forkjoin-ai/bitwise-genomes

Bitwise Genome Datasets

DNA sequences encoded as 2-bit binary with topological annotations. 4x smaller than FASTA. Searchable at 32 bases per CPU cycle.

Format

A = 00, C = 01, G = 10, T = 11
4 bases per byte. The DNA IS the binary.

Each .bw file contains:

• Magic header 0x4257 ("BW")
• Sequence length (4 bytes, big-endian)
• Header string (variable length)
• Packed 2-bit bases

Tools

bw -- DNA ripgrep

Install the Bitwise CLI to search these datasets:

cargo install --path .  # from the bitwise repo

# Search for a pattern across a gene
bw grep GGTGGCGTAGGC cancer-genes/fasta/KRAS.fasta

# Count mutations between reference and tumor
bw count reference.fasta tumor.fasta

# Compression stats
bw stats cancer-genes/fasta/BRCA1.fasta

Search speed: 90 million bases per second on a single CPU core.

Aeon FlowFrame Protocol

These datasets stream natively as Aeon FlowFrames:

stream_id = chromosome (1-25)
sequence  = genomic position
flags     = FORK | FOLD | VENT (structure type)
payload   = 2-bit packed bases

Wire = storage = memory. No serialization boundary.

helix.repair

Search these datasets live at helix.repair -- a DNA topology search engine powered by Bitwise encoding and 402 Lean theorems.

Datasets

cancer-genes/

20 clinically important cancer genes from NCBI RefSeq:

Gene	Accession	Bases	Bitwise Size	Function
TP53	NM_000546.6	2,512	628 B	Tumor suppressor ("guardian of the genome")
BRCA1	NM_007294.4	7,088	1,772 B	DNA repair (breast/ovarian cancer)
BRCA2	NM_000059.4	11,954	2,989 B	DNA repair (breast/ovarian/prostate)
KRAS	NM_004985.5	5,306	1,327 B	GTPase (pancreatic/lung/colorectal)
EGFR	NM_005228.5	9,905	2,477 B	Growth factor receptor (lung cancer)
BRAF	NM_004333.6	6,459	1,615 B	Kinase (melanoma/colorectal)
PIK3CA	NM_006218.4	9,259	2,315 B	PI3K catalytic (breast/endometrial)
PTEN	NM_000314.8	8,515	2,129 B	Phosphatase (glioblastoma/prostate)
APC	NM_000038.6	10,704	2,676 B	Wnt regulator (colorectal)
RB1	NM_000321.3	4,768	1,192 B	Retinoblastoma protein
MYC	NM_002467.6	3,721	931 B	Transcription factor (many cancers)
IDH1	NM_005896.4	2,318	580 B	Isocitrate dehydrogenase (glioma)
VHL	NM_000551.4	4,414	1,104 B	Von Hippel-Lindau (renal cancer)
ALK	NM_004304.5	6,240	1,560 B	Receptor tyrosine kinase (lung/lymphoma)
HER2	NM_004448.4	4,557	1,140 B	ERBB2 (breast cancer)
ATM	NM_000051.4	12,915	3,229 B	DNA damage response kinase
MGMT	NM_002412.5	4,678	1,170 B	DNA methyltransferase (glioblastoma)
TERT	NM_198253.3	4,039	1,010 B	Telomerase (many cancers)
JAK2	NM_004972.4	7,023	1,756 B	Janus kinase (myeloproliferative)
FLT3	NM_004119.3	3,826	957 B	FMS-like tyrosine kinase (AML)

Usage

With bw CLI

# Install
cargo install --path .

# Search for a mutation hotspot
bw grep GGTGGCGTAGGC datasets/cancer-genes/fasta/KRAS.fasta

# Pack FASTA to Bitwise binary
bw pack datasets/cancer-genes/fasta/TP53.fasta > TP53.bw

# Count mutations between sequences
bw count ref.fasta tumor.fasta

# Compression stats
bw stats datasets/cancer-genes/fasta/BRCA1.fasta

With WASM (JavaScript/TypeScript)

import { pack_bases, search_packed, mutation_count } from 'bitwise';

const packed = pack_bases(new TextEncoder().encode('ATGCTAGCATGC'));
const needle = pack_bases(new TextEncoder().encode('TAGC'));
const matches = search_packed(packed, 12, needle, 4);
// matches = [4]  -- found TAGC at position 4

Theory

Every dataset is backed by mechanized Lean 4 theorems (zero sorry):

• dna_is_folded_knot: DNA IS a folded knot (PsycheGrindExtended Pass 17)
• two_bit_four_per_byte: 4 bases per byte by construction (Pass 39)
• word_parallel_speedup: 32x search speedup (Pass 39)
• xor_detects_mutations: XOR = mutation detection (Pass 39)
• noncoding_is_void: non-coding DNA IS the void boundary (Pass 43)
• junk_not_junk: "junk" DNA carries MORE information (Pass 43)
• sigma_monotone_with_age: σ IS a molecular clock (GenomicVoidArchaeology)
• unwinding_theorem: history reconstructible from void (GenomicVoidArchaeology)

402 theorems total. The math proves the encoding. The encoding enables the search. The search reveals the biology.

Source

All sequences from NCBI RefSeq (public domain). Fetched via E-utilities API. Reproducible via scripts/fetch-and-convert.sh.

Related

• helix.repair -- DNA topology search engine
• Aunt Sandy -- Cancer genomics via Buleyean probability
• Gnosis -- Formal verification engine (402 Lean theorems)

License

Data: CC-BY-4.0 (sequences are public domain from NCBI) Code: MPL-2.0

hg38 -- Full Human Reference Genome

2.9GB FASTA → 736MB Bitwise binary. 25 chromosomes.

Chromosome	Bases	Bitwise Size
chr1	248,956,422	59 MB
chr2	242,193,529	58 MB
chr3	198,295,559	47 MB
chr4	190,214,555	45 MB
chr5	181,538,259	43 MB
chr6	170,805,979	41 MB
chr7	159,345,973	38 MB
chr8	145,138,636	35 MB
chr9	138,394,717	33 MB
chr10	133,797,422	32 MB
chr11	135,086,622	32 MB
chr12	133,275,309	32 MB
chr13	114,364,328	27 MB
chr14	107,043,718	26 MB
chr15	101,991,189	24 MB
chr16	90,338,345	22 MB
chr17	83,257,441	20 MB
chr18	80,373,285	19 MB
chr19	58,617,616	14 MB
chr20	64,444,167	15 MB
chr21	46,709,983	11 MB
chr22	50,818,468	12 MB
chrX	156,040,895	37 MB
chrY	57,227,415	14 MB
chrM	16,569	4.1 KB

Too large for GitHub. Reproduce locally:

# Download and convert (requires ~4GB disk)
bash scripts/fetch-and-convert-hg38.sh

# Or use Cloud Build
gcloud builds submit --config=cloudbuild-whole-genome.yaml --substitutions=_ASSEMBLY=hg38 .

Search speed: 90 million bases per second on a single CPU core.