Singapore researchers unveil AI framework for out-of-domain drug prediction

A Singapore team presented a test-time adaptation framework at ICLR 2026 that helps drug-discovery models make bioactivity predictions outside their training data. The approach is designed to improve accuracy on unseen proteins, scaffolds, and assay conditions without requiring access to full source datasets.

Why it matters: - Drug-discovery AI often breaks down when it sees unfamiliar proteins, molecular scaffolds, or assay settings. - TAB is designed to help models adapt in those out-of-domain cases without needing the original training data. - That could make predictive models more useful across companies and labs that cannot share proprietary datasets.

What happened: - A Singapore research team presented a new test-time adaptation framework at the International Conference on Learning Representations (ICLR) in April 2026. - The paper is titled "Test-Time Adaptation without Source Data for Out-of-Domain Bioactivity Prediction." - The co-authors are Yiming Yang, Zhiyuan Zhou, Yueming Yin, Associate Professor Hoi-Yeung Li and Associate Professor Adams Wai-Kin Kong. - The framework is called TAB, short for test-time adaptation for out-of-domain bioactivity prediction.

The details: - TAB targets two main failure modes in bioactivity prediction: overfitting to molecular geometry outside the binding pocket and reliance on shortcut patterns that do not cause binding. - The framework focuses the model on the actual protein pocket and the geometry of the molecule-protein interaction. - TAB randomly masks portions of the atomic structure to reduce noise and push the model toward binding-relevant features. - Monte Carlo dropout is used to measure prediction confidence during adaptation. - The model runs multiple predictions with slight randomness, then gives more weight to consistent outputs and less weight to unstable ones. - TAB includes a two-track self-supervised learning system. - One track teaches modified versions of the same drug-protein pair to stay close in representation, while different pairs are pushed apart. - A memory queue stores thousands of past examples to support that contrastive learning step. - The second track uses a momentum encoder, a slowly updated copy of the main model that acts as a stable reference point.

Between the lines: - The research is aimed at a core limitation of drug-discovery AI: models can look strong in lab settings but falter when the real-world target changes. - By reducing dependence on source data, TAB could lower one of the biggest practical barriers to deploying models across institutions. - The provisional patent application filed and assigned to Nanyang Biologics suggests the company sees commercial value in the method. - TAB is also part of Vecura, Nanyang Biologics' agentic AI platform for molecular discovery and life science research.

What's next: - The team says TAB showed stronger performance on DTIGN, SIU 0.6 and DrugOOD benchmarks under scaffold-, protein- and assay-based OOD settings. - On eight DTIGN subsets, TAB improved Pearson's R by 8.2% and Kendall's Tau by 5.8% on average over the best baseline. - Further validation will likely focus on whether the gains hold in broader drug-discovery workflows beyond benchmark datasets. - NYB, or Nanyang Biologics, says it develops discovery platforms including Vecura and the Vecurate library for pharmaceuticals, consumer health and cosmeceuticals.

The bottom line: - TAB is a test-time adaptation approach that aims to make drug-discovery AI more reliable when it leaves the comfort of its training data.