PSP - 蛋白质序列提取 Transformer 蛋白质语言模型 ESM2 特征

欢迎关注我的CSDN：https://spike.blog.csdn.net/
本文地址：https://spike.blog.csdn.net/article/details/132888139

蛋白质语言模型 ESM (Evolutionary Scale Modeling) 是一种利用深度学习技术来预测蛋白质结构和功能的方法。ESM 通过在大规模的蛋白质序列数据库上，训练一个自回归的神经网络，学习蛋白质的进化规律和序列-结构-功能的关系。ESM 可以根据给定的蛋白质序列，生成其对应的隐向量，表示其结构和功能的特征，还可以利用隐向量进行多种下游任务，如结构预测、功能注释、相互作用分析等。ESM 是一种强大而通用的蛋白质语言模型，为蛋白质科学提供了新的视角和工具。

ESM (Evolutionary Scale Modeling)，即进化尺度模型，包括 ESM-2、ESMFold、ESM-MSA-1b、ESM-1v、ESM-IF1(反向折叠)，即

• ESM-2，2022.8，SOTA 通用目的蛋白质语言模型 v2 版，其中 ESM-1v 是 v1 版本。
• ESMFold，2022.11，端到端的单序列 3D 结构预测
• ESM-MSA-1b，2021.6，MSA Transformer 语言模型
• ESM-IF1，2022.4，反向折叠模型

具体参考：ESM GitHub

1. 配置 Docker 环境

配置 TORCH_HOME 与 BOS 环境，即:

vim ~/.bashrc

export TORCH_HOME=[your folder]/torch_home/
alias bos='bcecmd/bcecmd --conf-path bcecmd/bceconf/ bos'

建议配置 TORCH_HOME，固定 PyTorch 模型缓存地址，即 torch_home/hub/checkpoints 中。

在 Docker Image 中，导入 ESM 环境：

conda create -n esmfold --clone miniconda3/envs/esmfold

需要安装的 Torch 相关包：

pip install -q torch-scatter -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
pip install -q torch-sparse -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
pip install -q torch-cluster -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
pip install -q torch-spline-conv -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
pip install -q torch-geometric

导出 Docker 环境：

# 提交 Tag
docker ps -a
docker commit [container id] esmfold:v1.0

# 准备远程 Tag
docker tag esmfold:v1.0 [your ip]/esmfold:v1.0

# 推送至远程
docker push [your ip]/esmfold:v1.0
# 从远程拉取
# docker pull [your ip]/esmfold:v1.0

2. 批量推理 ESM2 模型

配置 ESM 推理脚本：

set -xe
PROJECT_DIR="$(cd "$(dirname $0)" && pwd)/.."

source activate esmfold
export PATH="/usr/local/cuda-11.6/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-11.6/lib64:$LD_LIBRARY_PATH"
export TORCH_HOME=[your folder]/torch_home/

echo "${PROJECT_DIR}"

python "${PROJECT_DIR}/scripts/extract.py" esm2_t36_3B_UR50D \
  "${PROJECT_DIR}/mydata/all-1.fasta" \
  [your folder]/esm2_3B_feat/ \
  --toks_per_batch 1536 \
  --repr_layers -1 \
  --include per_tok contacts \
  --truncation_seq_length 1536 \
  --num_workers 8

测试 A100 显卡 80G，最大支持 1536 序列长度。

优化 scripts/extract.py 脚本，输出结果是序列 MD5 编码的特征，避免序列过长或名字重复：

1. 增加 num_workers，提升推理速度。
2. 替换 label 为蛋白质序列。
3. 增加断点处理，避免重复搜索

即

# ...
data_loader = DataLoader(
    dataset, collate_fn=alphabet.get_batch_converter(args.truncation_seq_length),
    batch_sampler=batches, num_workers=args.num_workers,
)
# ...
# result = {"label": label}
result = {
    
    "label": strs[i]}  # label 修改成序列
# ...
for i, label in enumerate(labels):
    args.output_file = args.output_dir / f"{
      
      label}.pt"
    if os.path.isfile(args.output_file):
        warnings.warn(f"The feat has processed. {
      
      args.output_file}")
        continue
# ...

注意不能使用 num_workers 否则程序无法运行。

日志：

python workspace/esm-by-chenlong/run/../scripts/extract.py esm2_t36_3B_UR50D workspace/esm-by-chenlong/mydata/all-1.fasta pdb_dataset/esm2_6b_feat/ --toks_per_batch 1536 --repr_layers -1 --include per_tok contacts --truncation_seq_length 1536 --num_workers 32
Transferred model to GPU
Read /nfs_beijing_ai/chenlong/workspace/esm-by-chenlong/run/../mydata/all-1.fasta with 27115 sequences
Processing 1 of 6668 batches (66 sequences)
Processing 2 of 6668 batches (61 sequences)
Processing 3 of 6668 batches (56 sequences)
Processing 4 of 6668 batches (52 sequences)
Processing 5 of 6668 batches (51 sequences)

注意序列尺寸 2048 导致显存溢出。

3. 准备 ESM2 输入 FASTA 数据

将 FASTA 文件夹中的全部 FASTA 文件组成1个文件，并且序列描述，转换成 Hash 编码，避免相同序列的特征重复生成特征，即：

• seq_encoder ：Hash 编码函数，同时也用于查找。
• load_feat：读取 feature 特征，支持显示数据和绘制图像。
• merge_fasta_folder：合并 FASTA 文件夹。

即：

#!/usr/bin/env python
# -- coding: utf-8 --
"""
Copyright (c) 2022. All rights reserved.
Created by C. L. Wang on 2023/9/13
"""
import argparse
import os
import sys
import warnings
from pathlib import Path

from tqdm import tqdm

p = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if p not in sys.path:
    sys.path.append(p)

from myutils.project_utils import traverse_dir_files_for_large, read_file, write_list_to_file


class Esm2FastaGenerator(object):
    """
    ESM2 工具类
    """
    def __init__(self):
        pass

    @staticmethod
    def seq_encoder(sequence):
        """
        将 seq 使用 hash 编码，避免重复生成
        """
        import hashlib
        return hashlib.md5(sequence.encode(encoding="utf-8")).hexdigest()

    @staticmethod
    def load_feat(path, is_print=False):
        """
        加载 ESM 特征文件，以及打印特征
        """
        import torch
        from torch import Tensor
        rep = torch.load(path)
        if is_print:
            print(f"[Info] rep: {
      
      rep.keys()}")
            for key in rep.keys():
                val = rep[key]
                if isinstance(val, str):
                    print(f"[Info] {
      
      key}: {
      
      val}")
                elif isinstance(val, dict):
                    for sub_key in val.keys():
                        print(f"[Info] {
      
      key}: {
      
      sub_key}: {
      
      val[sub_key].shape}")
                elif isinstance(val, Tensor):
                    print(f"[Info] {
      
      key}: {
      
      val.shape}")
                else:
                    print(f"[Info] {
      
      key}: {
      
      val}")

            # 绘制接触矩阵
            import matplotlib.pyplot as plt
            contacts_map = rep["contacts"]
            plt.matshow(contacts_map)
            plt.title("contacts_map")
            save_name = "contacts_map.png"
            plt.savefig(save_name, bbox_inches='tight', format='png')
            plt.show()
        return rep

    @classmethod
    def merge_fasta_folder(cls, folder_path, output_path):
        """
        合并 fasta 文件，用于 esm 推理
        """
        print(f"[Info] folder_path: {
      
      folder_path}")
        print(f"[Info] output_path: {
      
      output_path}")
        assert os.path.isdir(folder_path)
        path_list = traverse_dir_files_for_large(folder_path, ext="fasta")
        print(f"[Info] fasta: {
      
      len(path_list)}")
        seq_set = set()
        for path in tqdm(path_list, "[Info] fasta"):
            data_lines = read_file(path)
            n = len(data_lines)
            for i in range(1, n, 2):
                seq = data_lines[i]
                if seq:
                    seq_set.add(seq)
        seq_list = list(seq_set)
        print(f"[Info] seq unique: {
      
      len(seq_list)}")
        # create_empty_file(output_path)
        seq_lines = []
        header_set = set()
        for seq in tqdm(seq_list, "[Info] seq"):
            header = cls.seq_encoder(seq)
            header_set.add(header)
            seq_lines.append(f">{
      
      header}")
            seq_lines.append(seq)
        assert len(seq_lines) // 2 == len(header_set)
        write_list_to_file(output_path, seq_lines)
        print(f"[Info] over! {
      
      output_path}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "-f",
        "--folder-path",
        type=Path,
        required=True,
    )
    parser.add_argument(
        "-o",
        "--output-path",
        type=Path,
        required=True
    )
    args = parser.parse_args()

    folder_path = str(args.folder_path)
    output_path = str(args.output_path)
    if os.path.isfile(output_path):
        warnings.warn(f"The output file exists, append lines to it! {
      
      output_path}")
    # from root_dir import DATA_DIR
    # folder_path = os.path.join(DATA_DIR, "fasta")
    # output_path = os.path.join(DATA_DIR, "all.fasta")
    Esm2FastaGenerator.merge_fasta_folder(folder_path, output_path)


def main2():
    from root_dir import DATA_DIR
    feat_path = os.path.join(DATA_DIR, "fffd26f4307d76eec938ac9c2c93a698.pt")
    Esm2FastaGenerator.load_feat(feat_path, is_print=True)


if __name__ == '__main__':
    main()
    # main2()

输出的序列 ESM2 特征包括：

• label 序列描述
• representations 序列表征 LxH
• mean_representations 均值化表征 H
• bos_representations 起始 Token 表征 H
• contacts 序列接触表征 LxL

例如 序列长度是 65，ESM2 650M 的 Embeddings 是 1280，ESM2 3B 是 2560，即：

[Info] rep: dict_keys(['label', 'representations', 'contacts'])
[Info] label: MAKDSKAPVVEIFDERDGCTSAGSTGKASDAGEKGLLVKVSMQKVGYNAIMAKSVAASYMNK
[Info] representations: 36: torch.Size([62, 2560])
[Info] contacts: torch.Size([62, 62])

其中 序列长度 235 的 ESM2 3B 特征，约是 2.6M，序列长度 65 的 ESM2 650M 特征，约是 361 KB。

4. 测试 ESM2 推理脚本

推理脚本：

#!/usr/bin/env python
# -- coding: utf-8 --
"""
Copyright (c) 2022. All rights reserved.
Created by C. L. Wang on 2023/9/11
"""
import math
import os
import sys
import time

import torch

import esm

p = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if p not in sys.path:
    sys.path.append(p)

from myutils.project_utils import time_elapsed


class Esm2Infer(object):
    """
    推理 ESM2 特征
    """
    def __init__(self):
        print("[Info] 加载模型开始! ")
        s_time = time.time()
        self.model, self.alphabet = esm.pretrained.esm2_t33_650M_UR50D()
        print(f"[Info] vocab: {
      
      self.alphabet.to_dict()}")
        self.batch_converter = self.alphabet.get_batch_converter()
        self.model.eval()  # disables dropout for deterministic results
        print(f"[Info] 加载模型完成! 耗时: {
      
      time_elapsed(s_time, time.time())}")

    def predict(self, data_list):
        """
        数据示例:
        data_list = [
            ("protein1", "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"),
            ("protein2", "KALTARQQEVFDLIRDHISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
            ("protein2 with mask", "KALTARQQEVFDLIRD<mask>ISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
            ("protein3",  "K A <mask> I S Q"),
        ]
        """
        print(f"[Info] data_list: {
      
      len(data_list)}")
        batch_labels, batch_strs, batch_tokens = self.batch_converter(data_list)
        print(f"[Info] batch_labels: {
      
      batch_labels}")
        print(f"[Info] batch_tokens: {
      
      batch_tokens}")
        batch_lens = (batch_tokens != self.alphabet.padding_idx).sum(1)
        print(f"[Info] batch_lens: {
      
      batch_lens}")  # 有效维数

        # Extract per-residue representations (on CPU)
        with torch.no_grad():
            results = self.model(batch_tokens, repr_layers=[33], return_contacts=True)
        token_representations = results["representations"][33]

        # Generate per-sequence representations via averaging
        # NOTE: token 0 is always a beginning-of-sequence token, so the first residue is token 1.
        sequence_representations = []
        for i, tokens_len in enumerate(batch_lens):
            feat = token_representations[i, 1: tokens_len - 1]
            # embeddings = feat.mean(0)
            # print(f"[Info] idx: {i}, feat: {feat.shape}, embeddings: {embeddings.shape}")
            # sequence_representations.append(embeddings)
            sequence_representations.append(feat)
        return sequence_representations


def main():
    data_list = [
        ("protein1", "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"),
        ("protein2", "KALTARQQEVFDLIRDHISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
        ("protein2 with mask", "KALTARQQEVFDLIRD<mask>ISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
        ("protein3", "K A <mask> I S Q"),
    ]
    ei = Esm2Infer()
    ei.predict(data_list)


if __name__ == '__main__':
    main()