Agent-optimized GPU polygenic scoring at cohort scale

A polygenic score (PGS) is a weighted sum of allele dosages that can only be interpreted as a percentile within a reference distribution. Currently, risk evaluation across models requires scoring each individual against a whole reference panel, but this process is prohibitively expensive when performed on CPUs. To address this issue, researchers reformulated PGS calculation as a cohort-wide sparse matrix product on a GPU and optimized it using an autonomous coding agent.

This agent accessed both synthetic and real data for training, with the real data consisting of 1000 Genomes data.

The optimized scorer demonstrated significant improvements in speed compared to existing methods. In a synthetic benchmark, the scorer achieved a 10.6-fold speedup during the training course, reducing processing time from 184.9 seconds to just 17.4 seconds. Moreover, the scorer ran 71 times faster than pgsc_calc, the existing CPU-based method.

However, when tested on held-out real genomes, the synthetic-trained scorer was 15.3 times slower than the real-data-trained version and showed a significant divergence from pgsc_calc. The maximum absolute difference between the two scored 3.7x10^-1, whereas pgsc_calc had a much smaller difference of 3.1x10^-4.

The researchers conducted three replications of the synthetic loop from an identical prompt. They found that the agent optimization trajectory variance explained 91 to 96% of the total wall-clock variance, while data sampling accounted for 0 to 2.6% and repeated execution contributed 0.2 to 3.3%. Despite these findings, the researchers ultimately scored all 3,202 1000 Genome genotypes against all 5,385 models within 6.5 hours using a memory-bounded rebuild.

They released this scoring manifest, which includes ancestry-stratified reference distributions for the entire PGS Catalog, as a frozen resource.