NeuroMedix leverages advanced machine learning to accelerate drug design. Most drugs are derived from natural compounds and then optimized, rather than being synthesized from scratch in a lab. To support this approach, we use and refine the NP-VAE (Natural Product-oriented Variational Autoencoder) model developed by Ochiai et al. 2023, designed specifically to handle complex, biologically active natural compounds. Trained on 30,000 compounds from DrugBank and a dataset of natural compounds, NP-VAE overcomes limitations that traditional VAE models face in processing natural product data.

Our project focuses on Alzheimer’s Disease (AD), a neurodegenerative disorder affecting memory, cognition, and behavior, and the 7th leading cause of death worldwide. NeuroMedix refines the NP-VAE model and adapts it for FDA-approved drugs targeting AD, facilitating the discovery of new, synthesizable drug candidates.

The NP-VAE model operates in two stages:

1. Encoding: A molecule in SMILES format of a known drug compound is compressed into a latent representation.
2. Decoding: This latent representation is used to reconstruct or generate new molecular structures in SMILES format, enabling the discovery of potential drug analogues.

NeuroMedix introduces an interactive platform where scientists and pharmaceutical companies can enter an FDA-approved drug into a search bar and receive suggestions for novel, synthesizable drug candidates with similar structures. This scalable approach can be expanded to target other diseases, offering small pharmaceutical companies a streamlined path to developing new treatments.

The repository includes the following folders:

• model_v1: Custom-built VAE model for generating SMILES strings, designed to aid in the drug discovery process.
• model_v2: NP-VAE model, as developed by Ochiai et al. (2023), adapted for generating analogues of FDA-approved Alzheimer’s disease drugs.
• client: Website allowing users to input an FDA-approved drug and retrieve top 3 suggested analogues with structurally similar features, facilitating quick exploration of potential drug candidates.

Ochiai, T., Inukai, T., Akiyama, M. et al. "Variational autoencoder-based chemical latent space for large molecular structures with 3D complexity." Communications Chemistry, 6, 249 (2023). https://doi.org/10.1038/s42004-023-01054-6