Direct Coupling Analysis predicts point-mutation costs accurately but is a poor generative model of allosteric fitness

In the synthetic allosteric-network alignment of Bravi, Ravasio, Brito, and Wyart (claim-bravi-2020-applies-dca-to-evolved-allosteric-networks-as-synthetic-msa), Direct Coupling Analysis shows a sharp asymmetry between two of its uses. Discriminatively — ranking the fitness cost of a single point mutation — it is accurate: the DCA-inferred single-mutation cost map matched the true cost map with high correlation, "the comparison is excellent, as evident also from the high correlation revealed by the scatter plot." Generatively — sampling new sequences from the inferred model — it fails: "the mean obtained fitness is rather low," and the authors conclude "the generative power of DCA is limited in the context of allostery."

The paper's own compression of the result: "DCA predicts well the cost of point mutations but is a rather poor generative model." The two capacities come apart because a model can score the local effect of flipping one link correctly while still failing to reproduce the joint distribution needed to synthesize a whole functional configuration from scratch — the generative task loads on higher-order structure that the pairwise DCA statistics do not carry.

This bears on the vault's entity-direct-coupling-analysis cluster and its broader interest in energy-based and generative models: DCA is itself a maximum-entropy (Potts/Boltzmann) model, so a documented generative shortfall is a data point on where such pairwise energy-based inference stops working. It sits beside the distance-dependent failure in claim-dca-underestimates-long-range-epistasis-in-allosteric-materials; the two are distinct limits (generative fidelity vs. long-range epistasis), and the paper raises the possibility that a neural network could improve on both — a lead left in the capture, not promoted here.