ESM-2 for Protein Function Prediction

Please also see the more recent fine-tuned model AmelieSchreiber/esm2_t6_8M_finetuned_cafa5.

This model is not intended for protein function prediction, but rather as a checkpoint for further fine-tuning, especially with Low Rank Adaptation (LoRA). This is an experimental model fine-tuned from the esm2_t6_8M_UR50D model for multi-label classification. In particular, the model is fine-tuned on the CAFA-5 protein sequence dataset available here. More precisely, the train_sequences.fasta file is the list of protein sequences that were trained on, and the train_terms.tsv file contains the gene ontology protein function labels for each protein sequence. For more details on using ESM-2 models for multi-label sequence classification, see here. Due to the potentially complicated class weighting necessary for the hierarchical ontology, further fine-tuning will be necessary.

Fine-Tuning

The model was fine-tuned for 7 epochs at a learning rate of 5e-5, and achieves the following metrics:

Validation Loss: 0.0027,
Validation Micro F1: 0.3672,
Validation Macro F1: 0.9967,
Validation Micro Precision: 0.6052,
Validation Macro Precision: 0.9996,
Validation Micro Recall: 0.2626,
Validation Macro Recall: 0.9966

Using the model

First, download the train_sequences.fasta file and the train_terms.tsv file, and provide the local paths in the code below:

import os
import numpy as np
import torch
from transformers import AutoTokenizer, EsmForSequenceClassification, AdamW
from torch.nn.functional import binary_cross_entropy_with_logits
from sklearn.model_selection import train_test_split
from sklearn.metrics import f1_score, precision_score, recall_score
# from accelerate import Accelerator
from Bio import SeqIO

# Step 1: Data Preprocessing (Replace with your local paths)
fasta_file = "/Users/amelieschreiber/.cursor-tutor/projects/python/cafa5/cafa-5-protein-function-prediction/Train/train_sequences.fasta"
tsv_file = "/Users/amelieschreiber/.cursor-tutor/projects/python/cafa5/cafa-5-protein-function-prediction/Train/train_terms.tsv"

fasta_data = {}
tsv_data = {}

for record in SeqIO.parse(fasta_file, "fasta"):
    fasta_data[record.id] = str(record.seq)

with open(tsv_file, 'r') as f:
    for line in f:
        parts = line.strip().split("\t")
        tsv_data[parts[0]] = parts[1:]

# tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t6_8M_UR50D")
seq_length = 1022
# tokenized_data = tokenizer(list(fasta_data.values()), padding=True, truncation=True, return_tensors="pt", max_length=seq_length)

unique_terms = list(set(term for terms in tsv_data.values() for term in terms))

Second, downlowd the file go-basic.obo from here and store the file locally, then provide the local path in the the code below:

import torch
from transformers import AutoTokenizer, EsmForSequenceClassification
from sklearn.metrics import precision_recall_fscore_support

# 1. Parsing the go-basic.obo file
def parse_obo_file(file_path):
    with open(file_path, 'r') as f:
        data = f.read().split("[Term]")
        
    terms = []
    for entry in data[1:]:
        lines = entry.strip().split("\n")
        term = {}
        for line in lines:
            if line.startswith("id:"):
                term["id"] = line.split("id:")[1].strip()
            elif line.startswith("name:"):
                term["name"] = line.split("name:")[1].strip()
            elif line.startswith("namespace:"):
                term["namespace"] = line.split("namespace:")[1].strip()
            elif line.startswith("def:"):
                term["definition"] = line.split("def:")[1].split('"')[1]
        terms.append(term)
    return terms

parsed_terms = parse_obo_file("go-basic.obo")  # Replace `go-basic.obo` with your path

# 2. Load the saved model and tokenizer
model_path = "AmelieSchreiber/cafa_5_protein_function_prediction"
loaded_model = EsmForSequenceClassification.from_pretrained(model_path)
loaded_tokenizer = AutoTokenizer.from_pretrained(model_path)

# 3. The predict_protein_function function
def predict_protein_function(sequence, model, tokenizer, go_terms):
    inputs = tokenizer(sequence, return_tensors="pt", padding=True, truncation=True, max_length=1022)
    model.eval()
    with torch.no_grad():
        outputs = model(**inputs)
        predictions = torch.sigmoid(outputs.logits)
        predicted_indices = torch.where(predictions > 0.05)[1].tolist()
    
    functions = []
    for idx in predicted_indices:
        term_id = unique_terms[idx]  # Use the unique_terms list from your training script
        for term in go_terms:
            if term["id"] == term_id:
                functions.append(term["name"])
                break
                
    return functions

# 4. Predicting protein function for an example sequence
example_sequence = "MAYLGSLVQRRLELASGDRLEASLGVGSELDVRGDRVKAVGSLDLEEGRLEQAGVSMA"  # Replace with your protein sequence
predicted_functions = predict_protein_function(example_sequence, loaded_model, loaded_tokenizer, parsed_terms)
print(predicted_functions)