This article has been updated to reflect that Australian National University and the National Cancer Institute collaborated on the development of ENLIGHT-DP.
NEW YORK – Pangea Biomed on Wednesday published a study showing that its artificial intelligence-driven ENLIGHT-DP platform successfully predicted true responders to targeted treatments and immunotherapies based on an analysis of histopathology images.
The results of the study, published in Nature Cancer, are based on an analysis of data from five independent cohorts of cancer patients. The company has already been evaluating ENLIGHT-DP in clinical trials with drug developers. Now, based on the platform's performance in the published study, Pangea is planning follow-up analyses that could support regulatory filings on ENLIGHT-DP, with the goal of offering it in the next two years to physicians as a tool for guiding patient care.
Pangea researchers collaborated with scientists from the Australian National University and the US National Cancer Institute to develop ENLIGHT-DP and to analyze hematoxylin and eosin (H&E)-stained slides from patients with a range of cancers who participated in previously conducted and published clinical studies. They found that patients who received a treatment that the ENLIGHT-DP platform recommended were more than twice as likely to respond. Among patients who received treatments that aligned with ENLIGHT-DP recommendations, the response rate was 39.5 percent higher than the published baseline response from each study.
ENGLIGHT-DP is a two-step AI system combining ENLIGHT (Expression Networks for highlighting Tumor vulnerabilities), which is an algorithm that predicts response to targeted and immune therapies based on gene expression, and DeepPT, a deep-learning framework that predicts mRNA expression from slides.
ENLIGHT uses in vitro, preclinical, and clinical datasets to build genetic interaction maps that can identify gene pairs with significance for tumor vulnerability. Pangea has previously described the use of ENLIGHT to retrospectively predict patient response to therapy based on tumor transcriptomics derived from RNA sequencing data. And in March 2023, the Tel Aviv, Israel-based firm started working with Onconova Therapeutics to search for biomarkers related to the PLK1 pathway, which is inhibited by Onconova's investigational drug rigosertib.
With the addition of DeepPT, Pangea has extended ENLIGHT's capabilities. The researchers trained DeepPT to predict genome-wide tumor mRNA expression from formalin-fixed, paraffin-embedded whole slide images and their corresponding bulk gene expression profiles from patient samples in The Cancer Genome Atlas. They then validated DeepPT's predictions in two independent datasets. Now, in the current study, researchers have used DeepPT to predict expression profiles based on slides from patients in published studies, and applied ENLIGHT to generate patient therapy response predictions.
Pangea CEO Tuvik Beker said that the company wanted to see whether the DeepPT-inferred mRNA expression profiles could replace expression profiles derived from actual mRNA sequencing. "The results exceeded our wildest dreams," Beker said.
The researchers initially expected to find only a narrow scenario in which limited prediction capabilities could be demonstrated based only on pathology slides. "But for every dataset we could lay our hands on, we got predictions that were on par with ones obtained using state-of-the-art RNA sequencing," he said.
The slides used in the study came from patients with HER2-positive breast cancer treated with chemotherapy plus Genentech's Herceptin (trastuzumab); BRCA-positive pancreatic cancer patients treated with PARP inhibitors; a mixed group of lung, cervical, and head and neck cancer patients treated with Merck KGaA's investigational bifunctional fusion protein immunotherapy bintrafusp alfa; and ALK-positive non-small cell lung cancer patients treated with ALK inhibitors.
A key difference between ENLIGHT-DP's cancer treatment response predictive capabilities and other AI models, according to the study authors, is that ENLIGHT-DP does not require training on matched imaging and response data for each specific drug and indication, which can be difficult to obtain. In spite of the lack of drug- and indication-specific training, ENLIGHT-DP performed comparably to supervised models with such training, such as a classifier designed to predict clinical outcomes in non-small cell lung and gynecological cancer patients or a deep neural network model that predicts responses to anti-PD-1 therapies in patients with melanoma and lung cancer from histopathology images.
"We compared ENLIGHT-DP to a direct supervised model using a deep learning architecture very similar to DeepPT and showed that it didn't just match the direct model's performance but actually exceeded it," Beker said.
The main challenge the researchers faced in this study was figuring out why their approach worked as well as it did, even though DeepPT expression inference is inherently noisy. Beker said they concluded that although much of the detail that would be present in mRNA sequencing data is lost when inferred via imaging, the information that is preserved relates to the most robust features of the tumor, allowing DeepPT to make accurate response predictions.
According to Beker, if validated in additional cohorts, ENLIGHT-DP could be useful in situations where patients can't receive next-generation sequencing to inform treatment strategies because they have to get on treatment quickly, they lack sufficient tissue for molecular tumor profiling, or due to other constraints. "It can truly democratize precision oncology, bringing it to low- and middle-income countries where NGS isn't widely available," Beker said.
Pangea has already expanded the analysis performed in the Nature Cancer study and obtained similar results in blinded studies in two new patient cohorts. The company plans to carry out further studies in new cohorts in the coming months, submit the data with regulators and payors, and bring ENLIGHT-DP-based tests to the clinic.
"We are already utilizing ENLIGHT-DP in our collaborations with drug developers to bring new drugs to market more quickly and more effectively," Beker said. "We hope that within the next 24 months we will be able to also start offering it to physicians for routine analysis to guide more informed treatment decisions."