Traditional genomics operates on a fundamental simplification: the additive assumption. In this view, the risk of a disease or the value of a trait is the simple sum of individual genetic effects. While this linear approach has powered genome-wide association studies (GWAS) for decades, it often fails to account for "missing heritability"—the gap between observed inheritance and what our models can explain.
Researchers are increasingly moving toward deep learning to capture the non-linear, hierarchical interactions that define biological reality. By replacing linear summation with neural architectures, these models can detect epistasis (gene-gene interactions) and dominance effects that are invisible to standard statistical methods.
The Failure of the Additive Assumption
In a standard linear model, the impact of having two copies of a genetic variant is exactly double the impact of having one. However, biological systems rarely work this way. As Arnór I. Sigurdsson, Simon Rasmussen, and colleagues point out in Improved prediction of blood biomarkers using deep learning, linear models treat genetic variants as continuous variables, which prevents them from capturing recessive or dominant inheritance patterns.
More importantly, linear models evaluate variants independently. This ignores epistasis, where the effect of one gene is contingent on the presence of another. Lorin Crawford argues that this inability to model high-dimensional interactions is a primary driver of missing heritability (At the intersection of machine learning, biology, and health: an interview with Lorin Crawford). When the combined effect of two mutations is a "shout" rather than two "whispers," linear models only hear the whispers.
Capturing Interaction through Non-linearity
Deep learning models solve this by using non-linear activation functions (like ReLU or Sigmoid) across multiple layers. These layers act as hierarchical feature extractors. In the first layer, the model might identify individual SNP effects; in deeper layers, it begins to "see" how combinations of those SNPs create emergent phenotypes.
For example, Jiajie Peng and Jingyi Li developed DeepPRS, which uses bidirectional LSTM architectures to capture these dependencies (A Deep Learning-based Genome-wide Polygenic Risk Score for Common Diseases Identifies Individuals with Risk). This allows the model to identify high-risk individuals who lack common "red flag" markers but possess specific, risky interaction patterns across their genome. Similarly, Xiongbin Kang and Alexander Schönhuth used capsule networks in DiseaseCapsule to aggregate "votes" from distributed genetic features, uncovering non-linear signals in ALS that traditional GWAS missed (Predicting the prevalence of complex genetic diseases from individual genotype profiles using capsule networks).
Biologically Constrained Architectures
A common critique of deep learning in biology is the "black box" problem: a model might predict a trait accurately but offer no insight into the underlying mechanism. To solve this, researchers are building "biologically annotated" or "informed" neural networks.
In Multi-scale Inference of Genetic Trait Architecture using Biologically Annotated Neural Networks, Pinar Demetci, Gregory Darnell, and Lorin Crawford describe BANNs. These networks constrain their architecture based on known biological metadata, such as which SNPs belong to which genes and which genes belong to which pathways. This allows the model to remain interpretable—researchers can see which biological "nodes" the model is prioritizing—while still benefiting from the non-linear power of deep learning to detect epistasis.
A similar logic applies to GEARS, developed by Yusuf Roohani and Jure Leskovec. By integrating gene co-expression and Gene Ontology knowledge graphs, GEARS can predict the effects of multi-gene perturbations that were never seen in the training data (Predicting transcriptional outcomes of novel multigene perturbations with GEARS). It does this by representing perturbations as multidimensional embeddings that interact through non-linear modules, effectively simulating synergy and suppression.
From Prediction to Discovery
The shift toward non-linear modeling isn't just about better accuracy; it's about discovering new biology. Xiaopu Zhou and Yu Chen demonstrated that by "deconstructing" a deep learning model's internal risk scores, they could pinpoint specific immune-related pathways (like TNF-α signaling) as the primary drivers of Alzheimer's risk (Deep learning-based polygenic risk analysis for Alzheimer’s disease prediction).
By moving beyond the additive assumption, these researchers are beginning to map the true, interactive landscape of the genome. As Marie-Christine Fritzsche and colleagues note, this transition from simple linear summation to complex AI-driven prediction introduces new ethical and regulatory challenges, but it also offers the most promising path toward resolving the "missing heritability" of human disease (Ethical layering in AI-driven polygenic risk scores—New complexities, new challenges).