model/Protenix-pxd/protenix/data/compute_esm.py

# Copyright 2024 ByteDance and/or its affiliates.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import os

import pandas as pd
import torch
from esm import FastaBatchedDataset, pretrained
from tqdm.auto import tqdm

ESM_CONFIG = {
    "esm2-3b": {
        "type": "esm2",
        "model_path": "esm2_t36_3B_UR50D.pt",
        "emb_dim": 2560,
        "n_layers": 36,
    },
    "esm2-3b-ism": {
        "type": "esm2",
        "model_path": "esm2_t36_3B_UR50D_ism.pt",
        "emb_dim": 2560,
        "n_layers": 36,
    },  # https://www.biorxiv.org/content/10.1101/2024.11.08.622579v2
}


def _load_esm2_model(model_path):
    if os.path.exists(model_path):
        model, alphabet = pretrained.load_model_and_alphabet_local(model_path)
    else:
        model, alphabet = pretrained.load_model_and_alphabet(
            os.path.splitext(os.path.basename(model_path))[0]
        )
    return model, alphabet


def load_esm_model(model_name, local_esm_dir="release_data/checkpoint"):
    local_model_path = os.path.join(local_esm_dir, ESM_CONFIG[model_name]["model_path"])
    if os.path.exists(local_model_path):
        print("Try to load ESM language model from ", local_model_path)

    if "ism" in model_name and not os.path.exists(local_model_path):
        raise RuntimeError(
            f"esm2-3b-ism model: {local_model_path} does not exist \n"
            + "this model can not be download from fair-esm, \n"
            + "download it from https://af3-dev.tos-cn-beijing.volces.com/release_model/esm2_t36_3B_UR50D_ism.pt"
        )
    if model_name.startswith("esm2"):
        model, alphabet = _load_esm2_model(local_model_path)
    model.eval()
    if torch.cuda.is_available():
        model = model.cuda()

    return model, alphabet


def _check_files_exist(save_dir, labels):
    return all(
        [os.path.exists(os.path.join(save_dir, label + ".pt")) for label in labels]
    )


def compute_ESM_embeddings(
    model_name,
    model,
    alphabet,
    labels,
    sequences,
    save_dir,
    toks_per_batch=4096,
    truncation_seq_length=1022,
):
    if model_name.startswith("esm2"):
        embeddings = compute_esm2_embeddings(
            model,
            alphabet,
            labels,
            sequences,
            save_dir,
            toks_per_batch,
            truncation_seq_length,
        )
    return embeddings


# Adapt from Corso, Gabriele, et al. "Diffdock: Diffusion steps, twists, and turns for molecular docking."
# URL: https://github.com/gcorso/DiffDock/blob/main/utils/inference_utils.py
def compute_esm2_embeddings(
    model,
    alphabet,
    labels,
    sequences,
    save_dir,
    toks_per_batch=4096,
    truncation_seq_length=1022,
):
    dataset = FastaBatchedDataset(labels, sequences)
    batches = dataset.get_batch_indices(toks_per_batch, extra_toks_per_seq=1)
    data_loader = torch.utils.data.DataLoader(
        dataset,
        collate_fn=alphabet.get_batch_converter(truncation_seq_length),
        batch_sampler=batches,
    )
    repr_layer = model.num_layers
    embeddings = {}
    with torch.no_grad():
        for batch_idx, (labels, strs, toks) in enumerate(tqdm(data_loader)):
            print(
                f"Processing {batch_idx + 1} of {len(batches)} batches ({toks.size(0)} sequences)"
            )
            if _check_files_exist(save_dir, labels):
                continue
            if torch.cuda.is_available():
                toks = toks.to(device="cuda", non_blocking=True)
            out = model(toks, repr_layers=[repr_layer], return_contacts=False)
            representation = out["representations"][repr_layer].to(device="cpu")
            for i, label in enumerate(labels):
                truncate_len = min(truncation_seq_length, len(strs[i]))
                embeddings[label] = representation[i, 1 : truncate_len + 1].clone()
                save_path = os.path.join(save_dir, label + ".pt")
                torch.save(embeddings[label], save_path)
    return embeddings


def pdb_sequences_iterator(
    input_path="./scripts/msa/data/pdb_seqs/pdb_seq.csv",
    save_path="./scripts/msa/data/pdb_seqs/pdb_labels_seqs.csv",
    start_id=0,
    end_id=-1,
):
    if os.path.exists(save_path):
        df_seq = pd.read_csv(save_path)
    else:
        df = pd.read_csv(input_path)
        # Protein only
        df = df[df["mol_type"] == "protein"]
        # Sequence name
        df["pdb_entity_id"] = df["pdb_id"] + "_" + df["entity_id"].astype(str)
        # Group by 'seq'
        df_seq = (
            df.groupby("seq")["pdb_entity_id"]
            .apply(lambda x: ",".join(x))
            .reset_index()
        )
        # Use the first pdb_entity_id as the label
        df_seq["seq_label"] = df_seq["pdb_entity_id"].apply(lambda x: x.split(",")[0])
        assert df_seq["seq_label"].nunique() == len(df_seq)
        # Get a part id
        df_seq["part_id"] = df_seq["pdb_entity_id"].apply(lambda x: x[1:3])
        df_seq.to_csv(save_path)

    if end_id == -1:
        end_id = len(df_seq)
    df_seq = df_seq[start_id:end_id]

    part_counts = dict(df_seq["part_id"].value_counts())
    for part_id, count in part_counts.items():
        df_part = df_seq[df_seq["part_id"] == part_id]
        print(f"Part {part_id}: {len(df_part)} sequences.")
        yield part_id, df_part["seq_label"].tolist(), df_part["seq"].tolist()


def process_pdb_dataset(
    model_name, root_save_dir, pdb_seq_path, pdb_seq_label_path, start_id=0, end_id=-1
):

    model, alphabet = load_esm_model(model_name)
    seq_iterator = pdb_sequences_iterator(
        pdb_seq_path, pdb_seq_label_path, start_id, end_id
    )
    error_parts = []
    for part_id, labels, sequences in seq_iterator:
        save_dir = os.path.join(root_save_dir, f"{part_id}")

        if not os.path.exists(save_dir):
            os.makedirs(save_dir)
        print(f"[{part_id}] Generating ESM language model embeddings")
        lm_embeddings = compute_ESM_embeddings(
            model_name,
            model,
            alphabet,
            labels,
            sequences,
            save_dir,
            truncation_seq_length=4094,
            toks_per_batch=16384,
        )
        print(f"[{part_id}] Processed {len(lm_embeddings)} sequences in total. Done!")

    print("Error parts: ", error_parts)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model_name", type=str, choices=list(ESM_CONFIG.keys()))
    parser.add_argument("--start_id", type=int, default=0)
    parser.add_argument("--end_id", type=int, default=-1)
    args = parser.parse_args()

    save_dir = f"./esm_embeddings/{args.model_name}"
    pdb_seq_path = "./scripts/msa/data/pdb_seqs/pdb_seq.csv"
    pdb_seq_label_path = "./scripts/msa/data/pdb_seqs/pdb_labels_seqs.csv"

    if not os.path.exists(save_dir):
        print("Make dir: ", save_dir)
        os.makedirs(save_dir)
    process_pdb_dataset(
        args.model_name,
        save_dir,
        pdb_seq_path,
        pdb_seq_label_path,
        args.start_id,
        args.end_id,
    )


if __name__ == "__main__":
    main()