Sufficiency of sequence information for improving backbone generation in hybrid co-generation models

Determine whether integrating explicit sequence-generation processes into backbone structure generation within hybrid sequence-structure co-generation models, such as DiffAb, MultiFlow, and CarbonNovo, is sufficient to produce more plausible backbone structures than backbone-only methods like RFdiffusion, as measured by designability and diversity metrics.

Background

Hybrid co-generation models combine discrete sequence generation (via diffusion or flow matching) with continuous backbone generation in a shared neural network, aiming to co-generate both modalities. Examples include DiffAb and MultiFlow, which add a discrete sequence process to a backbone generation model, and CarbonNovo, which interleaves backbone diffusion with a structure-conditioned Markov random field for sequence.

Despite this integration, the authors note these approaches require paired sequence-structure data and do not directly generate side-chain atom positions. Crucially, they report that current hybrid methods have not demonstrated substantial gains in designability or diversity compared to backbone-only approaches like RFdiffusion, raising the question of whether adding sequence information is actually sufficient to improve backbone plausibility.