Predictive model developed to improve accuracy of blood-based comprehensive genomic profiling

A research group led by Professor Hidenori Kage from the Graduate School of Medicine, the University of Tokyo and Assistant Professor Hiroaki Ikushima from the University of Tokyo Hospital has announced the development of a machine learning model that estimates, prior to testing, the individual probability of success for blood-based comprehensive genomic profiling (CGP) in pancreatic cancer patients. The model has been implemented and made publicly available as a web application (https://pancreasliquidcgp.streamlit.app/). It enables patients and clinicians to consider when to use CGP-which can only be performed once in a lifetime under the current insurance system in Japan-and what to use as the sample. The findings are expected to contribute to more effective treatment of pancreatic cancer. The results were published in ESMO Open on February 12.

Pancreatic cancer is known to be one of the most rapidly progressing cancers and is generally resistant to conventional chemotherapy. Molecularly targeted therapies, which have fewer side effects, have come into wider use; however, most of these drugs are only effective against specific gene alterations, making it necessary to test for the presence of such alterations. However, CGP (one of the available tests) can only be used once in a lifetime under the current insurance system.

CGP can be performed either by directly collecting tumor tissue from the pancreas (or a metastatic organ) and analyzing the cancer genomic information within, or by detecting cancer-related gene alterations that have leaked from tumor cells into the bloodstream via a blood test-a method known as liquid biopsy.

Although the blood-based approach is minimally invasive, drawing blood at a time when an insufficient amount of tumor DNA has leaked from the cancer cells into the bloodstream can result in an inconclusive test.

The research group therefore set out to develop a machine learning model that in advance predicts the probability that blood-based CGP will yield a successful result. Using clinical data from 2,220 pancreatic cancer patients who underwent CGP between August 2021 and December 2023, the group built and trained a machine learning model to estimate the likelihood that tumor-derived DNA circulating in the blood could be detected by blood-based CGP.

The training data was drawn from the Center for Cancer Genomics and Advanced Therapeutics (C-CAT) at the National Cancer Center, which collects clinical and cancer genomic information from cancer patients across Japan.

The predictive model the group developed demonstrated good accuracy, and is expected to increase opportunities to provide patients with optimally tailored pancreatic cancer treatment.

Ikushima commented: "CGP, which is essential for recommending the most suitable cancer treatment for each individual patient, can only be performed once in a lifetime under Japan's current insurance system. By using the model we have established to consider when and how best to carry out that precious single opportunity, we hope to be able to deliver more effective cancer treatments with fewer side effects to more patients than ever before."

Journal Information
Publication: ESMO Open
Title: Prediction model for ctDNA detectability in liquid comprehensive genomic profiling of advanced pancreatic cancer based on Japanese real-world data
DOI: 10.1016/j.esmoop.2026.106069