HALO: AI co-pilot for scientific hypothesis generation in drug discovery

Scientific discovery is notoriously inefficient, with hypothesis generation often hindered by an enormous search space. Current AI systems can propose new candidates but lack interactive support for the reasoning process, leading to surface-level hypotheses. To solve this, researchers from UCLA (Youngseung Jeon, Kat Limqueco, JiaSyuan Chang, Xiang 'Anthony' Chen) developed HALO, a human-AI collaborative framework based on co-abduction — a novel approach that blends human intuition with AI's pattern recognition. HALO is specifically designed for molecular hypothesis generation in drug discovery, enabling users to cluster candidate molecules, identify underlying strategies, and synthesize multiple strategies into cohesive, promising hypotheses.

In expert studies involving 10 medicinal chemists, HALO significantly facilitated abductive reasoning: it improved efficient candidate observation, systematic strategy identification, and coherent multi-strategy composition. Participants using HALO produced higher-quality and more diverse candidate molecules than with conventional tools. The system will appear at the 39th Annual ACM Symposium on User Interface Software and Technology (UIST '26). By operationalizing co-abduction, HALO demonstrates a practical path for AI to augment, rather than replace, the deep reasoning scientists need for breakthrough discoveries.