Predicting protein function directly from data works better than using shape

Researchers found that choosing protein designs based on structural confidence can actually lead to worse results for metal binding.

Scientists can design proteins to perform specific tasks. They can change protein sequences to create new functions. This could impact manufacturing and the materials economy.

Many methods use protein structure to guide design. These methods can create plausible protein candidates. However, structural confidence is a poor filter for protein function.

The researchers studied lanmodulin, a protein with four EF-hand domains. They wanted to design variants that bind rare earth elements. They focused on terbium and dysprosium, which are hard to separate.

One group used a model trained on measured metal-binding affinity. This model used protein sequence, ligand identity, and assay conditions. Another group selected designs based on structural confidence.

The group using structural confidence had a 16.7% success rate for binding terbium. The group using the function-guided model had an 83.3% success rate. The odds of binding were 25-fold different between the two panels.

The function-guided designs also showed high selectivity. They outranked a randomly drawn ortholog 84% of the time across 149 orthologs. The function-trained model's predicted pKdp K_{\text{d}} gave the highest ROC AUC in every panel.

Structural confidence is a signal for plausibility. It is not a selection criterion for function. Direct supervision on measured function is what ranks designs.

The goal is to develop a general protein function model. This model would design novel proteins for novel tasks. It would work at and beyond the frontier of protein mechanism explanations.

Key facts

  • The function-guided panel had an 83.3% success rate for binding terbium.
  • The structure-confidence-selected panel had only a 16.7% success rate for binding terbium.
  • The odds of binding differed 25-fold between the two panels.
  • The function-guided designs outranked a randomly drawn ortholog 84% of the time across 149 orthologs.

All sections

Edition