A deep learning approach to quantify tumor microenvironment features associated with postoperative ctDNA status and outcomes in a phase III FOLFOX-based adjuvant colon cancer trial (N0147; Alliance).

Abstract

3525 Background: Tumor microenvironment (TME) features of colorectal cancer (CRC) may influence detection of circulating tumor DNA (ctDNA) and further refine prognosis. We applied a deep learning histopathologic algorithm to quantify TME features in stage III colon cancers and evaluated their association with postoperative ctDNA status among participants in a phase III FOLFOX-based adjuvant trial (NCCTG N0147). Methods: QuantCRC (Aiforia Technologies) was applied to digitized hematoxylin and eosin-stained whole-slide images of stage III colon adenocarcinomas to extract quantitative histopathologic features. Among cases where data met quality control (N=1817), associations between 15 distinct pathologist-defined features and clinical outcomes—time to recurrence (TTR), disease-free survival (DFS), and overall survival (OS)—were evaluated, stratified by postoperative ctDNA status assessed by a tissue-free assay (Guardant Reveal). Relationships were analyzed using univariate and multivariable Cox proportional hazards models that were adjusted for clinicopathologic and molecular features ( KRAS , BRAF V600E , MMR status). Interaction testing was performed. Results: We found statistically significant quantitative differences for histopathological features by ctDNA status. Features associated with ctDNA positivity(+) included lymphovascular invasion (LVI), and increased %high grade, %tumor budding, %necrosis, and %stroma. ctDNA+ cases had lower %inflammatory stroma and fewer %signet ring cells (all p< 0.006). Multivariable analyses, irrespective of ctDNA status, identified %tumor budding, %mucin, %signet ring cells, %inflammatory tumor bed, and %inflammatory stroma as significantly associated with adverse outcomes (TTR, DFS, OS). Within ctDNA+ cases, higher %immature tumor bed was associated with adverse outcome for all 3 variables. Among ctDNA-negative cases, significantly poorer outcomes (TTR, DFS, OS) were observed for tumors with lower tumor infiltrating lymphocyte (TIL) density and reduced %necrosis. After adjustment, significant interactions with ctDNA status were observed for LVI (p< 0.006) and %signet ring cells (p< 0.035) across all outcome variables, for %necrosis with DFS, OS (p<0.040), and for TILs with TTR (p<0.015). Analysis is ongoing for ctDNA tumor fraction as well as tumor genotyping data (739 genes; Guardant360). Conclusions: QuantCRC identifies TME features associated with postoperative ctDNA status that enable risk-stratification within ctDNA groups. Among ctDNA-negative patients, reduced intratumoral TIL density and reduced tumor necrosis identify high-risk subgroups with inferior clinical outcomes, and may have clinical utility.