In this episode, the fifth in our miniseries covering the mAbs journal article collection on artificial intelligence and machine learning in antibody development, we speak to Aubin Ramon, Postdoctoral Research Associate in the Sormanni Lab at Imperial College London (UK), about his paper in the collection: Prediction of protein biophysical traits from limited data: a case study on nanobody thermostability through NanoMelt.
Aubin addresses one of the most persistent challenges in applying AI to specialized fields like antibody therapeutics: the scarcity of high-quality training data. Through the development of NanoMelt, he demonstrates how a combination of sophisticated modeling approaches, general protein stability prediction models, and robust validation pipelines can achieve meaningful predictions with training datasets of just 600-700 sequences.
Together, we explore why data quality often matters more than quantity, the counterintuitive finding that general protein models can outperform antibody-specific ones, and how NanoMelt is already being applied in synthetic library design and therapeutic development, with exciting improvements on the horizon in NanoMelt 2.
[01:40] The data availability problem in AI and a novel solution for it
[06:05] Achieving accuracy with small data sets
[09:05] Is this model accurate enough to be truly useful?
[12:50] General vs highly specific models for themostabilty prediction
[15:40] The limits of just generating more data
[19:20] Adapting the model for more complex/full length antibodies
[21:05] Advice for using NanoMelt for your own work
[24:15] Applications of NanoMelt in current research and announcing NanoMelt2