Mixture-of-Prompts for Parameter-Efficient ADMET Prediction

Abstract Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is central to modern drug discovery, yet building separate, fully fine-tuned models for each end point is computationally costly and difficult to scale across many heterogeneous tasks. At the same time, recent molecular foundation models provide rich, transferable representations that invite more parameter-efficient adaptation strategies. In this work, we propose a Mixture-of-Prompts (MoP) framework for parameter-efficient single-task ADMET prediction on top of a frozen molecular encoder. For each end point, we keep the backbone fixed and introduce a small set of learnable prompt “experts” together with a lightweight gating network. Given the encoder representation of a molecule, the gate produces a dense mixture over the experts, whose pooled prompt embedding is concatenated with the frozen molecular embedding and fed into a shallow prediction head. This design allows each end point to express task-specific behavior through a compact prompt expert pool, while updating only a small fraction of the total parameters. We instantiate MoP on the standardized PharmaBench ADMET benchmark, training a separate MoP-adapted model for each of the 11 end points under both random and scaffold splits, and further evaluate it on external OpenADMET/ExpansionRx end points. Experiments show that MoP matches or improves upon strong baselines based on head-only adaptation, LoRA, and prefix tuning on many end points, with robust performance under scaffold-split evaluation and informative behavior under external distribution shift. These results suggest that modeling mixtures of prompt experts on top of frozen molecular foundation models provides a flexible and scalable route toward foundation-model-based ADMET modeling.

Authors

Institutions

Publication Details

Journal
ACS Omega
Published
2026-11-16
DOI
https://doi.org/10.1021/acsomega.6c09697
Primary Topic
Computational Drug Discovery Methods
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Mixture-of-Prompts for Parameter-Efficient ADMET Prediction

Gang Yu, Jian Huang, Bo Lin, Boping Li et al.
3 links
ACS Omega
Computational Drug Discovery Methods
article

Mixture-of-Prompts for Parameter-Efficient ADMET Prediction

Gang Yu, Jian Huang, Bo Lin, Boping Li, Jiayi Yin
article en

Abstract

Abstract Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is central to modern drug discovery, yet building separate, fully fine-tuned models for each end point is computationally costly and difficult to scale across many heterogeneous tasks. At the same time, recent molecular foundation models provide rich, transferable representations that invite more parameter-efficient adaptation strategies. In this work, we propose a Mixture-of-Prompts (MoP) framework for parameter-efficient single-task ADMET prediction on top of a frozen molecular encoder. For each end point, we keep the backbone fixed and introduce a small set of learnable prompt “experts” together with a lightweight gating network. Given the encoder representation of a molecule, the gate produces a dense mixture over the experts, whose pooled prompt embedding is concatenated with the frozen molecular embedding and fed into a shallow prediction head. This design allows each end point to express task-specific behavior through a compact prompt expert pool, while updating only a small fraction of the total parameters. We instantiate MoP on the standardized PharmaBench ADMET benchmark, training a separate MoP-adapted model for each of the 11 end points under both random and scaffold splits, and further evaluate it on external OpenADMET/ExpansionRx end points. Experiments show that MoP matches or improves upon strong baselines based on head-only adaptation, LoRA, and prefix tuning on many end points, with robust performance under scaffold-split evaluation and informative behavior under external distribution shift. These results suggest that modeling mixtures of prompt experts on top of frozen molecular foundation models provides a flexible and scalable route toward foundation-model-based ADMET modeling.

ACS Omega
Zhejiang University (CN)
National Natural Science Foundation of China, National Key Research and Development Program of China, Natural Science Foundation of Zhejiang Province
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon