For decades, quantitative genetics has relied heavily on the "infinitesimal model"—the assumption that complex traits and disease risks are the product of thousands of individual genetic variants, each contributing a small, independent, and purely additive effect. This assumption underpins traditional genome-wide association studies (GWAS) and standard Polygenic Risk Scores (PRSs), which calculate risk by simply summing up weighted allele counts.
However, biological systems are rarely purely linear. Genes operate within intricate biochemical networks where the impact of one mutation is frequently contingent on the state of another—a phenomenon known as epistasis. As Lorin Crawford highlights in At the intersection of machine learning, biology, and health: an interview with Lorin Crawford, relying strictly on additive linear models leaves a significant portion of broad-sense heritability unaccounted for, contributing to the long-standing "missing heritability" problem.
To bridge this gap, a major research direction has emerged focused on leveraging non-linear machine learning architectures, non-parametric statistics, and advanced optimization techniques to capture epistatic interactions across the genome.
Why Linear Models Miss Epistasis
To understand why traditional statistical frameworks fail, consider the mathematical intuition behind a standard linear regression model used in a GWAS. The model evaluates each single nucleotide polymorphism (SNP) independently, looking for a marginal correlation with the phenotype.
Now, imagine a simple non-linear genetic switch: a disease phenotype only manifests if a patient inherits a specific risk allele at both Locus A and Locus B. If a patient has the risk allele at Locus A but not Locus B, or vice versa, they remain healthy.
Locus A: [0] AND Locus B: [0] ==> Healthy (0)
Locus A: [1] AND Locus B: [0] ==> Healthy (0)
Locus A: [0] AND Locus B: [1] ==> Healthy (0)
Locus A: [1] AND Locus B: [1] ==> Disease (1)
If you calculate the independent marginal effect of Locus A across a population, the signal is heavily diluted or completely washed out because its success depends entirely on the state of Locus B. As Ping Zeng et al. demonstrate in Detecting epistasis with the marginal epistasis test in genetic mapping studies of quantitative traits, variants involved in these complex interactions often exhibit negligible or zero additive signals. Consequently, standard filtering protocols drop these variants long before interaction testing occurs.
Furthermore, as noted by Hongyu Zhao and Weichuan Yu in The complete compositional epistasis detection in genome-wide association studies, biological systems can exhibit even more deceptive patterns like XOR-like relationships, where disease risk is elevated if a specific genotype is present at exactly one of two loci, but not both. Linear models are fundamentally blind to these multi-locus dependencies.
The Non-Linear Toolkit
To capture these non-additive dynamics, researchers have introduced several distinct computational paradigms that can map complex genotype-to-phenotype landscapes.
1. Deep Learning and Compositional Embeddings
Deep neural networks are naturally suited for this problem because their multi-layer architectures inherently learn hierarchical feature combinations. Instead of treating SNPs as isolated scalar inputs, modern deep learning frameworks process them through non-linear activation functions and multi-layer feed-forward networks to capture high-dimensional dependencies.
For instance, Xiaopu Zhou et al. developed a deep learning framework for Alzheimer's disease prediction in Deep learning-based polygenic risk analysis for Alzheimer’s disease prediction. By replacing traditional linear summation with deep neural networks, their model automatically accounts for epistatic dependencies within specific genomic regions, significantly outperforming standard weighted risk scores.
Similarly, in Predicting the prevalence of complex genetic diseases from individual genotype profiles using capsule networks, Xiongbin Kang and Alexander Schönhuth leverage capsule networks to model non-linear interactions across the entire genome, successfully identifying "non-additive" genes that were previously hidden from view.
When dealing with multi-gene perturbations, Yusuf Roohani et al. introduced GEARS in Predicting transcriptional outcomes of novel multigene perturbations with GEARS. GEARS represents genetic perturbations as distinct multidimensional embeddings and composes them through non-linear modules. This design allows the model to predict emergent biological effects—such as synergy and epistasis—even for gene combinations that were never seen during training.
2. Non-Parametric Kernel Methods
Another powerful approach avoids explicit interaction modeling by mapping genetic data into high-dimensional feature spaces using kernels. In a Reproducing Kernel Hilbert Space (RKHS), complex non-linear relationships in the original low-dimensional space can be evaluated as linear relationships.
In Nonparametric Method for Genomics-Based Prediction of Performance of Quantitative Traits Involving Epistasis in Plant Breeding, Xiaochun Sun and Rita H. Mumm demonstrated that when epistasis is dominant, non-parametric kernel-based models consistently outperform traditional linear Bayesian methods. By mapping complex genetic patterns into high-dimensional spaces, these models capture broad-sense genetic variation without requiring explicit prior knowledge of which specific genes are interacting.
3. Tree-Based and Ensemble Methods
Decision trees are inherently capable of capturing conditional, non-linear interactions because their structure is built on sequential splits (e.g., "If SNP A is present, check if SNP B is also present").
To exploit this, Hong Xue et al. designed MegaSNPHunter in MegaSNPHunter: a learning approach to detect disease predisposition SNPs and high level interactions in genome wide association study. By using boosting trees to iteratively filter and rank SNPs based on their contribution to a global model rather than their marginal significance, the framework isolates high-level interactions that univariate statistical tests miss.
Similarly, Qianru Wang et al. introduced lo-siRF (low-signal signed iterative random forests) in Epistasis regulates genetic control of cardiac hypertrophy to extract low-signal, non-additive gene-gene interactions driving cardiac traits, proving that tree-based ensembles can uncover complex modifiers invisible to standard GWAS.
Taming the Combinatorial Explosion
If modeling non-linear interactions is so effective, why hasn't the field completely abandoned linear models? The answer lies in the curse of dimensionality.
If a study tracks 1 million SNPs, evaluating all possible pairwise interactions ($O(N^2)$) requires testing approximately 500 billion combinations. Testing three-way or higher-order interactions ($O(N^3)$) quickly becomes computationally impossible.
To overcome this bottleneck, modern epistatic modeling relies on two main strategies:
- Two-Stage Screening and Global Prioritization: Rather than performing an exhaustive search of all combinations, frameworks like the TS-SIS algorithm developed by Jiahan Li et al. in A fast algorithm for detecting gene–gene interactions in genome-wide association studies use a two-stage screening process to rapidly eliminate uninformative dimensions before applying rigorous interaction thresholding.
- Hardware-Accelerated Optimization: Researchers are rewriting the underlying linear algebra of genomic models to run on modern parallel hardware. As detailed by Hatem Ltaief et al. in Toward Capturing Genetic Epistasis From Multivariate Genome-Wide Association Studies Using Mixed-Precision Kernel Ridge Regression, leveraging low-precision hardware (such as INT8/FP8) on GPU supercomputers allows multivariate kernel ridge regression to scale to cohorts of millions of individuals, processing massive genomic matrices orders of magnitude faster than traditional CPU-bound software.
By combining the structural flexibility of machine learning with high-performance computing, this research direction transforms epistasis from an intractable computational headache into a measurable, predictable component of complex trait architecture.
Key Foundational Work
MegaSNPHunter: a learning approach to detect disease predisposition SNPs and high level interactions in genome wide association study (2009) — contributed 2 concept elements to this direction.