Direct Coupling Analysis captures short-range but substantially underestimates long-range epistasis in allosteric materials

The central finding of Bravi, Ravasio, Brito, and Wyart (claim-bravi-2020-applies-dca-to-evolved-allosteric-networks-as-synthetic-msa) is that Direct Coupling Analysis reads epistasis in a distance-dependent way. In their evolved allosteric networks, DCA's predicted epistasis between two links (∆∆Eij, equal by construction to the magnitude of its inferred couplings |Jij|) tracks the true epistasis (∆∆Fij) closely for pairs that are close together, but "strongly underestimates long-range epistasis."

The authors trace the failure to the statistics of the alignment itself: even where long-range epistasis is genuinely strong, the synthetic MSA shows no long-range statistical correlations for DCA to read. "The absence of long-range correlations suggests that it will be particularly challenging to capture long-range functional dependencies from low order statistics of the MSA alone." The diagnosis is structural rather than incidental — the pairwise, low-order statistics DCA is built on do not encode a coupling that expresses itself only across distance, so no amount of the same kind of data recovers it. The paper offers a toy Boolean AND/OR model as a mechanism for why cross-subpart epistasis specifically is the part that vanishes (left as a lead in the capture, not promoted here).

This is the claim the paper was built to explain — it was constructed to account for the empirical report (Anishchenko et al., PNAS 2017) that real allosteric proteins show "no statistical evidence for the existence of long-range direct couplings" under DCA. The prediction is then tested against real protein data in claim-pdz-dca-couplings-track-short-range-epistasis-more-than-long-range, and it sits beside the separate generative shortfall in claim-dca-predicts-mutation-costs-but-poor-generative-model-of-allostery. It connects to the vault's entity-direct-coupling-analysis cluster and, through the soft-mode reading of trained networks (claim-physical-networks-become-what-they-learn-soft-modes), to whether shared structure implies shared readability (backpropagation-gap).