Giulia KennedyPhD
VeracyteChapter 10 · Strand B · Publication

Afirma GSC and the Xpression Atlas

The second-generation Afirma: RNA sequencing and ensemble machine learning, validated in JAMA Surgery in 2018, finding at least a third more benign nodules than the original test, with the Xpression Atlas reading variants and fusions from the same sample.

MAY 2017 – SEP 2022CHIEF SCIENTIFIC OFFICER AND CHIEF MEDICAL OFFICER, VERACYTEVERACYTE, INC. · SOUTH SAN FRANCISCO, CALIFORNIA

Overview

Six years after Afirma launched, Veracyte rebuilt it. The Afirma Genomic Sequencing Classifier (GSC) replaced microarrays with whole-transcriptome RNA sequencing and a single classifier with an ensemble of machine learning algorithms. Pivotal validation data were announced at the AACE congress on May 4, 2017, and the company began moving patients to the new test that quarter.12

Contributions

Kennedy framed the work in the terms of the field it borrowed from: “We are employing the same machine learning methods that are being used in other fields such as social media and self-driving cars, but applying them to thyroid cancer diagnosis”.1 In October 2017, presenting at the American Thyroid Association meeting, she described the ensemble algorithms as deriving “from years of thoughtful scientific research, data analysis and statistical modeling”; the data showed the GSC could separate benign from cancerous Hürthle cell nodules, a long-standing difficulty.3 She is last author of the 2019 BMC Systems Biology paper that set out that solution.4

The clinical validation appeared in JAMA Surgery in 2018. In a blinded study of 191 Bethesda III/IV samples from 49 U.S. centres, the GSC identified at least one third more benign nodules than the original classifier, with 91% sensitivity; Kennedy is among the authors.567 Veracyte estimated the new test could save 70% of patients with benign nodules from surgery.2

The Xpression Atlas

The same sequencing data could be read a second way. The Afirma Xpression Atlas, launched at AACE in May 2018 with Kennedy presenting, reports expressed gene variants and fusions from the whole transcriptome of the same fine-needle aspirate. “Our RNA sequencing-based platform enables us to derive unprecedented amounts of rich genomic content from clinical samples,” she said at the launch.8 Later papers extended the platform to real-world molecular findings across tens of thousands of nodules (JCEM, 2021) and to a medullary thyroid carcinoma classifier (Thyroid, 2022).

From the record

Facts

  • On May 4, 2017 Veracyte announced pivotal clinical validation data for the Afirma Genomic Sequencing Classifier (GSC) at the AACE Annual Scientific and Clinical Congress, and in the second quarter of 2017 began transitioning from the original classifier to the RNA sequencing-based GSC.12
  • Veracyte stated that the GSC could save an estimated 70% of patients with benign nodules from unnecessary thyroid surgery.2
  • The JAMA Surgery validation (Patel et al., announced May 23, 2018) found the GSC identified at least one third more benign nodules among cytologically indeterminate nodules than the original classifier, with 91% sensitivity, in a blinded validation of 191 Bethesda III/IV samples from 49 U.S. centres.56
  • In October 2017 Veracyte presented data at the American Thyroid Association meeting showing the GSC ensemble machine learning algorithms could distinguish challenging subtypes, including benign from cancerous Hürthle cell nodules.3
  • The Afirma Xpression Atlas, which reports gene variants and fusions from the whole transcriptome of thyroid FNA samples, was launched at the AACE congress in May 2018, with Kennedy presenting at the launch event.8

Contributions

  • Co-author of the JAMA Surgery 2018 validation of the Afirma GSC and last author of the BMC Systems Biology 2019 paper on identifying Hürthle cell cancers with a trio of machine learning algorithms.74
  • Described the ensemble machine learning behind the GSC as deriving "from years of thoughtful scientific research, data analysis and statistical modeling".3

Research

  • Thyroid nodule genomic classification

    APR 2008 – DEC 2022

    The programme that defined Veracyte, where Kennedy was Chief Scientific Officer from 2008 and Chief Medical Officer from 2018. The Afirma Gene Expression Classifier was developed from 315 nodules (JCEM, 2010), launched in January 2011 and validated prospectively in 3,789 patients (NEJM, 2012). The RNA sequencing-based Afirma GSC followed in 2017, validated in JAMA Surgery (2018) with at least one third more benign calls than the original test, and was extended with the Xpression Atlas of variants and fusions, Hürthle cell and medullary thyroid carcinoma classifiers. By January 2018 Veracyte had performed 100,000 Afirma tests.

  • Machine learning for clinical diagnostics

    DEC 2010 – PRESENT

    The method that links every classifier Kennedy has helped build since 2010: training algorithms on high-dimensional genomic data from clinical samples and validating them against histopathology. She presented on Veracyte's use of machine learning to translate genomic data into commercial diagnostic tests at the Precision Medicine World Conference in January 2018, and described the Afirma GSC's ensemble machine learning algorithms as deriving "from years of thoughtful scientific research, data analysis and statistical modeling". The same approach underlies the Envisia and Percepta classifiers and PinkDx's whole-transcriptome vaginal swab test.

Technologies
  • Afirma Genomic Sequencing Classifier
  • Whole-transcriptome RNA sequencing
  • Ensemble machine learning classifiers

Publications