Protein_Sequence_Toxicity_Predictor

Overview

The Protein_Sequence_Toxicity_Predictor is a specialized sequence classification model designed to predict the potential toxicity level of an arbitrary peptide or protein sequence. This model is critical for early-stage drug discovery, synthetic biology, and material science, where quick assessment of a novel sequence's risk profile is essential.

It treats the amino acid sequence (e.g., "M-A-S-K...") as a sentence and the amino acids as tokens, adapting the BERT architecture for biological sequence understanding.

Model Architecture

The model is based on a modified BERT (Bidirectional Encoder Representations from Transformers) architecture, specifically BertForSequenceClassification.

• Base Model: A custom BERT-like model was trained from scratch on a large corpus of protein sequences (Uniref/Swiss-Prot), with a vocabulary size of 30, including the 20 standard amino acids, special tokens (CLS, SEP, PAD), and modified amino acids.
• Input: The input is a raw amino acid sequence (e.g., 'MVLSPADKTN...').
• Tokenizer: A simple BPE-like or character-level tokenizer is used, where each amino acid is treated as a single token.
• Classification: A standard linear classification layer outputs the probability distribution across three toxicity classes: Non_Toxic, Mildly_Toxic, and Highly_Toxic.
• Training: Fine-tuned on a curated dataset of experimentally verified toxic and non-toxic peptide sequences, including antimicrobial peptides and known allergens.

Intended Use

• Drug Design Safety: Rapidly screening candidate therapeutic peptides for potential adverse effects before expensive in vitro or in vivo testing.
• Synthetic Biology: Predicting the safety of novel proteins designed for industrial or research applications.
• Allergen Prediction: Initial assessment of newly discovered or engineered food proteins for potential allergenic risk.
• High-Throughput Screening: Integrating into automated pipelines for filtering large libraries of generated protein sequences.

Limitations

• Context Dependency: Toxicity is often highly dependent on the protein's 3D folding, concentration, and the organism/cell line exposed. This model, being based solely on the primary (1D) sequence, cannot fully account for all these factors.
• Novel Sequence Space: Performance on entirely novel sequences, especially those with non-canonical or heavily modified amino acids, may be lower than on sequences closely related to the training data.
• Threshold Bias: The definition and threshold between Mildly_Toxic and Highly_Toxic are based on the training dataset's experimental methodology, which may not generalize perfectly to all external assays.
• Sequence Length: The current configuration is optimized for typical peptide lengths ($\le 510$ tokens, corresponding to roughly 510 amino acids). Very large protein sequences (e.g., multi-domain proteins) will require truncation, potentially missing important structural information.