henhaobio
Sign in

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)GeneticsPeer-reviewed journal article

ATLAS: Population-Level Disease Locus Discovery via Differential Attention in Genomic Language Models

Yuqi LiuiD, Kaiwen DengiD, Yuhua YeiD, Zilin WangiD, Jiajie ZhaniD, Shicheng CheniD, Xinyue HuiD, An ChangiD, Huijun SheniD, Zhaorong LiiD, Shiping LiuiD, Xin JiniD, Kui CheniD, Xianzhi QiiD, Xiangmin XuiD, Haiqiang ZhangiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

Identifying disease-associated genetic variants remains a key challenge in genomics, especially in small cohorts or for rare and complex mutation types where genome-wide association studies (GWAS) often fall short. We introduce ATLAS, a population-level framework that leverages attention signals from pretrained genomic language models (gLMs) to detect disease-associated genes and loci directly from raw sequences--without requiring explicit variant calls or supervised training. ATLAS first performs gene-level differential attention analysis to prioritize candidate genes, followed by base-level analysis to localize disease-associated regions at single-haplotype resolution. We validate ATLAS on synthetic and {beta}-thalassemia datasets, demonstrating robust performance across diverse allele frequencies (down to 10%), cohort sizes (below 200 individuals per group), and genomic scales. Compared to GWAS, ATLAS achieves higher recall of known loci and captures haplotype-specific signals missed by traditional methods. Cross-model benchmarking further shows that precise localization depends on both model size and pretraining on diverse human genomes. In summary, ATLAS offers a scalable, sequence-native alternative to traditional statistical genetics.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (10 Feb 2026) · CC BY 4.0

Contrasting a pretrained genomic language model's attention between case and control haplotypes ranked HBB first on chromosome 11 and recovered more known β-thalassemia variants than GWAS, without per-variant tests.

GWAS need large cohorts to detect modest-effect variants, offer limited haplotype resolution and are not designed for indels, while most genomic language model (gLM) applications are supervised predictions on single individuals. ATLAS instead compares a pretrained gLM's attention patterns between case and control populations, a population-level use the paper describes as largely unexplored, requiring no explicit variant calls (it reads haplotype sequences) and no supervised training.

Key results

  1. SuggestedGene-level differential attention ranked HBB as the top gene on chromosome 11 on both haplotypes, with lower attention entropy in cases than controls; few other genes passed significance, and several of those are hemoglobin-related.

    HBB's −log10 adjusted p was 6.36× (haplotype 1) and 9.60× (haplotype 2) that of the second-ranked gene; entropy reduction in cases padj < 10^−32. Only 15 other genes reached padj < 0.05, of which HBG1, HBD, HBMS and OR52A1 are listed as thalassemia modifiers or hemoglobin regulators (Table 1). · Gene-level differential attention identifies…; Gene-wise Differential Attention Analysis; Tables

  2. SuggestedWithin HBB, attention-derived clusters overlapped more clinically reported variants than a GWAS run on the same cohort.

    Against the 31 IthaGenes variants present in the cohort, GWAS (p < 6.3 × 10^−9) found a single locus overlapping known variants, whereas ATLAS found 4 clusters covering 7 reported sites on haplotype 1 and 3 clusters covering 6 reported sites on haplotype 2. · Base-level discovery of disease-associated…; Thalassemia Cohort and Data Preprocessing; Benchmark with GWAS

  3. SuggestedAcross the 15 other candidate genes, ATLAS recovered every GWAS-significant locus, reported signals in 6 further genes, and produced narrow candidate windows with haplotype-specific patterns.

    All GWAS-significant loci (e.g., in OR52A1 and HBD) were co-identified, counted when a GWAS locus falls inside a cluster interval padded by 10 bp; clusters spanned 11.62 bp on average; asymmetric clusters between haplotypes were seen in HBD. · Base-level discovery of disease-associated…; D.2 GWAS Configuration

  4. SuggestedOn synthetic variants injected into chromosome 11 sequences, localization held up at low allele frequency, in very small cohorts and at sequence lengths up to 384 kb.

    Relative to the 4 kb, 100% allele frequency, 1,429-sample baseline, AUPRC, SNR and FRiW fell by < 10% at 10% allele frequency or 10 individuals per group, with cluster precision and recall above 0.80; from 4 to 384 kb, normalized weighted distance stayed < 0.4, and cluster precision/recall at 384 kb was ≈1.0. · Synthetic validation of base-level localization…; Synthetic Datasets; C. Synthetic Dataset Generation

  5. SuggestedLocalization quality depended on both model size and human-centric pretraining: the larger Genos model outperformed the smaller one, while the generalist Evo2-40B gave no matching advantage.

    On the β-thalassemia HBB task, Genos-v2 vs Genos-1.2B: AUPRC 0.173→0.265, SNR 3.24→4.94. Evo2-40B averaged AUPRC ≈0.093, similar to smaller baselines (Table 4). · Impact of model capacity and pretraining…; Tables

Methods, data and reproducibility

Haplotype-resolved sequences are run through a pretrained gLM (Genos), with final-layer attention averaged over heads and summed per position. Per-gene attention summaries (mainly Shannon entropy) are compared between cohorts using Mann-Whitney U tests with Benjamini-Hochberg FDR. Base-wise log2 fold changes that pass a length-adaptive threshold are then clustered with DBSCAN into candidate regions.

  • Cross-sectional β-thalassemia whole-genome sequencing cohort (409 carriers, 245 thalassemia intermedia, 775 thalassemia major), phased with Beagle4 and reconstructed into haplotype sequences on GRCh38 · human (sequences reconstructed on the GRCh38 reference) · n = 1,429 individuals
  • Protein-coding genes on chromosome 11 shorter than 512 kb, used for the gene-level scan · human (GRCh38 chromosome 11) · n = 1,031 genes (stated; length bins in Supplementary A.2 sum differently)
  • Synthetic datasets: target variants (8 SNPs up to 15 SNPs + 5 indels) inserted into 4, 20, 128 and 384 kb regions, with background from the β-thalassemia cohort; HBB excluded · human (background from the β-thalassemia cohort) · n = case:control sizes from 200:200 down to 10:10; allele frequency 100% to 10%
  • IthaGenes thalassemia variant database used as ground truth (512 HBB variants, 31 present in the cohort, 8 most prevalent as primary benchmark set) · human (thalassemia variants in HBB) · n = 31 variants in cohort; 8 primary
Methods & tools
Genos (Genos-10b-v1, Genos-1.2B, Genos-v2), Evo 2 (Evo2-40B), Luca series, FlashAttention-2 v2.8.3, DBSCAN, Beagle v4, VCFtools, PLINK2, Ensembl, IthaGenes, PyTorch 2.5.1
Software & environment
PyTorch 2.5.1+cu124, Python 3.12.2, FlashAttention-2 v2.8.3, CUDA Toolkit 12.4.
Assay conditions & compute
One node with 4× NVIDIA A40 (48GB) GPUs. Length thresholds were benchmarked on a single NVIDIA H100 (80GB): FlashAttention at 128k took 15.47 s and 42.9 GB per sequence; vanilla attention used 53.7 GB at 8,192 and ran out of memory at 32k.
Reproducibility gaps
The β-thalassemia data can only be requested from the corresponding author, and release needs the original data custodians' approval; no data accession is given.
Limitations (5)
  • Only binary phenotypes driven by major genes were studied; extension to complex, polygenic traits is left for future work. · stated by the authors
  • Analyses rely on reconstructed genomes that carry processing noise; the authors expect higher-fidelity sequencing to improve localization. · stated by the authors
  • ATLAS prioritizes candidate loci but does not establish causality; large-scale functional assays are needed to validate the predictions. · stated by the authors
  • Real-data validation covers one cohort, one Mendelian disease and chromosome 11 only, with no external replication; the text does not say how carriers (409), intermedia (245) and major (775) patients were split into case and control groups. · noted in this summary
  • Synthetic cases were homozygous mutant (1|1) or downsampled to reference (0|0), never heterozygous; controls were reference. GWAS quality control is described inconsistently (missingness < 15% in Methods vs --geno 0.01 in the appendix), which may affect the head-to-head comparison. · noted in this summary

Read this if you work on case-control locus discovery in small cohorts and want to know whether attention from a pretrained genomic language model can substitute for per-variant GWAS testing.

Citations
1
FWCI
196.45
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMcManus '23Huang '24Benegas '23Mountjoy '20Aparo '24Bowling '20
51% semantic matchGenomics
Justin N.J. McManus, Robert J. Lovelett, Daniel Lowengrub+1 more·Cell Genomics·Mar 1, 2023· 4 citations