Using AI to profile gene expression from cancer imaging

An artificial intelligence (AI) tool has been developed to analyse images of tumours and predict changes in gene expression within cancer cells.

Cancer is a complex disease, which often requires expensive genetic testing in order to prescribe the most appropriate treatment. To avoid these costly methods, a team from Stanford University, California, have created an AI tool that uses stained images of tumour biopsies to predict changes in gene activity in cancer cells.

'This kind of software could be used to quickly identify gene signatures in patients' tumours, speeding up clinical decision-making and saving the healthcare system thousands of dollars,' said Dr Olivier Gevaert, senior author of this study and associate professor of biomedical data science at Stanford University.

Published in Nature Communications, the AI-based tool called SEQUOIA was trained using deep learning. This training involved providing the model with 7584 tumour biopsy images from 16 different types of cancer, paired with known gene expression profiles. Through this process, SEQUOIA learned to recognise common patterns within the data. Subsequently, SEQUOIA could accurately predict gene expression profiles in new tumour samples across two independent cohorts, including seven different types of cancers across six tissues.

Development of an AI-based tool, such as SEQUOIA, takes time and requires continuous optimisation to ensure accuracy and reliability. Dr Gevaert said, 'It took a number of iterations of the model for it to get to the point where we were happy with the performance.'

Understanding gene expression profiles within a tumour is important, since the activation or silencing of specific genes can influence tumour behaviour. These changes impact the response of the tumour to certain chemotherapies, as well as its likelihood of spreading to other parts of the body. While other image-based tools for identifying gene expression profiles in cancer exist, SEQUOIA has demonstrated greater accuracy than current methods, providing more reliable predictions of changes in tumour gene activity.

Currently, the Food and Drug Administration uses the MammaPrint test for breast cancer. This sequences breast cancer tumours to assess gene expression and provides a 'risk score', indicating the likelihood of cancer recurrence. However, this sequencing-based approach is expensive. The team showed that SEQUOIA's image-based scores closely matched those of MammaPrint. Additionally, SEQUOIA's predicted scores aligned with clinical outcomes, accurately identifying high-risk patients with higher recurrence rates according to their gene expression profile.

In the future, the researchers hope to further improve SEQUOIA by training it on a larger number of samples from a wider variety of cancers. Dr Gevaert said, 'We've shown how useful this could be for breast cancer and we can now use it for all cancers and look at any gene signature that is out there.'