StackNAFLD: An Accurate Stacking Ensemble Learning Targeting NAFLD Treatment

Abstract

Nonalcoholic fatty liver disease (NAFLD) is a slow-progressing yet complex disease with multiple pathophysiological mechanisms that make it challenging to treat. In this study, we developed a machine learning (ML)-based stacking ensemble model to predict molecules that could inhibit NAFLD progression utilizing data from animal experiments. We systematically collected 75 agents from preclinical experiments and classified them as inducers and inhibitors based on each study end point. Then, we computed 12 sets of molecular fingerprints and trained them with three baseline ML models. After that, the stacked model was trained using the predictive features from the baseline models and validated with 5-fold cross-validation (5-CV) and leave-one-out cross-validation (LOOCV). We found that the stacked model outperformed its baseline model across various evaluation metrics, thereby improving the prediction of the NAFLD inhibitory activity. Additionally, we tested the robustness and applicability domain of the stacked model, ensuring that this model delivered a trustworthy prediction. Moreover, we highlighted key molecular features, such as carboxylic, alkene, or aromatic rings, underscoring their influence on the decision-making of the stacked model. In conclusion, we have provided an effective method for improving molecular property prediction by using the stacking ensemble learning approach. Furthermore, we hosted our software in an open-access GitHub repository for further reproducibility and use in the drug discovery pipeline.

1

Stacking ensemble learning model architecture used in this study.

2

Chemical distribution of the in vivo NAFLD data set using the t-SNE distribution plot.

3

5-CV AUROC scores of RF baseline models based on 13 molecular fingerprints.

4

5-CV AUPRC scores of RF baseline models based on 13 molecular fingerprints.

5

5-CV AUROC scores of SVM baseline models based on 13 molecular fingerprints.

6

5-CV AUPRC scores of SVM baseline models based on 13 molecular fingerprints.

7

5-CV AUROC scores of XGB baseline models based on 13 molecular fingerprints.

8

5-CV AUPRC scores of XGB baseline models based on 13 molecular fingerprints.

9

5-CV AUROC and AUPRC metrics of the stacked model. (A) and (B) represent the average AUROC and AUPRC metrics of the stacked model without the Combined fingerprint, while (C) and (D) show the corresponding metrics when the Combined fingerprint is included.

10

y-Randomization experiment of the StackNAFLD model.

11

Top feature from (A) StackNAFLD model and (B) SubFP with XGB baseline model.

12

Example of structure importance for NAFLD classification.