Malte Holmer

Doctoral student
Address
Office
Contact
Research Project
Explainability in AI-driven early-phase drug discovery with Network Balance Scaling
There are a variety of methods that support the medicinal chemist in the Design-Make-Test-Analyze (DMTA) cycle and provide an answer to the question, "What to make next?". In recent years, machine learning models (ML models) became state of the art for this task, as they make precise QSAR/QSPR predictions and therefore can be of great use. However, as the models improved, their complexity also increased, while their explainability decreased. For this problem, there are already approaches in the literature, with the most prominent method probably being Shapley Additive Explanations (SHAP).
In recent years, we have developed a method in our research group that offers an alternative to these classical explanation approaches while simultaneously improving the QSAR/QSPR predictions of any ML models. Network Balance Scaling (NBS) forms a network of molecules using Matched Molecular Pair Analysis (MMPA), where each node represents a molecule and the edges between them represent an MMP relationship. Properties are assigned to the nodes based on model predictions or empirical measurements, while the edges are characterized by the average property difference of the MMP transformation. Visualizing these networks allows the chemist to relate predictions of structural analogs to their measured counterparts based on MMP statistics between them and the measured value. Through an optimization algorithm, NBS balances property predictions with MMP statistics and empirical measurements, which results in correction terms for QSAR/QSPR predictions of any model. We are currently working on NBS 2, which stands out compared to its predecessor, mainly due to better usability and shorter run times. NBS 2 allows for iterative extension of already existing networks and therefore allows the medicinal chemist to use the tool exploratively.