IMPRESS: AI-coupled HPC workflows for protein design
Protein design campaigns are often run as multi-stage pipelines with a generative backbone model, an inverse-folding model, and a structure predictor chained together in that order. The stages differ by one to two orders of magnitude in cost per candidate, and the allocation of sampling effort across them is fixed before a campaign begins. Such a filter cascade can only discard candidates. It…
Protein design campaigns typically follow a multi-stage pipeline consisting of a generative backbone model, an inverse-folding model, and a structure predictor. These stages vary in cost per candidate by one to two orders of magnitude, and the distribution of sampling effort across them is predetermined before a campaign commences.
This filter cascade can only eliminate candidates, unable to redirect sampling toward regions of the design space that a model stack can access, and for targets with constraints differing significantly from the generative model's training distribution, it may yield no solution even with substantial sample counts. The authors introduce IMPRESS (Integrated Machine-learning for PRotein Structures at Scale), a framework where design pipelines function as stateful objects that can halt, continue, resample, or branch at runtime based on their accumulated results.
Adaptive decisions are made asynchronously concerning execution, allowing the campaign to refine itself without global synchronization. IMPRESS is applied to three tasks: backbone generation on discontinuous enzyme active-site motifs, de novo small molecule binder design, and PDZ binder redesign. Adaptive parameterization effectively addresses complex active-site motifs spanning three or more discontinuous residue islands that a fixed cascade cannot solve.
It also directs the realized task mix towards inexpensive sequence exploration, generating sequence ensembles that explore broadly yet remain close to their inputs. All comparisons in the study are based on in silico metrics. The authors hypothesize that adaptive parameterization will be most beneficial for targets farthest from the training distributions of the underlying models, with the same primitives applicable to incorporate experimental measurements as steering signals.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.