--- license: cc-by-nc-sa-4.0 --- # ProkBERT PhaStyle **Model Name**: neuralbioinfo/PhaStyle-mini **Model Type**: Genomic Language Model (ProkBERT-based) **Model Description**: ProkBERT PhaStyle is a fine-tuned genomic language model designed for phage lifestyle prediction. It classifies phages as either **virulent** or **temperate** directly from nucleotide sequences. The model is based on ProkBERT architecture and was trained on the **BACPHLIP dataset**, excluding *E. coli* sequences By leveraging transfer learning, ProkBERT PhaStyle is optimized for handling **fragmented sequences**, commonly encountered in metagenomic and metavirome datasets. The model provides a fast, efficient alternative to traditional methods without requiring complex preprocessing pipelines or curated databases. ### Key Points: - **Trained on BACPHLIP** dataset excluding *E. coli* sequences. - **Segment Length** for training: 512 base pairs. - **Output**: Binary classification (virulent or temperate). - **Model Parameters**: ~25 million parameters. --- ## Intended Use ProkBERT PhaStyle is designed for phage lifestyle prediction tasks, suitable for: - **Phage Therapy**: Identifying virulent phages for bacterial infection treatment. - **Microbiome Engineering**: Understanding the interaction between temperate and virulent phages in various microbiomes. - **Metagenomic Studies**: Classifying fragmented phage sequences from environmental or clinical samples. ### Inference Code ProkBERT PhaStyle requires the **ProkBERT tokenizer** and a **custom classification model** (`BertForBinaryClassificationWithPooling`). Below is a high-level overview of how to use the model in inference mode: ```python aaa ``` ```bash python bin/PhaStyle.py \ --fastain data/EXTREMOPHILE/extremophiles.fasta \ --out output_predictions.tsv \ --ftmodel neuralbioinfo/PhaStyle-mini \ --modelclass BertForBinaryClassificationWithPooling \ --per_device_eval_batch_size 196 ``` ### Datasets Used: - **BACPHLIP (without E. coli)**: 1,868 training sequences and 246 validation sequences. - **Guelin Collection**: 394 *Escherichia* phages (temperate and virulent types). - **EXTREMOPHILE Phages**: 16 phages isolated from extreme environments, including deep-sea, acidic, and arsenic-rich habitats. Each dataset was processed using **512bp segment lengths** to simulate fragmented metagenomic assemblies. --- ## Performance Results The performance of ProkBERT PhaStyle was evaluated on various datasets, including *Escherichia* and EXTREMOPHILE phages, using segment lengths of 512bp and 1022bp. The results are summarized below: ### Performance on Escherichia Test Set (512bp segments) | Method | Balanced Accuracy | MCC | Sensitivity | Specificity | |-----------------------|-------------------|-------|-------------|-------------| | ProkBERT-mini | 0.91 | 0.83 | 0.94 | 0.89 | | ProkBERT-mini-long | 0.90 | 0.82 | 0.96 | 0.85 | | ProkBERT-mini-c | 0.89 | 0.80 | 0.95 | 0.84 | | DNABERT-2-117M | 0.84 | 0.72 | 0.95 | 0.74 | | Nucleotide Transformer-50m | 0.85 | 0.72 | 0.92 | 0.78 | | Nucleotide Transformer-100m | 0.87 | 0.75 | 0.93 | 0.82 | | Nucleotide Transformer-500m | 0.88 | 0.78 | 0.96 | 0.80 | | DeePhage | 0.86 | 0.71 | 0.84 | 0.88 | | PhaTYP | 0.91 | 0.83 | 0.94 | 0.88 | ### Performance on Escherichia Test Set (1022bp segments) | Method | Balanced Accuracy | MCC | Sensitivity | Specificity | |-----------------------|-------------------|-------|-------------|-------------| | ProkBERT-mini | 0.94 | 0.88 | 0.97 | 0.91 | | ProkBERT-mini-long | 0.94 | 0.89 | 0.97 | 0.91 | | ProkBERT-mini-c | 0.93 | 0.87 | 0.97 | 0.89 | | DNABERT-2-117M | 0.90 | 0.80 | 0.95 | 0.85 | | Nucleotide Transformer-50m | 0.90 | 0.80 | 0.94 | 0.85 | | Nucleotide Transformer-100m | 0.92 | 0.83 | 0.94 | 0.89 | | Nucleotide Transformer-500m | 0.91 | 0.84 | 0.96 | 0.87 | | DeePhage | 0.91 | 0.82 | 0.94 | 0.88 | | PhaTYP | 0.92 | 0.84 | 0.96 | 0.87 | --- ### Performance on EXTREMOPHILE Test Set (512bp segments) | Method | Balanced Accuracy | MCC | Sensitivity | Specificity | |-----------------------|-------------------|-------|-------------|-------------| | ProkBERT-mini | 0.93 | 0.83 | 0.99 | 0.87 | | ProkBERT-mini-long | 0.93 | 0.82 | 1.00 | 0.86 | | ProkBERT-mini-c | 0.92 | 0.80 | 0.99 | 0.84 | | DNABERT-2-117M | 0.89 | 0.74 | 0.99 | 0.79 | | Nucleotide Transformer-50m | 0.91 | 0.79 | 0.98 | 0.84 | | Nucleotide Transformer-100m | 0.90 | 0.76 | 0.97 | 0.82 | | Nucleotide Transformer-500m | 0.91 | 0.78 | 0.99 | 0.82 | | DeePhage | 0.87 | 0.75 | 0.84 | 0.91 | | PhaTYP | 0.76 | 0.52 | 0.74 | 0.79 | ### Performance on EXTREMOPHILE Test Set (1022bp segments) | Method | Balanced Accuracy | MCC | Sensitivity | Specificity | |-----------------------|-------------------|-------|-------------|-------------| | ProkBERT-mini | 0.96 | 0.91 | 1.00 | 0.93 | | ProkBERT-mini-long | 0.96 | 0.90 | 1.00 | 0.92 | | ProkBERT-mini-c | 0.94 | 0.86 | 1.00 | 0.89 | | DNABERT-2-117M | 0.94 | 0.85 | 0.98 | 0.90 | | Nucleotide Transformer-50m | 0.93 | 0.83 | 0.99 | 0.87 | | Nucleotide Transformer-100m | 0.95 | 0.88 | 0.98 | 0.91 | | Nucleotide Transformer-500m | 0.96 | 0.89 | 1.00 | 0.91 | | DeePhage | 0.92 | 0.80 | 0.96 | 0.87 | | PhaTYP | 0.80 | 0.58 | 0.84 | 0.76 | --- ### Inference Speed and Running Times | Model | Execution Time (seconds) | Inference Speed (MB/sec) | |------------------------|--------------------------|--------------------------| | ProkBERT-mini-long | 132 | 0.52 | | ProkBERT-mini | 141 | 0.49 | | ProkBERT-mini-c | 146 | 0.47 | | DNABERT-2-117M | 284 | 0.23 | | Nucleotide Transformer-50m | 292 | 0.21 | | Nucleotide Transformer-100m | 313 | 0.20 | | Nucleotide Transformer-500m | 500 | 0.15 | | DeePhage | 159 | 0.43 | | PhaTYP | 2718 | 0.10 | | BACPHLIP | 7125 | 0.04 | For more detailed results, including additional metrics, please refer to the original research paper. --- ## Inference Speed and Running Times The computational performance of ProkBERT PhaStyle was evaluated using 1,000 randomly selected sequences from the BACPHLIP dataset. The evaluation was performed on a consistent hardware setup with NVIDIA Tesla A100 GPUs. The execution times and inference speeds of various models are summarized below: ### Execution Times (in seconds) | Model | Execution Time (seconds) | Inference Speed (MB/sec) | |--------------------------|--------------------------|--------------------------| | **ProkBERT-mini-long** | **132** | **0.52** | | ProkBERT-mini | 141 | 0.49 | | ProkBERT-mini-c | 146 | 0.47 | | DNABERT-2-117M | 248 | 0.25 | | Nucleotide Transformer-50m| 342 | 0.18 | | Nucleotide Transformer-500m| 502 | 0.12 | | DeePhage | 159 | 0.43 | | PhaTYP | 2,718 | 0.03 | | BACPHLIP | 7,125 | 0.01 | ## Limitations ProkBERT PhaStyle is specifically designed for **binary classification** of phage lifestyles (virulent vs. temperate) and does not handle non-phage sequences. It is recommended to use this model in conjunction with upstream pipelines that identify phage sequences. For large-scale inference, **GPU support** is strongly advised. # Citing this work If you use the data in this package, please cite: ```bibtex @Article{ProkBERT2024, author = {Ligeti, Balázs and Szepesi-Nagy, István and Bodnár, Babett and Ligeti-Nagy, Noémi and Juhász, János}, journal = {Frontiers in Microbiology}, title = {{ProkBERT} family: genomic language models for microbiome applications}, year = {2024}, volume = {14}, URL={https://www.frontiersin.org/articles/10.3389/fmicb.2023.1331233}, DOI={10.3389/fmicb.2023.1331233}, ISSN={1664-302X} } ```