Written by: Stephen Hsu
Primary Source: Information Processing, 12/27/2018
Our new paper describes over a dozen genomic predictors for common disease risk, constructed via machine learning on hundreds of thousands of genotypes. The predictors use anywhere from a few tens (e.g., 20 or 50) to thousands of SNPs to compute the risk PGS (Poly-Genic Score) for a specific disease.
The figure above (Atrial Fibrillation) shows out-of-sample testing of risk prediction (black dots with error bars) compared to theoretical prediction (red line). The theoretical prediction uses the empirical fact that cases and controls are normally-distributed in PGS score, with the two distributions shifted relative to each other. Cases have, on average, higher risk scores, and come to dominate in high PGS percentile bins. So, conditional on a high PGS risk score (e.g., 99th percentile PGS), the probability of the condition can be significantly elevated (e.g., ~8 times typical probability of developing atrial fibrillation).
We can identify, from SNP genotype alone, a subset of the population with unusual risk for conditions like Atrial Fibrillation or Diabetes or Breast Cancer or Prostate Cancer.
Just a year or two ago this would have seemed like science fiction to biomedical researchers…
Empirical validation of risk is limited by availability of out-of-sample populations for whom we have genotype and disease status. However, it is clear from the results that the theoretical models do a good job of predicting odds ratios once the properties of the case and control normal distributions (mean and standard deviation of PGS) are known.
These predictors only require data from an inexpensive ~$50 SNP array. Once the ~1 million SNPs on the array are measured *all* of the disease risks can be computed for an individual patient. It is only a matter of time before genotyping of this kind becomes Standard of Care in health systems around the world.
In the paper we also analyze the rate of improvement of prediction AUC as training sample size increases. With more data these predictors will become significantly more accurate — the relevant timescale is just a few years!
Louis Lello, Timothy Raben, Soke Yuen Yong, Laurent CAM Tellier, Stephen D. H. Hsu
We construct risk predictors using polygenic scores (PGS) computed from common Single Nucleotide Polymorphisms (SNPs) for a number of complex disease conditions, using L1-penalized regression (also known as LASSO) on case-control data from UK Biobank. Among the disease conditions studied are Hypothyroidism, (Resistive) Hypertension, Type 1 and 2 Diabetes, Breast Cancer, Prostate Cancer, Testicular Cancer, Gallstones, Glaucoma, Gout, Atrial Fibrillation, High Cholesterol, Asthma, Basal Cell Carcinoma, Malignant Melanoma, and Heart Attack. We obtain values for the area under the receiver operating characteristic curves (AUC) in the range 0.58 – 0.71 using SNP data alone. Substantially higher predictor AUCs are obtained when incorporating additional variables such as age and sex. Some SNP predictors alone are sufficient to identify outliers (e.g., in the 99th percentile of PGS) with 3-8 times higher risk than typical individuals. We validate predictors out-of-sample using the eMERGE dataset, and also with different ancestry subgroups within the UK Biobank population. Our results indicate that substantial improvements in predictive power are attainable using training sets with larger case populations. We anticipate rapid improvement in genomic prediction as more case-control data become available for analysis.