Bayesian neural networks for molecular discovery

Navigating the Maze of Molecular Discovery with Smarter, Uncertainty-Aware Graph Neural Networks:

Developing new molecules is similar to navigating a complex multistory labyrinth - it is time consuming, and is filled with a lot of trial and error.
When developing a new drug, the goal is to find the right balance between the properties of the molecule, such as its effectiveness and solubility.
Instead of relying solely on traditional engineering methods, scientists are now utilizing Graph Neural Networks (GNNs), which offer innovative ways to optimize these molecular properties by efficiently processing complex structural data.
By treating the molecules as mathematical graphs, these networks have substantially improved our ability of predict the properties of novel molecules.

However, there's a catch: GNNs can sometimes be "overconfident" in their predictions, unaware of what they don't know. It's like a GPS system that guides you confidently into a dead-end. Current approaches to add a layer of uncertainty quantification come with a high price. Either they take a hit to their predictive performance, or the computation becomes prohibitively expensive.

In this project, we aim to develop new methods for uncertainty-aware GNNs that don't compromise on performance or feasibility. These models can serve as 'oracles' for scientists, guiding them towards promising molecules for drug discovery. Moreover, they can flag areas where their predictions are uncertain, inviting further experimental verification. In essence, these new models create a feedback loop: research informs the model, and the model, in turn, informs research - making the whole process of drug discovery more efficient and reliable.