TL;DR: A team from MIT, the University of Bonn and Fraunhofer SCAI reported a curation framework called PRISMS that selects candidates for slow laboratory tests using comparative pairwise judgments instead of predicted absolute scores.

A traffic jam at the lab door

Modern computational methods can propose candidate molecules, materials and other research designs at an unprecedented scale, while capacity to build and test those proposals physically remains scarce. The paper that introduces the new approach calls this mismatch validation congestion. The idea is straightforward: proposals are abundant, benches and instruments are not.

The scale gap is large. Virtual screening libraries routinely contain 105 to 109 compounds, according to the account in Scientific Discovery under Validation Congestion via Multi-Fidelity Pairwise Rankings. Airfoil collections can contain hundreds of thousands of feasible shapes. Making and testing, by contrast, involves synthesis or fabrication plus measurement and safety checks. A single experimental round in drug discovery typically requires one to three months to complete.

That timing reshapes the research task. Scientific design discovery is described in that arXiv report as an iterative loop of proposing designs and validating selected designs to locate high-performing options. When generation is cheap and validation is slow, progress depends less on producing one more proposal and more on choosing a small set of promising proposals for the next costly test. The authors call that choice design curation.

Curation, in this definition, sends more promising designs to experiments in each round and therefore reduces the number of design-experiment loops needed for discovery. The illustration in the paper contrasts conventional curation, which predicts absolute scores from scarce test data, with the proposed route, which curates with selectively queried rankers operating at multiple fidelity levels.

Why absolute-score models stall early

The dominant computational route for curation trains a regression model on previously validated designs, then picks new candidates from predicted property values or associated uncertainty estimates. That approach can work when enough training labels from experiments are available. The difficulty, the authors explain, is that under congestion the labels needed for a reliable regressor are themselves expensive to acquire.

A connected overview of AI-assisted discovery makes a similar point about dependence on physical follow-through. How Does an AI Drug Compound Discovery Platform Find, Rank, and Validate New Drug Candidates? describes such platforms as linked software, scientific data and laboratory workflows that identify molecules that may act on a disease-related target. Its stated purpose is not to produce an approved medicine by itself, but to help teams search larger target and molecule spaces, prioritize experiments, model likely outcomes and decide what to synthesize or test next.

That overview stresses traceability. Strong programs preserve links between each prediction and its source data, including assay conditions, confidence measures, chemical identifiers and version history. The reason given is that two compounds with nearly identical structures can produce different results in different laboratories or cell systems. Historical measurements used for training may also contain errors, missing values, inconsistent protocols and publication bias, which is why the overview says computational predictions require staged validation.

Practical prediction covers many properties at once. Models estimate binding affinity, cellular activity, selectivity, solubility, metabolic stability, permeability and toxicity. Generative models can propose variations around an initial scaffold, while active-learning systems choose next compounds from uncertainty and expected information gain. A practical ranking score might combine predicted potency, expected selectivity, synthetic accessibility, assay availability, intellectual-property constraints and estimated pharmacokinetics. The overview presents the output as a portfolio of testable hypotheses rather than a single answer.

The comparative insight

The new paper reconsiders the task as comparative. The goal, it argues, is often not to forecast every design's exact property value but to determine which designs are better. Examples given are whether molecule A inhibits a target more strongly than molecule B, whether material A is stronger than material B, or whether airfoil A has a higher lift coefficient than airfoil B.

Useful curation signals therefore do not have to take the form of absolute measurements. They can instead take the form of pairwise rankings, or comparative judgments that one design is more promising than another. The authors state that such rankings are substantially easier to gather than absolute experimental values.

Rankings can come from heterogeneous knowledge sources. The report lists empirical rules of thumb, computational models and domain experts, spanning different levels of fidelity and cost. Inexpensive heuristics may supply abundant but noisy comparisons, while experienced scientists may supply more reliable judgments at higher cost. Computational assistance can also range from simple rules to agentic workflows.

The distinction matters because human and computational knowledge that is hard to express as a calibrated number may still be easy to express as a choice. A chemist may not know a precise inhibition value yet may judge one scaffold more promising than another. A docking calculation may misestimate free energy yet still order two compounds correctly. The paper treats those comparisons as first-class evidence for curation.

Virtual screening shows the ranking tradition

Ranking compounds computationally before laboratory testing already has a long history in drug research. Virtual screening — Edgepedia defines virtual screening as automatically evaluating very large libraries of compounds with computer programs. It describes the method as the in silico analog of high-throughput screening, considering orders of magnitude more compounds than an experimental assay can handle.

Two families are distinguished in that encyclopedia entry. Structure-based screening docks each compound into a three-dimensional model of the target binding site, while ligand-based screening ranks candidates by similarity to known ligands. Docking combines a sampling algorithm that generates candidate binding poses with a scoring function that approximates free energy for each pose. Ligand approaches replace the target structure with reference ligands and compare shape and pharmacophore features.

The workflow described there has four steps: identifying the ligand-binding site on the target structure, preparing the chemical library, docking each compound and ranking by predicted binding score, and rescoring or visually inspecting binding modes of top-ranked compounds, for example the top 1%. Target structures can come from X-ray crystallography, NMR, neutron scattering, homology modeling or molecular dynamics simulations, with preparation decisions about protonation states, water molecules and receptor flexibility.

Libraries are drawn from collections such as ZINC and the Enamine REAL space, whose lead-like section grew from 1.56 billion compounds in March 2021 to 3.93 billion by August 2023, according to that entry. Because most search algorithms are stochastic, practitioners test the protocol on a representative ligand subset before a full screen. Ranked compounds are then filtered for undesirable features, including promiscuous compounds flagged by PAINS filters and reactive frequent hitters caught by the ALARM-NMR filter.

Performance context in the same entry places prospective virtual screens at 1–40% hit rates, compared with 0.01–0.14% for experimental high-throughput screening. About 12% of top-scoring compounds show activity, with a median most-potent hit near 3 micromolar across 54 campaigns. The standard benchmark cited is DUD-E, with 22,886 actives and 1,411,214 decoys. One large-scale example docked 4.5 billion lead-like molecules in about one week for roughly $25,000. More than 60 docking programs were developed over roughly three decades after DOCK, as counted by a recent review cited there.

What PRISMS does differently

The proposed framework is named PRISMS. It uses pairwise rankings from one or more experts, potentially spanning multiple levels of expertise, to identify promising candidates without relying on data-hungry regressors. That description comes directly from the abstract of the October 1 paper.

PRISMS does not require a single expert. It can combine several rankers that differ in reliability and query cost. When experts differ in fidelity and cost, the system escalates pairwise queries from lower- to higher-fidelity rankers based on a Fisher-information criterion. In plain terms, cheap and noisy comparisons handle the broad sort, and expensive and careful comparisons are reserved for pairs where additional information is most valuable.

The escalation rule is central to the design. Rather than asking a senior scientist or a heavy simulation about every pair, the framework asks inexpensive sources first and moves a comparison upward only when the expected information gain justifies the added cost. The Fisher-information criterion provides the mathematical basis for that decision, quantifying how much a new comparison would sharpen the estimate of underlying design quality.

The authors position expertise broadly. Input can come from computational tools or human input at multiple fidelity levels, ranging from empirical rules of thumb to agentic workflows and experienced scientists. That breadth lets a laboratory combine what it already has — heuristics, docking scores, machine-learning rankers, automation-assisted review and specialist judgment — into one curation process.

The framework operates in two settings studied in the paper. One is iterative screening that selects designs from fixed drug discovery libraries. The other is optimization that generates new designs without restriction to a predefined library. Both settings keep physical validation as the scarce resource and use rankings to decide what deserves that resource next.

Fewer rounds to half the top hits

In iterative screening from fixed drug discovery libraries, PRISMS reached 50% top-10 discovery recall in about 42% fewer rounds than regression-only active learning. The comparison is reported in the abstract and frames recall as finding half of the ten best designs in the library.

The same test compared PRISMS with a ranking-based method that lacked selective escalation. Against that version, PRISMS reached the same 50% top-10 recall in about 15% fewer rounds. The gap isolates the contribution of the escalation step: rankings alone help, and choosing which ranker to ask, and when, helps further.

Fewer rounds carry practical weight because each round can take months. With one to three months typically required for a single experimental round in drug discovery, a reduction of about two-fifths in round count implies a substantial saving in calendar time for a screening campaign that would otherwise need many loops. The paper does not claim that laboratory work disappears; it claims that better curation sends stronger candidates to the laboratory sooner.

The fixed-library setting mirrors how many teams work today. A virtual collection is enumerated in advance, computational methods order or shortlist entries, a small fraction is synthesized or purchased for biochemical testing, and results inform the next shortlist. In that loop, early accuracy matters because early misses consume the same scarce test slots as early hits.

Stronger sets when designs are generated

In optimization that generates new designs without restriction to a predefined library, PRISMS achieved about 18.8% higher hypervolume than the Bayesian optimization baseline. Hypervolume here is the multi-objective measure used in the study to capture the quality and spread of generated designs.

The unrestricted setting differs from library screening because proposals are not limited to listed compounds. The system can propose novel structures, often by varying scaffolds or exploring continuous representations, then test selected proposals physically or in high-fidelity evaluation. Improvement in hypervolume indicates a frontier of designs that dominates more objective space than the baseline frontier.

The AI-platform overview describes related generative behavior. Generative models can propose variations around an initial scaffold, while active-learning systems select next tests from uncertainty and expected information gain. Thresholds remain project-specific: a fragment with moderate predicted affinity may still be valuable if soluble and chemically tractable, while a high-affinity molecule can be undevelopable because it crosses membranes poorly or reacts with multiple proteins.

Together, the two reported gains address both common workflows: choosing well from what is already enumerated, and inventing well beyond the list. The authors summarize the pattern by stating that multi-fidelity expert pairwise rankings can provide an efficient alternative to regression-driven curation in scientific design discovery.

Who reported the work and when

The report appeared on October 1, 2026 as arXiv:2610.01827, in the machine-learning category. Authors listed are Kevin Tirta Wijaya, Alston Lo, Michael Sun, Wojciech Matusik and Vahid Babaei. Affiliations given are the University of Bonn and Fraunhofer SCAI for Wijaya and Babaei, and the Massachusetts Institute of Technology for Lo, Sun and Matusik.

The paper frames validation congestion as a growing condition, not a temporary backlog. Virtual collections keep expanding — billions of synthesisable compounds are now listed in make-on-demand catalogues — while synthesis, assay, fabrication and safety review remain physical and time-bound. Curation therefore becomes a durable part of discovery infrastructure.

The study does not present PRISMS as a replacement for experiments. Predictions still meet staged laboratory confirmation, with starting points that may be a defined protein or pathway in target-first work or a cell, organoid or disease phenotype in phenotypic work. Main early readouts remain binding, inhibition, mechanism, cellular state or disease phenotype, with confirmation in orthogonal assays. The contribution is a different way to decide which hypotheses reach those assays first.