utils/inference_utils.py

from functools import cache
from pathlib import Path

from esm import FastaBatchedDataset, pretrained
from rdkit.Chem import AddHs
from torch_geometric.data import Dataset, HeteroData
import numpy as np
import torch
import prody as pr
import esm
import pandas as pd

from datasets.process_mols import generate_conformer, read_molecule, get_lig_graph_with_matching, moad_extract_receptor_structure
from datasets.parse_chi import aa_idx2aa_short, get_onehot_sequence


def get_sequences_from_pdbfile(file_path):
    sequence = None
    # prodyb package requires str input
    pdb = pr.parsePDB(str(file_path))
    seq = pdb.ca.getSequence()
    one_hot = get_onehot_sequence(seq)

    chain_ids = np.zeros(len(one_hot))
    res_chain_ids = pdb.ca.getChids()
    res_seg_ids = pdb.ca.getSegnames()
    res_chain_ids = np.asarray([s + c for s, c in zip(res_seg_ids, res_chain_ids)])
    ids = np.unique(res_chain_ids)

    for i, id in enumerate(ids):
        chain_ids[res_chain_ids == id] = i

        s_temp = np.argmax(one_hot[res_chain_ids == id], axis=1)
        s = ''.join([aa_idx2aa_short[aa_idx] for aa_idx in s_temp])

        if sequence is None:
            sequence = s
        else:
            sequence += (":" + s)

    return sequence

@cache
def process_protein(protein_string):
    input_path = Path(protein_string)
    # Check if the input is a path to a file
    if Path(protein_string).is_absolute() or len(Path(protein_string).parts) > 1:
        # Check if the input is a PDB file path
        if input_path.is_file() and input_path.suffix == '.pdb':
            # Extract sequence from PDB file
            return get_sequences_from_pdbfile(input_path), str(input_path)
        else:
            raise FileNotFoundError(f"File {protein_string} not found or not a PDB file")
    else:
        # Assume the input is already a FASTA sequence
        return protein_string, None


def compute_esm_embeddings(model, alphabet, labels, sequences):
    # settings used
    toks_per_batch = 4096
    repr_layers = [33]
    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
    )

    assert all(-(model.num_layers + 1) <= i <= model.num_layers for i in repr_layers)
    repr_layers = [(i + model.num_layers + 1) % (model.num_layers + 1) for i in repr_layers]
    embeddings = {}

    with torch.no_grad():
        for batch_idx, (labels, strs, toks) in enumerate(data_loader):
            print(f"Processing {batch_idx + 1} of {len(batches)} batches ({toks.size(0)} sequences)")
            if torch.cuda.is_available():
                toks = toks.to(device="cuda", non_blocking=True)

            out = model(toks, repr_layers=repr_layers, return_contacts=False)
            representations = {layer: t.to(device="cpu") for layer, t in out["representations"].items()}

            for i, label in enumerate(labels):
                truncate_len = min(truncation_seq_length, len(strs[i]))
                embeddings[label] = representations[33][i, 1: truncate_len + 1].clone()
    return embeddings


def generate_esm_structure(model, filename, sequence):
    model.set_chunk_size(256)
    chunk_size = 256
    output = None

    while output is None:
        try:
            with torch.no_grad():
                output = model.infer_pdb(sequence)

            with open(filename, "w") as f:
                f.write(output)
                print("saved", filename)
        except RuntimeError as e:
            if 'out of memory' in str(e):
                print('| WARNING: ran out of memory on chunk_size', chunk_size)
                for p in model.parameters():
                    if p.grad is not None:
                        del p.grad  # free some memory
                torch.cuda.empty_cache()
                chunk_size = chunk_size // 2
                if chunk_size > 2:
                    model.set_chunk_size(chunk_size)
                else:
                    print("Not enough memory for ESMFold")
                    break
            else:
                raise e
    return output is not None


class InferenceDataset(Dataset):
    def __init__(self,
                 df, out_dir,
                 lm_embeddings, receptor_radius=30, c_alpha_max_neighbors=None, precomputed_lm_embeddings=None,
                 remove_hs=False, all_atoms=False, atom_radius=5, atom_max_neighbors=None, knn_only_graph=False):

        super(InferenceDataset, self).__init__()
        self.receptor_radius = receptor_radius
        self.c_alpha_max_neighbors = c_alpha_max_neighbors
        self.remove_hs = remove_hs
        self.all_atoms = all_atoms
        self.atom_radius, self.atom_max_neighbors = atom_radius, atom_max_neighbors
        self.knn_only_graph = knn_only_graph

        self.df = df

        # generate LM embeddings
        if lm_embeddings and (precomputed_lm_embeddings is None or precomputed_lm_embeddings[0] is None):
            print("Generating ESM language model embeddings")
            model_location = "esm2_t33_650M_UR50D"
            model, alphabet = pretrained.load_model_and_alphabet(model_location)
            model.eval()
            if torch.cuda.is_available():
                model = model.cuda()
            df[['protein_sequence', 'protein_path']] = df['X2'].apply(process_protein).apply(pd.Series)
            labels, sequences = [], []
            for i in range(len(df)):
                s = df['protein_sequence'].iloc[i].split(':')
                sequences.extend(s)
                labels.extend([df['name'].iloc[i] + '_chain_' + str(j) for j in range(len(s))])
            # TODO improve efficiency for repeated X2 values
            lm_embeddings = compute_esm_embeddings(model, alphabet, labels, sequences)

            self.lm_embeddings = []
            for i in range(len(df)):
                s = df['protein_sequence'].iloc[i].split(':')
                self.lm_embeddings.append(
                    [lm_embeddings[f"{df['name'].iloc[i]}_chain_{j}"] for j in range(len(s))]
                )

        elif not lm_embeddings:
            self.lm_embeddings = [None] * len(self.complex_names)

        else:
            self.lm_embeddings = precomputed_lm_embeddings

        # generate structures with ESMFold
        if None in df['protein_path'].values:
            print("generating missing structures with ESMFold")
            model = esm.pretrained.esmfold_v1()
            model = model.eval().cuda()

            for i in range(len(df)):
                # TODO improve efficiency for repeated X2 values
                protein_sequence = df['protein_sequence'].iloc[i]
                protein_file = df['protein_path'].iloc[i]
                complex_name = df['name'].iloc[i]
                if protein_file is None:
                    protein_file = f"{out_dir}/{complex_name}/{complex_name}_esmfold.pdb"
                    if not Path(protein_file).is_file():
                        print("generating", df['protein_path'].iloc[i])
                        generate_esm_structure(model, protein_file, protein_sequence)
                        df['protein_sequence'].iloc[i] = protein_sequence

    def len(self):
        return len(self.df)

    def get(self, idx):
        name = self.df['name'].iloc[idx]
        protein_file = self.df['protein_path'].iloc[idx]
        ligand_description = self.df['X1'].iloc[idx]
        mol = self.df['mol'].iloc[idx]
        lm_embedding = self.lm_embeddings[idx]

        # build the pytorch geometric heterogeneous graph
        complex_graph = HeteroData()
        complex_graph['name'] = name

        if mol is not None:
            mol = AddHs(mol)
            generate_conformer(mol)
        else:
            print(f'Failed to read molecule {ligand_description}. Skipping...')
            complex_graph['success'] = False
            return complex_graph

        try:
            # parse the receptor from the pdb file
            get_lig_graph_with_matching(mol, complex_graph, popsize=None, maxiter=None, matching=False, keep_original=False,
                                        num_conformers=1, remove_hs=self.remove_hs)

            moad_extract_receptor_structure(
                path=protein_file,
                complex_graph=complex_graph,
                neighbor_cutoff=self.receptor_radius,
                max_neighbors=self.c_alpha_max_neighbors,
                lm_embeddings=lm_embedding,
                knn_only_graph=self.knn_only_graph,
                all_atoms=self.all_atoms,
                atom_cutoff=self.atom_radius,
                atom_max_neighbors=self.atom_max_neighbors)

        except Exception as e:
            print(f'Skipping {name} because of the error:')
            print(e)
            complex_graph['success'] = False
            return complex_graph

        protein_center = torch.mean(complex_graph['receptor'].pos, dim=0, keepdim=True)
        complex_graph['receptor'].pos -= protein_center
        if self.all_atoms:
            complex_graph['atom'].pos -= protein_center

        ligand_center = torch.mean(complex_graph['ligand'].pos, dim=0, keepdim=True)
        complex_graph['ligand'].pos -= ligand_center

        complex_graph.original_center = protein_center
        complex_graph.mol = mol
        complex_graph['success'] = True
        return complex_graph