NEW YORK – Lunit on Monday said that it has partnered with AstraZeneca to develop digital pathology tools that can predict the likelihood that a tumor harbors non-small cell lung cancer-driver mutations based on hematoxylin and eosin (H&E)-stained slide images.
The firms aim to develop a rapid AI-powered screening tool for predicting actionable mutations, beginning with EGFR, from whole slide images of tissue samples. The tool would initially be developed for research use only, but Lunit Oncology Group Head Ken Nesmith said the companies envision it as a pre-screening tool that could identify patients likely to have an EGFR mutation ahead of standard molecular testing, giving doctors an opportunity to prioritize next-generation sequencing results for certain patients before starting standard-of-care immunotherapy.
The new tool is to draw on Lunit's AI algorithm, dubbed Lunit Scope, which comprises two computer vision models — a cell detection model and a tissue segmentation model — which are themselves based on a deep learning convolutional neural network. Lunit Scope was trained on annotated whole slide images of lung cancer samples from the Prostate Lung Colorectal Ovarian cancer screening trial, the National Lung Screening Trial, and the biobanking company iSpecimen.
Lunit currently offers two tools based on this algorithm for research use, Lunit Scope IO and Lunit Scope PD-L1, that both aim to identify patients that could potentially benefit from immunotherapy.
In 2022, Seoul, South Korea-based Lunit published a study in The American Journal of Pathology demonstrating that Lunit Scope could analyze H&E slides of non-small-cell lung carcinoma samples from The Cancer Genome Atlas and generate immune and mutational profiles that classified them into three cancer immune phenotypes: inflamed, excluded, and desert.
In samples with the inflamed subtype, pathways associated with immune response and immune-related cell types were enriched. Those results align with previous studies suggesting that clinical response to immunotherapy is most robust in tumor cells with an inflamed subtype. KRAS and BRAF mutations and the MET splicing variant were found mostly in cells with the inflamed subtype.
Cells with an immune-excluded subtype showed enriched metabolic pathways, including glycolysis, fatty acid metabolism, and oxidative phosphorylation. Those cells were the most likely to have EGFR and PIK3CA mutations, which could account for the poor response to immunotherapy often seen in this subtype. Lastly, tumor cells with the immune-desert subtype had the lowest lymphocyte density and were enriched in certain pathways, including protein secretion, G2M checkpoint, and Wnt β catenin signaling pathways, among others.
That study underpins Lunit Scope IO, which classifies 16 types of cancer cells, including lung, head and neck, breast, melanoma, and others, into these three immune phenotypes.
The other Lunit Scope tool, Lunit Scope PD-L1, quantifies PD-L1 in tumor cells from whole slide images to identify patients eligible for PD-L1 therapy. In September, Lunit announced that it began a collaboration with Roche to integrate a tumor proportion score (TPS) algorithm for PD-L1 scoring by Lunit Scope PD-L1 into Roche's Navify digital pathology platform.
In 2023, Guardant Health announced it would incorporate Lunit's technology into the Guardant Galaxy suite of AI analytics for cancer testing, beginning with Lunit Scope PD-L1, under a partnership agreement inked in 2021. Guardant said it may also add Lunit's HER2 scoring and inflammatory scoring assays to its portfolio to investigate the platform's ability to identify patients eligible for HER2-targeted therapies and immune checkpoint inhibitors, respectively.
In this new collaboration with AstraZeneca, Lunit will deploy a new tool, the Lunit Scope Genotype Predictor, to assess the likelihood that a tumor harbors NSCLC driver mutations, beginning with EGFR mutations, based on H&E whole slide images. As a part of the collaboration, the partners plan to explore predicting additional biomarkers in the future.
"Genomic testing in lung [cancer] is resource-intensive and time-consuming, and it's too often bypassed because of the urgency to simply begin treatment," Nesmith said, adding that a tool like Lunit Scope Genotype Predictor could allow the clinician to prioritize molecular testing for patients whose tumors are identified as having a high likelihood of harboring an EGFR or other lung cancer driver mutation.
At the same time, Nesmith said, physicians could hold standard immunotherapy or other treatments until genomic profiling results are available. "Too often it's happening that the doctor goes straight to immunotherapy. If [next-generation sequencing] panel testing, which takes two or three weeks, shows an EGFR mutation, it's probably too late to change course because they're already on immunotherapy and unlikely to be taken off it."
Nesmith said Lunit and AstraZeneca are planning "extensive" real world validation studies to confirm the tool's predictive accuracy and assess its clinical impact. Until then, the tool will be developed for research use only. "This work will ensure the tool's reliability and effectiveness in informing treatment decisions for non-small cell lung [cancer] patients," Nesmith said.
AstraZeneca declined to comment for this story.
Carlo Bifulco, chief medical officer of Providence Genomics and director of translational molecular pathology at the Earle A. Chiles Research Institute, is part of a team developing GigaPath, a whole-slide foundation model for digital pathology designed for cancer subtyping that may also have utility for classifying genetic mutations in tumors based on slide images alone.
Bifulco described Lunit's approach as "valuable" and said it "demonstrates an important application of convolutional neural networks." He sees Lunit's technology as primarily a screening tool to enrich a cohort of patients that should be sequenced and could be particularly useful in settings where there are persistent gaps in access to NGS testing and in screening retrospective cohorts for clinical studies.
"In the long term, I believe this approach will excel in supporting [antibody-drug conjugates], T-cell engagers, bispecifics, et cetera, where multiplexing needs exceed what conventional immunohistochemistry [testing] or similar technologies can currently provide," Bifulco said. "Additionally, these tools could play a role in [quality assurance and quality control], identifying discrepancies or cases that warrant alternative molecular assessment methods."
Bifulco noted that the field of AI-based biomarker prediction lacks consensus benchmarks for validation and that establishing those should be a priority to ensure the comparability and robustness of models like Lunit's. Although he did not have concerns specific to Lunit's model, he said overfitting and model brittleness in general are ongoing challenges for AI in the biomarker prediction space.
Lunit is not the only organization pursuing AI solutions for predicting which patients have actionable mutations ahead of molecular testing. Owkin announced today that it is partnering with Proscia to incorporate its AI diagnostic algorithm MSIntuit CRC v2 into Proscia's precision medicine AI portfolio as a tool to pre-screen patients for microsatellite stable (MSS)/proficient mismatch repair (pMMR) colorectal tumors. The companies said this would allow physicians to expedite testing for those patients by immunohistochemistry (IHC) staining, PCR testing, or NGS.
Meanwhile, BioAI is using deep learning to identify biomarkers based on immunohistochemistry (IHC) testing or H&E slides, and recently partnered with Arbele to develop a companion diagnostic to identify colorectal cancer patients with CDH17 expression who might respond well to the company's antibody-drug conjugates.
Another research group from Guangzhou Medical University in China has introduced an AI algorithm, DeepGEM, to predict mutations associated with all types of lung cancer from images of tumor biopsies. The researchers envision the algorithm as an alternative pathway for patients to access targeted treatments in regions where conventional genomic testing is not available due to cost or where the expected turnaround time for testing is too long.