our technology
We develop our genomic classifiers using novel, machine-learning algorithms that are trained to answer specific clinical questions. Our approach is comparable to facial recognition software, which recognizes patterns of pixels that correspond with the image of a person’s face. In Veracyte’s case, instead of pixels, the technology recognizes patterns of genes that correspond with “clinical truth” (for example: a benign thyroid nodule).

Using the facial-recognition software metaphor, one would train the technology to distinguish between Person A and someone else, using high-quality images of Person A. One would train the software to recognize Person A’s eye, for example, by identifying not just the raw data (e.g., black and white pixels) that make up the eye, but also by recognizing the patterns of those pixels and how they fit together into a bigger picture that is specific to Person A.

Similarly, we have trained our Afirma Gene Expression Classifier to recognize benign thyroid nodules among those deemed “inconclusive”–not clearly benign or malignant–by traditional diagnostic approaches. We used hundreds of reference samples of such thyroid nodules for which the status–or “clinical truth”–was subsequently known
(based on surgical evaluation, which is considered the “gold standard”). Similar to the facial-recognition software example, we trained our algorithm to identify not just genes, but also gene patterns, that distinguish benign thyroid nodules from those that are malignant.

As genomic understanding has advanced, scientists now know that the presence or absence of a specific gene or genes does not necessarily signify cancer or another disease. Rather, these inputs are just pieces of the bigger picture. Potentially, this recent learning will further increase the significant role of machine-learning technology in diagnostics. As scientific understanding continues to progress, additional features that inform disease status, such as gene mutation or imaging data, can potentially help us define “clinical truth” with ever-more precision, enabling development of tests that provide even clearer answers to important clinical questions. Similar to how moving from low- to high-definition imaging technology can significantly improve facial recognition software.