Cost-Aware Multi-Fidelity Bayesian Optimization with Runtime-Gated Cascading Tiers for Autonomous in silico Dermatology Discovery

Abstract

Autonomous virtual-screening pipelines suffer a recurring failure mode: cheap predictors (Boltz-2 affinity, ADMET, xTB conformer scoring) accumulate thousands of candidate scores, yet expensive validation tiers (long molecular dynamics, absolute binding free energy, wet-lab) advance unevenly because the scheduler cannot reason simultaneously about *cost*, *information value*, *gate-policy*, and *runtime availability* of each tier. We describe a cost-aware multi-fidelity Bayesian optimization (BO) scheduler that operates over a unified evidence ledger (789 compound–target rows, 136 columns) and a 7-tier cascade ranging from Boltz-2 cofold (~minutes/pair, GPU) through OpenFold3 + AQAffinity (~minutes/pair, GPU) to wet-lab IC₅₀ / IVRT / IVPT (~₩100k+/sample, CRO). The acquisition score blends expected-improvement on missing axes against a tier-specific cost prior; scientific gates (`scientific_gates.yaml`) and curator directives veto advance for compounds failing safety, novelty, or applicability-domain criteria. To prevent the scheduler from pushing work onto a tier whose runtime is missing, we add a *runtime probe* that resolves each tier's executable, model checkpoint, and dependency before scoring; tiers with missing runtime are flagged as `runtime_blocked` rather than gate-blocked, and stalled work-items (`consecutive_runs ≥ 3` in identical signature) are recorded in a per-run queue-state JSON for downstream observability. We report the scheduler's behavior on a current dermatology workload: 752 compound–target pairs queued for the OpenFold3 advance, top acquisition score 0.937, distributed across 8 protein targets (TYR, DCT, MMP1, TYRP1, LOX, CTGF, TGFB1, PTGS2). The runtime-gate distinction proved decisive: when AQAffinity calibration was incomplete, 124 work-items were correctly held under `mmp1_zinc_pending` rather than spuriously advanced. We discuss the design as a reusable substrate for in silico dermatology pipelines, the gap between cheap-tier saturation and expensive-tier throughput, and the limitations of cost-prior elicitation when tier-specific runtimes evolve. **Keywords**: multi-fidelity Bayesian optimization, cost-aware acquisition, autonomous discovery, evidence ledger, scientific gates, runtime gating, dermatology, OpenFold3, AQAffinity, Boltz-2 --- --- Author: Cheongwoo Han (HanCheongWoo) | ORCID: 0009-0004-4805-8815 Affiliation: Genesis_Medicine Lab, Seoul, Republic of Korea / HAN PREDICT, Inc. / Recover Korean Medicine Clinic License: CC-BY 4.0 Funding: Self-funded by HAN PREDICT, Inc. (no external funding) Conflict of Interest: HCW is founder of HAN PREDICT, Inc. and consults for Recover Korean Medicine Clinic. No external funding for this work. Pre-clinical academic research stage; no commercial product is currently sold based on these results. Status: In silico only. No clinical claim. Wet-lab and IRB pending (filed 2026-04-27). Cross-reference: This work is also archived at Zenodo, DOI 10.5281/zenodo.20018356 (https://zenodo.org/record/20018356). Prior history: First submission. New v0.1 draft 2026-05-03 describing cost-aware multi-fidelity BO scheduler with runtime-gated cascading tiers; submitted to bioRxiv 2026-05-03.