Development and Validation of an Explainable Machine Learning-Based Prediction Model for Drug-Food Interactions from Chemical Structures

Abstract

Possible drug-food constituent interactions (DFIs) could change the intended efficiency of particular therapeutics in medical practice. The increasing number of multiple-drug prescriptions leads to the rise of drug-drug interactions (DDIs) and DFIs. These adverse interactions lead to other implications, e.g., the decline in medicament's effect, the withdrawals of various medications, and harmful impacts on the patients' health. However, the importance of DFIs remains underestimated, as the number of studies on these topics is constrained. Recently, scientists have applied artificial intelligence-based models to study DFIs. However, there were still some limitations in data mining, input, and detailed annotations. This study proposed a novel prediction model to address the limitations of previous studies. In detail, we extracted 70,477 food compounds from the FooDB database and 13,580 drugs from the DrugBank database. We extracted 3780 features from each drug-food compound pair. The optimal model was eXtreme Gradient Boosting (XGBoost). We also validated the performance of our model on one external test set from a previous study which contained 1922 DFIs. Finally, we applied our model to recommend whether a drug should or should not be taken with some food compounds based on their interactions. The model can provide highly accurate and clinically relevant recommendations, especially for DFIs that may cause severe adverse events and even death. Our proposed model can contribute to developing more robust predictive models to help patients, under the supervision and consultants of physicians, avoid DFI adverse effects in combining drugs and foods for therapy.

Keywords: DrugBank; FooDB; adverse food reaction; chemical informatics; drug–food interactions; drug–nutrient interactions; explainable artificial intelligence; machine learning; precision medicine; simplified molecular-input line-entry system.

Figure 1

The workflow of our study. First, we obtained the SMILES notations of drug and food constituents from DrugBank and FooDB databases. After pre-processing, we filtered out 1133 drugs and 4341 food compounds, making 2,382,903 drug–food pairs in the benchmark dataset. We subsequently used PyBioMed and RDKit packages in Python to extract 3780 features of each interacting drug–food pair. We applied a four-step feature selection process to the training set to find the 18 most important features. Five classification algorithms were applied to the training data via five-fold cross-validation. As XGBoost gave the best prediction outcome, we fine-tuned it using the validation set. Finally, we tested our optimum XGBoost model on the internal test set and one external test set containing 1922 drug-food pairs. Finally, we used the model to recommend some common drug–food compound combinations.

Figure 2

Confusion matrix of our optimal XGBoost model on the testing and the external test sets. On the testing set (left plot): The model most accurately detected positive and non-significant DFIs (recall 0.99 in both classes) while only recognizing 87% of negative DFIs. Likewise, on the external test set (right plot), the model recognized all positive DFIs and 99% of non-significant DFIs. Negative DFIs were recognized as acceptable, with 94% of those discriminated against.

Figure 3

The SHAP (SHapley Additive exPlanations) plot of eighteen optimal features. The red dots of MRVSA0, EstateVSA2, MRVSA9, MRVSA8 and blue dots of PEOEVSA5, MTPSA+MTPSA, VSAEstate10+VSAEstate10 gather on the right side of the x-axis, indicating that the high values and low values of these features, respectively, direct the model in recognizing the non-significant DFIs. High PEOEVSA5, EstateVSA7, slogPVSA9, MTPSA+MTPSA, and low values of MRVSA0, MRVSA9 help detect the negative DFIs. The positive DFIs are identified by the increasing values of PEOEVSA5, EstateVSA0*LabuteASA, EstateVSA1*VSAEstate8 and the decline of PEOEVSA9, EstateVSA7, EstateVSA2, slogPVSA9, MRVSA2, VSAEstate7+VSAEstate7, slogPVSA0, PEOEVSA12.