Artificial intelligence (AI) has been increasingly used in drug discovery research. Researchers are using machine learning applications like graph neural networks (GNNs) to predict how strongly a certain molecule binds to a target protein. However, a study by Prof. Dr. Jürgen Bajorath and his team found that GNNs largely remember known data and do not learn specific chemical interactions when predicting drug potency.
The scientists analyzed six different GNN architectures using a method called “EdgeSHAPer” to understand how the GNNs generated predictions. They found that most GNNs mainly remembered chemically similar molecules encountered during training and their binding data, rather than focusing on learning protein-ligand interactions.
This could mean that the predictions made by GNNs are overrated, as they can be made using simpler methods. However, the research also identified GNN models that showed potential for further improvement in learning interactions. The study reveals that AI is not magical and offers opportunities for improvement in future drug discovery research.
Prof. Bajorath believes that understanding how machine learning models arrive at their results is important and his team’s approach focuses on developing methods for explaining predictions of complex models. The study has been published in Nature Machine Intelligence, shedding more light on the black box of AI models.
>>Join our Facebook Group be part of community. <<