HaploHash

Haploblock-level functional annotation + privacy-preserving genomic hashes. Team 11 · 16–18 Sep 2026 · haploblocks.org

39,141 haploblocks (hg38) · 2,548 individuals · 5,096 haplotypes · 64-bit hash

39,141 haploblocks (hg38)   +   AlphaGenome Atlas | gnomAD
            \                          /
             \___ aggregate to block __/   top-1% statistic
                         |
               annotated haploblocks
                 /                \
           haplograph          64-bit hash
   nodes / edges / islands   S|CHROM|HAPLO|CLUSTER|VAR
      lift-weighted          re-identification attack

The task

1. Score every variant in every haploblock with AlphaGenome Atlas — interpret how genetic variation impacts biology at the molecular level.
2. Annotate the variants with additional biological information.
3. Hash variants, AVI scores and functional annotations for anonymisation.
4. Attack — reverse engineer our own anonymisation to see if it is secure.

What are haploblocks?

Recombination-cold blocks that are inherited as single units:

Phased haplotypes per block are merged into region-level FASTA and clustered with MMseqs2 — each cluster becomes one hash value:

What is AlphaGenome?

AlphaGenome reads the reference sequence plus one variant and predicts its impact on expression, splicing and regulation. AVI ranks every possible SNV genome-wide — PHRED 20 = "worse than 99% of all possible mutations".

Does AVI match real biology?

Missing heritability lives in rare variants with large effects (Manolio et al., Nature):

Consistency check — haploblocks ranked by AVI risk vs. real GTEx eQTLs:

Highest-risk blocks carry 3.2× more real eQTLs than lowest-risk blocks — 39,074 haploblocks in 10 risk groups, Spearman ρ = 0.39, p < 0.001.

Constraint vs AVI

Hypothesis: the most constrained haploblocks should be the most deleterious if mutated. We used AlphaGenome to screen perturbations across all blocks — including regions where no natural variation exists.

Result: the opposite sign. AVI rises with haplotype diversity, robust across every specification (+0.030, p = 5×10⁻⁶; +0.099, p = 10⁻⁹⁵) — likely an allele-frequency artifact: AlphaGenome is least reliable exactly where empirical variation is absent.

Biological annotations

GO enrichment on genes overlapping the blocks (dot size = count, colour = p.adj, x = GeneRatio):

Molecular function	Biological process	Cellular component

Example block — chr22:27,840,691–28,224,919 (PITPNB, TTC28 · 60 transcript isoforms):

• GO BP / MF: 6 / 11 significant terms — all PITPNB, lipid transport
• GO CC: 3 significant terms — all TTC28, mitotic spindle
• Reactome: 3 pathways — phospholipid / PI metabolism
• all hits driven by a single gene (needs to be validated)

Hashing & privacy

Each (block, cluster) pair encodes to one 64-bit word — strand, chromosome, haploblock, cluster, variants, AVI:

Field	Bits	Values
strand	1	hap1 / hap2
chrom	5	1–24
block	16	0–65,535
cluster	13	0–8,191
variants	10	0–1,023
avi	8	quantised PHRED
reserved	11	—
total	64	53 used

Re-identification exposure (a singleton cluster identifies a haplotype by one block alone):

Singleton clusters genome-wide	7.7 M
Blocks containing at least one singleton	91.5%
Blocks at which the average person is a singleton	3,027
Suppressing the 5,000 most-exposed blocks	removes 50.8% of singletons

Mitigation — two orders of magnitude apart:

Strategy	Cost
Suppress whole blocks until no singletons remain	91.5% of blocks lost
Per-cluster k-anonymity (k = 2)	3.9% of the dataset suppressed

AVI scores and annotations stay block-level; no individual allele states are exposed.

Repo map

Folder	What
annotate_atlas/	AlphaGenome Atlas scoring, chr22 pilot QC, GTEx comparison
intersect-constraint-AVI/	constraint vs AVI analysis
bio_annotations/	GO / Reactome enrichment
hash/	64-bit encoding (haplohash.py)
reid/	re-identification attack

Team

Who	Track
Mauricio Moldes	block filtering, gene counts
Alejandra Caballero, Mina	annotation databases, pruning
Markus Marandi	AlphaGenome atlas
Robert Campbell, David Bonet	effect sizes (burden / SKAT / ACAT)
Aditya Kumar Karna	encoding, hashing, hash attack

Data

1000G — 2,548 individuals, 26 populations, data.haploblocks.org. hg38, SNVs only. UKB pending.