Workshop: The 12th Computational Approaches for Cancer Workshop (CAFCW26)
Authors: Oleksandr Narykov, Yitan Zhu, and Thomas Brettin (Argonne National Laboratory); Yvonne Evrard (Leidos Biomedical Research, Frederick National Laboratory); Alexander Partin (Argonne National Laboratory); James Doroshow (National Cancer Institute); and Arvind Ramanathan and Rick Stevens (Argonne National Laboratory)
Abstract: Deciding which cell-line-to-clinic drug-response transfer gains are real, rather than artifacts of a favorable seed or an uneven cohort, is not one large computation but thousands of small independent ones: each claim must be scored one drug at a time, under leak-free cross-validation, against a matched negative control, and repeated over seeds. We present the HPC harness that makes this audit routine on a single multi-GPU node. The pipeline is IMPROVE-conformant - one data model and one central metrics module across cell lines, xenografts, organoids, and patients - and its unit of work is a (dataset, method, drug-shard, seed) tuple. Drugs are partitioned disjointly by p mod num_shards so workers never race; a flexible pinned-slot scheduler auto-detects the device count and keeps exactly one job per GPU, refilling each device the instant its job exits; and several small experiments are packed onto disjoint GPU subsets to saturate the node. The statistical nulls are the primary parallel workload-per-drug label permutations and a 5,000-resample paired bootstrap fan-out with the same sharding, and a memory-efficient attention kernel fits a genomic foundation-model probe into 16GB. The same scripts scale down to one GPU, so every result is bit-reproducible. Run at this scale with a null on every claim, the science is sobering: a feature-free base-rate predictor reproduces a published transfer model's discrimination (r=0.94 across four cohorts), the training loss beats the encoder (0.81 vs 0.42 rank correlation).
Back to The 12th Computational Approaches for Cancer Workshop (CAFCW26) Archive Listing
Back to Full Workshop Archive Listing