Abstract
We analyze dog genotypes (i.e., positions of dog DNA sequences that often vary between different dogs) in order to predict the corresponding phenotypes (i.e., unique observed characteristics). More specifically, given chromosome data from a dog, we aim to predict the breed, height, and weight. We explore a variety of linear and non-linear classification and regression techniques to accomplish these three tasks. We also investigate the use of a neural network (both in linear and non-linear modes) for breed classification and compare the performance to traditional statistical methods. We show that linear methods generally outperform or match the performance of non-linear methods for breed classification. However, we show that the reverse is true for height and weight regression. Finally, we evaluate the results of all of these methods based on the number of input features used in the analysis. We conduct experiments using different fractions of the full genomic sequences, resulting in input sequences ranging from 20 SNPs to ~200k SNPs. In doing so, we explore the impact of using a very limited number of SNPs for prediction. Our experiments demonstrate that these phenotypes in dogs can be predicted with as few as 0.5% of randomly selected SNPs (i.e., 992 SNPs) and that dog breeds can be classified with 50% balanced accuracy with as few as 0.02% SNPs (i.e., 40 SNPs).
Introduction
In the field of genomics, Genome-Wide Association Studies (GWAS) are frequently used to identify correlations between genetic markers—specifically Single Nucleotide Polymorphisms (SNPs)—and physical traits, or phenotypes. However, using these markers to build predictive models is a distinct and significant challenge. This research aims to determine which machine learning architectures are best suited for genomic data, specifically comparing complex non-linear models like deep neural networks against traditional linear statistical methods for predicting a dog's breed and physical dimensions (height and weight). A secondary objective was to investigate the "minimum viable data" required for reliable prediction, which has profound implications for reducing the cost of genetic testing in veterinary medicine and agriculture.
Methodology
The researchers explored a wide array of classification and regression techniques to address the prediction tasks:
- Breed Classification: The team tested Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forests, AdaBoost, and Multi-Layer Perceptrons (MLP).
- Phenotype Regression (Height/Weight): Techniques utilized included Elastic Net, XGBoost, Support Vector Regressors, and KNN Regressors.
- Neural Network Variants: A specific MLP was designed to toggle between linear and non-linear modes (using ReLU activation) to isolate and measure the impact of non-linearity on model performance.
A key part of the methodology involved subsampling. The models were not only run on the full dataset but were also tested across a range of input sizes, from a mere 20 SNPs up to the full set of approximately 200,000 SNPs.
Data and Visualization
The study utilized a dataset derived from the "Embark" genotyping array, consisting of 198,473 SNPs per dog. The sample included 482 purebred dogs representing 75 distinct breeds. Height and weight measurements were available for a subset of 337 dogs. To assist in visualization and analysis, dogs were grouped into "clades" based on their genetic lineage.
Discussion of Genetic Clustering
The researchers employed dimensionality reduction techniques, specifically Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE), to visualize how genetic data naturally clusters by breed and clade.
A striking insight from these visualizations and the subsequent model evaluations is the "split" in model performance between breed and physical traits. Breed identification appears to be a largely additive genetic problem, which explains why linear models like Logistic Regression are highly effective. In contrast, height and weight are influenced by more complex, non-linear interactions between genes, where models like KNN and XGBoost demonstrate superior performance. Furthermore, the selective breeding of purebred dogs has introduced significant genetic redundancy, allowing models to achieve high accuracy even when using only a tiny fraction of the available genetic data.
Key Results
- Breed Accuracy: Logistic Regression achieved a peak of 92% balanced accuracy using the full SNP set.
- Extreme Efficiency: The models achieved 50% balanced accuracy for breed classification using only 40 SNPs (0.02% of the data).
- Size Prediction: For height and weight, the KNN regressor achieved an $R^2$ value of approximately 0.8 using only 0.5% of the SNPs (992 markers).
- Linear Dominance: The top five methods for breed classification were all linear, confirming that complex non-linear architectures are often unnecessary for ancestry-related tasks in genomic data.
Limitations and Future Work
The primary limitation of this study is the relatively small sample size. With only 482 dogs, the dataset is small for training deep neural networks, which may explain why simpler models often outperformed more complex ones. Additionally, the study focused exclusively on purebred dogs; predicting traits in mixed-breed dogs would be significantly more challenging due to more complex genetic mixing.
Future work suggested by the authors includes:
- Intelligent SNP Selection: Instead of random sampling, using "informed" selection (picking SNPs known to be associated with specific traits) could further boost accuracy with even less data.
- Cross-Species Application: Applying these machine learning techniques to crops or other domestic animals to optimize breeding for nutrition or climate resilience.
- Mixed-Breed Modeling: Extending the architecture to handle the more "noisy" genetic data of non-purebred populations.
Summary updated