CPDD: Adapting an autoregressive DNA language model for bidirectional phage genome modeling

CPDD, a 147.5M-parameter bidirectional discrete diffusion model for bacteriophage genomes, restores 42.4% of visible incorrect bases in a mixed-corruption test.

• Why it matters: Many phage genome tasks depend on both sequence flanks, but standard left-to-right autoregressive DNA models are not naturally bidirectional for restoration and editing.
• What they did: CreatiPhage Diffusion DNA-LM (CPDD) adapts a pretrained 144.5M autoregressive Transformer into a bidirectional discrete diffusion model, trained with masked and GIDD-style substitutions across 10 kb, 32 kb, and whole-source contexts.
• The result: On a screened panel of 522 phage CDSs, CPDD improves random-mask restoration versus a comparable diffusion model and enables infA fitness prediction plus seven-class virion protein subtype classification from frozen representations.

AI summary of the abstract · may contain errors · not medical advice · read the paper.