Research map / Probabilistic and scientific modeling / Molecules and proteins
Flexible Kernels for Protein Property Prediction
Martin Jankowiak, Yerdos Ordabayev, Rudraksh Tuwani, Henry Neil Ward, Hunter Nisonoff, James M McFarland, Gevorg Grigoryan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BLOSUM-based non-linear kernel instead of plain RBF kernel (ablation) | LOCK-GP · 21 local landscapes | predictive performance | Improved |
| add regularizing priors on kernel exponents (ablation) | LOCK-GP · 21 local landscapes | predictive performance | Improved |
| use local residue-wise exponents instead of global ones (ablation) | LOCK-GP · 21 local landscapes | fit | Improved |
| assume local linearity (product with linear kernel) (ablation) | LOCK-GP · 21 local landscapes | uncertainty quality | Improved |
| switch between different BLOSUM substitution matrices (ablation) | LOCK-GP · 21 local landscapes | predictive performance | No clear effect |
| add structure-conditioned correlation matrices (CLOCK) over BLOSUM LOCK | GP kernel · Thermostability multi-landscape | Spearman R (N=50): 0.498 -> 0.755 | Improved |
| learn a global correlation matrix instead of BLOSUM50 | LOCK-GP · Thermostability multi-landscape | Spearman R | Improved |
| add CNN zero-shot RBF kernel to CLOCK-GP | CLOCK-GP · Thermostability multi-landscape | Spearman R | Improved |
| train CLOCK-GP on only 10 landscapes instead of Global CLOCK-GP on 280 | CLOCK-GP · Thermostability multi-landscape | Spearman R | Improved |
| use LOCK-GP instead of Kermut-GP (two foundation models) | LOCK-GP · Extrapolation (128 training points) | Pearson R: 16/21 datasets | Improved |
| use LOCK-GP instead of Tanimoto-GP | LOCK-GP · Extrapolation (512 training points) | MAE: 22% lower | Improved |
| use ESM-2 ridge regression instead of Tanimoto-GP | Ridge-ESM2 · Unseen mutations (96 training points) | Pearson R: 0.560 -> 0.490 | Hurt |
Closest papers by meaning
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.86)
- Multi-Scale Representation Learning for Protein Fitness Prediction (NeurIPS 2024 · similarity 0.86)
- Enhancing Protein Mutation Effect Prediction through a Retrieval-Augmented Framework (NeurIPS 2024 · similarity 0.85)
- Kernel-Based Evaluation of Conditional Biological Sequence Models (ICML 2024 · similarity 0.85)
- Kermut: Composite kernel regression for protein variant effects (NeurIPS 2024 · similarity 0.85)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
