Deep learning models for biological prediction often assume that structured prior knowledge (PK) improves performance. We systematically tested this assumption using the Graph-Enhanced Gene Activation and Repression Simulator (GEARS), a million-parameter hybrid model designed to predict cellular responses at the transcriptomic level to combinatorial genetic perturbations. By comparing curated PK networks (DoRothEA, CollecTRI), Gene Ontology-based graphs and random graphs, we found that structured PK did not consistently improve performance - random graphs performed just as well. This revealed a major limitation: GEARS does not effectively use biological priors, but only learns from training data, limiting its ability to generalise beyond its original dataset. To rigorously validate these findings, we conducted extensive ablation studies, retraining GEARS under different knowledge graph configurations, hyperparameter settings, and architectural modifications. This required a high-performance computing infrastructure capable of handling the immense computational demands of large-scale deep learning retraining. The de.NBI Cloud provided a customised and dedicated virtual machine with 2 RTX 6000 GPUs, 256GB RAM, 16 VCPUs, and 400GB of disk space, as well as a 1TB volume for storing trained models, allowing us to efficiently test multiple configurations, manage large datasets, and automate model retraining workflows. The scalability and flexibility of this infrastructure allowed us to systematically evaluate inductive biases by enabling the training of multiple configurations in parallel and seamlessly connecting to remote development tools for faster inference, plotting and prototyping. Our results highlight the need for more rigorous evaluations of hybrid models, as deep learning systems often appear to leverage structured knowledge, but may in fact rely on biases in the model architecture. Without the powerful infrastructure of de.NBI, such a systematic and large-scale investigation would not have been possible.
