Electron Density-Based Host-Guest Binder Optimization

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Mar 20 2024

In an article published in the journal Nature, researchers introduced a machine learning model trained on electron density for generating host-guest binders, outputted as a simplified molecular-input line-entry system (SMILES) format with over 98% accuracy.

They utilized a variational autoencoder to produce three-dimensional (3D) representations of electron density and electrostatic potentials, then optimized guest generation via gradient descent. The model's practical application yielded discoveries of previously validated and unreported guests for molecular host systems.

Background

Host-guest chemistry, involving molecules that form complexes like a lock and key, has diverse applications from catalysis to materials science. However, discovering new guest molecules for existing hosts or optimizing new host-guest systems is laborious. Prior methods relied on costly experimental iterations. Traditional digital representations of molecules, like SMILES strings, lacked 3D information essential for understanding host-guest interactions.

Although 3D volumes of molecules have been used in predicting properties, correlating them with clear molecular structures remained challenging. This study bridged this gap by leveraging electron density representations of host molecules decorated with electrostatic potential. By training a transformer model, the researchers efficiently converted 3D volume molecular descriptors into SMILES representations, enabling the generation of defined molecular structures usable by chemists.

They showed that optimizing volumetric shape and charge interactions between 3D descriptors led to the discovery of novel guest molecules for well-known host systems, cucurbit[n]urils, and metal–organic cages. This approach circumvented the need for prior knowledge of the host-guest system beyond the host's chemical structure.

Additionally, the authors highlighted the potential of transformer models and electron density representations in streamlining the discovery process. Thus, the researchers introduced a novel computational framework for accelerating the discovery of host-guest systems, overcoming the limitations of previous methods, and opening avenues for efficient molecular design and discovery.

Methods

The researchers utilized Python 3.9 with TensorFlow (TF) for machine learning models, mainly using TF 2.7 and later updated to version 2.10. The publicly available quantum machines nine (QM9) dataset was employed, comprising 133,885 molecules, focusing on their SMILES representations and XYZ coordinates. A script was developed to download the dataset, generate electron densities, and calculate electrostatic potentials for each molecule, saving them into TF record files.

A transformer model was implemented to convert electron densities into SMILES representations. Two strategies were tested to transform 3D data into two-dimensional (2D) matrices. Fitness functions were employed for optimization, aiming to maximize molecule size, minimize overlapping electron densities, and maximize interactions between host and guest electrostatic potentials using gradient descent.

The quality of generated SMILES libraries was benchmarked, indicating high internal diversity and novelty compared to the training set. Cucurbituril CB[6] and metal–organic cage [Pd₂1₄](BArF)₄ were used for experimental validation. Hydrogen (¹H) nuclear magnetic resonance (NMR) titration determined the association constant between hosts and various guests, with fitting models applied to the data.

• Cucurbituril CB[6] guest binding titrations: ¹H NMR titration was conducted in deuterium oxide (D₂O)/formic acid-d₂, with CB[6] and guest amines. Signals indicated fast exchange on the NMR timescale, with peak positions plotted against CB[6] concentration and fitted to a 1:1 binding model.
• Cage [Pd₂1₄](BArF)₄ guest binding titrations: ¹H NMR titration occurred in CD₂Cl₂, maintaining cage concentration constant. Signals also showed fast exchange, with peak positions plotted against guest concentration and fitted to a 1:1 binding model.

Overall, the methods encompassed data preparation, model implementation, optimization, benchmarking, and experimental validation, providing a comprehensive framework for computational-driven discovery in host-guest systems.

Results

The research developed a two-tier workflow for the computer-aided discovery of experimentally validated guests for cucurbituril CB[6] and metal–organic cage [Pd₂1₄]⁴⁺. Initially, an in silico workflow generated virtual libraries of potential guest molecules, utilizing a variational autoencoder (VAE) and a transformer model to translate 3D electron density volumes into SMILES representations. Then, an in vitro workflow experimentally tested the most promising candidates. The VAE, fully convolutional neural (FCN) network, and transformer model enabled the generation of guest molecules based on the electron data of a target host, approached as an optimization problem.

Gradient descent was utilized to find guests by maximizing molecular size, minimizing overlapping with the host, and maximizing electrostatic interactions. The algorithm successfully identified known guests for CB[6] and produced new candidates for both hosts. Experimental testing confirmed the binding affinities of generated guests for CB[6] and revealed promising but commercially limited options for [Pd₂1₄]⁴⁺.

Despite limitations regarding molecule size constraints inherent in the training data, the algorithm generated guests with suitable structural features for binding with the cage. Overall, the study demonstrated the efficacy of the computational approach in discovering and optimizing guest molecules for specific host molecules, laying the groundwork for further applications in host–guest chemistry.

Discussion

The study explored the use of self-referencing embedded strings (SELFIES) as an alternative molecular representation but found no improvement over SMILES notation. While the QM9 dataset was suitable for smaller host molecules like CB[6], the larger cavity of [Pd₂1₄]⁴⁺ required bigger guest molecules. To address this, a function was implemented to increase molecule size, but future research could aim to utilize datasets with larger molecules, such as GDB-17.

Additionally, the authors envisioned integrating ligand selection into the generative process and automating molecule synthesis using platforms like Chemputer robots. This approach aimed to create a closed-loop system, seamlessly connecting optimization and experimental testing in host-guest chemistry.

Conclusion

In conclusion, the authors introduced a computational framework using machine learning and optimization to discover and optimize guest molecules for host-guest systems. By leveraging electron density representations and transformer models, the approach enabled an efficient generation of guest molecules, validated experimentally for cucurbituril CB[6] and metal–organic cage [Pd₂1₄]⁴⁺. Future directions include integrating ligand selection and automating synthesis for a closed-loop discovery process.