Skip to main content
Premium Trial:

Request an Annual Quote

Deep Learning Framework Uses T-Cell Analysis for Early Cancer Detection

NEW YORK — Taking advantage of the immune system's response to cancer at the beginning stages of the disease, a team of UT Southwestern Medical Center scientists have developed a computational framework for early cancer detection based on changes in T-cell receptor populations.

In a study appearing in Science Translational Medicine on Wednesday, the researchers demonstrate the method's potential with a variety of different cancers and suggest that it could improve diagnostic accuracy when used in conjunction with other screening methods.

Advances in single-cell and high-throughput sequencing have enabled new techniques for cancer diagnosis including ones based on cell-free DNA (cfDNA) and circulating tumor cells in blood, yet these technologies have their limitations, the scientists wrote.

For instance, "cfDNA-based methods rely on preselected panels of cancer somatic mutations, and the identification of circulating tumor cells usually relies on a few epithelial biomarkers or cell morphological changes, which might be subjective and nonspecific," they noted. Recent data also suggest that most mutations observed in plasma cfDNA are derived from white blood cells rather than cancer, raising questions about specificity.

Because the adaptive immune system responds early to tumor antigens, changes in a patient's T-cell repertoire would likely undergo cancer-specific changes during tumor progression. However, most cancer antigens are unknown, the identification of cancer-associated T cells remains difficult, and there is no diagnostic method to monitor signals in the T-cell repertoire, the researchers noted. Cancer-associated T-cell receptors (csTCRs), however, may share common biochemical signatures allowing for their identification.

To that end, the team developed a deep convolutional neural network-based method — dubbed Deep CNN Model for Cancer-Associated TCRs, or DeepCAT — for the de novo prediction of caTCRs in blood samples. They trained the method by applying an algorithm for detecting tumor-infiltrating T-cell CDR3 sequences using RNA sequencing data to roughly 4,200 tumor RNA-seq samples in The Cancer Genome Atlas covering 32 cancer types.

DeepCAT was validated using cancer-specific or non-cancer TCRs obtained from multiple major histocompatibility complex I multimer-sorting experiments, demonstrating prediction power for TCRs specific to cancer antigens. The scientists then tested it on blood TCR-seq samples from 13 clinical studies, including eight cohorts of patients with early- or late-stage cancer and five cohorts of healthy or virus-infected individuals.

The scores generated by DeepCAT was able to identify cancer patients with near-perfect accuracy for breast pancreatic, ovarian, colorectal cancers, and melanoma. DeepCAT showed poorer performance for patients with glioblastoma, bladder, and lung cancer, but the scientists noted that patients in these cohorts had all undergone multiple pretreatments, which may have depleted proliferating lymphocytes in the immune system, lowered the content of effector T cells in the blood repertoire, and altered the estimation. "The blood samples collected without prior treatment, including the breast, pancreatic, and ovarian cancer cohorts, consistently yielded higher cancer scores," they wrote.

In terms of early cancer detection, DeepCAT was tested using blood samples from four independent cohorts of patients with treatment-naïve early-stage kidney, ovarian, pancreatic, and lung cancers and demonstrated superior sensitivity and specificity to current blood-based biomarkers. It also had the same, if not higher, accuracy as methods based on cfDNA, methylation, or circulating tumor DNA mutations.

The UT Southwestern investigators conceded that DeepCat has several limitations including the inability to determine the tissue of a cancer's origin and the possible impact of chronic inflammatory conditions on cancer score estimations, although these issues might be overcome through further refinement.

"The cancer score is not intended to replace the current diagnostic methods at this time," they concluded. "Rather, future efforts should be made to explore whether the combined use of the cancer score with existing screening modalities, such as breast mammogram, lung CT scan, gastrointestinal endoscopy, or pelvic ultrasonography, can improve diagnostic accuracy in patients."