50,000× faster molecular dynamics at the NVIDIA Open Models AI Codefest
At the Bitdeer AI & NVIDIA Codefest, the LynxKite team fine-tuned a physically grounded generative model on 1,000 ns molecular dynamics trajectories — targeting a 10,000× speed-up over classical MD and landing at 50,000×.
The challenge
The Open Models AI Codefest brought together teams exploring how open and fine-tuned AI models can be applied to real scientific and technical problems. Team LynxKite MD — Daniel Darabos, Lívia Babos and Derek Smith — took on a drug-discovery problem: using machine learning to make molecular dynamics insights faster, more accessible and more useful for prioritising candidate molecules.
Getting more out of every simulation
Molecular dynamics simulations produce huge amounts of trajectory data, yet much of it goes underused. The team set out to test whether AI can exploit those trajectories more fully, learning receptor-specific motions that matter for ligand binding, efficacy and drug discovery decisions.
Why open models matter
The range of entries showed how far open models reach: they are practical tools that can be adapted across domains, from financial services to life sciences. The team also drew on the NVIDIA ecosystem, including the BioNeMo suite, which brings modern AI methods to molecular modelling, protein science and drug discovery.
For LynxKite, the Codefest was a chance to test a focused idea, learn from other builders, and contribute to open-source workflows that could eventually make computational drug discovery faster, cheaper and more collaborative.
Thanks to Dr. Deb Goswami for organising the event.