Neighbor molecules improve prediction accuracy express
Predicting the properties of a new molecule is like guessing the flavor of an unfamiliar dish. The program works like an experienced chef: it “tastes” similar molecules whose properties are already known, and based on their “flavor,” it refines its guess. For example, for water or carbon dioxide, the algorithm finds studied analogs and adjusts boiling point or solubility.
The chef doesn't just copy neighbors: it evaluates how much to trust its own forecast. For this, a measure of uncertainty is used, similar to entropy in physics: the higher the uncertainty, the more the program relies on verified data from similar substances. Surprisingly, the method can work with contradictory information: it decides which “neighbor” to trust more by weighing the reliability of sources.
In the end, the error drops by nearly 20%—roughly as if a weather forecast became a fifth more accurate. And all this without lengthy retraining: the algorithm incorporates new data on the fly, speeding up the search for drugs and materials.
Original abstract
as written by the authors arXiv:2607.11091
A trained molecular property model can be refined at test time by correcting each prediction with the measured labels of the most similar training molecules, a retraining-free procedure we call neighbor fusion; evidential neural networks make it principled by using their aleatoric and epistemic uncertainty to parameterize a Bayesian update. Our main contribution, PG-EVIKAL, learns a property-distance metric to re-rank structurally similar neighbors by their property relevance before fusion, building on EVIKAL (scalar Kalman filter) and GP-EVIKAL (Gaussian process variant handling correlated neighbors). Evaluated on 16 molecular datasets, PG-EVIKAL reduces RMSE relative to the evidential model baseline on 14 of them, with a median reduction of 19.4%, and improves calibration; in sequential-assay scenarios it further incorporates newly measured molecules, refining predictions as they arrive without retraining. This work demonstrates that evidential uncertainty decomposition is not merely a calibration objective but an actionable inference resource that enables test-time refinement of molecular property predictions.
Lab measurement of a single molecular property can take hours, while computer prediction takes fractions of a second. The new method makes such calculations nearly 20% more accurate without requiring additional experiments.
The fantastical universal substance analyzer becomes more real: the algorithm refines molecular properties without a lab, relying on the experience of 'neighbors'.