Predicting ADMET Properties with Deep Learning: What Works and What Does Not

A molecule can bind its target with sub-nanomolar affinity and still fail as a drug. It might be metabolized too quickly in the liver (poor metabolic stability), fail to cross cell membranes (poor permeability), accumulate in off-target tissues (poor distribution), or cause cardiac toxicity by blocking the hERG potassium channel. These properties, collectively known as ADMET (absorption, distribution, metabolism, excretion, and toxicity), determine whether a potent molecule becomes a viable drug.

At Gimerny AI, ADMET prediction is integrated into every stage of our molecular design pipeline. GimernyMolecule generates molecules conditioned on ADMET constraints, ensuring that candidates are drug-like from the start rather than requiring extensive optimization afterward. This post describes our ADMET prediction suite, the architectures we use, and an honest assessment of where deep learning excels and where it still falls short.

The 21 Endpoints

Our ADMET prediction suite covers 21 endpoints across five categories. Absorption endpoints include Caco-2 permeability, PAMPA permeability, aqueous solubility (LogS), human intestinal absorption (HIA), and P-glycoprotein substrate/inhibitor status. Distribution endpoints cover plasma protein binding (PPB), volume of distribution (VDss), blood-brain barrier permeation (LogBB), and fraction unbound in plasma. Metabolism endpoints include metabolic stability (human liver microsomes and hepatocytes), CYP450 inhibition (five isoforms: 1A2, 2C9, 2C19, 2D6, 3A4), and CYP substrate identification. Excretion endpoints encompass renal clearance and total clearance prediction. Toxicity endpoints cover hERG channel inhibition, acute oral toxicity (LD50), Ames mutagenicity, hepatotoxicity (DILI), and cardiotoxicity (QT prolongation).

Architecture Comparison

We systematically compared four architectures across all 21 endpoints: Morgan fingerprints with gradient-boosted trees (XGBoost), message-passing neural networks (MPNNs), graph transformers with attention-based pooling, and 3D equivariant GNNs operating on conformer ensembles.

The results defy simple narratives. No single architecture dominates across all endpoints. For endpoints with large training datasets (>10K compounds) and clear structure-activity relationships, such as aqueous solubility and CYP inhibition, all four methods perform comparably, with the simpler fingerprint-XGBoost baseline within 2-3% AUC of the best neural network. For endpoints with smaller datasets (<5K compounds) or complex, non-local property dependencies, such as hERG inhibition and hepatotoxicity, the 3D equivariant GNN provides a meaningful advantage (5-12% AUC improvement), likely because these properties depend on molecular shape and electrostatics that 2D representations miss.

Our production system uses an ensemble approach: for each endpoint, we train all four architectures and combine predictions using a stacking meta-learner. This ensemble outperforms any single architecture on 19 of 21 endpoints.

What Works Well

Several ADMET properties can now be predicted with sufficient accuracy to drive design decisions. Aqueous solubility (LogS) achieves RMSE of 0.58 log units, accurate enough to rank compounds and flag insoluble candidates. CYP inhibition achieves AUC >0.92 for all five isoforms, reliable enough to filter out metabolic liabilities early. Caco-2 permeability achieves RMSE of 0.38 log units, adequate for distinguishing permeable from impermeable compounds. Lipophilicity (LogP/LogD) achieves RMSE of 0.42 log units, among the best-predicted molecular properties.

These endpoints share common traits: large, high-quality training datasets; clear physicochemical drivers (solubility correlates with molecular weight, polar surface area, and hydrogen bond donors); and relatively simple structure-property relationships.

What Does Not Work (Yet)

Other endpoints remain stubbornly difficult. Hepatotoxicity (DILI) achieves only 0.72 AUC. DILI is a multi-mechanism phenomenon involving reactive metabolite formation, mitochondrial dysfunction, bile salt transport inhibition, and immune-mediated responses. No single molecular representation captures all these mechanisms. Volume of distribution (VDss) achieves RMSE of 0.65 log units, insufficient for quantitative prediction. VDss depends on tissue binding, which is influenced by the full physiological context (blood flow, tissue composition) rather than molecular properties alone. Clearance prediction remains challenging because clearance is determined by enzymatic processes that depend on the concentration and activity of specific enzymes, which vary across individuals and are poorly captured by molecular features alone.

The Applicability Domain Problem

Perhaps the most important lesson is that average performance metrics tell an incomplete story. A model with 0.90 AUC on a held-out test set may have 0.98 AUC on compounds similar to the training set and 0.70 AUC on structurally novel compounds, which are precisely the ones you want to predict in a drug design setting.

We address this with explicit applicability domain estimation. For each prediction, we compute a confidence score based on the compound's distance to the training set in feature space (using the model's learned representations), prediction uncertainty from the ensemble (disagreement among the four base models), and presence of structural alerts (known problematic substructures not well-represented in training data).

Predictions with low confidence scores are flagged for experimental verification rather than trusted blindly. In practice, approximately 15-20% of predictions for novel chemotypes fall below our confidence threshold.

Integration into Generative Design

ADMET prediction is most powerful when integrated into the generative design loop rather than used as a post-hoc filter. In GimernyMolecule, ADMET predictors serve as reward signals in the multi-objective reinforcement learning framework. The diffusion model is guided to generate molecules that simultaneously optimize target binding, ADMET properties, and synthetic accessibility.

This approach produces molecules that are drug-like from the outset, rather than requiring multiple rounds of design-predict-redesign cycles. In our internal benchmarks, molecules generated with ADMET-integrated design require 40% fewer optimization rounds to reach candidate quality compared to target-only generation followed by ADMET filtering.

The honest truth about ADMET prediction is that it is a solved problem for some properties and an open research question for others. Knowing which is which, and communicating uncertainty clearly, is as important as the predictions themselves.