Researchers Develop AI-Based Blood Test for Early Ovarian Cancer Detection
January 29, 2024The molecular diversity among ovarian cancer (OC) patients poses a significant hurdle to early detection, hindering the identification of shared biomarkers. In a recent study titled “A personalized probabilistic approach to ovarian cancer diagnostics” published in Gynecologic Oncology, researchers from the Georgia Institute of Technology have overcome this challenge by employing machine learning (ML) on patient metabolic profiles. This approach identifies biomarker patterns for personalized OC diagnosis, addressing the complexity of early detection in diseases like OC, which progress rapidly with minimal clinical symptoms in the early stages. Notably, the average five-year survival rate for late-stage ovarian cancer patients is only 31%, emphasizing the critical importance of early detection, where the survival rate can exceed 90%.
The ML-based classifiers developed by Georgia Tech researchers demonstrated a 93% accuracy in distinguishing cancer from control samples across 564 patient samples from Georgia, North Carolina, Philadelphia, and Western Canada.
John McDonald, a professor in the School of Biological Sciences at Georgia Tech and the corresponding author of the study, emphasized the current focus on personalized therapy targeting individual genes. However, he noted that this approach may not be universally effective due to the molecular heterogeneity observed in patients with the same disease. McDonald cited the limited success of targeted immune therapies in ovarian cancer, working in only about 5% of patients due to their significant heterogeneity.
While acknowledging the value of profiling individual patients to address heterogeneity, McDonald stressed the importance of “knowing what you’re looking for.”
McDonald continued, stating, “You can do trial and error, but my belief is that artificial intelligence (AI) will bridge that gap by leveraging computers to analyze patterns.”
Given that only seven percent of the numerous metabolites circulating in human blood have been characterized, pinpointing the specific molecular processes within an individual’s metabolic profile remains challenging, limiting the pathway for therapeutic development. Nonetheless, ML-based predictive models can utilize a vast dataset of uncharacterized metabolites, accurately identified by mass spectrometry, to recognize metabolic patterns and aid in ovarian cancer (OC) diagnostics.
McDonald highlighted that historically, predictions in medicine were made based on correlations, without the need to understand cause-and-effect relationships. He emphasized that understanding why these metabolites change is not essential for making predictions.
While McDonald recognizes the ideal nature of a multi-omic ML approach for learning disease patterns, he acknowledged that scalability and complexity currently pose limitations. The team opted for metabolic profiles as an alternative, considering them a biological “end point” that captures the combined effects of molecular changes, including factors like protein level, epigenetic control, diet, lifestyle, and more.
Capitalizing on the accuracy of their ML approach, the research team devised a clinically valuable method for ovarian cancer (OC) diagnosis. In this approach, a patient’s individual metabolic profile is utilized to calculate the probability of developing OC. For instance, if a patient’s metabolic profile falls within a low likelihood cancer range, they would be recommended for yearly monitoring. Conversely, a patient scoring within a range where over 90% of previously diagnosed OC cases occurred would be advised for immediate screening and treatment.
McDonald emphasized the clinical informativeness of this probabilistic diagnostic approach compared to binary (yes/no) tests. The diagnostic test has progressed to trials with various clinics in Georgia, undergoing prospective evaluation of its validity and utility. Additionally, a startup named MyOncoDx is being established with the intention to market the technology upon completion of necessary trials and FDA approval. McDonald estimated that the test could become generally available in approximately a year or so.
