deepbind.py

"""
/*
Copyright (c) 2023, thewall.
All rights reserved.
BSD 3-clause license:
Redistribution and use in source and binary forms,
with or without modification, are permitted provided
that the following conditions are met:
  1. Redistributions of source code must retain the
     above copyright notice, this list of conditions
     and the following disclaimer.
  2. Redistributions in binary form must reproduce
     the above copyright notice, this list of conditions
     and the following disclaimer in the documentation
     and/or other materials provided with the distribution.
  3. Neither the name of the copyright holder nor the
     names of its contributors may be used to endorse or
     promote products derived from this software without
     specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
"""
import os
import datasets
import torch
from torch import nn
import torch.nn.functional as F
import numpy as np
import pandas as pd
from typing import List
from functools import partial

DEEPBIND_MODEL_CONFIG = datasets.load_dataset(path="thewall/deepbindweight", split="all")
SELEX_CONFIG = pd.read_excel(DEEPBIND_MODEL_CONFIG[0]['selex'], index_col=0)

class DeepBind(nn.Module):
    ALPHABET = "ATGCN"
    ALPHABET_MAP = {key: i for i, key in enumerate(ALPHABET)}
    ALPHABET_MAP["U"] = 1
    ALPHABET_COMPLEMENT = "TACGN"
    COMPLEMENT_ID_MAP = torch.IntTensor([1, 0, 3, 2, 4])

    def __init__(self, reverse_complement=True, num_detectors=16, detector_len=24, has_avg_pooling=True, num_hidden=1,
                 tokenizer=None):
        super(DeepBind, self).__init__()
        self.reverse_complement = reverse_complement
        self.num_detectors = num_detectors
        self.detector_len = detector_len
        self.has_avg_pooling = has_avg_pooling
        self.num_hidden = num_hidden
        self.build_embedding()
        self.detectors = nn.Conv1d(4, num_detectors, detector_len)
        if has_avg_pooling:
            self.avg_pool = nn.AvgPool1d(detector_len)
        self.max_pool = nn.MaxPool1d(detector_len)
        fcs = [nn.Linear(num_detectors * 2 if self.has_avg_pooling else num_detectors, num_hidden)]
        if num_hidden > 1:
            fcs.append(nn.ReLU())
            fcs.append(nn.Linear(num_hidden, 1))
        self.fc = nn.Sequential(*fcs)
        self.tokenizer = tokenizer if tokenizer is not None else self.get_tokenizer()

    @classmethod
    def get_tokenizer(cls):
        from tokenizers import Tokenizer, models, decoders
        tokenizer = Tokenizer(models.BPE(vocab=cls.ALPHABET_MAP, merges=[]))
        tokenizer.decoder = decoders.ByteLevel()
        return tokenizer

    @classmethod
    def complement_idxs_encode_batch(cls, idxs, reverse=False):
        return torch.stack(list(map(partial(cls.complement_idxs_encode, reverse=reverse), idxs)))

    @classmethod
    def complement_idxs_encode(cls, idxs, reverse=False):
        if reverse:
            idxs = reversed(idxs)
        return cls.COMPLEMENT_ID_MAP[idxs]

    def build_embedding(self):
        """ATGC->ACGT:0321"""
        embedding = torch.zeros(5, 4)
        embedding[0, 0] = 1
        embedding[1, 3] = 1
        embedding[2, 2] = 1
        embedding[3, 1] = 1
        embedding[-1] = 0.25
        self.embedding = nn.Embedding.from_pretrained(embedding, freeze=True)
        return embedding

    @property
    def device(self):
        return self.detectors.bias.device

    def _load_detector(self, fobj):
        # dtype = functools.partial(lambda x:torch.Tensor(eval(x))
        dtype = lambda x: torch.Tensor(eval(x))
        weight1 = self._load_param(fobj, "detectors", dtype).reshape(self.detector_len, 4, self.num_detectors)
        biases1 = self._load_param(fobj, "thresholds", dtype)
        # Tx4xC->Cx4xT
        self.detectors.weight.data = weight1.permute(2, 1, 0).contiguous().to(device=self.detectors.weight.device)
        self.detectors.bias.data = biases1.to(device=self.detectors.bias.device)

    def _load_fc1(self, fobj):
        num_hidden1 = self.num_detectors * 2 if self.has_avg_pooling else self.num_detectors
        dtype = lambda x: torch.Tensor(np.array(eval(x)))
        weight1 = self._load_param(fobj, "weights1", dtype).reshape(num_hidden1, self.num_hidden)
        biases1 = self._load_param(fobj, "biases1", dtype)
        self.fc[0].weight.data = weight1.T.contiguous().to(device=self.fc[0].weight.device)
        self.fc[0].bias.data = biases1.to(device=self.fc[0].bias.device)

    def _load_fc2(self, fobj):
        dtype = lambda x: torch.Tensor(np.array(eval(x)))
        weight2 = self._load_param(fobj, "weights2", dtype)
        biases2 = self._load_param(fobj, "biases2", dtype)
        assert not (weight2 is None and self.num_hidden > 1)
        assert not (biases2 is None and self.num_hidden > 1)
        if self.num_hidden > 1:
            self.fc[2].weight.data = weight2.reshape(1, -1).to(device=self.fc[2].weight.device)
            self.fc[2].bias.data = biases2.to(device=self.fc[2].bias.device)

    @classmethod
    def _load_param(cls, fobj, param_name, dtype):
        line = fobj.readline().strip()
        tmp = line.split("=")
        assert tmp[0].strip() == param_name
        if len(tmp) > 1 and len(tmp[1].strip()) > 0:
            return dtype(tmp[1].strip())

    @classmethod
    def load_model(cls, sra_id="ERR173157", file=None, ID=None):
        if file is None:
            if ID is None:
                data = SELEX_CONFIG
                ID = data.loc[sra_id]["ID"]
            file = os.path.join(DEEPBIND_MODEL_CONFIG['config'][0], f"{ID}.txt")
        keys = [("reverse_complement", lambda x: bool(eval(x))), ("num_detectors", int), ("detector_len", int),
                ("has_avg_pooling", lambda x: bool(eval(x))), ("num_hidden", int)]

        hparams = {}
        with open(file) as fobj:
            version = fobj.readline()[1:].strip()
            for key in keys:
                value = cls._load_param(fobj, key[0], key[1])
                hparams[key[0]] = value
            if hparams['num_hidden'] == 0:
                hparams['num_hidden'] = 1
            model = cls(**hparams)
            model._load_detector(fobj)
            model._load_fc1(fobj)
            model._load_fc2(fobj)
            print(f"load model from {file}")
            return model

    def inference(self, sequence: List[str], window_size=0, average_flag=False):
        if isinstance(sequence, str):
            sequence = [sequence]
        ans = []
        self.tokenizer.no_padding()
        for seq in sequence:
            inputs = torch.IntTensor(self.tokenizer.encode(seq).ids).unsqueeze(0).to(device=self.device)
            score = self.test(inputs, window_size, average_flag).item()
            ans.append(score)
        return ans

    @torch.no_grad()
    def batch_inference(self, sequences: List[str], window_size=0, average_flag=False):
        if isinstance(sequences, str):
            sequences = [sequences]
        self.tokenizer.enable_padding()
        encodings = self.tokenizer.encode_batch(sequences)
        ids = torch.Tensor([encoding.ids for encoding in encodings]).to(device=self.device)
        mask = torch.BoolTensor([encoding.attention_mask for encoding in encodings]).to(device=self.device)
        seq_len = mask.sum(dim=1)
        score = self.batch_scan_model(ids, seq_len, window_size, average_flag)
        if self.reverse_complement:
            rev_seq = self.complement_idxs_encode_batch(ids.cpu().long(), reverse=True)
            rev_seq = torch.Tensor(rev_seq).to(device=self.device).float()
            rev_score = self.batch_scan_model(rev_seq, seq_len, window_size, average_flag)
            score = torch.stack([rev_score, score], dim=-1).max(dim=-1)[0]
        return score.cpu().tolist()

    def batch_scan_model(self, ids, seq_len, window_size: int = 0, average_flag: bool = False):
        if window_size < 1:
            window_size = int(self.detector_len * 1.5)
        scores = torch.zeros_like(seq_len).float()
        masked = seq_len <= window_size
        for idx in torch.where(masked)[0]:
            scores[idx] = self.forward(ids[idx:idx + 1, :seq_len[idx]].int())
        if torch.all(masked):
            return scores
        fold_ids = F.unfold(ids[~masked].unsqueeze(1).unsqueeze(1), kernel_size=(1, window_size), stride=1)
        B, W, G = fold_ids.shape
        fold_ids = fold_ids.permute(0, 2, 1).reshape(-1, W)
        ans = self.forward(fold_ids.int())
        ans = ans.reshape(B, G)
        if average_flag:
            valid_len = seq_len - window_size + 1
            for idx, value in zip(torch.where(~masked)[0], ans):
                scores[idx] = value[:valid_len[idx]].mean()
        else:
            unvalid_mask = torch.arange(G).unsqueeze(0).to(seq_len.device) >= (
                        seq_len[~masked] - window_size + 1).unsqueeze(1)
            ans[unvalid_mask] = -torch.inf
            scores[~masked] = ans.max(dim=1)[0]
        return scores

    @torch.no_grad()
    def test(self, seq: torch.IntTensor, window_size=0, average_flag=False):
        score = self.scan_model(seq, window_size, average_flag)
        if self.reverse_complement:
            rev_seq = self.complement_idxs_encode_batch(seq.cpu().long(), reverse=True)
            rev_seq = torch.IntTensor(rev_seq).to(device=seq.device)
            rev_score = self.scan_model(rev_seq, window_size, average_flag)
            score = torch.cat([rev_score, score], dim=-1).max(dim=-1)[0]
        return score

    def scan_model(self, seq: torch.IntTensor, window_size: int = 0, average_flag: bool = False):
        seq_len = seq.shape[1]
        if window_size < 1:
            window_size = int(self.detector_len * 1.5)
        if seq_len <= window_size:
            return self.forward(seq)
        else:
            scores = []
            for i in range(0, seq_len - window_size + 1):
                scores.append(self.forward(seq[:, i:i + window_size]))
            scores = torch.stack(scores, dim=-1)
            if average_flag:
                return scores.mean(dim=-1)
            else:
                return scores.max(dim=-1)[0]

    def forward(self, seq: torch.IntTensor):
        seq = F.pad(seq, (self.detector_len - 1, self.detector_len - 1), value=4)
        x = self.embedding(seq)
        x = x.permute(0, 2, 1)
        x = self.detectors(x)
        x = torch.relu(x)
        x = x.permute(0, 2, 1)
        if self.has_avg_pooling:
            x = torch.stack([torch.max(x, dim=1)[0], torch.mean(x, dim=1)], dim=-1)
            x = torch.flatten(x, 1)
        else:
            x = torch.max(x, dim=1)[0]
        x = x.squeeze(dim=-1)
        x = self.fc(x)
        return x


if __name__ == "__main__":
    """
       AGGUAAUAAUUUGCAUGAAAUAACUUGGAGAGGAUAGC
       AGACAGAGCUUCCAUCAGCGCUAGCAGCAGAGACCAUU
       GAGGTTACGCGGCAAGATAA
       TACCACTAGGGGGCGCCACC

    To generate 16 predictions (4 models, 4 sequences), run
    the deepbind executable as follows:

       % deepbind example.ids < example.seq
       D00210.001   D00120.001   D00410.003   D00328.003
        7.451420    -0.166146    -0.408751    -0.026180
       -0.155398     4.113817     0.516956    -0.248167
       -0.140683     0.181295     5.885349    -0.026180
       -0.174985    -0.152521    -0.379695    17.682623
    """
    sequences = ["AGGUAAUAAUUUGCAUGAAAUAACUUGGAGAGGAUAGC",
                 "AGACAGAGCUUCCAUCAGCGCUAGCAGCAGAGACCAUU",
                 "GAGGTTACGCGGCAAGATAA",
                 "TACCACTAGGGGGCGCCACC"]
    model = DeepBind.load_model(ID='D00410.003')
    print(model.batch_inference(sequences))

    import random
    import time
    from tqdm import tqdm

    sequences = ["".join([random.choice("ATGC") for _ in range(40)]) for i in range(1000)]


    def test_fn(sequences, fn):
        start_time = time.time()
        for start in tqdm(range(0, len(sequences), 256)):
            batch = sequences[start: min(start + 256, len(sequences))]
            fn(batch)
        print(time.time() - start_time)


    # test_fn(sequences, model.inference)
    # test_fn(sequences, model.batch_inference)
    model = model.cuda()
    test_fn(sequences, model.batch_inference)
    test_fn(sequences, model.inference)
    test_fn(sequences, model.batch_inference)
    test_fn(sequences, model.inference)