ADMET-EvO: a self-evolving scientific agent for sustained research across heterogeneous tasks
Abstract
Scientific agents can move beyond automated model building by using accumulated evidence to revise both their questions and experimental strategies. The challenge is sustaining this adaptation across heterogeneous tasks without overfitting decisions to internal validation. Absorption, distribution, metabolism, excretion and toxicity (ADMET) prediction provides a demanding setting across diverse assays, datasets and chemical domains. We therefore developed ADMET-EvO, an evidence-gated agent that formalizes endpoints, generates falsifiable hypotheses and tests interventions across data, feature and model axes. It carries supported, rejected and inconclusive outcomes forward to guide each new cycle. Across the 22-task Therapeutics Data Commons (TDC) ADMET benchmark, ADMET-EvO achieved the highest task-normalized score of 96.77. Evidence-guided selection reduced cumulative fitting time by 72.2% within a predefined non-inferiority margin. It also formalized 43 toxicity-related tasks and constructed endpoint-specific predictors. Together, these results show how ADMET-EvO can accumulate evidence, revise its strategy and expand its research scope over time.
Related
- DEPICT enables sample-centric therapeutic prioritization across evolving and spatially heterogeneous tumor states
- Agents Propose, Code Verifies, Humans Decide: Verified Agentic Ingestion of Heterogeneous Scientific Tables into Relational Databases
- Proceedings of the 2025 Society for Research into Hydrocephalus and Spina Bifida Annual Scientific Meeting: Bridging Innovation and Collaboration in Hydrocephalus and Spina Bifida
- A research roadmap for grazinglands: leveraging scientific evidence to drive beef sustainability outcomes.
- Long-term biodiversity responses to anthropogenic change are spatially and taxonomically heterogeneous across coastal wetlands
BibTeX
@misc{zhou2026admetevo,
title = {{ADMET-EvO: a self-evolving scientific agent for sustained research across heterogeneous tasks}},
author = {Yiling Zhou and Yilin Wang and Jianmin Wang and Heqin Zhu and Zirui Wang and Chang-yu Hiesh and Kejun Ying and Jiaqi Wang and Yuzhi Xu and Tingjun Hou and Odin Zhang},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.10121},
eprint = {2609.10121},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.10121},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - ADMET-EvO: a self-evolving scientific agent for sustained research across heterogeneous tasks
AU - Yiling Zhou
AU - Yilin Wang
AU - Jianmin Wang
AU - Heqin Zhu
AU - Zirui Wang
AU - Chang-yu Hiesh
AU - Kejun Ying
AU - Jiaqi Wang
AU - Yuzhi Xu
AU - Tingjun Hou
AU - Odin Zhang
PY - 2026
DA - 2026/09/10
PB - arXiv
DO - 10.48550/arxiv.2609.10121
UR - https://arxiv.org/abs/2609.10121
AB - Scientific agents can move beyond automated model building by using accumulated evidence to revise both their questions and experimental strategies. The challenge is sustaining this adaptation across heterogeneous tasks without overfitting decisions to internal validation. Absorption, distribution, metabolism, excretion and toxicity (ADMET) prediction provides a demanding setting across diverse assays, datasets and chemical domains. We therefore developed ADMET-EvO, an evidence-gated agent that formalizes endpoints, generates falsifiable hypotheses and tests interventions across data, feature and model axes. It carries supported, rejected and inconclusive outcomes forward to guide each new cycle. Across the 22-task Therapeutics Data Commons (TDC) ADMET benchmark, ADMET-EvO achieved the highest task-normalized score of 96.77. Evidence-guided selection reduced cumulative fitting time by 72.2% within a predefined non-inferiority margin. It also formalized 43 toxicity-related tasks and constructed endpoint-specific predictors. Together, these results show how ADMET-EvO can accumulate evidence, revise its strategy and expand its research scope over time.
ER -