Biologically informed neural network models are robust to spurious interactions via self-pruning

Abstract

Computational models of cellular networks hold promise to uncover disease mechanisms and guide therapeutic strategies. Biology-informed neural networks (BINNs) is an emerging approach to create such models by combining the predictive power of deep learning with prior knowledge, a vital aspect of biological research. The architectures of BINNs enforces a network structure from which mechanism can ideally be inferred. However, a key challenge is to evaluate the reliability of these models, as cells are inherently complex, involving intricate and sometimes unknown interactions. Currently, analysis mainly focuses on selected pathways rather than a more comprehensive perspective. In this work we demonstrate an alternative holistic approach: we measure to which extent purposefully introduced spurious interactions are down-weighted by a BINN during training (self-pruning). The metric suggested RRA (Relative Residual Area) allows for direct distribution comparison with perfect self-pruning achieved at zero and a failure to self-prune if above one. To enable rapid testing, we updated LEMBAS (Large-scale knowledge-EMBedded Artificial Signaling-networks), our recurrent neural network framework for intracellular signaling dynamics, with full GPU acceleration. Our implementation achieves a >7-fold speedup compared to the original while preserving predictive accuracy. We evaluated self-pruning in 3 different datasets and found that when spurious interactions are introduced at random, the model prunes these to a larger extent than those from the prior knowledge network (PKN), provided the model is regularized with a sufficiently large L2 norm. This suggests that BINNs can be robust to uncertainty in the PKN. Implementation and applicationOur implementation of LEMBAS is freely available under a MIT license at https://github.com/AvlantNilssonLab/LEMBAS_GPU. The models and results to generate the figures can be downloaded through https://zenodo.org/records/17425598.