Escaping the forest: a sparse, interpretable, and foundational neural network alternative for tabular data

Tabular datasets are pervasive across biomedical research, powering applications from genomics to clinical prediction. Despite recent advances in neural architectures for tabular learning, there remains no consensus on models that balance performance, interpretability, and efficiency. Here, we introduce sTabNet, a meta-generative framework that automatically constructs sparse, interpretable neural architectures tailored to tabular data. The model integrates two key components. First, automated architecture generation leverages unsupervised, feature-centric Node2Vec random walks to define network connectivity, introducing a priori sparsity and improving generalisation while mitigating overfitting. Second, a dedicated attention layer jointly learns feature importance with model parameters during training, providing intrinsic interpretability. Evaluated across diverse biomedical tasks-including RNA-Seq classification, single-cell profiling, and survival prediction, sTabNet achieves performance on par with, or exceeding, leading tree-based models such as XGBoost, while remaining computationally efficient and CPU-trainable. Our experiments show that sTabNet generalises effectively across in-domain and out-of-domain datasets, yielding biologically consistent insights and surpassing post-hoc explainability methods such as SHAP in stability and clarity. Together, these results establish sTabNet as a foundational and versatile framework for data-efficient, interpretable neural learning on tabular data.