Running AlphaFold3 on Distributed High-Throughput Computing Infrastructure: Scaling Workloads and Enabling Ultra-Large Predictions
TL;DR
dHTC is established as a viable—and in some regimes superior— execution model for data-intensive structural biology workflows and provide a general blueprint for deploying large, data-intensive applications on distributed cyberinfrastructure.
Abstract
AlphaFold 3 (AF3) enables atomic-resolution prediction of biomolecular complexes, driving rapidly growing demand across the life sciences. However, its ∼ 750,GB reference database has effectively confined production deployments to systems with shared parallel filesystems, creating a major barrier for scalability. Distributed high-throughput computing (dHTC) platforms offer vast, heterogeneous compute capacity, but fundamentally lack the shared data infrastructure assumed by AF3. We present a data-aware deployment of AF3 for dHTC, implemented on the Center for High Throughput Computing (CHTC) and the Open Science Pool (OSPool). The workflow is decomposed into a CPU-bound data pipeline that executes on nodes with locally staged, scheduler-advertised databases, and a GPU-bound inference pipeline that opportunistically scales across distributed resources. Using CUDA Unified Virtual Memory (UVM), we extend inference beyond physical GPU limits, enabling predictions of ultra-large complexes that exceed device vRAM. By elevating dataset locality to a schedulable resource via HTCondor ClassAds, we eliminate prohibitive per-job data transfers and enable efficient, federated execution. Beyond scaling throughput, we demonstrate that dHTC can support previously infeasible workloads. Together, these results establish dHTC as a viable—and in some regimes superior— execution model for data-intensive structural biology workflows and provide a general blueprint for deploying large, data-intensive applications on distributed cyberinfrastructure.