AI Method Holds Promise for Precision Medicine
January 29, 2024STARVar Outperforms Other Tools in Diagnosing Real Patients
The researchers tested STARVar on genomic datasets from Saudi Arabian and international patients. It consistently ranked the correct disease variant highest among hundreds of candidates.
This demonstrates STARVar’s superior performance over tools relying on rigid symptom vocabularies. The algorithm leverages flexible text analysis to pinpoint culprit mutations.
The team applied STARVar to diagnose a Saudi girl with various symptoms. Out of nearly 800 variants, it accurately identified the pathogenic mutation in the MMP2 gene responsible for her condition.
This real-world case highlights STARVar’s clinical utility in interpreting genomic data in the context of complex symptoms.
STARVar is now freely available online. The researchers hope clinical geneticists will incorporate this novel analytic approach into diagnostic workflows.
By combining literature evidence, genomes, and crucially, nuanced symptom data, STARVar represents a powerful new AI asset for solving rare disease cases.
Researchers Develop Interpretable AI Models to Uncover Patterns in Biological Data
The autoencoder models aim to compress highly complex bioscience data into relevant patterns and characteristics. As Professor Mika Gustafsson explains, “We let the data speak for itself without steering the model based on prior knowledge. The autoencoder organized the data in ways that map to biological processes.”
By identifying key features in an unbiased manner, the models can uncover predictive signals related to individual environmental exposures where limited training data exists. The researchers can then use these learned patterns to build classifiers for diverse personalized factors.
Gustafsson’s team strives to create interpretable AI models unlike inscrutable “black boxes.” As he states, “We want to understand what the model tells us about disease biology, not just classify who is ill. By interpreting the data, we gain insights into the underlying causes.”
This approach of opening the AI hood and peeking inside holds promise for elucidating the mechanisms of health and disease. The researchers’ long-term goal is to translate patterns in big data into biological knowledge and actionable precision medicine strategies.
The study published in Briefings in Bioinformatics demonstrates the potential of AI to self-organize complex data into meaningful representations. The authors believe interpretable deep learning will become an increasingly valuable tool for accelerating discoveries in biomedicine and healthcare.
The research was funded by the Swedish Research Council, Wallenberg AI/Autonomous Systems and Software Program, and the SciLifeLab & Wallenberg National Program for Data-Driven Life Science.
