Workshops Archive

Addressing Challenges in Collaborative Computational Cancer Policy Research


Workshop: The 12th Computational Approaches for Cancer Workshop (CAFCW26)

Authors: Jonathan Ozik and Nicholson Collier (Argonne National Laboratory, University of Chicago); Sayaka Ishizawa (MD Anderson Cancer Center); Maryam Eghbalizarch (MD Anderson Cancer Center, Kennesaw State University); and Iakovos Toumazis (MD Anderson Cancer Center)

Abstract: Two computational artifacts underpin personalized lung cancer screening: a natural history model (NHM) that reproduces disease progression (CHRONOS-Lung), and a decision-analytic framework that converts individual risk into a screening schedule. ENGAGE provides the latter as a partially observable Markov decision process (POMDP), solved once per risk profile across a O(10^8) combinatorial space. Both are written in Julia and were developed at MD Anderson. Since 2020, successive ALCC allocations have let our team run large-scale calibration and analysis campaigns for multiple cancer modeling groups on leadership computing, which made the ENGAGE campaigns conceivable. We review four themes from moving these models onto HPC.

The first was dependency and portability. The POMDP solves relied on a commercial solver, so we moved to the open, JuMP-compatible HiGHS. Scale then exposed fragility local runs never showed, with 230 of 1,000 solves failing at default tolerances. Sweeping three tolerances across 64 combinations identified relaxed settings we adopted. Separately, the Julia projects pinned no manifests, so library drift triggered recompilation on compute nodes, a limited issue locally but a large one when many concurrent compilations fire at once. Pinning environments and amortizing JIT and library-load costs across node-local task batches removed that overhead.

The second was scale. The POMDP campaign produces an enormous number of small files in Julia-specific binary formats, a pattern parallel filesystems handle poorly and one aggravated by sharing an allocation and file systems with other cancer modeling groups. We ran in batches of two million POMDP solves, chaining an archival tar job that triggered the subsequent batch, staged through the scratch file system, and served the resulting archives through a Globus collection. Input handling had similar issues. The solves re-read the same jld2 files repeatedly, so at our scale 55,680 of them contended for the same files, causing I/O deadlock and sharply longer runtimes until we rewrote them to cache the reads.

The third was pluggability. CHRONOS-Lung, the sex-, histology-, and nodule-attenuation-specific NHM, comprises 12 Julia models sharing a 15-state monthly-cycle structure with Gompertz tumor growth and size-dependent stage transitions. Its original calibration was a bespoke extreme-elitist genetic algorithm in Julia, in which the two fittest of 100 individuals produced every child. Because the EMEWS model exploration framework treats both the NHM and the exploration algorithm as pluggable tasks, we could investigate alternatives efficiently, using EMEWS queues for Python to let the open-source Python DEAP library drive the Julia NHM over MPI. We finally adopted a (mu + lambda) scheme with tournament selection, which produced robust fits across the multiple calibration targets.

The fourth was statistical. Calibration had used 1,000-person cohorts for expediency. Scaling studies let us move to 100,000, spending concurrency on reduced uncertainty and better estimator quality rather than wall-clock alone. The calibrated models reproduced their SEER targets closely and validated against four independent external cohorts.

Throughout, the governing question was how much domain code must change to meet HPC constraints and how much the workflow layer can absorb on its behalf.


Back to The 12th Computational Approaches for Cancer Workshop (CAFCW26) Archive Listing

Back to Full Workshop Archive Listing