Source
Jumper et al. (2021), LIT-583, “MSA depth and cross-chain contacts”. Read as NOTE-321.
The account
Co-evolving positions in an alignment indicate which residues touch. Given enough sequences, the network can read out enough contacts to place the fold, and the rest of the network then refines geometry from learned physical regularities. Below a threshold, the contacts are too noisy to place the fold, and refinement has nothing correct to refine. The result is a threshold rather than a slope.
The threshold is LIT-583’s measurement, from its analysis of MSA depth: accuracy drops sharply below a median alignment depth of about 30 sequences and gains little above about 100 (Figure 5a), and removing both metagenomic databases costs 6.1 GDT, mostly from a few targets losing more than 20. The two-stage reading is the authors’ hypothesis for that shape, not a separate experiment.
What it explains
- The sharp drop below about 30 sequences and the plateau above about 100 (Figure 5a)
- Why removing metagenomic databases hurts only the few targets they push across the threshold
- Why single-sequence protein language models (LIT-505) are the natural fix. They supply the coarse signal from learned priors instead of an alignment
Where it is weak
- Offered as a hypothesis. The paper does not separate the two stages
- The intermediate-structure trajectories (Figure 4b) show some targets settling early and others late, which fits but does not test it