iterative_alignment_experiment_unconditioned/sample_esm_first_round.py

import os
import torch
import numpy as np
import pandas as pd
import h5py
import re
from omegaconf import OmegaConf
import h5py
import lightning as L
from pera.nn import BidirectionalModel, sample_components_from_bidirectional_transformer, sample_perturbations, sample_embedding_perturbations
from esm.tokenization.sequence_tokenizer import EsmSequenceTokenizer
from Bio.Seq import Seq

device = torch.device("cuda:0")

sequence_tokenizer = EsmSequenceTokenizer()

import argparse

# set up parser
parser = parser = argparse.ArgumentParser(description="Calculating the log-likelihood of a sequence")
parser.add_argument('--target', type=str, required=True, help='Dataset as a string')
parser.add_argument('--num_samples', type=int, required=False, default=384, help='Number of samples to process (default: 100000)')
parser.add_argument('--version_number', type=str, required=False, default=1, help='Version number as a string')
parser.add_argument('--replicate', type=int, required=False, default=1, help='Replicate number as an integer')
args = parser.parse_args()

target = args.target
num_samples = args.num_samples
replicate = args.replicate

cfg_filename = "./config.yaml"
network_filename = "/scratch/groups/rotskoff/sebastian/era/protein_era/models/esm3clm/esm3_clm.pt"
save_folder_name = f"{target}/base_model_{num_samples}"

cfg = OmegaConf.load(cfg_filename)
sampling_temperature=1
OmegaConf.update(cfg, "train.lightning_model_args.sampling_temperature", sampling_temperature)
esm_model = BidirectionalModel(cfg["nn"]["model"], 
                                cfg["nn"]["model_args"],
                                **cfg["train"]["lightning_model_args"]).to(device)
esm_model.load_model_from_ckpt(network_filename)
esm_model.eval()
print("")
mask_token_sequence = cfg["nn"]["model_args"]["residue_token_info"]["mask"]
bos_token_sequence = cfg["nn"]["model_args"]["residue_token_info"]["bos"]
eos_token_sequence = cfg["nn"]["model_args"]["residue_token_info"]["eos"]
pad_token_sequence = cfg["nn"]["model_args"]["residue_token_info"]["pad"]


os.makedirs(save_folder_name, exist_ok=True)

data = target #  "GB1", "ParD2", "TEV", "TrpB3F", "TrpB3I", "TrpB4"
data_root_path = "/scratch/groups/rotskoff/sebastian/era/protein_era/data"

sequence_tokenizer = EsmSequenceTokenizer()

if data.startswith("TrpB"):
    df = pd.read_csv(f"{data_root_path}/TrpB/scale2max/{data}.csv")
    with open(f"{data_root_path}/TrpB/TrpB.fasta", "r") as file:
        parent_sequence_decoded = file.readlines()[1].strip()
    
elif data == "DHFR":
    df = pd.read_csv(f"{data_root_path}/{data}/scale2max/{data}.csv")
    with open(f"{data_root_path}/{data}/{data}.fasta", "r") as file:
        nucleotide_seq = file.readlines()[1].strip()
    nucleotide_seq = Seq(nucleotide_seq)
    parent_sequence_decoded = str(nucleotide_seq.translate())  # Translate to amino acid sequence
    
else:
    df = pd.read_csv(f"{data_root_path}/{data}/scale2max/{data}.csv")
    with open(f"{data_root_path}/{data}/{data}.fasta", "r") as file:
        parent_sequence_decoded = file.readlines()[1].strip()
        
if data != "GB1":        
    muts = df["muts"].iloc[0]
else:
    muts = df["muts"].iloc[100000]

numbers = re.findall(r'\d+', muts)
mask_indices = list(map(int, numbers))
num_masks_per_sequence = num_samples // 4
num_to_generate_per_mask = 4


parent_sequence = torch.tensor(sequence_tokenizer.encode(parent_sequence_decoded, 
                                                            add_special_tokens=True), device=device).unsqueeze(0).long()
sequence_length = parent_sequence.shape[1]


all_masked_sequences = []
all_unmasked_sequences_decoded = []
all_unmasked_sequences = []
all_logps = []


while len(all_unmasked_sequences_decoded) < num_samples:
    
    print(len(all_unmasked_sequences_decoded))

    masked_sequences = parent_sequence.clone().repeat(num_to_generate_per_mask, 1)
    masked_sequences[:, mask_indices] = mask_token_sequence




    
    sequence_id = torch.ones((num_to_generate_per_mask, sequence_length), device=device).long() * 1
    
    structure_tokens = torch.ones((num_to_generate_per_mask, sequence_length), device=device).long() * 4096
    structure_tokens[:, 0] = 4098
    structure_tokens[:, -1] = 4097

    coords = torch.inf * torch.ones((num_to_generate_per_mask, sequence_length, 3, 3), device=device)
    
    average_plddt = torch.ones((num_to_generate_per_mask), device=device)

    per_res_plddt = torch.zeros((num_to_generate_per_mask, sequence_length), device=device)
    ss8_tokens = torch.zeros((num_to_generate_per_mask, sequence_length), device=device).long()
    sasa_tokens = torch.zeros((num_to_generate_per_mask, sequence_length), device=device).long()

    function_tokens = torch.zeros((num_to_generate_per_mask, sequence_length, 8), device=device).long()
    residue_annotation_tokens = torch.zeros((num_to_generate_per_mask, sequence_length, 16), device=device).long()




    with torch.no_grad():
        unmasked_sequences = sample_components_from_bidirectional_transformer(transformer_model=esm_model,
                                                                                masked_sequence_tokens=masked_sequences,
                                                                                structure_tokens=structure_tokens,
                                                                                average_plddt=average_plddt,
                                                                                per_res_plddt=per_res_plddt,
                                                                                ss8_tokens=ss8_tokens,
                                                                                sasa_tokens=sasa_tokens,
                                                                                function_tokens=function_tokens,
                                                                                residue_annotation_tokens=residue_annotation_tokens,
                                                                                bb_coords=coords,
                                                                                sequence_id=sequence_id,
                                                                                mask_token_sequence=mask_token_sequence,
                                                                                bos_token_sequence=bos_token_sequence,
                                                                                eos_token_sequence=eos_token_sequence,
                                                                                pad_token_sequence=pad_token_sequence,
                                                                                inference_batch_size=1)


        
        masked_indices = (masked_sequences == mask_token_sequence).float()
        logits = esm_model.nn(sequence_tokens=masked_sequences,
                                structure_tokens=structure_tokens,
                                average_plddt=average_plddt,
                                per_res_plddt=per_res_plddt,
                                ss8_tokens=ss8_tokens,
                                sasa_tokens=sasa_tokens,
                                function_tokens=function_tokens,
                                residue_annotation_tokens=residue_annotation_tokens,
                                sequence_id=sequence_id,
                                bb_coords=coords)["sequence_logits"].detach()
        logps = torch.nn.functional.log_softmax(logits/sampling_temperature, dim=-1)
        logps = torch.gather(logps, dim=-1, index=unmasked_sequences.unsqueeze(-1)).squeeze(-1)
        logps = (logps * masked_indices).sum(-1).detach()
        
        decoded_seqs = [sequence.replace(" ", "") for sequence in sequence_tokenizer.batch_decode(unmasked_sequences[:, 1:-1])]
        for seq, logp, masked_seq, unmasked_seq in zip(decoded_seqs, logps, masked_sequences, unmasked_sequences):
            if seq in all_unmasked_sequences_decoded:
                continue
            else:
                all_unmasked_sequences_decoded.append(seq)
                all_logps.append(logp)
                all_masked_sequences.append(masked_seq)
                all_unmasked_sequences.append(unmasked_seq)



all_unmasked_sequences_decoded = all_unmasked_sequences_decoded[:num_samples]
all_masked_sequences = all_masked_sequences[:num_samples]
all_unmasked_sequences = all_unmasked_sequences[:num_samples]
all_logps = all_logps[:num_samples]
    
all_masked_sequences = torch.stack(all_masked_sequences, dim=0)
all_unmasked_sequences = torch.stack(all_unmasked_sequences, dim=0)
all_logps = torch.stack(all_logps, dim=0)



to_save = {"parent_sequence": parent_sequence,
            "all_masked_sequences": all_masked_sequences,
            "all_unmasked_sequences": all_unmasked_sequences,
            "all_unmasked_sequences_decoded": all_unmasked_sequences_decoded,
            "all_logps": all_logps}
torch.save(to_save, f"{save_folder_name}/trpb_{replicate}.pt")