Skip to main content
Back to timeline
bioRxivSource publication:

Protein-design pipelines learn to change course mid-run: IMPRESS reallocates sampling at runtime and solves five-island active sites a fixed cascade misses

Synopsis

Protein design usually chains diffusion backbone generation, inverse folding and structure prediction into a fixed pipeline whose sampling budget is set before the campaign starts, so filters can only discard candidates and cannot steer sampling toward regions the model stack can reach. IMPRESS makes each design pipeline a stateful object that halts, continues, resamples or branches at runtime from its own accumulated backbones, sequences and folds, with decisions evaluated asynchronously and without global synchronization. On the 41-motif AME enzyme-ligand benchmark, the adaptive workflow returns high-quality solutions for motifs with three or more discontinuous residue islands that the fixed cascade does not solve at all, including better than 40% of five-island motifs.

Source-provided article image: IMPRESS: AI-coupled HPC workflows for protein design
Figure 1 · bioRxiv · Page 6

Interpretation

The sampling budget of a protein-design pipeline is no longer fixed before the campaign starts; the pipeline's own accumulated intermediate results decide what happens next at runtime. The prior arrangement was a filter cascade with constant fan-out: generate a set number of backbones, design a set number of sequences per backbone, fold each sequence, and keep whatever passes confidence and geometry thresholds. On the AME benchmark of 41 enzyme-ligand motifs taken from experimentally determined bound structures, the adaptive workflow returns high-quality solutions for motifs with three or more discontinuous residue islands, which the fixed cascade does not solve at all.

Each design pipeline is a stateful object holding its own history, and at explicit decision points it may halt, continue, resample a previous stage, or spawn child pipelines that inherit the parent's state and then adapt independently. Decisions are evaluated asynchronously with respect to execution, so a campaign refines itself without a global synchronization barrier and tasks with very different resource profiles stay in flight together. The framework executes through RADICAL-AsyncFlow and the RHAPSODY AI-HPC runtime, which place heterogeneous inference and simulation tasks on heterogeneous resources, and was developed and tested on NAIRR allocations on the Delta and Anvil supercomputers.

In small-molecule binder design, the adaptive rules shift the realized task mix toward inexpensive sequence exploration and away from expensive folding, and the ensemble converges over the campaign to a higher mean pLDDT with fewer low-confidence structures. The nonadaptive reference routes tasks from a fixed table, and thresholds are numerically identical between the adaptive and nonadaptive arms; the arms differ only in whether they act on the accumulated measurements. In PDZ binder redesign, an input-matched analysis covering 53 of 70 PDZ input candidates found the adaptive ensemble's sequence-space entropy not statistically distinguishable from the reference, though the sequences show signs of targeted exploration.

Perspective

The demonstrated scope is three in silico use cases: backbone generation on discontinuous enzyme active-site motifs, de novo small-molecule binder design, and PDZ binder redesign, with all comparisons between in silico metrics and production runs on HPC. The authors expect adaptive parameterization to matter most for targets furthest from the underlying models' training distributions, and expect the same primitives to accommodate experimental measurement as a steering signal.

The metrics used to steer the workflows are also typically the metrics used to evaluate them, and this circularity bounds what the results can establish; the adaptive campaigns combine multiple metrics and mechanisms, so the results do not decisively prove that a particular metric corresponds to improved objective protein quality. The reported small-molecule campaigns terminated on the scheduler wall clock before the task budget was reached, and in the PDZ case the adaptive and reference ensembles' sequence-space entropy was not statistically distinguishable.