Spaces:
Running
Running
File size: 8,484 Bytes
e76b79a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 | # Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import pathlib
import string
import torch
from esm import Alphabet, FastaBatchedDataset, ProteinBertModel, pretrained, MSATransformer
import pandas as pd
from tqdm import tqdm
from Bio import SeqIO
import itertools
from typing import List, Tuple
import numpy as np
def remove_insertions(sequence: str) -> str:
""" Removes any insertions into the sequence. Needed to load aligned sequences in an MSA. """
# This is an efficient way to delete lowercase characters and insertion characters from a string
deletekeys = dict.fromkeys(string.ascii_lowercase)
deletekeys["."] = None
deletekeys["*"] = None
translation = str.maketrans(deletekeys)
return sequence.translate(translation)
def read_msa(filename: str, nseq: int) -> List[Tuple[str, str]]:
""" Reads the first nseq sequences from an MSA file, automatically removes insertions.
The input file must be in a3m format (although we use the SeqIO fasta parser)
for remove_insertions to work properly."""
msa = [
(record.description, remove_insertions(str(record.seq)))
for record in itertools.islice(SeqIO.parse(filename, "fasta"), nseq)
]
return msa
def create_parser():
parser = argparse.ArgumentParser(
description="Label a deep mutational scan with predictions from an ensemble of ESM-1v models." # noqa
)
# fmt: off
parser.add_argument(
"--model-location",
type=str,
help="PyTorch model file OR name of pretrained model to download (see README for models)",
nargs="+",
)
parser.add_argument(
"--sequence",
type=str,
help="Base sequence to which mutations were applied",
)
parser.add_argument(
"--dms-input",
type=pathlib.Path,
help="CSV file containing the deep mutational scan",
)
parser.add_argument(
"--mutation-col",
type=str,
default="mutant",
help="column in the deep mutational scan labeling the mutation as 'AiB'"
)
parser.add_argument(
"--dms-output",
type=pathlib.Path,
help="Output file containing the deep mutational scan along with predictions",
)
parser.add_argument(
"--offset-idx",
type=int,
default=0,
help="Offset of the mutation positions in `--mutation-col`"
)
parser.add_argument(
"--scoring-strategy",
type=str,
default="wt-marginals",
choices=["wt-marginals", "pseudo-ppl", "masked-marginals"],
help=""
)
parser.add_argument(
"--msa-path",
type=pathlib.Path,
help="path to MSA in a3m format (required for MSA Transformer)"
)
parser.add_argument(
"--msa-samples",
type=int,
default=400,
help="number of sequences to select from the start of the MSA"
)
# fmt: on
parser.add_argument("--nogpu", action="store_true", help="Do not use GPU even if available")
return parser
def label_row(row, sequence, token_probs, alphabet, offset_idx):
wt, idx, mt = row[0], int(row[1:-1]) - offset_idx, row[-1]
assert sequence[idx] == wt, "The listed wildtype does not match the provided sequence"
wt_encoded, mt_encoded = alphabet.get_idx(wt), alphabet.get_idx(mt)
# add 1 for BOS
score = token_probs[0, 1 + idx, mt_encoded] - token_probs[0, 1 + idx, wt_encoded]
return score.item()
def compute_pppl(row, sequence, model, alphabet, offset_idx):
wt, idx, mt = row[0], int(row[1:-1]) - offset_idx, row[-1]
assert sequence[idx] == wt, "The listed wildtype does not match the provided sequence"
# modify the sequence
sequence = sequence[:idx] + mt + sequence[(idx + 1) :]
# encode the sequence
data = [
("protein1", sequence),
]
batch_converter = alphabet.get_batch_converter()
batch_labels, batch_strs, batch_tokens = batch_converter(data)
wt_encoded, mt_encoded = alphabet.get_idx(wt), alphabet.get_idx(mt)
# compute probabilities at each position
log_probs = []
for i in range(1, len(sequence) - 1):
batch_tokens_masked = batch_tokens.clone()
batch_tokens_masked[0, i] = alphabet.mask_idx
with torch.no_grad():
token_probs = torch.log_softmax(model(batch_tokens_masked.cuda())["logits"], dim=-1)
log_probs.append(token_probs[0, i, alphabet.get_idx(sequence[i])].item()) # vocab size
return sum(log_probs)
def main(args):
# Load the deep mutational scan
df = pd.read_csv(args.dms_input)
# inference for each model
for model_location in args.model_location:
model, alphabet = pretrained.load_model_and_alphabet(model_location)
model.eval()
if torch.cuda.is_available() and not args.nogpu:
model = model.cuda()
print("Transferred model to GPU")
batch_converter = alphabet.get_batch_converter()
if isinstance(model, MSATransformer):
data = [read_msa(args.msa_path, args.msa_samples)]
assert (
args.scoring_strategy == "masked-marginals"
), "MSA Transformer only supports masked marginal strategy"
batch_labels, batch_strs, batch_tokens = batch_converter(data)
all_token_probs = []
for i in tqdm(range(batch_tokens.size(2))):
batch_tokens_masked = batch_tokens.clone()
batch_tokens_masked[0, 0, i] = alphabet.mask_idx # mask out first sequence
with torch.no_grad():
token_probs = torch.log_softmax(
model(batch_tokens_masked.cuda())["logits"], dim=-1
)
all_token_probs.append(token_probs[:, 0, i]) # vocab size
token_probs = torch.cat(all_token_probs, dim=0).unsqueeze(0)
df[model_location] = df.apply(
lambda row: label_row(
row[args.mutation_col], args.sequence, token_probs, alphabet, args.offset_idx
),
axis=1,
)
else:
data = [
("protein1", args.sequence),
]
batch_labels, batch_strs, batch_tokens = batch_converter(data)
if args.scoring_strategy == "wt-marginals":
with torch.no_grad():
token_probs = torch.log_softmax(model(batch_tokens.cuda())["logits"], dim=-1)
df[model_location] = df.apply(
lambda row: label_row(
row[args.mutation_col],
args.sequence,
token_probs,
alphabet,
args.offset_idx,
),
axis=1,
)
elif args.scoring_strategy == "masked-marginals":
all_token_probs = []
for i in tqdm(range(batch_tokens.size(1))):
batch_tokens_masked = batch_tokens.clone()
batch_tokens_masked[0, i] = alphabet.mask_idx
with torch.no_grad():
token_probs = torch.log_softmax(
model(batch_tokens_masked.cuda())["logits"], dim=-1
)
all_token_probs.append(token_probs[:, i]) # vocab size
token_probs = torch.cat(all_token_probs, dim=0).unsqueeze(0)
df[model_location] = df.apply(
lambda row: label_row(
row[args.mutation_col],
args.sequence,
token_probs,
alphabet,
args.offset_idx,
),
axis=1,
)
elif args.scoring_strategy == "pseudo-ppl":
tqdm.pandas()
df[model_location] = df.progress_apply(
lambda row: compute_pppl(
row[args.mutation_col], args.sequence, model, alphabet, args.offset_idx
),
axis=1,
)
df.to_csv(args.dms_output)
if __name__ == "__main__":
parser = create_parser()
args = parser.parse_args()
main(args)
|