Artificial Intelligence-Driven De Novo Peptide Drug Design: Architectures, Pipelines, and Translational Challenges

Abstract

In peptide drug design, function often emerges from the conformational ensemble encoded by a sequence and from how that ensemble engages a target. From this perspective, de novo peptide design can be viewed as an inverse problem of the sequence–structure–function relationship: starting from predefined functional and pharmacological objectives, one infers the required structural constraints and interfacial features and then generates candidate sequences that satisfy them. Achieving this goal requires a deeper mechanistic understanding of how peptide conformational ensembles, molecular recognition, and functional determinants jointly govern activity, as well as the ability to encode these principles in computational representations and models that are amenable to learning, explicit constraint, and optimization. Recent advances in artificial intelligence (AI)-driven approaches, including protein language models (pLMs), diffusion and latent diffusion models (LDM), conditional flow matching (CFM), and AI-assisted prediction and screening, are beginning to shift the field beyond sequence or structure generation toward function-oriented and controllable peptide design. At the same time, these advances have highlighted persistent bottlenecks in sequence–structure co-generation, conformational control in flexible systems, and interpretable multi-objective optimization. As de novo peptide design methods rapidly proliferate, technical routes are diversifying and benchmarking standards remain fragmented. This review therefore surveys recent progress in AI-driven generative architectures, control strategies, and emerging directions, with the aim of distilling reusable principles and providing a structured reference for reliable, translationally relevant functional peptide design.