Extending AIMNet2 to Macrocyclic Peptides Through Data-Efficient Continual Training
4 affiliations
- 1Department of Chemistry
- 2Carnegie Mellon University
- 3Computational Biology Department, School of Computer Science
- 4Department of Materials Science and Engineering
Abstract
Macrocyclic peptides are among the most promising therapeutic modalities, yet no general-purpose machine learned interatomic potential (MLIP) has been validated for their conformational analysis. We show that AIMNet2 matches the accuracy of MACE-OFF23(S) on the MPCONF196 macrocyclic benchmark, despite macrocycles being entirely absent from its training data, at three-to sixfold lower computational cost. To close the remaining accuracy gap, we introduce a continual training protocol that integrates fewer than 25,000 domain-specific conformers (<0.1% of the training corpus) while preserving accuracy on non-macrocyclic molecules, reducing macrocyclic conformational energy RMSE by 25–27%. Applied to cyclosporine A as an independent test case, the continually trained model reproduces DFT conformational energies with sub-kcal mol −1 accuracy (RMSE 0.78 kcal mol −1 ) and correct rank ordering of all metastable conformers, outperforming all other tested MLIPs and semi-empirical methods. This continual training strategy, together with expanded non-covalent interaction coverage, has been adopted in the AIMNet2 (2025) release, which unifies small-molecule, macrocyclic, and intermolecular interaction treatment in a single production potential.
Related
- DEL-Flow: AI-Driven Structural Generation and Activity Validation of MDM2-Targeting Macrocyclic Peptides
- Integration of Acoustic Ejection Mass Spectrometry Enables Efficient High Throughput Purification Workflow for Peptides in Discovery Space
- Tokenising quantum data for label-efficient training of molecular models
- Safety Reassured: Macrocyclic GBCA Shows No Increased Risk of Parkinsonism.
- Mechanical macrocyclic effects: stimuli-responsive dual emission by self-inclusion and mutual exclusion
BibTeX
@misc{gao2026extending,
title = {{Extending AIMNet2 to Macrocyclic Peptides Through Data-Efficient Continual Training}},
author = {Runtian Gao and Roman Zubatyuk and Olexandr Isayev},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15002334/v2},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15002334/v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - Extending AIMNet2 to Macrocyclic Peptides Through Data-Efficient Continual Training
AU - Runtian Gao
AU - Roman Zubatyuk
AU - Olexandr Isayev
PY - 2026
DA - 2026/09/10
PB - ChemRxiv
DO - 10.26434/chemrxiv.15002334/v2
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15002334/v2
AB - Macrocyclic peptides are among the most promising therapeutic modalities, yet no general-purpose machine learned interatomic potential (MLIP) has been validated for their conformational analysis. We show that AIMNet2 matches the accuracy of MACE-OFF23(S) on the MPCONF196 macrocyclic benchmark, despite macrocycles being entirely absent from its training data, at three-to sixfold lower computational cost. To close the remaining accuracy gap, we introduce a continual training protocol that integrates fewer than 25,000 domain-specific conformers (<0.1% of the training corpus) while preserving accuracy on non-macrocyclic molecules, reducing macrocyclic conformational energy RMSE by 25–27%. Applied to cyclosporine A as an independent test case, the continually trained model reproduces DFT conformational energies with sub-kcal mol −1 accuracy (RMSE 0.78 kcal mol −1 ) and correct rank ordering of all metastable conformers, outperforming all other tested MLIPs and semi-empirical methods. This continual training strategy, together with expanded non-covalent interaction coverage, has been adopted in the AIMNet2 (2025) release, which unifies small-molecule, macrocyclic, and intermolecular interaction treatment in a single production potential.
ER -