CASTER: Predicting Drug Interactions with Chemical Substructure Representation

Kexin Huang; Cao Xiao; Trong Nghia Hoang; Lucas M. Glass; Jimeng Sun

Source-linked AI summary

CASTER: Predicting Drug Interactions with Chemical Substructure Representation

Kexin Huang, Cao Xiao, Trong Nghia Hoang, Lucas M. Glass, Jimeng Sun

arXiv:1911.06446v2Nov 15, 2019cs.LGq-bio.QMstat.ML

TL;DR

Accurate, interpretable drug–drug interaction prediction from chemical structures remains challenging. CASTER uses mechanism-informed chemical substructures and dictionary learning, achieving more accurate and interpretable predictions than prior generic-representation approaches.

Problem
DDI prediction lacks specialized chemical representations that focus on interaction-relevant substructures rather than entire molecular structures.
Method
CASTER mines frequent chemical substructures and uses dictionary learning coefficients to represent and explain predicted drug interactions.
Results
CASTER provided more accurate and interpretable DDI predictions than previous approaches using generic drug representations.
Takeaways & Limitations
CASTER supports mechanism-informed, substructure-based DDI prediction with explanations tied to the relevance of individual molecular subgraphs.
Takeaways & Limitations
CASTER’s explanations are not intended to interpret the chemical process regulating drug interactions.

Abstract

from arXiv · show

Adverse drug-drug interactions (DDIs) remain a leading cause of morbidity and mortality. Identifying potential DDIs during the drug design process is critical for patients and society. Although several computational models have been proposed for DDI prediction, there are still limitations: (1) specialized design of drug representation for DDI predictions is lacking; (2) predictions are based on limited labelled data and do not generalize well to unseen drugs or DDIs; and (3) models are characterized by a large number of parameters, thus are hard to interpret. In this work, we develop a ChemicAl SubstrucTurE Representation (CASTER) framework that predicts DDIs given chemical structures of drugs.CASTER aims to mitigate these limitations via (1) a sequential pattern mining module rooted in the DDI mechanism to efficiently characterize functional sub-structures of drugs; (2) an auto-encoding module that leverages both labelled and unlabelled chemical structure data to improve predictive accuracy and generalizability; and (3) a dictionary learning module that explains the prediction via a small set of coefficients which measure the relevance of each input sub-structures to the DDI outcome. We evaluated CASTER on two real-world DDI datasets and showed that it performed better than state-of-the-art baselines and provided interpretable predictions.

1 Introduction

CASTER addresses limitations in DDI prediction by representing drugs through informative chemical substructures, leveraging unlabeled chemical-structure data for generalizability, and explaining outcomes with sparse relevance coefficients.

Adverse DDIs cause substantial morbidity, mortality, and medical costs, making accurate knowledge during drug design important for patients and pharmaceutical companies.
DDI prediction lacks specialized drug representations because interactions often involve only a few functional substructures, while other substructures are less relevant.
CASTER uses chemical-structure data from drugs and other compounds, including unlabeled drug-food pairs, to improve the generalizability of drug representations.
CASTER improves interpretability by generating a small set of coefficients measuring each input substructure’s relevance to DDI outcomes.Drug pairs are projected onto a subspace defined by generalizable embeddings of frequent substructures.

3. Interpretable prediction. · 2 Method

CASTER combines mechanism-inspired drug representations, deep auto-encoding with labelled and unlabelled data, and prediction methods designed to be accurate and interpretable. Its interpretability explains how predictions are made, not the chemical processes regulating drug interactions.

2 Method: CASTER is presented as a framework for predicting drug-drug interactions from functional drug representations.The framework follows problem settings summarized earlier in the paper.
2 Method: The drug representation is inspired by the mechanism of drug interactions.This functional representation is described as a dedicated methodological component.
2 Method: A deep auto-encoding method uses labelled and unlabelled data to embed drug representations into a parsimonious latent space.The latent space is intended to generalize to new drugs.
2 Method: CASTER’s interpretation explains how predictions are made rather than the chemical process regulating drug interactions.The paper explicitly limits the intended scope of interpretability.
2 Method: The method therefore distinguishes prediction interpretation from mechanistic explanation of drug interactions.This distinction defines what the interpretability component is and is not intended to provide.
2 Method: The resulting representation is leveraged to generate accurate and interpretable predictions.The framework links representation learning with both predictive accuracy and interpretability.

2.1 Problem Settings

The problem is to predict whether two drugs interact from their chemical-structure representations. CASTER learns drug embeddings and maps drug pairs to interaction probabilities using reported and unreported pairs.

Problem formulation: DDI prediction takes two drugs or their representations as input and outputs whether an interaction exists.Each drug can be encoded as a SMILES string describing its atoms and bonds in depth-first traversal order.
Problem formulation: The dataset contains n drug SMILES strings, m reported interacting pairs, and τ drug pairs not reported with interactions.The unreported pairs are defined as U = (S × S) \ I.
Prediction objective: The model learns G : S × S → [0, 1], mapping each drug pair to the probability that the drugs interact.This probability-valued mapping represents the chance of interaction between the two input drugs.
Prediction objective: Each drug is embedded as z ∈ R^d through E : S → R^d, after which G maps the pair of embeddings to its interaction outcome.The embedding and prediction mappings are implemented in the deep embedding and interaction modules described in later sections.

2.2 Generating Functional Representations

CASTER generates functional drug-pair representations by mining recurring chemical substructures from SMILES sequences. Each pair becomes a multi-hot vector over frequent substructures, which outperforms classic fixed-sized fingerprints for DDI prediction.

Generating Functional Representations: A chemical substructure is frequent when its consecutive SMILES substring occurs above the practitioner-specified threshold η.The substring corresponds to a depth-first traversal representation of a molecular subgraph.
Generating Functional Representations: CASTER mines recurring chemical substructures by identifying frequent SMILES substrings across the molecular database.Depth-first traversal representations allow recurring subgraphs across drugs to be extracted efficiently.
Generating Functional Representations: Each drug pair is represented by a k-dimensional multi-hot vector indicating which frequent substructures occur in both drugs.The representation is generated for drug pairs in both labelled and unlabelled datasets.
Generating Functional Representations: The functional representation maps drug pairs to frequent substructures and has better predictive value than classic fixed-sized drug fingerprints in DDI prediction.Different substructures may interact with each other and contribute to drug interactions.

2.3 Latent Feature Embedding

CASTER learns a low-dimensional latent embedding of functional drug-pair representations using an encoder–decoder architecture. Its unsupervised reconstruction objective uses unlabelled drug pairs to improve features for predictors trained on limited labelled data.

Latent embedding: The module embeds functional representations of drug–drug or drug–food pairs into a d-dimensional latent space, where d ≪ k.The embedding is generated by a neural-network encoder parameterized by weights W_e and biases b_e.
Reconstruction: A decoder neural network maps each latent embedding back to its functional representation for reconstruction.The decoder is parameterized by weights W_d and biases b_d.
Reconstruction: Encoder and decoder parameters are learned by minimizing a reconstruction loss over the functional representations.This optimization jointly learns the parameters of both neural networks.
Unsupervised learning: The embedding procedure is completely unsupervised because its reconstruction objective requires only unlabelled drug pairs.Using unlabelled data helps extract features that improve predictors trained on a much smaller labelled dataset.

2.4 Interpretable Prediction with Dictionary Learning

CASTER uses dictionary learning to represent drug-pair functional features with projection coefficients that support prediction and interpretation. Each coefficient measures the relevance of an associated molecular subgraph to the predicted interaction.

CASTER replaces simple logistic regression with dictionary learning to help practitioners understand its predictions and identify substructures potentially leading to interactions.
Functional representations are encoded into latent vectors and projected onto a learned basis, producing coefficients r that form a dictionary representation for each drug pair.The projection uses an L2-norm regularization term to control projection complexity.
A neural network uses the dictionary representation r to compute the probability that a labelled drug pair has an interaction.
Each coefficient r_i indicates the statistical importance of its associated molecular subgraph C_i for the drug pair’s interaction, explaining CASTER’s prediction rationale.

3 Experiment · 3.1 Experimental Setup · CASTER Setup.

The experiments assess CASTER’s predictive accuracy, the value of unlabelled data under limited labels, and the interpretability of its dictionary module. Evaluation uses two labelled DDI datasets, additional unlabelled drug-compound pairs, defined metrics, selected hyperparameters, and repeated train-validation-test splits.

3 Experiment: The experiment examines whether CASTER improves DDI accuracy, benefits from unlabelled data with limited labels, and supports interpretation through its dictionary module.These correspond to the study’s three experimental questions: Q1, Q2, and Q3.
Datasets.: The unlabelled data include 9,675 drugs and 24,738 food constituents, with 220,000 randomly generated drug-drug pairs and 220,000 drug-food pairs.The drugs include experimental, nutraceutical, and investigational types, while food constituent SMILES strings come from FooDB.
Metrics.: Prediction accuracy and efficiency are measured with ROC-AUC, PR-AUC, F1 Score, and # Parameters.ROC-AUC and PR-AUC summarize ranking performance, while # Parameters measures model size.
CASTER Setup.: CASTER uses frequency threshold η = 50, yielding k = 1,722 frequent substructures, with latent dimension d = 50.The encoder and decoder use three-layer perceptrons with hidden size 500, while the predictor uses six layers with specified hidden sizes.
Evaluation Strategies.: Datasets are randomly split into training, validation, and testing sets in a 7:1:2 ratio, with five independent runs using different random splits.Early stopping uses validation ROC-AUC, and the validation-selected model is evaluated on the test set.

3.2 Q1: CASTER achieves higher accuracy in DDI prediction · 3.3 Q2: CASTER leverages unlabelled data successfully to improve prediction performance · 3.4 Q3: CASTER generates interpretable prediction – A case study

Across three questions, CASTER is evaluated for predictive accuracy, use of unlabelled data, and interpretability. It outperforms strong baselines, improves consistently with more unlabelled data, and identifies chemically meaningful substructures behind predicted DDIs.

3.2 Q1: CASTER achieves higher accuracy in DDI prediction: CASTER is compared with Logistic Regression, Nat.Prot, Mol2Vec, MolVAE, and DeepDDI for end-to-end DDI prediction.The baselines use SPM-derived representations, similarity heuristics, fingerprint embeddings, variational autoencoders, or neural predictors.
3.2 Q1: CASTER achieves higher accuracy in DDI prediction: CASTER provides more accurate DDI prediction than other strong baselines on the BIOSNAP and DrugBank datasets.Table 2 reports results for both datasets, with the first and second rows corresponding to BIOSNAP and DrugBank, respectively.
3.3 Q2: CASTER leverages unlabelled data successfully to improve prediction performance: With only a small labelled set, increasing unlabelled data consistently improves CASTER’s DDI prediction accuracy on both datasets.The experiment varies unlabelled-data size while evaluating performance on the test set.
3.4 Q3: CASTER generates interpretable prediction – A case study: CASTER assigns coefficients to input substructures, showing their relevance to predicted DDIs, a feature missing from other baselines.These coefficients support interpretation of which chemical substructures contribute to the predicted interaction.
3.4 Q3: CASTER generates interpretable prediction – A case study: For Sildenafil and Isosorbide Mononitrate, CASTER predicts interaction with a 0.7928 confidence score.The case study examines interaction with a nitrate-based drug.
3.4 Q3: CASTER generates interpretable prediction – A case study: The case study demonstrates that CASTER can suggest sparse and reasonable substructure cues likely responsible for DDIs.Its coefficients are presented as evidence for understanding the interaction mechanism.

4 Conclusion

CASTER is an end-to-end dictionary learning framework for DDI prediction that uses a specialized, chemical-mechanism-inspired drug representation. Empirically, it produced more accurate and interpretable predictions than approaches using generic drug representations.

CASTER is an end-to-end dictionary learning framework with a specialized representation inspired by the chemical mechanism of drug interactions.
CASTER provided more accurate and interpretable DDI predictions than previous approaches using generic drug representations.
Future work will extend CASTER to chemical sub-graph embedding.

A Implementation Details

The implementation preprocesses SMILES with RDKit, omits unprocessable strings, and evaluates DrugBank using a five-fold experimental split. This preprocessing reduces the available drugs and positive DDI samples in both datasets.

Preprocessing: RDKit converts SMILES strings to canonical form, and data points that cannot be processed into Mol format are omitted.This preprocessing step removes invalid or unsupported SMILES entries before modeling.
Dataset preprocessing: 1,850 drugs and 221,523 positive DDIs remain for DrugBank after preprocessing, compared with 1,322 drugs and 41,520 positive DDIs for BIOSNAP.DrugBank decreases from 2,159 to 1,850 drugs and 222,127 to 221,523 positive samples; BIOSNAP decreases from 1,514 to 1,322 drugs and 48,514 to 41,520 positive samples.
Evaluation protocol: DrugBank experiments divide the dataset into five exclusive folds and report average and standard deviation scores across five independent runs.This split scheme reduces experimental time while leveraging the large DrugBank dataset.