Source
Abramson et al. (2024), LIT-584. Read as NOTE-320.
The practice
The recommendation originates in LIT-584. Abramson et al. found AlphaFold 3’s diffusion head inventing compact structure in disordered regions, added AlphaFold-Multimer predictions to its training data, and report that this cross-distillation “greatly reduced” hallucination. The support is that phrase and an Extended Data figure this record has not read. What the main text does show is that the hallucinations which remain are low-confidence but not visibly disordered.
- Expect a generative head to fill gaps with plausible structure. The data under-determines disordered regions, and a sampler resolves that by inventing structure
- Distill from a model that regresses to the mean there. AlphaFold- Multimer’s predictions put disorder in an extended, recognizable form. Adding them to training teaches the generative model to do the same
- Rank against it at inference too. AF3 adds a solvent-accessible surface term to its ranking to favor extended over compact inventions
Conditions
- The evidence is the source’s word and an Extended Data figure that was not read here
- It imports the teacher’s convention, not ground truth. Extended loops are how AF2-style models draw disorder, not what disordered proteins look like