Pangea Biomed’s AI predicts cancer treatment response from histopathology images

What if physicians could predict which cancer treatments will work best for individual patients using just their tumor images? The precision oncology company Pangea Biomed made strides in that direction with the publication of its AI-powered ENLIGHT-DP method in Nature Cancer. This technology taps deep learning, an ever-more-popular type of machine learning with a large number of layers, to analyze standard tumor images and predict patient treatment response, potentially bypassing the need for time-consuming and expensive genetic sequencing.

Predicting gene expression from tumor images

ENLIGHT-DP itself was developed through a collaboration between Pangea Biomed and researchers from the Australian National University and the National Cancer Institute, who also jointly conducted the study on the technology.

ENLIGHT-DP works in two steps. First, a deep learning framework known as DeepPT analyzes the tumor image, learning to identify subtle patterns in the way cells are organized and how they appear based. Or as the Nature Cancer paper noted, it “predicts genome-wide tumor mRNA expression from slides” and “successfully predicts transcriptomics in all 16 The Cancer Genome Atlas cohorts tested…”

The resulting visual clues are then translated into a prediction of the tumor’s gene expression — essentially, which genes are turned ‘on’ or ‘off’. The study in Nature Cancer found that DeepPT significantly outperformed existing state-of-the-art methods in predicting gene expression from images. For example, DeepPT achieved a mean of median correlations of 0.41 across nine cancer types when considering the top 1,000 predicted genes, nearly doubling the performance of HE2RNA (0.16) and SEQUOIA (0.26) (Hoang et al., 2024). Two independent datasets provided further validation, demonstrating its ability to generalize beyond the initial training data to unseen data.

Matching patients to effective treatments

In the second step, researchers predicted the genetic information into ENLIGHT, an unsupervised ML approach that analyzes the complex interplay between drug mechanisms and the tumor’s unique genetic profile to predict the likelihood of response to a specific therapy. The Nature Cancer paper explains that the ENLIGHT platform predicts “response to targeted and immune therapies from the inferred expression values.”

One of the most promising aspects of ENLIGHT-DP is its potential to dramatically improve treatment response rates. In the study, patients who received ENLIGHT-matched treatments showed a 2.28 times higher likelihood of responding to therapy compared to those who were not matched, representing a 39.5% increase in response rate. Additionally, the study found that DeepPT accurately captures the prognostic value of certain gene signatures.

Already a real-world impact

While the paper had impressive statistical findings, Tuvik Beker, CEO of Pangea Biomed emphasized the real-world impact. “I would just like to mention that it’s not just theoretical work,” he said. “The ENLIGHT technology has already saved the lives of patients in advanced stages of disease by offering them treatment options that would otherwise not have been considered.”

Beker believes the tech signals a new era in precision oncology — one where widespread testing and personalized treatment matching become the standard of care. “I really think it heralds a new era in precision oncology where everyone could get easily tested and matched with the most effective therapy,” he added.

Developing digital pathology-based biomarkers

“This new approach actually allows us to develop digital pathology-based biomarkers for any drug that is properly characterized in terms of its mechanism of action,” Beker said. This capability could inform the use and development of existing and novel cancer therapies, ultimately supporting the use of more effective treatment strategies and improved patient outcomes.

“Perhaps the most important thing is that this method allows one to get around the key problem that plagues both digital pathology and other AI technologies for response prediction,” Beker said. “Normally to predict response, you need matched datasets of pretreatment features and post-treatment outcomes. That’s the hardest type of data to get.”

A promise to make oncology R&D spending more efficient

By eliminating the need for expensive and time-consuming genetic sequencing, ENLIGHT-DP could significantly reduce R&D costs and accelerate time-to-market for novel cancer therapies. This accessibility was a key driver in the technology’s development. “Why this was compelling was evident,” says Beker. “It’s just the level of success that was surprising.”

The speed at which ENLIGHT-DP can analyze standard pathology slides could potentially offer results significantly faster than the 4–6 week turnaround time for next-generation sequencing.

Validated across six cancer types, four treatments, and five independent patient cohorts, ENLIGHT-DP offers a versatile solution for diverse oncology needs. This broad applicability makes it particularly unique in the field. “I’m not aware currently of any other response prediction technology in oncology that has shown positive results on such a wide array of cancer types as well as different drugs,” Beker says.

---

Filed Under: Uncategorized
Tagged With: cancer treatment prediction, deep learning, gene expression prediction, Pangea Biomed, personalized therapy matching, precision oncology, tumor image analysis
 

---