Transcript
0:00ADME and machine learning
Lesson 7.5 of our prediction module: machine learning for ADME. ADME is absorption, distribution, metabolism and excretion, the four processes that set the drug concentration in the body. Machine learning is a computer method that learns from examples the relation between a chemical structure and a measured ADME value. The training data are the compounds with measured values.
0:33Descriptors, fingerprints and the model
Molecular descriptors are numbers calculated from the structure, such as the molecular weight and the polar surface area. A fingerprint is a row of bits, and the bits record which small substructures the molecule contains. The model is the fitted relation between these numbers and the measured value.
0:59Test set and applicability domain
A test set is a group of compounds that the model does not learn from. The agreement of the model's predictions with their measured values gives the accuracy. Chemical space is the set of all possible structures, and a map of it places similar compounds close together. The applicability domain is the part of chemical space where the model predicts reliably. The error rises as the similarity of a compound to its nearest training compounds falls.
1:36How a model learns an ADME endpoint
This lesson has three parts. The first part is how a model learns an ADME endpoint. An endpoint is the measured property. A laboratory measures one assay, for example solubility or microsomal stability, for thousands of compounds. The model is fitted so that its predicted values match these measured values.
2:03How the accuracy is measured
The second part is how the accuracy is measured. A random split takes the test compounds at random. A time split trains on the compounds measured before a date and tests on the later ones. The core of a molecule is its ring systems and the chains that link them. A scaffold split holds back whole cores. A random split gives an optimistic accuracy, and a time split gives one closer to the accuracy on compounds not yet measured.
2:40Comparison of splitting methods
One study compared 14 models on 8 data sets with 10 types of split. Scaffold splits gave results close to the random split. Splits that group compounds by similarity and hold back whole clusters were the hardest. The result on a random split predicted the result on these clustered splits poorly.
3:06Where models do well and where they fail
The third part is where the models do well and where they fail. One public platform has models for 41 ADME and toxicity data sets. The coefficient of determination measures how closely the predictions follow the measured values, and 1 is perfect. For 5 of its 10 data sets with numeric values, this coefficient was above 0.6. Models fail in three situations. The first is new chemical space, where the test compounds are not similar to any training compound. The second is an assay with few measured compounds or a large measurement error. In one industrial comparison, the same endpoint measured in two laboratories agreed about as closely as a model agreed with one laboratory. The third is an endpoint that depends on complex biology. Clearance is the volume of plasma from which the body removes all of the drug per unit of time. It depends on enzymes, transporters, blood flow and binding. In repeated random splits, one public model predicted 52.5 percent of 650 low clearance compounds within twofold, and 10.4 percent of 144 high clearance compounds. High clearance compounds were also rare in the training data.
4:45Rat liver microsomal stability
A study of rat liver microsomal stability shows how a model performs on compounds measured later. Liver microsomes are vesicles formed from the internal membranes of liver cells, and these membranes hold the cytochrome P450 enzymes. The assay holds each compound with the microsomes for 15 minutes and measures how much remains. The half-life in the assay is the time in which half of the compound disappears. The curated set holds 20,216 compounds: 11,534 unstable, with a half-life below 30 minutes, and 8,682 stable, above 30 minutes. For six control compounds in every plate, the smallest difference that the assay resolved was about twofold.
5:48Balanced accuracy and the time split
The authors trained models on these compounds and scored them by balanced accuracy. Balanced accuracy is the average of the two class fractions that a model predicts correctly. In a time split, a model trained on the 6,448 compounds measured up to 2014 predicted the 2,576 compounds measured in 2015. For nine of the ten largest drug discovery projects, the balanced accuracy rose above 0.75 in at least one of three years. For one of these nine projects the accuracy fell in the third year, and the authors assume that new cores caused the fall. A model trained on that project alone predicted it better.
6:45Summary
In summary, a model learns an ADME endpoint from measured compounds and predicts it from the structure. A random split gives an optimistic accuracy, and a time split comes closer to the accuracy on compounds not yet measured. Models fail on compounds far from their training data, on sparse or noisy assay data, and on endpoints that depend on complex biology. In the rat microsomal data set, nine of the ten largest projects reached a balanced accuracy above 0.75 in at least one of three years. For one of these, the authors assume that new cores lowered it. This concludes lesson 7.5 on machine learning for ADME.
Sources
- Lombardo F, Desai PV, Arimoto R, et al. In silico absorption, distribution, metabolism, excretion, and pharmacokinetics (ADME-PK): utility and best practices. An industry perspective from the International Consortium for Innovation through Quality in Pharmaceutical Development. J Med Chem. 2017. PMID 28609624
- Siramshetty VB, Shah P, Kerns E, et al. Retrospective assessment of rat liver microsomal stability at NCATS: data and QSAR models. Sci Rep. 2020. PMID 33244000
- Sheridan RP, Feuston BP, Maiorov VN, et al. Similarity to molecules in the training set is a good discriminator for prediction accuracy in QSAR. J Chem Inf Comput Sci. 2004. PMID 15554660
- Seal S, Trapotsi MA, Mahale M, et al. PKSmart: an open-source computational model to predict intravenous pharmacokinetics of small molecules. J Cheminform. 2025. PMID 41013647
- Sheridan RP. Time-split cross-validation as a method for estimating the goodness of prospective prediction. J Chem Inf Model. 2013. PMID 23521722
- Bemis GW, Murcko MA. The properties of known drugs. 1. Molecular frameworks. J Med Chem. 1996. PMID 8709122
- Fooladi H, Vu TNL, Mathea M, et al. Evaluating machine learning models for molecular property prediction: performance and robustness on out-of-distribution data. J Chem Inf Model. 2025. PMID 40947919
- Swanson K, Walther P, Leitz J, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics. 2024. PMID 38913862
- Broccatelli F, Trager R, Reutlinger M, et al. Benchmarking accuracy and generalizability of four graph neural networks using large in vitro ADME datasets from different chemical spaces. Mol Inform. 2022. PMID 35156325
- Sheridan RP, Karnachi P, Tudor M, et al. Experimental error, kurtosis, activity cliffs, and methodology: what limits the predictivity of quantitative structure-activity relationship models? J Chem Inf Model. 2020. PMID 32207612
Illustrations adapted from Servier Medical Art (smart.servier.com), CC BY 4.0.
