HALO: Interactive Co-abductive Reasoning in Scientific Hypothesis Generation
Youngseung Jeon, Kat Limqueco, JiaSyuan Chang, Xiang 'Anthony' Chen
cs.HC
2026-07-21
HALO's human-AI workflow (cluster, identify strategies, synthesize) lifted improved properties per drug molecule from 2.25 to 3.16 and raised diversity, in a 10-chemist study.
Lead optimization in drug discovery starts from a molecule with some activity and tries to rework its structure so it passes on absorption, toxicity, blood-brain barrier permeability and a stack of other properties at once. Each site a chemist touches opens 10^4 to 10^6 analogues, and the trade-offs have to be negotiated by hand. Generative AI can dump candidate molecules by the thousand, but it does nothing for the step that actually matters: turning those candidates into a hypothesis worth pursuing. Chemists are left doing the reasoning manually, and the hypotheses stay surface-level.
HALO's premise is that generation is not the bottleneck; organizing generated candidates into something you can reason about is.
The authors model hypothesis generation as co-abduction, extending abductive reasoning (inferring the best explanation for an observation) into a human-AI loop with three stages:
The components are off-the-shelf: GeLLM4O-C for molecule generation, ADMET-AI for property evaluation, and LLMs for strategy explanation and synthesis. The baseline uses the same generative model with the three co-abductive modules turned off, so the comparison isolates the value of the collaboration framework; model capability is held constant.
Ten medicinal chemists with 8 to 15 years of experience (mean 10.9) took part in a within-subjects design: HALO and baseline, 30 minutes each, seven candidate molecules delivered.
| Metric | Baseline | HALO |
| Avg. improved properties | 2.25 ± 0.48 | 3.16 ± 0.37 |
| Avg. pairwise similarity (lower = more diverse) | 58.27% ± 11.01% | 46.07% ± 7.68% |
Both differences reach p = 0.0039. On a 7-point Likert scale HALO scored significantly higher across efficient observation, systematic strategy identification, and coherent synthesis. Each session produced on average 1.8 (SD 0.92) abductive leaps, insights logged via an Insight button; after a leap, AI calls dropped sharply (from 9.11 to 1.44 on average) and manual editing rose, suggesting the framework pushes chemists toward reasoning on their own.
The contribution is not a stronger generative model; it is the observation that the real bottleneck in AI drug discovery is interaction. The cluster-explain-synthesize pattern is portable in principle to other AI-for-science settings like materials design or retrosynthetic planning. For anyone building AI drug-discovery tools, this is a reusable product-design skeleton.
Quality is scored entirely by ADMET-AI, with no wet-lab validation; the authors note wet-lab costs around $2,500 per compound, impractical here. The 30-minute tasks and N = 10 sample are small, and real optimization usually takes over a week, so generalizability is uncertain. Abductive leaps were instrumented only in the experimental group, with no baseline comparison.