Abstract
Combining two clinically approved drugs has potential to improve treatment for common disease. But, with many thousands of combinations possible, clinically testing all pairs of drugs, with all common diseases, is not feasible. Here, we propose DRACO, a new machine learning method for discovering therapeutic drug combinations by leveraging knowledge about which health conditions each drug has been tested for. Our graph neural network model is able to generalize from a curated set of drug combinations to discover new combination therapies. We showcase our approach’s power to answer the question: given a drug and a health condition, what second drug would create an effective combination? Of our top three predicted candidate combinations, 74% have been previously reported. In the harder task of distinguishing the small number of reported combinations from millions of possible combination-condition triplets, our model ranks 94% of unseen triplets at the highest 0.1%. We expect DRACO to be a useful tool for proposing new therapies across thousands of diseases.
Introduction
The pursuit of synergistic drug combinations is essential for treating complex, multi-factorial diseases. While "cocktail" therapies can enhance efficacy and mitigate drug resistance, the search space for these combinations is vast. With approximately 2,000 FDA-approved drugs and over 3,000 common health conditions, the number of possible drug-drug-condition triplets reaches into the billions.
Traditional computational approaches often focus on specific therapeutic areas, such as oncology, or rely on narrow datasets like cell-line sensitivity. There remains a critical need for a generalized framework capable of operating across the entire "phenome"—the complete set of human clinical conditions. DRACO (Drug combinAtion for COnditions) addresses this gap by framing drug combination discovery as a link prediction problem within a biological knowledge graph.
Methodology
The DRACO framework utilizes a Multi-Layer Perceptron (MLP) architecture built upon pre-trained Knowledge Graph Embeddings (KGE).
Embedding Integration and Fine-Tuning
The model leverages the Drug Repurposing Knowledge Graph (DRKG) to provide a foundational understanding of biological entities. To incorporate drugs and conditions not present in the original graph, the authors employed a two-stage strategy. First, new embeddings were trained to align with the fixed DRKG vector space. Second, these embeddings were jointly fine-tuned to optimize performance for the specific task of combination prediction.
Model Architecture
The core of DRACO is an MLP that takes a triplet $(d_1, d_2, c)$ as input. To ensure the model is agnostic to the order of drugs (maintaining the property that $d_1 + d_2$ is equivalent to $d_2 + d_1$), the model computes scores for both permutations and aggregates them. This symmetry constraint is vital for biological realism.
Negative Sampling
A significant challenge in training such models is the absence of "negative" data—combinations known to be ineffective. DRACO utilizes a 100:1 negative-to-positive sampling ratio. Crucially, it employs "hard negatives," where a known drug pair is matched with an incorrect condition. This forces the model to learn specific therapeutic relationships rather than simply identifying frequently used drugs.
Results and Discussion
DRACO was evaluated on its ability to predict effective "partner" drugs and classify known therapeutic triplets against a background of millions of possibilities.
Predictive Accuracy in Drug Ranking
In the "second-drug" prediction task—identifying a suitable partner drug given one drug and a target condition—DRACO demonstrated remarkable precision. The model placed the correct partner drug within the top three candidates for 74% of the test cases.
Phenome-Wide Generalization
When subjected to a stringent "needle in a haystack" test involving 2 million potential triplets for a specific condition, DRACO ranked 94% of known effective combinations within the top 0.1% of all possibilities. Furthermore, the model exhibited strong zero-shot generalization; on a completely held-out set of 2,552 combinations, it assigned high potential scores to 21% of true combinations, significantly outperforming random baselines.
Conclusion
DRACO represents a significant step forward in phenome-wide drug discovery. By transferring knowledge from single-drug indications to multi-drug combinations, the model successfully navigates the astronomical search space of potential therapies. Future iterations of the model may incorporate multi-omics data and explicit safety/toxicity predictions to further refine the pipeline for clinical translation.
Summary updated