Federated learning-based protein language models with Apheris on AWS

This article discusses federated learning-based protein language models using Apheris on AWS, focusing on privacy-preserving AI model training for healthcare and life sciences organizations.

• Enables fine-tuning of protein language models (ESM-2) across multiple sites without sharing raw data
• Uses Amazon EKS, Amazon S3, and NVIDIA FLARE for secure, distributed model training
• Implements Low-Rank Adaptation (LoRA) to reduce trainable parameters by 98%
• Demonstrated effectiveness through experiments with balanced and imbalanced datasets
• Preserves data privacy while improving model performance through collaborative learning

The solution allows healthcare organizations to collaborate on AI model development while maintaining data privacy and intellectual property protection.