MIT researchers developed a method using sparse autoencoders to understand how protein language models make predictions. These models, similar to large language models but trained on amino acid sequences, are widely used for drug discovery and vaccine development but operate as black boxes. The sparse autoencoder technique expands neural network representations from hundreds to thousands of nodes, allowing individual features to be isolated and interpreted. Using AI assistant Claude, researchers can now identify which specific protein characteristics each node encodes, making the models more interpretable and potentially improving their application in biological research.
Sort: