Personalized Federated Learning and AI for Drug Discovery: Challenges, Applications, and Cloud Solutions

Speaker Xu Chi, a senior engineer at Huawei, introduces his research on computer‑assisted drug design and new‑drug development, focusing on how personalized federated learning can empower AI in pharmaceutical research.

1. Challenges of drug screening – Over 4,500 diseases lack effective treatments; new‑drug development is costly, high‑risk, and time‑consuming, with average R&D investment rising from $800 million to $2.6 billion and development cycles exceeding 12 years.

2. Drug‑development workflow – Includes target discovery, pre‑clinical studies, clinical trials, and regulatory approval, requiring iterative design, screening, and optimization of compounds.

3. AI in drug discovery – AI techniques (supervised learning, reinforcement learning, generative models, explainable AI, graph computing) enable rapid data integration, target prediction, protein‑structure modeling, and biomarker discovery, dramatically shortening discovery timelines (e.g., InsilicoMedicine identified a candidate in 46 days).

4. Huawei Cloud Medical AI Platform – Built on Ascend+Kunpeng clusters and ModelArts, offering integrated algorithms, tools, and automated pipelines for end‑to‑end drug research.

5. Collaborative achievements – Joint projects with the Institute of Drug Research (iPhord protein‑structure prediction), drug‑repurposing, AutoOmics biomarker discovery, and large‑scale virtual screening using 15,000‑core Huawei cloud resources, reducing screening time from 30 days to 1 day.

6. Personalized federated learning – Describes federated learning as a distributed approach that protects proprietary drug data while enabling collaborative model training; introduces FedAMP, a personalized federated algorithm that outperforms standard FedAvg by weighting contributions based on model similarity and data quality.

7. Federated learning service workflow – Four steps: federation creation, partner invitation, agent deployment, and federation launch, allowing participants to monitor status and results in real time.

8. Large‑scale virtual screening cloud service – Provides visualized, high‑throughput screening of millions of compounds against targets (e.g., COVID‑19 proteins) via the cloud platform.

The session concludes with a Q&A on secure aggregation and a call for audience engagement.