app.py

import obonet
import pandas as pd
import streamlit as st
from transformers import EncoderDecoderModel, PreTrainedTokenizerFast

from io import StringIO
from Bio import SeqIO

from inference import run_inference, prepare_inputs, post_process_outputs, join_predictions_with_terms

@st.cache_resource
def load_model(model_path):
    return EncoderDecoderModel.from_pretrained(model_path).eval().requires_grad_(False)

@st.cache_data
def get_terms_df():
    graph = obonet.read_obo('http://purl.obolibrary.org/obo/go/go-basic.obo')
    data = {
        "term": [],
        "name": []
    }

    for term, obj in graph.nodes(data=True):
        data['term'].append(term)
        data['name'].append(obj['name'])

    return pd.DataFrame(data)

def load_tokenizers(protein_tokenizer_path, term_tokenizer_path):
    protein_tokenizer = PreTrainedTokenizerFast.from_pretrained(protein_tokenizer_path)
    text_tokenizer = PreTrainedTokenizerFast.from_pretrained(term_tokenizer_path)

    return protein_tokenizer, text_tokenizer


textarea_placeholder = "Input your sequences in fasta format"
sample_protein = """>sp|A0A0C5B5G6|MOTSC_HUMAN Mitochondrial-derived peptide MOTS-c OS=Homo sapiens OX=9606 GN=MT-RNR1 PE=1 SV=1
MRWQEMGYIFYPRKLR
"""

tokenizers_path = ('khairi/ProtFormer', 'khairi/ProtNLA-V1')
aspects = {
    '<molecular_function>': 'Molecular Function',
    '<cellular_component>': 'Cellular Component',
    '<biological_process>': 'Biological Process',
}

model = load_model("khairi/ProtNLA-V1")
protein_tokenizer, text_tokenizer = load_tokenizers(tokenizers_path[0], tokenizers_path[1])

model.config.vocab_size = len(text_tokenizer)
terms_df = get_terms_df()

fasta_sequences = st.text_area("Input Sequences (fasta format):", value=sample_protein, placeholder=textarea_placeholder)
aspect = st.selectbox(label="Gene Ontology Aspect", options=aspects.keys(), format_func=lambda x: aspects[x])

predict = st.button(label='Run')

if predict:
    fasta_io = StringIO(fasta_sequences) 
    records = SeqIO.parse(fasta_io, "fasta") 
    for record in records:
        sequence = str(record.seq)
        inputs, prompt = prepare_inputs(protein_tokenizer=protein_tokenizer, term_tokenizer=text_tokenizer,
            aspect=aspect, protein_sequence=sequence)
        
        outputs = run_inference(model, inputs, prompt, 5)
        predictions = post_process_outputs(model, outputs, text_tokenizer)
        predictions = join_predictions_with_terms(predictions, terms_df)

        st.header('Predictions')
        st.subheader(f'Protein Sequence: {sequence}')
        st.subheader(f'Sequence Length: {len(record.seq)}')
        st.subheader(f'GO Aspect: {aspects[aspect]}')
        st.write(predictions)