proks P Bruining GJ Activating mutations in the gene encoding the ATP-sensitive potassium-channel subunit Kir6.2 and permanent neonatal diabetes N Engl J Med 2004 350 1838 1849 15115830
Nichols CG Koster JC Diabetes and insulin secretion: Whither KATP? Am J Physiol Endocrinol Metab 2002 283 E403 E412 12169432
|
17570749
|
PMC1274281
|
CC BY
|
2021-01-05 10:39:25
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e345
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020345
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 1625067110.1371/journal.pmed.0020354Research ArticleDermatologyRheumatologyImmunology and AllergyRheumatologyConnective Tissue DiseaseDermatologyPaclitaxel Modulates TGFβ Signaling in Scleroderma Skin Grafts in Immunodeficient Mice Paclitaxel in the Treatment of SclerodermaLiu Xialin
1
Zhu Shoukang
1
Wang Tao
1
Hummers Laura
2
Wigley Fredrick M
2
Goldschmidt-Clermont Pascal J
1
Dong Chunming
1
*1Division of Cardiology, Department of Medicine, Duke University Medical Center, Durham, North Carolina, United States of America,2Division of Rheumatology, Johns Hopkins University, Baltimore, Maryland, United States of AmericaHuizinga Tom Academic EditorLeiden University Medical CentreThe Netherlands*To whom correspondence should be addressed. E-mail: [email protected]
Competing Interests: The authors have declared that no competing interests exist.
Author Contributions: CD had the original idea for the study, and designed and oversaw the study. XL and SZ performed most of the experimental procedures and data analysis and participated in the study design. TW performed some experiments. LH and FMW recruited the patients and collected the skin biopsies. PJGC helped develop the concept of the study, and coordinated and supervised the experiments. XL, PJGC, and CD contributed to the writing of the report.
12 2005 1 11 2005 2 12 e35416 5 2005 25 8 2005 Copyright: © 2005 Liu et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.
Manipulating Microtubules in Systemic Sclerosis
Tweaking Microtubules to Treat Scleroderma
Background
Systemic sclerosis (SSc) is characterized by excessive fibrosis and obliterative vascular lesions. Abnormal TGFβ activation is implicated in the pathogenesis of SSc. Aberrant TGFβ/Smad signaling can be controlled by stabilization of microtubules with paclitaxel.
Methods and Findings
SSc and healthy human skin biopsies were incubated in the presence or absence of paclitaxel followed by transplantation into severe combined immunodeficient mice. TGFβ signaling, fibrosis, and neovessel formation were evaluated by quantitative RT-PCR and immunohistochemical staining. Paclitaxel markedly suppressed Smad2 and Smad3 phosphorylation and collagen deposition in SSc grafts. As a result, the autonomous maintenance/reconstitution of the SSc phenotype was prevented. Remarkably, SSc grafts showed a 2-fold increase in neovessel formation relative to normal grafts, regardless of paclitaxel treatment. Angiogenesis in SSc grafts was associated with a substantial increase in mouse PECAM-1 expression, indicating the mouse origin of the neovascular cells.
Conclusion
Low-dose paclitaxel can significantly suppress TGFβ/Smad activity and lessen fibrosis in SCID mice. Transplantation of SSc skin into SCID mice elicits a strong angiogenesis—an effect not affected by paclitaxel. Although prolonged chemotherapy with paclitaxel at higher doses is associated with pro-fibrotic and anti-angiogenic changes, the findings described here indicate that low-dose paclitaxel may have therapeutic benefits for SSc via modulating TGFβ signaling.
In a mouse model of Systemic sclerosis low-dose paclitaxel can suppress an important disease-related pathway, that involves TGFβ and Smad, and lessen fibrosis .
==== Body
Introduction
Systemic sclerosis (SSc) is an autoimmune disease characterized by excessive deposition of extracellular matrix (ECM) proteins and obliterative vascular lesions in the skin and internal organs [1]. Although the causative factors of this disease remain to be characterized, the main pathobiological features of SSc comprise three interactive components: autoimmune attack, vascular damage, and a lesion in fibroblasts [2,3]. The fibrotic process consists of massive deposition of connective tissue, mostly collagens, which is frequently responsible for the failure of many organs in patients with SSc. Consequently, an array of antifibrotic agents has been developed to: (1) reduce synthesis, excretion, or polymerization of collagen fibrils; (2) enhance collagenase activity; and (3) neutralize cytokines capable of stimulating collagen synthesis, such as TGFβ (transforming growth factor-beta), interleukin-4, and interleukin-6. To date, several of these antifibrotic agents have been tested in clinical trials [4]. These include D-penicillamine, colchicine, interferon gamma, and relaxin. Unfortunately, none of these medicines have any proven efficacy in retarding the fibrotic process [5].
TGFβ is a multifunctional regulatory cytokine that is involved in a large number of cellular activities. TGFβ induces matrix accumulation and tissue fibrosis associated with SSc [4]. In addition, TGFβ promotes endothelial cell apoptosis (a property that may be amplified by the presence of anti-endothelial cell autoantibodies found in SSc), and inhibits smooth muscle cell apoptosis. It also regulates T lymphocyte–mediated immune reactions [6]. Thus, TGFβ is believed to play a central role in the pathogenesis of SSc by activating tissue fibroblasts directly or indirectly through endothelial cells, by regulating lymphocyte function, and by affecting endothelial and smooth muscle cell survival and thrombus formation. Therefore, it is conceivable that inhibition of the aberrant TGFβ signaling may be a promising therapeutic strategy for SSc.
TGFβ initiates its diverse cellular responses by binding to and activating specific cell-surface receptors that have intrinsic serine/threonine kinase activity. The activated TGFβ receptors stimulate the phosphorylation of receptor-regulated Smad2 and Smad3 proteins (R-Smads), which in turn form complexes with Smad4 that accumulate in the nucleus and regulate the transcription of target genes. Inhibitory Smad7 acts in an opposing manner to the R-Smads, inhibiting TGFβ signaling [7]. We have previously demonstrated that endogenous Smad-2, −3, and −4 bind microtubules (MTs) in several cell lines, and the binding provides a negative regulatory mechanism to control TGFβ activity. Disruption of the MT network by chemical agents, such as nocodazole and colchicine, leads to ligand-independent Smad nuclear accumulation and transcription of TGFβ-responsive genes, and increases TGFβ-induced Smad activity [8]. We have also found that inhibitory Smad7 is selectively decreased, and we and others have shown that R-Smad3 is increased, in SSc skin fibroblasts, resulting in uncontrolled TGFβ activity that may be, at least in part, responsible for the aberrant ECM deposition observed in SSc [9]. Indeed, in a hybrid human–severe combined immunodeficient (SCID) mouse skin xenotransplant model, we were able to maintain/reconstitute the SSc phenotype using skin biopsies from SSc patients, indicating that the altered balance between inhibitory Smad7 and R-Smads, due to Smad7 deficiency and Smad3 up-regulation, may represent an intrinsic and persistent defect in tissue fibroblasts that can maintain or even induce SSc lesions autonomously in the absence of altered circulatory or systemic factors [9]. The aim of this study was to determine if MT stabilization with low-dose paclitaxel could inhibit TGFβ/Smad signaling, ameliorating the fibrotic changes associated with SSc.
Methods
Participant Characteristics
Skin biopsy (6-mm punch) was obtained in an area above the elbow considered to have grossly intact skin thickness as determined by clinical palpation of patients with SSc and in the same location of control participants. Thirty-two patients, of whom 26 were female and 6 were male, aged 38–64 y of age (average 48 y) with diffuse cutaneous SSc were studied. The median duration of skin disease was 5.5 y (2–10 y). Concomitant treatment of SSc patients included the immunosuppressant mycophenolate mofetil, angiotensin-converting enzyme inhibitors, calcium channel blockers, and proton-pump inhibitors. These patients were recruited from the Johns Hopkins University Scleroderma Center. All patients met the American College of Rheumatology criteria for the diagnosis of SSc [10]. Patients with overlap syndromes (e.g., lupus) were excluded. Twelve normal participants, including nine women and three men with an average age of 44 y (range 33–58 y) were also analyzed. All patients and volunteers gave their informed consent, and the study was approved by the Johns Hopkins University Human Subjects Institutional Review Board Committee.
Skin Transplantation
SCID mice (C.B-17/lcrHsd-scid) were purchased from Jackson Laboratory (Bar Harbor, Maine, United States). Mice aged 6–8 wk and weighing 18–22 g were used for transplantation. Briefly, the skin biopsy patch was trimmed into an oval shape depleted of fat tissue, and placed in 2 mM paclitaxel (Taxol; Sigma, St. Louis, Missouri, United States) or in PBS for 30 min at 4 °C. The tissue was rinsed with PBS just before transplantation. On the dorsolateral surface of the recipient SCID mouse, an oval graft bed approximately 6 mm in diameter was created to fit the graft, leaving the deep fascia layer intact. The trimmed skin patch was transplanted onto the graft bed by suturing the skin patch into the defect with 8–0 suture around the margin of the patch. A total of 25 transplants, including 17 SSc grafts and eight normal grafts, were included in the study (Table 1). The grafts, together with a small ring of the native skin, were harvested at 30 d following transplantation. Nineteen non-transplanted biopsy samples (15 SSc and four normal) were included as nonsurgical controls. Upon sacrifice, the underside of the skin was photographed for angiogenesis study, and the tissue was then divided into two segments: One segment was fixed in 10% neutral buffered formalin and embedded in paraffin, and the other segment was stored in RNAlater (Ambion, Austin, Texas, United States) for RNA expression studies. All animals were cared for in compliance with the “Principles of Laboratory Animal Care and the Guide for the Care and Use of Laboratory Animals,” prepared by the Institute of Laboratory Animal Resources and published by the National Institutes of Health (NO86 to 23, revised 1985).
Table 1 The Summary of Skin Samples Used in the Study
Immunohistochemistry and Histology
SSc and normal specimens were processed by 10% formalin-fixation and paraffin-embedding. Immunohistochemistry (IHC) for Smad2, Smad3 (Santa Cruz Biotechnology, Santa Cruz, California, United States), phospho-Smad2 (Upstate Biotechnology, Lake Placid, New York, United States), phospho-Smad3 (a generous gift from Dr. Peter ten Dijke, Leiden University, Leiden, The Netherlands), collagen-1 (Santa Cruz Biotechnology), and PAI-1 (American Diagnostica, Greenwich, Connecticut, United States) was performed using Vector's ABC kits (Vector Laboratories, Burlingame, California, United States) on 3-μm consecutive serial sections. Briefly, after deparaffinization, slides were quenched in 3% H2O2 for 10 min to block endogenous peroxidase and washed in PBS. Sections were incubated with the primary antibody for 1 h and then with biotinylated secondary antibody followed by ABC reagents. Color development was achieved by incubating diaminobenzidine (DAB) as a substrate. Slides were counterstained with Mayer's hematoxylin. Preincubation of the primary antibody with specific blocking peptides or substitution of the primary antibody with an irrelevant IgG served as negative controls. Smad2- and Smad3-positive cells were counted in at least six high-power fields in each sample by two independent observers (CMD and XL). A minimum of 500 cells was counted. Percent positive cells were calculated as the number of positive cells/total number of cells × 100. Cells positive for phospho-Smad2 and phospho-Smad3 were counted in a similar fashion. Among the Smad2- and Smad3-positive cells, the percentage of those stained for phospho-Smad2 and phospho-Smad3 was calculated as the number of phospho-Smad2– and phospho-Smad3–positive cells/the number of cells positive for Smad2 and Smad3, respectively, × 100.
Sections in series with IHC were stained with H & E (hematoxylin/eosin), and Verhoeff's van Gieson elastin and Mason trichrome (VVM). Each section was examined for the presence, extent, and distribution of collagen, elastic fibers, and other matrix proteins.
Angiogenesis Assessment
The number of microvessels was counted from three to five randomly selected high-power fields (40× magnification) in histology slides stained with H & E. Neovessel formation was also evaluated macroscopically by counting the number of vessels on the underside of the grafts and skin biopsies in a 6-mm field—the entire graft. Angiogenic activity was compared among different groups, including SSc skin biopsy, SSc grafts, normal skin biopsy, and normal skin grafts.
TaqMan Real-Time Reverse Transcription-PCR
RNA was isolated from the skin using RNeasy Mini kit (Quiagen, Chatsworth, California, United States). One μg total RNA was used for the synthesis of first strand cDNA using the SUPERSCRIPT Preamplification System (Life Technologies, Rockville, Maryland, United States). PCR was optimized for the quantitation of alpha2(I) collagen (COL1A2; mouse and human share the same sequence), human PECAM-1 (platelet endothelial cell adhesion molecule-1), and mouse PECAM-1 with specific primers and probes. A sequence detector (ABI Prism 7700, PE Applied Biosystems, Foster City, California, United States) was used to measure the amplified product in direct proportion to the increase in fluorescence emission continuously during the PCR amplification. For each sample, a threshold cycle (Ct) value was calculated from each amplification plot, representing the PCR cycle number at which the fluorescence was detectable above an arbitrary threshold. To normalize Ct of the target gene copies to 18S rRNA, ΔCt was calculated as Ct (target) − Ct (18S rRNA). For each sample, the level of COL1A2, human PECAM-1, and mouse PECAM-1 was calculated as 2−ΔCt. Each sample was tested in triplicate and repeated twice.
Statistical Analysis
Data are presented as mean ± SEM. Analysis of variance was performed to compare differences among different groups. A p-value < 0.05 was considered statistically significant.
Results
Paclitaxel Inhibits TGFβ/Smad Signaling and Collagen Deposition in SSc Grafts
We demonstrated previously that because of the Smad7 deficiency and increased Smad3 expression in SSc fibroblasts, transplantation of SSc skin retained the SSc phenotype in SCID mice, indicating that the altered balance between positive and inhibitory Smads may represent an intrinsic defect in tissue fibroblasts that can maintain or even induce SSc lesions autonomously in the absence of altered circulatory or systemic factors [9]. Since MT instability enhances TGFβ/Smad signaling pathway [8], we reasoned that stabilization of MTs with paclitaxel might dampen the exacerbated TGFβ signaling in the SSC grafts and prevent the maintenance/reconstitution of SSc phenotype. IHC using anti-phospho-Smad2 antibody revealed enhanced Smad2 phosphorylation, a marker for TGFβ signaling, in SSc tissue, relative to normal skin specimens (87% ± 11% versus 23% ± 15%, p < 0.01), when both pre- and post-transplant tissues were analyzed together, whereas the level of total Smad2 remained comparable between normal versus SSc skin tissue. Remarkably, pre-transplant incubation of SSc skin with paclitaxel substantially suppressed the level of Smad2 phosphorylation, approaching that of normal grafts (28% ± 19% versus 21% ± 13%, p > 0.05; (Figure1A–1F). Minimal, if any, effects of paclitaxel on Smad2 activity in normal skin grafts were detected (normal grafts + paclitaxel, 26% ± 18% versus normal grafts, 21% ± 13%, p > 0.05). Furthermore, the level of total Smad2 and Smad7 (data not shown) remained unchanged.
Figure 1 Paclitaxel Treatment Suppresses Smad2 Phosphorylation
Strong nuclear staining for phospho-Smad2 (A) and total Smad2 (B) is observed in an SSc skin graft. Smad2 phosphorylation (C) is rare in normal graft, which expresses abundant Smad2 (D). Smad2 phosphorylation is suppressed with paclitaxel (Taxol) treatment (E), without affecting total Smad2 (F) in SSc skin graft. Tx, transplantation.
A similar increase in Smad3 phosphorylation was detected in SSc specimens compared to normal skin tissue (83% ± 11% versus 20% ± 15%, p < 0.01). In addition, the level of total Smad3 was increased in SSc versus normal skin tissue (81.5% ± 10% versus 55% ± 21%, p < 0.01)—an observation that is consistent with our previous study. Importantly, pre-transplant incubation of SSc skin substantially suppressed the level of Smad3 phosphorylation to that of normal grafts (27% ± 14% versus 24% ± 13%, p > 0.05). Furthermore, the expression level of total Smad3 was also downregulated by paclitaxel treatment in SSc (81.5% ± 10% versus 57. 5% ± 12%, p = 0.05) (Figure S1A–S1F). There was no detectable effect of paclitaxel on Smad3 activation in normal skin grafts (30% ± 16% versus 25% ± 11%, p > 0.05). These findings indicate that preincubation of SSc skin with paclitaxel can effectively offset Smad7 deficiency and Smad3 up-regulation–induced augmented TGFβ signaling in SSc skin grafts, without affecting the total Smad2 and Smad7 expression level. Moreover, when TGFβ signaling is not perturbed in normal skin grafts, paclitaxel does not exert detectable effects.
The progressive accumulation of ECM in the skin and internal organs is a hallmark of SSc, of which collagen type I is the major constituent. Indeed, collagen type I metabolites have been used as markers to evaluate disease activity in SSc [11]. To examine whether paclitaxel treatment affected collagen deposition in SSc skin grafts, we performed quantitative TaqMan real-time reverse transcription-PCR (TRT-PCR) for COL1A2. Instructively, the expression of COL1A2—a gene whose promoter contains multiple Smad-binding elements (SBE)—was reduced by 4.5-fold (p < 0.01) with paclitaxel treatment in SSC grafts, reaching a level that approximated COL1A2 mRNA expression in normal skin grafts. By contrast, paclitaxel had hardly any effects on COL1A2 mRNA expression in normal skin grafts (Figure 2).
Figure 2 Paclitaxel Decreases COLA2 mRNA Expression in SSc Grafts
TRT-PCR analysis shows that the expression of COL1A2 is reduced 4.5-fold with paclitaxel (Taxol) treatment in SSC grafts, reaching a level equivalent to that of normal skin grafts. By contrast, paclitaxel has no effects on COL1A2 expression in normal skin grafts.
To confirm the TRT-PCR findings, we performed histological VVM staining to evaluate total collagen deposition and IHC to examine collagen-1 expression level. As shown in Figure 3, there was extensive deposition of total collagen and other ECM proteins, including elastic fibers, in the entire dermis of SSc grafts, which was markedly reduced with paclitaxel treatment. Strong, intense collagen-1 staining was demonstrated in untreated SSc grafts, relative to much weaker staining in paclitaxel-treated SSc tissue, which was comparable to normal skin grafts treated with and without paclitaxel (Figure 3). These data indicate that low-dose paclitaxel prevents the maintenance/reconstitution of the SSc phenotype in SCID mice, and this effect may be mediated by stabilizing MT-Smad complex and subsequent inhibition of TGFβ/Smad signaling.
Figure 3 Paclitaxel Decreases Collagen Deposition in SSc Grafts
VVM staining shows abundant, thick collagen bundles in the SSc graft (A). The amount of collagen is markedly reduced, and the collagen fibers become finer with paclitaxel (Taxol) treatment (B), comparable to that seen in the normal skin graft (C). Similarly, IHC staining demonstrates intense collagen-1 staining in the SSc graft (D), which is substantially decreased with paclitaxel treatment (E) to a level similar to that seen in the normal skin graft (F).
Paclitaxel Does Not Affect the Enhanced Local Angiogenesis in SSc Skin Grafts
SSc skin lesions are characterized by obliterative microvascular lesions and decreased capillary density, suggesting excessive endothelial injury and/or a deregulated, insufficient angiogenic response [12]. Three basic mechanisms could account for the defective vascular repair in SSc: (1) lack of signals produced by the skin to recruit progenitor cells from the bone marrow, (2) appropriate skin recruitment signals but failure of the bone marrow to mount an adequate repair process, and (3) appropriate skin recruitment signals and adequate bone marrow endothelial progenitor cell (EPC) supply but excessive destruction of EPCs upon their mobilization by immune system. To determine if the defective angiogenic response might be related to lack of signals produced by the skin to recruit progenitor cells from the bone marrow and whether paclitaxel would adversely affect the angiogenic process in SCID mice, we examined neovessel formation following SSc and normal skin transplantation. The EPC supply from the mouse bone marrow is assumed to be nonlimiting. Macroscopic examination of the underside of the 6-mm skin grafts revealed more pronounced angiogenesis in SSc grafts than normal skin transplants (Figure 4A–4D). There were on average 30 vessels in the SSc graft in three high-power fields (40×), as compared with 15 vessels in normal skin grafts (p < 0.01) (Figure 4E). To further confirm these findings and determine the origin of the neovascular cells, we performed TRT-PCR for human and mouse PECAM-1, an endothelial cell marker. If bone marrow failure and/or autoimmunity-mediated EPC destruction, but not secretion of mobilizing factors by the skin, represented the major bottleneck for a defect repair process in SSc, an aggressive angiogenic response with mouse progenitor cells and, therefore, increased mouse PECAM-1, in the SSc grafts would be expected. As shown in Figure 4F, mouse PECAM-1 was almost undetectable in SSc and normal skin biopsies before transplantation, reflecting the specificity of the primers and probe for the mouse gene. Substantial increase in mouse PECAM-1 mRNA expression was observed in SSc and normal skin grafts. The amplitude of the increase in mouse PECAM-1, however, was significantly greater in SSc than in normal grafts (p < 0.01). Human PECAM-1 was expressed in much lower levels in SSc than in normal skin tissue (p < 0.01) before transplantation, indicative of vascular deficiency in SSc. The level of human PECAM-1 did not change significantly before and after transplantation (Figure 4G). Collectively, these data indicate that the neovascular cells are of recipient mouse origin and that exhaustion of EPC supply from the patient's bone marrow and/or autoimmunity-mediated targeting and destruction of EPCs after mobilization may contribute to the vascular lesion associated with SSc.
Figure 4 SSc Skin Grafting Stimulates Angiogenesis in SCID Mice
Macroscopic analysis of the vessels in the underside of normal (A and B) and SSc (C and D) skin grafts reveals more pronounced angiogenesis in SSc grafts; paclitaxel (Taxol) treatment has no effect on angiogenesis at the macroscopic level (B and D versus A and C). Microscopic analysis shows that the number of vessels in three high-power fields of the SSc grafts is greater than that of normal skin grafts; paclitaxel has no effect on angiogenesis at the microscopic level (E). TRT-PCR for mouse PECAM-1 demonstrates a substantial increase in mouse PECAM-1 mRNA expression in SSc and normal skin grafts regardless of paclitaxel treatment; the amplitude of the increase is significantly greater in SSc than in normal skin grafts (F). Human PECAM-1 is expressed in much lower levels in SSc than in normal skin tissues, indicative of vascular deficiency in SSc; the lack of difference in the expression level of human PECAM-1 in SSc and normal skin tissues before and after transplantation indicates that the neovascular cells are derived from the recipients (G).
The putative activity of paclitaxel to induce endothelial cell or EPC apoptosis and/or to inhibit the proliferation of these cells with subsequent blockade of angiogenesis represented a potential untoward side effect for the use of paclitaxel in the treatment of SSc. To exclude this possibility, we compared the number of vessels in SSc and normal skin grafts treated with and without paclitaxel. As shown in Figure 4A–4E, treatment with paclitaxel did not adversely affect the angiogenic process. A similar increase in neovessel formation was observed in paclitaxel-treated versus nontreated SSc grafts, as compared with pre-transplantation skin biopsies from the same patients. Furthermore, paclitaxel incubation had little, if any, effect on the expression level of PECAM-1 mRNA, in particular mouse PECAM-1, in SSc and normal skin grafts (Figure 4F and 4G). These findings support the notion that low-dose paclitaxel can be used to modulate TGFβ/Smad signaling and treat the fibrotic lesion, without adversely affecting the vascular component of SSc pathobiology.
Discussion
In the present study, we found that the autonomous maintenance/reconstitution of the SSc phenotype in the hybrid SCID mouse transplant model was substantially prevented by pre-transplant incubation of the SSc skin with 2 mM paclitaxel (taxol), an MT-stabilizing agent. Furthermore, SSc grafts showed a 2-fold increase in neovessel formation, as compared with normal skin grafts, regardless of paclitaxel treatment. The angiogenic process in SSc grafts was associated with a substantial increase in mouse, but not human, PECAM-1 expression, pointing the origin of the neovascular cells to the recipient mice, perhaps the bone marrow. The angiogenesis data suggest that the skin of patients with SSc has preserved ability to trigger and support an angiogenic response. Collectively, these findings indicate that low-dose paclitaxel may potentially help keep in check the fibrotic process associated with SSc, without adversely affecting the vascular component of the disease.
Members of the TGFβ superfamily play a central role in fibrosis, contributing to the influx and activation of inflammatory cells and fibroblasts and their subsequent elaboration of ECM [4]. TGFβ propagates its signal mainly via a signal transduction network involving receptor serine/threonine kinases at the cell surface and their substrates, the Smad proteins. Upon phosphorylation and oligomerization, R-Smads move into the nucleus to regulate transcription of target genes [7]. Recent findings indicate that the aberrant ECM synthesis by cultured SSc fibroblasts is due, at least in part, to the constitutively enhanced activation of the TGFβ signaling, which may result from the elevated levels of TGFβ receptor type I, inappropriate overexpression/activation of Smad2 and Smad3, and/or decreased Smad7 expression. Indeed, evidence obtained from Smad3-deficient mice shows that TGFβ-induced pro-fibrotic activities are mainly mediated by Smad3 [13]. Smad3-deficient inflammatory cells and fibroblasts do not respond to the chemotactic effects of TGFβ and do not auto-activate TGFβ [14,15]. Furthermore, Smad3-deficient mice are resistant to radiation-induced cutaneous fibrosis, bleomycin-induced pulmonary fibrosis, carbon tetrachloride–induced hepatic fibrosis, and glomerular fibrosis induced by type 1 diabetes caused by streptozotocin [16,17]. We have demonstrated that Smad7, the inhibitory Smad specific for TGFβ signaling, is selectively decreased, whereas Smad3 expression is increased in SSc fibroblasts. TGFβ signaling events, including phosphorylation of Smad2 and Smad3 and transcription of PAI-1 gene, are increased in SSc fibroblasts, relative to normal fibroblasts. Importantly, the imbalance between Smad7 and Smad3 itself can maintain or induce the SSc phenotype in SCID mice [9]. Furthermore, we have previously shown that MTs serve as a negative regulator for TGFβ/Smad signaling by forming a complex with endogenous Smad2, Smad3, and Smad4, sequestering the R-Smads away from the TGFβ receptor in several cell types [8]. Stabilization of MTs by low-dose paclitaxel can dampen to a normal level the exacerbated TGFβ signaling due to MT instability and block TGFβ-induced inhibition of myogenesis in C2C12 myoblasts [18]. In the present study, we provide evidence indicating that transient incubation of SSc skin with paclitaxel before transplantation into SCID mice substantially suppressed the phosphorylation of Smad2 and Smad3, two homologous Smad proteins that transduce signals from TGFβ and activin. Remarkably, paclitaxel treatment efficiently blocks the autonomous reconstitution and maintenance of the SSc phenotype in SCID mice. These data are consistent with our previous observations, supporting the notion that TGFβ/Smad signaling is regulated by the dynamic stability of MTs, which is sensitive to low-dose MT stabilizing agents, like paclitaxel.
Prolonged chemotherapeutic treatment with paclitaxel has been associated with scleroderma-like changes, albeit in only a small fraction of patients. It is noteworthy that the anti-tumor effect of paclitaxel is mediated via the inhibition of cell proliferation and requires a much higher dosage. The inhibition of TGFβ/Smad signaling, however, can be achieved with very-low-dose paclitaxel. We and others have demonstrated that low-dose paclitaxel had minimal, if any, detectable effects on cell proliferation and other cellular activities, including fibrosis. Intriguingly, low-dose paclitaxel has been shown to inhibit collagen-induced arthritis and other autoimmune disorders in various animal models[19,20]. The low-dose paclitaxel treatment in our human–SCID mouse skin transplant model resulted in marked inhibition of TGFβ/Smad signaling, as evidenced by the decreased phosphorylation of Smad2 and Smad3, and lessened fibrosis. Our data indicate that under our experimental conditions in the SCID mouse, low-dose paclitaxel does not induce scleroderma skin changes. This does not, however, refute the potential linkage between higher doses of paclitaxel used for cancer therapy in humans and skin fibrosis.
The structural and functional vascular and microvascular abnormalities, including Raynaud's phenomenon, represent one of the most important pathological features of SSc [12]. Indeed, microvascular damage and consequent loss of blood supply is found in all involved organs and leads to underperfusion and chronic ischemia, which may play an important role in organ dysfunction and even in the pathogenesis of in SSc. The fibrotic changes may represent a default pathway resulting from vascular failure. Endothelial apoptosis caused by viral infection, immune reactions to viral or environmental factors, reperfusion injury, or anti-endothelial antibodies, is considered a precipitating event in the genesis of vascular lesions in SSc [21]. Recent evidence, however, indicates that vascular repair, particularly that mediated via adult stem cells/EPCs from the bone marrow, plays an important role in maintaining vascular homeostasis and angiogenesis in a variety of disease states [22,23]. It is hypothesized that the vessel wall can deal fairly well with multiple circulating and local noxious stimuli as long as the bone marrow–derived repair capacity remains intact [24]. Indeed, many autoimmune processes might target the repair pathways that are needed to maintain the homeostasis of involved tissues [25]. Adequate vascular repair entails adequate supply of competent progenitor cells in, and their efficient mobilization from the bone marrow, as well as the effective homing to, and subsequent differentiation of these progenitor cells within the vessel wall. Any dysregulation in these processes can tilt the balance of vascular repair and injury in favor of injury and vascular lesion formation. It was postulated that the inadequate angiogenic response in SSc was due to reduced expression of angiogenic factors, such as vascular endothelial growth factor (VEGF), and their receptors. It was recently shown, however, that both VEGF and its receptors (VEGFR1 and VEGFR2) were up-regulated in SSc skin specimens compared with healthy controls [26]. In addition, VEGF protein was significantly increased in blood samples from patients with SSc, reaching levels observed in patients with numerous malignant diseases [27]. Thus, there appears to exist a proper, if not increased, activation of the VEGF/VEGF-receptor axis—key to EPC mobilization—in patients with SSc [26]. One could not, however, rule out the possibility that other, yet to be identified, factors might be missing from the skin and blood of SSc patients, factors that are required to mount and support a successful angiogenic response.
The human skin–SCID mouse transplantation model provides a remarkable opportunity to determine whether lack of signals from the skin of SSc patients to recruit EPCs from the bone marrow might play a role in the vascular lesion formation in SSc, since the bone marrow of the SCID mice is considered intact in terms of EPC supply and there is an absence of immune-mediated destruction of EPCs. Extensive analysis of vascular formation following SSc and normal skin transplantation both at microscopic and macroscopic levels revealed a robust angiogenic response in the SSc grafts, at least twice that seen in normal skin grafts. TRT-PCR for human and mouse PECAM-1 demonstrated a substantial increase in mouse PECAM-1 mRNA expression in SSc and normal skin grafts. Furthermore, the amplitude of the increase in mouse PECAM-1 was significantly greater in SSc than in normal grafts. In contrast, human PECAM-1 was expressed in much lower levels in SSc than in normal skin tissues before transplantation, indicative of vascular deficiency in SSc, and the expression level of human PECAM-1 did not change before and after transplantation in SSc and normal skin tissues, indicating that the neovascular cells are derived from the mouse recipients rather than the human donors. These data indicate that the signals from SSc skin, if anything, are stronger in stimulating the mobilization of EPCs from the bone marrow. Moreover, the robust angiogenic activity observed in SSC relative to normal grafts argues against the possibility that the angiogenic response in the SCID mouse is solely due to a wound-healing effect in response to grafting procedures.
It remains to be determined, however, if the EPC supply in the bone marrow or autoimmunity-induced targeting and destruction of EPCs after their mobilization or both serve as the culprit in undermining the vascular repair process, contributing to vascular lesion formation in SSc. Using an animal model of atherosclerosis, we have demonstrated that exhaustion of selected progenitor cell populations, including EPCs and their supporting cells, a process that is accelerated by risk factors, can lead to the inability of the bone marrow to mount a successful vascular repair process, contributing to the initiation and progression of atherosclerotic vascular lesion formation [22,28]. Circulating EPCs and CD34+/KDR+ precursor cells were reduced in patients with atherosclerotic coronary artery disease. The reduction of these cells represented significant risk factors for atherosclerosis, even after adjusting for most classic risk factors, including age, sex, hypertension, diabetes, smoking, family history, and low-density lipoprotein cholesterol levels. Furthermore, factors that reduce cardiovascular risk, such as statins or exercise, elevate EPC levels, which contribute to enhanced endothelial repair. Hence, a reduced circulating EPC level has been proposed as a significant risk factor for cardiovascular disease. There is conflicting evidence regarding the level of circulating EPCs in SSc patients. Specifically, Kuwana et al. [25] showed that the levels of circulating EPCs, defined by the expression of CD34, CD133, and VEGFR2, were decreased in SSc patients. By contrast, Del Papa et al. [29] found that circulating EPCs—cells positive for CD34 and CD133—were increased in SSc patients, particularly in the early stages of the disease. Although the discrepancy between these studies might be due to the differing definitions of EPCs and different disease stages of SSc patients, it underscores the idea that further investigation is warranted in delineating the relative contribution of inadequate bone marrow EPC supply and excessive destruction of circulating EPCs to the imbalance between vascular injury and repair in SSc.
The established mechanism that confers the antitumor effects of paclitaxel relates to its antiproliferative and antiangiogenic activity when used at large doses and for a prolonged period of time [30]. Indeed, the putative antiangiogenic effect of paclitaxel represented our primary concern for its use in the treatment of SSc. Remarkably, when paclitaxel- and sham-treated SSc and normal grafts were analyzed for their angiogenic response at microscopic and macroscopic levels, similar numbers of vessels were observed, indicating that low-dose paclitaxel does not affect neovessel formation that is likely mediated via EPCs recruited from the bone marrow.
In conclusion, we have demonstrated that SSc skin treated with low-dose paclitaxel can significantly suppress the exacerbated TGFβ/Smad activity of SSc skin and lessen the fibrotic changes upon transplantation into SCID mice. We have found that transplantation of SSc skin into SCID mice elicits effective angiogenesis—an effect that is significantly stronger than with normal skin grafting. Importantly, the neovascular cells are almost exclusively derived from the recipients, perhaps originating from the mouse bone marrow. These observations should shed light on SSc disease pathogenesis and provide evidence for the development of novel therapeutic strategies. The fact that low-dose paclitaxel suppresses fibrosis without dysregulating angiogenesis suggests that fibrosis might be the result of a “default” pathway that develops autonomy once the SSc tissue becomes depleted of blood vessels.
Supporting Information
Figure S1 Paclitaxel Treatment Suppresses Smad3 Phosphorylation
Strong nuclear staining for phospho-Smad3 (A) and total Smad3 (B) is observed in an SSc skin graft. Smad3 phosphorylation (C) is rare in normal graft, which also expresses a low level of total Smad3 (D). Smad3 phosphorylation and perhaps total Smad 3 are suppressed with paclitaxel (Taxol) treatment in SSc (E and F).
(9.5 MB TIF).
Click here for additional data file.
Patient Summary
Background
Systemic sclerosis or scleroderma (SSc) is the name for a group of progressive diseases, all of which involve the abnormal growth of connective tissue. SSc is triggered when the body's immune system turns against the body, causing abnormal production of collagen that can be limited to the skin or extend to the blood vessels and internal organs.
Why Was This Study Done?
Various genes have been suggested as being involved in the production of collagen. One of the key genes in this pathway, and in SSc, is TGFβ, which affects the activity of a number of other genes. A part of a cell's internal structure, known as microtubules, affects TGFβ activity. Previous work has shown that microtubules can be affected by drugs, such as the one in this study, paclitaxel, which stabilizes the microtubules.
What Did the Researchers Do and Find?
They looked at the effect of paclitaxel on human skin samples that had been taken from people with SSc, or from people with normal skin, and then transplanted into mice. This transplantation is a good way of studying the effects of drugs in this disease without having to test them directly on people. The researchers found that paclitaxel affected the activity of TGFβ and the related genes, and ultimately decreased the amount of collagen that would usually be found in the skin of people with SSc.
What Do These Findings Mean?
These findings suggest that paclitaxel would be worth investigating further as a useful drug in the treatment of people with SSc. One concern, however, is that in people who have been treated with high doses of this drug for other conditions, such as some types of cancer, the opposite effect of that shown here has been found: increased collagen has been produced. Therefore, before it can be certain that this drug is safe to use, further work will need to be done to determine what the different effects of the drug are at high and low doses, and what a safe dose is in humans.
Where Can I Get More Information Online?
MedlinePlus has many links to pages of information on SSc:
http://www.nlm.nih.gov/medlineplus/scleroderma.html
The Scleroderma Foundation is a non-profit organization based in the United States that provides information on scleroderma for patients, and supports research:
http://www.scleroderma.org/
The Scleroderma Research Foundation, which helped support this work:
http://www.srfcure.org/
This work was supported by a grant from the Scleroderma Research Foundation to PJGC. The authors wish to thank Ms. Ederick Forbes for her technical assistance. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Citation: Liu X, Zhu S, Wang T, Hummers L, Wigley FM, et al. (2005) Paclitaxel modulates TGFβ signaling in scleroderma skin grafts in immunodeficient mice. PLoS Med 2(12): e354.
Abbreviations
Ctthreshold cycle
ECMextracellular matrix
EPCendothelial progenitor cell
IHCimmunohistochemistry
MTmicrotubule
R-Smadreceptor-regulated Smad
SCIDsevere combined immunodeficient
SScsystemic sclerosis
TRT-PCRTaqMan Real-Time Reverse Transcription-PCR
VEGFvascular endothelial growth factor
VVMVerhoeff's van Gieson elastin and Mason trichrome
==== Refs
References
LeRoy EC A brief overview of the pathogenesis of scleroderma (systemic sclerosis) Ann Rheum Dis 1992 51 286 288 1550420
Kissin EY Korn JH Fibrosis in scleroderma Rheum Dis Clin North Am 2003 29 351 369 12841299
Kahaleh MB LeRoy EC Autoimmunity and vascular involvement in systemic sclerosis (SSc) Autoimmunity 1999 31 195 214 10739336
Branton MH Kopp JB TGF-beta and fibrosis Microbes Infect 1999 1 1349 1365 10611762
Wigley FM Sule SD Novel therapy in the treatment of scleroderma Expert Opin Investig Drugs 2001 10 31 48
Pollman MJ Naumovski L Gibbons GH Vascular cell apoptosis: Cell type-specific modulation by transforming growth factor-beta1 in endothelial cells versus smooth muscle cells Circulation 1999 99 2019 2026 10209007
Derynck R Zhang YE Smad-dependent and Smad-independent pathways in TGF-beta family signalling Nature 2003 425 577 584 14534577
Dong C Li Z Alvarez R Feng XH Goldschmidt-Clermont PJ Microtubule binding to Smads may regulate TGF beta activity Mol Cell 2000 5 27 34 10678166
Dong C Zhu S Wang T Yoon W Li Z Deficient Smad7 expression: A putative molecular defect in scleroderma Proc Natl Acad Sci U S A 2002 99 3908 3913 11904440
Subcommittee for scleroderma criteria of the American Rheumatism Association Diagnostic and Therapeutic Criteria Committee Preliminary criteria for the classification of systemic sclerosis (scleroderma) Arthritis Rheum 1980 23 581 590 7378088
Scheja A Wildt M Wollheim FA Akesson A Saxne T Circulating collagen metabolites in systemic sclerosis. Differences between limited and diffuse form and relationship with pulmonary involvement Rheumatology (Oxford) 2000 39 1110 1113 11035131
Sgonc R The vascular perspective of systemic sclerosis: Of chickens, mice and men Int Arch Allergy Immunol 1999 120 169 176 10592461
Lakos G Takagawa S Chen S Ferreira AM Han G Targeted disruption of TGF-beta/Smad3 signaling modulates skin fibrosis in a mouse model of scleroderma Am J Pathol 2004 165 203 217 15215176
Yang X Letterio JJ Lechleider RJ Chen L Hayman R Targeted disruption of SMAD3 results in impaired mucosal immunity and diminished T cell responsiveness to TGF-beta EMBO J 1999 18 1280 1291 10064594
Feinberg MW Shimizu K Lebedeva M Haspel R Takayama K Essential role for Smad3 in regulating MCP-1 expression and vascular inflammation Circ Res 2004 94 601 608 14752027
Zhao J Shi W Wang YL Chen H Bringas P Smad3 deficiency attenuates bleomycin-induced pulmonary fibrosis in mice Am J Physiol Lung Cell Mol Physiol 2002 282 L585 593 11839555
Flanders KC Smad3 as a mediator of the fibrotic response Int J Exp Pathol 2004 85 47 64 15154911
Zhu S Goldschmidt-Clermont PJ Dong C Transforming growth factor-beta-induced inhibition of myogenesis is mediated through Smad pathway and is modulated by microtubule dynamic stability Circ Res 2004 94 617 625 14739161
Brahn E Tang C Banquerigo ML Regression of collagen-induced arthritis with taxol, a microtubule stabilizer Arthritis Rheum 1994 37 839 845 7911665
Cao L Sun D Cruz T Moscarello MA Ludwin SK Inhibition of experimental allergic encephalomyelitis in the Lewis rat by paclitaxel J Neuroimmunol 2000 108 103 111 10900343
LeRoy EC Systemic sclerosis. A vascular perspective Rheum Dis Clin North Am 1996 22 675 694 8923590
Rauscher FM Goldschmidt-Clermont PJ Davis BH Wang T Gregg D Aging, progenitor cell exhaustion, and atherosclerosis Circulation 2003 108 457 463 12860902
Jiang Y Jahagirdar BN Reinhardt RL Schwartz RE Keene CD Pluripotency of mesenchymal stem cells derived from adult marrow Nature 2002 418 41 49 12077603
Goldschmidt-Clermont PJ Loss of bone marrow-derived vascular progenitor cells leads to inflammation and atherosclerosis Am Heart J 2003 146 S5 12 14564300
Kuwana M Okazaki Y Yasuoka H Kawakami Y Ikeda Y Defective vasculogenesis in systemic sclerosis Lancet 2004 364 603 610 15313361
Distler O Distler JH Scheid A Acker T Hirth A Uncontrolled expression of vascular endothelial growth factor and its receptors leads to insufficient skin angiogenesis in patients with systemic sclerosis Circ Res 2004 95 109 116 15178641
Distler O Del Rosso A Giacomelli R Cipriani P Conforti ML Angiogenic and angiostatic factors in systemic sclerosis: Increased levels of vascular endothelial growth factor are a feature of the earliest disease stages and are associated with the absence of fingertip ulcers Arthritis Res 2002 4 R11 12453314
Edelberg JM Tang L Hattori K Lyden D Rafii S Young adult bone marrow-derived endothelial precursor cells restore aging-impaired cardiac angiogenic function Circ Res 2002 90 E89 E93 12039806
Del Papa N Colombo G Fracchiolla N Moronetti LM Ingegnoli F Circulating endothelial cells as a marker of ongoing vascular disease in systemic sclerosis Arthritis Rheum 2004 50 1296 1304 15077314
Thatte U Bagadey S Dahanukar S Modulation of programmed cell death by medicinal plants Cell Mol Biol (Noisy-le-grand) 2000 46 199 214 10726985
|
16250671
|
PMC1274282
|
CC BY
|
2021-01-05 10:39:26
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e354
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020354
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 1625561910.1371/journal.pmed.0020359Policy ForumInfectious DiseasesEpidemiology/Public HealthHealth PolicyInfectious DiseasesMicrobiologyPublic HealthA Systematic Analytic Approach to Pandemic Influenza Preparedness Planning Policy ForumBarnett Daniel J *Balicer Ran D Lucey Daniel R Everly George S JrOmer Saad B Steinhoff Mark C Grotto Itamar Daniel J. Barnett and Ran D. Balicer made an equal contribution to the development of this manuscript. Daniel J. Barnett is at the Johns Hopkins Center for Public Health Preparedness, Department of Environmental Health Sciences, Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland, United States of America. Ran D. Balicer is in the Epidemiology Department, Faculty of Health Sciences, Ben-Gurion University of the Negev, Be'er Sheva, Israel. Daniel R. Lucey is in the Department of Microbiology and Immunology, Georgetown University School of Medicine, Washington, District of Columbia, United States of America. George S. Everly, Jr., is at the Johns Hopkins Center for Public Health Preparedness, Department of Psychiatry, Johns Hopkins School of Medicine, Baltimore, Maryland, United States of America. Saad B. Omer is in the Department of International Health, Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland, United States of America. Mark C. Steinhoff is in the Department of International Health and the Department of Epidemiology, Johns Hopkins Bloomberg School of Public Health, and the Department of Pediatrics, Johns Hopkins School of Medicine. Itamar Grotto is in the Epidemiology Department, Faculty of Health Sciences, Ben-Gurion University of the Negev, Be'er Sheva, Israel.
Competing Interests: The codevelopment of this manuscript by the Johns Hopkins Center for Public Health Preparedness has been supported in part through cooperative agreement U90/CCU324236-01 with the Centers for Disease Control and Prevention (http://www.cdc.gov/). All aspects of all authors' work were independent of the funding source.
*To whom correspondence should be addressed. E-mail: [email protected] 2005 1 11 2005 2 12 e359Copyright: © 2005 Barnett et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.The World Health Organization warns that a flu pandemic is inevitable, and possibly imminent. Barnett and colleagues discuss a tool called the Haddon Matrix that could help in pandemic influenza planning.
==== Body
The prospect of a pandemic with avian influenza is an urgent concern for public health leaders worldwide [1]. As pathogenic avian influenza A (H5N1) strains (Figure 1) continue to spread in East Asia, with recently reported expansion to Siberia and westward regions in Russia [2,3] as well as to migratory birds [4,5], the risk for reassortment of avian and human strains increases. Evidence cited by the World Health Organization in May 2005 suggests that H5N1 may be adapting to humans, thus potentially setting the stage for the next influenza pandemic [6].
Figure 1 Colorized Transmission Electron Micrograph of Avian Influenza A H5N1 Viruses Grown in MDCK Cells
The viruses are gold, and the MDCK cells are green.
(Photo: CDC/C. Goldsmith, J. Katz, and S. Zaki)
Animal data suggest that the current H5N1 strain appears to be even more deadly than the original 1997 Hong Kong avian influenza, a finding that correlates well with the observed human case fatality rates [7]. As of August 5, 2005, there have been 112 human cases of H5N1 in East Asia resulting in 57 deaths (case fatality rate = 51%) [8,9]. Also concerning are recent findings that in China and Indonesia the virus has infected pigs, a possible “mixing vessel” for both avian and human influenza viruses, thus providing an opportunity for reassortment from which a pandemic human strain could emerge [10,11]. Research suggesting that cats could host or transmit the H5N1 infection [12] adds to a worrisome picture of multispecies transmission that can elevate the risk of reassortment [9]. This epizootic outbreak in Asia is not expected to wane in the short term [9].
Influenza pandemics can have devastating impacts. The Spanish flu of 1918 was particularly destructive (Figures 2 and 3), resulting in a higher death total in less than two years than in all of World War I [13]. Although earlier accounts suggested the mortality from the 1918 pandemic was 20 million to 40 million, more recent assessments including new estimates from Africa and Asia suggest that a more realistic figure is 50–100 million [14]. The high rates of infection with the pandemic virus meant that even an average case fatality rate lower than 3% resulted in this large number of deaths [13,15]. A 1918-type influenza pandemic today is projected to cause 180–360 million deaths globally (including 1.7 million deaths in the United States) [1], with transmission of the disease lasting at least two years [16].
Figure 2 Emergency Hospital during 1918 Influenza Epidemic, Camp Funston, Kansas, United States
(Photo: Image “NCP 1603,” National Museum of Health and Medicine, Armed Forces Institute of Pathology, Washington, D.C.)
Figure 3 Historic Chart Showing Mortality Rates in America and Europe during 1918 and 1919
(Photo: Image “Reeve 3143,” National Museum of Health and Medicine, Armed Forces Institute of Pathology, Washington, D.C.)
The Next Pandemic: “Inevitable, and Possibly Imminent”
In light of recent episodes of human infection with H5N1 virus, the World Health Organization reiterated its 1997 call for all countries to prepare for the next pandemic, which it termed “inevitable, and possibly imminent” [17], and updated its own pandemic plan in April 2005 [18]. In the United States, it has been argued that of the 12 disaster scenarios recently assessed by the US Department of Homeland Security, pandemic influenza is the most likely and perhaps the most deadly [19]. A draft form of the US pandemic influenza plan was made public in August 2004 [20], and an updated plan is anticipated by September 2005.
The urgent need for comprehensive pandemic influenza planning is profound: an influenza pandemic starting today may have major international consequences, including global economic and political destabilization, an overwhelming of health care resources, and panic [21]. Current international plans [18,22], while useful, could benefit from enhanced detail [21] and organization; moreover, pandemic influenza plans have usually been national in scope and, in most countries, are only in a draft form and lack legal status [23].
The Haddon Matrix
An analytic approach for traffic safety injury epidemiology and prevention was developed by Dr. William Haddon, Jr. in the 1960s [24], and has since been termed “the Haddon matrix.” This matrix provides a multidimensional approach to understanding the contributing factors to injury before, during, and after an event [25]. The current version of the matrix is a grid with four columns, or axes, that represent contributing factors to injury (host, agent/vector, physical environment, sociocultural environment) and three rows that correspond to the time phases of a given form of injury (pre-event, event, and post-event) [26]. By compartmentalizing an injury into dimensions of time and contributing factors, the matrix can break a complex problem into more manageable segments. For each of the 12 cells, a decision analysis or prioritization can be used to select policies or actions with greatest feasibility or influence [27].
Although the Haddon matrix may seem unfamiliar to some infectious disease scientists, it incorporates familiar analytic elements in a systematic way. The four columns represent the classical epidemiologic triad of host, agent, and environment (physical and sociocultural). The three rows are equivalent to primary, secondary, and tertiary prevention of disease outbreaks. Indeed, Haddon himself used his analytic matrix to describe an outbreak of polio [24], and this matrix has been recently applied to other public-health emergency preparedness challenges such as SARS [28].
Applying the Matrix to Pandemic Influenza Preparedness
Comprehensive public health emergency preparedness and response efforts require effective pre-event (preventive), event (mitigation), and post-event (consequence management) strategies. By identifying the factors that may modify the outcome in each of these phases, one can prescribe the appropriate measures necessary to tackle each factor.
To this end, we specifically applied the Haddon matrix to pandemic influenza planning and response (Table 1), systematically identifying relevant factors in each phase (pre-event, event, post-event) and on each axis (human, agent/vector, physical environment, sociocultural environment). We then identified factors that may be associated with opportunities for public health intervention, and marked these factors in bold within the matrix (consistent with an approach described by Runyan [27]).
Table 1 The Haddon Matrix and Pandemic Avian Influenza
Items in bold are potential targets for public health intervention.
Table 1 Continued
The table shows that in all phases of an influenza pandemic, opportunities for public health intervention include a number of contributing human, physical environment, and sociocultural factors, but generally not agent/vector factors, since viruses generally cannot be modified easily as injury-causing devices. Importantly, the pre-event, event, and post-event rows of the matrix reflect the phase of a pandemic in which public health preparedness and response measures will take their effects; however, planning for each of these measures must occur before the pandemic begins.
The use of the Haddon matrix in the table as an analytic and planning tool for pandemic influenza is illustrated below by its application to readiness efforts in two different countries: Thailand, focusing on pre-event factors; and Israel, focusing on event factors. We chose Thailand as an example because of its regional susceptibility and the proactive nature of its anti-H5N1 planning efforts to date. We selected Israel as an example of a country outside of East Asia that has taken steps to address this potential global crisis. For both countries, we demonstrate the application of the matrix by addressing selected factors within each axis.
Pandemic Influenza Planning in Thailand
Thailand has had experience with H5N1 infections in both humans and animal populations—including chickens, ducks, birds, fighting roosters, and tigers—since January 2004. By October 2004, a total of 17 patients with H5N1 infection were identified, of whom 12 had died. The initial success of Thailand's national program against H5N1 avian influenza that began in the autumn of 2004 is evidenced by the fact that no human cases have been found between October 2004 and August 2005. Thus, Thailand's experience may offer practical lessons in preparing for an avian influenza-related human pandemic.
Through the lens of pre-event Haddon matrix factors, one can identify the strengths in Thailand's preparedness efforts, as well as opportunities for further enhancements. Selected examples of the pre-event axes for Thailand's pandemic influenza readiness efforts are described below.
Pre-event human factors
Thailand has developed surveillance and laboratory testing algorithms for influenza-like-illness in humans and animals, including definitions for “suspect,” “probable,” “confirmed,” “excluded,” and “on investigation” cases of H5N1. With written guidance from national authorities [29], public health workers, veterinary health workers, village health volunteers, and others [30] participated in an ongoing surveillance campaign nationwide beginning in October 2004 [31].
Pre-event risk communication to at-risk populations are also important. In the scenario of pandemic influenza, effective pre-event risk communication can reduce event-phase risk communication barriers [32]. An array of appropriate information on avian influenza and potential pandemic human influenza has been disseminated by the Thai Ministry of Public Health [33].
Pre-event agent/vector factors
Strain pathogenicity to its avian and human hosts is the major pre-event agent/vector factor. Most cases of human H5N1 infection have resulted from contact with infected chickens, fighting roosters, or ducks [9], with some ducks possibly being asymptomatic [34]. Regarding human pathogenicity, an autopsy of a patient from Thailand, one of the few involving H5N1 infection [35,36] reported that the virus can replicate in the human intestine as well as the lung [37] perhaps helping to explain the finding of diarrhea in some patients in Thailand and Vietnam [37–41].
Pre-event physical environment factors
Thailand has established a multifaceted communication system, including websites for human and animal-related H5N1 updates and standard protocols. Provincial health offices were directed by the Ministry of Public Health to form Surveillance and Rapid Response Teams at the provincial and district levels [42].
Hospital infection control infrastructure and protocols are also crucial. Patients meeting criteria for possible H5N1 infection “should be isolated and placed in a single room according to the standard precautions of the Ministry of Public Health” [42]. Even if the patient's initial rapid test for influenza A is negative “the patient must be treated with antivirals immediately” [42] in an effort to increase survival [41].
The availability of avian strain-specific vaccines is another significant factor. Webster and Hulse observed that Thailand's investigation of flu vaccines for “open range” (noncommercial) poultry represents a “prudent” policy shift that should be replicated in other countries in East Asia [43]. H5N1 vaccine studies in humans have not yet been initiated in Thailand.
Pre-event sociocultural factors
One of the most significant factors is political and social willingness to acknowledge and report disease dissemination. Initially, the Thai government was criticized for underplaying the existence and the magnitude of avian influenza in Thailand [44] but it has since taken significant proactive steps to address this urgent challenge. Between January 2004 and July 2005, a total of 59 official reports on surveillance for Highly Pathogenic Avian Influenza have been submitted to the World Organization for Animal Health by Thailand [45], and detailed reports were promptly published [35,36,38,41,46–48]. On September 28, 2004, the first media report of a probable case of person-to-person transmission appeared in Thailand [49] and was rapidly published [36]. On September 29 of that year, a national campaign against the H5N1 virus was declared by the Prime Minister of Thailand, with involvement by the Thai Cabinet [50,51]. These resulting efforts seem to have had a substantial impact, as detailed above.
Budget (preparedness resource allocation) is also important. The Thai National Strategic Plan for Avian Influenza and Plan for Pandemic Preparedness 2005–2007 was initiated with a budget of 4,026 million Thai baht (∼US$105 million) [52,53]. Thailand has been reported recently to have approved funding for the future purchase of up to 100,000 treatments of oseltamivir [54].
An in-place culling policy played a significant role. The culling of ducks (with farmer compensation) reduced the flocks that were positive for H5N1 from around 40% infected in October 2004 to almost undetectable levels in March 2005 [43].
Collaboration between human and veterinary health authorities was vital. Efforts are ongoing to closely link public health and animal health responses to H5N1 [52,53]. Surveillance combines epidemiologically linked testing for animals and humans [55]. In addition, Thailand interacts frequently with the World Health Organization regarding clinical H5N1 issues, and with the World Organization for Animal Health in reporting on animal surveillance for H5N1 [56].
Pandemic Influenza Planning in Israel
Applying the various influencing factors listed in the event phase of the Haddon matrix to the unique Israeli setting leads to several important insights regarding local pandemic preparedness, as shown in the following examples.
Event human factors
Israel has not initiated, as of yet, training activities for health care professionals directed specifically at pandemic preparedness, although such activities are planned to take place. Nevertheless, Israeli health care professionals, particularly frontline health care workers, are well experienced with terrorism-related mass casualty emergencies. Continuous training of the various components of the health care system for bioterrorism threats likely serves to enhance these workers' ability to deal with naturally occurring epidemic threats; these health care teams were shown to have increased likelihood of reporting to duty during a crisis [57]. Simulation-assisted medical training may be useful in increasing health care workers' compliance with personal protective equipment and infection control protocols, as has been shown in the preparation of Israeli medical teams to respond to chemical warfare casualties [58]. Upcoming tabletop exercises will test and refine current national contingency plans, while full-scale drills may be required to test certain practical and logistical aspects of antiviral drug dissemination.
Event agent/vector factors
Agent/vector factors listed in the matrix are expected to determine much of the local impact of the pandemic, but they generally cannot be influenced by preparedness and mitigation efforts. As these factors will remain unknown until the first stages of the pandemic, Israeli preparedness planners have taken into account a wide range of scenarios with different attack and mortality rates [59] in addressing issues such as surge capacity. For instance, a highly transmissible pandemic may render isolation and quarantine efforts largely futile [60] while a less transmissible strain, as witnessed in previous pandemics [61] may enable a containment approach more similar to that taken during the SARS epidemic (while accounting for considerable differences such as the incubation time or the impact of infectious asymptomatic cases). A highly pathogenic strain, perhaps more pathogenic than the 1918 strain (considering current case fatality rates of H5N1 human cases), will require the unparalleled ability to rapidly mobilize medical equipment and personnel to meet the increased demands for care in both primary and secondary care facilities. However, a less pathogenic strain may require measures similar to those taken during severe seasonal influenza epidemics.
Event physical environment factors
The availability of an effective immunization will be crucial. The importance of the recently published successful preliminary results of phase-I human H5N1 vaccine trials cannot be overestimated [62]. Nevertheless, both the safety and efficacy of the new vaccine remain to be assessed, and the effectiveness of this vaccine against a reassortant pandemic strain is currently difficult to predict.
Research efforts to produce active or passive immunization that will be universally effective against any influenza strain are currently underway in Israel and elsewhere. Once available, such modalities hold great promise for mitigation of future pandemics in their first stages [63]. Another type of immunotherapy that may be considered during an event is the use of immunoglobulins isolated from recovered patients to treat the ill or protect the exposed.
Stockpiled antivirals and antibiotics are important to Israel's preparedness. The Israeli Ministry of Health has successfully used cost-benefit analyses [59] to persuade decision makers to invest the funds necessary for the rapid creation of a national antiviral stockpile, and several strategies for the use of these drugs during the pandemic are considered [64]. The antiviral oseltamivir was found to be effective in mice against the newest strains of avian influenza currently sweeping through East Asia, suggesting that higher doses and prolonged courses of this drug may be required [7]. These findings, if validated in humans, may need to be factored into stockpiling planning efforts.
Event sociocultural factors
Israel has ensured that a legal and ethical framework for implementation of response measures exists. Including pandemic influenza in the list of “dangerous communicable diseases” defined by Israeli law will allow the Ministry of Health to uphold extreme measures such as involuntary quarantine and isolation, if needed.
Prioritizing target groups for antiviral drugs and vaccines, expected to be in short supply, requires the addressing of complex ethical, legal, social, and political considerations. The choice of which groups to prioritize would derive, in part, from the prioritizing of the various goals in using these drugs. If the focus is on reducing all mortality, different groups may be prioritized than if the main attempt is to reduce social disruption. A national ethics committee was recently appointed to address these issues.
Conclusions
By offering phase-specific insights into pandemic influenza planning, the Haddon matrix bridges injury-prevention epidemiology with global infectious disease preparedness and response. In the process, this analytic tool sheds light on opportunities for prevention, mitigation, and consequence management strategies to address a global public health threat.
In the face of the challenges described, the Haddon matrix analysis of pandemic influenza planning in Thailand and Israel reflects its applicability as a systematic tool for identifying urgent national and international pandemic avian influenza readiness needs. The scalability of the matrix also allows its use at the level of a county or city, as well as within institutions. At each of these levels, the matrix may facilitate the enhancement of preparedness plans, needs assessments, best practice identification, and resource distribution strategies.
Although the national examples above have selectively focused on pre-event factors in Thailand and event factors in Israel, the Haddon matrix can be also used to augment existing post-event phase plans. For example, the psychology of post-event reactions [65] must be addressed through ongoing mental health support and follow up and by effective post-event risk communication. The public health infrastructure may face the dual challenge of helping populations, including health care providers themselves, to be psychologically prepared for the next wave of a pandemic—perhaps worse than the first wave, as was the case in the 1918 pandemic [13]—while trying to recover from the first wave.
The Haddon matrix has limitations that must be recognized to ensure appropriate implementation. Importantly, the matrix is not a stand-alone planning tool; rather, the results of any Haddon-matrix–based analysis must be operationalized in the form of policies and procedures to achieve their desired effects on the factors included in the matrix. Moreover, the matrix is not static; the contents within its cells can and should be modified according to changing disease dynamics and situational challenges to maintain its usefulness in an evolving crisis.
Furthermore, even before a crisis, the choice of contents for each cell is not absolute, and open to the subjective interpretation of those who are preparing the matrix. Consequently, the table presented in this article should be regarded as a starting planning framework, not a final checklist. Also, while many of the items in the Haddon matrix cells may be measurable, the matrix itself is only a planning instrument—not an evaluation tool.
The known potential for an avian influenza pandemic offers not only challenges but also unprecedented opportunities for advance planning at all levels of public health in the international community [66]. This planning window may be rapidly closing, however [21]. As an efficient yet comprehensive analytic approach, the Haddon matrix lends itself to the types of rapid and complex decision making necessary to plan for and respond more effectively to an urgent pandemic health threat.
Citation: Barnett DJ, Balicer RD, Lucey DR, Everly GS Jr, Omer SB, et al. (2005) A systematic analytic approach to pandemic influenza preparedness planning. PLoS Med 2(12): e359.
Abbreviation
H5N1avian influenza A
==== Refs
References
Osterholm MT Preparing for the next pandemic N Engl J Med 2005 352 1839 1842 15872196
Russian News and Information Agency Bird flu spreads from Western Siberia to South Urals 2005 August 17 Available: http://en.rian.ru/russia/20050817/41175461.html . Accessed 19 August 2005
Coulombier D Paget J Meijer A Ganter B Highly pathogenic avian influenza reported to be spreading into western Russia. EuroSurveillance Weekly 2005 August 10 Available: http://www.eurosurveillance.org/ew/2005/050818.asp#1 . Accessed 19 August 2005
Liu J Xiao H Lei F Zhu Q Qin K Highly pathogenic H5N1 influenza virus infection in migratory birds Science 2005 309 1206 16000410
Chen H Smith GJD Zhang SY Qin K Wang J Avian flu: H5N1 virus outbreak in migratory waterfowl Nature 436 191 192
World Health Organization WHO intercountry-consultation 2005 Influenza A/H5N1 in humans in Asia May 6–7, 2005 Manila, Philippines Available: http://www.who.int/csr/resources/publications/influenza/WHO_CDS_CSR_GIP_2005_7_04.pdf . Accessed 27 June 2005
Yen HL Monto AS Webster RG Govorkova EA Virulence may determine the necessary duration and dosage of oseltamivir treatment for highly pathogenic A/Vietnam/1203/04 influenza virus in mice J Infect Dis 2005 192 665 672 16028136
World Health Organization Communicable disease surveillance and response (CSR): Avian influenza 2005 Available: http://www.who.int/csr/disease/avian_influenza/en/ . Accessed 19 August 2005
Centers for Disease Control and Prevention Recent avian influenza outbreaks in Asia 2005 August 5 Available: http://64.233.161.104/search?q=cache:AfzN0eTmN04J:www.cdc.gov/flu/avian/outbreaks/asia.htm+August+5,+2005+H5N1+avian+influenza+deaths&hl=en . Accessed 19 August 2005
Cyranoski D Bird flu spreads among Java's pigs Nature 2005 435 390 391
Cyranoski D Bird flu data languishes in Chinese journals Nature 2004 430 955
Kuiken T Rimmelzwaan G van Riel D van Amerongen G Baars M Avian H5N1 influenza in cats Science 2004 306 241 15345779
Kolata G Flu: The story of the great influenza pandemic of 1918 and the search for the virus that caused it 1999 New York Farrar, Straus, and Giroux 330
Johnson NP Mueller J Updating the accounts: Global mortality of the 1918–1920 “Spanish” influenza pandemic Bull Hist Med 2002 76 105 115 11875246
Barry JM The great influenza: The epic story of the deadliest plague in history 2004 New York Viking Penguin 546
World Health Organization Avian influenza: Assessing the pandemic threat 2005 Available: http://www.who.int/csr/disease/influenza/H5N1-9reduit.pdf . Accessed 29 June 2005
[Anonymous] World is ill-prepared for “inevitable” flu pandemic Bull World Health Organ 2004 82 317 318 15309817
World Health Organization WHO global influenza preparedness plan: The role of WHO and recommendations for national measures before and during pandemics 2005 Available: http://www.who.int/csr/resources/publications/influenza/en/WHO_CDS_CSR_GIP_2005_5.pdf . Accessed 27 June 2005
Lipsitch M Pandemic flu: We are not prepared Med Gen Med 7 2005 Available: http://www.medscape.com/viewarticle/502709 . Accessed 27 June 2005
United States Department of Health and Human Services Pandemic influenza response and preparedness plan 2004 Available: http://www.dhhs.gov/nvpo/pandemicplan/ . Accessed 7 July 2005
Osterholm MT Preparing for the next pandemic Foreign Aff 84 2005 Available: http://www.foreignaffairs.org/20050701faessay84402/michael-t-osterholm/preparing-for-the-next-pandemic.html . Accessed 27 June 2005
World Health Organization WHO checklist for influenza pandemic preparedness planning 2005 Available: http://www.who.int/csr/resources/publications/influenza/CDS_CSR_GIP_2005_4.pdf . Accessed 26 July 2005
Abbott A Avian flu special: What's in the medicine cabinet? Nature 2005 26 407 409
Haddon W The changing approach to the epidemiology, prevention, and amelioration of trauma: The transition to approaches etiologically rather than descriptively based Am J Public Health Nations Health 1968 58 1431 1438 5691377
Runyan CW Introduction: Back to the future—Revisiting Haddon's conceptualization of injury epidemiology and prevention Epidemiol Rev 2003 25 60 64 12940231
Haddon W Advances in the epidemiology of injuries as a basis for public policy Public Health Rep 1980 95 411 421 7422807
Runyan CW Using the Haddon matrix: Introducing the third dimension Inj Prev 1998 4 302 307 9887425
Barnett DJ Balicer RD Blodgett D Fews AL Parker CL The application of the Haddon matrix to public health readiness and response planning Environ Health Perspect 2005 113 561 566 15866764
Bureau of General Communicable Diseases. Department of Disease Control MOPH Thailand Avian influenza surveillance in human as at November 4, 2004 2005 Available: http://thaigcd.ddc.moph.go.th/AI_case_report_041104.html . Accessed 30 June 2005
Department of Livestock Development. Ministry of Agriculture of Cooperatives Thailand Overall operation 2003 Available: http://www.dld.go.th/home/bird_flu/AI_resp.html . Accessed 30 June 2005
Bureau of Epidemiology Department of Disease Control Ministry of Public Health Avian influenza surveillance in humans as of July 4, 2005 2005 Available: http://thaigcd.ddc.moph.go.th/AI_case_report_050704.html . Accessed 7 July 2005
United States Department of Health and Human Services Draft pandemic influenza preparedness and response plan 2005 Annex 9: Communication and education. Available: http://www.hhs.gov/nvpo/pandemicplan/annex9.communication.pdf . Accessed 29 June 2005
Bureau of General Communicable Diseases. Department of Disease Control MOPH Thailand Avian influenza (bird flu) control 2005 Available: http://thaigcd.ddc.moph.go.th/Bird_Flu_main_en.html . Accessed 30 June 2005
World Health Organization Avian influenza—Situation in Asia: Altered role of domestic ducks 2005 Available: http://www.who.int/csr/don/2004_10_29/en/index.html . Accessed 30 June 2005
Puthavathana P Auewarakul P Charoenying PC Sangsiriwut K Pooruk P Molecular characterization of the complete genome of human influenza H5N1 virus isolates in Thailand J Gen Virol 2005 86 423 433 15659762
Ungchusak K Auerwarakul P Dowell SF Kitphati R Auwanit W Probable person-to-person transmission of avian influenza (H5N1) N Engl J Med 2005 352 333 340 15668219
To KF Chan PK Chan KF Lee WK Lam WY Pathology of fatal human infection associated with avian influenza A H5N1 virus J Med Virol 2001 63 242 246 11170064
Uiprasertkul M Puthavathana P Sangsiriwut K Pooruk P Srisook K Influenza A H5N1 replication sites in humans 2005 Available: http://www.cdc.gov/ncidod/EID/vol11no07/04-1313.htm . Accessed 7 July 2005
de Jong MD Bach VC Phan TQ Vo MH Tran TT Fatal avian influenza A (H5N1) in a child presenting with diarrhea followed by coma N Engl J Med 2005 352 686 691 15716562
Tran TH Nguyen TL Nguyen TD Luong TS Pham PM Avian influenza A (H5N1) in 10 patients in Vietnam N Engl J Med 2004 350 1179 1188 14985470
Chotpoitayasunondh T Ungchusak K Hanshaoworakul W Chunsuthiwat S Sawanpanyalert P Human disease from influenza A (H5N1), Thailand, 2004 Emerg Infect Dis 2005 11 201 209 15752436
Ministry of Public Health Avian influenza: Prevention and control measures in humans continuing activities from November 2004 to February 2005 2005 Available: http://thaigcd.ddc.moph.go.th/AI_control_measure_041108.html . Accessed 15 July 2005
Webster R Hulse D Controlling avian flu at the source Nature 2005 435 415 416 15917777
Sipress A Thailand concedes missteps on bird flu The Washington Post 2004 January 29 Available: http://www.washingtonpost.com/ac2/wp-dyn?pagename=article&contentId=A58049-2004Jan28¬Found=true . Accessed 19 August 2005
World Organization for Animal Health Update on avian influenza in animals in Asia (type H5) 2005 Available: http://www.oie.int/downld/a_ai-asia.htm . Accessed 7 July 2005
Grose C Chokephaibulkit K Avian influenza virus infection of children in Vietnam and Thailand Pediatr Infect Dis J 2004 23 793 794 15295239
Chokephaibulkit K Uiprasertkul M Puthavathana P Chearskul P Auewarakul P A child with avian influenza A (H5N1) infection Pediatr Infect Dis J 2005 24 162 166 15702046
Centers for Disease Control and Prevention Cases of influenza A (H5N1)—Thailand, 2004 Morb Mortal Wkly Rep 2004 53 100 103
Sathirawattanakul D Bird flu alert: Human transmission probable The Nation 2004 September 29 Available: http://www.nationmultimedia.com/search/page.arcview.php?clid=2&id=106795&usrsess . Accessed 6 July 2005
Songklin P Sathirawattanakul D Grappling with fear 2004 September 30 Cabinet given bird-flu deadline. New York: The Nation. Available: http://www.nationmultimedia.com/search/page.arcview.php?clid=2&id=106814&usrsess Accessed 6 July 2005
Avian Influenza Control Operating Centre Department of Livestock Development Situation of highly pathogenic avian influenza (HPAI) of H5N1 subtype re-occurrence and control measures in Thailand (3 July–30 September 2004) 2004 Available: http://www.dld.go.th/home/bird_flu/return/HPAI(3Jul-30Sep04).html . Accessed 14 July 2005
Ungchusak K Concerns raised by pandemic influenza 2005 Available: http://www.who.int/csr/disease/influenza/ungchusak.pdf . Accessed 27 June 2005
Tourism Authority of Thailand Crisis Communication Centre Updates 2005 http://www.tatnews.org/ccc/2480.asp . Accessed 30 June 2005
Sipress A Countries hit by bird flu have little medicine to treat humans The Washington Post 2005 July 6 Available: http://www.washingtonpost.com/wp-dyn/content/article/2005/07/05/AR2005070501422.html?nav=rss_world/asia . Accessed 19 August 2005
Bureau of General Communicable Diseases. Department of Disease Control MOPH Thailand Standard operating protocol dealing with patients with avian influenza surveillance 2004 Available: http://thaigcd.ddc.moph.go.th/Bird_Flu_main_en.html . Accessed 6 July 2005
World Organization for Animal Health Update on avian influenza in animals in Asia (type H5) 2005 Available: http://www.oie.int/downld/a_ai-asia.htm . Accessed 6 July 2005
Shapira Y Marganitt B Roziner I Shochet T Bar Y Willingness of staff to report to their hospital duties following an unconventional missile attack: A state-wide survey Isr J Med Sci 1991 27 704 711 1757251
Vardi A Levin I Berkenstadt H Hourvitz A Eisenkraft A Simulation-based training of medical teams to manage chemical warfare casualties Isr Med Assoc J 2002 4 540 544 12120468
Balicer RD Huerta M Davidovitch N Grotto I Cost-benefit of stockpiling drugs for influenza pandemic Emerg Infect Dis 2005 11 1280 1282 16102319
Fraser C Riley S Anderson RM Ferguson NM Factors that make an infectious disease outbreak controllable Proc Natl Acad Sci U S A 2004 101 6146 6151 15071187
Mills CE Robins JM Lipsitch M Transmissibility of 1918 pandemic influenza Nature 2004 432 904 906 15602562
Enserink M Avian influenza: ‘Pandemic vaccine’ appears to protect only at high doses Science 2005 309 996
Fedson DS Preparing for pandemic influenza: An international policy agenda for vaccine development J Public Health Policy 2005 26 4 29 15906873
Balicer RD Huerta M Grotto I Tackling the next influenza pandemic BMJ 2004 328 1391 1392 15191958
Everly GS Lating JM Personality-guided therapy of posttraumatic stress disorder 2003 Washington (District of Columbia) American Psychological Association 267
Fauci AS Race against time Nature 2005 435 423 424 15917781
|
16255619
|
PMC1274283
|
CC BY
|
2021-01-05 11:13:37
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e359
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020359
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 10.1371/journal.pmed.0020402SynopsisGenetics/Genomics/Gene TherapyOpthalmologyGene TherapyOphthalmologyTackling Inherited Blindness Synopsis11 2005 1 11 2005 2 11 e402Copyright: © 2005 Public Library of Science.2005This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.
Pharmacological and rAAV Gene Therapy Rescue of Visual Functions in a Blind Mouse Model of Leber Congenital Amaurosis
==== Body
Imagine the eye is like a camera. The shutter, like the iris of the eye, opens and closes to let in the right amount of light. The lens helps focus light on the film. And the film is like the retina. Regardless of the quality of the camera, if the film is faulty, the developed pictures may be distorted or blurred. In this way, untreatable degenerative diseases of the retina, which affect millions of people worldwide, lead to varying degrees of irreversible blindness. These degenerative eye disorders include retinitis pigmentosa, which affects 1.5 million people, and age-related macular degeneration, which is a leading cause of blindness in North America. The list of inherited retinal dystrophies (degenerations) is long and includes Best disease, choroideremia, cone–rod dystrophy, congenital stationary night blindness, and Leber congenital amaurosis (LCA).
LCA is a collection of diseases all characterized by severe loss of vision at birth from retinal dysfunction. It is a leading cause of congenital blindness. Currently, there is no treatment for LCA; however, it is known that LCA can be caused by mutations in the gene encoding RPE65, a key protein involved in the production and recycling of the chromophore 11-cis-retinal (11-cis-RAL) in the eye. 11-cis-RAL is an integral part of rhodopsin and cone visual pigments, pigments essential for our vision. About 15% of patients with LCA have mutations in RPE65. Humans with this form of LCA and Rpe65-deficient mice models both have severely impaired rod and cone function.
Armed with this knowledge, scientists are honing in on various therapeutic strategies for genetic eye diseases. These strategies include somatic gene therapy, infusion of protective proteins, and embryonic cell transplantation. The hope is that such interventions will converge and lead to treatments that slow down or prevent the blindness characteristic of many degenerative eye diseases.
Pharmacological and rAAV Gene Therapy Rescue of the Retinoid Cycle
There have been several attempts to restore vision in patients with LCA using interventions such as calcium channel blockers and intraocular injection of neurotrophic factors. In most cases, the effects of these treatments lasted less than a month; hence, repeated administrations were required. Another approach is to bypass the biochemical block in mice without functional Rpe65 using synthetic cis-retinoids administered orally; such treatments have induced dramatic improvement in photoreceptor physiology.
Also, somatic gene therapy has been very successful in many animal models of retinal degeneration. In this issue of PLoS Medicine, Krzysztof Palczewski and colleagues attempted to combine two approaches to restore visual function with intraocular gene therapy and oral pharmacologic treatment with novel retinoid compounds in lecithin retinol acyl transferase (LRAT)–deficient mice. LRAT is a key enzyme involved in storage of vitamin A in the form of retinyl esters in structures known as retinosomes. In mice without LRAT, no 11-cis-RAL chromophore is produced, and visual function is severely impaired. Lrat mutations have been detected in a subset of patients with LCA.
The team found that gene therapy using intraocular injection of recombinant adeno-associated virus carrying the Lrat gene successfully restored electroretinographic and pupillary light responses in Lrat−/− mice. Production of 11-cis-RAL was also restored. Pharmacological intervention with orally administered pro-drugs 9-cis-retinyl acetate and 9-cis-retinyl succinate also caused long-lasting restoration of retinal function in Lrat-deficient mice. Combining interventions produced markedly increased levels of visual pigment, and 1,000-fold improvements in pupillary light response and electroretinogram sensitivity. Direct comparison of each treatment was difficult, but both therapies provide efficient recovery of higher order visual responses. One advantage of oral retinoid treatment was its ease of administration compared with the subretinal injections required for viral vectors. Another factor was that the orally administered compounds were not stored in the liver for long, and were quickly oxidized and secreted. Pharmacological treatment could also be given multiple times; several low-dose treatments show cumulative effects. The main disadvantage of oral treatment was the potential for long-term systemic toxicity compared with vector targeting of LRAT to the RPE, which needs to be examined in future studies.
Interestingly, the researchers observed that chromophore supplementation and somatic gene therapy were optimally effective in combination, particularly when chromophore supplementation was continued at low doses for longer periods of time. The authors suggest that the combined approach might be more suitable for treating a wider age range of patients. Although much more preclinical testing is required, it is likely that pharmacologic and somatic gene therapeutic approaches could be used together if such testing proves safe and successful in human trials. The authors speculate that treatment of patients with oral retinoids could begin in infancy to avoid amblyopia while also avoiding the difficulties associated with surgery in very young patients. For older patients, a long-lasting drug-free treatment might be achieved by surgical introduction of viral vectors.
|
0
|
PMC1274284
|
CC0
|
2021-01-05 10:39:21
|
no
|
PLoS Med. 2005 Nov 1; 2(11):e402
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020402
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 10.1371/journal.pmed.0020403SynopsisDermatologyRheumatologyConnective Tissue DiseaseDermatologyImmunology and AllergyRheumatologyManipulating Microtubules in Systemic Sclerosis Synopsis12 2005 1 11 2005 2 12 e403This is an open-access article distributed under the terms of the Creative Commons Public Domain Declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.2005This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.
Paclitaxel Modulates TGFβ Signaling in Scleroderma Skin Grafts in Immunodeficient Mice
==== Body
Systemic sclerosis (scleroderma or SSc) is the name for a group of progressive diseases, all of which involve the abnormal growth of connective tissue. They are chronic degenerative disorders in which there is widespread vascular deterioration and tissue loss. The recognition of SSc dates back many years. Indeed, the characteristic stretched thickening of the skin was probably first described by Hippocrates. A more definitive description of the condition was made by Carlo Curzio in 1753, who described a patient as having wood-like skin with “tight eyelids, difficulty in opening her mouth, coldness of her skin.”
The etiology of SSc is still not completely understood, but appears to be autoimmune. Downstream of the immune activation, the molecules that have been implicated are the profibrotic cytokines, such as transforming growth factor-beta (TGFβ), interleukin-4 (IL-4), platelet-derived growth factor (PDGF), and connective tissue growth factor, all of which can cause fibrosis. In addition to the profibrotic effects, TGFβ and PDGF may also contribute to vasculopathy. Furthermore, the vascular changes in SSc skin lesions are associated with anti-endothelial cell autoantibodies.
One problem in the research on SSc has been finding an appropriate animal model. One such model has now been developed: the transplantation of skin samples from patients with SSc into immunodeficient mice. Without the interference of an effective murine immune system, it is possible to test the effect of various interventions on the patients' skin samples.
Suppression of Smad2 phosphorylation by paclitaxel in in SSc skin graft
In a paper in PLoS Medicine, Chunming Dong and colleagues used this model to investigate the effect of paclitaxel on SSc skin samples. Paclitaxel is an attractive drug because it stabilizes microtubules, which affect the propagation of TGFβ signaling. TGFβ is a multifunctional regulatory cytokine involved in a large number of cellular activities. It initiates its effects by binding to and activating specific cell-surface receptors. These activated TGFβ receptors stimulate a family of proteins (R-Smads), some of which act to stimulate further TGFβ signaling, and others of which act to inhibit it. The microtubules regulate the R-Smad access to and activation by TGFβ receptors.
Dong and colleagues were able to show that in this model of SSc, paclitaxel markedly suppressed the activation of two of the Smads, Smad2 and Smad3, and, hence, collagen deposition in SSc grafts. They also showed that the SSc grafts had increased neovessel formation relative to normal grafts, regardless of paclitaxel treatment, thus indicating that paclitaxel did not have a negative effect on angiogenesis. In fact, they found that the neovascular cells were likely derived from the mouse hosts.
Do these findings indicate that paclitaxel is a good drug for SSc? Certainly more work needs to be done before this can be said. A particular concern that paclitaxel can itself lead to fibrosis at high doses when used, for example, in the treatment of certain tumors (in contrast to low doses used here) needs to be assessed further. However, SSc is a difficult progressive disease, and results such as these should be considered seriously as avenues for future therapies.
|
0
|
PMC1274285
|
CC0
|
2021-01-05 10:39:28
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e403
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020403
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 10.1371/journal.pmed.0020404SynopsisImmunologyInfectious DiseasesMicrobiologyVirologyHematologyPathologyVaccinesInfectious DiseasesA Persistent Immune Response to an Acute Virus Synopsis12 2005 1 11 2005 2 12 e404Copyright: © 2005 Public Library of Science.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.
Prolonged Activation of Virus-Specific CD8+ T Cells after Acute B19 Infection
==== Body
Human parvovirus B19 (B19) can cause a wide range of conditions, which can depend to a large extent on an individual's immunologic and hematologic status. In a normal host, parvovirus infection can be asymptomatic or cause a range of clinical syndromes, from erythema infectiosum (“slapped cheek” disease) to chronic arthritis. Hydrops fetalis and fetal death are complications of intrauterine B19 infection, and patients who are immunocompromised or who have hematologic disorders are at risk of aplastic anemia.
To date, it has been believed that clearance of acute infection is associated with the lifelong emergence of antiviral IgG, but there is increasing evidence for an important role for cellular immune responses, which suggests there might be more to the way the body deals with this virus. One previous study detected CD8+ T cell responses—which kill virus-infected cells by cytokine secretion—in three asymptomatic seropositive individuals.
A better understanding of this aspect of the body's immune response could have important implications not only for vaccine development and treatment for B19, but also for other viral infections. The long-lasting CD8+ T lymphocyte response means that parvovirus-based vectors could be considered in vaccine strategies for other infections. The identification of B19 epitopes for CD8+ T lymphocytes also offers the chance to analyze the role of such effector cells in chronic arthritis, a disease in which B19 has been implicated. Better understanding of this response might also allow researchers to use B19 as a model for analyzing immunological memory, immunodominance, and the interplay between cellular and humoral immune responses to a common clinically relevant human pathogen.
In this month's PLoS Medicine, Adiba Isa and colleagues describe evolution of long-lived CD8+ immune responses against B19 in 11 adults with primary B19 infection. The phenotype of CD8+ T cells in acute B19 infection has not been studied before. Normally, the symptoms of this virus are short lived, but the immune responses showed here indicate sustained activity many months after initial infection. The team studied two groups of people: 11 who had been recently infected and five who had the virus many years ago. CD8+ T cell responses were mapped using a screening system, which took advantage of the B19's compact and stable viral genome. The researchers used human leukocyte antigen (HLA)–peptide multimeric complexes to detect CD8+ T cell responses during acute B19 infection.
The researchers believe their results show a new style of host–virus relationship in which an acute human viral infection induces persistent activated CD8+ T cell responses. They found that these responses continued to increase, in some cases for many months, long after acute symptoms had resolved—something not seen with other viruses. For example, responses to HIV are strong in acute infection but typically decline as the virus is controlled.
Alongside the expansion of antiviral responses was the continued change in the B19 CD8+ T cells, as indicated by a range of markers. The evolution of markers could represent a maturation pathway, said the authors, driven by restimulation in vivo with antigen.
This T cell response to B19 infection indicated the persistence of antigen long after the resolution of acute infection. However, the authors said the status of the virus postinfection is still not understood. For example, in this study, PCR analysis found B19 DNA in the blood early on during infection, but assays were negative after 6–12 months when T cell populations remained active. The most likely reason for these immune responses is low-level replication at a tissue site for weeks or months after infection, suggest the authors. But this hypothesis can only be checked by more sensitive PCR assays.
In addition, the relationship between joint or bone marrow pathology and T cell responses seen was not clear. In the patients with arthritis, the most active CD8+ T cell responses were seen at stages where joint symptoms had resolved.
Altogether, this study is the first demonstration that a virus, not considered a true or classical persistent infection, can lead to a persistent activated CD8+ T cell response. It suggests that B19 persists after acute infection, provoking sustained activated CD8+ T cell responses, which might then play a role in viral clearance. A better understanding of cytotoxic T lymphocytes could improve understanding of the role of T cells in acute and persistent infections and be of great value in vaccine design and immunotherapy.
|
0
|
PMC1274286
|
CC BY
|
2021-01-05 11:13:38
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e404
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020404
|
oa_comm
|
==== Front
PLoS MedPLoS MedpmedplosmedPLoS Medicine1549-12771549-1676Public Library of Science San Francisco, USA 10.1371/journal.pmed.0020406SynopsisGenetics/Genomics/Gene TherapyDiabetes/Endocrinology/MetabolismPredicting the Development of Type 2 Diabetes Synopsis12 2005 1 11 2005 2 12 e406This is an open-access article distributed under the terms of the Creative Commons Public Domain Declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.2005This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration, which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.
Genetic Prediction of Future Type 2 Diabetes
==== Body
Type 2 diabetes has been loosely defined as “adult onset” diabetes, but as diabetes becomes more common, cases are being diagnosed in younger people and children. In determining the risk of developing diabetes, environmental factors, such as food intake and exercise, are known to have an important role; most people with type 2 diabetes are either overweight or obese. Inherited factors are also important, but the genes involved remain poorly defined. In rare forms of diabetes, mutations of one gene can result in disease, whereas in type 2 diabetes, many genes are thought to be involved. One difficulty in understanding the genetic role is that genes associated with diabetes might show only a subtle variation in their sequence, and these variations may be extremely common. Hence, it can be very hard to link such common gene variations, known as single nucleotide polymorphisms (SNPs), with increased risk of developing diabetes.
One method of finding these diabetes genes is by whole-genome linkage studies in which associations between parts of the genome and risk of developing diabetes are looked for. Studies so far have identified several candidate genes associated with type 2 diabetes, although many results have been difficult to replicate. The list of genes for which there is good evidence from meta-analyses includes genes encoding for PPARG, calpain 10, Kir 6.2, and insulin receptor substrate-1 (IRS1).
These genes have a variety of effects; PPARG P12A polymorphism is associated with enhanced insulin sensitivity and protects against type 2 diabetes. Two SNPs in the gene encoding for cystein protease calpain 10 (CAPN10) confer increased susceptibility to insulin resistance and type 2 diabetes. Kir 6.2 is involved in glucose-stimulated insulin secretion in pancreatic cells. And carriers of a polymorphism in the IRS1 gene have been shown to have reduced islet insulin content in pancreatic islets.
Genetic prediction of type 2 diabetes in the Botnia study
In this issue of PLoS Medicine, Valeriya Lyssenko and colleagues from Lund University sought to consolidate previous work by studying the predictive value of these variants for type 2 diabetes side by side in the largest study of its kind to date. They investigated the effect of these gene variants in 2,293 nondiabetic people aged 18–70 years old in western Finland—the Botnia study—over a median of six, range 2–12, years. In addition, they also studied the uncoupling protein 2 gene (UCP2)—a polymorphism in the promoter of this gene (UCP2 −866G>A) (rs659366) has been associated in some, but not all, studies with increased risk of type 2 diabetes and impaired insulin secretion.
The study took place from 1990 to 2002, and enrolled patients from five health centers in western Finland who were asked to have health checks every two to three years. Six percent (132) of people developed type 2 diabetes. The key finding was that variants in the PPARG and CAPN10 genes increased future risk for type 2 diabetes, particularly in individuals with other risk factors. In individuals with a high risk of developing diabetes—with a fasting plasma glucose (FPG) of 5.6 millimoles per liter and body mass index (BMI) of 30 kilograms per square meter—the hazard ratio increased to 21.2 for the combination of the PPARG PP and CAPN10 SNP43/44 GG/TT genotypes compared with those with low-risk genotypes with normal FPG and BMI less than 30 kilograms per square meter.
The researchers found that replacing the family history with the PPARG and CAPN10 variants in a predictive model (particularly in combination) gave almost the same strong prediction of subsequent type 2 diabetes. These genotypes also influenced the relationship between BMI and FPG, that is, in carriers of risk genotypes, there was a steeper increase in FPG for any given BMI.
The authors argue that the comparison of all the key gene variants side by side in one large study adds substantially to previous papers that have examined the effect of single gene variants on the risk of conversion to type 2 diabetes in interventional trials.
However, it is important to understand the effect of these variants on the risk of disease in a large, prospective observational study before studying additive or synergistic effects with interactions such as lifestyle changes, they said. One of the problems of other studies has been that results have been different between different subgroups.
Although this study has limited power, as the largest of its kind it suggests that genetic variants in candidate genes can predict future type 2 diabetes, particularly in association with conventional risk factors such as obesity and abnormal glucose tolerance. With accumulating data from prospective studies, it should be possible to define whether there will be a future role for genetic prediction of type 2 diabetes or whether these variants will influence response to prevention or treatment.
|
0
|
PMC1274287
|
CC0
|
2021-01-05 10:39:24
|
no
|
PLoS Med. 2005 Dec 1; 2(12):e406
|
utf-8
|
PLoS Med
| 2,005 |
10.1371/journal.pmed.0020406
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 1626119210.1371/journal.pcbi.0010050plcb-01-05-03Research ArticleGenomic Variability within an Organism Exposes Its Cell Lineage Tree Genomic Variability Exposes Cell LineageFrumkin Dan 1Wasserstrom Adam 1Kaplan Shai 1Feige Uriel 2Shapiro Ehud 12*1 Department of Biological Chemistry, Weizmann Institute of Science, Rehovot, Israel
2 Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
Bonhoeffer Sebastian EditorSwiss Federal Institute of Technology, Switzerland* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 15 9 2005 1 5 e5030 6 2005 13 9 2005 Copyright: © 2005 Frumkin et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.What is the lineage relation among the cells of an organism? The answer is sought by developmental biology, immunology, stem cell research, brain research, and cancer research, yet complete cell lineage trees have been reconstructed only for simple organisms such as Caenorhabditis elegans. We discovered that somatic mutations accumulated during normal development of a higher organism implicitly encode its entire cell lineage tree with very high precision. Our mathematical analysis of known mutation rates in microsatellites (MSs) shows that the entire cell lineage tree of a human embryo, or a mouse, in which no cell is a descendent of more than 40 divisions, can be reconstructed from information on somatic MS mutations alone with no errors, with probability greater than 99.95%. Analyzing all ~1.5 million MSs of each cell of an organism may not be practical at present, but we also show that in a genetically unstable organism, analyzing only a few hundred MSs may suffice to reconstruct portions of its cell lineage tree. We demonstrate the utility of the approach by reconstructing cell lineage trees from DNA samples of a human cell line displaying MS instability. Our discovery and its associated procedure, which we have automated, may point the way to a future “Human Cell Lineage Project” that would aim to resolve fundamental open questions in biology and medicine by reconstructing ever larger portions of the human cell lineage tree.
Synopsis
The human body is made of about 100 trillion cells, all of which are descendants of a single cell, the fertilized egg. The quest to understand their path of descent, called a cell lineage tree, is shared by many branches of biology and medicine, including developmental biology, immunology, stem cell research, brain research, and cancer research. So far, science has been able to determine the cell lineage tree of tiny organisms only, worms with a thousand cells or so. Our team has discovered that the mutations accumulated in each cell in our body during its normal development from the zygote carry sufficient information to reconstruct, in principle, cell lineage trees for large organisms, including humans. Inspired by this discovery, we developed an automated procedure for the reconstruction of cell lineage trees from DNA samples. A direct application of these results may include the analysis of the development of cancer. The results may also inspire a future “Human Cell Lineage Project,” whose aim would be to reconstruct an entire human cell lineage tree.
Citation:Frumkin D, Wasserstrom A, Kaplan S, Feige U, Shapiro E (2005) Genomic variability within an organism exposes its cell lineage tree. PLoS Comput Biol 1(5): e50.
==== Body
Introduction
A multicellular organism develops from a single cell, the zygote, through numerous binary cell divisions and cell deaths. Consequently, at any given time, the lineage relations between the cells of the organism can be represented by a rooted labeled binary tree called the cumulative cell lineage tree (Figure 1A). For any sample of cells, such as cells from a specific organ or tissue, the lineage relations can also be represented by a tree, called the cell sample lineage tree, which is partial to the cumulative tree (Figure 1B). Such a tree was reconstructed for the 959 somatic cells of Caenorhabditis elegans by direct observation of cell divisions [1], a technique that can be used for lineage analysis of small transparent organisms. Understanding the cell lineage trees of higher organisms, especially human, is a fundamental challenge of many branches of biology [2–10] and medicine [11–15]. Development of higher organisms is, however, less deterministic than that of C. elegans, and therefore the cell lineage trees of individuals of the same species may vary considerably.
Figure 1 Cell Lineage Concepts
(A) Multicellular organism development can be represented by a rooted labeled binary tree called the organism cumulative cell lineage tree. Nodes (circles) represent cells (dead cells are crossed), and each edge (line) connects a parent with a daughter. The uncrossed leaves, marked blue, represent extant cells.
(B) Any cell sample (A–E) induces a subtree, which can be condensed by removing nonbranching internal nodes and labeling the edges with the number of cell divisions between the remaining nodes. The resulting tree is called the cell sample lineage tree.
(C) A small fraction of a genome accumulating substitution mutations (colored) is shown. Lineage analysis utilizes a representation of this small fraction, called the cell identifier. Phylogenetic analysis reconstructs the tree from the cell identifiers of the samples. If the topology of the cell sample lineage tree is known, reconstruction can be scored.
(D) Coincident mutations, namely two or more identical mutations that occur independently in different cell divisions (blue mutation in A and B), and silent cell divisions, namely cell divisions in which no mutation occurs (D–F), may result in incorrect (red edge) or incomplete (unresolved ternary red node) lineage trees. Excessive mutation rates might result in successive mutations (not shown), which cause the lineage information to be lost.
Lineage relations among cells have been studied using a variety of clonal assays [2,3,6,8,10,16–24]. Such assays act by detecting the progeny of a single founder cell, which has been marked by a heritable marker. Some assays mark the founder cell by an invasive technique such as injection of a tracer molecule [16,18] or retroviral infection [10], which may interfere with the normal growth and biological function of the marked cell population. Other noninvasive clonal assays are based on spontaneous mutations in the founder cell, for example, the loss or gain of large genomic fragments [19], mitochondrial DNA mutations [20], T-cell receptor gene recombination [21], and changes in the number of microsatellite (MS) repeat units [15,24]. Epigenetic changes have also been used for clonal assays [22] and for determining stem cell growth dynamics [23]. A clonal assay provides limited lineage information because it determines only whether certain cells are descendants of the founder cell.
Genetic variability has also been used for reconstructing lineage trees of several tissue samples extracted from the same individual. In one study [25], tissue samples from breast cancer patients were analyzed for loss of heterozygosity and mutations in mitochondrial DNA, and the result of this analysis was fed into a phylogenetic algorithm, yielding tissue lineage trees. In a different study [26], lineage trees of colorectal cancer and adenoma tissue samples were reconstructed from mutations in MS loci. These studies applied clustering algorithms to genetic variability among heterogeneous tissue samples. However, the meaning of the output of such an algorithmic analysis is not necessarily clear, as the lineage relations among tissue samples, where one or more tissue samples contain cells of heterogeneous lineage, are normally not amenable to simple representation via binary trees. On the other hand, the lineage relations among single cells or discrete cell clones can be represented naturally via rooted labeled binary trees, and the question of whether such a particular mathematical tree faithfully represents the lineage relations among a given set of cells or cell clones sampled from an individual is well-defined and has a simple yes/no answer (see Text S1). Genomic variability in immunoglobulin genes has also been used to create mutational lineage trees [27] in the study of the dynamics of selection in the immune system. Because this work analyzed mutations in a functional gene, which determines the ability of the cell to undergo clonal expansion, the shape of the mutational lineage trees reflects primarily selection forces and does not necessarily correlate to the cell lineage tree.
In this paper we show that cell lineage trees can be reconstructed from genomic variability caused by somatic mutations, and that somatic mutations in higher organisms contain sufficient information to allow precise reconstruction of the organism cell lineage tree. We describe a hybrid in vitro/in silico automated procedure for reconstructing cell lineage trees from DNA samples, and demonstrate its effectiveness and precision in a controlled environment.
Results
Somatic Genomic Variability Encodes Cell Lineage
Somatic mutations are sufficiently rare for common wisdom to say that “the genome is the same in every cell in the body” except for some white blood cells [28] and except for cancer [26]. We discovered that somatic mutations accumulated during normal development of a higher organism, including human and mouse, implicitly encode its entire cell lineage tree with very high precision.
Figure 1C shows how accumulated somatic mutations, each viewed as an implicit clonal marker, may encode a cell lineage tree. As we show, sufficiently many markers enable the inference of the cell lineage tree, as closely related cells tend to share more markers than distantly related cells. If in each cell division each daughter cell acquires a new mutation, and all mutations are unique and persistent, then the organism cell lineage tree can be precisely reconstructed from this mutation information, using known phylogenetic algorithms [29].
However, precise reconstruction may be hampered by three factors: coincident mutations (Figure 1D), silent cell divisions (Figure 1D), and successive mutations (see proof of theorem 1 in the Materials and Methods). Still, known phylogenetic algorithms can produce useful lineage information in spite of these problems if the mutations carry sufficient information with a sufficiently high “signal-to-noise” ratio. Although phylogenetic analysis algorithms were originally developed for reconstructing lineage trees of species [30], they are also applicable to cell lineage tree reconstruction. It is conceivable, though, that they would be outperformed by algorithms designed specifically for this new task. Such an algorithm may use a more accurate model of MS mutation behavior and make use of the precise root information, which may be obtained in organism cell lineage trees, in order to reconstruct the tree more accurately.
In principle, any mutation information may assist lineage tree reconstruction. We focus on mutations in MSs for the following reasons: (1) MS slippage mutations, which insert or delete repeated units in an MS, are thought to occur during DNA replication [31] and hence are coupled to cell division; (2) MS mutations occur at relatively high rates [31] and offer a broad range of rates to choose from; (3) MS mutations are believed to occur independently at different loci, usually without affecting phenotype, and are unlikely to be selected against somatically since most are found in noncoding genomic sequences; (4) MSs are highly abundant in human, mouse, and many other organisms [31]; and (5) animals with mutations in key mismatch repair (MMR) genes display very high mutation rates in MSs [32,33] in all tissues and are available for experimentation and analysis. MMR-deficient humans [32] and mice [33] have been shown to develop normally, albeit with a high incidence of cancer. Genetic variability of MS loci has been used for linkage analysis [34], individual identification [35], phylogenetic analysis of species [36], and genealogical history analysis of populations [37].
Theoretical Potential of the Method
In order to asses the theoretical potential of lineage analysis using genomic MSs, we obtained data regarding human and mouse MSs and performed calculations and computer simulations based on these data. We searched the human and mouse genomes for MSs and found about 1.5 million loci interspersed on all chromosomes and containing a variable number of tandem repeats. Based on this data, and on published data regarding human and mouse MS mutation rates [38], we calculated that in each cell division in wild-type humans and mice, each daughter cell acquires on average approximately 50 new mutations in MS loci. Based on this information, we were able to prove theorem 1 (see below), which implies that in human and mouse lineage trees with a maximum depth of 40 cell divisions (which can contain up to 1012 leaves and correspond, under reasonable assumptions, to a newborn mouse; Figure 2) and with any topology, the complete cell lineage tree can be reconstructed with no errors with a probability greater than 99.95%.
Figure 2 Simulation of MS Mutations and Reconstruction Score on Random Trees
Two types of random trees with 32 leaves were generated, and MS stepwise mutations were simulated. Results of simulations of wild-type human using different numbers of MS loci are shown. The white line marks the perfect score limit (according to the Penny and Hendy tree comparison algorithm [29]). The results show that it is possible to accurately reconstruct the correct tree for trees of depth equivalent to human newborn and mouse newborn (marked by blue and green dots, respectively) using the entire set of MS loci. A mathematical analysis proves that any tree of depth 40 (equivalent to mouse newborn) can be reconstructed with no errors. Simulations with MS mutation rates of MMR-deficient organisms demonstrate that cell lineage reconstruction is possible with as few as 800 MS loci (the white line indicates the 0.95 score). The quality of reconstruction depends on the topology of the tree and its maximal depth, which together influence the signal-to-noise ratio.
The Possibility of Precise Reconstruction of Cell Lineage Trees Not Deeper Than 40 Cell Divisions
We prove that our approach has the potential of reconstructing without error condensed trees of sets of cells that are many orders of magnitude larger than anything achieved in the past. Our proof is based on certain assumptions regarding the DNA contents and nature of mutations in the human genome. These assumptions are stated explicitly in this manuscript, and (to the best of our knowledge) are in agreement with existing biological literature.
In describing our theoretical results, we prefer simplicity over achieving the best possible results based on our assumptions. In particular, the “triplet algorithm” that we present for reconstructing the condensed tree does not make use of all information available to it, and there is a lot of slack in the analysis. One may expect that more sophisticated algorithms coupled with tighter analysis will allow one to extend the family of trees that can be inferred using our methods.
We have chosen not to try to strengthen our theoretical results at this point for the following reasons. First, we believe that the vast potential of our method is made clear already by the analysis that we provide here. Second, and perhaps more importantly, it may be premature to enter into a lengthy theoretical analysis before establishing more firmly the biological assumptions on which the analysis is based. The biological assumptions that we make here may eventually turn out to be too optimistic in some respects (e.g., that mutation events are, or can be viewed as being, statistically independent), and too pessimistic in other respects (e.g., that the only significant source of variability in the human genome is MSs). Hence, there is not much point in performing tedious and time-consuming analysis based on current biological assumptions.
We now describe the mutation model that we assume in the analysis, which we call the “uniform model.” We use the following notation: m, number of MS loci; p, probability of mutation per MS locus per cell division; d, maximum depth of leaves in the cumulative tree; n, number of extant cells. The main assumption for the uniform model is that all mutation events are statistically independent. The simplifying assumptions (which simplify the presentation, but can be relaxed without qualitatively changing the results) are the following: (1) the identifier (a vector representing MS lengths) of the root is known and used as a reference; (2) both daughter cells of the root lead to extant cells; (3) all loci have the same mutation probability; and (4) mutations are stepwise, with equal probability for +1 and −1. The numerical values that are used in our analysis of the uniform model (roughly corresponding to known information about wild-type human MSs) are m = 2 × 106 and p = 2.5 × 10−5.
This completes the description of the uniform model. Our analysis addresses the following question: assuming that one could read with no error the identifiers of all extant cells, can one (with high probability, over the events of random mutations) reconstruct the underlying condensed tree with no error at all? The answer depends on the “shape” of the extant tree. We present some ranges of parameters for which reconstruction is possible. Specifically, we take d = 40, and n as being arbitrary (but of course, no more than 240). Note that n may be “in the same ball park” as the number of extant cells in a human (which is believed to be around 247).
Theorem 1: Assuming the uniform model, with probability above 0.9995 (over the random mutation events), the genetic information in human cells suffices in order to reconstruct without error the condensed version of any extant tree of depth at most 40, regardless of the number of extant cells and the shape of the tree.
Proof: See Materials and Methods. It is not easy to extend this analytical result to trees deeper than 40 cell divisions; however, we have good reasons to believe that the theorem does not represent a singular data point, as shown by our computer simulations.
Computer Simulations
We performed simulations on two types of randomly generated cell lineage trees (Figure 2), and simulated wild-type and MMR-deficient mutational behavior on hypothetical organisms that develop according to the pattern of these trees. The topology of cell sample lineage trees with 32 randomly chosen cells was then reconstructed based on analysis of mutations in a set of MSs. In wild-type human and mouse, using the entire set of genomic MSs yields accurate reconstruction in trees with a depth of several hundred cell divisions, corresponding to adult mice and newborn humans. Highly accurate reconstruction is achieved even when using a small fraction of the genomic MS loci (e.g., a tree of depth up to 400 cell divisions can be reconstructed with more than 90% accuracy using 10% of the genomic MS loci; see Figure 2). In MMR-deficient organisms, a few hundred MSs are sufficient for accurate reconstruction of complex cell lineage trees (Figure 2). MS loci that have excessive mutation rates should be avoided as they increase the likelihood of coincident and successive mutations. Simulations were performed for samples with up to 100 cells, because of the computational requirements of the phylogenetic analysis algorithm, but the results suggest that reconstruction scores may not decrease as the size of the cell sample increases.
Silent Cell Divisions in a Newborn Human
In order to assess the extent of silent cell divisions, which might act as a limiting factor for reconstructing human cell lineage trees, we calculated the probability for a silent cell division and estimated the total number of cells in the tree for a newborn human. We found that in a single cell division the probability of a daughter cell acquiring no new mutations is less than 10−21. For estimating the total number of cells in the tree we created a model of human embryonic development that overestimates the number of cells and cell divisions, and thus can serve as a theoretical upper bound on the size of the cumulative cell lineage tree of a newborn human. We found that in the model, in more than 99.9% of newborns, there is at least one new mutation in each daughter cell in each cell division. This suggests that during human prenatal development, even a single silent cell division is unlikely to occur. As mentioned above, coincident mutations may cause erroneous tree reconstruction, but because there are no data on the topology and depth of newborn cumulative cell lineage trees, it is difficult to estimate their effect in this model.
Experiments in Plants
C. elegans, with its known cell lineage tree, may have provided an excellent in vivo control for our cell lineage inference procedure, except that its genome does not contain a sufficient number of MSs to allow precise reconstruction. Plants are a good model system for lineage studies at the tissue level because of their nearly invariant pattern of cell division and lack of cell migration ([39], p. 299), which results in a correlation between their physical and lineage distances. Our first evidence supporting this correlation was obtained from analysis of MS variability in wild-type Robinia pseudoacacia trees, which have been shown to have somatic mutations in an MS locus [40]. (Analysis of multiple MS loci in R. pseudoacacia is not currently possible since its genome has not been sequenced.) We located a tree with somatic MS variability and extracted DNA from 28 tissue samples. We found that 25 samples contained the same genotype, which was considered normal, and three samples contained a mutant genotype. These three samples were physically clustered on the same small branch, which contained only mutant samples (Figure 3A). This demonstrates that spontaneous somatic MS mutations in wild-type plants can be used as clonal markers. However, in some plants, such as Pinus strobes, the somatic mutations in MSs are rare [41]. In order to obtain sufficient mutations we therefore used MMR-deficient Arabidopsis thaliana for further experimental analysis.
Figure 3 Analysis of Whole Organisms
(A) Photograph and scheme of the R. pseudoacacia tree used for the lineage experiment. All three identically mutated samples (red) come from the same small branch.
(B) A. thaliana plant used for the experiment. The location of each sample is indicated.
(C) Transverse scheme of the A. thaliana plant showing all sampled stem (rectangles) and cauline leaf (ovals) tissues. Mutations that occurred in two or more samples are depicted by colored circles.
Past work on A. thaliana analyzed genetic mutations that result in albino sectors [42,43], showing that tissues from the same organ are more likely to be clonally related than tissues from different organs [42], and that a small radial angle in the transverse plane between two samples increases the probability that they are clonally related [43].
In order to analyze genomic variability within a plant, we grew an MMR-deficient A. thaliana AtMSH2::TDNA mutant SALK_002708 [44], extracted DNA from 23 different tissue samples (Figure 3B), and amplified 22 MS loci for each DNA sample. We found that samples that shared a similar mutation in a particular MS locus tended to be physically adjacent (Figure 3C). This correlation between genetic and physical distance between the samples was statistically significant. Analysis of two other A. thaliana plants found far fewer somatic mutations and a weak correlation between genetic and radial distances (See Text S2 for details).
An Automated Procedure for Lineage Reconstruction from DNA Samples
We developed a procedure that takes as input a set of DNA samples, primers for MS loci, information on expected MS sizes, and information on PCR and capillary electrophoresis multiplexing compatibility between MS loci, and outputs a reconstructed cell lineage tree (with edge lengths) correlated with the DNA samples (Figure 4). The procedure involves common lab protocols, including PCR and capillary electrophoresis, and known algorithms, including a phylogenetic analysis algorithm. The resulting tree provides, in addition to the inferred topology, also depth information, representing the inferred number of cell divisions that occurred along each edge in the lineage tree, and confidence information. The procedure is oblivious to the DNA source and quality, which may be from clones of a single cell, from tissue samples, or from single cells. In this work we used DNA samples extracted from cell clones (with automatic signal analysis) and from heterogeneous tissue samples (with manual signal analysis). In principle, the procedure can work with DNA amplified from single cells [45], if provided in sufficient quantity. However, amplifying DNA from a single cell, reliably and in a sufficient quantity, is a technical challenge that has yet to be overcome. Therefore, at the moment, we focus on potential applications that can utilize cell clones. The DNA of cell clones represents the DNA of the founder cells sufficiently reliably to allow precise reconstruction of the lineage tree among those founder cells (see Materials and Methods and Text S3). However, not all cell types can be grown to clones of sufficient size for our current method to be applied without further DNA amplification.
Figure 4 Automated Procedure for Lineage Tree Reconstruction
The procedure accepts biological samples and PCR primers as input, and outputs a reconstructed lineage tree. It consists of a series of seven consecutive steps (numbered), during which the physical biological samples are “transformed” into digital data, which are then analyzed algorithmically. We built a hybrid in vitro/in silico automated system that performs steps 2–7 of the procedure (outlined), and used it to process DNA from tissue samples and single-cell clones. Incorporation of whole genome amplification techniques in the future may enable processing of single cells as well. For a detailed specification of the procedure, see Protocol S1.
We accomplished the procedure in a hybrid in vitro/in silico automated system, which operates as follows. A predetermined set of n MS loci is amplified from each sample by multiplex PCR and run on a capillary machine, yielding several histograms for each sample. A programmable laboratory robot augmented with a PCR machine performs the liquid handling for PCR, the PCR itself, and the preparation of the samples for the capillary machine. Sample analysis by the capillary machine produces histograms, which typically show two main peaks representing the allelic value of each MS, as well as a stutter pattern that is typical of PCR of MSs [38]. A computer program, developed by us, that utilizes a signal processing algorithm (see Protocol S1) resolves the stutter pattern and assigns relative allelic values to all the MS loci in all samples. To aid analysis, we select MS loci that are expected to produce little stutter upon PCR amplification. Subsequently, each sample is assigned an identifier—a vector of 2n elements corresponding to the 2n analyzed alleles. Each element (called relative allelic value) is a whole number equal to the difference between the number of repeats of that allele and the number of repeat units of the corresponding allele in an arbitrary reference sample (allelic crossover may occur; see Text S4). Finally, a computer program applies a phylogenetic algorithm to the set of sample identifiers and produces a reconstructed tree associated with the DNA samples.
Cultured Cell Trees
To quantitatively evaluate the cell lineage tree reconstruction procedure, we cultured ex vivo cell trees with known topologies and well-estimated edge lengths, called cultured cell trees (CCTs). We constructed three CCTs (A–C; Figure 5A–5C) using human adenocarcinoma cells (culture LS174T, European Collection of Cell Cultures), which have a mutation in a key MMR gene [46] and high MS mutation rates [47]. We chose a set of 51 MS loci of various repeat types and various numbers of repeats (see Table S1 and Text S3B for selection criteria). DNA samples obtained from the root and leaf nodes were fed into the cell lineage tree reconstruction procedure, yielding a reconstructed tree for each CCT (see Table S2 for all cell identifiers). Reconstructions were performed using the neighbour-joining (NJ) [29] phylogenetic algorithm. In all cases the topology of the CCT was reconstructed precisely; thus, the correct topology was found out of a total of (A) 135,135, (B) 34,459,425, and (C) 13,749,310,575 possible topologies for the 8, 10, and 12 leaves of CCTs A, B, and C, respectively (Figure 5A–5C). The edge lengths in the reconstructed trees were in linear correlation to the actual number of cell divisions in the CCTs (Figure 5D; R
2 = 0.955). Furthermore, reconstructions of the CCTs without using root identifiers (Figure S1) yielded perfect scores for CCTs A and C, and a score of 7/8 for CCT B (the incorrect edge is colored red in Figure S1), suggesting that accurate reconstruction is feasible from the extant cells alone. Finally, for each CCT we found a minimal set of loci that yielded correct reconstruction using NJ (Figure 5A; colored contours in Figure S2), and analyzed how many loci were needed, on average, for precise reconstruction. We found that CCT A, being simpler than trees B and C, indeed requires fewer loci (Figure 5E). These results suggest that in MMR-deficient organisms, complex lineage trees may be reconstructed using a small set of MS loci. CCTs serve as a controlled system for phylogenetic analysis, and also may provide exact numerical data regarding the rates, nature, and correlation of mutations, allowing the assessment of the validity of MS mutation models.
Figure 5 CCT Model System
(A–C) A cell sample lineage tree with a predesigned topology is created by performing single-cell bottlenecks on all the nodes of the tree. Lineage analysis is performed on clones of the root and leaf cells. Three CCTs (A–C) were created using LS174T cells that display MS instability. All topologies were reconstructed precisely. Edge lengths are drawn in proportion to the output of the algorithm. Gray edges represent correct partitions according to the Penny and Hendy tree comparison algorithm [29], and their width represents the bootstrap value [29] (n = 1,000) of the edge. A minimal set of loci yielding perfect reconstruction was found for each CCT (each colored contour represents a different mutation shared by the encircled nodes; see also Figure S2).
(D) There is a linear correlation (R
2 = 0.955) between reconstructed and actual node depths.
(E) Reconstruction scores of CCTs A–C using random subsets of MS loci of increasing sizes (average of 500).
Discussion
We discovered that somatic mutations in higher organisms carry enough information to enable precise reconstruction of the entire organism cell lineage tree. We demonstrated the practical utility of the discovery by developing a prototype automated procedure for the reconstruction of cell lineage trees from DNA samples.
In the short term, small-scale projects utilizing this discovery and its associated procedure may aim to gain preliminary understanding of partial lineage trees associated with different organs or systems, by analyzing cell samples containing only dozens or hundreds of cells. In addition, analysis of the development of cancer using this method may provide immediate benefits. Cancer analysis may not require the perfection of single-cell methods, since clonal tissue samples may be obtainable from solid tumors.
In the longer term, with the improvement of DNA sequencing technologies [48], these results may inspire the initiation of a “Human Cell Lineage Project,” whose aim would be to reconstruct an entire human cell lineage tree. A precursor project, which may face fewer hurdles, would be a “Mouse Cell Lineage Project.” Both projects would require multidisciplinary teams, with members familiar with different organs or biological subsystems, but either project would benefit from the teams working on the same individual organism, since accumulated mutation information regarding the same individual could greatly improve the precision of the overall tree reconstruction process. Still, as in the Human Genome Project, diversity would be needed to separate incidental from essential properties of the organism cell lineage tree.
Materials and Methods
Number of MS loci and estimation of the number and rate of MS mutations.
We downloaded the human (build 35) and mouse (build 33) genomes from UCSC Genome Bioinformatics (http://hgdownload.cse.ucsc.edu/downloads.html). We wrote a MATLAB (MathWorks, Natick, Massachusetts, United States) program for searching MSs in any sequenced genome and used it to search for all mono- to hexanucleotide MSs in human and mouse that were nine uninterrupted repeats or longer. For any repeat unit (e.g., AAG) its frame shifts (AGA, GAA) were not searched, so results are a slight underestimate. See Tables S3 and S4 for data.
For estimation of the number of MS mutations in each cell division in human, we obtained from the literature [49] the approximated rates of mutations in human MSs per human generation, as a function of the length of the MS (number of uninterrupted tandem repeat units). Although the rate of mutations in MS loci is also dependent to a great extent on the specific locus examined [31], in our analysis we assumed as a first approximation that the average mutation rate is the mutation rate obtained from [49]. Mutation rates in [49] are given per human generation. These mutation rates were transformed to rates per cell division by dividing them by 186.5, which is the average of the approximated number of cell divisions in human male and female generations (350 and 23, respectively; see [49]). Because the mutation rates of MSs with 9–15 repeat units seem to increase exponentially with MS length, we used MATLAB to calculate a linear fit of the logarithm of these mutation rates. From this linear fit we obtained the mutation rates for MSs with 9–15 repeat units. The linear fit gives
The mutation-rate function for MSs with 9–15 repeat units is therefore
where e is the basis of the natural logarithm. For all MSs with less than nine repeat units, we made a conservative assumption that their rate of mutation is zero. Because of the lack of information regarding mutation rates in MSs with more than 15 repeat units, we made another conservative assumption that the mutation rates of all such loci are the same as for loci with 15 repeat units. Therefore, our estimated mutation rates for short and long MSs most likely represent an underestimate of the actual rate. We sorted the human MSs according to their length, and for each length we computed the expected number of mutations acquired by a daughter cell in a single cell division by multiplying the mutation rate by the number of MSs. The total number of expected MS mutations was computed by summing the expected number of mutations in each length category. See complete data in Table S3.
In contrast to the information regarding human MS mutation rates, there are fewer published data regarding mouse MS mutation rates, and data from different sources may be inconsistent. Comparison of data from several studies of human [50,51] and mouse [52,53] germ line mutations reveals that the rate of MS mutations per organism generation in mouse is 1–10 times lower than the human rate. In a model of MS mutations based on equilibrium distributions of MS repeat lengths [54], the rate of mouse MS mutations per organism generation is about five times higher than the corresponding human rate. Mice have faster life cycles than humans, and consequently the number of cell divisions per mouse generation is smaller than in humans. It is estimated that mice have approximately 6.5-fold fewer cell divisions per organism generation than humans [55]. Incorporating the data on MS mutations per organism generation and the estimated numbers of cell divisions, the rate of MS mutations per cell division in mice seems to be from about 1.5 times lower to about 30 times higher than the corresponding human rate. In our calculation of the expected number of mutations in mice, we make a conservative assumption that the rate of mutations is equal to the corresponding rate in humans (described above). Based on this assumption and the numbers of mouse MS alleles, we calculate the expected number of mutations in a similar fashion to the corresponding calculations in human MS (described above). See data in Table S4.
Our mathematical analysis, simulations, and the reconstruction of CCTs assume a uniform MS mutation rate across tissue types, as there is no sufficient knowledge at present to assign different somatic MS mutation rates to different tissue types.
Proof of theorem 1.
The simplifying assumptions underlying the uniform model make the calculation of some key quantities (which we call “signal,” “noise,” and “loss,” as indication of their effect on our eventual reconstruction algorithm) rather straightforward. We provide such calculations now (omitting some details). For signal, the expected number of mutations per edge of the extant tree is m × p = 50. The probability that the number of mutations on an edge is t < 50 behaves roughly like 50t × e−50/t! (here e is the basis of the natural logarithm). For noise, we estimate the number of accidental coincident mutations as follows. Given two branches of length b, the expected number of coincident mutations along these branches is roughly (50b)2/2m. When b = 40 (which is the maximum that we shall consider in this manuscript), this number is one. The probability that there are t coincident mutations in two branches of length 40 behaves roughly like 1/e × 1/t! For loss, successive mutations at the same locus may result in a loss of the signal. We estimate the extent of this loss as follows. Essentially, the worst case is when two leaves share a path of length d/2 = 20 (on which they are expected to have 1,000 common mutations), and then continue separately up to depth 40. They will each have roughly 1,000 more mutations, so one of the common mutations is expected to be undone. On the other hand, they are expected to share 1/4 of a coincident mutation, which somewhat compensates for the lost mutation. It turns out that for our analysis (with d = 40) the effect of lost mutations is negligible, and we will ignore it altogether. (Ignoring lost signal is further justified by the fact that the worst case for lost signal appears in tree configurations that are very different from those that give the worst case for coincident mutations.)
We now describe the reconstruction algorithm that we analyzed. This is a new algorithm, which we call the “triplet algorithm,” designed to facilitate the proof. This algorithm is chosen here because its analysis is simple, but we do not necessarily advocate its use in practice. We suspect that similar (and perhaps even better) results are true for other algorithms as well.
The basic primitive of the triplet algorithm is a “triplet subroutine.” Given identifiers for three cells (say, A, B, and C), the triplet subroutine counts for every pair of cells the number of common mutations, namely, the number of loci in which the two cells have the same label, and moreover, this label is different from the corresponding label of the root. The pair of cells that maximize this count (say, A and B) are output by the triplet subroutine. We say that the triplet subroutine is “successful” if the pair of cells that it outputs is the one that has the longer common branch (or equivalently, the deeper common ancestor).
The triplet subroutine will be successful unless there is some value of t such that there were at most t mutations along the branch common only to A and B, and at least t accidental coincident mutations between B and C (or between A and C). The probability of this event is roughly 50t × e−51 × (1/t!)2, which is maximized when t = 7, giving roughly 2.2 × 10−18. Summing over all values of t (using the fact that there is an exponential drop as we move away from t = 7), the total probability of not being successful is around 10−17. Hence one can execute the triplet subroutine on 1017 arbitrary (not just random!) triplets, and still be likely to be successful in all executions.
We now describe the triplet algorithm. View every cell as a vertex in an auxiliary graph G. In an execution of a triplet subroutine that outputs cells A and B (say, on input A, B, and C), put an edge between A and B. As long as there are more than two connected components in the graph G, pick three vertices from three different components and execute a triplet subroutine on them, thereby adding an edge to the graph and decreasing the number of connected components by one. After m − 1 steps, two connected components remain. Each one of them necessarily corresponds to a subtree of depth at most d − 1. The condensed version of each of the subtrees can be inferred separately by repeating the above procedure. Hence, at most (and in most cases, less than) d × m successful executions of the triplet subroutine suffice.
We have just shown that for every extant tree of depth d and m extant cells, d × m consecutive successful executions of the triplet subroutine guarantee that the triplet algorithm outputs the true condensed tree. This suffices in order to prove our theorem, because 1017 is much larger than 40 × 240. In fact, the ratio between these numbers is such that the probability that the tree is constructed with no error is greater than 0.9995. This completes our proof.
Computer simulations.
The simulations demonstrate, first, that human wild-type MS mutations enable accurate reconstruction of cell lineage trees and, second, that with higher mutation rates, as in MMR-deficient cells, cell lineage trees can be accurately reconstructed with no more than 800 MS loci (a kit containing 800 primer pairs for human MS amplification is commercially available).
The simulation proceeds as follows. A random tree is generated according to chosen topology type, maximal depth, and number of leaves. MS mutations are simulated according to number of loci and mutation rates, and leaf identifiers are generated. A lineage inference algorithm reconstructs the lineage tree. The inferred tree is compared with the generated tree, and the result is scored. Mutations and tree inference are performed ten times for each generated tree. For each depth, five random trees from each topology type are generated.
Since we do not usually have prior knowledge of the real organism's tree topology and branch lengths, we simulated two types of random trees that reflect topology space variability to a reasonable extent (see Figure 2). For the simulation, we generated trees with 32 leaves and various depth limits that reflect limits on the number of cell divisions from the root (i.e., the zygote). Each tree had 32 leaves, a large enough number to reflect the tree topology. Increasing the number of leaves does not necessarily increase inference difficulty since it adds information. In our simulations, increasing the number of leaves from 32 to 100 did not affect the average score.
Type I random trees have a random binary path, and are generated as follows. Generate LEAF _ NUMBER of unique nonoverlapping binary strings, each of a random length of up to MAX _ DEPTH bits. Each string represents a path in a binary tree leading to a leaf. In such a tree the least common ancestor of any pair of leaves is usually relatively close to the root. There is only a (1/2)n chance that two random binary strings of length n will be the same; therefore, within a tree of 32 samples and depth 100, most leaves will split within the first six tree levels. Such a tree is difficult to infer since the mutation signal on the tree internal edges is low and the mutation noise (i.e., coincident mutations on long paths) is high.
Trees of type II are generated by random node addition, as follows. The tree is initialized with two paths of random length up to MAX _ DEPTH that start at the root and lead to two leaves. An iteration adds a leaf by randomly picking a path, and on it randomly picking an internal node as the source of the new path. The depth of the new leaf is determined randomly between (new internal node + 1) and MAX _ DEPTH. Leaves are added until LEAF _ NUMBER is obtained. The procedure generates a variety of tree topologies with branches at various depths. The procedure often produces nonbalanced trees. In this family of trees, increasing the maximum depth does not always result in an increase of noise over signal since internal branches are often deep.
Mutations in MS loci were simulated by a stepwise model. The root was assigned a vector of zeros of the size of the simulated ALLELE _ NUMBER. An ALLELE _ MUTATION _ RATE, which is the probability of each MS locus mutating in a single cell division, was chosen. Then in each simulated cell division, each locus could mutate by increasing or decreasing the repeat number according to its assigned probability. Starting with the root identifier we generated the identifier of all tree nodes and leaves by simulating mutations as described above.
In simulated MMR-deficient cells we used mutation rates not higher than one per 100 cell divisions, according to Table 1. A future algorithm might use loci of various mutation rates, for example, using slow and reliable MS loci to obtain the coarse topology and then using the faster loci to infer local subtrees for which slow loci do not provide enough information.
Table 1 Mutation Rates in MMR-Deficient Human Simulations
Lineage tree inference was done with the NJ algorithm [29], which uses a distance matrix as input. We used the “equal or not” distance function, which increases the distance between two identifiers by one for each locus that differs. Inference using a maximum parsimony algorithm was also tried, with similar results.
The generated tree and the reconstructed tree were compared using Penny and Hendy's topological distance algorithm [29] (implemented using MATLAB). In this algorithm, the removal of each internal edge partitions the root and leaves into two groups. We assigned a score equal to the ratio of partitions, which was equal in the two trees to the total number of partitions. This scoring is rather strict as it might drop considerably with even single leaf misplacement.
Simulations were performed for samples with up to 100 cells, because of the computational resources required by the phylogenetic analysis algorithm.
Silent cell divisions in humans.
The probability for no mutations in each daughter cell in each length category was calculated by the formula
The probability for no mutations in all loci was calculated by multiplying the probabilities for no mutations in all length categories (see data in Table S5).
In order to estimate the total number of cells in a human neonatal cell lineage tree, we developed a model of human wild-type development. This is an overestimate model, which is intended to contain a larger depth and a larger number of cell divisions than the real (unknown) tree. In this model, development starts from a single cell, the zygote, in a series of 46 binary cell divisions, producing a binary tree with a full depth of 46, which has approximately 1014 leaves and approximately 1014 internal nodes, and hence has approximately 2 × 1014 nodes altogether (according to published data the adult human has about 1014 cells). This series of 46 divisions lasts for 23 d because each cell cycle is exactly 12 h long (according to published data the cell cycle in early human embryogenesis is 12–24 h). From this point on, there are additional 486 cycles of 12 h in 243 d until birth. In each cycle, each cell divides with a probability of 0.5 and dies with a probability of 0.5. Therefore, the number of living cells remains relatively constant from day 23 to day 266 (birth) at about 1014, and in each day 2 × 1014 cells are produced. The total number of cells produced during this process, and therefore the total number of nodes in the complete cell lineage tree at birth, is approximately 2 × 1014 + 486 × 1014 = 4.9 × 1016.
The probability for at least one new MS mutation in every cell in the human neonatal cell lineage tree was calculated by the formula (probability for no MS mutations in a single daughter cell)total number of cells. This calculation gives
Experiments in plants.
DNA from A. thaliana and R. pseudoacacia was extracted using the Extract-N-Amp kit (Sigma, St. Louis, Missouri, United States), and amplified according to the kit instructions. A list of primers for R. pseudoacacia and A. thaliana is given in Tables S6 and S7, respectively. Amplified products were run on a capillary electrophoresis machine (ABI Prism, Avant-3100, Applied Biosystems, Foster City, California, United States). Mutations were determined by manual comparison of capillary histograms.
Individuals of A. thaliana AtMSH2::TDNA mutant SALK_002708 (seeds kindly provided by J. Leonard) were grown in a growth room under long-day conditions. All A. thaliana plants were verified as mutant as described in [44].
CCTs.
Primers for most MS loci were designed using Primer3 (http://frodo.wi.mit.edu/cgi-bin/primer3/primer3_www.cgi) with the following parameters changed from default: Primer size = 20,22,27 (minimum, optimal, maximum); Primer Tm = 62 °C, 65 °C, 68 °C; Max Tm difference = 2.5 °C; CG clamp = 1. Some primers were taken from STRbase (http://www.cstl.nist.gov/biotech/strbase/).
CCTs were created using LS174T human colon adenocarcinoma cells, which were obtained from the European Collection of Cell Cultures (Salisbury, United Kingdom) and were grown in medium containing EMEM (Eagle's minimum essential medium, in Earle's balanced salt solution, GIBCO, San Diego, California, United States), 2 mM glutamine, 1% nonessential amino acids, 10% fetal bovine serum, and 1% penicillin-streptomycin. We estimated that LS174T cells divide every 1.5 d according to the frequency of routine plate passages. We created CCTs as follows. Initially, a single cell was isolated from a cell stock and was defined as the tree root. This cell was allowed to proliferate for a desired number of cell divisions (passages were performed when required). Then, two cells were isolated from the root progeny, and were defined as its daughter cells in the tree. This procedure was continued for each daughter cell, creating the granddaughter cells, etc., until the entire tree was grown. The tree root and leaf cells were cloned in plates, and lineage analysis was performed on DNA obtained from these clones. Lineage analysis performed on clones is expected to yield the same results as analysis on the founder cells of the clones (see Text S3A). Clones from single cells were created as follows: (1) trypsinizing and lifting cells from semiconfluent plates, (2) thrusting the cells ten times through a 1-μm mesh (Sefar, Heiden, Switzerland), (3) verifying by microscope that 99% or more of the cells were not attached to other cells, (4) diluting the cells with ratios ranging between 1:5,000 and 1:100,000 and spreading the cells on new plates, (5) waiting for single cells to form small islands (about 2–3 wk), and (6) lifting islands to new plates using cloning cylinders (Sigma). DNA was extracted from clones of all cells corresponding to nodes of the CCTs using Wizard SV Genomic Purification System (Promega, Fitchburg, Wisconsin, United States). Cells from all nodes of the CCTs were frozen in liquid nitrogen using a freezing medium containing 90% fetal bovine serum and 10% DMSO. Lineage reconstruction from CCT DNA samples (root and leaves only) was performed according to the automated procedure, as described in Protocol S1.
Figure S2 shows reconstructed trees for CCTs A–C without using the root for reconstruction. All reconstructions were performed using NJ (with the “equal or not” distance function). The unrooted trees outputted by NJ were rooted at the midpoint of the longest path from among all possible pathways between any two leaf nodes. Reconstructions of CCTs A and C were perfect, and a score of 7/8 was achieved for CCT B (Figure S1). The identifier of the zygote or root of a tree may be deduced in one of the following manners: (1) deduction from parental identifiers or (2) deduction from the most common allele. Deduction from parental identifiers (deduction of zygote identifier) is performed as follows. When performing lineage analysis on tissues from MMR-deficient organisms, the organism should be produced by a cross between two animals heterozygous for a mutation in an MMR gene (e.g., Mlh1
+/−), with each parent from a different inbred line. Animals that are heterozygous for a mutant MMR gene have normal or slightly elevated MS mutation rates. A cross between two such animals produces (with a frequency of 1:4) an animal that is homozygous for the mutant gene, with greatly elevated MS mutation rates. In order to deduce the identifier of the root (zygote) of such an organism, which is used in an experiment, the identifiers of its parents should be obtained. Because the parents come from inbred lines, they are homozygous at each MS locus and therefore deducing the identifier of the zygote is straightforward. The deduced identifier is very close (and when analyzing a few hundred MS loci may be identical) to the actual identifier because somatic MS mutations in the parents are very rare. It is important to note that this procedure deduces the identifier of the zygote of the organism, which may or may not be identical to the root of the reconstructed tree.
Deduction from the most common allele (deduction of root identifier) is performed as follows. In this procedure, the most common allele is determined for each MS locus in the population of sampled cells, and this value is assigned to the root identifier. Thus, the root identifier consists of the most common values in the cell population. In balanced trees that are not too deep, the deduced identifier will be very close (and may be identical) to the actual root identifier. However, in unbalanced (nonsymmetric) trees, this procedure will result in the deduced identifier being “tilted” towards the larger branch, and in deep trees the deduced identifier may differ from the actual identifier in MS loci that accumulate mutations in a nonsymmetric fashion. For example, in an MS locus that is biased towards MS contraction, the deduced identifier value may be smaller than the actual value. It is important to note that this procedure deduces the identifier of the root of the tree, which is not necessarily the zygote of the organism.
Supporting Information
Figure S1 Reconstructed Trees for CCTs A–C without Using the Root for Reconstruction
(56 KB JPG)
Click here for additional data file.
Figure S2 Example Minimal Sets of Loci Yielding Perfect Reconstruction of CCTs B and C
(93 KB JPG)
Click here for additional data file.
Protocol S1 Specification for the Automated Procedure for Lineage Reconstruction from DNA Samples
A complete description of the protocol for reconstructing lineage trees from DNA samples, including the capillary histogram signal analysis algorithm and tree reconstruction and scoring algorithms.
(474 KB DOC)
Click here for additional data file.
Table S1 List of MS Loci Used for the CCT Model System
(89 KB DOC)
Click here for additional data file.
Table S2 Cell Identifiers for CCTs A–C
(88 KB DOC)
Click here for additional data file.
Table S3 Estimated Number of MS Mutations in Each Cell Division for Human
(27 KB DOC)
Click here for additional data file.
Table S4 Estimated Number of MS Mutations in Each Cell Division for Mouse
(26 KB DOC)
Click here for additional data file.
Table S5 Silent Cell Divisions in Human
(27 KB DOC)
Click here for additional data file.
Table S6 List of MS Loci Used for R. pseudoacacia
(28 KB DOC)
Click here for additional data file.
Table S7 List of MS Loci Used for A. thaliana
(50 KB DOC)
Click here for additional data file.
Text S1 Lineage Analysis at the Single Cell and Tissue Levels
(48 KB DOC)
Click here for additional data file.
Text S2 Full A. thaliana Results
(165 KB DOC)
Click here for additional data file.
Text S3 Reconstruction Using Cell Clones and MS Selection Criteria
(A) Reconstructing cell lineage trees from DNA extracted from cell clones.
(B) Selection criteria for MSs.
(29 KB DOC)
Click here for additional data file.
Text S4 Ignoring the Effect of Allelic Crossovers
(27 KB DOC)
Click here for additional data file.
We thank A. Levy for pointing us to MSs and MMR-deficiency, which are the basis of our current implementation; Z. Livneh for advice and support; K. Katzav for the design and preparation of the figures; J. Leonard for mutant A. thaliana seeds; R. Hadar-Gabay and the Jerusalem Botanical Gardens for assistance in R. pseudoacacia experiments; J. Japha for pointing out locations of R. pseudoacacia trees; R. Adar and G. Linshiz for ongoing support; and G. Bejerano, Y. Benenson, A. Eldar, B. Geiger, A. Regev, and E. Segal for critical review and suggestions. This work was supported by the Israeli Science Foundation, by a research grant from Dr. M. Roshwald, by the Moross MD Center, and by the R. and A. Belfer Institute of Mathematics and Computer Science.
Competing interests. A patent application may be made on the results reported.
Author contributions. DF, AW, and ES conceived and designed the experiments. DF, AW, and SK performed the experiments. DF, AW, SK, UF, and ES analyzed the data, contributed reagents/materials/analysis tools, and wrote the paper.
A previous version of this article appeared as an Early Online Release on September 15, 2005 (DOI: 10.1371/journal.pcbi.0010050.eor).
Abbreviations
CCTcultured cell tree
MMRmismatch repair
MSmicrosatellite
NJneighbor joining
==== Refs
References
Sulston JE Schierenberg E White JG Thomson JN 1983 The embryonic cell lineage of the nematode Caenorhabditis elegans
Dev Biol 100 64 119 6684600
Stern CD Fraser SE 2001 Tracing the lineage of tracing cell lineages Nat Cell Biol 3 E216 E218 11533679
Clarke JD Tickle C 1999 Fate maps old and new Nat Cell Biol 1 E103 E109 10559935
Noctor SC Flint AC Weissman TA Dammerman RS Kriegstein AR 2001 Neurons derived from radial glial cells establish radial units in neocortex Nature 409 714 720 11217860
Ardavin C Martinez del Hoyo G Martin P Anjuere F Arias CF 2001 Origin and differentiation of dendritic cells Trends Immunol 22 691 700 11739000
Anderson DJ Gage FH Weissman IL 2001 Can stem cells cross lineage boundaries? Nat Med 7 393 395 11283651
Kim KM Shibata D 2002 Methylation reveals a niche: Stem cell succession in human colon crypts Oncogene 21 5441 5449 12154406
Dor Y Brown J Martinez OI Melton DA 2004 Adult pancreatic beta-cells are formed by self-duplication rather than stem-cell differentiation Nature 429 41 46 15129273
Alvarez-Buylla A Garcia-Verdugo JM Tramontin AD 2001 A unified hypothesis on the lineage of neural stem cells Nat Rev Neurosci 2 287 293 11283751
Walsh C Cepko CL 1992 Widespread dispersion of neuronal clones across functional regions of the cerebral cortex Science 255 434 440 1734520
Bernards R Weinberg RA 2002 A progression puzzle Nature 418 823 12192390
Yamamoto N Yang M Jiang P Xu M Tsuchiya H 2003 Determination of clonality of metastasis by cell-specific color-coded fluorescent-protein imaging Cancer Res 63 7785 7790 14633704
Hope KJ Jin L Dick JE 2004 Acute myeloid leukemia originates from a hierarchy of leukemic stem cell classes that differ in self-renewal capacity Nat Immunol 5 738 743 15170211
Weigelt B Glas AM Wessels LF Witteveen AT Peterse JL 2003 Gene expression profiles of primary breast tumors maintained in distant metastases Proc Natl Acad Sci U S A 100 15901 15905 14665696
Tang M Pires Y Schultz M Duarte I Gallegos M 2003 Microsatellite analysis of synchronous and metachronous tumors: A tool for double primary tumor and metastasis assessment Diagn Mol Pathol 12 151 159 12960697
Ben-Yair R Kahane N Kalcheim C 2003 Coherent development of dermomyotome and dermis from the entire mediolateral extent of the dorsal somite Development 130 4325 4336 12900449
Zernicka-Goetz M Pines J McLean Hunter S Dixon JP Siemering KR 1997 Following cell fate in the living mouse embryo Development 124 1133 1137 9102300
Dubertret B Skourides P Norris DJ Noireaux V Brivanlou AH 2002 In vivo imaging of quantum dots encapsulated in phospholipid micelles Science 298 1759 1762 12459582
Shelley Hwang E Nyante SJ Yi Chen Y Moore D DeVries S 2004 Clonality of lobular carcinoma in situ and synchronous invasive lobular carcinoma Cancer 100 2562 2572 15197797
Parrella P Xiao Y Fliss M Sanchez-Cespedes M Mazzarelli P 2001 Detection of mitochondrial DNA mutations in primary breast cancer and fine-needle aspirates Cancer Res 61 7623 7626 11606403
van Dongen JJ Langerak AW Bruggemann M Evans PA Hummel M 2003 Design and standardization of PCR primers and protocols for detection of clonal immunoglobulin and T-cell receptor gene recombinations in suspect lymphoproliferations: Report of the BIOMED-2 Concerted Action BMH4-CT98–3936 Leukemia 17 2257 2317 14671650
Tan SS Faulkner-Jones B Breen SJ Walsh M Bertram JF 1995 Cell dispersion patterns in different cortical regions studied with an X-inactivated transgenic marker Development 121 1029 1039 7743919
Yatabe Y Tavare S Shibata D 2001 Investigating stem cells in human colon by using methylation patterns Proc Natl Acad Sci U S A 98 10839 10844 11517339
Fujii H Matsumoto T Yoshida M Furugen Y Takagaki T 2002 Genetics of synchronous uterine and ovarian endometrioid carcinoma: Combined analyses of loss of heterozygosity, PTEN mutation, and microsatellite instability Hum Pathol 33 421 428 12055677
Morandi L Pession A Marucci GL Foschini MP Pruneri G 2003 Intraepidermal cells of Paget's carcinoma of the breast can be genetically different from those of the underlying carcinoma Hum Pathol 34 1321 1330 14691919
Tsao JL Tavare S Salovaara R Jass JR Aaltonen LA 1999 Colorectal adenoma and cancer divergence. Evidence of multilineage progression Am J Pathol 154 1815 1824 10362806
Dunn-Walters DK Belelovsky A Edelman H Banerjee M Mehr R 2002 The dynamics of germinal centre selection as measured by graph-theoretical analysis of mutational lineage trees Dev Immunol 9 233 243 15144020
Gilbert S 2000 Developmental biology, 6th ed Sunderland (Massachusetts) Sinauer Associates 749 p.
Graur D, Wen-Hsiung, L 2000 Fundamentals of molecular evolution, 2nd ed Sunderland Sinauer Associates 481 p.
Brown JR Douady CJ Italia MJ Marshall WE Stanhope MJ 2001 Universal trees based on large combined protein sequence data sets Nat Genet 28 281 285 11431701
Ellegren H 2004 Microsatellites: Simple sequences with complex evolution Nat Rev Genet 5 435 445 15153996
Vilkki S Tsao JL Loukola A Poyhonen M Vierimaa O 2001 Extensive somatic microsatellite mutations in normal human tissue Cancer Res 61 4541 4544 11389087
Wei K Kucherlapati R Edelmann W 2002 Mouse models for human DNA mismatch-repair gene defects Trends Mol Med 8 346 353 12114115
Hearne CM Ghosh S Todd JA 1992 Microsatellites for linkage analysis of genetic traits Trends Genet 8 288 294 1509520
Butler J 2001 Forensic DNA typing London Academic Press 322 p.
Schlotterer C 2001 Genealogical inference of closely related species based on microsatellites Genet Res 78 209 212 11865709
Bowcock AM Ruiz-Linares A Tomfohrde J Minch E Kidd JR 1994 High resolution of human evolutionary trees with polymorphic microsatellites Nature 368 455 457 7510853
Shinde D Lai Y Sun F Arnheim N 2003 Taq DNA polymerase slippage mutation rates measured by PCR and quasi-likelihood analysis: (CA/GT)n and (A/T)n microsatellites Nucleic Acids Res 31 974 980 12560493
Meyerowitz EM Somerville CR 1994 Arabidopsis Plainview (New York) Cold Spring Harbor Laboratory Press 1,300 p.
Lian C Oishi R Miyashita N Hogetsu T 2004 High somatic instability of a microsatellite locus in a clonal tree, Robinia pseudoacacia
Theor Appl Genet 108 836 841 14625672
Cloutier D Rioux D Beaulieu J Schoen DJ 2003 Somatic stability of microsatellite loci in Eastern white pine, Pinus strobus L Heredity 90 247 252 12634808
Woodrick R Martin PR Birman I Pickett FB 2000 The Arabidopsis embryonic shoot fate map Development 127 813 820 10648239
Furner IJ Pumfrey JE 1992 Cell fate in the shoot apical meristem of Arabidopsis thaliana
Development 115 755 764
Leonard JM Bollmann SR Hays JB 2003 Reduction of stability of Arabidopsis genomic and transgenic DNA-repeat sequences (microsatellites) by inactivation of AtMSH2 mismatch-repair function Plant Physiol 133 328 338 12970498
Hellani A Coskun S Benkhalifa M Tbakhi A Sakati N 2004 Multiple displacement amplification on single cell and possible PGD applications Mol Hum Reprod 10 847 852 15465849
Deng G Chen A Hong J Chae HS Kim YS 1999 Methylation of CpG in a small region of the hMLH1 promoter invariably correlates with the absence of gene expression Cancer Res 59 2029 2033 10232580
Shibata D Peinado MA Ionov Y Malkhosyan S Perucho M 1994 Genomic instability in repeated sequences is an early somatic event in colorectal tumorigenesis that persists after transformation Nat Genet 6 273 281 8012390
Hood L Galas D 2003 The digital code of DNA Nature 421 444 448 12540920
Brinkmann B Klintschar M Neuhuber F Huhne J Rolf B 1998 Mutation rate in human microsatellites: Influence of the structure and length of the tandem repeat Am J Hum Genet 62 1408 1415 9585597
Kwiatkowski DJ Henske EP Weimer K Ozelius L Gusella JF 1992 Construction of a GT polymorphism map of human 9q Genomics 12 229 240 1339384
Petrukhin KE Speer MC Cayanis E Bonaldo MF Tantravahi U 1993 A microsatellite genetic linkage map of human chromosome 13 Genomics 15 76 85 8432553
Dallas JF 1992 Estimation of microsatellite mutation rates in recombinant inbred strains of mouse Mamm Genome 3 452 456 1643307
Dietrich W Katz H Lincoln SE Shin HS Friedman J 1992 A genetic map of the mouse suitable for typing intraspecific crosses Genetics 131 423 447 1353738
Kruglyak S Durrett RT Schug MD Aquadro CF 1998 Equilibrium distributions of microsatellite repeat length resulting from a balance between slippage events and point mutations Proc Natl Acad Sci U S A 95 10774 10778 9724780
Drake JW 1999 The distribution of rates of spontaneous mutation over viruses, prokaryotes, and eukaryotes Ann N Y Acad Sci 870 100 107 10415476
|
16261192
|
PMC1274291
|
CC BY
|
2021-01-05 09:19:22
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e50
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010050
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 10.1371/journal.pcbi.0010051plcb-01-05-09Message from ISCBThe ISCB: Growing and Evolving in Step with Science Gribskov Michael Michael Gribskov is president of the International Society for Computational Biology. E-mail: [email protected]
10 2005 28 10 2005 1 5 e51Copyright: © 2005 Michael Gribskov.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited. Citation:Gribskov M (2005) The ISCB: Growing and evolving in step with science. PLoS Comput Biol 1(5): e51.
==== Body
In 1993, the Intelligent Systems for Molecular Biology conference was successfully launched as an annual meeting for scientific exchange within the nascent interdisciplinary science of computational biology. Demand grew stronger each year for an umbrella organization that extended its reach beyond a once-a-year conference. The International Society for Computational Biology (ISCB) was formed in 1997 to provide computational biologists and bioinformaticians worldwide with a community of peers with whom they could interact year-round in their mutual quest to advance the understanding of living systems through computation. Over the past 13 years, the ISCB has grown and evolved along with the fields of computational biology and bioinformatics. As the society's president, I am proud to report that our membership now comprises nearly 2,000 members in over 50 countries.
Each year, PLoS Computational Biology, as the official journal of the ISCB, will publish the bylaws of the society. These are the legally registered rules by which the elected and appointed leaders of the society, along with the membership at large, must abide. It is my hope that ISCB members will take the time to read these rules, and to understand, thereby, the legal framework in which ISCB operates. Clearly, as ISCB grows, these rules will be adapted to changing circumstances—knowing the current rules is a first step to making any improvements to them. We also believe that our bylaws may be useful to other groups trying to form similar organizations.
International Society for Computational Biology Bylaws
Article I: Office
Section 1: Principal office.
The principal office of the Corporation shall be at San Diego Supercomputer Center, UCSD, 9500 Gilman Drive, La Jolla, CA 92093-0505
Section 2: Other offices.
The Corporation may also have an office or offices in such other place or places as the business of the Corporation may require and the Board of Directors may from time to time appoint.
Article II: Members
Section 1:Annual meeting.
The annual meeting of the members of the Corporation shall be held on a day duly designated by the Board of Directors either within or without the United States if not a legal holiday, and if a legal holiday then the next succeeding day not a legal holiday, for the transaction of such corporate business as may come before the meeting.
Section 2: Special meetings.
Special meetings of the members may be called at any time for any purpose or purposes by the Chairman of the Board, the President, a Vice President, or a majority of the Board of Directors, and shall be called forthwith by the Chairman of the Board, the President, a Vice President, the Secretary or any Director of the Corporation upon the request in writing of a majority of all the members entitled to vote on the business to be transacted at such meeting. Such request shall state the purpose or purposes of the meeting. Business transacted at all special meetings of members shall be confined to the purpose or purposes stated in the notice of the meeting.
Section 3: Place of holding meetings.
All meetings of members shall be held within or without the United States at a place designated by the Board of Directors. The members may hold their meetings in person, by conference telephone, by E-mail over the Internet or other similar electronic media, or by any combination of the foregoing.
Section 4: Notice of meetings.
Written notice of each meeting of the members shall be mailed, postage prepaid by the Secretary or person appointed by the President or sent by E-mail over the Internet or similar electronic media by the Secretary or person appointed by the President, to each member of record entitled to vote thereat at his or her post office address or E-mail address, as it appears upon the books of the Corporation, at least ten (10) days before the meeting. Each such notice shall state the place, E-mail information if the meeting will be held partly or completely by electronic means over the Internet or other electronic media, day, and hour at which the meeting is to be held and, in the case of any special meeting, shall state briefly the purpose or purposes thereof.
Section 5: Quorum.
The presence in person, by E-mail over the Internet or similar electronic media, or by proxy (each of which shall constitute “Attendance” for all purposes of these bylaws), of one third of the members of the Corporation shall constitute a quorum at all meetings of the members except as otherwise provided by law, by the Certificate of Incorporation or by these bylaws If less than a quorum shall be in Attendance at the time for which the meeting shall have been called, the meeting may be adjourned from time to time by a majority vote of the members in Attendance, without any notice other than by announcement at the meeting, until a quorum shall be in Attendance. At any adjourned meeting at which a quorum shall be in Attendance, any business may be transacted which might have been transacted if the meeting had been held as originally called.
Section 6: Conduct of meetings.
Meetings of members shall be presided over by the President of the Corporation or, if he is not in Attendance, by a Vice President, or, if none of said officers is in Attendance, by a chairman to be elected at the meeting. The Secretary of the Corporation, or if he is not in Attendance, any Assistant Secretary shall act as secretary of such meetings; in the absence of the Secretary and any Assistant Secretary, the presiding officer may appoint a person to act as Secretary of the meeting.
Section 7: Voting.
At all meetings of members every member entitled to vote thereat shall have one (1) vote. Such vote may be made either in person, by conference call, by E-mail over the Internet or similar electronic media, or by proxy appointed by an instrument in writing subscribed by such member or his or her duly authorized attorney, bearing a date not more than three (3) months prior to said meeting, unless said instrument provides for a longer period. Such proxy shall be dated, but need not be sealed, witnessed or acknowledged. All elections shall be had and all questions shall be decided by a majority of the votes cast at a duly constituted meeting, except as otherwise provided by law, in the Certificate of Incorporation or by these bylaws.
If the chairman of the meeting shall so determine, a vote by ballot may be taken upon any election or matter, and the vote shall be so taken upon the request of ten percent (10%) or more of all of the members entitled to vote on such election or matter. In either of such events, the proxies and ballots shall be received and be taken in charge and all questions touching the qualification of voters and the validity of proxies and the acceptance or rejection of votes, shall be decided by members appointed by the chairman of said meeting.
Section 8: Identity of members.
The members of the Corporation shall be composed of those individuals who have filled out a registration form and paid their dues. Individuals shall retain their status as members so long as they pay any and all annual dues imposed by the Corporation upon its members.
Section 9: Members' voting rights.
The members of the Corporation shall nominate from the membership, no later than their annual meeting in the manner set forth in the Procedure for Nomination, up to four (4) Directors who shall also be concurrently nominated to hold one of the Nominated Offices as further defined in Article IV, Section 1 below. Said nominations shall be considered strong recommendations from the members to the Board of Directors to elect these individuals as officers and Directors of the Corporation (the “Director/Officers”). Upon their election by the Directors as set forth in Article III, Section 4 below, these four Director/Officers shall serve as officers of the Corporation for one year in the case of President-elect, and two years for all other officers including President. The four Director/Officers shall also concurrently serve as Directors of the Corporation during their full term as officers plus one year following completion of their term or terms of office, unless removed by the Directors pursuant to the provisions of Article III, Section 4(iv) below. The membership shall follow such rules pertaining to the nomination of officers as are promulgated by the Board of Directors in the Procedure For Nomination adopted as of March 2003 as attached to these bylaws as Exhibit A, or as later amended.
Section 10: Directors as members.
The Board of Directors of the Corporation shall be members. A Director who fails to keep his or her membership current during the course of his or her term will forfeit voting rights until current year registration is rectified, or resign from the Board of Directors if not registered by April 1 of the membership year.
Article III: Board of Directors
Section 1: General powers.
The property and business of the Corporation shall be managed under the direction of the Board of Directors of the Corporation.
Section 2: Number.
The number of Directors, including Director/Officers, shall be twenty eight (28) or such other number, but not less than three (3) nor more than thirty (30), as may be designated from time to time by resolution of a majority of the entire Board of Directors.
Section 3: Term of office.
The term of office of each Director, other than a Director/Officer, shall be for three (3) years; the term of office of each Director/Officer shall be for the duration of the term or terms of office of the officer plus one additional year after the final term as officer.
Section 4: Filling of vacancies.
In the case of any vacancy in the Board of Directors, including Director/Officers, the remaining Directors may elect a successor at a regular or special meeting of Directors, by affirmative vote of the majority thereof.
In the case of any vacancy in the Director/Officers, the Directors may, but are not required to, ask the members to make nominations under the Procedure for Nomination.
If the number of Directors is increased as provided in these bylaws, the additional Directors so provided for shall be elected by an affirmative vote of a majority of the entire Board of Directors already in office to hold office for a three (3) year term.
Notwithstanding the foregoing provisions, any Director, including a Director/Officer, may be removed from office as a Director with or without cause by the affirmative vote of a majority of the Directors entitled to vote at any regular or special meeting of Directors called for that purpose. A Director who is a Director/Officer shall be automatically removed as an officer of the Corporation upon his or her removal as a Director.
Section 5: Place of meeting.
The Board of Directors may hold their meetings and have one or more offices, and keep the books of the Corporation, either within or outside the State of California, at such place or places as they may from time to time determine by resolution or by written consent of all the Directors. The Board of Directors may hold their meetings in person, by conference telephone, by E-mail over the Internet or other similar electronic media, or by any combination of the foregoing.
Section 6: Regular meetings.
Regular meetings of the Board of Directors may be held at such time and place as shall from time to time be determined by resolution of the Board, provided that written notice of each meeting of the Board of Directors shall be mailed, postage prepaid by the Secretary or person appointed by the President or sent by E-mail over the Internet or similar electronic media by the Secretary or person appointed by the President, to each Director and Director/Officer at his or her post office address or E-mail address, as it appears upon the books of the Corporation, at least ten (10) days before the meeting. The annual meeting of the Board of Directors shall be held within ten days prior to or following the annual meeting of members. Any business may be transacted at any regular meeting of the Board.
Section 7: Special meetings.
Special meetings of the Board of Directors shall be held whenever called by any member of the Board of Directors. The Secretary or person appointed by the President shall give notice of each special meeting of the Board of Directors, by mailing the same at least three (3) days prior to the meeting or by E-mailing over the Internet or similar electronic media, the same at least two (2) days before the meeting, to each Director; but such notice may be waived by any Director. Unless otherwise indicated in the notice thereof, any and all business may be transacted at any special meetings. At any meeting at which every Director shall be in Attendance, even though without notice, any business may be transacted and any Director may in writing or by E-mail over the Internet or similar electronic media, waive notice of the time, place and objectives of any special meeting.
Section 8: Quorum.
The presence in person, by conference call, by E-mail over the Internet or similar electronic media, or by proxy appointed by an instrument in writing subscribed by such Director or his or her duly authorized attorney, bearing a date not more than three (3) months prior to said meeting, unless said instrument provides for a longer period (such proxy shall be dated, but need not be sealed, witnessed or acknowledged) (each of which shall constitute “Attendance” for all purposes of these bylaws) of a majority of the whole number of Directors shall constitute a quorum for the transaction of business at all meetings of the Board of Directors. If at any meeting less than a quorum shall be in Attendance, a majority of those in Attendance may adjourn the meeting from time to time. The act of a majority of the Directors in Attendance at any meeting at which there is a quorum shall be the act of the Board of Directors, except as may be otherwise specifically provided by law, by the Certificate of Incorporation or by these bylaws.
Section 9: Required vote.
An affirmative vote of a majority of those in Attendance shall be necessary for the passage of any resolution.
Section 10: Compensation of directors.
Directors, including Director/Officers, shall not receive any stated salary for their services as such, but, at the discretion and unanimous approval by the Executive Committee, a fixed sum may be allowed for Attendance at each regular or special meeting of the Board and such compensation shall be payable whether or not a meeting is adjourned because of the absence of a quorum. This sum may be disbursed to all Directors, including Director/Officers, or to individual Directors or Director/Officers with extenuating circumstances regarding lack of institutional reimbursement of costs for Attendance at each regular or special meeting. If the sum to any or all of the Directors and Director/Officers exceeds $5000, it must be approved by majority vote of the Board. Nothing herein contained shall be construed to preclude any Director or Director/Officer from serving the Corporation in any other capacity, and receiving compensation therefore.
Section 11: Standing committees.
The Board of Directors shall elect a septe Standing Committee for the conference coordination for the annual ISMB meeting which is responsible for oversight of all ISCB conference activities sponsored by and affiliated with the Corporation. Another Standing Committee shall be elected to assume responsibility for publications of the society. A Task Force shall be selected for the election of officers and for the setting of standards for voting at meetings held on the Internet or by other electronic media and shall consist of such Directors and members as determined by majority vote of the Directors. The Standing Committees shall be selected by the Board of Directors at any meeting of the Board of Directors. Any and all Task Forces may be selected by the Board of Directors at any duly constituted meeting of the Directors or by the Committee Chair(s) at any committee meeting.
Section 12: Other committees.
The Board of Directors may, by resolution passed by a majority of the whole Board, designate one or more other committees, each committee to consist of one or more of the Directors of the Corporation, which, to the extent provided in the resolution, shall have and may exercise the powers of the Board of Directors, and may authorize the seal of the Corporation to be affixed to all papers which may require it. Such committee or committees shall have such names as may be determined from time to time by resolution adopted by the Board of Directors.
Section 13: Task forces.
The Board of Directors and committee chairs may, by resolution passed by a majority of the whole Board or at the direction of the committee chairs, designate one or more Task Forces, each Task Force to consist of one or more of the Directors of the Corporation or one or more of the committee members for which the Task Force shall be formed. Task Forces shall be formed for the purposes of task-specific time-limited work in order to make recommendations to the Board of Directors or committee chairs. Task Forces shall not have nor exercise the powers of the Board of Directors, nor authorize the seal of the Corporation to be affixed to any papers which may require it. Such Task Forces shall have such names as may be determined from time to time by resolution adopted by the Board of Directors or at the direction of the committee chairs for which they have formed.
Article IV: Officers
Section 1: Election, tenure and compensation.
The officers of the Corporation shall be a President-elect, a President, a Vice President, a Secretary, and a Treasurer (the “Nominated Officers”), and also such other officers including a Chairman of the Board and/or one or more Vice Presidents and/or one or more assistants to the foregoing officers as the Board of Directors from time to time may appoint for the proper conduct of the business of the Corporation, including but not limited to an Executive Officer to assist in day to day business matters and a Vice Chair of Conferences to assist in managing the annual conference (collectively the “Appointed Officers”). The Nominated Officers shall be nominated by the members and Directors according to the procedures set forth in the Procedure For Nomination. The Nominated Officers shall serve for two years as officers. Any two or more of the above offices, except those of President and Secretary, may be held by the same person, but no officer shall execute, acknowledge or verify any instrument in more than one capacity if such instrument is required by law or by these bylaws to be executed, acknowledged or verified by any two or more officers. If compensation or salary is paid to officers or Director/Officers of the Corporation it shall be fixed by resolutions adopted by the Board of Directors.
In the event that any office other than an office required by law, shall not be filled by the members, or, once filled, subsequently becomes vacant, then such office and all references thereto in these bylaws shall be deemed inoperative unless and until such office is filled in accordance with the provisions of these bylaws.
Except where otherwise expressly provided in a contract duly authorized by the Board of Directors, all officers of the Corporation shall be subject to removal at any time by the affirmative vote of a majority of the Board of Directors or the whole membership at a special meeting of the Board of Directors or the members respectively, duly called according to the rules set forth in these bylaws. All agents and employees of the Corporation shall be subject to removal at any time by the affirmative vote of a majority of the whole Board of Directors, and any agents and employees, other than those elected by the membership, shall hold office at the discretion of the Board of Directors or of the officers appointing them.
Section 2: Powers and duties of the chairman of the board.
The Chairman of the Board shall preside at all meetings of the Board of Directors unless the Board of Directors shall by a majority vote of a quorum thereof elect a chairman other than the Chairman of the Board to preside at meetings of the Board of Directors. The Chairman may sign and execute all authorized bonds, contracts or other obligations in the name of the Corporation; and he or she shall be ex-officio a member of all standing committees.
Section 3: Powers and duties of the president.
The President shall be the chief executive officer of the Corporation and shall have general charge and control of all its business affairs and properties. He or she shall preside at all meetings of the members.
The President may sign and execute all authorized bonds, contracts or other obligations in the name of the Corporation. The President shall have signature power and the authority to assign signature power to the Executive Officer and/or Vice Chair of Conferences to sign checks in amounts up to $5,000. All checks for amounts over $5,000 for costs not associated with contracts and expenses previously approved by the Board of Directors shall require approval of the Board of Directors and the signature of two officers or one officer and one Appointed Officer. The President shall have the general powers and duties of supervision and management usually vested in the office of president of a corporation. The President shall be ex-officio a member of all the Standing committees. He shall do and perform such other duties as may, from time to time, be assigned to him by the Board of Directors.
In the event that the membership does not take affirmative action to fill the office of Chairman of the Board, the President shall assume and perform all powers and duties given to the Chairman of the Board by these bylaws.
Section 4: Powers and duties of the president-elect.
The President-elect may sign and execute all authorized bonds, contracts, or other obligations in the name of the Corporation. The President-elect shall have such other powers and shall perform such other duties as may be assigned to the President-elect by the Board of Directors or by the President. In case of the absence or disability of the President, the duties of that office shall be performed by the President-elect, and the taking of any action by the President-elect in place of the President shall be conclusive evidence of the absence or disability of the President.
Section 5: Powers and duties of the vice president.
The Board of Directors may appoint more than one Vice President. Any Vice President (unless otherwise provided by resolution of the Board of Directors) may sign and execute all authorized bonds, contracts, or other obligations in the name of the Corporation. Each Vice President shall have such other powers and shall perform such other duties as may be assigned to the Vice President by the Board of Directors or by the President. In case of the absence or disability of the President and President-elect, the duties of the office of President shall be performed by any Vice President, and the taking of any action by any such Vice President in place of the President shall be conclusive evidence of the absence or disability of the President.
Section 6: Secretary.
The Secretary shall handle all voting matters, whether at actual meetings, telephonic meetings or meetings held on the Internet or other electronic media; he or she shall give, or cause to be given, notice of all meetings of members and Directors and all other notices required by law or by these bylaws, and in case of his or her absence or refusal or neglect to do so, any such notice may be given by any person thereunto directed by the President, or by the Directors or members upon whose written request the meeting is called as provided in these bylaws. The Secretary shall record all the proceedings of the meetings of the members and of the Directors in books provided for that purpose, and he or she shall perform such other duties as may be assigned to him or her by the Directors or the President. He or she shall have custody of the seal of the Corporation and shall affix the same to all instruments requiring it, when authorized by the Board of Directors or the President, and attest the same. In general, the Secretary shall perform all the duties generally incident to the office of Secretary, subject to the control of the Board of Directors and the President.
Section 7: Treasurer.
The Treasurer shall oversee the Executive Officer's and/or Vice Chair of Conferences' custody of all the funds and securities of the Corporation, and he or she shall oversee the Executive Officer's and/or Vice Chair of Conferences' full and accurate account of receipts and disbursements in books belonging to the Corporation. The Treasurer shall oversee the Appointed Officers' deposit of all moneys and other valuables in the name and to the credit of the Corporation in such depository or depositories as may be designated by the Board of Directors. He or she shall have the power to sign checks under his or her signature in amounts up to $5,000. All checks for amounts over $5,000 shall require the signature of two officers or one officer and one Appointed Officer.
The Treasurer shall oversee the Appointed Officers' disbursement of the funds of the Corporation as may be ordered by the Board of Directors and may require the Appointed Officers to make proper vouchers for such disbursements. The Treasurer shall render to the President and the Board of Directors, whenever either of them so requests, an account of all of the Appointed Officers' transactions and of the financial condition of the Corporation.
The Treasurer shall give the Corporation a bond, if required by the Board of Directors, in a sum, and with one or more sureties, satisfactory to the Board or Directors, for the faithful performance of the duties of his or her office and for the restoration to the Corporation in case of his or her death, resignation, retirement or removal from office of all books, papers, vouchers, moneys, and other properties of whatever kind in his or her possession or under his or her control belonging to the Corporation.
The Treasurer shall perform all the duties generally incident to the office of the Treasurer, subject to the control of the Board of Directors and the President.
Section 8: Assistant secretary.
The Board of Directors may appoint an Assistant Secretary or more than one Assistant Secretary. Each Assistant Secretary shall (except as otherwise provided by resolution of the Board of Directors) have power to perform all duties of the Secretary in the absence or disability of the Secretary and shall have such other powers and shall perform such other duties as may be assigned to him by the Board of Directors or the President. In case of the absence or disability of the Secretary, the duties of the office shall be performed by any such Assistant Secretary, and the taking of any action by any such Assistant Secretary in place of the Secretary shall be conclusive evidence of the absence or disability of the Secretary.
Section 9: Assistant treasurer.
The Board of Directors may appoint an Assistant Treasurer or more than one Assistant Treasurer. Each Assistant Treasurer shall (except as otherwise provided by resolution of the Board of Directors) have power to perform all duties of the Treasurer in the absence or disability of the Treasurer and shall have such other powers and shall perform such other duties as may be assigned to him by the Board of Directors or the President. In case of the absence or disability of the Treasurer, the duties of the office shall be performed by any Assistant Treasurer, and the taking of any action by any such Assistant Treasurer in place of the Treasurer shall be conclusive evidence of the absence or disability of the Treasurer.
Article V: Corporate Seal
Section 1: Seal.
In the event that the President shall direct the Secretary to obtain a corporate seal, the corporate seal shall be circular in form and shall have inscribed thereon the name of the Corporation, the year of its organization and the word “Delaware”. Duplicate copies of the corporate seal may be provided for use in the different offices of the Corporation but each copy thereof shall be in the custody of the Secretary of the Corporation or of an Assistant Secretary of the Corporation nominated by the Secretary.
Article VI: Bank Accounts and Loans
Section 1: Bank accounts.
Such officers or agents of the Corporation as from time to time shall be designated by the Board of Directors shall have authority to deposit any funds of the Corporation in such banks or trust companies as shall from time to time be designated by the Board of Directors and such officers or agents as from time to time shall be authorized by the Board of Directors to withdraw any or all of the funds of the Corporation so deposited in any such bank or trust company, upon checks, drafts or other instruments or orders for the payment of money, drawn against the account or in the name or behalf of this Corporation, and made or signed by such officers or agents; and each bank or trust company with which funds of the Corporation are so deposited is authorized to accept, honor, cash and pay, without limit as to amount, all checks, drafts or other instruments or orders for the payment of money, when drawn, made or signed by officers or agents so designated by the Board of Directors until written notice of the revocation of the authority of such officers or agents by the Board of Directors shall have been received by such bank or trust company. There shall from time to time be certified to the banks or trust companies in which funds of the Corporation are deposited, the signature of the officers or agents of the Corporation so authorized to draw against the same. In the event that the Board of Directors shall fail to designate the persons by whom checks, drafts and other instruments or orders for the payment of money shall be signed, as hereinabove provided in this Section, all of such checks, drafts and other instruments or orders for the payment of money shall be signed by the President or a Vice President and countersigned by the Secretary or Treasurer or an Assistant Secretary or an Assistant Treasurer or Appointed Officer of the Corporation.
Section 2: Loans.
Such officers or agents of this Corporation as from time to time shall be designated by the Board of Directors shall have authority to effect loans, advances or other forms of credit at any time or times for the Corporation from such banks, trust companies, institutions, corporations, firms or persons as the Board or Directors shall from time to time designate, and as security for the repayment of such loans, advances, or other forms of credit to assign, transfer, endorse and deliver, either originally or in addition or substitution, any or all stocks, bonds, rights and interests of any kind in or to stocks or bonds, certificates of such rights or interests, deposits, accounts, documents covering merchandise, deposits and accounts receivable and other commercial paper and evidences of debt at any time held by the Corporation; and for such loans, advances or other forms of credit to make, execute and deliver one or more notes, acceptances or written obligations of the Corporation on such terms, and with such provisions as to the security or sale or disposition thereof as such officers or agents shall deem proper; and also to sell to, or discount or rediscount with, such banks, trust companies, institutions, corporations, firms or persons any and all commercial paper, bills receivable, acceptances and other instruments and evidences of debt at any time held by the Corporation, and to that end to endorse, transfer and deliver the same. There shall from time to time be certified to each bank, trust company, institution, corporation, firm or person so designated the signatures of the officers or agents so authorized; and each such bank, trust company, institution, corporation, firm or person is authorized to reply upon such certification until written notice of the revocation by the Board of Directors of the authority of such officers or agents shall be delivered to such bank, trust company, institution, corporation, firm or person.
Article VII: Reimbursements
Any payments made to an officer or other employee of the Corporation, such as salary, commission, interest or rent, or entertainment expense incurred by him or her, which shall be disallowed in whole or in part as a deductible expense by the Internal Revenue Service, shall be reimbursed by such officer or other employee of the Corporation to the full extent of such disallowance. It shall be the duty of the Directors, as a Board, to enforce payment of each such amount disallowed. In lieu of payment by the officer or other employee, subject to the determination of the Board of Directors, proportionate amounts may be withheld from his or her future compensation payments until the amount owed to the Corporation has been recovered.
Article VIII: Miscellaneous Provisions
Section 1: Fiscal year.
The fiscal year of the Corporation shall end on the last day of December.
Section 2: Notices.
Whenever, under the provisions of these bylaws, notice is required to be given to any Director, officer or member it shall not be construed to mean personal notice, but such notice shall be given in writing, by E-mail over the Internet or similar electronic media, by mail, by depositing the same in a post office or letter box, in a postpaid sealed wrapper, addressed to each member, officer or Director at such address as appears on the books of the Corporation, or in default of any other address, to such Director, officer or member at the general post office in the town of La Jolla, California, and such notice shall be deemed to be given at the time the same shall be thus mailed. Any member, Director or officer may waive any notice required to be given under these bylaws.
Section 3: Waiver, consent.
Any notice required to be given under these bylaws or otherwise may be waived by the Director, officer or member to whom such notice is required to be given and the Attendance of any person at a meeting shall constitute waiver of notice thereof as to such person. Any action which may be taken at a meeting of the Directors, officers or members may be taken without a meeting if a consent in writing, setting forth the action so taken, shall be signed by all of the Directors, officers or members entitled to vote with respect to the subject matter thereof. Such consent shall have the same force and effect as a unanimous vote of the Directors, officers or members, as the case may be.
Article IX: Amendments
Section 1. Amendment of bylaws.
Any member can propose an amendment of the bylaws by submitting the change to the President. If a majority of the Board of Directors adopts the amendment it shall be adopted.
Article X: Indemnification
Section 1: Indemnification of directors and officers.
The Corporation shall indemnify and advance expenses to a Director, officer or Appointed Officer of the Corporation in connection with a proceeding to the fullest extent permitted by and in accordance with the General Corporation Law of the State of Delaware.
Section 2: Indemnification of employees and agents.
With respect to an employee or agent, other than a Director, officer or Appointed Officer of the Corporation, the Corporation may, as determined by the Board of Directors of the Corporation, indemnify and advance expenses to such employee or agent in connection with a proceeding to the extent permitted by and in accordance with the General Corporation Law of the State of Delaware.
September 16, 2003: Bylaws amended by majority vote of the Board of Directors
|
0
|
PMC1274292
|
CC BY
|
2021-01-05 09:18:23
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e51
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010051
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 1626119410.1371/journal.pcbi.0010053plcb-01-05-05Research ArticleSNPdetector: A Software Tool for Sensitive and Accurate SNP Detection An Automated SNP and Mutation Detection Tool Zhang Jinghui 1*Wheeler David A 2Yakub Imtiaz 2Wei Sharon 2Sood Raman 3Rowe William 1Liu Paul P 3Gibbs Richard A 2Buetow Kenneth H 11 Laboratory of Population Genetics, National Cancer Institute, National Institutes of Health, Bethesda, Maryland, United States of America
2 Human Genome Sequencing Center and Department of Molecular and Human Genetics, Baylor College of Medicine, Houston, Texas, United States of America
3 Genetics and Molecular Biology Branch, National Human Genome Research Institute, National Institutes of Health, Bethesda, Maryland, United States of America
Marth Gabor EditorBoston College, United States of America* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 1 5 e5319 5 2005 21 9 2005 Copyright: © 2005 Zhang et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Identification of single nucleotide polymorphisms (SNPs) and mutations is important for the discovery of genetic predisposition to complex diseases. PCR resequencing is the method of choice for de novo SNP discovery. However, manual curation of putative SNPs has been a major bottleneck in the application of this method to high-throughput screening. Therefore it is critical to develop a more sensitive and accurate computational method for automated SNP detection. We developed a software tool, SNPdetector, for automated identification of SNPs and mutations in fluorescence-based resequencing reads. SNPdetector was designed to model the process of human visual inspection and has a very low false positive and false negative rate. We demonstrate the superior performance of SNPdetector in SNP and mutation analysis by comparing its results with those derived by human inspection, PolyPhred (a popular SNP detection tool), and independent genotype assays in three large-scale investigations. The first study identified and validated inter- and intra-subspecies variations in 4,650 traces of 25 inbred mouse strains that belong to either the Mus musculus species or the M. spretus species. Unexpected heterozgyosity in CAST/Ei strain was observed in two out of 1,167 mouse SNPs. The second study identified 11,241 candidate SNPs in five ENCODE regions of the human genome covering 2.5 Mb of genomic sequence. Approximately 50% of the candidate SNPs were selected for experimental genotyping; the validation rate exceeded 95%. The third study detected ENU-induced mutations (at 0.04% allele frequency) in 64,896 traces of 1,236 zebra fish. Our analysis of three large and diverse test datasets demonstrated that SNPdetector is an effective tool for genome-scale research and for large-sample clinical studies. SNPdetector runs on Unix/Linux platform and is available publicly (http://lpg.nci.nih.gov).
Synopsis
Single nucleotide polymorphisms (SNPs) are an abundant and important class of heritable genetic variations, and many of them contribute to genetic diseases. Accurate and automated detection of SNPs as heterozygous alleles in fluorescence-based sequencing traces from diploid DNA samples is challenging because of the low signal-to-noise ratio in the data, and because of sequencing artifacts associated with the various DNA sequencing chemistries.
The authors of this publication have developed a new computer program, SNPdetector, that improves upon existing software tools. The main design principle of SNPdetector was to model the process of human visual inspection of experienced analysts. The new tool is able to cut down significantly on both false positive and false negative discovery rates. Good performance can be achieved, without the need for retraining, in substantially different datasets such as SNP discovery in human resequencing data, mutation discovery in zebra fish candidate genes, and discovery of inter- and intra-subspecies variations in inbred mouse strains. The results demonstrate that this software tool is suitable for the automation of SNP discovery in diploid sequencing traces, and permits a substantial reduction of costly and laborious visual data analysis.
Citation:Zhang J, Wheeler DA, Yakub I, Wei S, Sood R, et al. (2005) SNPdetector: A software tool for sensitive and accurate SNP detection. PLoS Comput Biol 1(5): e53.
==== Body
Introduction
Identification of genetic variations and mutations is important for the discovery of genetic predisposition to complex diseases. Although a wide variety of methods are available for de novo single nucleotide polymorphism (SNP) discovery [1], DNA sequencing is the method of choice for high-throughput screening studies. DNA sequencing may follow either a random shotgun strategy [2–5] or a directed strategy using PCR amplification of specific target regions of interest [6]. As the high-density haplotype map of the human genome [7] nears completion, the demand for large-scale SNP surveys seeking genetic mutations linked to or causative of a wide variety of human diseases (such as diabetes, heart disease, and cancer) is expected to greatly increase [8].
Direct sequencing of PCR-amplified genomic fragments from diploid samples results in mixed sequencing templates. Therefore, one of the most challenging issues in SNP discovery by this method is to distinguish bona fide heterozygous allelic variations from sequencing artifacts, which can give rise to two overlapping fluorescence peaks similar to true heterozygotes. Currently, PolyPhred [9] is the most widely used SNP discovery software for such an analysis. It reports a heterozygous allele only when the site shows a decrease of about 50% in peak height compared to the average height for homozygous individuals. However, inspection of the computational results by human analysts is often required to ensure a low false positive rate, a labor-intensive process.
To provide a sensitive and accurate method for SNP detection in fluorescence-based resequencing, we developed a new software tool, SNPdetector, aiming to “computerize” the manual review process. We report SNPdetector's application in three large-scale genetic variation studies and compare its results with those obtained by human inspection, by PolyPhred, and by experimental validation. In the first study, resequencing was used to validate mouse SNPs discovered by whole-genome shotgun sequencing. The second study identifies novel SNPs in the ENCODE regions of the human genome [10], and the third study aims to discover mutations induced by ENU in 1,236 zebra fish.
Results
System Design of SNPdetector
SNPdetector processes one PCR amplicon at a time with the following four main steps (Figure 1). (1) Run the program Phred [11] to derive base calls, quality scores, and primary and secondary peak information for each trace file. (2) Align sequence reads obtained by resequencing to a reference sequence using SIM [12], a program that implements the Smith–Waterman algorithm. This ensures that all PCR reads are optimally aligned even when there is substantial sequence variation. The ends of the alignments are trimmed, and the user can choose to filter low-quality reads and/or high-quality reads with poor alignments (usually a signal of misassembly). (3) Identify high-quality sequence variations using neighborhood quality standard (NQS) [3], which requires a variation site and each base in its flanking window to exceed a user-defined quality score threshold. NQS was originally developed for automated SNP identification in haploid samples, which is similar to finding SNPs in diploid samples that have homozygous minor alleles. (4) Identify heterozygous genotypes and evaluate the validity of all SNPs. This last step determines the genotype for each sample by analyzing trace files of both forward and reverse orientations. It screens potential systematic sequencing errors by “computerizing” techniques such as horizontal and vertical scanning employed by experienced SNP inspectors. The implementation details are described in Materials and Methods.
Figure 1 Schematic Diagram of the Principal Steps in the Analysis of Sequencing Variants Found by SNPdetector
Paralellograms are analytical modules (usually C programs), and rectangles are input and output data. Programs obtained from the public domain are displayed in italics while those developed in this work are shown in bold. SNPdetector requires the following three sets of input data: (1) a template sequence file, (2) the forward and the reverse sequencing primers, and (3) the trace files. The output includes a list of high-quality SNPs and their genotype calls in each subject.
During the development of SNPdetector, the mouse resequencing data and a subset of the zebra fish resequencing data were used as training data for developing filters for false positive calls and to determine the lower bound of the signal thresholds for identifying true positive variation. The human resequencing data in the ENCODE regions were not included in training but were used as an independent testing dataset to evaluate the accuracy of the software after training had taken place. The results of the three studies presented here were obtained using the same software configuration.
Validation and Detection of SNPs in Inbred Mouse Strains
In this investigation we attempted to validate 151 mouse SNPs on Chromosome 16 that were originally discovered by shotgun sequencing of seven laboratory inbred strains [13,14]. We designed 93 sets of forward and reverse PCR primers to assay 40 kb of genomic sequence in 25 inbred strains (details in Materials and Methods of [13]).
During the development of SNPdetector, this dataset was used as the primary training data for identifying sequencing artifacts that are likely to produce false positive SNP calls. The mouse data were chosen for the following reasons. (1) Mouse inbred strains are expected to be either completely homozygous or to have an extremely low rate of heterozygosity as a result of their breeding history. This expectation has been confirmed experimentally by intra-strain variation analysis [14]. Therefore, the vast majority of the heterozygous genotype calls are false positives resulting from sequencing artifacts. (2) In a previous study we analyzed the high-resolution haplotype structure as well as the phylogeny of the mouse strains in the resequenced regions [13]. The knowledge of inbred mouse genetic architecture thus acquired provides an additional reference for resolving ambiguities in SNP validity assessment. (3) The 151 SNPs discovered by genomic shotgun sequencing provide independent verification for SNPs discovered by PCR resequencing.
When using the option to include all sequence reads, SNPdetector identified a total of 1,178 SNPs in all 25 strains. For each SNP, the assembled trace data were manually reviewed using the program Consed [15]. Manual inspection found 11 SNPs to be invalid. SNPdetector found all but two of the SNPs originally discovered by genomic shotgun sequencing. Manual inspection revealed that the two missing SNPs reside in regions where sequences “stutter,” one (dbSNP rs4171354) caused by a polynucleotide track and the other (dbSNP rs4139636) by a simple tandem repeat (STR). Thus these two “SNPs” represent sequence variations resulting from slipped strand extension of DNA polymerase in PCR rather than genetic polymorphisms. Excluding low-quality reads, SNPdetector found 1,019 SNPs, 1,009 of which are valid.
We ran the same dataset using the Phred/Phrap/PolyPhred 5.0.2 package and manually reviewed 187 putative SNPs with score ≥ 30 that were found only by PolyPhred. In the alignments produced by SIM and Phrap, gap locations can vary if they reside in polynucleotide repeats or tandem repeats. As a result, SNPdetector and PolyPhred may produce slightly different locations for a substitution variation adjacent to an insertion/deletion (indel) polymorphism. These discrepancies were manually resolved by comparing the alignments of Phrap and SIM. There were 34 additional valid SNPs in the PolyPhred output, which makes a total of 1,201 valid SNPs when combined with the valid SNPs found by SNPdetector.
Using the 1,201 valid SNPs as a benchmark of the total number of valid SNPs in the mouse resequencing data, we analyzed the error rate for SNPdetector and PolyPhred 5.0.2. The results are summarized in Table 1. SNPdetector, with the run time option of including all sequence reads, has the lowest false positive and false negative rates (0.93% and 2.58%, respectively). In the output of PolyPhred, the results obtained from score ≥ 97 have the lowest false positive rate (5.31%) and those from score ≥ 30 have the lowest false negative rate (14.73%).
Table 1 Comparison of the Results Obtained by SNPdetector and PolyPhred (Version 5.0.2) in Mouse Resequencing
Discovery of Heterozygous Alleles in the Wild-Derived Inbred Mouse Strain CAST/Ei
In the mouse sequencing data described above, SNPdetector identified 11 putative SNPs with heterozygous genotypes in at least one of the 25 inbred strains. As part of the quality assurance process, we manually reviewed all of these SNPs. Six of the putative SNPs were false positives arising from background noise. However, manual inspection detected no sequencing artifact in the remaining five markers. The markers are located at two genomic loci. One of the two loci encodes the EphA6 gene, and the two wild-derived inbred strains, MOLF/Ei and SPRET/Ei, are heterozygous at three putative SNP sites. The other locus encodes Bach1, a heme-binding transcription factor. CAST/Ei, a wild-derived inbred strain of the Mus. mus. castaneus subspecies, was heterozygous at two putative SNP sites located at the 3′ UTR of Bach1. None of the five markers has homozygous minor alleles.
To determine the validity of the unexpected heterozygosity in the inbred strains, we redesigned sequencing primers to assay DNA samples of the original animals as well as additional animals of CAST/Ei (three animals), MOLF/Ei (one animal), and SPRET/Ei (one animal) strains. The observed heterozygosity at the EphA6 locus turned out to be an artifact of genomic duplication coupled with polymorphisms of MOLF/Ei and SPRET/Ei strains at the sequencing primer binding sites. However, no genomic duplication was found for the Bach1 locus using the current mouse assembly (March 2005 release of NCBI build 34). Genotypes of the four CAST/Ei animals are summarized in Table 2. One CAST/Ei animal was homozygous at the first SNP site while all other sites were heterozygous (Figure S1). The minor alleles in the CAST/Ei heterozygote were the same as those represented in the human orthologous sequence.
Table 2 Heterozygosity at the Bach1 Locus in Animals of Wild-Derived Inbred Strain CAST/Ei
SNP Discovery in Human
We used SNPdetector for SNP discovery as part of the HapMap project [7]. In the SNP discovery phase 48 unrelated individuals from four populations were chosen for resequencing: 16 from the Centre d'Etude du Polymorphisme Humain collection [16]; 16 from Yoruba individuals from Ibadan, Nigeria; eight from Japanese individuals from Tokoyo, Japan; and eight from Han Chinese individuals from Beijing, China. Cell lines for each are available from the Coriell Institute for Medical Research (http://locus.umdnj.edu/nigms/products/hapmap.html).
A total of 11,241 candidate SNPs were found across all regions, of which approximately half (51.9%) were novel, compared to data in build 121 of dbSNP. In all, 80% of the SNPs had a minor allele frequency greater than 0.05.
Nearly 6,000 of the SNPs identified by SNPdetector were selected for genotyping in expanded assay panels using a variety of commercial and academic genotyping platforms as part of the International HapMap Project. The larger panels consisted of the following samples: 90 from the Centre d'Etude du Polymorphisme Humain collection; 90 from Yoruba individuals from Ibadan, Nigeria; 45 from Han Chinese individuals from Beijing, China; and 45 from Japanese individuals from Tokoyo, Japan; the 16 individuals used in the SNP discovery phase were a subset of the genotype panel. Candidate SNPs that turned out to be monomorphic across all populations were considered to be false positives. The false positive rate ranged from 2% on Chromosome 12 to 5.8% on Chromosome 7 (Table 3). The overall SNP validation rate was 95.5%.
Table 3 Genotyping of Candidate SNPs Identified by SNPdetector in Human ENCODE Regions
To compare the performance of SNPdetector with the other SNP detection programs, we reanalyzed a subset of ENCODE data (61 amplicons on Chromosome 18) using PolyPhred 5.0.2 and NovoSNP [17] (a new SNP detection software package). We did not run this analysis on the entire ENCODE dataset because for computational SNPs that do not have genotype data, we had to manually review sequence traces to assess their validity.
A total of 85 valid SNPs were found by at least one of the three programs, 71 of which were verified by experimental genotyping, and 14 by manual review alone. The results, summarized in Table 4, show that the false positive and the false negative rates of SNPdetector are much lower than those of the other two programs for this dataset: the combined false positive and false negative error of SNPdetector is approximately half of the lowest error rate in PolyPhred and one tenth of that in NovoSNP.
Table 4 Comparison of SNPdetector with PolyPhred 5.0.2 and NovoSNP on a Subset of ENCODE Data
Detection of ENU-Induced Mutations in 1,236 Zebra Fish
A total of 26 pairs of forward and reverse primers were designed to identify ENU-induced mutations in 1,236 zebra fish in several candidate genes. The fish population is expected to have common and rare polymorphisms in addition to mutations. Each mutation is expected to be present as a heterozygote in only one fish, resulting in minor allele frequency of 4.0 × 10−4 in the overall population.
To detect mutations at such low frequency requires high sensitivity of the computational tool. Therefore, a subset of the zebra fish data was used as the training data for developing modules that distinguish weak signals from sequencing artifact.
We then ran SNPdetector to discover all candidate genetic variations in the entire dataset. Those that had only one heterozygote across the entire population were considered to be putative mutations. These putative mutations were manually reviewed and subjected to repeated sequencing. SNPdetector identified all eight verified mutations. A total of 102 SNPs with minor allele frequencies ranging from 0.2% to 50% were also identified. To find all mutations using PolyPhred 5 requires setting a score threshold of six (the highest PolyPhred score is 99). At such a low threshold, the majority of the variations identified are expected to be false positives.
Sequence Coverage of the Mouse, Human, and Zebra Fish Datasets
We analyzed the sequence coverage to estimate the overall false negative rate resulting from rejection of low-quality bases by SNPdetector. Sequence coverage refers to the percentage of the total bases that are accepted for SNP identification by the program. In this analysis, each aligned base was subjected to the acceptability test employed by SNPdetector, which evaluates its short-range and long-range quality score distribution as well as secondary peak profile (details in Materials and Methods).
We calculated the read-based coverage and the sample-based coverage; the latter combines the forward and the reverse reads from the same sample (details in Materials and Methods). The results are summarized in Table 5. In the read-based coverage analysis, 89%–91% of the total bases were accepted; 3%–4% of the total bases were rejected because of stutter, showing that stutter accounts for 30%–40% of all rejections. Q20 bases (e.g., bases with Phred quality score ≥ 20) constituted 89%–90% of the total bases, while their percentage in the accepted bases was higher (in the range 95%–98%), indicating that a good proportion of rejected bases are of low quality. However, quality score alone does not determine the status of a base. In the read in Figure 2, four bases with very low quality scores (in the range of nine to 11) were accepted because they had no secondary peak background (Figure 2A) while one Q20 base was rejected because of its background noise (Figure 2B).
Table 5 Sequence Coverage Analysis of the Three Datasets
Figure 2 Rejected and Accepted Bases in a Sequence Trace
The Phred quality scores are indicated at the top. The quality scores for rejected bases are labeled in red. Accepted bases are marked by rectangular boxes.
(A) A subregion of polyA bubble showing that low-quality bases with no secondary peaks are accepted by SNPdetector.
(B) A subregion showing that a Q20 base is rejected because of its high secondary peak even though the majority of neighboring bases have high-quality scores.
When we combined forward and reverse reads from the same sample to calculate coverage, 94%–95% of the total bases in the three datasets were accepted. The sample-based coverage gives a more accurate estimate of false negative rate resulting from lack of coverage than the read-based coverage because (1) in all three datasets each sample was sequenced in both orientations; (2) SNPdetector analyzes both the forward and reverse reads from the same sample to obtain the genotype; and (3) sequencing artifacts such as stutter have a complementary pattern in the forward and the reverse reads, e.g., stutters in one orientation usually have non-stutter bases in the opposite orientation.
Discussion
We have demonstrated the ability of SNPdetector to accurately call SNPs in resequencing reads from PCR templates with very low false negative rates (2%–6%) and acceptable false positive rates (1%–9%). In the test data analyzed here, the error rate of SNPdetector is much lower than that of the two alternative methods: PolyPhred (version 5.0.2) and NovoSNP (see Table 4). InSNP [18] is another recently developed SNP analysis software package. We did not reanalyze the test data using this tool because its main function is to support interactive human inspection rather than perform automated data analysis. The false positive rate of InSNP was reported to be in the range of 93% to 95% [18].
SNPdetector is able to find SNPs or mutations of very low frequency because it does not rely on multiple instances of a minor allele to evaluate SNP validity. In our experience of manual SNP review, we have observed that sequencing artifacts such as stuttering, bubbling, and spilling usually occur in multiple samples at the same locus (Figure 3). In the case of stuttering, the sequence artifact can be attributed to sequence repeat content (i.e., polynucleotides or STRs) or indel polymorphism. Thus, noise that reduces the accuracy of SNP detection can be systematic and highly reproducible. Multiple observations of a genotype are considered confirmatory only if they were derived from sequence reads of the same sample in opposite orientations, because complementary bases are assayed in the forward and reverse sequence reactions.
Figure 3 A PolyA Bubble That Occurs in Multiple Samples
The bubble was recognized as a sequencing artifact by SNPdetector, and no SNP was called even though the alternative adenine residue (in the highlighted column) appeared in two samples with an average Phred quality score of 20. In addition, all three traces in this region have a polyG spill at the right, with a secondary guanine peak spanning four residues; and a polyT spill at the left, with a secondary thymine peak spanning three residues.
The sensitivity of SNPdetector enabled us to discover unexpected heterozygosity in the inbred strain CAST/Ei. Of the 1,167 mouse SNPs, two located at the 3′ UTR of Bach1 were heterozygous in CAST/Ei strain while the remaining 24 strains were homozygous. This discovery was confirmed by repeated sequencing of additional animals of CAST/Ei strain using a different pair of sequencing primers. All but one of the genotypes were heterozygous (see Table 2), suggesting that maintaining heterozygosity at this locus might be critical to CAST/Ei. Though heterozygosity of noncoding DNA was previously shown in recombinant inbred strains [19], this is the first case to our knowledge in which heterozygosity is observed in the mRNA transcript of a well-established inbred strain. Maintenance of heterozygosity is expected to be accompanied by reduced fecundity, and CAST/Ei is known to have smaller litter size than other inbred strains [20].
The three studies presented here include regions of very high SNP density. For example, among the regions with the highest SNP density, one 622-bp zebra fish amplicon contains 11 SNPs and one 854-bp mouse amplicon contains 26 SNPs. The genetic divergence in these regions can lead to the generation of multiple contigs if we attempt to assemble all the sequence reads, and the errors in the alignments of an assembly can become a major source of SNP detection error. In the mouse study, one of the strains is SPRET/Ei. It belongs to M. spretus, not M. musculus; the other 24 strains belong to the latter species. In the data analyzed here, the variation rate between M. spretus and M. musculus is approximately one every 50 bp, indicating SNPdetector can be useful for identifying inter-species variations of highly related organisms.
During the development of SNPdetector, we used the mouse resequencing data as the training dataset because heterozygotes in inbred mouse strains are almost always false positive as a result of mouse breeding history. This allowed us to investigate potential sources of false positive heterozygous allele calls and develop filters for these sequencing artifacts. The initial design of SNPdetector had included an option to allow a user to set the threshold on the quality measurement of individual genotype. However, with effective filtering, a low-quality threshold increased the sensitivity but not the false positive rate of SNP detection, making user intervention unnecessary.
In the human resequencing data, the assessment of SNP validity was based on experimental genotyping in those cases where the data were available. We resorted to visual inspection of trace data only when there were no genotyping data (see Table 4). However, we noticed that in some cases genotype data were inconsistent with the result of visual analysis. For example, in the test data presented in Table 4, four visually apparent SNPs were scored as monomorphic in genotyping. These were rare SNPs with one heterozygote in the resequencing population; each SNP had both forward and reverse sequence coverage (Figure S2). On the other hand, two visually rejected SNPs were scored as polymorphic in genotyping; neither was found by SNPdetector or PolyPhred. Taking into account the results of visual analysis, the false positive and false negative rates of SNPdetector would be 4.55% and 3.45%, respectively, much lower values than those in Table 4. Clearly, the error rates of the genotyping assays, though not yet available, must be taken into account when used as a standard for assessing the accuracy of SNP discovery.
The current version of SNPdetector is able to find indel polymorphisms when there are homozygous minor alleles. The stutters caused by heterozygous indels are detected but not decoded, partly because of the difficulty in distinguishing the stutters caused by indel polymorphisms from those caused by polynucleotide runs or STRs. Additionally, the “.poly” files generated by Phred for peak analysis only include primary and secondary peak information. However, the secondary peaks may not always represent one of the two reads downstream of a heterozygous indel if there are sequencing artifacts in the region. We are currently evaluating the possibility of revising Phred to export quality scores of the secondary peaks to facilitate the decoding of heterozygous indels.
The current version of the program does not implement sequence assembly because, with the successful completion of the Human Genome Project and genome projects in other species, high-quality reference genomic sequences are readily available for human and other model organisms. An assembly module could be easily incorporated if SNPdetector were used to analyze an organism lacking a high-quality reference sequence.
At this level of accuracy, the success rate of the resequencing reactions is now the limiting factor in screens for identifying novel SNPs and mutations. SNPdetector runs on Unix and Linux and is publicly available by anonymous ftp (http://lpg.nci.nih.gov).
Materials and Methods
Human ENCODE resequencing.
PCR amplicons were designed to tile five human genome ENCODE [16] regions (ENm110 on 7p15.2, ENr321 on 8q24.11, ENr232 on 9q34.11, ENr123 12 q12, and ENr213 on 18 q12.1), each 500 kb in length. In total, 4,190 PCR reactions were carried out on each individual to amplify the 2.5 Mb of genomic sequence.
PCR reactions were run in 6-fold multiplex reactions arranged so that consecutive amplicons were never in the same reaction. To the 5′ end of each primer pair were attached specific sequencing primers so the reaction mixture could be sequenced directly upon completion of the amplification. DNA sequences were tested for fidelity to the intended amplicon sequence by comparison to the human genome using BLAST. Sequencing reads that failed to make their best match to the genome between the sequencing primers were rejected from SNP analysis. In all, 258,909 sequences met this quality criterion and went on to SNPdetector. SNPdetector aligned the DNA sequences to the reference (NCBI build 34 of the human genome), and then the program called SNPs. From this analysis, 11,241 candidate SNPs were identified, including 1,571 homozygous indels. Approximately one-third of the candidate SNPs were not eligible for genotyping because they failed to meet criteria for assay design. These criteria exclude SNPs lying in palindromes, AT- or GC-rich regions, or repeated DNA (see http://www.hapmap.org/downloads/assay-design_protocols.html).
Programming language and system requirements.
SNPdetector was implemented in C and Perl. It currently runs on Unix and Linux platforms.
Identification of low-quality or misassembled reads.
Each subject sequence is aligned to the reference sequence; the two ends of each read are trimmed until there is a 20-bp window with 95% or greater identity to the reference sequence at each end. Sequences that lack such a window are not included in SNP detection.
The trimmed alignments are then used to evaluate the read quality. A read is considered poor quality if it does not have a subregion (at least 30 bp) with high sequence quality (Phred quality score ≥ 30 in 90% of the bases) as well as high sequence identity (95%) to the reference sequence. We did not require high average quality across the entire sequence because such measures tend to exclude stuttered reads (caused by STR, polynucleotide, or indel polymorphism). The average quality score for a stuttered read is usually low but the bases upstream of the stutter can be of sufficient quality for SNP analysis. If a sequence has a high quality (defined previously), highly divergent (<70% identity to the reference sequence) subregion, then the read is considered a “misassembled” sequence not suitable for SNP detection.
The parameters were derived from empirical analysis of genomic regions with high SNP density. However, they are adjustable in the pipeline, and the user can opt out of the low-quality/misassembly check.
Modification of NQS for detecting SNP and indel polymorphisms.
Prior to heterozygote detection, we implemented a modified version of NQS to identify putative SNP and indel polymorphisms that have homozygous minor alleles in the resequencing sample. First, the many-to-one alignments computed by SIM [12] were converted to an M × N multiple alignment by projecting insertions in subject sequences as deletions in the reference sequence. M corresponds to the alignment length; N corresponds to the number of subject sequences. A position in M is considered a putative variation site if it has more than two qualified alleles across N samples.
A residue is considered a qualified substitution allele if it and each base in its 4-bp flanking regions has Phred quality score ≥ 15. This minimum quality score is an adjustable parameter in the program. If a residue fails this standard NQS check but resides in a 9-bp window where the quality score of each base exceeds 25, then it is qualified by this expanded window NQS. The latter criterion ensures that a high-quality residue is included even if there is sequencing artifact on one side of the 4-bp flanking region.
A higher-quality threshold is applied to qualify a putative indel because errors in base calling tend to generate false indel polymorphisms. The minimum quality score of NQS is 25 instead of 15. If a putative indel resides in a polynucleotide repeat, then the entire repeat as well as the 4-bp flanking region of the repeat is required to pass the NQS quality check because in a repeat polymorphism the gap location is arbitrary. The expanded window NQS is not used for indel allele qualification because alignment artifacts resulting from base calling errors often produce false indel alleles.
If a sample has forward and reverse reads and its sequence in the opposite direction is identical in the 10-bp regions flanking either side of the putative allele, the quality scores from the two reads are combined for the NQS check. When there is a discrepancy in the forward and reverse reads of a putative indel allele, the indel allele is disqualified.
Putative substitutions and indels identified by NQS are subjected to further evaluation in the process described below. Those that fail in subsequent test are not listed in the output.
Ratio of primary to secondary peak in heterozygote detection.
The zebra fish resequencing data had more than 1,200 subjects sequenced in both forward and reverse orientations. The large sample size in this study allowed us to inspect the distribution of secondary-to-primary-peak ratio as noise in “dirty” homozygotes (e.g., homozygotes with a secondary peak background) and as signal in true heterozygotes. We found that it was not uncommon for a homozygote to have a secondary peak approximately 20% of the primary peak height. On the other hand, it was uncommon for a true heterozygote to have a secondary peak less than 30% of the primary peak. Therefore, in the default setting we used the 30% threshold as the lower bound for detecting putative heterozygotes; a secondary peak below 20% of the primary peak was considered background noise in a dirty homozygote in the initial genotype assessment.
Genotype quality classification.
The genotype of a sequence read is classified into one of the following six categories: high, med, low1, low2, low3, and reject. The quality class is determined by analyzing the Phred quality score of the variant site and its 4-bp flanking side. The threshold of each quality class is based on previous empirical analysis of NQS accuracy [3,4] (e.g., the minimum quality score for each base in the flanking regions is 15) as well as comparison with the validated SNPs in the training datasets. The initial assignment of a genotype quality class may be modified by the subsequent processing described below under “Horizontal and Vertical Scan.”
In Phred, one of the four parameters for discriminating errors from correct base calls is “uncalled/called ratio,” e.g., the ratio of the height of the largest uncalled peak to the smallest called peak within a 7-bp window around the current site [11]. In many cases, the uncalled secondary peak at a heterozygote site is the largest uncalled peak in this 7-bp window. As a result, the Phred quality scores of a heterozygote and its flanking bases can be much lower than those of a homozygote with a similar peak profile. Such an example is shown in Figure 4. The most dramatic Phred quality score drop is found at the heterozygote site and its immediate 1-bp neighbors. We define these three bases as a “heterozygote Phred quality score drop unit” (HQDU). The flanking region is analyzed under two conditions: (1) using the 4-bp flanking region around the current site without taking into account the HQDUs at the site and within its flanking region, and (2) using only those bases that do not belong to HQDUs at the site and within its flanking region. If there are fewer than four such bases within 20 bp of the site, then the flanking region is considered invalid (score set to zero). The 20-bp constraints ensure that a region that consists entirely of HQDUs (as in the case of stutter) is ignored. The maximum of conditions 1 and 2 is used to represent the flanking region score of the site because condition 1 is more accurate if the maximum uncalled peak (defined by Phred) in the 7-bp window is not the secondary peak of a heterozygote but a sequencing artifact (such as bubble). On the other hand, a score derived from condition 2 is more accurate if a heterozygote is the maximum uncalled peak in the region. An example of skipping HQDUs in the flanking region analysis is shown in Figure 4.
Figure 4 Sequence Traces of a SNP Cluster with Three Consecutive SNPs
The top is a homozygous sample and the bottom a heterozygous one. The Phred quality score is labeled on top of each base. In the heterozygous sample, the three HQDPs around the three heterozygotes are labeled with red lines at the bottom. The flanking bases used for calculating genotype quality class of the highlighted heterozygote in the middle are marked by rectangular boxes, which do not include any HQDPs. The flanking bases used to assess background noise in the flanking region are labeled with brackets at the bottom.
Once the flanking region is defined, the genotype score is calculated as follows. If each base at the flanking region exceeds a Phred quality score of 15, 25, or 40, then the flanking region is assigned a score of one, two, or three, respectively. If the variation site exceeds a Phred quality score of 15, 25, or 40, then it is assigned a score of one, two, or three, respectively. If the average quality score of the flanking region exceeds 25 or 40, then the score is incremented by one or two, respectively. If the average is below 15 (low), then a penalty of −1 is imposed. The combined total score is then used to derive the initial classification group as follows: score < 0 → reject; score = [0,1] → low3; score = 2 → low2; score = [3,4] → low1; score = [5,6] → med; and score ≥ 7 → high. Thus, a genotype of class “reject” has Phred quality score below 15 at the site as well as the flanking region; such a site will not be used for SNP detection.
The initial genotype quality class can be turned into the class reject in horizontal or vertical scan analysis. In these scans, a putative heterozygote may also be reclassified as a dirty homozygote or vice versa. When a genotype changes its status, the quality class is also recalculated.
Genotype noise assessment.
To evaluate noise within the 4-bp flanking region of a putative heterozygote or homozygote, the program checks the secondary peak of each base in the flanking region. As above, define p = (secondary_peak_area/primary_peak_area) × 100 (i.e., percent of primary peak area occupied by secondary peak). If each base in the flanking region passes the test of p = 0, p ≤ 10, or p ≤ 20, then the flanking region is considered to have no, little, or limited noise, respectively. A site with a p > 70 secondary peak in the flanking region is skipped to avoid penalizing a putative heterozygote in a SNP cluster (see an example in Figure 4). The same test is applied to measure the noise level at the site of a homozygote. For a putative heterozygote, the higher the p-value, the stronger the signal. A classification of no, little, and limited noise is awarded to putative heterozygote sites with p ≥ 80, p ≥ 50, and p ≥ 30, respectively.
Horizontal and vertical scan.
Manual SNP inspection usually involves a horizontal scan of the same trace and a vertical scan across multiple traces at a putative SNP site. The horizontal scan assesses whether the signal is distinguishable from local noise, while the vertical scan determines whether the signal is distinguishable from the noise in the other samples.
To model the horizontal scan, SNPdetector first identifies short-range (5 bp) and long-range (>50 bp) features indicative of potential problems in a sequence read. These are (1) regions with low sequence similarity, (2) regions with low sequence quality, and (3) regions with a high secondary peak background. Long-range features, when occurring downstream of a STR (computed with the program Ptrfinder [21]) or a polynucleotide track (≥8 bp), or an indel polymorphism identified in the NQS analysis, usually indicate stuttering in PCR amplification [22], and stutters are disqualified (e.g., genotype quality class set to reject). Short-range features accompanied by specific types of flanking sequence indicate potential artifacts—e.g., spilling (i.e., a background of secondary polynucleotide track extended from a neighboring primary peak; see Figure 3), bubbling (i.e., reads embedded underneath a polynucleotide blob; Figure 3), or factitious indels resulting from base-calling errors. Details of the parameters used for horizontal scan are summarized in Table 6.
Table 6 Parameters Used in Horizontal Scan
To model the vertical scan of a human inspector, SNPdetector first identifies high-quality homozygotes with no secondary peak (e.g., secondary-to-primary-peak-area ratio is zero) in the 10-bp flanking region. These “clean” homozygotes are then used to find “dirty” homozygotes. A dirty homozygote is determined to be present if one of the following conditions is true: (1) there is a discrepancy between forward and reverse reads of the same sample, e.g., a clean homozygote is found in sequence read of one orientation while the read in the opposite orientation has a secondary peak; or (2) a sequence read has a low secondary peak (secondary-to-primary-peak-area ratio < 10) and no reduction of its primary peak compared to that of a clean homozygote derived from the same orientation. To adjust for baseline differences in peak area measurements in different traces, we used a relative score, i.e., the ratio of the primary peak area at the putative site to that of its immediate homozygous neighbors. The unclassified traces are processed in ascending order of their secondary-to-primary-peak-area ratio, and each is compared to reads of clean or dirty homozygotes to determine (1) whether its secondary peak area is comparable to the secondary-to-primary-peak-area ratio found in dirty homozygotes, and (2) whether the reduction of its primary peak is comparable to those observed in dirty-to-clean homozygotes. A sequence read deemed indistinguishable from a classified homozygote is considered to represent a dirty homozygote and is included in the analysis of the remaining data. A more stringent threshold is used for reads with short-range features indicative of sequence artifacts identified during the horizontal scan. The genotype quality class and noise class are recomputed for putative heterozygotes reclassified as dirty homozygotes.
The complete list of parameters used for the horizontal and the vertical scan is listed in Table S1.
Analysis of sequence coverage of the three datasets.
Sequence coverage refers to the percentage of the total bases that are accepted for SNP detection by the software. Empty trace files, reads with unacceptable quality (details under “Identification of Low-Quality or Misassembled Reads”), and reads that fail to make their best match to the reference sequences are considered assay failures. They are not used as an input for SNPdetector and, as a result, are not included in the coverage analysis.
The total number of resequenced bases in a read includes every base that is aligned to the genomic interval spanned by the forward and the reverse sequence primers. An aligned base is considered acceptable if it passes “horizontal scan,” described in the previous section. The numbers of Q20 bases in the total bases and the accepted bases are recorded. The number of bases rejected due to stuttering is also recorded.
In all three datasets, each sample was sequenced in both the forward and the reverse orientations, giving a 2× redundancy in sequence coverage for each sample.
In addition to the read-based coverage described above, we also analyzed the sample-based coverage, which combines the reads from both the forward and the reverse orientations to calculate the total bases and the accepted bases for each sample. At each position in the genomic interval spanned by the forward and the reverse primers, bases from all reads of the same sample are evaluated. An accepted base in one read always overwrites a rejected base in another. For bases with the same status, the higher quality is recorded for the Q20 analysis.
Run time parameter of Phrap and PolyPhred 5.0.2.
We ran Phred/Phrap/PolyPhred using the template genomic sequences and traces derived from resequencing. Each template sequence was converted into a reference trace using the program SudoPhred. Each base of the reference template sequence was assigned a Phred quality score of 59. Using the default parameters of Phrap, an amplicon with a high SNP rate can be split into multiple contigs. For example, in the mouse resequencing data, 95% of the amplicons were assembled into multiple contigs even after excluding the traces derived from the strain SPRET/Ei (which belongs to the species M. spretus). Using the parameters “–repeat_stringency 0.55 –forcelevel 2,” we were able to obtain a single contig in 95% of the amplicons. Therefore, we used this setting in all datasets.
We activated options in PolyPhred to use the template genomic sequence as the reference sequence and to combine forward and reverse reads from the same individual for genotype calls (the “–source” option).
In the results generated by PolyPhred 5.0.2, approximately 50% of the high-quality SNPs (score = 99) were monomorphic in their genotype calls (i.e., PolyPhred's own assessment was monomorphic even though it called a SNP). We visually analyzed 30 such cases, and were able to confirm that they were all monomorphic sites. We developed a filter to remove these monomorphic “SNPs” from the PolyPhred output.
Analysis of false positive and false negative rates.
We calculated false positive rate to measure the specificity of the three programs tested in this study using the following formula: the number of false positive SNPs divided by the number of SNPs discovered. This formula is referred to as the false discovery rate [23]. We calculated false negative rate to measure sensitivity using the following formula: the number of known missed SNPs divided by the number of all true SNPs.
Supporting Information
Figure S1 Trace Chromatogram of Four CAST/Ei Animals at Bach1 Locus (SNP1 in Table 2)
All but the second animal (animal B in Table 2) are heterozygous.
(88 KB PDF)
Click here for additional data file.
Figure S2 An Example of Discrepancy between Visual Analysis and Genotyping Result
The second and the third traces are the reverse and the forward reads from an individual identified as a heterozygote by visual analysis. The minor allele T was only found in this individual. The top is a sequence of a homozgygote control. Both SNPdetector and PolyPhred found this SNP (PolyPhred score = 99). However, the genotype result is monomorphic at this site.
(104 KB PDF)
Click here for additional data file.
Table S1 SNPdetector Parameters Used to Make Genotype and SNP Calls
(113 KB PDF)
Click here for additional data file.
We thank Dr. Milton English, Ms. Lin Lei, and Ms. Maria Anderson from National Human Genome Research Institute for their help in ENU mutant collection and Drs. Dongying Wu and Jun Yu from Beijing Genomics Institute for performing sequencing reactions. We thank Mr. Michael Gandolph from National Cancer Institute (NCI) for sequencing the additional mouse strains and Dr. Myung-Soo Lyu from NCI for supplying additional mouse animals. We thank Drs. Kent Hunter and Robert Williams for discussions about the heterozygous mouse genotypes. We are grateful to Ms. Jose Gortier, Mr. Steve Haneline, and Mr. Haneline's group of manual data analysts, who shared their manual review experiences with us. This research was supported by the Intramural Research Program of the Center for Cancer Research, NCI, National Institutes of Health.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. JZ conceived and designed the experiments, developed the software tool, and performed testing. WR implemented the software. JZ, DAW, IY, SW, RS, and PPL analyzed the data. JZ, PPL, RAG, and KHB contributed reagents/materials/analysis tools. JZ, DAW, and RS wrote the paper.
Abbreviations
HQDUheterozygote Phred quality score drop unit
indelinsertion/deletion
NQSneighborhood quality standard
SNPsingle nucleotide polymorphism
STRsimple tandem repeat
==== Refs
References
Yeung AT Hattangadi D Blakesley L Nicolas E 2005 Enzymatic mutation detection technologies Biotechiques 38 749 758
Sachidanandam R Weissman D Schmidt SC Kakol JM Stein LD 2001 A map of human genome sequence variation containing 1.42 million single nucleotide polymorphisms Nature 409 928 933 11237013
Altshuler D Pollara VJ Cowles CR Van Etten WJ Baldwin J 2000 An SNP map of the human genome generated by reduced representation shotgun sequencing Nature 407 513 516 11029002
Mullikin JC Hunt SE Cole CG Mortimore BJ Rice CM 2000 An SNP map of human chromosome 22 Nature 407 516 520 11029003
Marth GT Korf I Yandell MD Yeh RT Gu Z 1999 A general approach to single-nucleotide polymorphism discovery Nat Genet 23 452 456 10581034
Crawford DC Carlson CS Rieder MJ Carrington DP Yi Q 2004 Haplotype diversity across 100 candidate genes for inflammation, lipid metabolism, and blood pressure regulation in two populations Am J Hum Genet 74 610 622 15015130
The International HapMap Consortium 2003 The International HapMap Project Nature 426 789 796 14685227
Crawford DC Akey DT Nickerson DA 2005 The patterns of natural variation in human genes Annu Rev Genomics Hum Genet 6 287 312 16124863
Nickerson DA Tobe VO Taylor SL 1997 PolyPhred: Automating the detection and genotyping of single nucleotide substitutions using fluorescence-based resequencing Nucleic Acids Res 25 2745 2751 9207020
The Encode Consortium 2004 The ENCODE (ENCyclopedia Of DNA Elements) Project Science 306 636 640 15499007
Ewing B Green P 1998 Base-calling of automated sequencer traces using phred. II. Error probabilities Genome Res 8 186 194 9521922
Huang XQ Hardison RC Miller W 1990 A space-efficient algorithm for local similarities Comput Appl Biosci 6 373 381 2257499
Zhang J Hunter KW Gandolph M Rowe WL Finney RP 2005 A high-resolution multistrain haplotype analysis of laboratory mouse genome reveals three distinctive genetic variation patterns Genome Res 15 241 249 15687287
Wade CM Kulbokas EJ 3rd Kirby AW Zody MC Mullikin JC 2002 The mosaic structure of variation in the laboratory mouse genome Nature 420 574 578 12466852
Gordon D Abajian C Green P 1998 Consed: A graphical tool for sequence finishing Genome Res 8 195 202 9521923
Dausset J Cann H Cohen D Lathrop M Lalouel JM 1990 Centre d'Etude du Polymorphisme Humain (CEPH): Collaborative genetic mapping of the human genome Genomics 6 575 577 2184120
Weckx S Del-Favero J Rademakers R Claes L Cruts M 2005 NovoSNP, a novel computational tool for sequence variation discovery Genome Res 15 436 442 15741513
Manaster C Zheng W Teuber M Wachter S Doring F 2005 InSNP: A tool for automated detection and visualization of SNPs and InDels Hum Mutat 26 11 19 15931688
Williams RW Gu J Qi S Lu L 2001 The genetic structure of recombinant inbred mice: High-resolution consensus maps for complex trait analysis Genome Biol 2 RESEARCH0046 11737945
Le Roy I Roubertoux PL Jamot L Maarouf F Tordjman S 1998 Neuronal and behavioral differences between Mus musculus domesticus (C57BL/6JBy) and Mus musculus castaneus (CAST/Ei) Behav Brain Res 95 135 142 9754885
Collins JR Stephens RM Gold B Long B Dean M 2003 An exhaustive DNA micro-satellite map of the human genome using high performance computing Genomics 82 10 19 12809672
Shinde D Lai Y Sun F Arnheim N 2003 Taq DNA polymerase slippage mutation rates measured by PCR and quasi-likelihood analysis: (CA/GT)n and (A/T)n microsatellites Nucleic Acids Res 31 974 980 12560493
Storey JD Tibshirani R 2003 Statistical significance for genomewide studies Proc Natl Acad Sci U S A 100 9440 9445 12883005
|
16261194
|
PMC1274293
|
CC BY
|
2021-01-05 09:18:23
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e53
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010053
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 1626119510.1371/journal.pcbi.0010054plcb-01-05-08Research ArticleUltrasensitization: Switch-Like Regulation of Cellular Signaling by Transcriptional Induction UltrasensitizationLegewie Stefan 1*Blüthgen Nils 1Schäfer Reinhold 2Herzel Hanspeter 11 Institute for Theoretical Biology, Humboldt University, Berlin, Germany
2 Laboratory of Molecular Tumor Pathology, Charité, Berlin, Germany
Miyano Satoru EditorUniversity of Tokyo, Japan* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 26 9 2005 1 5 e549 5 2005 26 9 2005 Copyright: © 2005 Legewie et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Cellular signaling networks are subject to transcriptional and proteolytic regulation under both physiological and pathological conditions. For example, the expression of proteins subject to covalent modification by phosphorylation is known to be altered upon cellular differentiation or during carcinogenesis. However, it is unclear how moderate alterations in protein expression can bring about large changes in signal transmission as, for example, observed in the case of haploinsufficiency, where halving the expression of signaling proteins abrogates cellular function. By modeling a fundamental motif of signal transduction, the phosphorylation–dephosphorylation cycle, we show that minor alterations in the concentration of the protein subject to phosphorylation (or the phosphatase) can affect signal transmission in a highly ultrasensitive fashion. This “ultrasensitization” is strongly favored by substrate sequestration on the catalyzing enzymes, and can be observed with experimentally measured enzymatic rate constants. Furthermore, we show that coordinated transcription of multiple proteins (i.e., synexpression) within a protein kinase cascade results in even more pronounced all-or-none behavior with respect to signal transmission. Finally, we demonstrate that ultrasensitization can account for specificity and modularity in the regulation of cellular signal transduction. Ultrasensitization can result in all-or-none cell-fate decisions and in highly specific cellular regulation. Additionally, switch-like phenomena such as ultrasensitization are known to contribute to bistability, oscillations, noise reduction, and cellular heterogeneity.
Synopsis
Hormones and other external stimuli induce cellular transitions such as cell division or differentiation by regulating gene expression. Hormone-induced cellular transitions are known to occur in a switch-like fashion: while weak background stimuli are rejected, cellular transitions proceed fully as soon as a threshold hormone concentration is exceeded. Earlier studies have described several mechanisms whereby such a switch-like behavior can be realized in intracellular communication via signal transduction networks, which convert hormonal signals into alterations in gene expression.
The authors demonstrate how switch-like behavior can be further enhanced downstream of hormone-induced gene expression. They show that even minor (hormone-induced) alterations in gene expression can dramatically affect the activity of intracellular signal transduction networks, and thereby modify cellular behavior. This phenomenon has been termed “ultrasensitization.”
Ultrasensitization can explain the pronounced dosage sensitivity observed for many disease-associated signal transduction proteins: for example, the mutation of one of two alleles (gene copies), resulting in a 2-fold reduction of gene expression, can already initiate disease progression. Although such sensitivity towards mutations is potentially harmful, the fact that cells nevertheless exhibit ultrasensitization suggests that somehow cells benefit from ultrasensitization. The authors illustrate how ultrasensitization improves the specificity and efficiency of cell-to-cell communication and contributes to cellular memory.
Citation:Legewie S, Blüthgen N, Schäfer R, Herzel H (2005) Ultrasensitization: Switch-like regulation of cellular signaling by transcriptional induction. PLoS Comput Biol 1(5): e54.
==== Body
Introduction
Cellular signal transduction exhibits two layers of regulation: upstream stimuli such as extracellular peptide hormones activate intracellular signaling intermediates (e.g., mitogen-activated protein kinases), which in turn induce intracellular responses (e.g., activated transcription factors) as indicated in Figure 1A. This type of regulation, which usually operates on the time scale of minutes, will be termed “fast regulation” in this paper. Fast regulation involves the posttranslational modification of pre-existing protein pools in order to transduce signals to the nucleus. Responses induced by fast regulation (Response 1), such as activated transcription factors, often in turn alter the total abundance of their own activators (Intermediate 1) or that of intermediates in heterologous cascades (Intermediate 2), e.g., owing to induced mRNA/protein synthesis or to degradation (Figure 1A). For example, signal transduction pathways in the immune system alter the concentration of their own constituents by transcriptional positive or negative feedback to bring about sensitization [1] or desensitization [2]. Likewise, cyclic guanosine monophosphate signaling affects the heterologous mitogenic cascades via “transcriptional crosstalk” by inducing the MKP-1 phosphatase and cyclin-dependent kinase inhibitors [3]. Owing to the long half-lives of most mRNAs [4] and proteins [5], this type of regulation operates on the time scale of hours in most (but not all) cases and thus will be referred to as “slow regulation” here.
Figure 1 Slow and Fast Regulation of Cellular Signal Transduction
(A) Schematic representation of cellular signal transduction. Upstream stimuli (e.g., hormones) result in altered gene expression by eliciting rapid intracellular responses such as transcription factors (fast regulation). The resulting changes in protein expression often in turn affect cellular signal processing of upstream inputs via transcriptional feedback or crosstalk (slow regulation).
(B) Schematic representation of a phosphorylation–dephosphorylation cycle, where the kinase K and the phosphatase P catalyze the (de)phosphorylation of the substrate, S. Hormonal stimulation (i.e., fast regulation) was modeled by altering the total kinase concentration, Ktot = K + S0K, and the steady-state concentration of free phosphorylated substrate, S1, was taken as the response. The impact of slow regulation was modeled by varying the concentration of the substrate (Stot = S0 + S0K + S1 + S1P) or that of the phosphatase (Ptot = P + S1P).
Figure 1A shows a simplified view of signal transduction: upstream stimuli such as hormones (i.e., fast regulation) induce gene expression (i.e., slow regulation), which in turn alters cellular signal processing. Previous theoretical and experimental studies (e.g., [6,7]) have mainly analyzed how fast regulation influences slow regulation, while downstream effects, i.e., the impact of slow regulation on fast regulation, are less well investigated. However, slow regulation is probably equally important, since the expression of phosphoproteins is altered during a variety of physiological processes such as differentiation [8], development [9], apoptosis [10], long-term potentiation [11], the cell cycle [12], and the circadian rhythm [13]. Furthermore, the deregulated expression of wild-type phosphoproteins has been shown to be correlated with diseases such as diabetes [8] and cancer [14].
At a first glance, one may expect that altered expression of Intermediate 1 (i.e., slow regulation) affects steady-state signal transmission via Intermediate 1 (i.e., fast regulation) in a linear fashion (Figure 1A). However, recent research suggests that strong nonlinearity is observed at least in some cases: a variety of tumor-suppressor genes involved in cellular signal transduction do not follow Knudson's two-hit hypothesis, i.e., they do not require a homozygotic loss of both alleles to support tumor progression. Instead, loss of a single copy, i.e., halving protein expression, is sufficient to abrogate tumor-suppressor function and this phenomenon has been termed haploinsufficiency [15,16]. This suggests that the expression of signaling proteins (i.e., slow regulation) affects signal transmission (i.e., fast regulation) in a highly switch-like fashion. As increased signal transmission elicited by transcriptional induction has been referred to as sensitization [1], such ultrasensitive regulation will be referred to here as “ultrasensitization.”
Available experimental data suggest that ultrasensitization is physiologically advantageous: receptor-tyrosine kinases, which elicit different cellular responses, are known to induce broadly overlapping sets of immediate early genes, although the amplitude of the immediate early gene-induction is receptor-specific [17,18]. Ultrasensitization allows cells to discriminate such minor differences in immediate early gene expression (i.e., in slow regulation), and thus confers specificity to receptor-tyrosine kinase signaling. In addition, even saturating hormone concentrations induce/repress the vast majority of target mRNAs less than 10-fold (e.g., [19]). Thus, the stimulus-response of hormone-induced transcription exhibits significant basal activation, so that mRNA induction (i.e., slow regulation) is relatively insensitive towards extracellular hormone concentrations [20]. In such cases, ultrasensitization can dramatically increase the cellular effects of extracellular hormone administration and may help to establish all-or-none cell-fate decisions. Finally, strong nonlinearities such as ultrasensitization are known to contribute to bistability, oscillations, noise reduction, and cellular heterogeneity (see Discussion).
Phosphorylation is the most common mode of eukaryotic information transfer, and it has been estimated that one third of all cellular proteins are phosphorylated [21]. As several haploinsufficient tumor suppressors encode phosphoproteins (e.g., p53, BRCA1, H2AX, and pRb), or phosphatases (PTEN), we were interested in whether ultrasensitization can occur in a simple phosphorylation–dephosphorylation cycle (Figure 1B). We show here that pronounced ultrasensitization is possible with experimentally measured kinetic constants, particularly in the parameter range where the substrate–enzyme ratio is such that the majority of the substrate subject to (de)phosphorylation is sequestered by the catalyzing enzymes. Furthermore, we show that coordinated transcription of multiple proteins (i.e., synexpression) within a kinase cascade can result in even more pronounced ultrasensitization. Finally, we demonstrate that ultrasensitization can account for specificity and modularity in the regulation of cellular signal transduction.
Results
Ultrasensitization in a Phosphorylation–Dephosphorylation Cycle
As outlined in the Introduction, we were interested in the means by which altered protein expression (i.e., slow regulation) affects signal transmission (i.e., fast regulation) by a simple phosphorylation–dephosphorylation cycle depicted in Figure 1B. Protein (de)phosphorylation was modeled using the irreversible Michaelis-Menten mechanism, where each elementary step is described by mass-action kinetics. As indicated in Figure 1B, the free unphosphorylated substrate, S0, reversibly associates with the free kinase, K, and the resulting kinase–substrate complex, S0K, may irreversibly form the free phosphorylated substrate, S1, thereby recycling the free kinase, K. Likewise, the free phosphorylated substrate, S1, is subject to dephosphorylation by the phosphatase and this occurs via formation of the substrate–phosphatase complex, S1P (see Protocol S1 for the differential equations). Importantly, our modeling approach takes substrate sequestration on the enzyme–substrate complexes, S0K and S1P, into account, which is in contrast to the well-known Michaelis-Menten approximation for enzyme catalysis, where the substrate concentration is assumed to significantly exceed that of the enzyme. The impact of upstream stimuli such as extracellular hormones (i.e., fast regulation) was modeled by altering the total kinase concentration, Ktot = K + S0K, and the steady-state concentration of the free phosphorylated substrate, S1, was taken as the response of the system. As we assume that the phosphatase competes with downstream effectors for a single docking site in the substrate (e.g., [22]), the substrate–phosphatase complex, S1P, does not contribute to the response.
To understand how protein expression (i.e., slow regulation) affects signal transmission (i.e., fast regulation), the steady-state stimulus response of the model depicted in Figure 1B was plotted for varying substrate expression, Stot = S0 + S0K + S1 + S1P. Then, the steady-state activation levels (S1) for a given stimulus (Ktot) before and after induction of the substrate (Stot) were compared. As indicated by the right arrow in Figure 2, a 5-fold induction of the substrate expression can bring about a 94-fold increase in the steady-state activation level upon strong stimulation (Ktot >> Ptot). Ultrasensitization is even more pronounced upon intermediate stimulation, since then a 5-fold altered substrate expression can bring about a 325-fold increase in the steady-state activation level. By contrast, nonlinearity disappears for very weak stimulation, i.e., for Ktot → 0, where the absolute activation levels are negligible (Figure 2).
Figure 2 Ultrasensitization in a Phosphorylation–Dephosphorylation Cycle
Stimulus-response of the phosphorylation–dephosphorylation cycle depicted in Figure 1B for varying substrate expression levels, Stot = S0 + S0K + S1 + S1P, on a double-logarithmic scale. The relative alterations in the response, S1, for a given stimulus, Ktot = K + S0K, elicited by a 5-fold change in substrate expression are indicated next to the vertical arrows. Parameters chosen: kon,K = koff,P = kcat,P = 0.01; koff,K = kcat,K = 1; kon,P = 1.6; Ptot = 1.25.
Thus, we proved that protein expression (i.e., slow regulation) can affect signal transmission (i.e., fast regulation) in a highly nonlinear fashion provided that stimulation is sufficiently strong. In other words, ultrasensitization is possible in a phosphorylation–dephosphorylation cycle, and transfection experiments in living cells support this conclusion [23].
Ultrasensitization Due to Substrate Sequestration
To gain further insight into ultrasensitization in a phosphorylation–dephosphorylation cycle, we analyzed steady-state signal transmission upon strong stimulation (Ktot >> Ptot), since an analytical expression for the response, S1, could be obtained in this limit (Protocol S1):
Here, the Michaelis-Menten constant of the phosphatase, KM,P, and the threshold substrate concentration, Stot,T, where ultrasensitization occurs (see below), are given by:
The rate constants kon, koff, and kcat describe the individual steps of kinase (K) or phosphatase (P) catalysis as indicated in Figure 1B, and Ptot = P + S1P is the total phosphatase expression level.
As shown in Figure 3, a large relative change in signal transmission according to Equation 1 from S1 << Stot to S1 ≈ Stot can be observed upon minor increases in substrate expression. As this dramatic relative change occurs in addition to the absolute increase in Stot, substrate expression, Stot (slow regulation), affects signal transmission via S1 (fast regulation) in a highly ultrasensitive fashion (ultrasensitization), particularly if:
If this condition holds, signal transmission is negligible (i.e., S1 << Stot) for weak substrate-expression levels, Stot < Stot,T (Figure 3), since then virtually all substrate is sequestered on the enzyme–substrate complexes, S0K and/or S1P (Protocol S1). By contrast, signal transmission begins to rise in an ultrasensitive fashion (ultrasensitization) as soon as the substrate expression exceeds the threshold, i.e., as soon as Stot > Stot,T. This can be explained as follows: the threshold, Stot,T, equals the maximal amount of substrate that can be sequestered on the enzyme–substrate complexes, S0K and/or S1P (Protocol S1), so that substrate sequestration no longer prevents signal transmission if Stot > Stot,T. Ultrasensitization finally vanishes for very strong substrate expression (Stot >> Stot,T), since then the amount of sequestered substrate is negligible, so that substrate expression affects signal transmission in a linear fashion (i.e., S1 ≈ Stot).
Figure 3 Ultrasensitization Due to Substrate Sequestration
The normalized maximal response of the phosphorylation–dephosphorylation cycle depicted in Figure 1B is plotted as a function of substrate expression on a semilogarithmic scale for the limit of strong stimulation (according to Equation 1), where Ktot >> Ptot. The threshold, Stot,T, (see Equation 2) was varied as indicated, while the Michaelis-Menten constant of the phosphatase, KM,P, was kept constant and assumed to be unity. The scheme on the top indicates the mechanism of ultrasensitization: for weak substrate-expression, most of the substrate is sequestered on the enzyme–substrate complexes, S0K and S1P, while signal transmission via S1 occurs as soon as substrate expression, Stot, exceeds the threshold, Stot,T.
Thus, we showed that pronounced ultrasensitization due to substrate sequestration on the catalyzing enzymes requires Equation 3 to be fulfilled (Figure 3). As outlined in the Discussion, Equation 3 is in accordance with experimentally measured data, so that ultrasensitization due to sequestration is expected to be observed in vivo. It should be noted that numerical studies revealed that ultrasensitization due to substrate sequestration under the regime of Equation 3 does not require very strong stimulation (Ktot >> Ptot), which was assumed to derive analytical expressions, but rather can be observed provided that (Protocol S1):
Equation 4 can be considered to be the general requirement for any ultrasensitization to occur in a phosphorylation–dephosphorylation cycle, since otherwise kinase activity (i.e., the stimulus) is too weak to elicit significant accumulation of the active species, S1. As outlined above, ultrasensitization refers to a large relative increase in signal transmission as the substrate expression level is increased, and thus requires that the majority of substrate is active (i.e., that S1 ≈ Stot) for sufficiently large Stot. For such high substrate expression levels, substrate sequestration on the catalyzing enzymes is negligible, and the overall velocities of phosphorylation (S0 → S1) and dephosphorlyation (S1 → S0) can be approximated by Vmax,K and Vmax,P. According to Goldbeter and Koshland [24], strong signal transmission (i.e., S1 ≈ Stot) for large Stot, and thus ultrasensitization, can be observed only if the kinase velocity exceeds that of the phosphatase, i.e., if Equation 4 holds (Protocol S1). This result confirms our earlier observation that ultrasensitization vanishes upon weak stimulation (see Figure 2).
Ultrasensitization Due to Activity Switching
Even though ultrasensitization due to substrate sequestration is preserved for intermediate stimulus levels (see above), this does not explain why ultrasensitization can be more pronounced for intermediate stimuli when compared to strong stimulation (Figure 2). For intermediate stimulus levels, an additional mechanism, which is independent of substrate sequestration, can bring about enhanced sensitization. Provided that Equation 4 holds, increasing substrate expression induces an “activity switch” from high overall phosphatase activity (S1 → S0) to high overall kinase activity (S0 → S1) if kinase catalysis is significantly less saturated than phosphatase catalysis, i.e., if:
As already mentioned, ultrasensitization refers to a large relative increase in signal transmission as the substrate expression level is increased. While Equation 4 is required for strong signal transmission (S1 ≈ Stot) for large Stot (see above), Equation 5 ensures that signal transmission vanishes (S1 << Stot) if Stot is small: for low substrate expression levels, where enzyme saturation is negligible, the overall phosphorylation and dephosphorylation velocities at steady state can be approximated by linear kinetics with the first-order rate constants Vmax,K/KM,K and Vmax,P/KM,P (Protocol S1). If Equations 4 and 5 hold, the phosphatase activity outnumbers that of the kinase (i.e., Vmax,P/KM,P >> Vmax,K/KM,K) provided that the stimulus, Ktot, is not too strong, so that signal transmission does not occur (S1 << Stot). Hence, a pronounced relative change from weak to strong signal transmission (ultrasensitization) can be observed as the substrate expression level is increased if both Equations 4 and 5 hold. Importantly, this ultrasensitization due to activity switching is independent of substrate sequestration.
Corresponding numerical results are shown in Figure 4 for a phosphorylation–dephosphorylation cycle with weak substrate sequestration, where intermediary stimulus Vmax,K results in pronounced ultrasensitization due to activity switching. Ultrasensitization disappears for stronger stimulus levels, since then the overall kinase activity outnumbers that of the phosphatase regardless of the substrate expression level (see above). In other words, ultrasensitization due to activity switching is restricted to intermediate stimulus levels, which explains why ultrasensitization in Figure 2 is optimal upon intermediate stimulation.
Figure 4 Ultrasensitization Due to Activity Switching
The normalized response of the phosphorylation–dephosphorylation cycle depicted in Figure 1B is plotted as a function of substrate expression on a semilogarithmic scale for varying stimulus levels. To relate the plot to analytical results given in the main text (Equation 4), the stimulus, Ktot, is expressed as Vmax,K and given in times of Vmax,P. For the parameters chosen (kon,K = 0.02; koff,K = kcat,K = koff,P = 1; kon,P = 2; kcat,P = Ptot = 0.1), substrate sequestration is insignificant (i.e., Equation 3 does not hold) and the kinase is significantly less saturated than the phosphatase (Equation 5). The scheme on the top indicates the mechanism of ultrasensitization: increasing substrate expression induces a switch from high overall phosphatase activity (S1 → S0) to high overall kinase activity (S0 → S1).
Ultrasensitization Due to Synexpression within a Kinase Cascade
Often multiple intermediates in a signaling cascade of phosphorylation–dephosphorylation cycles (Figure 5A) are transcribed coordinately in a so-called synexpression group. For example, the insulin receptor and its downstream substrate, IRS-1, are both upregulated during adipocyte differentiation [8,25], and this dramatically enhances insulin sensitivity in adipocytes when compared to fibroblastoid precursors. In addition, multiple components of the yeast pheromone-sensing pathway are coordinately induced in a transcriptional positive feedback loop following pheromone stimulation [19,26].
Figure 5 Ultrasensitization Due to Synexpression within a Signaling Cascade
(A) Schematic representation of a signaling cascade subject to synexpression. An increase in the regulator, r, was assumed to result in a proportional increase in the expression of intermediates S and T, both of which are subject to covalent modification by (de)phosphorylation.
(B) Ultrasensitization due to synexpression within a signaling cascade measured numerically by plotting the normalized response as a function of the regulator concentration, r, where Stot = Ttot = r (solid line). To show that synexpression enhances ultrasensitization, the case where the regulator, r, affects transcription of T only is also shown for Stot = 10 (dashed line). Similar results were obtained for other values of Stot or other stimulus strengths (data not shown). Parameters chosen: koff,1 = koff,5 = kcat,2 = kcat,6 = koff,3 = koff,7 = kcat,8 = PS,tot = PT,tot = 1; kon,1 = 0.02; kon,5 = 0.2; kon,3 = 2.1; kon,7 = 2; kcat,4 = 1.1; Ktot = 10.
As the impact of such synexpression groups on signal transmission has not yet been investigated, we analyzed a minimal model, where a transcriptional regulator, r, simultaneously induces the expression of the protein S and its downstream substrate, T, both of which are subject to covalent modification by phosphorylation (Figure 5A). For simplicity, we assumed that an increase in the regulator, r, results in a proportional increase in the expression levels of both cascade intermediates, i.e., in Stot and Ttot. Numerical simulations of the model depicted in Figure 5A demonstrate that synexpression of S and T (solid line in Figure 5B) significantly enhances ultrasensitization when compared to expression of T alone (dashed line in Figure 5B). For example, a 3-fold increase in the regulator, r, (indicated by vertical dotted lines) results in a 24-fold increase in the non-normalized response, T1, if synexpression is assumed, while only a 9-fold increase is observed in a system devoid of synexpression (Figure 5B). It should be noted that this enhanced sensitization in Figure 5B does not result from zero-order ultrasensitivity [24] of the response, T1, with respect to intermediate S1, but rather is an inherent property of cascades subject to synexpression.
The latter conclusion could be confirmed analytically for a simplified kinase cascade model, where substrate sequestration and enzyme saturation were neglected (Protocol S2). These analytical studies (Protocol S2) revealed that ultrasensitization due to synexpression is observed regardless of the parameters chosen, although the degree of ultrasensitivity is parameter-dependent. As expected, the more cascade stages are coordinately affected by the transcriptional regulator, r, the more pronounced is ultrasensitization due to synexpression (Protocol S2). Because the absolute levels of phosphoproteins within kinase cascades often differ substantially, i.e., Stot ≠ Ttot in Figure 5A [27,28], we were interested in how such differential expression affects ultrasensitization due to synexpression. It turned out that ultrasensitization due to synexpression is most pronounced if the absolute concentrations increase along the cascade, i.e., if Stot < Ttot in Figure 5A (Protocol S2), as previously reported for the mitogen-activated protein kinase cascade [27,28].
Our modeling studies presented in this section explain why coordinated upregulation of the insulin receptor and its downstream substrate, IRS-1, during adipocyte differentiation results in dramatically enhanced insulin sensitivity (see above). Importantly, ultrasensitization due to synexpression is not restricted to the regulatory mode depicted in Figure 5A, but is also observed if multiple deactivators of a signaling cascade are synexpressed (Protocol S2) as, for example, observed in fibroblast growth-factor signaling pathways [29]. Available experimental evidence supports ultrasensitization due to synexpression, since titration with pervanadate, a general inhibitor of protein tyrosine phosphatases, increases Mek phosphorylation in an ultrasensitive fashion [30], most likely by simultaneously activating multiple tyrosine-phosphorylated proteins including receptors, adaptors, and Mek itself.
Ultra(de)sensitization Can Bring About Specificity and Modularity
In the previous sections, we showed that altered substrate expression, Stot, can affect signal transmission via phosphorylation–dephosphorylation cycles (see Figure 1B) in a highly ultrasensitive fashion. Likewise, an increase in the phosphatase expression level, Ptot = P + S1P, can decrease signal transmission upon sufficiently strong stimulation in a switch-like manner (Figure 6) as suggested by transfection studies in living cells [23]. This ultrasensitivity, which will be referred to here as “ultradesensitization,” is particularly pronounced if phosphatase catalysis exhibits strong saturation (see Protocol S1), i.e., if:
Here, we propose that ultradesensitization due to induced phosphatase expression results in highly specific regulation of cellular signal transmission provided that Equation 6 holds. Consider the scheme depicted in the upper-right corner of Figure 6, where the unphosphorylated substrate, S, is phosphorylated by three kinases, K1, K2, and K3. According to Figure 6, signal transmission via K1 (red line) upon strong stimulation is specifically switched off by a relatively minor increase in the phosphatase expression level, Ptot = P + S1P (indicated by horizontal arrow), while signaling via K2 (green line) and K3 (blue line) is essentially unaffected. In other words, ultradesensitization due to phosphatase expression specifically switches off individual signaling crosstalk interactions in a binary fashion if Equation 6 holds. The differential sensitivity towards phosphatase expression shown in Figure 6 is due to the fact that catalysis (i.e., kcat,K in Figure 1B) by K1 is slower than that by K2 and K3. Further increasing the phosphatase expression sequentially deactivates signaling via K2 and K3 as indicated by the vertical dotted lines in Figure 6. Thus, four specific and binary regulatory states can be realized within two orders of magnitude of phosphatase expression, which is the maximal range of transcriptional induction or repression in vivo (e.g., [19]).
Although similar specificity can, in principle, also be achieved by reduced expression of the kinases, K1 through K3, phosphatase regulation is advantageous for two reasons. First, induced phosphatase expression (as indicated by the vertical arrow in Figure 6) abolishes K1-mediated phosphorylation of S, while leaving other actions of K1 unaffected, and thereby improves specificity in cellular regulation. Second, transcriptional regulation of phosphatase expression can result in modularization and thereby simplification of cellular regulation: weak phosphatase induction (horizontal arrow in Figure 6) simultaneously downregulates all phosphorylation events that behave like the reaction catalyzed by K1 (highly sensitive module), while leaving less sensitive modules (blue and green lines in Figure 6) unaffected.
Figure 6 Ultra(de)sensitization Can Bring About Specificity and Modularity
The response of a phosphorylation–dephosphorylation cycle (see Figure 1B) upon strong stimulation is shown as a function of the phosphatase expression level (Ptot = P + S1P) for varying the ratio of the turnover numbers, kcat,P/kcat,K. Equation 1 was used for plotting, because this expression also applies independently of the relative kinase and phosphatase expression levels provided that Ktot >> Stot and Ktot >> KM,K (see Protocol S1). The plots shown correspond to the scheme depicted in the upper-right corner, where the free unphosphorylated substrate, S, is phosphorylated by the three kinases, which differ in their turnover numbers, kcat,K: K1 (red line; kcat,P/kcat,K = 100), K2 (green line; kcat,P/kcat,K = 20), and K3 (blue line; kcat,P/kcat,K = 3). As indicated by the vertical dotted lines, ultradesensitization may result in four binary regulatory states depending on the phosphatase expression level, particularly if the phosphatase is strongly saturated with its substrate (Equation 6), which is what we assumed here (KM,P = 0.01; Stot = 1).
Importantly, similar conclusions regarding selective and modular regulation of phosphorylation events (shown in Figure 6) also hold if a promiscuous phosphatase dephosphorlyates multiple phosphorylation–dephosphorylation cycles [31–34] with high affinity (Equation 6). In this case, a minor induction in phosphatase expression is predicted to specifically deactivate all highly sensitive phosphorylation–dephosphorylation cycles (red line in Figure 6), while leaving less sensitive cycles essentially unaffected (blue and green lines in Figure 6). Transfection studies support such differential sensitivity of cellular signaling pathways towards phosphatase expression [35,36]. In addition, experimental evidence is consistent with modularization of cellular regulation, since minor alterations in PTB-1B expression were shown to simultaneously affect both cytokine- and insulin-mediated signaling [32].
Discussion
Previous research on signal transduction has mainly focused on how hormonal stimulation (fast regulation) alters gene expression (slow regulation). In comparison, the impact of gene expression on the processing of hormonal inputs has been less well investigated, although both types of regulation are intimately coupled (see Figure 1A). In the work described here, we used computational methods to show that minor alterations in the expression of signaling proteins can result in large changes in steady-state signal transmission. In other words, we showed that slow regulation of signal transduction can affect fast regulation in a highly nonlinear fashion. Because increased signal transmission elicited by transcriptional induction has been referred to as sensitization [1], we have termed such nonlinearities ultrasensitization.
Experimental data suggest that gene expression is relatively insensitive towards extracellular stimulation (e.g., [19]), so that physiological hormone concentrations will affect cellular protein expression only to a minor extent. Biochemical networks exhibiting ultrasensitization transduce such minor changes into large alterations in cellular function, so that ultrasensitization is expected to enhance the effects of extracellular hormone administration. In addition, subthreshold alterations in protein expression (e.g., Stot < Stot,T in Equation 1) will be neutralized, so that ultrasensitization is predicted to contribute to all-or-none cell-fate decisions.
If the expression of a protein subject to covalent modification by phosphorylation exhibits stochastic variations (reviewed in [37,38]), ultrasensitization can contribute to cellular heterogeneity, which is known to be important in some types of cell differentiation [37]: provided that the mean expression level is in the range of ultrasensitization (i.e., Stot ≈ Stot,T in Equation 1), some cells will be highly sensitive towards extracellular growth factors, whereas others will be essentially insensitive. In a related manner, it has been suggested that increased stochasticity in gene expression contributes to haploinsufficiency of tumor-suppressor genes [39]. In this model, transient reduction of tumor-suppressor function constitutes a window of opportunity for cancer progression. Ultrasensitization might enhance such stochastic effects by transducing relatively minor alterations in gene expression into dramatic changes in tumor-suppressor function. Importantly, ultrasensitization also explains other types of dosage sensitivity, which have been implicated in a variety of diseases. For example, relatively minor overexpression of oncogenic phosphoproteins [14] is predicted to dramatically affect mitogenic signaling, thereby contributing to carcinogenesis. Finally, strong nonlinearities such as ultrasensitization result in bistability or oscillations when combined with positive [40] or negative [41] feedback.
By analyzing steady-state signal transmission for a given stimulus as a function of protein expression, we showed that alterations in substrate expression can bring about ultrasensitization in a phosphorylation–dephosphorylation cycle (see Figure 2) provided that stimulation is sufficiently strong (Equation 4). Importantly, the kinetic scheme depicted in Figure 1B also applies for other types of covalent modification (e.g., acetylation) and for the activation cycles of small G proteins so that ultrasensitization is expected to be a widespread phenomenon in cellular information transfer. Our analytical results demonstrate that ultrasensitization upon strong stimulation requires substrate sequestration on the catalyzing enzymes (Equation 3) to be fulfilled. More specifically, substrate sequestration on the kinase and/or the phosphatase completely prevents signal transmission for weak substrate-expression levels (see Figure 3). As the amount of sequestered substrate cannot exceed the threshold, Stot,T (Equation 2), signal transmission begins to rise in an ultrasensitive fashion (ultrasensitization) as soon as the substrate expression level marginally exceeds the threshold expression level, i.e., if Stot > Stot,T (Figure 3).
Even though transfection experiments suggest that ultrasensitization occurs in vivo [23], we asked whether the predicted requirement for ultrasensitization due to substrate sequestration (Equation 3) is in accordance with available enzyme kinetic data. Unfortunately, Michaelis-Menten constants of phosphatases are often measured using nonphysiological substrates such as p-nitrophenyl phosphate (i.e., the Michaelis-Menten constants are likely to be lower in vivo), and only very few quantifications of cellular phosphatase concentrations were performed in particular cell types. Nevertheless, available data reveal some candidate phosphatases such as PP1B (KM,P = 0.04–10 μM; Ptot = 0.5 μM) and PP2B (KM,P = 2–20 μM; Ptot = 20 μM), whose cellular concentration Ptot was reported to exceed the Michaelis-Menten constant KM,P [42]. Likewise, the cellular concentration of PP2A, which has been estimated to be as much as 0.25% of the total cellular protein [43], exceeds the Michaelis-Menten constant KM,P = 10 μM [44].
Even though available data suggest that other phosphatases such as PTP-1B (KM,P = 0.6–8 μM; Ptot = 0.02 μM) and PP2C (KM,P = 0.1–0.3 μM; Ptot = 0.01 μM) do not fulfill the condition KM,P ≤ Ptot [42], it should be kept in mind that subcellular targeting, which has been shown for a variety of phosphatases (e.g., [33,34]), is known to dramatically increase the effective phosphatase concentrations. For example, phosphatases usually exhibit 2- to 10-fold higher concentrations in the nucleus than in the cytoplasm [45]. Likewise, it has been estimated that the effective concentration of signaling proteins is increased by a factor of 1,000 if both the enzyme and the substrate are localized at the membrane [46]. Finally, the threshold Stot,T (i.e., Equation 3) also depends on the ratio of the catalytic rate constants kcat,P/kcat,K (see Equation 2), which was experimentally measured for the Erk phosphorylation–dephosphorylation cycle and was shown to be 4.1–23.8 for Erk-dephosphorylation by MKP-3, and even higher (41.3–238.1) for Erk-dephosphorylation by HePTP [47,48]. Thus, experimental data suggest that ultrasensitization due to substrate sequestration occurs in vivo. Accordingly, association of phosphoproteins with catalyzing enzymes was reported to inhibit cellular signal transmission [49–51].
As shown in Figure 2, ultrasensitization can be even more pronounced for intermediate stimulus levels when compared to very strong stimulation. In this case, an additional mechanism, which is independent of substrate sequestration, contributes to pronounced ultrasensitization upon intermediate stimulation (see Figure 4): increased substrate expression induces an ultrasensitive switch from high overall phosphatase activity (S1 → S0) to high overall kinase activity (S0 → S1) if Equations 4 and 5 hold true. This ultrasensitization due to activity switching is likely to be physiologically relevant, since it simply requires that kinase catalysis is significantly less saturated than phosphatase catalysis (Equation 5), in addition to sufficiently strong stimulation (Equation 4).
In addition, we proved that coordinated expression of multiple intermediates within a kinase cascade (see Figure 5A) results in ultrasensitization due to synexpression (Figure 5B), which is more pronounced the greater the number of intermediates that are expressed coordinately (see Protocol S2). Available experimental evidence supports ultrasensitization due to synexpression, since titration with pervanadate, a general inhibitor of protein tyrosine phosphatases, increases Mek phosphorylation in an ultrasensitive fashion [30], most likely by simultaneously activating multiple tyrosine-phosphorylated proteins, including receptors, adaptors, and Mek itself. Our modeling studies demonstrate that coordinated expression of multiple cascade proteins, e.g., owing to genomic organization in operons, allows more efficient control over cellular signal transduction, when compared to expression of a single rate-limiting protein. This might be one of the reasons why functionally related proteins are frequently expressed in so-called synexpression groups [52].
Finally, we showed that induced expression of phosphatases also affects cellular information transfer, i.e., fast regulation, via phosphorylation–dephosphorylation cycles (see Figure 1B) in a highly ultrasensitive fashion (ultradesensitization), particularly if phosphatase catalysis is strongly saturated (Equation 6). Importantly, ultradesensitization can account for specificity and modularity in cellular signal transduction: a minor increase in phosphatase expression may coordinately switch off highly sensitive phosphorylation events, which we collectively referred to as the highly sensitive module (red line in Figure 6), while less sensitive modules (green and blue lines in Figure 6) are essentially unaffected. Thus, depending on the phosphatase expression level, cellular signal transduction can exhibit multiple binary (on/off) regulatory states (vertical dotted lines in Figure 6). The modular model of phosphatase action allows highly flexible regulation of cellular signal transduction, since each phosphatase has a characteristic set of substrates, which partially overlaps with that of other phosphatases, so that most phosphoproteins are dephosphorylated by multiple phosphatases [31]. Thus, depending on the phosphatase that is induced transcriptionally, each phosphoprotein might be coregulated with different phosphoproteins; that is, the modular organization differs for each phosphatase.
The paper provides the theoretical framework for more quantitative experimental measurements in order to elucidate the regulation of cellular signal transduction. More specifically, quantitative measurements of how protein expression affects cellular information transfer are highly desirable to understand how haploinsufficiency, specificity, efficiency, bistability, oscillations, noise suppression, and cellular heterogeneity arise. We propose that the predictions made in this paper should be tested experimentally by controlling transcription with recombinant vectors or by the use of transfectants with defined transgene copy numbers. Importantly, it is not sufficient to analyze the phosphorylation status of the protein studied as a measure of signal transmission, since then the potentially inactivating substrate sequestration on the phosphatase (S1P-complex in Figure 1B) is not taken into account. Thus, kinase-activity assays like that previously described for Erk activity, which uses Erk-induced MBP-phosphorylation as the readout [53], are needed. We propose to test our predictions experimentally by analyzing the impact of phosphatase expression on Erk-mediated signal transmission for other reasons as well. First, induction of Erk phosphatases owing to transcriptional feedback [29] or crosstalk [3] is known to be physiologically important. Second, the Erk–phosphatase complex (S1P in Figure 1B) is known to be catalytically inactive (e.g., [22]), which favors ultrasensitization due to substrate sequestration. Third, ultrasensitization (according to Equation 3) is also likely to be observed owing to the low Michaelis-Menten constant of MKP-3 [48] and to the fact that kcat,P >> kcat,K (see above).
As it might be difficult to achieve coordinated expression of multiple phosphoproteins in transfection experiments, ultrasensitization due to synexpression should be further tested using general inhibitors of kinase or phosphatase action. For example, an experimental investigation on the impact of hypoxia, i.e., oxygen depletion, on cellular survival signaling might be physiologically interesting. Hypoxia results in a dramatic decrease in the cellular adenosine triphosphate concentration from ~1 mM to ~20 μM [54], a value well below the Michaelis-Menten constant for adenosine triphosphate of many kinases, so that the phosphorylation rates of kinases are globally lowered [55,56]. Our analytical results (Protocol S2) predict that survival signaling mediated by protein kinases exhibits a sharp all-or-none response with respect to cellular adenosine triphosphate levels (i.e., the oxygen concentration), so that extracellular survival factors no longer rescue cells from entering into apoptosis as soon as the oxygen concentration falls below a critical level.
Our results also emphasize that the impact of gene expression on hormone-induced signal transduction should be further studied theoretically to gain insight into how other signaling modules respond to altered expression of their constituents. For example, explicit calculations (data not shown) reveal that steady-state signal transmission via proteins subject to multisite phosphorylation is much more sensitive towards induced phosphatase expression when compared to the single-site mechanism assumed in Figure 1B. Such further analyses might be done by using metabolic control analysis (reviewed in [57]), which has been recently applied to understand the impact of gene expression on cellular signaling [58,59]. Signals in living cells are often transient (e.g., owing to receptor downregulation). Although our additional calculations reveal that ultrasensitization is preserved for such transient signals (Protocol S3), further investigations are needed. Interestingly, strong substrate sequestration in a phosphorylation–dephosphorylation cycle (i.e., Stot < Stot,T in Equation 1) does not always completely abolish signal transduction, but results in transient signals owing to adaptation (F. Bruggeman and N. Blüthgen, personal communication), if the rate of kinase catalysis (kcat,K) is significantly faster than that for substrate sequestration on the phosphatase (kon,P). In this case, ultradesensitization due to altered phosphatase expression mainly affects the signal duration rather than the signal amplitude (data not shown), so that further analysis of ultra(de)sensitization may provide insight into how cells regulate mitogenesis versus differentiation [60].
Materials and Methods
Numerical simulations were done using the MATLAB computing environment (The Mathworks, Natick, Massachusetts, United States). Analytical results were confirmed using Maple 7 (Waterloo Maple, Waterloo, Ontario, Canada). Computer codes are available from the authors upon request.
Supporting Information
Protocol S1 Ultrasensitization in a Phosphorylation–Dephosphorylation Cycle—Mathematical Derivations
(107 KB PDF)
Click here for additional data file.
Protocol S2 Ultrasensitization Due to Synexpression within a Kinase Cascade—Mathematical Derivations
(50 KB PDF)
Click here for additional data file.
Protocol S3 Ultrasensitization is Preserved upon Transient Stimulation
(283 KB PDF)
Click here for additional data file.
Accession Numbers
The Swiss-Prot (http://www.ebi.ac.uk/swissprot) accession numbers for the proteins discussed in this paper are BRCA1 (P38398), Erk (P27361), H2AX (P16104), HePTP (P35236), insulin (P01308), insulin receptor (P06213), IRS-1 (P35568), MBP (P02686), Mek (Q02750), MKP-1 (P28562), MKP-3 (Q16828), p53 (P04637), PP1B (P62140), PP2A (P67775), PP2B (P16298), PP2C (P35813), pRb (P06400), PTB-1B (P18031), and PTEN (P60484).
This work was supported by the Deutsche Forschungsgemeinschaft (SFB 618) and the German Ministry for Research and Education. We thank Frank Bruggeman, Branka Cajavec, Thomas Höfer, René Hoffmann, Michael Ronellenfitsch, and Christine Sers for useful discussions.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. SL conceived and performed initial simulations. SL, NB, RS, and HH discussed preliminary results, suggested further calculations, and wrote the paper.
A previous version of this article appeared as an Early Online Release on September 26, 2005 (DOI: 10.1371/journal.pcbi.0010054.eor).
==== Refs
References
Hu X Herrero C Li WP Antoniv TT Falck-Pedersen E 2002 Sensitization of IFN-gamma Jak-STAT signalling during macrophage activation Nat Immunol 3 859 866 12172544
Ilangumaran S Ramanathan S Rottapel R 2004 Regulation of the immune system by SOCS family adaptor proteins Semin Immunol 16 351 365 15541651
Pilz RB Casteel DE 2003 Regulation of gene expression by cyclic GMP Circ Res 93 1034 1046 14645134
Yang E van Nimwegen E Zavolan M Rajewsky N Schroeder M 2003 Decay rates of human mRNAs: Correlation with functional characteristics and sequence attributes Genome Res 13 1863 1872 12902380
Futcher B Latter GI Monardo P McLaughlin CS Garrels JI 1999 A sampling of the yeast proteome Mol Cell Biol 19 7357 7368 10523624
Swameye I Muller TG Timmer J Sandra O Klingmuller U 2003 Identification of nucleocytoplasmic cycling as a remote sensor in cellular signaling by databased modeling Proc Natl Acad Sci U S A 100 1028 1033 12552139
Schoeberl B Eichler-Jonsson C Gilles ED Muller G 2002 Computational modeling of the dynamics of the MAP kinase cascade activated by surface and internalized EGF receptors Nat Biotechnol 20 370 375 11923843
Sesti G Federici M Hribal ML Lauro D Sbraccia P 2001 Defects of the insulin receptor substrate (IRS) system in human metabolic disorders FASEB J 15 2099 2111 11641236
Mellstrom B Achaval M Montero D Naranjo JR Sassone-Corsi P 1991 Differential expression of the jun family members in rat brain Oncogene 6 1959 1964 1719462
Widmann C Gibson S Johnson GL 1998 Caspase-dependent cleavage of signalling proteins during apoptosis. A turn-off mechanism for anti-apoptotic signals J Biol Chem 273 7141 7147 9507028
Miller S Yasuda M Coats JK Jones Y Martone ME 2002 Disruption of dendritic translation of CaMKIIalpha impairs stabilization of synaptic plasticity and memory consolidation Neuron 36 507 519 12408852
Whitfield ML Sherlock G Saldanha AJ Murray JI Ball CA 2002 Identification of genes periodically expressed in the human cell cycle and their expression in tumors Mol Biol Cell 13 1977 2000 12058064
Akhtar RA Reddy AB Maywood ES Clayton JD King VM 2002 Circadian cycling of the mouse liver transcriptome, as revealed by cDNA microarray, is driven by the suprachiasmatic nucleus Curr Biol 12 540 550 11937022
Blume-Jensen P Hunter T 2001 Oncogenic kinase signalling Nature 411 355 365 11357143
Fero ML Randel E Gurley KE Roberts JM Kemp CJ 1998 The murine gene p27Kip1 is haplo-insufficient for tumour suppression Nature 396 177 180 9823898
Santarosa M Ashworth A 2004 Haploinsufficiency for tumour suppressor genes: When you don't need to go all the way Biochim Biophys Acta 1654 105 122 15172699
Fambrough D McClure K Kazlauskas A Lander ES 1999 Diverse signalling pathways activated by growth factor receptors induce broadly overlapping, rather than independent, sets of genes Cell 97 727 741 10380925
Hazzalin CA Mahadevan LC 2002 MAPK-regulated transcription: A continuously variable gene switch? Nat Rev Mol Cell Biol 3 30 40 11823796
Roberts CJ Nelson B Marton MJ Stoughton R Meyer MR 2000 Signalling and circuitry of multiple MAPK pathways revealed by a matrix of global gene expression profiles Science 287 873 880 10657304
Legewie S Blüthgen N Herzel H 2005 Quantitative analysis of ultrasensitive responses FEBS J 272 4071 4079 16098190
Cohen P 2001 The role of protein phosphorylation in human health and disease. The Sir Hans Krebs Medal Lecture Eur J Biochem 268 5001 5010 11589691
Tanoue T Adachi M Moriguchi T Nishida E 2000 A conserved docking motif in MAP kinases common to substrates, activators and regulators Nat Cell Biol 2 110 116 10655591
Lammers R Bossenmaier B Cool DE Tonks NK Schlessinger J 1993 Differential activities of protein tyrosine phosphatases in intact cells J Biol Chem 268 22456 22462 7693671
Goldbeter A Koshland DE Jr 1981 An amplified sensitivity arising from covalent modification in biological systems Proc Natl Acad Sci U S A 78 6840 6844 6947258
Rosen ED Spiegelman BM 2000 Molecular regulation of adipogenesis Annu Rev Cell Dev Biol 16 145 171 11031233
MacKay VL Li X Flory MR Turcott E Law GL 2004 Gene expression analyzed by high-resolution state array analysis and quantitative proteomics: Response of yeast to mating pheromone Mol Cell Proteomics 3 478 489 14766929
Ferrell JE Jr 1996 Tripping the switch fantastic: How a protein kinase cascade can convert graded inputs into switch-like outputs Trends Biochem Sci 21 460 466 9009826
Yeung K Seitz T Li S Janosch P McFerran B 1999 Suppression of Raf-1 kinase activity and MAP kinase signalling by RKIP Nature 401 173 177 10490027
Tsang M Dawid IB 2004 Promotion and attenuation of FGF signaling through the Ras-MAPK pathway. Sci STKE 2004: pe17 PMID 15082862
Zhao Z Tan Z Diltz CD You M Fischer EH 1996 Activation of mitogen-activated protein (MAP) kinase pathway by pervanadate, a potent inhibitor of tyrosine phosphatases J Biol Chem 271 22251 22255 8703041
Tonks NK Neel BG 2001 Combinatorial control of the specificity of protein tyrosine phosphatases Curr Opin Cell Biol 13 182 195 11248552
Fukada T Tonks NK 2003 Identification of YB-1 as a regulator of PTP1B expression: Implications for regulation of insulin and cytokine signalling EMBO J 22 479 493 12554649
Janssens V Goris J 2001 Protein phosphatase 2A: A highly regulated family of serine/threonine phosphatases implicated in cell growth and signalling Biochem J 353 417 439 11171037
Bollen M 2001 Combinatorial control of protein phosphatase-1 Trends Biochem Sci 26 426 431 11440854
Tanoue T Moriguchi T Nishida E 1999 Molecular cloning and characterization of a novel dual specificity phosphatase, MKP-5 J Biol Chem 274 19949 19956 10391943
Tanoue T Yamamoto T Maeda R Nishida E 2001 A Novel MAPK phosphatase MKP-7 acts preferentially on JNK/SAPK and p38 alpha and beta MAPKs J Biol Chem 276 26629 26639 11359773
Fiering S Whitelaw E Martin DI 2000 To be or not to be active: The stochastic nature of enhancer action BioEssays 22 381 387 10723035
Rao CV Wolf DM Arkin AP 2002 Control, exploitation and tolerance of intracellular noise Nature 420 231 237 12432408
Cook DL Gerber AN Tapscott SJ 1998 Modeling stochastic gene expression: Implications for haploinsufficiency Proc Natl Acad Sci U S A 95 15641 15646 9861023
Ferrell JE Jr Machleder EM 1998 The biochemical basis of an all-or-none cell fate switch in Xenopus oocytes Science 280 895 898 9572732
Kholodenko BN 2000 Negative feedback and ultrasensitivity can bring about oscillations in the mitogen-activated protein kinase cascades Eur J Biochem 267 1583 1588 10712587
Brown GC Kholodenko BN 1999 Spatial gradients of cellular phospho-proteins FEBS Lett 457 452 454 10471827
Goldberg Y 1999 Protein phosphatase 2A: Who shall regulate the regulator? Biochem Pharmacol 57 321 328 9933020
Price NE Mumby MC 2000 Effects of regulatory subunits on the kinetics of protein phosphatase 2A Biochemistry 39 11312 11318 10985776
Bollen M Beullens M 2002 Signalling by protein phosphatases in the nucleus Trends Cell Biol 12 138 145 11859026
Kholodenko BN Hoek JB Westerhoff HV 2000 Why cytoplasmic signalling proteins should be recruited to cell membranes Trends Cell Biol 10 173 178 10754559
Mansour SJ Candia JM Matsuura JE Manning MC Ahn NG 1996 Interdependent domains controlling the enzymatic activity of mitogen-activated protein kinase kinase 1 Biochemistry 35 15529 15536 8952507
Zhou B Wang ZX Zhao Y Brautigan DL Zhang ZY 2002 The specificity of extracellular signal-regulated kinase 2 dephosphorylation by protein phosphatases J Biol Chem 277 31818 31825 12082107
Klingmuller U Lorenz U Cantley LC Neel BG Lodish HF 1995 Specific recruitment of SH-PTP1 to the erythropoietin receptor causes inactivation of JAK2 and termination of proliferative signals Cell 80 729 738 7889566
Machide M Kamitori K Kohsaka S 2000 Hepatocyte growth factor-induced differential activation of phospholipase c gamma 1 and phosphatidylinositol 3-kinase is regulated by tyrosine phosphatase SHP-1 in astrocytes J Biol Chem 275 31392 31398 10896658
Okamura H Garcia-Rodriguez C Martinson H Qin J Virshup DM 2004 A conserved docking motif for CK1 binding controls the nuclear localization of NFAT1 Mol Cell Biol 24 4184 4195 15121840
Niehrs C Pollet N 1999 Synexpression groups in eukaryotes Nature 402 483 487 10591207
Sturgill TW Ray LB Anderson NG Erickson AK 1991 Purification of mitogen-activated protein kinase from epidermal growth factor-treated 3T3-L1 fibroblasts Methods Enzymol 200 342 351 1720187
Dagher PC 2000 Modeling ischemia in vitro: Selective depletion of adenine and guanine nucleotide pools Am J Physiol Cell Physiol 279 C1270 C1277 11003607
Braunton JL Wong V Wang W Salter MW Roder J 1998 Reduction of tyrosine kinase activity and protein tyrosine dephosphorylation by anoxic stimulation in vitro Neuroscience 82 161 170 9483512
Kobryn CE Mandel LJ 1994 Decreased protein phosphorylation induced by anoxia in proximal renal tubules Am J Physiol 267 C1073 C1079 7943270
Heinrich R Schuster S 1996 The regulation of cellular systems New York Chapman & Hall 372 p.
Hornberg JJ Bruggeman FJ Binder B Geest CR de Vaate AJ 2005 Principles behind the multifarious control of signal transduction. ERK phosphorylation and kinase/phosphatase control FEBS J 272 244 258 15634347
Lee E Salic A Kruger R Heinrich R Kirschner MW 2003 The roles of APC and axin derived from experimental and theoretical analysis of the Wnt pathway PLoS Biol 1 e10. DOI: 10.1371/journal.pbio.0000010 . 14551908
Marshall CJ 1995 Specificity of receptor tyrosine kinase signalling: Transient versus sustained extracellular signal-regulated kinase activation Cell 80 179 185 7834738
|
16261195
|
PMC1274294
|
CC BY
|
2021-01-05 09:19:22
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e54
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010054
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 1626119610.1371/journal.pcbi.0010055plcb-01-05-07Research ArticleDissimilatory Metabolism of Nitrogen Oxides in Bacteria: Comparative Reconstruction of Transcriptional Networks Regulation of Nitrogen Oxides MetabolismRodionov Dmitry A 14*Dubchak Inna L 2Arkin Adam P 3Alm Eric J 3Gelfand Mikhail S 1451 Institute for Information Transmission Problems, Russian Academy of Sciences, Moscow, Russia
2 Genomics Division, Lawrence Berkeley National Laboratory, Berkeley, California, United States of America
3 Physical Biosciences Division, Lawrence Berkeley National Laboratory, Berkeley, California, United States of America
4 State Scientific Center GosNIIGenetika, Moscow, Russia
5 Department of Bioengineering and Bioinformatics, Moscow State University, Moscow, Russia
Miyano Satoru EditorUniversity of Tokyo, Japan* To whom correspondence should be addressed. E-mail: [email protected] 2005 28 10 2005 29 9 2005 1 5 e552 3 2005 29 9 2005 Copyright: © 2005 Rodionov et al.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Bacterial response to nitric oxide (NO) is of major importance since NO is an obligatory intermediate of the nitrogen cycle. Transcriptional regulation of the dissimilatory nitric oxides metabolism in bacteria is diverse and involves FNR-like transcription factors HcpR, DNR, and NnrR; two-component systems NarXL and NarQP; NO-responsive activator NorR; and nitrite-sensitive repressor NsrR. Using comparative genomics approaches, we predict DNA-binding motifs for these transcriptional factors and describe corresponding regulons in available bacterial genomes. Within the FNR family of regulators, we observed a correlation of two specificity-determining amino acids and contacting bases in corresponding DNA recognition motif. Highly conserved regulon HcpR for the hybrid cluster protein and some other redox enzymes is present in diverse anaerobic bacteria, including Clostridia, Thermotogales, and delta-proteobacteria. NnrR and DNR control denitrification in alpha- and beta-proteobacteria, respectively. Sigma-54-dependent NorR regulon found in some gamma- and beta-proteobacteria contains various enzymes involved in the NO detoxification. Repressor NsrR, which was previously known to control only nitrite reductase operon in Nitrosomonas spp., appears to be the master regulator of the nitric oxides' metabolism, not only in most gamma- and beta-proteobacteria (including well-studied species such as Escherichia coli), but also in Gram-positive Bacillus and Streptomyces species. Positional analysis and comparison of regulatory regions of NO detoxification genes allows us to propose the candidate NsrR-binding motif. The most conserved member of the predicted NsrR regulon is the NO-detoxifying flavohemoglobin Hmp. In enterobacteria, the regulon also includes two nitrite-responsive loci, nipAB (hcp-hcr) and nipC (dnrN), thus confirming the identity of the effector, i.e. nitrite. The proposed NsrR regulons in Neisseria and some other species are extended to include denitrification genes. As the result, we demonstrate considerable interconnection between various nitrogen-oxides-responsive regulatory systems for the denitrification and NO detoxification genes and evolutionary plasticity of this transcriptional network.
Synopsis
Comparative genomics is the analysis and comparison of genomes from different species. More then 100 complete genomes of bacteria are now available. Comparative analysis of binding sites for transcriptional regulators is a powerful approach for functional gene annotation. Knowledge of transcriptional regulatory networks is essential for understanding cellular processes in bacteria. The global nitrogen cycle includes interconversion of nitrogen oxides between a number of redox states. Despite the importance of bacterial nitrogen oxides' metabolism for ecology and medicine, our understanding of their regulation is limited. In this study, the researchers have applied comparative genomic approaches to describe a regulatory network of genes involved in the nitrogen oxides' metabolism in bacteria. The described regulatory network involves five nitric oxide−responsive transcription factors with different DNA recognition motifs. Different combinations of these regulators appear to regulate expression of dozens of genes involved in nitric oxide detoxification and denitrification. The reconstructed network demonstrates considerable interconnection and evolutionary plasticity. Not only are genes shuffled between regulons in different genomes, but there is also considerable interaction between regulators. Overall, the system seems to be quite conserved; however, many regulatory interactions in the identified core regulatory network are taxon-specific. This study demonstrates the power of comparative genomics in the analysis of complex regulatory networks and their evolution.
Citation:Rodionov DA, Dubchak IL, Arkin AP, Alm EJ, Gelfand MS (2005) Dissimilatory metabolism of nitrogen oxides in bacteria: Comparative reconstruction of transcriptional networks. PLoS Comput Biol 1(5): e55.
==== Body
Introduction
Interconversion of nitrogen species between a number of redox states forms the biogeochemical nitrogen cycle, which has multiple environmental impacts and industrial applications. Bacteria can utilize soluble nitrogen oxides, nitrate and nitrite, as terminal electron acceptors in oxygen-limiting conditions. Two dissimilar pathways of nitrate respiration, ammonification and denitrification, involve formation of a common intermediate, nitrite, but end in different products, ammonia and gaseous nitrogen oxides or dinitrogen, respectively (Figure 1). At the first step, nitrite is formed by one of three different types of nitrate reductases: soluble assimilatory Nas, membrane-associated respiratory Nar, and periplasmic dissimilatory Nap. The next step of ammonification is conversion of nitrite into ammonia by either respiratory cytochrome c nitrite reductase NrfA or detoxifying siroheme-containing enzyme NirBD [1]. In contrast, during denitrification, nitrite is reduced to nitric oxide (NO), nitrous oxide, and, finally, dinitrogen, using nitrogen oxide reductases NirK (or NirS), NorB, and NosZ, respectively [2].
Figure 1 The Bacterial Inorganic Nitrogen Cycle
The ammonification, denitrification, detoxification, nitrogen fixation, and nitrification pathways are shown by colored solid lines with genes names involved in the pathway. The dashed black line shows possible non-enzymatic interconversions of nitrogen oxides. The dotted line shows additional formation of NO and hydroxylamine during nitrite ammonification.
NO is a signaling and defense molecule in animals, but bacteria are sensitive to high NO concentrations due to its reactivity and membrane permeability [3]. NO and hydroxylamine, two toxic intermediates in 6-electron reduction of nitrite, could be formed during nitrite ammonification [4,5]. In addition to a classical NO reductase (NorB) present in denitrifying species, two other bacterial NO detoxification enzymes have been characterized: an NO reductase (flavorubredoxin NorVW in Escherichia coli) [6] and an NO dioxygenase (flavohemoglobin Hmp or Fhp in E. coli, Bacillus subtilis, Ralstonia eutropha, and Pseudomonas species) [7–9].
An unusual redox enzyme, called the hybrid cluster protein (Hcp) or “prismane protein,” has been extensively studied in strictly anaerobic (Desulfovibrio species) and facultative anaerobic (E. coli, Salmonella typhimurium, Acidothiobacillus ferrooxidans, Shewanella oneidensis) bacteria, where it is induced mostly during conditions of nitrite or nitrate stress, suggesting a role in nitrogen metabolism [10–14]. In the latter bacteria, the hcp gene always forms a possible operon with NADH oxidoreductase hcr, whose product catalyzes reduction of Hcp in the presence of NADH [11]. Until recently, the in vivo electron-accepting substrate of Hcp was unknown, and based on the crystal structure, NO was assumed to be a good candidate for this role [12,15]. In vitro studies demonstrated oxygen-sensitive hydroxylamine reductase activity of the E. coli Hcp protein, suggesting its possible role in detoxification of reactive by-products of nitrite reduction [16]. More recently, the requirement of the hcp gene for in vivo hydroxylamine reduction was observed in Rhodobacter capsulatus E1F1 [17].
Expression of the denitrification genes is known to be activated by nitrogen oxides and low oxygen tension [18]. Both in denitrifying and ammonifying γ-proteobacteria, the nitrate/nitrite signal is processed by the two-component sensor-regulator NarX-NarL and its paralog NarQ-NarP in E. coli that control the respiratory nitrate reductase operon nar and the nitrite ammonifying loci nir and nrf [19]. Various transcription factors of the FNR family have been described as NO-sensing regulators of denitrification: DNR in Pseudomonas species, NNR in Paracoccus denitrificans, and NnrR in Rhodobacter sphaeroides and Bradyrhizobium japonicum [18]. The DNR/NNR and NnrR proteins cluster phylogenetically in separate subgroups, separately from other family members including FNR, a global regulator of anaerobiosis in facultative anaerobic bacteria [20].
Another NO-responsive transcriptional factor, σ54-dependent NorR, activates expression of the NO reductases norVW in E. coli and norAB in R. eutropha [21,22]. Three tandem upstream activator sites with the core consensus GT-(N7)-AC were identified as NorR-binding sites observed in both promoter regions. Analysis of the adjacent regions of additional norR orthologs in bacterial genomes revealed similar tandem NorR-binding sites upstream of the norA and norB genes in Ralstonia species, norVW in Salmonella species, hmp in Vibrio cholerae and Pseudomonas aeruginosa, and hcp in V. vulnificus [23].
A nitrite-sensitive transcriptional repressor, named NsrR, has been identified in lithoautotrophic β-proteobacterium Nitrosomonas europeae, where it regulates expression of the copper-nitrite reductase nirK [24]. Co-localization of the nsrR ortholog and the hcp gene in R. capsulatus E1F1 suggested that NsrR and nitrite could be involved in the regulation of hydroxylamine assimilation in this α-proteobacterium [17]. NsrR is a member of the Rrf2 family of transcriptional regulators, which includes a putative Rrf2 regulator for a redox operon in D. vulgaris [25], the iron-responsive repressor RirA, which controls iron uptake in rhizobia [26], and the IscR repressor for the Fe-S cluster assembly operon in E. coli [27].
Despite this diversity of regulatory systems, our understanding of the regulation of the nitrogen oxides metabolism in bacteria is very limited. For example, NO- and nitrite-dependent activation of expression of hmp in E. coli and B. subtilis, hcp-hcr (nipAB) and dnrN (nipC) in S. typhimurium, and norB and aniA (nirK) in Neisseria gonorrhoeae has been described [28–30], but specific transcriptional factors involved in this control are not yet known. In this study, we analyzed regulation of the nitrosative stress and denitrification genes in available bacterial genomes using comparative genomics approaches [31,32] and predicted a large number of new regulatory elements for these genes. In addition to a complete description of the previously known NorR, DNR, and NnrR regulons, we report identification of a novel FNR-like regulator, named HcpR, for the hcp and other redox-related genes in anaerobic bacteria. Starting from very limited data, we were able to identify the NsrR-binding motif and describe the NsrR regulons in sequenced γ- and β-proteobacteria, as well as in the Bacillus and Streptomyces species. Combining published experimental and newly obtained comparative data, we have reconstructed the NO- and nitrite-dependent transcriptional regulatory network for dissimilatory metabolism of nitrogen oxides in bacteria.
Results
HcpR: Recognition Motifs and Core Regulon
A member of the CRP/FNR family of transcription factors, HcpR has been initially identified as the regulator of the hcp gene encoding the hybrid cluster protein and the frdX encoding a ferredoxin-like protein in the Desulfovibrio species, anaerobic metal-reducing δ-proteobacteria [33]. The consensus of the candidate HcpR binding sites, wTGTGAnnnnnnTCACAw, is similar to the CRP consensus of E. coli.
Close hcpR orthologs were detected in other δ-proteobacteria, namely two Geobacter species, Desulfotalea psychrophila and Desulfuromonas. However, the same CRP-like motifs were not present in these genomes. As the analysis of the regulator multiple alignment revealed a substitution in the helix-turn-helix motif involved in DNA recognition that could cause this change (see “Co-evolution of regulators and their recognition motifs” for details), and since the considered species have multiple hcp and frdX paralogs, we applied the motif detection procedure to a set of corresponding upstream regions and obtained a new FNR-like palindromic motif with consensus sequence wyTTGACnnnnGTCAArw, which has a notable distinction from the CRP-like motif in the third position (not G, but T). This recognition motif was observed upstream of most hcp and frdX paralogs in the studied δ-proteobacteria, as well as upstream of some additional genes in Desulfuromonas and Geobacter species (Figure 2 and Table S1).
Figure 2 Genomic Organization of Genes Regulated by HcpR
Yellow circles with numbers denote candidate HcpR sites with different consensus sequences. These numbers correspond to the HcpR profile numbers in Figure 3.
Figure 3 Maximum Likelihood Phylogenetic Tree of the FNR/CRP Family of Transcriptional Regulators
The third column contains sequences of helix-turn-helix motifs in the proteins. Two specificity-determining positions correlated with DNA motifs are colored (R180 and E181 in proteins correlate with G3 and A6 in DNA sites, respectively). The fourth column includes sequence logos for presumably homogeneous and large site sets and sequence consensi for small sets of DNA sites and for well-established motifs of other factors (FNR, CRP, CooA, NtcA, ArcR). The last column indicates the name of a search profile constructed in this study.
Close orthologs of hcpR from δ-proteobacteria are present in two cyanobacteria, Anabaena variabilis and Synechocystis sp. (Avar17201 and slr0449 in Figure 3), where they are divergently transcribed with the hcp and norB genes, respectively. Both these genes are preceded by candidate HcpR sites with consensus sequence TTGACnnnnGTCAA, and no other similar sites were found in the genomes of these two cyanobacteria (Figure 2).
To analyze possible regulation of the hcp genes in other taxonomic groups of bacteria, we considered their gene neighborhoods and found that genes for FNR-like regulators are often co-localized with hcp in most Clostridium species, Bacteroides, Thermotogales, and Treponema denticola (Figure 2). On the phylogenetic tree of the FNR/CRP protein family (Figure 3), all such regulators form a separate branch, named HcpR2, and additional representatives of this branch always co-occur with hcp genes in bacterial genomes. By applying the motif recognition procedure to a set of hcp upstream regions from HcpR2-containing genomes, we identified a conserved DNA motif with consensus GTAACnnnnGTTAC.
Other types of DNA motifs were observed upstream of the hcp genes in Clostridium thermocellum, C. difficile, and Porphyromonas gingivalis, and upstream of the hcp gene in Thermoanaerobacter tengcongensis. In the latter species the hcp gene has a CRP-like regulatory site and is preceded by the crp2 gene, which is an ortholog of the B. subtilis fnr gene, making it likely that crp2 regulates hcp (Figure 2). The predicted HcpR2 regulons in most Bacteroides species, P. gingivalis, Fusobacterium nucleatum, T. denticola, and Thermotogales contain only hcp genes (Figure 2).
DNR and NnrR Core Regulons
In two denitrifying Pseudomonas species, P. stutzeri and P. aeruginosa, expression of the nir, nor, and nos genes is regulated by the NO-responsive FNR-like transcriptional activator DNR that binds to a DNA motif similar to the consensus FNR box, TTGATnnnnATCAA [34,35]. By a combination of similarity search and phylogenetic analysis of the CRP/FNR protein family (Figure 3), we identified DNR orthologs in the genomes of various denitrifying β-proteobacteria, including three Ralstonia and two Burkholderia species, C. violaceum and Thiobacillus denitrificans (Figure 4). To identify the DNR recognition motif in denitrifying species, we selected the upstream regions of denitrification genes encoding nitrite, NO, and nitrous oxide reductases from genomes containing DNR orthologs and applied the motif detection procedure. The resulting FNR box-like motif with consensus CTTGATnnnnATCAAG was identified upstream of most denitrification genes (nirS, nirK, norB, nosZ) (Table S2).
Figure 4 Genomic Organization of Genes Regulated by NsrR, NorR, and DNR in β-Proteobacteria
Magenta, green, and blue circles denote candidate NsrR, NorR, and DNR sites, respectively. Candidate σ54 promoters associated with NorR sites are shown by angle arrows. Experimentally known sites of NorR and DNR are marked by “s.” Additional sites of the NarP and FNR factors are indicated by purple squares and black triangles, respectively.
No orthologs of the HcpR, DNR, NsrR (below), and NorR (below) regulators were identified in α-proteobacterial genomes. The only exception seems to be R. capsulatus E1F1, whose genome contains an nsrR ortholog close to the hcp gene within the nitrate assimilation nas gene cluster [17]. However, in denitrifying species, including R. sphaeroides and B. japonicum, the FNR-like transcriptional factor NnrR activates expression of nitrite and NO reductases and of the nnrS gene [36–38]. Orthologs of nnrR were identified in six α-proteobacteria, all of which also possess the nir and nor genes involved in denitrification (Figure 5). The NnrR orthologs form a separate branch on the phylogenetic tree of the CRP/FNR family (Figure 3). To analyze the NnrR regulon, the motif detection procedure was applied to a training set of the nir, nor, nos, and nnrS upstream regions from α-proteobacteria. The conserved NnrR recognition motif with consensus ctTTGcgnnnncgCAAag was identified upstream of most denitrification genes (Table S3). The same candidate NnrR sites have been previously identified in R. sphaeroides by a comparison of the nir, nor, and nnrS upstream regions and then confirmed for the latter gene by site-directed mutagenesis [36].
Figure 5 Genomic Organization of Genes Regulated by NnrR and NsrR in α-Proteobacteria
Orange and magenta circles denote candidate NnrR and NsrR sites, respectively. Experimentally known NnrR sites are marked by “s.”
Co-Evolution of Regulators of the CRP/FNR Family and Their Recognition Motifs
The HcpR recognition motifs identified in several bacteria demonstrated some diversity, which could be correlated with changes in the regulator DNA-binding helix-turn-helix domain. In particular, the CRP-like motif wTGTGAnnnnnnTCACAw of Desulfovibrio species differs from the FNR-like motif wyTTGACnnnnGTCAArw in other δ-proteobacteria in the third position (not T, but G). Examination of multiple alignment of the CRP/FNR-like proteins revealed one specific amino acid (R180) in the HTH motif involved in DNA recognition, which has changed from arginine (like in E. coli CRP and Desulfovibrio HcpR) to Ser or Pro in other δ-proteobacteria (see the HcpR1 branch of the phylogenetic tree in Figure 3). Similarly, the difference between the HcpR2 motif GTAACnnnnGTTAC and the motifs of δ-proteobacteria is consistent with substitution of Glu-181 in the DNA recognizing HTH domain to Pro in the HcpR2 proteins (Figure 3).
The structure of CRP in complex with its DNA operator has been determined [39]. Three positions (1ber, chain A residues 180, 181, and 185) interact with the DNA target site, and mutagenesis studies have shown that point mutations at these positions alter the specificity of the protein [40]. We systematically analyzed HcpR, HcpR2, DNR, and NnrR sites identified here, as well as several known consensus sequences for other CRP/FNR-family regulators and observed a correlation of two specificity-determining positions, R180 and E181, and contacting bases in a DNA recognition motif, G3 and A6, respectively (Figure 3).
A different substitution is observed in three bacterial species, where position 181 in the HcpR2 protein is filled by either Ser (Clostridium thermocellum and C. difficile), or Gln (Porphyromonas gingivalis). In agreement with these replacements, the candidate HcpR2 motifs in these species differ from the common recognition motif detected for most HcpR2-containing genomes (Figure 3). Finally, the Crp2 regulator in T. tengcongensis has CRP-like regulatory motif and is orthologous to the FNR regulator of B. subtilis.
The phylogenetic tree of the CRP/FNR regulators (Figure 3) represents four main groups of proteins analyzed in this study: DNR, NnrR, HcpR1, and HcpR2. It also includes several well-studied family members with established DNA-binding consensuses. All respective branches on the tree are deeply rooted, and thus their phylogenetic relationships to each other are not well resolved and differ from results of a more extensive phylogenetic analysis of the CRP/FNR-like transcriptional regulators [41]. Nevertheless, in both trees the HcpR1 and DNR branches cluster together, whereas HcpR1 and HcpR2 form two quite distinct branches on the phylogenetic tree (Figure 3). All these proteins lack the canonical FNR-type cysteine motif, thus excluding their binding of the oxygen-labile Fe-S cluster [41,42].
NsrR: Recognition Motifs and Core Regulon
The above analysis suggests that HcpR controls the hcp genes in strictly anaerobic bacteria. However, a large number of facultative anaerobic bacteria possessing the hcp gene lack hcpR orthologs. In E. coli, S. typhimurium, and S. oneidensis, hcp is expressed only under anaerobic conditions in the presence of nitrite or nitrate [10–12]. In an attempt to explain a possible molecular mechanism of this induction, we first aligned the upstream regions of the hcp genes from eight enterobacteria and identified two highly conserved regions (Figure S1). The upstream potential recognition motif resembles the consensus sequence of the FNR-binding site and thus most likely is involved in the anaerobic induction of the hcp-hcr operon by FNR [12]. The second potential DNA motif, an imperfect inverted repeat with consensus gATGyAT-(N5)-ATrCATc located downstream of the FNR site, is likely the binding site for a regulatory protein that responds to nitrogen oxides.
Construction of a recognition rule and search in complete E. coli genome identified similar sites in upstream regions of the hypothetical gene dnrN (ytfE) and the hmp gene encoding NO-detoxifying flavohemoglobin. Importantly, both these candidate sites are highly conserved in multiple alignments of dnrN and hmp upstream regions from related enterobacteria (Figures S2 and S3). Search in the S. oneidensis genome identified the same DNA motif upstream of hcp-hcr, SO4302 (dnrN), and SO2805 (nnrS), the latter encoding a hypothetical heme-copper-containing membrane protein [36]. The hmp gene is absent in this genome (Figure 6).
Figure 6 Genomic Organization of Genes Regulated by NsrR, NorR, and DNR in γ-Proteobacteria
Magenta, green, and blue circles denote candidate NsrR, NorR, and DNR sites, respectively. Candidate σ54 promoters associated with NorR sites are shown by angle arrows. Experimentally known sites of NorR and DNR are marked by “s.” Additional sites of the NarP and FNR factors are indicated by purple squares and black triangles, respectively.
Identification of the conserved palindromic motif suggests that some common transcription factor co-regulates the hcp-hcr, dnrN, hmp, and nnrS genes in enterobacteria and Shewanella species. Since bacterial transcription factors often directly regulate adjacent genes [43], we analyzed gene neighborhoods of genes preceded by the predicted sites (Figures 4 and 6). In many proteobacteria, including Vibrionales, Acinetobacter sp., Chromobacterium violaceum, Ralstonia, and Bordetella species, as well as in Gram-positive bacilli and actinobacteria, the flavohemoglobin gene hmp is positionally clustered with a hypothetical transcriptional factor from the Rrf2 protein family. The characterized members of the PF02082 family are the Rrf2 repressor for the electron transport operon hmc in Desulfovibrio vulgaris [25], the iron-sulfur cluster repressor IscR in E. coli [27], the iron-responsive regulator RirA in rhizobia [26], and the nitrite-sensitive repressor NsrR for the nitrite reductase operon nirK in Nitrosomonas europeae [24]. Phylogenetic analysis (DAR, unpublished data) demonstrated that all representatives of the Rrf2 protein family associated with hmp genes appear to be orthologs of the N. europeae NsrR protein, and thus we tentatively assign this name to the entire subfamily. Orthologs of the nitrite-sensitive repressor NsrR were identified in all β- and most γ-proteobacteria, being absent only in Pasteurellaceae, Pseudomonadales, and V. cholerae. We predict that this transcriptional factor actually binds the identified DNA motif upstream of nitrite/NO-induced genes in enterobacteria and Shewanella.
To further analyze the NsrR regulon, we constructed a recognition rule for the NsrR sites and used it to scan the genomes of γ- and β-proteobacteria (Table S4; Figures 4 and 6). The flavohemoglobin gene hmp has an upstream NsrR site in most of these genomes, excluding Pseudomonadales, V. cholerae, and Polaromonas sp., where it is a member of the NO-responsive regulon NorR (see below). The nnrS gene, another well-conserved member of the NsrR regulon, was found in some genomes within a possible operon with nsrR or hmp (Figures 4 and 6). The norB gene encoding an NO reductase in denitrifying bacteria is preceded by NsrR sites in the Neisseria species, C. violaceum, Polaromonas sp., Ralstonia solanacearum, and two Burkholderia species, B mallei and B. pseudomallei. Another key enzyme of the denitrification, the copper-containing nitrite reductase NirK, is predicted to be a member of the NsrR regulon in the Neisseria species, C. violaceum, and N. europeae (Figure 4), and in the latter bacterium it was recently shown to be a target of this nitrite-sensitive repressor [24]. In addition to γ-proteobacteria, the hcp gene was found under NsrR regulation in a β-proteobacterium (B. cepacia), and an α-proteobacterium (R. capsulatus E1F1).
Orthologs of nsrR have been also found in the complete genomes of most Bacillus and Streptomyces species, where they are clustered with the flavohemoglobin gene hmp. The only exception is B. subtilis, which has a stand-alone nsrR ortholog, yhdE. The predicted NsrR-binding motif appears to be well conserved in these Gram-positive bacteria, and candidate sites were observed only upstream the hmp genes. Multiple experimental studies in B. subtilis showed nitrite- or NO-dependent induction of expression of hmp; however, the mechanism of this control was not known [9,29]. The experimentally mapped hmp promoter in B. subtilis significantly overlaps with the predicted tandem NsrR sites [44].
The obtained data suggest that the nitrite-responsive NsrR regulon has a wide phylogenetic distribution. Its most conserved member is the NO-detoxifying flavohemoglobin Hmp, which is present both in Gram-negative and Gram-positive bacteria. Most other regulon members are involved in the nitrosative stress and denitrification. The identified NsrR recognition motif, a palindrome with consensus gATGyAT-(N5)-ATrCATc, is well conserved in most analyzed bacteria (Figure 7). The only exception is the NsrR recognition motif in Neisseria species, where symmetrical positions G4 and C16 are replaced by T and A, respectively.
Figure 7 Sequence Logos for the Identified NsrR-Binding Sites in Various Bacterial Taxa
NorR Regulon
In Vibrionales, the hcp-hcr operon is preceded by a gene that encodes a homolog of the NO-responsive regulator NorR, named NorR2. NorR is a σ54-dependent transcriptional activator that regulates expression of the NO reductase operons, norVW in E. coli and norAB in R. eutropha [21,22]. NorR binds to three tandem operator regions, inverted repeats with degenerate consensus GT-(N7)-AC, which are localized upstream of the σ54 promoter site [23]. By applying the motif detection procedure to the hcp promoter regions from four Vibrionales genomes, we identified two tandem palindromic sites with consensus GATGT-(N7)-ACATC (Figure S4). These likely binding sites for the NorR2 protein are localized immediately upstream of candidate σ54 promoters well conforming to the consensus, and thus could be involved in the NO-dependent activation of the hcp-hcr operon (Table S5). In addition, the norR2-hcp-hcr gene loci in Vibrionales contain a single NorR site without an associated σ54 promoter located upstream of the norR2 gene. This site could be involved in the negative autoregulation of the NorR2 expression (Figure 6).
To analyze analogous NO-responsive regulons in other species, we performed exhaustive similarity search and identified norR-like genes in only a limited number of β- and γ-proteobacteria (Figures 4 and 6). An E. coli-like arrangement of the divergently transcribed norR and norVW genes with conserved tandem NorR-binding sites and a σ54 promoter was found in S. typhimurium, two Erwinia species, and two Vibrio species. Other NO-detoxification genes possibly regulated by candidate NorR sites are the NO-reductase norB in Ralstonia spp. and Shewanella putrefaciens, and the NO dioxygenase hmp gene in V. cholerae, Pseudomonas spp., Polaromonas sp., and B. fungorum. In all these cases except P. stutzeri, the norR gene is clustered with the target genes on the chromosome. In the unfinished genome of P. stutzeri, the candidate tandem NorR sites followed by candidate σ54 promoters were found upstream of the hmp and dnrN genes, but the norR gene was not found in the sequenced portion of the genome. In addition, we found that V. cholerae has a second target for NorR in the genome, the hypothetical gene nnrS, which was identified as a member of various NO/nitrite-responsive regulons in other proteobacteria (NsrR, DNR, NnrR, see below).
The consensus sequences of NorR and NorR2 recognition motifs identified in various taxonomic groups have only a limited number of universally conserved positions (Figure 8). Positions G5 and T6 and complementary positions A14 and C15 are the most conserved ones throughout the NorR family, being only partially replaced in Polaromonas sp. (A5). Noteworthy, in some Vibrio species, two norR paralogs are present, norR1 and norR2, which are associated with the norVW and hcp-hcr operons, respectively. The NorR1 and NorR2 consensus sequences differ significantly in four positions (C7, G13 for NorR1 and G2, C18 for NorR2), allowing for discrimination of the target sites by the NorR paralogs in these species.
Figure 8 Sequence Logos for the Identified NorR-Binding Sites in Various Species of Proteobacteria
Complex Regulation of Hybrid Cluster Protein Genes
Differences in the predicted mode of regulation of the hybrid cluster proteins (Table 1), which are present in diverse bacterial and archaeal species, are well traced on the phylogenetic tree of this protein family (Figure 9). Indeed, the hcp gene is regulated by HcpR (highlighted in yellow) in many anaerobic bacteria, by NsrR in facultative anaerobic enterobacteria, and some β- and α-proteobacteria (in magenta), by NorR in most Vibrionales (in green), and by DNR in A. ferrooxidans and Thermochromatium tepidum (in blue). It often forms operons with either the NADH oxidoreductase hcr (in γ-proteobacteria) or the ferredoxin-like gene frdX (in δ-proteobacteria and Clostridium spp.), suggesting functional linkage between the Hcp and Hcr/FrdX proteins. In addition to predicted NsrR sites, the hcp-hcr operons in enterobacteria are also preceded by candidate binding sites of the anaerobic activator FNR, suggesting their induction during anaerobiosis (Figure S1). We also investigated the regulatory regions of hcp in two Pasteurellaceae lacking all above-mentioned nitrogen oxides regulators (Actinobacillus pleuropneumoniae and Mannheimia succiniciproducens) and found a strong candidate binding site of NarP, a response regulator from the nitrate/nitrite-specific two-component regulatory system NarQ-NarP. The NarP regulon in E. coli contains mainly genes from the nitrate/nitrite respiration pathway [18], whereas in the above two Pasteurellaceae, the NarP regulon is extended to include the detoxification genes hcp-hcr, dnrN, and norB. Additional analysis using the NarP recognition rule revealed candidate NarP sites upstream of some NsrR-regulated genes in γ-proteobacteria and Neisseria species (Figure 6). In agreement with these findings, the nitrate- and nitrite-induced transcription from the hcp promoter in E. coli was found to depend on the response regulator proteins NarL and NarP [45]. These observations show that the nitrite induction of the NO-detoxification genes in different genomes can be achieved by multiple transcriptional factors.
Table 1 Predicted Members of Regulatory and Metabolic Networks of the Nitrogen Oxides Dissimilatory Metabolism
Table 1 Continued
Figure 9 Maximum Likelihood Phylogenetic Tree of the Hybrid Cluster (Prismane) Proteins
Genes predicted to be regulated by the nitrogen oxides–related factors are highlighted by respective colors. Additionally, predicted FNR-regulated genes are denoted by black circles. Genes positionally linked to the NADH oxidoreductase hcr and the hypothetical ferredoxin frdX genes are shown by red and blue lines, respectively. Archaeal genes are shown by pointed lines.
Additional Members of the Regulons
The main regulatory interactions analyzed in this study are shown in Figure 10 and Table 1. The core regulon members, that is, genes regulated by the nitrogen-oxide responsive factors NsrR, HcpR, DNR, NnrR, and NorR in many genomes, are the hybrid cluster protein gene hcp, the NO-detoxifying flavohemoglobin hmp, two hypothetical genes dnrN and nnrS, NO reductase operon norVW, and multiple denitrification genes, nir, nor, and nos, encoding nitrite, NO, and nitrous oxide reductases, respectively. The core regulon members can be regulated by different regulators in various genomes. Further, some genes may be regulated by several regulators simultaneously. All considered regulons also contain a large number of additional members, which are summarized in Table 1.
Figure 10 Regulatory Interactions between Genes for Dissimilatory Nitrogen Oxides Metabolism in Bacteria
(A) Distribution of regulators and regulated genes. The number of cases when a gene is regulated by a specific transcription factor is indicated by the length of a colored bar in the histogram. The white bar in the histogram shows the cases when the gene is present in a genome possessing at least one of the studied regulons, but is not regulated by any of them.
(B) Combined regulatory network. Arrows denote regulatory interactions, with the thickness reflecting the frequency of the interaction in the analyzed genomes. Experimentally established (for DNR, NnrR, NsrR, and NorR) and predicted based on the regulon content (for HcpR) signal molecules are shown in filled ovals, and the protein family for each transcription factor is shown below.
Though the main target of the HcpR regulators are genes encoding the hybrid cluster protein Hcp and the hypothetical ferredoxin-containing protein FrdX, the HcpR regulons are significantly extended in δ-proteobacteria and clostridia (Table S1, Figure 2, and Table 1). In Desulfuromonas and Geobacter species, they include the nitrite reductase nrfHA, the NADH dehydrogenase ndh, and the nitrate reductase nar operon. In Desulfovibrio species, the HcpR regulon is extended to include the apsBA and sat loci involved in the sulfate reduction pathway. Among hypothetical genes, the predicted HcpR regulons often contain ferredoxin-, hemerythrin-, or cytochrome c-like genes. For instance, the CTC0897-CTC0898 operon of C. tetani encoding a permease and a ferredoxin-like protein, respectively, is likely regulated by the divergently transcribed paralog of HcpR2 (CTC0896).
Additional members of the NsrR regulon were identified in all enterobacteria (Figure 6). Two hypothetical transporter genes, that are homologous to the tellurite resistance gene tehB and the nitrite extrusion gene narK, could be involved in the protection from the nitrosative stress by excretion of nitrite from cytoplasm. In support of these observations, the NsrR regulons were found to include a narK-like transporter gene (nasA) in β-proteobacterium C. violaceum and a tehB homolog in Photobacterium profundum. The V. fischeri NsrR regulon includes a homolog of the eukaryotic alternative oxidase Aox. In Legionella pneumophila, a non-denitrifying γ-proteobacterium without hmp, hcp, dnrN, and nnrS genes, the glbN-lpg2536 operon encoding a heme-containing cyanoglobin and a hypothetical ferredoxin reductase was found to be a sole NsrR target. We suggest that both these genes could be involved in the detoxyfication process by mediating NO oxidation (similar to the flavohemoglobin Hmp).
Denitrifying bacteria like the Pseudomonas and Burkholderia species contain additional members of the DNR and NnrR regulons (Figures 4 and 5). For instance, the hemN gene encoding an O2-independent coproporphyrinogen III oxidase involved in the protoheme biosynthesis and relevant for denitrification [46] is preceded by a strong DNR site in the Burkholderia species but has candidate NnrR-binding sites in some α-proteobacteria. In Brucella melitensis, the NnrR regulon includes the nosA gene, encoding an outer membrane copper receptor, shown to be required for the assembly of the copper-containing nitrous oxide reductase in P. stutzeri [47]. Finally, a gene of unknown function (COG4309) is probably co-transcribed with the NnrR-regulated nor genes in three rhizobial species, where as in three β-proteobacteria this gene is a predicted member of the NsrR regulon and it is often co-transcribed with norB (Figure 4). These observations indicate that the COG4309 family could be relevant for function of the NO reductase.
Two close dnr paralogs, most likely resulting from a recent duplication, were found in A. ferrooxidans. One is divergently transcribed with the hcp-COG0543 operon, which is preceded by a strong candidate DNR site. The second paralog clusters with the cgb-COG0543-COG0446 operon, which also is preceded by a candidate DNR site. The product of the first gene in this operon is similar to a single-domain hemoglobin in Campylobacter jejuni, whish is indispensable for defense against NO and nitrosating agents [48]. Thus we predict that the recently duplicated DNR paralogs in A. ferrooxidans regulate two different NO-detoxifying systems.
Discussion
The results of this study demonstrate considerable variability of the metabolic and regulatory systems for nitrogen oxides (Figure 10). Many genes change the regulators in different genomes (Table 1). However, overall, the system seems to be quite conserved. Genes involved in denitrification, such as nir, nor, and nos, are mainly regulated by two NO-responsive transcriptional activators from the FNR/CRP family, NnrR in α-proteobacteria and DNR in β-proteobacteria, and Pseudomonas spp. Three different nitrogen oxides-responsive transcription factors appear to regulate genes required for defense against the nitrosative stress: the σ54-dependent activator NorR from the NtrC family in some γ- and β-proteobacteria (present in at least 18 species), the Rrf2 family NsrR repressor widely distributed in proteobacteria and firmicutes (39 species), and the FNR-like transcription factor HcpR in diverse anaerobic bacteria (22 species). The primary targets of the newly identified regulator HcpR are the hybrid-cluster protein Hcp, which has a protective role in nitrite stress conditions, and the associated ferredoxin-like proteins. NorR usually regulates cytoplasmic NO reductase norVW and sometimes another membrane-bound NO reductase (norB) and the NO dioxygenase hmp. The NsrR regulon almost always includes the hmp and hcp genes, as well as one or both genes of unknown function, dnrN and nnrS.
On the whole, the NsrR, NorR, and DNR regulons are differentially distributed in γ- and β-proteobacteria, and the former prevails over the other two. All three regulons are present only in three Ralstonia species, where NsrR controls the NO-detoxifying flavohemoglobin, whereas NorR and DNR regulate the denitrification genes (nir, nor, and nos). In addition, the NorR and NsrR factors co-occur in ten other non-denitrifying species, complementing each other in the control of the nitrosative stress genes. Finally, NorR and DNR regulators were found only in P. aeruginosa, and NsrR and DNR co-occur in four denitrifying β-proteobacteria.
Some regulatory interactions in the identified core regulatory network seem to be taxon-specific (thin lines in Figure 10B; see also “Complex regulation of hybrid cluster protein genes” above). They include NsrR-regulated norVW and nos in P. profundum, norB and nirK in various β-proteobacteria; NorR-regulated dnrN in P. stutzeri and nnrS in P. aeruginosa; HcpR-regulated dnrN in B. fragilis and Desulfuromonas, norB in Synechocystis sp.; and DNR-regulated nnrS and hmp in P. stutzeri, hcp in A. ferrooxidans and T. tepidum. The extensions of the NsrR regulon include the denitrification genes nirK and norB in Neisseria species, Burkholderia spp., and C. violaceum. The former is of particular interest as, in contrast to the latter two lineages, the Neisseria species lack the DNR regulator, assuming a lineage-specific substitution of both the transcription factor and its binding sites. Indeed, in the Neisseria species, the complete denitrification pathway including nitrite, NO, and nitrous oxide reductases, as well as dnrN and the two-component regulatory system narQP, seems to be regulated by NsrR. The hypothesis that NsrR mediates regulation of denitrification genes in Neisseria is further supported by the observation that in N. gonorrhoeae, the norB gene is strongly induced by NO independently of FNR and NarP [30].
Not only genes are shuffled between regulons in different genomes, but there may exist considerable interaction between regulators. Firstly, in some species the DNR regulon overlaps with other nitrogen oxide-responsive regulons. The upstream regions of norB and nirK in C. violaceum, COG4309-norB in R. solanacearum, and nnrS2 in Burkholderia species contain two candidate regulatory sites, a downstream NsrR site and an upstream DNR site (Figure 4), yielding both positive regulation by the NO-responsive activator DNR and nitrite-induced de-repression by the NsrR repressor. Secondly, the NO-detoxifying gene hmp in P. stutzeri is preceded by three candidate NorR sites at positions −192, −173, and −148 (relative to the translational start site), a strong DNR site at position −116, and a putative σ54 promoter at position −91 (Figure 6). This arrangement of regulatory elements indicates dual positive control of the hmp expression by different NO-responsive activators, σ54-dependent NorR and σ70-dependent DNR. Finally, in many cases genes are regulated by two additional regulators, the oxygen-responsive factor FNR (hcp-hcr, hmp, and narK in enterobacteria) and the nitrite/nitrate-sensitive two-component system NarQ/NarP (hcp-hcr, dnrN, and hmp in enterobacteria, nnrS in Vibrionales and Shewanella spp.). More complex regulatory interactions are observed in Neisseria spp., where NsrR regulates the NarQ/NarP system, whereas the common upstream region of the divergently transcribed genes norB and nirK contains two candidate NsrR sites, a candidate NarP site in the middle, and an FNR-binding site immediately upstream of nirK, the latter being involved in the anaerobic induction of this gene [49].
Various aspects of the described regulatory network for the nitrogen oxides metabolism are verified by a large number of independent experimental observations. Upregulation of the hcp gene in response to growth on nitrate or nitrite was reported in S. oneidensis, E. coli, S. typhimuruim, and D. vulgaris [10–12,14]; the same pattern of regulation was observed for dnrN in S. typhimuruim and P. stutzeri [12,34]. Our prediction of positive regulation of hcp-frdX and negative regulation of the sulfate reduction genes apsBA and sat [33] was validated in a macroarray hybridization study, where hcp was upregulated 255-fold with 5 mM nitrite, whereas aprAB and sat were downregulated nearly 10-fold in the same conditions [14]. In addition, nitrite induced transcription of thiosufate reductase phsA (a candidate member of the HcpR regulon) and inhibited the membrane-bound electron transport complex qmoABC, located just downstream of the apsBA genes and thus also possibly repressed by HcpR.
The flavohemoglobin gene hmp is induced by NO and nitrite in E. coli and B. subtilis [28,29], and the mechanism of the hmp regulation by the nitrite repressor NsrR predicted in this study is conserved in these diverse bacteria. Another NO-mediated mechanism of hmp regulation in E. coli by the O2-responsive repressor FNR was proposed [50]. At that, NO was found to inactivate FNR anaerobically, restoring the hmp expression. However, our data indicate that, in contrast to the candidate NsrR site, the FNR binding site is conserved in only closely related bacteria E. coli and S. typhimurium (Figure S3). Finally, NO induces the norB expression in N. gonorrhoeae, but it was found to be independent of known nitrogen oxide-responsive regulators [30]. Here we describe possible co-regulation of all denitrification genes in the Neisseria species by the nitrite-sensitive repressor NsrR. Recently, transcriptional regulation of the flavohemoglobin gene fhp (hmp ortholog) by the NO-responsive regulator FhpR (NorR ortholog) has been demonstrated in P. aeruginosa [51].
A recent paper by Elvers et al. [52] describes a new nitrosative stress-responsive regulon in ɛ-proteobacterium C. jejuni regulated by a member of the CRP/FNR family. This regulator (named NssR) is homologous (26% identity) to the HcpR2 factor from T. maritima described in this study. It has an FNR-like recognition motif (TTAACnnnnGTTAA) and specificity-determining positions A180 and Q181, conforming to the correlation between these two positions and contacting bases in the DNA motif observed in this study. NssR positively regulates expression of the Campylobacter globin gene cgb, encoding a single-domain hemoglobin that mediates resistance to NO and nitrosative stress [48]. Thus the regulation of the cgb gene by HcpR in A. ferrooxidans predicted in this study is in agreement with verified NssR-mediated activation of the cgb ortholog in C. jejuni.
While this study was being completed, we obtained a personal communication from S. Spiro from Georgia Institute of Technology, Atlanta, Georgia, United States, that the NsrR ortholog in E. coli (YjeB) is a NO-sensitive transcriptional repressor of several nitrosative stress responsive genes including hmp, ytfE (dnrN), and ygbA. Moreover, a common inverted repeat sequence coincided with the defined here NsrR recognition motif was shown to be involved in NsrR-mediated repression of the ytfE gene.
The mechanisms of regulation of nitrogen oxides metabolism in bacteria are of great importance both for ecology and medicine. Nitrifying and denitrifying microorganisms are significant sources of both nitric and nitrous oxides production in the atmosphere, and thus have a great impact in the greenhouse effect [53]. Nitrate has become a pollutant of groundwater and surface water. NO and reactive nitrogen intermediates are also part of the arsenal of antimicrobial agents produced by macrophages [54]. Therefore, nitrosative stress tolerance genes, which are inducible after invasion, provide a strong advantage for pathogenic bacteria that need to resist the host defense system. Here we tentatively characterized the transcriptional regulatory network for the genes involved in these significant metabolic processes. Overall, although each particular prediction made in this study may require experimental verification, the emerging overall picture seems to be rather consistent and robust.
Materials and Methods
Complete and partial bacterial genomes were downloaded from GenBank [55]. Preliminary sequence data were also obtained from the Institute for Genomic Research (http://www.tigr.org), the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk/), and the DOE Joint Genome Institute (http://jgi.doe.gov). The gene identifiers from GenBank are used throughout. Genome abbreviations are listed in Table 2. Protein similarity search was done using the Smith-Waterman algorithm implemented in the GenomeExplorer program [56]. Orthologous proteins were defined by the best bidirectional hits criterion and named by either a common name of characterized protein or by an identifier in the Clusters of Orthologous Groups (COG) database for uncharacterized proteins [57]. Distant homologs were identified using PSI-BLAST [58]. The SEED tool, which combines protein similarity search, positional gene clustering, and phylogenetic profiling of genes, was applied for comparative analysis and annotation of multiple microbial genomes (see the “Nitrosative stress and Denitrification” subsystems at http://theseed.uchicago.edu/FIG/index.cgi). The phylogenetic trees were constructed by the maximum likelihood method implemented in PHYLIP [59] using multiple sequence alignments of protein sequences produced by CLUSTALX [60]. In addition, the InterPro [61], and PFAM [62] databases were used to verify protein functional and structural annotations.
Table 2 List of Genome Abbreviations Used in This Study
Table 2 Continued
For identification of candidate regulatory motifs, we started from sets of potentially co-regulated genes (using previous experimental and general functional considerations). A simple iterative procedure implemented in the program SignalX (as described previously in [63]) was used for construction of common transcription factor–binding motifs in sets of upstream gene fragments. Each genome encoding the studied transcription factor was scanned with the constructed profile using the GenomeExplorer software (see the detailed description at http://bioinform.genetika.ru/projects/reconstruction/index.htm), and genes with candidate regulatory sites in the upstream regions were selected. The threshold for the site search was defined as the lowest score observed in the training set. Dependent on the DNA motif and the number of sites in the training set, such threshold could be too strict or too permissive. It seems that the threshold choice was adequate in our cases, as little clear false positives were encountered, and, on the other hand, most functionally relevant genes were found to belong to at least one of the studied regulons (Table 3 and Results and Discussion sections). The upstream regions of genes that are orthologous to genes containing regulatory sites of any of studied nitrogen-related factors were examined for candidate sites even if these were not detected automatically with a given threshold. Among new candidate members of a regulon, only genes having candidate sites conserved in at least two other genomes were retained for further analysis. We also included new candidate regulon members that are functionally related to the nitrogen oxides metabolism. This procedure allowed us to reject a small number of false positive sites identified after scanning of microbial genomes (Table 3). Sequence logos for derived regulatory motifs were drawn using WebLogo package v.2.6 [64] (http://weblogo.berkeley.edu/).
Table 3 Distribution of Predicted Regulatory Sites in Bacterial Genomes
Supporting Information
Figure S1 Multiple Sequence Alignment of the Upstream Regions of the hcp-hcr Operons from Enterobacteria
(22 KB DOC)
Click here for additional data file.
Figure S2 Multiple Sequence Alignment of the Upstream Regions of the dnrN Genes from Enterobacteria
(20 KB DOC)
Click here for additional data file.
Figure S3 Multiple Sequence Alignment of the Upstream Regions of the hmp Genes from Enterobacteria
(26 KB DOC)
Click here for additional data file.
Figure S4 Multiple Sequence Alignment of the Upstream Regions of the hcp-hcr Operons from Vibrio Species
(24 KB DOC)
Click here for additional data file.
Table S1 Candidate HcpR-Binding Sites
(40 KB XLS)
Click here for additional data file.
Table S2 Candidate DNR-Binding Sites
(34 KB XLS)
Click here for additional data file.
Table S3 Candidate NnrR-Binding Sites
(29 KB XLS)
Click here for additional data file.
Table S4 Candidate NsrR-Binding Sites
(24 KB XLS)
Click here for additional data file.
Table S5 Candidate NorR-Binding Sites and Corresponding σ54 Promoters
(23 KB XLS)
Click here for additional data file.
Accession Numbers
The Pfam (http://www.sanger.ac.uk/Software/Pfam/) accession numbers for products discussed in this paper are: hypothetical transcriptional factor from Rrf2 protein family (PF02082) and eukaryotic alternative oxidase Aox (PF01786).
Complete and partial bacterial genomes were downloaded from GenBank [55].
This study was partially supported by grants from the Howard Hughes Medical Institute (55000309 to MSG), the Russian Fund of Basic Research (04–04–49361 to DAR), the Russian Science Support Fund (MSG), and the Russian Academy of Sciences (Programs “Molecular and Cellular Biology” and “Origin and Evolution of the Biosphere”). ILD, EJA, and APA were in part supported by a US Department of Energy Genomics: GTL grant (DE-AC03-76SF00098, to APA). We thank Anna Gerasimova and Dmitry Ravcheev for the FNR and NarP recognition profiles, and Andrey Mironov and Sergey Stolyar for useful discussions.
Competing interests. The authors have declared that no competing interests exist.
Author contributions. DAR and MSG conceived and designed the experiments. DAR performed the experiments. DAR, ILD, APA, and EJA analyzed the data. DAR, ILD, APA, EJA, and MSG wrote the paper.
A previous version of this article appeared as an Early Online Release on September 29, 2005 (DOI: 10.1371/journal.pcbi.0010055.eor).
Abbreviation
NOnitric oxide.
==== Refs
References
Simon J 2002 Enzymology and bioenergetics of respiratory nitrite ammonification FEMS Microbiol Rev 26 285 309 12165429
Shapleigh JP 2000 The denitrifying prokaryotes Dworkin M Falkow N Rosenberg H Schleifer KH Stackebrandt E The prokaryotes: An evolving electronic resource for the microbiological community, 3rd edition New York Springer-Verlag
Poole RK 2005 Nitric oxide and nitrosative stress tolerance in bacteria Biochem Soc Trans 33 176 180 15667299
Rudolf M Einsle O Neese F Kroneck PM 2002 Pentahaem cytochrome c nitrite reductase: Reaction with hydroxylamine, a potential reaction intermediate and substrate Biochem Soc Trans 30 649 653 12196156
Kuznetsova S Knaff DB Hirasawa M Setif P Mattioli TA 2004 Reactions of spinach nitrite reductase with its substrate, nitrite, and a putative intermediate, hydroxylamine Biochemistry 43 10765 10774 15311938
Gomes CM Giuffre A Forte E Vicente JB Saraiva LM 2002 A novel type of nitric-oxide reductase. Escherichia coli flavorubredoxin J Biol Chem 277 25273 25276 12101220
Gardner PR Gardner AM Martin LA Salzman AL 1998 Nitric oxide dioxygenase: An enzymic function for flavohemoglobin Proc Natl Acad Sci U S A 95 10378 10383 9724711
Cramm R Siddiqui RA Friedrich B 1994 Primary sequence and evidence for a physiological function of the flavohemoprotein of Alcaligenes eutrophus
J Biol Chem 269 7349 7354 8125952
LaCelle M Kumano M Kurita K Yamane K Zuber P 1996 Oxygen-controlled regulation of the flavohemoglobin gene in Bacillus subtilis
J Bacteriol 178 3803 3808 8682784
Beliaev AS Thompson DK Khare T Lim H Brandt CC 2002 Gene and protein expression profiles of Shewanella oneidensis during anaerobic growth with different electron acceptors OMICS 6 39 60 11881834
van den Berg WA Hagen WR van Dongen WM 2000 The hybrid-cluster protein (‘prismane protein') from Escherichia coli . Characterization of the hybrid-cluster protein, redox properties of the [2Fe-2S] and [4Fe-2S-2O] clusters and identification of an associated NADH oxidoreductase containing FAD and [2Fe-2S] Eur J Biochem 267 666 676 10651802
Kim CC Monack D Falkow S 2003 Modulation of virulence by two acidified nitrite-responsive loci of Salmonella enterica serovar Typhimurium Infect Immun 71 3196 3205 12761099
Dominy CN Deane SM Rawlings DE 1997 A geographically widespread plasmid from Thiobacillus ferrooxidans has genes for ferredoxin-, FNR-, prismane- and NADH-oxidoreductase-like proteins which are also located on the chromosome Microbiology 143 3123 3136 9353917
Haveman SA Greene EA Stilwell CP Voordouw JK Voordouw G 2004 Physiological and gene expression analysis of inhibition of Desulfovibrio vulgaris hildenborough by nitrite J Bacteriol 186 7944 7950 15547266
Cooper SJ Garner CD Hagen WR Lindley PF Bailey S 2000 Hybrid-cluster protein (HCP) from Desulfovibrio vulgaris (Hildenborough) at 1.6 A resolution Biochemistry 39 15044 15054 11106482
Wolfe MT Heo J Garavelli JS Ludden PW 2002 Hydroxylamine reductase activity of the hybrid cluster protein from Escherichia coli
J Bacteriol 184 5898 5902 12374823
Cabello P Pino C Olmo-Mira MF Castillo F Roldan MD 2004 Hydroxylamine assimilation by Rhodobacter capsulatus E1F1. requirement of the hcp gene (hybrid cluster protein) located in the nitrate assimilation nas gene region for hydroxylamine reduction J Biol Chem 279 45485 45494 15322098
Zumft WG 2002 Nitric oxide signaling and NO dependent transcriptional control in bacterial denitrification by members of the FNR-CRP regulator family J Mol Microbiol Biotechnol 4 277 286 11931559
Stewart V 1993 Nitrate regulation of anaerobic respiratory gene expression in Escherichia coli
Mol Microbiol 9 425 434 8412692
Korner H Sofia HJ Zumft WG 2003 Phylogeny of the bacterial superfamily of Crp-Fnr transcription regulators: Exploiting the metabolic spectrum by controlling alternative gene programs FEMS Microbiol Rev 27 559 592 14638413
Tucker NP D'Autreaux B Studholme DJ Spiro S Dixon R 2004 DNA binding activity of the Escherichia coli nitric oxide sensor NorR suggests a conserved target sequence in diverse proteobacteria J Bacteriol 186 6656 6660 15375149
Pohlmann A Cramm R Schmelz K Friedrich B 2000 A novel NO-responding regulator controls the reduction of nitric oxide in Ralstonia eutropha
Mol Microbiol 38 626 638 11069685
Busch A Pohlmann A Friedrich B Cramm R 2004 A DNA region recognized by the nitric oxide-responsive transcriptional activator NorR is conserved in beta- and gamma-proteobacteria J Bacteriol 186 7980 7987 15547270
Beaumont HJ Lens SI Reijnders WN Westerhoff HV van Spanning RJ 2004 Expression of nitrite reductase in Nitrosomonas europaea involves NsrR, a novel nitrite-sensitive transcription repressor Mol Microbiol 54 148 158 15458412
Keon RG Fu R Voordouw G 1997 Deletion of two downstream genes alters expression of the hmc operon of Desulfovibrio vulgaris subsp. vulgaris Hildenborough Arch Microbiol 167 376 383 9148780
Todd JD Wexler M Sawers G Yeoman KH Poole PS 2001 RirA, an iron-responsive regulator in the symbiotic bacterium Rhizobium leguminosarum
Microbiology 148 4059 4071
Schwartz CJ Giel JL Patschkowski T Luther C Ruzicka FJ 2001 IscR, an Fe-S cluster-containing transcription factor, represses expression of Escherichia coli genes encoding Fe-S cluster assembly proteins Proc Natl Acad Sci U S A 98 14895 14900 11742080
Poole RK Anjum MF Membrillo-Hernandez J Kim SO Hughes MN 1996 Nitric oxide, nitrite, and Fnr regulation of hmp (flavohemoglobin) gene expression in Escherichia coli K-12 J Bacteriol 178 5487 5492 8808940
Moore CM Nakano MM Wang T Ye RW Helmann JD 2004 Response of Bacillus subtilis to nitric oxide and the nitrosating agent sodium nitroprusside J Bacteriol 186 4655 4664 15231799
Householder TC Fozo EM Cardinale JA Clark VL 2000 Gonococcal nitric oxide reductase is encoded by a single gene, norB , which is required for anaerobic growth and is induced by nitric oxide Infect Immun 68 5241 5246 10948150
Osterman A Overbeek R 2003 Missing genes in metabolic pathways: A comparative genomics approach Curr Opin Chem Biol 7 238 251 12714058
Gelfand MS Novichkov PS Novichkova ES Mironov AA 2000 Comparative analysis of regulatory patterns in bacterial genomes Brief Bioinform 1 357 371 11465053
Rodionov DA Dubchak I Arkin A Alm E Gelfand MS 2004 Reconstruction of regulatory and metabolic pathways in metal-reducing delta-proteobacteria Genome Biol 5 R90 15535866
Vollack KU Zumft WG 2001 Nitric oxide signaling and transcriptional control of denitrification genes in Pseudomonas stutzeri
J Bacteriol 183 2516 2526 11274111
Arai H Mizutani M Igarashi Y 2003 Transcriptional regulation of the nos genes for nitrous oxide reductase in Pseudomonas aeruginosa
Microbiology 149 29 36 12576577
Bartnikas TB Wang Y Bobo T Veselov A Scholes CP 2002 Characterization of a member of the NnrR regulon in Rhodobacter sphaeroides 2.4.3 encoding a haem-copper protein Microbiology 148 825 833 11882718
Kwiatkowski AV Shapleigh JP 1996 Requirement of nitric oxide for induction of genes whose products are involved in nitric oxide metabolism in Rhodobacter sphaeroides 2.4.3 J Biol Chem 271 24382 24388 8798693
Mesa S Bedmar EJ Chanfon A Hennecke H Fischer HM 2003
Bradyrhizobium japonicum NnrR, a denitrification regulator, expands the FixLJ-FixK2 regulatory cascade J Bacteriol 185 3978 3982 12813094
Schultz SC Shields GC Steitz TA 1991 Crystal structure of a CAP-DNA complex: The DNA is bent by 90 degrees Science 253 1001 1007 1653449
Luscombe NM Thornton JM 2002 Protein-DNA interactions: Amino acid conservation and the effects of mutations on binding specificity J Mol Biol 320 991 1009 12126620
Korner H Sofia HJ Zumft WG 2003 Phylogeny of the bacterial superfamily of Crp-Fnr transcription regulators: Exploiting the metabolic spectrum by controlling alternative gene programs FEMS Microbiol Rev 27 559 592 14638413
Crack J Green J Thomson AJ 2004 Mechanism of oxygen sensing by the bacterial transcription factor fumarate-nitrate reduction (FNR) J Biol Chem 279 9278 9286 14645253
Korbel JO Jensen LJ von Mering C Bork P 2004 Analysis of genomic context: Prediction of functional associations from conserved bidirectionally transcribed gene pairs Nat Biotechnol 22 911 917 15229555
LaCelle M Kumano M Kurita K Yamane K Zuber P 1996 Oxygen-controlled regulation of the flavohemoglobin gene in Bacillus subtilis
J Bacteriol 178 3803 3808 8682784
Filenko NA Browning DF Cole JA 2005 Transcriptional regulation of a hybrid cluster (prismane) protein Biochem Soc Trans 33 195 197 15667305
Zumft WG 1997 Cell biology and molecular basis of denitrification Microbiol Mol Biol Rev 61 533 616 9409151
Lee HS Abdelal AH Clark MA Ingraham JL 1991 Molecular characterization of nosA , a Pseudomonas stutzeri gene encoding an outer membrane protein required to make copper-containing N2O reductase J Bacteriol 173 5406 5413 1885521
Elvers KT Wu G Gilberthorpe NJ Poole RK Park SF 2004 Role of an inducible single-domain hemoglobin in mediating resistance to nitric oxide and nitrosative stress in Campylobacter jejuni and Campylobacter coli
J Bacteriol 186 5332 5341 15292134
Lissenden S Mohan S Overton T Regan T Crooke H 2000 Identification of transcription activators that regulate gonococcal adaptation from aerobic to anaerobic or oxygen-limited growth Mol Microbiol 37 839 855 10972806
Cruz-Ramos H Crack J Wu G Hughes MN Scott C 2002 NO sensing by FNR: Regulation of the Escherichia coli NO-detoxifying flavohaemoglobin, Hmp EMBO J 21 3235 3244 12093725
Arai H Hayashi M Kuroi A Ishii M Igarashi Y 2005 Transcriptional regulation of the flavohemoglobin gene for aerobic nitric oxide detoxification by the second nitric oxide-responsive regulator of Pseudomonas aeruginosa
J Bacteriol 187 3960 3968 15937158
Elvers KT Turner SM Wainwright LM Marsden G Hinds J 2005 NssR, a member of the Crp-Fnr superfamily from Campylobacter jejuni , regulates a nitrosative stress-responsive regulon that includes both a single-domain and a truncated haemoglobin Mol Microbiol 57 735 750 16045618
Levine JS Augustsson TR Anderson IC Hoell JM Jr 1984 Tropospheric sources of NOx: Lightning and biology Atmos Environ 18 1797 1804 11540827
Fang FC 1997 Perspectives series: Host/pathogen interactions. Mechanisms of nitric oxide-related antimicrobial activity J Clin Invest 99 2818 2825 9185502
Benson DA Karsch-Mizrachi I Lipman DJ Ostell J Wheeler DL 2005 GenBank Nucleic Acids Res 33 D34 38 15608212
Mironov AA Vinokurova NP Gelfand MS 2000 GenomeExplorer: Software for analysis of complete bacterial genomes Mol Biol 34 222 231
Tatusov RL Natale DA Garkavtsev IV Tatusova TA Shankavaram UT 2001 The COG database: New developments in phylogenetic classification of proteins from complete genomes Nucleic Acids Res 29 22 28 11125040
Altschul SF Koonin EV 1998 Iterated profile searches with PSI-BLAST—A tool for discovery in protein databases Trends Biochem Sci 23 444 447 9852764
Felsenstein J 1981 Evolutionary trees from DNA sequences: A maximum likelihood approach J Mol Evol 17 368 376 7288891
Thompson JD Gibson TJ Plewniak F Jeanmougin F Higgins DG 1997 The CLUSTAL_X windows interface: Flexible strategies for multiple sequence alignment aided by quality analysis tools Nucleic Acids Res 25 4876 4882 9396791
Apweiler R Attwood TK Bairoch A Bateman A Birney E 2001 The InterPro database, an integrated documentation resource for protein families, domains and functional sites Nucleic Acids Res 29 37 40 11125043
Bateman A Birney E Cerruti L Durbin R Etwiller L 2002 The Pfam protein families database Nucleic Acids Res 30 276 280 11752314
Gelfand MS Koonin EV Mironov AA 2000 Prediction of transcription regulatory sites in Archaea by a comparative genomic approach Nucleic Acids Res 28 695 705 10637320
Crooks GE Hon G Chandonia JM Brenner SE 2004 WebLogo: A sequence logo generator Genome Res 14 1188 1190 15173120
|
16261196
|
PMC1274295
|
CC BY
|
2021-01-05 09:18:23
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e55
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010055
|
oa_comm
|
==== Front
PLoS Comput BiolPLoS Comput. BiolpcbiplcbploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, USA 1626119710.1371/journal.pcbi.0010057plcb-01-05-06EditorialTen Simple Rules for Getting Published Bourne Philip E Philip E. Bourne is Editor-in-Chief of PLoS Computational Biology. E-mail: [email protected] 2005 28 10 2005 1 5 e57Copyright: © 2005 Philip E. Bourne.2005This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited. Citation:Bourne PE (2005) Ten simple rules for getting published. PLoS Comput Biol 1(5): e57.
==== Body
The student council (http://www.iscbsc.org/) of the International Society for Computational Biology asked me to present my thoughts on getting published in the field of computational biology at the Intelligent Systems in Molecular Biology conference held in Detroit in late June of 2005. Close to 200 bright young souls (and a few not so young) crammed into a small room for what proved to be a wonderful interchange among a group of whom approximately one-half had yet to publish their first paper. The advice I gave that day I have modified and present as ten rules for getting published.
Rule 1: Read many papers, and learn from both the good and the bad work of others.
It is never too early to become a critic. Journal clubs, where you critique a paper as a group, are excellent for having this kind of dialogue. Reading at least two papers a day in detail (not just in your area of research) and thinking about their quality will also help. Being well read has another potential major benefit—it facilitates a more objective view of one's own work. It is too easy after many late nights spent in front of a computer screen and/or laboratory bench to convince yourself that your work is the best invention since sliced bread. More than likely it is not, and your mentor is prone to falling into the same trap, hence rule 2.
Rule 2: The more objective you can be about your work, the better that work will ultimately become.
Alas, some scientists will never be objective about their own work, and will never make the best scientists—learn objectivity early, the editors and reviewers have.
Rule 3: Good editors and reviewers will be objective about your work.
The quality of the editorial board is an early indicator of the review process. Look at the masthead of the journal in which you plan to publish. Outstanding editors demand and get outstanding reviews. Put your energy into improving the quality of the manuscript before submission. Ideally, the reviews will improve your paper. But they will not get to imparting that advice if there are fundamental flaws.
Rule 4: If you do not write well in the English language, take lessons early; it will be invaluable later.
This is not just about grammar, but more importantly comprehension. The best papers are those in which complex ideas are expressed in a way that those who are less than immersed in the field can understand. Have you noticed that the most renowned scientists often give the most logical and simply stated yet stimulating lectures? This extends to their written work as well. Note that writing clearly is valuable, even if your ultimate career does not hinge on producing good scientific papers in English language journals. Submitted papers that are not clearly written in good English, unless the science is truly outstanding, are often rejected or at best slow to publish since they require extensive copyediting.
Rule 5: Learn to live with rejection.
A failure to be objective can make rejection harder to take, and you will be rejected. Scientific careers are full of rejection, even for the best scientists. The correct response to a paper being rejected or requiring major revision is to listen to the reviewers and respond in an objective, not subjective, manner. Reviews reflect how your paper is being judged—learn to live with it. If reviewers are unanimous about the poor quality of the paper, move on—in virtually all cases, they are right. If they request a major revision, do it and address every point they raise both in your cover letter and through obvious revisions to the text. Multiple rounds of revision are painful for all those concerned and slow the publishing process.
Rule 6: The ingredients of good science are obvious—novelty of research topic, comprehensive coverage of the relevant literature, good data, good analysis including strong statistical support, and a thought-provoking discussion. The ingredients of good science reporting are obvious—good organization, the appropriate use of tables and figures, the right length, writing to the intended audience—do not ignore the obvious.
Be objective about these ingredients when you review the first draft, and do not rely on your mentor. Get a candid opinion by having the paper read by colleagues without a vested interest in the work, including those not directly involved in the topic area.
Rule 7: Start writing the paper the day you have the idea of what questions to pursue.
Some would argue that this places too much emphasis on publishing, but it could also be argued that it helps define scope and facilitates hypothesis-driven science. The temptation of novice authors is to try to include everything they know in a paper. Your thesis is/was your kitchen sink. Your papers should be concise, and impart as much information as possible in the least number of words. Be familiar with the guide to authors and follow it, the editors and reviewers do. Maintain a good bibliographic database as you go, and read the papers in it.
Rule 8: Become a reviewer early in your career.
Reviewing other papers will help you write better papers. To start, work with your mentors; have them give you papers they are reviewing and do the first cut at the review (most mentors will be happy to do this). Then, go through the final review that gets sent in by your mentor, and where allowed, as is true of this journal, look at the reviews others have written. This will provide an important perspective on the quality of your reviews and, hopefully, allow you to see your own work in a more objective way. You will also come to understand the review process and the quality of reviews, which is an important ingredient in deciding where to send your paper.
Rule 9: Decide early on where to try to publish your paper.
This will define the form and level of detail and assumed novelty of the work you are doing. Many journals have a presubmission enquiry system available—use it. Even before the paper is written, get a sense of the novelty of the work, and whether a specific journal will be interested.
Rule 10: Quality is everything.
It is better to publish one paper in a quality journal than multiple papers in lesser journals. Increasingly, it is harder to hide the impact of your papers; tools like Google Scholar and the ISI Web of Science are being used by tenure committees and employers to define metrics for the quality of your work. It used to be that just the journal name was used as a metric. In the digital world, everyone knows if a paper has little impact. Try to publish in journals that have high impact factors; chances are your paper will have high impact, too, if accepted.
When you are long gone, your scientific legacy is, in large part, the literature you left behind and the impact it represents. I hope these ten simple rules can help you leave behind something future generations of scientists will admire.
|
16261197
|
PMC1274296
|
CC BY
|
2021-01-05 09:18:23
|
no
|
PLoS Comput Biol. 2005 Oct 28; 1(5):e57
|
utf-8
|
PLoS Comput Biol
| 2,005 |
10.1371/journal.pcbi.0010057
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2411620212610.1186/1471-2105-6-241Research ArticleA comparison of RNA folding measures Freyhult Eva [email protected] Paul P [email protected] Vincent [email protected] The Linnaeus Centre for Bioinformatics, Uppsala University, Uppsala, Sweden.2 Dept. of Evolutionary Biology, University of Copenhagen, Universitetsparken 15, 2100 Copenhagen Ø, Denmark.3 School of Computing Sciences, University of East Anglia, Norwich, NR4 7TJ, UK.2005 3 10 2005 6 241 241 3 5 2005 3 10 2005 Copyright © 2005 Freyhult et al; licensee BioMed Central Ltd.2005Freyhult et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
In the last few decades there has been a great deal of discussion concerning whether or not noncoding RNA sequences (ncRNAs) fold in a more well-defined manner than random sequences. In this paper, we investigate several existing measures for how well an RNA sequence folds, and compare the behaviour of these measures over a large range of Rfam ncRNA families. Such measures can be useful in, for example, identifying novel ncRNAs, and indicating the presence of alternate RNA foldings.
Results
Our analysis shows that ncRNAs, but not mRNAs, in general have lower minimal free energy (MFE) than random sequences with the same dinucleotide frequency. Moreover, even when the MFE is significant, many ncRNAs appear to not have a unique fold, but rather several alternative folds, at least when folded in silico. Furthermore, we find that the six investigated measures are correlated to varying degrees.
Conclusion
Due to the correlations between the different measures we find that it is sufficient to use only two of them in RNA folding studies, one to test if the sequence in question has lower energy than a random sequence with the same dinucleotide frequency (the Z-score) and the other to see if the sequence has a unique fold (the average base-pair distance, D).
==== Body
Background
Noncoding RNAs (ncRNAs) are sequences that are transcribed from DNA that function as RNA rather than being translated to protein. Many of the known ncRNAs, such as transfer RNA (tRNA), ribosomal RNA (rRNA), spliceosomal RNA (snRNA), and microRNAs (miRNA), have key functions in the cell. Moreover, various new families of ncRNAs are emerging, and, as indicated in recent studies in mouse [1] and 10 human chromosomes [2], many more transcripts are for ncRNAs than was previously expected. In the late 1980's Maizel and co-workers proposed the use of thermodynamic stability to identify noncoding RNAs in sequence data [3-5]. Since then, there has been a great deal of discussion concerning whether or not ncRNA sequences support secondary structure features that are significantly different from those of random sequences. In particular, following some contradictory results concerning the stability of messenger RNAs (mRNA) presented in [6-8], in [9] it was concluded that ncRNAs have more stable structures than random sequences, but that the difference is not significant enough to be of use in identifying novel RNAs in sequence data on its own (see also [10]). Even so, more recent findings suggest that thermodynamic stability can be used to identify novel members of special families of RNAs [11], and that stability coupled with comparative genomics data is a useful tool for identifying ncRNAs in general [12].
To shed more light on the above findings, we present a large scale investigation for how well ncRNA sequences fold compared with random sequences. In particular, we investigate six measures for how well an RNA sequence folds (normalised energy (dG), Z-score (Z) and p-value (p) of minimal free energy (MFE), Shannon entropy (Q), average base pair distance (D), and valley index (VI), for definitions see the Methods section), and compare the behaviour of these measures over a large range of Rfam ncRNA families (see Table 1), including many of the families that appeared in the studies mentioned above.
Table 1 The data sets used in this study. The first column contains a short name describing the data set, that is later used in text and figures. The RNA families/data sets can contain several types of sequences, such as the RNase family that contains both RNase P and RNase MRP. The different sequence types, or family members, are given in column two. In column three and four the number of family members (NFM) and the total number of sequences (NS) are given, respectively. The last two columns in the table give the mean and standard deviation of the sequence length and %GC-content.
short name (full name) family members NS length % GC
miRNA (microRNA) all 38 miRNAs in Rfam 135 82.82 ± 16.01 46.01 ± 7.44
intron group I and II 107 148.03 ± 113.72 43.24 ± 10.33
RNase RNase P and MRP 147 320.89 ± 37.80 56.70 ± 9.88
SRP (signal recognition particle) bacterial and eukaryotic/archae SRP 77 187.14 ± 100.47 58.27 ± 10.44
rRNA (ribosomal RNA) small subunit and 5S rRNA 578 380.06 ± 196.50 50.93 ± 8.37
snRNA (small nuclear spliceosomal RNA) all 8 spliceosomal snRNAs in Rfam 82 135.22 ± 39.40 47.19 ± 6.69
riboswitch lysine, s-box (SAM riboswitch), cobalamin 154 175.87 ± 49.59 51.38 ± 10.33
tmRNA 59 345.03 ± 32.18 45.48 ± 10.07
regulatory IRE, IRES, SECIS, HIV primer binding site, VARNA 17 80.62 ± 56.24 49.78 ± 10.21
tRNA (transfer RNA) 565 73.16 ± 5.41 46.94 ± 12.02
telomerase 17 442.53 ± 41.23 64.49 ± 6.98
snoRNA (small nucleolar RNA) all 177 guide snoRNAs in Rfam 412 97.60 ± 39.64 43.38 ± 7.43
Hh1 (Hammerhead ribozyme (type I)) 16 54.44 ± 24.08 49.49 ± 8.03
mRNA (messenger RNA) 32 329.94 ± 90.33 49.98 ± 8.47
shuffled (control data set) 130 199.86 ± 154.01 50.33 ± 10.61
Methods
Data sets
All data sets, except the protein control and the ribosomal RNA data sets, were obtained from Rfam 6.1 [13]. Rfam seed alignments were used to select a collection of RNA families, which are specified in Table 1. The rRNA data set consists of a large representative subset of the eukaryotic SSU rRNA sequences in the European rRNA database (see additional file 1 for further details).
For each class of RNA we obtained an alignment of sequences, which we filtered so that it had no more than 80% sequence identity. This was done using the program weight, that is part of the Sean Eddy "squid" utilities (downloaded 2004 from ).
In addition to the 13 data sets specified in Table 1, two control data sets were included; a protein control data set, consisting of 32 small protein coding sequences, and a set of shuffled RNA sequences (see additional file 1 for further details). The shuffled data set consists of 10 sequences from each of the 13 RNA data sets that were permuted, preserving dinucleotide frequencies [14], resulting in 130 sequences.
RNA folding statistics
Several quantities have been proposed for predicting how well an RNA molecule folds. In this paper we consider the following: The normalised minimal free energy (MFE) per base-pair (dG), the Z-score (Z), the p-value (p), the Shannon entropy (Q), the average base-pair distance (D), and the valley index (VI). We now present formal definitions for each of these measures. Let x = x1 ⋯ xL, denote an RNA sequence of length L, so that xi is either A, C, G or U for each 1 ≤ i ≤ L.
The normalised energy, dG, is arrived at from a free energy minimisation procedure. It is defined as
dG(x):=E(x)L,
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGKbazcqWGhbWrcqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceeGaa8hEaGabaiaa+LcacaGF6aGaa4xpamaalaaabaGaemyrauKaeiikaGIaa8hEaiaa+LcaaeaacqWGmbataaGaeiilaWcaaa@433B@
where E(x) is the minimal free energy (MFE) for sequence x, as computed using RNAfold [15]. This program implements the folding algorithm presented in [16].
The Z-score and the p-value compare the MFE of the sequence x to the MFEs of permuted versions of x having identical dinucleotide composition. These compositions are preserved due to the importance of stacked base-pairs in the calculation of MFE [8]. For each sequence in this study, 500 shuffled sequences were generated using a mono- and dinucleotide frequency preserving procedure implemented in the program shuffle that is part of the Sean Eddy "squid" utilities.
The Z-score [17] is the number of standard deviations by which the MFE of x deviates from the mean MFE of the set Xshuffled(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaaimaacaWFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacfiGaa43Caiaa+HgacaGF1bGaa4Nzaiaa+zgacaGFSbGaa4xzaiaa+rgaaeqaaOGaeiikaGsegyvzYrwyUfgaiyqacaqF4bGaeiykaKcaaa@4AB6@ of shuffled sequences [6,8,9,17]. It is defined as
Z(x):=E(x)−<Xshuffled(x)>σ(Xshuffled(x)),
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaacqWGAbGwcqGGOaakryGvLjhzH5wyaGqbbiaa=HhacqGGPaqkcqGG6aGocqGH9aqpdaWcaaqaaiabdweafjabcIcaOiaa=HhacqGGPaqkcqGHsislcqGH8aapimaacaGFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacgiGaa03Caiaa9HgacaqF1bGaa0Nzaiaa9zgacaqFSbGaa0xzaiaa9rgaaeqaaOGaeiikaGIaa8hEaiabcMcaPiabg6da+aqaaGGaciab8n8aZjabcIcaOiaa+HfadaWgaaWcbaGaa03Caiaa9HgacaqF1bGaa0Nzaiaa9zgacaqFSbGaa0xzaiaa9rgaaeqaaOGaeiikaGIaa8hEaiabcMcaPiabcMcaPaaacqGGSaalaaa@66BB@
where <·> and σ(·) denote the mean and the standard deviation of the MFEs of the sequences in Xshuffled(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaaimaacaWFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacfiGaa43Caiaa+HgacaGF1bGaa4Nzaiaa+zgacaGFSbGaa4xzaiaa+rgaaeqaaOGaeiikaGsegyvzYrwyUfgaiyqacaqF4bGaeiykaKcaaa@4AB6@.
The p-value of x is the fraction of sequences in Xshuffled(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaaimaacaWFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacfiGaa43Caiaa+HgacaGF1bGaa4Nzaiaa+zgacaGFSbGaa4xzaiaa+rgaaeqaaOGaeiikaGsegyvzYrwyUfgaiyqacaqF4bGaeiykaKcaaa@4AB6@ having MFE lower than x or, expressed differently, the area under the distribution function to the left of the MFE of x. It is defined as
p(x):=MN,
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGWbaCcqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceeGaa8hEaGabaiaa+LcacqGG6aGocqGH9aqpdaWcaaqaaiabd2eanbqaaiabd6eaobaacqGGSaalaaa@4061@
where M is the number of sequences in Xshuffled(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaaimaacaWFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacfiGaa43Caiaa+HgacaGF1bGaa4Nzaiaa+zgacaGFSbGaa4xzaiaa+rgaaeqaaOGaeiikaGsegyvzYrwyUfgaiyqacaqF4bGaeiykaKcaaa@4AB6@ with MFE lower than the MFE of x, and N is the number of shuffled sequences, |Xshuffled(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegm0B1jxALjhiov2DaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaaimaacaWFybWaaSbaaSqaaeHbnf2C0vMCJfMCKbacfiGaa43Caiaa+HgacaGF1bGaa4Nzaiaa+zgacaGFSbGaa4xzaiaa+rgaaeqaaOGaeiikaGsegyvzYrwyUfgaiyqacaqF4bGaeiykaKcaaa@4AB6@|.
In vivo, RNAs commonly exist in an ensemble of structures. The distribution of these structures can be modelled by a Boltzmann distribution. Using this setup, it is possible to efficiently compute the partition function, Z, for the ensemble S(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbbjxAHXgaiqaacaWFtbGaeiikaGsegyvzYrwyUfgaiuqacaGF4bGaeiykaKcaaa@3CC3@ of secondary structures corresponding to an RNA sequence x [18]. In particular, the probability of a structure Sα ∊ S(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbbjxAHXgaiqaacaWFtbGaeiikaGsegyvzYrwyUfgaiuqacaGF4bGaeiykaKcaaa@3CC3@ (which we regard as a set of base-pairs) is given by P(Sα)=e−Eα/RTZ
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaacciGae8xSdegabeaakiabcMcaPiabg2da9maalaaabaGaemyzau2aaWbaaSqabeaacqGHsislcqWGfbqrdaWgaaadbaGae8xSdegabeaaliabc+caViabdkfasjabdsfaubaaaOqaaiabdQfaAbaaaaa@3D85@, where Z=∑Sα∈S(x)e−Eα/RT
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamrtHrhAL1wy0L2yHvtyaeHbnfgDOvwBHrxAJfwnaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaacqWGAbGwcqGH9aqpdaaeqaqaaiabdwgaLnaaCaaaleqabaGaeyOeI0Iaemyrau0aaSbaaWqaaGGaciab=f7aHbqabaWccqGGVaWlcqWGsbGucqWGubavaaaabaGaem4uam1aaSbaaWqaaiab=f7aHbqabaWccqGHiiIZt0uy0HwzTfgDPnwy3aqeh0uy0HwzTfgDPnwy3aacfaGae4NKWpLaeiikaGYexLMBbXgBcf2CPn2qVrwzqf2zLnharGGvLjhzH5wyaGGbbiaa9HhacqGGPaqkaeqaniabggHiLdaaaa@6095@
, Eα is the free energy of Sα, R = 8.31451 Jmol-1K-1 is the molar gas constant, and T is the temperature, which we take as 310.15 K (37°C). The base-pair probability pij (the probability that xi pairs with xj) is then given by pij=∑Sα∈S(x)P(Sα)δijα
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamrtHrhAL1wy0L2yHvtyaeHbnfgDOvwBHrxAJfwnaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakeaacqWGWbaCdaWgaaWcbaGaemyAaKMaemOAaOgabeaakiabg2da9maaqababaGaemiuaaLaeiikaGIaem4uam1aaSbaaSqaaGGaciab=f7aHbqabaGccqGGPaqkcqWF0oazdaqhaaWcbaGaemyAaKMaemOAaOgabaGae8xSdegaaaqaaiabdofatnaaBaaameaacqWFXoqyaeqaaSGaeyicI48enfgDOvwBHrxAJf2naeXbnfgDOvwBHrxAJf2naGqbaiab+jj8tjab+bW9OiabcIcaOmXvP5wqSXMqHnxAJn0BKvguHDwzZbqeiyvzYrwyUfgaiyqacaqF4bGaeiykaKcabeqdcqGHris5aaaa@68B5@
is 1 if xi and xj is a base-pair in Sα, and 0 otherwise.
We use the implementation of McCaskill's algorithm in RNAfold to compute base-pair probabilities. The normalised Shannon entropy of x [19] is then defined as
Q(x):=−∑i<jpijlog2(pij)L.
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGrbqucqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceeGaa8hEaGabaiaa+LcacqGG6aGocqGH9aqpdaWcaaqaaiabgkHiTmaaqababaGaemiCaa3aaSbaaSqaaiabdMgaPjabdQgaQbqabaGccyGGSbaBcqGGVbWBcqGGNbWzdaWgaaWcbaGaeGOmaidabeaakiabcIcaOiabdchaWnaaBaaaleaacqWGPbqAcqWGQbGAaeqaaOGaeiykaKcaleaacqWGPbqAcqGH8aapcqWGQbGAaeqaniabggHiLdaakeaacqWGmbataaGaeiOla4caaa@553C@
We can also use base-pair probabilities to compute the average base pair distance between all structures in S(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbbjxAHXgaiqaacaWFtbGaeiikaGsegyvzYrwyUfgaiuqacaGF4bGaeiykaKcaaa@3CC3@, <dBP> as follows (I.Hofacker, pers. commun.). (Version 1.5beta of RNAfold output this measure as "ensemble diversity".) The base-pair distance, dBP(Sα, Sβ) between two structures Sα and Sβ on x is defined as the number of base-pairs not shared by the structures Sα and Sβ (see e.g. [20]). Hence, if |Sα| is the number of base-pairs in Sα, i.e. |Sα|= ∑i<jδijα'
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqGG8baFcqWGtbWudaWgaaWcbaacciGae8xSdegabeaakiabcYha8jabb2da9iaaykW7daaeqaqaaiab=r7aKnaaDaaaleaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceiGaa4xAaiaa+PgaaeaacqWFXoqycaGFNaaaaaqaaiabdMgaPjabgYda8iabdQgaQbqab0GaeyyeIuoaaaa@4B45@
, where i and j lie between 1 and L, then the base-pair distance between structures Sα and Sβ equals
dBP(Sα,Sβ)= |Sα∪Sβ|−|Sα∩Sβ|=|Sα|+|Sβ|−2|Sα∩Sβ|=∑i<j(δijα+ δij−2δijδij).
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakqaaeeqaaiabdsgaKnaaBaaaleaacqWGcbGqcqWGqbauaeqaaOGaeiikaGIaem4uam1aaSbaaSqaaGGaciab=f7aHbqabaGccqGGSaalcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPiabg2da9iaaykW7caaMc8UaeiiFaWNaem4uam1aaSbaaSqaaiab=f7aHbqabaGccqWIQisvcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcYha8jabgkHiTiabcYha8jabdofatnaaBaaaleaacqWFXoqyaeqaaOGaeSykIKKaem4uam1aaSbaaSqaaiab=j7aIbqabaGccqGG8baFcqqG9aqpcqqG8baFcqWGtbWudaWgaaWcbaGae8xSdegabeaakiabcYha8jabgUcaRiabcYha8jabdofatnaaBaaaleaacqWFYoGyaeqaaOGaeiiFaWNaeyOeI0IaeGOmaiJaeiiFaWNaem4uam1aaSbaaSqaaiab=f7aHbqabaGccqWIPisscqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcYha8bqaaiabg2da9maaqafabaGaeiikaGIae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaaiab=f7aHbaakiabgUcaRaWcbaGaemyAaKMaeyipaWJaemOAaOgabeqdcqGHris5aOGaaGPaVlaaykW7cqWF0oazdaqhaaWcbaGaemyAaKMaemOAaOgabaWaaSbaaWqaaiab=j7aIbqabaaaaOGaeyOeI0IaeGOmaiJae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaamaaBaaameaacqWFXoqyaeqaaaaakiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaadaWgaaadbaGae8NSdigabeaaaaGccqGGPaqkcqGGUaGlaaaa@94E2@
In particular,
<dBP>=12∑Sα,Sβ∈S(x)[P(Sα)P(Sβ)∑i<j(δijα+δijβ−2δijαδijβ)].
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqGH8aapcqWGKbazdaWgaaWcbaGaemOqaiKaemiuaafabeaakiabg6da+iabg2da9maalaaabaGaeGymaedabaGaeGOmaidaamaaqafabaGaei4waSLaemiuaaLaeiikaGIaem4uam1aaSbaaSqaaGGaciab=f7aHbqabaGccqGGPaqkcqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPmaaqafabaGaeiikaGIae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaaiab=f7aHbaakiabgUcaRiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqWFYoGyaaGccqGHsislcqaIYaGmcqWF0oazdaqhaaWcbaGaemyAaKMaemOAaOgabaGae8xSdegaaOGae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaaiab=j7aIbaaaeaacqWGPbqAcqGH8aapcqWGQbGAaeqaniabggHiLdaaleaacqWGtbWudaWgaaadbaGae8xSdegabeaaliabcYcaSiabdofatnaaBaaameaacqWFYoGyaeqaaSGaeyicI48exLMBbXgBcf2CPn2qVrwzqf2zLnharyqqYLwySbaceaGaa43uaiabcIcaOeHbwvMCKfMBHbacfeGaa0hEaiabcMcaPaqab0GaeyyeIuoakiabcMcaPiabc2faDjabc6caUaaa@8222@
Since |Sα|= ∑i<jδijα'
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqGG8baFcqWGtbWudaWgaaWcbaacciGae8xSdegabeaakiabcYha8jabb2da9iaaykW7daaeqaqaaiab=r7aKnaaDaaaleaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceiGaa4xAaiaa+PgaaeaacqWFXoqycaGFNaaaaaqaaiabdMgaPjabgYda8iabdQgaQbqab0GaeyyeIuoaaaa@4B45@
, <dBP> can thus be rewritten as
<dBP> = 12∑Sα,Sβ∈S(x)[P(Sα)P(Sβ)∑i<j(δijα+δijβ−2δijαδijβ)] =12∑i<j[∑Sα∈S(x)P(Sα)δijα︸pij∑Sβ∈S(x)P(Sβ)︸1+∑Sα∈S(x)P(Sα)︸1∑Sβ∈S(x)P(Sβ)δijβ︸pij −2∑Sα∈S(x)P(Sα)δijα︸pij∑Sβ∈S(x)P(Sβ)δijβ]︸pij =12∑i<j[pij+pij+2pijpij] =∑i<j(pij−pij2).
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamrtHrhAL1wy0L2yHvtyaeHbnfgDOvwBHrxAJfwnaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaWaaeGaeaaakqaabeqaaiabgYda8iabdsgaKnaaBaaaleaacqWGcbGqcqWGqbauaeqaaOGaeyOpa4JaaGPaVlaaykW7cqGH9aqpcaaMc8UaaGPaVpaalaaabaGaeGymaedabaGaeGOmaidaamaaqafabaGaei4waSLaemiuaaLaeiikaGIaem4uam1aaSbaaSqaaGGaciab=f7aHbqabaGccqGGPaqkcqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPmaaqafabaGaeiikaGIae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaaiab=f7aHbaakiabgUcaRiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqWFYoGyaaGccqGHsislcqaIYaGmcqWF0oazdaqhaaWcbaGaemyAaKMaemOAaOgabaGae8xSdegaaOGae8hTdq2aa0baaSqaaiabdMgaPjabdQgaQbqaaiab=j7aIbaakiabcMcaPiabc2faDbWcbaGaemyAaKMaeyipaWJaemOAaOgabeqdcqGHris5aaWcbaGaem4uam1aaSbaaWqaaiab=f7aHbqabaWccqGGSaalcqWGtbWudaWgaaadbaGae8NSdigabeaaliabgIGioprtHrhAL1wy0L2yHDdarCqtHrhAL1wy0L2yHDdaiuaacqGFsc=ucqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaebcwvMCKfMBHbacgeGaa0hEaiabcMcaPaqab0GaeyyeIuoaaOqaaiaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7cqGH9aqpdaWcaaqaaiabigdaXaqaaiabikdaYaaadaaeqbqaaiabcUfaBnaayaaabaWaaabuaeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8xSdegabeaakiabcMcaPiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqWFXoqyaaaabaGaem4uam1aaSbaaWqaaiab=f7aHbqabaWccqGHiiIZcqGFsc=ucqGGOaakcaqF4bGaeiykaKcabeqdcqGHris5aaWcbaGaemiCaa3aaSbaaWqaaiabdMgaPjabdQgaQbqabaaakiaawIJ=aaWcbaGaemyAaKMaeyipaWJaemOAaOgabeqdcqGHris5aOWaaGbaaeaadaaeqbqaaiabdcfaqjabcIcaOiabdofatnaaBaaaleaacqWFYoGyaeqaaOGaeiykaKcaleaacqWGtbWudaWgaaadbaGae8NSdigabeaaliabgIGiolab+jj8tjabcIcaOiaa9HhacqGGPaqkaeqaniabggHiLdaaleaaiyaacaaFXaaakiaawIJ=aiabgUcaRmaayaaabaWaaabuaeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8xSdegabeaakiabcMcaPaWcbaGaem4uam1aaSbaaWqaaiab=f7aHbqabaWccqGHiiIZcqGFsc=ucqGFaCpkcqGGOaakcaqF4bGaeiykaKcabeqdcqGHris5aaWcbaGaeGymaedakiaawIJ=amaayaaabaWaaabuaeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqWFYoGyaaaabaGaem4uam1aaSbaaWqaaiab=j7aIbqabaWccqGHiiIZcqGFsc=ucqGGOaakcaqF4bGaeiykaKcabeqdcqGHris5aaWcbaGaemiCaa3aaSbaaWqaaiabdMgaPjabdQgaQbqabaaakiaawIJ=aaqaaiaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlaaykW7caaMc8UaaGPaVlabgkHiTiabikdaYmaayaaabaWaaabuaeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8xSdegabeaakiabcMcaPiab=r7aKnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqWFXoqyaaaabaGaem4uam1aaSbaaWqaaiab=f7aHbqabaWccqGHiiIZcqGFsc=ucqGGOaakcaqF4bGaeiykaKcabeqdcqGHris5aaWcbaGaemiCaa3aaSbaaWqaaiabdMgaPjabdQgaQbqabaaakiaawIJ=amaayaaabaWaaabuaeaacqWGqbaucqGGOaakcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPiab=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@FAB4@
Thus normalising by length, the average base-pair distance is given by
D(x):=∑i<j(pij−pij2)L.
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGebarcqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceeGaa8hEaGabaiaa+LcacaGF6aGaa4xpamaalaaabaWaaabeaeaacqGGOaakcqWGWbaCdaWgaaWcbaGaemyAaKMaemOAaOgabeaakiabgkHiTiabdchaWnaaDaaaleaacqWGPbqAcqWGQbGAaeaacqaIYaGmaaGccqGGPaqkaSqaaiabdMgaPjabgYda8iabdQgaQbqab0GaeyyeIuoaaOqaaiabdYeambaacqGGUaGlaaa@5044@
The last measure that we consider in this study is the valley index (VI) [21]. It can be regarded as an approximation to D (see below), and is meant to measure the number of "valleys" in the RNA folding landscape of x.
Formally it is defined as follows: List the suboptimal structures of x according to their free energies so that Sopt, an MFE structure for x, is first and S1,..., Sn are the next n structures on x with Eopt ≤ E1 ≤ ⋯ ≤ En. Put Ssubopt = {Sopt, S1,..., Sn}, and define
VI(x):=∑Sα,Sβ∈SsuboptdBPnorm(Sα,Sβ)w(α)w(β)∑Sα,Sβ∈Ssuboptw(α)w(β),
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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nfadaWgaaadbaGaem4CamNaemyDauNaemOyaiMaem4Ba8MaemiCaaNaemiDaqhabeaaaSqab0GaeyyeIuoaaaGccqGGSaalaaa@8B0A@
where w(α)=e−(Eα−Eopt)/RT
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWG3bWDcqGGOaakiiGacqWFXoqycqGGPaqkcqGH9aqpcqWGLbqzdaahaaWcbeqaaiabgkHiTiabcIcaOiabdweafnaaBaaameaacqWFXoqyaeqaaSGaeyOeI0Iaemyrau0aaSbaaWqaaiabd+gaVjabdchaWjabdsha0bqabaWccqGGPaqkcqGGVaWlcqWGsbGucqWGubavaaaaaa@4342@ is the Boltzmann factor, and
dBPnorm(Sα,Sβ):=dBP(Sα,Sβ)L.
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGKbazdaWgaaWcbaGaemOqaiKaemiuaaLaemOBa4Maem4Ba8MaemOCaiNaemyBa0gabeaakiabcIcaOiabdofatnaaBaaaleaaiiGacqWFXoqyaeqaaOGaeiilaWIaem4uam1aaSbaaSqaaiab=j7aIbqabaGccqGGPaqkcqGG6aGocqGH9aqpdaWcaaqaaiabdsgaKnaaBaaaleaacqWGcbGqcqWGqbauaeqaaOGaeiikaGIaem4uam1aaSbaaSqaaiab=f7aHbqabaGccqGGSaalcqWGtbWudaWgaaWcbaGae8NSdigabeaakiabcMcaPaqaaiabdYeambaacqGGUaGlaaa@4F09@
Note that our definition of VI differs slightly from the Kitagawa et al.'s definition since we use normalised base-pair distance, dBPnorm, rather than the coarse-grained tree metric in their study. The suboptimal structures S1,..., Sn are randomly sampled with probabilities equal to their Boltzmann weights using the program RNAsubopt [22]. We sample 300 structures resulting in between 16 (regulatory) and 300 (telomerase) unique structures.
In principle, the valley index for an RNA with a low number of valleys in the folding landscape should be low, whereas an RNA with a multi-valley folding landscape should have a correspondingly higher index. Note that the sums in the definition on VI are taken over all structures in a set of suboptimal structures within a certain energy distance from the MFE. If the energy distance is increased this set of structures will eventually include all the sequences in the ensemble S(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbbjxAHXgaiqaacaWFtbGaeiikaGsegyvzYrwyUfgaiuqacaGF4bGaeiykaKcaaa@3CC3@. In this situation, in view of the definition of w(α) it follows that the valley index of x can be rewritten as
VI(x)=∑Sα,Sβ∈S(x)dBPnorm(Sα,Sβ)(e−(Eα−Eopt)/RT)(e−(Eβ−Eopt)/RT)∑Sα,Sβ∈S(x)(e−(Eα−Eopt)/RT))(e−(Eβ−Eopt)/RT)=∑Sα,Sβ∈S(x)dBPnorm(Sα,Sβ)(e−Eα/RT)(e−Eβ/RT)∑Sα,Sβ∈S(x)(e−Eα/RT)∑Sβ∈S(x)(e−Eβ/RT)=∑Sα,Sβ∈S(x)dBPnorm(sα,Sβ)(e−Eα/RT)(e−Eβ/RT)Z2=∑Sα,Sβ∈S(x)∑P(Sα)P(Sβ)dBPnorm(Sα,Sβ)=2D(x).
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamrtHrhAL1wy0L2yHvtyaeHbnfgDOvwBHrxAJfwnaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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HhacqGGPaqkaeqaniabggHiLdaakeaadaaeqaqaaaWcbaGaem4uam1aaSbaaWqaaiab+f7aHbqabaWccqGGSaalcqWGtbWudaWgaaadbaGae4NSdigabeaaliabgIGiolab9jj8tjabcIcaOiaa=HhacqGGPaqkaeqaniabggHiLdGccqGGOaakcqWGLbqzdaahaaWcbeqaaiabgkHiTiabdweafnaaBaaameaacqGFXoqyaeqaaSGaei4la8IaemOuaiLaemivaqfaaOGaeiykaKYaaabeaeaacqGGOaakcqWGLbqzdaahaaWcbeqaaiabgkHiTiabdweafnaaBaaameaacqGFYoGyaeqaaSGaei4la8IaemOuaiLaemivaqfaaOGaeiykaKcaleaacqWGtbWudaWgaaadbaGae4NSdigabeaaliabgIGiolab9jj8tjabcIcaOiaa=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HhacqGGPaqkaSqaaiabdofatnaaBaaameaacqGFXoqyaeqaaSGaeiilaWIaem4uam1aaSbaaWqaaiab+j7aIbqabaWccqGHiiIZcqqFsc=ucqGGOaakcaWF4bGaeiykaKcabeqdcqGHris5aOGaeiOla4caaaa@A28E@
Thus, 12VI(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaadaWcaaqaaiabigdaXaqaaiabikdaYaaacqWGwbGvcqWGjbqscqGGOaaktCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbwvMCKfMBHbaceeGaa8hEaiabcMcaPaaa@3E25@ can be thought of as an approximation of D(x) in case the set Ssubopt (x) used in the computation of VI(x) is a proper subset of S(x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaeHbbjxAHXgaiqaacaWFtbGaeiikaGsegyvzYrwyUfgaiuqacaGF4bGaeiykaKcaaa@3CC3@.
Results and discussion
Comparison of measures
The six measures that we investigated are correlated to varying degrees; see Table 2 and Figure 1. The measures Q and D are highly correlated (correlation coefficient = 0.98), which could be due to the fact that they are both computed using McCaskill base pair probabilities, pij. Also, as expected, the Z-score and p-value are strongly correlated, but not in a linear fashion (see Figure 1). We see that the Z-scores are more sensitive for low values than the p-values (e.g. all Z-scores below -3 correspond to a p-value of 0.0), and so Z-scores are more informative.
Table 2 Correlations between measures. Correlation coefficients between the different measures, values above 0.5 are in bold.
dG 1.00
Z 0.62 1.00
p 0.48 0.74 1.00
Q 0.33 0.52 0.51 1.00
D 0.31 0.51 0.48 0.98 1.00
VI 0.19 0.23 0.19 0.29 0.33 1.00
length -0.26 -0.19 -0.10 0.32 0.28 -0.17 1.00
%GC -0.78 -0.13 -0.12 -0.08 -0.06 -0.05 0.18 1.00
G/C ratio 0.03 -0.05 -0.02 0.01 0.01 0.06 0.00 -0.14 1.00
dG Z p Q D VI length %GC G/C ratio
Figure 1 Correlations between measures. Correlations between all the different measures for all the data sets are shown. The diagonal figures show the distributions of the measures.
The statistic dG is weakly correlated to all other measures. However, it is interesting to note that dG is negatively correlated to %-GC. This is to be expected since GC base pairs have lower energy than the other possible base-pairings. The miRNA family is an exception to this rule, since it has low dG values, but an average %-GC of about 50%, see Figure 1.
Table 2 shows that the correlation between VI and the other measures is low over all families. However, Figure 1 indicates that for a subset of all the sequences the correlation between VI and Q or D is very strong. This is also confirmed by computing the correlation coefficients for the 15 RNA families separately. miRNA, SRP, tRNA, telomerase, and Hh1 show strong correlations (> 0.65) between VI and Q or D, whereas the corresponding correlations for rRNA, snRNA, riboswitch, regulatory, and snoRNA are weak (< 0.3).
Comparison between RNA families
In general, we deem an RNA sequence to have a stable secondary structure if the measures dG, Z, and p are significantly lower than the corresponding values for the shuffled control data sets. To check whether this was the case for the different data sets, we applied a Mann-Whitney rank sum test [23]. This test compares two data sets and computes the probability that the two data sets are sampled from the same distribution. Unlike the t-test, the Mann-Whitney test is distribution free since it compares the ranks of the data values instead of the data values themselves.
At a significance level of 99% the Mann-Whitney test indicated that Z and p are higher for the shuffled data set than for any of the real RNA data sets, except for mRNA and Hh1. The same held for the normalised energy dG, except for the tmRNA, tRNA, regulatory and snoRNA families. This result agrees with those observed in [10], that ncRNAs have significantly lower Z-score than unstructured sequences. This can also be seen in Figure 2.
Figure 2 Box and whisker plots of dG, Z, p, Q, D, and VI. Box and whisker plots displaying medians, quartiles and range of the measures dG, Z, p, Q, D, and VI. The lines of the box are at the lower quartile, median, and upper quartile values. The box width is proportional to the number of sequences in the data set. The whisker lines extend from each end of the box to the most extreme data value or have a maximal length of 1.5 times the box height. Data points beyond the ends of the whiskers are marked by +.
The measures Q and D can be used to indicate whether a sequence folds into a unique secondary structure or into several alternative structures [24]. The riboswitch data set consists of sequences known to have alternative structures, and so we expected the values of Q and D to be rather high for this data set. We did find this to be the case, but surprisingly they were also as high or even higher for other data sets (see Figure 2).
The high values of Q and D obtained for the mRNA and shuffled data sets is probably due to the fact that these RNAs are unstructured, and hence there are many alternative possible structures. This could also explain the values of Q and D for tmRNA, since tmRNAs are to a large extent mRNA-like (large parts of such molecules are unstructured). Other RNA families like tRNA and RNAse have tertiary interactions that aren't included in secondary structure, which explains their relatively high Q- and D-values. The interaction of rRNAs and snoRNAs with proteins and other RNAs most likely stabilise their native structures, even though alternative structures are possible.
The values of our measures for the telomerase sequences were unexpected. Telomerase has low energy per base, yet it has a rather high Z-score compared to the other ncRNAs. The high stability of this molecule is most likely due to an unusual sequence composition; the telomerase sequences have a high %-GC level, 65% (see Figure 3). The high values of Q and D suggest that the telomerase sequences have alternative structures.
Figure 3 Box and whisker plots of length, %GC, and G/C ratio. Box and whisker plots displaying medians, quartiles and range of the sequence length, %GC, and G/C ratio for all our test data sets. The lines of the box are at the lower quartile, median, and upper quartile values. The box width is proportional to the number of sequences in the data set. The whisker lines extend from each end of the box to the most extreme data value or have a maximal length of 1.5 times the box height. Data points beyond the ends of the whiskers are marked by +.
The miRNAs have very stable structures, indicated by low Z and dG, especially in view of their %GC level (~50%). This has previously been observed in [11]. The miRNAs also have low values of Q, D, and VI, indicating a unique structure.
Comparison with previous studies
Seffens and Digby [6] examined 51 mRNA sequences and observed that they have lower folding energy than shuffled versions of the sequences preserving mono- but not dinucleotide frequencies. Shortly after, Workman and Krogh examined 46 of the 51 mRNAs and showed that they do not have lower folding energy than shuffled versions of the sequences, when the dinucleotide frequencies are preserved [8]. In our study, in which sequences were shuffled so as to preserve both mono- and dinucleotide frequencies, we confirm that mRNAs do not have lower folding energy than shuffled sequences. In [8] a small sample of rRNA and tRNA sequences were also investigated and it was indicated that rRNA, but not tRNA has lower folding energy than dinucleotide shuffled sequences. Our study, with significantly more data, agrees with their findings for rRNA, but differs for tRNAs, which we found to have significantly lower Z-scores than shuffled sequences. Rivas and Eddy [9] argue that secondary structure alone is generally not significant for the detection of ncRNA, but note that ncRNAs have slightly lower folding energies than shuffled sequences. Note that in [9] sequences are shuffled preserving mononucleotides only, whereas in our study we shuffled sequences preserving dinucleotide frequencies. Rivas and Eddy computed Z-scores for a large set of tRNAs, and even though we adopt a different shuffling procedure, our results for tRNA are in good agreement with Rivas and Eddy's findings.
Kitagawa et al. [21] observed that five snRNAs have low folding energies compared to shuffled sequences. Our studies confirm this observation, and in general we found that snRNA sequences have lower folding energies than shuffled sequences with the same dinucleotide frequency. Kitagawa et al. also computed VI values for the same five snRNAs, and observed that the values varied considerably (indicating that some have uni-valley landscapes while other have multi-valley landscapes). Although we used a variant of VI, we also found that the VI value varies considerably for different snRNA sequences.
Bonnet et al. observed that miRNAs have considerably lower folding energy than dinucleotide shuffled sequences, unlike tRNA and rRNA [11]. Our studies confirm this observation, although Bonnet et al. investigated shorter regions of the rRNA, while we investigated full rRNA sequences.
In our study, we found the mean Z-scores (and p-values) to be significantly lower for ncRNAs (except the Hammerhead type I family) than for the shuffled sequences (although the Z-scores for mRNA were not lower). This is in agreement with recent results presented in [10], where it is shown that non-coding RNAs have lower Z-scores than coding RNAs for a selection of RNA families (tRNA, Hammerhead type III, a regulatory element (SECIS), SRP, snRNA (U1 and U2), mRNA (divided into coding sequence and 5'- and 3'-untranslated regions)).
Conclusion
We have studied six previously defined measures for predicting how well an RNA molecule is expected to fold (dG, Z, p, Q, D, and VI), and applied them to a large collection of RNAs from the Rfam database. We found all of these measures to be correlated to some degree. The measures Z and p are strongly correlated, but Z is more sensitive than p. Since dG is a measure of MFE it is strongly correlated to the nucleotide composition of the sequence, and so a low dG does not necessarily imply a stable structure. Hence, it is probably sufficient to use Z as opposed to p and dG. For the families that we used in this study, we found the mean Z-scores (and p-values) to be significantly lower for ncRNAs than for the shuffled sequences.
The three measures Q, D and VI can be regarded as measures of the ruggedness of the RNA folding landscape. Both Q and D are computed from the partition function and are thus strongly correlated, and so either of them is probably sufficient for measuring ruggedness. The valley index VI can be viewed as an approximation of the average base-pair distance D (see Methods section), and so there is no advantage in computing VI, especially since it is slow to compute, whereas D can be computed efficiently. RNA families having high values of D (and Q) were either unstructured RNA sequences, long RNA sequences that fold with the help of proteins, or RNAs with alternative folds or pseudoknot structures.
Thus, in summary, we expect that rather than using all of dG, Z, p, Q, D, and VI to predict how well an RNA molecule folds, that it is sufficient to use only Z and D (or Q). Our studies suggest that a combination of Z-score and D value might be useful for identifying well-defined RNA structures, such as the miRNAs (in agreement with results presented in [11]), and, based on our results, we expect that variations of these measures (such as the alignment Z-scores described in [12]), will provide a useful tool for the general problem of RNA structure identification.
Authors' contributions
EF was involved in selecting the data sets from Rfam and implementing the analyses. PPG developed the ideas presented in the paper and was involved in selecting the data sets from Rfam and implementing the analyses. VM was involved in developing the ideas presented in the paper. All authors contributed to the writing of this manuscript. All authors read and approved the final manuscript.
Supplementary Material
Additional File 1
Data sets. This zip-file contains all the sequences we have used for this study.
Click here for file
Acknowledgements
The authors thank Ivo Hofacker for documentation regarding the average base pair distance and Peter Clote for providing access to his work [10] before publication. PPG was supported by a Carlsberg Foundation Grant (21-00-0680).
==== Refs
Suzuki M Hayashizaki Y Mouse-centric comparative transcriptomics of protein coding and non-coding RNAs Bioessays 2004 26 833 843 15273986 10.1002/bies.20084
Cheng J Kapranov P Drenkow J Dike S Brubaker S Patel S Long J Stern D Tammana H Helt G Sementchenko V Piccolboni A Bekiranov S Bailey DK Ganesh M Ghosh S Bell I Gerhard D Gingeras T Transcriptional Maps of 10 Human Chromosomes at 5-Nucleotide Resolution Science 2005 308 1149 54 15790807 10.1126/science.1108625
Le S Chen J Currey K Maizel JJV A program for predicting significant RNA secondary structures Comput Appl Biosci 1988 4 153 159 2454711
Le SY Chen JH Maizel JV Thermodynamic stability and statistical significance of potential stem-loop structures situated at the frameshift sites of retroviruses Nucleic Acids Res 1989 17 6143 6152 2549508
Chen J Le S Currey K Maizel J A computational procedure for assessing the significance of RNA secondary structure Comput Appl Biosci 1990 6 7 18 1690072
Seffens W Digby D mRNAs have greater negative folding free energies than shuffled or codon choice randomized sequences Nucleic Acids Res 1999 27 1578 1584 10075987 10.1093/nar/27.7.1578
Schultes EA Hraber PT LaBean TH Estimating the contributions of selection and self-organization in RNA secondary structure J Mol Evol 1999 49 76 83 10368436
Workman C Krogh A No evidence that mRNAs have lower folding free energies than random sequences with the same dinucleotide distribution Nucleic Acids Res 1999 27 4816 4822 10572183 10.1093/nar/27.24.4816
Rivas E Eddy SR Secondary structure alone is generally not statistically significant for the detection of noncoding RNAs Bioinformatics 2000 16 583 605 11038329 10.1093/bioinformatics/16.7.583
Clote P Ferré F Kranakis E Krizanc D Structural RNA has lower folding energy than random RNA of the same dinucleotide frequency RNA 2005 15840812
Bonnet E Wuyts J Rouze P Van de Peer Y Evidence that microRNA precursors, unlike other non-coding RNAs, have lower folding free energies than random sequences Bioinformatics 2004 20 2911 2917 15217813 10.1093/bioinformatics/bth374
Washietl S Hofacker IL Stadler PF Fast and reliable prediction of noncoding RNAs Proc Natl Acad Sci USA 2005 102 2454 2459 15665081 10.1073/pnas.0409169102
Griffiths-Jones S Bateman A Marshall M Khanna A Eddy SR Rfam: an RNA family database Nucleic Acids Res 2003 31 439 441 12520045 10.1093/nar/gkg006
Altschul SF Erickson BW Significance of nucleotide sequence alignments: a method for random sequence permutation that preserves dinucleotide and codon usage Mol Biol Evol 1985 2 526 538 3870875
Hofacker IL Fontana W Stadler PF Bonhoeffer LS Tacker M Schuster P Fast folding and comparison of RNA secondary structures Monatshefte für Chemie 1994 125 167 188 10.1007/BF00818163
Zuker M Stiegler P Optimal computer folding of large RNA sequences using thermodynamics and auxiliary information Nucleic Acids Res 1981 9 133 148 6163133
Le SY Maizel JV Jr A method for assessing the statistical significance of RNA folding J Theor Biol 1989 138 495 510 2480496
McCaskill JS The equilibrium partition function and base pair binding probabilities for RNA secondary structures Biopolymers 1990 29 1105 1119 1695107 10.1002/bip.360290621
Huynen M Gutell R Konings D Assessing the reliability of RNA folding using statistical mechanics J Mol Biol 1997 267 1104 1112 9150399 10.1006/jmbi.1997.0889
Moulton V Zuker M Steel M Pointon R Penny D Metrics on RNA secondary structures J Comput Biol 2000 7 277 292 10890402 10.1089/10665270050081522
Kitagawa J Futamura Y Yamamoto K Analysis of the conformational energy landscape of human snRNA with a metric based on tree representation of RNA structures Nucleic Acids Res 2003 31 2006 2013 12655018 10.1093/nar/gkg288
Wuchty S Fontana W Hofacker IL Schuster P Complete suboptimal folding of RNA and the stability of secondary structures Biopolymers 1999 49 145 165 10070264 10.1002/(SICI)1097-0282(199902)49:2<145::AID-BIP4>3.0.CO;2-G
Mann H Whitney D On a test whether one of two random variables is stochastically larger than the other Ann Math Statist 1947 18 50 60
Mathews DH Using an RNA secondary structure partition function to determine confidence in base pairs predicted by free energy minimization RNA 2004 10 1178 1190 15272118 10.1261/rna.7650904
|
16202126
|
PMC1274297
|
CC BY
|
2021-01-04 16:27:46
|
no
|
BMC Bioinformatics. 2005 Oct 3; 6:241
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-241
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2511622566710.1186/1471-2105-6-251Research ArticleDifferentiation of regions with atypical oligonucleotide composition in bacterial genomes Reva Oleg N [email protected]ümmler Burkhard [email protected] Klinische Forschergruppe, OE6711, Medizinische Hochschule Hannover, Carl-Neuberg-Strasse 1, D-30625 Hannover, Germany2 Danylo Zabolotny Institute of Microbiology and Virology of the National Academy of Science of Ukraine, Dep. of Antibiotics, 154 Zabolotnogo Str., D03680, Kyiv GSP, Ukraine2005 14 10 2005 6 251 251 7 6 2005 14 10 2005 Copyright © 2005 Reva and Tümmler; licensee BioMed Central Ltd.2005Reva and Tümmler; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Complete sequencing of bacterial genomes has become a common technique of present day microbiology. Thereafter, data mining in the complete sequence is an essential step. New in silico methods are needed that rapidly identify the major features of genome organization and facilitate the prediction of the functional class of ORFs. We tested the usefulness of local oligonucleotide usage (OU) patterns to recognize and differentiate types of atypical oligonucleotide composition in DNA sequences of bacterial genomes.
Results
A total of 163 bacterial genomes of eubacteria and archaea published in the NCBI database were analyzed. Local OU patterns exhibit substantial intrachromosomal variation in bacteria. Loci with alternative OU patterns were parts of horizontally acquired gene islands or ancient regions such as genes for ribosomal proteins and RNAs. OU statistical parameters, such as local pattern deviation (D), pattern skew (PS) and OU variance (OUV) enabled the detection and visualization of gene islands of different functional classes.
Conclusion
A set of approaches has been designed for the statistical analysis of nucleotide sequences of bacterial genomes. These methods are useful for the visualization and differentiation of regions with atypical oligonucleotide composition prior to or accompanying gene annotation.
==== Body
Background
The number of sequenced prokaryotic genomes increases rapidly each year. Their comprehensive analysis requires the development of new high-throughput computational methods. The analysis of oligonucleotide usage biases has been recognized to be practical for the recognition of pathogenicity islands [1,2] and elucidation of origins of orphan sequences [3-5]. Recently we have developed methods for the global analysis of oligonucleotide usage (OU) in complete sequences of bacterial chromosomes, plasmids and phages [6]. The patterns of deviations of oligonucleotide frequencies from expectations were shown to be genome signatures reflecting to some extent the phylogenetic links between microorganisms [3,4,7,8].
The usage of oligonucleotides in bacterial sequences is not random. Frequencies of the oligonucleotide words (further – words) depend strongly on their physicochemical properties such as base stacking energy, propeller twist angle, bendability, position preference and protein deformability [6]. Oligonucleotide usage in bacterial genomes is strongly influenced by codon usage [9], however, there are further, yet unknown mechanisms of word selection [10].
To characterize OU in a sequence, the concept of OU patterns has been introduced [6]. Disparity of frequencies of words and their reverse complements termed as pattern skew (PS) and variance of oligonucleotide frequencies (OUV) are attributes of each OU pattern and the distance (D) expresses the difference between two OU patterns. These OU parameters are independent of the length of the sequence and hence allow the comparison of windows of different sequence length ([6] and see 'Materials and methods'). This study applied OU statistics to visualize and discern gene islands of different functional classes. The developed methods are of importance for structural, functional and comparative genomics.
Results and discussion
Types of OU patterns, abbreviations and nomenclature
Counts of words of different lengths N from 2 to 7-mer were analyzed in this work applying different schemes of normalization. Different types of OU patterns were abbreviated as type_N-mer. Types were "n0" for non-normalized, "n1" for normalized by mononucleotide frequencies, "n2" for normalized by dinucleotides and so on. For example, the non-normalized tetranucleotide usage pattern is denoted as n0_4 mer, trinucleotide usage pattern normalized by dinucleotides is n2_3 mer, pentanucleotide usage pattern normalized by trinucleotides is n3_5 mer. Each OU pattern is characterized by three statistical parameters: D – distance between two patterns of the same type (in this work we used distances D between local and global genome patterns); PS – pattern skew, distance between the two patterns of the direct and reverse strands of the same DNA sequence; and OUV – oligonucleotide usage variance. Correspondingly, the nomenclature is as follows: distance between a local n0_4 mer pattern and the corresponding global pattern – D:n0_4 mer; pattern skew of a n0_3 mer pattern – PS:n0_3 mer; variance of a n3_7 mer pattern – OUV:n3_7 mer. Two subtypes of normalization of local OU patterns were defined: normalized by frequencies of component words in the current genomic fragment (internal normalization, i) and in the complete sequence of the genome (global normalization, g). For example, internal and global OUV determined for a local n1_4 mer pattern were OUV:n1i_4 mer and OUV:n1g_4 mer, respectively. Internal normalization was always used in this study with the exception of the chapter "Identification of horizontally transferred elements" where the distances between OUV:n1i_4 mer and OUV:n1g_4 mer are analyzed. To simplify nomenclature, the index i was skipped in the pattern type abbreviation in all other chapters.
OU constraints in bacterial DNA
OUV values of OU patterns from n0_7 mer to n6_7 mer were calculated for the complete genome sequences of Bacillus subtilis 168, Escherichia coli K12 and Pseudomonas putida KT2440 (Fig. 1). OUV of n0_7 mer patterns depends strongly on GC-content getting minima in genomes with a GC content of about 50% such as in E. coli (Fig. 1) and maxima in AT-rich and, especially, GC-rich organisms, probably because OU is more strongly biased in GC-rich sequences [6,11]. Normalization of OU by mononucleotide frequency significantly removes this bias caused by GC-content (Fig. 1 and see ref. [6]). OUV n1_7 mer, however, is still high (Fig. 1). OUV decreases continuously with increase of the word length of internal normalization getting close to zero for n5 and n6 normalization of heptanucleotide usage (Fig. 1). This observation suggests that most OU constraints are caused by mononucleotide frequency and di-, tri- and tetranucleotide combinations while biases in frequencies of longer oligonucleotide words are probably just an extension of constraints of shorter component words.
Figure 1 OUV of different heptanucleotide usage patterns from n0_7 mer to n6_7 mer determined for complete bacterial genomes.
Local variations of OU patterns
To analyze local variations of OU in bacterial genomes, the sliding window approach was used. 163 bacterial chromosomes of eubacteria and archaea published in the NCBI database were analyzed. Local OU patterns were calculated for 8 kb genome fragments with 2 kb sliding windows [6]. Fig. 2 shows the distances D of local n0_4 mer patterns in three selected bacterial genomes: E. coli K12, P. putida KT2440 and B. subtilis 168 chromosomes. Genomic regions termed the 'core sequences' were characterized by OU patterns being similar to the global pattern of the chromosome. However, multiple genomic loci with alternative OU patterns that can make up more than 10% of the whole genome [11] were also detected in the three tested bacterial genomes (Fig. 2). Locally deviant OU patterns were found to comprise of heterogeneous subsets of parasitic and recent foreign DNA, ancient genes for ribosomal constituents (RNAs and proteins), multidomain genes and non-coding sequences with multiple tandem repeats.
Figure 2 Distances D between local n0_4 mer patterns and the global n0_4 mer patterns in the A)E. coli K12; B)P. putida KT2440 and C)B. subtilis 168 chromosomes. Local patterns were calculated for the sequence fragments of 8 kbp with sliding windows of 2 kbp. The 90% confidence interval of D values is depicted by horizontal lines. The loci with D-values exceeding the genomic confidence interval are considered as gene islands. The abscissa indicates the coordinates of the bacterial chromosomes as they were published in the NCBI database [27].
These functionally and evolutionarily unrelated subsets of atypical genomic loci were differentiated by the other OU statistical parameters: OUV and PS. These parameters often exhibited extreme values in detected atypical regions, however, their profiles were not congruent to each other. For example, consider the two adjacent gene islands in the P. putida KT2440 genome from 160 kbp to 240 kbp (Fig. 3). The first region (coordinates 170,815 – 180,000 bp) comprises of two tandem operons for ribosomal RNAs (rrnA-rrnA') [12], while the second 26,045 bp sequence covers the largest P. putida gene PP0168 encoding the surface adhesion protein [11]. Both regions were recognized by alternative OU patterns (maximal D:n0_4 mer were 59% and 37.5%, respectively, see Figs. 2 and 3). Notably, OUV:n1_4 mer has its genomic minimum (0.08) in the first region but its genomic maximum (0.88) in the second region, whereas PS:n0_4 mer is maximal (74.7%) in the first region and it is closer to the average level (47.5%) in the second region. This example illustrates that the combination of several OU pattern parameters may be useful for the differentiation of unrelated gene subsets.
Figure 3 Curves of D:n0_4 mer, PS:n0_4 mer and OUV:n1_4 mer in a locus of the P. putida KT2440 genome covering two regions with atypical OU: rrnA-rrnA* gene cluster and a long multidomain gene PP0168 encoding the surface adhesion protein. Local OU patterns were analyzed in 5 kbp sliding windows with steps of 1 kbp. Curves are specified by a color code: blue for D, green for PS and brown for OUV. Protein coding genes are shown by red bars and genes for ribosomal RNAs are shown in black. The abscissa indicates the coordinates of the locus in the chromosome. The upper horizontal line shows the upper boundary of the 95% confidence interval of intragenomic deviation of D values. The lower horizontal line separates genes by their direction of transcription.
The application of this procedure to a whole genome is shown in Fig. 4 for the cases of P. putida KT2440 and Mycobacterium leprae TN. Dots corresponding to the genome fragments were plotted in accordance with their D:n0_4 mer, OUV:n1_4 mer and PS:n0_4 mer values. The majority of fragments that represent the core genome clusters in one area. Three outlier groups detected in P. putida KT2440 and in the majority of other tested genomes were termed sections (Fig. 4A). Section I is heterogeneous and includes long intergenic regions, clusters of short hypothetical genes, laterally transferred elements and genes for ribosomal RNAs. The OU patterns of section I are characterized by low OUV and high PS. The operons for ribosomal RNAs exhibited the highest PS values (depicted by red dots, see Fig. 4). Genes for ribosomal proteins are localized in section II. This separation of ribosomal protein genes from the bulk genome was observed in most analyzed bacterial chromosomes but in some slow-growing microorganisms such as M. leprae these genes were not distinct from the core sequence (Fig. 4B). This observation is consistent with the notion that the codon usage in genes encoding ribosomal proteins is separate from the rest of genes in fast-growing bacteria but indistinguishable in slow-growing bacteria [13]. The differential codon usage of fast-growing bacteria has the consequence that ribosomal protein mRNA transcripts utilize other tRNA pools than the other mRNA species for the most abundant amino acids and hence the synthesis of the translational machinery is uncoupled from all other translational demands of the cell [14].
Figure 4 Dot-plot presentation of 8 kb genomic fragments of A)P. putida KT2440 and B)M. leprae TN chromosomes. Fragments of 8 kbp were generated with a sliding window 2 kbp. Each dot represents the D:n0_4 mer, OUV:n1_4 mer and PS:n0_4 mer values of one fragment. The latter parameter is depicted by a color code represented by the bar in the right part of the figure. The grey lines indicate borders of the inner quartiles of values for the corresponding OU statistical parameters.
Section III encompasses the regions with outermost OUV (approximately 3 to 15 standard deviations of genomic OUV) and locus-specific OU patterns (large D values). The genetic repertoire covered by these loci is represented in Table 1. These regions typically comprise of one or more large multidomain genes of over 4 kbp in length or non-coding sequences with multiple tandem repeats. Examples are genes coding for surface proteins (P. putida KT2440, Staphylococcus aureus N315, Xylella fastidiosa Temecula 1), hemagglutinins and hemolysins (Acinetobacter sp., Bordetella bronchiseptica RB50, Pseudomonas aeruginosa PA01, Pseudomonas syringae DC3000, X. fastidiosa Temecula 1 and Yersinia pestis KIM), fatty-acid synthetases (Corynebacterium efficiens YS-314) and genes for proteins with an overrepresentation of a few amino acids (Mycobacterium tuberculosis H37Rv, Streptomyces coelicolor A3(2)). Many bacterial chromosomes lack these genetic elements. It seems that these genes or mulidomain regions are species specific. For example, consider the M. leprae genome lacking such genetic elements (Fig. 4B) in comparison with the closely related M. tuberculosis H37Rv (Table 1). The genetic elements of section III were not observed in the following tested genomes: Aeropyrum pernix K1, Agrobacterium tumefaciens C58, Aquifex aeolicus VF5, Archaeglobus fulgidus DSM4304, Azoarcus sp. EbN1, Bacillus anthracis Ames, B. subtilis 168, Bdellovibrio bacteriovorus HD100, Borrelia burgdorferi B31, Campylobacter jejuni NCTC 11168, E. coli K12, Enterococcus faecalis V583, Francisella tularensis Schu 4, Haemophilus influenzae KW20, Halobacterium sp. NRC1, Helicobacter pylori J99, Lactococcus lactis IL1403, Mesorhizobium loti MAFF303099, Prochlorococcus marinus CCMP1375, Pyrococcus furiosus DSM 3638, Salmonella enterica Ty2, Shigella flexneri 2457T, Streptococcus pneumoniae R6, S. pyogenes MGAS8232, Treponema pallidum Nichols.
Table 1 Genetic repertoire of loci characterized by atypical tetranucleotide usage patterns and extreme OUV (section III in Fig. 4) identified in bacterial chromosomes
Genome Genes and the encoded protein Start* Length (bp) ΔD† ΔOUV‡
Acinetobacter sp. putative hemagglutinin/hemolysin-related protein 923,008 11,136 3.11 4.13
non-coding multiple repeats TTTAGAAA 2,448,000 5.600 2.24 17.33
Bordetella bronchiseptica RB50 BB1186: putative hemolysin 1,268,967 10,041 5.13 4.12
Bradyrhizobium japonicum USDA110 blr325: unknown 3,592,327 17,058 3.17 4.65
bll356: unknown 3,930,196 10,326 6.23 5.02
bll371: unknown 4,106,955 12,387 4.39 4.95
bll547: unknown 6,017,600 12,633 5.04 6.16
Corynebacterium efficiens YS-314 fasA: fatty-acid synthase I 962,711 8,919 2.85 3.85
fasB: fatty-acid synthase II 2,541,750 9,069 2.88 5.42
Deinococcus radiodurans R1 chromosome 1 DR1461-1462: hypothetical proteins 1,465,188 10,000 2.19 8.27
non-coding tandem repeats CCCGCCC 519,833 8,415 7.06 8.42
E. coli O157:H7 Z0609, Z0615: RTX family exoproteins 581,356 20,160 1.82 9.43
Mycobacterium tuberculosis H37Rv Rv0272c-Rv0279c hypothetical Gly-, Ala-rich proteins 328.573 10,499 1.52 9.15
Rv0297-Rv0304c: hypothetical Gly-, Ala-, Asn-rich proteins 361,332 11,431 8.79 7.91
Rv0355c: Asn-rich protein 424,775 9,903 8.31 10.91
Rv0573c-Rv0578c: hypothetical Gly-rich proteins 665,849 10,066 0.60 4.72
Rv0742-Rv0747: hypothetical Gly-rich proteins 832,979 7,876 1.24 3.97
Rv1060-Rv1068c: hypothetical Gly-, Ala-rich proteins 1,183,506 8,641 1.04 5.54
Rv1084-Rv1092c: hypothetical proteins 1,207,634 11.395 2.19 6.44
multiple repeats CCGCCGCCA 1,630,636 7,592 2.33 8.84
Rv2490c-Rv2494: hypothetical Gly-rich proteins 2,801,252 7,482 2.60 5.50
Pseudomonas aeruginosa PAO1 PA1874: hypothetical protein 2,036,441 7,407 2.61 5.61
P. putida KT2440 PP0168: Thr-rich surface adhesion protein 194,494 26,046 2.58 6.97
PP0806: surface adhesion protein 926,690 18,930 1.17 4.39
P. syringae DC3000 PSPTO3229: filamentous hemagglutinin 3,629,677 18,825 2.34 7.87
Rhodopirellula baltika 1 RB3077: putative cyclic nucleotide binding protein 1,588,083 18,024 1.62 6.19
RB4375: large polymorphic membrane protein, probable extracellular nuclease; 2,242,933 9,171 3.23 7.09
RB11769: probable aggregation factor core protein MAFp3 6,335,006 24,522 5.25 6.31
Rhodopseudomonas palustris CGA009 conserved hypothetical protein 1,459,664 9,891 2.61 3.38
conserved hypothetical protein 1,475,303 13,008 2.89 4.18
Sulfolobus solfataricus P2 non-coding tandem repeats GAATTGAAAG 1,228,221 12,238 1.94 15.25
1,253,000 5,000 1.50 8.67
1,305,242 5,000 1.89 12.39
Staphylococcus aureus N315 ebhA – ebhB: large surface anchored proteins 1,437,928 20,142 4.04 10.07
SA2447: similar to streptococcal hemagglutinin 2,755,253 6,816 3.03 9.29
Streptomyces coelicolor A3(2) SC8F4.01c: Ala/Glu-rich protein 586,509 3.981 2.16 5.40
SC2H4.02: hypothetical protein 6,836,057 6,552 2.86 4.80
Xanthomonas campestris ATCC33913 yapH: putative autotransporter adhesin 2,374,740 11,886 3.22 6.61
Xylella fastidiosa Temecula 1 non-coding sequence, multiple 1,183,606 11,095 1.31 9.81
repeats (GGT)n 1,447,312 11,139 1.37 10.91
pspA1: hemagglutinin 2,082,143 10,134 1.06 9.78
pspA2: hemagglutinin 2,501,956 10,374 1.41 11.79
Yersinia pestis KIM irp1-2: yersiniabactin peptide/polyketide synthetase; 2,654,642 15,867 4.27 6.05
yapH: putative autotransporter adhesin 3,747,888 11,133 2.66 8.60
y3579: putative filamentous hemagglutinin 3,961,333 9,888 3.31 4.32
* left coordinate of the locus in the chromosomal sequence;
† deviation of the D:n0_4 mer value calculated for the locus from the mean genomic D:n0_4 mer in standard deviations;
‡ deviation of the OUV:n1_4 mer value calculated for the locus from the mean genomic OUV:n1_4 mer in standard deviations;
Section I is heterogeneous. The genes for ribosomal RNAs are discerned from the other genes in section I by their extremely high PS of 60 – 70% that are usually the highest values in the genome. For further differentiation of the gene classes in section I, the next chapter describes the strategy to apply further OU statistical parameters to identify the subgroup of horizontally acquired elements.
Identification of horizontally transferred elements
Identification of laterally acquired elements in chromosomal sequences is of great importance because genomic islands often comprise pathogenicity and catabolic versatility determinants [15,16]. Two types of normalization of local OU patterns, – internal and global (see above), – were applied to visualize horizontally transferred gene islands within a genome sequence. The reason for introduction of these additional parameters was to improve the discrimination of foreign inserts in genome sequences. In core sequences, where the mononucleotide content is virtually the same as in the complete genome, results of internal and global normalization are identical in contrast to the laterally transferred loci characterized by an alternative mononucleotide content (in terms of GC-content, G/C-skew and A/T-skew). Correspondingly, values of OUV:n1i_4 mer and OUV:n1g_4 mer should merge in core sequences but widely diverge in gene islands (Fig. 5A). This concept was proven for genomes with known gene islands: SKIN element in Bacillus subtilis 168 [17], phage related gene islands in P. putida KT2440 [11] and in Salmonella enterica Ty2 [18], pathogenicity island LEE in E. coli O157:H7 [19], IS-elements, pathogenicity and prophage islands in Shigella flexneri 2457T [20], ISFtu1 element in Francisella tularensis Schu4 [21], cag pathogenicity island in Helicobacter pylori 26695 [2] and 67 kbp gene island in X. fastidiosa 9a5c [22]. All mentioned gene islands were successfully localized from the comparison of local with global OU patterns, however, no large foreign regions were observed in sequences of Bradyrhizobium japonicum and Mesorhizobium loti chromosomes, which both contain large symbiotic gene islands [23,24]. It looks as if these gene islands had been acquired a long time ago and hence their OU patterns adapted to the host genome OU signatures by genome amelioration [4,25].
Figure 5 Gene islands in the P. putida KT2440 genome identified by discordant OUV:n1i_4 mer and OUV:n1g_4 mer values A) in a local gene map and B) globally in the complete genome. Genome fragments of 8 kbp were generated with a sliding window in step of 2 kbp. Red bars in figure A indicate protein coding genes and black bars-hypothetical genes. The horizontal line in the part A separates genes by direction of transcription. The yellow-shaded 8 kbp long fragment in A corresponds to the red dot indicated by an arrow in B.
An example for the identification of a laterally acquired gene island is shown in Fig. 5. The island in the chromosome of P. putida KT2440 has significantly divergent OUV:n1i_4 mer and OUV:n1g_4 mer values and D:n0_4 mer values beyond the 95% confidence interval of the complete chromosome (Fig. 5A). Since OUV:n1i_nmer and OUV:n1g_nmer in local patterns and the difference thereof are automatically calculated by the program, the method may be used for the high-throughput identification of horizontally transferred elements in bacterial genomes. Whereas OUV:n1i_4 mer and OUV:n1g_4 mer values are strongly correlated in the bulk P. putida genome, all islands show up by high OUV:n1g_4 mer and low OUV:n1i_4 mer values (Fig. 5B).
Informative assignments of the OU statistical parameters
The objective of our work was to analyze the informative assignment and applicability of different statistical parameters of OU. Di-, tri- and tetranucleotide usage patterns are charged with most information content (see Fig. 1). The optimal word length will provide maximal information about the question of interest. First, one has to consider the minimal sequence length that gives reliable OU statistics. The threshold values of the minimum length of sequence were calculated to be 0.3, 1.2, 5 and 20 kbp for di-, tri-tetra- and pentanucleotides, respectively [6]. However, to be informative, the window should of course be not too long, because otherwise short range fluctuations of OU will vanish. We recommend that the window should not be longer than 10-fold of its minimal length. Tetranucleotide (and, sometimes, pentanucleotide) usage patterns are more appropriate for the global analysis of sequences. A long sliding window silences signals from the local repeats and structural biases at the level of individual genes so that the characteristics of whole operons and gene islands become apparent. For a more detailed analysis of chromosomal loci or short genomes of bacterial plasmids and phages, tri- and dinucleotide usage patterns may be more appropriate. For example, in Fig. 6 the mosaic structure of the plasmid pKLC102 was recovered by investigation of local trinucleotide usage patterns (genomic fragments were segregated by 1.2 kbp sliding windows in steps of 200 bp). Three peaks of high D values depict recombination sites of the plasmid where additional genetic elements (transposons, integrons and gene cassettes) may be inserted [26]. A region with extremely high OUV:n1_3 mer corresponds to the putative replication origin of the plasmid [26].
Figure 6 Structural analysis of the complete sequence of the plasmid pKLC102 by local trinucleotide usage patterns. Local OU patterns were analyzed in 1.2 kbp sliding windows with steps of 0.2 kbp. The scale indicates the coordinates of the plasmid sequence and separates genes by their direction of transcription. Red bars depict protein coding genes and black bars hypothetical genes. Grey bars along the D and OUV axes depict the 3-sigma ranges of fluctuation of D:n0_3 mer and OUV:n1_3 mer in a randomly generated sequence of the same length and mononucleotide contents as pKLC102.
To check whether the local fluctuations of OU parameters are statistically valid, a sequence of 100 kbp of mononucleotide content similar to pKLC102 was randomly generated. The ranges of 3-sigma fluctuation of D:n0_3 mer and OUV:n1_3 mer in the random sequence are depicted in Fig. 6 by vertical grey bars along the corresponding D and OUV axes. In the real sequences these values vary over a significantly larger range with the mean value of D smaller and the mean OUV higher than in the randomly generated sequence. (The plasmid pKLC102 sequence and the randomly generated sequence are included in the additional files as examples of source data files pKLC102.fts and random.fts, respectively.)
Normalization of OU by the internal component words changes the information assignment of OU biases. The three parameters D, PS and OUV were calculated for n0_4 mer, n1_4 mer, n2_4 mer and n3_4 mer local patterns for the pKLC102 genome and a part of the E. coli K12 chromosome from 1 Mbp to 2 Mbp. The former one is an example of a mosaic genome, and the latter one represents a regular bacterial chromosome. Correlation coefficients were calculated for respective OU statistical parameters determined for non-normalized and normalized local OU patterns. The correlation coefficients varied between 0.10 and 0.89 for pKLC102 and between 0.46 and 0.94 for E. coli (Table 2). This data demonstrates that n0, n1, n2 and n3 of 4 mer local patterns measure different characteristics of a sequence. In other words, the statistical parameters with different types of normalization provide non-redundant information that can be exploited for a refined analysis of genome organization. In case of tetranucleotide usage analysis four types of patterns exist: n0_4 mer, n1_4 mer, n2_4 mer and n3_4 mer. Each pattern type can be characterized by three parameters, D, PS and OUV that provide in total a comprehensive set of 12 non-redundant parameters for the nucleotide sequence analysis. Moreover, two subtypes of normalized OU patterns were introduced above, – with internal and global normalization, – that results in a total set of 21 non-redundant tetranucleotide usage statistical parameters each suitable for the refinement of functional gene classes in a raw nucleotide sequence.
Table 2 Correlation coefficients between D, PS and OUV of n0_4 mer local patterns with those of the corresponding n1, n2 and n3 normalized patterns
Parameters Normalization type
n1_4 mer n2_4 mer n3_4 mer
plasmid pKLC102, window 5,000 bp, step 2,500 bp
D:n0_4 mer 0.85* 0.82 0.40
PS:n0_4 mer 0.40 0.60 0.10
OUV:n0_4 mer 0.89 0.83 0.39
1 Mbp-2 Mbp locus of E. coli K12 chromosome, window 10,000 bp, step 5,000 bp
D:n0_4 mer 0.94 0.84 0.63
PS:n0_4 mer 0.88 0.75 0.53
OUV:n0_4 mer 0.61 0.46 0.35
*Values in the cells of the table indicate the correlation coefficients between respective OU statistical parameters D, PS and OUV determined for n0 patterns and the normalized patterns n1, n2 and n3. For example, 0.85 is the correlation coefficient between series of values D:n0_4 mer and D:n1_4 mer determined for overlapping 5 kbp fragments of pKLC102.
Conclusion
Bacterial genomes are not homogeneous but contain polymorphic blocks including horizontally transferred gene islands, non-coding sequences, long multidomain genes and ancient conserved gene clusters. The structural polymorphism of bacterial genomes may be effectively analyzed by local OU pattern signatures. A set of statistical approaches has been designed to perform this structural analysis of nucleotide sequences of bacterial genomes. These methods are useful for the visualization of regions with atypical oligonucleotide composition. The combination of the informative parameters that are 21 in case of tetranucleotide usage analysis, facilitates the prediction of gene classes. Moreover, many other subtypes of OU patterns may be additionally introduced. To this end, OU statistical analysis provides a valuable toolbox for the functional classification of regions and genes of interest prior to common-practice gene annotation.
A command line version of the Python program to apply the OU statistics methods mentioned above is available as additional file. To run the program, first the Python interpreted language program must be downloaded from the Web-site and installed on the computer. The source DNA sequence (or sequences) should be saved in FASTA format in text file(s) with .FST file name extensions. Users may choose the OU statistical parameters to be calculated and the parameters of the sliding window by setting corresponding command line arguments. Many different OU parameters may be determined by a single run of the program and all FST files in the target folder will be processed continuously in a batch. For each source data file an output file in TXT format will be saved in the same folder. The full list of arguments and description of how to use the program are documented in the readme.doc file provided in the additional files. The program is fast enough to calculate all set of OU parameters mentioned in this paper for a complete bacterial genome of average length in 10–20 min depending on the computer performance.
Several general conclusions about OU in bacteria can be drawn from this report. First, most OU constraints are hidden in di-, tri- and tetranucleotide combinations that vanish with increasing word length (see Fig. 1). For example, in case of a hexamer the four possible heptamer words will have the same likelihood to occur next in the sequence. Hence, i)the analysis of the oligonucleotide distribution of up to 4-mers is sufficient to uncover all OU constraints in the sequence; and ii)neighbor effects are limited to dipeptides so that protein evolution is not skewed by oligonucleotide biases. Second, D and PS values are correlated in local patterns (see the examples for D:n0_4 mer and PS:n0_4 mer in Fig. 3 and 4). This observation is in accordance with the general trend in bacterial sequences to keep parity of frequencies of words and their reverse complements, in other words- a trend towards minimal PS [6]. OU parity is most pronounced for the OU pattern of the whole chromosome, whereas fluctuations of OU in local patterns lead to an increased PS. The exceptions are the laterally transferred elements with their island-specific OU signature. In this case, large D values of the local OU patterns may be associated with low PS (see blue and green dots in section I in Fig. 4).
Methods
Sequences of 163 bacterial chromosomes including eubacterial and archaeal genomes published in the NCBI database [27] were analyzed in this study.
The OU statistical parameters-variance of word deviations (OUV); distances between patterns (D); pattern skew between leading and lagging strand (PS) were calculated by applying the algorithms described previously [6]. In a sequence of Lseq nucleotides we calculated numbers of occurrence of overlapping N-long oligonucleotide words. There are 4N possible combinations of nucleotides and the total number of words in a sequence corresponds to the sequence length Lseq. OU pattern was denoted as a matrix of deviations Δ[ξ1...ξN]
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqqHuoardaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDbqabaaaaa@3977@ of observed from expected counts for all possible words of the length N:
Δ[ξ1...ξN]=(C[ξ1...ξN]|obs−C[ξ1...ξN]|e)/C[ξ1...ξN]|0
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@6E5A@
where ξn is any nucleotide A, T, G or C at the position 1, 2, 3, ... N in the N-long word; C[ξ1...ξN]|obs is the observed count of the word, [ξ1...ξN]; C[ξ1...ξN]|e is the expected count and C[ξ1...ξN]|0 is a standard count estimated from the assumption of an equal distribution of words in the sequence: (C[ξ1...ξN]|0=Lseq×4−N
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDjabcYha8jabicdaWaqabaGccqGH9aqpcqWGmbatdaWgaaWcbaGaem4CamNaemyzauMaemyCaehabeaakiabgEna0kabisda0maaCaaaleqabaGaeyOeI0IaemOta4eaaaaa@476E@).
OU parameters of words of length N were normalized by shorter words n (0 ≤ n <N) as follows:
C[ξ1...ξN]|e=C[ξ1...ξN]|0
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDjabcYha8jabdwgaLbqabaGccqGH9aqpcqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDjabcYha8jabicdaWaqabaaaaa@4BE3@ if OU is not normalized, or C[ξ1...ξN]|e=C[ξ1...ξN]|n
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDjabcYha8jabdwgaLbqabaGccqGH9aqpcqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDjabcYha8jabd6gaUbqabaaaaa@4C5A@ if OU is normalized by empirical frequencies of all shorter words of the length n. The normalization was performed as follows. First at all, we calculated observed frequencies F[ξ1...ξn]
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGgbGrdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOBa4gabeaaliabc2faDbqabaaaaa@3966@ of n-long words in the sequence. Each word of length N can be represented as a consecutive set of N - n + 1 overlapping component words of length n. For example, a pentamer ATGGC can be expressed as a set of 4 overlapping dimers: AT, TG, GG and GC. In a general case of a N-long word, a component word [ξ1...ξn] reduces the set of available options for the next word in the sequence to 4 possible oligonucleotides: [ξ2...ξn, A], [ξ2...ξn, T], [ξ2...ξn, G] and [ξ2...ξn, C]. The relative frequencies of these words are:
F[ξ2...ξn,ξn+1]×[(F[ξ2...ξn,A]+F[ξ2...ξn,T]+F[ξ2...ξn,G]+F[ξ2...ξn,C])]−1
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGgbGrdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabikdaYaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaWexLMBbXgBcf2CPn2qVrwzqf2zLnharyGvLjhzH5wyaGabaiaa=5gaaeqaaSGaeiilaWIaeqOVdG3aaSbaaWqaaiaa=5gacqGHRaWkcqaIXaqmaeqaaSGaeiyxa0fabeaakiabgEna0kabcUfaBjabcIcaOiabdAeagnaaBaaaleaacqGGBbWwcqaH+oaEdaWgaaadbaGaeGOmaidabeaaliabc6caUiabc6caUiabc6caUiabe67a4naaBaaameaacaWFUbaabeaaliabcYcaSiaa=feacqGGDbqxaeqaaOGaey4kaSIaemOray0aaSbaaSqaaiabcUfaBjabe67a4naaBaaameaacqaIYaGmaeqaaSGaeiOla4IaeiOla4IaeiOla4IaeqOVdG3aaSbaaWqaaiaa=5gaaeqaaSGaeiilaWIaa8hvaiabc2faDbqabaGccqGHRaWkcqWGgbGrdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabikdaYaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaa8NBaaqabaWccqGGSaalcaWFhbGaeiyxa0fabeaakiabgUcaRiabdAeagnaaBaaaleaacqGGBbWwcqaH+oaEdaWgaaadbaGaeGOmaidabeaaliabc6caUiabc6caUiabc6caUiabe67a4naaBaaameaacaWFUbaabeaaliabcYcaSiaa=neacqGGDbqxaeqaaOGaeiykaKIaeiyxa01aaWbaaSqabeaacqGHsislcqaIXaqmaaaaaa@8BF7@
whereby the F values are the observed frequencies of the particular word of length n in the complete sequence and ξ is any nucleotide A, T, G or C. The expected count of a word [ξ1...ξN] of length N in a Lseq long sequence normalized by frequencies of n-mers (n <N) was calculated as follows:
C[ξ1...ξN]|n=Lseq×F[ξ1...ξn]×∏i=2N−n+1(F[ξi...ξi+n−2,ξi+n−1]∑XA,T,G,CF[ξi...ξi+n−2,X])
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqWGdbWqdaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaWexLMBbXgBcf2CPn2qVrwzqf2zLnharyGvLjhzH5wyaGabciaa=5eaaeqaaSGaeiyxa0LaeiiFaWNaemOBa4gabeaakiabg2da9iabdYeamnaaBaaaleaacqWGZbWCcqWGLbqzcqWGXbqCaeqaaOGaey41aqRaemOray0aaSbaaSqaaiabcUfaBjabe67a4naaBaaameaacqaIXaqmaeqaaSGaeiOla4IaeiOla4IaeiOla4IaeqOVdG3aaSbaaWqaaiaa=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@A127@
For further processing of OU statistics, the words were sorted by their Δ[ξ1...ξN] and the ranks of words instead the real values of deviations of observed from expected counts were used. The rank values (from 1 to 256 in the case of tetranucleotide analysis) were assigned to the words in accordance with their Δ[ξ1...ξN]
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqqHuoardaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDbqabaaaaa@3977@ values by ordering the words from the most overrepresented one (the greatest Δ[ξ1...ξN]
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqqHuoardaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDbqabaaaaa@3977@ to the least represented one (the lowest Δ[ξ1...ξN]
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqqHuoardaWgaaWcbaGaei4waSLaeqOVdG3aaSbaaWqaaiabigdaXaqabaWccqGGUaGlcqGGUaGlcqGGUaGlcqaH+oaEdaWgaaadbaGaemOta4eabeaaliabc2faDbqabaaaaa@3977@. This approach made the OU statistical parameters free from any dependence on the sequence length, provided that the sequence has a minimum length Lmin so that in a random sequence of the same length Lmin 95% of all words of length N occur at least ten times (see above and [6]). Hence, local OU patterns that meet these requirements could be compared with the global pattern.
The distance D between two patterns was calculated as the sum of absolute distances between ranks of identical words (w, in a total 4N different words) in patterns i and j as follows:
D(%)=100×∑w4N|rankw,i−rankw,j|−DminDmax−Dmin
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@6530@
PS is a particular case of D where patterns i and j were calculated for the same DNA but for direct and reversed strands, respectively. Dmax = 4N(4N - 1)/2 and Dmin = 0 when calculating a D, or, in a case of PS calculation, Dmin = 4N if N is an odd number or Dmin = 4N - 2N if N is an even number [6].
The definition of OUV was provided in our previous paper [6].
The random sequence was generated by a in-house program using the Python randomizer [28].
List of abbreviations
OU – oligonucleotide usage;
OUV – oligonucleotide usage variance;
PS – pattern skew;
D – distance between two OU patterns of an identical type.
Authors' contributions
ONR did Python programming. Both authors contributed equally to all other presented data.
Supplementary Material
Additional File 1
There is an additional ZIP archive file OligoWords for BMC Bioinf.zip comprising following documents: OligoWords1.1.exe.py - a command line version of the program implemented in Python2.2 [28]. readme.doc - description of the project in Word97 format. pKLC102.fst- sequence of the plasmid pKLC102 [26] in FASTA format that may be used as a source data file for the program OligoWords1.1.exe.py (see readme.doc). random.fst - a randomly generated sequence comparable with one of the plasmid pKLC102 by length and mononucleotide content. The file is in FASTA format that may be used as a source data file for the program OligoWords1.1.exe.py (see readme.doc).
Click here for file
Acknowledgements
This work was supported by the DFG-sponsored Europäisches Graduiertenkolleg 653.
==== Refs
Noble PA Citek RW Ogunseitan OA Tetranucleotide frequencies in microbial genomes Electrophoresis 1998 19 528 535 9588798 10.1002/elps.1150190412
Pride DT Blaser MJ Identification of horizontally acquired elements in Helicobacter pylori and other prokaryotes using oligonucleotide difference analysis Genome Let 2002 1 2 15 10.1166/gl.2002.003
Abe T Kanaya S Kinouchi M Ichiba Y Kozuki T Ikemura T Informatics for unveiling hidden genome signatures Genome Res 2003 13 693 702 12671005 10.1101/gr.634603
Pride DT Meinersmann RJ Wassenaar TM Blaser MJ Evolutionary implications of microbial genome tetanucleotide frequency biases Genome Res 2003 13 145 155 12566393 10.1101/gr.335003
Teeling H Waldmann J Lombardot T Bauer M Glöckner FO TETRA: a web-service and a stand-alone program for the analysis and comparison of tetranucleotide usage patterns in DNA sequences BMC Bioinformatics 2004 5 163 15507136 10.1186/1471-2105-5-163
Reva ON Tümmler B Global features of sequences of bacterial chromosomes, plasmids and phages revealed by analysis of oligonucleotide usage patterns BMC Bioinformatics 2004 5 90 15239845 10.1186/1471-2105-5-90
Karlin S Global dinucleotide signatures and analysis of genomic heterogeneity Curr Opin Microbiol 1998 1 598 610 10066522 10.1016/S1369-5274(98)80095-7
Karlin S Mrazek J Campbell A Compositional biases of bacterial genomes and evolutionary implications J Bacteriol 1997 179 3899 3913 9190805
Gorban AN Popova TG Zinovyev AY Four basic symmetry types in the 7-cluster structure of microbial genomic sequences In Silico Biol 2005 5 0025
Weinel C Ussery DW Ohlsson H Sicheritz-Ponten T Kiewitz C Tümmler B Comparative genomics of Pseudomonas aeruginosa PAO1 and Pseudomonas putida KT2440: orthologs, codon usage, REP elements and oligonucleotide motif signatures Genome Letters 2002 1 175 187 10.1166/gl.2002.021
Weinel C Nelson KE Tümmler B Global features of the Pseudomonas putida KT2440 genome sequence Environ Microbiol 2002 4 809 818 12534464 10.1046/j.1462-2920.2002.00331.x
Weinel C Tümmler B Hilbert H Nelson KE Kiewitz C General method of rapid Smith/Birnstiel mapping adds for gap closure in shotgun microbial genome sequencing projects: application to Pseudomonas putida KT2440 Nucleic Acids Res 2001 29 E110 11713330 10.1093/nar/29.22.e110
Carbone A Zinovyev A Képès Codon adaptation index as a measure of dominanting codon bias Bioinformatics 2003 19 2005 2015 14594704 10.1093/bioinformatics/btg272
Kiewitz C Weinel C Tümmler B Genome codon index of Pseudomonas aeruginosa : a codon index that utilizes whole genome sequence data Genome Letters 2002 1 61 70 10.1166/gl.2002.008
Hacker J Kaper JB Pathogenicity islands and the evolution of microbes Annu Rev Microbiol 2000 54 641 679 11018140 10.1146/annurev.micro.54.1.641
van der Meer JR Sentchilo V Genomic islands and the evolution of catabolic pathways in bacteria Curr Opin Biotechnol 2003 14 248 254 12849776 10.1016/S0958-1669(03)00058-2
Sato T Kobayashi Y The ars operon in the skin element of Bacillus subtilis confers resistance to arsenate and arsenite J Bacteriol 1998 180 1655 1661 9537360
Deng W Liou SR Plunkett G 3rdMayhew GF Rose DJ Burland V Kodoyianni V Schwartz DC Blattner FR Comparative genomics of Salmonella enterica serovar Typhi strains Ty2 and CT18 J Bacteriol 2003 185 2330 2337 12644504 10.1128/JB.185.7.2330-2337.2003
Perna NT Mayhew GF Posfai G Elliott S Donnenberg MS Kaper JB Blattner FR Molecular evolution of a pathogenicity island from enterohemorrhagic Escherichia coli O157:H7 Infect Immun 1998 66 3810 3817 9673266
Wei J Goldberg MB Burland V Venkatesan MM Deng W Fournier G Mayhew GF Plunkett G 3rdRose DJ Darling A Complete genome sequence and comparative genomics of Shigella flexneri serotype 2a strain 2457T Infect Immun 2003 71 2775 2786 12704152 10.1128/IAI.71.5.2775-2786.2003
Larsson P Oyston PC Chain P Chu MC Duffield M Fuxelius HH Garcia E Halltorp G Johansson D Isherwood KE The complete genome sequence of Francisella tularensis, the causative agent of tularemia Nat Genet 2005 37 153 159 15640799 10.1038/ng1499
Simpson AJ Reinach FC Arruda P Abreu FA Acencio M Alvarenga R Alves LM Araya JE Baia GS Baptista CS The genome sequence of the plant pathogen Xylella fastidiosa. The Xylella fastidiosa Consortium of the Organization for Nucleotide Sequencing and Analysis Nature 2000 406 151 157 10910347 10.1038/35018003
Kaneko T Nakamura Y Sato S Asamizu E Kato T Sasamoto S Watanabe A Idesawa K Ishikawa A Kawashima K Complete genome structure of the nitrogen-fixing symbiotic bacterium Mesorhizobium loti DNA Res 2000 7 331 338 11214968 10.1093/dnares/7.6.331
Kaneko T Nakamura Y Sato S Minamisawa K Uchiumi T Sasamoto S Watanabe A Idesawa K Iriguchi M Kawashima K Complete genomic sequence of nitrogen-fixing symbiotic bacterium Bradyrhizobium japonicum USDA110 DNA Res 2002 9 189 97 12597275 10.1093/dnares/9.6.189
Lawrence JG Ochman H Amelioration of bacterial genomes: rates of change and exchange J Mol Evol 1997 44 383 397 9089078
Klockgether J Reva O Larbig K Tümmler B Sequence analysis of the mobile genome island pKLC102 of Pseudomonas aeruginosa C J Bacteriol 2004 186 518 534 14702321 10.1128/JB.186.2.518-534.2004
NCBI Genome Sequence Database
The Python home site
|
16225667
|
PMC1274298
|
CC BY
|
2021-01-04 16:27:46
|
no
|
BMC Bioinformatics. 2005 Oct 14; 6:251
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-251
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2521622567410.1186/1471-2105-6-252SoftwareCoaSim: A flexible environment for simulating genetic data under coalescent models Mailund Thomas [email protected] Mikkel H [email protected] Christian NS [email protected] Peter JM [email protected] Jesper N [email protected] Leif [email protected] Bioinformatics Research Center, University of Aarhus, Høegh Guldbergsgade 10, 8000 Århus C, Denmark2 Bioinformatics ApS, Høegh Guldbergsgade 10, 8000 Århus C, Denmark3 Department of Computer Science, University of Aarhus, Høegh Guldbergsgade 10, 8000 Århus C, Denmark2005 14 10 2005 6 252 252 12 7 2005 14 10 2005 Copyright © 2005 Mailund et al; licensee BioMed Central Ltd.2005Mailund et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Coalescent simulations are playing a large role in interpreting large scale intra-specific sequence or polymorphism surveys and for planning and evaluating association studies. Coalescent simulations of data sets under different models can be compared to the actual data to test the importance of different evolutionary factors and thus get insight into these.
Results
We have created the CoaSim application as a flexible environment for Monte Carlo simulation of various types of genetic data under equilibrium and non-equilibrium coalescent processes for a variety of applications. Interaction with the tool is through the Guile version of the Scheme scripting language. Scheme scripts for many standard and advanced applications are provided and these can easily be modified by the user for a much wider range of applications. A graphical user interface with less functionality and flexibility is also included. It is primarily intended as an exploratory and educational tool
Conclusion
CoaSim is a powerful tool because of its flexibility and ease of use. This is illustrated through very varied uses of the application, e.g. evaluation of association mapping methods, parametric bootstrapping, and design and choice of markers for specific questions
==== Body
Background
These years witness popularity of coalescent based inference of evolutionary parameters, either through exact or approximate methods in order to understand within-species, and in particular human, evolution. This is fuelled by large scale efforts to collect data from different human populations, e.g. the HapMap [1] and ENCODE projects in humans. Important aims are to link genetic variation to phenotypic traits, e.g. complex disease, in order to identify the causal variants, as well as the human genetic history and dispersal.
The coalescent models are good null models to test specific hypotheses against, and therefore it is often useful to simulate genetic data under different coalescent models.
Coalescent simulation of data sets and subsequent calculation of some summary statistics on the data set can be used to obtain the distribution of this summary statistic under a given scenario, and this can be compared to the value of the summary statistic calculated from real data. Parametric bootstrapping can be done by estimating parameters from data under a given (coalescent) model, apply coalescent simulations under some model using the observed value of the parameters and see if the value observed is compatible with the simulation model. The efficiency of new methods can also be evaluated: Methods for inference of recombination rate and its variation, the importance of gene conversion versus recombination, as well as methods for association mapping under different genetic models have to rely on some heuristic approximations of the full likelihood models which are generally too computationally complex. Validating these approximations and inferring the importance of misspecification of the model of analysis employed is usually done using some variant of coalescent simulations.
Various applications for the simulation of the coalescent process are already available (e.g. [2-6]). The present application aims at attaining maximum flexibility in model specification. Population size changes, bottlenecks and demographics (migration rate matrix) can be specified precisely, there is a choice of mutation models (including microsatellite mutation models), recombination and gene conversion can be specified precisely and are allowed to vary over the region. Furthermore, complex disease models including arbitrary interactions between genes, varying penetrance and environmental effects, can also be specified. Case-control data sets under these disease models and any ascertainment scheme can then be generated for subsequent evaluation of association mapping methods aiming at detecting disease causing variants.
Implementation
Installation
CoaSim can be obtained at , where instructions for the installation are also provided. CoaSim is written in C++ and is available as source code and in a binary version for Linux operating system as rpm-files.
An introduction to the simulator is provided as a "getting started" manual, where also the GUI version is described. The GUI is an intuitive interface to the simulator, allowing novice users to immediately start the simulations of sequences. Markers are added manually and the simulation stages can be monitored (see Figures 1, 2, 3). For more advanced purposes, users should turn to the guile-scheme version of CoaSim. Controlling the simulations through scheme makes CoaSim a very flexible and powerful simulation tool.
Figure 1 Screen shots from CoaSim GUI showing the input dialog.
Figure 2 Screen shot from GUI showing the simulation status dialog.
Figure 3 Screenshot from GUI showing simulated data.
Running the application
The CoaSim Guile manual gives a number of examples of simple and advanced usage of the guile scheme version. Scripts that allow the iterative simulation of population samples under various demographic scenarios and disease models are introduced. An example demographic scenario is illustrated in Figure 4. Population merging times, sizes, migration rates etc. needs to be specified in the Scheme script. An example script that simulates the situation in Figure 4 is shown in Figure 5. The advanced usage of the tool for extracting summary statistics such as the mean of the total branch length of the simulated ARGs, the mean of the total tree height, the mean number of recombination nodes and coalescent nodes are all exemplified in scripts accompanying the distribution.
Figure 4 Example graphical representation of a demographic scenario with population splits, migration, and growth. f is the size of a population in units of 2N, β = 2Nb is the growth rate. Above the dotted arrows are the backwards migration rates, again scaled in units of 2N. Scheme code implementing the complete model is shown in Figure 5.
Figure 5 Example Scheme code that implements the demographic scenario of Figure 4. The Scheme code specifies both the population structures and the migration rates between populations, and simulates a sample of 100 individuals from population P1 and 50 from each of populations P3 and P4, with the merging of populations P3 and P4 into p2 at time 1.5 and the merging of populations P1 and P2 at time 3. The merge of P1 and P2 is followed by a bottleneck followed by a period of constant population size of f*2N = 10N.
The output format of the sequence or marker data sets (such as Hudson's ms format) can be specified in the scheme script.
Results and discussion
Coalescent models
For an introduction to coalescent theory, see for example [7]. Basically, the ancestral history of a sample is simulated by repeated drawing of random numbers from competing exponential distributions corresponding to different evolutionary forces such as coalescence, migration, recombination, gene conversion. Coalescent intensities depend on the population size in each of demes where sampled genes are present. The number of demes, their sizes and the migration matrix are allowed to vary arbitrarily as specified by the input to the simulations. When an event (coalescence, migration, recombination or gene conversion) occurs, the state of the sample is updated and the intensities of the exponential distributions for the next event calculated. The process is continued until all parts of the gene have found a most recent common ancestor. The resulting graph, termed the ancestral recombination graph (ARG), is then used to generate data under different mutation models as specified by the user. Data can be generated either under the infinite sites or finite sites model of mutation. The marker positions can either be specified or chosen randomly, with uniform probability of distribution within the interval. It is possible either to condition on a mutation rate or on a given number of segregating sites for replicated simulations. The finite sites model can have any number K of states for a given position.
There are three kinds of built-in markers:
(1) Trait markers are binary polymorphisms (presence or absence of a trait, such as disease state) with a simple mutation model: after simulating the ARG, a mutation is placed uniformly at random on the tree local to the marker position, nodes below the mutation will have the mutant allele while all others will have the wild-type allele. A range of accepted mutant-frequencies can be specified and a simple rejection-sampling scheme is used to ensure it: if, after placing the mutation, the number of mutant leaves is not within the range, the ARG is rejected and the simulation restarted. This insures an unbiased collection of ARGs with a binary marker within a given frequency range.
(2) SNP markers resemble trait markers in that they are binary polymorphisms, and use the same mutation model as the trait-markers. They differ from the trait-markers in how the mutant-frequency is ensured: If, after the mutation has been placed, the number of mutant leaves does not fall within the accepted range, the mutation is re-placed, but the ARG is not rejected and re-simulated. This places a bias on the markers, but one that resembles the ascertainment bias seen in association studies, where SNPs are chosen to have frequencies in certain ranges. An unbiased sampling of mutations within the required frequency range can be obtained by constructing a new ARG and repeating the simulation process, until the required frequencies are attained.
(3) Microsatellite markers have a specified number of alleles, k, and a different mutation model than traits and SNPs. For microsatellite markers, each edge in the local tree at the marker is considered in turn and, with the likelihood of mutation being an exponential distribution that depends on the specified mutation rate and the length of the edge. If mutation occurs, a randomly chosen allele from 0 to k-1 is placed on the child node; if no mutation occurs, the child node gets a copy of the allele at the parent node.
In addition to the built-in marker types, Scheme scripting allows specification of most conceivable mutation models, the stepwise mutation model for microsatellites is included as an example.
Trait markers can be marked as being involved in disease susceptibility under a specified, but arbitrary model of disease risk as a function of biallelic variants at a number of sites specified a priori. A multidimensional matrix specifies disease risk for any combination of variants. Thus, from the simulated data and the disease model, case-control data of any size can be generated. Diploid data with known or unknown phase can be easily generated by calling the relevant function from the Scheme script as specified in the manual.
Splitting the sequences into cases and controls based solely on the allele at a trait marker or combination of alleles in a more complex disease determination is not always appropriate since it assumes full penetrance. However, in many complex diseases, penetrance is incomplete and dependent on environmental factors and general genetic background, and the same disease phenotypically can be caused by a different genetic pathway. Hence, it is possible to control the likelihood of a given individual belonging to the case or control group by specifying the probabilities given the possible genotypes.
Population size changes, bottlenecks, merging of populations (splitting of populations when viewed forward in time) and migration between sub-populations can vary arbitrarily as specified through the definitions of epochs in the Scheme script controlling the simulation (see Figures 4 and 5).
Speed issues
We have measured execution times for a series of simulations. Simulating 10,000 SNP markers with a minor allele frequency of 10% for 10,000 chromosomes with a recombination rate of 100 took 85 seconds on a 3.0 GHz Pentium 4, 1 GB RAM machine. Increasing the recombination rate to 1000 caused the execution time to increase to 153 seconds.
We also simulated a scenario that in likely to find applications in simulating interacting disease loci. In this scenario, 1000 sequences were simulated for a setup where two trait markers which frequencies range between 20% to 40% were located in a region of 1000 SNP markers with a minor allele frequency of 10%,. This simulation took 28 seconds.
Improvements to existing packages
Hudsons program ms provides many of CoaSims features for population growth, migration, recombination, gene conversion, arbitrary ascertainment schemes, recombination rate variation across the chromosome, stepwise mutation models, trait markers and cases and controls specified by the interaction between genes and genotype-specific penetrance probabilities. A difference is the ease of extracting the desired output. Using ms, the user has to post-process the output using custom made scripts, whereas CoaSim provides the required functionalities. With ms, it is not possible to specify the position of markers, to employ a finite sites mutation model or to simulate linked microsatellites.
Some example applications of CoaSim
We have used CoaSim for three applications that demonstrate some of its flexibility.
1. Estimation of recombination rate and effective population size in Iceland from microsatellite data. In this study, microsatellite diploid, phase unknown, markers were simulated under different mutation models and for different rates of recombination. The expected decay of different measures of linkage disequilibrium with recombination rate could then be estimated and compared to the expected decay in a large microsatellite data set from Iceland. This allowed us to estimate the effective population size of the Icelandic population and to investigate whether the population bears any sign of recent population growth [8].
2. Case-control data sets were generated under a simple disease model (single locus dominance) but with various rates of heterogeneity and penetrance. The data with disease causing mutation removed was then directly piped into the GeneRecon program [9, 10] that attempts to infer the disease marker position from all the simulated markers. This allowed us to investigate the effect of marker density, penetrance, heterogeneity, number of cases/controls etc ([9], T. Mailund et al., unpublished results).
3. The effect of single recombination event. Genealogies were simulated for a given recombination rate but only genealogies with exactly one recombination event were used for simulation of marker (SNP) data sets. These data sets were then used as input to programs aimed at estimating the recombination rate, allowing us to estimate the variance in effect on data that a single recombination event can have and relate this to number of markers, population growth etc (M. H. Schierup et al. unpublished results).
Conclusion
The CoaSim software package was designed for flexibility and with adaptations and extensions of the various Scheme scripts provided and the user manual, a wide range of situations can be accommodated. It provides the user with a much wider range of demographical models, of marker types and disease models without much loss of user friendliness compared to competing software.
Availability and requirements
Project name: CoaSim 4.0
Project home page:
Source codes for Guile and GUI versions are supplied [see Additional files]
Operating system(s): Linux Fedora 1–4, Redhat 9, and MacOSX
Programming language: C++ and scheme
Other requirements: Guile Scheme version 1.6. The GUI requires QT version 3.3
License: GNU
Authors' contributions
TM, MHS, CNSP, JNM, LS, planned the project. TM, JNM and PJMM wrote the software, TM wrote the documentation. LS and TM tested the software. MHS, TM, LS and CNSP wrote the paper. All authors read and approved the final manuscript.
Supplementary Material
Additional File 1
Click here for file
Additional File 2
Click here for file
Additional File 3
Click here for file
Additional File 4
Click here for file
Acknowledgements
We thank Thomas Bataillon for useful comments to the manuscript. TM received support from ISIS Katrinebjerg, MHS, LS and CNSP acknowledges support from the Danish Natural Sciences Research Council.
==== Refs
The International HapMap Project Nature 2003 426 789 796 14685227 10.1038/nature02168
Posada D Wiuf C Simulating haplotype blocks in the human genome Bioinformatics 2003 19 289 290 12538254 10.1093/bioinformatics/19.2.289
Wiuf C Hein J The coalescent with gene conversion Genetics 2000 155 451 462 10790416
Wiuf C Posada D A coalescent model of recombination hotspots Genetics 2003 164 407 417 12750351
Spencer CC Coop G SelSim: a program to simulate population genetic data with natural selection and recombination Bioinformatics 2004 20 3673 3675 15271777
Hudson RR Generating samples under a Wright-Fisher neutral model of genetic variation Bioinformatics 2002 18 337 338 11847089 10.1093/bioinformatics/18.2.337
Hein J Schierup MH Wiuf C Gene genealogies, variation and evolution - A primer in coalescent theory
2004 Oxford University Press
Rafnar T Thorlacius S Steingrimsson E Schierup MH Madsen JN Calian V Eldon BJ Jonsson T Hein J Thorgeirsson SS The Icelandic Cancer Project--a population-wide approach to studying cancer Nat Rev Cancer 2004 4 488 492 15170451 10.1038/nrc1371
Mailund T Pedersen CNS Bardino J Vinter B H.H. K Initial experiences with GeneRecon on MiG 2005
|
16225674
|
PMC1274299
|
CC BY
|
2021-01-04 16:27:46
|
no
|
BMC Bioinformatics. 2005 Oct 14; 6:252
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-252
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2531622567610.1186/1471-2105-6-253Research ArticleProfNet, a method to derive profile-profile alignment scoring functions that improves the alignments of distantly related proteins Ohlson Tomas [email protected] Arne [email protected] Stockholm Blolnformatlcs Center, Stockholm University, SE-106 91 Stockholm, Sweden2005 14 10 2005 6 253 253 9 5 2005 14 10 2005 Copyright © 2005 Ohlson and Elofsson; licensee BioMed Central Ltd.2005Ohlson and Elofsson; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Profile-profile methods have been used for some years now to detect and align homologous proteins. The best such methods use information from the background distribution of amino acids and substitution tables either when constructing the profiles or in the scoring. This makes the methods dependent on the quality and choice of substitution table as well as the construction of the profiles.
Here, we introduce a novel method called ProfNet that is used to derive a profile-profile scoring function.
The method optimizes the discrimination between scores of related and unrelated residues and it is fast and straightforward to use. This new method derives a scoring function that is mainly dependent on the actual alignment of residues from a training set, and it does not use any additional information about the background distribution.
Results
It is shown that ProfNet improves the discrimination of related and unrelated residues. Further it can be used to improve the alignment of distantly related proteins.
Conclusion
The best performance is obtained using superfamily related proteins in the training of ProfNet, and a classifier that is related to the distance between the structurally aligned residues. The main difference between the new scoring function and a traditional profile-profile scoring function is that conserved residues on average score higher with the new function.
==== Body
Background
Alignment of proteins is one of the fundamental methods in bioinformatics. Alignments are used to detect homology and to study evolutionary events. The ability to align distantly related proteins can be improved significantly by the inclusion of evolutionary information [1,2] or predicted features [3]. Although significant improvements of alignment qualities has been seen recently in CASP [4], it is not clear how much the improved performance is due to an improvement of alignment methodologies and how much is due to increased number of sequences and structures that can be used to span the distance between a query protein and a target structure. However, in a recent study we have shown that the average alignment quality, as measured by MaxSub [5], improved by 10% at the family level and 50% at the superfamily level by the use of profile-profile scoring instead of sequence-profile scoring [6]. These findings are comparable to the ones found in a number of recent studies [7-9].
Profile-profile alignments can be implemented in several different ways [10-16]. The fundamental difference between different profile-profile alignment methods lie in how they calculate the score between two profile vectors. A profile, as defined in this study, can be seen as a set of vectors where each vector contains the frequency of each amino acid at a particular position in a multiple sequence alignment. In traditional sequence-profile alignments the score is calculated by extracting (the log of) the probability for an amino acid in this vector. However, in profile-profile alignments it is necessary to compare two vectors and this can be done in several different ways, including; calculating the sum of pairs, the dot product or a correlation coefficient between the two vectors. In addition, information about the background frequency and substitution probabilities can be used. Although, it has been shown that profile-profile methods using a probabilistic model seem to be superior to other methods [6,8], it is quite likely that better profile-profile scoring functions could be developed.
Here we present ProfNet, a method to develop novel profile-profile scoring functions. ProfNet is based on the ability to separate related from unrelated residue pairs, and it uses an artificial neural network (ANN) trained to identify pairs of residues from structurally aligned proteins. We show that ProfNet provides significantly better identification of related residues than prob_score [17] and that it also can be used to provide a slight improvement of the alignment of distantly related proteins. Another advantage of this approach is that it makes it trivial to add additional information to the scoring function.
Results
It could be expected that a good profile-profile scoring function should provide high scores if two profile vectors have similar amino acid distributions that differs from the background distribution. In addition, the score should include information about what amino acids are more likely to be exchanged with each other. In an earlier study we found that one profile-profile scoring method, prob_score [17], performed these tasks quite well [6]. However, it is quite possible that a better function could be found. In order to develop such a function we have develop the method ProfNet that separates residue pairs that should and should not be aligned. Here, it is assumed that residue pairs aligned in a structural alignment should also be aligned by the profile-profile scoring function, while residue pairs belonging to proteins from different folds should not be aligned at all. Therefore, the scoring function was trained to identify pairs of structurally aligned residues. Finally, the ability to correctly align protein pairs using this novel scoring function was tested.
Identification of related residues
A set of artificial neural networks (ANNs) were trained to distinguish related, by structural alignments, and non-related residue pairs. The first set of networks were trained using a simple representation where all aligned residues were considered to be related and a set of unrelated residues were chosen from randomly selected positions in unrelated protein pairs. The ANNs were trained using different datasets containing proteins of varying degrees of similarity. The performance of the different ProfNet versions was measured using the Matthews correlation coefficient (MCC) and the number of standard deviations separating the related and non-related residue pairs, i.e. the Z-score. In table 1 it can be seen that the MCC values for prob_score drops from 0.51 for family to 0.17 for superfamily and to 0.13 for fold related scores. No large difference in performance between the identification of superfamily and fold related pairs can be found, indicating that the physiochemical aspects of protein similarity is of greatest importance at this level of similarity.
Table 1 MCC-values and the corresponding Z-scores for prob_score and ProfNet versions trained on different datasets. The ProfNet versions were trained on profile vector pairs from unrelated proteins and protein positions related at family (ProfNet_fam), superfamily (ProfNet_su), fold (ProfNet_fold), and all SCOP levels (family, superfamily and fold) (ProfNet_all). The training of ProfNet_S was done using superfamily related profile vector pairs as positive examples, and classified by the S-score instead of the binary classifiers used in the other cases. The results are shown for protein pairs related on family, superfamily and fold level. The best results are shown in bold.
MCC Z-score
training fam su fold fam su fold
prob_score 0.51 0.17 0.13 1.53 0.69 0.35
ProfNet_fam 0.51 0.18 0.14 1.69 0.72 0.42
ProfNet_su 0.49 0.19 0.16 1.69 0.81 0.52
ProfNet_fold 0.26 0.12 0.13 0.89 0.51 0.47
ProfNet_all 0.50 0.18 0.16 1.84 0.81 0.50
ProfNet_S 0.45 0.18 0.17 1.58 0.79 0.56
For ProfNet_fam (which uses family related data in the training) a slight improvement over prob_score was seen. The Z-scores increased by 5–20% while the MCC values show a marginal increase at the superfamily and fold levels. In contrast to prob_score the ProfNet_fam scores for family related residues have a non-Gaussian distribution, see figure 1. For ProfNet_su (which uses superfamily related data in the training) the separation for distantly related residues got noticeable better at only a marginal lost performance at the family level, but when using only proteins from different superfamilies, but similar folds in the ANN training, (ProfNet_fold) the results are worse at all levels indicating that the evolutionary information is lost here. Using a combination of proteins from all SCOP levels (ProfNet_all) did perform similar to ProfNet_su, most likely because the superfamily set contain a set of residues related at a similar level. In figure 1 it can be seen that ProfNet_fold hardly separates the pairs at all, while all other scoring functions clearly separates the family-related, and to some extent also the superfamily-related from the non-related pairs.
Figure 1 The distribution of scores from family, superfamily, fold related and randomly chosen profile vectors for prob_score and the five different ProfNet versions. The ProfNet versions were trained on profile vector pairs from unrelated proteins and proteins related at family (ProfNet_fam), superfamily (ProfNet_su), fold (ProfNet_fold), and a combination of family, superfamily and fold (ProfNet_all). The S-score training (ProfNet_S) was done using superfamily related vector pairs as positive examples, and classified by the S-score instead of the binary classifiers used in the other cases. All graphs show a Gaussian distribution, except for the family related scores in ProfNet_fam, which instead seems to follow an extreme-value like distribution. In each plot, the fraction of residues within a certain score range is plotted against the score. The exact values of the Y-axis have been left out for clarity.
In the above tests, all aligned positions were treated equally. However, certainly some of the aligned positions in the structural alignment are more closely aligned than others. Therefore, we also used a continuous function related to the distance between the two residues after the structural superposition. To measure the distance between two residues we used the S-score [18]. This ProfNet version, called ProfNet_S, performed quite well at the superfamily and fold levels, but did not distinguish the family related residues optimally. The Z-score for the fold related scores shows an improvement over prob_score by 60% and all the curves show a Gaussian like distribution, see Table 1 and Figure 1.
A ROC-plot was constructed the same way as in Edgar 2004 [19] from the data used in the MCC analysis, figure 2. It can be seen that ProfNet_S is slightly better on superfamily level for low error rates, and clearly better at fold level.
Figure 2 ROC plot based on the score for pairs of related and unrelated profile positions for prob_score and the S-score trained ProfNet. For each score the log of the error rate is plotted against the sensitivity for proteins related at superfamily, and fold level. The performance on family level was similar for the methods and was therefore left out for clarity.
Alignment quality
Although the identification of related residues might have some practical value [20], the real benefit from an improved scoring function would be if it could improve alignments and/or the detection of related proteins. It has earlier been shown that the alignment accuracy is increased by the use of profile-profile comparisons [7]. In an earlier study we noted a correlation between the ability to separate residues and the alignment quality if the gap-penalties were optimized for each scoring function individually [6]. In Table 2 it can be seen that the best ProfNet versions performed on par with prob_score on the ability to align proteins related on family or superfamily level, while a small (10%) increase in alignment qualities could be observed for the ProfNet_S and ProfNet_su versions for fold related proteins. The three ProfNet versions that provide the best alignments have the best identification of related residues. Taken this into account there seems to be some truth in our assumption that there should be a relationship between the ability to separate related from unrelated residues and aligning proteins. These results imply that some of the information needed to optimally align distantly related proteins are better captured by ProfNet than by prob_score. Furthermore, in figure 3, it can be seen that the ProfNet alignments produce more correct models at a given error rate than prob_score. A slightly improved performance can be seen for ProfNet on superfamily level for error rates > 0.03 and at error rates > 0.1, at the fold level.
Table 2 Alignment quality results for prob_score and the ProfNet versions trained on different datasets. The ProfNet versions were trained on profile vector pairs from unrelated proteins and protein positions related at family (ProfNet_fam), superfamily (ProfNet_su), fold (ProfNet_fold), and all SCOP levels (ProfNet_all). The training of ProfNet_S was done using superfamily related profile vector pairs as positive examples, and classified by the S-score instead of the binary classifiers used in the other cases. The average MaxSub scores are listed for a test sets with proteins related the family, superfamily or fold levels. The best results are shown in bold.
training fam su fold
prob_score 0.56 0.20 0.063
ProfNet_fam 0.57 0.20 0.064
ProfNet_su 0.57 0.20 0.070
ProfNet_fold 0.55 0.17 0.057
ProfNet_all 0.57 0.20 0.067
ProfNet_S 0.57 0.20 0.072
Figure 3 ROC plot based on protein model quality as measured by the MaxSub score for prob_score and the S-score trained ProfNet. For each score the log of the error rate is plotted against the sensitivity for proteins related at superfamily, and fold level. The performance on family level was similar for the methods and was therefore left out for clarity.
Unfortunately, we did not see any significant improvement on the ability to detect related proteins using these novel scoring functions. The failure to increase fold recognition indicates that there still is work to do to find the optimal profile-profile scoring function. Quite likely, the construction of the negative training set was not done optimally.
Discussion
Both prob_score and ProfNet provide a score for two profile vectors, that should be related to the similarity between the two (profile) positions. In the following sections we will compare these two functions, where ProfNet_S is used as a representative of the ProfNet method.
The correlation coefficient between prob_score and ProfNet is 0.68, indicating that the main features are similar but that there also exists differences. To understand the differences, the score from the scoring functions were examined for residues with varying degrees of conservation. The conservation was measured by the frequency of the most frequent amino acid in the profile vector. It should be noted that the frequency is not the directly observed frequency from the multiple sequence alignment but instead calculated from the PSI-BLAST profiles. Further, the residue pairs were sorted into four groups, pairs with identical conserved amino acids and pairs where the conserved amino acids in the two vectors had a positive, zero, or negative BLOSUM62 [21] score. In figure 4 it can be seen that for ProfNet the average scores for all groups increase with increased conservation, while for prob_score only the score for identical conserved residues increase. In table 3 the average score for the six groups, using a 30% conservation cutoff is shown. As expected both scoring functions score identical residues highest, while pairs of conserved unrelated residues score lower. However, it is notable that, on average, ProfNet provide higher scores than prob_score for all conserved pairs, regardless of the relationship between the two residues. ProfNet actually provides similar scores to a pair of conserved residues with negative BLOSUM scores as to one conserved and one non-conserved residue. Clearly, being conserved increases, for some reason, the chances to be structurally aligned.
Figure 4 Average score for different classes of conserved residues. The classes were clustered by scores from vectors where the conserved residues in the vectors i) were identical, ii) had a positive BLOSUM62 score, iii) had a BLOSUM62 score of zero, and iv) had a negative BLOSUM62 score. The cutoff for the conserved residues are shown on the X-axis and the Z-score is shown on the Y-axis. A residue is considered conserved to a certain degree if it has a value in the profile vector above the cutoff. The solid bold lines are the scores for ProfNet_S, while the dotted lines are the scores for prob_score.
Table 3 Average Z-scores for prob_score and ProfNet_S for the scores for different types of conserved residue pairs. ProfNet_S was trained on superfamily related profile vector pairs and using the S-score as a classifier. The pairs are grouped into pairs with identical residues, positive, zero and negative BLOSUM scores. Finally, the Z-scores for a pair containing one conserved and one non-conserved residue and two non-conserved residues are shown. The highest scores are shown in bold.
prob_score ProfNet_S
identical res 3.09 2.50
pos BLOSUM 1.46 1.99
zero BLOSUM 0.44 1.10
neg BLOSUM -0.91 0.22
cons-non cons 0.21 0.15
non cons-non cons 1.06 0.53
To further investigate the differences in the scoring of conserved residues, substitution tables were derived from prob_score and ProfNet. The scores of the two tables were transformed into Z-scores and plotted against each other in figure 5. Here, it can be seen that prob_score ranks the residue pairs similar to BLOSUM62, giving the highest scores to identical pairs, while ProfNet on the other hand does not rank the residue pairs the same way. The correlation coefficient between the BLOSUM62 matrix and the scores from ProfNet was 0.75, compared to 0.95 for prob_score, see table 4. Figure 5 shows the same tendency that was observed in table 3, i.e. that ProfNet score most of the conserved residue pairs higher than prob_score. Clearly during the training of ProfNet other features than the BLOSUM62 classification has been learned. In figure 5 some outliers exist that might aid the explanation of the differences. ProfNet score pairs containing either a Cys or a Trp high while these pairs are scored low by prob_score. Trp and Cys are among the least frequent residues and a conservation of 30% (which is used as a cutoff for a conserved residue) might actually correspond to a higher degree of conservation than for a more common residue. Therefore, these scores could be explained by the general trend that the ProfNet scores increase with conservation. Another interesting outlier is the residue pair Ile-Val that is scored higher than many of the identical residue pairs by ProfNet. This indicates that the structural alignment might put more emphasize on physicochemical similarity than an evolutionary similarity.
Figure 5 The Z-scores of the "substitution tables" generated by prob_score and ProfNet_S plotted against each other. The average Z-score for each residue pair is shown. The residues are written using their one-letter code.
Table 4 Correlation between different substitution tables and the profile-profile scores. Three tables, BLOSUM62, GONNET and JTT, are derived from sequence alignments while SDM is a structure based table. Three tables derived from ProfNet are also included, using the all (trained on data from all SCOP levels), su (trained on superfamily related data) and S-score (trained on superfamily related data and using the S-score as a classifier instead of a binary classifier) versions. STRUCTAL is a substitution table constructed from the residue matches found in the structurally aligned superfamily-related training set. The highest correlations are shown in bold.
subst. table prob_score ProfNet_S
BLOSUM62 0.95 0.75
GONNET 0.93 0.75
JTT 0.87 0.70
SDM 0.89 0.73
STRUCTAL 0.85 0.82
ProfNet_su 0.83 0.96
ProfNet_all 0.83 0.95
ProfNet_S 0.80 1
The scores from the substitution tables GONNET [22], JTT [23] and SDM [24] were also compared with the ProfNet derived substitution table, see table 4. The first two substitution tables are based on sequence alignments, while SDM is a structurally derived substitution table, i.e. based on structural alignments. Overall, prob_score showed a higher correlation to the substitution tables than to ProfNet, and ProfNet showed higher correlation with prob_score than with the substitution tables. This shows that ProfNet capture some of the substitution table information and some of the conservation information used in prob_score. It can also be seen that prob_score and ProfNet show comparable correlation with a substitution table created directly from the structurally aligned superfamily-related dataset.
Future development
Here, we have only used the most obvious information from the profiles, i.e. the frequencies in the profile vectors for the development of the scoring function. One possible advantage of the ProfNet method is that it is easy to include other types of information, such as gap-information and predicted features, into the scoring functions.
Conclusion
A novel method, ProfNet, to derive a profile-profile scoring function is shown to improve the discrimination between related and unrelated residue residues pairs. Further, ProfNet can be used to marginally improve the alignment quality of proteins related at the fold level. One benefit of this method is that it is easy to use and fast to evaluate, while one drawback is that a good and well balanced training set has to be used, and it is slower than prob_score. When choosing the training set, it seems as if the family related set is too focused on sequence similarity while the fold related training set on the other hand does not seem to include enough closely related pairs. The superfamily related training set could be seen as an intermediate, where the network will learn the features in the residue pairs that are essential when scoring unseen residue pairs. It was also found that using a binary classifier is not the best way to classify the training data, but instead some continuous classifier could be used. When using the superfamily related training data and ProfNet_S we see an improvement over prob_score by 31% in MCC (60% in Z-score) and 14% in average alignment quality for the fold related proteins. Interestingly, ProfNet clearly scores all conserved residues higher than prob_score does.
Methods
Profiles
We used the log-odds profiles obtained after ten iterations of PSI-BLAST [25] version 2.2.2, using an E-value cutoff of 10-3 and all other parameters at default settings. The search was performed against nrdb90 from EBI [26]. The frequency profiles, used in prob_score, were back-calculated from the log-odds profiles obtained from PSI-BLAST as in [6]. The profiles used in ProfNet were created by a transformation of the log-odds profiles using a simple transformation as in PSI-PRED [27], trans formed_score(x)=1(1+exp−x)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWG0baDcqWGYbGCcqWGHbqycqWGUbGBcqWGZbWCcaaMc8UaemOzayMaem4Ba8MaemOCaiNaemyBa0MaemyzauMaemizaqMaei4xa8Laem4CamNaem4yamMaem4Ba8MaemOCaiNaemyzauMaeiikaGIaemiEaGNaeiykaKIaeyypa0ZaaSaaaeaacqaIXaqmaeaacqGGOaakcqaIXaqmcqGHRaWkieGacqWFLbqzcqWF4baEcqWFWbaCdaahaaWcbeqaaiabgkHiTiabdIha4baakiabcMcaPaaaaaa@5515@, where x is the value from the log-odds profile. In this study a profile is a matrix of dimensions 20xL, where L is the length of the target or query sequence. The term "profile vector" also known as "profile column" is a 20 × 1 dimensional vector with values corresponding to the occurrence of each amino acid, as calculated from the PSI-BLAST log-odds profiles, in a certain position in the profile.
Scoring of two profiles
The input to the ProfNet scoring function is two transformed profile vectors, see above. The score between two profiles was calculated by first filling the dynamic programming matrix using ProfNet as a scoring function. After the matrix is filled, standard dynamic programming is used, with affine gap penalties. The number of calculations for each cell in the dynamic programming matrix for ProfNet is hn × in, where hn = # hidden nodes in the ANN and in = # input nodes (= 2 × 20), typically 20 × 40. The number of calculations for each cell in the dynamic programming matrix for prob_score is 2 × r × (1 + x), where r = # residues in the alphabet (= 20), and x is the number of calculations for a logarithm, i.e. 2 × 20 × (1 + x). In our implementation ProfNet is almost three times slower than prob_score.
In Mittelman et. al. [17] it is shown that probabilistic scoring functions is significantly better than other scoring functions and in Wang & Dunbrack 2004 [8], it is stated that with optimized gap penalties, most scoring functions behave similarly to one another in alignment accuracy. Taking all this into account, we choose to use the probabilistic scoring function prob_score instead of for example COMPASS or PICASS03 [17], since it was used in our previous study, where it was shown to be one of the best methods.
Training sets
A subset of SCOP [28] version 1.57, class a to e, where no two protein domains have more than 75% sequence identity was used in the training of the artificial neural networks. For the positive training examples, protein pairs were structurally aligned using STRUCTAL [29] and all pairs of residues within 3 Å separation were used, while a set of negative training examples was created from randomly selected residue pairs from proteins of different folds. For the positive and negative data sets no more than 15 aligned positions from the same protein pair were used.
In an attempt to clarify what dataset to use in the ANN training we used five different datasets. The datasets consist of pairs of profile vectors corresponding to the aligned residues between protein pairs from the same family, superfamily (where no two proteins came from the same family), fold (where no two proteins came from the same superfamily), and a combination of family, superfamily and fold as positive examples and using randomly chosen vector pairs from unrelated protein positions as negative examples. The ANNs were trained to classify the profile vector pairs as related or unrelated (0 or 1). We also trained an ANN with the superfamily related set as positive examples, and trained to classify the profile vector pair according to the S-score [18]S−score=11+rmsd2/5
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGtbWucqGHsislcqWGZbWCcqWGJbWycqWGVbWBcqWGYbGCcqWGLbqzcqGH9aqpdaWcaaqaaiabigdaXaqaaiabigdaXiabgUcaRiabdkhaYjabd2gaTjabdohaZjabdsgaKnaaCaaaleqabaGaeGOmaidaaOGaei4la8IaeGynaudaaaaa@421C@. The rmsd is calculated between the Cα atoms of the aligned residues.
The ratio between the positive and negative examples was not adjusted, instead all examples were used in the training as this was shown to produce the best alignment quality results for the superfamily related training set (data not shown). The size of the datasets ranges from 20 000 examples for the fold related and the negative dataset to 100 000 for the S-score trained examples.
Matthews correlation coefficient
When comparing how well a method can separate positive and negative examples, such as the scores for related and unrelated profile positions, Matthews Correlation coefficient [30] (MCC) is a useful fitness measure. MCC takes into account both over-prediction and under-prediction and imbalanced data sets. It is defined as, MCC=tp×tn−fn×fp(tn+fn)(tn+fp)(tp+fn)(tp+fp)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGnbqtcqWGdbWqcqWGdbWqcqGH9aqpdaWcaaqaaiabdsha0jabdchaWjabgEna0kabdsha0jabd6gaUjabgkHiTiabdAgaMjabd6gaUjabgEna0kabdAgaMjabdchaWbqaamaakaaabaGaeiikaGIaemiDaqNaemOBa4Maey4kaSIaemOzayMaemOBa4MaeiykaKIaeiikaGIaemiDaqNaemOBa4Maey4kaSIaemOzayMaemiCaaNaeiykaKIaeiikaGIaemiDaqNaemiCaaNaey4kaSIaemOzayMaemOBa4MaeiykaKIaeiikaGIaemiDaqNaemiCaaNaey4kaSIaemOzayMaemiCaaNaeiykaKcaleqaaaaaaaa@6201@. True positives (tp) are correctly predicted related scores, true negatives (tn) are correctly predicted unrelated scores, false negatives (fn) wrongly predicted related scores and false positives (fp) wrongly predicted unrelated scores. The MCC score is in the interval (-1,1), where one shows a perfect separation, and zero is the expected value for random scores. Three subsets (family, superfamily, and fold level) of the SCOP version 1.57 dataset that were not used in the training were used to calculate the MCC-values for each method.
Artificial neural network training
The artificial neural networks (ANNs) were trained on 80% of the dataset, where a protein is only present in either the training or the test set. The neural network package Netlab in MatLab was used for the ANN training [31,32]. A linear activation function was chosen, and the training was carried out using the scaled gradient algorithm. Given two residues that should be aligned according to the training data, the ANN functions extracted their respective residue vectors from the transformed PSI-BLAST profiles, see above. The training of the ANNs was done using a grid search over the number of hidden nodes and number of training cycles. After the initial grid search, the search procedure was tuned to the area that produced the best results. At least 49 sets of parameters were tested for each ANN. The ANN-based scoring functions were chosen by selecting the ANN with the highest MCC-value and the minimum number of training cycles and hidden nodes. In the next step the ANN were used for the alignment quality test. The ProfNet scoring functions were implemented into the Palign [1,33] package
In summary, the ANNs were trained to identify related and unrelated profile vectors. The ANN use two transformed profile vectors, as described above, as input. The network should output a high score if two vectors are related and a low score otherwise. The ANNs that use a binary classifier outputs a value in the range (-0.6, 1.7), and the ANN that use the continuous S-score as a classifier output scores in the range (0, 1). In a sense, the network is trying to find a function that best can explain and correlate the training examples, i.e. the 40 numbers from the two profile vectors, and their output values. For the S-score trained network, the training examples classification are related to the rmsd between the Cα atoms of the two residues that are aligned. With this strategy, the ANN is trained to predict the distance between the two residues, and hence if they should be aligned or not.
Alignment quality
This dataset was also constructed from the same subset of SCOP version 1.57, class a to e, where no two protein domains have more than 75% sequence identity. From this dataset we included no more than 5 proteins from the same superfamily and no more than one model per domain target, we used in total 799 family, 672 superfamily and 602 fold related protein pairs. Among the superfamily related proteins, no proteins from the same family were included, and among the fold related proteins, no proteins from the same superfamily were included.
Throughout this study, only local alignments were used. For each alignment we created a model of the query protein and compared the structure of this model with the correct structure. We used MaxSub [5] which finds the largest subset of Cα atoms of a model that superimpose well over the experimental model. We only report the MaxSub score because we noted in our earlier study [6] that the results obtained using other methods, such as LGscore [18], were almost identical. The parameters for the best MaxSub scores on superfamily and fold level are not always the same, therefore we show the results for a choice of parameters with MaxSub scores reasonably high at all levels.
Parameters
The gap- and shift-parameters has to be optimized to get a good performance in the alignment quality test. The shift value is added to the score, so that an average score is negative, and the gap-opening (GO) and gap-extension (GE) is used to penalize for including a gap in the sequence. The gap-parameters in the alignment quality test were optimized with the constraint that the gap-extension penalty should be 5 or 10% of the gap-opening penalty. Ideally other ratios should be tried as well but as this would take too long time and we found reasonably good results, using the GO/GE ratio described above we did not spend any more time on the optimization. In addition this ratio between GO and GE has been seen to perform well in many other scoring schemes such as PSI-BLAST and prob_score. By using this rule we only had to search two two-dimensional parameter landscapes. We searched a grid of G0 = (0.1,0.2...,1.5) and shift = (-0.5, -0.45,...,1.5) for the ProfNet methods, and G0 = (0.2,0.3...,3.5) and shift = (-0.5,-0.45,...,1.5) for prob_score. The set of parameters with the best MaxSub score was then chosen.
ROC plot
In the two ROC-plots, the error rate is plotted against the sensitivity (= tp/(tp + fn)). In figure 2 the error rate and sensitivity was calculated from scores of related and unrelated profile positions, i.e. from the MCC analysis data. In figure 3, the alignment quality was calculated for the superfamily related set that was used in the alignment quality test and a negative dataset. The negative dataset consists of 1000 unrelated protein pairs from SCOP version 1.57, class a to e, where no two protein domains have more than 75% sequence identity.
Substitution matrices
In the comparison of ProfNet and prob_score, a conserved residue was defined as a residue with "frequency" (calculated from the converted PSI-BLAST profiles) above a certain cutoff in the profile frequency vector. In a non-conserved vector, no residue has a frequency above 0.10. To analyze how the methods score conserved residues, the average score was calculated between the conserved residues related at superfamily level from the test set used in the MCC test. To make the comparison more straightforward, the scores were transformed into Z-scores according to Z-score(x) = (x - μ)/σ, where μ is the average score over many randomly chosen examples and σ is the standard deviation. From these scores, substitution table-like matrices were derived for the methods. All different ProfNet versions produced the same outliers (data not shown).
Authors' contributions
Tomas Ohlson wrote the code for the analysis, designed the test set and performed all experiments. Arne Elofsson participated in the design of the study. Both authors collaborated in writing the manuscript.
Acknowledgements
This work was supported by grants from the Swedish Natural Sciences Research Council. We wish to thank Bob MacCallum and Björn Wallner for valuable support and discussions.
==== Refs
Elofsson A A study on how to best align protein sequences Proteins 2002 15 330 339 11835508 10.1002/prot.10043
Wallner B Fang H Ohlson T Frey-Skött J Elofsson A Using evolutionary information for the query and target improves fold recognition Proteins 2004 54 342 350 14696196 10.1002/prot.10565
Rost B Sander C Prediction of protein secondary structure structure at better than 70% accuracy J Mol Biol 1993 232 584 599 8345525 10.1006/jmbi.1993.1413
Moult J Hubbard T Bryant SH Fidelis K Pedersen JT Critical assesment of methods of proteins structure predictions (CASP): Round II Proteins (Suppl) 1997 1 2 6 10.1002/(SICI)1097-0134(1997)1+<2::AID-PROT2>3.0.CO;2-T
Siew N Elofsson A Rychlewski L Fischer D MaxSub: An automated measure to assess the quality of protein structure predictions Bioinformatics 2000 16 776 785 11108700 10.1093/bioinformatics/16.9.776
Ohlson T Wallner B Elofsson A Profile-profile methods provide improved fold-recognition: A study of different profile-profile alignment methods Proteins 2004 57 188 197 15326603 10.1002/prot.20184
Marti-Renom M Madhusudhan M Sali A Alignment of protein sequences by their profiles Protein Sci 2004 13 1071 1087 15044736 10.1110/ps.03379804
Wang G Dunbrack RL Scoring profile-to-profile sequence alignments Protein Sci 2004 13 1612 1626 15152092 10.1110/ps.03601504
Edgar R Sjolander K A comparison of scoring functions for protein sequence profile alignment Bioinformatics 2004 20 1301 1308 14962936 10.1093/bioinformatics/bth090
Fischer D Altman R, Dunker A, Hunter L, Klien T Hybrid Fold Recognition: Combining sequence derived properties with evolutionary information Pacific Symposium on Biocomputing 2000 5 World Scientific 116 127
Rychlewski L Jaroszewski L Li W Godzik A Comparison of sequence profiles. Strategies for structural predictions using sequence information Protein Sci 2000 9 232 241 10716175
von Öhsen N Sommer I Zimmer R Altman RB, Dunker AK, Hunter L, Jung TA, Klein TE Profile-profile alignments: a powerful tool for protein structure prediction Pacific Symposium on Biocomputing 2003 252 263 12603033
Yona G Levitt M Within the twilight zone: A sensitive profile-profile comparison tool based on information theory J Mol Biol 2002 315 1257 1275 11827492 10.1006/jmbi.2001.5293
Sadreyev R Grishin N COMPASS: A Tool for comparison of multiple protein alignments with assessment of statistical significance J Mol Biol 2003 326 317 336 12547212 10.1016/S0022-2836(02)01371-2
Edgar R Sjolander K SATCHMO: sequence alignment and tree construction using hidden Markov models Bioinformatics 2003 19 1404 1411 12874053 10.1093/bioinformatics/btg158
Pei J Sadreyev R Grishin NV PCMA: fast and accurate multiple sequence alignment based on profile consistency Bioinformatics 2003 19 427 428 12584134 10.1093/bioinformatics/btg008
Mittelman D Sadreyev R Grishin N Probabilistic scoring measures for profile-profile comparison yield more accurate short seed alignments Bioinformatics 2003 19 1531 1539 12912834 10.1093/bioinformatics/btg185
Cristobal S Zemla A Fischer D Rychlewski L Elofsson A A study of quality measures for protein threading models BMC Bioinformatics 2001 2 11545673
Edgar R MUSCLE: a multiple sequence alignment method with reduced time and space complexity BMC Bioinformatics 2004 5 15318951
Tress M Jones D Valencia A Predicting reliable regions in protein alignments from sequence profiles J Mol Biol 2003 330 705 718 12850141 10.1016/S0022-2836(03)00622-3
Henikoff S Henikoff JG Amino acid substitution matrices from protein blocks Proc Natl Acad Sci 1992 10915 10919 1438297
Gonnet G Cohen M Benner S Exhaustive matching of the entire protein sequence database Science 1992 257 1609 1610 1482492
Jones D Taylor W Thornton J The rapid generation of mutation data matrices from protein sequences Comput Appl Biosci 1992 8 275 282 1633570
Prlic A Domingues F Sippl M Structure derived substitution matrices for alignment of distantly related sequences Protein Engineering 2000 13 10964983
Altschul S Madden T Schaffer A Zhang J Zhang Z Miller W Lipman D Gapped BLAST and PSI-BLAST: a new generation of protein database search programs Nucleic Acids Res 1997 25 3389 3402 9254694 10.1093/nar/25.17.3389
Holm L Sander C Removing near-neighbour redundancy from large protein sequence collections Bioinformatics 1998 14 423 429 9682055 10.1093/bioinformatics/14.5.423
Jones D Protein secondary structure prediction based on position-specific scoring matrices J Mol Biol 1999 292 195 202 10493868 10.1006/jmbi.1999.3091
Murzin A Brenner S Hubbard T Chothia C SCOP: a structural classification of proteins database for the investigation of sequences and structures J Mol Biol 1995 247 536 540 7723011 10.1006/jmbi.1995.0159
Gerstein M Levitt M Comprehensive assessment of automatic structural alignment against a manual standard, the scop classification of proteins Protein Sci 1998 7 445 456 9521122
Matthews B Comparison of predicted and observed secondary structure, of T4 phage lysozyme Biochim Biophys Acta 1996 405 442 451 1180967
Bishop CM Neural Networks for Pattern Recognition 1995 Great Clarendon St, Oxford OX2 6DP, UK.: Oxford University Press
Nabney I Bishop C NetLab: Netlab neural network software 1995
Elofsson A Ohlson T palign
|
16225676
|
PMC1274300
|
CC BY
|
2021-01-04 16:27:46
|
no
|
BMC Bioinformatics. 2005 Oct 14; 6:253
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-253
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2591623232110.1186/1471-2105-6-259Research ArticleHuman promoter genomic composition demonstrates non-random groupings that reflect general cellular function McNutt Markey C [email protected] Ron [email protected] Wenwu [email protected] Irene [email protected] Wendy J [email protected] Idalia [email protected] Cynthia M [email protected] GVR [email protected] Kevin [email protected] The Advanced Technology Center, Laboratory of Receptor Biology and Gene Expression, National Cancer Institute, Bethesda, Maryland 20892-4605, USA2 The University of Texas Southwestern Medical Center at Dallas, TX, USA3 Bristol-Myers Squibb, Syracuse, NY, USA2005 18 10 2005 6 259 259 17 5 2005 18 10 2005 Copyright © 2005 McNutt et al; licensee BioMed Central Ltd.2005McNutt et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The purpose of this study is to determine whether or not there exists nonrandom grouping of cis-regulatory elements within gene promoters that can be perceived independent of gene expression data and whether or not there is any correlation between this grouping and the biological function of the gene.
Results
Using ProSpector, a web-based promoter search and annotation tool, we have applied an unbiased approach to analyze the transcription factor binding site frequencies of 1400 base pair genomic segments positioned at 1200 base pairs upstream and 200 base pairs downstream of the transcriptional start site of 7298 commonly studied human genes. Partitional clustering of the transcription factor binding site composition within these promoter segments reveals a small number of gene groups that are selectively enriched for gene ontology terms consistent with distinct aspects of cellular function. Significance ranking of the class-determining transcription factor binding sites within these clusters show substantial overlap between the gene ontology terms of the transcriptions factors associated with the binding sites and the gene ontology terms of the regulated genes within each group.
Conclusion
Thus, gene sorting by promoter composition alone produces partitions in which the "regulated" and the "regulators" cosegregate into similar functional classes. These findings demonstrate that the transcription factor binding site composition is non-randomly distributed between gene promoters in a manner that reflects and partially defines general gene class function.
==== Body
Background
Amidst a continuous bombardment of diverse stimuli from the external environment, metazoan organisms have adopted multiple strategies to respond specifically and decisively to a myriad of extracellular events. The biological map that determines this is encoded within the gene regulatory regions of the genome. Deciphering the inherent language in these encrypted codes is a major challenge of the post-genomic era. The search, retrieval and examination of the upstream regulatory sequences of eukaryotic genes coupled with empirical determination of their transcriptional regulatory function has yielded a wealth of potentially useful information relevant to the sequence-specific codes used to dynamically coordinate the spatial, temporal, and kinetic assembly of gene regulatory complexes at specific genes [1]. Cells must orchestrate this coordinated gene expression in order to efficiently execute the multitude of cellular programs that direct specific functions.
Essential components of controlling networks that modulate cellular programming are the regulatory sequences or transcription factor binding sites (TFBS). TFBSs comprise the basic unit of information stored within the upstream genomic regions located near the transcription start site (TSS) of most genes [1,2]. These typically 8–15 bp nucleotide sequences interact specifically with the DNA-binding domains of several hundred different transcription factors. Since it is widely accepted that the TFBS arrangement and composition of these upstream regulatory regions are the fundamental determinants of gene expression, many software applications and computational approaches have been developed to sort and identify TFBSs in the regulatory regions of genes determined to have similar patterns of expression [3-5]. One popular approach is based on a software algorithm that compares the potential binding site base frequencies against an established database of empirically determined nucleotide frequencies derived from published biological studies [5]. The resulting position weight matrixes (PWM) are then used to search for and characterize potential binding sites dependent on their statistical similarities to known TFBSs. A major goal of this approach is to analyze co-occurring TFBS frequencies in the regulatory regions of similarly regulated genes as a means of defining transcriptional pathways or networks that orchestrate the co-expression. Most biologists measure steady state RNA levels as an indicator of gene expression. Thus, the linkages between TFBS occurrence and gene expression will undoubtedly be imperfect due to the fact that: 1) steady-state levels of expressed mRNA are a combined result of both active transcription and mRNA turnover; 2) indirect regulation of transcription factors by post-translational modification or other transcriptional components is a common control mechanism in metazoan biology; and 3) most PWM libraries are derived from empirical data sets and therefore have limited inclusiveness [1,6-8]. Nonetheless, focused and global analysis of gene promoter composition has the potential of yielding important insight into gene regulation.
Recent efforts to define a common vocabulary to describe the function of all genes through the use of established Gene Ontology terms has provided a standardized approach of analyzing genes, clustered by any objective criteria, with respect to their cellular function [9,10]. Combining the analysis of gene promoter composition with gene ontology annotation provides a novel and innovative means through which linkages between gene regulatory networks and programs of cellular function can be identified and defined.
In this study, we analyzed the transcription factor binding site composition of 1400 bp promoter regions defined as 1200 bp upstream and 200 bp downstream of the transcription start site of 7,298 genes previously characterized in a recent microarray study of the kinetic patterns of gene expression in a mitogen-stimulated human leukemic T-cell line [11]. Though the composition of TFBSs in these 7,298 genes show very poor correlation with the measured global kinetic patterns of steady-state gene expression, independent partitional clustering of the TFBS composition within these 1400 bp regions "in silico" produced definable non-random gene groups for which distinct classes of ontology terms were found more frequently than expected by random chance. Moreover, analysis of the TFBSs that were most significant for distinguishing these gene groups revealed strong correlations between the ontology terms of the transcription factors predicted to bind the controlling gene regulatory regions and the ontology terms of the clustered genes themselves; thus, establishing a functional link defined by the ontology of the regulated gene and that of its regulators (TFBS-associated transcription factors). Refinement of this approach may provide a general means of defining the regulatory genomic templates upon which transcriptional networks are integrated to control specific programs of gene expression and cellular behavior.
Results
The process of T-cell activation has been a widely applied model system for the study of stimulus-evoked transcriptional control [12]. Prior analysis of this system has shown that many of the molecular signaling pathways initiated during T-cell activation converge on RAS-dependent effectors coupled with integrated secondary messenger signaling mediated by increased calcium influx. Thus, a common means of achieving robust activation of lymphoid cell lines is through pharmacological manipulations brought about by the addition of phorbol ester and calcium ionophore to resting cells. Several recent studies have profiled time dependent changes in steady-state gene expression of mitogen-induced T-cells to search for transcriptional pathways that appeared to be disproportionately effected based on a time series analysis of the data [11,13]. The fundamental linkage between transcriptional pathways and the expressed gene is the presence of recognition motifs or TFBS within the upstream regulatory region or promoters of the pathway-influenced gene. Thus, we sought to ask whether this logic could be extended to tease out biologically significant associations between TFBS frequencies and kinetic patterns of gene expression from a previously published study of the human T-cell line Jurkat [11]. This microarray data set contains steady-state mRNA profiles measured at 0, 1, 2, 6, 12 and 24 hours following stimulation. The data set was first filtered to remove uncharacterized and poorly annotated genes (see Methods). The hybridization data from the remaining 7,298 genes was then analyzed by K-means clustering to group or classify those genes with similar kinetic patterns of mitogen-induced expression (Figure 1a). As demonstrated in Figure 1b, the expression profiles of the 7,298 genes analyzed in phorbol ester and ionomycin stimulated Jurkat T-cells can be separated into 4 kinetic clusters.
Figure 1 Cluster analysis of gene expression data set from mitogen stimulated T-cells compared to promoter TFBS composition. (a) K-means cluster analysis of cDNA expression profiles of phorbol ester and ionomycin stimulated Jurkat T-cells collected at 0, 1, 2, 6, 12, and 24 hours after stimulation [11]. Total genes in each cluster is indicated in parentheses. (b) Centroid plot representing average kinetic profiles of the four clusters at the six measured time intervals. (c) Principal component analysis (PCA) of TFBS frequencies in the genomic sequences extracted from the 7,298 genes profiled in Figure 1a. (1200 base pairs upstream and 200 base pairs down stream from the start of transcription). Prior to analysis, each gene was color-coded by its respective cluster shown in Figure 1a (red = cluster/group 1, no change, green = cluster/group 2 early elevated expression, blue = cluster/group 3, repressed expression, and yellow = cluster/group 4 late elevated expression). (d) The extracted promoter sequences of each gene were then compared with respect to TFBS composition alone by K-means clustering. Nine out of sixteen clusters contained more the 4 genes (indicated as groups 1,4,6,9,10,12,13,15, and 16).
The first cluster (group 1, red) was the largest (4539) and represented genes that were essentially unchanged by mitogen stimulation. The second cluster (group 2, green) contained 175 genes and represents genes whose expression was induced early, within the first 2 hours of stimulation. The third cluster (group 3, blue) contained 990 members and represents genes that were relatively repressed by mitogen stimulation. The fourth cluster (group 4, yellow) contained 1594 members and represents genes whose expression rose late (post 6 hours) following mitogen stimulation. Given the rather broad differences between the groups and the known mitogen and calcium sensitivity of the AP-1, NF-kappa B and NFAT transcription factors pathways, it was expected that many promoters of the induced gene clusters (particular group 2) would show an asymmetric enrichment for TFBSs that bind AP-1, NF-kappa B or NFAT [12]. To address this hypothesis, 1400 bp of genomic sequence (1200 bp upstream and 200 bp downstream of the TSS) were extracted from the 7,298 genes using the ProSpector Promoter inspection tool (see Methods). These regions (referred to as promoter regions) were then scored for the presence of 164 different motifs based on the TRANSFAC 6.0 position weight matrices using the MatInspector algorithms described by Quandt et al [5]. Matrix and core thresholds were set at 0.75 and 1.0 respectively. The TFBS composition of the genes were then compared by principal component analysis (PCA), where the cluster classifications of the genes based on kinetic expression pattern were color coded (red = cluster/group 1; green = cluster/group 2, blue = cluster/group 3, and yellow = cluster/group 4). The genes were then grouped by the relative promoter frequencies of the 164 motifs applying 0.75/1.0 matrix/core PWM thresholds. In this presentation, the original 164 PWM motif vector space of the genes is reduced to 3 principal component vectors each representing a summed linear contribution from all 164 motifs [14,15]. As shown in Figure 1c, the clustering of the promoter TFBS frequencies produces a diffuse pattern that shows no correlation with the kinetic categories derived from gene expression data in Figure 1b.
These data indicate that broad kinetic grouping by gene expression alone fails to show strong correlations with transcription factor binding site composition. Though dramatic, this conclusion is not unexpected. Recent studies suggest that the correlation between steady-state mRNA levels and active transcription is at best 50%, since steady-state mRNA is the net result of not only nascent transcription, but also mRNA turnover [7]. Accordingly, promoter composition is likely to have a significantly stronger correlation with active transcription than with mRNA stability. Nonetheless, future studies aimed at generating a finer partitioning of the kinetic categories through the use of multiple conditions (e.g. different modes of stimulation) will be better prone to generate more selective gene groups with higher conditional correlation between TFBS composition and patterns of gene expression.
In clear contrast however, when the TFBS compositions of the 7,298 genes were analyzed independent of gene expression data by K-means clustering, sixteen distinct and stable clusters could be identified (Figure 1d). Seven of these clusters contained 4 or less genes and were discarded. The remaining 9 major partitions were composed of clusters containing from 271 (Cluster four) to 1266 (Cluster sixteen) genes (Figure 1d).
To determine whether there were any functional differences between the gene classes shown in Figure 1d, the genes in each cluster were analyzed for preferential enrichment or depletion of ontology terms using the GoMiner web-based software package [16]. GoMiner facilitates biological interpretation of gene lists using a quantitative statistical output that identifies gene ontology terms that are asymmetrically distributed between gene clusters. Over- and under-represented terms are ranked by a two-sided p-value from the Fisher's exact T-test [16]. The top 40 gene ontology terms for each gene cluster are shown in Table One. On first inspection, it is clear that each of the 9 gene clusters have distinct differences in gene ontology terms. Cluster one appears dominated by cell cycle and DNA replication terms. The immune response, defense response and cell communication terms appear to be a major discriminating feature with prominent asymmetric distribution across the gene clusters. Development, morphogenesis and differentiation terms are also major class separating terms in the gene clusters.
Table 1 Distribution of Ontology terms within Gene Clusters. The gene clusters identified in Figure 1d were analyzed for asymmetric distribution of ontology terms using the Gominer Software [16]. The top 40 gene ontology terms for each cluster ranked by significance scoring (Fishers exact T-test) are shown. Total numbers of genes in each cluster are indicated in parentheses. Statistical ranking of asymmetrically distributed gene ontology terms is represented by an estimated p-value (Fisher's Exact T-test).
Cluster One (1187)
P-Value Ontology Term
0.0003 DNA dependent DNA replication
0.0003 mitotic cell cycle
0.0008 DNA replication
0.001 structural constituent of cytoskeleton
0.0014 metabolism
0.0015 proteolysis and peptidolysis
0.0016 cell cycle
0.0016 hydrolase activity
0.0016 S phase of mitotic cell cycle
0.0021 protein metabolism
0.0024 protein catabolism
0.0028 DNA replication and chromosome cycle
0.0029 small ribosomal subunit
0.0031 intracellular
0.0031 extracellular
0.004 DNA replication factor C complex
0.0059 nucleic acid binding activity
0.0059 ATP dependent helicase activity
0.006 transmembrane receptor protein phosphatase activity
0.006 transmembrane receptor protein tyrosine phosphatase activity
0.0061 cell proliferation
0.0063 mitochondrial inner membrane
0.0065 extracellular space
0.0065 macromolecule catabolism
0.0071 protein phosphatase activity
0.0073 nucleobase, nucleoside, nucleotide and nucleic acid metabolism
0.0074 replication fork
0.0078 protein amino acid dephosphorylation
0.0078 dephosphorylation
0.0078 protein-ligand dependent protein catabolism
0.0081 mitochondrial ribosome
0.009 inner membrane
0.0092 mitochondrion
0.0095 cellular_component unknown
0.0111 helicase activity
0.0113 organellar ribosome
0.0123 N-linked glycosylation
0.0123 di-, tri-valent inorganic cation homeostasis
0.014 proton-transporting ATP synthase complex
0.014 spindle
Cluster Nine (724)
P-Value Ontology Term
0.0003 mitochondrion
0.0005 metabolism
0.0008 intracellular
0.0018 biosynthesis
0.0022 complement activation, alternative pathway
0.003 complement activation
0.0044 complement activity
0.0047 sugar binding activity
0.0047 carbohydrate binding activity
0.006 humoral defense mechanism (sensu Vertebrata)
0.0067 plasma membrane
0.007 cell adhesion molecule activity
0.0071 1-phosphatidylinositol 3-kinase complex
0.0071 membrane attack complex
0.0071 hydrolase activity, acting on acid anhydrides, catalyzing transmembrane movement of substances
0.0071 phosphatidylinositol 3-kinase activity
0.0079 ATP-binding cassette (ABC) transporter activity
0.0098 cell adhesion
0.0099 chemotaxis
0.0099 taxis
0.0125 cell-cell adhesion
0.013 mitochondrial membrane
0.0151 lectin
0.0156 G-protein coupled receptor protein signaling pathway
0.0176 cellular_component unknown
0.0187 P-P-bond-hydrolysis-driven transporter activity
0.02 thyroid hormone generation
0.02 lipid raft
0.02 ethanol oxidation
0.02 ethanol metabolism
0.02 flowering
0.02 thyroid hormone metabolism
0.02 aldo-keto reductase activity
0.02 alcohol dehydrogenase activity, iron-dependent
0.02 alcohol dehydrogenase activity, metal ion-independent
0.02 T-cell differentiation
0.02 negative regulation of Wnt receptor signaling pathway
0.02 fluid secretion
0.022 homophilic cell adhesion
0.0266 humoral immune response
Cluster Four (271)
P-Value Ontology Term
0.0002 cytoplasm
0.001 transcription
0.0012 regulation of transcription, DNA-dependent
0.0013 regulation of transcription
0.0015 transcription, DNA-dependent
0.0029 immune response
0.0029 nucleus
0.0034 transferase activity, transferring sulfur-containing groups
0.0034 solute:sodium symporter activity
0.005 defense response
0.0051 phenol metabolism
0.0051 catecholamine metabolism
0.0051 organic acid transporter activity
0.0053 cell communication
0.0055 response to biotic stimulus
0.0059 protein modification
0.0063 protein kinase CK2 activity
0.0069 solute:cation symporter activity
0.0071 response to external stimulus
0.0084 negative regulation of transcription
0.0093 biogenic amine metabolism
0.0093 adherens junction
0.0096 cAMP-dependent protein kinase activity
0.0096 cyclic-nucleotide dependent protein kinase activity
0.0096 casein kinase activity
0.0097 transcription from Pol II promoter
0.0099 secretin-like receptor activity
0.0099 neurotransmitter:sodium symporter activity
0.0099 neurotransmitter transporter activity
0.0099 biogenic amine biosynthesis
0.0103 protein amino acid phosphorylation
0.0106 G-protein coupled receptor activity
0.0112 neurogenesis
0.0119 transmembrane receptor protein serine/threonine kinase signaling pathway
0.0128 phosphorylation
0.0139 small GTPase mediated signal transduction
0.0141 protein kinase activity
0.0151 brain development
0.016 frizzled receptor signaling pathway
0.016 frizzled receptor activity
Cluster Ten (815)
P-Value Ontology Term
<.0001 nucleobase, nucleoside, nucleotide and nucleic acid metabolism
<.0001 nucleus
<.0001 intracellular
<.0001 extracellular space
<.0001 extracellular
<.0001 RNA binding activity
<.0001 nucleic acid binding activity
0.0001 plasma glycoprotein
0.0001 oxidoreductase activity, acting on the CH-NH2 group of donors, oxygen as acceptor
0.0003 oxidoreductase activity, acting on the CH-NH2 group of donors
0.0003 molecular_function
0.0003 alpha-type channel activity
0.0004 response to external stimulus
0.0004 channel/pore class transporter activity
0.0005 chymotrypsin activity
0.0005 RNA metabolism
0.0007 trypsin activity
0.0011 metabolism
0.0014 immune response
0.0015 defense response
0.0016 RNA processing
0.0025 response to biotic stimulus
0.0028 cell surface receptor linked signal transduction
0.0028 integral to membrane
0.0031 regulation of transcription
0.0032 transcription
0.0037 signal transducer activity
0.0039 translation regulator activity
0.004 regulation of transcription, DNA-dependent
0.004 membrane
0.0042 voltage-gated ion channel activity
0.0044 ligand-dependent nuclear receptor activity
0.0044 potassium channel activity
0.0044 steroid hormone receptor activity
0.0045 ion transport
0.005 small GTPase mediated signal transduction
0.0051 nucleoplasm
0.0052 cation channel activity
0.0054 digestion
0.0058 ligand-regulated transcription factor activity
Cluster Six (474)
P-Value Ontology Term
0.0002 development
0.0002 extracellular matrix structural constituent
0.0003 muscle development
0.0004 muscle contraction
0.0007 intramolecular isomerase activity
0.0013 cell differentiation
0.0014 mitochondrion
0.002 cellular process
0.002 organogenesis
0.0022 cell adhesion
0.0027 cytoskeleton
0.0032 oncogenesis
0.0032 structural constituent of cytoskeleton
0.0033 cell communication
0.0036 morphogenesis
0.0037 troponin complex
0.0037 NGF/TNF (6 C-domain) receptor activity
0.0042 circulation
0.0046 structural molecule activity
0.0048 actin cytoskeleton
0.0049 cell motility
0.005 muscle fiber
0.0056 photoreceptor activity
0.0056 G-protein coupled photoreceptor activity
0.0056 collagen type I
0.011 intermediate filament cytoskeleton
0.011 intermediate filament
0.0125 transcription cofactor activity
0.0128 extracellular matrix structural constituent conferring tensile strength activity
0.0128 sarcomere
0.0128 myofibril
0.0128 collagen
0.0139 response to stress
0.0149 hydrolase activity
0.016 intramolecular isomerase activity, interconverting aldoses and ketoses
0.016 phosphagen metabolism
0.016 neurofilament
0.016 galactose binding lectin
0.016 inactivation of MAPK
0.0176 striated muscle thin filament
Cluster Twelve (619)
P-Value Ontology Term
<.0001 cell communication
0.0001 signal transduction
0.0078 development
0.0103 phosphate metabolism
0.0103 phosphorus metabolism
0.0159 neurogenesis
0.016 cell adhesion
0.0179 intracellular signaling cascade
0.0196 amino acid transport
0.0311 small GTPase mediated signal transduction
0.0384 coreceptor activity
0.0464 heme-copper terminal oxidase activity
0.0464 acute-phase response
0.0464 regulation of metabolism
0.0476 cell-cell signaling
0.085 beta3-adrenergic receptor activity
0.085 purine ribonucleoside catabolism
0.085 purine ribonucleoside metabolism
0.085 pentose catabolism
0.085 pentose metabolism
0.085 ribose catabolism
0.085 adenosine metabolism
0.085 manganese ion transport
0.085 ADP-sugar diphosphatase activity
0.085 bile acid biosynthesis
0.0858 cellular respiration
0.094 organelle organization and biogenesis
0.0966 alcohol catabolism
0.1096 xenobiotic metabolism
0.1096 neuropeptide signaling pathway
0.1105 meiosis
0.1136 deaminase activity
0.1198 synaptic transmission
0.1215 transmission of nerve impulse
0.1314 monovalent inorganic cation transporter activity
0.1491 chloride transport
0.1627 internalization receptor activity
0.1627 regulation of mitotic cell cycle
0.1627 cAMP metabolism
0.1627 regulation of cell volume
Cluster Thirteen (1208)
P-Value Ontology Term
0.0002 mitochondrion
0.0004 intracellular
0.0008 metabolism
0.0012 extracellular
0.0026 DNA repair
0.0031 immune response
0.0041 extracellular space
0.0045 phosphatidylinositol transporter activity
0.0061 cytosolic large ribosomal subunit (sensu Eukarya)
0.0065 defense response
0.0069 nucleobase, nucleoside, nucleotide and nucleic acid metabolism
0.0071 RNA binding activity
0.0078 large ribosomal subunit
0.0083 heme biosynthesis
0.0083 sex determination
0.0095 G-protein coupled receptor protein signaling pathway
0.0097 integral to membrane
0.0098 biosynthesis
0.0101 integral to plasma membrane
0.0109 mitotic cell cycle
0.0147 pigment biosynthesis
0.0147 post Golgi transport
0.015 nucleus
0.0157 cyclohydrolase activity
0.0157 protein amino acid methylation
0.0157 RNA-nucleus export
0.0157 transferase activity, transferring pentosyl groups
0.0169 porphyrin biosynthesis
0.0169 chromatin remodeling complex
0.0169 heme metabolism
0.0178 plasma membrane
0.0192 S phase of mitotic cell cycle
0.0193 coenzymes and prosthetic group biosynthesis
0.0209 cell surface receptor linked signal transduction
0.021 ion transport
0.0233 trypsin activity
0.0234 pigment metabolism
0.0236 inorganic anion transport
0.0266 apoptosis regulator activity
0.0268 nucleic acid binding activity
Cluster Fifteen (725)
P-Value Ontology Term
0.0007 blood vessel development
0.0007 angiogenesis
0.001 phosphotransferase activity, alcohol group as acceptor
0.0013 nuclear localization sequence binding activity
0.0017 protein kinase activity
0.002 response to pest/pathogen/parasite
0.0023 protein serine/threonine kinase activity
0.0024 kinase activity
0.0025 cellular process
0.0028 cell migration
0.0038 actin polymerization and/or depolymerization
0.0048 spermatid development
0.0048 NLS-bearing substrate-nucleus import
0.0048 galactosyltransferase activity
0.0051 signal transduction
0.0053 protein tyrosine kinase activity
0.0059 embryogenesis and morphogenesis
0.006 neurogenesis
0.007 immune response
0.0073 cell-matrix adhesion
0.0073 nucleotide binding activity
0.0077 Golgi apparatus
0.0079 transferase activity, transferring phosphorus-containing groups
0.0088 phosphate metabolism
0.0088 phosphorus metabolism
0.0091 protein amino acid phosphorylation
0.0097 response to wounding
0.0097 response to biotic stimulus
0.0106 phosphorylation
0.0111 RAN protein binding activity
0.0112 morphogenesis
0.0113 development
0.0113 purine nucleotide binding activity
0.012 actin filament-based process
0.0121 importin, beta-subunit
0.0121 actin modulating activity
0.0121 actin monomer binding activity
0.0121 regulation of actin polymerization and/or depolymerization
0.0124 cytoskeleton organization and biogenesis
0.0137 cell communication
Cluster Sixteen (1266)
P-Value Ontology Term
0.0004 immune response
0.0008 oncogenesis
0.0009 defense response
0.0042 ionic insulation of neurons by glial cells
0.0125 inflammatory response
0.0245 histogenesis and organogenesis
0.0261 sarcomere alignment
0.0261 phagocytosis, engulfment
0.0261 negative regulation of osteoclast differentiation
0.0261 regulation of osteoclast differentiation
0.0261 negative regulation of cell differentiation
0.0261 NO mediated signal transduction
0.0326 activation of NF-kappaB-inducing kinase
0.0327 oogenesis
0.0453 cell activation
0.0483 humoral immune response
0.0491 protein modification
0.0575 regulation of cell differentiation
0.0673 cell cycle
0.07 biotin metabolism
0.0806 phosphate metabolism
0.0806 phosphorus metabolism
0.0888 sensory organ development
0.0888 G-protein signaling, adenylate cyclase activating pathway
0.1073 pattern specification
0.1111 gametogenesis
0.119 peptide receptor activity
0.1221 microtubule-based process
0.1251 phosphate transport
0.1251 glutathione conjugation reaction
0.1251 G-protein chemoattractant receptor activity
0.1256 phagocytosis
0.1256 carbohydrate kinase activity
0.1299 regulation of transcription
0.1309 fatty acid metabolism
0.1435 antimicrobial humoral response (sensu Invertebrata)
0.1435 protein amino acid phosphorylation
0.1454 NIK-I-kappaB/NF-kappaB cascade
0.1507 protein phosphatase type 2C activity
0.1507 heavy metal ion transport
To determine which TFBSs were most important for discriminating the different gene clusters, the 164 motifs were ranked for significance in each cluster by ANOVA assigned significance based on discriminatory power. The significance ranking was derived from the p-value output for each motif in each cluster and then converted to a color score based on the ranking (1–164 = low-high = blue-red). A contour heat diagram showing the differential ranking of the 164 motifs in each of the 9 clusters is shown in Figure 2a. As apparent from this heat diagram, the TFBS patterns of the majority of the clusters produce very distinct signatures (Figure 2a).
Figure 2 Significance ranking of TFBSs in respective clusters. (a) The TFBSs in the sequences of each cluster were sorted and ranked by ANOVA analysis to determine those sites that best discriminated the different clusters. The TFBSs in each cluster were then assigned ranks (1–164) according to their significance (p-value) from the ANOVA analysis. Highest ranking in red, lowest in blue. (b) Partitioning of gene promoter composition with more stringent PWM matrix similarity thresholds reduces the number of clusters identified by K-means analysis. Shown are four of six clusters containing greater than 4 genes. (groups 1, 2, 5 and 6). (c) Analysis of the most discriminating TFBSs in the four clusters in Figure 2b by ANOVA, as in Figure 2a.
When the TFBS composition of the 7,298 promoter regions was scored using more stringent PWM thresholds optimized to yield fewer potential false positive predictions (Methods), a surprising decrease in the diversity of the clustering pattern for the promoters was observed (Figure 2b). The analysis predicted 6 clusters from which two were discarded for having fewer than 4 genes. As expected, a heat diagram of the 164 TFBS rankings in each of the clusters showed considerably less distinct signatures (Figure 2c). Moreover, the ontology terms associated with the 4 clusters were much less distinct with a higher total ratio of redundant terms (see supplemental data, Table two). This tendency for more relaxed PWM similarity thresholds to generate greater diversity in predicted promoter composition suggests that the inherent or "perceived" degenerate nature of transcription factor binding sites serves to broaden the potential "categories" or strategies of gene regulation [17]. This suggests high thresholds, though reducing the number of potential false positives, have the severe negative effect of overlooking real binding sites [17].
A logical prediction in this study is that the associated biological function of the transcription factors that regulate the gene groups should share some similarity with the function of the genes that they regulate. In other words, the "regulator" should show similar function to the "regulated". To ask this question, we looked for any correlation between the ontology terms of the transcription factors (TFs) predicted most likely to bind to the promoter regions of the gene clusters (TFO) and the ontology terms of the gene clusters themselves (GCO). Accordingly, a list of transcription factors known to recognize the most discriminating TFBSs for each cluster was generated (total of 55 TF genes, Figures 3, 4, 5, 6, 7). The top 10 TFBSs were segregated into over-represented (RED) or under-represented (GREEN) groups for their respective gene clusters. The list of genes encoding the transcription factors that bind the top 10 TFBSs in each cluster was then compiled based on TRANSFAC 6.0 annotation. The GoMiner software was then used to rank the gene ontology terms based on the statistical significance of their occurrence within the transcription factor clusters (TFO). The top gene ontology terms with a ranking p-value less than 0.05 (determined by GoMiner) were then extracted and listed depending on whether they were over-represented (RED) or under-represented (GREEN) in each TFO. These lists were then compared with the top gene ontology terms of each respective gene cluster (gene cluster ontology, GCO) with p-values less then 0.05. The over-represented transcription factor ontology terms (TFO) found to share similarity with terms in the respective gene cluster ontologies (GCO) (within two branches of the ontology clade) are displayed in bold capitals letters (Figures 3, 4, 5, 6, 7, right column).
Figure 3 Analysis of Ontology term distribution. The top 20 best discriminating gene ontology terms in each cluster were sorted for over-representation (RED) and under-representation (Green) and compared to the top 10 discriminating TFBSs for each cluster as determined by ANOVA (Figure 2). The top 10 over-represented (Red) and under-represented (Green) TFBSs for each cluster are shown. The transcription factors that recognize the TFBSs were grouped and then analyzed for asymmetric distribution of ontology terms using GoMiner (TF ontology terms, right). Transcription factor genes that are known to bind the over-represented TFBSs (TF Genes, enriched) are shown enclosed in boxes. Transcription factor ontology terms that overlap the gene cluster ontology terms within 2 branches of the ontology clade are shown in bold. Those terms with exact matches in the gene cluster ontologies are indicated with an asterisk. The numbers in parentheses indicate the total number of ontology terms associated with each respective cluster. The numbers in brackets indicated those ontology terms with a significance measurement p-value < 0.05 (Fisher Exact T-test). Representative genes from Clusters one, six and thirteen are shown in supplemental Table 3.
Figure 4 Analysis of Ontology term distribution. The top 20 best discriminating gene ontology terms in each cluster were sorted for over-representation (RED) and under-representation (Green) and compared to the top 10 discriminating TFBSs for each cluster as determined by ANOVA (Figure 2). The top 10 over-represented (Red) and under-represented (Green) TFBSs for each cluster are shown. The transcription factors that recognize the TFBSs were grouped and then analyzed for asymmetric distribution of ontology terms using GoMiner (TF ontology terms, right). Transcription factor genes that are known to bind the over-represented TFBSs (TF Genes, enriched) are shown enclosed in boxes. Transcription factor ontology terms that overlap the gene cluster ontology terms within 2 branches of the ontology clade are shown in bold. Those terms with exact matches in the gene cluster ontologies are indicated with an asterisk. The numbers in parentheses indicate the total number of ontology terms associated with each respective cluster. The numbers in brackets indicated those ontology terms with a significance measurement p-value < 0.05 (Fisher Exact T-test). Representative genes from Clusters one, six and thirteen are shown in supplemental Table 3.
Figure 5 Analysis of Ontology term distribution. The top 20 best discriminating gene ontology terms in each cluster were sorted for over-representation (RED) and under-representation (Green) and compared to the top 10 discriminating TFBSs for each cluster as determined by ANOVA (Figure 2). The top 10 over-represented (Red) and under-represented (Green) TFBSs for each cluster are shown. The transcription factors that recognize the TFBSs were grouped and then analyzed for asymmetric distribution of ontology terms using GoMiner (TF ontology terms, right). Transcription factor genes that are known to bind the over-represented TFBSs (TF Genes, enriched) are shown enclosed in boxes. Transcription factor ontology terms that overlap the gene cluster ontology terms within 2 branches of the ontology clade are shown in bold. Those terms with exact matches in the gene cluster ontologies are indicated with an asterisk. The numbers in parentheses indicate the total number of ontology terms associated with each respective cluster. The numbers in brackets indicated those ontology terms with a significance measurement p-value < 0.05 (Fisher Exact T-test). Representative genes from Clusters one, six and thirteen are shown in supplemental Table 3.
Figure 6 Analysis of Ontology term distribution. The top 20 best discriminating gene ontology terms in each cluster were sorted for over-representation (RED) and under-representation (Green) and compared to the top 10 discriminating TFBSs for each cluster as determined by ANOVA (Figure 2). The top 10 over-represented (Red) and under-represented (Green) TFBSs for each cluster are shown. The transcription factors that recognize the TFBSs were grouped and then analyzed for asymmetric distribution of ontology terms using GoMiner (TF ontology terms, right). Transcription factor genes that are known to bind the over-represented TFBSs (TF Genes, enriched) are shown enclosed in boxes. Transcription factor ontology terms that overlap the gene cluster ontology terms within 2 branches of the ontology clade are shown in bold. Those terms with exact matches in the gene cluster ontologies are indicated with an asterisk. The numbers in parentheses indicate the total number of ontology terms associated with each respective cluster. The numbers in brackets indicated those ontology terms with a significance measurement p-value < 0.05 (Fisher Exact T-test). Representative genes from Clusters one, six and thirteen are shown in supplemental Table 3.
Figure 7 Analysis of Ontology term distribution. The top 20 best discriminating gene ontology terms in each cluster were sorted for over-representation (RED) and under-representation (Green) and compared to the top 10 discriminating TFBSs for each cluster as determined by ANOVA (Figure 2). The top 10 over-represented (Red) and under-represented (Green) TFBSs for each cluster are shown. The transcription factors that recognize the TFBSs were grouped and then analyzed for asymmetric distribution of ontology terms using GoMiner (TF ontology terms, right). Transcription factor genes that are known to bind the over-represented TFBSs (TF Genes, enriched) are shown enclosed in boxes. Transcription factor ontology terms that overlap the gene cluster ontology terms within 2 branches of the ontology clade are shown in bold. Those terms with exact matches in the gene cluster ontologies are indicated with an asterisk. The numbers in parentheses indicate the total number of ontology terms associated with each respective cluster. The numbers in brackets indicated those ontology terms with a significance measurement p-value < 0.05 (Fisher Exact T-test). Representative genes from Clusters one, six and thirteen are shown in supplemental Table 3.
A qualitative comparison of the gene cluster ontology terms and their respective transcription factor ontology terms reveals several similarities in the over-represented terms. Cluster one shows significant overlap of ontology terms for cell division. Cluster six shows overlapping terms for cell communication. Cluster twelve contained overlapping terms for cellular metabolism. Cluster thirteen shows a puzzling anti-correlation with response to external stimuli. Cluster fifteen shows overlapping terms with morphogenesis and development. When these correlations are tested for significance by the method of hyper-geometric distribution, Clusters one and six shows statistically significant correlation. Within Cluster one, gene cluster and transcription factor ontology terms for cell cycle regulation overlapped significantly (p = 1.08E-04). Within Cluster six, there was substantial overlap for cell communication ontology terms (p = 0.0030). Both cell cycle regulation and cell communication encompass fundamental and highly conserved processes in mammalian cells. Less than a third of clusters showed statistically significant correlations between gene group and transcription factor ontology terms. Nonetheless, given the unbiased manner in the which the gene lists and TF lists were generated and the small number of TF genes used to generate that TFO terms (55) compared to the number used to generate the GCO terms (7298), this approach shows substantial promise for identifying functional correlations between the transcriptional pathways and the genes regulated by them. It is reasonable to anticipate that these correlations will strengthen as the number and quality of the PWMs expand and the transcription factor gene ontology annotation improves in number and accuracy (see Discussion).
Discussion
Changes or alteration in gene expression are often linked to influences at the regulatory elements within the promoter regions of the targeted genes. The transcription factors that bind these regulatory elements form the final controlling functional link to the signaling pathways that are triggered and integrated as the cell adapts to environmental change. Thus, collective control of these integrated pathways forms the major conduit that governs changes and patterns of cellular behavior. These relationships are particularly applicable to metazoan systems.
Transcriptional control in metazoan cells is the culmination of multiple signal-induced transcriptional pathways, where the collective influence of more than one transcription factor and pathway hold sway on the ultimate expression of targeted genes. This combinatorial logic provides a means through which a finite number of transcriptional pathways can converge to produce seemingly infinite patterns of gene regulatory control. Deciphering this logic and how it links downstream function to upstream signaling requires expanded methods of interpreting promoter composition. By classifying patterns of promoter composition and linking these classifications to functional categories, of both the regulated genes and the transcription factors that regulate them, this approach provides a rational method for identifying meaningful relationships between promoter composition and gene function.
Though only 2 of 9 clusters showed a statistically significant correlation between ontology terms of the clusters and the transcription factors (Figures 3, 4, 5, 6, 7), an inspection of the ontology terms of several of the gene clusters in comparison to the transcription factors reveals numerous relationships that have been well established in the literature, though not reflected in the currently available ontological annotation for the factors. Cluster one is dominated by E2F transcription factors that are well known to exert control over genes involved in cell cycle regulation. Therefore, the overlap between Cluster one gene and transcription factor ontology terms for cell cycle regulation are significant (p-value = 1.08E-04). Cluster Four showed no matches in the most significant ontology terms, however, the significant potential regulators of this cluster include AP-2 transcription factors, which have broad roles in vertebrate development including control of apoptosis and cell cycle [18]. Moreover, AP-2 factors have been also found to control receptor tyrosine kinase expression and other factors involved in the negative regulation of gene expression [19,20]. EGR1 and ZNF42 factors are widely known to regulate genes important for mitogenesis and differentiation [21-23]. Thus, a more expanded annotation of these terms would have shown greater correlation with the top ontology terms in both Cluster four and Cluster fifteen. These include cell communication, negative regulation of transcription, protein modification, protein tyrosine kinase activity, embryogenesis, morphogenesis, signal transduction and angiogenesis (Figures 3 and 6). Even though Cluster six shows statistically significant overlap between its ontology terms and those of its potential regulating transcription factors, many seemingly obvious matches could not be found in the annotation of some of the potential regulators. In particular, there is a significant absence of ontology terms for oncogenesis, morphogenesis and cellular differentiation for the NF-kappa B family subset of the top discriminating transcription factors. Control of these cellular process are well described for NF-kappa B [24]. The transcription factors in Cluster nine show no overlap; yet, it is dominated by octamer binding sites and several reports indicate octamer family members have a role in the control of expression of cellular adhesion molecules and other participants in wound healing [25,26]. In Cluster ten, the roles for TCF3 and TCF8 in early B-cell differentiation, immunoglobin expression and T-cell function certainly should have produced an overlap with the gene cluster ontology terms for immune response and defense response [27-29].
Another dominant factor that will improve the deduced linkages between ontologies of the regulating transcription factors and the regulated genes will be improvements in the accuracy of predicting TFBS occurrence. Multiple difficult factors have to be addressed. The first is accurate prediction of the promoter regions themselves. In this work, we define the promoter region in terms of the start of transcription (TSS) and retrieve sequence 200 bp downstream and 1200 bp upstream of this position. Using the ProSpector search engine, the TSS is extracted from RefSeq annotation provided by the USCS genome assembly [30]. More precise identification of TSS is available from recent curated databases containing empirically derived TSS positions such as MPromDb , OMGProm , and DBTSS [31,32]. These resources will certainly improve on the accuracy of the promoter identification as their inventories continue to grow from the current 8,793 (DBTSS) and 13,780 (MPromDb) human genes. Nonetheless, a comparison of the promoter sequences queried from ProSpector and those from MPromDb showed a greater than 80% overlap in more than 80% of the mutually retrieved sequences (data not shown). It should be noted that metazoan promoter regions are highly complex and have multiple different TSS positions and consequently multiple promoters [33]. Many of these alternate promoters are tissue specific [33]. This feature unavoidably confounds the approach and is not adequately addressed in current promoter analysis tools. In addition to alternate promoters, metazoan gene regulatory regions are influenced by distant enhancer regions, locus control regions and a complicated tissue-specific interplay between transcriptional co-activator complexes and transcription factors [1,34-36]. These complex factors probably account for the better performance of promoter prediction tools on yeast data sets in comparison to higher eukaryotes [37].
A particularly difficult problem with the use of PWMs to annotate gene regulatory regions is the unavoidable occurrence of "false positive" and "false negatives". This is predominantly the case when searching for new TFBSs in uncharacterized genetic regions. Figures 2b and 2c show that using a high PWM threshold has the negative result of reduced promoter discrimination and potentially high levels of true "false negatives". At the heart of the matter is how we discriminate true false negatives and positives. It is indisputable that this can only be done through empirical validation and verification. The thresholds set for the PWM analysis in figures 2b and 2c were too high to detect known sites for CREB, AP1, NF-kappa B and NFAT in the IL2 promoter and failed to retrieve any of the 5 known sites for NFAT in the IL4 promoter [38-43]. Thus, the use of high thresholds is inappropriate. For empirically uncharacterized gene regulatory regions, there is no way to discriminate between correct identifications, true false positives or false negatives. High frequency occurrence of motifs in some promoters should be met with some skepticism, but it is important to keep in mind that our understanding of transcription factor interaction with the genome continues to evolve. The interaction between transcription factors and TFBSs is not static, but highly dynamic and repetitive [44]. Thus, gradients of high and low affinity binding sites for classes of factors within a single gene regulatory locus could be physiologically relevant.
The presence of GC rich regions and CpG islands creates an important issue that requires consideration. These types of regions contain high densities of binding sites for factors such as Sp1, AP2, and EGR2/3. Though greater than 80% of promoters of are thought to contain CpG islands [45,46], differences in their presence, length or position will lead to background noise in the analysis. Recently developed approaches are able to address this problem through the use of background models representing either the entire genome (which is still subject to GC rich asymmetry because of the preferential concentration at transcription start sites) or random/unselected groups of promoter regions [47,48]. The method described in this current study ranks motifs not by their PWM score, but by using ANOVA to discriminate across the opposing clusters. By this approach, the aggregate of the opposing clusters serves as the background model for discriminatory significance of the TFBSs within each group. Though the presence of GC rich regions contribute significant noise to the analysis, this problem does not overwhelm the approach since it robustly discriminates true differences in promoter composition and correctly groups genes of known ontology with those containing mutual TFBSs that have been empirically validated (see supplemental Table 3).
Only Clusters fifteen, thirteen and four show high ranks for GC-containing TFBSs (Figures 3 and 6). At the very least, this indicates that there is a non-random distribution of GC rich regions amongst promoters. Nonetheless, the relative contribution of GC rich tracts or CpG islands to this distribution cannot be determined by our method. As expected, the clusters with high ranks for GC-containing TFBSs are in close apposition (Figure 1d). The fact that they show rather low correlation between GCO and TFO reflects the noise due to the high occurrence of GC rich regions. Still, it must be emphasized that CpG islands represent legitimate sites for factors like Sp1, EGR3/2 and AP2. Thus, the detection of such binding sites is likely to be physiologically important and their clustering patterns may contain biological information that will increase in importance as our understanding of transcription factor ontology is refined. Interestingly, recent studies suggest that genes lacking CpG islands tend to be expressed with a higher degree of tissue specificity and contain more GO terms consistent with "signal transducer", confirming the speculation that many CpG islands are associated with house-keeping gene function [46].
Unlike the yeast studies of Tavazoie et al [49], our approach failed to show any correlation between gene expression patterns at the RNA level and promoter TFBS composition. This result could be due in large part to the differences in yeast and metazoan gene regulatory regions as discussed above. In addition, post-transcriptional regulation of RNA stability is likely to be much more complex in metazoans than yeast. Another very important consideration is that Tavazoie et al chose to study the cell cycle, a time series of cellular behavior that is rich in various distinct molecular programs. It may be that the use of time-dependent changes following mitogen stimulation is too broad and lacks sufficient variability and distinction of gene expression to provide the discriminatory power necessary for the analysis of gene regulatory regions. Recently, another group has made an elegant application of GO terms to predict biological function by promoter composition [50]. By this approach, Bluthgen et al used predetermined TFBS combinations of known biological significance to extract genes with similar biological function based on overlapping promoter composition. This approach is very promising, confirms our central hypothesis, and shares similarity with a previously reported method where the biological significance of TFBS combinations, derived from kinetic profiles of transcriptional regulator occupancy via chromatin immuno-precipitation, was used to identify similarly regulated genes [14]. Our analysis differs from Bluthgen et al in that it does not depend on prior knowledge. In contrast, it begins with neither a pre-selected TFBS framework nor any other biological information. The possibility of identifying previously unrecognized ontologies linking the targeted genes with their targeting transcription factors remains preserved.
Conclusion
The examination of the upstream regulatory sequences of eukaryotic genes has the potential of yielding a wealth of information that will unravel the transcriptional control codes that govern spatial and temporal changes in gene expression. Combining multivariate analysis of promoter composition with classification by gene ontology provides a method that defines functional links between regulated genes and the genes that regulate them. The above are just a few of the examples where expanded ontology term annotation for the transcription factors based on current literature and improved promoter annotation methods will enhance the functional correlation between the regulator and the regulated. Just as important, these examples also point out how this method may also aid in identifying previously unrecognized functions for known transcription factors through the identification of "mutual ontology" terms. Ultimately, it will be the broader refinement and expansion of both PWMs for TFBSs and the functional vocabulary for all genes (in particular those gene encoding transcription factors), that will have a significant impact on improving the utility of this approach.
Methods
ProSpector database
To aid in the extraction and analysis of human promoter regions, a web-based resource named ProSpector (PROmoter inSPECTion) was developed [51]. The ProSpector website operates through an Apache web server and a MySQL relational database. The user interface of the website is written in PHP. The website provides a search tool for retrieving oriented human gene promoter regions by gene name (HUGO), gene description, gene symbol (HUGO), UniGene cluster, RefSeq ID, or LocusLink ID. The search feature is facilitated by keeping local copies of NCBI databases, including UniGene, RefSeq, and LocusLink. The actual promoter sequences are retrieved through a tool developed by the Genome Analysis Unit of the National Cancer Institute [52]. This tool retrieves promoters by extracting regions 5' of gene transcription start sites identified by RefSeq annotation. Transcription start is defined by RefSeq mRNAs aligned with genomic chromosomal contigs from the UCSC/NCBI Assembly – hg12/Build 30. ProSpector also allows extracted promoters to be analyzed for putative transcription factor binding sites using the MatInspector algorithms described by Quandt et al [5] and a subset of the position weight matrices (PWM) from the TRANSFAC 6.0 public database of transcription factors [53]. Briefly, base composition at putative TFBS is calculated as a vector score (Ci(i)) where:
Ci(i)=(100/ln5)×[∑b∈A,C,G,T,gap[P(i,b)×lnP(i,b)]+ln5] (1)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@7707@
P(i, b) being the relative frequency of base (b) at position (i) calculated from the PWM. Core similarity is used to quickly screen for potential binding sites with a high similarity to the most conserved region of the PWM and is determined from the four consecutive bases in the PWM with the highest Ci and calculated using:
core_sim=[∑j=mm+3score(b,j)]/[∑j=mm+3max_score(j)] (2)0≤core_sim≤1
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@7F77@
The score for base (b) and position (j) is simply the matrix value of base (b) at position (j) as defined by the PWM and the max score is the highest value in the matrix at position (j). Likeness or similarity to the TFBS PWM is calculated independent of the core by:
mat_sim=[∑j=1nCi(j)×score(b,j)]/[∑j=1nCi(j)×max_score(j)] (3)0≤mat_sim≤1
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@87C9@
A subset of 164 PWMs representative of the major families of human transcription factors contained in TRANSFAC 6.0 was employed in this study.
Optimization of TRANSFAC thresholds
When promoter sequences are analyzed for potential transcription factor binding sites, they are selected based on their similarity to TRANSFAC position weight matrices. This similarity is represented by its matrix similarity (mat_sim, Equation 3). By setting a threshold, only potential binding sites with an equal or larger similarity are selected. Because each position weight matrix is of differing inherent degeneracy, an optimized matrix threshold was generated for each matrix to provide an alternative method to minimize the number of potential false positive binding sites. To optimize the matrix thresholds, 1,000,000 bases of random sequence were analyzed with each matrix at intervals of successively higher matrix thresholds. The optimum threshold was arbitrarily defined for each matrix to be the point at which the matrix was detected as scoring only one binding site per 1000 bp of the random DNA.
Microarray analysis
Analysis was performed on previously published microarray data [11] to generate a list of genes separated into groups based on specific steady-state mRNA expression levels. In these studies, the Jurkat human T-cell line was stimulated with phorbol ester and ionomycin. mRNA was isolated from cells at 0, 1, 2, 6, 12, and 24 hours after stimulation. The prefiltered hybridization signals (provided by Dieh et al [11]) were normalized and filtered to remove any spots marked as "bad" on any of the twelve arrays. The ratio expression data was then log transformed and standardized to a control set of hybridization signals from mRNA isolated from untreated Jurkat T-cells at each time point. The promoter regions (defined as 1200 bp upstream and 200 bp downstream from the transcription start site) of the resulting list of Unigene clusters were then retrieved using the ProSpector website. UniGene clusters for which no sequence could be found were discarded in this step. A final list of 7298 UniGene clusters and promoter regions remained. For analysis of gene expression data, the method of K-means clustering was used [54]. The optimum number of clusters for the data set (four) was determined using the method described by Davies and Bouldin [55].
Promoter analysis
The genes analyzed in the microarray analysis were also analyzed for putative transcription factor binding using the ProSpector website . The gene promoter regions spanning 1200 bp upstream of transcription start and 200 bp downstream of transcription start were analyzed with all 164 position weight matrices. The analysis was performed twice: once applying a common matrix threshold of 0.75 and again with the optimized matrix thresholds. In both analyses, a threshold of 1 for the core similarity was used. The results of the analysis were analyzed by segregating the genes based on the number of each position weight matrix that scored a binding site in the promoter. As in the microarray analysis, the method of Davies and Bouldin was used to determine the optimal number of clusters according to promoter composition. K-means clustering was used to segregate the genes. In the analysis with a blanket 0.75 matrix threshold, the data was found through Davies-Bouldin analysis to separate into 16 groups. These clusters remained stable after the introduction of normalized Gaussian noise up to two fold standard deviation. Seven of the groups were small, containing 1 to 4 genes, and were discarded as outliers. The final result was 9 groups. The analysis was also conducted using optimized thresholds (see above). This analysis segregated into 6 groups. Again, groups with 4 or less genes were discarded as outliers leaving 4 groups derived from the optimized threshold TFBS analysis. For each analysis, clustered genes were analyzed using ANOVA to determine the transcription factor binding site motifs most significant in differentiating the clusters. Principal component analysis (PCA), ANOVA (one-way) and K-means clustering analysis were performed using Partek Pro 5.0 (Partek Corp.).
Gene ontology
The GoMiner gene ontology tool was used to rank and categorize the gene ontology terms that were more significantly enriched beyond random assignment in each gene cluster [16]. Gene ontology terms were ranked according to p-values (Fisher's Exact T-test) generated by GoMiner [16]. Testing comparing ten random gene groups (1000 each) showed only random grouping of ontology terms versus the total population (data not shown). GoMiner was also used to identify those ontology terms that were enriched in the genes groups encoding the transcription factors known to associate with the most significant TFBSs identified in each respective cluster by ANOVA analysis. Significance testing for shared gene cluster ontology (GCO) terms and transcription factor ontology (TFO) terms was done by estimating the random probability of observing significant ontology terms in both TFO and GCO. The number of TFO terms were considered to be N, of which n are significant, and there are m significant GCO's found in TFO list of which k are also significant TFO's. The probability of obtaining k terms was tested if n values are randomly drawn from all TFO terms in which m are GCO's. The random process describing the null-hypothesis (no preferential overlap of GO terms) is described by the hypergeometric distribution which can be calculated by the phyper function of the R-Statistical software [56].
Implementation
The ProSpector promoter retrieval and annotation tool is available for open access at . ProSpector is compatible with Internet Explorer, Mozilla, Firefox, Opera and Safari.
Authors' contributions
MCM conceived of, designed, and implemented the ProSpector promoter annotation tool, carried out the promoter and ontology term analysis and assisted in the drafting of the manuscript. RT carried out statistics, promoter and ontology term analysis and assisted in the drafting of the manuscript. WC, IC, WJF, IM and CMH provided advice and assisted in the preparation of the manuscript. GVRC provided advice and performed much of the statistical analysis. KG conceived of the project, provided advice and opinions essential to the development of the analysis and helped draft the manuscript. All authors read and approved the final manuscript.
Supplementary Material
Additional file 1
Supplemental Table 2, ontology terms associated with the 4 clusters in Figure 2
Click here for file
Additional file 2
Supplemental Table 3, representative genes from Clusters one, six and thirteen
Click here for file
Acknowledgements
This research was supported by the Intramural Research Program of the NIH, National Cancer Institute.
==== Refs
Michelson AM Deciphering genetic regulatory codes: a challenge for functional genomics Proc Natl Acad Sci U S A 2002 99 546 548 11805309 10.1073/pnas.032685999
Harbison CT Gordon DB Lee TI Rinaldi NJ Macisaac KD Danford TW Hannett NM Tagne JB Reynolds DB Yoo J Jennings EG Zeitlinger J Pokholok DK Kellis M Rolfe PA Takusagawa KT Lander ES Gifford DK Fraenkel E Young RA Transcriptional regulatory code of a eukaryotic genome Nature 2004 431 99 104 15343339 10.1038/nature02800
Liu Y Wei L Batzoglou S Brutlag DL Liu JS Liu XS A suite of web-based programs to search for transcriptional regulatory motifs Nucleic Acids Res 2004 32 W204 W207 15215381
Liu XS Brutlag DL Liu JS An algorithm for finding protein-DNA binding sites with applications to chromatin-immunoprecipitation microarray experiments Nat Biotechnol 2002 20 835 839 12101404
Quandt K Frech K Karas H Wingender E Werner T MatInd and MatInspector: new fast and versatile tools for detection of consensus matches in nucleotide sequence data Nucleic Acids Res 1995 23 4878 4884 8532532
Hochheimer A Tjian R Diversified transcription initiation complexes expand promoter selectivity and tissue-specific gene expression Genes Dev 2003 17 1309 1320 12782648 10.1101/gad.1099903
Fan J Yang X Wang W Wood WHIII Becker KG Gorospe M Global analysis of stress-regulated mRNA turnover by using cDNA arrays Proc Natl Acad Sci U S A 2002 99 10611 10616 12149460 10.1073/pnas.162212399
Schones DE Sumazin P Zhang MQ Similarity of position frequency matrices for transcription factor binding sites Bioinformatics 2005 21 307 313 15319260 10.1093/bioinformatics/bth480
Harris MA Clark J Ireland A Lomax J Ashburner M Foulger R Eilbeck K Lewis S Marshall B Mungall C Richter J Rubin GM Blake JA Bult C Dolan M Drabkin H Eppig JT Hill DP Ni L Ringwald M Balakrishnan R Cherry JM Christie KR Costanzo MC Dwight SS Engel S Fisk DG Hirschman JE Hong EL Nash RS Sethuraman A Theesfeld CL Botstein D Dolinski K Feierbach B Berardini T Mundodi S Rhee SY Apweiler R Barrell D Camon E Dimmer E Lee V Chisholm R Gaudet P Kibbe W Kishore R Schwarz EM Sternberg P Gwinn M Hannick L Wortman J Berriman M Wood V de la CN Tonellato P Jaiswal P Seigfried T White R The Gene Ontology (GO) database and informatics resource Nucleic Acids Res 2004 32 D258 D261 14681407 10.1093/nar/gkh066
Ashburner M Mungall CJ Lewis SE Ontologies for biologists: a community model for the annotation of genomic data Cold Spring Harb Symp Quant Biol 2003 68 227 235 15338622 10.1101/sqb.2003.68.227
Diehn M Alizadeh AA Rando OJ Liu CL Stankunas K Botstein D Crabtree GR Brown PO Genomic expression programs and the integration of the CD28 costimulatory signal in T cell activation Proc Natl Acad Sci U S A 2002 99 11796 11801 12195013 10.1073/pnas.092284399
Kane LP Lin J Weiss A Signal transduction by the TCR for antigen Curr Opin Immunol 2000 12 242 249 10781399 10.1016/S0952-7915(00)00083-2
Lin Z Fillmore GC Um TH Elenitoba-Johnson KS Lim MS Comparative microarray analysis of gene expression during activation of human peripheral blood T cells and leukemic Jurkat T cells Lab Invest 2003 83 765 776 12808112
Smith JL Freebern WJ Collins I De Siervi A Montano I Haggerty CM McNutt MC Butscher WG Dzekunova I Petersen DW Kawasaki E Merchant JL Gardner K Kinetic profiles of p300 occupancy in vivo predict common features of promoter structure and coactivator recruitment Proc Natl Acad Sci U S A 2004 101 11554 11559 15286281 10.1073/pnas.0402156101
Alter O Brown PO Botstein D Singular value decomposition for genome-wide expression data processing and modeling Proc Natl Acad Sci U S A 2000 97 10101 10106 10963673 10.1073/pnas.97.18.10101
Zeeberg BR Feng W Wang G Wang MD Fojo AT Sunshine M Narasimhan S Kane DW Reinhold WC Lababidi S Bussey KJ Riss J Barrett JC Weinstein JN GoMiner: a resource for biological interpretation of genomic and proteomic data Genome Biol 2003 4 R28 12702209 10.1186/gb-2003-4-4-r28
Moses AM Chiang DY Kellis M Lander ES Eisen MB Position specific variation in the rate of evolution in transcription factor binding sites BMC Evol Biol 2003 3 19 12946282 10.1186/1471-2148-3-19
Hilger-Eversheim K Moser M Schorle H Buettner R Regulatory roles of AP-2 transcription factors in vertebrate development, apoptosis and cell-cycle control Gene 2000 260 1 12 11137286 10.1016/S0378-1119(00)00454-6
Bar-Eli M Gene regulation in melanoma progression by the AP-2 transcription factor Pigment Cell Res 2001 14 78 85 11310795 10.1034/j.1600-0749.2001.140202.x
Benson LQ Coon MR Krueger LM Han GC Sarnaik AA Wechsler DS Expression of MXI1, a Myc antagonist, is regulated by Sp1 and AP2 J Biol Chem 1999 274 28794 28802 10497252 10.1074/jbc.274.40.28794
Adamson E de BI Mittal S Wang Y Hayakawa J Korkmaz K O'Hagan D McClelland M Mercola D Egr1 signaling in prostate cancer Cancer Biol Ther 2003 2 617 622 14688464
Gaboli M Kotsi PA Gurrieri C Cattoretti G Ronchetti S Cordon-Cardo C Broxmeyer HE Hromas R Pandolfi PP Mzf1 controls cell proliferation and tumorigenesis Genes Dev 2001 15 1625 1630 11445537 10.1101/gad.902301
Hromas R Davis B Rauscher FJIII Klemsz M Tenen D Hoffman S Xu D Morris JF Hematopoietic transcriptional regulation by the myeloid zinc finger gene, MZF-1 Curr Top Microbiol Immunol 1996 211 159 164 8585946
Lin A Karin M NF-kappaB in cancer: a marked target Semin Cancer Biol 2003 13 107 114 12654254 10.1016/S1044-579X(02)00128-1
Copertino DW Jenkinson S Jones FS Edelman GM Structural and functional similarities between the promoters for mouse tenascin and chicken cytotactin Proc Natl Acad Sci U S A 1995 92 2131 2135 7534412
Iademarco MF McQuillan JJ Dean DC Vascular cell adhesion molecule 1: contrasting transcriptional control mechanisms in muscle and endothelium Proc Natl Acad Sci U S A 1993 90 3943 3947 7683412
Greenbaum S Zhuang Y Regulation of early lymphocyte development by E2A family proteins Semin Immunol 2002 14 405 414 12457613 10.1016/S1044532302000751
Genetta T Ruezinsky D Kadesch T Displacement of an E-box-binding repressor by basic helix-loop-helix proteins: implications for B-cell specificity of the immunoglobulin heavy-chain enhancer Mol Cell Biol 1994 14 6153 6163 8065348
Postigo AA Dean DC Independent repressor domains in ZEB regulate muscle and T-cell differentiation Mol Cell Biol 1999 19 7961 7971 10567522
2005
Palaniswamy SK Jin VX Sun H Davuluri RV OMGProm: a database of orthologous mammalian gene promoters Bioinformatics 2005 21 835 836 15531605 10.1093/bioinformatics/bti119
Suzuki Y Yamashita R Nakai K Sugano S DBTSS: DataBase of human Transcriptional Start Sites and full-length cDNAs Nucleic Acids Res 2002 30 328 331 11752328 10.1093/nar/30.1.328
Landry JR Mager DL Wilhelm BT Complex controls: the role of alternative promoters in mammalian genomes Trends Genet 2003 19 640 648 14585616 10.1016/j.tig.2003.09.014
Levine M Tjian R Transcription regulation and animal diversity Nature 2003 424 147 151 12853946 10.1038/nature01763
Butler JE Kadonaga JT The RNA polymerase II core promoter: a key component in the regulation of gene expression Genes Dev 2002 16 2583 2592 12381658 10.1101/gad.1026202
Butler JE Kadonaga JT Enhancer-promoter specificity mediated by DPE or TATA core promoter motifs Genes Dev 2001 15 2515 2519 11581157 10.1101/gad.924301
Tompa M Li N Bailey TL Church GM De Moor B Eskin E Favorov AV Frith MC Fu Y Kent WJ Makeev VJ Mironov AA Noble WS Pavesi G Pesole G Regnier M Simonis N Sinha S Thijs G van Helden J Vandenbogaert M Weng Z Workman C Ye C Zhu Z Assessing computational tools for the discovery of transcription factor binding sites Nat Biotechnol 2005 23 137 144 15637633 10.1038/nbt1053
Serfling E Avots A Neumann M The architecture of the interleukin-2 promoter: a reflection of T lymphocyte activation Biochim Biophys Acta 1995 1263 181 200 7548205
Shapiro VS Truitt KE Imboden JB Weiss A CD28 mediates transcriptional upregulation of the interleukin-2 (IL-2) promoter through a composite element containing the CD28RE and NF-IL-2B AP-1 sites Mol Cell Biol 1997 17 4051 4058 9199340
Butscher WG Powers C Olive M Vinson C Gardner K Coordinate transactivation of the interleukin-2 CD28 response element by c-Rel and ATF-1/CREB2 J Biol Chem 1998 273 552 560 9417115 10.1074/jbc.273.1.552
Sun YL Glimcher LH Hoey T Novel nfat sites that mediate activation of the interleukin-2 promoter in response to t-cell receptor stimulation Mol Cell Biol 1995 15 6299 6310 7565783
Powell JD Lerner CG Ewoldt GR Schwartz RH The -180 site of the IL-2 promoter is the target of CREB/CREM binding in T cell anergy J Immunol 1999 163 6631 6639 10586058
Rockman MV Hahn MW Soranzo N Loisel DA Goldstein DB Wray GA Positive selection on MMP3 regulation has shaped heart disease risk Curr Biol 2004 14 1531 1539 15341739 10.1016/j.cub.2004.08.051
Phair RD Misteli T High mobility of proteins in the mammalian cell nucleus Nature 2000 404 604 609 10766243 10.1038/35007077
Gardiner-Garden M Frommer M CpG islands in vertebrate genomes J Mol Biol 1987 196 261 282 3656447 10.1016/0022-2836(87)90689-9
Yamashita R Suzuki Y Sugano S Nakai K Genome-wide analysis reveals strong correlation between CpG islands with nearby transcription start sites of genes and their tissue specificity Gene 2005 350 129 136 15784181 10.1016/j.gene.2005.01.012
Stepanova M Tiazhelova T Skoblov M Baranova A A comparative analysis of relative occurrence of transcription factor binding sites in vertebrate genomes and gene promoter areas Bioinformatics 2005 21 1789 1796 15699025 10.1093/bioinformatics/bti307
Tullai JW Schaffer ME Mullenbrock S Kasif S Cooper GM Identification of transcription factor binding sites upstream of human genes regulated by the phosphatidylinositol 3-kinase and MEK/ERK signaling pathways J Biol Chem 2004 279 20167 20177 14769801 10.1074/jbc.M309260200
Tavazoie S Hughes JD Campbell MJ Cho RJ Church GM Systematic determination of genetic network architecture Nat Genet 1999 22 281 285 10391217 10.1038/10343
Bluthgen N Kielbasa SM Herzel H Inferring combinatorial regulation of transcription in silico Nucleic Acids Res 2005 33 272 279 15647509 10.1093/nar/gki167
Matys V Fricke E Geffers R Gossling E Haubrock M Hehl R Hornischer K Karas D Kel AE Kel-Margoulis OV Kloos DU Land S Lewicki-Potapov B Michael H Munch R Reuter I Rotert S Saxel H Scheer M Thiele S Wingender E TRANSFAC: transcriptional regulation, from patterns to profiles Nucleic Acids Res 2003 31 374 378 12520026 10.1093/nar/gkg108
Dougherty ER Barrera J Brun M Kim S Cesar RM Chen Y Bittner M Trent JM Inference from clustering with application to gene-expression microarrays J Comput Biol 2002 9 105 126 11911797 10.1089/10665270252833217
Davies DL Bouldin DW A cluster separation measure IEEE Transactions on Pattern Analysis and Machine Intelligence 1979 1 224 227
|
16232321
|
PMC1274301
|
CC BY
|
2021-01-04 16:27:27
|
no
|
BMC Bioinformatics. 2005 Oct 18; 6:259
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-259
|
oa_comm
|
==== Front
BMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 1471-2105-6-2611624204410.1186/1471-2105-6-261Methodology ArticleA method of precise mRNA/DNA homology-based gene structure prediction Churbanov Alexander [email protected] Mark [email protected] Daniel [email protected] Hesham [email protected] Department of Computer Science, College of Information Science and Technology, University of Nebraska at Omaha, Omaha, NE 68182-0116, USA2005 21 10 2005 6 261 261 8 3 2005 21 10 2005 Copyright © 2005 Churbanov et al; licensee BioMed Central Ltd.2005Churbanov et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Accurate and automatic gene finding and structural prediction is a common problem in bioinformatics, and applications need to be capable of handling non-canonical splice sites, micro-exons and partial gene structure predictions that span across several genomic clones.
Results
We present a mRNA/DNA homology based gene structure prediction tool, GIGOgene. We use a new affine gap penalty splice-enhanced global alignment algorithm running in linear memory for a high quality annotation of splice sites. Our tool includes a novel algorithm to assemble partial gene structure predictions using interval graphs. GIGOgene exhibited a sensitivity of 99.08% and a specificity of 99.98% on the Genie learning set, and demonstrated a higher quality of gene structural prediction when compared to Sim4, est2genome, Spidey, Galahad and BLAT, including when genes contained micro-exons and non-canonical splice sites. GIGOgene showed an acceptable loss of prediction quality when confronted with a noisy Genie learning set simulating ESTs.
Conclusion
GIGOgene shows a higher quality of gene structure prediction for mRNA/DNA spliced alignment when compared to other available tools.
==== Body
Background
A vast amount of genomic data, including most of the human genome [1], is now available in publicly accessible databases, and the deposition of additional data continues at a rapid pace. Genomic data requires meticulous interpretation and annotation for meaningful information to be extracted. Genes, the most important functional blocks in the human genome, require exact structural annotation for future biological experiments such as reverse genetics and microarray experiments.
Most of the human genes have piecewise structure with a number of exons separated by introns. Introns are excised from original gene transcripts (pre-mRNA) to form mature mRNA. By aligning mRNA with originating genomic clones, we can reconstruct gene structure.
Several fast and efficient tools, such as BLAST [2] and BLASTX [3], were introduced in the early 90's to search databases for homologous blocks, an essential component of all gene structural prediction algorithms. An original method of gene structure prediction based on a set of protein-DNA blocks [4], implemented in GeneBuilder, was followed by Procrustes implementation [5]. Later, there were numerous implementations exploiting the idea of homology-based gene structure prediction, including GeneSeqer with SplicePredictor [6], AAT [7], EbEST [8], ESTMAP [9], TAP [10], Sim4 [11], Spidey [12], GrailEXP Galahad [13], BLAT [14] and est2genome [15]. Other genome annotation software is described in [16,17].
Homology-based methods of gene structure prediction, referred to as spliced alignment, are often classified according to the homology type they employ (DNA/DNA, DNA/mRNA, DNA/Protein, etc.) [16]; frequently, programs employ more than one homology type. The purpose of a spliced alignment algorithm is to explore all possible assemblies of potential exons (blocks) to find a chain of exons which best fits an mRNA target sequence.
In this paper we discuss GIGOgene, a gene structure prediction tool. GIGOgene, like existing spliced alignment software [11,16], can deal with repeating domains, paralogs and pseudogenes. In addition, GIGOgene is capable of combining structural prediction of a gene from partial gene models that span across several genomic clones. The key to GIGOgene higher precision, in the case of mRNA/DNA spliced alignment, is in the use of new splice-enhanced affine gap penalty global alignment for noise-tolerant recovery of exon-intron boundaries, including non-canonical splice sites, with simultaneous prediction of short exons. GIGOgene uses a filtering step to remove suboptimal blocks for better prediction quality.
Implementation
Before we proceed with formal description of methods, we need to define a High-scoring Segment Pair (HSP), otherwise known as a block. In the context of this paper, an HSP is a statistically significant alignment between segments (subsequences) in DNA and mRNA obtained from a BLASTN result. Parameters characterizing HSPs include location in the mRNA query and in the DNA target sequence, and different quality values such as expectation value (E), percent identity, and score.
Below, we provide a brief description of the steps in our gene structural prediction process. Some of these steps are self-explanatory, while others are considered in detail in the following subsections:
Step 1 Align curated mRNA sequence(s) with DNA target sequence database using BLASTN [2].
Step 2 Parse the BLASTN output and select genomic clones that score above 200 bits with an expectation value of no more than 1e1. Through experimentation we have determined that these values are optimal for recovery of most of the essential HSP sets needed for further analysis. These values can be easily adjusted.
Step 3 By pairwise comparative analysis of an HSP set for each selected genomic clone, exclude HSPs with mRNA segments totally within other larger mRNA HSP segments. The longest HSP is assumed to contain the true exonic boundaries; shorter subHSPs usually result from paralogous and pseudogenic matches.
Step 4 Disambiguate the HSP sequences for all the selected clones, as discussed [see Subsection Algorithm for an unambiguous HSP sequences allocation]. The result of this step is a set of unambiguous HSP sequences.
Step 5 Build an interval graph of overlapping unambiguous HSP sequences. The interval graph captures intersection relations of nodes (unambiguous HSP sequences) as we put edges between nodes when nodes belong to different genomic clones, while their mRNA composite segments intersect. Edges between HSP sequences from the same genomic clone are not allowed.
Step 6 Occasionally, short exons missed by the BLASTN algorithm or dust low-complexity filtering result in interrupted unambiguous HSP sequences. Their fragments reside in different interval graph nodes marked with the same genomic clone and transcript. We merge these nodes to form longer, original, uninterrupted unambiguous HSP sequences. An intuitive interpretation of this step is in Figure 1.
Figure 1 Schematic example of HSP sequence restoration.
Step 7 Compact the interval graph, as discussed [see Subsection Joining unambiguous HSP sequences]. This results in the biggest composite genomic clone containing the maximum number of possible exons.
Step 8 Use splice-enhanced affine gap penalty global alignment to identify possible intron/exon boundaries in the composite genomic clone, as discussed [see Subsection Splice-enhanced affine gap penalty global alignment].
Step 9 Extract intron and exon segments from the composite genomic clone and print a report.
Algorithm for an unambiguous HSP sequences allocation
It is well-known that genes, or parts of genes, are duplicated during the course of evolution. This can result in ambiguities during the assembly of a complete gene structure from HSP sequences, as illustrated in Figure 2.
Figure 2 BLASTN alignment structure interpretation. Matches to pseudogene(s) and paralog(s) may be misinterpreted as resulting from original gene; repeating domains in mRNA further confuse gene structure prediction by causing cross-matches.
The problem arises when a segment in an mRNA transcript matches multiple segments in a genomic clone. To address this our algorithm for finding an unambiguous HSP sequence (a chain of putative exons) adheres to the following biological principles:
1. Transcripts are always linear. Thus, we require the set of HSPs to be sequential (we refer to this as the sequential rule).
2. Splicing of pre-mRNA does not introduce any alternations in the order of exons.
3. Alternative splicing does not affect the order of exons in a gene.
4. The similarity of homologous fragments of a gene gradually decreases due to sporadic mutations. As a result, HSPs from the real gene usually have higher scores than HSPs from corresponding pseudogene(s) or paralog(s), as schematically shown in Figure 2. We thus reject unambiguous HSP sequences with average percent identities below a certain threshold; a threshold of 97% produced good results in our experiments with mRNAs. Threshold value could be easily adjusted, if needed, to find gene structure with distant homologs, such as Expressed Sequence Tags (ESTs).
5. Pre-mRNA splicing results in mature mRNA, with exons arranged side by side. In the case of an HSP sequence containing potential exons, we require the entire mRNA transcript to be covered with segments continuously, without breaks.
Disambiguation of an HSP set is shown in Figure 3.
Figure 3 Idea behind the disambiguating algorithm. We distinguish HSP sequences matching real gene, pseudogene(s) and paralog(s), eliminating HSPs from repeating domains.
For the purposes of the disambiguating algorithm we build a bipartite graph structure, where segments are nodes and HSPs are edges connecting the nodes. A dynamic programming disambiguation procedure with an affine gap penalty (Figure 4) is then used to disambiguate the HSPs into a linear sequence. Modifying our early system prototype [18], we changed the criteria for solution optimality (we originally estimated solution quality based on average HSP sequence identity).
Figure 4 HSP sequence disambiguating algorithm.
In our experiments, we noticed cases in which HSPs from the real gene match have a smaller identity compared to HSPs originating from paralogs and pseudogenes. Thus, the disambiguation procedure finds the unambiguous HSP sequence covering the longest mRNA segment with the minimum number of HSPs. For the HSP sequences of equal length with the same number of HSPs, we compare the maximum total weight where the weight of an HSP is a tradeoff between its identity and size:
weight = size·m100-x (1)
Here x is the BLASTN-assigned percent identity for an HSP, size is the HSP length, and m <1 is the decay rate to ensure substantial weight loss for identity lower than the threshold value. The value m = 0.85 produced good results in our experiments. Weight function (1) characterizes the importance of any given HSP in a global solution.
The disambiguation procedure can be represented as a series of transitions between states (Figure 5). State Iy is visited when a sequence of genomic segments is interrupted. This subtracts D, and E for every additional genomic segment missed, from the total weight. If a continuously overlapping transcript-side sequence of segments is broken, the total weight is nullified by visiting state Ix. Weight is gained at state M with a normal transcript-side overlapping sequence of HSPs.
Figure 5 State Diagram of the disambiguating algorithm. Here Wi,j is the weight (1) of an HSP containing transcript segment i and genomic clone segment j. States correspond to weight matrices of partial solutions in dynamic programming.
For this algorithm we must allocate a score matrix F of dimensionality 2 × (# of RNA segments) × (# of DNA segments) and a matrix C of the same size to record the intermediate HSP sequences in the dynamic programming procedure. For the convenience of indexing we introduce aliases M ← F0 and Iy ← F1. We ignore matrix Ix as being unnecessary. The boolean function CONNECTED(i, j) determines the overlap between segments i and j in the transcript.
To generate the final set of unambiguous HSP sequences for a given BLASTN result, the HSP sequences are restored from matrix C using the recursive algorithm shown in Figure 6. At the end of the disambiguating procedure we disregard HSP sequences of average identity lower than threshold value. As explained below, the resulting set of unambiguous HSP sequences can then be optimally connected using an interval graph.
Figure 6 Solution recovery.
Joining unambiguous HSP sequences
To construct a complete HSP sequence out of several smaller overlapping ones, we build an interval graph. Each node in the graph is an unambiguous HSP sequence originating from the disambiguating algorithm discussed [see Subsection Algorithm for an unambiguous HSP sequences allocation]. In order for the nodes to be connected by an edge, they must contain overlapping HSP sequences coming from different genomic clones. To join the nodes, a Floyd-Warshall all-pairs-longest-path algorithm [19] known to run in O(n3) time is used (Figure 7). Joining nodes provides both a larger HSP sequence and the ability to join two genomic clones at a common point.
Figure 7 Modified Pairwise Floyd-Warshall.
If an attempt is made to connect nodes with overlapping HSP sequences from the same genomic clone, the program backs up and searches for other possible optimal unambiguous connections for different clones (see the algorithm in Figure 8). This backing-up modification adds at most O(n2) for each step in the pairwise algorithm for a dense graph, resulting in an O(n5) procedure.
Figure 8 Finding the best combination of HSP sequences to connect.
Although the produced graphs may have different degrees of density, in our experiments they were not sparse enough to use Johnson's modification [19], which runs in O(V2 log V + VE).
To connect the nodes, we solve the following maximization problem:
Dk,i,j={Combination HSP sequences i and j, if k=0,argmaxDk−1,i,j,Dk−1,i,k∪Dk−1,k,j(max(SIZE(Dk−1,i,j),SIZE(Dk−1,i,k∪Dk−1,k,j))),if k>0.
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@C70A@
Figure 7 shows the dynamic programming procedure, after we initialize matrix D. Upon completion of the procedure, we extract the maximum element from matrix Dn and recover the solution. The COMBINATORIALCONNECT function used to find the best combination of HSP sequences is shown in Figure 8.
Splice-enhanced affine gap penalty global alignment
In order to identify precise intron/exon boundaries in a genomic clone, a modified Needleman-Wunch global alignment algorithm with affine gap penalty is used to create a spliced alignment between segments of query and target sequences.
The basic Needleman-Wunch algorithm provides a scattered (i.e. frequently interrupted) mRNA/DNA alignment pattern, with no clear indication of exon/intron boundaries. With affine gap penalties, we penalize the score each time we break an alignment [20]; this provides an alignment clustered within putative exons, but usually without precise indications of exon/intron boundaries. The addition of sensor information (GT/AG, AT/AC or similar rules [21]) results in precise gene structural prediction.
Our implementation is a modification of the affine gap penalty algorithm [20] and can be explained in terms of transitions between states in a Hidden Markov Model (HMM) [20]. Specifically, there are thirteen matrices of size n × m introduced, corresponding to states as shown in Figure 9. The matrices are reduced to arrays of size 2 × 13 × m, since we need only two rows in the scoring matrix F and backtracking matrices [22,23].
Figure 9 State Diagram of the modified spliced alignment algorithm. See the text for an explanation of the various states.
In our algorithm, we introduce the match state M (Figure 9), which uses the BLASTN scoring matrix. The state Iy corresponds to a gap penalty in the mRNA transcript, while the other states correspond to forming gaps in the genomic clone and have nucleotide-specific score deductions. Gap-opening matrices dA and dG express score preferences to open a gap with either nucleotide A or G, respectively; d is a generic gap-opening penalty; and e is a generic gap-extending penalty. Typically, the cost of extending a gap is set to be five to ten times lower than the cost for opening a gap. Gap-extension penalty matrices eA, eC, eG and eT express scoring preference to extend gaps with nucleotides A, C, G and T, respectively.
In order to save running time, we use anchors – short nucleotide segments from mRNA and DNA expected to contain exon/intron boundary fragments with donor/acceptor signals. A normal anchor does not have mismatches in state M. If a mismatch is encountered, it may mean a short exon is present. If necessary, the anchor can be expanded and the spliced alignment rerun with two full exons and intron between to identify possible short exons or address sequencing errors, as discussed [see Subsection Advantages of using splice-enhanced affine gap penalty global alignment in gene structure prediction].
According to our model, introducing or extending a gap is straightforward using the GT/AG rule. The penalty becomes severe if we try inserting a gap without the rule; we would rather use higher-extension-penalty state Ix for short gaps frequently resulting from sequencing errors. The AT/AC rule works in much the same way, except with a higher score penalty.
We implement the affine gap penalty spliced alignment algorithm in a linear memory of size S(m + n) and running time O(n × m), where n is the size of a DNA fragment and m is the size of an mRNA fragment.
These are the steps in implementing the algorithm:
1. Run the spliced alignment ALIGN(0...n, 0...m) to find indexes of u and v (the split points for a recursive call). Here u is the vertical median index, and v is the horizontal index of a point where the optimal traceback intersects the median.
2. Restore the matrix context for the recursive calls and prior-state information for proper backtracking.
3. Make the recursive calls for nucleotide segments ALIGN(0...u, 0...v) and ALIGN(u...n, v...m), etc.
4. If either of the nucleotide segments' lengths in a recursive call is less than or equal to 1, call the ordinary spliced alignment for these pieces to get the alignment states.
More detailed explanation of the spliced alignment algorithm we use is in [18].
Advantages of using splice-enhanced affine gap penalty global alignment in gene structure prediction
There are several advantages of using splice-enhanced affine gap penalty global alignment, discussed [see Subsection Splice-enhanced affine gap penalty global alignment], for gene structure prediction:
• ability to recover canonical and non-canonical splice sites;
• noise-tolerant prediction of splice sites;
• ability to recover short exons;
• ability to handle low complexity regions in genomic DNA, if sorted out by dust filtering.
A splice site usually happens on the boundaries of HSPs, but in most cases mRNA segments of neighboring HSPs overlap with no clear indication of a splice site. Recovery of a splice site could be formulated as a combinatorial problem of finding optimal exon/intron boundaries in HSPs' overlap vicinity. A dynamic programming approach, such as the splice-enhanced affine gap penalty global alignment we use, allows us to consider all possible rearrangements around HSPs' overlap to pick optimal splice sites in polynomial time.
The process of splice site recovery is schematically shown in Figure 10. Segments of mRNA and DNA sequences used for splice site prediction are called anchors. To accelerate the gene structure prediction process we use small (30 nt) anchors by default.
Figure 10 Recovery of optimal splice site.
The ideal variant of splice site recovery is shown in Figure 10. In a small number of cases we have misalignment, as shown schematically in Figure 11. Misalignment may occur if:
Figure 11 Recovery of suboptimal splice site.
• neighboring HSPs overlap too much, so that we can't reliably identify a splice site with small anchors;
• there is a sequencing error adjacent to a splice site;
• short exons are present.
All of these cases require additional application of splice-enhanced affine gap penalty global alignment with anchors expanded to include the entire HSP segments for maximum error tolerance, as shown in Figure 11.
As an example of a successful anchor expansion, consider an HSP sequence interrupted by undetected short exon(s) (Figure 12). After combining interrupted unambiguous HSP sequences, as described [see Section Implementation], the small exons are recovered by processing the expanded anchors with our spliced alignment procedure.
Figure 12 Recovery of small exons.
Similarly, low-complexity regions may interrupt HSP sequencing in BLASTN results due to dust filtering. In this case, interrupted unambiguous HSP sequences are combined and the gap will be closed by sequence matching, resulting in a monolithic exon (Figure 13).
Figure 13 Handling of BLASTN HSP interrupted by low-complexity filtering.
Results
Experiments with Genie learning set
GIGOgene was tested, along with Spidey, est2genome, Sim4, Galahad and BLAT on 462 mRNA transcripts of the human Genie multi-exon annotated learning set . We used transcripts corresponding to mRNA or CDS features in the Genie learning set annotation.
Sensitivity (ESn) and specificity (ESp) were calculated according to the formulas
ESn=TEAE (3)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGfbqrcqWGtbWucqWGUbGBcqGH9aqpdaWcaaqaaiabdsfaujabdweafbqaaiabdgeabjabdweafbaacaWLjaGaaCzcamaabmaabaGaeG4mamdacaGLOaGaayzkaaaaaa@398C@
ESp=TEPE (4)
MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGfbqrcqWGtbWucqWGWbaCcqGH9aqpdaWcaaqaaiabdsfaujabdweafbqaaiabdcfaqjabdweafbaacaWLjaGaaCzcamaabmaabaGaeGinaqdacaGLOaGaayzkaaaaaa@39B0@
Here TE is the number of accurately predicted exon boundaries, AE is the number of annotated exon boundaries in the Genie learning set, and PE is the number of predicted exon boundaries. Only internal exonic boundaries were considered. Results are summarized in Table 1.
Table 1 Comparative exon-level sensitivity and specificity for different programs on human Genie learning set
TE AE PE ESn ESp
Galahad 4744 4909 4790 96.64% 99.04%
Spidey 4827 4909 4847 98.33% 99.59%
EST2genome 4742 4909 4752 96.60% 99.79%
Sim4 4837 4909 4845 98.53% 99.83%
BLAT 4832 4909 4902 98.43% 98.57%
GIGOgene 4864 4909 4865 99.08% 99.98%
This test is designed as evidence of general prediction quality of different gene structure annotation tools. GIGOgene has the highest sensitivity and specificity in this case, which highlights advantages of the approach we use.
Experiments with micro-exon detection
We followed the Sim4 prediction compensating procedure described in [24] to identify human genes containing canonical micro-exons (3–12nt in our case). This way we were able to annotate 44 genes in the human DNA phase 3 database. Table 2 compares performance of different programs for a micro-exonic set of genes.
Table 2 Micro-exon gene set comparative level sensitivity and specificity for different programs
TE AE PE ESn ESp
Galahad 1220 1422 1278 85.79% 95.46%
Spidey 1251 1422 1334 87.97% 93.78%
EST2genome 1270 1422 1318 89.31% 96.36%
Sim4 1278 1422 1326 89.87% 96.38%
BLAT 1375 1422 1424 96.69% 96.56%
GIGOgene 1420 1422 1422 99.86% 99.86%
This study shows that the GIGOgene program has the highest structural prediction sensitivity and specificity in this case. BLAT recovered 96.69% true exonic boundaries in the micro-exonic set, while other programs had fraction of true splice sites recovered no more than 90%, i.e. they most likely miss micro-exon(s) from their prediction.
Experiments with non-canonical splice sites
According to [25] approximately 98.71% of all splice sites are reported to be canonical, 0.56% are in the biggest group of GC-AG non-canonical splices sites, and the remaining 0.76% consist of small groups of size no more than 0.05% each. Following the description in [25] we parsed the Human SpliceDB database of EST supported, corrected and GenBank High Throughput Genome sequencing projects (HTG) supported pairs of non-canonical splice sites. Then we aligned the pairs to the human RefSeq database using BLASTN to extract transcripts containing verified non-canonical splice sites. Found transcripts were BLAST-aligned to the NCBI human phase 3 DNA database to match corresponding gene-containing clones. We splice-aligned found transcripts and corresponding genomic clones using GIGOgene. A manual check on 108 gene structural predictions identified no problems on the GIGOgene side. A comparable performance study for other programs is shown in Table 3.
Table 3 Non-canonical gene set comparative level sensitivity and specificity
TE AE PE ESn ESp
Galahad 2764 2896 2818 95.44% 98.08%
Spidey 2857 2896 2893 98.65% 98.76%
EST2genome 2788 2896 2888 96.27% 96.54%
Sim4 2868 2896 2893 99.03% 99.14%
BLAT 2880 2896 2987 99.45% 96.42%
GIGOgene 2896 2896 2896 100.00% 100.00%
In this study est2genome made a mistake in annotating virtually every non-canonical splice site while reinforcing canonical splice rule. Although BLAT was very sensitive in this experiment, it makes mistakes occasionally.
Simulated EST experiment
In order to research the EST-related performance of different programs we introduced 4% noise in the Genie experiment discussed [see Subsection Experiments with Genie learning set]. Noise was equiprobably distributed between random nucleotide insertions, deletions and substitutions. Results of a simulated EST experiment are presented in Table 4.
Table 4 Noisy Genie experiment
TE AE PE ESn ESp
Galahad 4531 4909 4655 92.30% 97.34%
Spidey 3547 4909 4759 72.26% 74.53%
EST2genome 4704 4909 4737 95.82% 99.30%
Sim4 4775 4909 4833 97.27% 98.80%
BLAT 3898 4909 17338 79.41% 22.48%
GIGOgene 4446 4909 4767 90.57% 93.27%
With simulated EST study our program performed worse than Sim4 and est2genome, about as well as Galahad, and substantially better than BLAT and Spidey, the programs that were specifically designed for mRNA/DNA spliced alignment. The reason for substantial quality loss with GIGOgene is in splice site annotation strategy. If we get a number of nucleotide inserts between exon boundaries in mRNA, they can be easily interpreted as micro-exon(s) with non-canonical splice sites, rather than reinforcing the GT-AG rule in a genomic clone as Sim4 and EST2genome do. That is why these two applications have rather poor performance in micro-exonic testing [see Subsection Experiments with micro-exon detection], where they sacrifice micro-exons to reinforce canonical splice rule.
Run-time comparison
In Table 5 we compare running time for different programs required to annotate the set of micro-exon containing genes mentioned [see Subsection Experiments with micro-exon detection].
Table 5 Comparative time in seconds required by Pentium IV computer to annotate a set of genes containing micro-exons. BLASTN running time is included in GIGOgene timing.
Sim4 Spidey BLAT Galahad GIGOgene EST2genome
3.705 sec. 11.419 sec. 16.029 sec. 170.333 sec. 1504.444 sec. 5323.904 sec.
Run time comparison on the set of micro-exons indicates that our program runs faster than est2genome but slower than other tools we have looked at. By using splice-enhanced affine gap penalty global alignment we traded execution time for quality, compare to simpler heuristics used to predict splice sites in other tools.
Chromosome 22 experiment
For this experiment we chose human chromosome 22 whole draft sequence NC_000022.8 from NCBI Genbank. A total of 506 transcripts were mapped to the chromosome by parsing human RefSeq flatfiles, but only 430 transcripts have corresponding genes annotated in NCBI Genbank.
We report running time for all 506 transcripts mapped to chromosome 22. For the GIGOgene program it took 12 hours 16 minutes 42 seconds to parse BLASTN results, while BLASTN took 9 days 18 hours 9 minutes 3 seconds to align transcripts to the chromosome (without dust filtering). Such a long running time could be explained by extensive low-complexity domains duplicated across the chromosome. BLASTN with low-complexity filtering took only 13 hours, 33 minutes and 18 seconds, but the following GIGOgene gene structural prediction was inferior to the results reported in Table 6. BLAT annotation took 12 hours 8 minutes 39 seconds (without dust filtering).
Table 6 Chromosome 22 prediction quality for 430 mapped transcripts with structural annotation
TE AE PE ESn ESp
BLAT 7025 7088 8003 99.11% 87.78%
GIGOgene 7036 7088 7071 99.27% 99.51%
We report exon-level comparative performance of BLAT and GIGOgene in Table 6.
Results of BLAT and GIGOgene comparison on Chromosome 22 whole draft sequence annotation agree well with the previously observed tendency: with GIGOgene, gene structural prediction takes longer, compared to BLAT, and has higher prediction quality.
Conclusion
Using a homology-based approach, we have designed a program for eukaryotic gene structural annotation. In case of mRNA/DNA spliced alignment we have been able to improve on exon-level sensitivity and specificity by addressing several possibilities of error. Program domain is limited to mRNA/DNA spliced alignment with a reasonable fraction of sequencing errors. Experiments on running time position our tool as a relatively slow utility for annotating specific cases of gene structural prediction.
Several published spliced alignment algorithms were mentioned [see Section Background]. Our splice-enhanced affine gap penalty global alignment in some ways similar to the spliced alignment of protein/DNA blocks described in the Procrustes paper [5]. The key differences in our implementation is that it works in linear memory and is effective in annotation of both canonical and non-canonical splice sites. Compared to protein-DNA alignment, it has finer granularity, which translates to smaller possibility for incorrect structural prediction, especially for micro-exons. We can also annotate both CDS and UTR regions, while protein-DNA homology programs, such as Procrustes [5] and Genomescan [26], are limited to CDS region only.
The stand-alone program version, web implementation interface, test results and manual for GIGOgene are available at .
Availability and requirements
Project name: Good In Good Out gene structural prediction tool (GIGOgene)
Project home page: .
Operating system: Platform independent
Programming language: Java
Other requirements: Java 1.4.1 or higher
License: GNU Lesser General Public Licence
Authors' contributions
AC and DQ conceptualized the project and set up computational facility. DQ implemented BLASTN SAX parser and made many valuable suggestions through the progress of our study. AC implemented and evaluated GIGOgene Java code. MP extensively edited the manuscript and made many important changes. HA helped to conceptualize the tool, provided general support and gave final approval of the version to be published. All authors read and approved the final manuscript.
Acknowledgements
We would like to thank members of the Bioinformatics Group at the University of Nebraska at Omaha who provided useful feedback on our progress and program. This work was supported by the NIH grant number P20 RR16469 from the INBRE program of National Center for Research Resource.
==== Refs
IHGSC Initial sequencing and analysis of the human genome Nature 2001 409 860 921 11237011
Altschul WG Miller W Myers E Lipman D Basic Local Alignment Search Tool Journal of Molecular Biology 1990 215 403 410 2231712
Gish W States DJ Identification of protein coding regions by database similarity search Nature Genetics 1993 3 266 272 8485583
Rogozin IB Milanesi L Kolchanov NA Gene structure prediction using information on homologous protein sequence Comput Appl Biosci 1996 12 161 170 8872383
Gelfand MS Mironov AA Pevzner PA Gene recognition via spliced sequence alignment Proc Natl Acad Sci USA 1996 93 9061 9066 8799154
Usuka J Zhu W Brendel V Optimal spliced alignment of homologous cDNA to a genomic DNA template Bioinformatics 2000 16 203 211 10869013
Huang X Adams MD Zhou H Kerlavage AR A tool for analyzing and annotating genomic sequences Genomics 1997 46 37 45 9403056
Jiang J Jacob HJ EbEST: an automated tool using expressed sequence tags to delineate gene structure Genome Res 1998 8 268 275 9521930
Milanesi L Rogozin IB ESTMAP: a system for expressed sequence tags mapping on genomic sequences IEEE Trans Nanobioscience 2003 2 75 78 15382662
Kan Z Rouchka EC Gish WR States DJ Gene structure prediction and alternative splicing analysis using genomically aligned ESTs Genome Res 2001 11 889 900 11337482
Florea L Hartzell G Zhang Z Rubin GM Miller W A computer program for aligning a cDNA sequence with a genomic DNA sequence Genome Research 1998 8 967 974 9750195
Wheelan SJ Church DM Ostell JM Spidey: A tool for mRNA-to-Genomic Alignment Genome Research 2001 11 1952 1957 11691860
Xu Y Uberbacher EC Automated Gene Identification in Large-Scale Genomic Sequences Journal of Computational Biology 1997 4 325 338 9278063
Kent WJ BLAT-the BLAST-like alignment tool Genome Research 2002 12 656 664 11932250
Mott R EST_GENOME: a program to align spliced DNA sequences to unspliced genomic DNA CABIOS 1997 13 477 478 9283765
Mathé C Sagot MF Schiex T Rouzé P Current methods of gene prediction, their strengths and weaknesses Nucleic Acids research 2002 30 4103 4117 12364589
Miller W Comparison of genomic DNA sequences: solved and unsolved problems Bioinformatics 2001 17 391 397 11331233
Tchourbanov A Quest D Ali H Pauley M Norgren R A New Approach for Gene Annotation Using Unambiguous Sequence Joining Proceedings of the Computational Systems Bioinformatics (CSB'03) 2003 IEEE Computer society 353 362
Cormen TH Leiserson CE Rivest RL Stein C Introduction to Algorithms 2001 2 MIT Press
Durbin R Eddy S Krogh A Mitchison G Biological sequence analysis 1998 Cambridge University Press
Burge CB Padgett RA Sharp PA Evolutionary fates and origins of U12-type introns Molecular Cell 1998 2 773 785 9885565
Hirschberg DS A linear-space algorithm for computing maximal common subsequences Communications of the ACM 1975 18 341 343
Myers EW Miller W Optimal alignments in linear space Computer Applications in the Bio-sciences 1988 4 11 17
Volfovsky N Haas BJ Salzberg SL Computational Discovery of Internal Micro-Exons Genome Research 2003 13 1216 1221 12799353
Burset M Seledtsov IA Solovyev VV Analysis of canonical and non-canonical splice sites in mammalian genomes Nucleic Acids Research 2000 28 4364 4375 11058137
Yeh RF Lim LP Burge CB Computational inference of homologous gene structures in the human genome Genome Research 2001 11 803 816 11337476
|
16242044
|
PMC1274302
|
CC BY
|
2021-01-04 16:27:46
|
no
|
BMC Bioinformatics. 2005 Oct 21; 6:261
|
utf-8
|
BMC Bioinformatics
| 2,005 |
10.1186/1471-2105-6-261
|
oa_comm
|
==== Front
BMC CancerBMC Cancer1471-2407BioMed Central London 1471-2407-5-1321622569710.1186/1471-2407-5-132Research ArticleA prospective evaluation of treatment with Selective Internal Radiation Therapy (SIR-spheres) in patients with unresectable liver metastases from colorectal cancer previously treated with 5-FU based chemotherapy Lim L [email protected] P [email protected] D [email protected] JD [email protected] R [email protected] D [email protected] A [email protected] W [email protected] M [email protected] The Royal Melbourne Hospital, Parkville, Victoria, Australia2 Cabrini Hospital Malvern, Victoria, Australia3 The Canberra Hospital, Canberra, ACT, Australia2005 15 10 2005 5 132 132 4 2 2005 15 10 2005 Copyright © 2005 Lim et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
To prospectively evaluate the efficacy and safety of selective internal radiation (SIR) spheres in patients with inoperable liver metastases from colorectal cancer who have failed 5FU based chemotherapy.
Methods
Patients were prospectively enrolled at three Australian centres. All patients had previously received 5-FU based chemotherapy for metastatic colorectal cancer. Patients were ECOG 0–2 and had liver dominant or liver only disease. Concurrent 5-FU was given at investigator discretion.
Results
Thirty patients were treated between January 2002 and March 2004. As of July 2004 the median follow-up is 18.3 months. Median patient age was 61.7 years (range 36 – 77). Twenty-nine patients are evaluable for toxicity and response. There were 10 partial responses (33%), with the median duration of response being 8.3 months (range 2–18) and median time to progression of 5.3 mths. Response rates were lower (21%) and progression free survival shorter (3.9 mths) in patients that had received all standard chemotherapy options (n = 14). No responses were seen in patients with a poor performance status (n = 3) or extrahepatic disease (n = 6). Overall treatment related toxicity was acceptable, however significant late toxicity included 4 cases of gastric ulceration.
Conclusion
In patients with metastatic colorectal cancer that have previously received treatment with 5-FU based chemotherapy, treatment with SIR-spheres has demonstrated encouraging activity. Further studies are required to better define the subsets of patients most likely to respond.
==== Body
Background
Colorectal cancer (CRC) is the most common GI malignancy accounting for 4718 deaths in Australia [1] and almost 437,000 deaths worldwide annually making it the most third most common malignancy in the developed world [2]. Around fifty to sixty percent of these patients will develop liver metastases, and in approximately 20% of cases the liver is the only site of disease at death[3]. Surgical resection of all apparent disease is possible in selected patients, however for the majority of patients with metastatic CRC the standard approach remains systemic chemotherapy.
Selective Internal Radiation (SIR) spheres (Sirtex Medical, Sydney, Australia) are a new radiotherapeutic treatment for liver metastases. These resin microspheres contain yttrium, a high energy beta-emitting isotope, and are embolised into the hepatic artery where they become lodged within the microvasculature of the tumour. The treatment is relatively selective as hepatic tumours derive their blood supply almost exclusively from the hepatic artery whereas normal liver parenchyma is supplied predominantly by the portal circulation. Animal studies suggest that SIR spheres allow on average 200–300 Gy to be delivered to liver tumours [4]. In contrast the delivery of standard external beam radiation therapy to the whole liver is limited by the ability of the normal parenchyma to tolerate only 30–35 Gy, an insufficient dose to produce a significant anti-tumour effect [5].
Encouraging results have been reported following previous studies of SIR spheres in metastatic colorectal cancer. In a series of 21 chemonaive patients with colorectal liver metastases who were randomised to receive intravenous 5FU alone or intravenous 5FU plus SIR spheres, the combination demonstrated a higher response rate and significantly improved progression free survival compared to chemotherapy alone [6]. In a larger study of 74 patients combining SIR-spheres with hepatic artery chemotherapy superior response rates and time to progression over treatment with chemotherapy alone were seen in patients with colorectal cancer [7]. With the exception of hepatocellular carcinoma[8] results in other tumour types have not been so encouraging [9].
We report here the first prospective series conducted to better define the efficacy and safety of SIR-spheres(Yttrium90) in patients with colorectal cancer and liver metastases that have previously received 5-FU based chemotherapy. No financial support was received from SIRTex for the purposes of this study.
Methods
Data for consecutive patients with metastatic colorectal cancer treated with Sirtex microspheres were collected prospectively across 3 Australian centres from Jan 2002 and March 2004. During this period of the time both oxaliplatin and irinotecan were not reimbursable in Australia as part of first-line therapy for patients with metastatic colorectal cancer outside of a clinical trial. These agents were available for patients that had progressed following initial 5-FU based treatment.
Patients were informed of the available evidence regarding SIR-spheres treatment. Patients that elected to proceed with treatment were informed that data would be collected prospectively as part of a research project. Toxicities and protocols were outlined in accordance with the manufacturer's guidelines and were common across all participating centres. Patients were considered eligible if they had liver metastases from colorectal cancer with histological confirmation of their primary tumour. All patients were required to have measurable disease within the liver. Extra-hepatic disease was allowed if the liver was the dominant site of disease.
Patients with an expected survival of less than 3 months, documented brain metastases, or a poor performance status (ECOG >2) were excluded. Adequate hepatic, renal and liver function was required including a normal clotting profile, an albumin >30 g/L, bilirubin <20 umol/L and no evidence of liver decompensation such as ascites or portal hypertension was permitted. Patients with portal vein thrombosis were also excluded from this study.
Previous treatment with chemotherapy was allowed provided this had been more than 2 months prior to planned treatment with SIR-spheres. Patients received bolus 5FU chemotherapy concurrent with the SIR-spheres, as a radiosensitiser, and subsequently in responding patients at the investigator's discretion.
Pre treatment workup and disease evaluation
All patients underwent a standard pre-treatment evaluation as per manufacturers guidelines. An hepatic angiogram was performed to define arterial anatomy prior to treatment and to permit a nuclear medicine scan with radio labeled Tc 99 m-MAA (macro-aggregated albumin) in order to exclude patients at high risk of lung (radiation pneumonitis) or GI toxicity (gastric/duodenal ulceration) due to hepato-systemic shunting or aberrant vasculature. The fraction of extra-hepatic shunting was determined for each patient as a percentage. Patients who had shunting of greater than 20% were excluded from the study. Shunting of between 12 and 20% resulted in a reduction in the dosage of spheres delivered. The final dose of SIR-spheres administered in units of GBq was calculated according to each patient's body surface area and percentage of tumour involvement of liver (an assessment made by the radiologist after viewing the baseline CT scan). Across the three participating centres, the workup and delivery of SIR spheres was performed by a total of four experienced interventional radiologists with two centres having all treatment administered by a single interventional radiologist.
Following a single treatment of SIR-spheres, two weeks after the MAA scan, patients were routinely followed and assessed at monthly intervals. Acute toxicity was assessed initially and at subsequent clinical visits. All toxicities were graded according to National Cancer Institute (NCI) Common Toxicity Criteria, and performance status according to Eastern Cooperative Oncology Group (ECOG) criteria. Disease evaluation was performed by CT imaging at 2 months and bi-monthly thereafter until disease progression. A complete response (CR) was defined as the disappearance of all target lesions. A partial response (PR) was defined as >30 percent decrease in the sum of the longest dimension of the target lesions at 2 months. All responses were confirmed on repeat imaging. Progressive disease (PD) was defined as a >20 percent increase in the sum of the longest recorded target lesion(s) within the liver or the development of new or progressive extrahepatic disease. Stable disease (SD) was defined as a decrease not sufficient to qualify for a PR nor an increase not sufficient to qualify for PD. The results of all staging investigations and toxicity assessments were collected and analysed prospectively. The duration of response was determined from the date of the first response evaluation at 2 months until progressive disease.
Results
Patient characteristics
A total of 30 patients underwent treatment between January 2002 and March 2004.(see table 1). There were 22 males and 8 females with a median age of 61.7 years (range 36–77). Ninety percent of patients had an ECOG performance status of 0 or 1. Twenty percent of patients (6) had low-volume extra-hepatic disease, the remainder of the patients had liver-only disease. All patients had failed 5FU chemotherapy and 22 (73%) patients had failed either oxaliplatin or irinotecan containing regimens, with 14 (46%) having progressed through both. 21 of the 30 patients received 5-FU concurrent with the SIR-spheres.
Of the 30 patients treated, 29 are evaluable for disease response. One patient presented shortly after treatment with worsening liver function tests, liver failure and death considered by the investigator to likely be from rapid disease progression. This was not confirmed with imaging. Overall, there were 10 partial responses (33% of all 30 patients treated). Of note ongoing responses were seen in a number of patients, with continued reduction in the size of liver lesions occurring out to 12 months from treatment. One patient with an initial partial response had achieved a complete response at six months. The median duration of response is 8.3 months (range between 2–18+ months) at a median follow up time (as of July 2004) of 18.3 months for all patients. Eight patients in total had stable disease at 2 months (27%) and the remaining 12 (40%) had disease progression (n = 11) or were not evaluable (n = 1). The median time to progression for all patients was 5.3 mths, however, patients who achieved a partial response within the liver as a group had a median progression free survival of 9.2 mths. See table 2
All responses occurred in patients with disease confined to the liver (n = 24) and no responses were seen in patients with a performance status of 2 (n = 3). In the 8 patients that had received only prior 5-FU there were 6 responses (75%). In the patients that had received oxaliplatin and/or irinotecan (n = 22) there were 5 responses (23%). This included responses in 3 of 14 (21%) patients that had previously received all standard chemotherapy options. No other factors were apparently predictive of response, including the bulk of disease, in this study.
Toxicity
Overall toxicity assessments were carried out according to NCI common toxicity criteria at all three centres. The findings were consistent across all study sites with between 2–8 weeks of lethargy, anorexia, nausea and RUQ pain being observed to a variable extent in most patients. However, in most cases, these side effects were mild and self-limiting following treatment with standard anti-emetics and analgesics medication. Three patients reported severe nausea and lethargy for 2 weeks following treatment and moderate symptoms were reported in several patients up to 1 month following treatment.
Serious treatment related toxicities clearly related to treatment with SIR-spheres were recorded in 4 patients (13%) who had gastric/duodenal ulceration confirmed on gastroscopy. In 3 of these patients, SIR-spheres were seen in biopsies taken from ulcerated mucosa (Figure 3) and in two patients there was rapid improvement following treatment with proton pump inhibitors. However, one patient had severe ongoing and disabling pain, anorexia and nausea that continued until the time of death 3 months later and another patient experienced ongoing symptoms beyond 3 months, despite the use of proton pump inhibitors.
Other serious toxicity potentially related to SIR-spheres included a single patient who was admitted to hospital with acute right upper quadrant pain and marked deterioration in her liver function tests one month following treatment with SIR-spheres. The clinical assessment was that the patient likely had radiation hepatitis. Her symptoms settled with conservative management.
Discussion
This prospective evaluation documents the experience in our three institutions of using Selective Internal Radiation spheres to treat liver metastases from colorectal cancer in patients that had previously received 5-FU chemotherapy. Our series adds to the published experience documenting the activity of this treatment in selected patients, along with the potential for significant toxicity. For the period of this study oxaliplatin and irinotecan were available in Australia only for patients that had progressed following initial 5-FU based treatment and this is reflected in the study design.
Overall in our experience, treatment with SIR-spheres demonstrated promising activity in pre-treated patients with liver metastases from colorectal cancer. Partial responses were seen in six of the eight patients that had failed 5FU alone. This includes one patient that was able to undergo potentially curative resection of residual liver disease after further response to systemic chemotherapy and she remains disease free 22 months later. The response rate in our series is similar to that reported in the previous small randomized study where this combination was used first-line [7]. The progression free and overall survival data for our study are shown in Figure 3. Due to the small patients numbers this activity may partly reflect chance or patient selection. However, this is encouraging efficacy as the alternative treatment for these patients would have been irinotecan alone or oxaliplatin plus 5-FU where responses are typically seen in less than 20% of patients [10,11].
Significant response rates were also seen in patients that had progressed through several lines of chemotherapy. Notably, fourteen of these patients had previously received both irinotecan and oxaliplatin, and the response rate was maintained in this group. The only treatment other option open to these patients would be cetuximab, a monoclonal antibody directed at the epidermal growth factor receptor (EGFR). As a single agent responses are seen in about 10% of patients [12] and response rates are higher when concurrent irinotecan is administered [12]. However, not all patients are suitable for cetuximab as approximately 25% of patients do not express the EGFR[13]. Based on our results SIR-spheres appear to be a good option for patients with colorectal cancer who have liver only disease and maintain a good performance status following progression on all standard chemotherapy drugs. The suggestion from our results is that benefit from the addition of SIR-spheres will be greater if used earlier, but this will need to be confirmed in larger studies.
The toxicity data from our experience was consistent with findings from earlier trials involving SIR-spheres. The 13% severe gastric/duodenal ulceration rate is significant and is consistent with a recent larger study reporting a 12% GI ulceration rate [14]. Gastrointestinal ulceration occurred despite strictly adhered to protocols of pre-treatment workup and treatment only by experienced interventional radiologists. The product information recommends routine use of a H-2 antagonist prophylactically the day before the procedure and for a month afterwards in view of the known association of peptic ulceration with SIR-spheres treatment [15]. Although this was not part of our treatment protocol, the prophylactic use of proton-pump inhibitors or H-2 antagonists should be considered in patients treated with SIR-spheres. In our analysis there were no apparent indicators of which patients were likely to experience toxicity. In particular this did not appear related to disease bulk or patient performance status.
Conclusion
In summary, this series has demonstrated that treatment with SIR-spheres produces encouraging responses in pretreated patients with colorectal cancer. Further studies in this group of patients should be pursued. On the basis of our results, thought should be given to stratifying patients according to the presence or absence of extrahepatic disease and according to performance status. We have also demonstrated the potential for significant toxicity, and such treatment should only be conducted in centres with experienced interventional radiologists.
The results of ongoing phase I/II trials that are exploring the combination of SIR-spheres with multi-agent chemotherapy (including irinotecan and oxaliplatin containing regimens) are eagerly awaited. Ultimately however, there remains a need for large phase III clinical trials evaluating the efficacy of standard therapy plus SIR-spheres compared to standard therapy alone. Such trials will define the true value of SIR-spheres and will permit insight into optimal patient selection. Finally the financial costs of this new treatment are not insignificant and future studies will also need to clearly demonstrate cost effectiveness as well as efficacy in an environment where the costs of treating patients with colorectal cancer are rapidly escalating [16].
Pre-publication history
The pre-publication history for this paper can be accessed here:
Figures and Tables
Figure 1 Survival curves for the 30 patients treated.
Figure 2 (A) Progression free survival for the two groups of patients, those that had previously received 5-FU or xeloda, and those that had received at least one oxaliplatin or irinotecan containing regimen. (B) Overall survival for the two groups of patients, those that had previously received 5-FU or xeloda, and those that had received at least one oxaliplatin or irinotecan containing regimen.
Figure 3 Biopsy of gastric mucosa showing inflammation (gastritis) from several SIR spheres, clearly visible.
Table 1 Patient Characteristics
Total Number 30
Male 22 (73%)
Female 8 (27%)
ECOG 0 16 (53%)
ECOG 1 11 (37%)
ECOG 2 3 (10%)
Extra-hepatic disease 7 (20%)
Failed 5FU alone 8 (27%)
Failed 5FU + CPT-11 and Oxaliplatin 14 (46%)
Failed 5FU + either CPT-11 or Oxaliplatin 8 (27%)
Table 2 Results At 2 months by patient group (*1 patient died prior to the first evaluation and has been classified as progressive disease).
Partial Response (%) Stable Disease (%) Progressive Disease (%)
Overall (n = 30) 10 (33%) 8 (27%) 12 (40%)*
Failed 5FU alone (n = 8) 6 1 1*
Failed 5FU and CPT-11 and Oxal (n = 14) containing regimens 3 2 8
Failed 5FU and either CPT-11 or Oxaliplatin (n = 8) containing regimens 2 2 4
Extra-Hepatic disease (n = 7) 0 2 5
ECOG 2 (n = 3) 0 1 2
==== Refs
Australian Institute of Health and Welfare (AIHW), Australasian Association of Cancer Registries (AACR) Cancer in Australia 2000 2003 Canberra: AIHW
Pisani P Parkin DM Bray F Ferlay J Estimates of the worldwide mortality from 25 cancers in 1990 Int J Cancer 1999 83 18 29 10449602 10.1002/(SICI)1097-0215(19990924)83:1<18::AID-IJC5>3.0.CO;2-M
Weiss L Grundmann E Torhorst J Harveit F Moberg I Eder M Fenoglio-Preiser C Napier J Haematogenous Metastasis patterns in colonic carcinoma: An analysis of 1541 necropsies J Pathol 1986 150 195 203 3806280 10.1002/path.1711500308
Ingold JA Reed GB Kaplan HS Bagshaw MA Radiation hepatitis AM J Roentgenol Radium Ther Nucl Med 1965 03 200 208
Lawrence TS Robertson JM Anscher MS Jirtle RL Ensminger WD Fajardo LF Hepatic toxicity resulting from cancer treatment Int J Radiat Oncol Biol Phys 1995 31 1237 48 7713785 10.1016/0360-3016(94)00418-K
Gray BN Anderson JE Burton MA van Hazel G Codde J Morgan C Regression of liver metastases following treatment with yttrium-90 microspheres Aust N Z J Surg 1992 62 105 10 1586298
Gray B Van Hazel G Hope M Burton M Moroz P Anderson J Gebski V Randomised trial of SIR-Spheres plus chemotherapy vs. chemotherapy alone for treating patients with liver metastases from primary large bowel cancer Ann Oncol 2001 12 1711 20 11843249 10.1023/A:1013569329846
Lau WY Ho S Leung TW Chan M Ho R Johnson PJ Li AK Selective internal radiation therapy for nonresectable hepatocellular carcinoma with intraarterial infusion of 90yttrium microspheres Int J Radiat Oncol Biol Phys 1998 40 583 92 9486608 10.1016/S0360-3016(97)00818-3
Lim L Gibbs P Yip D Shapiro JD Dowling R Smith D Little A Bailey W Liechtenstein M A prospective study of treatment with Selective Internal Radiation Therapy (SIR spheres) in patients with unresectable primary or secondary hepatic malignancies Internal Medicine Journal (in press-to be published in April 2005 edition of Internal Medicine Journal)
Rougier P Van Custem E Bajetta E Randomised trial of irinotecan versus fluorouracil by continuous infusion after fluorouracil failure in patients with metastatic colorectal cancer Lancet 1998 352 1407 1412 9807986 10.1016/S0140-6736(98)03085-2
Rothenberg ML Oza AM Bigelow RH Superiority of oxaliplatin and fluorouracil-leucovorin compared with either therapy alone in patients with progressive colorectal cancer after irinotecan and fluorouracil-leucovorin: Interim results of a phase III trial J Clin Oncol 2003 21 2059 2069 12775730 10.1200/JCO.2003.11.126
Cunningham D Humblet Y Siena S Khayat D Bleiberg H Santoro A Bets D Mueser M Harstrick A Verslype C Chau I Van Cutsem E Cetuximab monotherapy and cetuximab plus irinotecan in irinotecan-refractory metastatic colorectal cancer N Engl J Med 2004 351 337 45 15269313 10.1056/NEJMoa033025
Saltz L Meropol NJ Loehrer PJ Needle M Kopit J Mayer RJ Phase II Trial of Cetuximab in Patients With Refractory Colorectal Cancer That Expresses the Epidermal Growth Factor Receptor J Clin Oncol 2004 22 1201 1208 14993230 10.1200/JCO.2004.10.182
Stubbs RS Cannan RJ Mitchell AW Selective internal radiation therapy with 90yttrium microspheres for extensive colorectal liver metastases J Gastrointest Surg 2001 5 294 302 11360053 10.1016/S1091-255X(01)80051-2
Sirtex Medical Limited SIR-Spheres Product Monograph, Sydney 2002
Scrag D The price tag on progress chemotherapy for colorectal cancer N Engl J Med 2004 351 317 9 15269308 10.1056/NEJMp048143
|
16225697
|
PMC1274303
|
CC BY
|
2021-01-04 16:03:04
|
no
|
BMC Cancer. 2005 Oct 15; 5:132
|
utf-8
|
BMC Cancer
| 2,005 |
10.1186/1471-2407-5-132
|
oa_comm
|
==== Front
BMC CancerBMC Cancer1471-2407BioMed Central London 1471-2407-5-1331622570310.1186/1471-2407-5-133Research ArticleMicroheterogeneity of transthyretin in serum and ascitic fluid of ovarian cancer patients Gericke Beate [email protected] Jens [email protected] Jalid [email protected] Sophie [email protected]önsgen Dominique [email protected] Alexander [email protected] Florian J [email protected] Department of Physiology and Pathophysiology, Institute of Nutritional Science, University of Potsdam, Arthur-Scheunert-Allee 114-116; D-14558 Nuthetal, Germany2 Department of Obstetrics and Gynecology, Campus Virchow-Klinikum, Charité Universitätsmedizin Berlin, Augustenburger Platz 1, 13353 Berlin, Germany3 Interdisciplinary Center for Mass Spectrometry of Biopolymers, University of Potsdam, Karl-Liebknecht-Str. 24-25, D-14476 Golm, Germany2005 17 10 2005 5 133 133 6 4 2005 17 10 2005 Copyright © 2005 Gericke et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Transthyretin (TTR), a traditional biomarker for nutritional and inflammatory status exists in different molecular variants of yet unknown importance. A truncated form of TTR has recently been described to be part of a set of biomarkers for the diagnosis of ovarian cancer. The main aim of the study was therefore to characterize differences in microheterogeneity between ascitic fluid and plasma of women affected with ovarian cancer and to evaluate the tumor site as the possible source of TTR.
Methods
Subjects were 48 women with primary invasive epithelial ovarian cancer or recurrent ovarian carcinoma. The control group consisted of 20 postmenopausal women. TTR and retinol-binding protein (RBP) levels were measured by enzyme-linked immunoassay (ELISA) and C-reactive protein (CRP) levels by a high-sensitivity latex particle turbidimetric assay. The molecular heterogeneity of TTR was analysed using immunoprecipitation and matrix-associated laser desorption ionization time-of-flight mass spectrometry (MALDI-TOF-MS). Presence of TTR in tumor tissue was determined with indirect peroxidase immunostaining.
Results
TTR and RBP (μg/ml) levels in serum were 148.5 ± 96.7 and 22.5 ± 14.8 in affected women compared to 363.3 ± 105.5 and 55.8 ± 9.3 in healthy postmenopausal women (p < 0.01). In ascitic fluid, levels were 1.02 ± 0.24 and 4.63 ± 1.57 μg/ml, respectively. The mean levels of TTR and RBP in serum showed a tendency to decrease with the severity of the disease and were lower in affected women whose CRP levels were > 40 mg/ml (p = 0.08 for TTR; p < 0.05 for RBP). No differences in TTR microheterogeneity were observed between TTR isolated from serum of affected and healthy women or from ascitic fluid. TTR occurred rather consistently in four variants. Mass signals were at 13758 ± 7, 13876 ± 13 (greatest intensity), 13924 ± 21 and 14062 ± 24 Da, representing native, S-cysteinylated, S-cysteinglycinylated and glutathionylated TTR, respectively. Serum of healthy and affected women as well as ascitic fluid contained the truncated fragment of TTR (12828 ± 11 Da). No immunoreactive TTR was observed in the tumor sites.
Conclusion
The severity of the cancer associated catabolism as well as the inflammation status affect serum TTR and RBP levels. Neither TTR nor its truncated form originates from tumor tissue and its occurrence in ascites may well reflect the filtration from blood into ascitic fluid.
==== Body
Background
TTR, formerly called prealbumin, belongs to a group of proteins including thyroxine-binding globulin and albumin which bind to and transport thyroid hormones in the blood. TTR is also involved in the metabolism of vitamin A as it binds RBP, the specific plasma transport protein for retinol. First identified in 1942 by Kabat et al. [1] in serum and cerebrospinal fluid, TTR has been described as a so-called visceral protein that is synthesized in the liver in response to nutritional supply. TTR plasma levels can thus be used as a sensitive biochemical parameter of subclinical malnutrition, because both the synthesis of proteins as well as energy intake are reflected in its serum levels. Serum levels of TTR, however, are also affected by acute and chronic diseases associated with an acute-phase response. Under these conditions, liver activity is converted to the synthesis of acute-phase response proteins, resulting in a drop in visceral proteins despite adequate nutritional supply [2-5].
Epithelial ovarian cancer is the leading cause of death from gynaecologic malignancies in western countries [6]. Most patients are first diagnosed at an advanced stage with diffuse peritoneal metastasis outside the pelvis (FIGO stage III or IV). Tumor stage is one of the most important prognostic factors in ovarian cancer [7]. The 5-year survival rate for FIGO stage III ranges from 25 to 45%, whilst for patients diagnosed at FIGO stage I or II survival rates are between 85 and 95% [8]. Therefore, various strategies have been investigated to increase the detection rate of early ovarian cancer. A systematic review by Fung and co-workers [9] on the screening of postmenopausal women for ovarian cancer showed that for every 10,000 women participating in an annual screening program with cancer associated antigen 125 (CA125) over 3 years, 800 have had an ultrasound scan due to an elevated CA125, 30 underwent surgery because of an abnormal ultrasound, whilst only 6 women were diagnosed with ovarian cancer of whom only three where detected at an early stage. Therefore, despite the combination of CA125 monitoring and ultrasound this strategy remains insufficient as a screening tool.
Unfortunately, most other biomarkers also have low sensitivity and specificity and little predictive value [10]. Application of new technologies for detection of ovarian cancer could have an important effect on public health [11], but to achieve this goal, specific and sensitive molecular markers are essential [12,13]. This need is especially urgent in women, who have a high risk of ovarian cancer due to family or personal history of cancer, and for women with a genetic predisposition to cancer due to abnormalities in predisposition genes [14]. Besides protein profiling, the determination of changes in the microheterogeneity of a variety of proteins has been suggested as an approach to biomarker discovery [15].
In plasma, the non-glycosylated TTR is present as a tetramer of non-covalently bound monomers of ~14 kDa. Physiologically its importance is related to the transport of thyroid hormones and retinol. The association of TTR and RBP is a prerequisite for the homeostatic control of plasma and retinol levels. In addition to mutations on protein level, TTR exists in different isoforms [16,17]. The isoforms results when the cyteine residue on position 10 (Cys10) makes a mixed disulfide with the amino acid cysteine the peptide cysteinyl-glycine, or the peptide glutathione. The possible importance of this as a risk factor for the onset of senile systemic amyloidosis remains to be elucidated [18,19]. Additionally, Cys10 adducts of S-homocysteine of TTR have been detected in plasma of humans with hyper-homocysteinemia [20]. Recently, a new truncated variant of TTR together with apolipoprotein A1 and a cleaved fragment of inter-α-trypsin inhibitor heavy chain H4 were described as an efficient set of new biomarkers for ovarian cancer in women [21].
In human primary hepatic cancer, the mRNA of TTR, which is normally highly expressed in the liver, is strikingly suppressed [22]. Not only in hepatic cancer, but also for broncho-pulmonary carcinoid cancers, TTR was concluded to be a useful marker [23]. Nothing however is known with regard to ovarian cancer as a source of TTR itself or cleavage products thereof. Since we have recently shown that TTR is present in ascitic fluid [24], it cannot be excluded that its presence might not only arise from an increased permeability for blood constituents into ascitic fluid due to an intensified vascularization [25], but may also reflect the secretion of products synthesized by the malignant ovary cells due to the intimate contact between tumor site and ascitic fluid.
The study was thus conducted to primarily characterize possible differences in microheterogeneity of TTR arising from posttranslational modifications and/or products of protein degradation or proteolysis between serum and ascitic fluid of women with ovarian cancer and also to evaluate the affected ovary as a possible site of TTR expression.
Methods
Subjects
The study was conducted on 48 patients (average age 53 ± 11.9; mean ± SD) with ovarian cancer admitted to the Department of Gynecology and Obstetrics, Charité, Campus Virchow-Klinikum, Berlin, Germany. All patients underwent primary surgery with median laparotomy, hysterectomy, adenectomy, omentectomy and pelvic and paraortal lymphadenectomy to achieve maximal tumor reduction. According to the classifications of the International Federation of Gynecology and Obstetrics (FIGO), the treated group consisted of four patients with stage Ic, two with stage IIc, 28 with stage III and 14 with stage IV. Of the 48 women, 25 were suffering from primary ovarian carcinoma and 23 had recurrent ovarian carcinoma. The controls were 20 healthy postmenopausal women (average age 58 ± 1.0; mean ± SD). The postmenopausal status was proved by the assessment of LH and FSH. The study protocol was approved by the hospitals and University of Potsdam Ethics Committee. All samples and relevant clinical data were obtained from the Tumor Bank Ovarian Cancer (TOC). Informed consent was obtained from each participant.
Ascitic fluid was collected under sterile conditions from the patients with ovarian carcinoma and centrifuged at 1500 × g for 20 min at 4°C. The supernatants were stored at -80°C. Serum was separated from freshly drawn blood at the same time of paracentesis and stored at -80°C until assayed.
Tissue samples were fixed in 4% PBS-buffered paraformaldehyde for 24 h and embedded in paraplast. Each of the samples was stained routinely with haematoxylin and eosin (H+E) and diagnosed. All tumor samples were reviewed by a pathologist.
Determination of TTR, RBP and CRP levels
Levels of TTR and RBP in serum and ascitic fluid were determined by ELISA using polyclonal rabbit anti-human antibodies (DakoCytomation, Hamburg, Germany) [17]. CRP levels in serum were measured with a high sensitivity latex turbimetric immunoassay using a latex-coupled monoclonal mouse anti-human antibody (Olympus AU 600, Biomed, Germany). The sensitivity of this assay was 0.005 mg/dl. The 90th percentile of normal CRP distribution was 0.3 mg/dl.
Immunoprecipitation of TTR and subsequent analysis by MALDI-TOF-MS
TTR from serum and ascitic fluid of 20 randomized representative women was prepared by immunoprecipitation. The subgroup consisted of two patients with FIGO stage Ic or IIc, 13 with stage III and 5 with stage IV. Briefly, 15 μl of serum or ascitic fluid was treated with an equal amount of a polyclonal rabbit anti human antibody (DakoCytomation). The mixture was incubated for two hours at 37°C and then centrifuged at 15.000 × g for 15 min at room temperature. The supernatant was removed and the immunoprecipitated complex of TTR and antibody was then washed with high performance liquid chromatography grade water.
To determine the disulfide linkage of TTR adducts, the immunoprecipitated TTR was treated with dithiothreitol (DTT). DTT solution, 100 mM in buffer (100 mM NH4CO3, pH 8.8) was added to the solution at a ratio of 1:1 (DTT solution volume/TTR solution volume). The mixture was incubated for 2 h at room temperature and precipitated samples were subsequently subjected to MALDI-TOF-MS.
MALDI mass spectra of the precipitated TTR from serum and ascitic fluid were obtained using a Reflex II MALDI-TOF mass spectrometer (Bruker-Daltonik, Bremen, Germany). MALDI-TOF MS of serum samples was performed in linear mode at 20 k acceleration voltage using sinapic acid as matrix. For ionization, a nitrogen laser (337 nm, 3 ns pulse width, 3 Hz) was used. The samples were prepared in a two step procedure: First, 0.5 μl serum were deposited on the target. Secondly, 0.5 μl saturated sinapinic acid solution was placed on serum drop and dried. This step was repeated. The matrix solution contained 1 mg sinapinic acid and equal amounts (25 μl) 1% trifluoroacetic acid and acetonitrile. For optimisation of the mass spectra, the laser was aimed either at the central area of the sample or at the outmost edge of the crystal rim. All spectra were measured using external calibration.
Immunohistochemistry of TTR
For indirect peroxidase immunostaining of TTR, slides were deparaffinized, rehydrated in a decreased series of alcohol to water and exposed for 60 min in 0.5% hydrogen peroxide in methanol in order to deactivate endogenous peroxidases. Non-specific antibody binding was blocked for 30 min in Tris-buffered saline (TBS, pH 7.6) containing 5% bovine serum albumin (BSA; Sigma, Taufkirchen, Germany). The primary human anti-TTR antibody (DakoCytomation) was diluted 1:100 in 1% bovine serum albumin (BSA) in TBS. After overnight incubations at 4°C, the sections were treated with peroxidase-coupled swine anti-rabbit IgG (DakoCytomation) diluted 1:100 in 1% BSA in TBS for 30 min. The antigen-antibody binding sites were visualized by incubating the sections in a solution of diaminobenzidine tetrahydrochloride (DAB; Sigma) containing 0.01% hydrogen peroxide in 0.1 M imidazole buffer (pH 7.1). Counterstaining was performed with Papanicolaou hematoxylin. Negative controls, which included the omission of the primary antibodies, revealed no significant labelling. A positive control (liver) was included in each individual staining process. The sections were examined and photographed with an Olympus BX-50 microscope equipped with a ColorView 12 CCD video camera (SIS, Münster, Germany). Images were processed using analySIS™ 3.0 software (SIS).
Statistical procedures
Values are expressed as means and standard deviations (SD). Unpaired t tests were performed to compare serum values with ascitic fluid or to compare between the groups using standard methods software (SPSS package, version 10.0). P < 0.05 was regarded as statistically significant.
Results
TTR and RBP levels in serum and ascitic fluid
Results of serum and ascitic fluid levels of TTR and RBP are shown in Table 1. In women with cancer, serum levels of both TTR and RBP were lower compared to healthy controls (p < 0.01). In the more severe stages of ovarian cancer the levels showed a tendency to be even lower. Within the cancer group, increased levels of CRP (cut-off > 40 mg/ml) in serum were associated with lower levels of TTR (p = 0.08) and RBP (p < 0.05) (Fig. 1). TTR and RBP levels in ascitic fluid were substantially lower compared to serum (p < 0.01). No obvious differences of TTR and RBP concentration in ascitic fluid between FIGO stages were observed.
TTR microheterogeneity in serum and ascitic fluid
Using the combination of immunoprecipitation and subsequent MALDI-TOF-MS we were able to show that no obvious differences exists in the microheterogeneity of TTR between serum of affected and healthy women as well as in ascitic fluid. TTR monomer occurred rather consistently in four major variants in the range where TTR and its conjugated forms should normally appear (m/z 13,700 – 14,100). The results are summarized in Table 2. In the mass spectra of serum and ascitic fluid (Fig. 2 and Tab. 2) peaks dominated at m/z 13,875.8 ± 12.8 and 13,876.9 ± 13.3 respectively. Three additional mass spectra were recorded. The mass differences between these variants were similar in serum and ascitic fluid (Tab. 2). The molecular mass of 13,757.7 ± 7.1 Da corresponded to the native, unmodified TTR. The other peaks in serum representing Cys10 adducts for S-cysteine (TTR- Cys10-S-S-Cys, mass = 13,875.8 ± 12.8 Da), S-cysteinylglycine (TTR- Cys10-S-S-CysGly, mass = 13,923.6 ± 21.0) and S-glutathione (TTR- Cys10-S-S-SG, mass = 14,062.1 ± 24.7). The shift in the mass spectrum of TTR after treatment with DTT, towards the native form of TTR, indicates that the adducts are formed via the disulfide linkage at Cys10 (Fig. 2). Additionally, in serum and ascitic fluid a smaller mass signal could be observed with varying intensity at a molecular mass of 12828 ± 11 Da.
Immunohistochemistry of TTR
In order to assess the expression of TTR within ovarian cancer tissue we performed immunohistochemical staining using a polyclonal TTR antibody in paraffin embedded sections. TTR immunoreactivity was previously tested in human liver sections and revealed cytoplasmic staining within hepatocytes (data not shown). In ovarian cancer tissues diffuse TTR immunostaining was only observed within blood vessels, haemorrhages or plasma insudations (Fig. 3). No TTR labelling however was seen within the epithelial cells of any cancer specimen.
Discussion
Epithelial ovarian cancer is the leading cause of death from gynaecologic malignancies in western countries [26,27]. The tumor stage at time of diagnosis and the postoperative residual tumor mass are important prognostic factors and are unequivocally related to overall survival [26]. Other prognostic factors are identified mostly in small series and are the source of controversial discussion in the relevant literature.
Serum TTR is traditionally a valid marker for nutritional status in general and in cancer patients it has gained considerable interest with regard to the use as an early diagnostic marker in ovarian cancer [21]. As depletion of nutritional reserves and a subsequent significant weight loss can lead to an increased risk of morbidity, reduced chemotherapy response, and shorter survival in patients with cancer, TTR is a valid prognostic marker [28]. Interestingly however, TTR and RBP levels of serum are affected not only by the nutritional status of the individual but are also reduced during the acute phase response associated with inflammation [29]. Additionally to quantitative aspects, the TTR molecule in serum exists in numerous variants due either to genetic differences or because of modification on one readily accessible cystein within the molecule. The microheterogeneity is affected by different metabolic aspects such as oxidative stress or homocystein levels [20,30]. Nothing is known however with regard to possible variation due to metabolic alterations in cancer.
The results of the present study confirm previous results for cancer patients in general and especially for patients with ovarian cancer, regarding the greatly reduced serum levels of TTR and RBP [31]. Interestingly however, the intensity of the disease has no significant influence on serum levels, indicating that it is a general phenomenon possibly associated with cancer induced cachexia which is already present at early stages. For drawing a general conclusion, this group (stage I/II) was too small in sample size. On the other hand, when the differing inflammatory statuses were considered, obvious differences were observed between the cancer patients for serum levels of TTR and RBP. Using 40 mg/l as cut-off for C-reactive protein (CRP) TTR and RBP serum levels were reduced in those individuals with increased CRP values. This clearly supports observations showing that the inflammation status greatly reduces TTR and RBP serum levels as a consequence of a reduced synthesis of this negative acute-phase protein in the liver [5].
Using immunological procedures we were recently able to show the presence of TTR in ascitic fluid from women with ovarian cancer, however no quantitative data, especially with regard to cancer stages, is as yet available [32]. In accordance to our previous semi-quantitative study, TTR in ascitic fluid was more than 100-fold lower when compared to its serum levels. This ratio is much lower in comparison to the one observed for RBP (Tab. 1). Based on the difference between their molecular masses, 55 kDa for the hetero-tetramer TTR and 21 kDa for RBP, one would expect a different ascites/serum ratio, as an inverse correlation exists between ascites/serum ratio and the mean of the molecular weight of various proteins [5]. From this observation one could assume that RBP and TTR are not transferred individually but rather as the complex usually present in serum [33]. In general, results support the hypothesis that the concentration of TTR and RBP in ascitic fluid is the result of a passive transfer from serum into the ascitic fluid. The accumulation of these and other serum constituents is mainly attributed to the increased capillary permeability caused by an increase in permeability-inducing factors such as the vascular endothelial growth factor (VEGF) [25].
These observations and the fact that no obvious differences in microheterogeneity between TTR from serum and ascitic fluid can be observed, both with regard to the known modification at the Cys10 and the recently described truncated form, it can be assumed that all TTR in ascitic fluid originates through a passive transfer from serum. This is further supported by the observation that the tumor site itself does not express any immunoreactive TTR. It can not be excluded however, that the tumor site or components in the ascitic fluid may have proteolytic properties that possibly result in the unobserved modifications of TTR or other proteins.
With regard to the microheterogeneity of TTR in serum and ascitic fluid, the results support and confirm previous studies undertaken by us and others with regard to molecular variants of TTR in serum [17,34-36]. As in these studies, TTR in serum and ascitic fluid was dominant in four variants. The 118 Da larger variant is the S-cysteinylated form of the native TTR whilst the signal at 14,062 Da can be attributed to the S-glutathionylated TTR form [30,34,37]. As TTR contains only one cysteine residue (Cys10), the adduct must result when the Cys10 residue forms a mixed disulfide with the amino acid cysteine, the dipeptide cysteinylglycine or the tripeptid glutathione. The shift in the mass spectrum of TTR variants towards the native TTR molecule mass in serum and ascites fluid after treatment with DTT indicates that the adducts are formed via the disulfide linkage at Cys10. In addition to this, we confirmed in serum of healthy and affected women as well as in ascitic fluid the presence of a smaller immunoreactive form of TTR with a molecular mass of 12,830 Da, which was recently identified as a truncated form of TTR lacking the NH2-terminal 10 amino acids [21]. Its presence in both serum and ascitic fluid supports once again the idea of passive transfer from serum into ascitic fluid during its accumulation.
Conclusion
Results show that, although the microheterogeneity of TTR itself and the occurrence of possible immunoreactive fragments thereof in serum and ascites fluid is unaffected by the cancer. Absolute levels of TTR as well as RBP in serum are negatively affected by the disease and by inflammatory processes associated with the cancer. It cannot be excluded that other metabolic effects yet to be defined might interact with the cancerous process. Thus, to fully validate the specificity of TTR or any of its fragments as a biomarker for ovarian cancer, a careful selection of controls has to be implemented, including consideration of nutritional status and the presence of inflammatory processes especially the possible influence of various hepatic diseases.
List of abbreviations used
BSA (bovine serum albumin)
CA125 (cancer associated antigen 125)
CRP (C-reactive protein)
Cys 10 (solt cysteine residue on position 10 of each TTR subunit)
Da (Dalton)
DTT (dithioretiol)
EAM (energy-absorbing molecule)
ELISA (enzyme-linked immunoassay)
FIGO (International Federation of Gynecology and Obstertrics)
MALDI (matrix assisted laser desorption and ionization – time-of-flight – mass spectrometry)
MW (molecular weight)
RBP (retinol-binding protein)
SD (standard deviation)
TBS (Tris-buffered saline)
TOC (Tumor Bank Ovarian Cancer)
TTR (transthyretin)
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
BG, JR, JS and FJS participated in the conception, design, data analysis and the writing of the manuscript. SH participated in the analysis and interpretation of MS. DK and AM participated in sample validation and data analysis. All authors have read and approved the last version of the manuscript.
Pre-publication history
The pre-publication history for this paper can be accessed here:
Acknowledgements
We thank the "Medizinisches Labor Potsdam", Dr. Martin Kern, for the assistance in the determination of CRP. We also thank E. Meyer for technical assistance in the histological analysis.
Figures and Tables
Figure 1 Correlation between levels of CRP and TTR in serum (p = 0.08) and between levels of CRP and RBP in ascitic fluid (p < 0.05).
Figure 2 Mass spectra resulting from MALDI-TOF-MS after immunoprecipitation of TTR in plasma (A) and ascites (B) obtained from women with ovarian cancer. The untreated TTR (I) and the TTR after treatment with dithioretiol (II) are shown.
Figure 3 Histological sections from ovarian malignoma subjected to staining with H+E (A-C) or immunodetection of TTR (D-E). Diffuse TTR immunostaining was only detectable within blood vessels (arrows) or plasma insudations (asterisks) (D). No immunoreactivity was observed within epithelial cells of any tumour specimen (D-F)). Negative controls, which included the omission of the primary antibody, revealed no significant labelling (G-I).
Table 1 Comparison of levels (mean ± SD) of TTR and RBP in serum and asciticfluid in relation to FIGO-stages and levels of CRP in ovarian cancer patients
serum (μg/mL) ascitic fluid (μg/mL) percentage of serum in ascitic fluid
TTR RBP TTR RBP TTR RBP
Healthy control n = 20 363.3 ± 105.5 55.8 ± 9.3 - - - -
All Stages ovarian cancer 148.5 ± 96.7 22.5 ± 14.8 1.022 ± 0.239 4.632 ± 1.572 1.0 ± 0.6 26.2 ± 14.9
Stage I/II ovarian cancer 162.5 ± 69.4 34.2 ± 22.3 0.966 ± 0.375 3.580 ± 0.849 0.8 ± 0.6 14.5 ± 9.8
Stage III ovarian cancer 155.6 ± 107.6 19.5 ± 7.5 1.074 ± 0.185 4.800 ± 1.711 1.0 ± 0.6 27.2 ± 11.8
Stage IV ovarian cancer 129.0 ± 96.0 17.4 ± 7.3 0.965 ± 0.241 4.627 ± 1.672 1.2 ± 0.7 32.1 ± 23.3
Ovarian cancer CRP > 40 mg/l 116.0 ± 94.3 14.2 ± 4.3 1.047 ± 0.213 4.736 ± 1.886 1.4 ± 0.7 35.4 ± 17.2
Ovarian cancer CRP < 40 mg/l 166.3 ± 92.1 24.1 ± 12.5 1.005 ± 0.260 4.503 ± 1.415 0.8 ± 0.5 22.3 ± 11.1
All values for ascitic fluid (if available) are significantly different (p < 0.01) from serum value.
Table 2 Molecular mass of immunoprecipitated TTR (Da, mean ± SD) assigned to different forms1 of TTR between serum and ascitic fluid of 20 representative women with ovarian cancer. Values in brackets represent mass differences of modified TTR in relationship to the native form of TTR (Da, mean ± SD).
TTR cysTTR cysglycTTR glutTTR
serum 13757.7 ± 7.1 13875.8 ± 12.8 (118.1 ± 13.7) 13923.6 ± 21.0 (166.8 ± 20.9) 14062.1 ± 24.7 (306.4 ± 23.9)
ascitic fluid 13752.7 ± 16.5 13876.9 ± 13.3 (124.2 ± 13.7) 13926.8 ± 11.1 (176.8 ± 1.8) 14042.5 ± 27.0 (294.6 ± 22.7)
1 TTR = native TTR; cysTTR = TTR + cysteinylation; cysglycTTR = TTR + cysteinylglycine; glutTTR = TTR + glutathione; no significant differences between molecular masses of serum and ascitic fluid could be observed
==== Refs
Kabat EA Moore D Landow H An electrophoretic study of the protein components in cerebrospinal fluid and their relationship to serum proteins J Clin Invest 1942 21 571 577
Ingenbleek Y Young V Transthyretin (prealbumin) in health and disease: nutritional implications Annu Rev Nutr 1994 14 495 533 7946531 10.1146/annurev.nu.14.070194.002431
Lasztity N Biro L Nemeth E Pap A Antal M Protein status in pancreatitis – transthyretin is a sensitive biomarker of malnutrition in acute and chronic pancreatitis Clin Chem Lab Med 2002 40 1320 1324 12553437 10.1515/CCLM.2002.227
Power DM Elias NP Richardson SJ Mendes J Soares CM Santos CR Evolution of the thyroid hormone-binding protein, transthyretin Gen Comp Endocrinol 2000 119 241 255 11017772 10.1006/gcen.2000.7520
Abraham K Muller C Gruters A Wahn U Schweigert FJ Minimal inflammation, acute phase response and avoidance of misclassification of vitamin A and iron status in infants – importance of a high-sensitivity C-reactive protein (CRP) assay Int J Vitam Nutr Res 2003 73 423 430 14743546
Landis SH Murray T Bolden S Wingo PA Cancer statistics, 1999 CA Cancer J Clin 1999 49 8 31 10200775
Sehouli J Drescher FS Mustea A Elling D Friedmann W Kuhn W Nehmzow M Opri F Klare P Dietel M Lichtenegger E Granulosa cell tumor of the ovary: 10 years follow-up data of 65 patients Anticancer Res 2004 24 1223 1229 15154651
Pecorelli S Odicino F Favalli G Ovarial cancer: best timing and applications of debulking surgery Ann Oncol 2000 11 141 144 11079131 10.1023/A:1011128032107
Fung MF Bryson P Johnston M Chambers A Screening postmenopausal women for ovarian cancer: a systematic review J Obstet Gynaecol Can 2004 26 717 728 15307976
Hu W Wu W Kobayashi R Kavanagh JJ Proteomics in cancer screening and management in gynecologic cancer Curr Oncol Rep 2004 6 456 462 15485615
Ozols RF Outcome issues in ovarian cancer Oncology (Huntingt) 1995 9 135 139
Jacobs IJ Skates SJ MacDonald N Menon U Rosenthal AN Davies AP Woolas R Jeyarajah AR Sibley K Lowe DG Oram DH Screening for ovarian cancer: a pilot randomised controlled trial Lancet 1999 353 1207 1210 10217079 10.1016/S0140-6736(98)10261-1
Menon U Jacobs IJ Recent developments in ovarian cancer screening Curr Opin Obstet Gynecol 2000 12 39 42 10752515 10.1097/00001703-200002000-00007
Petricoin EF Ardekani AM Hitt BA Levine PJ Fusaro VA Steinberg SM Mills GB Simone C Fishman DA Kohn EC Liotta LA Use of proteomic patterns in serum to identify ovarian cancer Lancet 2002 359 572 577 11867112 10.1016/S0140-6736(02)07746-2
Schweigert FJ Characterization of protein microheterogeneity using mass spectrometry based immunoassays Brief Funct Genomic Proteomic 2005 15975260
Bernstein LH Ingenbleek Y Transthyretin: its response to malnutrition and stress injury. clinical usefulness and economic implications Clin Chem Lab Med 2002 40 1344 1348 12553442 10.1515/CCLM.2002.232
Schweigert FJ Wirth K Raila J Characterization of the microheterogeneity of transthyretin in plasma and urine using SELDI-TOF-MS immunoassay Proteome Sci 2004 2 5 15341658 10.1186/1477-5956-2-5
Quan D Cohen JA Clinical variant of familial amyloid polyneuropathy Muscle Nerve 2002 26 417 420 12210373 10.1002/mus.10208
Takaoka Y Ohta M Miyakawa K Nakamura O Suzuki M Takahashi K Yamamura K Sakaki Y Cysteine 10 is a key residue in amyloidogenesis of human transthyretin Val30Met Am J Pathol 2004 164 337 345 14695346
Sass JO Nakanishi T Sato T Sperl W Shimizu A S-homocysteinylation of transthyretin is detected in plasma and serum of humans with different types of hyperhomocysteinemia Biochem Biophys Res Commun 2003 310 242 246 14511677 10.1016/j.bbrc.2003.08.089
Zhang Z Bast RC JrYu Y Li J Sokoll LJ Rai AJ Rosenzweig JM Cameron B Wang YY Meng XY Berchuck A Van Haaften-Day C Hacker NF de Bruijn HW van der Zee AG Jacobs IJ Fung ET Chan DW Three biomarkers identified from serum proteomic analysis for the detection of early stage ovarian cancer Cancer Res 2004 64 5882 5890 15313933 10.1158/0008-5472.CAN-04-0746
Gu JR Jiang HQ He LP Li DZ Zhou XM Dai WL Qian LF Chen YQ Schweinfest C Papas T Transthyretin (prealbumin) gene in human primary hepatic cancer Sci China B 1991 34 1312 1318 1666289
Suresh UR Wilkes S Hasleton PS Prealbumin in the diagnosis of bronchopulmonary carcinoid tumours J Clin Pathol 1991 44 573 575 1713221
Schweigert FJ Steinhagen B Raila J Siemann A Peet D Büscher U Concentrations of carotenoids, retinol and alpha-tocopherol in plasma and follicular fluid of women undergoing IVF Hum Reprod 2003 18 1259 1264 12773456 10.1093/humrep/deg249
Aslam N Marino CR Malignant ascites: new concepts in pathophysiology, diagnosis, and management Arch Intern Med 2001 161 2733 2737 11732940 10.1001/archinte.161.22.2733
Sehouli J Mustea A Könsgen D Lichtenegger W Conventional and experimental prognostic factors in ovarian cancer Zentralbl Gynakol 2004 126 315 322 15478050 10.1055/s-2004-832305
Cannistra SA Cancer of the ovary N Engl J Med 2004 351 2519 2529 15590954 10.1056/NEJMra041842
Slaviero KA Read JA Clarke SJ Rivory LP Baseline nutritional assessment in advanced cancer patients receiving palliative chemotherapy Nutr Cancer 2003 46 148 157 14690790 10.1207/S15327914NC4602_07
Ingenbleek Y Young VR Significance of transthyretin in protein metabolism Clin Chem Lab Med 2002 40 1281 1291 12553432 10.1515/CCLM.2002.222
Ando Y Suhr O Yamashita T Ohlsson PI Holmgren G Obayashi K Terazaki H Mambule C Uchino M Ando M Detection of different forms of variant transthyretin (Met30) in cerebrospinal fluid Neurosci Lett 1997 238 123 126 9464635 10.1016/S0304-3940(97)00868-9
Mahlck CG Grankvist K Plasma prealbumin in women with epithelial ovarian carcinoma Gynecol Obstet Invest 1994 37 135 140 8150370
Schweigert FJ Raila J Sehouli J Büscher U Accumulation of selected carotenoids, alpha-tocopherol and retinol in human ovarian carcinoma ascitic fluid Ann Nutr Metab 2004 48 241 245 15331882 10.1159/000080457
Blomhoff R Transport and metabolism of vitamin A Nutr Rev 1994 52 S13 23 8202278
Terazaki H Ando Y Suhr O Ohlsson PI Obayashi K Yamashita T Yoshimatsu S Suga M Uchino M Ando M Post-translational modification of transthyretin in plasma Biochem Biophys Res Commun 1998 249 26 30 9705825 10.1006/bbrc.1998.9097
Kiernan UA Tubbs KA Gruber K Nedelkov D Niederkofler EE Williams P Nelson RW High-throughput protein characterization using mass spectrometric immunoassay Anal Biochem 2002 301 49 56 11811966 10.1006/abio.2001.5478
Kiernan UA Tubbs KA Nedelkov D Niederkofler EE McConnell E Nelson RW Comparative urine protein phenotyping using mass spectrometric immunoassay J Proteome Res 2003 2 191 197 12716133 10.1021/pr025574c
Lim A Sengupta S McComb ME Theberge R Wilson WG Costello CE Jacobsen DW In vitro and in vivo interactions of homocysteine with human plasma transthyretin J Biol Chem 2003 278 49707 49713 14507924 10.1074/jbc.M306748200
|
16225703
|
PMC1274304
|
CC BY
|
2021-01-04 16:03:05
|
no
|
BMC Cancer. 2005 Oct 17; 5:133
|
utf-8
|
BMC Cancer
| 2,005 |
10.1186/1471-2407-5-133
|
oa_comm
|
==== Front
BMC Cell BiolBMC Cell Biology1471-2121BioMed Central London 1471-2121-6-361622566910.1186/1471-2121-6-36Research ArticleRegulation of actin cytoskeleton architecture by Eps8 and Abi1 Roffers-Agarwal Julaine [email protected] Jennifer B [email protected] Jeffrey R [email protected] Department of Genetics, Cell Biology, and Development, University of Minnesota, 6-160 Jackson Hall, 321 Church St SE, Minneapolis, MN 55455, USA2005 14 10 2005 6 36 36 24 5 2005 14 10 2005 Copyright © 2005 Roffers-Agarwal et al; licensee BioMed Central Ltd.2005Roffers-Agarwal et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
The actin cytoskeleton participates in many fundamental processes including the regulation of cell shape, motility, and adhesion. The remodeling of the actin cytoskeleton is dependent on actin binding proteins, which organize actin filaments into specific structures that allow them to perform various specialized functions. The Eps8 family of proteins is implicated in the regulation of actin cytoskeleton remodeling during cell migration, yet the precise mechanism by which Eps8 regulates actin organization and remodeling remains elusive.
Results
Here, we show that Eps8 promotes the assembly of actin rich filopodia-like structures and actin cables in cultured mammalian cells and Xenopus embryos, respectively. The morphology of actin structures induced by Eps8 was modulated by interactions with Abi1, which stimulated formation of actin cables in cultured cells and star-like structures in Xenopus. The actin stars observed in Xenopus animal cap cells assembled at the apical surface of epithelial cells in a Rac-independent manner and their formation was accompanied by recruitment of N-WASP, suggesting that the Eps8/Abi1 complex is capable of regulating the localization and/or activity of actin nucleators. We also found that Eps8 recruits Dishevelled to the plasma membrane and actin filaments suggesting that Eps8 might participate in non-canonical Wnt/Polarity signaling. Consistent with this idea, mis-expression of Eps8 in dorsal regions of Xenopus embryos resulted in gastrulation defects.
Conclusion
Together, these results suggest that Eps8 plays multiple roles in modulating actin filament organization, possibly through its interaction with distinct sets of actin regulatory complexes. Furthermore, the finding that Eps8 interacts with Dsh and induced gastrulation defects provides evidence that Eps8 might participate in non-canonical Wnt signaling to control cell movements during vertebrate development.
==== Body
Background
Remodeling of the actin cytoskeleton is critical for mediating changes in cell shape, migration, and adhesion. Actin filament architecture is regulated by a large group of actin binding proteins that modulate actin assembly, disassembly, branching, and bundling [1]. Actin organization is also regulated by growth factor signals that stimulate the activity of Rho family GTPases, which mediate actin remodeling and formation of stress fibers, filopodia, and membrane ruffles [2]. Although much has been learned about the general properties of actin binding proteins, the mechanisms by which these proteins control actin architecture in vivo are poorly understood.
Eps8 (EGF receptor pathway substrate 8) was originally identified as a substrate of the EGF receptor [3] and is the founding member of a multigene family of Eps8-like proteins named Eps8L1, Eps8L2, and Eps8L3 [4,5]. Eps8 is thought to transduce growth factor signals by acting as a scaffold protein to support the formation of multi-protein signaling complexes that promote the activation of Rho family GTPases. Consistent with this model, studies in Eps8 null fibroblasts showed that Eps8 is required for growth factor-induced Rac activation as well as Rac-dependent actin remodeling and membrane ruffling [6]. Eps8 is a critical component of a complex that contains the p85 regulatory subunit of phosphoinositide 3-kinase, Abi1, and Sos1, which acts as a guanine nucleotide exchange factor (GEF) for Rac [6,7]. Eps8 interacts directly with Abi1 through its SH3 domain, which possesses a novel peptide binding specificity [8], and this binding is thought to relieve auto-inhibition of Eps8 [9].
Eps8 also directly binds actin, suggesting that it may function by localizing Rac to sites of actin remodeling [10]. Eps8 binds actin through its C-terminal effector domain and expression of the effector region in serum-starved cells elicits Rac-dependent actin remodeling and membrane ruffling [10]. Studies using deletion mutants of Eps8 show that the C-terminal effector domain is required for localizing Eps8 to membrane ruffles and the transduction of signals to Rac [10]. A recent study revealed that C-terminal fragments of Eps8 also possess actin barbed-end capping activity in vitro and can substitute for capping protein in actin-based motility assays, suggesting a mechanism by which Eps8 might regulate actin filament dynamics in vivo [9]. Interestingly, full-length Eps8 on its own lacks capping activity in vitro, but can block actin polymerization in the presence of Abi1 [9]. The capping activity of Eps8 does not require Rac indicating that Eps8 can modulate actin dynamics through Rac-dependent and -independent mechanisms. Together, these data implicate Eps8 as a key regulator of actin filament dynamics and suggest that its activity is modulated through association with distinct sets of interacting regulatory proteins.
Eps8 has also been shown to bind Dishevelled (Dsh) [11], a key regulator of canonical and non-canonical Wnt signaling [12,13]. Dsh is required for the establishment of cell polarity and directed migration during gastrulation in vertebrates [14-16]. The mechanism by which Dsh controls cell polarity and migration is unclear, but is hypothesized to involve the modulation of actin dynamics through activation of RhoA and Rac [17,18]. The ability of Eps8 to bind both Dsh and actin and stimulate Rac activation suggests that Eps8 may play an important role in regulating Dsh function during gastrulation, but this possibility has not been investigated.
In this study, we utilized cultured mammalian cells and Xenopus embryos as model systems to investigate the mechanism by which Eps8 regulates actin filament architecture in vivo. Our results provide evidence that Eps8 can stimulate the assembly of distinct types of actin-based structures in cells and that the morphology of the actin structures induced by Eps8 is dependent on its interactions with Abi1. In addition, we show that Eps8 can recruit actin regulatory proteins, such as N-WASP and Dsh, to actin filaments and that mis-expression of Eps8 impairs cell movements during gastrulation in Xenopus embryos. Together, these data suggest that the role of Eps8 in modulating actin organization is multifaceted and is dependent on its participation in several potentially distinct multi-protein actin regulatory complexes.
Results
Enhanced formation of filopodia-like structures in cells expressing Eps8
To gain insights into the role Eps8 plays in regulating actin filament architecture, we examined the effect of increasing Eps8 levels on actin remodeling in mammalian cultured cells. For these studies, we utilized the mouse melanoma cell line B16F1 [19], the human breast cancer cell line MDA-MB231 [20], and the MDA-MB231BO cell line, which is a highly metastatic, bone seeking clone of the parental line [21]. These cells were chosen because they are highly motile and express a variety of cellular protrusions including lamellipodia and filopodia. Control B16F1, MDA-MB231, and MDA-231BO cells stained for actin are shown in Figure 1. We found that expression of a c-myc epitope tagged version of mouse Eps8 (Eps8-myc) in B16F1 cells elicited the formation of filopodia-like structures, which stained brightly with phalloidin (Figure 1D–I). The filopodia-like structures extended from lateral and dorsal regions of the cell and Eps8 localized along the length of these protrusions and was enriched at their tips (Figure 1F, inset). Similar results were seen in MDA-MB231 (Figure 1J–L) and MDA-MB231BO (Figure 1M–O) breast cancer cells. More than 90% of the transfected cells displayed the actin phenotype shown. We also observed the formation of long, snake-like actin cables in approximately 50% of the MDA-MB231BO cells, which were typically not seen in either B16F1 cells or the parental MDA-MB231 cells.
Figure 1 Eps8 induced actin remodeling in cultured cells. Phalloidin staining of the actin cytoskeleton in untransfected (A) B16F1, (B) MDA-MB231, and (C) MDA-MB231BO cells. Cells possess few filopodia-like structures extending from lateral and dorsal surfaces and do not possess cytoplasmic actin cables. Actin structures induced by Eps8 in (D-I) B16F1, (J-L) MDA-MB231, and (M-O) MDA-MB231BO cells. Distribution of Eps8-myc (D,G,J,M) revealed by 9e10 anti-c-myc antibody and actin (E,H,K,N) revealed by phalloidin staining in fixed cells. Right column (F,I,L,O) shows merged images with Eps8-myc in red and actin in green. The boxed region in (F) is enlarged in the inset. Eps8 induces the formation of filopodia-like structures in B16F1 and MDA-MB231 cells and localizes to filopodia-like structures, ruffles, and actin cables in MDA-MB231BO cells. In B16F1 cells, Eps8 localizes along the length of the filopodia-like structures (arrowheads in G-I) and is enriched at their tips (arrowheads, inset in F). Scale bar is equal to 10 μm in (A-F) and (J-O) and 5 μm in (G-I).
Abi1 modulates Eps8-dependent actin remodeling
To test whether Abi1 can modulate the activity of Eps8 in cultured cells we examined the effect of co-expressing Eps8 and Abi1 on actin architecture. Similar to data reported previously [9], simultaneous expression of Eps8-myc and Abi1-GFP in B16F1, MDA-MB231, and MDA-MB231BO cells resulted in remodeling of the actin cytoskeleton characterized by formation of cable-like actin bundles within the cytoplasm (Figure 2A–I). The actin cables were typically found at the ventral surface of the cell and displayed few branches. Eps8 and Abi1 co-localized along the length of the actin cables. Interestingly, Abi1 was not enriched with Eps8 in filopodia-like structures (Figure 2, arrowheads in A-C), suggesting that Abi1 may not contribute to Eps8-function at the plasma membrane. More than 95% of transfected cells displayed the actin phenotype shown. Expression of Abi1 alone (data not shown) or an Abi1 mutant (Abi1DY) unable to bind Eps8 [22] failed to stimulate actin cable formation (Figure 2J–L), indicating that the ability of Eps8 to induce actin cables is dependent on its interaction with Abi1.
Figure 2 Abi1 modulates the morphology of actin structures induced by Eps8. Actin structures induced by Eps8 and Abi1 in (A-C) B16F1, (D-F) MDA-MB231, and (G-I) MDA-MB231BO cells. Distribution of Eps8-myc (A,D,G,J) revealed by 9e10 anti-c-myc antibody, Abi1-GFP (B,E,H), and Abi1DY-GFP (K) revealed by GFP, and actin (C,F,I,L) revealed by phalloidin staining in fixed cells. (A-I) Simultaneous expression of Eps8 and Abi1 induces the formation of actin cables in B16F1, MDA-MB231, and MDA-MB231BO. Eps8 and Abi1 co-localize in association with actin cables (arrows) but Abi1 is not enriched with Eps8 in filopodia (arrowheads in A-C). (J-L) Formation of actin cables in B16F1 cells is dependent on the interaction of Eps8 and Abi1. Abi1DY does not co-localize with Eps8 and does not induce actin cable formation. Scale bar is equal to 10 μm.
Eps8 induces actin remodeling in Xenopus embryos
To further examine the role of Eps8 in regulating actin architecture, we utilized Xenopus animal cap explants, which provide a powerful system for analyzing protein localization and function in vivo. Animal caps explants are dissected from blastula stage embryos and consist of an outer polarized epithelium and 2–3 layers of non-epithelial deep cells. We found that expression of Eps8 has different effects on actin organization in superficial epithelial cells versus deep cells. In control explants, actin filaments are enriched at apical cell-cell junctions in superficial epithelial cells (Figure 3A) and at the cortex of deep cells facing the blastocoel (Figure 3E). In superficial epithelial cells, Eps8 expression caused an accumulation of actin filaments at sites of cell-cell contact in apparent association with adherens junctions (Figure 3B–D, arrowheads). In contrast, Eps8 expression induced the formation of cable-like actin structures within the cytoplasm of deep cells (Figure 3F–H, arrows) and modified the morphology of actin filaments at the cell cortex (Figure 3F–H, arrowheads). The morphology and length of the actin structures in deep cells was variable; long, unbranched filaments were observed in cortical regions in association with the free membrane domain that faces the blastocoel, whereas thick actin bundles were often seen throughout the cytoplasm. Staining of animal caps with anti-myc antibodies showed that Eps8-myc localized along the length of actin filaments in both deep and superficial cells (Figure 3D,H; co-localization appears yellow). Thus, Eps8 associates with actin filaments and can dramatically affect the organization of actin in Xenopus animal cap cells as it does in cultured cells.
Figure 3 Eps8-induced actin remodeling in Xenopus embryos. Distribution of actin (A,C,D,E,G,H) revealed by phalloidin staining and Eps8-myc (B,D,F,H) revealed by 9e10 anti-c-myc antibody in animal cap cells of Xenopus embryos. In control animal caps, actin filaments are enriched at cell-cell junctions in superficial epithelial cells (A) and the cortex of deep cells (E). (B-D) In superficial epithelial cells, Eps8 expression causes an enrichment of actin filaments at cell-cell junctions (arrowheads). (F-H) In deep cells facing the blastocoel, Eps8 expression induces the formation of actin cables (arrowheads). (D,H) Eps8 is red and actin is green in merged images. Scale bar is equal to 10 μm.
Abi1 modulates the activity of Eps8 in Xenopus embryos
To test whether Abi1 can regulate Eps8 function in Xenopus embryos, we co-expressed Eps8-myc and Abi1-GFP in animal cap cells and analyzed the localization of Eps8, Abi1, and actin by confocal microscopy. We found that when expressed alone, Abi1-GFP localized to small aggregates found throughout the cytoplasm and did not affect actin organization (data not shown). In contrast, simultaneous expression of Eps8-myc and Abi1-GFP induced the formation of star-like actin structures in superficial epithelial cells of the animal cap (Figure 4A–C). Actin stars were found at the apical surface and consisted of actin-containing spikes radiating from a central actin foci or short bundle. The actin stars did not appear to protrude from the apical surface and Eps8 and Abi1 co-localized with actin in the stars. Since Eps8 and Abi1 facilitate signaling through Rac in cultured cells we tested whether actin star formation was dependent on Rac. In control animal caps, endogenous Rac was enriched at the cell cortex in association with cell-cell junctions (data not shown). In animal caps expressing Eps8 and Abi1, Rac was not recruited to the actin stars (Figure 4D–F) suggesting that Rac activity is not required for actin star formation. In agreement with this idea, expression of dominant negative Rac (RacN17) failed to inhibit Eps8/Abi1-induced actin star formation (data not shown). Thus, Abi1 modulates Eps8 activity in Xenopus and Eps8 and Abi1 can stimulate actin remodeling in a Rac-independent manner.
Figure 4 Abi1 modulates Eps8-induced actin remodeling in a Rac-independent manner. Distribution of Eps8-myc (A,D), Abi1GFP (B), actin (C) revealed by phalloidin staining, and Rac (E,F) in Xenopus animal cap cells. (A-C) Simultaneous expression of Eps8 and Abi1 induce the formation of actin stars at the apical surface of superficial epithelial cells. (D-F) Endogenous Rac is not recruited to Eps8/Abi1-induced actin stars. (F) Eps8 is red and Rac is green in the merged image. Scale bar is equal to 10 μm in A,B and D-F and to 5 μm in C.
Recruitment of Actin Regulatory Proteins to Eps8/Abi1-induced actin structures in Xenopus
Eps8 has been shown to possess Abi1-dependent barbed-end capping activity in vitro [9], suggesting that the effects we observed in Xenopus may be due to increased capping of actin filaments. To test this idea, we analyzed whether expression of capping protein induced similar changes in actin organization. Capping protein (CP) is an α/β heterodimer that is thought to provide the major barbed-end capping activity in eukaryotic cells [23,24]. In these experiments, animal caps expressing both the α and β subunits of CP were examined for changes in actin filament distribution. In addition, since both the α and β subunits were GFP-tagged, their expression was confirmed by Western blot analysis using anti-GFP antibodies (data not shown). We found that expression of CP had no effect on actin organization in animal cap cells (data not shown). In addition, we found that expression of capping protein did not block the formation of Eps8/Abi1-induced actin stars, although low levels of capping protein were found to co-localize with the actin stars (Figure 5A–D, arrowhead). Thus, the formation of actin stars does not directly correlate with enhanced capping protein activity nor does enhanced capping protein activity affect Eps8/Abi1-induced remodeling of the actin cytoskeleton.
Figure 5 Regulation of Eps8/Abi1-induced actin remodeling in Xenopus. Distribution of Eps8-myc (A,E,I,M; red in D,H,L,P), actin (B,F,J,N; green in D,H,L,P), CP-GFP (C, blue in D), N-WASP-GFP (G, blue in H), FP4-mito-GFP (K, blue in L), and Xvasp-GFP (O, blue in P). (A-D) CP does not block formation of actin stars. (E-H) N-WASP is recruited to Eps8/Abi1-induced actin stars. Actin star formation is not altered in response to inhibition of Ena/VASP activity (I-L) or increased levels of Xvasp (M-P). Scale bar is equal to 10 μm in A-L and 5 μm in M-P.
WASP/Scar proteins play an important role in stimulating actin filament nucleation by the Arp2/3 complex [25-27]. To test whether the formation of actin stars involves recruitment of WASP proteins we analyzed the distribution of N-WASP-GFP in animal cap cells expressing Eps8 and Abi1. N-WASP co-localized with Eps8 and actin (Figure 5E–H), indicating that WASP proteins are recruited to Eps8/Abi1-induced actin structures. We also tested whether N-WASP activity is required for Eps8/Abi1-induced actin star formation by co-expressing Eps8, Abi1 and a dominant negative form of N-WASP (N-WASP-CA). We found that N-WASP-CA expression did not significantly alter the actin structures induced by Eps8 and Abi1 (data not shown). These data suggest that Eps8 and Abi1 can recruit actin nucleators to specific sites in the cell, although N-WASP function may not be strictly required for Eps8/Abi1-induced actin remodeling.
Members of the Ena/VASP family are critical regulators of actin filament dynamics and are thought to antagonize actin filament capping at the leading edge of migrating cells [28]. Given this central role, we tested whether increased or decreased Ena/VASP activity would affect Eps8/Abi1-induced actin star formation. Expression of a dominant negative protein (FP4-mito-GFP, [28,29]) that specifically neutralizes the function of all Ena/VASP proteins was used to knockdown Ena/VASP activity whereas expression GFP-tagged Xenopus VASP (Xvasp) was used to increase Ena/VASP activity. The ability of the FP4-mito dominant negative to mis-localize Ena/VASP proteins in Xenopus was confirmed by showing that it caused the redistribution of endogenous Ena from the cell periphery to the mitochondria surface (data not shown). We found that neither FP4-mito-GFP (Figure 5I–L) nor Xvasp-GFP (Figure 5M–P) had an effect on the presence of Eps8/Abi1-induced actin stars. In addition, Xvasp-GFP did not co-localize with the actin stars, indicating that Ena/VASP proteins are not recruited to these actin structures (Figure 5M–P).
Eps8 recruits Dsh to the membrane and actin filaments
Previous studies have reported that Eps8 can bind the Wnt signaling protein Dsh [11], which is required for the transduction of both canonical and non-canonical Wnt signals [13]. Since Dsh is required for cell polarization and convergent extension movements during gastrulation [14-16,30,31], we hypothesized that the formation of an Eps8/Dsh complex may be important for regulating Dsh localization and function during gastrulation. To test this idea, we asked whether Eps8 interacts with Dsh in animal cap cells. When expressed alone, Dsh-GFP displays a punctate cytoplasmic distribution in animal cap explants (Figure 6A) [32]. Expression of Eps8 caused a dramatic redistribution of Dsh-GFP to the plasma membrane and cytoplasmic actin filaments where it co-localized with actin and Eps8 (Figure 6B–F). In superficial epithelial cells, Dsh was recruited to cell-cell junctions (Figure 6B,C; arrow) and in deep cells Dsh was recruited to cytoplasmic actin cables (Figure 6D–F; arrow) and the cell cortex (Figure 6D–F; arrowhead). Furthermore, we found that epitope-tagged forms of Dsh, Eps8, and Abi1 co-localize in animal cap cells (Figure 6G–I) suggesting that they can form a tri-complex in vivo. These data provide evidence that Eps8 interacts with and may regulate the distribution and/or function of Dsh through recruitment of Dsh to the membrane and actin filaments.
Figure 6 Dsh is recruited to the plasma membrane and actin filaments in response to Eps8 expression. Distribution of Dsh-GFP (A,B,D, green in F), actin (C), Eps8-myc (E,G, red in F), Abi1-GFP (H), and Dsh-flag (I) in Xenopus animal cap cells. (A) Localization of Dsh-GFP in control animal caps. (B,C) In response to Eps8 expression, Dsh is recruited to the membrane and cell-cell junctions in superficial epithelial cells where it co-localizes with actin (arrows). (D-F) In deep cells facing the blastocoel, Eps8 induced the recruitment of Dsh to cytoplasmic actin cables and the cell cortex where Dsh co-localized with Eps8 (co-localization appears yellow). (G-I) Co-localization of Eps8, Abi1, and Dsh in animal caps cells (arrows mark site of co-localization). Scale bars are equal to 10 μm.
Identification and developmental expression of Xenopus Eps8
Eps8 can interact with Dsh and is thought to play an important role in regulating actin remodeling in motile cells, raising the possibility that Eps8 might be a key regulator of cell movements during gastrulation in vertebrate embryos. To begin to address the role of Eps8 during embryonic development, we performed in silico analyses to identify the Xenopus ortholog of Eps8. Searches of the TIGR (TC263683) and NCBI (MGC81285; Image 6631907) databases led to the identification of cDNAs that encode Xenopus Eps8 (XEps8). The predicted XEps8 protein shows a high degree of sequence identity with both mouse and human Eps8 and contains the conserved PTB, SH3, and C-terminal effector domains. The developmental expression of XEps8 transcripts was determined by RT-PCR. We found that XEps8 transcripts are provided maternally and are present throughout development (Figure 7A). We also found that XEps8 is expressed in isolated dorsal and ventral marginal zone tissue of gastrula stage embryos and that levels of XEps8 are higher in dorsal marginal regions compared to ventral regions (Figure 7A). Finally, we probed blots of embryonic lysates with anti-XEps8 polyclonal antibodies and found that XEps8 protein appears as a doublet and is present in unfertilized eggs, gastrula, and neurula stage embryos (Figure 7B). These analyses show that XEps8 is expressed at the relevant time and place to regulate cell movements during gastrulation.
Figure 7 Mis-expression of Eps8 results in gastrulation defects. (A) RT-PCR analysis of the developmental expression of XEps8. ODC serves as a control for RNA isolation and reverse transcription. (B) Western blot probed with anti-XEps8 antibodies show that XEps8 protein is provided maternally and is present in gastrula, neurula, and tailbud stage embryos. (C) Control and (D) Eps8-injected embryos at stage 12. In control embryos the blastopore is well formed and has progressed vegetally (arrowheads). In contrast, Eps8-injected embryos display severe buckling of tissue above the blastopore (arrowheads) and a disorganized blastopore lip that is delayed and malformed. (E) Control stage 37/38 embryos. (F) Eps8-injected embryos show a range of phenotypes including microcephaly, cyclopia, and shortening and arching of the A-P axis. Top = low dose (50 pg); middle = intermediate dose (200 pg); bottom = high dose (1 ng). (G,H) Histological analysis of Eps8-injected embryos shows that Eps8 expression causes a broadening of the notochord (no) and disorganization of the neural tube (nt) and somites (so).
To test the requirement for XEps8 during development, we utilized a morpholino (MO) antisense oligonucleotide targeted to the 5'-untranslated region to specifically knockdown levels of XEps8 protein during development. We found that the XEps8 MO could specifically block the expression of a myc-tagged version of XEps8, but injection of the XEps8 MO into 4-cell stage embryos resulted in embryos with no apparent phenotype (data not shown). The lack of a knockdown phenotype is not surprising since Eps8-/- mice also displayed no obvious phenotype [6]. Since Eps8 is a member of a multi-gene family, we searched TIGR and NCBI databases for additional Xenopus Eps8 genes and found evidence for a second XEps8 gene as well as three XEps8-like genes. Therefore, the lack of a phenotype in XEps8 knockdown embryos is likely due to the expression of multiple XEps8 family members, including XEps8L1, XEps8L2, and XEps8L3, during early development (Roffers-Agarwal and Miller, unpublished results). Thus, assessing the role of Eps8 proteins in Xenopus will require novel knockdown techniques capable of simultaneously and specifically inhibiting the activity of multiple gene products during early development.
Expression of Eps8 disrupts cell movements during gastrulation
Since knockdown experiments produced negative results, we performed mis-expression experiments to test whether altering Eps8 activity would affect cell movements during gastrulation. Synthetic mRNA encoding mouse Eps8-myc or GFP as a control was injected into the equatorial region of both dorsal blastomeres at the 4-cell stage and resulting embryos were then examined for developmental abnormalities. Defects in Eps8-injected embryos were first apparent at stage 10.5 (early gastrula). At this stage, control embryos formed a well-defined dorsal lip indicative of the onset of gastrulation movements and involution of dorsal mesoderm. In contrast, Eps8-injected embryos showed a delay in the formation of the dorsal lip and when observed, the lip was disorganized (data not shown). By stage 12, Eps8-injected embryos displayed a severe delay in blastopore closure and buckling of tissue above the blastopore (Figure 7D). Eps8-injected embryos eventually complete gastrulation and tadpoles displayed a phenotype including a shortened and arched anterior-posterior axis and head defects (Figure 7F). The defects caused by Eps8 are dose dependent; low doses (50 pg) of Eps8 result in cyclopia and a shortened A-P axis, moderate doses (200 pg) show varying degrees of cyclopia, microcephaly, and shortening and arching of the A-P axis, and high doses (1 ng) result in varying degrees of anencephaly, shortening and arching of the A-P axis, and spina bifida. Control, GFP-injected embryos appeared normal at all stages examined (Figure 7C,E). These data are consistent with the idea that Eps8-induced actin re-organization leads to defects in cell movements during gastrulation in Xenopus.
The gross morphological defects caused by dorsal expression of Eps8 could be the result of defects in convergent extension or inhibition of mesoderm development, both of which would give superficially similar phenotypes. In order to distinguish between these two possibilities we performed histological analysis on injected embryos (Figure 7G,H). Histological sections of Eps8-injected embryos demonstrated that notochord, somites, and neural tissue are all present, showing that expression of Eps8 does not globally perturb specification of mesodermal or neural cell fates. Instead, expression of Eps8 resulted in broadening of the notochord along the mediolateral axis and morphological defects in the neural tube and somites. The widening of the notochord is consistent with the idea that expression of Eps8 impairs convergent extension movements of the axial mesoderm.
Analysis of activin-induced elongation of animal cap explants provides a powerful assay for studying the cell movements associated with gastrulation [14,33,34]. In these experiments, the animal pole region of an embryo is removed at the blastula stage and placed in culture. Untreated animal cap explants and caps expressing Eps8 differentiate into atypical epidermis and remain rounded (Figure 8A,B) whereas addition of recombinant activin induces mesodermal differentiation, convergent extension movements, and elongation of uninjected explants (Figure 8C). We found that expression of Eps8 inhibits activin-induced elongation of animal cap explants (Figure 8D). The failure of activin-induced animal caps to elongate was not caused by a block in mesoderm induction since both Xbra (pan-mesoderm) and XmyoD (paraxial mesoderm) were expressed in control and Eps8-injected animal caps following activin treatment (Figure 8E).
Figure 8 Eps8 blocks elongation of activin treated animal caps. (A,B) Control and Eps8-injected animal cap explants remain rounded in the absence of activin. (C,D) Control explants elongate extensively in the presence of activin, whereas Eps8 expression inhibits elongation. (E) RT-PCR analysis shows that Eps8 expression does not block activin-mediated induction of the mesodermal markers Xbra and MyoD. ODC is a control for mRNA isolation, reverse transcription, and gel loading.
Discussion
Here, we have investigated how Eps8 regulates actin filament architecture and how this activity impacts cell movements during gastrulation. Our results, together with previous studies, provide evidence that Eps8 plays multiple roles in regulating the actin cytoskeleton and that these functions are influenced by the participation of Eps8 in multi-protein actin regulatory complexes.
Based on in vitro studies, Eps8 is hypothesized to promote capping of actin barbed-ends in an Abi1-dependent manner [9]. Our findings suggest that in addition to its proposed role as a barbed end capping protein, Eps8 might play additional roles in regulating actin organization in vivo. This idea is supported by the observation that Eps8 expression resulted in enhanced formation of actin-rich filopodia-like structures in cultured cells and enhanced formation of actin bundles and accumulation of actin at cell-cell junctions in Xenopus embryos. The presence of the filopodia-like structures on the dorsal surface of cells suggests that they are protrusive in nature and do not represent retraction structures, which are typically associated with sites of cell adhesion. Additional studies examining the dynamics of these Eps8-induced structures will help clarify the origin and nature of these structures. In addition, we found that Abi1 modulated Eps8 activity, promoting the formation of actin cables in cultured cells and actin stars in Xenopus, suggesting that Eps8 can regulate actin dynamics through Abi1-dependent and -independent mechanisms. Consistent with this idea, Abi1 did not co-localize with Eps8 at the tips of the filopodia-like structures in cultured cells suggesting that additional regulators of Eps8 remain to be identified.
The correlation between Eps8 expression and enhanced formation of filopodia-like structures and actin cables is consistent with the idea that Eps8 may regulate actin filament elongation in vivo. Regulation of barbed-end elongation and filopodia formation is thought to involve a balance between barbed-end capping and anti-capping activities. Proteins such as CP are hypothesized to block elongation and favor formation of a dendritic network [35], whereas proteins including Ena/VASP proteins, which antagonize capping, are hypothesized to promote actin filament elongation and filopodia formation [28,36,37]. Our work examining the regulation of Eps8 activity by CP, N-WASP, and Ena/VASP in Xenopus yielded largely negative results, however, making it difficult to discern the relative contribution of Eps8 capping activity versus other potential modes of activity in the regulation of actin architecture. Further biochemical analyses will help elucidate the molecular mechanism(s) by which Eps8 regulates actin dynamics in vivo.
Previous work [6,7,9,38] and our results show that the ability of Eps8 to modulate actin organization is regulated by its interaction with distinct binding partners such as Abi1. We found that Abi1 can modulate Eps8 activity in cultured cells and Xenopus embryos. Abi1 binds to the SH3 domain of Eps8 [38,39] and it has been proposed that this binding may alter the conformation or activity of the adjacent actin-binding domain of Eps8 [9]. The mechanism by which Abi1 might regulate Eps8 activity remains unclear, but may involve recruitment of additional regulatory factors such as Dsh, Sos1, and Rac to the Eps8/Abi1 complex [7,38]. In addition, our work shows that N-WASP is recruited to Eps8/Abi1-induced actin stars suggesting that the Eps8/Abi1 complex interacts either directly or indirectly with actin nucleating factors. This idea is supported by the observation that Eps8 can facilitate actin-based motility of N-WASP-coated beads in vitro in the presence of Arp2/3, ADF/cofilin, and profilin [9]. Further studies will be required to examine how Abi1 modulates Eps8 activity and how Eps8 works with Abi1 and other regulatory factors to control actin organization in vivo.
Eps8 has been shown to bind Dsh [11], a component of the Wnt signaling pathway that is required for transduction of canonical Wnt/β-catenin and non-canonical signals [12,13]. Here, we have shown that Eps8 expression recruits Dsh to actin filaments and the cell membrane in Xenopus. These data are significant because the role of Dsh in non-canonical Wnt/Polarity signaling is thought to be dependent on its localization to the membrane and its ability to affect cell polarity and migration through regulation of the actin cytoskeleton [14-18]. Dsh activity during gastrulation is dependent on both RhoA and Rac, and the formin homology protein DAAM1 is required for Dsh-mediated activation of RhoA [17,18]. However, a link between Dsh and Rac has not been identified. The Eps8/Abi1/Sos1 complex is required for growth factor stimulated activation of Rac [6], suggesting that Eps8 might provide an important link between Dsh, Rac, and the actin cytoskeleton during development. Consistent with this idea, expression of Eps8 impaired cell movements during gastrulation and Eps8, Abi1, and Dsh co-localize in Xenopus suggesting that these proteins can form a tri-complex in vivo. Interestingly, we did not observe an effect of Eps8 on Dsh-mediated induction of Wnt/β-catenin target genes (siamois and Xnr3, JRA and JRM unpublished results) indicating that Eps8 does not participate in canonical Wnt/β-catenin signaling. Unfortunately, our attempts to analyze the requirement for Eps8 in Xenopus were unsuccessful due to the expression of multiple Eps8 family members during early development. Thus, additional studies are necessary to determine the potential role of Eps8 in the transduction of non-canonical Wnt signals and the potential role of Eps8 family members during gastrulation in vertebrates.
Conclusion
How might Eps8 regulate the actin cytoskeleton in vivo? Our findings together with data from previous studies support the idea that Eps8 might regulate actin architecture in multiple ways. Eps8 can bind to both barbed ends and the sides of actin filaments [9,10] and it is possible that these different modes of actin binding mediate distinct effects on actin architecture in cells. Barbed-end capping activity might regulate actin filament dynamics and stabilize existing filaments whereas an alternative activity might promote the formation and maintenance of actin arrays required for protrusive force generation and cellular structures such as microvilli and filopodia. This idea is consistent with our observation that Eps8 is enriched at the tips of filopodia-like structures and localizes along the length of the filopodia-like structures and actin cables. This model is also in agreement with the observation that Eps8 localizes to microvilli in the intestinal epithelium of C. elegans and knockdown of Eps8 is associated with defects in microvilli formation [40]. The formation of actin cables in cells expressing Eps8 and Abi1 and actin clusters in Xenopus embryos suggests that Abi1 is a critical modulator of Eps8's activity as an actin regulatory protein. The finding that Eps8 expression impairs cell movements during gastrulation provides further support for this view and underscores the idea that the proper balance of actin assembly, disassembly, and organization is essential for controlling morphogenetic movements during development. Thus, Eps8 has emerged as a critical regulator of actin filament dynamics and further analysis of Eps8 and its binding partners will help shed light on the mechanisms that mediate actin-based motility in vivo.
Methods
Expression constructs, antibodies, and cells
Constructs used were: mouse Eps8-myc pCS2+ (Eps8 cDNA provided by Dr. P.P. DiFiore, European Institute of Oncology, Milan [6]), human Abi1-GFP pCS2+ (Abi1 cDNA provided by Dr. Ann Marie Pendergast, Duke University; [41]), human capping protein α-GFP and β-GFP pCS2+ (capping protein cDNAs provided by Dr. Dorothy Schafer, University of Virginia [42]), FP4-mito-GFP pCS2+ (FP4-mito cDNA provided by Dr. Frank Gertler, MIT; [29]), RacN17 pCS2+ (RacN17 cDNA provided by Dr. Jennifer Westendorf, University of Minnesota), Xenopus Dsh-GFP pCS2+ [32], and Xenopus Dsh-flag pCS2+ [43]. Xenopus N-WASP-mRFP pCS2+, Xenopus N-WASP-CA-mRFP pCS2+ and Xenopus VASP-GFP pCS2+ were constructed by PCR using full length IMAGE cDNAs obtained from ATCC. Details of vector construction are available upon request. Primary antibodies used were: mouse anti-c-myc 9e10 [44], mouse anti-flag (Sigma), rabbit anti-Eps8 (Santa Cruz Biotechnology), and anti-Rac (Transduction Labs). Anti-XEps8 rabbit polyclonal antibodies were raised against a peptide corresponding to the carboxyl terminal (NH2-SDSGVESFDEGNSH-COOH) conjugated to KLH (Sigma Genosys). A cysteine residue was added to the amino terminus of the peptide to facilitate conjugation to KLH. Secondary antibodies used were: Alexa568 goat anti-mouse, Alexa647 goat anti-mouse, and goat anti-rabbit Alexa568 and Alexa647 (all secondary antibodies were from Molecular Probes). Alexa568 phalloidin (Molecular Probes) was used to visualize F-actin. Cells used were B16F1 mouse melanoma cells (ATCC), MDA-MB231, and MDA-MB231BO (provided by Dr. Douglas Yee, University of Minnesota).
Cell culture, transfections, and imaging
B16F1, MDA-MB231, and MDA-MB231BO cells were grown in DMEM (CellGro) supplemented with 10% FBS (HyClone) at 5% CO2. For transfections, cells were plated on acid washed coverslips and transfected with Lipofectamine (Invitrogen). For imaging, cells were washed once with PBS and fixed in 4% formaldehyde in CSK buffer (10 mM Hepes pH 7.5, 150 mM sucrose, mM EGTA, 0.1% Triton X-100) for 15 min. at room temperature. Alternatively, cells were permeabilized with 0.1% Triton X-100 in PEM buffer (10 mM Pipes pH 7.4, 1 mM EDTA, 1 mM MgCl2) for 30 seconds and fixed with pre-warmed 4% paraformaldehyde in PEM buffer for 30 min. at 37°C. Fixed cells were then washed three times in PBS + 0.1% Triton X-100 (PBST), and incubated in PBST, 2% BSA, 10% normal goat serum (NGS) to prevent non-specific binding of antibodies. Staining with primary and secondary antibodies was performed in PBST, 2% BSA, 10% NGS for 2 hours at room temperature. Images were collected using a Zeiss spinning disc confocal microscope and digital images were processed using Adobe Photoshop.
Embryos, microinjections, imaging, RT-PCR, and Western blotting
Xenopus laevis eggs were fertilized in vitro and subsequently de-jellied in 2% cysteine (Sigma Chemical). Embryos were reared in 1/3× Marc's Modified Ringer's (MMR). Embryos were staged according to Nieuwkoop and Faber [45]. Animal cap explants were prepared using a Gastromaster microsurgery instrument (Xenotek Engineering) and cultured in 1× Steinberg's in the presence of 10 ng/ml Activin (R&D Systems). For microinjections, embryos were placed in a solution of 4% Ficoll in 1/3× MMR and injected using a Harvard Apparatus microinjector and a Narashige micromanipulator. Injected embryos were reared in 4% Ficoll in 1/3× MMR supplemented with 10 μg/ml Gentamicin (Invitrogen) to stage 9 then washed and reared in 1/3× MMR + Gentamicin. Capped mRNA for injections was synthesized using mMessage Machine (Ambion) and purified using NucAway columns (Ambion).
For imaging, embryos and explants were fixed in 4% formaldehyde in CSK buffer at room temperature for 30 min., washed three times in PBST, and incubated in PBST, 2% BSA, 10% NGS to prevent non-specific binding of antibodies. Staining with primary and secondary antibodies was performed in PBST, 2% BSA for 2 hours at room temperature. Actin was visualized with Alexa568 phalloidin (Molecular Probes). Images were captured with a Zeiss spinning disk confocal microscope and digital images were processed with Adobe Photoshop.
RT-PCR analysis was performed using total RNA isolated from Xenopus explants. cDNA was prepared using SuperScript II reverse transcriptase (Invitrogen) and PCR was performed using GoTaq polymerase (Promega). Primers used for RT-PCR: XEps8: forward: 5'-attccctgagatgttgctccg-3', reverse: 5'-tagcagcagcgatttgccc-3'; XmyoD: forward: 5'-agctccaactgctccgacggcatgaa-3', reverse: 5'-aggagagaatccagttgatggaaca-3'; Xbra: forward: 5'-ggatcgttatcacctctg-3', reverse: 5'-gtgtagtctgtagcagca-3'; and ODC: forward: 5'-gccattgtgaagactctctccattc-3', reverse: 5'-ttcgggtgattccttgccac-3'.
Protein lysates for Western blots were prepared by homogenizing embryos in ice-cold lysis buffer (20 mM Tris pH 7.5, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 0.5% Triton X-100) supplemented with protease inhibitors (1 mM PMSF, 1 mM pepstatin, 10 μg/ml leupeptin, and 10 μg/ml aprotinin). Homogenates were cleared by centrifugation at 14,000 rpm for 10 min. at 4°C. SDS sample buffer was added to the cleared lysate and boiled for 4 min. prior to separation by SDS-PAGE. Approximately one embryo equivalent was loaded per lane on 10% gels (BioRad). Proteins were blotted to PVDF membrane (BioRad), blots were blocked in 5% milk in TBS + 0.1% Tween, and probed with anti-XEps8 antibodies (1:2000) for two hours at room temperature. Visualization was performed using a horseradish peroxidase conjugated anti-rabbit secondary antibody (Jackson ImmunoLabs) and enhanced chemiluminescence (Pierce).
List of abbreviations used
Eps8: Epidermal Growth Factor substrate 8, Abi1: Abl Interacting Protein 1, CP: capping protein, Dsh: Dishevelled, RT-PCR: Reverse Transcriptase Polymerase Chain Reaction
Authors' contributions
JRA participated in analysis Eps8 and Abi1 regulation of actin in B16F1 cells, participated in analysis of Eps8, Abi1, and Dsh interactions in Xenopus, performed Eps8/Abi1 interactions with capping protein, N-WASP, and Ena/VASP proteins, and performed RT-PCR study of analysis of Eps8 activity in Wnt signaling. JBX carried out histological analysis of Eps8-injected embryos. JRM participated in experiments analyzing Eps8 and Abi1 regulation of actin in B16F1 cells, performed studies in MDA-MB231 cells, conducted studies examining the role of Eps8 in gastrulation, coordinated the study, and drafted the paper. All authors read and approved the final manuscript.
Acknowledgements
The authors wish to thank members of the Miller lab for critical reading of the manuscript. We are grateful to Drs. Pier Paolo Di Fiore, Anne Marie Pendergast, Dorothy Schafer, Jennifer Westendorf, and Lorene Lanier for cDNA constructs. We thank our colleague Dr. Lorene Lanier for helpful insights and comments on the manuscript. This work was supported by a grant from the National Science Foundation to J.R.M. (0315767).
==== Refs
Revenu C Athman R Robine S Louvard D The co-workers of actin filaments: from cell structures to signals Nat Rev Mol Cell Biol 2004 5 635 646 15366707 10.1038/nrm1437
Raftopoulou M Hall A Cell migration: Rho GTPases lead the way Dev Biol 2004 265 23 32 14697350 10.1016/j.ydbio.2003.06.003
Fazioli F Minichiello L Matoska V Castagnino P Miki T Wong WT Di Fiore PP Eps8, a substrate for the epidermal growth factor receptor kinase, enhances EGF-dependent mitogenic signals Embo J 1993 12 3799 3808 8404850
Tocchetti A Confalonieri S Scita G Di Fiore PP Betsholtz C In silico analysis of the EPS8 gene family: genomic organization, expression profile, and protein structure Genomics 2003 81 234 244 12620401 10.1016/S0888-7543(03)00002-8
Offenhauser N Borgonovo A Disanza A Romano P Ponzanelli I Iannolo G Di Fiore PP Scita G The eps8 family of proteins links growth factor stimulation to actin reorganization generating functional redundancy in the Ras/Rac pathway Mol Biol Cell 2004 15 91 98 14565974 10.1091/mbc.E03-06-0427
Scita G Nordstrom J Carbone R Tenca P Giardina G Gutkind S Bjarnegard M Betsholtz C Di Fiore PP EPS8 and E3B1 transduce signals from Ras to Rac Nature 1999 401 290 293 10499589 10.1038/45822
Innocenti M Frittoli E Ponzanelli I Falck JR Brachmann SM Di Fiore PP Scita G Phosphoinositide 3-kinase activates Rac by entering in a complex with Eps8, Abi1, and Sos-1 J Cell Biol 2003 160 17 23 12515821 10.1083/jcb.200206079
Kishan KV Scita G Wong WT Di Fiore PP Newcomer ME The SH3 domain of Eps8 exists as a novel intertwined dimer Nat Struct Biol 1997 4 739 743 9303002 10.1038/nsb0997-739
Disanza A Carlier MF Stradal TE Didry D Frittoli E Confalonieri S Croce A Wehland J Di Fiore PP Scita G Eps8 controls actin-based motility by capping the barbed ends of actin filaments Nat Cell Biol 2004 6 1180 1188 15558031 10.1038/ncb1199
Scita G Tenca P Areces LB Tocchetti A Frittoli E Giardina G Ponzanelli I Sini P Innocenti M Di Fiore PP An effector region in Eps8 is responsible for the activation of the Rac- specific GEF activity of Sos-1 and for the proper localization of the Rac-based actin-polymerizing machine J Cell Biol 2001 154 1031 1044 11524436 10.1083/jcb.200103146
Inobe M Katsube K Miyagoe Y Nabeshima Y Takeda S Identification of EPS8 as a Dvl1-associated molecule Biochem Biophys Res Commun 1999 266 216 221 10581192 10.1006/bbrc.1999.1782
Veeman MT Axelrod JD Moon RT A second canon. Functions and mechanisms of beta-catenin-independent Wnt signaling Dev Cell 2003 5 367 377 12967557 10.1016/S1534-5807(03)00266-1
Wharton KA Jr Runnin' with the Dvl: proteins that associate with Dsh/Dvl and their significance to Wnt signal transduction Dev Biol 2003 253 1 17 12490194 10.1006/dbio.2002.0869
Tada M Smith JC Xwnt11 is a target of Xenopus Brachyury: regulation of gastrulation movements via Dishevelled, but not through the canonical Wnt pathway Development 2000 127 2227 2238 10769246
Wallingford JB Harland RM Xenopus Dishevelled signaling regulates both neural and mesodermal convergent extension: parallel forces elongating the body axis Development 2001 128 2581 2592 11493574
Wallingford JB Rowning BA Vogeli KM Rothbacher U Fraser SE Harland RM Dishevelled controls cell polarity during Xenopus gastrulation Nature 2000 405 81 85 10811222 10.1038/35011077
Habas R Dawid IB He X Coactivation of Rac and Rho by Wnt/Frizzled signaling is required for vertebrate gastrulation Genes Dev 2003 17 295 309 12533515 10.1101/gad.1022203
Habas R Kato Y He X Wnt/Frizzled activation of Rho regulates vertebrate gastrulation and requires a novel Formin homology protein Daam1 Cell 2001 107 843 854 11779461 10.1016/S0092-8674(01)00614-6
Fidler IJ Biological behavior of malignant melanoma cells correlated to their survival in vivo Cancer Res 1975 35 218 224 1109790
Cailleau R Young R Olive M Reeves WJ Jr Breast tumor cell lines from pleural effusions J Natl Cancer Inst 1974 53 661 674 4412247
Yoneda T Williams PJ Hiraga T Niewolna M Nishimura R A bone-seeking clone exhibits different biological properties from the MDA-MB-231 parental human breast cancer cells and a brain-seeking clone in vivo and in vitro J Bone Miner Res 2001 16 1486 1495 11499871
Mongiovi AM Romano PR Panni S Mendoza M Wong WT Musacchio A Cesareni G Di Fiore PP A novel peptide-SH3 interaction Embo J 1999 18 5300 5309 10508163 10.1093/emboj/18.19.5300
Cooper JA Schafer DA Control of actin assembly and disassembly at filament ends Curr Opin Cell Biol 2000 12 97 103 10679358 10.1016/S0955-0674(99)00062-9
Wear MA Cooper JA Capping protein: new insights into mechanism and regulation Trends Biochem Sci 2004 29 418 428 15362226 10.1016/j.tibs.2004.06.003
Bompard G Caron E Regulation of WASP/WAVE proteins: making a long story short J Cell Biol 2004 166 957 962 15452139 10.1083/jcb.200403127
Higgs HN Pollard TD Regulation of actin filament network formation through ARP2/3 complex: activation by a diverse array of proteins Annu Rev Biochem 2001 70 649 676 11395419 10.1146/annurev.biochem.70.1.649
Weaver AM Young ME Lee WL Cooper JA Integration of signals to the Arp2/3 complex Curr Opin Cell Biol 2003 15 23 30 12517700 10.1016/S0955-0674(02)00015-7
Bear JE Svitkina TM Krause M Schafer DA Loureiro JJ Strasser GA Maly IV Chaga OY Cooper JA Borisy GG Gertler FB Antagonism between Ena/VASP proteins and actin filament capping regulates fibroblast motility Cell 2002 109 509 521 12086607 10.1016/S0092-8674(02)00731-6
Bear JE Loureiro JJ Libova I Fassler R Wehland J Gertler FB Negative regulation of fibroblast motility by Ena/VASP proteins Cell 2000 101 717 728 10892743 10.1016/S0092-8674(00)80884-3
Sokol SY Analysis of Dishevelled signalling pathways during Xenopus development Curr Biol 1996 6 1456 1467 8939601 10.1016/S0960-9822(96)00750-6
Wallingford JB Harland RM Neural tube closure requires Dishevelled-dependent convergent extension of the midline Development 2002 129 5815 5825 12421719 10.1242/dev.00123
Yang-Snyder J Miller JR Brown JD Lai CJ Moon RT A frizzled homolog functions in a vertebrate Wnt signaling pathway Curr Biol 1996 6 1302 1306 8939578 10.1016/S0960-9822(02)70716-1
Howard JE Smith JC Analysis of gastrulation: different types of gastrulation movement are induced by different mesoderm-inducing factors in Xenopus laevis Mech Dev 1993 43 37 48 8240971 10.1016/0925-4773(93)90021-O
Kwan KM Kirschner MW Xbra functions as a switch between cell migration and convergent extension in the Xenopus gastrula Development 2003 130 1961 1972 12642499 10.1242/dev.00412
Mejillano MR Kojima S Applewhite DA Gertler FB Svitkina TM Borisy GG Lamellipodial versus filopodial mode of the actin nanomachinery: pivotal role of the filament barbed end Cell 2004 118 363 373 15294161 10.1016/j.cell.2004.07.019
Han YH Chung CY Wessels D Stephens S Titus MA Soll DR Firtel RA Requirement of a vasodilator-stimulated phosphoprotein family member for cell adhesion, the formation of filopodia, and chemotaxis in dictyostelium J Biol Chem 2002 277 49877 49887 12388544 10.1074/jbc.M209107200
Lebrand C Dent EW Strasser GA Lanier LM Krause M Svitkina TM Borisy GG Gertler FB Critical role of Ena/VASP proteins for filopodia formation in neurons and in function downstream of netrin-1 Neuron 2004 42 37 49 15066263 10.1016/S0896-6273(04)00108-4
Innocenti M Tenca P Frittoli E Faretta M Tocchetti A Di Fiore PP Scita G Mechanisms through which Sos-1 coordinates the activation of Ras and Rac J Cell Biol 2002 156 125 136 11777939 10.1083/jcb.200108035
Biesova Z Piccoli C Wong WT Isolation and characterization of e3B1, an eps8 binding protein that regulates cell growth Oncogene 1997 14 233 241 9010225 10.1038/sj.onc.1200822
Croce A Cassata G Disanza A Gagliani MC Tacchetti C Malabarba MG Carlier MF Scita G Baumeister R Di Fiore PP A novel actin barbed-end-capping activity in EPS-8 regulates apical morphogenesis in intestinal cells of Caenorhabditis elegans Nat Cell Biol 2004 6 1173 1179 15558032 10.1038/ncb1198
Stradal T Courtney KD Rottner K Hahne P Small JV Pendergast AM The Abl interactor proteins localize to sites of actin polymerization at the tips of lamellipodia and filopodia Curr Biol 2001 11 891 895 11516653 10.1016/S0960-9822(01)00239-1
Schafer DA Welch MD Machesky LM Bridgman PC Meyer SM Cooper JA Visualization and molecular analysis of actin assembly in living cells J Cell Biol 1998 143 1919 1930 9864364 10.1083/jcb.143.7.1919
Cheyette BN Waxman JS Miller JR Takemaru K Sheldahl LC Khlebtsova N Fox EP Earnest T Moon RT Dapper, a Dishevelled-associated antagonist of beta-catenin and JNK signaling, is required for notochord formation Dev Cell 2002 2 449 461 11970895 10.1016/S1534-5807(02)00140-5
Evan GI Lewis GK Ramsay G Bishop JM Isolation of monoclonal antibodies specific for human c-myc proto-oncogene product Mol Cell Biol 1985 5 3610 3616 3915782
Nieuwkoop PD Faber J Normal Table of Xenopus laevis (Daudin) 1967 Amsterdam: North-Holland
|
16225669
|
PMC1274305
|
CC BY
|
2021-01-04 16:31:30
|
no
|
BMC Cell Biol. 2005 Oct 14; 6:36
|
utf-8
|
BMC Cell Biol
| 2,005 |
10.1186/1471-2121-6-36
|
oa_comm
|
==== Front
BMC Chem BiolBMC Chemical Biology1472-6769BioMed Central London 1472-6769-5-31621612210.1186/1472-6769-5-3Research ArticleCovalent attachment of the plant natural product naringenin to small glass and ceramic beads Lu Yuhua [email protected] Niloufer G [email protected] Erich [email protected] Department of Plant Cellular and Molecular Biology and Plant Biotechnology Center, The Ohio State University, Columbus, OH 43210, USA2005 10 10 2005 5 3 3 24 5 2005 10 10 2005 Copyright © 2005 Lu et al; licensee BioMed Central Ltd.2005Lu et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
Natural products have numerous medicinal applications and play important roles in the biology of the organisms that accumulate them. Few methods are currently available for identifying proteins that bind to small molecules, therefore the discovery of cellular targets for natural products with pharmacological activity continues to pose a significant challenge in drug validation. Similarly, the identification of enzymes that participate in the biosynthesis or modification of natural products remains a formidable bottleneck for metabolic engineering. Flavonoids are one large group of natural products with a diverse number of functions in plants and in human health. The coupling of flavonoids to small ceramic and glass beads provides a first step in the development of high-throughput, solid-support base approaches to screen complex libraries to identify proteins that bind natural products.
Results
The utilization of small glass and ceramic beads as solid supports for the coupling of small molecules was explored. Initial characterization of the beads indicated uniform and high capacity loading of amino groups. Once the beads were deemed adequate for the linking of small molecules by the coupling of NHS-fluorescein followed by microscopy, chemical hydrolysis and fluorometry, the flavonoid naringenin was modified with 1,4-dibromobutane, followed by the attachment of aminopropyltriethoxysilane. After NMR structural confirmation, the resulting 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin was attached to the ceramic beads.
Conclusion
Our results demonstrate that ceramic and glass beads provide convenient solid supports for the efficient and facile coupling of small molecules. We succeeded in generating naringenin-coupled ceramic and glass beads. We also developed a convenient series of steps that can be applied for the solid-support coupling of other related flavonoids. The availability of solid-support coupled naringenin opens up new opportunities for the identification of flavonoid-binding proteins.
==== Body
Background
Natural products are small molecules synthesized by bacteria, algae, fungi and plants, as part of their biotic and abiotic interactions with the environment. They provide a valuable source of pharmaceuticals, with 45% of the most popular drugs derived from natural products [1]. However, the identification of molecular targets for natural compounds with pharmacological activity remains a significant bottleneck in the process of drug validation. Plants provide a formidable source of natural products (phytochemicals) with pharmacological activity [2]. Over 100,000 phytochemicals have already been identified from the small fraction of the plant kingdom that has so far been surveyed [3]. These phytochemicals can be classified into large groups that include the alkaloids, the terpenoids, and the phenylpropanoids.
Flavonoids, derived from phenylpropanoids, are widely distributed throughout the plant kingdom and are abundantly present in many fruits and leaves [4]. They are characterized by the presence of two benzene rings linked by a pyrane or pyrone ring (ring C). Based on the position and modifications of the A, B and C rings, the 4,000+ flavonoids so far known can be classified into several sub-classes including the flavonols, the flavones, the isoflavones and the red/purple anthocyanin pigments. Flavonoids are potent antioxidants [5] and have important activities as dietary anti-carcinogens and anti-inflammatory compounds [6-10]. Flavonoids are also significant to plants, serving as signal molecules in various developmental processes [11]. Several studies have investigated the effect of flavonoids on the activity of various enzymes. For example, flavonoids possess protein kinase [12-15] and P-glycoprotein [16] inhibitory activities. However, with a few exceptions, in vivo animal and plant cellular targets for flavonoids are largely unknown.
Few methods are currently available for identifying proteins that bind to small molecules of plant origin, for example flavonoids. The development of high-throughput methods for the identification of flavonoid-binding proteins could significantly advance our understanding of the mechanisms by which flavonoids modulate plant hormone transport [17], contribute to plant male fertility [18], serve as allelochemicals [19] and facilitate the identification of missing metabolic enzymes in the flavonoid biosynthetic pathway. One promising approach is the generation of flavonoid derivatives that contain a benzophenone chromophore for use as a photoaffinity reagent [20]. While benzophenone-modified flavonoids can potentially permit the identification of flavonoid-binding proteins, they are inadequate for the isolation of significant quantities of proteins, necessary for their mass spectrometry identification. The isolation of flavonoid-binding proteins would be simplified by the availability of flavonoids covalently-linked to solid supports that would permit the affinity-purification of flavonoid-binding proteins from complex protein mixtures or protein libraries. Few methods are currently available that explore the possibility of linking select flavonoids to solid supports.
Here, we describe the efficient coupling of the flavanone naringenin, a central intermediate in the flavonoid biosynthetic pathway [21], to small ceramic and glass beads. The method developed can be easily adapted for the coupling of other related flavonoids to beads. The naringenin-coupled beads provide a powerful chemical genetic tool to probe biological systems.
Results and discussion
Ceramic and glass beads provide high-capacity solid supports
Many beaded solid supports are available, primarily developed towards combinatorial chemistry efforts [22]. The possibility to use small ceramic and hollow glass beads for the conjugation of small molecules was investigated here, because of their low cost, reduced size and compatibility with most organic and inorganic solvents. Scanning electron microscopy (SEM) experiments (Fig. 1) showed that the ceramic beads have an average particle volume of 5 μm3 and a diameter of approximately 2 μm, about 5 times smaller than the glass beads (Table 1). These dimensions make them much smaller than the usual solid supports utilized in solid phase synthesis. The SEM analyses (Fig. 1A, B) also confirmed that both the ceramic and glass beads are not porous. This is an important property to consider given the ultimate goal to use the beads to identify proteins that bind to the small molecule, which requires a minimal hindrance of the small molecule by the bead matrix.
Figure 1 Scanning electron micrograph of ceramic and glass beads. Scanning electron micrograph of a representative (A) ceramic bead and (B) glass bead.
Table 1 Comparison of amino loading on the ceramic beads and glass beads
Ceramic beads Glass beads
Fluorescein loading (μmol/g) 0.34 0.03
Diameter of bead (μm) 2 10
Density of bead (g/cc) 2.5 1.1
Numbers of -NH2 group loading per bead 4 × 106 1 × 107
Surface area on bead for each – NH2 group (Å2) 250 3000
The presence of Si-OH groups on the surface of the beads allows the rapid conjugation of bifunctional organosilanes [23,24]. To investigate the capacity of the loading of functional groups onto the surface of the ceramic and glass beads, silanisation of the beads was performed using 3-aminopropyltriethoxysilane (Fig. 2A). After washing and drying completely, the amino-modified beads were stored at 4°C until use.
Figure 2 Modification and fluorescein-loading of ceramic beads. (A) Reaction scheme for the loading of amino groups on the ceramic and glass beads and the subsequent reaction with NHS-fluorescein. (B) Fluorescence micrographs of ceramic beads and glass beads.
In order to determine the effective loading of amino groups on the beads, both naïve (unmodified) and amino-modified glass and ceramic beads were incubated with NHS-fluorescein in the dark at 4°C for 24 hrs. After completion of the reaction, the beads were washed with water and dried. Naïve glass or ceramic beads displayed no fluorescence, nor did they fluoresce after incubation with NHS-fluorescein, while amino-modified glass and ceramic beads fluoresced after reaction with fluorescein (Fig. 2B). To quantitatively analyze the loading of amino groups on the surface of the beads, fluorescein was cleaved off by treatment with 1N hydrochloric acid. After the solution was neutralized to pH 9, the released fluorescence was quantified by fluorometry. For the ceramic beads, the loading of fluorescein, which represents the number of amino groups on the beads available for reaction, was established to be 0.34 μmol/g, which is the minimum possible loading of amino groups on the beads. Since the density of ceramic beads is 2.5 g/cc and the average diameter of bead is 2 μm, it was estimated that there is one amino group every approximately 250 Å2 of surface area, with a loading capacity of around 4 × 106 amino groups per ceramic bead (Table 1). Similar calculations estimated the capacity of the glass beads at 1 × 107 amino-groups per bead. To further investigate the distribution of the amino loading on the ceramic and glass beads, FACS flow cytometry was utilized. After the reaction with NHS-fluorescein, over 90% of the ceramic and glass beads displayed fluorescence (data not shown). Thus, not only were the beads efficiently loaded with amino groups, but this loading was also uniformly distributed among the beads. Together, these results indicate that both the ceramic and the glass beads met the desired requirements for serving as solid supports for the coupling of flavonoids, and the amino-modified beads have further potential for solid phase synthesis.
Modification and coupling of naringenin to the solid support
Initially, it was envisioned that the amino-loaded ceramic beads could be used for the direct coupling of naringenin, in the presence of a convenient linker such as dimethyl suberimidate (DMS). This approach, however, resulted to be impractical because technical limitations prevented the quantification of the amount of naringenin coupled to the beads, and it was feared that non-reacted amino groups on the beads or imidoester groups on the DMS linker would react with components of the biological system (e.g., proteins) that were intended to be probed.
Therefore, a different strategy was utilized, where the linker was first linked to the organosilane group on one end and naringenin on the other (Fig. 3), followed by the coupling of the entire moiety to the beads. One equivalent of 1,4-dibromobutane was first reacted with naringenin (Fig. 3), resulting in the formation of 7-(4-bromo-n-butoxy)naringenin (1) with little amount of 7,4'-di(4-bromo-n-butoxy) naringenin as a by-product [25]. After separation by chromatography on a silica column and structural elucidation by NMR (see Material and methods), the 7-(4-bromo-n-butoxy)naringenin was reacted with aminopropyltriethoxysilane, to give compound (2). The compound was purified by chromatography on silica gel and confirmed to be 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin by NMR. Once the identity of compound (2)was confirmed, its coupling to sodium hydroxide-treated ceramic beads in 95% methanol was performed overnight. Separation of the beads from non-conjugated compound (2)was accomplished by filtration and washing.
Figure 3 Modification and loading of naringenin to ceramic and glass beads. Reaction scheme for the loading of naringenin onto the ceramic and glass beads. Compound (1) corresponds to 7-(4-bromo-n-butoxy) naringenin and compound (2) corresponds to 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin.
To ensure that compound (2)(Fig. 3) was coupled to the beads, advantage was taken from the reactivity of NHS-fluorescein with the -NH- group present in 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (Fig. 3). The fluorescence detected is indicative of the coupling of the 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin to the ceramic and glass beads (Fig. 4).
Figure 4 Fluorescence micrographs of compound (2)-loaded beads. Compound (2)-loaded (A) ceramic beads and (B) glass beads treated with NHS-fluorescein.
To further quantify the amount of 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin coupled to the beads, the molar extinction coefficient (ε) of compound (2)was first calculated at 314 nm, corresponding to one of the absorption peaks of this compound (Fig. 5). The ε value of 17,244 cm-1M-1 was used to calculate the amount of compound (2)that remained in solution after a coupling reaction was carried with a slight molar excess of 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (assuming a bead capacity of 0.34 μmol/g). The loading of compound (2)was established to be 0.465 μmol per gram of beads, corresponding to approximately 5.5 × 106 molecules of 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin per ceramic bead. Using a similar approach, the coupling of 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin to the glass beads was established to be 0.00016 μmol/g, corresponding to 5.3 × 104 molecules per glass bead.
Figure 5 Absorption properties of compound (2) (A) Absorption spectra of naringenin and compound (2). (B) Calculation of the molar extinction coefficient (ε) for compound (2) at 314 nm.
Conclusion
The development of a simple and reliable method to link flavonoids to solid supports, in this particular case, small glass and ceramic beads, is described here. Interestingly, ceramic beads provided a much more robust solid support for coupling naringenin than the glass beads. These naringenin-coupled ceramic beads provide a convenient tool for the identification and isolation of naringenin-binding proteins. Although these studies describe only the coupling of naringenin to the beads through the 7-OH group in ring A, the possibility to selectively protect the various -OH groups offers opportunities for presenting the different groups in naringenin for recognition by proteins.
Methods
Microscopy
Scanning electron micrographs were obtained on a Philips XL Series Scanning Electron Microscope (SEM). Fluorescence micrographs were obtained on a Nikon Eclipse E600 (Nikon, Japan) microscope equipped with a OSRAM HBO Mercury Short Arc Lamp (λex 485 nm, λem 515 nm, Mercury 100 W, CHIU Technical Corporation). Fluorescence was measured in a FLEX Station™ Instrument (λex 491 or 496 nm, λem 518 or 519 nm, Molecular Devices) with the SOFTmax@PRO program.
1H, 13C NMR and UV spectra
1H and 13C NMR spectra were recorded on a Bruker NMR spectrometer (DRX-250 and DRX-500) in acetone-d6 (Aldrich). UV absorption spectra were performed on a Cary 50 Bio UV-Visible spectrophotometer. Data was analyzed by the Varian Cary WinUV Scan application.
Chromatography methods
Column chromatography was performed on silica gel (Aldrich, 200–400 mesh) using the indicated solvents. TLC analyses were carried out using silica gel plates (Aldrich).
Amino-modification of ceramic beads
Five grams of ceramic beads (3M™ Zeeospheres™ Ceramic Microspheres) or glass beads (Aldrich) were shaken in 20 mL of 10% sodium hydroxide solution overnight, and subsequently washed with water (40 mL), 1% hydrochloric acid (40 mL), water (40 mL) and methanol (40 mL), three times each. Then, the ceramic beads were shaken in 30 mL of 3% aminopropyltriethoxysilane solution made in 95% methanol at 4°C overnight, followed by washing with methanol (40 mL) and water (40 mL), three times each. After air-drying, the amino-modified ceramic beads were stored at 4°C.
Fluorescein-coupled ceramic beads
Amino-modified ceramic beads (0.3 g) were added to 1 mL of 50 mM sodium bicarbonate buffer (pH ~8.5). In a dark room, 0.2 mg of NHS-fluorescein (Pierce Biotechnology, Inc.) were dissolved in 200 μL of DMSO, and then the NHS-fluorescein solution was added to the sodium bicarbonate suspension of beads and mixed well. The suspension was protected from light and shaken at 4°C for 24 hrs. After the liquid was removed by centrifugation and washed extensively with water, the fluorescein-bound ceramic beads were dried and stored at -20°C for future use. To compare with the control, bare (not amino group loaded) ceramic beads were treated under the same conditions. To test the fluorescence of bare and modified beads, all beads were separately shaken in a mixture of 200 μL of DMSO and 1 mL of 50 mM sodium bicarbonate buffer and dried.
Hydrolysis of fluorescein from ceramic beads
Fifty mg of fluorescein-loaded ceramic beads were shaken in 1 mL of 1 N hydrochloric acid solution at 4°C for 4 days in the dark. After filtration through a 0.45 μm cellulose filter and washed with water, the filtrate was neutralized with 50 mM sodium bicarbonate buffer. To compare with the control, all bare beads incubated with or without fluorescein were treated under the same conditions. Fluorescence was measured on 200 μL of serial cleavage solution for each kind of beads in a 96 well cell microplate from both the top and bottom faces. A pure NHS-fluorescein solution was prepared as a calibration standard to calculate the absolute amount of fluorescein released form the beads.
Preparation of 7-(4-bromo-n-butoxy)naringenin (1)
Naringenin (136 mg, 0.5 mmol, 1.0 eq) was dissolved in acetone (5 mL), then potassium carbonate (69 mg, 0.5 mmol, 1.0 eq) was added and heated to reflux for 30 minutes. To this suspension, 1,4-dibromobutane (107 mg, 0.5 mmol, 1.0 eq) was added and stirred at reflux for 8 hours. After cooling to room temperature, the mixture was concentrated on a rotary evaporator to become a thick slurry. The slurry was re-dissolved in ethyl acetate (30 mL), washed with distilled water (three times with 10 mL) and brine (10 mL), then evaporated. Purification was performed by silica-gel column chromatography (3:1 hexanes/ethyl acetate) to give pure 7-(4-bromo-n-butoxy)naringenin as a white solid (132 mg, 65% yield): 1H NMR (acetone-d6, 250 MHz) δ 1.86–1.97 (m, 2H), 2.00–2.09 (m, 2H), 2.74 (dd, J = 17.0, 3.0 Hz, 1H), 3.19 (dd, J = 17.0, 12.8 Hz, 1H), 3.57 (t, J = 6.5 Hz, 2H), 4.11 (t, J = 6.3 Hz, 2H), 5.46 (d, J = 12.5 Hz, 1H), 6.03 (s, 1H), 6.04 (s, 1H), 6.89 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 8.48 (s, 1H), 12.11 (s, 1H); 13C NMR (acetone-d6, 62.9 MHz) δ 28.34, 30.18, 34.27, 43.46, 68.38, 79.98, 94.94, 95.88, 103.72, 116.17, 128.99, 130.67, 158.68, 164.15, 164.95, 168.09, 197.53.
Preparation of 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (2)
7-(4-bromo-n-butoxy)naingenin (50 mg, 0.12 mmol, 1.0 eq) and pyridine (28 mg, 0.36 mmol, 3.0 eq) were dissolved in dry DCM (1 mL), 3-aminopropyltriethoxysilane (32 mg, 0.14 mmol, 1.2 eq) was then added and shaken for 48 hrs. After removal of DCM, the residue was purified by silica-gel column chromatography (2:1 hexanes/ethyl acetate) to give pure 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin as a pale yellow solid (15 mg, 22% yield): 1H NMR (acetone-d6, 500 MHz) δ 0.72 (t, J = 8.2 Hz, 2H), 1.16 (t, J = 6.7 Hz, 9H), 1.42 (m, 2H), 1.79 (m, 2H), 1.89 (m, 2H), 2.9 (dd, J = 16.8, 12.2 Hz, 1H), 3.20 (dd, J = 16.9, 3.0 Hz, 1H), 3.48 (m, 2H), 3.54 (m, 2H), 3.80 (q, J = 7.2 Hz, 6H), 4.00 (t, J = 6.3 Hz, 2H), 5.12 (dd, J = 12.2, 2.8 Hz, 1H), 5.77 (d, J = 2.3 Hz, 1H), 5.88 (d, J = 2.3 Hz, 1H), 6.88 (d, J = 8.6 Hz, 2H), 7.36 (d, J = 8.5 Hz, 2H), 8.58 (s, 1H). This is an easy, but not efficient procedure. We successfully completed this reaction independently twice, however, we failed in other opportunities to obtain the desired product for unexplained reasons.
Coupling of naringenin to ceramic beads
Ceramic beads (30 mg) were first treated with 10% sodium hydroxide solution and washed with water, hydrochloric acid and methanol as above, then shaken with 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (20 mg) in 95% methanol (1 mL) for 24 hrs. After filtration, the naringenin-coupled ceramic beads were washed three times each with methanol, water and methanol, then vacuum dried. The obtained beads were treated with NHS-fluorescein and detected by fluorescence microscope, as described above.
Determination of naringenin loading on ceramic beads
7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (0.164 mg, 0.3 μmol) was shaken with ceramic beads (500 mg) in 3 mL of 95% methanol for 24 hrs. After filtration, 1 mL of the filtrate was measured by absorbance at 314 nm. 7-(4-(3-(triethoxysilyl)propylamino)butoxy) naringenin (compound 2) was dissolved in 95% methanol (concentration from 128 μM to 12.8 μM) and measured as a calibration. Based on Beer's law, the molar extinction coefficient (ε) was determined to be 17244 cm-1M-1.
Authors' contributions
YL carried out all the chemical experiments described, NGI helped with the fluorescence microscopy and fluorometry. EG conceived the project and participated in the design and coordination of the study.
Table 2 Comparison of naringenin loading on the ceramic and the glass beads
Ceramic beads Glass beads
Naringenin Loading (μmol/g) 0.465 0.00016
Numbers of naringenin loading on per bead 5.5 × 106 5.3 × 104
Surface area on bead for each naringenin molecule (Å2) 180 5.7 × 105
Acknowledgements
We thank Juan Grotewold for the suggestion to use the 3 M ceramic beads and Ben Bowen for many helpful discussions over the years. This research was supported by a grant from the National Science Foundation (MCB-0437318) to EG.
==== Refs
Muller-Kuhrt L Putting nature back into drug discovery Nat Biotechnol 2003 21 602 12776140 10.1038/nbt0603-602
Rates SM Plants as source of drugs Toxicon 2001 39 603 613 11072038 10.1016/S0041-0101(00)00154-9
Verpoorte R Verpoorte R and Alfermann AW Plant secondary metabolism Metabolic engineering of plant secondary metabolism 2000 Dordrecht, Kluwer Academic Publishers 1 29
Miean KH Mohamed S Flavonoid (myricetin, quercetin, kaempferol, luteolin, and apigenin) content of edible tropical plants J Agric Food Chem 2001 49 3106 3112 11410016
Rice-Evans CA Miller NJ Paganga G Antioxidant properties of phenolic compounds Trends Plant Sci 1997 2 152 159 10.1016/S1360-1385(97)01018-2
Fiala ES Reddy BS Weisburger JH Naturally occurring anticarcinogenic substances in foodstuffs. Ann Rev Nutrit 1985 5 295 321 2992547 10.1146/annurev.nu.05.070185.001455
Di Carlo G Mascolo N Izzo AA Capasso F Flavonoids: old and new aspects of a class of natural therapeutic drugs. Life Sci 1999 65 337 353 10421421 10.1016/S0024-3205(99)00120-4
Middleton E Kimball ES The flavonoids as potential therapeutic agents Immunopharmaceuticals 1996 , CRC Press 227 257
Middleton EJ Kandaswami C Theoharides TC The effects of plant flavonoids on mammalian cells: implications for inflammation, heart disease, and cancer Pharmacol Rev 2000 52 673 751 11121513
Middleton EJ Kandaswami C Effects of flavonoids on immune and inflammatory cell functions Biochem Pharmacol 1992 43 1167 1179 1562270 10.1016/0006-2952(92)90489-6
Taylor LP Grotewold E Flavonoids as developmental regulators Curr Op Plant Biol 2005 8 317 323 10.1016/j.pbi.2005.03.005
Ferriola PC Cody V Middleton EJ Protein kinase C inhibition by plant flavonoids. Kinetic mechanisms and structure-activitiy relationships. Biochem Pharmacol 1989 38 1617 1624 2730676 10.1016/0006-2952(89)90309-2
Geahlen RL Koonchanok NM McLaughlin JL Pratt DE Inhibition of protein-tyrosine kinase activity by flavanoids and related compounds J Nat Prod 1989 52 982 986 2607357
Cushman M Nagarathnam D Burg DL Geahlen RL Synthesis and protein-tyrosine kinase inhibitory activities of flavonoid analogues J Med Chem 1991 34 798 806 1995903 10.1021/jm00106a047
Gamet-Payrastre L Manenti S Gratacap MP Tulliez J Chap H Payraste B Flavonoids and the inhibition of PKC and PI 3-kinase. Gen Pharmac 1999 32 279 286 10.1016/S0306-3623(98)00220-1
Di Pietro A Conseil G Perez-Victoria JM Dayan G Baubichon-Cortay H Trompier D Steinfels E Jault JM de Wet H Maitrejean M Comte G Boumendjel A Mariotte AM Dumontet C McIntosh DB Goffeau A Castanys S Gamarro F Barron D Modulation by flavonoids of cell multidrug resistance mediated by P-glycoprotein and related ABC transporters Cell Mol Life Sci 2002 59 307 322 11915946 10.1007/s00018-002-8424-8
Murphy A Peer WA Taiz L Regulation of auxin transport by aminopeptidases and endogenous flavonoids. Planta 2000 211 315 324 10987549 10.1007/s004250000300
Taylor LP Hepler PK Pollen germination and tube growth Annu Rev Plant Physiol Plant Mol Biol 1997 48 461 491 15012271 10.1146/annurev.arplant.48.1.461
Bais HP Vepachedu R Gilroy S Callaway RM Vivanco JM Allelopathy and exotic plant invasion: from molecules and genes to species interactions Science 2003 301 1377 1380 12958360 10.1126/science.1083245
Tanaka H Stohlmeyer MM Wandless TJ Taylor LP Synthesis of flavonol derivatives as probes of biological processes Tetrahedron Letters 2000 41 9735 9739 10.1016/S0040-4039(00)01767-6
Winkel-Shirley B Flavonoid biosynthesis. A colorful model for genetics, biochemistry, cell biology and biotechnology Plant Physiol 2001 126 485 493 11402179 10.1104/pp.126.2.485
Yu Z Bradley M Solid supports for combinatorial chemistry Curr Opin Chem Biol 2002 6 347 352 12023116 10.1016/S1367-5931(02)00327-7
Chen JC von Lintig FC Jones SB Huvar I Boss GR High-efficiency solid-phase capture using glass beads bonded to microcentrifuge tubes: immunoprecipitation of proteins from cell extracts and assessment of ras activation Anal Biochem 2002 302 298 304 11878811 10.1006/abio.2001.5572
Beier M Hoheisel JD Versatile derivatisation of solid support media for covalent bonding on DNA-microchips Nucleic Acids Res 1999 27 1970 1977 10198429 10.1093/nar/27.9.1970
Borges MN Messeder JC Figueroa-Villar JD Synthesis, anti-Trypanosoma cruzi activity and micelle interaction studies of bisguanylhydrazones analogous to pentamidine Eur J Med Chem 2004 39 925 929 15501541 10.1016/j.ejmech.2004.07.001
|
16216122
|
PMC1274306
|
CC BY
|
2021-01-04 16:30:50
|
no
|
BMC Chem Biol. 2005 Oct 10; 5:3
|
utf-8
|
BMC Chem Biol
| 2,005 |
10.1186/1472-6769-5-3
|
oa_comm
|
==== Front
BMC Dev BiolBMC Developmental Biology1471-213XBioMed Central London 1471-213X-5-231623617310.1186/1471-213X-5-23Research ArticleNovel splice variants associated with one of the zebrafish dnmt3 genes Smith Tamara HL [email protected] Christine C [email protected] Aizeddin A [email protected] Ross A [email protected] Dept. of Biochemistry, Memorial University, St. John's, NL, Canada2 Faculty of Medicine, University of Manitoba, Winnipeg, MB., Canada3 Department of Pediatrics & Child Health and Biochemistry and Medical Genetics, University of Manitoba, Winnipeg, MB., Canada2005 19 10 2005 5 23 23 8 6 2005 19 10 2005 Copyright © 2005 Smith et al; licensee BioMed Central Ltd.2005Smith et al; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background
DNA methylation and the methyltransferases are known to be important in vertebrate development and this may be particularly true for the Dnmt3 family of enzymes because they are thought to be the de novo methyltransferases. Mammals have three Dnmt3 genes; Dnmt3a, Dnmt3b, and Dnmt3L, two of which encode active enzymes and one of which produces an inactive but necessary cofactor. However, due to multiple promoter use and alternative splicing there are actually a number of dnmt3 isoforms present. Six different dnmt3 genes have recently been identified in zebrafish.
Results
We have examined two of the dnmt3 genes in zebrafish that are located in close proximity in the same linkage group and we find that the two genes are more similar to each other than they are to the other zebrafish dnmt3 genes. We have found evidence for the existence of several different splice variants and alternative splice sites associated with one of the two genes and have examined the relative expression of these genes/variants in a number of zebrafish developmental stages and tissues.
Conclusion
The similarity of the dnmt3-1 and dnmt3-2 genes suggests that they arose due to a relatively recent gene duplication event. The presence of alternative splice and start sites, reminiscent of what is seen with the human DNMT3s, demonstrates strong parallels between the control/function of these genes across vertebrate species. The dynamic expression levels of these genes/variants suggest that they may well play a role in early development and this is particularly true for dnmt3-2-1 and dnmt3-1. dnmt3-2-1 is the predominantly expressed form prior to zygotic gene activation whereas dnmt3-1 predominates post zygotic gene activation suggesting a distinct developmental role for each.
==== Body
Background
The epigenetic modification of DNA by the addition of a methyl group to the 5 position of cytosine is an important mechanism for control of gene expression in vertebrates. This is particularly true during development where DNA methylation is thought to have a role in genome imprinting [1,2], X inactivation [3] and lineage determination [4]. Methylation has been most intensely studied in mammals where the levels have been shown to be quite dynamic during early development, decreasing soon after fertilization and increasing again by the gastrula stage [5,6]. The importance of this de-methylation/re-methylation cycle to the developmental process has been clearly demonstrated by perturbations of that methylation that generally leads to embryonic lethality [7,8]. Given the importance of methylation in sustaining normal early developmental processes, the enzymes that add and maintain that methylation are of significant interest. The dnmt3 family of methyltransferases that are thought to be important in de novo methylation (that is the addition of methyl groups to previously unmethylated sequences) are of particular interest in this context. There are three members of this family in mammals; two have catalytic activity, Dnmt3a and Dnmt3b; and the third, Dnmt3L, is important as a cofactor, particularly for the methylation of imprinted loci [9]. Functionally, however, the dnmt3 family is not limited to just three products because both the Dmnt3a and b transcripts can be alternatively spliced to generate a number of different RNAs. Dnmt3a has two splice variants differing in the 5' region whereas dnmt3b has a number of possible splicing products [10]. These variations in the dnmt3 proteins may allow for a greater diversity in the function and/or targets of these enzymes.
Methylation in zebrafish has recently been the focus of a number of reports, and methylation has been found to be dynamic during its early development [11]. Also, as in mammals, blocking re-methylation in zebrafish results in abnormal development and death [12].
The zebrafish actually has at least twice the mammalian number of dnmt3 genes; six have been submitted to databases so far (GenBank numbers AB196914, AB196915, AB196916, AB196917, AB196918, AB196919) [19]. The significance of the increase in dnmt3 gene copy number in zebrafish is unknown.
We have isolated and analysed a number of the zebrafish dnmt3 gene sequences and have identified two dnmt3 sequences that are located very close together in a single linkage group. The very close proximity of the two sequences provides an interesting opportunity to examine how the expression of these genes is controlled since one copy has a very limited upstream promoter region relative to the other.
Results and Discussion
We used a dnmt3 sequence already present in the zebrafish EST database (GenBank number AF135438) to identify and isolate the complete cDNA sequences of four of the dnmt3 genes found in the zebrafish. Three of these are located in the same linkage group (linkage group 23) and two of them very closely juxtaposed to each other (Figure 1). The very close proximity of those two genes has some interesting implications with respect to their origin and the control of their expression, given the much more limited potential promoter region of one relative to the other. We, therefore, undertook a closer examination of the two genes, which we named dnmt3-1 and dnmt3-2.
Figure 1 The dnmt3 genes in zebrafish. A) The Welcome Trust Sanger Institute library numbers are provided for all genes while the genomic size and distances are indicated for genes 1, 2, and 3. B) A more detailed view of the genomic structures of Gene 1 and Gene 2. Boxes represent exons and adjoining lines represent introns.
From the end of the polyA addition site of gene 1 to the beginning of our cloned sequence for gene 2 (probably not actually beginning at the cellular transcriptional start site) consists of only 1428 base pairs. Since there is only a small amount of 5' sequence that is associated with the dnmt3-2 gene this limits the control of the expression of this gene to a small and easily manipulated region. Analysis of this region suggests that it is a TATA-less promoter with a number of potential transcription factor binding sites including AP1 and SP1 binding sites which have also been reported for mammalian Dnmt3s [13,14]. The sequence of the cloned genes, dnmt3-1 and dnmt3-2 revealed open reading frames that could encode polypeptides of 1447 and 1297 amino acids, respectively. Comparison of the sequences of these two genes to zebrafish genomic maps present in the Genbank database allows for an analysis of the genomic structure. That structure along with the relative position of the two genes is shown in figure 1. The two genes are very similar in sequence; 72% at the nucleotide level and 74% identical at the amino acid level, with large regions being more than 80% identical (figure 2). This is in contrast to only 19–28% similarity at the nucleotide level, and 36–46% amino acid similarity when compared to the other dnmt3 sequences present in the zebrafish genome. This trend is also true for the conserved methyltransferase motifs. For instance, the PWWP motif of gene 1 and gene 2 are 88% and 84% similar at the amino acid and nucleotide levels, respectively, but considerably less similar to the other dnmt3 sequences (e.g. dnmt3-2 vs gene 4, accession #196918, has 64% and no significant similarity at the amino acid and nucleotide levels, respectively) (BLAST, NCBI)).
Figure 2 Homology between Gene 1 and Gene 2 at the nucleotide level. Percentages above 80% are shaded, and specific homology percentages are provided below the figure (numbers obtained by BLAST alignment, NCBI). Arrows span the nucleotides that give rise to the identified motifs labelled above the respective arrows. Note: All identified motifs are characteristic of the C-5 methyltransferases, with the exception of the CH domain.
Recent additions to the sequence databases included two zebrafish sequences that appear to correspond to the same two genes and were named dnmt3 and dnmt5 respectively (GenBank numbers AB196914, AB196916). Our sequencing data corroborate the sequences submitted to the databases except for a few minor variations in regions with triplet repeats which may be an artefact of polymerase slippage in cloning or represent real triplet repeat differences that exist in the gene.
The high homology between dnmt3-1 and dnmt3-2 relative to other zebrafish dnmt3's, as well as their close proximity, suggests that these genes represent a duplication event. Postlethwait et al. [15] provides support for a model where two polyploidization events occurred in a common ancestor of zebrafish and mammals. However, there are often additional multigene members in zebrafish. Postlethwait et al. [15] argues that either chromosome duplication or another tetraploidization event in the zebrafish lineage is the most likely mechanisms by which these additional members arose. The tight clustering seen here, however, suggests that, at least in this instance, tandem gene duplication has occurred.
The most interesting aspect of our analyses is that at least one of the genes, dnmt3-2, includes at least two start sites and a number of splice variants. These were initially identified in cDNA libraries generated from 1–2 cell embryos and RACE-PCR and were later confirmed by RT-PCR in a number of early embryonic zebrafish stages as well as somatic tissues (figure 3). This demonstrates that they are all expressed at least to the level of RNA. Densitometric analysis revealed that the transcript levels are not equivalent and that the relative levels of the different genes and isoforms fluctuate independently between the stages examined (Figure 4). All genes and variants examined are expressed in early embryonic stages, though dnmt-3-2-1 appears to be the most significant prior to zygotic gene activation (zygotic gene activation occurs at ~3 hours). All transcripts demonstrated declining levels leading up to this event, suggesting maternal supply turnover. Following zygotic gene activation however, there appears to be a marked shift towards dnmt3-1 being the most highly expressed. Additionally, there appears to be tissue dependent differences in expression levels (Figure 5). These differences in expression profiles for the different gene products and isoforms suggest that they are regulated independently and each may be playing distinct and separate roles during the development of the zebrafish.
Figure 3 dnmt3 isoforms. A) RT-PCR followed by agarose gel electrophoresis and hybridization with a biotin-labelled probe. Stages/tissues used are labelled on the left side of figure. The first lane in all cases contains a doublet representing the two splice variants of dnmt3-2 differing in size by 78 bp. The second lane shows the alternate translational start site variant of dnmt3-2. The third lane is the product of the primers specific for gene 1 and the last lane is a control reaction loaded on each gel to allow comparisons between gels. The amount of reaction loaded was not the same in all lanes but was varied to produce more equivalent band intensities for more accurate quantification. Controls lacking reverse transcriptions produced no amplification products (not shown). B) RT-PCR of a constitutively expressed gene, max, for each RNA used serving as an internal standard for quantification. Lanes 1–6 show the max amplicons generated from the samples used in panel (A), ovaries through to brain. Lane 7 contains size markers.
Figure 4 Expression Summary. A) Graph showing data from figure 3 developmental stages corrected for differences in amounts loaded, and normalized to max to correct for concentration differences as well as the control for exposure differences (see methods). B) Graph showing data from figure 3 somatic tissues corrected as above and demonstrating the relative expression levels of the three transcripts in those tissues.
Figure 5 The various transcripts produced from Gene 2. The genomic structure is presented on the top line, with the 5' region of interest magnified below illustrating the various alternative splicing products. Transcript 1 (line 2) differs from the others by an alternate transcriptional start site and a missing exon 2. Transcripts two (line 3) and three (line 4) lack the first exon and are alternatively spliced in the second exon (the one missing from transcript 1). All splice variants occur upstream of the AUG start site.
The shortest of these variants, dnmt3-2-1, corresponds to the dnmt5 sequence in the database. The two novel variants reported here differ in size from that sequence by 187 (dnmt3-2-2) and 265 (dnmt3-2-2b) base pairs. These variants are actually associated with the gene having the most restricted promoter region. A schematic of the three products is shown in figure 5.
There are several interesting aspects of these dnmt3-2 variants. To begin with, although the splicing difference between variant dnmt3-2-2 and dnmt3-2-2b appears to involve the same 3' splice junction it has a different 5' splice junction, meaning that one of those splice sites is located within the exon of the other variant. However, both of the junctions still abide by the GT/AG rule for splice junctions.
The second interesting aspect of these splice variants is that all of them are 5' to the initiator AUG. Therefore, none of them actually affect the amino acid sequence. This suggests that either the splicing differences are trivial or they play a regulatory role in the translation or localization or some other aspect of the various splice variants. The latter possibility is a more reasonable assumption since, parsimoniously, it seems unreasonable to assume that this RNA would be alternatively spliced in a variety of ways for no biologically relevant reason. This situation is not unique to zebrafish dnmt3 genes. Similar splice variants in the 5'untranslated region have also been reported for human DNMT3s [13].
Conclusion
We have isolated and analysed several of the dnmt3 genes from the zebrafish. In this report we have focused on two of the genes that are located in close proximity in a single linkage group and we find that the two genes are considerably more similar to each other than they are to the other zebrafish dnmt3 genes. This suggests that they arose as a result of a relatively recent gene duplication event. We have also found evidence for the existence of several different splice variants and alternative splice sites associated with one of the two genes, reminiscent of what is seen with the human DNMT3s. Expression analyses of these genes and variants demonstrate that are dynamic during development with distinct patterns that suggest they are independently controlled and, possibly, have different functions in development.
Methods
Total RNA was isolated from ovarian tissue using the phenol/chloroform method of Chomczynski and Sacchi [16]. Ovarian Poly A+ RNA (FastTrack 2.0 kit, Invitrogen Inc. Carlsbad, CA) was used for first strand cDNA synthesis using the BD SMART™ RACE cDNA Amplification Kit (Clontech, Palo Alto, CA). Using the zebrafish dnmt3 EST (GenBank number AF135438), a gene specific primer, GSP1 (see Table 1 for primer sequences), was designed to amplify, along with the universal primer, the relevant cDNA using PCR conditions as described by the manufacturer. Based on a resulting sequence that corresponded to four different regions of the genomic map (The Sanger Institute Welcome Trust zebrafish sequencing project), a series of gene specific primers were designed for the various genes.
Table 1 Primers used
Primer name Sequence
GSP1 5'- GACAGGACCCTGAATGGACGTCGCT
GSP2 5'- GAGAGAGCACTGAGATGTCAG
GSP3 5'- CCAGAAATCTGTTGGAGACATTACACC
GSP4 5'-AAGGCAGTATGGAGTCTGTCTGCA
GSP5 5'-CAGTCATGGCAATGTCTTTCC
GSP6 5'-ATGTATGTCCTGTGAGGAGGAAC
PCR products were fractionated on 0.8% agarose gels, visualized with ethidium bromide, excised from the gel and cloned into pCR 2.1 vectors (TOPO TA cloning kit, Invitrogen Inc. Carlsbad, CA). The cloned products were then purified (Wizard®Plus Minipreps, Promega Inc. Madison, WI), and sequenced (Cortec DNA Service Laboratories Inc., Kingston, ON).
Reverse Transcription-Polymerase Chain Reaction (RT-PCR) was used to determine the relative expression levels of gene 1, gene 2, and its variants in tissues. Total RNA from zebrafish ovarian tissue, 1–2 cell embryos, 64 cell embryos, 6 hour embryos, muscle tissue, and brain tissue was isolated as described above, and the integrity checked by ethidium bromide staining. The RNA was then reverse transcribed using M-MLV Reverse Transcriptase (Invitrogen Inc. Carlsbad, CA) using primers specific for the various genes and variants. GSP2 was used for first strand cDNA synthesis of Gene 1 in conjunction with GSP3, generating a predicted amplicon of 521 bp. Primer GSP4 designed to anneal to all three gene 2 variants was used with GSP5 for dnmt3-2-1 to produce a 420 bp amplicon and with GSP6 to produce two amplicons of 597 bp and 675 bp from gene 2 variants dnmt3-2-2, and dnmt3-2b. In addition, RT-PCR was conducted to generate a 440 bp amplicon with GSP7 and GSP8, primers specific for a constitutively expressed gene, max [17]. PCR reactions were set up as described by the manufacturer, except that 2 ul of cDNA template were used for each reaction. PCR conditions were designed to ensure that all amplifications were within the logarithmic phase. Those conditions were; 94°C for 1 min, 25 cycles of 94°C 30 sec, 59°C for 30 sec, 72°C for 1 min, and a 72°C for1 min final extension for all primer sets except max which was only amplified for 14 cycles. Controls lacking RT were run for each RNA sample.
RT-PCR products were separated on a 1.5 % agarose gel, transferred to nylon membrane (Roche, Indianapolis, IN) and visualized by hybridization with a biotin labeled sequence designed to hybridize to gene 1, gene 2, and the variants (North2South Biotin labeling kit, Pierce Biotechnology Inc. Rockford, Il). Densitometric analysis of autoradiographs was performed to determine the relative expression levels of the genes and their isoforms at the above mentioned zebrafish developmental stages and tissues. Samples could be compared on different blots by using a control sample present on each autoradiograph, and samples were calibrated using the endogenous control max.
Zebrafish care and feeding was performed essentially as described by Westerfield [18]. All experimentation was done with the approval of the Canadian Council on Animal Care.
Authors' contributions
THLS completed the isolation of several of the zebrafish dnmt3 genes and performed the expression analysis identifying the isoforms. CCD first identified that numerous dnmt3 genes exist and isolated a portion of two of them. AAM initiated this work and isolated the first dnmt3 sequence from the fish. RAM conceived the study and isolated portions of a number of the dnmts. All had intellectual and design input into portions of the experimentation. All authors read and approved the manuscript.
Acknowledgements
The authors wish to thank Denise Flint for her editorial assistance. This work was supported by grants from the Natural Sciences and Engineering Research council of Canada.
==== Refs
Efstratiadis A Parental imprinting of autosomal mammalian genes Curr Opin Genet Dev 1994 4 265 280 8032205 10.1016/S0959-437X(05)80054-1
Reik W Allen N Genomic imprinting: Imprinting with and without methylation Curr Biol 1994 4 145 7 7953517 10.1016/S0960-9822(94)00034-5
Riggs A Pfeifer G X-chromosome inactivation and cell memory Trends Genet 1992 8 169 174 1369742
Cedar H DNA methylation and gene activity Cell 1988 3 3 4 3280142 10.1016/0092-8674(88)90479-5
Monk M Boubelik M Lehnert S Temporal and regional changes in DNA methylation in the embryonic, extraembryonic and germ cell lineages during mouse embryo development Development 1987 99 371 82 3653008
Santos F Hendrich B Reik W Dean W Dynamic reprogramming of DNA methylation in the early mouse embryo Dev Biol 2002 241 172 82 11784103 10.1006/dbio.2001.0501
Li E Bestor T Jaenisch R Targeted mutation of the DNA methyltransferase gene results in embryonic lethality Cell 1992 69 915 926 1606615 10.1016/0092-8674(92)90611-F
Okano M Bell D Haber D Li E DNA methyltransferases Dnmt3a and Dnmt3b are essential for de novo methylation and mammalian development Cell 1999 99 247 57 10555141 10.1016/S0092-8674(00)81656-6
Hata K Okana M Lei H Li E Dnmt3L cooperates with the Dnmt3 family of de novo DNA methyltransferases to establish maternal imprints in mice Devel 2002 129 1983 1993
Chen T Ueda Y Dodge J Wang Z Li E Establishment and maintenance of genomic methylation patterns in mouse embryonic stem cells by Dnmt3a and Dnmt3b Mol Cell Biol 2003 23 5594 5605 12897133 10.1128/MCB.23.16.5594-5605.2003
Mhanni AA McGowan RA Global changes in genomic methylation levels during early development of the zebrafish embryo Dev Gene Evol 2004 214 412 417 10.1007/s00427-004-0418-0
Martin C Laforest L Akimenko M Ekker M A role for DNA methylation in gastrulation and somite patterning Dev Biol 1999 206 189 205 9986732 10.1006/dbio.1998.9105
Yanagisawa Y Ito E Yuasa Y Maruyama K The human methyltransferases DNMT3A and DNMT3B have two types of promoters with different CpG contents Biochim Biophys Acta 2002 1577 457 465 12359337
Detich N Szyf M Szyf M Regulation of DNA methyltransferases in cancer DNA Methylation and Cancer Therapy 2005 Eurekah.com and Kluwer Academic/Plenum Publishers 125 141
Postlewait J Yan Y-L Gates M Horne S Amores A Brownlie A Donovan A Egan E Force A Gong Z Goutel C Fritz A Kelsh R Knapik E Liao E Paw B Ransom D Singer A Thomson M Abduljabbar T Yelick P Beier D Joly J-S Larhammar D Rosa F Westerfield M Zon L Johnson S Talbot W Vertebrate genome evolution and the zebrafish gene map Nat Genet 1998 18 345 349 9537416 10.1038/ng0498-345
Chomczynski P Sacchi N Single-step method of RNA isolation by acid guanidinium thiocyanate-phenol-chloroform extraction Anal Biochem 1987 162 156 159 2440339 10.1016/0003-2697(87)90021-2
Schreiber-Agus N Chin L Chen K Torres R Thomson C Sacchettini J Depinho R Evolutionary relationships and functional conservation among vertebrate MAX-associated proteins: the zebra fish homolog of Mxi1 Oncogene 1994 9 3167 3177 7936639
Westerfield M The Zebrafish Book 1995 Eugene, OR: University of Oregon Press
Shimoda N Yamakoshi K Miyake A Takeda H Identification of a gene required for de novo DNA methylation of the zebrafish no tail gene Dev Dyn 2005 233 1509 1516 15937923 10.1002/dvdy.20455
|
16236173
|
PMC1274307
|
CC BY
|
2021-01-04 16:30:32
|
no
|
BMC Dev Biol. 2005 Oct 19; 5:23
|
utf-8
|
BMC Dev Biol
| 2,005 |
10.1186/1471-213X-5-23
|
oa_comm
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.