Inspiration
In analyzing public health data, we realized that income would be a serious bottleneck for people's health, even when policy is done correctly. Therefore, in this project we tried to serve low-income individuals by making the process of producing medicines less intensive, and therefore cheaper.
What it does
First, we created a GNN, to predict public health scores based on data from LiveData Technologies, and with the help of RapidFire AI's visualization tools. Next, we made a GNN to predict chemical reactants from products, which will simplify research and development. By making the pathway to synthesize a drug optimal, we can ensure that it is mass produced as quickly as possible.
How we built it
We used PyTorch, specifically PyGeometric tensor objects, to encode data about public health by district. Our nodes were the districts and their independent evaluations, while our edges were the relations between districts. This allowed us to make intelligent predictions about how changes in one district could affect another district.
Additionally, we did a very similar thing for molecules, but this time using PyGeometric tensor objects to encode molecule structure. Then, the GNN would predict which reactants could synthesize that product.
Challenges we ran into
Some datasets for molecule analysis were too extensive, with more than 40,000 chemical reactions. We don't have the computing power to process all of the data in those cases, so we had to resort to random sampling on smaller numbers of reactions.
Accomplishments that we're proud of
We were able to reach roughly 50% accuracy on reactant predictions given a certain product. This is especially impressive when the products are huge, and the way the product can be broken down or synthesized is highly unclear.
What's next for GNNs for Public Health Solutions
In the future, we want to expand our models by training them on more data, and pairing them with other models like Google's AlphaFold. This will allow us to understand large scale molecule - molecule interactions, and to build the future of medicine.
Log in or sign up for Devpost to join the conversation.