Cyclic peptides containing backbone N-methylated amino acids (BNMeAAs) and d-amino acids (d-AAs) exhibit exceptional metabolic stability, membrane permeability, and oral bioavailability, representing a pivotal middle space scaffold between biologics and s small-molecule drugs. To optimize pharmacokinetic and physicochemical properties, various non-canonical amino acids (NCAAs) are frequently introduced; however, these modifications pose significant challenges for three-dimensional structure prediction. While Rosetta simplecycpepₚredict (SCP) can generate high-accuracy structures across a broad chemical space based on energy functions, it requires extensive conformational sampling, resulting in prohibitive computational costs. Recently, deep learning approaches such as AlphaFold3, Boltz-2, HighFold2, and NCPepFold have advanced cyclic peptide prediction but offer limited support for BNMeAAs and d-AAs. In previous work, HighFold-MeD was proposed, which distilled massive conformational data generated by Rosetta SCP into the AlphaFold2 framework, significantly accelerating prediction for modified cyclic peptides. Building upon this, the present study introduces HighFold-MeD2, an enhanced model based on Boltz-2. By leveraging the sequence, modification information, and cyclic constraints as inputs, HighFold-MeD2 utilizes the Pairformer-Diffusion architecture of Boltz-2 to learn the generation of all-atom coordinates from random initialization to Rosetta target structures. The predicted structures are subsequently refined via Amber force field energy minimization to resolve local steric clashes. Unlike its AlphaFold2-based predecessor, which relied on labor-intensive hard-coding for non-natural amino acids, HighFold-MeD2 achieves seamless integration through universal Chemical Component Dictionary (CCD) representations. It not only demonstrates superior prediction precision over HighFold-MeD and state-of-the-art baselines but also validates the feasibility of utilizing CADD-generated conformations to fine-tune all-atom diffusion models, establishing a highly accessible and extensible paradigm for cyclic peptide drug design.
Cao et al. (2026) studied this question.