Predicting Canine Cancer and Cardiac Disease Risk with Revolutionary Direct-to-Consumer Genetic Testing

In recent years, direct-to-consumer genetic testing has gained widespread adoption for both people and their animal companions. This trend accelerated during the COVID-19 pandemic, when at-home health testing became widely adopted. Although these tests do not always achieve the same level of analytical rigor as laboratory-based methods, their convenience has significantly expanded access to health-related insights.

At SmartDx, we are building on this paradigm by bringing advanced, AI-powered genetic risk assessment to the veterinary space, enabling animal companions to benefit from broader access to testing and more meaningful insights into disease risk.

The Limitations of Most At-Home Health Tests

Most existing assessments focus on identifying individual mutations associated with disease. While this information can be useful, it provides only a partial picture. A single gene rarely drives disease risk. Instead, it reflects complex interactions across the genome. Given the high prevalence of potentially concerning variants in canine populations, mutation detection alone is insufficient for accurate risk assessment. That’s why pet parents need an intelligent, integrative approach that can evaluate how genetic factors work together to influence disease outcomes.

Why SmartDx Is the Future of Direct-to-Consumer Genetic Testing

SmartDx addresses this gap by offering machine-learning driven risk prediction for cancer and cardiac diseases. By analyzing genomic data holistically rather than in isolation, our models provide a more nuanced and clinically meaningful assessment of disease risk. Machine learning, particularly deep neural networks, has become state-of-the-art across many phenotype classification applications. However, training high-performing machine learning models requires large, well-annotated datasets.

To maximize clinical relevance and population impact, SmartDx focuses on diseases that are both prevalent and broadly distributed across breeds. Our models target common canine cancers such as hemangiosarcoma, lymphoma, and osteosarcoma, as well as select cardiac conditions. Because these diseases aren’t confined to specific breeds, our assessments apply to a wide and diverse canine population.

Leveraging Today’s Machine Learning to Power Tomorrow’s Healthcare

Today’s direct-to-consumer genetic testing provides convenient access to high-level health information, but most stop at reporting the presence or absence of specific mutations.

Interpreting disease risk from this information alone is challenging due to the complex genetic architecture underlying most conditions. SmartDx advances beyond this limitation by applying machine-learning algorithms that integrate genetic relationships across the genome, delivering deeper insight into a dog’s likelihood of developing common cancers and cardiac diseases.

Morse is SmartDx’s proprietary preprocessing algorithm designed to efficiently transform raw genomic data into a model-ready format. Morse extracts genotype information from sequenced genomes and converts these data into a compact binary representation. SmartDx’s machine-learning risk assessment engine, Mew, then analyzes this binarized representation. Morse serves as a critical bridge between sequencing and modeling, ensuring consistent, high-throughput processing of genomic data derived from samples collected using DNA Genotek Performagene kits and sequenced on the Thermo Fisher Axiom Canine HD array.

Our direct-to-consumer genetic testing panels predict lifetime disease outcome (risk), not current disease status. Using sequenced genomic data, each model estimates whether a dog is likely to develop a specific condition over the course of its natural life, enabling proactive health planning rather than reactive care.

Exploring SmartDx’s Model Development

All SmartDx models are trained and evaluated on datasets balanced with a 50:50 ratio of diagnosed-to-control dogs. To mitigate class imbalance, we apply oversampling methods, while a proprietary interpolation strategy handles incomplete genotype calls. Control dogs reach advanced age without developing the diseases represented in each panel, enabling high-confidence labeling of lifetime disease outcomes.

Model development was further enhanced through semi-supervised learning applied to more than 14,000 unlabeled canine genomes, expanding dataset diversity and improving model generalizability. Self-supervised learning techniques enable these unlabeled samples to be directly incorporated into training, allowing them to contribute meaningfully to performance gains. This strategy is particularly important in domains where fully labeled datasets are scarce, supporting robust and accurate risk prediction despite incomplete phenotype annotation. Model training follows established machine learning best practices, with:

• 70% of labeled data used for training
• 10% for validation
• 20% held out for independent testing

Collectively, these datasets span more than 300 canine breeds, helping to minimize breed-specific bias in lifetime disease risk prediction in our direct-to-consumer genetic testing.

Cancer Panel

Description
The cancer panel predicts whether a dog will develop any of six common cancer types over their natural lifespan. Each disease-specific model was trained independently in a one-vs-rest format, treating dogs with other cancers and healthy controls as the negative class.

Test-Set Metrics (n = 510)

Cardiac Panel

Description
The cardiac panel predicts whether a dog is genetically predisposed to develop specific heart conditions over the course of their natural lifetime. Each disease-specific model was trained independently in a one-vs-rest format, treating dogs with other cardiac diseases and healthy controls as the negative class.

Test-Set Metrics (n = 30 per disease)