Miriam Babukhian

Biologics such as antibodies and other engineered proteins are revolutionising how we treat cancer and are other serious diseases. But they still face a major challenge: sometimes these therapies attach to healthy tissues, causing harmful side effects and slowing progress. Current ways of detecting this risk rely on long, costly lab studies and animal testing, which often don’t fully reflect human biology and raise ethical concerns.

New advances in artificial intelligence (AI), such as transformers that power ChatGPT, now make it possible to model biology faster and more accurately. Therefore, this project explores whether AI, streamlined into a single pipeline, can make off-target screening of biologics both faster and more comprehensive by comparing a therapy’s intended protein target against the entire set of human proteins.

Our approach combines three steps: generating virtual variants of the target, scanning them for similarities across all human proteins, and using 3D models to predict when a therapy might bind to the wrong target. By integrating explainable AI, we can also highlight which regions of a protein are most likely to drive these unwanted effects.

The result will be a ranked safety profile for each new therapy, giving researchers clearer guidance before lab testing. This could cut costs, reduce animal use, and deliver safer treatments to patients more quickly.

‍