Published November 10, 2026 | Version v2

chain to loop: cost-aware value-of-information scheduling for cyclic-peptide discovery

  • 1. BioMate AI
  • 2. Department of Gastrointestinal Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston
  • 3. Department of Structural Biology, St. Jude Children's Research Hospital, Memphis, TN

Description

Drug discovery screens candidate molecules through a cascade of assays rather

than a single test — from cheap property predictors through docking, molecular

dynamics and free-energy perturbation to synthesis and a wet-lab measurement

— whose costs differ by orders of magnitude. As these pipelines become

automated, a platform must decide not only which candidate to test next but

which assay to spend the next unit of budget on — a scheduling question that is

usually ignored. What little evidence exists is measured across repeated runs of a

single method rather than across the laboratories or targets to which a result

must generalise. Here we ask whether a cost-aware value-of-information (VOI)

policy can do better: at each step it chooses the candidate-and-assay pair that

most reduces advance-or-kill uncertainty per unit of cost. The preliminary

experimental results are promising. On two real cyclic-peptide discovery cascades

— PAMPA permeability for 715 molecules across 27 laboratories, and 𝐾i binding

affinity for 1,964 compounds across 40 targets — VOI scheduling beat the

standard cascade that runs each candidate from the cheapest assay to the dearest

(with the sealed endpoint assigned the expensive role by construction, a stylised

cost model), raising the cost-weighted top-10 AUC by 0.201 for permeability and

by a positive margin for binding (0.080), each beating a random baseline.

Significance is judged by a cluster bootstrap over whole laboratories and targets

rather than individual molecules — inference to the population of such units

rather than a held-out test on unseen ones. Choosing what to measure next, not

only what to make, turns a fixed validation cascade into an adaptive one — and as

discovery moves toward self-driving laboratories, deciding where to spend each

experiment may come to matter as much as the molecules proposed.

Files

MethodPaperB_VOI_nature-style (3).pdf

Files (2.9 MB)

Name Size
md5:648d4d3dc9b1dd882f2a843268eb63d7
2.9 MB Preview Download

Additional details

Dates

Updated
2026-11-10