#32 · ML / AI methods
Locus-resolved tools scale variant interaction, colocalisation and SV detection
Most items here are computational tools, and the common thread is locus-resolved analysis of genomic variation. GWAIS-Web performs genome-wide SNP-SNP interaction screening in case-control GWAS with GPU/FPGA acceleration, and gpu-coloc reports a roughly 1000-fold speedup for genetic colocalisation across millions of association signals. GrassSV combines depth-of-coverage and contig-based recognition to detect major structural variant classes from high-throughput DNA-seq data. Aardvark adds a basepair score for haplotype comparison, reducing representation bias for small variants. ReSeT selects reference genomes using a facility-location objective with an inter-taxon discrimination term and an ambiguity index, benchmarked on viral datasets. What is new is the extension of locus-resolved analysis to pairwise interaction and colocalisation at million-scale, alongside benchmarking of variant representation.
Published papers
Is a Win-Win possible? Achieving pareto-optimal privacy-utility balance in fine-tuned genome language model embeddings against embedding reconstruction attacks.
Benchmarks 13 transformer models (general-purpose and genomic-specialized) against position-specific embedding reconstruction attacks, finding tokenization,…
Primer design through submodular function estimation.
PRISM casts multiplex PCR primer design as constrained submodular maximization with local search, balancing viral genome coverage against primer-dimer risk and…
Preprints
Refnd: Preventing Data Leakage in Relational Datasets
Refnd is a loglinear-time HNSW proximity-graph splitting algorithm, with a Relational Generative Process formalization, that prevents data leakage in…