Manuscript · 8 May 2026

Rational compute allocation via mutual information and uncertainty

Early screening treated as a compute-allocation problem: scarce computation goes to the candidates whose evaluation is expected to reduce uncertainty the most.

Open manuscript (PDF) Open poster (PDF)

ManuscriptPosterMachine learning

Abstract#

Rather than sending every generated molecule through the most expensive modelling stages, the workflow uses uncertainty-aware and information-theoretic descriptors to decide which candidates deserve deeper analysis. The study frames model confidence, evidence disagreement and expected information gain as explicit ranking signals that can be combined with physicochemical, structural and biological descriptors. This creates a practical bridge between high-throughput approximation and targeted downstream simulation.

MImutual information for prioritising informative candidates
UQuncertainty as a routing feature, not only an error bar
MLdescriptors aligned with downstream simulation cost
Triagepipeline-ready data products for iterative selection

Visual summary#

Pareto construction summary
Figure 1. Pareto construction. Prioritisation framed as a trade-off among informative but imperfect screening signals.
Screening depth summary
Figure 2. Screening depth. Compute allocation adjusted to the expected value of learning from each candidate.
Uncertainty summary
Figure 3. Uncertainty. Uncertain predictions become routing signals for candidates that need better evidence.
Data realism summary
Figure 4. Data realism. Separating useful ranking information from overconfident claims about biological activity.