AMP-DiT: Antimicrobial Peptide Design with AMP-classifier Conditional Diffusion Transformers

Existing AMP generation methods largely rely on protein language models such as ESM-2, which are trained on full-length proteins rather than peptide-specific distributions, leading to representation mismatch. Moreover, current approaches struggle to generate sequences that align with AMP classifier signals, despite activity prediction being the most critical objective in AMP design. We introduce AMP-DiT, a conditional discrete diffusion framework that generates peptides directly in sequence space, avoiding reliance on pre-trained protein language models. Our method uses classifier conditioning from \textit{Macrel} to bias generation toward high antimicrobial activity, improving functional quality during sampling. Conditioning on a single predictor generalizes across multiple AMP prediction models, indicating that AMP-DiT captures underlying antimicrobial features rather than overfitting. Overall, AMP-DiT outperforms existing AMP design methods across almost all available AMP-classifier metrics while maintaining diversity.