Geometric Generative Modelling for Protein Design and Synthetic Biology

Generative modelling techniques like diffusion and flow matching have revolutionised fields like image generation and are about to do the same in the scientific domain, particularly in synthetic biology where they are used to generate novel molecules useful as therapeutics, catalysts and many more applications. However, the success rates of these models remain very low, and their controllability remains weak, reducing their practical impact.

This project aims to create powerful and practically useful generative models of biomolecules for applications in protein design and beyond. The main question this project intends to answer is

What is the most efficient way to model and design a biomolecule computationally so that it has a high chance of fulfilling the design constraints in the real world?

Efficiency here involves both computational efficiency in terms of compute resources such as time and memory, but also experimental efficiency in terms of the success rate these designed molecules show once they are manufactured and tested in the laboratory. Major themes that will be explored are

1. Learning/modelling molecular representations that are both computationally efficient as well as performant in the task at hand; this will involve both

representation in terms of granularity (do we model a molecule atom-by-atom or in a coarse-grained fashion?) as well as in terms of geometrical invariances (do we learn necessary equivariances or do we enforce them by architecture constraints?)

2. Improving fundamental neural network architectures and training procedures for generative modelling on biomolecules. Many current approaches are directly transferred from image generation, and initial investigations show that these are often suboptimal.

3. Enhancing controllability and raw distribution modelling capacity of generative models. In generative modelling for other modalities like images, metadata like

image captions are used to subdivide the hard general distribution modelling problem into many smaller ones. Doing something similar in the domain of

biomolecules promises to improve performance and allow practitioners to use these models a lot more effectively and tuned towards their needs.

4. Including experimental constraints: in many cases, downstream laboratory experiments have practical constraints that are important to consider in model

design, whether that is as part of which data is available as input to the model or what are the experimental limitations in terms of synthesizing the predicted

molecules. Current models often fail to be integrated into a rapid feedback loop by neglecting these considerations, including them into model design promises to make these models more impactful in practice.

5. Improving benchmarking and evaluation of models: Since the use of generative models in biology started only recently, it is not fully clear how to evaluate in

practice which model is the best, especially given the limited experimental evaluation that is performed (see point above). Therefore, one focus of this thesis

will be developing better in-silico benchmarking metrics and tools to guide model development in a more streamlined way.

This project falls within the EPSRC "healthcare technologies" research area and will be carried out with collaborators in Oxford (such as Professor Charlotte Deane) as well as collaborators abroad.