Extending AIMNet2 to Macrocyclic Peptides through Data-Efficient Continual Training

Abstract

Macrocyclic peptides are among the most promising therapeutic modalities, yet no general-purpose machine-learned interatomic potential (MLIP) has been systematically benchmarked for their conformational analysis. We show that on the MPCONF196 macrocyclic benchmark, the four-model AIMNet2 ensemble outperforms MACE-OFF23(S), despite macrocycles being entirely absent from its training data. A single AIMNet2 model is 3- to 5-fold faster and needs 5- to 63-fold less GPU memory than the MACE-OFF23 models at their default float64-precision. The four-model ensemble is about as fast as MACE-OFF23 but retains a 3- to 32-fold memory advantage even against MACE-OFF23 in float32. The larger MACE-OFF23(M) and MACE-OFF23(L) models remain more accurate. To narrow this gap, we introduce a continual training protocol that requires new quantum-chemical labels for fewer than 25,000 domain-specific conformers (<0.13% of the training corpus) while preserving conformational-energy accuracy on nonmacrocyclic molecules. On a sequence-disjoint holdout of unseen CREMP macrocyclic peptides, the resulting model, AIMNet2-macro, reduces conformational energy RMSE by 25 to 27% and DFT force RMSE by 20% (25% on backbone atoms) and improves lowest-energy conformer recovery relative to the original AIMNet2. It retains the conformational-energy accuracy of the original model on the nonmacrocyclic drug-like molecules of CONFORGE and also improves accuracy on the structurally distinct macrocycles of MPCONF196, although this smaller improvement is not statistically significant. Among the ten training strategies compared, only continual training on the combined corpus (AIMNet2-macro) achieves this improvement without degrading small-molecule accuracy. AIMNet2-macro is therefore the best balanced, though not the most accurate, of the models tested. As an independent case study, AIMNet2-macro reproduces DFT conformational energies of six cyclosporine A conformers with an RMSE of 0.78 kcal mol–1 and the correct rank ordering, the lowest error among the MLIPs and semiempirical methods evaluated on this system. The continual training strategy has been adopted in the AIMNet2 (2025) release, which unifies small-molecule, macrocyclic, and intermolecular-interaction treatment in a single production potential.