Mobile Menu

Predicting biomarkers for immunotherapy response using machine learning

The complex nature of the tumour microenvironment poses a great challenge for the extraction of biomarkers of immune response and immunotherapy efficacy. To overcome this challenge, a recent study used RNA-sequencing (RNA-seq) data to derive sequence-based signatures of the tumour microenvironment. The researchers used machine learning to extract mechanistic biomarkers of antitumour immune responses that effectively predict patient response to immune-checkpoint blockers.


In recent years, immunotherapy has revolutionised cancer treatment. Immune-checkpoint blockers (ICBs) boost the patient’s immune system to effectively recognise and attack cancerous cells. Patients treated with immune-based therapies have shown promising results, especially in terms of long-term patient survival and curative potential. However, a major drawback is that only a minority of patients achieve complete response. Additionally, high immunological toxicity and considerable costs per patient are limiting the use of immune-checkpoint blockers. Therefore, it is important for clinicians to identify biomarkers, sparing unnecessary and potentially harmful treatments in patients that are unlikely to respond to ICBs.

The tumour microenvironment and biomarkers for immunotherapy

Different mechanisms in the tumour microenvironment (TME) are involved in mediating the immune response and affect the efficacy of ICB therapies. The first important aspect of the TME is the cell type composition. Different types of TME cells can have a pro- or antitumour role in regulating cancer progression and response to treatment. Another important aspect is the inter- and intracellular regulation of cellular functions that are responsible for shaping the anticancer immune response. These aspects should be taken into account when providing a comprehensive description of the TME.

A holistic approach to deriving biomarkers of immune response can inform clinicians on the efficacy of ICB treatments in patients. Different emerging omics technologies have allowed researchers to take snapshots of the TME in bulk tumours, in single cells and from images of tumour slides. The combination of these tools with computational approaches has the potential to provide a complete picture of the TME, which can shed light on how complex cellular and intercellular mechanisms orchestrate the immune response.

However, these technologies are not yet widely available, and the accompanying computational tools are still in their infancy and have shown difficulty in training the models. For precision medicine to be improved, researchers urgently need different approaches to derive a comprehensive description of the TME and how it regulates the immune response in individual patients, using available patient data.

Using RNA-seq data to predict biomarkers for immunotherapy

In this study, the researchers used an approach based on RNA-seq data combined with different types of prior knowledge to derive a holistic description of the tumour microenvironment. The team searched the TME using the Bayesian efficient multiple-kernel learning model and RNA-seq data from 7,550 cancer patients across 18 solid cancers contained within The Cancer Genome Atlas (TCGA). RNA-seq datasets are publicly available, but the information on which patients respond to ICB therapy is only available for a small subset of patients and cancer types, therefore this their method aimed to overcome this data problem. The researchers chose several substitute immune responses from the same datasets to be used as an indicator of the effectiveness of ICBs. By using substitute immune responses in the training process the researchers were able to overcome the issues with applying computational models for biomarker identification in immunotherapy.

They used the machine learning model to look for associations between the derived system-based features and the immune response, estimated using 14 predictors (proxies) that were obtained from recent publications. The proxies were considered as different tasks to be predicted by their machine learning model and the researchers used multi-task learning algorithms to learn all tasks jointly. An overall description of the approach can be seen in figure 1, which was taken directly from the article by Lapuente-Santana et al.

Significantly, the researchers found that their machine learning model outperforms biomarkers currently used in the clinical setting to assess ICB treatments, demonstrating that their method could be used to successfully assess patients that would respond well to immunotherapy treatments.

Figure 1. Overview of the researcher’s approach. (A) The first step of their approach involved deriving five systems-based signatures of the TME based on the integration of RNA-seq data and different sources of prior knowledge. (B) Next, the researchers produced a cancer-specific median correlation of each of the 10 scores of immune response with all 14 other scores. (C) Finally, the cancer specific models were trained on TCGA data. This produced systems-based signatures of the TME and scores of immune responses were derived by combining RNA-seq data and prior publications. These were used respectively as algorithm inputs and outputs. The trained models were used to define system biomarkers of immune response.


This study used computational models to assess RNA-seq data, from a large dataset of patients and tumour subtypes, to predict biomarkers for immunotherapy response. Their model was shown to outperform biomarkers that are currently being used by clinicians to assess immunotherapy response, demonstrating the success and utility of their approach. Moreover, the researchers hope to utilise emerging single-cell methodologies, which will allow them to look at the TME from a different angle to provide complementary insights into intra- and extracellular interactions, which could help them to adapt their framework further.

Image credit: kjpargeter – Freepik

More on these topics

Cancer / Immunotherapy / Machine Learning

Share this article