Improving Variant Effect Prediction by Steering Sparse Mechanistic Features in Protein Language Models biorxiv.org paper View original →