omim.py

import os
import pandas as pd
import gzip
from datasets import GeneratorBasedBuilder, DatasetInfo, SplitGenerator, Split, BuilderConfig, Features, Value
from Bio import SeqIO
from typing import Dict, Any

class VepOmimConfig(BuilderConfig):
    def __init__(self, sequence_length=2048, fasta_path=None, **kwargs):
        super().__init__(**kwargs)
        self.sequence_length = sequence_length
        self.fasta_path = fasta_path

    @property
    def sequence_length(self):
        return self._sequence_length

    @sequence_length.setter
    def sequence_length(self, value):
        self._sequence_length = value

    @property
    def fasta_path(self):
        return self._fasta_path

    @fasta_path.setter
    def fasta_path(self, value):
        self._fasta_path = value

class VepOmimSplit(GeneratorBasedBuilder):
    BUILDER_CONFIG_CLASS = VepOmimConfig
    BUILDER_CONFIGS = [
        VepOmimConfig(name="default", sequence_length=2048, fasta_path=None)
    ]
    DEFAULT_CONFIG_NAME = "default"

    def _info(self):
        return DatasetInfo(
            features=Features({
                "ref_forward_sequence": Value("string"),
                "alt_forward_sequence": Value("string"),
                "label": Value("int32"),
                "chromosome": Value("string"),
                "position": Value("int32"),
                "ref": Value("string"),
                "alt": Value("string"),
                "consequence": Value("string"),
            })
        )

    def _split_generators(self, dl_manager):
        data_files = {
            "omim": "omim.csv",
        }
        downloaded_files = dl_manager.download(data_files)
        return [
            SplitGenerator(name="omim", gen_kwargs={"filepath": downloaded_files["omim"]}),  # type: ignore
        ]

    def _load_fasta_sequences(self, fasta_path: str) -> Dict[str, str]:
        """加载 fasta 序列到内存字典，支持 gzip 压缩"""
        sequences = {}
        if fasta_path.endswith('.gz'):
            with gzip.open(fasta_path, 'rt') as f:
                for record in SeqIO.parse(f, 'fasta'):
                    sequences[record.id] = str(record.seq)
        else:
            with open(fasta_path, 'r') as f:
                for record in SeqIO.parse(f, 'fasta'):
                    sequences[record.id] = str(record.seq)
        return sequences

    def _generate_examples(self, filepath: str):
        df = pd.read_csv(filepath)
        config: VepOmimConfig = self.config  # type: ignore
        seq_len = config.sequence_length
        fasta_path = config.fasta_path
        if fasta_path is None:
            raise ValueError("You must provide fasta_path when loading the dataset!")
        
        # 加载所有序列到内存
        sequences = self._load_fasta_sequences(fasta_path)
        
        for idx, row in df.iterrows():
            chrom = str(row['chromosome'])
            if not chrom.startswith('chr'):
                chrom = 'chr' + chrom
            
            if chrom not in sequences:
                raise ValueError(f"Chromosome {chrom} not found in fasta. Available: {list(sequences.keys())[:5]}...")
            
            pos = int(row['position'])
            ref = str(row['ref'])
            alt = str(row['alt'])
            half = seq_len // 2
            start = max(0, pos - half - 1)  # 0-based indexing
            end = pos + half - 1
            
            seq = sequences[chrom][start:end]
            seq_list = list(seq)
            center_idx = half
            ref_seq = seq_list.copy()
            ref_seq[center_idx] = ref
            ref_seq = ''.join(ref_seq)
            alt_seq = seq_list.copy()
            alt_seq[center_idx] = alt
            alt_seq = ''.join(alt_seq)
            yield idx, {
                "ref_forward_sequence": ref_seq,
                "alt_forward_sequence": alt_seq,
                "label": int(row["label"]),
                "chromosome": str(row["chromosome"]),
                "position": int(row["position"]),
                "consequence": row["consequence"],
                "ref": row["ref"],
                "alt": row["alt"]
            }