TL;DR
Researchers have demonstrated that deep learning (DL) significantly outperforms traditional linear models in predicting blood and urine biomarkers from genetic data. By moving beyond the "additive assumption," their models capture complex non-linear relationships, non-additive genetic effects, and gene-gene interactions (epistasis), particularly in lipid-related traits.
What Problem Does This Paper Address?
Most current methods for calculating Polygenic Scores (PGS)—which predict a person's trait or disease risk based on their DNA—rely on a major simplification: additivity. They assume that the effect of each genetic variant (SNP) simply adds up to a final score.
However, biology is rarely that simple. This "additive model" misses two critical phenomena:
- Non-additive allelic effects: Where having two copies of a variant isn't exactly twice as impactful as having one (e.g., dominance or recessivity).
- Epistasis: Where the effect of one gene depends on the presence of another.
The authors set out to see if Deep Learning could bridge this gap by identifying these "hidden" non-linear patterns in the UK Biobank data.
Methods
The team developed a deep learning architecture called Genome-local-net (GLN).
- Architecture: GLN uses locally connected layers and residual blocks designed to handle high-dimensional genotype data efficiently.
- Encoding: Unlike linear models that treat genotypes as 0, 1, or 2, the DL models used one-hot encoding. This allows the model to treat each genotype (AA, AG, GG) as a distinct category, naturally capturing non-additive (dominant/recessive) effects.
- Comparison: They benchmarked GLN against state-of-the-art linear methods:
bigstatsrandsnpnet-2.0(LASSO). - Interpretability: They used SHAP values to identify which SNPs were driving the model's decisions and then used Ordinary Least Squares (OLS) to statistically confirm the non-linear and interaction effects the DL model discovered.
Datasets
The study utilized the UK Biobank, one of the world's largest genetic and health resources.
- Samples: ~488,000 individuals.
- Features: ~803,000 SNPs (filtered to ~487,000 after quality control).
- Targets: 34 different blood and urine biomarkers (e.g., cholesterol, bilirubin, testosterone).
- Diversity: While trained on UK-origin data, the models were evaluated across diverse populations, including non-UK European, African, and Asian cohorts.
Key Insights and Ideas
- Beyond the Straight Line: A significant portion of the DL model's edge came from capturing non-linear patterns in basic covariates like age and sex. For example, the way cholesterol levels change with age isn't a simple straight line, and it differs between men and women.
- Capturing Epistasis: The models successfully identified strong interaction effects between different SNPs. For Lipoprotein(a), the DL model found a complex web of interactions that linear models simply cannot see.
- The "GWAS Set" Advantage: Interestingly, when the researchers restricted the models to only use SNPs already known to be significant from GWAS (~9,000 SNPs), the DL models performed equally well or better than linear models for all 34 traits. This suggests DL is more efficient at squeezing information out of known genetic signals.
- Cross-Population Consistency: The performance gains seen in the DL models were consistent across different ethnic backgrounds, suggesting that the non-linear biological mechanisms the model learned are likely universal rather than population-specific.
Key Results
- Lipid Traits Win Big: The DL models showed the greatest improvement in lipid-related traits. For total bilirubin, the DL model achieved an $R^2$ gain of 0.093 over the best linear model.
- Quantifiable Gains: For the "GWAS set" of SNPs, the difference in $R^2$ (a measure of how much variance the model explains) increased by 0.034 to 0.102 for key biomarkers.
- Lipoprotein(a) Performance: In Asian cohorts, the DL model's $R^2$ gain for Lipoprotein(a) reached as high as 0.12.
- Non-Additive Discovery: The authors highlighted specific SNPs, like
rs10455872(LPA gene) andrs887829(UGT1A1 gene), where the DL model correctly utilized dominant and recessive patterns that linear models were forced to average out.
Limitations and Future Work
- Overfitting on Large Feature Sets: When using the "full set" of 500k SNPs, the DL models occasionally overfit on certain traits (like HDL cholesterol and HbA1c), performing slightly worse than linear models. This suggests that better SNP pre-selection is still necessary.
- Interpretability Gap: While SHAP values help, connecting a complex neural network's weights back to specific biological pathways remains a challenge.
- Multi-modal Integration: The authors suggest that the real future of this work lies in combining genomics with other data types—like proteomics, medical imaging, and electronic health records—to create truly personalized medicine.
Summary updated