Adapting Evidential Neural Networks to Test-Time Neighbor Fusion Improves Molecular Property Prediction

2026-07-13

AI Summary

Existing molecular property prediction models struggle to leverage experimentally measured labels of known similar molecules during testing, limiting both predictive accuracy and calibration. This work proposes PG-EVIKAL, the first method to treat uncertainty from evidential neural networks as an actionable inference resource, enabling adaptive refinement at test time without retraining. By employing a property-aware distance metric to re-rank structurally similar neighbors and fusing their labels via Bayesian updating—augmented with scalar Kalman filtering (and its Gaussian process extension, GP-EVIKAL)—the approach achieves dynamic correction. Evaluated across 16 benchmark datasets, PG-EVIKAL significantly reduces RMSE in 14 cases (median reduction: 19.4%) while simultaneously improving calibration, and further supports continual integration and prediction refinement for new molecules in sequential experimentation settings.

Abstract

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.

Problem

• molecular property prediction
• test-time refinement
• neighbor fusion
• uncertainty quantification
• evidential neural networks

Innovation

• evidential neural networks
• test-time neighbor fusion
• uncertainty decomposition
• property-distance metric
• sequential assay