AI Model Reads Molecules Atom by Atom to Predict Dangerous Drug Combinations

Every year, adverse drug events send more than 1.5 million people in the United States to emergency departments and contribute to nearly half a million hospitalizations, and drug interactions account for a substantial share of that burden. In primary care settings, drug interactions have been estimated to cause close to three percent of all hospitalizations, translating to over 200,000 admissions annually, while population-based analyses show that exposure to contraindicated drug combinations can raise the risk of emergency hospitalization by two to three times. With nearly half of U.S. adults taking at least one prescription medication and many taking five or more, the number of potentially interacting drug pairs grows combinatorially with every additional prescription. A new study published in Frontiers in Drug Safety and Regulation now reports a deep learning system, called InfoMedex, that predicts specific types of drug-drug interactions directly from molecular structure, and—crucially—explains which parts of a molecule drive its predictions.

The research team, led by Andrew Disharoon of the Medical University of South Carolina and Aaron J. Masino of Clemson University, assembled a dataset of 377,628 drug-pair samples drawn from DrugBank and the RxNav Drug and Drug Interaction dataset, with chemical structures retrieved through the PubChem API and therapeutic target information from the Therapeutic Target Database. Rather than relying on a small set of broad interaction categories, the team processed the free-text interaction descriptions in these databases—for example, sentences like “the metabolism of (R)-warfarin can be increased when combined with acetaminophen”—and extracted 168 normalized interaction templates by masking out the drug names. Each sample therefore consists of a drug pair, structural and target information for both compounds, and a set of binary labels indicating which of the 168 interaction types apply. Careful preprocessing removed formulation-specific terms such as “hydrochloride” or “maleate” so that, for instance, metoprolol succinate and metoprolol tartrate are treated as the same base drug, and a synonym list unified names like acetaminophen and paracetamol. Interactions with fewer than 50 samples were dropped to ensure enough data for reliable training.

A central design choice distinguishes InfoMedex from earlier approaches: the way molecules are turned into sequences a neural network can read. Instead of treating SMILES strings—compact text encodings of molecular structure—as ordinary character strings, the team converted them to Atom-in-SMILES notation, in which each token explicitly encodes an atom together with its local bonding environment, including ring membership, aromaticity, and attached hydrogens. This chemically meaningful tokenization preserves structural context that character-level splitting or compressed molecular fingerprints can lose, and it is particularly well suited to transformer architectures, where positional embeddings and self-attention can capture relationships between distant substructures. Prior work has shown that Atom-in-SMILES representations better capture local chemical context than character-level tokenization, facilitating substructure-level reasoning.

The architecture itself is a hybrid of transformer and convolutional components processing three input streams. Tokenized Atom-in-SMILES sequences for the drug of interest and for the interacting drug are each converted into dense embeddings of dimension 512 with learned positional encodings. Gene features derived from the Therapeutic Target Database, one-hot encoded across 1,143 features, pass through a two-layer feed-forward network into the same embedding space and are replicated across the sequence length. The three streams are concatenated into a unified sequence of size 3,072 and processed by a transformer encoder with eight attention heads, pre-layer normalization, and padding masks to keep padding tokens out of attention computations. A convolutional stack of four 1D layers with 124 filters and kernel size 7, ReLU activations, a dropout layer, and a linear classification head then produces logits for all 168 interaction labels. The full model contains roughly 70.35 million learnable parameters and was trained with sigmoid focal loss, an objective designed for severe class imbalance, reaching peak validation performance after an average of 22.8 epochs.

The performance results are striking, though they come with an important caveat. Under the primary evaluation, in which drug pairs are split randomly but individual drugs may appear in both training and test sets, the model achieved a micro-averaged area under the precision-recall curve of 0.857 and a macro mean AUPRC of 0.825 across the 168 labels. Accuracy averaged 0.997 and AUC-ROC 0.998, while precision-recall analysis showed the model maintaining precision above 70 percent at recall values as high as 80 percent. Given that only 0.603 percent of the more than 6.3 million test predictions were positive labels, the researchers deliberately relied on precision-recall curves rather than ROC curves, which prior studies have shown to be essentially uninformative under extreme class imbalance. Stratified analysis across frequency quantiles—from interactions positive in as few as 0.013 percent of samples up to the most common at 13.4 percent—showed median precision above 0.8 across all quantiles. For three very rare interaction types the model’s predicted probabilities failed to exceed the 0.5 decision threshold, but the probability distributions for positive and negative samples were largely separated, suggesting appropriate model uncertainty rather than outright failure.

That caveat emerged from a secondary, more demanding evaluation. When the team held out 10 percent of the drugs—86 compounds appearing as the drug of interest in at least one sample—and excluded every sample containing them, mean AUPRC fell from 0.825 to 0.341 across 142 evaluable labels. In other words, InfoMedex generalizes well to unseen drug pairs involving drugs it has encountered elsewhere in training, but predicting interactions for entirely unseen molecules remains substantially harder. This distinction matters enormously for the model’s ultimate use case: flagging risks for newly approved drugs that have no interaction history at all. The authors are candid that this limitation should temper expectations for near-term deployment on novel compounds.

Perhaps the most intriguing contribution is the explainability analysis. Using Integrated Gradients, a gradient-based attribution method implemented in the Captum library, the team computed the contribution of each atom token to selected predictions, aggregating per-sample ranks across many partner drugs to obtain a robust estimate of which molecular regions matter most. Because a given drug’s token sequence and input position are fixed across samples, the averaging marginalizes only over the partner drug, yielding a well-defined estimand: the typical within-drug importance ranking of each token over the empirical distribution of partners. Rank aggregation was chosen over magnitude-based averaging because it is robust to samples carrying disproportionate attribution mass, and interquartile ranges were reported to measure rank stability.

The case studies focused on two well-characterized cytochrome P450 inhibitors. For bupropion, a CYP2D6 inhibitor represented by 29 atom tokens, the highest-importance tokens clustered on the carbonyl-centered motif—the ketone carbonyl carbon had a mean rank of 1.88—and on the bulky N-alkyl tert-butylamine region, including the quaternary carbon at mean rank 6.42 and the basic nitrogen. Both patterns align with known pharmacology: the carbonyl is the starting point for reductive metabolites that show stronger CYP2D6 inhibition than the parent drug, and CYP2D6 substrates and inhibitors frequently present a basic nitrogen that contributes to binding in the enzyme’s tight active site. The meta-chloro phenyl region also ranked prominently, consistent with evidence that this motif discourages aromatic hydroxylation and redirects metabolism toward CYP2B6, which produces hydroxybupropion—the metabolite responsible for roughly two-thirds of bupropion’s in vivo CYP2D6-inhibiting activity.

Ritonavir, the HIV protease inhibitor deployed clinically precisely because of its potent CYP3A4 inhibition, presented a different picture. Represented by 83 atom tokens with 18 points of rotation, the molecule drew the model’s strongest emphasis to the central scaffold, where the hydroxyl token had the highest mean rank of 1.39 with a narrow interquartile range of 1.0 to 1.75 across 18 partner drugs, followed by an adjacent stereocenter. These atoms sit near the benzyl and phenyl bulk that provides steric stabilization, helping hold the drug in a binding-ready configuration within the CYP3A4 cavity. Structural studies have shown ritonavir is a type II ligand that coordinates the heme iron through a thiazole nitrogen while packing hydrophobic groups tightly into the active site, effectively locking the molecule into a high-affinity pose with a slow off-rate. The model also emphasized carbonyl-linked oxygen tokens and the methylene linker connecting the heteroaryl ring to the tertiary amine arm, features that plausibly govern how the bulky aromatic half and the heme-facing heteroaryl half engage the enzyme. Notably, some highly ranked tokens were non-atom structural signposts such as parentheses, which the authors suggest the model uses to guide its structural reading of the molecule rather than as chemically interpretable features.

Taken together, the results position InfoMedex as a scalable, interpretable complement to post-marketing surveillance and formal interaction studies, of which only a limited subset of clinically relevant combinations can feasibly be evaluated under FDA guidance. The ablation finding that randomly perturbing the target features did not significantly change performance metrics suggests the current one-hot gene encodings underexploit mechanistic context, and the authors propose graph neural networks or cross-attention architectures as avenues to better fuse drug-target data. The held-out-drug results temper enthusiasm for novel compounds, and the explainability analysis covered only interactions with well-studied CYP-mediated mechanisms. Still, the demonstration that a transformer reading molecules atom by atom can both predict 168 interaction types with high precision and point to the specific chemical motifs responsible marks a meaningful step toward AI systems that do not just flag drug interactions but can generate mechanistic hypotheses about why they occur—exactly the kind of transparent reasoning that clinical and regulatory adoption will demand.

Subject of Research: Deep learning prediction and explainability of drug-drug interactions using Atom-in-SMILES tokenization and a multimodal CNN-transformer architecture

Article Title: InfoMedex: drug-drug interaction prediction via a multimodal CNN-transformer model

Article References: Disharoon, A., Pasupuleti, S., Thurston, C., Fiore, M., McKenzie, S., Smith, A., & Masino, A. J. (2026). InfoMedex: drug-drug interaction prediction via a multimodal CNN-transformer model. Frontiers in Drug Safety and Regulation, 6, Article 1841035. https://doi.org/10.3389/fdsfr.2026.1841035

Image Credits: AI Generated

DOI: 10.3389/fdsfr.2026.1841035

Keywords: drug-drug interactions, deep learning, transformer, Atom-in-SMILES, explainable AI, Integrated Gradients, cytochrome P450, pharmacovigilance, DrugBank, bupropion, ritonavir, precision medicine