app.py

import gradio as gr, numpy as np

import torch
from transformers import EsmTokenizer,EsmForMaskedLM

model_name = "facebook/esm2_t6_8M_UR50D"
tokenizer = EsmTokenizer.from_pretrained(model_name)

device = torch.device("cpu")

model_mlm = EsmForMaskedLM.from_pretrained(model_name).to(device)

# 2. Define the PLL Calculation Function
def predict_ppp(sequence) -> float:
    """
    Calculates the ESM2 Pseudolog-Likelihood (PLL) for a single amino acid sequence.

    PLL = sum_i( log P(x_i | x_{~i}) )
    """
    # Tokenize the sequence
    # This automatically adds <CLS> and <EOS> tokens
    input_ids = tokenizer(sequence, return_tensors='pt')['input_ids']

    # The true sequence length (excluding special tokens)
    L = len(sequence)

    # The mask indices correspond to the AA sequence positions
    # We ignore the first (<CLS>) and last (<EOS>) tokens.
    mask_indices = torch.arange(1, L + 1)

    # Accumulator for the log-likelihood sum
    pll_sum = 0.0

    # Iterate over each position in the sequence to mask it
    for i in mask_indices:
        # Create a copy of the input_ids
        masked_input = input_ids.clone()

        # Mask the current residue (token ID for MASK is 1)
        masked_input[0, i] = tokenizer.mask_token_id

        # Get model logits (unnormalized log-probabilities)
        with torch.no_grad():
            outputs = model_mlm(masked_input)
            logits = outputs.logits # shape: (batch_size, seq_len, vocab_size)

        # Extract the log-probabilities for the prediction at the masked position
        # We use log_softmax to get log-probabilities
        log_probs = torch.log_softmax(logits[0, i], dim=-1)

        # Get the token ID of the *actual* residue at the masked position
        target_token_id = input_ids[0, i].item()

        # Get the log-probability of the actual residue
        log_prob_of_target = log_probs[target_token_id].item()

        # Add to the sum
        pll_sum += log_prob_of_target
        L = len(sequence)
        ppp = np.exp(-pll_sum / L)

    return ppp

demo = gr.Interface(
    fn=predict_ppp,
    inputs=[
        gr.Textbox(label="Enter Protein Amino Acid Sequence (1-letter code)",
        placeholder="ACDEFGHIKLMNPQRSTVWY"),
           ],
    outputs="text",
    title="Nano Protein Language Model for Pseudo Perplexity (PPP) prediction of a protein sequence",
    description="Enter an amino acid sequence (using the 1-letter code) to predict its Pseudo Perplexity (PPP)",
    examples=[
        ["MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"], # Example sequence
            ]
)
demo.launch()