Niigata University scientists have developed an AI tool based on histopathological characteristics of tumour tissues to predict tumour mutational burden in patients with colorectal cancer and determine the efficacy of immunotherapy.
In the field of oncology, biomarkers are critical indicators of patient responsiveness to specific drugs and therapeutic approaches. In recent years, several emerging candidate biomarkers have shown promising results in identifying the efficacy of immune checkpoint inhibitors (ICIs), a class of immunotherapy that is especially effective against colorectal cancer (CRC). These biomarkers include programmed death-1 protein ligand (PD-L1), density of tumour-infiltrating lymphocytes (TILs), and tumour mutational burden (TMB).
Among the three, an elevated level of TMB (TMB-H), which refers to high numbers of somatic mutations per coding area of a tumour genome in cancer cells, is known to be a predictive biomarker for ICI therapies for various solid cancers. TMB-H is usually determined through gene panel testing. However, in clinical practice, it is not feasible to perform gene panel testing for all cancer patients as it is expensive, time-consuming, and not easily available.
To simplify the process of detecting TMB-H, a group of scientists from Niigata University explored the histopathological characteristics of TMB-H in CRC patients and developed a convolutional neural network-based algorithm that is capable of predicting TMB-H CRC directly from tumour tissues stained with haematoxylin and eosin (H&E). Their model shows promising benefits to reduce the burden of diagnostics on pathologists, and speed up the process of gathering the necessary information needed to determine the patients’ responsiveness to ICIs, serving as a promising alternative to gene panel tests.
Dr. Yoshifumi Shimada and colleagues first gathered data from two CRC cohorts, the Japanese CRC (JP-CRC) and The Cancer Genome Atlas (TCGA) CRC cohorts, both of which showed elevated levels of TMB. Using data from the JP-CRC, they examined the histopathological characteristics of TMB-H CRC tissues that have been stained with H&E and discovered that TILs are significantly associated with TMB-H CRC.
They later combined their findings with the TCGA CRC cohort to develop their convolutional neural network-based TMB-H prediction model. Their Convoluted Neural Network model employs the Inception V3 learning model and was built by transforming and normalising digital information of neoplastic and non-neoplastic images of TMB-H colorectal cancer tumour tissues from the JP-CRC-cohort.
“We have developed artificial intelligence to predict genetic alterations in colorectal cancer by deep learning using haematoxylin and eosin slides. This artificial intelligence is important in solving the cost problems associated with genetic analysis and facilitating personalised medicine in colorectal cancer,” said Dr. Shimada.
Besides their novel Convoluted Neural Network model, they also devised a way to predict TMB-H CRC by only using TIL information derived from H&E slides of tumour tissues. However, because the individuals involved in this study were not patients treated with ICIs, they were unable to draw definite conclusions about how well patients respond to ICI therapies after TMB-H diagnosis. Furthermore, additional studies need to be done to validate whether or not TIL can be used as a predictive biomarker to measure the efficacy of ICIs.
Nevertheless, their findings are expected to provide a cost- and time-saving alternative to gene panel testing, potentially helping clinicians to readily identify CRC patients who demonstrate TMB-H and determine whether they can benefit from ICI therapies. [APBN]
Source: Shimada et al. (2021). Histopathological characteristics and artificial intelligence for predicting tumour mutational burden-high colorectal cancer. Journal of gastroenterology, 56(6), 547-559.