Deep learning is reshaping structural biology, but adoption across Australian labs is impacted by scarce compute, fragmented tooling, and reproducibility gaps. Australian BioCommons is working with the Australian Structural Biology Computing community to deliver coordinated infrastructure and capability for deep learning-enabled structural biology.
The program comprises three pillars:
(1) Shared access to models and compute. A curated, continually updated model zoo (sequence-to-structure predictors, multimer/complex modellers, generative design) deployed on national Tier 1 and Tier 2 HPC and cloud with fair-use allocation tailored to academic workloads;
(2) Reproducible, portable workflows. Containerised pipelines with full provenance (workflow languages, versioned reference datasets, automated validation reports) enabling benchmarkable, end-to-end runs from sequence to candidate designs.
(3) Capability and community. An active community initiative with targeted training, a national helpdesk, and domain-specific “recipes” co-designed with structural biologists for complex structural modelling and design.
This work is guided by the Australian Structural Biology Computing community (join at https://australian-structural-biology-computing.github.io/) and the community’s published roadmap (Zenodo: https://zenodo.org/records/15786982).
Early pilots demonstrate complete, auditable runs on national systems, producing outputs suitable for publication and translation while preserving data ownership and IP clarity. We will present the current architecture, governance, exemplar results, and onboarding pathways for Australian researchers. By reducing adoption friction and standardising best practice, this initiative enables Australian researchers to apply state-of-the-art deep learning to protein structure and function problems at scale.