Towards a Next-Generation GEC with Improved Specificity: Feasibility Analysis Using Machine Learning on a Combination of Gene Expression, Variants and Fusions From a Single Novel Sequencing Platform

Friday, September 23
9:45 AM MT

Interview with Giulia C. Kennedy, Ph.D., Chief Scientific Officer, Veracyte

Dr. Giulia Kennedy is the chief scientific officer of Veracyte. Her vision and research were crucial to the development of Veracyte’s genomic tests to help reduce diagnostic ambiguity and unnecessary surgeries, including the Afirma® Gene Expression Classifier (GEC). Here, Dr. Kennedy discusses findings from recently completed research evaluating the feasibility of an enhanced version of Veractye’s Afirma GEC molecular classifier.


Q: Why did you do this study?

Dr. Kennedy: Years ago when we developed the Afirma GEC molecular test our primary goal was to achieve high sensitivity so that we could help to accurately rule out cancer in patients with indeterminate cytopathology. Now that we have demonstrated – through both extensive clinical research and real-world clinical practice – that we have achieved this goal, we wanted to focus on enhancing the specificity of the test. We developed a novel approach that we believed would help us increase specificity without compromising the test’s sensitivity, and this study was the initial test of this approach.


Q: Can you describe your approach?

Dr. Kennedy: We theorized that we could use sophisticated machine learning techniques combined with the rich data provided by a robust RNA sequencing platform to improve the Afirma GEC’s specificity. We used machine learning when we initially developed the Afirma GEC, but now we’re integrating data and algorithms that are much richer and more complex, including gene variant and fusion information.


Q: And what did you find?

Dr. Kennedy: We found that by combining the rich gene expression, variant and fusion information provided by an RNA sequencing platform with advanced algorithms, we were able to train the Afirma GEC to recognize benign thyroid nodules with a high level of specificity while keeping sensitivity high. Interestingly, we’re finding that gene alteration data on its own is not very helpful in differentiating between benign and cancerous nodules. But by using machine learning, we believe we can unleash the power of this information to be clinically useful.


Q: What are the implications of your findings for people with thyroid nodules?

Dr. Kennedy: The biology of thyroid nodules is very complex, making the problem of indeterminate nodules incredibly challenging from a molecular diagnostics standpoint, but also for physicians and their patients. By creating a molecular test that can help resolve these indeterminates with both high sensitivity and specificity, we will help prevent even more people with indeterminate nodules from undergoing surgery on nodules that turn out to be benign.


Q: What is the next step?

Dr. Kennedy: We will apply the same methods used in this initial feasibility study to a larger set of training samples. Then, once we have cross-validated the results from these two studies, we will develop an independent test set. Throughout the development of this enhancement we’ve worked with physicians who are key thought leaders in thyroid cancer, and we will continue to engage them as we move to these next steps.