#30 · ML / AI methods

Sparse and learned models move genetic association beyond single-locus additive effects

The four items share a shift from single-locus, marginal modeling toward multivariable statistical and machine learning frameworks applied to genome-wide markers or sequence. SSLPRS fits a polygenic risk score using a Spike-and-Slab LASSO prior and coordinate-ascent optimization on GWAS summary statistics, targeting sparse effect estimation. A benchmark of nine ML classifiers on gene-specific breast cancer missense variants identifies Extra Trees as the best performer, with interpretability and calibration analyses. STRAND introduces a 1B-parameter short-range transformer for exomic nucleotide decoding, evaluated on variant effect prediction and ClinVar. The end-to-end preprint instead maps SNP markers to images and text via latent encoding, validated on a strawberry diversity panel. What is new is the emphasis on sparsity, calibration, and learned representations, though the objectives differ: risk estimation, pathogenicity classification, sequence modeling, and direct phenotype-like prediction. Only the first directly addresses association testing.