smftools 0.1.7__py3-none-any.whl → 0.2.3__py3-none-any.whl

smftools/informatics/basecalling.py

@@ -0,0 +1,67 @@
+import subprocess
+from pathlib import Path
+
+def canoncall(model_dir, model, pod5_dir, barcode_kit, bam, bam_suffix, barcode_both_ends=True, trim=False, device='auto'):
+    """
+    Wrapper function for dorado canonical base calling.
+
+    Parameters:
+        model_dir (str): a string representing the file path to the dorado basecalling model directory.
+        model (str): a string representing the the dorado basecalling model.
+        pod5_dir (str): a string representing the file path to the experiment directory containing the POD5 files.
+        barcode_kit (str): A string reppresenting the barcoding kit used in the experiment.
+        bam (str): File path to the BAM file to output.
+        bam_suffix (str): The suffix to use for the BAM file.
+        barcode_both_ends (bool): Whether to require a barcode detection on both ends for demultiplexing.
+        trim (bool): Whether to trim barcodes, adapters, and primers from read ends.
+        device (str): The device to use. 'auto' is default, which can detect device to use. Can also specify metal, cpu, cuda.
+    
+    Returns:
+        None
+            Outputs a BAM file holding the canonical base calls output by the dorado basecaller.
+    """
+    output = bam + bam_suffix
+    command = ["dorado", "basecaller", "--models-directory", model_dir, "--kit-name", barcode_kit, "--device", device, "--batchsize", "0"]
+    if barcode_both_ends:
+        command.append("--barcode-both-ends")
+    if not trim:
+        command.append("--no-trim")
+    command += [model, pod5_dir]
+    command_string = " ".join(command)
+    print(f"Running {command_string}\n to generate {output}")
+    with open(output, "w") as outfile:
+        subprocess.run(command, stdout=outfile)
+
+def modcall(model_dir, model, pod5_dir, barcode_kit, mod_list, bam, bam_suffix, barcode_both_ends=True, trim=False, device='auto'):
+    """
+    Wrapper function for dorado modified base calling.
+
+    Parameters:
+        model_dir (str): a string representing the file path to the dorado basecalling model directory.
+        model (str): a string representing the the dorado basecalling model.
+        pod5_dir (str): a string representing the file path to the experiment directory containing the POD5 files.
+        barcode_kit (str): A string representing the barcoding kit used in the experiment.
+        mod_list (list): A list of modification types to use in the analysis.
+        bam (str): File path to the BAM file to output.
+        bam_suffix (str): The suffix to use for the BAM file.
+        barcode_both_ends (bool): Whether to require a barcode detection on both ends for demultiplexing.
+        trim (bool): Whether to trim barcodes, adapters, and primers from read ends
+        device (str): Device to use for basecalling. auto, metal, cpu, cuda.
+    
+    Returns:
+        None
+            Outputs a BAM file holding the modified base calls output by the dorado basecaller.
+    """
+    import subprocess
+    output = bam + bam_suffix
+    command = ["dorado", "basecaller", "--models-directory", model_dir, "--kit-name", barcode_kit, "--modified-bases"]
+    command += mod_list
+    command += ["--device", device, "--batchsize", "0"]
+    if barcode_both_ends:
+        command.append("--barcode-both-ends")
+    if not trim:
+        command.append("--no-trim")
+    command += [model, pod5_dir]
+    print(f'Running: {" ".join(command)}')
+    with open(output, "w") as outfile:
+        subprocess.run(command, stdout=outfile)

smftools/informatics/bed_functions.py

@@ -0,0 +1,366 @@
+from pathlib import Path
+import os
+import subprocess
+from typing import List, Optional, Union
+import pysam
+import pybedtools
+import pyBigWig
+
+import numpy as np
+import pandas as pd
+import concurrent.futures
+from concurrent.futures import ProcessPoolExecutor
+
+import matplotlib.pyplot as plt
+
+from ..readwrite import make_dirs
+
+def _bed_to_bigwig(fasta: str, bed: str) -> str:
+    """
+    BED → bedGraph → bigWig
+    Requires:
+      - FASTA must have .fai index present
+    """
+
+    bed = Path(bed)
+    fa = Path(fasta)  # path to .fa
+    parent = bed.parent
+    stem = bed.stem
+    fa_stem = fa.stem
+    fai = parent / f"{fa_stem}.fai"
+
+    bedgraph = parent / f"{stem}.bedgraph"
+    bigwig = parent / f"{stem}.bw"
+
+    # 1) Compute coverage → bedGraph
+    print(f"[pybedtools] generating coverage bedgraph from {bed}")
+    bt = pybedtools.BedTool(str(bed))
+    # bedtools genomecov -bg
+    coverage = bt.genome_coverage(bg=True, genome=str(fai))
+    coverage.saveas(str(bedgraph))
+
+    # 2) Convert bedGraph → BigWig via pyBigWig
+    print(f"[pyBigWig] converting bedgraph → bigwig: {bigwig}")
+
+    # read chrom sizes from the FASTA .fai index
+    chrom_sizes = {}
+    with open(fai) as f:
+        for line in f:
+            fields = line.strip().split("\t")
+            chrom = fields[0]
+            size = int(fields[1])
+            chrom_sizes[chrom] = size
+
+    bw = pyBigWig.open(str(bigwig), "w")
+    bw.addHeader(list(chrom_sizes.items()))
+
+    with open(bedgraph) as f:
+        for line in f:
+            chrom, start, end, coverage = line.strip().split()
+            bw.addEntries(chrom, int(start), ends=int(end), values=float(coverage))
+
+    bw.close()
+
+    print(f"BigWig written: {bigwig}")
+    return str(bigwig)
+
+def _plot_bed_histograms(
+    bed_file,
+    plotting_directory,
+    fasta,
+    *,
+    bins=60,
+    clip_quantiles=(0.0, 0.995),
+    cov_bin_size=1000,       # coverage bin size in bp
+    rows_per_fig=6,          # paginate if many chromosomes
+    include_mapq_quality=True,   # add MAPQ + avg read quality columns to grid
+    coordinate_mode="one_based",  # "one_based" (your BED-like) or "zero_based"
+):
+    """
+    Plot per-chromosome QC grids from a BED-like file.
+
+    Expects columns:
+      chrom, start, end, read_len, qname, mapq, avg_base_qual
+
+    For each chromosome:
+      - Column 1: Read length histogram
+      - Column 2: Coverage across the chromosome (binned)
+      - (optional) Column 3: MAPQ histogram
+      - (optional) Column 4: Avg base quality histogram
+
+    The figure is paginated: rows = chromosomes (up to rows_per_fig), columns depend on include_mapq_quality.
+    Saves one PNG per page under `plotting_directory`.
+
+    Parameters
+    ----------
+    bed_file : str
+    plotting_directory : str
+    fasta : str
+        Reference FASTA (used to get chromosome lengths).
+    bins : int
+        Histogram bins for read length / MAPQ / quality.
+    clip_quantiles : (float, float)
+        Clip hist tails for readability (e.g., (0, 0.995)).
+    cov_bin_size : int
+        Bin size (bp) for coverage plot; bigger = faster/coarser.
+    rows_per_fig : int
+        Number of chromosomes per page.
+    include_mapq_quality : bool
+        If True, add MAPQ and avg base quality histograms as extra columns.
+    coordinate_mode : {"one_based","zero_based"}
+        One-based, inclusive (your file) vs BED-standard zero-based, half-open.
+    """
+    os.makedirs(plotting_directory, exist_ok=True)
+
+    bed_basename = os.path.basename(bed_file).rsplit(".bed", 1)[0]
+    print(f"[plot_bed_histograms] Loading: {bed_file}")
+
+    # Load BED-like table
+    cols = ['chrom', 'start', 'end', 'read_len', 'qname', 'mapq', 'avg_q']
+    df = pd.read_csv(bed_file, sep="\t", header=None, names=cols, dtype={
+        'chrom': str, 'start': int, 'end': int, 'read_len': int, 'qname': str,
+        'mapq': float, 'avg_q': float
+    })
+
+    # Drop unaligned records (chrom == '*') if present
+    df = df[df['chrom'] != '*'].copy()
+    if df.empty:
+        print("[plot_bed_histograms] No aligned reads found; nothing to plot.")
+        return
+
+    # Ensure coordinate mode consistent; convert to 0-based half-open for bin math internally
+    # Input is typically one_based inclusive (from your writer).
+    if coordinate_mode not in {"one_based", "zero_based"}:
+        raise ValueError("coordinate_mode must be 'one_based' or 'zero_based'")
+
+    if coordinate_mode == "one_based":
+        # convert to 0-based half-open [start0, end0)
+        start0 = df['start'].to_numpy() - 1
+        end0   = df['end'].to_numpy()   # inclusive in input -> +1 already handled by not subtracting
+    else:
+        # already 0-based half-open (assumption)
+        start0 = df['start'].to_numpy()
+        end0   = df['end'].to_numpy()
+
+    # Clip helper for hist tails
+    def _clip_series(s, q=(0.0, 0.995)):
+        if q is None:
+            return s.to_numpy()
+        lo = s.quantile(q[0]) if q[0] is not None else s.min()
+        hi = s.quantile(q[1]) if q[1] is not None else s.max()
+        x = s.to_numpy(dtype=float)
+        return np.clip(x, lo, hi)
+
+    # Load chromosome order/lengths from FASTA
+    with pysam.FastaFile(fasta) as fa:
+        ref_names = list(fa.references)
+        ref_lengths = dict(zip(ref_names, fa.lengths))
+
+    # Keep only chroms present in FASTA and with at least one read
+    chroms = [c for c in df['chrom'].unique() if c in ref_lengths]
+    # Order chromosomes by FASTA order
+    chrom_order = [c for c in ref_names if c in chroms]
+
+    if not chrom_order:
+        print("[plot_bed_histograms] No chromosomes from BED are present in FASTA; aborting.")
+        return
+
+    # Pagination
+    def _sanitize(name: str) -> str:
+        return "".join(ch if ch.isalnum() or ch in "-._" else "_" for ch in name)
+
+    cols_per_fig = 4 if include_mapq_quality else 2
+
+    for start_idx in range(0, len(chrom_order), rows_per_fig):
+        chunk = chrom_order[start_idx:start_idx + rows_per_fig]
+        nrows = len(chunk)
+        ncols = cols_per_fig
+
+        fig, axes = plt.subplots(
+            nrows=nrows, ncols=ncols,
+            figsize=(4.0 * ncols, 2.6 * nrows),
+            dpi=160,
+            squeeze=False
+        )
+
+        for r, chrom in enumerate(chunk):
+            chrom_len = ref_lengths[chrom]
+            mask = (df['chrom'].to_numpy() == chrom)
+
+            # Slice per-chrom arrays for speed
+            s0 = start0[mask]
+            e0 = end0[mask]
+            len_arr = df.loc[mask, 'read_len']
+            mapq_arr = df.loc[mask, 'mapq']
+            q_arr = df.loc[mask, 'avg_q']
+
+            # --- Col 1: Read length histogram (clipped) ---
+            ax = axes[r, 0]
+            ax.hist(_clip_series(len_arr, clip_quantiles), bins=bins, edgecolor="black", alpha=0.7)
+            if r == 0:
+                ax.set_title("Read length")
+            ax.set_ylabel(f"{chrom}\n(n={mask.sum()})")
+            ax.set_xlabel("bp")
+            ax.grid(alpha=0.25)
+
+            # --- Col 2: Coverage (binned over genome) ---
+            ax = axes[r, 1]
+            nb = max(1, int(np.ceil(chrom_len / cov_bin_size)))
+            # Bin edges in 0-based coords
+            edges = np.linspace(0, chrom_len, nb + 1, dtype=int)
+
+            # Compute per-bin "read count coverage": number of reads overlapping each bin.
+            # Approximate by incrementing all bins touched by the interval.
+            # (Fast and memory-light; for exact base coverage use smaller cov_bin_size.)
+            cov = np.zeros(nb, dtype=np.int32)
+            # bin indices overlapped by each read (0-based half-open)
+            b0 = np.minimum(np.searchsorted(edges, s0, side="right") - 1, nb - 1)
+            b1 = np.maximum(np.searchsorted(edges, np.maximum(e0 - 1, 0), side="right") - 1, 0)
+            # ensure valid ordering
+            b_lo = np.minimum(b0, b1)
+            b_hi = np.maximum(b0, b1)
+
+            # Increment all bins in range; loop but at bin resolution (fast for reasonable cov_bin_size).
+            for lo, hi in zip(b_lo, b_hi):
+                cov[lo:hi + 1] += 1
+
+            x_mid = (edges[:-1] + edges[1:]) / 2.0
+            ax.plot(x_mid, cov)
+            if r == 0:
+                ax.set_title(f"Coverage (~{cov_bin_size} bp bins)")
+            ax.set_xlim(0, chrom_len)
+            ax.set_xlabel("Position (bp)")
+            ax.set_ylabel("")  # already show chrom on col 1
+            ax.grid(alpha=0.25)
+
+            if include_mapq_quality:
+                # --- Col 3: MAPQ ---
+                ax = axes[r, 2]
+                # Clip MAPQ upper tail if needed (usually 60)
+                ax.hist(_clip_series(mapq_arr.fillna(0), clip_quantiles), bins=bins, edgecolor="black", alpha=0.7)
+                if r == 0:
+                    ax.set_title("MAPQ")
+                ax.set_xlabel("MAPQ")
+                ax.grid(alpha=0.25)
+
+                # --- Col 4: Avg base quality ---
+                ax = axes[r, 3]
+                ax.hist(_clip_series(q_arr.fillna(np.nan), clip_quantiles), bins=bins, edgecolor="black", alpha=0.7)
+                if r == 0:
+                    ax.set_title("Avg base qual")
+                ax.set_xlabel("Phred")
+                ax.grid(alpha=0.25)
+
+        fig.suptitle(
+            f"{bed_basename} — per-chromosome QC "
+            f"({'len,cov,MAPQ,qual' if include_mapq_quality else 'len,cov'})",
+            y=0.995, fontsize=11
+        )
+        fig.tight_layout(rect=[0, 0, 1, 0.98])
+
+        page = start_idx // rows_per_fig + 1
+        out_png = os.path.join(plotting_directory, f"{_sanitize(bed_basename)}_qc_page{page}.png")
+        plt.savefig(out_png, bbox_inches="tight")
+        plt.close(fig)
+
+    print("[plot_bed_histograms] Done.")
+
+def aligned_BAM_to_bed(aligned_BAM, out_dir, fasta, make_bigwigs, threads=None):
+    """
+    Takes an aligned BAM as input and writes a BED file of reads as output.
+    Bed columns are: Record name, start position, end position, read length, read name, mapping quality, read quality.
+
+    Parameters:
+        aligned_BAM (str): Path to an input aligned_BAM to extract to a BED file.
+        out_dir (str): Directory to output files.
+        fasta (str): File path to the reference genome.
+        make_bigwigs (bool): Whether to generate bigwig files.
+        threads (int): Number of threads to use.
+
+    Returns:
+        None
+    """
+    threads = threads or os.cpu_count()  # Use max available cores if not specified
+
+    # Create necessary directories
+    plotting_dir = out_dir / "bed_cov_histograms"
+    bed_dir = out_dir / "beds"
+    make_dirs([plotting_dir, bed_dir])
+
+    bed_output = bed_dir /  str(aligned_BAM.name).replace(".bam", "_bed.bed")
+
+    print(f"Creating BED-like file from BAM (with MAPQ and avg base quality): {aligned_BAM}")
+
+    with pysam.AlignmentFile(aligned_BAM, "rb") as bam, open(bed_output, "w") as out:
+        for read in bam.fetch(until_eof=True):
+            if read.is_unmapped:
+                chrom = "*"
+                start1 = 1
+                rl = read.query_length or 0
+                mapq = 0
+            else:
+                chrom = bam.get_reference_name(read.reference_id)
+                # pysam reference_start is 0-based → +1 for 1-based SAM-like start
+                start1 = int(read.reference_start) + 1
+                rl = read.query_length or 0
+                mapq = int(read.mapping_quality)
+
+            # End position in 1-based inclusive coords
+            end1 = start1 + (rl or 0) - 1
+
+            qname = read.query_name
+            quals = read.query_qualities
+            if quals is None or rl == 0:
+                avg_q = float("nan")
+            else:
+                avg_q = float(np.mean(quals))
+
+            out.write(f"{chrom}\t{start1}\t{end1}\t{rl}\t{qname}\t{mapq}\t{avg_q:.3f}\n")
+
+    print(f"BED-like file created: {bed_output}")
+
+    def split_bed(bed):
+        """Splits into aligned and unaligned reads (chrom == '*')."""
+        bed = str(bed)
+        aligned = bed.replace(".bed", "_aligned.bed")
+        unaligned = bed.replace(".bed", "_unaligned.bed")
+        with open(bed, "r") as infile, open(aligned, "w") as aligned_out, open(unaligned, "w") as unaligned_out:
+            for line in infile:
+                (unaligned_out if line.startswith("*\t") else aligned_out).write(line)
+        os.remove(bed)
+        return aligned
+
+    print(f"Splitting: {bed_output}")
+    aligned_bed = split_bed(bed_output)
+
+    with ProcessPoolExecutor() as executor:
+        futures = []
+        futures.append(executor.submit(_plot_bed_histograms, aligned_bed, plotting_dir, fasta))
+        if make_bigwigs:
+            futures.append(executor.submit(_bed_to_bigwig, fasta, aligned_bed))
+        concurrent.futures.wait(futures)
+
+    print("Processing completed successfully.")
+
+def extract_read_lengths_from_bed(file_path):
+    """
+    Load a dict of read names that points to the read length
+
+    Params:
+        file_path (str): file path to a bed file
+    Returns:
+        read_dict (dict)
+    """
+    import pandas as pd
+    columns = ['chrom', 'start', 'end', 'length', 'name']
+    df = pd.read_csv(file_path, sep='\t', header=None, names=columns, comment='#')
+    read_dict = {}
+    for _, row in df.iterrows():
+        chrom = row['chrom']
+        start = row['start']
+        end = row['end']
+        name = row['name']
+        length = row['length']
+        read_dict[name] = length
+
+    return read_dict

smftools/informatics/binarize_converted_base_identities.py

@@ -0,0 +1,172 @@
+def binarize_converted_base_identities(base_identities, strand, modification_type, bam, device='cpu', deaminase_footprinting=False, mismatch_trend_per_read={}, on_missing="nan"):
+    """
+    Efficiently binarizes conversion SMF data within a sequence string using NumPy arrays.
+
+    Parameters:
+        base_identities (dict): A dictionary returned by extract_base_identities. Keyed by read name. Points to a list of base identities.
+        strand (str): A string indicating which strand was converted in the experiment (options are 'top' and 'bottom').
+        modification_type (str): A string indicating the modification type of interest (options are '5mC' and '6mA').
+        bam (str): The bam file path
+        deaminase_footprinting (bool): Whether direct deaminase footprinting chemistry was used.
+        mismatch_trend_per_read (dict): For deaminase footprinting, indicates the type of conversion relative to the top strand reference for each read. (C->T or G->A if bottom strand was converted)
+        on_missing (str): Error handling if a read is missing
+        
+    Returns:
+        dict: A dictionary where 1 represents a methylated site, 0 represents an unmethylated site, and NaN represents a site without methylation info.
+        If deaminase_footprinting, 1 represents deaminated sites, while 0 represents non-deaminated sites.
+    """
+    import numpy as np
+
+    if mismatch_trend_per_read is None:
+        mismatch_trend_per_read = {}
+
+    # Fast path
+    if modification_type == "unconverted" and not deaminase_footprinting:
+        return {k: np.full(len(v), np.nan, dtype=np.float32) for k, v in base_identities.items()}
+
+    out = {}
+
+    if deaminase_footprinting:
+        valid_trends = {"C->T", "G->A"}
+
+        for read_id, bases in base_identities.items():
+            trend_raw = mismatch_trend_per_read.get(read_id, None)
+            if trend_raw is None:
+                if on_missing == "error":
+                    raise KeyError(f"Missing mismatch trend for read '{read_id}'")
+                out[read_id] = np.full(len(bases), np.nan, dtype=np.float32)
+                continue
+
+            trend = trend_raw.replace(" ", "").upper()
+            if trend not in valid_trends:
+                if on_missing == "error":
+                    raise KeyError(
+                        f"Invalid mismatch trend '{trend_raw}' for read '{read_id}'. "
+                        f"Expected one of {sorted(valid_trends)}"
+                    )
+                out[read_id] = np.full(len(bases), np.nan, dtype=np.float32)
+                continue
+
+            arr = np.asarray(bases, dtype="<U1")
+            res = np.full(arr.shape, np.nan, dtype=np.float32)
+
+            if trend == "C->T":
+                # C (unconverted) -> 0, T (converted) -> 1
+                res[arr == "C"] = 0.0
+                res[arr == "T"] = 1.0
+            else:  # "G->A"
+                res[arr == "G"] = 0.0
+                res[arr == "A"] = 1.0
+
+            out[read_id] = res
+
+        return out
+
+    # Non-deaminase mapping (bisulfite-style for 5mC; 6mA mapping is protocol dependent)
+    bin_maps = {
+        ("top", "5mC"):    {"C": 1.0, "T": 0.0},
+        ("bottom", "5mC"): {"G": 1.0, "A": 0.0},
+        ("top", "6mA"):    {"A": 1.0, "G": 0.0},
+        ("bottom", "6mA"): {"T": 1.0, "C": 0.0},
+    }
+    key = (strand, modification_type)
+    if key not in bin_maps:
+        raise ValueError(f"Invalid combination of strand='{strand}' and modification_type='{modification_type}'")
+
+    base_map = bin_maps[key]
+
+    for read_id, bases in base_identities.items():
+        arr = np.asarray(bases, dtype="<U1")
+        res = np.full(arr.shape, np.nan, dtype=np.float32)
+        # mask-assign; unknown characters (N, -, etc.) remain NaN
+        for b, v in base_map.items():
+            res[arr == b] = v
+        out[read_id] = res
+
+    return out
+
+    # if mismatch_trend_per_read is None:
+    #     mismatch_trend_per_read = {}
+
+    # # If the modification type is 'unconverted', return NaN for all positions if the deaminase_footprinting strategy is not being used.
+    # if modification_type == "unconverted" and not deaminase_footprinting:
+    #     #print(f"Skipping binarization for unconverted {strand} reads on bam: {bam}.")
+    #     return {key: np.full(len(bases), np.nan) for key, bases in base_identities.items()}
+
+    # # Define mappings for binarization based on strand and modification type
+    # if deaminase_footprinting:
+    #     binarization_maps = {
+    #         ('C->T'): {'C': 0, 'T': 1},
+    #         ('G->A'): {'G': 0, 'A': 1},
+    #     }
+
+    #     binarized_base_identities = {}
+    #     for key, bases in base_identities.items():
+    #         arr = np.array(bases, dtype='<U1')
+    #         # Fetch the appropriate mapping
+    #         conversion_type = mismatch_trend_per_read[key]
+    #         base_map = binarization_maps.get(conversion_type, None)
+    #         binarized = np.vectorize(lambda x: base_map.get(x, np.nan))(arr)  # Apply mapping with fallback to NaN
+    #         binarized_base_identities[key] = binarized
+
+    #     return binarized_base_identities
+    
+    # else:
+    #     binarization_maps = {
+    #         ('top', '5mC'): {'C': 1, 'T': 0},
+    #         ('top', '6mA'): {'A': 1, 'G': 0},
+    #         ('bottom', '5mC'): {'G': 1, 'A': 0},
+    #         ('bottom', '6mA'): {'T': 1, 'C': 0}
+    #     }
+
+    #     if (strand, modification_type) not in binarization_maps:
+    #         raise ValueError(f"Invalid combination of strand='{strand}' and modification_type='{modification_type}'")
+
+    #     # Fetch the appropriate mapping
+    #     base_map = binarization_maps[(strand, modification_type)]
+
+    #     binarized_base_identities = {}
+    #     for key, bases in base_identities.items():
+    #         arr = np.array(bases, dtype='<U1')
+    #         binarized = np.vectorize(lambda x: base_map.get(x, np.nan))(arr)  # Apply mapping with fallback to NaN
+    #         binarized_base_identities[key] = binarized
+
+    #     return binarized_base_identities
+    # import torch
+
+    # # If the modification type is 'unconverted', return NaN for all positions
+    # if modification_type == "unconverted":
+    #     print(f"Skipping binarization for unconverted {strand} reads on bam: {bam}.")
+    #     return {key: torch.full((len(bases),), float('nan'), device=device) for key, bases in base_identities.items()}
+
+    # # Define mappings for binarization based on strand and modification type
+    # binarization_maps = {
+    #     ('top', '5mC'): {'C': 1, 'T': 0},
+    #     ('top', '6mA'): {'A': 1, 'G': 0},
+    #     ('bottom', '5mC'): {'G': 1, 'A': 0},
+    #     ('bottom', '6mA'): {'T': 1, 'C': 0}
+    # }
+
+    # if (strand, modification_type) not in binarization_maps:
+    #     raise ValueError(f"Invalid combination of strand='{strand}' and modification_type='{modification_type}'")
+
+    # # Fetch the appropriate mapping
+    # base_map = binarization_maps[(strand, modification_type)]
+    
+    # # Convert mapping to tensor
+    # base_keys = list(base_map.keys())
+    # base_values = torch.tensor(list(base_map.values()), dtype=torch.float32, device=device)
+
+    # # Create a lookup dictionary (ASCII-based for fast mapping)
+    # lookup_table = torch.full((256,), float('nan'), dtype=torch.float32, device=device)
+    # for k, v in zip(base_keys, base_values):
+    #     lookup_table[ord(k)] = v
+
+    # # Process reads
+    # binarized_base_identities = {}
+    # for key, bases in base_identities.items():
+    #     bases_tensor = torch.tensor([ord(c) for c in bases], dtype=torch.uint8, device=device)  # Convert chars to ASCII
+    #     binarized = lookup_table[bases_tensor]  # Efficient lookup
+    #     binarized_base_identities[key] = binarized
+
+    # return binarized_base_identities