A Composable and Modular Framework for Protein Structure Prediction on HPC
Current AI protein structure prediction models involve multi-stage processing that combines deep neural networks with bioinformatics tools such as multiple sequence alignment (MSA). Researchers increasingly rely on intermediate or penultimate-layer activations from these models for downstream tasks including contact prediction, binding-site identification, and model interpretability. We describe the VizFold plugin, a modular framework that can be extended toward end-to-end composable pipelines. We demonstrate feasibility through standardized hook-based tracing for ESMFold and Boltz-2, archive validation, and reproducible deployment on an HPC cluster using managed caches, modules, quotas, and Slurm workflows. The framework extracts attention maps from user-selected layers and exports intermediate representations in a backend-specific run bundle (Boltz) or a canonical archive tree (ESMFold), with shared trace text conventions and explicit provenance metadata suitable for cross-model comparison. We provide step-by-step documentation for instrumentation and deployment so that other groups can reproduce or extend the pipeline on their own clusters. The framework and instrumentation code are open-source and available at https://github.com/AI2Science/vizfold-foundation.