Abstract
Improving drug discovery efficiency is a core and long-standing challenge in drug discovery. For this purpose, many graph learning methods have been developed to search potential drug candidates with fast speed and low cost. In fact, the pursuit of high prediction performance on a limited number of datasets has crystallized them, making them lose advantage in repurposing to new data generated in drug discovery. Here we propose a flexible method that can adapt to any dataset and make accurate predictions. The proposed method employs an adaptive pipeline to learn from a dataset and output a predictor. Without any manual intervention, the method achieves far better prediction performance on all tested datasets than traditional methods, which are based on hand-designed neural architectures and other fixed items. In addition, we found that the proposed method is more robust than traditional methods and can provide meaningful interpretability. Given the above, the proposed method can serve as a reliable method to predict molecular interactions and properties with high adaptability, performance, robustness and interpretability. This work would take a solid step forward to the purpose of aiding researchers to design better drugs with high efficiency.
Introduction
In the race to accelerate drug discovery, graph learning has become a primary tool for predicting molecular interactions and properties. However, most current models suffer from "crystallization"—they are meticulously hand-designed and tuned for specific, limited datasets. When these models encounter new, real-world data, their performance often declines because their fixed architectures and hyperparameters lack flexibility.
Achieving state-of-the-art results currently requires significant deep learning expertise to manually tune optimizers, loss functions, and message-passing layers. The Graph Learning based Adaptive Machine (GLAM) addresses this by creating a system that adapts itself to any dataset without human intervention, making high-powered molecular modeling accessible and robust.
Methodology
The core of GLAM is an automated pipeline that treats model design as a multi-stage selection process, navigating a vast configuration space of architectures and hyperparameters.
The Adaptive Pipeline
- Configuration Sampling: The system starts with a massive configuration space that includes choices for neural architectures (number of layers, types of message-passing), hyperparameters (learning rates, batch sizes), optimizers, and loss functions.
- Low-Fidelity Training: To optimize computational resources, GLAM samples many configurations and trains them for just a few epochs.
- High-Fidelity Training: The top-performing configurations from the low-fidelity stage are selected for full-scale training using multiple random seeds.
- Predictor Blending: Instead of picking a single "winner," GLAM blends the best-performing predictors. This ensemble approach is a key driver of the model’s stability and accuracy.
General Architectures
The researchers developed two general architectures: a single-graph version for molecular properties (such as solubility) and a pair-graph version for interactions (such as drug-protein or drug-drug binding).
Experimental Results
The authors evaluated GLAM on 14 representative datasets covering a wide spectrum of pharmacology, including molecular interactions (LIT-PCBA, BindingDB, DrugBank) and molecular properties (ESOL, Lipophilicity, FreeSolv, BACE, BBBP, SIDER, Tox21, ToxCast).
Superior Performance and Stability
GLAM achieved the best scores on all 14 tested datasets. Compared to the previous best-reported results, GLAM reduced prediction errors by an average of 18.7%. In drug-drug interaction tasks on the DrugBank dataset, GLAM achieved a ROC-AUC of 0.911, significantly outperforming established baselines like CASTER (0.861) and DeepDDI (0.844).
Robustness and Interpretability
Using a new algorithm called PASP (Property-Slightly-Affected Structure Perturbation), the authors demonstrated that while standard models like GCN and GAT saw performance drops when molecular structures were slightly tweaked, GLAM’s error rate remained much lower and more stable.
Furthermore, GLAM offers chemical intuition. By visualizing the hidden states of the message-passing blocks, the researchers showed the model naturally "attends" to chemically relevant groups—for example, correctly identifying hydroxyl and amino groups as the primary drivers of hydrophilicity.
Conclusion
GLAM demonstrates that an automated search can consistently outperform models manually tuned by human experts. By being more granular in its optimization than a human could practically be, GLAM sets a new standard for adaptability in drug discovery. Future work will explore "adaptive feature input" to further boost performance for niche tasks and expand the framework to other scientific fields like materials science.
Summary updated