sequence_processor.py

"""
sequence_processor.py
=====================
Biological translation engine for DeepCRISPR.
Converts raw CRISPR guide RNA / DNA sequences into the 306-dimension
embedding format required by the Mega Model.

Strategy:
  1. If HuggingFace `transformers` + `torch` are installed, uses a
     pre-trained DNA language model (DNABERT-2 or similar) to generate
     learned embeddings, then projects to 306 dimensions.
  2. Otherwise, falls back to a deterministic k-mer frequency + positional
     encoding method that produces a reproducible 306-dim vector from
     any DNA string — no GPU or pip install required.

Author: Mujahid
"""

import pandas as pd
import numpy as np
import hashlib
import math
from typing import List, Union


# ────────────────────────────────────────────────────────────────────────────
# Public API
# ────────────────────────────────────────────────────────────────────────────

def extract_306_embeddings(raw_sequence: str) -> pd.DataFrame:
    """
    Convert a raw DNA/RNA sequence (or multiple sequences separated by
    newlines) into a DataFrame with exactly 306 embedding columns.

    Parameters
    ----------
    raw_sequence : str
        One or more DNA/RNA sequences, one per line.
        Valid characters: A, T/U, G, C, N (case-insensitive).

    Returns
    -------
    pd.DataFrame
        Shape (n_sequences, 306) with columns matching the Mega Model's
        expected feature names (loaded from mega_feature_importance.csv
        if available, otherwise emb_0 … emb_305).
    """
    sequences = _parse_sequences(raw_sequence)
    if not sequences:
        raise ValueError("No valid DNA/RNA sequences found in input.")

    # Try transformer-based embeddings first
    try:
        embeddings = _transformer_embeddings(sequences)
    except Exception:
        embeddings = _kmer_embeddings(sequences)

    # Build DataFrame with correct column names
    col_names = _get_column_names()
    return pd.DataFrame(embeddings, columns=col_names)


# ────────────────────────────────────────────────────────────────────────────
# Sequence Parsing
# ────────────────────────────────────────────────────────────────────────────

_VALID_BASES = set("ATUGCNatugcn")


def _parse_sequences(raw: str) -> List[str]:
    """Parse, clean and validate DNA/RNA sequences from raw text."""
    lines = raw.strip().split("\n")
    sequences = []
    for line in lines:
        seq = line.strip().upper().replace("U", "T")  # RNA → DNA
        # Skip FASTA headers and empty lines
        if not seq or seq.startswith(">"):
            continue
        # Remove whitespace and non-base characters
        seq = "".join(c for c in seq if c in "ATGCN")
        if len(seq) >= 10:  # Minimum viable guide length
            sequences.append(seq)
    return sequences


# ────────────────────────────────────────────────────────────────────────────
# Column Names (match mega_feature_importance.csv)
# ────────────────────────────────────────────────────────────────────────────

def _get_column_names() -> List[str]:
    """Get the 306 feature names from mega_feature_importance.csv or fallback."""
    import os
    fi_path = os.path.join(
        os.path.dirname(os.path.abspath(__file__)),
        "mega_feature_importance.csv"
    )
    try:
        fi_df = pd.read_csv(fi_path)
        names = fi_df.iloc[:, 0].tolist()
        return names[:306]
    except Exception:
        return [f"emb_{i}" for i in range(306)]


# ────────────────────────────────────────────────────────────────────────────
# Strategy 1: Transformer-Based Embeddings (if available)
# ────────────────────────────────────────────────────────────────────────────

def _transformer_embeddings(sequences: List[str]) -> np.ndarray:
    """
    Use a HuggingFace DNA language model to produce 306-dim embeddings.
    Raises ImportError / Exception if transformers/torch not available.
    """
    from transformers import AutoTokenizer, AutoModel
    import torch

    model_name = "zhihan1996/DNABERT-2-117M"
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
    model.eval()

    all_embeddings = []
    for seq in sequences:
        inputs = tokenizer(seq, return_tensors="pt", padding=True, truncation=True,
                           max_length=512)
        with torch.no_grad():
            outputs = model(**inputs)
        # Mean-pool token embeddings → single vector
        hidden = outputs.last_hidden_state.mean(dim=1).squeeze().numpy()
        # Project / pad / truncate to exactly 306 dimensions
        emb_306 = _project_to_306(hidden)
        all_embeddings.append(emb_306)

    return np.array(all_embeddings)


# ────────────────────────────────────────────────────────────────────────────
# Strategy 2: K-mer Frequency + Positional Encoding (fallback)
# ────────────────────────────────────────────────────────────────────────────

def _kmer_embeddings(sequences: List[str]) -> np.ndarray:
    """
    Deterministic embedding: combines k-mer frequencies (k=1..6) with
    positional encodings and a hash-based expansion to fill 306 dims.
    Fully reproducible, no dependencies beyond numpy.
    """
    all_embeddings = []
    for seq in sequences:
        features = []

        # ── 1) Mono/di/tri/tetra/penta/hexa-nucleotide frequencies ──
        bases = ["A", "T", "G", "C"]
        for k in range(1, 7):
            kmers = {}
            for i in range(len(seq) - k + 1):
                kmer = seq[i:i + k]
                if all(c in "ATGC" for c in kmer):
                    kmers[kmer] = kmers.get(kmer, 0) + 1
            total = sum(kmers.values()) or 1
            # Sorted k-mer frequencies
            all_kmers = _generate_kmers(bases, k)
            for km in all_kmers:
                features.append(kmers.get(km, 0) / total)

        # ── 2) Sequence-level statistics ──
        n = len(seq)
        gc = (seq.count("G") + seq.count("C")) / max(n, 1)
        features.extend([
            gc,                             # GC content
            1 - gc,                         # AT content
            n / 100.0,                      # Normalized length
            seq.count("N") / max(n, 1),     # N fraction
        ])

        # ── 3) Positional encoding of first 20 bases ──
        for pos in range(20):
            base = seq[pos] if pos < n else "N"
            one_hot = [1.0 if base == b else 0.0 for b in "ATGC"]
            features.extend(one_hot)

        # ── 4) Hash-based expansion to exactly 306 ──
        features = np.array(features, dtype=np.float64)
        emb_306 = _project_to_306(features)
        all_embeddings.append(emb_306)

    return np.array(all_embeddings)


def _generate_kmers(bases: List[str], k: int) -> List[str]:
    """Generate all possible k-mers from the given bases, sorted."""
    if k == 1:
        return sorted(bases)
    shorter = _generate_kmers(bases, k - 1)
    return sorted([s + b for s in shorter for b in bases])


# ────────────────────────────────────────────────────────────────────────────
# Projection helper
# ────────────────────────────────────────────────────────────────────────────

def _project_to_306(vec: np.ndarray) -> np.ndarray:
    """
    Deterministically project / pad / truncate a vector to exactly 306 dims.
    Uses a seeded random projection matrix for expansion.
    """
    target = 306
    vec = np.array(vec, dtype=np.float64).flatten()

    if len(vec) == target:
        return vec
    elif len(vec) > target:
        return vec[:target]
    else:
        # Expand using a deterministic hash-based projection
        result = np.zeros(target, dtype=np.float64)
        result[:len(vec)] = vec
        # Fill remaining dims with sinusoidal combinations of existing features
        for i in range(len(vec), target):
            seed_val = vec[i % len(vec)]
            result[i] = math.sin(seed_val * (i + 1) * 0.1) * 0.5
        return result