IDBS Data Management Blog

Say what you’re not saying, don’t say it, say what you didn’t say

Last time I blogged that modelling is not limited to software engineering, play and simulation; but is universal in human endeavour. I mentioned that considering accuracy is important but not sufficient in assessing a model. What other considerations are there?

My favourite lens for looking at a model is abstraction. In philosophical terminology, abstraction is about grouping concepts together at decreasing levels of detail. So, a duck is a duck and no other thing is a duck (no matter how it looks or walks or sounds); but applying abstraction allows us to talk about birds and say useful things, which might be rather exasperating if we had to list every bird in the world to say them. This kind of classification is a particular feature of object-oriented programming languages (which may or may not be a good thing).

Leaving it out

A modeller, not saying

However another way of considering abstraction is to pause before asking what a model is saying, and ask: what is this model not saying?

The model of biological change that we call evolution has incredible empirical support, so that its application has great explanatory and predictive power (some would even say that we don’t apply it enough). Strangely, though, it seems to cause an awful lot of consternation to those who subscribe to another model called creationism.

Why strange? At first sight, both of these models deal with how the world came to be the way it is. But evolution models a process, and has nothing whatsoever to say about how that process began, or why it began, or who began it. Conversely, creationism says nothing about how its proposed agent went about his craft (well, usually). He just did it. Apples and oranges.

Putting it back in

Any critical analysis or use of a model has to carefully stick to assessing or building upon what it actually models. This might sound simple, but humans find it remarkably tricky. We are fond of making cultural and doctrinal assumptions and applying intuitions without knowing about it. (In a black alley, a black cat spies a black rat. How?*)  Unfortunately this is not only inevitable, it’s usually necessary.

Why so? Models almost always rely on background information. Of particular interest in computer science and artificial intelligence is the notion of semantics: the meaning of symbols. Tell a robot to fetch you a cuppa, and it may suffer the same semantic confusion as is now affecting US readers: a cuppa what?

I’ve really enjoyed seeing the Cancer Research UK Stratified Medicine Programme getting so much coverage around the globe. This initiative will blaze the trail for the wider adoption of genetic testing to support diagnosis and treatment of various cancers including breast, colorectal, lung, prostate, ovarian and skin cancer.

The 2 year program will see 9,000 samples and associated clinical data systematically captured and genetically tested for known cancer variants with a view to building a comprehensive warehouse of cancer data. This will then form a research resource to better understand the genetic basis for diagnosis and disease treatment, and in the future support clinical decision-making.

The Cancer Research UK project is part of a larger Stratified Medicines Innovation Platform funded by the Technology Strategy Board and other bodies to advance personalised medicine in the UK.

We at IDBS are proud to be leading one of these projects, designed to support industry and academic collaboration in stratified medicines based on high quality, longitudinal patient information and associated genetics. The Acropolis project will focus on the secure infrastructure to analyse and share this data, leading to improved disease understanding and patient outcomes.

We wish the project team well because this is by no means a trivial undertaking and highlights many of the difficulties faced by academic medical centres around the world in bringing together critical data to support translational medicine research.

The project team face many challenges in bringing these disparate data sets together, not least of which is capturing patient data from the clinical systems it is stored in. This information is frequently distributed across electronic medical records, laboratory systems, PACs and cancer information systems requiring sophisticated informatics to extract the source data and format it consistently. Information governance and protected health information compliance are also key areas to address as is integration with biobanks and incorporation of genetic tests results from standardised procedures.

We will be cheering our sister project on and continuing to provide support to the Cancer Research UK team as we begin the journey towards personalised medicine in the UK.

Last Friday’s report in Reuters that less than 1% of US cancer patients get into clinical trials for new treatments tells a much wider story: that clinical trial and the real world clinical environment are often very different things.

The conclusion of the academics was:

In addition to profoundly low overall cancer trial accrual, vast underrepresentation by age, cancer stage, and site continue to exist. The generalizability of these trials to a real world perspective remains an open question. Physicians, payers, the National Cancer Institute, and other stakeholders need to develop broader cancer trials to benefit the millions of patients with cancer in the United States.

If any clinical trial is made up of demographically similar, symptom-selected patients cohorts who ‘qualify’ for the study, whilst the population the therapeutic is designed to treat is a demographically, genomically and symptomatically-diverse group, it is hardly surprising that clinical trial data often do not hold up after approval and launch.

Whilst this remains inconveniently true, it does no good to indulge in hand-wringing. Far better to look at how real world clinical data can be used to supplement clinical trials and better understand patient diversity and response. Clinical trials can then be adapted to treat what in reality are a diverse set of clustered patient cohorts with separate therapeutics, targeted at patients using biomarkers and diagnostics.

Underpinning this drive towards ‘stratified medicine’ or ‘precision medicine’ is ensuring we can bring together the symptom, demographic AND genomic data and use them properly to select multiple cohorts of patients based on all available knowledge to fully understand the underlying disease, not just applying a subset of clinical and lab data.

None of this is easy but the FDA knows it is coming. Two years ago Lawrence Lesko, PhD, FCP, director of clinical pharmacology and biotherapeutics said:

“For precision medicine, the disease must be diagnosed to the gene,” and he emphasized that vast amounts of data generated by human genome research in recent years creates an opportunity for drug makers to better target medicines to smaller populations of patients. But at the same time, the challenge for drug makers is like turning a large field of long-range radar antennae inward into the body. As Lesko points out, “Who is going to interpret the data?” 

Once again, the ability to handle the data is key enabling technology.

IDBS have proven we can do this with real world data and provide the key to unlock insights into better trials and more predictable patient outcomes.