Source
Jumper et al. (2021), LIT-583, and Abramson et al. (2024), LIT-584. Read as NOTE-321 and NOTE-320.
The practice
The practice starts with LIT-583. Jumper et al. trained AlphaFold 2 to regress the lDDT-Cα its own prediction would score (pLDDT) and the TM-score (pTM). Over 10,795 chains these track the truth at r = 0.76 and r = 0.85, and the same paper used the confidence to filter its self-distillation set. LIT-584 carries the head into a generative model: AlphaFold 3 trains it on a cheap rollout of the diffusion sampler, ranks seeds by it, and antibody–antigen quality keeps rising with the number of seeds ranked by ipTM.
- Supervise a confidence head with the real metric. Train it against the metric the prediction will actually be scored on (lDDT per residue, TM-score per chain), computed on the model’s own prediction during training
- For a sampler, train it on a sample. AF3 runs a cheap, large-step rollout of the diffusion sampler during training so the head sees a complete structure
- Use it to spend inference compute. Draw several seeds or samples and keep the one the head rates best. Add hard penalties for constraint violations (chirality, clashes) that the head does not capture
- Use it to curate training data (SOTA-354)
Conditions
- Calibration is shown in aggregate (r = 0.76 for pLDDT). A single confident prediction can still be wrong
- The head inherits the model’s blind spots. AF3’s hallucinated regions are low-confidence, but not recognizably disordered