anthology-of-the-sota

LIT-584: Accurate structure prediction of biomolecular interactions with AlphaFold 3

2 min read

Properties6
StatusActive
Tagsbiomolecular-modeling, generative-modeling, model-architecture
ExtendsLIT-583 — Highly accurate protein structure prediction with AlphaFold
Cited as “introduced by” bySOTA-352 — When a structure predictor is made generative, add a regression model's predictions to the training data so disordered regions are learned as disorder rather than invented structure
Cited as “paper” byNOTE-320 — AlphaFold 3
Cited as “source” bySOTA-352 — When a structure predictor is made generative, add a regression model's predictions to the training data so disordered regions are learned as disorder rather than invented structure, SOTA-353 — Train a head that predicts the model's own accuracy on each output, against the metric you care about, and use it to rank samples and filter predictions

Graph View

Abramson, Adler, Dunger, Evans, Green, Pritzel, Ronneberger, Willmore et al., Google DeepMind and Isomorphic Labs (Nature 630, 2024) — DOI 10.1038/s41586-024-07487-w

Key takeaways

  • Architecture: a pairformer (48 blocks) with reduced MSA processing replaces the Evoformer. A diffusion module predicts raw atom coordinates, replacing AF2’s frames and torsions. Neither needs invariance or equivariance to global rotations, which “simplif[ies] the machine learning architecture”
  • Results (Figure 1c): on PoseBusters (428 protein–ligand complexes, run with a model trained to a 2019 cutoff), it beats AutoDock Vina (P = 2.27×10⁻¹³), although Vina is given the solved pocket and AF3 is not. It beats RoseTTAFold All-Atom (P = 4.45×10⁻²⁵). It is also better than RoseTTAFold2NA on protein–nucleic acid complexes under 1,000 residues, and than AlphaFold-Multimer v2.3 on protein–protein and antibody interfaces. The success rates are plotted, not stated in the text
  • Hallucination: a generative model invents plausible structure in disordered regions. Training data enriched with AlphaFold-Multimer predictions, where disorder appears as extended loops, “greatly reduced” this (Extended Data Figure 1, not read)
  • Confidence under diffusion: training runs a cheap “mini-rollout” of the sampler to get a full structure, so the confidence head can be trained against that structure’s true error
  • Sampling: five seeds × five diffusion samples, ranked by predicted confidence with chirality and clash penalties. For antibody interfaces, quality keeps rising with more seeds, up to 1,000 (Figure 5a)
  • Limits: 4.4% chirality violations on PoseBusters, clashes in very large protein–nucleic complexes, hallucination that is low-confidence but not ribbon-like, and one static conformation. For example it always predicts cereblon closed, as bound, even when no ligand is given (apo)

Standing in the anthology

Filed from #163 (“AlphaFold”). It extends AlphaFold 2 (LIT-583). It sources SOTA-352 (cross-distillation against hallucination) and is a second source for SOTA-353 (confidence-ranked sampling).

On the headline comparisons. The PoseBusters comparison is fairer to the baseline than to AF3, since Vina gets the pocket. The success percentages are only in figures, so none are quoted here. The code and weights were released under restricted terms, which limits independent checks.

Read — NOTE-320.

Backlinks