My father is a flavorist, so I grew up in Iran, where my parents have a factory to produce flavor and fragrances. My final project is something that it may not leverage from Ginkgo platform, so I decided to propose a project based on my family background.
The goal of this project is to create new synthetic fragrances. The goal to create the chemical and DNA sequence of a library for existing commercial fragrances, and based on those library use machine learning to generate new fragrances.- Use Mass Spectrometry and gas chromatography to create the library of existing chemical compound analysis ( this could be an open source library). There already existing databases for FlavorDB a database of flavor molecules
- Based on existing chemical compounds, look for enzymes that could produce analysed chemical compounds.
- Look for gene sequence that could create those enzymes.
- Use Machine Learning methods ( such as TSNE, VAE, PCR, UMAP) based on both sequencing and chemical compound library for clustering and dimensionality reductions
- Based on hidden layers, estimate what consider as “desirable” enzyme
- Use VAE ( or recent attention based generative models) to create new sequences that could map for enzymes that produce novel flavors.
- Use Ginkgo platform to engineer novel flavor in yeast by fermentation.
- Find a parameter and tune the model to create a “desirable smell”.
There are many methods for clustering and dimensionality reduction algorithms such as PCA, UMAP, TSnE or even VAEs. Although, there are many recent generative (attention-based) models, VAE could work and it has already been used for music generation at Google ( Google Magenta) project
