[Paper Review] Benchmarking 80 binary phenotypes from the openSNP dataset using deep learning algorithms and polygenic risk score tools

Muhammad Muneeb, David B. Ascher+3 more|arXiv (Cornell University)|Mar 6, 2026

Genetic Associations and Epidemiology0 citations

TL;DR

The study benchmarks 80 binary phenotypes from openSNP using 29 ML, 80 DL, and 3 PRS tools with 675 clumping/pruning configurations, reporting average 5-fold AUC to compare performance between ML/DL and PRS approaches.

ABSTRACT

Genotype-phenotype prediction plays a crucial role in identifying disease-causing single nucleotide polymorphisms and precision medicine. In this manuscript, we benchmark the performance of various machine/deep learning algorithms and polygenic risk score tools on 80 binary phenotypes extracted from the openSNP dataset. After cleaning and extraction, the genotype data for each phenotype is passed to PLINK for quality control, after which it is transformed separately for each of the considered tools/algorithms. To compute polygenic risk scores, we used the quality control measures for the test data and the genome-wide association studies summary statistic file, along with various combinations of clumping and pruning. For the machine learning algorithms, we used p-value thresholding on the training data to select the single nucleotide polymorphisms, and the resulting data was passed to the algorithm. Our results report the average 5-fold Area Under the Curve (AUC) for 29 machine learning algorithms, 80 deep learning algorithms, and 3 polygenic risk scores tools with 675 different clumping and pruning parameters. Machine learning outperformed for 44 phenotypes, while polygenic risk score tools excelled for 36 phenotypes. The results give us valuable insights into which techniques tend to perform better for certain phenotypes compared to more traditional polygenic risk scores tools.

Motivation & Objective

• Benchmark genotype-phenotype prediction on openSNP-derived binary phenotypes.
• Compare machine learning, deep learning, and polygenic risk score tools for case-control classification.
• Systematically vary SNP pre-selection and PRS parameters to assess performance impact.
• Highlight phenotypes where environmental factors limit genetic prediction and discuss transfer learning considerations.

Proposed method

• Preprocess openSNP phenotype data to extract 80 binary phenotypes and convert to Plink format.
• Quality-control genotype data with thresholds (MAF 0.01, HWE 1e-6, genotype rate 0.01, missing rate 0.7) and remove duplicates.
• For ML/DL: use p-value thresholding from GWAS to select SNPs (50–10000 SNPs variants) and train 29 ML and 80 DL models with various hyperparameters; report average 5-fold AUC.
• For PRS: generate GWAS-based base files from training data, apply quality control, perform pruning and clumping with 675 parameter combinations across Plink, PRSice, and Lassosum tools; convert PRS to binary for evaluation.
• Compare performance across ML/DL and PRS tools using AUC, and synthesize best-performing models by phenotype.

Figure 1: A workflow of genotype-phenotype prediction using ML/DL and PRS. A case/control classification flowchart using ML/DL and PRS tools. First, clean phenotype data and extract binary phenotypes from the openSNP dataset. Second, merge the genotype data for each phenotype, convert the dataset to

Experimental results

Research questions

• RQ1 Which phenotypes are better predicted by ML/DL models versus polygenic risk score tools?
• RQ2 How do SNP pre-selection thresholds and clumping/pruning parameters influence predictive performance across methods?
• RQ3 Which specific models/hyperparameters yield the best AUC for individual phenotypes?
• RQ4 Are there phenotypes where environmental factors limit genetic prediction, and can transfer learning mitigate data scarcity?

Key findings

• ML/DL outperformed PRS tools for 44 phenotypes, while PRS tools excelled for 36 phenotypes.
• ANN was the best DL algorithm for 26 phenotypes; XGBoost was the best ML algorithm for 11 phenotypes.
• Best DL hyperparameters: dropout 0.2, optimizer Adam, batch size 1, epochs 50 (23 phenotypes).
• PRS tools showed Plink often performing best among PRS tools, with clumping threshold 0.1 yielding better results; Lassosum benefited from pruning parameter variation.
• For some phenotypes (e.g., Type IIDiabetes, Seborrhoeic Dermatitis, Aphantasia, Eczema, Hypertension, Plantar Fasciitis, Fibromyalgia) ML/DL achieved AUC ≥ 80%; for Scoliosis, Restless leg syndrome, Misophonia, Hypertriglyceridemia, Bone Mineral Density, PRS achieved AUC ≥ 80%.
• Overall conclusion: best model depends on data quality, phenotype architecture, and hyperparameters; openSNP data limits interpretation; transfer learning discussed as a strategy.

Figure 2: This diagram shows the AUC for each phenotype obtained from the ML/DL algorithms and group phenotypes on the number of SNPs that yield the best results.