Predicting the Host-guest Binding Gibbs Free Energy for Guests
Host-guest chemistry involves noncovalent interactions between host and guest molecules and has been applied in areas such as adaptive materials, sensing, and drug delivery. However, the discovery of new hosts and the identification of suitable guests remain challenging because selective binding is governed by subtle and often competing noncovalent interactions that are difficult to predict. To overcome this predictive limitation, computational framework was developed to predict Gibbs free energies of host-guest binding using a hybrid machine-learning model. The model was trained and tested on 4155 host-guest systems manually curated from the literature and combines descriptors derived from semiempirical quantum-mechanical calculations with three-dimensional molecular fingerprints. Evaluation on a test set yielded an R2 of 0.83 and a mean absolute error of 3.1 kJ per mol. The computational efficiency of the model enabled prediction times of approximately one to five minutes per host-guest system with an average size of 174 atoms, supporting the framework’s applicability to high-throughput screening for host-guest discovery.
Authors
- Mario Öeren (ORCID: https://orcid.org/0000-0003-4292-5557)
- Riina Aav (ORCID: https://orcid.org/0000-0001-6571-7596)
- Rauno Reitalu (ORCID: https://orcid.org/0009-0006-4635-034X)
Institutions
- Tallinn University of Technology (EE)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-17
- DOI
- https://doi.org/10.26434/chemrxiv.15012395/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint