#science fair#project#programming#ai
0
LeishNN: AI based Drug Discovery for Leishmania
2021-04-21 · 2 min readUsing neural networks to develop novel molecules to target Leishmaniasis. A documentation of my science fair project!
Yay! The Contra Costa Science Fair just ended, and I ended up getting 3rd place! I had a ton of fun on my project, and I learned a lot on neural networks, especially generative semi-supervised architectures. Check out the video, or the explanation below to learn more about it!A little bit about my project is that, Leishmaniasis is a neglected tropical parasitic disease that kills millions of people in poor underdeveloped countries. Those that suffer from this disease cannot afford expensive medicines as the R&D cost is too high leading to limited commercial development. The goal of this project is to utilize AI to build an open-source neural network to generate novel molecules for leishmania. Utilizing AI for drug discovery would decrease R&D costs and being open source will allow development or other neglected diseases.The target chosen was Methionyl-tRNA synthetase, which plays a key role in protein synthesis in the parasite. This target also has a lot of prior structure activity data, which allows for training the model.The network used was adapted from GENTRL (Generative Tensorial Reinforcement Learning), outlined in Dr. Zhavoronkov's paper, where they used GENTRL to successfully create a novel molecule for DDR1 Kinase. The model uses a variational auto-encoder to determine features in the data, and then uses reinforcement learning to tune the decoder based on pre-trained self-organizing maps. The rewards I used were log P, synthetic accessibility, and log IC50 values. The datasets used were MOSES leads, general ligase, and specific target data.After training, the model was able to generate promising novel molecules. The final structures had a maximum potency of 93%, and a 2.8 accessibility score, implying significant synthesizability for a bioactive molecule. I showed the final output to an expert discovery chemist, who said the molecules looked feasible to create, and were sufficiently complex to warrant bioactivity.Woot woot! You can check out my code at the GitHub repository, or download a pdf of my board! I hope to either build on this with testing, or make something even cooler next year!!!Thanks for reading! Liked the story? Click the heart
Created with ☕ by @neelr