Matching Patients with Treatments: Developing Clinically Meaningful Diagnostics To Inform Cancer-Immunotherapy Selection

HTTPS://STOCK.ADOBE.COM

Immune-checkpoint inhibitors (ICIs) represent a promising approach to treating malignant cancers. For instance, monoclonal antibodies (mAbs) can be designed to disrupt interactions between programmed–cell-death receptor 1 (PD-1) proteins expressed on T cells and associated ligand 1 (PD-L1) proteins on the surfaces of tumor cells. Such interactions are known to activate immune cascades that result in T-cell dysfunction and exhaustion, enabling cancer cells to slow down T-cell activity. In theory, applying a PD-1 or PD-L1 inhibitor should reverse the immune checkpoint and “release the brake” on T-cell responses, promoting cancer destruction. PD-1/PD-L1 blockade forms the basis for seven US Food and Drug Administration (FDA)–approved biologics, and two other approved products apply the same principle to cytotoxic T-lymphocyte–associated protein 4 (CTLA-4) (1). Several related targets and modalities are under clinical evaluation, too (2, 3).

ICIs offer hope to patients as first-, second-, and later-line treatments for a breadth of indications, including Hodgkin lymphoma and skin, lung, renal, and bladder cancers. However, oncologists often have difficulty predicting which patients will respond to treatment. When administered as monotherapies for relapsed/recurrent conditions, PD-1 and PD-L1 inhibitors achieve different and sometimes lackluster objective response rates (ORRs) across indications: ~70% in Hodgkin lymphoma, ~40% in skin cancers, ~20% in lung cancers, ~25% in renal cancer, 13–23% in bladder cancer, and 13–16% in head and neck squamous-cell carcinoma (HNSCC) (3, 4). Moreover, patients who respond initially to ICIs sometimes experience relapse, with surviving cancer cells having acquired resistance mechanisms (3, 4).

Combined with the high costs of immunotherapy, such clinical outcomes underscore the need to improve in vitro diagnostics (IVDs) for patient stratification and treatment planning. In January 2024, Ofer Sharon, MD (chief executive officer of OncoHost), spoke with me about why available methods for estimating ICI suitability fail to provide the kind of information that would help oncologists to evaluate treatment options. He emphasized that researchers and IVD developers must change how they think about biomarkers for treatment suitability. Especially considering inconclusiveness about PD-1/PD-L1 expression levels as predictors of ICI treatment success (3–10), scientists must devise approaches that can account for the interplay of cancer, the immune system, and immunotherapy. OncoHost’s approach to managing such complexity is to leverage artificial intelligence and/or machine learning (AI/ML) for analysis of proteomic data from patient blood plasma. By evaluating differential protein-expression patterns and identifying biological processes associated with tumor resistance, the company’s diagnostic platform can provide oncologists with accurate, clinically useful information about different ICI-based approaches.

Sharon is a physician and entrepreneur with 20 years of experience in clinical and commercial product development for start-ups in the health-technology, biotechnology, and medical-device industries. Before joining OncoHost, he served as medical director for AstraZeneca in Israel, new-technologies scout for MedImmune, and medical director for Merck Sharp & Dohme (MSD) in Israel. He has cofounded several healthcare companies, focusing especially on bioinformatics and ML platforms for early intervention in disease progression. Sharon holds an MD from Tel Aviv University in Israel.

A Need for New Diagnostic Approaches

What problem with IVDs does OncoHost intend to address? Our goal is to change how scientists think about biomarkers for immunotherapy. Many companies develop antibodies as cancer immunotherapies. We seek to understand which patients will benefit from ICIs and which patients will have no benefit or experience adverse reactions.

Antibodies are complex in nature and participate in many biological signaling pathways. The traditional way to consider biomarkers for cancer treatment is “linear.” Let’s say that we have developed an antibody to PD-1 or PD-L1. The basic assumption has been that if a drug targets PD-L1, then measuring PD-L1 levels in a tissue sample will indicate whether a patient will benefit from that drug. The same logic has been applied to CTLA-4, lymphocyte-activation gene 3 (LAG-3) proteins, and immunoreceptor tyrosine-based inhibitory motif domain (TIGIT) proteins.

Unfortunately, researchers have discovered that such thinking is incorrect (5–10). The reason why is simple to understand but difficult to address. Immunotherapy primes a patient’s immune system, which is a complicated network of many cell types, each of which can have multiple kinds of interactions. For instance, consider how differently immune cells behave in normal tissues and tumors. All of those biological interactions influence whether a patient responds to treatment. It is oversimplistic to think that checking one biomarker will be enough to decipher all of that complexity.

Hence comes the concept of multicomponent or composite biomarkers, which OncoHost is exploring to enhance evaluation of immunotherapy success. Of course, the big question is how to integrate and analyze all of the data points that are required to develop a composite biomarker. We use ML. To understand the many complicated biological interactions involved in cancer treatment, we need mathematical tools that can ingest significant amounts of data and identify patterns in them.

What kinds of IVDs are available for PD-1/PD-L1 measurement and prediction of ICI success, and how effective are they? Five kits have received regulatory approval for such testing. They are developed by different companies and are not harmonized. All five are based on immunohistochemistry (IHC). Clinicians perform a biopsy to obtain a slice of tumor tissue. Then, a pathologist stains the sample and looks at the material under a microscope to count cells that express the biomarker.

Such tools have limited effectiveness. Today, clinicians generally do not use measurement of PD-L1 expression to decide whether a patient should receive ICIs. The reason is simple: PD-L1 expression, as measured by available tests, is not a predictive biomarker for treatment success (5–10). For some kits, accuracy levels are just above the level of a coin toss. Of course, we can treat PD-L1 expression as a prognostic biomarker rather than a predictive one because patients whose tumors express extremely high PD-L1 levels generally respond better to PD-L1 inhibitors than do patients whose tumors have extremely low levels. But we do not know where the cut-off lies as far as treatment success. When results show moderate expression levels, we cannot tell whether a patient will benefit from an ICI approach.

Machine Learning and Proteomic Pattern Recognition

How would you describe your company’s PROphet technology? The platform combines ML, proteomic analysis, systems biology, and bioinformatics to identify protein-expression patterns in samples of blood plasma. Initially, we developed the technology to estimate treatment outcomes for patients with non–small-cell lung cancer (NSCLC) (see the “Extending ML-Based Proteomic Analysis” box on the next page).

First, we collect a patient blood sample, either in clinic or by mobile phlebotomy. Our platform requires ~200 µL of plasma. A standard tube for blood collection holds about 4 mL of plasma, so a standard blood draw will yield much more than the platform needs. The collected material is sent to our centralized laboratory in North Carolina, US. The facility is US Clinical Laboratory Improvements Amendments (CLIA)–approved and US Commission on Office Laboratory Accreditation (COLA)–accredited. There, the plasma undergoes proteomic analysis using a SomaScan platform from SomaLogic. That system enables us to measure ~7000 types of protein per sample, ~1500 types of which are stable and repeatable enough for our analytical purposes.

Next, we apply our PROphet ML algorithm to identify expression patterns in the proteomic data, checking especially for ~400 resistance-associated proteins (RAPs). We have developed a classifier that analyzes the proteomic data alongside PD-1/PD-L1 biomarker information and clinical knowledge about patient outcomes. Thus, we can differentiate between patients who are likely and unlikely to respond to treatment with PD-L1 inhibitors, both as a monotherapy and alongside chemotherapy. After analysis, the algorithm provides a score for each RAP to indicate its predictive value for patient response to immunotherapy. Those individual scores are combined, and the final score informs clinicians about the optimal treatment approach. Patients who are likely to respond to ICIs as monotherapies or tandem treatments receive a positive report; patients who are unlikely to respond well to treatment receive a negative report.

The PROphet algorithm assigns each RAP a different weight. Some RAPs are more important and impactful than others are. We also have designed the algorithm to determine the source of a given protein — e.g., whether it comes from diseased or normal tissue or from an immune response. By focusing on RAPs that come from tumor and immune-system activity (of course, some overlaps manifest here), we can map biological-signaling pathways for a given patient. That enables prediction of treatment outcomes (11, 12) and risks for adverse events (13).

The PROphet system has exhibited high accuracy (>90%) and analytical validity over three blinded clinical studies with patient pools of increasing sizes (11, 12). When identifying patients with both positive PROphet reports and high PD-L1 levels (>50%), we determined that such patients would respond in the same way to a combination of immunotherapy and chemotherapy as to immunotherapy alone. Thus, for that group of patients, we could spare them from chemotherapy and its high costs, potential adverse events, and reduction in quality of life — without compromising on clinical efficacy. We also determined that patients who showed high PD-L1 expression levels but received a negative PROphet report would live about 3× longer if treated with a combination of immunotherapy and chemotherapy as opposed to immunotherapy alone.

Effectively, we use a sophisticated mathematical algorithm as a useful clinical tool. Currently, oncologists do not have meaningful ways to determine whether to prescribe immunotherapy alone or with chemotherapy. Our platform addresses that need.

Why does the PROphet platform use patient blood rather than another sample type? How does blood plasma provide the kind of information that you need for sophisticated proteomic analysis? Using blood plasma was among the first requirements for our research and development (R&D) team. My rationale is twofold. First, blood collection is easy on patients. It is accessible, noninvasive, and repeatable. Blood draws can be performed in every clinic with few errors by clinicians and minimal risk for patients. I believe that when developing innovative biotechnology, you always must consider how it will be used in a clinical setting. If your amazing assay accurately evaluates patient responses but requires biopsies, then clinicians cannot use it in daily practice.

My second reason is technical. Plasma is a “soup” with many ingredients. When you analyze it, you measure cells, proteins, and residual DNA from all over a patient’s body, not only from a tumor. The downside is that it is difficult to distinguish each “ingredient” and identify its source. Analysts are interested in tumors. If the protein being measured comes from a healthy cell rather than a tumor, it might not work in my algorithm. The upside, however, is that the mixture provides a holistic picture of interactions between a patient’s immune system and cancer, which is not always one tumor.

Consider patients with metastases. A primary tumor might manifest in lung tissue while others might present in the liver and brain. Those metastases are unlikely to be genetically and proteomically identical to the primary tumor. Different locations entail different biology. A biopsy of the primary tumor will provide useful data but will miss the story of tumor activity in the brain and liver. However, plasma will include proteins from all of those tissues. Using blood plasma serves as a “downstream-from-a-tumor” approach that is accessible, easy to integrate into a clinical workflow, and holistic.

What kinds of computational capabilities do you need to perform such evaluations? The OncoHost team uses many specialized mathematical models. Ideally, data scientists would apply AI/ML to identify patterns in information from large clinical cohorts or from extensive databases. If you have a million data points, then a good data scientist can identify a signal in that information. However, the OncoHost team needed to develop mathematical models to extract signals from small sets of patient data. You cannot draw blood from 1 million patients with metastatic lung cancer; you must obtain material from your own clinical trial or from a biobank, which usually contains hundreds rather than millions of samples. For our purposes, we had access to samples from ~2000 patients. So we needed a mathematical approach to isolate signals from small cohorts.

Better Matches of Patients and Therapies

How effectively do you think the IVD field is using ML and related tools? We are just at the beginning of such work. In precision medicine and precision oncology specifically, most available AI/ML tools are for image processing — e.g., for counting cells that overexpress a given protein. That is where the field’s most mature companies operate.

Few companies are working on prognostic tools for treatment planning. Our company applies ML to understand resistance mechanisms of cancer, using proteomics as a tool for clinical guidance. Biodesix similarly uses proteomics to predict clinical responses. What differentiates OncoHost, I believe, is our interest in both treatment-response prediction and clinical utility. In that sense, we more closely resemble companies that work in genomics, which seek specific mutations in patient gene sequences to guide treatment. Such tools are more advanced than are those for proteomics because genomics has had more time in basic research and in the market. But genomics is less relevant for immunotherapy-response prediction than it is for highly targeted therapies (e.g., gene therapies).

What else could developers of diagnostic tools do to help clinicians select therapies for their patients? Leaders must make several critical decisions when starting up a diagnostics company. First, you must choose an optimal direction for your business, with two possible pathways to follow. One path is to develop tools for pharmaceutical companies, creating technologies that will enhance patient stratification for clinical trials. That would help in identifying the right patients for a given treatment and improving study success rates. The second pathway is to develop diagnostics for oncologists and patients.

The latter pathway requires a distinct mindset. Clinicians want to treat their patients, and patients are seeking treatment. Certification and predictive value have different weight in that situation than they do in the pharmaceutical-company context. Even if a diagnostic predicts that the odds of a positive treatment outcome are 2%, the oncologist and patient might go ahead with treatment, especially if that approach is the last option. Two percent is greater than zero.

So if you are developing a tool for clinicians and patients, then you need to think differently about what you can do for them. You must think about clinical activity from the beginning, considering how and why an oncologist would use your technology. The “why” question still raises other questions — e.g., what clinical benefit would your diagnostic provide to physicians, patients, and (in some regions) insurance companies? You must account for all such elements. Otherwise, your tool could be extremely accurate, but nobody will use it.

The industry needs many more companies that develop clinically useful diagnostics for prediction of treatment outcomes. Using current approaches to clinical-trial design, a significant number of studies fail because sponsors have not identified the right patients for treatment. And if your loved one is diagnosed with cancer, then you just want to find the most effective treatment possible. I believe that companies and investors need to place more focus on matching patients to treatments — at least at the same level that they set for identifying targets and developing drugs.

References

1 Shiravand Y, et al. Immune Checkpoint Inhibitors in Cancer Therapy. Curr. Oncol. 29(5) 2022: 3044–3060; https://doi.org/10.3390/curroncol29050247.

2 Liu J, et al. PD-1/PD-L1 Checkpoint Inhibitors in Tumor Immunotherapy. Front. Pharmacol. 12, 2021: e731798; https://doi.org/10.3389/fphar.2021.731798.

3 Xu-Monette ZY, et al. PD-1/PD-L1 Blockade: Have We Found the Key To Unleash the Antitumor Immune Response? Front. Immunol. 8, 2017: 1597; https://doi.org/10.3389/fimmu.2017.01597.

4 Pitt JM, et al. Resistance Mechanisms to Immune-Checkpoint Blockade in Cancer: Tumor-Intrinsic and -Extrinsic Factors. Immunity 44, 2016: 1255–1269; https://doi.org/10.1016/j.immuni.2016.06.001.

5 Shen X, Zhao B. Efficacy of PD-1 or PD-L1 Inhibitors and PD-L1 Expression Status in Cancer: Meta-Analysis. BMJ 362, 2018: k3529; https://doi.org/10.1136/bmj.k3529.

6 Bravaccini S, Bronte G, Ulivi P. TMB in NSCLC: A Broken Dream? Int. J. Mol. Sci. 22, 2021: 6536; https://doi.org/10.3390/ijms22126536.

7 Havel JJ, Chowell D, Chan TA. The Evolving Landscape of Biomarkers for Checkpoint Inhibitor Immunotherapy. Nat. Rev. Cancer 19, 2019: 133–150; https://doi.org/10.1038/s41568-019-0116-x.

8 Indini A, Rijavec E, Grossi F. Circulating Biomarkers of Response and Toxicity of Immunotherapy in Advanced Non-Small Cell Lung Cancer (NSCLC): A Comprehensive Review. Cancers (Basel) 13, 2021: 1794; https://doi.org/10.3390/cancers13081794.

9 Gibney GT, Weiner LM, Atkins MB. Predictive Biomarkers for Checkpoint Inhibitor-Based Immunotherapy. The Lancet Oncol. 17(12) 2016: e542–E551; https://doi.org/10.1016/S1470-2045(16)30406-5.

10 Jørgensen JT. Companion Diagnostic Assays for PD-1/PD-L1 Checkpoint Inhibitors in NSCLC. Exp. Rev. Mol. Diagn. 16(2) 2016: 131–133; https://doi.org/10.1586/14737159.2016.1117389.

11 NCT04056247. Predicting Responsiveness in Oncology Patients Based on Host Response Evaluation During Anti-Cancer Treatments (PROPHETIC). ClinicalTrials.gov 29 August 2023; https://clinicaltrials.gov/study/NCT04056247.

12 Christopoulos P, et al. Plasma Proteome–Based Test for First-Line Treatment Selection in Metastatic Non–Small Cell Lung Cancer. JCO Precision Oncol. 8, 2024: 555; https://doi.org/10.1200/PO.23.00555.

13 Naidoo J, et al. Pre-Treatment Plasma Proteomics-Based Predictive Biomarkers for Immune Related Adverse Events in Non-Small Cell Lung Cancer. BMJ J. ImmunoTher. Cancer 11(s1) 2023: 1229; https://doi.org/10.1136/jitc-2023-SITC2023.1229.

Brian Gazaille, PhD, is managing editor of BioProcess International, part of Informa Connect Life Sciences; [email protected]; 1-212-600-3594. Ofer Sharon, MD, is a physician and chief executive officer of OncoHost Ltd., Hamelacha Street 17, Floor 1, Binyamina-Giv’at Ada, Israel; https://oncohost.com.