Summary
In protein structure prediction, uncertainty metrics are used to evaluate the correctness of a structural model at various levels. They have been applied to protein design tasks in ways analogous to energy functions, prompted in part by the observation that these networks have learned an implicit energy function — even though they cannot be used directly for design.
Related notes
- Diffusion-based protein structure prediction methods double as energy methods comparable to traditional force fields
- Protein structure prediction and design metrics don’t correlate with expression probability
- Confidence metrics for diffusion-based structure prediction methods can be improved with minimal changes to conditioning representations
- Including structure prediction confidence while training inverse folding improves sequence diversity but not sequence recovery
- Most ML quality metrics cannot effectively predict enzyme activity after controlling for similarity to native
- pLDDT correlates with number of homologous sequences provided during runtime
- Protein structure prediction and design confidence metrics do not correlate with binding affinity
- pLDDT and PAE inversely correlated with protein dynamics in dynamic naturally occurring proteins, but not de novo proteins
- pLDDT is inversely correlated with CDRH3 length
- Protein folding neural networks cannot predict protein stability
- Self-consistency perplexity is correlated with pLDDT
- AlphaFold3 ipTM can distinguish between antibody binders and nonbinders
- Inverse folding sequence perplexities correlate with Rosetta energies, forward folding TM-scores, and sequence recovery
- PAE weakly correlates with Ab-Ag binding