Bayesian regression model combining docking and deep generative networks published in JCAMD

We've published new research in the Journal of Computer-Aided Molecular Design that could speed up the hunt for new medicines. Our team developed a hybrid AI model that's better at designing potential drug molecules than traditional computational methods.
Here's how it works: we combined two powerful AI approaches - one that generates new molecular structures and another that predicts how well those molecules will bind to their target protein. By linking these together with reinforcement learning, our system can efficiently explore vast chemical possibilities and focus on the most promising candidates.
When we tested this approach on DDR1 kinase (a protein involved in cancer), the results were impressive. Our hybrid model found molecules with much higher binding scores than conventional similarity-based methods. Even better, it discovered chemically diverse compounds that looked quite different from known inhibitors but still bound effectively to the target.
What makes this especially exciting is that our approach reduces one of the biggest limitations in AI drug discovery - the need for massive datasets. Instead of just making slight tweaks to existing drugs, our system can venture into unexplored chemical territory to find genuinely novel compounds.
While there's still more work ahead, this hybrid strategy could help researchers identify promising new treatments for diseases where current drugs aren't working well enough.

Full Text