Explainable AI: Designing better molecules to become drugs.

With explainable AI in the form of symbolic regression, we can get an improved understanding of what drives drug properties in conjunction with accurate drug property predictions. Here we use the QLattice to generate siRNA activity models from publicly available data to create insights that can be used to design active siRNAs. Importantly, the workflow presented here can also be used to predict any desired outcome such as toxicity, duration or even be used for disease understanding and target identification.

1. We generate Model A: a highly accurate activity model for unmodified siRNAs.
   • The model reveals that there are specific dimers at the first and last positions of the duplex that are more or less beneficial.­
   • However, an siRNA with unfavourable dimers at named positions can still be active with the right target binding energy (whole ΔG).
   • Nevertheless, an siRNA with the worst possible first and last dimers cannot be saved by an optimal binding energy.
2. Interestingly, fully chemically modified siRNAs do not fit model A indicating that other features drive the activity for these.
3. Model B is therefore generated using the same input features as used for Model A but trained on data from chemically modified siRNAs. It is showing improved predictability for modified siRNAs.
4. In general Model A has a poor performance on modified siRNAs and Model B has a poor performance on unmodified siRNAs suggesting different main drivers of activity for these two classes of siRNAs.