Predictive biomarkers in the age of spatial biology

Blood samples

Listen to this article on the DDW Podcast:

By Ori Zelichov MD, VP Clinical Development at Nucleai.

Immunotherapies, namely immune checkpoint inhibitors (ICI), have revolutionised cancer treatment providing durable clinical benefits for patients with metastatic solid tumours. While ICIs are now considered a backbone treatment in many cancer types, only a small subset of patients (20-30%) benefit and respond to those drugs, highlighting the need for better predictive biomarkers for patient selection. Compared with targeted therapies, which are guided by the identification of the target mutation in next-generation sequencing (NGS), PD-1 and PD-L1 inhibitors are most commonly selected using the quantification of PD-L1 expression on biopsy slides. While NGS provides a clear binary result, PD-L1 scoring is a continuous and subjective measurement. The difference between these two biomarkers, as well as the distinct mechanisms of actions of targeted therapies and immunotherapies, creates different response dynamics of the two drugs: Patients positive to NGS-biomarkers have high response rates to targeted therapies but only for a short time. However, patients positive to PD-L1 tend to have low response rates to immunotherapies but see longer benefits from the drug. The lack of accurate predictive biomarkers for immune therapies is a substantial pain for both cancer patients and oncologists. Furthermore, immuno-oncology (IO) costs $200K per year per patient so billions are spent on drugs that don’t work. For pharma companies developing the next generation of ICI drugs, 95% of trials will fail without good biomarkers to enroll only patients with high likelihood of response. How can we identify better predictive biomarkers for immunotherapy that will allow patients to benefit from high response rates that are also durable over time?

The new generation of biomarkers

It is clear today that the analysis of the immune system and the tumour microenvironment (TME) are critical for predicting a response to ICI and that the tumour-centric approach that makes targeted therapies successful is not sufficient for immunotherapies. The birth of precision oncology can be related to the introduction of next-generation sequencing (NGS) technologies and the approval of targeted therapies for subsets of patients with certain genomic alterations. To date, more than 40 targeted therapies were approved by the FDA with companion diagnostics that are based on genomic testing. As a result, genomics has become the holy grail of personalised medicine. Naturally, with the introduction of immunotherapy drugs, 20 years later, the scientific community had hoped to find the predictive biomarkers for these drugs using the same methods of gene sequencing. However, it turned out that, as opposed to targeted therapies, response to immunotherapy drugs is not affected by the tumour’s genomic profile but by the maturity and contexture of the immune system. As immunotherapies don’t work directly on the tumour but by unleashing the immune system, it makes sense that the analysis of the immune system and the TME are critical for predicting response to ICI. The TME consists of the interplay between individual cells in the tumour, immune cells, blood vessels, fibroblasts and the extracellular matrix. The density, type, location and even the organisation of immune cells within solid tumours were shown to be highly important in predicting clinical outcome in patients treated with immunotherapies.

Biomarkers that integrate immune cell and TME analysis

Several peer-reviewed papers were published in 2021 and provided relevant and interesting observations on the importance and potential clinical benefits of biomarkers that integrate immune cell and TME analysis. In the Journal of Thoracic Oncology, M Shiwasawa et al.1 presented a score that combines PD-L1 and tumour-infiltrating lymphocyte (TIL) density and showed that it was better correlated with ICI response than PD-L1 scoring alone. The score was retrospectively analysed on 228 Non-Small Cell Lung Cancer patients treated with ICI. In Nature Cancer, L Vanhersecke et al.2 showed the association between the presence of mature tertiary lymphoid structures (TLS) and ICI efficacy cross-cancer. In a retrospective study of 328 patients treated with IO for different cancer types, the presence of mature TLS’ was associated with improved objective response rates, progression-free survival, and overall survival, independent of PD-L1 expression status and CD8+ T cell density. In another study, Eli Lilly 3 presented outcome data from the Phase III ORIENT-11 study on its PD-1 drug, Sintilimab, in NSCLC patients. While the efficacy of the drug was high when combined with chemo, the researchers also tested the predictive role of TILs (analysed by RNAseq) in predicting response in the study. What they found was that high or medium immune cell infiltration was strongly associated with improved PFS in the combo group, in contrast to absent or low immune cell infiltration. While these papers are promising, TME analysis is complex and the discovery and deployment of new TME-based markers for immunotherapy are difficult. How can one analyse the complex TME data from tumours? While genomic testing can identify biomarkers for targeted therapies and Immunohistochemistry (IHC) is used for PD-L1 scoring for immunotherapy prediction, TME analysis needs another approach.

Deep learning

The best source for TME data is tumour biopsies that contain the cellular profile as well as the spatial distribution of tumour cells and immune cells in the tumour and its environment. But analysing pathology slides is not simple – DNA, RNA, proteins, radiology, and patient medical records are all digitised and represented in bits. Computational analysis of this data is being utilised for biomarker discovery, as well as drug development, and clinical trial design. Pathology tissue biopsies, on the other hand, are analogue by nature and therefore, underutilised in medical research today. As a result, we have robust insights on how diseases behave at a molecular level, but still lack the understanding of how they work on a spatial level. Extracting and calculating all of these data points from a biopsy is not achievable by the human eye. Computer vision algorithms on the other hand, can accurately quantify the number of different cell types in the biopsy, their locations and calculate the interaction patterns between them. More importantly, by integrating this spatial analysis with the corresponding clinical outcome, deep learning can identify subtle patterns that are associated with response to a certain treatment and can uncover novel visual prognostic or predictive biomarkers. If we could only analyse biopsies using computers, the spatial characterisation of tissues would complement genomics, proteomics, and radionics data and produce the next generation of biomarkers, as well as diagnostics and drugs. The concept of AI spatial analysis is promising in biomarker discovery but AI algorithms are only as powerful as the data behind them. Machine learning can really only start unlocking new information that can help develop drugs, diagnose patients and determine the right course of treatment for each patient when large and high-quality biology data sources are accessed. Fortunately, as tumour biopsies are taken as a routine in the process of cancer diagnosis, there are millions of (analogue) glass slides that are underutilised in the pharma R&D and can be used for new discoveries by AI. Digital scanners can nowadays take analogue glass slides and turn them into digital images that can be analysed by computers. While there are many companies that apply AI algorithms on these images, they mostly try to mimic pathologists, and make their work faster and more accurate. For biomarker discovery, a much more detailed and complex analysis of both the images and clinical data are needed. In Nucleai, we have been able to digitise and analyse tens of thousands of slides to date, where each slide size can reach several gigabytes and contain hundreds of thousands of cells.

Hundreds of human interpretable features from each slide can be generated, instead of the simple information extracted from the slides today by the human eye. While this database is being used to discover novel biomarkers and drug targets, it also allows pharma companies to improve clinical trial design and develop better drugs. The data originates from laboratories and health systems around the world, creating a robust and heterogenous platform of image analysis models. The speed and generalisability of the models provide a catalogue of novel spatial insights. These findings, when combined with genomics and outcomes data, provide a necessary layer of information to the drug discovery and clinical trial processes. Incorporating spatial biology into drug development, though underutilised by pharma companies today, can deliver value and improve the efficacy of treatments for patients.

Disease prognosis and drug response

As mentioned, fusing spatial biology data with phenotype data is where the ‘magic’ happens. It allows us to understand how spatial biology affects disease prognosis and drug response. For example, in a retrospective analysis of 90 NSCLC patients treated with first-line single-agent pembrolizumab, we identified a spatial immune signature in H&E slides that could identify patients with durable clinical benefit. The interim results of this biomarker discovery study, conducted by Nucleai, showed the predictive power of spatial analysis of TILs in NSCLC. The novel predictive biomarker, composed of different histological features, such as the proximity between tumour cells and tumour-infiltrating lymphocytes (TILs), was validated on a cohort of 43 patients. Patients with positive Nucleai scores had a significantly higher median Overall Survival (OS) (Not reached vs.17.8m) and two-year OS (70.8% vs. 33%, p=0.02) than patients with negative Nucleai score. But these slides are the tip in the iceberg. There are millions of other glass slides that are buried in pathological labs and can be used for spatial biology research. Analysis of spatial information from pathology slides using AI algorithms is expected to revolutionise cancer research and bring the next generation immuno-oncology biomarkers and drugs to the clinic.

Volume 23, Issue 2 – Spring 2022

About the author

Ori Zelichov MD serves as Vice President of Clinical Development at Nucleai, where he leads the company’s scientific research and clinical studies. Prior to joining Nucleai, Zelichov was the Medical Director of Novellus, where he led product and clinical strategy for a novel targeted therapy, now being tested in clinical trials. Zelichov received his MD. from Ben Gurion University, completed his internship at Tel Aviv Medical Center, and graduated from the healthcare innovation program of MIT and Harvard Medical School.

References

1.https://www.jto.org/article/S1556-0864(21)02373-X/fulltext

2.https://www.nature.com/articles/s43018-021-00232-63.

3. https://www.jto.org/article/S1556-0864(21)02330-3/fulltext

Related Articles

Join FREE today and become a member
of Drug Discovery World

Membership includes:

  • Full access to the website including free and gated premium content in news, articles, business, regulatory, cancer research, intelligence and more.
  • Unlimited App access: current and archived digital issues of DDW magazine with search functionality, special in App only content and links to the latest industry news and information.
  • Weekly e-newsletter, a round-up of the most interesting and pertinent industry news and developments.
  • Whitepapers, eBooks and information from trusted third parties.
Join For Free