It’s mid-July and blazing hot here in Massachusetts. Luckily, we’re entering the summer vacation season and a break from the seemingly endless stream of biotech and oncology conferences that began in earnest last September, culminating in June with ASCO and EHA. We may also see a pause in the rush of biotech IPOs. There is always an interesting dynamic at play between progress as reported at medical conferences and the attractiveness of stock offerings – and this year the energy firing that dynamic is a bit unusual, especially across the immuno-oncology (IO) landscape.
As has been widely reported, ASCO was disappointing for next generation IO players. ‘Next generation’ refers to those companies developing assets that target diverse and novel immune regulatory targets, beyond anti-CTLA4 and anti-PD-(L)-1 antibodies. Essentially all companies bringing forward IPOs in the IO space have next-gen aspirations.
Why was ASCO disappointing? In part the answer is obvious: numerous expert reviews (eg. ours, from 2015: https://www.nature.com/articles/nrd4591) had promoted the compelling story that IO combinations would further improve the treatment of ever more patients in even more cancer indications. Such hypotheses drove intense investment in biotech companies, and the development of novel drugs targeting the diverse pathways of interest. During 2017 and 2018, IO combo hypotheses began reading out in clinical trial data, and the early results were underwhelming. This data wave culminated at ASCO in June.
The examples have by now been widely discussed: the collapse of the IDO inhibitor class with failures in late stage studies, the early defeat of an agonist anti-ICOS antibody, the realization that none of the many agonist antibodies to TNF superfamily receptors (4-1BB, OX40, GITR, CD27, etc.) were going to be quick wins, the modest activity of anti-CSF1R antibody, a miss from the VISTA program, and so on.
But the data were disappointing in a distinctly different way that impacts all programs, whether in trials or not. Simply put, we are due for a reality check regarding the quality of evidence used to support novel IO combination hypotheses. There are three (at least) aspects to consider:
1) IO animal models have limited utility. I think we all want to see better in vivo models of IO efficacy, and not just the same old PD-1 or PD-L1 combos in the usual mouse models. The usual mouse models are those that are partially responsive to anti-CTLA4, anti-PD-1 or anti-PD-L1, eg. tumors that grow out from the cell lines MC38, CT26, B16 and a few others when they are implanted into a mouse subcutaneously. Figure 1 is an example of the type of results we often see.
Figure 1. Activity of an IDO inhibitor in combination with anti-PD-1.
A few things to note are the modest IDO monotherapy activity (diamonds), and the modest improvement over anti-PD-L1 alone when the combination is given (circles), the fact that this is only a delay in tumor growth kinetics (by a few days at best) and a small survival advantage (2/15 mice in the combo arm survived). This study was conducted by top tier investigators with superb reputations (Spranger, Gajewsky and colleagues, the paper is at doi: 10.1186/2051-1426-2-3) – this was the best result they could get, and it is considered a positive outcome. This is the reality of traditional IO combo studies in mice: the fact is that these models have very limited value for predicting success in the clinical trial setting. The alternatives – PDX models and perhaps 3D-culture models – can be difficult to source and expensive to run.
By the way, the IDO inhibitor story may not be dead yet, as post-hoc analyses suggest that the leading programs (from NewLink and Incyte) had sub-optimal exposure and saturation of target. Also, novel compounds with a different mode of action (heme-displacement) are in development. So, we’ll see.
2) Correlation of target expression with survival data has limited value. This is an interesting area for dissection. The general argument put forward is that expression of some target of interest is associated with better or worse outcome in one or more cancer indications. The word ‘associated’ gives away the issue, as we are talking here about the correlation of expression data and patient outcome. The CSF1R story was born in such data – first the observation that myeloid cells, notable TAMS and MDSCs, were associated with poor prognosis in a large number of cancers, then the observation that CSF1R expression was also correlated with outcome, as was expression of the ligand for CSF1R, M-CSF. Here is an example of these kind of data, in this case from lung cancer patients, published earlier this year (doi:10.1038/s41598-017-18796-8):
Figure 2. CSF1R levels and overall survival in lung cancer
If you read the methods for this paper – and I’m just using this as an example, not to pick on this particular study – what you’ll see right away is that the measurement of analytes (M-CSF and CSF1R) was performed using a single biopsy or surgical resection sample. There are no longitudinal data (successive samples taken during treatment). As we learned in the old days of transcript profiling of autoimmune disease, longitudinal data are essential – the Type 1 interferon story in lupus being a poster child of the pitfalls encountered in non-longitudinal analysis.
The myeloid cell data are even more interesting – here is an example looking at myeloid clusters in NSCLC patients (smokers) and correlation with survival (from doi: 10.1371/journal.pone.0065121):
Figure 3. Myeloid cell clusters and survival
There are a lot of these types of studies in many different cancer indications. So now you have a pretty clear narrative: CSF1R+ myeloid cells present in tumors are associated with poor outcome for those patients. That’s the observation. The hypotheses that flow from the observation are various: anti-tumor efficacy will result from depleting tumor-resident myeloid cells, or from interfering with this myeloid cell accumulation, or from changing the phenotype of the myeloid cell population in the tumor. The first hypothesis was tested with the anti-CSF1R antibody cabiralizumab which potently depletes all myeloid cells across a range of tumor indications but has no to limited efficacy either as monotherapy or in IO combination therapy. The second and third hypotheses were tested in glioma using a CSF1R inhibitor, PLX3397, thought to prevent myeloid cell accumulation and polarize myeloid cells to a pro-inflammatory phenotype. This drug did not show efficacy (doi: 10.1093/neuonc/nov245).
As with the IDO story, the CSF1R story is not yet fully told, and the early data in at least one tumor type (pancreatic cancer) was interesting. But the near-term conclusion is that the foundational observation (myeloid cells in tumors are associated with poor prognosis) did not produce successful hypotheses regarding CSF1R. Why not? The answer again lies in the way these analyses are done – using single point surgical resection or biopsy tissue to set the level of the analyte. This is not a criticism: there are no other readily available samples, unless the disease continues to disseminate (ie. there are metastases). But here’s the rub: metastases are themselves an indicator of advancing disease and cannot be used as an independent measure. And this raises the larger point: when the variable is survival, many parameters will correlate with outcome, and it’s very hard indeed to distinguish cause from effect – which bring us to biomarkers.
3) Good IO biomarkers are elusive. Anyone following IO has seen the confusion regarding PD-L1 expression as a biomarker of response to anti-PD-1 and anti-PD-L1 antibody therapies. Enough said. But is PD-L1 expression a uniquely complex example of biomarker use? Or is it representative of the level of background noise seen in the complex immune response to cancer that IO therapies attempt to exploit? I favor the latter assessment. A quick definition of ‘biomarker’ as used here: measurement of an analyte that is hypothesized to be associated with response to therapy. A biomarker can be used as inclusion criteria in clinical trial enrollment, or as an exploratory endpoint to assess clinical responses. The depletion of myeloid cells in the anti-CSF1R trial mentioned above is one example. A textbook attempt at the use of a biomarker in a novel IO combo trial is seen in the early development of JTX-2011, the agonist anti-ICOS antibody. The indications examined in the ICONIC clinical trial were chosen based on ICOS mRNA expression. Expression on infiltrating T cells in tumors was analyzed across ~30 cancer indications and confirmed using immunohistochemistry. Based on the frequency of high ICOS expression, ICONIC enrolled patients with lung cancer (NSCLC), Head & Neck cancer, triple negative breast cancer, gastric cancer, and melanoma. Of note, most of these indications are known to be responsive to immune checkpoint therapy with the anti-PD-(L)-1 class of antibodies, and some respond to the anti-CTLA4 classes of antibodies. In “ICOS-high” indications the rationale for increasing the overall response rate to therapy by adding an agonist anti-ICOS antibody to an immune checkpoint blocking antibody (anti-PD-1 or anti-CTLA4) looked pretty straightforward. The ICONIC trial was presented at ASCO to near universal disappointment as the response data were not higher than immunotherapy alone. As to the biomarker data, that becomes hard to judge in the absence of a clear efficacy signal, but Jounce made the best of it, suggesting that ICOS expression “appears to associate with anti-tumor activity” which is a pretty soft statement. As with IDO and CSF1R, the anti-ICOS story is not done, and more clinical data will follow. However, the clinical hypothesis that anti-ICOS combination immunotherapy with anti-PD(L)-1 or anti-CTLA4 therapies would increase patient response rates was not supported in this early study.
So what to make of all this? The response to the negative news flow has been predictable, with some of these companies losing significant share value. Several large pharmaceutical companies have taken the opportunity to declare a sharpened focus: Novartis’ Jay Bradner, for example, stated that the company will only develop new therapeutics that demonstrate monotherapy efficacy preclinically (back to those mouse models, but it’s a step), and Astra Zeneca’s Pascal Soriot has declared the company will target patients earlier in disease.
It is interesting to consider the sheer number of ongoing combination trials, a number that is tracked in real time by Beacon Intelligence (beacon-intelligence.com). As of last week the numbers looked like this:
Figure 4. IO combination clinical trials
It’s a huge number of ongoing trials. There are lots of interesting therapeutics to read out in the near term, including such classes as STING agonists, the adenosine pathway inhibitors (A2AR, A2AR/A2BR, CD39 and CD73), anti-CD47 antibodies, arginase inhibitors and on and on. Given the sheer number of trials it is timely to wonder if we are just blindly pushing ahead. It is certain that we are not learning all of the available lessons – one only has to look at the (investor) excitement generated by data from very early trials of cytokines (IL-2, IL-10) and TLR agonists. Some of these agents will certainly fail in later stage clinical trials.
The brings us back to where we started and the notion that there is an odd energy dynamic between the ongoing reporting of clinical trial results and the march of immuno-oncology IPOs to market. Given the recent spate of negative news, why do IO IPOs continue to draw investor interest?
Let’s take two examples from the recent IPO pool.
Scholar Rock is a company with a completely novel asset platform targeting the TGF beta family of proteins – this IPO was very well received even though the company is early in clinical execution, especially in IO (the lead program is for muscle wasting diseases). TGF beta is emerging as a critical biologic signal of resistance to immunotherapy. Right upfront it is important to note that the caveats in assigning causality to resistance pathways are similar to assigning causality to association with survival – these are findings based on single source pretreatment cancer biopsy or resection samples, and the patient outcome is followed over time (but the analytes are not typically followed longitudinally). But TGF beta is a different kind of beast, notable for its’ roles in immunosuppression and wound healing, ie. resolving inflammation in support of tissue remodeling. The basic biology lends itself to exploitation by cancer cells as they build an immunosuppressive microenvironment and remodel that environment, often leading to aggressive growth and metastasis. TGF beta activity also triggers TGF beta expression in a feed-forward loop. Surface Oncology, who came to market with an anti-CD47 antibody in Phase 1 (neither novel nor first) and an anti-CD73 antibody program (again neither novel nor first) was also well received – in large measure because their anti-IL-27 program is highly novel and associated with signaling through the PBAF complex – which has in turn been linked to the efficacy of T cell mediated anti-tumor immunity in studies from diverse labs.
Companies like these bring forward a new wave of therapeutic hypotheses to test, and that is attractive in an environment where the old wave of therapeutic hypotheses appear to be stumbling. Of course, the investment hypothesis, at least in the long term, is that they will succeed.
Two more comments: 1) re-examining and critiquing lines of evidence in IO drug development should be an iterative process, and 2) this is a process that can be applied to other therapeutic modalities – oncolytic viral therapy development and cellular therapy for solid tumors, to name just two.