Precision Medicine: the importance of correctly matching the tool to the task

his year, on April 16th the Personalized Medicine World Conference (PMWC) made its debut in the UK. Attendee’s benefited from new insights behind the US and UK government genome initiatives, as well as some of the personal interactions (i.e. some ‘genomic gossip’) that drove the launch of these pioneering projects. Given their high profile nature much is expected from these genome- sequencing initiatives. There were many important reflections made during the opening presentations in Oxford, by Professor’s Peter Donnelly, Eric Green and Sir John Bell, points that I believe will shape not only the lasting impact of, but also the perception of, the first era of personalised medicine.

Several of the nuggets of wisdom imparted during the opening speeches particularly caught my attention. Firstly, a point was made regarding

managing public expectations and opinions. Analogy was provided through reference to European ‘public’ perception of the genetically

modified crops (GMC) industry. This is a key example of how a novel technology can encounter insurmountable barriers in the face of adverse

(or misplaced) perceptions. Perceptions retained for a generation or more. This immediately made me wonder, what if the current wave of

initiatives over-promise what can be delivered, will ‘faith’ in personalised medicine among front-line medical staff, investors and the public

result in a GMC scenario?

A second impression I took from the opening presentations of the Personalized Medicine World Conference was that there was no consistent view

of what ‘personalised’ healthcare would look like or how the evidence base for creating such a personalised approach can be generated. For some,

it was articulated in a manner suggesting it was a natural extension of ‘evidence based medicine’ whereby statistical association of factors mined

from newly emerging ‘big data’ would refine the granularity of treatment plans or contribute to prediction of risk of future disease. For others it

was more firmly related to DNA sequencing and identification of the underlying molecular defects that co-associated with an individual’s tumor;

personalised diagnosis to be precise.

by Professor James A. Timmons PhD, King’s College London, UK

T

Without the human genome project we would also not have microarrays, a technology that delivers highly reproducible tumor stratification and drug class response predictors.

31

by Professor James A. Timmons PhD, King’s College London, UK

What’s the evidence base(s)

The outcomes of the human genome project3

represent a key platform for many subsequent

and ongoing precision medicine efforts. For

example, Foundation medicine Inc (I have no

connection with this company) report that

‘three times more actionable alterations’ are

detected using massively parallel DNA sequencing

than traditional histopathology4. Without the

human genome project we would also not have

microarrays, a technology that delivers highly

reproducible tumor stratification5 and drug

class response predictors6 using a single small

sample of RNA. Thus it seems logical to extend

the genome sequencing efforts to expand our

knowledge of individual DNA aberrations

associated with various types of cancer and

other major diseases. Yet the nucleic acid that

has proven to be most successful in delivering

precision medicine to oncology is, beyond any

doubt, RNA and not DNA7–10.

Are we backing the wrong base in the large-

scale public precision medicine initiatives,

when industry is successfully turning RNA into

FDA and NICE approved products? I believe so

and the reasons for this are somewhat

predictable. However first let’s reflect on the

Foundation Medicine DNA sequencing 4 results.

In this multi cancer type analysis ~1.5 events

per patient were defined as being actionable.

How many actionable events were restricted to

a narrow class of established mutations where

drug choice can be ruled in or out is unclear.

Referring to data in this manner is like

referring to the total number of FDA approved

genomic diagnostics without referring to the

fact that such tests cover an extremely narrow

spectrum of usage or utility.

Could DNA analysis be utilized in the near

future to offer the breast cancer patient

a chemotherapy normally provided to a

pancreatic cancer patient? Certainly not and

this raises a point of concern regarding the

need to manage expectations and the earlier

point about GMC’s. To implement the first

wave of precision medicine, we need solutions

that are integrated (Figure 1) into current and

near-future capacities not a vision of medicine

in 2250.

Without the human genome project we would also not have microarrays, a technology that delivers highly reproducible tumor stratification and drug class response predictors.

PR

EC

ISIO

N T

OO

LS

MolecularDiagnosis

Molecular drug signatures matched back to patient profiles

Patient profiles used to match to or screen for new treatments

MolecularPrognostic

MultipleTreatment options

Treatment Responsepredictor

Discovery of newtreatments

In contrast, a RNA signature of the patients

tumor at least holds the chance of being

contrasted immediately with >80 FDA approved

chemotherapeutics, or those in development,

and a novel option being provided today.

Industrialization of this RNA approach is

technically feasible and it will begin to stack

the odds in the favor of finding a novel drug

match to a novel tumor profile, using existing

small molecule drugs or compounds.

I would also question whether more complete

sequencing of an individual’s DNA, has any

real potential to yield prognostics for most of

the ‘common’ diseases, such as dementia or

diabetes, never mind match the emerging

ability of RNA diagnostics/prognostics to

deliver precision medicine for Oncology.

Indeed, it was very revealing that all speakers

presenting DNA sequencing strategies at the

PMWC, illustrated progress [sic] of DNA

approaches by presenting conventional case

As of today, RNA signatures are already used to

develop drug-response predictors for approved

drugs and those under development11,12. And,

as mentioned above, RNA signatures are used

to provide prognostic information and match

tumor profiles to drug response predictors.

precision medicine strategy will need to evolve

in a joined-up manner, delivering ‘data’ that

can be ‘interpreted’ by all phases of the patient

treatment and recovery process.

In short, while discovering a subset of pancreatic

cancer patients have a specific deletion in a

novel kinase is valuable, in the absence of an

alternative drug treatment, this information

represents basic research and not something

useful to the patient being profiled (Figure 2).

DNAUnproven to guide drug deverlopmentSingle gene drug-targets not for chronic complex disease

MolecularDiagnosis

Diagnosis/prognosis

Limited impacton treatment

selection

‘smart’ devicedata

Diagnosis?Impact ontreatmentselection?

Limited feasibility to guide drug developmentor individualize/match drug to patient biology

examples of rare childhood disease (and

without doubt doing some good regarding

the management of the patient). Genome wide

association studies (GWAS), and more recently

exome sequencing, have failed to provide

diagnostic models of common chronic diseases

(such dementia, diabetes or obesity) because

the phenotyping in these studies is too naïve or

inaccurate, or because the statistical models for

analyzing DNA sequence associations remains

too primitive or because the disease is largely

non-genetic.

In contrast to DNA, multi-gene RNA

signatures represent integrated models that

combine non-linear interactions and

continuous variables into genuine diagnostics

that impact on patient treatment decisions,

today. The real edge that RNA has over DNA,

is that RNA abundance integrates genetic,

epigenetic and environmental influences,

capturing more variance, and so making it

rather obvious why it is the nucleic acid choice

for delivering precision medicine today and

tomorrow, even if it is superseded some time

in the future. Nevertheless, it is currently

computationally intractable to explore the

entire transcriptome and while data may

contain a ‘diagnostic’ signature, it is not always

discoverable. The same will be true for all large

scale OMIC or ‘big-data’ projects.

Thus while there is no doubt in my mind that

the US Precision medicine initiative and the UK

100,000 Genome Project will yield new knowl-

edge for rare diseases and a much more refined

map of the cancer genomes, I worry that this

cannot be mapped sufficiently well to open

up the required range of treatment options.

This brings the discussion back to managing

expectations. The political backing to the recent

Precision Medicine can be a double-edged

Are we backing the wrong ‘base’ in the large-scale publicly funded precision medicine initiatives, when industry is successfully turning RNA into FDA and NICE approved products?

gold-standard phenotyping was unfeasible in

large clinical cohorts) to identify factors that

might cause disease. There have been some

blindingly good successes. The association

between lung cancer and smoking is so strong,

that one can conclude causality13. Yet the

relative risk of developing lung cancer from

smoking ranges from <2 up to >40-fold increase

across continents reflecting differing genetics

and differences in data recording methodologies.

Thus if something that is so apparently causal

(smoking and lung cancer) suffers from

methodological challenges to see the ‘signal

from noise’ then how might we gather

sufficiently accurate and informative data on

weaker interactions, to provide individual level

assessments of risk or personalised solutions?

Smart technology combined with measures

of medical outcomes (or ‘health’) based on

patient records14 is considered one promising

‘big data’ solutions to this problem. The

question remains, does greater granularity

(and noise) lead to better predictive models for

sword. Both public and private investment

in this field will demand early success. The

time-lag between genomic discovery and novel

therapeutic development (the latter is typically

measured in decades) will create the impres-

sion that OMIC driven personalised medicine

has been over-hyped.

There are also many technical hurdles to

over-come. At the Oxford PMWC meeting,

Personalis Inc. founder Euan Ashley (also

director of the clinical genomics services at

Stanford) highlighted existing and major

technical hurdles when trying to generate a

clinical grade whole genome sequence. The

topic represented yet another genomic

‘elephant in the room’ at the Oxford meeting.

He explained that new high-throughput

sequencing assays fail to provide information

on large regions within exons, so much so

they can’t find known disease causing regions.

He flashed up Personalis’s technology solution to

improving the quality of exome sequence data,

and it appeared for the examples given (albeit

briefly) at least 25% of each exon regions have

poor coverage, even after getting the ‘Per-

sonalis treatment’. Going ‘genome-wide’ for

the individual may still take some time, while

establishing the impact of such information a

lot longer, when it comes to chronic adulthood

diseases (the diseases which are undermining

the financial stability of many western health

care systems). Dr Ashley introduced another

aspect of the personalised medicine revolu-

tion, namely the hope that mobile technology

gathered ‘big data’ and data mining of medical

records could reveal novel predictors of health,

including the launch of Apple’s health app.

What is big data and how does it differ

from traditional evidence based medicine?

If you want to generate a really impressive

p-value from two parameters that have a genu-

ine interaction, then make your sample size as

big as possible. This is precisely what we have

seen from the Nature-published ‘mega’ GWAS

studies where even the smallest of biological

interactions can become ‘GWAS significant’.

Population epidemiology has long relied on

approximate clinical measures (because

the individual patient or simply more data and

less chance to find a predictive model? (recall,

exploring exhaustively multi-level interactions

are beyond conventional computing).

Let’s consider the ‘health app’, which uses

detection features (or battery usage) within a

smart phone to track your physical movements

and provide approximations of physiological

traits (e.g. heart rate). Anyone that has run a

clinical trial, where accelerometer and heart

rate data is gathered using near clinical grade

devices, will tell you that the best data are

noisy (we are doing this in the metapredict.eu

exercise and pre-diabetes study). Too noisy

to relate physical activity reliably to accurately

measured metabolic and physiological capacities?

Probably. The idea that a smart phone that

approximates your movements (or your 6yr old

son’s movements when he takes it from you to

play angry-birds) will provide a window into

personalised analysis of your physical status is

naïve (figure 3).

Proven to guide

treatment

selection

ContinuousOMIC

variable e.g RNA

Diagnosis

/

prognosis

Proven to match

drug to disease

Limited feasibility to guide drug developmentor individualize/match drug to patient biology

Enab

les d

rug-

repurp

osing

Professor Timmons has spent equal time in academic

and industrial science. Following training in the physiologi-

cal and pharmaceutical sciences he led lead-discovery and

lead-development programs in cardiovascular and

metabolic disease at Pfizer and Organon. After establishing

a translational medicine project in collaboration with

Karolinska Institute, Jamie moved to Sweden to oversee

the completion of this muscle genomics project and

then ran a group in the Centre for Genomics and

Bioinformatics working with RNA based technologies.

Awarded his first chair aged 35yr, he is scientific leader for

an FP7 project that is developing novel ‘machine learning’

based diagnostics for ‘pre-diabetes’, as well as predictive

models for cardiovascular adaptation in humans. In 2012,

he set-up XRGenomics LTD which has produced the first

accurate RNA diagnostic of biological ageing and he acts

as a scientific advisor to companies operating in the

personalised medicine arena. His unique multi-disciplinary

training allows him to view the new world of Precision

Medicine from multiple perspectives.

Enabling Precision Medicine

There is no doubt in my mind that diagnosis

and treatment solutions need to be advanced

in parallel to avoid disappointment (Figure 4).

While laboratory-based genomic technologies

are advancing each year, as is our ability to

capture and store ‘big data’ - our ability to

model such data, or develop novel treatment

strategies lags far behind. We lack computational

capacity to exhaustively explore the search

space of even modest ‘big data’ (e.g. a gene-chip

transcriptomic profile). While the speed at

which we can develop and validate molecular

therapeutics remains firmly rooted in the

20th century. There are pragmatic and obvious

solutions.

Firstly, greater effort should be placed on

producing molecular (RNA) signatures of all

chemical entities known to be safe in humans.

This should include profiling all of the

numerous compounds that failed (due to lack

of efficacy for the original disease) in Phase

II or III during the past few decades of drug

developments. Such an idea is a simple

extension of the connectivity database

initiative16 but requires a broader public-

private cooperation along the lines being

achieved within the EU funded IMI programs

and other more recent collaborative programs.

In parallel with this, a substantial investment

in dramatically extending the use of transcriptome

profiling in large population cohorts and

intervention trials, using high-throughput

chip based technologies e.g. the 384 GeneTitan

platform. This not only creates novel molecular

insight into disease but allows for the

computational matching of poorly studied

but otherwise ‘safe’ small molecular drugs

(currently sitting unused in Pharmaceutical

companies) and disease or phenotype response

‘figure prints’. This is a robust and emerging

strategy for matching tumor characteristics

with chemotherapy6,12,17 and predicting

which patients would benefit from treatment

and what their other options might be. These

are examples of RNA based precision and

personalised medicine tools, while progress

with metabolomics provides additional options

which may transpire to be as powerful as RNA

profiling.

A final comment would be that the

composition of the teams leading the

precision medicine initiatives must be

broadened. Data gathering companies,

data-logistics and gene-sequencing gurus

represent an essential but narrow aspect

of the experience required to meet the

political expectation of Precision Medicine.

We need to integrate into these teams, those

experienced in drug development (clinical),

along with the wealth of experience within

the fields of molecular diagnostics (technical

validation) and disease specialists (to help

disseminate the novel approaches to

colleagues). An interesting challenging is the

structure of the medical profession. Clinical

experts are siloed, rarely attending medical

congresses out-with their clinical specialties.

I recently witnessed this when attending a

meeting focused on dementia, that referred to

aspects of cardiovascular and metabolic disease

that have long since been dismissed as overtly

simplistic within those latter disciplines. As I

have tried to emphasize throughout this article,

the Precision Medicine revolution needs to

impact, in parallel, across the entire range of

activities that come together to deliver health-

care and not sequentially (pardon the pun),

one phase at a time.

It is naïve for many reasons beyond the sen-

sitivity of the parameters being measured. To

personalize the relationship between physical

activity and the metabolic and physiological

outcomes, complex molecular phenotyping is

required – simply put we all respond differently

to all types of physical intervention15 and a

given unit of physical activity, blood pressure

or glucose excursion yields a ‘personal’ set of

outcomes. Thus interpretation of ‘big data’

requires calibration to the individual’s biology.

Current ‘big data’ gadgets replicate traditional

epidemiology through the creation of data

models that relate behaviors to population

averages. Smarter thinking is required to make

use of the data. For example, utilization of

continuous data gathering technologies and

‘big data’ medical record analysis may yield

numerous insights in the correct setting.

Modeling links between changes in behavior

and periods of ill-health may provide

forewarning that an individual my benefit

from taking some preventative treatment days

prior to symptoms emerging (coupled with

internet based surveillance data of winter flu?).

Some types of calibration might emerge

from sufficient ‘big data’ modeling but

whatever diagnostic emerges from modeling

smart technology data, it will need to

fulfill specificity and sensitivity criteria and be

calibrated. These terms have still to enter the

routine vocabulary of ‘big data’ – at the

moment its all ‘Cloud analytics’ – referring

mainly to the storage and transfer problem. In

short, ‘big data’ analytics represents another

potential strategy for precision medicine, but

where and when it should be best deployed

to improve medical diagnosis remains to be

evaluated. Further, the aim is to provide

personalised guidance or diagnosis, and while

this can be done by studying a person over

time, its unclear how n=1 treatment can be

validated, when both disease and time wait for

nowo(man). In short, we should not forget

that association today is not necessarily

evidence that can be used tomorrow.

37

References

1. Messerli FH. Chocolate Consumption, Cognitive Function, and Nobel Laureates. N Engl J Med. 2012;367(16):1562-1564. doi:10.1056/NEJMon1211064.2. Dewey FE, Grove ME, Pan C, et al. Clinical interpreta-tion and implications of whole-genome sequencing. JAMA. 2014;311(10):1035-45. doi:10.1001/jama.2014.1717.3. Consortium IHG. Initial sequencing and analysis of the human genome. Nature. 2002;420(February):520-562.4. Frampton GM, Fichtenholtz A, Otto G a, et al. Develop-ment and validation of a clinical cancer genomic profiling test based on massively parallel DNA sequencing. Nat Biotechnol. 2013;31(11):1023-31. doi:10.1038/nbt.2696.5. Naoi Y, Kishi K, Tsunashima R, et al. Comparison of efficacy of 95-gene and 21-gene classifier (Oncotype DX) for predic-tion of recurrence in ER-positive and node-negative breast cancer patients. Breast Cancer Res Treat. 2013;140(2):299-306. doi:10.1007/s10549-013-2640-9.6. Mulligan JM, Hill L a, Deharo S, et al. Identification and validation of an anthracycline/cyclophosphamide-based chemotherapy response assay in breast cancer. J Natl Cancer Inst. 2014;106(1):djt335. doi:10.1093/jnci/djt335.7. Paik S, Shak S, Tang G, et al. A multigene assay to predict recurrence of tamoxifen-treated, node-nega-tive breast cancer. N Engl J Med. 2004;351(27):2817-2826. doi:10.1080/14733400500093379.8. Cusumano PG, Generali D, Ciruelos E, et al. European inter-institutional impact study of MammaPrint. The Breast. 2014;23(4):423-428. doi:10.1016/j.breast.2014.02.011.9. Albain KS, Barlow WE, Shak S, et al. Prognostic and predictive value of the 21-gene recurrence score assay in postmenopausal women with node-positive, oestrogen-receptor-positive breast cancer on chemotherapy: a retrospective analysis of a randomised trial. Lancet Oncol. 2010;11(1):55-65. doi:10.1016/S1470-2045(09)70314-6.10. Dowsett M, Sestak I, Lopez-Knowles E, et al. Comparison of PAM50 risk of recurrence score with oncotype DX and IHC4 for predicting risk of distant recurrence after endocrine therapy. J Clin Oncol. 2013;31(22):2783-2790. doi:10.1200/JCO.2012.46.1558.11. Geeleher P, Cox NJ, Huang RS. Clinical drug response can be predicted using baseline gene expression levels and in vitro drug sensitivity in cell lines. Genome Biol. 2014;15(3):R47. doi:10.1186/gb-2014-15-3-r47.12. Knudsen S, Jensen T, Hansen A, et al. Development and validation of a gene expression score that predicts response to fulvestrant in breast cancer patients. PLoS One. 2014;9(2):e87415. doi:10.1371/journal.pone.0087415.13. Lee PN, Forey BA, Coombs KJ. Systematic review with meta-analysis of the epidemiological evidence in the 1900s relating smoking to lung cancer. BMC Cancer. 2012;12(1):385. doi:10.1186/1471-2407-12-385.14. Roque FS, Jensen PB, Schmock H, et al. Using electronic patient records to discover disease correlations and stratify patient cohorts. PLoS Comput Biol. 2011;7(8). doi:10.1371/journal.pcbi.1002141.15. Timmons JA. Variability in training-induced skeletal muscle adaptation. J Appl Physiol. 2010. Available at: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=21030666.16. Lamb J, Crawford ED, Peck D, et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Science (80- ). 2006;313(5795):1929-1935. Available at: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=17008526.17. Baggerly K a., Coombes KR. Deriving chemosensitivity from cell lines: Forensic bioinformatics and reproduci-ble research in high-throughput biology. Ann Appl Stat. 2009;3(4):1309-1334. doi:10.1214/09-AOAS291.

PR

EC

ISIO

N T

OO

LS

“And it ought to be remembered that there is nothing more difficult to take in hand, more perilous to conduct, or

more uncertain in its success, than to take the lead in the introduction of a new order of things.”

Niccolò Machiavelli, The Prince

P E R S O N A L I Z Ethejournalofprecisionmedicine.com

The Journal of Precision Medicine will publish

bi-monthly in 2016. Our next issue will be

published in September, so to ensure regular

copies please subscribe online and

choose your preference for either print

or digital versions or both!