This ongoing project was spearheaded as part of a research visit to the Translational Immunology Lab headed by A/Prof Vivek Naranbhai at Monash University, Melbourne in early 2026. Now, as part of the THETA-TEAM, we continue our collaboration to develop and implement computational methods to design novel drug candidates to treat cancer.

Ever since the release of AlphaFold2, the field of structural biology has changed dramatically. There has been a boom in the release of new and improved structure prediction models that allow users to solve the structure of a large proportion of previously unsolved proteins. These models can even be used to “hallucinate” new proteins, and research has pushed for the development of methods for conditional generation of specific proteins or small molecules that serve a specific function.

However, most of such de novo protein design pipelines are either closed-source or closed-scope: they are frequently developed to excel within a narrow set of boundaries, are difficult to extend, or are computationally prohibitive to deploy at scale.

In this project, we set out to design and implement a high-level programming language for computational drug design. This pipeline is based on PyTorch / JAX and implements target and/or binder requirements and constraints as re-usable loss terms that can be combined to optimise for drug candidates that fulfil different types of purposes. Another benefit is that it is simple to extend these losses to multiple targets, enabling the user to incorporate both specificity and generalised binding as learnable properties. For optimisation, we have access to several high performance computing clusters, including M3 Massive (Monash University cluster) and NCI Gadi, the Australian national computing cluster. Here, we can run the model in parallel on multiple GPUs by calculating gradient for multiple diffusion sample on a dedicated GPUs.

We have used this pipeline to generate different binders that for example recognise and bind to conserved sites amongst multiple targets and even to specifically recognise and bind to phosphorylated sites or even single point mutations. In the lab, we compared our method to a previous SOTA approach (BindCraft) where it would take 2-3 weeks to design, inspect, and digitally evaluate 100 candidates to select for binding specificity given a positive and negative target. In contrast, our pipeline was able to directly optimise for specificity and deliver 100 candidates within 48 hours.

I am currently working on designing novel protein-based binders to detect specific mutated forms of well-known oncoproteins. I am also working on enhancing the pipeline to improve the performance of our antibody and antibody-binder design pipelines. In these projects, I work closely with the bio-medical staff at the Translational Immunology lab and am able to meaningfully contribute to discussion regarding experimental setup and evaluation given my history as a biomedical scientist. I am an active part of the THETA-TEAM, an international team encompassing a diverse range of experts including AI scientists, structural biologists, immunologists, and medical doctors, to bring together strong and varied backgrounds to tackle complex challenges in the intersection between AI and health.