# CHA2: CHemistry Aware Convex Hull Autoencoder Towards Inverse Molecular Design

**Authors:** Mohammad Sajjad Ghaemi, Hang Hu, Anguang Hu, Hsu Kiang Ooi  
**arXiv:** [2302.11000](https://arxiv.org/abs/2302.11000)  
**Date:** 2023-02-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.11000v1/

## Key points

- The paper addresses the challenge of optimizing molecular design and discovering novel chemical structures
- Autoencoders are used to reduce the search space into a lower-dimensional latent volume
- The continuous property of the latent space enables flexible representation for inverse design and discovery of novel molecules
- Using a convex hull surrounding top molecules with high QEDs helps ensnare a tight subspace in the latent representation
- The method is validated using QM9 dataset and SELFIES representation to calibrate the autoencoder
- The suggested method effectively carries out inverse molecular design and leads to unfolding novel chemical structures
- Potential applications include drug discovery and material science.

## Summary

The paper titled "CHA2: CHemistry Aware Convex Hull Autoencoder Towards Inverse Molecular Design" addresses the challenge of optimizing molecular design and discovering novel chemical structures. This task is difficult due to the vast combinatorial design space of discrete molecular structures, making it nearly impossible to explore comprehensively. To overcome this challenge, the authors propose reducing the search space using autoencoders into a lower-dimensional latent volume. Autoencoders are deep learning techniques that consist of an encoder and a decoder. The encoder reduces the discrete molecular structure into a latent space, while the decoder inverts it back to the molecular design. The continuous property of the latent space provides a flexible representation for inverse design and enables discovery of novel molecules. However, exploring this latent space requires insights to generate new structures. To address this issue, the authors suggest using a convex hull surrounding top molecules with high quantitative estimates of drug-likeness scores (QEDs). This approach helps ensnare a tight subspace in the latent representation and allows for efficient discovery of novel molecules with high QEDs. To validate their method, they use QM9 dataset as training data along with Self-Referencing Embedded Strings (SELFIES) representation to calibrate the autoencoder. They demonstrate that their suggested method effectively carries out inverse molecular design and leads to unfolding novel chemical structures. Overall, this paper presents an innovative approach using autoencoders and convex hulls to optimize molecular design and discover novel chemical structures with potential applications in drug discovery and material science.