Contribution
A structure predictor that reaches atomic accuracy in most cases, validated blind at CASP14 and on post-cutoff PDB chains.
Key insight
Reason jointly over the alignment and pairwise geometry, then build coordinates end to end. The pair representation is updated with triangle-consistent operations, and the whole network is iterated on its own output.
Key results
- CASP14 backbone: 0.96 Å median against 2.8 Å for the next best
- pLDDT vs lDDT-Cα: r = 0.76. pTM vs TM-score: r = 0.85 (10,795 chains)
- Self-distillation on about 350,000 predicted structures improves the network (Figure 4a, magnitude plotted)
- Alignment depth below about 30 sequences: accuracy drops sharply. Above about 100: small gains. Without both metagenomic databases: −6.1 GDT, with a few targets losing more than 20
Limitations
- Needs an alignment. Orphan and designed proteins are the weak case
- Poor on chains shaped by partners in heteromeric complexes
- Ablations only as plots, from three seeds of the baseline
- Supplementary methods not read