Researchers have developed a new artificial intelligence system that predicts whether a clinical trial will succeed by mapping the complex web of relationships between drugs, diseases, genes, and patient eligibility criteria. The model, called HGGT, short for Heterogeneous Gated Graph Transformer, outperformed existing methods across all three phases of drug testing, offering a potential tool to cut the staggering costs of drug development.
Clinical trials are notoriously risky, with only about 10% of candidate drugs advancing from preclinical testing to regulatory approval. Each approved drug costs an estimated $2 billion to bring to market, and many failures occur late in the process after massive investments. The new model addresses a key gap: previous prediction tools treated trial-related factors in isolation, missing how a drug’s target gene, the disease mechanism, and even the wording of eligibility criteria interact to influence outcomes.
HGGT uses a gated graph transformer architecture, a type of deep learning that treats clinical trials as a network of connected nodes. These nodes include trials, diseases, drugs, genes, targets, and abstracts. Unlike standard graph neural networks that apply equal weight to all connections, HGGT dynamically learns which relationships matter most for a given trial. For example, it can identify when a strong genetic link between a drug target and a disease is more predictive of success than the trial’s design features. The model was tested on a benchmark dataset called TOP, which was expanded with genetic and abstract data from the TRIALPANORAMA database, creating an augmented version named TOP-gene.
In head-to-head comparisons, HGGT achieved the highest scores on three key metrics across Phase I, Phase II, and Phase III trials: PR-AUC (precision-recall area under the curve), F1 score, and ROC-AUC (receiver operating characteristic area under the curve). These metrics measure how accurately the model distinguishes trials that will succeed from those that will fail. The results suggest that capturing heterogeneous, multi-scale interactions between biological and clinical entities is more effective than previous approaches that relied on simpler, homogeneous graphs or fixed relationship weights.
The findings also reinforce the importance of genetic evidence in trial success. Trials lacking strong support from human genetic data or genetically modified animal models are more likely to be terminated for lack of efficacy. Conversely, trials are more likely to be stopped for safety reasons if the drug target gene is either highly restricted in the human population or widely expressed across tissues. By integrating this genetic context into predictions, HGGT offers a more complete picture of why some trials fail.
Looking ahead, the researchers aim to refine HGGT for real-world use in trial design and resource allocation. The model could help sponsors prioritize which drug candidates to advance, identify trials at high risk of failure before they begin, and optimize eligibility criteria to improve patient recruitment. While further validation in prospective settings is needed, this approach signals a shift toward AI-driven, biology-aware decision-making in drug development, potentially accelerating the path from laboratory discoveries to approved therapies.