ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction

Dongmin Bang, Juyeon Kim, Haerin Song, Sun KimBioinformatics 2025 · DOI

TL;DR

This paper introduces ADME-DL, a novel framework that significantly improves drug-likeness prediction by incorporating pharmacokinetic (PK) principles into molecular foundation models (MFMs). By training MFMs sequentially on Absorption, Distribution, Metabolism, and Excretion (ADME) tasks, ADME-DL creates richer molecular embeddings that better capture a compound's biological behavior, leading to more accurate predictions of its potential as a drug.

What Problem Does This Paper Address?

The early stages of drug discovery involve sifting through vast numbers of potential drug candidates to identify those most likely to succeed. Traditional methods often rely solely on structural properties, neglecting crucial pharmacokinetic (PK) factors like absorption, distribution, metabolism, and excretion (ADME), which are vital for a drug's efficacy and safety in the body. While recent AI models have improved molecular representation, they often treat ADME properties independently or fail to capture their inherent sequential and interdependent nature. This oversight limits their ability to accurately predict drug-likeness, leading to wasted resources on compounds that might fail later in development due to poor PK profiles. ADME-DL aims to bridge this gap by explicitly modeling the PK flow and interdependencies within ADME processes.

Methods

ADME-DL employs a two-step, model-agnostic framework:

  1. Sequential ADME Multi-Task Learning (MTL): This is the core innovation. Instead of treating all ADME tasks as independent, ADME-DL trains MFMs (like GraphMVP, MolCLR, and Transformers) on 21 ADME endpoints in a specific sequence: Absorption (A) → Distribution (D) → Metabolism (M) → Excretion (E). This sequential approach mimics the natural PK lifecycle of a drug and allows upstream tasks to inform downstream ones, creating an "ADME-aware" embedding space. To handle conflicts between tasks within a category (e.g., multiple absorption tasks), they use the PCGrad algorithm.
  2. Drug-Likeness Prediction: The ADME-aware embeddings generated in Step 1 are then fed into a simple Multi-Layer Perceptron (MLP) classifier. This MLP is trained to distinguish between approved drugs and non-drug compounds, leveraging the enriched representations for improved accuracy.

The framework is "model-agnostic" because it can be applied to various MFMs, allowing researchers to integrate their preferred molecular representation models.

Datasets

The study utilizes several publicly available datasets:

  • ADME Datasets: 21 ADME-associated endpoints from the Therapeutics Data Commons (TDC) database. These cover Absorption (8 tasks), Distribution (3 tasks), Metabolism (8 tasks), and Excretion (2 tasks).
  • Drug-Likeness Datasets:
    • Positive Set: Approved drugs from the DrugMAP database.
  • Negative Sets: Large chemical libraries of non-approved compounds from ZINC20, PubChem, and ChEMBL.
  • Benchmark Sets: Three combined datasets were created: DrugMAP-ZINC, DrugMAP-PubChem, and DrugMAP-ChEMBL, each containing approved drugs and 10,000 non-drugs. These were further split into five subsets for 5-fold cross-validation.

Key Insights and Ideas

  • PK Flow Matters: The sequential A→D→M→E order is not arbitrary; it reflects the biological reality of how drugs behave in the body and significantly improves model performance compared to random or naïve MTL.
  • Task Interdependencies are Crucial: Naïve MTL can lead to "task interference," where learning one task degrades performance on another. ADME-DL's sequential approach mitigates this by respecting causal relationships.
  • Model-Agnostic Enhancement: The benefits of sequential ADME training are not limited to a specific MFM architecture; it consistently improves performance across various models, highlighting the generalizability of the approach.
  • Richer Embeddings: The ADME-aware embeddings capture more than just structural features; they encode biological context, leading to more nuanced and accurate drug-likeness predictions.
  • Predictive Power Across Drug Discovery Stages: The model's drug-likeness scores naturally decrease as compounds move from approved drugs to investigational and AI-generated molecules, reflecting the progression of drug development and refinement.

Key Results

  • ADME-DL achieved up to a +2.4% improvement in drug-likeness prediction performance over state-of-the-art baselines.
  • Performance enhancement across tested MFMs ranged up to +18.2%.
  • In case studies with clinically annotated drugs, ADME-DL's predictions aligned more closely with expert annotations and assigned higher drug-likeness scores to known drugs compared to a single-task learning (STL) baseline.
  • The t-SNE visualization of the embedding space showed improved clustering and differentiation when using ADME-DL, particularly for bioavailability groups.

Limitations and Future Work

  • Limitations: The paper acknowledges that while ADME-DL shows promise, the current framework focuses on ADME properties. The authors suggest that future work could expand to include other critical endpoints like toxicity (ADMET), which present even more complex interdependencies.
  • Future Work:
    • Integrate toxicity endpoints and other ADMET attributes.
  • Explore more complex dependency modeling for these additional attributes.
  • Further refine the framework for more efficient and accurate drug development pipelines.

Summary updated

BibTeX
@article{bang2025adme,
  author = {Dongmin Bang and Juyeon Kim and Haerin Song and Sun Kim},
  title = {{ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction}},
  journal = {Bioinformatics},
  year = {2025},
  doi = {10.1093/bioinformatics/btaf259},
}
Figure