Difficulties in interpreting machine learning (ML) models and their predictions limit the practical applicability of and confidence in ML in pharmaceutical research. There is a need for agnostic approaches aiding in the interpretation of ML models regardless of their complexity that is also applicable to deep neural network (DNN) architectures and model ensembles. To these ends, the SHapley Additive exPlanations (SHAP) methodology has recently been introduced. The SHAP approach enables the identification and prioritization of features that determine compound classification and activity prediction using any ML model. Herein, we further extend the evaluation of the SHAP methodology by investigating a variant for exact calculation of Shapley values for decision tree methods and systematically compare this variant in compound activity and potency value predictions with the model-independent SHAP method. Moreover, new applications of the SHAP analysis approach are presented including interpretation of DNN models for the generation of multi-target activity profiles and ensemble regression models for potency prediction.

In qualitative or quantitative studies of structure−activity relationships (SARs), machine learning (ML) models are trained to recognize structural patterns that differentiate between active and inactive compounds. Understanding model decisions is challenging but of critical importance to guide compound design. Moreover, the interpretation of ML results provides an additional level of model validation based on expert knowledge. A number of complex ML approaches, especially deep learning (DL) architectures, have distinctive black-box character. Herein, a locally interpretable explanatory method termed Shapley additive explanations (SHAP) is introduced for rationalizing activity predictions of any ML algorithm, regardless of its complexity. Models resulting from random forest (RF), nonlinear support vector machine (SVM), and deep neural network (DNN) learning are interpreted, and structural patterns determining the predicted probability of activity are identified and mapped onto test compounds. The results indicate that SHAP has high potential for rationalizing predictions of complex ML models.

Screening of compound libraries against panels of targets yields profiling matrices. Such matrices typically contain structurally diverse screening compounds, large numbers of inactives, and small numbers of hits per assay. As such, they represent interesting and challenging test cases for computational screening and activity predictions. In this work, modeling of large compound profiling matrices was attempted that were extracted from publicly available screening data. Different machine learning methods including deep learning were compared and different prediction strategies explored. Prediction accuracy varied for assays with different numbers of active compounds, and alternative machine learning approaches often produced comparable results. Deep learning did not further increase the prediction accuracy of standard methods such as random forests or support vector machines. Target-based random forest models were prioritized and yielded successful predictions of active compounds for many assays.

Compound activity prediction is a major application of machine learning (ML) in pharmaceutical research. Conventional single-task (ST) learning aims to predict active compounds for a given target. In addition, multitask (MT) learning attempts to simultaneously predict active compounds for multiple targets. The underlying rationale of MT learning is to guide and further improve modeling by exploring and exploiting related prediction tasks. For MT learning, deep neural networks (DNNs) are often used, establishing a link between MT and deep learning. In this work, ST and MT strategies for ML methods including DNN were compared in the systematic prediction of highly potent and weakly potent protein kinase inhibitors. A total of 19 030 inhibitors with activity against 103 human kinases were used for modeling. Given its composition, the data set provided many related prediction tasks. DNN, support vector machine, and random forest ST and MT models were derived and compared. Clear trends emerged. Regardless of the method, MT learning consistently outperformed ST modeling. Overall MT-DNNs achieved the highest prediction accuracy, but advantages over other MT-ML methods were only marginal. Furthermore, feature weights were extracted from models to evaluate correlation between different prediction tasks.