Title: Building a Generative Protein Evolution Engine with DPLM-Evo

Abstract: The structural and functional properties of proteins are defined by gradual evolutionary processes governed by biophysical and functional constraints. While protein language models extract complex evolutionary patterns from extensive sequence data, discrete diffusion-based protein language models (DPLMs) have emerged as powerful tools for both analysis and generation. Nevertheless, current DPLMs predominantly utilize masked diffusion, a mechanism that conflicts with the fundamental biological principle that proteins evolve via accumulated edits rather than materializing from masked states. This reliance results in a lack of explicit pretraining objectives for substitution and insertion/deletion (indel) operations, thereby restricting the models' capacity for optimization-style post-editing and flexible guided generation.

To overcome these challenges, we introduce DPLM-Evo, an evolutionary discrete diffusion framework designed to explicitly predict substitution, insertion, and deletion operations during the denoising phase. DPLM-Evo decouples the latent alignment space, which is upsampled in length, from the variable-length space of observed sequences, rendering indel-aware generation computationally feasible. Furthermore, to ensure that substitutions mirror genuine evolutionary processes, we incorporate a contextualized evolutionary noising kernel that generates mutation patterns informed by biological context.

In benchmark evaluations, DPLM-Evo demonstrates enhanced sequence understanding and delivers state-of-the-art performance in mutation effect prediction on ProteinGym within a single-sequence context. Additionally, the framework supports variable-length simulated evolution and facilitates the post-editing and optimization of existing proteins through explicit edit trajectories.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC