This is a repository copy of Pan-cancer image-based detection of clinically actionable genetic alterations.

Purpose: Gastroesophageal adenocarcinoma (GEA) has a poor prognosis and few therapeutic options. Utilizing a 73gene plasma-based next-generation sequencing (NGS) cellfree circulating tumor DNA (ctDNA-NGS) test, we sought to evaluate the role of ctDNA-NGS in guiding clinical decisionmaking in GEA. Experimental Design: We evaluated a large cohort (n ¼ 2,140 tests; 1,630 patients) of ctDNA-NGS results (including 369 clinically annotated patients). Patients were assessed for genomic alteration (GA) distribution and correlation with clinicopathologic characteristics and outcomes. Results: Treatment history, tumor site, and disease burden dictated tumor-DNA shedding and consequent ctDNA-NGS maximum somatic variant allele frequency. Patients with locally advanced disease having detectable ctDNA postoperatively experienced inferior median disease-free survival (P ¼ 0.03). The genomic landscape was similar but not identical to tissue-NGS, reflecting temporospatial molecular heterogeneity, with some targetable GAs identified at higher frequency via ctDNA-NGS compared with previous primary tumor-NGS cohorts. Patients with known microsatellite instabilityhigh (MSI-High) tumors were robustly detected with ctDNA-NGS. Predictive biomarker assessment was optimized by incorporating tissue-NGS and ctDNA-NGS assessment in a complementary manner. HER2 inhibition demonstrated a profound survival benefit in HER2-amplified patients by ctDNA-NGS and/or tissue-NGS (median overall survival, 26.3 vs. 7.4 months; P ¼ 0.002), as did EGFR inhibition in EGFR-amplified patients (median overall survival, 21.1 vs. 14.4 months; P ¼ 0.01). Conclusions: ctDNA-NGS characterized GEA molecular heterogeneity and rendered important prognostic and predictive information, complementary to tissue-NGS. See related commentary by Frankell and Smyth, p. 6893

The Cancer Genome Atlas (TCGA) is one of the largest biorepositories of digital histology. Deep learning (DL) models have been trained on TCGA to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. However, we demonstrate that these features vary substantially across tissue submitting sites in TCGA for over 3,000 patients with six cancer subtypes. Additionally, we show that histologic image differences between submitting sites can easily be identified with DL. Site detection remains possible despite commonly used color normalization and augmentation methods, and we quantify the image characteristics constituting this site-specific digital histology signature. We demonstrate that these site-specific signatures lead to biased accuracy for prediction of features including survival, genomic mutations, and tumor stage. Furthermore, ethnicity can also be inferred from site-specific signatures, which must be accounted for to ensure equitable application of DL. These site-specific signatures can lead to overoptimistic estimates of model performance, and we propose a quadratic programming method that abrogates this bias by ensuring models are not trained and validated on samples from the same site.

Background: Patients with HPVþ oropharyngeal squamous cell carcinoma were assigned to dose and volume de-escalated radiotherapy (RT) or chemoradiotherapy (CRT) based on response to induction chemotherapy in an effort to limit treatmentrelated toxicity while preserving efficacy. Patients and methods: Patients were classified as low-risk (T3, N2B, 10 pack-year history) or high-risk (T4 or N2C or >10 PYH). After three cycles of carboplatin/nab-paclitaxel, response was assessed using Response Evaluation Criteria in Solid Tumors 1.1. Low-risk patients with 50% response received 50 Gray (Gy) RT (RT50) while low-risk patients with 30%-50% response or high-risk patients with 50% response received 45 Gy CRT (CRT45). Patients with lesser response received standardof-care 75 Gy CRT (CRT75). RT/CRT was limited to the first echelon of uninvolved nodes. The primary end point was 2-year progression-free survival compared with a historic control of 85%. Secondary end points included overall survival and toxicity. Results: Sixty-two patients (28 low risk/34 high risk) were enrolled. Of low-risk patients, 71% received RT50 while 21% received CRT45. Of high-risk patients, 71% received CRT45. With a median follow-up of 29 months, 2-year PFS and OS were 95% and 100% for low-risk patients and 94% and 97% for high-risk patients, respectively. The overall 2-year PFS was 94.5% and within the 11% noninferiority margin for the historic control. Grade 3þ mucositis occurred in 30%, 63%, and 91% of the RT50, CRT45, and CRT75 groups, respectively (P ¼ 0.004). Rates of any PEG-tube use were 0%, 31%, and 82% for RT50, CRT45, and CRT75 groups, respectively (P < 0.0001). Conclusions: Induction chemotherapy with response and risk-stratified dose and volume de-escalated RT/CRT for HPVþ OPSCC is associated with favorable oncologic outcomes and reduced acute and chronic toxicity. Further evaluation of induction-based de-escalation in large multicenter studies is justified. Clinical trial registration: Clinical trials.gov identifier: NCT02258659.

IMPORTANCE Postoperative chemoradiation is the standard of care for cancers with positive margins or extracapsular extension, but the benefit of chemotherapy is unclear for patients with other intermediate risk features. OBJECTIVE To evaluate whether machine learning models could identify patients with intermediate-risk head and neck squamous cell carcinoma who would benefit from chemoradiation.