Interpretation of machine learning models using shapley values: application to compound potency and multi-target activity predictions

Abstract

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.

Keywords: Black box character; Compound activity; Compound potency prediction; Feature importance; Machine learning; Model interpretation; Multi-target modeling; Shapley values; Structure–activity relationships.

Fig. 1

Comparison of kernel and tree SHAP. For 10 activity classes, distributions of correlation coefficient (r) values for kernel and tree SHAP calculations, corresponding to approximated and exact SHAP values, respectively, are reported in boxplots. Black horizontal lines indicate median values. Results are shown for classification (activity prediction, top) and regression (potency value prediction, bottom) models using RF (blue) and ExtraTrees (red)

Fig. 2

Interpretation of GB-based compound potency prediction. a For an exemplary prediction, a feature importance ranking is shown including features with positive (red) and negative (blue) contributions to the prediction of the high potency value. Sequential arrows on the left are proportional to the feature contributions or SHAP values (shown on the pK_i scale). The summation of the expected value (7.7, gray) and all feature contributions yield the predicted pK_i value (9.6). Numbers in white preceded by # indicate top-ranked features. b From the top to the bottom, top-ranked features with positive contributions are iteratively mapped onto the test compound

Fig. 3

Interpretation of RF-based compound potency prediction. a Feature contributions to an exemplary prediction are depicted according to Fig. 2. The expected value (8.4, gray) and all feature contributions yield the predicted pK_i value (10.3). b Top-ranked features with positive contributions are mapped onto the test compound. c Top-5 and -10 ranked features are mapped onto three analog from the same series

Fig. 4

Comparative interpretation of RF- and GBM-based potency prediction. In a, positive (red) and negative (blue) feature contributions are compared for RF- and GBM-based regression models. The predicted pK_i values are shown in bold and different colors for RF (yellow) and GBM (orange). White numbers give indices of top-ranked features. b The top-1 and top-5 ranked features are mapped onto the compound. These features are common to both models

Fig. 5

SHAP-based interpretation of MT-DNN predictions. Each output neuron facilitates activity prediction of a different target (T_x). A SHAP-based explanation model is generated for each node/target. For a given test compound, each output prediction is rationalized

Fig. 6

Interpretation of MT-DNN activity predictions. In a and b SHAP analysis is shown for two inhibitors that were highly potent against two kinases. The top-1 and top-3 positive and negative features are mapped onto the compound and colored according to their contributions

Fig. 7

Rationalizing an error of a GB regression model. For a compound with high potency against muscarinic acetylcholine receptor M3, the potency value was ~ 1000-fold under-predicted by the GB model. SHAP identifies features making strong positive contributions to the prediction that are mapped onto the compound (left). By contrast, features with strongest negative contributions to potency prediction are absent in the compound. The corresponding atom environments are shown on the right

Fig. 8

Rationalizing an MT-DNN classification error. A highly potent inhibitor of ribosomal protein S6 kinase 1 was incorrectly predicted to be weakly potent against this target. SHAP analysis identifies a variety of features with strong negative contributions to the prediction. The top-6 ranked negative features are mapped onto the inhibitor. Three of these features are present in the compound, but the three others are absent. The corresponding atom environments are displayed