Learning the Next Experiment: A Decision Architecture for Closed-Loop Drug Discovery under Cost, Uncertainty, and Scientific Value Download PDF

Closed-loop drug discovery seeks to connect computational inference, experiment selection, physical execution, measurement, and model updating into an iterative learning process. Yet the central decision in such a loop—what experiment should be performed next—cannot be resolved reliably by maximizing predicted activity, model confidence, expected improvement, or any other single performance measure. Candidate experiments differ not only in their likelihood of producing desirable outcomes but also in their capacity to reduce uncertainty, distinguish competing mechanisms, broaden chemical or biological coverage, reveal model failure, consume scarce resources, and delay other decisions. This article develops a proposed Next-Experiment Decision Architecture for organizing these non-equivalent considerations. The architecture treats experiment selection as a staged scientific decision comprising admissibility assessment, multidimensional value characterization, constrained portfolio assembly, and accountable authorization, execution, and adaptive control. It separates prediction from experimental confirmation, confidence from calibrated uncertainty, structural novelty from mechanistic learning, and data availability from decision suitability. It further introduces an evidence ledger that preserves model, assay, procedural, and human-decision provenance across iterations. The synthesis indicates that closed-loop value depends on the relationship between an experiment and the unresolved decision state, rather than on an intrinsic acquisition score. Important limitations include context dependence, incomplete uncertainty models, difficulty comparing heterogeneous value dimensions, and the absence of prospective validation across discovery stages and automated platforms. The proposed architecture is therefore not a deployment-ready decision tool. It is a conceptual and methodological framework intended to support testable experiment-selection policies, transparent failure handling, and more scientifically bounded implementation of active learning in pharmaceutical discovery.