A histology-based AI biomarker was able to estimate personalized relative benefit of particular adjuvant chemotherapy regimens among patients with resected pancreatic ductal adenocarcinoma, according to study findings published in the Journal of Clinical Oncology.

Investigators developed and validated a deep learning model that analyzes routine histology slides to predict the relative benefit of adjuvant gemcitabine vs modified FOLFIRINOX (mFOLFIRINOX). The resulting biomarker, termed PANCprAId, identified patient subgroups with differential benefit from the two regimens in an external validation cohort from the randomized phase III PRODIGE-24/CCTG PA6 trial.

“Our results suggest that histology-based models can identify patients who may benefit from gemictabine or mFOLFIRINOX, supporting a more biology-driven treatment strategy,” the study authors, including corresponding author Remy Nicolle, PhD, of Centre de Recherche sur l'Inflammation (CRI), Université Paris Cité, Paris, wrote.

Study Details

The investigators trained separate deep learning models to predict disease-free survival following adjuvant gemcitabine or mFOLFIRINOX using digitized whole-slide images from hematoxylin, eosin, and saffron–stained resection specimens. The development cohort comprised 231 patients who underwent curative-intent pancreatectomy and subsequently received either gemcitabine (n = 177) or mFOLFIRINOX (n = 54). The treatment-specific models were then integrated into the PANCprAId algorithm, which estimated the relative benefit of one regimen compared with the other for an individual patient.

External validation was performed in 313 assessable patients from the randomized PRODIGE-24/CCTG PA6 trial, including 137 patients who were treated with gemcitabine and 176 treated with mFOLFIRINOX. Survival analyses used stratified Cox proportional hazards models, Kaplan-Meier estimates, and interaction testing to determine whether the biomarker predicted differential treatment benefit.

Model Methods

Tumor and nontumor tissue were tiled and then classified using the PACpAInt framework. Tissue features were extracted with the pretrained foundation model UNI, and the features were then aggregated and processed with attention-based multiple-instance learning.

Deep learning models were trained on an open-source pipeline to predict disease-free survival based on validation Cox loss. Stratification of sensitivity or resistance to each regimen was assessed on median predicted risk based on the training cohort.

The investigators also generated weight-score heatmaps to highlight the spatial distribution of attention and predicted scores that most impacted the model’s prognostication.

Key Results

Among patients who received gemcitabine, those whose tumors were predicted to respond less favorably had significantly shorter disease-free survival than those whose tumors were predicted to respond more favorably, with a hazard ratio (HR) of 1.69 (95% confidence interval [CI] = 1.04–2.73, P = .03). Among patients treated with mFOLFIRINOX, those whose tumors were predicted to respond less favorably experienced significantly worse disease-free survival than those whose tumors were predicted to respond more favorably (HR = 2.02, 95% CI = 1.40–3.00, P < .001). Each model predicted outcomes only in patients receiving its corresponding treatment and showed no significant prognostic value in patients treated with the alternate regimen.

When the two models were combined into PANCprAId, the biomarker identified patients predicted to derive greater benefit from gemcitabine (favGEM; 15%) or mFOLFIRINOX (favFFX; 85%). In the favFFX subgroup, patients assigned to mFOLFIRINOX achieved longer median disease-free survival than those receiving gemcitabine (21.4 vs 11.1 months, HR = 2.00, 95% CI = 1.48–2.67, P < .001). In the favGEM subgroup, median disease-free survival was 33.5 months with gemcitabine vs 23.6 months with mFOLFIRINOX (HR = 0.48, 95% CI = 0.23–1.02, P = .09).

The investigators concluded: “Histology-based deep learning can derive a predictive biomarker of relative benefit from adjuvant [gemcitabine] vs mFOLFIRINOX in resected [pancreatic ductal adenocarcinoma]."

Associate editor of the Journal of Clinical Oncology Ravi B. Parikh, MD, MPP, FACP, who is also Editor-in-Chief of ASCO AI in Oncology, suggested that AI-based pathology biomarkers, such as PANCprAId, should be incorporated into future phase III studies as prespecified correlatives or stratification factors.

DISCLOSURES: The study was supported by the Ligue Contre le Cancer and the Institute National du Cancer. For full disclosures of the study authors as well as code availability, visit ascopubs.org.

Insights

In an accompanying editorial, Gabriel A. Brooks, MD, MPH, of Dartmouth Cancer Center, and Raghav Sundar, MD, PhD, of the Yale School of Medicine, raised the findings of Beaufils et al in the context of a prior study of a histology-based AI biomarker for predicting chemotherapy in advanced pancreatic cancer (see here for more on this prior study).

They overall found the results of the studies of both AI biomarkers to be exciting. “The provocative independent findings from this team, in a related but distinct clinical setting, suggest that histomorphology-based biomarkers for pancreatic cancer are likely to be more than just a flash in the pan,” Drs. Brooks and Sundar wrote, in regards to the Beaufils et al study.

However, they also raised questions about future development, validation, and costs for AI-derived biomarkers. “From a methodological standpoint, what constitutes adequate validation for AI-derived biomarkers? Should regulatory standards for approval differ between transparent, mechanistically grounded biomarkers and black box AI predictions? Regarding ownership and intellectual property, how should the tension between commercial interests—which incentivize innovation through proprietary protection—and scientific principles of transparency and reproducibility be resolved? Who truly owns insights derived from banked patient tissue samples and real-world clinical data, and what obligations do commercial developers have to patients and to the scientific community? Finally…[w]ill AI-based biomarkers democratize precision medicine by reducing the expense of molecular profiling, or will proprietary pricing create new disparities in access to personalized cancer therapies?”

While histomorphology-based AI biomarkers may be of interest, many questions must be answered before they can be accepted and adopted in clinical practice.

DISCLOSURES: For full disclosures of the study authors, visit ascopubs.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.