biometal

ARM-native bioinformatics library with streaming architecture and evidence-based optimization

What Makes biometal Different?

Stream data directly from networks and analyze terabyte-scale datasets on consumer hardware without downloading.

• Constant ~5 MB memory regardless of dataset size (99.5% reduction)
• 16-25× speedup using ARM NEON SIMD on Apple Silicon
• Network streaming from HTTP/HTTPS sources (no download needed)
• Evidence-based design (1,357 experiments, 40,710 measurements)

Quick Start

Installation

Rust:

[dependencies]
biometal = "1.2"

Python:

pip install biometal-rs  # Install
python -c "import biometal; print(biometal.__version__)"  # Test

Note: Package is biometal-rs on PyPI, but imports as biometal in Python.

Basic Usage

Rust:

use biometal::FastqStream;

// Stream FASTQ with constant memory (~5 MB)
let stream = FastqStream::from_path("dataset.fq.gz")?;

for record in stream {
    let record = record?;
    // Process one record at a time
}

Python:

import biometal

# Stream FASTQ with constant memory (~5 MB)
stream = biometal.FastqStream.from_path("dataset.fq.gz")

for record in stream:
    # ARM NEON accelerated (16-25× speedup)
    gc = biometal.gc_content(record.sequence)
    counts = biometal.count_bases(record.sequence)
    mean_q = biometal.mean_quality(record.quality)

📚 Documentation

• 📖 Comprehensive Docs - DeepWiki AI-assisted documentation
• 📓 Interactive Tutorials - Jupyter notebooks with real workflows
• 🦀 API Reference - Full Rust documentation
• 🐍 Python Guide - Python-specific documentation
• 📐 Architecture - Technical design details
• 📊 Benchmarks - Performance analysis
• ❓ FAQ - Frequently asked questions

📓 Interactive Tutorials

Learn biometal through hands-on Jupyter notebooks (5 complete, ~2.5 hours):

👉 Browse all tutorials →

🚀 Key Features

Streaming Architecture

• Constant ~5 MB memory regardless of dataset size
• Analyze 5TB datasets on laptops without downloading
• 99.5% memory reduction vs. traditional approaches

ARM-Native Performance

• 16-25× speedup using ARM NEON SIMD
• Optimized for Apple Silicon (M1/M2/M3/M4)
• Automatic scalar fallback on x86_64

Network Streaming

• Stream directly from HTTP/HTTPS (no download)
• Smart LRU caching + background prefetching
• Access public data (ENA, S3, GCS, Azure)

Operations Library

• Core operations: GC content, base counting, quality scores
• K-mer operations: Extraction, minimizers, spectrum (v1.1.0)
• QC operations: Trimming, filtering, masking (v1.2.0)
• BAM/SAM parser: Production-ready with 4× speedup via parallel BGZF (Nov 8, 2025)
  • 4.54 million records/sec throughput
  • 43.0 MiB/s compressed file processing
  • Constant ~5 MB memory (streams terabyte-scale alignments)
  • Python bindings (v1.3.0): CIGAR operations, SAM writing, alignment metrics
• 40+ Python functions for bioinformatics workflows

Performance Highlights

Platform Support

Test count includes 354 core library + 70 BAM/SAM parser tests

Evidence-Based Design

biometal's design is grounded in comprehensive experimental validation:

• 1,357 experiments (40,710 measurements, N=30)
• Statistical rigor (95% CI, Cohen's d effect sizes)
• Full methodology: apple-silicon-bio-bench
• 6 optimization rules documented in OPTIMIZATION_RULES.md

Roadmap

v1.0.0 (Released Nov 5, 2025) ✅ - Core library + network streaming v1.1.0 (Released Nov 6, 2025) ✅ - K-mer operations v1.2.0 (Released Nov 6, 2025) ✅ - Python bindings for Phase 4 QC BAM/SAM (Integrated Nov 8, 2025) ✅ - Native streaming alignment parser with parallel BGZF (4× speedup)

v1.3.0 (In Development) - Python BAM bindings with CIGAR operations and SAM writing

Next (Planned):

• Complete tag parsing (extended types from Phase 1)
• BAI/CSI index support (random access)
• Additional alignment statistics

Future (Community Driven):

• Extended operations (alignment, assembly)
• Additional formats (VCF, BCF, CRAM)
• Metal GPU acceleration (Mac-specific)

See CHANGELOG.md for detailed release notes.

Mission: Democratizing Bioinformatics

biometal addresses barriers that lock researchers out of genomics:

1. Economic: Consumer ARM laptops ($1,400) deliver production performance
2. Environmental: ARM efficiency reduces carbon footprint
3. Portability: Works across ARM ecosystem (Mac, Graviton, Ampere, RPi)
4. Data Access: Analyze 5TB datasets on 24GB laptops without downloading

Example Use Cases

Quality Control Pipeline

import biometal

stream = biometal.FastqStream.from_path("raw_reads.fq.gz")

for record in stream:
    # Trim low-quality ends
    trimmed = biometal.trim_quality_window(record, min_quality=20, window_size=4)

    # Length filter
    if biometal.meets_length_requirement(trimmed, min_len=50, max_len=150):
        # Mask remaining low-quality bases
        masked = biometal.mask_low_quality(trimmed, min_quality=20)

        # Check masking rate
        mask_rate = biometal.count_masked_bases(masked) / len(masked.sequence)
        if mask_rate < 0.1:
            # Pass QC - process further
            pass

K-mer Extraction for ML

import biometal

# Extract k-mers for DNABert preprocessing
stream = biometal.FastqStream.from_path("dataset.fq.gz")

for record in stream:
    # Extract overlapping k-mers (k=6 typical for DNABert)
    kmers = biometal.extract_kmers(record.sequence, k=6)

    # Format for transformer models
    kmer_string = " ".join(kmer.decode() for kmer in kmers)

    # Feed to DNABert - constant memory!
    model.process(kmer_string)

Network Streaming

import biometal

# Stream from HTTP without downloading
# Works with ENA, S3, GCS, Azure public data
url = "https://example.com/dataset.fq.gz"
stream = biometal.FastqStream.from_path(url)

for record in stream:
    # Analyze directly - no download needed!
    # Memory: constant ~5 MB
    gc = biometal.gc_content(record.sequence)

BAM Alignment Analysis (v1.3.0)

import biometal

# Stream BAM file with constant memory
reader = biometal.BamReader.from_path("alignments.bam")

for record in reader:
    # Access alignment details
    print(f"{record.name}: MAPQ={record.mapq}, pos={record.position}")

    # Analyze CIGAR operations
    for op in record.cigar:
        if op.is_insertion() and op.length >= 5:
            print(f"  Found {op.length}bp insertion")

    # Calculate alignment metrics
    ref_len = record.reference_length()
    query_len = record.query_length()
    print(f"  Reference: {ref_len}bp, Query: {query_len}bp")

# Convert BAM to SAM with filtering
writer = biometal.SamWriter.create("output.sam")
writer.write_header(reader.header)

for record in reader:
    if record.is_primary and record.mapq >= 30:
        writer.write_record(record)

writer.close()

FAQ

Q: Why biometal-rs on PyPI but biometal everywhere else? A: The biometal name was taken on PyPI, so we use biometal-rs for installation. You still import as import biometal.

Q: What platforms are supported? A: Mac ARM (optimized), Linux ARM/x86_64 (portable). Pre-built wheels for common platforms. See docs/CROSS_PLATFORM_TESTING.md.

Q: Why ARM-native? A: To democratize bioinformatics by enabling world-class performance on consumer hardware ($1,400 MacBooks vs. $50,000 servers).

More questions? See FAQ.md

Contributing

We welcome contributions! See CLAUDE.md for development guidelines.

biometal is built on evidence-based optimization - new features should:

1. Have clear use cases
2. Be validated experimentally (when adding optimizations)
3. Maintain platform portability
4. Follow OPTIMIZATION_RULES.md

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Citation

If you use biometal in your research:

@software{biometal2025,
  author = {Handley, Scott},
  title = {biometal: ARM-native bioinformatics with streaming architecture},
  year = {2025},
  url = {https://github.com/shandley/biometal}
}

For the experimental methodology:

@misc{asbb2025,
  author = {Handley, Scott},
  title = {Apple Silicon Bio Bench: Systematic Hardware Characterization},
  year = {2025},
  url = {https://github.com/shandley/apple-silicon-bio-bench}
}

Status: v1.3.0 in development 🚧
Latest: Python BAM bindings with CIGAR operations and SAM writing
Tests: 424 passing (354 library + 70 BAM parser)
Performance: 4.54M records/sec, 43.0 MiB/s throughput
Python Functions: 50+ (including full BAM support)
Evidence Base: 1,357 experiments, 40,710 measurements