Hugging Face's logo

neuralbioinfo
/
prokbert-mini-long

Fill-Mask
Transformers
PyTorch
megatron-bert
prokbert
bioinformatics
genomics
sequence embedding
genomic language models
nucleotide
dna-sequence
promoter-prediction
phage
Inference Endpoints
Model card Files Community
1
prokbert-mini-long / README.md
ligeti's picture
ligeti
Update README.md
6d78479 verified
preview code
|
raw
history blame contribute delete
15.6 kB
metadata
license: cc-by-nc-4.0
tags:
  - prokbert
  - bioinformatics
  - genomics
  - sequence embedding
  - genomic language models
  - nucleotide
  - dna-sequence
  - promoter-prediction
  - phage

ProkBERT-mini Model

ProkBERT-mini-long (also prokbert-mini-k6s2) is part of the ProkBERT family of genomic language models, specifically designed for microbiome applications. This model model can provide extended context size up to 4kb sequence by implementing the LCA tokenization with k-mer=6 and shift 2. This model showed comparable performance compare to other family member models.

Simple Usage Example

The following example demonstrates how to use the ProkBERT-mini-long model for processing a DNA sequence:

from transformers import MegatronBertForMaskedLM
from prokbert.prokbert_tokenizer import ProkBERTTokenizer

# Tokenization parameters
tokenization_parameters = {
    'kmer': 6,
    'shift': 2
}
# Initialize the tokenizer and model
tokenizer = ProkBERTTokenizer(tokenization_params=tokenization_parameters, operation_space='sequence')
model = MegatronBertForMaskedLM.from_pretrained("neuralbioinfo/prokbert-mini-long")
# Example DNA sequence
sequence = 'ATGTCCGCGGGACCT'
# Tokenize the sequence
inputs = tokenizer(sequence, return_tensors="pt")
# Ensure that inputs have a batch dimension
inputs = {key: value.unsqueeze(0) for key, value in inputs.items()}
# Generate outputs from the model
outputs = model(**inputs)

Model Details

Developed by: Neural Bioinformatics Research Group

Architecture: ProkBERT-mini-k6s1 is based on the MegatronBert architecture, a variant of the BERT model optimized for large-scale training. The model employs a learnable relative key-value positional embedding, mapping input vectors into a 384-dimensional space.

Tokenizer: The model uses a 6-mer tokenizer with a shift of 2 (k6s2), specifically designed to handle DNA sequences efficiently.

Parameters:

Parameter Description
Model Size 26.6 million parameters
Max. Context Size 4096 bp
Training Data 206.65 billion nucleotides
Layers 6
Attention Heads 6

Intended Use

Intended Use Cases: ProkBERT-mini-long is intended for bioinformatics researchers and practitioners focusing on genomic sequence analysis, including:

  • sequence classification tasks
  • Exploration of genomic patterns and features

Segmentation and Tokenization in ProkBERT Models

Preprocessing Sequence Data

Transformer models, including ProkBERT, have a context size limitation. ProkBERT's design accommodates context sizes significantly larger than an average gene but smaller than the average bacterial genome. The initial stage of our pipeline involves two primary steps: segmentation and tokenization. For more details about tokenization, please see the following notebook: Tokenization Notebook in Google Colab.

For more details about segmentation, please see the following notebook: Segmentation Notebook in Google Colab.

Segmentation

Segmentation is crucial for Genomic Language Models (GLMs) as they process limited-size chunks of sequence data, typically ranging from 0 to 4kb. The sequence is divided into smaller parts through segmentation, which can be either contiguous, splitting the sequence into disjoint segments, or random, involving randomly sampling segments of length L.

The first practical step in segmentation involves loading the sequence from a FASTA file, often including the reverse complement of the sequence.

Segmentation process:
Segmentation Process

Tokenization Process

After segmentation, sequences are encoded into a vector format. The LCA method allows the model to use a broader context and reduce computational demands while maintaining the information-rich local context.

Basic Steps for Preprocessing:

  1. Load Fasta Files: Begin by loading the raw sequence data from FASTA files.
  2. Segment the Raw Sequences: Apply segmentation parameters to split the sequences into manageable segments.
  3. Tokenize the Segmented Database: Use the defined tokenization parameters to convert the segments into tokenized forms.
  4. Create a Padded/Truncated Array: Generate a uniform array structure, padding or truncating as necessary.
  5. Save the Array to HDF: Store the processed data in an HDF (Hierarchical Data Format) file for efficient retrieval and use in training models.

Installation of ProkBERT (if needed)

For setting up ProkBERT in your environment, you can install it using the following command (if not already installed):

try:
    import prokbert
    print("ProkBERT is already installed.")
except ImportError:
    !pip install prokbert
    print("Installed ProkBERT.")

Training Data and Process

Overview: The model was pretrained on a comprehensive dataset of genomic sequences to ensure broad coverage and robust learning.

Training Process:

  • Masked Language Modeling (MLM): The MLM objective was modified for genomic sequences for masking overlapping k-mers.
  • Training Phases: The model underwent initial training with complete sequence restoration and selective masking, followed by a succeeding phase with variable-length datasets for increased complexity.

Evaluation Results for ProkBERT-mini-long

Model L Avg. Ref. Rank Avg. Top1 Avg. Top3 Avg. AUC
ProkBERT-mini-long 128 3.9432 0.2164 0.4781 0.9991
ProkBERT-mini-long 256 3.5072 0.2470 0.5258 0.9992
ProkBERT-mini-long 512 3.3026 0.2669 0.5435 0.9992
ProkBERT-mini-long 1024 3.2082 0.2768 0.5589 0.9992

Masking performance of the ProkBERT family.

Evaluation of Promoter Prediction Tools on E-coli Sigma70 Dataset

Tool Accuracy MCC Sensitivity Specificity
ProkBERT-mini 0.87 0.74 0.90 0.85
ProkBERT-mini-c 0.87 0.73 0.88 0.85
ProkBERT-mini-long 0.87 0.74 0.89 0.85
CNNProm 0.72 0.50 0.95 0.51
iPro70-FMWin 0.76 0.53 0.84 0.69
70ProPred 0.74 0.51 0.90 0.60
iPromoter-2L 0.64 0.37 0.94 0.37
Multiply 0.50 0.05 0.81 0.23
bTSSfinder 0.46 -0.07 0.48 0.45
BPROM 0.56 0.10 0.20 0.87
IBPP 0.50 -0.03 0.26 0.71
Promotech 0.71 0.43 0.49 0.90
Sigma70Pred 0.66 0.42 0.95 0.41
iPromoter-BnCNN 0.55 0.27 0.99 0.18
MULTiPly 0.54 0.19 0.92 0.22

The ProkBERT family models exhibit remarkably consistent performance across the metrics assessed. With respect to accuracy, all three tools achieve an impressive

Metric ProkBERT-mini ProkBERT-mini-c ProkBERT-mini-long Promotech Sigma70Pred iPromoter-BnCNN MULTiPly
Accuracy 0.81 0.79 0.81 0.61 0.62 0.61 0.58
F1 0.81 0.78 0.81 0.43 0.58 0.65 0.58
MCC 0.63 0.57 0.62 0.29 0.24 0.21 0.16
Sensitivity 0.81 0.75 0.79 0.29 0.52 0.66 0.57
Specificity 0.82 0.82 0.83 0.93 0.71 0.55 0.59

Promoter prediction performance metrics on a diverse test set. A comparative analysis of various promoter prediction tools, showcasing their performance across key metrics including accuracy, F1 score, MCC, sensitivity, and specificity.

Evaluation on phage recognition benchmark

method L auc_class1 acc f1 mcc recall sensitivity specificity tn fp fn tp Np Nn eval_time
DeepVirFinder 256 0.734914 0.627163 0.481213 0.309049 0.345317 0.345317 0.909856 4542 450 3278 1729 5007 4992 7580
DeepVirFinder 512 0.791423 0.708 0.637717 0.443065 0.521192 0.521192 0.889722 4510 559 2361 2570 4931 5069 2637
DeepVirFinder 1024 0.826255 0.7424 0.702678 0.505333 0.605651 0.605651 0.880579 4380 594 1982 3044 5026 4974 1294
DeepVirFinder 2048 0.853098 0.7717 0.743339 0.557177 0.6612 0.6612 0.8822 4411 589 1694 3306 5000 5000 1351
INHERIT 256 0.75982 0.6943 0.67012 0.393179 0.620008 0.620008 0.76883 3838 1154 1903 3105 5008 4992 2131
INHERIT 512 0.816326 0.7228 0.651408 0.479323 0.525248 0.525248 0.914973 4638 431 2341 2590 4931 5069 2920
INHERIT 1024 0.846547 0.7264 0.659447 0.495935 0.527059 0.527059 0.927825 4615 359 2377 2649 5026 4974 3055
INHERIT 2048 0.864122 0.7365 0.668595 0.518541 0.5316 0.5316 0.9414 4707 293 2342 2658 5000 5000 3225
MINI 256 0.846745 0.7755 0.766462 0.552855 0.735623 0.735623 0.815505 4071 921 1324 3684 5008 4992 6.68888
MINI 512 0.924973 0.8657 0.859121 0.732696 0.83046 0.83046 0.89998 4562 507 836 4095 4931 5069 16.3681
MINI 1024 0.956432 0.9138 0.911189 0.829645 0.879825 0.879825 0.94813 4716 258 604 4422 5026 4974 51.3319
MINI-C 256 0.827635 0.7512 0.7207 0.51538 0.640974 0.640974 0.861779 4302 690 1798 3210 5008 4992 7.33697
MINI-C 512 0.913378 0.8466 0.834876 0.69725 0.786453 0.786453 0.905109 4588 481 1053 3878 4931 5069 17.6749
MINI-C 1024 0.94644 0.8937 0.891564 0.788427 0.869479 0.869479 0.918175 4567 407 656 4370 5026 4974 54.204
MINI-LONG 256 0.777697 0.71495 0.686224 0.437727 0.622404 0.622404 0.807792 8065 1919 3782 6234 10016 9984 6.10304
MINI-LONG 512 0.880831 0.81405 0.798001 0.632855 0.744879 0.744879 0.881338 8935 1203 2516 7346 9862 10138 12.1307
MINI-LONG 1024 0.9413 0.88925 0.884917 0.781465 0.847195 0.847195 0.931745 9269 679 1536 8516 10052 9948 30.5088
MINI-LONG 2048 0.964551 0.929 0.927455 0.85878 0.9077 0.9077 0.9503 9503 497 923 9077 10000 10000 94.404
Virsorter2 512 0.620782 0.6259 0.394954 0.364831 0.247617 0.247617 0.993884 5038 31 3710 1221 4931 5069 2057
Virsorter2 1024 0.719898 0.7178 0.621919 0.51036 0.461799 0.461799 0.976478 4857 117 2705 2321 5026 4974 3258
Virsorter2 2048 0.816142 0.8103 0.778724 0.647532 0.6676 0.6676 0.953 4765 235 1662 3338 5000 5000 5737

Column Descriptions

  • method: The algorithm or method used for prediction (e.g., DeepVirFinder, INHERIT).
  • L: Length of the genomic segment.
  • auc_class1: Area under the ROC curve for class 1, indicating the model's ability to distinguish between classes.
  • acc: Accuracy of the prediction, representing the proportion of true results (both true positives and true negatives) among the total number of cases examined.
  • f1: The F1 score, a measure of a test's accuracy that considers both the precision and the recall.
  • mcc: Matthews correlation coefficient, a quality measure for binary (two-class) classifications.
  • recall: The recall, or true positive rate, measures the proportion of actual positives that are correctly identified.
  • sensitivity: Sensitivity or true positive rate; identical to recall.
  • specificity: The specificity, or true negative rate, measures the proportion of actual negatives that are correctly identified.
  • fp: The number of false positives, indicating how many negative class samples were incorrectly identified as positive.
  • tp: The number of true positives, indicating how many positive class samples were correctly identified.
  • eval_time: The time taken to evaluate the model or method, usually in seconds.

Ethical Considerations and Limitations

As with all models in the bioinformatics domain, ProkBERT-mini-k6-s1 should be used responsibly. Testing and evaluation have been conducted within specific genomic contexts, and the model's outputs in other scenarios are not guaranteed. Users should exercise caution and perform additional testing as necessary for their specific use cases.

Reporting Issues

Please report any issues with the model or its outputs to the Neural Bioinformatics Research Group through the following means:

Reference

If you use ProkBERT-mini in your research, please cite the following paper:

@ARTICLE{10.3389/fmicb.2023.1331233,
    AUTHOR={Ligeti, Balázs and Szepesi-Nagy, István and Bodnár, Babett and Ligeti-Nagy, Noémi and Juhász, János},
    TITLE={ProkBERT family: genomic language models for microbiome applications},
    JOURNAL={Frontiers in Microbiology},
    VOLUME={14},
    YEAR={2024},
    URL={https://www.frontiersin.org/articles/10.3389/fmicb.2023.1331233},
    DOI={10.3389/fmicb.2023.1331233},
    ISSN={1664-302X},
    ABSTRACT={...}
}