I'm James. I work on deep learning for drug discovery at Synteny. A year ago I knew nothing about protein folds. I moved into biology because I want this work to help people in the long run, not just climb a benchmark. I recently fine-tuned Boltz-2 for binding-affinity prediction. That went to MLSB 2025. “We can only see a short distance ahead, but we can see plenty there that needs to be done.” — Turing, 1950
Before that, I was at CVSSP working with Prof. Mark Plumbley on AutoML for audio. My thesis is Towards Optimal Architectures for Audio Machine Learning, and covers neural architecture search, filter pruning, and DARTS. Partway through, I spent six months as a research intern at Meta Reality Labs, working on diffusion models for audio.
Most of my work comes back to one question. What does a neural network really need to do, and how much of it can we take away? This looks different when you're upsampling a piano recording, compressing a CNN under a FLOP budget, or fitting a structural biology model onto a single GPU. The question is the same. Further back, I read Computer Science and Maths at Cambridge (Pembroke, 2021), where my dissertation won the CST Dissertation Prize.