Federated Discrete Denoising Diffusion Model for Molecular Generation with OpenFL

TL;DR

The paper addresses privacy barriers in drug discovery where data are siloed and centralized training is often impractical. It integrates a discrete denoising diffusion model (DiGress) with OpenFL to enable privacy-preserving federated training for graph-based molecular generation across collaborators, while a regressor guides property-driven generation via markers such as $HOMO$ and $\mu$. Empirical results on the QM9 dataset show the federated model achieving performance comparable to centralized training across $NLL$, $MAE$, $Validity$, and $Uniqueness$, with only modest absolute differences. This work demonstrates the practicality of federated learning with confidential computing for AI-driven drug discovery and outlines pathways to scale to larger datasets and more robust privacy-preservation techniques.

Abstract

Generating unique molecules with biochemically desired properties to serve as viable drug candidates is a difficult task that requires specialized domain expertise. In recent years, diffusion models have shown promising results in accelerating the drug design process through AI-driven molecular generation. However, training these models requires massive amounts of data, which are often isolated in proprietary silos. OpenFL is a federated learning framework that enables privacy-preserving collaborative training across these decentralized data sites. In this work, we present a federated discrete denoising diffusion model that was trained using OpenFL. The federated model achieves comparable performance with a model trained on centralized data when evaluating the uniqueness and validity of the generated molecules. This demonstrates the utility of federated learning in the drug design process. OpenFL is available at: https://github.com/securefederatedai/openfl

Figures (5)

• Figure 1: OpenFL Workflow API: A global model is sent to collaborators for validation and local training. Each collaborator validates and updates the model with local data, then sends the updated model back to the aggregator. The aggregator combines these updates to form a new global model, completing one federation round.
• Figure 2: Model training loss across rounds
• Figure 3: Model validation loss across rounds
• Figure 4: Sampling metrics across rounds
• Figure 5: Example of three molecules generated by the federated model. The diffusion chains illustrate generation from a noisy graph to a plausible molecule