We are thrilled to announce the latest publication from our lab, titled "Uncertainty Quantification with Graph Neural Networks for Efficient Molecular Design" in Nature Communications. This work demonstrates a novel integration of uncertainty quantification (UQ) and graph neural networks (GNNs) for optimizing molecular properties in expansive chemical spaces.
In this study, we develop a framework that combines directed message passing neural networks (D-MPNNs) with genetic algorithms (GAs), and incorporates UQ to guide molecular design across single- and multi-objective tasks. By introducing a probabilistic improvement optimization (PIO) strategy, our method addresses the key challenge of maintaining predictive reliability under domain shifts—a common limitation in computational-aided molecular design (CAMD).
Key Findings:
-
Uncertainty-Informed Optimization:
The proposed PIO method significantly improves optimization success rates compared to traditional, uncertainty-agnostic approaches, particularly in multi-objective tasks involving complex trade-offs.
-
Superior Performance in Hit Rate:
In multiple benchmarks (e.g., Tartarus and GuacaMol), PIO achieved higher Top-100 hit rates than both direct objective maximization (DOM) and expected improvement (EI) methods, with up to 92% success rate in protein-ligand design and 83-90% in multi-objective median molecule generation.
-
Robust Multi-Objective Design:
Unlike conventional scalarization methods, PIO effectively balances diverse design goals, such as maximizing similarity and polarity while minimizing logP, enabling better chemical diversity and functionality.
-
Well-Calibrated Uncertainty:
The surrogate models showed strong agreement between predicted and actual uncertainty, validated by calibration error curves with AUCE < 0.1, confirming the soundness of the UQ estimates in practice.
This research provides a new direction for machine learning in chemical design—enabling reliable exploration in vast chemical spaces, especially when predictive uncertainty is crucial. The PIO approach offers a scalable, uncertainty-aware strategy to discover high-performing molecules for applications in materials science, pharmaceuticals, and catalysis.
For more details, access the full publication here.