And other thoughts on treating cancer!

1           Philosophical Preamble

I wrote the first version of this article when looking forward to visiting my elderly mother who turned 87 in January 2017 (more about that later). She remembers when she was at school one of her friends was dying of cancer. At the time her family bemoaned the doctor who told them that if their daughter had only had this disease some ten years later (circa 1950) she would be cured. My mother recalls how, in those days, cancer was viewed as a disease which, by the time her children were born, would be consigned to the history books in terms of its lethality.

Well her youngest child, at the time of writing, had just turned 50 and as Imagen Therapeutics develops techniques to help improve the prognosis of cancer sufferers, I am painfully aware of the misplaced optimism of the 1950s. Yet this optimism has continued unabated through the decades. Cancer research charities are always trumpeting bullish messages such as “together we shall beat cancer”. In Manchester one of the campaigns in 2017 had a picture of the news reporter Jon Snow looking over his glasses with the caption “Now you can be a rebel with a cause”. Cancer, in this context, is presented as an evil wraith that the general public can rally against by handing over cash to fuel intensive research into its ultimate destruction.

There has of course been some wonderful progress made in relation to treating certain cancers. What is also encouraging is that modern-targeted therapies, which have been developed by the pharmaceutical industry, can sometimes work miraculously: bringing those who would have had weeks to live a quality extension to their lives which sometimes can last many months or even years. Yet these remarkable successes are so often isolated events or outliers within large clinical trials which, overall, are deemed as failures. Given that cancer has been so extensively studied over so many years, and has had literally billions of dollars thrown at it, a valid question to ask is what has gone wrong? Why has a research effort of lesser proportions brought us laptops with more power than supercomputers of the 1990s, yet a woman diagnosed with ovarian cancer stage III/IV will most likely undergo rounds of cytotoxic therapy which will make her feel ill and at best buy her perhaps an extra 6 to 18 months?

A possible answer to this question might lie in an underlying paradigm that permeates the discipline of biology. Many have read about the classic struggle Galileo had with the medieval church over the geocentric model of the universe. While this battle is often framed in terms of scientific reason triumphing over primitive superstition, the truth is a little more nuanced. In science a dominate paradigm acts like an interpretive lens making the experimental data fit consistently within the assumptions of the paradigm itself. In relation to the geocentric model, it was possible to explain the complex motion of the planets by introducing a deferent-and-epicycle model which effectively could account for the changing speeds of orbits and even their change in motion which occurs during the earth’s yearly cycle (Geocentric Model). Furthermore, the adherence to the geocentric model was not just based on poor theology but appeared to be supported by experimental data. For example, if the earth did move, then it was argued that one ought to be able to observe the shifting of the fixed stars due to stellar parallax. While the ancient Greeks understood that an alternative hypothesis, for the lack of stellar parallax, could be a moving earth in a universe where the stars were much further away than the planets, they rejected this idea on the basis that the distances required were unlikely to be possible. Finally, and somewhat unfortunately, the apparent constant luminosity from Venus appeared to support a geocentric model rather than a heliocentric one.

It was Thomas Kuhn who first coined the idea of a paradigm crisis where science itself interprets data within a theoretical framework often de-emphasizing information that potentially calls the theoretical framework into question. This process can continue for many years until finally there is enough contrary data and predictive failure to force “a crisis” in the theoretical framework. At that point it sometimes happens that a new framework is proposed which can more completely explain the experimental data. At this point the scientific community adopts the new framework or paradigm leading to what is known as “a paradigm shift”. I realise I am touching on the deeper philosophical debate around whether the Kuhnian model entails scientific relativism but for the purposes of this article it does not matter. The point simply is that the operation of science, in relation to experimental data, is always in the context of an interpretive framework.

So, in relation to cell biology, what is the paradigm that I argue is in crisis? It is simply the assumption that biological systems are “easy” to understand. At this point many will take issue arguing that of course no one in their right mind thinks that biological systems are simple to comprehend. Yet while we may pay lip service to the idea of biological complexity, we behave as though we really don’t take it all that seriously. (The catch phrase really should be “it isn’t biology” rather than “it isn’t rocket science”)

The problem of underestimating biological complexity is not new. Humanity has consistently made this error for centuries. For instance, even as late as the beginning of the 19th century, a great many scientists still believed in spontaneous generation. It seemed so obvious that life could arise from non-living material like the spontaneous generation of maggots from rotting meat. Louis Pasteur’s early work demonstrating that wine could not be produced in sterilized containers, drew stern criticism from his fellow scientist Pouchet. This led to the French Academia of Sciences offering a price of 2,500 francs to whoever could conceive of an experiment that produced conclusive evidence either for or against the theory of spontaneous generation. Pasteur won the prize when he demonstrated that nothing grew in sterilized broths that were enclosed in containers which had filter caps preventing the entry of bacteria. Pasteur wrote:

“Never will the doctrine of spontaneous generation recover from the mortal blow of this simple experiment. There is no known circumstance in which it can be confirmed that microscopic beings came into the world without germs, without parents similar to themselves.”

Yet fast forward less than 50 years to the Origin of the Species and the paradigm of simplistic biology still held sway. At the time of Darwin, the problem of life’s origin was not the molecular machinery of living organisms (this was at the time completely unknown) but rather the macro complexity of anatomical structures such as the human eye. So, the genius of Darwin’s work was that it gave the scientific community a powerful theory to explain the complexity of living things as they currently understood biological complexity; which was in terms of their current understanding of animal physiology.

Today we have uncovered two chemical languages within living entities. The first is RNA which is translated into the second which is protein. While we have also managed to identify many of the molecular machines, the proteins that constitutes the equipment of life, we still struggle to understand how these entities interrelate in the cell as a whole. We have succeeded at creating incredibly complex wiring diagrams of protein networks, yet our comprehension of the network as a whole is still tenuous.

As always when something is beyond us, we tend to fall into the trap of framing every problem in terms of what we do understand. This is a little bit like the man who comes out of a pub to find another man searching intently for something on the ground. The first man turns to the other and says “Hey mate, have you lost something?” The other replies, “As a matter of fact I have, I have lost my watch.” “Don’t worry”, says the first man, “I will help you look for it.” After 20 minutes of both men carefully searching the area the first man turns to the second and asks: “Hey mate, are you sure you lost your watch here because we have been looking for over quarter of an hour and have covered every inch of ground and we still can’t find it.” “No”, says the other man, “I lost it 5 miles down the road.” Exasperated, the first man protests: “Then why are we looking for it here?” The second man straightens and with a look of complete surprise says: “Because the light is better here of course!”

2           Limitations of using Genetic Analysis to Personalise Cancer Therapy

Experimentally we have had real success with DNA sequencing and so have been able to retrieve the blueprints of many living things including ourselves. Yet in doing so we now have the propensity to try and understand cell biology purely in terms of the genetic blue print even though DNA is two levels removed from the main cellular machinery – the proteins. This genecentric way of thinking is particularly strong in academic cancer research because we understand cancer to be a genetic disease. While there may ultimately be a genetic model for cancer, where the behaviour of any particular malignancy will be explained by the unique genetic changes inherent within it - interacting with the unique genetics for that particular patient, the difficulty of developing such a model to personalize therapy is severely underestimated by many biologists and clinicians alike. For instance, many scientists assume that the linking of complex genetic signatures to drug selection is a problem of “big data”. The technique often employed in these situations is to build computer-based neural networks (the heart of Artificial Intelligence: AI) where genetic signatures are correlated against clinical outcome data. These data form what is known as a training set where the output weightings of the many nodes of the neural network are optimized to the point that it can correlate a particular input signature with a particular clinical outcome. While such approaches have merit, and can be particularly powerful, they become much more difficult to implement as the complexity of the input data increases. In cancer, the amount of genetic variation between each individual, overlaid with the number of mutational changes that are common to any individual’s cancer, means that the input data is extremely high dimensional, possessing many degrees of freedom. This makes it very difficult to find a true global minimum in the network’s cost function and so the network parameters remain poorly tuned even with a relatively large clinical outcome data training set. Thus, the common approach of using some form of black-box algorithm to convert genetic signatures into definitive clinical predictions may never be fully realized because the amount of genetic change will vastly exceed the number of clinical observations required to adequately train the model. What is even worse is if one aims specifically to create a computer model to help choose which patients should receive which treatments then by definition the clinical training set must contain a large number of examples where patients have been treated with all the possible drug therapies that could be considered for cancer treatment. Yet most cancers are treated with relatively few therapies so clinical outcome data set, on which to train an AI model, simply does not exist.

However, at this point it is important to stress that I am not arguing that we should abandon genetic research. As always, the picture is never that black and white. Firstly, there are great examples, where specific mutations do strongly correlate with a particular drug treatment and, with further research, we are bound to find more. Secondly, related to the paradigm argument this article has been discussing, another big mistake is to keep thinking we can second-guess what research will be useful in terms of what we ultimately fund. I am sure that even if the current bold NHS project to sequence 100,000 human genomes (Large NHS Study) does not fully realise the personalized medicine dream, it will still produce lots of useful data. The issue I have is that academics (especially those who review grant applications) should not put “all their eggs” in the genomic basket. Given we still have a lot to learn about living systems, it surely makes sense to take a multifaceted approach to the problem of personalized chemotherapy.

3           Limitations of using Biomarker Analysis to Personalize Cancer Therapy

Likewise, large pharmaceutical companies also do not fare much better. Their speciality is the creation of chemical compounds that can interact with cellular molecular targets: the most common type of interaction being a chemical entity that disables the function of its target protein. Given this is probably one of the sharpest tools large pharma possesses, it is not too surprising that a very psychologically attractive model for cancer research has been the idea that cancer cells possess oncogenic addiction, which means they will contain an Achilles heel in relation to drug treatment. The idea is that there will be a particular molecular target, which if blocked, will deny the cancer the signals it requires to undergo its dysregulated growth. Yet how does one find the molecular target? Well the answer at first appears deceptively simple. Certainly, with breast and testicular cancer, it is known that blocking sex hormone receptors will often effectively treat these cancer types. It was therefore argued that a good target for oncogenic addiction would be a signal transduction protein that is up-regulated in the particular cancer under study. One classic example is the overexpression of Epidermal Growth Factor Receptor (EGFR) that occurs in about 60% of non-small cell lung carcinoma (Hirsch et al., 2003). Drug companies like AstraZeneca and others produced very specific inhibitors for this receptor and then trialled these agents in non-small cell lung carcinoma. The good news was that the oncogenic addiction model did turn out to be correct: the bad news is it was correct in only about 5-8% of patients, not the 60% of those over-expressing EGFR (Thatcher et al., 2005).

Enter the discipline of System Biology. System biologists tend to be engineers who have become bored with solid state electronics. When they see the complex signal transduction pathways of a cancer cell, they naturally compare it to the wiring diagram of a modern computer. The learning from man-made electronics is that the behaviour of the overall system cannot be easily reduced to any single electronic component. Rather the components themselves integrate into the emergent properties of the system as a whole. Applying this analogy to the signal transduction systems of cancer cells is helpful because it suggests that the idea of hitting a single protein node to abrogate malignancy will be, in many cases, a flawed concept. Thus, the discipline of Systems Biology potentially explains why EGFR inhibitor only work in a small number of lung cancer sufferers. The problem is that while 60% of non-small cell lung carcinomas might over-express the EGFR target, it is a much smaller percentage where the configuration of the pathway is such that the cell is oncogenically addicted to the survival signals from the EGFR protein.

Yet this finding also spells difficulty for those who want to personalize chemotherapy based on the over-expression of a target biomarker. No-one would now use the expression level of EGFR in non-small cell lung carcinoma as a selection criterion for EGFR inhibitor treatment. Yet there is one high-profile biomarker example which is used as a standard test for the stratification of patients in relation to whether they receive the drug Herceptin. Herceptin is a very expensive therapeutic antibody that blocks a homologue of EGFR known as ErbB2 or Her2. Given the ferocious price tag of this treatment, National Institute of Clinical Excellence, United Kingdom (NICE) recommends that women are only put on Herceptin if their cancer cells overexpress Her2 or have Her2 gene amplification (Nice Guidelines). This seems entirely sensible because how could a cancer, that is not overexpressing the molecular target, possibly be differentially sensitive to an agent that targets it? Yet in breast cancer it turns out that about 35% of women who overexpress Her2 will respond to Herceptin (Vogel et al., 2002). This means that the protein biomarker test has a false positive rate of approximately 65% some of which can be explained by limitations in the immunocytochemistry assay (Vogel, 2010). There is also literature which demonstrates that some women who do not overexpress Her2 can respond to Herceptin (Paik et al., 2008). So even in a protein biomarker test that has been adopted by the National Health Service, United Kingdom (NHS), the false positive rate is approximately 65% (only 35% of Her2 positive women will respond to Herceptin) and the false negative rate is not zero. Much of the literature that discusses “central HER2-negative” cancer response to Herceptin proposes that the extra effect is due to some interaction of the therapeutic antibody with the host immune system, independent of the antibody’s molecular target (Nuti et al., 2011). While this hypothesis may be valid, another plausible and simple option is that so-called “central HER2-negative” Herceptin-responsive cancer cells, while having low levels of HER2, are exquisitely sensitive to the survival signals their Her2 receptors provide because the protein is tightly linked to the main survival pathway of the cancer cells. Thus, a small level of protein expression, which will be scored on an immunohistochemistry slide by the technician as either negative or a single “+” is enough to make the cells responsive to Herceptin when this protein signal is abrogated.

4           Limitation of PDX Mouse Models (Avatars)

Another exciting development in the world of personalized chemotherapy has been the xenografting of patient cancer tissue into immunocompromised mice. These mice are poetically known as Avatars and the basic idea is to create enough of them to treat with several different chemotherapies to see which one is most effective. The big attraction of PDX mouse models is the cancer is being placed back into an in vivo setting where it also gets to interact fully with stromal cells. Therefore, most in the field agree that testing chemotherapy in this setting is by far the closest system to testing the drugs in the patient. This is why xenograft mice are favoured by the pharmaceutical industry for determining whether a particular drug is going to work in a particular patient who has been recruited into a clinical trial. There is no doubt that these types of approaches offer real promise and will play an important role both in cancer research and possible future personalized or stratified chemotherapy. This is especially so given the development of new immunotherapies. While traditional PDX mice cannot assess these therapies, as they do not have a functioning immune system, it should be possible in the future create transgenic mice which recapitulate the patient’s own immune system. Such models potentially could become very important, especially if immunotherapy becomes more common in cancer treatment.

However, PDX suffers from two important limitations. The first is cost. As discussed in section 6, page 10, it is actually important that a test which aims to personalize chemotherapy is able to simultaneously examine a large number of available therapies. This is very difficult using PDX because each mouse can only be treated with a single treatment. Therefore, if one wanted to test even 10 different therapies that would require at least 30 mice (assuming each drug is tested in triplicate). Also, often it is important to test treatments at different doses which requires even more mice. Given generating and treating these mice is extremely expensive, the concept of using PDX mouse in a general healthcare system simply is not financially plausible. Even many private patients are likely to have to curtail the number of avatars that are produced forcing the PDX mouse assessment to examine just a few selected therapies.

Another issue is that it takes at least 4 – 6 months (Information on PDX, Champions Oncology) to get a result from a PDX mouse experiment. Unfortunately, many patients with aggressive cancers don’t have that long to live so by the time the PDX experiment is completed, the patient may very well have died or their cancer progressed to the point where they are not strong enough to tolerate further chemotherapy.

5           High Content Analysis

Walk the corridors of many modern pharma and you will discover that early drug discovery is trying to introduce phenotypic screens to sit alongside more traditional assays. The key technology for phenotypic cell screening is automated fluorescence microscopy or High Content Analysis (HCA). Returning to paradigms for a moment, even here we have some interesting history in terms of technological blind spots. When I was doing my PhD in the 90s the focus, in terms of microscopy, was producing better and better images. Universities were often submitting large grants to secure confocal microscopes and later 2 photon confocal microscopes. Yet in all that excitement, we biologists may have missed the obvious and that was there was a ton of information in a fluorescence microscope image which was simply not being mined. The problem was that images were analysed manually by lab technicians or PhD students. While the human visual system is powerful at image recognition, our minds are poor at consistently extracting data from a large number of complex images. If we consider a cell, which has been stained with several different fluorophores, the image contains a wealth of information relating to many different biological systems. If this information is to be extracted from only a single image, then such a task can be completed by a well-trained biologist in a matter of hours. Alternatively, if one has a fluorescence images with a single stain (as a PhD student I had to count thousands of cell nuclei stained with the dye acridine orange and determine which ones were apoptotic), then it is also possible to get a lab scientist to spend an afternoon scoring a single feature from a dozen microscope slides or so. What becomes exceedingly difficult is when one wants to simultaneously capture the information in a multi-stained fluorescent microscopy experiment which extends over thousands of images. Even an experimentalist, who is game for such a Herculean task, will not be able to perform their analysis over a long period of time with the required analytical consistency. Also, the experimentalist can only judge staining intensities on a very crude and relative scale (a good example being the plus scale scores used by technicians determining the intensity of immuno-staining).

In contrast, microscopy images, which are captured using a CCD camera and converted into pixel-based images, can be directly accessed by computer software and numerical values applied to the staining intensity profiles. Additionally, because it is relatively easy to develop automated image analysis software which can correctly assign Region of Interest (ROIs) to cell images, the automated analysis of cell images is highly amenable to modern computing. Hence the creation of High Content Screening machines, which consist of a good quality fluorescent microscope linked to a computer so that computer can simultaneously control the movement of the microscope stage, image acquisition and image analysis, results in a new platform technology that offers something unique in relation to the tools we have to dissect cellular biology.

In mathematical terms this means that HCA allows us to perform experiments which have both a large independent and dependent variable space. The independent variable space are all the experimental manipulations that are performed as part of the experiment. Large pharma love techniques that offer a large independent variable space because they often want to screen a large number of chemical entities at multiple concentrations in order to mine chemical libraries for potential new drug candidates. The dependent variable space is all the measurements that can be extracted from the experiment. As discussed above, because HCA is extracting numerical information from fluorescence microscopy images, the number of measurements that can be returned by the computer platform can, in some cases, exceed well over 20 variables. Because HCA is a platform technology, it is possible to configure it for many different types of biological experiment. While it is commonly used to examine cellular signalling processes, it is also possible to design experiments which are focussing on overall cellular behaviour rather than the inner working of a cell signalling pathway. HCA is very amenable to these types of experiment because contained within the images is cellular morphology data which is directly influenced by the functions of the intracellular protein pathways.

Imagen Therapeutics was originally called Imagen Biotech and we started trading in 2007 as a company that provided HCA services to large pharma and small biotech. In order to remain distinctive from our competitors, we tended to focus on developing phenotypic assays which could be used by large pharma to help them understand their research molecules more effectively. The application of a phenotypic screen can offer important information which is not apparent using traditional signal transduction assays alone. For example, two chemical entities which have been shown to act on the same protein target might diverge in the phenotypic data they produce if their off-target secondary interactions are different. Also, a chemical entity might have been shown in a biological assay to strongly inhibit a protein target in vitro, but when the same entity is applied to a whole cellular system it might fail for a variety of reasons. Spotting this failure early in the drug discovery process helps pharma quickly eliminate molecules which ultimately will not make it as viable future drugs.

While phenotypic assays can be complex (two phenotypic assays we support measure cell stickiness and the production of neutrophil nets respectively. Both assays can be used to study inflammation independent of the signal transduction pathways which govern this process), there are also examples of relatively simple ones. Probably the simplest phenotypic assay is the measurement of cell death because the morphological change in cells when they die is easy to detect experimentally, especially using fluorescence microscopy. Cell death is a very important measure when it comes to cancer research because the aim of cancer therapy is to kill cancer cells while leaving the host cells unharmed. Given its central importance, it is not surprising that assays aimed at assessing cell death were used before the birth of HCA. A very common assay used by large pharma is known as the MTT assay. The MTT assay measures the activity of NADH-dependent oxidoreductase cellular enzymes which, under defined conditions, reflect the number of cells present. Treatment of a cell population with a compound which is either cytotoxic or cytostatic, will result in fewer cells in the treated population than an equivalent control and this decrease in cell number will be detected by less conversion of colourless MTT to formazan which is purple in colour. While this assay can be operated in a high throughput manner, it is not nearly as robust as its HCA equivalent because it is not directly measuring cell death. Rather it is measuring a surrogate for cell number and this distinction, as will be discussed later in this paper, is important.

6           Early Attempts to Personalise Chemotherapy Using Chemosensitivity Assays

It is not at all surprising that before the birth of large scale genetic screening and the isolation of a cancer biomarkers, an early attempt at personalising chemotherapy was attempted by applying assays such as MTT to cancer samples taken directly from the patient. The use of these assays peaked in the 1990s and despite some positive results, the overall conclusion was that they did not significantly improve patient outcomes when compared to patients whose treatment was determined solely by the oncologist.

Another common criticism levelled at these early attempts to personalize cancer drug treatment was that the patient’s cancer cells will not respond the same way in the laboratory test as they do in the patient. This has led to a significant number of oncology specialists being dismissive of trying to apply modern HCA to the personalization of chemotherapy as they see it as simply a rehash of the work that was done in the 20th century. However, a large ASCO review (Schrag et al., 2004) showed that in many of the studies the patients’ response rates were significantly greater in the assay-directed versus the physicians own choice group. The authors’ main criticism of this early body of work was the inability to design properly controlled randomised clinical trials. While this critique has some validity, it needs to be recognised that the demonstration of the potential worth of a diagnostic assay does not easily fit the paradigm of a classical clinical trial which has been developed in order to test whether a new drug has utility at treating a particular medical condition. For example, even a trial that simply randomly assigns the patients to either an assay-directed or physician’s choice intervention group suffers the difficult challenge that a confounding variable in these studies is how resistant the test tumour is to treatment in general. It is easy to demonstrate that the chances the clinician will select a treatment that is beneficial to the patient is inversely proportional to how resistant the tumour is to drug therapy in general. This also applies to any potential in vitro test.

Yet conceptually, it makes most sense to apply a personalisation strategy to cancers which carry a poor prognosis. Yet poor prognosis entails that the standard set of therapies, normally employed against the cancer in question, are not working. Hence in highly resistant cancers, where early attempts at personalisation are often attempted, you are asking of your clinician and of any chemosensitivity assay to determine which is the best therapy to choose out of a total set of therapies which do not work. There can obviously be no statistical power in such a study design. The ASCO study alluded to this problem in comparing the historical context of early chemosensitivity assays to the situation at the present time. It argued that while the authors felt there was of lack of conclusive evidence in these early trials, it recognised that these early studies were hampered by a relatively small number of treatment options. It concluded that now, more than ever, it is important to try and personalize cancer therapy, given the rise in signal transduction-targeted therapies that, in general, work spectacularly well in a minority of patients. In short, a personalized chemotherapy diagnostic is only as good as the therapies available to treat a particular malignancy. If the cancer is highly resistant to treatment then a personalised medicine assay, even if perfect, will offer little positive clinical benefit. However, in this case, it will be useful at eliminating all the therapies that will not work for the patient and which, by their cytotoxic nature, will make the general condition of the patient worse.

Additionally, in cancers where there is a mixture of patients, some quite responsive to standard therapies while others resistant to treatments, it is important that the overall treatment sensitivity index of these patients is matched across the clinical trial groups. For instance, if the assay-directed group contain a higher proportion of more difficult to treat patients then the probability that the assay-directed group shows no additional benefit (or is even negative) over the physicians’ own treatment choice group is high. Conversely, if the assay-directed group contain more patients who are intrinsically easier to treat, then it is likely that the beneficial effect of the in vitro assay will be exaggerated.

Another interesting question is how good does an in vitro assay have to be before it has a significant effect on survival curves in a clinical trial? Obviously, no laboratory test is perfect and the two measures used to evaluate a diagnostic assay are specificity and sensitivity. These two measures directly relate to the assay’s false positive and negative rate respectively. An assay which demonstrates either a low false positive or false negative rate will be useful in cancer treatment. If the assay has a low false positive rate then it means that when a drug does work in the assay it is most likely to work in the patient. This is true even if the assay’s overall sensitivity is low (assay has a high false negative rate). Similarly, an assay which has a very low false negative rate would be valuable at eliminating therapies that will definitely not benefit the patient even if its false positive rate means it is not reliable enough to suggest an alternative treatment. Obviously, the most desirable assay is one that has both a low false negative and positive rate which then means the assay is very specific and sensitive. What is fascinating is that while many critique the early work in the 1990s for not designing proper clinical trials, it is also noteworthy that even if these trials had of clearly defined a chemosensitivity assay’s specificity and sensitivity, there is no real consensus on how good these parameters must be before the assay is considered clinically useful. As discussed in section 3, page 6, the current Her 2 test for herceptin carries a false positive rate of 65% yet it is still a standard test used by the NHS. Ultimately, an assay which has either a specificity or sensitivity higher than chance (50%) will add value to no test but the size of the clinical trial required to demonstrate utility will become larger as the assay’s specificity and sensitivity approach randomness. Likewise, if it is known in advance that an assay’s discriminatory power is in just a single parameter (either sensitivity or specificity) then the design of the clinical trial to test the utility of the assay will be different depending on which parameter one is trying to assess. Obviously, it would not be sensible to select a patient’s treatment based on assay data if it is already known that an assay’s false positive rate is above 30% but it would be right to eliminate treatments if the same assay’s false negative rate was less than 10% (this is the very situation in relation to the Her2 test for the selection of breast cancer treatment).

7           The application of an HCA Cell Death Assay to Personalize Cancer Treatment

Unlike the chemosensitivity assays of the 1990s, we measure cell death directly by assessing the loss of cell membrane integrity (indicative of secondary necrosis) using HCA. We can to do this with cells growing in 2D monolayers or 3D spheroids. Our largest study is in advanced ovarian cancer culturing patient cancer cells that have been derived from ascites fluid drains.

Consistent with the poor prognosis of this condition, we found that our assay predicted that most patients would not respond to the standard chemotherapies that are used to treat this disease (Figure 1). The implication of this finding, in relation to designing a clinical trial to test the utility of our assay has already been discussed in the previous section.

The unequivocal nature of our in vitro assay, when the drugs are ineffective, is particularly useful to us because it means we can with confidence conclude when a drug has some anticancer activity in our in vitro assay. This also means it is relatively easy for us to partition assay data into complete negative responses (like the majority of those shown in Figure 1) versus dose response curves where the drug is clearly giving an in vitro response (Figure 2).

7.1        What does a response in our Assay Mean?

The first challenge to address when a positive response is observed in an in vitro assay is could such a response simply be due to an in vitro assay artefact? It is possible to address this question if the assay has been used to test a significant number of similar patients. In the extreme that a drug response is simply a function of an in vitro assay artefact, then all patient data relating to that drug should show similar positive responses. Alternatively, if one observes data similar to that for bortezomib in Figure 2, then it is much more likely that the response observed in the assay does correlate intrinsically to the sensitivity of the patient to the drug in question because different patient cells are giving different responses.

While this observation clearly helps us spot true assay false positive artefacts, it does not help us in deciding what response amongst those observed in the assay will correlate to a clinically useful response for the patient. For example, in Figure 2 perhaps all responses from the black curve above will respond to bortezomib or perhaps only patients who display a response similar to the green curve will be the ones that will benefit from bortezomib treatment.

At Imagen Therapeutics we have developed a method for dealing with this challenge. In many ways we are borrowing a concept from genetic analysis and that is to contextualise the output data in relation to the whole patient population. In relation to dose response curves, it is possible to collapse a dose response curve into a measurement which allows it to be ordered against other dose curves that have been obtained from other patients.

7.2        How to interpret the Calibrated Report Data

From this ordering it is then possible to determine where a patient’s response falls in relation to all the other patients that have been tested. This becomes useful when there is clinical trial response data for a given drug against a specific cancer. For example, if clinical trial data suggests that drug X is only effective in 10% of all patients treated, and the Imagen Therapeutic report indicates that the response data for drug X observed is consistent with it being in the top 30%, then one would not consider prescribing drug X to the patient for which the Imagen Therapeutic report was generated. Alternatively, if another drug (drug Y) is shown to work in a quarter of all cancers and the Imagen Therapeutics report data demonstrates that the in vitro dose response data for Drug Y corresponds to a dose response profile in the top 10% of all patients with a similar cancer type, then it is likely that drug Y will work for that particular patient on which the test has been performed.

The obvious problem presents when there is no clinical trial data for a particular drug on the cancer under investigation. However even in this scenario, one can still use the calibrated data, along with clinical judgement, to help reach an overall treatment strategy. For example, if the drug under consideration is a modern targeted therapy, and it is known in other cancer types to be effective only in approximately 10% of cases, then if our test scores the drug response in the top 2%, the data would be consistent with the hypothesis that this targeted-therapy will be efficacious in the test patient. Alternatively, if the drug response was only in the top 15% of the cancer training set then it would be much more prudent to choose either an alternative targeted therapy or even a standard cytotoxic; especially if the cytotoxic therapies are also showing activity in our test.

8           Bortezomib Revisited

It is clear from our initial work that bortezomib gives very different responses in different patients (Figure 2). Depending on what turns out to be a meaningful in vitro response, the responses in Figure 2 could suggest that only 2 out of 6 patients will have a strong response to this drug and at least one patient (red curve) out of the 6 will definitely receive no benefit.

What is fascinating is when this drug was tested in a 1st phase clinical trial; out of 15 patients, two of them had complete responses and 5 of them had partial responses, as judged by the RECIST scoring, giving an overall response rate of 47% (Aghajanian et al., 2005). Yet despite this promising start, a phase 2 clinical trial, this time using bortezomib in combination with doxorubicin, failed to reach statistical significance leading the authors to conclude that this treatment regimen was not useful in ovarian cancer (Parma et al., 2012). In this second study 24% of patients in the platinum sensitive group achieved an objective response with a median duration of 4.8 months. Disease stability was achieved in 30% of patients in both the platinum sensitive and platinum resistant groups. While these data showed that some patients were responding to the treatment combination, the overall benefit, compared to standard of care, was marginal when the patient population was taken as a whole. The authors’ conclusion was not to recommend further studies with bortezomib in combination with doxorubicin for the treatment of ovarian cancer. In both these studies however, bortezomib was never tested as a single agent but combined with a traditional chemotherapy.

The general approach of combining a modern chemotherapy with a traditional one is scientifically deeply problematic. Patient survival data is a population measure so even if, in any given individual, only one of the drugs in the combination is working, the combination must always do better than the monotherapy, when it comes to Kaplan Meier survival analysis, because more patients in the combination population will be getting at least one correct treatment than patients on the traditional cytotoxic alone. In other words, it is impossible to de-convolve whether the combination group is surviving longer because of the combination or because of each monotherapy working in different patients within the combination group. Yet with no test that adequately segments patients based on their response to individual therapies, pharmaceutical companies and oncologists are understandably unwilling to test modern chemo-drugs in monotherapy mode because when the modern therapies do not work they offer no clinical benefit. The unfortunate consequence however is that our overall understanding of how these therapies will work in responsive patients is not developing because these experiments, in general, are not being conducted in monotherapy mode.

The justification of combining bortezomib with a traditional cytotoxic in the phase 2 clinical trial was based around stabilizing I-kB. However, it should be noted that the original idea of blocking protein degradation was that it would target cellular processes that required rapidly changing levels in protein expression such as the cyclins during the cell cycle. Combining bortezomib with a cytotoxic, which is likely to effectively slows cell cycle progression, could mean that any positive benefits of bortezomib are reduced because the cell’s requirement for rapid cyclin degradation becomes less as the rate of cell cycle progression is decreased in the cancer cells. In short, it may be a na?ve assumption that combining a modern targeted therapy with a traditional cytotoxic will never interfere with the action of the modern therapy. Given the negative clinical trial data of combining bortezomib with doxorubicin we decided to include this combination in our ChemoFinder assay to test the hypothesis that the negative data may have been the result of doxorubicin interfering with the anticancer activity of bortezomib.

Figure 3 shows the results from this study. The first thing to note is that the error bars in bortezomib monotherapy modes for both 2D and 3D are large consistent with the variable response rate of individual patient cancer cells to bortezomib treatment. However, despite this variation, it is also clear that bortezomib, when combined with doxorubicin is less effective than bortezomib alone both in 2D culture (Figure 3 left panel) and in 3D culture: the presence of doxorubicin completely inhibits the cytotoxicity of bortezomib (Figure 3 right panel). It is data like these that should serve as a cautionary warning to the assumption that combining multiple therapies at worst will be neutral but hopefully additive or even synergistic. The sobering reality is that the combination may prevent one, or even both, of the therapies from working at their optimal capacity. We also observed a similar effect on some drugs in glioblastoma when they were combined with x-irradiation (Yu et al 2016).

8.1        Learning from bacteriology?

One potential weakness of modern-targeted therapy is that the usefulness of the therapy is time-limited to how quickly the cancer can evolve an escape mutation to evade the “cross hairs” of the targeted therapy. A classic example of this problem occurs in malignant melanoma where up to 50% of patients will respond to a BRAF inhibitor (Sullivan & Flaherty, 2013a; Sullivan & Flaherty, 2011). In some patients this treatment acts like a cure with patients who have been riddled with secondary melanomas throughout their body regaining what appears to be full health within weeks of commencing BRAF inhibitor therapy. Sadly, this remarkable treatment suffers a definite time limit as malignant melanoma will almost always acquire a secondary upstream NRAS mutation which makes further treatment of the disease with a BRAF inhibitors even worse for the patient (Sullivan & Flaherty, 2013b). This problem of cancers escaping one pathway block by evolving a bypass mechanism has led many cancer scientists to propose that hitting multiple pathways simultaneously might in fact be the ultimate solution to maintaining a therapeutic efficacy in relation to killing malignancies. While this idea certainly is not one that should be quickly dismissed, especially given the logic of System Biology as discussed in section 3, page 6, our experience with multi-drug resistant bacteria may serve as a cautionary tale in relation to striking the cancer at many multiple targets simultaneously. The danger of this approach maybe that we inadvertently put too much selection pressure on the tumour resulting in a multi-drug resistant clone being selected which does not respond to any chemotherapy. However, what is fascinating about the NRAS escape mutation in malignant melanoma, is that if the patient can be given a break from BRAF inhibitors, and they can be kept alive for a meaningful length of time, then often their cancers will be re-sensitized to BRAF inhibition (Schmid et al., 2015). This surprising finding suggests that in the future, personalized chemotherapy could be used to find the current Achilles heel of a given patient’s cancer. Once this target is identified, the patient is then given a very specific monotherapy which targets this single signalling node. If, or most likely when, this specific therapy becomes ineffective, due to evolution of resistance within the patient’s cancer, the new resistant tumour is tested again in order to identify a new agent to which it is sensitive. The patient is then switched to this monotherapy, receiving treatment until the cancer evolves to make this second treatment no longer effective. At this point the cancer is tested a third time and the next most effective treatment selected. This process will continue hopefully allowing the oncologist to find a treatment to which the cancer is sensitive. If the cancer is only ever exposed to single or perhaps dual targeted agents, then at some point the original drug sensitivities of the cancer will likely re-emerge in the same way a Braf sensitivity re-establishes itself in patients who are given a Braf therapy “holiday”. In short, the strategy is to use the personalized medicine test to “chase the cancer’s Achilles heel” and if ultimately this chase results in the recycling of old therapies, then all the better.

 9           Concluding remarks

As I step down as a director of Imagen Therapeutics, my final contribution to their excellent endeavour is this article which I began at the end of 2016. It is now July 2018 and as I reread this article, I was struck with sadness that my opening sentence, about visiting my elderly Mother in 2017, never materialised. My poor Mum passed away after a brain bleed that resulted from a fall. Although, I boarded a flight back to Australia as soon as my sister gave me the grave news, I simply could not make it home in time. Why did she develop a brain bleed from a relatively trivial fall? Mum was suffering from a slow progressing cancer called myelofibrosis which meant that any minor injury could result in a lethal bleed due to a severe lack of platelets. Although, I sadly did not see my Mother again, the time I spent in Australia allowed me to catch up with one of my best friends (who was also the best man at my wedding) who sadly was also battling cancer. Little did either of us know but only a month afterwards he too would be joining Mum. While thinking about the demise of my friend still brings deep sadness, Imagen Therapeutics was able to test his cancer cells about 6 months before his death. His test was compromised because the sample got held up in UK customs for a few days. The results we obtained were sobering in that his cells did not respond to any of the therapies we tested. I remember saying to the Chief Scientific Officer, Gareth Griffiths, that I hope the customs delay had created some weird assay artefact. While this hope did not materialise, in a strange sort of a way the result gave me comfort in the knowledge that no matter what drug they would have used to treat my friend, his cause was lost. I do not have to live with the “what if they had used drug X” question since most drug classes were tested on his cancer in vitro.

Our hatred of cancer has given us a false confidence in thinking about how best to defeat it. Cancer is not a single disease but a large collection of many molecular aetiologies all leading to a similar outcome of dysregulated cell growth and death. To make further progress against the multiple molecular dysfunctions that result in cancer, we are going to have to become a little less confident in our current knowledge and a little bit more circumspect in accepting that perhaps, in relation to cancer, we know less than we think.

 Reference List

Aghajanian, C., Dizon, D. S., Sabbatini, P., Raizer, J. J., Dupont, J., & Spriggs, D. R. (2005). Phase I trial of bortezomib and carboplatin in recurrent ovarian or primary peritoneal cancer. J.Clin.Oncol. 23, 5943-5949.

Hirsch, F. R., Varella-Garcia, M., Bunn, P. A., Jr., Di Maria, M. V., Veve, R., Bremmes, R. M., Baron, A. E., Zeng, C., & Franklin, W. A. (2003). Epidermal growth factor receptor in non-small-cell lung carcinomas: correlation between gene copy number and protein expression and impact on prognosis. J.Clin.Oncol. 21, 3798-3807.

Nuti, M., Bellati, F., Visconti, V., Napoletano, C., Domenici, L., Caccetta, J., Zizzari, I. G., Ruscito, I., Rahimi, H., Benedetti-Panici, P., & Rughetti, A. (2011). Immune effects of trastuzumab. J.Cancer 2, 317-323.

Paik, S., Kim, C., & Wolmark, N. (2008). HER2 status and benefit from adjuvant trastuzumab in breast cancer. N.Engl.J.Med. 358, 1409-1411.

Parma, G., Mancari, R., Del Conte, G., Scambia, G., Gadducci, A., Hess, D., Katsaros, D., Sessa, C., Rinaldi, A., Bertoni, F., Vitali, A., Catapano, C. V., Marsoni, S., van, d., V, & Colombo, N. (2012). An open-label phase 2 study of twice-weekly bortezomib and intermittent pegylated liposomal doxorubicin in patients with ovarian cancer failing platinum-containing regimens. Int.J.Gynecol.Cancer 22, 792-800.

Schmid, S., Siano, M., Joerger, M., Rodriguez, R., Muller, J., & Fruh, M. (2015). Response to dabrafenib after progression on vemurafenib in a patient with advanced BRAF V600E-mutant bronchial adenocarcinoma. Lung Cancer 87, 85-87.

Schrag, D., Garewal, H. S., Burstein, H. J., Samson, D. J., Van Hoff, D. D., & Somerfield, M. R. (2004). American Society of Clinical Oncology Technology Assessment: Chemotherapy Sensitivity and Resistance Assays. J.Clin.Oncol. 22, 3631-3638.

Sullivan, R. J. & Flaherty, K. (2013a). MAP kinase signaling and inhibition in melanoma. Oncogene 32, 2373-2379.

Sullivan, R. J. & Flaherty, K. T. (2011). BRAF in Melanoma: Pathogenesis, Diagnosis, Inhibition, and Resistance. J.Skin Cancer 2011, 423239.

Sullivan, R. J. & Flaherty, K. T. (2013b). Resistance to BRAF-targeted therapy in melanoma. Eur.J.Cancer 49, 1297-1304.

Thatcher, N., Chang, A., Parikh, P., Rodrigues, P. J., Ciuleanu, T., von Pawel, J., Thongprasert, S., Tan, E. H., Pemberton, K., Archer, V., & Carroll, K. (2005). Gefitinib plus best supportive care in previously treated patients with refractory advanced non-small-cell lung cancer: results from a randomised, placebo-controlled, multicentre study (Iressa Survival Evaluation in Lung Cancer). Lancet 366, 1527-1537.

Vogel, C. L., Cobleigh, M. A., Tripathy, D., Gutheil, J. C., Harris, L. N., Fehrenbacher, L., Slamon, D. J., Murphy, M., Novotny, W. F., Burchmore, M., Shak, S., Stewart, S. J., & Press, M. (2002). Efficacy and safety of trastuzumab as a single agent in first-line treatment of HER2-overexpressing metastatic breast cancer. J.Clin.Oncol. 20, 719-726.

Vogel, U. F. (2010). Confirmation of a low HER2 positivity rate of breast carcinomas - limitations of immunohistochemistry and in situ hybridization. Diagn.Pathol. 5, 50.

 Yu, K., Pathmanaban, O., Dutkog-Wozdz, J., Beyit, D., Benson, R.S.P., Griffiths G.J., McBain, C., Peers, I., Ceuppens, P., Kamaly-Asl, I., & Bigger, B. (2016). The use of non-standard chemotherapeutic agents in patient derived cell lines in Glioblastoma and Glioblastoma variants using High Throughput Content Screening. J Neuro Oncology (In Preparation)