id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
65eeac0a66c1381729b9206e
| 10 |
We obtained a MM-GB/SA score, which measures the binding free energy in ligand-protein complex at single point using the MMGB/SA method in the Prime program (Prime MMGBSA v3.000) . The docked poses were minimized around amino acids within 3Å from the inhibitors using the local optimization feature in Prime and the energies of the complexes were calculated using the OPLS4 force field and GB/SA continuum solvent model (Figure ). The binding free energy (Gbind) was then estimated using the equation:
|
65eeac0a66c1381729b9206e
| 11 |
where ΔEMM was the difference in energy between the ligand-protein complex and the sum of the energies of the ligand and free protein, using the OPLS-2005 force field; ΔGsolv was the difference in the GB/SA solvation energy between the ligand-protein complex and the sum of the solvation energies for the ligand and free protein; and ΔGSA was the difference in the surface area energy between the ligand-protein complex and the sum of the surface area energies for the ligand and free protein. Corrections for entropic changes were not applied. The dielectric constants were set according to solvent water (VSGB2.1 solvation model).
|
65eeac0a66c1381729b9206e
| 12 |
To compare the model validity between GLIDE docking and MM-GB/SA score against pIC50 value of MATE1 inhibitory activity, we performed a linear regression (Figure ) and calculated R 2 -values. The equation was as follows; where SSres means sum of squared of residuals, and SStot means total sum of squares. This calculation was performed in Microsoft Excel (version Office 365) manually.
|
65eeac0a66c1381729b9206e
| 13 |
The dataset was randomly divided into training (5/6 of all) and test datasets (1/6 of all). For machine learning, this selection process of training and test dataset were repeated six times to perform hold-out tests six times without replacement in test dataset for different runs. The holdout compound ID in each run was included in Table .
|
65eeac0a66c1381729b9206e
| 14 |
For machine learning modeling, chemical descriptors ECFP4 (Figure ), and compound SMILES themselves were used for RF and Message Passing Neural Network (MPNN), respectively. We firstly investigated the machine learning models (RF and MPNN) using only ECFP4 fingerprints (Figure ). Then, in order to investigate the effect of parallel usage of different type variables as explanatory ones, we built the models with both ECFP4 and MM-GB/SA scores (Figure ).
|
65eeac0a66c1381729b9206e
| 15 |
Regarding RF, the Python (ver. 3.7.10) scikit-learn (version 0.24.2) library RandomForest Regressor function was used, and the parameters were set as defaults . Regarding MPNN, the Python (ver. 3.7.10) Chemprop (version 1.3.1) library chemprop function was used, and the parameters were set as defaults . 2.7. Model Validation 5-fold CV was performed using the KFold function in scikit-learn (versions 0.24.2) with parameters: n_splits = 5 and shuffle = True. Furthermore, hold-out test was performed with a separate test dataset. To avoid the biased dataset composition, we repeated these process 6 times (Figure ).
|
65eeac0a66c1381729b9206e
| 16 |
In the 5-fold cross validation, AUC-ROC and Accuracy were calculated to quantify model performance. In the hold-out test, in addition to AUC-ROC and accuracy, sensitivity, specificity, Youden's index, MCC (Matthews Correlation Coefficient), false rate, F-measure were calculated using the confusion matrixes in Microsoft Excel (version Office 365) in order to examine the extrapolation of the model in detail.
|
65eeac0a66c1381729b9206e
| 17 |
The distribution of Tanimoto similarity of 5-nearest neighbours (NN) was calculated by ECFP4 using the RDKit (version 2020.09.01) 37 Chem functions. We set each bins according to the Tanimoto similarity as less than 0.2, 0.2 to 0.3, 0.3 to 0.4, and more than 0.4 (since lower and larger thresholds were not populated with sufficient numbers of data points) and plotted the percentage of compounds predicted correctly and incorrectly against these bins to investigate whether we can see the applicability domain through the predictivity of models built in this study.
|
65eeac0a66c1381729b9206e
| 18 |
To estimate the important amino acids for the inhibition of MATE1 activity, we calculated the Structural Interaction Fingerprints (SIFt) between the ligands and MATE1 using the Interaction Fingerprints in Maestro (Schrodinger Suite 2021-2) . The interaction patterns between the protein and the inhibits and non-inhibitors were compared. We drew a Venn-diagram for the number of interacted amino-acids against inhibitors and non-inhibitors using the matplotlib_venn (versions 0.11.10) venn2 function. We picked the amino-acids which interacted with inhibitors more than 10 times, then we calculated the ratio of the number of interactions with amino-acids among inhibitors vs non-inhibitors.
|
65eeac0a66c1381729b9206e
| 19 |
First, to investigate the chemical space of compounds collected in this study, UMAP analysis was performed comparing our compounds to FDA approved drugs. The result of this analysis is shown in Figure where it can be seen that the chemical space of compounds used in this study for predictive model building were consistent with that of FDA compounds. Hence we can conclude that the model to be built in this study could have a wide range of applicability in chemical space.
|
65eeac0a66c1381729b9206e
| 20 |
Next, in order to examine the validity of docking pose used in this study, we checked the docking pose of cimetidine. Because this compound is a clinical MATE inhibitor , although no bound X-ray structure is published, we can compare the docking poses obtained in our work to a previous study . The results are shown in Figure and. The NH moiety of the azole formed a hydrogen bond with Glu389, the azole ring can also form a π-interaction with Tyr299, and the cyanoguanidine moiety formed a hydrogen bond with Gln49. The interactions with Glu389 and Tyr299 were consistent with the previous report; and although the interaction with Gln49 was not reported, the position of the compound was consistent with the previous study . Hence we concluded that the quality of our docking poses passed this plausibility check by comparison with a previous study. . Regarding the MM-GB/SA regression model (Figure ), if the predicted value is below 10 μmol/L, the predicted class is considered positive. Conversely, if the predicted value exceeds 10 μmol/L, the predicted class is considered negative. When applying the classification threshold overall accuracy and ROC-AUC were relatively high at 0.684 and 0.742, respectively, and its specificity was 0.742; precision and sensitivity were somewhat lower at 0.660 and 0.614, respectively. However, considering the best AUC of the previous best ML model was 0.78 [J Med Chem. 2013;56(3): 781-795] , we can see that MM-GB/SA scores have the large impact for the prediction of MATE1 inhibitory activity.
|
65eeac0a66c1381729b9206e
| 21 |
Regarding the RF model using ECFP4 fingerprints, the ROC-AUC and averaged accuracy were very high (0.824 and 0.765, respectively), and comparing those in 5-fold CV (0.810 and 0.741) the overfitting was considered to be avoided. The precision and specificity were also high (0.786 and 0.876, respectively). However, since the sensitivity was relatively low (0.630), the concern with this model was to miss potential MATE1 inhibitors.
|
65eeac0a66c1381729b9206e
| 22 |
The RF model using both ECFP4 and MM-GB/SA score as input features resembles in its metrics the RF model using only ECFP4 fingerprints. Specifically, this model had high values for ROC-AUC (0.833), averaged accuracy (0.765), precision (0.779), specificity (0.862), and a relatively low value for sensitivity (0.640). Nevertheless, the ROC-AUC was 0.833, which is both a better value than that reported in previous studies, as well as the best value obtained here. The ROC-AUC and averaged accuracy of 5-fold CV of this model were 0.823 and 0.741, which were similar to those in the hold-out test set, and hence this model had little concern regarding overfitting.
|
65eeac0a66c1381729b9206e
| 23 |
Regarding the conventional MPNN model, the ROC-AUC was 0.695 which was lower than the MM-GB/SA regression model, and the accuracy was 0.684 which was identical to it. Respective values from 5-fold CV were 0.710 and 0.641, which indicates no significant overfitting. The precision and specificity were lowest in this study (at 0.637 and 0.655, respectively). However, the sensitivity was higher (0.716) than those of other models (0.614 to 0.640). Consequently, we can see that this MPNN model has lower predictivity in overall metrics; however, its high sensitivity is beneficial not to miss any inhibitors, which is a very important aspect in practical project work.
|
65eeac0a66c1381729b9206e
| 24 |
We next investigated the applicability domain of two RF models, based on only ECFP4 fingerprints, and based on both ECFP4 and MM-GB/SA scores,by 5-NN analysis using Tanimoto similarity of ECFP4 fingerprints. According to one practical implementation, if more compounds are predicted correctly in a higher Tanimoto similarity bin in a 5-NN analysis, we can say this model has an interpretable applicability domain, which we can use to estimate prediction confidence for a new compound in a given 5NN similarity bin . The results of this analysis are shown in Figure where it can be seen that in RF model using only ECFP4, the percentage of correctly predicted compounds increases according to the Tanimoto similarity, from 71% in the 0.1 to 0.2 of Tanimoto similarity bin (TSB), over 73% (0.2 to 0.3 of TSB) and 84% (0.3 to 0.4 of TSB), to 85% (more than 0.4 of TSB). The second results of this analysis are shown in Figure where it can be seen that in RF model using both ECFP4 and MM-GB/SA score, the percentage of correctly predicted compounds also increases according to the Tanimoto similarity, from 71% (0.1 to 0.2 of TSB), over 75% (0.2 to 0.3 of TSB) and 79% (0.3 to 0.4 of TSB), to 85% (more than 0.4 of TSB). Therefore based on these results, we can conclude that for the RF model, regardless of the explanatory variables, we see a correlation of model performance with proximity to the training set. Finally the result of MM-GB/SA regression model is shown in Figure where it can been that the percentage of correctly predicted compounds was largely independent of the Tanimoto similarity, ranging from 71% (0.1 to 0.2 of TSB), over 65% (0.2 to 0.3 of TSB), and 71% (0.3 to 0.4 of TSB) to 72% (more than 0.4 of TSB). Hence the performance of the MM-GB/SA regression model is independent of similarity to the training set, which is due to MM-GB/SA based on physics-based (instead of data-driven) model . From the practical viewpoint of drug discovery process, even if we do not have any experimental results, MM-GB/SA model secure us about 70% predictivity; we can say that according to the position of the chemical space of the new compounds we should select the kinds of models (data-driven, physics-based, or parallel use of them).
|
65eeac0a66c1381729b9206e
| 25 |
Next, in order to investigate the important amino-acids, we estimated the amino-acids of MATE1 transporter which interact with inhibitors through SIFts. The result of this analysis is shown in Figure where it can be seen that several amino-acids emerged frequently as contacted aminoacids, out of a total set of 570 amino acids present. The number of amino-acids that interact with inhibitors was 58, and those with non-inhibitors was 55. Interestingly 50 amino-acids were overlapped (Figure ). When we focused on amino-acids that interact with more than 10 inhibitors, we can find that Gln49, Ph53, Asn82, Trp274, Tyr277, Glu278, Tyr299, Ile303, and Tyr306 are involved in forming interactions (Figure ). Most of these amino-acids are related to the contact with non-inhibitors. Although only Glu278 had the larger number of contact in inhibitors than non-inhibitors, the difference of it was not so large (63 inhibitors and 41 noninhibitors). Hence, we can conclude that the path of MATE1 for the inhibitors and non-inhibitors are mostly the same.
|
65eeac0a66c1381729b9206e
| 26 |
In order to investigate the interactions of compounds with MATE1 further, we scrutinized the noninhibitor compounds which are known as in vitro typical substrates in FDA guidance 10 , 1-methyl-4-phenylpyridinium, tetraethylammonium, creatinine, metformin. The contacted amino-acids with these substrates were Gln49, Glu273, Trp274, Tyr277, Tyr299, Ala302, Ile303, Tyr306, Met307, Glu389. Since most of these amino-acids (Gln49, Trp274, Tyr277, Tyr299, Ile303, Tyr306) were the same with the inhibitor contacting amino-acids, we can conclude that substrate and inhibitor overall share rather similar binding space in MATE1. From the viewpoint of drug discovery using in silico technology, this indicates the difficulty to distinguish between classes based on the docking position, and this leads to the low predictivity of MM-GBSA score. On the other hand, from the viewpoint of drug discovery using experimental science, this is consistent with the fact that the IC50 values of each inhibitor when using different substrates are similar ; and practically this can release the experimental scientists from bothering of selecting substrates for MATE1.
|
65eeac0a66c1381729b9206e
| 27 |
In this work we aimed to integrate ligand-based and structure-based information to arrive at a MATE1 inhibition model which can be used for IC50 prediction. We found that overall best predictivity was observed in the RF model (including ECFP4 and MM-GB/SA input variables) with an ROC-AUC of 0.833. On the one hand RF models have good precision and specificity; on the other hand MPNN models have good sensitivity. Additionally, fingerprint-based models performed better with increasing proximity to the training set, which was shown not to be the case for a MM-GB/SA regression model due to its physics-based nature. Via Structural Interaction Fingerprint analysis, we identified residues relevant for inhibitor are mostly the same with non-inhibitors including substrates, such as Gln49, Trp274, Tyr277, Tyr299, Ile303, Tyr306. This is consistent with the fact that the IC50 values of each inhibitor when using different substrates are similar, and practically this can release the experimental scientists from bothering of selecting substrates for MATE1. Hence we were in the current study able to build a fit-for-purpose classification model for MATE1 inhibitory activity based on dose-response data based on both ECFP4 fingerprints and MM-GB/SA score for the first time, which will be useful in practice to identify and eliminate more compounds related to MATE1-based DDI in the early drug discovery stages.
|
616ec04ef718dffe58e3e06b
| 0 |
D-serine is a physiological co-agonist of the N-methyl D-aspartate (NMDA) type of glutamate receptor, a key excitatory neurotransmitter receptor in the brain. D-Serine in the brain is synthesized from its L-isomer by serine racemase and is metabolized by the D-amino acid oxidase (DAO, DAAO), a flavoenzyme that catalyzes the oxidative degradation of D-amino acids including D-serine to the corresponding α-keto acids. The function of the NMDA receptor requires the presence of both the agonist (glutamate) and the co-agonist (D-serine, glycine, and/or D-alanine). Importantly, D-serine has been reported to be the predominant NMDA co-agonist in the forebrain and linked directly to schizophrenia. 1 D-serine concentrations in serum and cerebrospinal fluid have been reported to be decreased in schizophrenia patients, 2 and oral administration of D-serine improved symptoms of schizophrenia when used as an adjuvant to typical and atypical antipsychotics. 3 Thus, it is plausible to explore pharmaceutical inhibition of DAO function as putative novel therapeutics to treat the positive (psychotic), negative and cognitive symptoms in schizophrenia.
|
616ec04ef718dffe58e3e06b
| 1 |
The simplest DAO inhibitor benzoic acid was reported in 1956. 4 Since the early 2000s, many small molecule DAO inhibitors have been reported in the literature (Figure ). 5 They all mimic the substrate D-serine and bind to the catalytic site of DAO. The early inhibitors 6 can all be characterized as aryl carboxylic acids or corresponding acid-bioisosteres with low molecular weight. Although they are potent and highly ligand efficient, they lack the vectors that are needed for optimization of potency and physicochemical properties. To that end in 2013, Astellas reported a new class of DAO inhibitors which contain a tail group reaching into a hydrophobic pocket perpendicular to the head group. 7 Takeda also worked on a similar chemical series which culminated in the discovery of their clinical candidate TAK-831 . The kojic acid derivatives were also explored by a Johns Hopkins research group. 8 In addition, Sunovion reported a new class of DAO inhibitors (10) that stabilize an activesite lid-open conformation, although the lead compounds suffer from poor pharmacokinetic and brain penetration properties. 9
|
616ec04ef718dffe58e3e06b
| 2 |
A few DAO inhibitors have entered into clinical development. Currently, only SyneuRx is actively developing NaBen® (sodium salt of 1) in a phase II/III clinical trial for refractory schizophrenia in adults. 10 Sepracor was developing SEP-227900 for neuropathic pain around 2010. Takeda was developing TAK-831 (luvadaxistat, 8) in phase 2 clinical trials 11 for the treatment of schizophrenia, which was the subject of a license agreement with Neurocrine in 2020. 12 In March 2021, Neurocrine reported topline data from the Phase II INTERACT study in adults with negative symptoms of schizophrenia treated with luvadaxistat (NBI-1065844/TAK-831). Although luvadaxistat did not meet its primary endpoint in the study, as measured by the change from baseline on the PANSS NSFS at Day 84, Luvadaxistat met secondary endpoints of cognitive assessment, which merit further clinical evaluation. 13 The improvement of cognitive function for TAK-831 in schizophrenic patients is consistent with improvement of cognitive performance in rodent models. For example, another DAO inhibitor SEP-227900 increased D-serine in the cerebellum of rats in a dose dependent manner, and pretreatment of rats with this DAO inhibitor increased memory of the test object in the novel object recognition model in rats, suggesting improved cognitive function. 14 There are many published DAO co-crystal structures in the literature. Figure shows the co-crystal structure of human DAO enzyme with a hydroxy pyridazinone ligand , which was one of the most potent DAO inhibitors reported by both Takeda and Astellas. 7 Overall, the ligand adopts an L-shaped conformation in the binding site. The hydroxy pyridazinone head group is stacked between the flavin ring of FAD (flavin adenine dinucleotide) and Tyr224. The hydroxy-carbonyl moiety forms a salt bridge with Arg283, and the N-H forms an H-bond with Gly313. On the other side, the phenyl ring sticks into a relatively hydrophobic pocket and stacks with Tyr224 to form a πedge interaction. Intrigued by the target biology and therapeutic potential for treatment of cognitive impairment in schizophrenia or other neurological disorders, we initiated a program to identify novel DAO inhibitors with best-in-class properties. The program leveraged the Schrödinger physics-based modeling technology, specifically, a human DAO Free Energy Perturbation (FEP+) model which was developed on the basis of published SAR data. 15 Key protein-ligand interactions presented in the co-crystal structures were taken into account as novel ligands were designed by multiple internal medicinal and computational chemists. The designs were further evaluated with the hDAO FEP+ model, and the top ideas were prioritized for synthesis. Among them (Table ), 12, 13 and 14 16 showed good hDAO biochemical potency, which is consistent with the FEP+ model prediction. 17 Thanks to their low molecular weights, all three compounds have good ligand efficiency (LE) and lipophilic ligand efficiency (LLE). Initially SAR work indicated that various substituents can be tolerated on the phenyl ring of both dihydropyrazine dione (12, DHP dione) and N-hydroxyl pyrimidine dione (13, NHP dione) hit classes. For instance, the CF3 group of 12 can be replaced with a chlorine to yield 15 which shares similar DAO potency. Similarly, a chlorine can be incorporated at the para-position of 13 to afford 16, which is slightly more potent than 13. In order to understand the in vivo pharmacokinetic (PK) properties of the hits, especially their ability to cross the blood-brain barrier (BBB), compounds 14, 15, and 16 were dosed as a cassette in mice along with 17 18 (Table ) as a reference compound. To understand the binding interactions of the DHP dione chemical series, a co-crystal structure of 12 was obtained via a soaking experiment with the hDAO apo crystal. As shown in Figure , compound 12 binds to the hDAO enzyme in a fashion very similar to ligand 11. The dihydropyrazine dione head group is stacked between the flavin ring and Tyr224. The hydroxy-carbonyl moiety forms a salt bridge with Arg283, and the NH forms H-bond with Gly313. On the tail side, the 4-trifluoromethyl phenyl sticks into the hydrophobic pocket. Unlike the acidic hydroxy pyridazinone head group in compound 11, the pKa of 12 was measured at 9.7, 20 which would imply a pKa penalty in binding to DAO, as only the anionic form can actively bind to DAO. In addition, the head group of 12 is pseudosymmetric with two possible anionic tautomers, and substitution on the head group can impact the tautomer distribution. Fortunately, quantum mechanics (QM) calculations suggested that the active tautomer is strongly favored for 12, by 0.8 kcal/mol. Although the DHP dione chemical series is relatively weaker than the hydroxy pyridazinone 21 chemical series due to the higher pKa of the head group, it may benefit from other properties such as pharmacokinetics and brain penetration.
|
616ec04ef718dffe58e3e06b
| 3 |
Initial SAR exploration was focused on the aromatic tail region of compound 12. Both rational design by medicinal chemists and computational enumeration by Schrödinger's AutoDesigner algorithm were applied to generate a diverse set of design ideas. The large number of designs were filtered by molecular properties, a CNS MPO, a druglikeness MPO and synthetic tractability, etc. The top scoring designs were progressed into FEP+ calculations to predict hDAO inhibitory potency. The compounds with favorable predicted hDAO potency were selected for synthesis at Charles River Laboratories (CRL). Additionally, active compounds were tested in the MDCK-MDR1 assays to assess cell permeability and efflux ratio (ER).
|
616ec04ef718dffe58e3e06b
| 4 |
Compound 17 was also included in Table as a reference compound, which was measured 17 nM in the hDAO biochemical assay. It is slightly right shifted in the hDAO cell assay, but about 4-fold left shifted in the mouse DAO cell assay. In the DHP dione chemical series, para-substitution on the phenyl ring is beneficial to potency, as the unsubstituted analog 18 is much less active. At the para-position, CN substitution (19) can also be tolerated in addition to Cl, while the methoxy analog 20 is less potent. From the mono Cl-substituted analogs 15, 21, and 22, para-substitution is the most preferred, while ortho-substitution is not tolerated. Compound 19 can be substituted with a fluorine ortho to the cyano group as in analog 23, while 3,5-dichloro substituted analog 24 is less active when compared to the mono-substituted analog 21. The tail region tolerates other hetero aromatic rings such as pyridine (25) and bicyclic aromatic rings such as quinoline (26) with some loss of potency. Polar groups can also be tolerated in this region as exemplified by compound 27 and 28. It is worth noting that both compound 27 and 28 were designed by the AutoDesigner algorithm featuring uncommon yet drug-like functionalities. In terms of hDAO FEP+ model performance, the majority of the prediction is within 1 log unit of the experimental IC50 value. Compared to 17, most analogs showed lower but moderate cell permeability and low efflux ratio in the MDCK-MDR1 assay, which may partially account for the near 10-fold shift in the hDAO cell assay. To ensure that the compound activity is not an artifact from their redox potential, the horseradish peroxidase assay (HRP) was developed as a counter screen. All compounds in Table were shown to be clean up to 10 µM in the HRP assay. SAR of the linker region was also explored (Table ). In terms of the linker length, the 2-carbon linker ( ) is superior to the 3-carbon linker (33) according to the FEP+ predictions. 22 It is also better than the 1-carbon linker based on the matched pair of 15 and 34. Analogs with fluorinated linker (i.e. 35) are also interesting, as fluorinesubstituted linkers were predicted to lower the pKa of the head group. However, compound 35 failed in the synthesis due to chemical stability issues. Linkers with hetero atoms were also explored. Although analogs with an oxygen linker failed in synthesis, the sulfur-linked analogs are stable enough for further SAR development. Encouragingly, both 36 and 37 are about 3-fold more potent than 12 in the hDAO biochemical assay. The enhancement in biochemical potency may partly be attributed to lower pKa's of the head groups in 36 and 37. Compound 36 has a measured pKa of 8.5, and compound 37 9.2. Compared to 12 with a pKa of 9.7, 36 and 37 are more favored to form the bio-active anionic structures. In addition, they both show a lower cell shift when compared to 12, possibly due to moderately higher cell permeability as measured in the MDCK-MDR1 assay. The tail SAR of the sulfur-linked analogs 36 and 37 largely resembles that of 12 (Table ). Further exploration of the tail SAR of sulfur-linked analogs led to significant potency improvement in the hDAO biochemical and cell assays. In order to further enhance compound inhibitory potency against hDAO, fused ring designs in the linker region were also assessed by the FEP+ model. One of the ideas that stood out was 54, which was designed by cyclizing the linker of 37 to form a fused 1,4-oxathiane ring. It was predicted that cyclization would lead to a gain in potency resulting at least in part from stabilization of the linker and tail piece. FEP+ predicted this compound to be a 3 nM inhibitor in the hDAO assay (Figure ). To our delight, the compound showed an IC50 of 25 nM in the assay, a 3fold improvement from 37. Thus, cyclized analogs with the best substituents from the chemical series were prepared. Most of these analogs showed significant improvement when compared to their acyclic counterparts in the hDAO biochemical assay. However, there was less improvement in the human and mouse cell DAO assays due to larger cell shift for the cyclized analogs. The binding mode of the cyclized analogs was confirmed by X-ray cocrystal structure of 59 (Figure ), which very much resembles the FEP+ snapshot of compound 54 binding to hDAO. The biggest changes are in the tail region due to different substituents at the para-position. Another interesting design on the cyclized analogs is quaternary methyl adduct 61. While this methyl addition was not initially predicted by our FEP+ model to lead to any gain in potency when compared to the des-methyl analog 55, the racemic quaternary methyl 61-rac 23 was tested to be about 5-fold more potent than 55-rac. That is because we did not have more closely related starting references for our FEP+ model at the time of the original prediction for 61. Subsequently we troubleshooted the FEP+ model by using the more closely related des-methyl analog 55 as the starting reference and observed that the methyl group displaces a high-energy water molecule concurrent with a predicted gain of potency (Figure ). The quaternary methyl group was incorporated into other analogs and resulted in roughly 2-fold improvement in DAO biochemical and cellular assays. Notably, compound 63 showed inhibition potency near 100 nM in both human and mouse DAO cell assays. In an effort to explore new opportunities for potency enhancement, careful examination of the DAO catalytic site revealed a subpocket just beyond the tail region, which was not explored by other groups (Figure ). In order to design into this subpocket, we employed our AutoDesigner algorithm to enumerate novel design ideas using compound 54 as a template. Initially over 198 million design ideas were generated by the algorithm, which were filtered by an array of criteria such as molecular properties, CNS and drug-like MPO's, and synthetic complexity.
|
616ec04ef718dffe58e3e06b
| 5 |
After GLIDE docking into the hDAO crystal structure, the surviving compounds were evaluated by the hDAO FEP+ model for potency. Only three top compounds were selected for synthesis, among which compound 66 stood out as a single digit nM hDAO inhibitor on the project. With just one round of synthesis, we were able to confirm that the subpocket is a viable design space to further enhance compound binding potency to the hDAO enzyme, which opens up much needed new SAR space for this target.
|
616ec04ef718dffe58e3e06b
| 6 |
In order to demonstrate the therapeutic potential of the DHP dione chemical class, the team next tried to identify a suitable candidate to probe PK/PD relationship in vivo. As mentioned earlier in the SAR, most analogs showed moderate permeability and low efflux ratio in the MDCK-MDR1 cell line. They showed excellent stability in the human and mouse liver microsome assay. The compounds have also shown good stability in human and mouse hepatocytes, as no significant turnover was observed for most compounds under the assay conditions employed. The in vivo drug metabolism and pharmacokinetic (DMPK) properties were assessed in cassettes of five compounds each, including 17 as the reference. Cassette administration is an extremely useful approach to generate in vivo PK data quickly in a cost effective and animal sparing fashion. A cassette dosing strategy also enabled direct comparison of drug brain penetrability among a set of compounds within the same set of animals. In practice, cassette doses were prepared for both intravenous (IV) and oral (PO) administration utilizing a standard dose formulation for each route throughout the project.
|
616ec04ef718dffe58e3e06b
| 7 |
Table shows mouse plasma PK of a few compounds in the chemical series. Most analogs showed low to moderate clearance and normal volume of distribution in mice, which resulted in good half-life values. They are also well absorbed when dosed orally with oral bioavailability generally over 40%. Not surprisingly, compound 66 showed reduced and less favorable oral bioavailability, possibly due to multiple rotatable bonds in the structure.
|
616ec04ef718dffe58e3e06b
| 8 |
Compound 46 was deprioritized due to lower free drug fraction in the brain. To enable selection of a PK/PD candidate, high dose oral PK studies were carried out at 10 and 100 mg/kg for both compounds. Compound 37 demonstrated good dose linearity in brain, while 42 showed sub dose proportionality at 100 mg/kg (Figure ).
|
616ec04ef718dffe58e3e06b
| 9 |
A B Ratio AUC0-inf Ratio AUC0-inf = 5.2 Modeling of the PK and theoretical enzyme occupancy (Equation ) after a single 100 mg/kg dose identified 37 to be the optimal compound to progress into a PK/PD study with a 150 mg/kg BID, Q4hr dosing regimen. This study design, in conjunction with the measured mouse cell IC50, the concentration of 37 in the cerebellum and the corresponding free fraction in this tissue was predicted to provide enzyme occupancy and coverage commensurate with an in vivo biomarker response (see Figure ). Projected tissue concentrations at 15, 50 and 150 mg/kg were calculated following a linear extrapolation of the measured values obtained from the 100 mg/kg dosing cohort illustrated in Figure . These data were used in Equation 1 to generate the %tEO profiles in Figure . Both 37 and 42 have been extensively screened in vitro for potential off-targets. In the Eurofins Safety/Diversity panel (Table ), COX2 is the only off-target for both 37 and 42, representing about 93-fold in vitro selectivity for 37 and 132-fold for 42. In addition, the compounds have also been screened against six additional CNS targets at Eurofins, and none of them showed significant activity at 10 µM on the six off-targets. No significant inhibition of the major human CYP enzymes (<40%, 3A4, 2D6, 2C9, 2C19, 2C8, 1A2) was observed for either compound at 10 µM. In addition, there was a complete absence of any cytotoxicity signal for either compound when they were tested at 100 µM in a HepG2 assay that measured 72-hour ATP production and 24-hour Glu/Gal mitotoxicity. Following ethical review and approval of the study protocol, the PK/PD assessment was undertaken to measure the modulation of D-serine levels in the cerebella of mice following administration of the test compound at one dose using the regimen described above. Two cohorts of animals were tested (compound and vehicle) using 33 animals in total (n=8/group for 37 and n=3/timepoint for vehicle). In both cases plasma and cerebella samples were collected following animal "take-down" at 4-hour, prior to 2nd dosing, 6-hour, and 10-hour after the initial dosing. The levels of D-serine in plasma and brain tissue were quantitatively determined using a chiral LC-MS/MS method, ensuring both adequate sensitivity and selectivity. In addition, CSF was sampled from the animals at the 10-hour timepoint to determine the free, unbound levels of 37.
|
616ec04ef718dffe58e3e06b
| 10 |
The bioanalytical results obtained from the PK/PD study are shown in Figure . As can be seen based on a mouse cell EC50 of ~150 ng/mL and brain tissue binding of 94.8%, free drug exposures exceeding the mouse cell EC50 were observed at 10 hours in the plasma, cerebellum and CSF. A significant increase of D-serine levels compared to vehicle was also observed in both the plasma and cerebellum at all three time points measured (Figure ). In addition, a parallel study was run to assess the receptor occupancy (RO) in the cerebellum with the Takeda tracer compound PGM019260 following the protocol published in Neurochemistry Research 2017 . As shown in Figure , the study confirmed significant RO of compound 37 in the PK/PD study, as projected by PK modeling (Figure ). The results of the PK/PD study are summarized in Table . Based upon the data, the PK/PD study with compound 37 has successfully demonstrated the pharmacological potential of hDAO inhibitors from the DHP dione chemical series. In parallel with the PK/PD study, compound 37 was also assessed in a catalepsy model using the same dosing regimen (150 mg/kg p.o. BID, Q4hr) that had generated the positive response in the PK/PD study. During this study, plasma samples were taken and used to assess the prolactin levels at 6 hours post the first dose, which was predicted to be around Cmax. As shown in Figure , no catalepsy or increase in prolactin levels was observed in this study. Plasma and brain concentrations of 37 were determined indicating that levels were similar to those achieved in the PK/PD study (data not shown). This study confirms that 37 is well tolerated in vivo at exposure levels required to evoke the desired PD responses.
|
616ec04ef718dffe58e3e06b
| 11 |
In summary, we have discovered a novel class of small molecule inhibitors against the human D-amino Acid Oxidase (DAO). Different from the earlier lead compounds, this chemical class features a non-acidic dihydropyrazine dione head moiety. Starting from hit compound 12, SAR work in the linker region led to the discovery of thioether linker analogs which showed enhanced DAO potency with desirable PK and brain penetration properties. With tool compound 37, we were able to demonstrate PK/PD in an in vivo mouse model at drug exposure levels devoid of any adverse events. Continued SAR work has led to compounds with significant improvement in both DAO biochemical and cellular potency.
|
616ec04ef718dffe58e3e06b
| 12 |
We have leveraged Schrödinger's computational modeling technology extensively to accelerate the program execution. Free energy perturbation (FEP+) technology was applied to prioritize compounds based on prospective binding potency predictions. Overall, the FEP+ models have performed well in predicting compounds binding potency to the hDAO enzyme. As shown in Figure , compound experimental hDAO inhibitory potency correlates well with prospectively predicted potency across the three chemical series. 24 Of the ~11000 ideas designed and profiled in silico, we synthesized 208 compounds and only 20 of these were unexpectedly inactive (>10 uM), demonstrating that the physics-based methods allowed us to quickly prioritize compounds of interest and deprioritize compounds that did not meet project objectives. In addition to structure-based design by seasoned medicinal chemists and modelers, we have also applied computational enumeration with our AutoDesigner algorithm to generate novel design ideas. Most notably, this effort has helped to identify a novel subpocket for further SAR development on the project. As is common for CNS programs, the challenge is to balance compound potency with desirable PK/brain penetration properties. While a working model to predict PK/brain penetration has been elusive on this project, we will continue to apply the Schrödinger computational modeling technology along with drug-likeness and CNS MPO filters to prioritize compounds for synthesis. Further optimization work toward a development candidate will be reported in due course.
|
616ec04ef718dffe58e3e06b
| 13 |
The D-amino acid oxidase (DAO) assays are fluorescence-based assays, in which the hydrogen peroxide (H2O2) generated from the reaction of D-serine with DAO and Flavine Adenine Dinucleotide (FAD), is linked to oxidation of Amplex Red in the presence of horseradish peroxidase (HRP). The Amplex Red reagent reacts with H2O2 in a 1:1 stoichiometry to produce the red-fluorescent oxidation product, resorufin, which is measured fluorometrically.
|
616ec04ef718dffe58e3e06b
| 14 |
The human DAO biochemical assay was performed using reagents at the following final assay concentrations: 1 nM recombinant full-length human DAO protein, D-Serine at Km concentration (10 mM), 50 µM FAD (excess), 50 µM Amplex Red and 0.1 U/mL HRP in the presence of compound or DMSO vehicle (1%). All reagents were made up in assay buffer containing 20 mM Tris, pH 7.4 + 0.1% BSA. The final assay volume was 25 µL/well.
|
616ec04ef718dffe58e3e06b
| 15 |
Briefly, 10 µL of a working solution containing 2.5 nM hDAO (TECC-1280-14AA, Takeda) and 125 µM FAD (F6625, Sigma) in assay buffer was added to all the wells in the assay ready plate (containing 250 nL compound / DMSO vehicle per well) except for the negative control wells. 10 µL of 125 µM FAD (working solution) was added to the negative control wells containing 250 nL of DMSO vehicle. The plates were incubated at 25˚C for 20 minutes (pre-incubation of compound with human DAO).
|
616ec04ef718dffe58e3e06b
| 16 |
10 µL of a working solution containing 125 µM Amplex Red and 0.25 U/mL HRP (A22188, ThermoFisher Scientific) in assay buffer was then added to all the wells. The reaction was initiated by the addition of 5 µL of 50 mM D-Serine (S4250, Sigma-Aldrich) to all the wells. The plate was incubated for 4 hours in the dark at 25˚C before measuring fluorescence in each well using the Envision plate reader with excitation at 530 nm and emission at 595 nm. Concentration response curves were generated using ActivityBase (IDBS). IC50 values were determined by plotting % Inhibition vs Log10 compound concentration using a sigmoidal fit with a variable slope (four parameter fit).
|
616ec04ef718dffe58e3e06b
| 17 |
Briefly, 5 µL of 250 µM FAD and 5µl of 50mM D-Serine were added to all the wells in the assay ready plate (containing 250 nL compound / DMSO vehicle per well). 5 µL of a working solution containing 250 µM Amplex Red and 0.5 U/mL HRP was added to all the wells except for the negative control wells. 5 µL of 250 µM Amplex Red was added to the negative control wells containing 250 nL of DMSO vehicle. 10 µL of 5 µM H2O2 was added to all the wells. The plate was incubated for 10 minutes in the dark at 25˚C before measuring fluorescence in each well using the Envision plate reader with excitation at 530 nm and emission at 595 nm.
|
616ec04ef718dffe58e3e06b
| 18 |
The human DAO cell assay routinely employed a CHO-K1 clone, which was stably transfected with a mammalian expression plasmid containing the human DAO nucleotide encoding the full-length human DAO protein. This cell line was originally generated as described in . The human DAO CHO-K1 stable cell line was routinely cultured in Gibco Ham's F-12 Nutrient Mix (31765-027, ThermoFisher Scientific) containing 10% FBS (10082-147, ThermoFisher Scientific) and 500 µg/mL Geneticin™ Selective Antibiotic (10131-027, ThermoFisher Scientific).
|
616ec04ef718dffe58e3e06b
| 19 |
The human DAO cell assay was performed using the following final assay conditions: 25,000 human DAO CHO-K1 stable cells/well, 50 mM D-Serine, 50 µM Amplex Red and 0.125 U/mL HRP. All cells and reagents were made up in 10 mM HEPES buffer (15630-056, ThermoFisher Scientific). The final assay volume was 25 µL/well. The human DAO CHO-K1 stable cells were trypsinised, resuspended in complete medium and centrifuged at 1200 rpm for 4 minutes at room temperature. The cell pellet was then washed in 10 mM HEPES buffer and centrifuged at 1200 rpm for 4 minutes at room temperature. The resulting cell pellet was resuspended in 10 mM HEPES buffer at 1.25x10 6 cells/mL. 25,000 human DAO CHO-K1 stable cells (20 µL in 10 mM HEPES buffer) were added to all the wells in the assay ready plate (containing 250 nL compound / DMSO vehicle per well). 5µl of a working solution containing 250 mM D-Serine, 250 µM Amplex Red and 0.625 U/mL HRP in assay buffer was added to all the wells except for the negative control wells. 5 µL of a working solution of 250 µM Amplex Red and 0.625 U/mL HRP in assay buffer was added to the negative control wells. The plate was incubated for 30 minutes in the dark at 25˚C before measuring fluorescence in each well using the Envision plate reader with excitation at 530 nm and emission at 595 nm. Dose response curves were generated using ActivityBase (IDBS). IC50 (Point of Inflection) values were determined by plotting % Inhibition vs Log10 compound concentration using a sigmoidal fit with a variable slope (four parameter fit).
|
616ec04ef718dffe58e3e06b
| 20 |
The mouse DAO cell assay routinely employed CHO-K1 cells, which were transiently transfected with an expression plasmid containing the mouse DAO nucleotide encoding the full-length mouse DAO protein. The CHO-K1 cell line was routinely cultured in Gibco Ham's F-12 Nutrient Mix containing 10% FBS. The mouse DAO cell assay was performed using the following final assay conditions: 35,000 CHO-K1 cells transiently transfected with mouse DAO/ pcDNA3.1+C_(K)-DYK expression plasmid, 50 mM D-Serine, 50 µM Amplex Red and 0.125 U/mL HRP. All cells and reagents were made up in 10 mM HEPES buffer. The final assay volume was 25 µL/well.
|
616ec04ef718dffe58e3e06b
| 21 |
The mouse DAO transiently transfected CHO-K1 cells were trypsinised, resuspended in complete medium and centrifuged at 1200 rpm for 4 minutes at room temperature. The cell pellet was then washed in 10mM HEPES buffer and centrifuged at 1200 rpm for 4 minutes at room temperature. The resulting cell pellet was resuspended in 10 mM HEPES buffer at 1.75x10 6 cells/mL.
|
616ec04ef718dffe58e3e06b
| 22 |
35,000 mouse DAO CHO-K1 transiently transfected cells (20 µL in 10 mM HEPES buffer) were added to all the wells in the assay ready plate (containing 250 nL compound / DMSO vehicle per well). 5 µL of a working solution containing 250 mM D-Serine, 250 µM Amplex Red and 0.625 U/mL HRP in assay buffer was added to all the wells except for the negative control wells. 5 µL of a working solution of 250 µM Amplex Red and 0.625 U/mL HRP in assay buffer was added to the negative control wells. The plate was incubated for 30 minutes in the dark at 25˚C before measuring fluorescence in each well using the Envision plate reader with excitation at 530 nm and emission at 595 nm.
|
616ec04ef718dffe58e3e06b
| 23 |
Male C57Bl/6NCrl mice, Inbred, SPF-Quality, Charles River, Germany between 8 and 10 weeks of age, ranging from 20 to 40 grams were used to study the pharmacokinetics of test compounds. On arrival and following randomization animals were housed individually in polycarbonate cages equipped with water bottles, unless contraindicated by study procedures (such as pharmacokinetic blood sampling) or clinical signs. Pelleted rodent diet (SM R/M-Z from SSNIFF® Spezialdiäten GmbH, Soest, Germany) was provided ad libitum throughout the study, except during designated procedures. The compounds were administered to the mice via a single intravenous (slow bolus) injection to the tail vein using a vehicle comprising DMSO, PEG200 and Water . Terminal blood samples were collected via aorta puncture following inhalation anaesthesia into K2EDTA tubes and stored on wet ice. Oral cohorts were dosed by gavage using a vehicle of 0.5% (w/v) methylcellulose and 0.1% (v/v) Tween80 in water with bloods collected using a similar procedure. Whole blood was processed to plasma by centrifugation (3000g for 10 minutes at 5°C) within 30 minutes of collection. Plasma samples were transferred into 96 well plates (matrix tubes) and stored at < -75°C. Following termination, brains were collected from the animals and the cerebella separated. Both tissues were rinsed with saline, weighed and stored at ≤ -75 oC prior to analysis using LC-MS-MS.
|
616ec04ef718dffe58e3e06b
| 24 |
Plasma and brain samples were extracted by protein precipitation using acetonitrile containing an appropriate internal standard. Specific reaction monitoring transitions were identified using automated instrumental optimization procedures for each compound studied, to ensure adequate linearity of response and define the upper and lower limits of quantitation. Samples were injected (SIL-30AC Autosampler, Schimadzu, Kyoto, Japan) onto a reverse phase chromatography system (A: 0.1% formic acid in ultrapure water; B: 0.1% formic acid in acetonitrile, Waters Corporation Acquity® UPLC column HSS T3 1.8μ). Analysis was performed using an API 5000 triple quadrupole mass spectrometer fitted with an electrospray ionisation source (AB Sciex, Ontario, ON, Canada). Pharmacokinetic analysis was performed with IDBS E-WorkBook v10 using mean data, non-compartmental analysis and the nominal dose of test item administered to the study animals.
|
616ec04ef718dffe58e3e06b
| 25 |
The stability of the test compounds (1 µM) was measured following incubation at 37 °C with hepatic microsomes (0.5 mg protein/mL for all species) in the presence of the cofactor, NADPH. Incubates were prepared in duplicate, with aliquots removed at 0, 5, 10, 20 and 40 minutes and reactions terminated and compound extracted by the addition of acetonitrile containing an analytical internal standard. The disappearance of parent compound was monitored by LC-MS/MS and the half-life determined over the time-course of incubation. The half-life values were used to calculate their in vitro intrinsic clearance expressed as µL/min/mg protein.
|
616ec04ef718dffe58e3e06b
| 26 |
The stability of test compounds (1 µM) were measured following incubation at 37 °C with cryopreserved hepatocytes in suspension at a cell density of 0.5 million cells per mL. Incubates were prepared in duplicate with aliquots removed at seven time points over a period of 120 minutes and reactions terminated and compound extracted by the addition of acetonitrile containing an analytical internal standard. The disappearance of the parent compounds were monitored by LC-MS/MS and half-life values determined over the course of the incubation. The half-life values obtained were used to calculate their in vitro intrinsic clearance expressed as µL/min/million cells.
|
616ec04ef718dffe58e3e06b
| 27 |
MDR1-MDCK cells were seeded into 24 well Transwell plates and cultured for 3 days to form monolayers. The test compounds were prepared at 10 µM in Hanks' Balanced Salt Solution containing 25 mM HEPES and loaded into the donor compartments of Transwell plates bearing the cell monolayers (pH 7.4 for both donor and receiver compartments). Lucifer Yellow was added to the apical buffer in all wells to assess integrity of the cell monolayer. Duplicate wells were prepared and incubated at 37°C in a CO2 incubator. Samples were removed at time zero and 60 minutes and test compound analysed by LC-MS/MS. Concentrations of Lucifer Yellow in the samples were measured using a fluorescence plate reader. The apparent permeability (Papp) values of test compound were determined for both the apical to basal (A>B) and basal to apical (B>A) permeation and the efflux ratio (B>A: A>B) determined.
|
616ec04ef718dffe58e3e06b
| 28 |
In the PK/PD study, adult male C57Bl/6 mice with 7-8 weeks of age were dosed orally with compound 37 as a suspension in 1% Tween 80 in 0.5% methylcellulose at 150 mg/kg BID q4h. Terminal tissue collection was conducted at 4, 6, and 10 hours after treatment (11 mice/timepoint). Mice being euthanized for the 4-hour group were euthanized before 2nd dosing. At each collection timepoint, mice were euthanized by CO2 asphyxiation and blood was collected via cardiac puncture into vials containing K + EDTA anticoagulant. Then, brains were extracted, and cerebellum dissected, separated into 2 equal parts then placed into pre-weighed 1.5ml tubes. Terminal CSF was collected for 10h treatment group only. Upon collection, all tissue samples and CSF were weighed, snap frozen in liquid nitrogen and stored at -80 °C for analysis.
|
616ec04ef718dffe58e3e06b
| 29 |
The receptor occupancy study followed very similar protocol as the PK/PD study. In addition to treatment group with compound 37 and the vehicle group, a third group of C57Bl/6 mice (n = 12, 4 at each timepoint) were dosed IV with tracer compound PGM019260 at 60 μg/kg in 10% DMSO in 0.5% 90% HP-β-CD, 20 minutes prior to the defined takedown time. Terminal tissue collection was conducted at 4, 6, and 10 hours after treatment (14 mice/timepoint). At each collection timepoint, mice were euthanized by CO2 asphyxiation and brains were extracted and dissected in cerebellum and prefrontal cortex tissue samples and placed into pre-weighed 2 ml tubes. Upon collection, all tissue samples were weighed, snap frozen in liquid nitrogen and stored at -80 °C for analysis.
|
616ec04ef718dffe58e3e06b
| 30 |
UPLC-MS was performed on a Waters Acquity I-Class with Waters Diode Array Detector coupled to a Waters SQD2 single quadrupole mass spectrometer using an Waters HSS C18 column (1.8 µm, 100 × 2.1 mm) or on a Waters DAD + Waters SQD2, single quadrupole UPLC-MS spectrometer using an Acquity UPLC BEH Shield RP18 1.7um 100 x 2.1mm (plus guard cartridge), maintained at 40 °C. The columns were initially held at 5% acetonitrile/water (with 0.1% formic acid or 10 mM ammonium bicarbonate in each mobile phase), followed by a linear gradient of 5-100% and then held at 100%.
|
616ec04ef718dffe58e3e06b
| 31 |
1 H Nuclear magnetic resonance (NMR) spectroscopy was carried out using a Bruker instrument operating at 400 MHz using the stated solvent at around room temperature unless otherwise stated. In all cases, NMR data were consistent with the proposed structures. Characteristic chemical shifts (δ) are given in parts-per-million using conventional abbreviations for designation of major peaks: e.g. s, singlet; d, doublet; t, triplet; q, quartet; dd, doublet of doublets; dt, doublet of triplets; m, multiplet; br, broad.
|
616ec04ef718dffe58e3e06b
| 32 |
Preparative HPLC purification was performed by reverse phase HPLC using a Waters Fractionlynx preparative HPLC system (2525 pump, 2996/2998 UV/VIS detector, 2767 liquid handler) or an equivalent HPLC system such as a Gilson Trilution UV directed system. The Waters 2767 liquid handler acted as both auto-sampler and fraction collector. The columns used for the preparative purification of the compounds were a Waters Sunfire OBD Phenomenex Luna Phenyl Hexyl or Waters Xbridge Phenyl at 10 µm 19 × 150 mm or Waters CSH Phenyl Hexyl, 19 × 150, 5 µm column. Appropriate focused gradients were selected based on acetonitrile and methanol solvent systems under either acidic or basic conditions. The modifiers used under acidic/basic conditions were formic acid or trifluoroacetic acid (0.1% V/V) and ammonium bicarbonate (10 mM) respectively. The purification was controlled by Waters Fractionlynx software through monitoring at 210-400 nm, and triggered a threshold collection value at 260 nm and, when using the Fractionlynx, the presence of target molecular ion as observed under API conditions. Collected fractions were analyzed by LCMS (Waters Acquity systems with Waters SQD).
|
616ec04ef718dffe58e3e06b
| 33 |
The diastereomeric separation of compounds was achieved by Supercritical Fluid Chromatography (SFC) using a Waters Thar Prep100 preparative SFC system (P200 CO2 pump, 2545 modifier pump, 2998 UV/VIS detector, 2767 liquid handler with Stacked Injection Module). The Waters 2767 liquid handler acted as both auto-sampler and fraction collector. Appropriate isocratic methods were selected based on methanol, ethanol or isopropanol solvent systems under un-modified or basic conditions. The standard SFC method used was modifier, CO2, 100 mL/min, 120 Bar backpressure, 40 °C column temperature. The modifier used under basic conditions was diethylamine (0.1% V/V). The modifier used under acidic conditions was either formic acid (0.1% V/V) or trifluoroacetic acid (0.1% V/V). The SFC purification was controlled by Waters Fractionlynx software through monitoring at 210-400 nm and triggered at a threshold collection value, typically 260 nm. Collected fractions were analyzed by SFC (Waters/Thar SFC systems with Waters SQD). The fractions that contained the desired product were concentrated by vacuum centrifugation.
|
616ec04ef718dffe58e3e06b
| 34 |
Copper iodide (0.019 g, 0.1 mmol) and bis(triphenylphosphine)Pd(II) dichloride (0.1 g, 0.15 mmol) were added and the mixture was stirred at room temperature for 18 hrs. The mixture was diluted with ethyl acetate (30 mL) and filtered through a plug of Celite TM . Water (50 mL) was added to the filtrate and the organic layer extracted. The aqueous layer was extracted with further portions of ethyl acetate (2 x 30 mL). The organic layers were combined, dried over magnesium sulfate, filtered and concentrated. The crude material was purified by flash column chromatography (5% ethyl acetate in cyclohexane isocratic) to afford the title compound as a yellow solid (0.29 g, 94% Step D: 5-(3,5-Dichlorophenethyl)-2,3-dimethoxypyrazine 10% Palladium hydroxide on carbon (0.1 g) was added to a solution of 5-((3,5-dichlorophenyl)ethynyl)-2,3dimethoxypyrazine (0.28 g, 0.9 mmol) and 1-methyl-1,4-cyclohexadiene (1.7 g, 18 mmol) and the mixture stirred at 65 °C for 3 h. A further portion of 1-methyl-1,4-cyclohexadiene (1.7 g, 18 mmol) was added and the mixture was stirred at 65 °C for a further 3 hrs. The solution was then cooled and filtered through a plug of Celite TM . The Celite TM was washed with a portion of ethyl acetate (10 mL), dichoromethane (10 mL) and methanol (10 mL). The filtrate was concentrated in vacuo and the crude residue partitioned between water (10 mL) and ethyl acetate (10 mL). The organic layer was washed with brine (10 mL) and the combined aqueous layers washed with ethyl acetate
|
616ec04ef718dffe58e3e06b
| 35 |
(1.6 g, 9 mmol) in one portion and the reaction mixture was stirred at room temperature for 48 hrs. A further portion of N-bromosuccinimide (1.6 g, 9 mmol) was added and the solution was stirred at room temperature for a further 48 hrs. The reaction mixture was concentrated under reduced pressure and the resulting crude material purified by flash column chromatography (0-30% diethyl ether in cyclohexane) to elute the desired product as a white solid (0.65 g, 21%). ¹H NMR (400 MHz, CDCl3): δ 7.70 (s, 1 H), 7.50-7.29 (m, 10 H), 5.45-5.40 (m, 4 H).
|
616ec04ef718dffe58e3e06b
| 36 |
Step E: 5-( hour. The mixture was further diluted with water (10 mL) and ethyl acetate (10 mL) and the aqueous layer extracted with ethyl acetate (3 x 40 mL). The combined organic layers were then washed with brine, dried over magnesium sulfate, filtered and concentrated. The crude material was purified by flash column chromatography (10-30% ethyl acetate in cyclohexane) to afford the title compound as a pale-yellow oil (0.35 g, 83%). ¹H NMR (400 MHz, CDCl3): δ 7. .
|
616ec04ef718dffe58e3e06b
| 37 |
To a stirred solution of phenethyl)pyrazine (0.042 g, 0.121 mmol) and Pd(dppf)2Cl2 (0.0089 g, 0.012 mmol) in toluene (1 mL) was added dimethyl zinc (0.24 mL, 2 M solution in toluene, 0.48 mmol) and the reaction mixture was heated at 80 °C for 27 hrs before a further portion of Pd(dppf)2Cl2 (0.0089 g, 0.0121 mmol) was added and the reaction was stirred at 80 °C a further 24 hrs. The mixture was cooled and water (2 mL) added. Ethyl acetate (5 mL) was added and the organic layer extracted. The aqueous layer was washed with further ethyl acetate (2 x 5 mL). The organic layers were combined, dried over magnesium sulfate, filtered and concentrated. The crude material was purified by flash column chromatography (0-5% ethyl acetate in cyclohexane)
|
616ec04ef718dffe58e3e06b
| 38 |
to afford the title compound as a pale-yellow solid (0.019 g, 48% Acetone cyanohydrin (0.015 g, 0.173 mmol) solubilized in 0.3 mL of 1-butanol was then added drop-wise over 4 hrs. After the addition was completed the reaction was stirred at room temperature for 18 hrs. Saturated NaHCO3 (2 mL) was then added and the reaction mixture was stirred for 10 min. The aqueous solution was extracted with ethyl acetate (2 x 5 mL). The organic layers were combined, dried over magnesium sulfate, filtered and concentrated. The crude material was purified by flash column chromatography (0-20% ethyl acetate in cyclohexane) to afford the title compound as a white solid (0.039 g, 80%). ¹H NMR (400 MHz, CDCl3): δ 7.56-7.52 (m, 2 H), .
|
616ec04ef718dffe58e3e06b
| 39 |
before being cooled to 0 °C. A solution of S- benzyl) methanesulfonothioate (665 mg, 2.46 mmol) in dry THF (2 mL) was added dropwise over 10 min. The reaction mixture was stirred at 0 °C for 30 minutes, allowed to warm to room temperature and stirred for 3 hrs. Saturated aqueous ammonium chloride solution (15 mL) was added followed by water (10 mL). The mixture was extracted with ethyl acetate (100 mL, 20 mL, 20 mL) and the combined organic layers were washed with brine (20 mL), dried over magnesium sulfate, filtered and concentrated under reduced pressure. The crude material was purified by flash column chromatography (10 -20% dichloromethane in cyclohexane) to yield the title compound as a yellow oil (269 mg, 39%). ¹H NMR (400 MHz, CDCl3): δ 8.15 (s, 1 H), .
|
616ec04ef718dffe58e3e06b
| 40 |
To a suspension of sodium hydride (60% in mineral oil, 188 mg, 7.81 mmol) in dry dioxane (3 mL) under nitrogen at room temperature was added dry methanol (0.32 mL, 7.81 mmol) dropwise over 10 min. The reaction mixture was stirred at room temperature for 1 hour. A solution of 2,3-dichloro-5-((4-(trifluoromethyl)benzyl)thio)pyrazine (265 mg, 0.781 mmol) in dry dioxane (2 mL) was added over 5 min. and the mixture was stirred at room temperature for 18 hrs. Saturated aqueous ammonium chloride solution (15 mL) was added followed by water (10 mL). The mixture was extracted with ethyl acetate and the combined organic layers were washed with brine (20 mL), dried over magnesium sulfate, filtered and concentrated under reduced pressure. The crude material was purified by flash column chromatography (10 -20% dichloromethane in cyclohexane) to yield the title compound as The following compounds were synthesized following the same procedure as 36. To a solution of 3,4-difluorobenzyl mercaptan (665 mg, 4.15 mmol) in dry acetonitrile (8 mL) under nitrogen was added sodium tert-butoxide (997 mg, 10.38 mmol). The mixture was stirred at room temperature for 5 min. A solution of 2-chloro-6-methoxypyrazine (500 mg, 3.46 mmol) was added and the mixture was heated at 85 °C for 90 min. The reaction mixture was cooled to room temperature, diluted with water (20 mL) and extracted with ethyl acetate (2 x 30 mL). The combined organic layers were washed with brine (20 mL), passed through a phase separation and concentrated under reduced pressure. The crude material was purified by flash column chromatography (15% ethyl acetate in cyclohexane) to yield the title compound as a red oil (708 mg, 76%). ¹H NMR (400 MHz, CDCl3): δ 8.02 (s, 1 H), 7.90 (s, 1 H), 7.26-7.20 (m, 1 H), 7.12-7.06 (m, 2 H), 4.35 (s, 2 H), 3.95 (s, 3 H).
|
616ec04ef718dffe58e3e06b
| 41 |
Step C: and sodium borohydride (690 mg, 18.3 mmol) were added and the reaction was stirred for 0.5 hrs. The reaction was brought to -20 °C and quenched with saturated ammonium chloride solution (10 mL). The reaction was diluted with ethyl acetate (20 mL). The organics were separated and the aqueous further extracted with ethyl acetate (2 x 20 mL).
|
616ec04ef718dffe58e3e06b
| 42 |
hrs. The reaction was diluted with ethyl acetate (5 mL) and water (5 mL). The organics were separated and the aqueous further extracted with ethyl acetate (2 x 5 mL). The combined organics were washed with brine, dried over sodium sulfate, filtered and concentrated under reduced pressure. The crude material was purified by flash column chromatography (0 -20% ethyl acetate in cyclohexane) to yield the title compound as a pale-yellow oil (160 mg, 92%). ¹H NMR (400 MHz, CDCl3): δ 7.64 (s, 1 H), 7.61 .
|
616ec04ef718dffe58e3e06b
| 43 |
The reaction was cooled to room temperature. The crude material was purified by flash column chromatography (0 -50% ethyl acetate in cyclohexane) to yield the title compound (0.92 g, 56% Step A: 2-Chloro-4-mercaptobenzonitrile To 2-chloro-4-fluorobenzonitrile (0.50 g, 3.2 mmol) was added sodium sulphide (0.27 g, 3.5 mmol) and in DMF (2.5 mL) and the reaction was stirred at room temperature for 2 hrs. 1 M Sodium hydroxide was added and the mixture washed with dichloromethane (5 mL). The aqueous layer was acidified to pH 1-2 with 1 M hydrogen chloride solution (2 mL) and extracted with dichloromethane (2 x 5 mL). The combined organic layers were washed with brine (5 mL), dried over magnesium sulfate, filtered and concentrated under reduced pressure to provide a crude residue. To the residue was added 10 % hydrogen chloride solution (2 mL) and the mixture was cooled with an ice-water bath. Zinc dust (2.00 g, 30.6 mmol) was added and the mixture was stirred for 1 hour. Ethyl acetate (5 mL) was added and the mixture was stirred for an additional 30 min. The organic layer was separated and washed with water (10 mL) and brine (10 mL), dried over magnesium sulfate, filtered and concentrated under reduced pressure to provide the desired product as an orange oil (0.48 g, 88% Step B: 4-(((5,6-Dimethoxypyrazin-2-yl)methyl)thio)phthalonitrile
|
616ec04ef718dffe58e3e06b
| 44 |
The organics were separated and the aqueous was extracted with dichloromethane (2 x 10 mL). The combined organics were passed through a phase separator and concentrated under reduced pressure. The crude material was purified by flash column chromatography (0 -30% ethyl acetate in cyclohexane) to yield the title compound as an off-white solid (0.14 g, 98%). ¹H NMR (400 MHz, CDCl3): δ 7.63 (d, J = 6. .
|
64e02cbd694bf1540c9921b9
| 0 |
Biomolecules exhibit dynamic behavior and adopt a distribution of conformations influenced by factors including sequence, temperature, pressure, ligand binding, and solution conditions. Traditionally, structural biology has predominantly focused on single-conformation models. However, it is broadly appreciated that most biomolecules must move to function. One of the earliest experimental demonstrations of protein structure plasticity came from NMR studies of aromatic-ring flips of the small protein bovine pancreatic trypsin inhibitor, where it was observed that conformational fluctuations allow rapid rotations of aromatic rings buried in the hydrophobic core . Recent advances in experimental and computational methods illustrate the importance of multiple-conformation modeling for understanding biomolecule functions. In particular, as the machine learning (ML) methods of AlphaFold2 (AF2) ,
|
64e02cbd694bf1540c9921b9
| 1 |
RosettaFold , OpenFold , ESMFold , RaptorX , and other advanced techniques have reached the stage where single structure prediction of small proteins is robust and reliable, and a current frontier is multiple-state modeling . Establishing consistent ontologies and formats for representing such multiple-state models is crucial for supporting and advancing this important area of structural biology. This perspective addresses the significance and handling of multiple-conformation models of biomolecules. We begin with some key definitions. Conformers refer to atomic structures capable of interconversion without making or breaking covalent bonds. Conformational "collections" are defined here as sets of these conformers. Note that this is a looser definition than the concept of "thermodynamic statistical ensemble" in statistical mechanics, which describes the Boltzmann distribution of conformations contributing to ensemble-averaged measurable parameters. Conformational "states" are distinct conformational models (or collections of conformational models) that, in principle, can be distinguished experimentally. Structural models can be categorized as "single-state" or "multiple-state" based on the nature of the experimental data or the theoretical inference. Multiple-state models may constitute a pair of conformers, pairs of conformational collections, or collections of many conformers of, for example, disordered polymer chains. The distinction between states, and the enumeration of the number of states, is determined by the interpretation used in modeling the data. Terms used for such multiple-state models in the literature include alternative conformations, multi-conformer models, switched folds, metamorphic states, chameleonic states, conformational ensembles, and conformational excited states. As various methods for interpreting experimental data in terms of multiple-state structural models are evolving, there is a current pressing need for standardized representations of such models and their corresponding data in structural databases and across the structural biology community.
|
64e02cbd694bf1540c9921b9
| 2 |
Modeling multiple conformations is fundamental to understanding biomolecule functions, as dynamics determines their ability to carry out these functions. In this perspective, we focus on representing diverse conformational states of proteins, but similar challenges also apply to nucleic acids such as DNA and RNA . Conformational dynamics underlie enzyme function and are especially important for membrane protein activities as receptors and transporters of ions, metabolites, and drugs. Protein-protein interfaces may also exhibit multiple conformational states, as observed, for example, in dimers of the influenza A virus non-structural protein NS1 and between domains of MHC class I molecules .
|
64e02cbd694bf1540c9921b9
| 3 |
Biomolecules often undergo conformational changes when interacting with binding partners or substrates, encompassing both induced fit and conformational selection mechanisms. These structural changes can occur at binding sites or be distributed across the structure. The significant role of allostery in enzyme function, where an allosteric modulator molecule binds to sites distal to the active site, has been recognized for decades . Recent advances combining experimental data with advanced modeling methods have revealed structural details of allosteric mechanisms . Evolutionary coupling (EC) based on sequence covariance analysis has also been utilized to enhance enzyme activities by perturbing allosteric networks with mutations distant from their active sites .
|
64e02cbd694bf1540c9921b9
| 4 |
Structural heterogeneity is also relevant to de novo protein design, and has been successfully used to create cyclic chameleon peptides that switch between exposed hydrophobic and hydrophilic surfaces to provide membrane permeability , two-state hinge proteins , and fold-switching metamorphic proteins . Membrane protein transporters have also been the subject of multiple-state de novo design efforts, such as the Zn 2+ -transporting four-helix bundle transmembrane protein Rocker .
|
64e02cbd694bf1540c9921b9
| 5 |
Recent reviews discuss the experimental methods that provide structural information on multiple conformational states of biomolecules . Crystallographic studies, using either X-ray or neutron diffraction, may capture different states in different crystal forms. Electron density can also often be fit to multiple conformations within a single crystal. Room-temperature (or higher temperature) X-ray crystallography may avoid structural bias from cryogenic cooling and reveal motions crucial for catalysis, ligand binding, and allosteric regulation . Other experimental data types, such as small-angle Xray scattering (SAXS) and electron microscopy (cryoEM) data, can frequently only be fit to multi-conformer models. Additionally, Förster resonance energy transfer (FRET), Double Electron-Electron Resonance (DEER) spectroscopy, and chemical cross-linking data have been used to model multiple conformational states since they can characterize interprobe distance distributions . With all experimental data, multiple-state fitting may be indicated and (cross)-validated only when a single-state model is inadequate.
|
64e02cbd694bf1540c9921b9
| 6 |
Nuclear Magnetic Resonance (NMR) spectroscopy is a valuable tool for studying the dynamic behavior of biomolecules. It employs properties such as nuclear relaxation and chemical exchange saturation transfer to determine interconversion rates and populations of conformations . NMR parameters reflect conformational averaging on parameter-specific timescales: for 1 H chemical shift, slow exchange (conformational lifetime >> ~ 1 ms) yields distinct resonances for individual states, fast exchange (conformational lifetime << ~ 1 ms) results in population-weighted average resonance frequencies, and intermediate exchange leads to characteristic resonance lineshapes. The distinction between slow and fast exchange depends on the difference in the resonance frequencies of the individual states and the relative populations; e.g., for a two-state system with 1 H chemical shift differences between 0.1 and 1 ppm at 800 MHz, intermediate exchange corresponds to rates from ca. 10 to 250,000 s -1 . Chemical shift refocusing experiments like Carr-Purcell-Meiboom-Gill (CPMG) relaxation dispersion and T1rho relaxation experiments, and saturation transfer experiments (e.g., chemical exchange by saturation transfer, CEST), can provide quantitative information about conformational exchange on the intermediate or slow chemical shift timescale and can be used to characterize sparsely-populated states that cannot otherwise be observed in NMR spectra . NOESY and residual dipolar coupling (RDC) data can reveal multiple conformations in fast dynamic equilibrium, and may provide structural restraints for modeling each state . Paramagnetic effects in metal-containing biomolecules provide ensemble-averaged distance restraints and can also determine ensemble-averaged relative orientations of structural domains . In many cases, the structures of the conformations in dynamic exchange are modeled by fitting back-calculated NMR parameters (e.g., chemical shifts) to ensemble-averaged NMR data.
|
64e02cbd694bf1540c9921b9
| 7 |
to study the various conformational states of biomolecules . Integrating experimental data, such as NMR and X-ray crystallographic data, with MD simulations has led to improved conformational heterogeneity evaluation. MD has been combined with X-ray crystallography data to generate multiplestate models that have improved fit to X-ray data compared to single-structure models . However, conventional MD simulations often fall short in capturing slower motions, particularly allosteric conformational changes. Approaches have also been developed to integrate experimental NMR data with MD simulations. By incorporating time-averaged distance restraints from NOE data , MD simulations can better model conformational distributions consistent with experimental data, as in the case of the DNA-binding loops of E. coli tryptophan repressor . Bayesian inference and ensemble fitting approaches, which leverage experimental data alongside MD simulations, can also generate multiple-state models . Similarly, chemical shift data have been used to guide or interpret computational methods. For example, NMR chemical shift perturbation analysis using programs like CHESCA has been used to characterize allosteric conformational switching upon ligand binding by chemical shift covariance analysis . Alternative conformational state modeling with Ohm, a structural perturbation propagation method, for a set of ~ 20 allosterically-modulated proteins was observed to provide excellent predictions of CHESCA-based chemical shift changes . Accelerated MD methods have also proven effective in modeling multiple conformational states of proteins, which align well with NMR data .
|
64e02cbd694bf1540c9921b9
| 8 |
Although AF2 was not trained to model protein dynamics, in some cases, it can provide information about the individual states in dynamic equilibrium. For example, comparison of AF2 and NMR models for CASP14 target T1027, Gaussia luciferase, against NMR data suggested that the AF2 prediction model corresponds to just one of the multiple conformations in the NMR sample . Subsequently, the proclivity of AF2 to model one of multiple conformational states was also reported for a collection of ~100 apo / holo protein structure pairs . Recently, methods have emerged to extend AF2 and other machine learning networks to model alternative protein conformations explicitly. These extensions involve leveraging ECs that distinguish multiple conformations , employing multiple templates with diverse conformations , using shallow multiple-sequence alignments , or by perturbing the neural network weights to generate conformational diversity. SPEACH_AF utilizes in silico mutations as input to AF2 to model conformational switching in soluble and membrane proteins .
|
64e02cbd694bf1540c9921b9
| 9 |
The advancements in experimental methods and modeling techniques for determining multiple conformational states of biomolecules necessitate improved methods for representing and archiving information about conformations in dynamic equilibrium. This can be challenging, as definitions of alternative conformational "states" depend on the timescale of the experimental data and/or the modeling methods used. Apart from the well-known Protein Data Bank (PDB) and Nucleic Acid Database (NDB), several other databases (ACMS, CoDNaS, D3PM, GLOCON, and MultiComp) primarily store and annotate data on alternative conformations obtained from diverse crystal structures, as reported in the PDB. These databases, along with other important structural databases that primarily archive single-state models but can also provide data about multiple conformational states, are listed in Table , along with their URLs.
|
64e02cbd694bf1540c9921b9
| 10 |
Clearly, multiple conformational state information is important in biology, and there is a need for consistent representation of such information in databases. In the following sections, we discuss some challenges in representing collections of molecular models derived from NMR data. While some points are specific to NMR structures and data, most are relevant for representing biomacromolecule structures obtained via various experimental or predictive modeling techniques.
|
64e02cbd694bf1540c9921b9
| 11 |
Solution NMR structures are typically represented as collections of coordinate sets, where each model in the collection is independently generated by fitting experimental data to a single conformer. This is done multiple times under different initialization conditions, generating a collection of conformers. The singlestate model representation commonly used for NMR structures encodes information about which regions of the protein structure are "well-defined" by the NMR data and which regions are not. Less-well-defined segments of the structure often (but not always) correspond to regions undergoing conformational dynamics. In the simplest case, the coordinates for the collection of conformer models are deposited to the PDB, accompanied by restraints, while the chemical shifts, peaks lists, and raw FID data are deposited to the Biological Magnetic Resonance Bank (BMRB).
|
64e02cbd694bf1540c9921b9
| 12 |
This single-state model representation can be confusing for scientists using NMR structures. The individual conformer models in the single-state representation are not meant to describe actual conformers contributing to the Boltzmann distribution of states present in the sample; rather the conformer collection provides information about the consistency of modeling. Although each conformer is considered to be a good fit to the data, the coordinate uncertainty in different regions can only be assessed by analyzing the conformer collection as a whole. The single-state model representation also does not provide a statistically reliable estimate of the atomic coordinate precision based on experimental measurement uncertainties, although Bayesian methods have been proposed for this purpose . Despite its limitations, the prevailing convention for biomolecular structure modeling using NMR data continues to be the singlestate model representation. When using such models, it is crucial to distinguish regions that are welldefined from the experimental data versus less precise regions where atomic positions are highly uncertain. This distinction is critical for the correct application of structure validation methods, which generally apply only to well-defined regions . Accordingly, it is important that the single-state model representation is conveyed in a simple way to users of NMR structures.
|
64e02cbd694bf1540c9921b9
| 13 |
In X-ray crystallography, single-state models use B-factors (which include various data quality effects) to describe the uncertainty of atomic positions. In a similar way, a single-state conformer collections generated from NMR data can be represented by a single representative conformer, along with information about coordinate uncertainty (Figure ). The wwPDB uses the medoid conformer as the representative structure, defined as the single conformer in the collection most like all the other conformers . Tools like Dihedral Angle Order Parameter (DAOP) , FindCore , and CYRANGE assess well-defined and not-well-defined residue ranges across the NMR conformer collection. The cutoffs used by these tools are based on standardized conventions. Presently, the wwPDB has adopted CYRANGE conventions to annotate well-defined residue ranges in the NMR structure validation report. PDBStat also uses these tools and writes information about well-ordered residues as well as atom-specific coordinate variances into a mmCIF (or conventional PDB format) coordinate file, allowing graphical rendering of this information onto a single representative conformer (e.g., the medoid conformer) (Figure ). After aligning the models with respect to a core atom set, coordinate uncertainties are converted using the Debye-Waller equation to effective "NMR B-factors" , indicating the variability and uncertainty in atomic positions across the conformer collection. It is unfortunate that these annotations are not more widely used compared to the widespread adoption of the analogous concept of predicted LDDT (pLDDT) scores reported for AlphaFold2 models . These annotations are essential for the informed use of NMR structures.
|
64e02cbd694bf1540c9921b9
| 14 |
Different modeling approaches are employed depending on the timescale of conformational averaging, such as slow or fast chemical shift exchange (Figure ). These include generating single-state or multistate models for each set of slowly exchanging resonances , deconvoluting information in spectra of rapidly-exchanging systems , fitting to spectral features of intermediate exchange , and matching chemical shifts of slowly-exchanging states to chemical shifts predicted from known structures . Various methods of data interpretation will generate different numbers of chemical shift lists, restraint lists, and coordinate sets, which need to be accounted for in creating the archived data structure.
|
64e02cbd694bf1540c9921b9
| 15 |
The community does not yet have a consensus scheme for deposition of multiple-state models and supporting data (e.g., FID, chemical shift, and restraint data). Figure illustrates examples of multiplestate models in the PDB. Some were deposited as pairs of separate PDB files, while in other cases, the multiple-state models were concatenated in a single PDB file. For the separate entries, although comments describing the relationships of these pairs of PDB files may be included in their header files, it is not always clear if these multiple conformations were derived from a single or multiple sets of experimental data, and information on relative populations may be lost. Conformational collections for three of the multiple-state models illustrated in Figure were deposited as a single PBD file. In two other cases (5tm0 and 2lwa), the sets (2 or 3, respectively) of structures were reported with separate chain IDs within a single PDB file. In contrast, the individual multiple-state models reported in the single PDB file 7r95
|
64e02cbd694bf1540c9921b9
| 16 |
were not distinguished by any specific designator. In X-ray crystal structures, alternative local conformations are often represented with distinct 'AltLoc IDs', the alternative location indicators, where atoms are assigned unique letters to represent different conformations . These AltLocs can range from single atoms to sets of connected or non-connected residues, and have the benefit of including relative populations of conformers as well as the positions for each atom with alternative coordinates.
|
64e02cbd694bf1540c9921b9
| 17 |
Despite progress in automating the assignment and creation of PDB files with AltLoc annotations, challenges persist regarding ease of interpretation and how these annotations are used. For some software, only one of the alternative conformations is used, and there are compatibility issues with other analysis software. In NMR studies, like those shown in Figure , multiple-state models are usually refined using multiple complete copies of the entire molecule for each state. This does not easily align with the AltLoc ID usage, which also does not currently support association with the corresponding alternative chemical shift assignments.
|
64e02cbd694bf1540c9921b9
| 18 |
To address these inconsistencies and improve data organization, it is crucial to establish standardized conventions for archiving multiple-state models and their relative populations in the PDB and other structural databases. Additionally, it is essential to ensure that the underlying experimental data, including raw FID data, are archived along with the model coordinates . This will allow for future reinterpretation of data and regeneration of models as methods improve. As illustrated in Figure , for cases of multiple coordinate sets derived from NMR data on fast-exchanging systems, there are (i) a set of raw data, (ii) a single set of chemical shifts, but potentially (iii) two (or more) sets of restraints, and (iv) two (or more) sets of atomic coordinates. In the case of multiple coordinate sets derived from slowly exchanging systems, there is again (i) a single set of data, but (ii) multiple sets of chemical shifts, one associated with each member of the slowly exchanging system, as well as potentially (iii) two (or more) sets of restraints, and (iv) two (or more) sets of atomic coordinates. This data organization required for representing multiple-state NMR structures is not currently supported by public biomolecular structure databases.
|
64e02cbd694bf1540c9921b9
| 19 |
The issues of representing single-state and multiple-state models also impact the representation of biomacromolecular structures based on cryoEM, X-ray crystallography, FRET, chemical cross-linking data, and other experimental methods. Issues of data structures needed to represent these models and data are analogous to those discussed above for solution NMR data. However, their details are beyond the scope of this perspective. Ensuring consistency in representing multiple conformational states modeled from various experimental and computational methods is critical for developing integrated structural biology methods and advancing dynamic modeling techniques.
|
64e02cbd694bf1540c9921b9
| 20 |
Standardized ontologies and formats for representing conformational collections, thermodynamic ensembles, and multiple-state models are crucial for effective communication and integration of structural data across the scientific community. Of course, the format of data deposition is inextricably related to the viewing capabilities of the software tools used to display such structures. Recent advances in experimental and computational methods, including machine learning (ML), provide exciting new opportunities for modeling and characterizing multiple conformations. Experimental distance restraint data can also be used as input for training ML-based structure prediction methods and will certainly impact ML-based methods for modeling multiple conformational states. As combined experimental and computational methods develop, models of the multiple conformational states of proteins and nucleic acids will enable biochemical, biophysical, and biological studies. The ability to consistently represent and archive information about conformations in dynamic equilibrium will facilitate research and enhance our understanding of biomolecular function. entries for NMR structures typically consist of a collection of ~20 conformers obtained through restrained modeling with NMR data. "Well-defined" parts of the structure can be determined using conventions encoded in programs like CYRANGE , FindCore2 , and dihedral angle order parameter DAOP software, and colored to indicate well-defined (blue) and not-well-defined (red) residues (left side, pdb 2kcd). The PDBStat program provides tools to create an image of the protein annotated with this information about model convergence. The medoid structure, determined by aligning models using well-defined heavy (or backbone N, Ca, C') atoms, is the first entry in a new file containing coordinates for the conformer collection in mmCIF (or conventional) PDB format. This file includes perresidue tags (q=1 for well-defined residues, q=0 for others). In addition, atom-specific coordinate variances are determined from the average atomic root mean-squared fluctuation (RMSF) across the conformer collection , and reported as effective "NMR B-factors", representing positional uncertainties across the collection. These mmCIF (or conventional) PDB format files can then be used to visualize well-defined regions (left) or atomic variances (shown schematically in three ways, top and right side) using programs like PyMOL. Representations showing atomic variances by coloring or by scaling the size of the ribbon are shown. For multiple domain structures, variance matrix analysis is used to parse the coordinates into well-defined units, which are then analyzed separately .
|
64e02cbd694bf1540c9921b9
| 21 |
, inhibitor-bound dengue virus NS2B/NS3 protease (two binding modes, 2m9p and 2m9q) , pro-islet amyloid polypeptide in detergent micelles (6ucj and 6uck) , E. coli tryptophan repressor (two states combined as pdb ID 5tm0) , and the membrane-bound SARS-CoV-2 spike protein HR1 ectodomain (two states combined as 7r95) . Also shown is the three-state model of influenza hemagglutinin fusion peptide A (combined as 2lwa) (C and D). Schematic representations for the data organization required for deposition of multiple-state models into the PDB and BMRB: (C) Fast or intermediate exchange between the conformers yields population-averaged chemical shifts, resulting in one or more sets of restraints and corresponding PDB coordinate sets. (D) Slow exchange between conformers leads to distinct chemical shifts, with multiple chemical shift entries from a single NMR dataset that are then used to calculate multiple sets of PDB coordinates. As in the fast/intermediate exchange case, a single set of chemical shifts arising from one slow-exchange ensemble may generate multiple restraints, leading to multiple coordinate sets, depending on the data analysis method. Multiple-state models may also be generated from a single restraint set in certain cases.
|
65771ebafd283d7904c36eda
| 0 |
dataset is generated, its further usage in applications such as machine learning tends to be limited due to the inherent imbalance of HTS screens, with inactive samples far outnumbering active ones. This stark imbalance biases machine learning algorithms to the inactive class, reducing their potential to identify further active compounds via machine learning. To avoid this issue, identifying the most important samples in an HTS dataset could be used to undersample and balance the datasets. To tackle all these limitations in HTS pipelines, we introduce a variety of machine learning data valuation approaches to the field of drug development pipelines. Data valuation methods estimate each training instance's impact on the optimum found by a machine learning or deep learning model. Some samples carry high amounts of information helpful for a machine learning model, some samples carry very little useful information, and some samples actually carry harmful information that actively worsens the models performance. Data valuation methods can identify these importance scores using various machine learning models and data valuation concepts. Some methods calculate the importance of a training sample on the prediction of a separate test set, while other methods score a training sample's importance on its own prediction. These compound-specific importance scores can be utilized to guide screening efforts, identify false positives, and remove uninformative compounds, tackling significant hurdles in current HTS pipelines. Within this study, we benchmarked gradient boosting-based influence functions such as MVS-A and CatBoost's object importance, the deep neural network-based approach TracIn, as well as data valuation using reinforcement learning (DVRL), and a KNN approximation of the cooperative game theory concept of Shapley values. To evaluate these data valuation approaches for the application of active learning and false positive detection, we curated a selection of 25 HTS datasets with different sizes, class imbalance, biological targets, assay technology, and false positive rates based on the Multifidelity PubChem Bioassay (MF-PCBA) benchmark. Additionally, to analyze the data valuation approach for undersampling, the MolData repository was utilized. Our results highlight the potency of data valuation methods to identify informative compounds in HTS applications effectively. The implications of these developments are far-reaching, offering the potential to expedite the drug discovery pipeline and to make substantial contributions to pharmaceutical research. This study provides a critical step forward in integrating machine learning with HTS, potentially transforming the drug development landscape. Therefore, to allow the scientific community to apply and improve upon the methods and applications used in this study, we created a benchmarking platform that allows straightforward application and modification of the data valuation methods to a wide range of datasets. The platform provides a modular toolbox for dataset loading and processing, allowing easy integration of your own data, as well as the base data valuation methods and applications used herein.
|
65771ebafd283d7904c36eda
| 1 |
Data valuation pertains to assessing and quantifying the value or impact of individual data points within a training dataset. This assessment is crucial, as not all data contribute equally to the learning process or the performance of a machine learning model. Some data points may carry a wealth of information, significantly enhancing model accuracy or aiding in generalization. Conversely, others may contribute little to no useful information, or worse, negatively impact model performance due to noise, outliers, or biases. In the context of machine learning, data valuation methods serve several key functions. Firstly, they allow for the identification of the most informative and influential data points, which can be leveraged to optimize the training process. This is particularly useful in scenarios where data acquisition or processing is costly, enabling a more efficient use of resources. Secondly, data valuation techniques can aid in diagnosing and improving model robustness by identifying and mitigating the impact of harmful or misleading data. Finally, in applications like active learning, data valuation can guide the selection of data points most likely to improve model performance when labeled, thus streamlining the learning process.
|
65771ebafd283d7904c36eda
| 2 |
KNN Shapley is an approximation of the cooperative game theory concept of Shapley Values (SVs). Classical SVs are calculated by computing the utility of a machine learning model of every possible subset of the training set with and without a sample of interest, calculating the difference, and averaging it over all subsets. However, this calculation is impractical for large datasets, as the number of subset calculations scales exponentially. KNN Shapley applies the concept of SVs only within the neighborhood of a validation set sample. As a result, many subsets can be ignored when calculating Shapley values, leading to greater computational efficiency. When applying the KNN Shapley values to a binary classification problem, the minority class samples are expected to be assigned higher values, as they are less prevalent, and their effect on the subset prediction is, therefore, greater. This is especially true for HTS data due to the high imbalance. Within the active class, true positives tend to be more influential than false positives, as their influence is computed on a separate test set. A true positives' absence worsens the prediction of true positives in the test set due to their likely similarity. In contrast, the absence of false positives does not negatively impact the test set prediction, resulting in false positives being assigned lower Shapley values.
|
65771ebafd283d7904c36eda
| 3 |
CatBoost Object Importance implements another approach for importance calculations on a test sample prediction called FastLeafInfluence. Its foundation, LeafInfluence, approximates the concept of Leave-One-Out by changing the weights of the leaves where a training sample would end up instead of leaving it out and retraining. LeafInfluence calculates the change in loss on the prediction of a test sample dependent on the perturbation of the weights of the leaves associated with the training sample of interest. The FastLeafInfluence scores of HTS data can be interpreted analogously to KNN Shapley.
|
65771ebafd283d7904c36eda
| 4 |
Data Valuation using Reinforcement Learning (DVRL) utilizes Reinforcement Learning, a subfield of machine learning where an agent learns by interacting with an environment over time. At each step, the agent chooses an action to interact with the environment and is given feedback, either a reward if the action was favorable towards the agent's goal or a punishment otherwise. DVRL uses this general framework to compute influence values for the training data. The agent, here the data value estimator, is fed with a set of training samples and assigns each training sample a sampling probability. The higher a training sample's assigned value, the more likely it will be sampled and fed to the environment, in this case, the predictor. The predictor is trained on the sampled training instances and evaluated on a separate validation set. The reward function then sends a reward or punishment to the agent, dependent on the performance change compared to previous iterations. Over time, the data value estimator learns which samples will result in the maximum reward, assigning them the highest values. These optimal values can be interpreted as the importance scores. Their interpretation is analogous to the previous methods.
|
65771ebafd283d7904c36eda
| 5 |
Tracing Gradient Descent (TracIn) implements a method of tracing training sample influence on the change in loss on a test sample when training deep neural networks. In an idealized notion, TracIn would quantify the change in loss on a test sample while updating the weights of one training sample. However, this idealized notion is impractical, due to its necessity to know the test sample at training time and the fact that gradient descent is usually done on a group of training samples at a time, not on individual samples. To adapt to these limitations, a model's state can be saved using checkpoints, exporting the model's weights at a certain point during training. The influence score is then calculated by the dot product of the loss gradients of training and test samples, weighted by the learning rate, and summed across all checkpoints. However, TracIn can be further adapted to compute self-influence, meaning the influence of a training sample on its own prediction. For this, the training sample gradients are only multiplied by themselves, denoting one training sample's impact on its own prediction. As mentioned previously, actives tend to have higher influence values than inactives, as their low abundance causes high gradients for their own prediction. However, unlike previous data valuation approaches, TracIn will assign higher influence values for false positives than true positives. This is because the influence calculation, regardless of test-or self-influence, always includes a training sample's gradient on its own prediction. Most true positives can be predicted without their own presence in the training set due to their similarity to other true positives, resulting in comparatively low self-influence scores. False positives, on the other hand, have very large gradients on their own prediction, as without their presence in the training set, they will likely be classified as inactive. Minimal Variance Sampling Analysis (MVS-A) is another GBM-based data valuation approach developed to compute importance scores for training samples during training, similar to TracIn. However, TracIn's approach of tracing gradients dependent on their change of weights is not directly applicable to GBM models, as they only have leaf weights. CatBoost's feature importance calculations use these leaf weights. MVS-A, instead, hypothesizes that monitoring changes in the decision tree structure provides better inductive bias for anomaly detection. Therefore, MVS-A adapts the original TracIn approach to use Minimal Variance Sampling (MVS) probabilities to define influence. MVS was originally developed to optimize tree splits during GBM model construction, sampling the most important samples to most accurately represent the entire dataset. This fits well as a notion of importance. Additionally, MVS-A treats each tree of a gradient boosting model as a checkpoint, similar to the checkpoints in TracIn. The scores are conceptually similar to the TracIn self-influence approach, as, similarly to TracIn, the loss function gradient and hessian of each sample are multiplied by itself, quantifying the change on their own prediction. MVS-A's importance scores are interpreted analogously to TracIn.
|
65771ebafd283d7904c36eda
| 6 |
Our study adopts an active learning approach for screening compound libraries, focusing on identifying active compounds more efficiently. Traditionally, active learning involves iteratively validating small batches and using a machine learning model to predict actives in the remaining library Figure (upper panel). However, this method typically favors samples likely to be active, which can skew the model's learning. To improve this, we integrate data valuation techniques to identify key samples in each batch and use a regression model to estimate the importance of the remaining library samples. This strategy ensures the selection of both important active and inactive samples for subsequent batches, fostering a balanced dataset that enhances the model's accuracy and understanding of class boundaries. This refined approach streamlines the discovery of active compounds and bolsters the overall learning process. To assess the potential of this pipeline, its ability to identify active samples in a library was compared to the performance of the state-of-the-art greedy approach, which was implemented using a LightGBM Classifier. This comparison and subsequent optimization were done on one dataset that represented an average over all datasets regarding size and imbalance. The initial results showed comparable performance for one data valuation method, MVS-A, while all methods did not show comparable results Figure .
|
65771ebafd283d7904c36eda
| 7 |
The greedy approach clearly outperforms random sampling at each iteration, identifying over 50% of all actives present in the dataset after screening only 10.5% of its samples. Data Valuation Using Reinforcement Learning (DVRL) performs similarly to random sampling. This might be due to imperfect model optimization, as DVRL is a very complex approach involving two separate machine learning models with many important parameters that allow tuning. However, parameter tuning in a real-world application might not be possible, as no clean training set would be present. Therefore, these influence methods have to perform well with little to no optimization to allow application on any HTS data. KNN Shapley performs similarly to DVRL. This is likely due to the limitations of a KNN algorithm when given a very small part of the compound library, as KNN algorithms perform poorly at generalization on small training sets.
|
65771ebafd283d7904c36eda
| 8 |
TracIn performs slightly better than KNN Shapley but is still outperformed by MVS-A and Cat-Boost. This is likely due to the different underlying machine learning models. While TracIn uses deep neural networks (DNNs) to identify important samples, MVS-A and CatBoost use gradient boosting algorithms. The discrepancy in the active learning performance is likely caused by the low prediction performance of DNNs, given such a small initial training set. Meanwhile, gradient boosting algorithms can already show strong predictions even with a small training set size. The discrepancy between MVS-A and CatBoost can be attributed to their difference in computing importance given the gradient boosted trees. While MVS-A focuses on the influence of a sample on the tree structure, CatBoost uses a sample's influence on leaf weights. These results suggest that analyzing splitting decisions provides more accurate influence scores compared to leaf weight analysis. Furthermore, these results further reaffirm that methods utilizing GBMs outperform deep learning approaches on small training sets. We hypothesized the bottleneck could be the importance score prediction of the separate regression model.
|
65771ebafd283d7904c36eda
| 9 |
Consequently, various regression models were tested in the influence active learning pipeline on the same dataset as before, always using MVS-A for influence calculation, as MVS-A has shown to have the best active learning performance (Fig. ). Among them, the sklearn implementation of a Gaussian process regressor showed the best active learning performance, improving the overall pipeline. Finally, the optimized pipeline was benchmarked against the state-of-the-art on the remaining 24 datasets. As seen in Figure , the influence-based active learning approach significantly outperforms the greedy benchmark at all steps across all datasets.
|
65771ebafd283d7904c36eda
| 10 |
This improvement amounts to retrieving, on average, 8% to 12% more actives, equivalent to approximately 200 to 300 additional active samples, dependent on the sampling percentage. Overall, the influence-based active learning approach retrieves, on average, 45.7% ± 0.2% of all actives within evaluating 10.5% of the datasets. This performance improvement is likely due to the influence-based method sampling a more diverse set of compounds, also targeting influential inactive samples, improving the model's prediction in the subsequent step. However, it is noteworthy that the state-of-the-art approach surpasses the influence-based method significantly in terms of computational speed, primarily due to the adoption of the GPR. As shown in Figure , when coupled with a rapid regression model like LightGBM, the influence-based approach performs similarly to the benchmark greedy approach. The computational time increase can largely be attributed to the GPR regression. These results are highly promising, showcasing the potential of influence-based methods to enhance active learning strategies. They offer a pathway for researchers to streamline their drug discovery pipelines, potentially revolutionizing the efficiency of the process.
|
65771ebafd283d7904c36eda
| 11 |
Next, the data valuation functions were applied to the problem of false and true positive prediction. Given an HTS dataset, the data valuation method of choice was trained exclusively with the primary screening data (Figure ). To estimate the effectiveness in separating true and false positives, the resulting importance ranking was compared to the confirmatory screen data. Performance was measured using relative top-10 precision, enrichment factor, and BEDROC scores. PAINS generally performed poorly regarding false positive prediction, as seen in Figure , performing worse than random. This is probably due to the limited number of fragments present in the filters, which only detect a tiny subset of false positives. Additionally, PAINS might detect some frequent hitters, which would be considered true positives in this approach, as these compounds would be active in the confirmatory screen. Lastly, PAINS produces simple false or true positive labels instead of a ranking, unlike the influence methods, making the performance metrics less suited for PAINS evaluation. Due to the low performance of the fragment filter approach, all influence methods show significant performance improvements. However, clear differences can be seen between the different approaches. KNN Shapley, arguably the simplest approach out of all influence methods, as it is approximated by a KNN with K = 5 and uses subsampling of the training set, already shows an improvement of around 10% in average relative precision over random sorting. When looking at individual datasets, KNN Shapley performs best regarding relative precision for 6 out of 25 datasets. KNN Shapley is the only influence method that, using the validation set for influence prediction, effectively outperforms random sorting. The reason for that is likely the small number of neighbors utilized in this approach. In general, false positives should have a negative impact on the prediction of inactive samples, as they might wrongfully suggest activity for a given structure. For influence methods that use a large number of samples, such as DVRL or CatBoost, the effect of a single false positive on the prediction of an inactive sample tends to be very small, as thousands of other correctly labeled inactives define a new sample's activity. For KNN with five neighbors, however, a false positive provides 20% of the information the model has to classify the new sample, therefore having a much larger impact on the classification of inactives compared to other methods. Especially in strongly imbalanced datasets, the presence of just one active sample in a neighborhood can suggest activity for a new sample. The false positive's removal from that neighborhood will positively impact the prediction of a new inactive sample, thereby assigning the false positive a much larger negative influence compared to other models, where the impact of one false positive is undermined by the large number of correctly labeled inactives. This is especially impressive when factoring in the computational costs of the different influence methods. KNN Shapley vastly outperforms the other methods when it comes to computation time, as seen in Figure .
|
65771ebafd283d7904c36eda
| 12 |
TracIn performs very well, both compared to PAINS and random sorting, with a performance increase in relative precision of around 14%. TracIn only performs best on three datasets regarding relative precision but is almost always close to the best-performing method. Interestingly, the performance is very similar between all three TracIn approaches across all metrics. This is expected for TracIn and TracIn pos, as their only difference is the removal of inactive samples from the validation set. Surprisingly, TracIn-self performs equally well. This behavior is likely caused by the fact that all TracIn approaches use self-prediction gradients. The only difference is that TracIn-self squares the self-gradients, while the validation set approaches multiply the self-gradients with the gradients on all validation set predictions. Still, if the majority of the influence score stems from the self-gradients, TracIn-self performs so well because it analyzes the effect of removing a training sample on its own prediction. While correctly labeled samples are not reliant on their own presence in the training set, as their label can be inferred from the other training samples, false positives, due to their discrepancy between label and actual activity, cannot be predicted using other training samples. Therefore, its gradient on its own prediction is very high, which also explains why all TracIn methods perform similarly, as the self-influence gradient dominates the influence score. This effect is especially exciting as it allows influence calculation, and therefore false positive prediction, without the need of a validation set, as TracIn-self already performs equally well or better than the validation set approaches. In a practical application, this allows false positive detection directly on primary HTS data without validating a small subset in the laboratory first. Regarding computational time, TracIn takes comparatively long, with around 270 seconds per replicate. This time is mainly spent on training the deep neural network and exporting the weights at every checkpoint. It is important to mention that once the FNN is trained and checkpoints are exported, all three TracIn methods can be applied to those saved checkpoints. The TracIn computations are relatively short, with only a few seconds, meaning the difference in calculating one approach or all three is negligible compared to the FNN training time. This behavior can also be seen in Figure ), showing a noticeable increase in computational time for transcription 5 due to the increased training time of the FNN. CatBoost, unlike TracIn, shows a substantial difference in performance between CatBoost-self and CatBoost-test across all metrics. However, it is important to mention that the implementation of CatBoost-self does not rely on the influence of one sample solely on its own prediction, as done in TracIn-self, but on all primary actives. Figure ) shows that CatBoost-test actually performs worse than random sorting for almost half of all datasets, while CatBoost-self almost always scores above 0.0. The reason CatBoost-test performs so poorly is likely the same as for the performances of KNN Shapley and DVRL, with the influence on inactive prediction being too low to be detectable. CatBoost-self, however, allows false positive prediction with similar performances compared to KNN Shapley, but without the need for a validation set. This might be because removing false positives close to activity cliffs increases prediction performance for correctly labeled active samples, resulting in low influence scores for false positives in the CatBoost-self approach. This effect might be lost in noise when calculating the influence of a false positive on all samples, which is why it is only detectable via CatBoost-self but not with CatBoost-test. While CatBoost-self does have longer computational time than KNN Shapley, as seen in Figure , it vastly outperforms TracIn due to its usage of gradient boosting instead of FNNs. MVS-A shows the best overall false positive prediction performance across all metrics, most notably in relative precision, where it performs best for 13 out of 25 datasets. However, TracIn and MVS-A generally perform very similarly. This is because they conceptually measure the same effect via different implementations. Both identify the impact of a given training sample on its own prediction, which works well for false positive detection for reasons explained previously. MVS-A has a slight edge over TracIn regarding performance, which might be because detecting changes in tree structure allows for better influence score assignments compared to gradient changes in the last two layers of an FNN. The results shown in Figure were calculated using ECFPs as molecular representation. Figure shows the same experiments using RDKit descriptors, which resulted in very similar results. Overall, all influence methods significantly outperformed the state-of-the-art fragment filter approach, and most methods displayed substantial false positive prediction improvements over random sorting.
|
65771ebafd283d7904c36eda
| 13 |
Previously, false positives were identified either via their high influence on their own prediction or their low influence on the prediction of other samples, dependent on the influence method. Conversely, the opposite approach can be used to identify true positives. True positive samples will have a high influence on the prediction of other samples, as the machine learning model uses these true positives to recognize patterns. Their self-influence, however, will be much lower than false positives, as their activity can also be inferred from other true positives, and a true positive's prediction is therefore not strongly dependent on its own presence in the training set.
|
65771ebafd283d7904c36eda
| 14 |
Ranking actives according to their PubChem activity score is used as a benchmark method for true positive detection, and all significance tests are done compared to the Score method. Similar to false positive precision, the true positive precision is displayed as relative precision, where the true positive rates were subtracted from the precision scores prior to plotting. The PubChem activity score method outperforms random sorting on average when looking at relative true positive precision in Figure ). Figure ) shows that the Score method actually outperforms all other methods on 5 out of 25 datasets, justifying its usage for HTS true positive detection. The influence functions behave very similarly to the false positive prediction results. Their computational time is also independent of true or false positive prediction, as the label of interest does not change the methodology of the influence approaches.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.