Can classical statistics and deep learning converge on explainable, causally driven target discovery?

Understanding the molecular causes of complex diseases remains one of the most pressing challenges in biomedicine. Despite large-scale genome-wide association studies mapping thousands of risk loci, identifying which genetic variants truly drive disease remains difficult. Traditional statistical genetics has laid a strong foundation for variant discovery, but it often struggles to capture nonlinear interactions and cannot fully integrate the breadth of the interconnected multi-omics data. In recent years, deep learning approaches have shown promise in bridging these gaps: modelling high-order genetic interactions, uncovering latent biological structure, and enabling multi-layered data integration. However, most current deep learning models for genomics remain exploratory in nature, and issues such as susceptibility to overfitting, difficulties in interpretability, and the general lack of standardized evaluation frameworks have limited their widespread adoption for genomics research. In this review, we explore how traditional statistical and deep learning methods can be applied to uncover causal mechanisms in complex disease. We critically compare these two frameworks for their advantages and limitations in detecting genetic associations and prioritizing causal associations. Towards the end, we propose a future direction centred around hybrid models that blend the scalability of deep learning with the inferential power of statistical genetics. Our goal is to guide researchers in developing next-generation computational tools to uncover the molecular basis of complex diseases and accelerate the translation of genetic findings into effective treatments.