New methods for reviewing 
mechanistic evidence 
Systematic review guidelines for integrating evidence 
from human, animal and other mechanistic studies 
which link diet, nutrition and physical activity to cancer 
Richard Martin 
School of Social and Community Medicine, University of Bristol
Why systematically 
synthesize mechanistic data? 
Having confidence in the evidence 
 
Extensive mechanistic data link diet with cancer 
– Biological plausibility; informs interventions 
 
Of 52 observational claims tested in trials published in JAMA, 
JNCI & NEJM none were confirmed & 10% contradictory 
 
< 10% highly promising basic science discoveries enter 
clinical use 
 
Evidence distorted (falsely inflated) by 
– Irrelevance (population, exposure, outcomes) 
– Weak internal validity (randomisation, allocation concealment, 
blinding, drop outs, co-morbidities) 
– Publication bias (98% of animal studies ‘significant’)
Selective reporting bias in cancer prognostic studies 
Meta-analysis of association of TP53 status with risk of death at 2 years 
Kyzas P A et al. JNCI J Natl Cancer Inst 2005;97:1043-1055
Our aims 
To develop a detailed protocol for comprehensive 
systematic reviews of mechanistic studies 
 
Systematic reviews 
– Allow objective appraisal of evidence 
– Reduce false-positive & false-negative results 
– Identify sources of bias, improving study quality 
 
This project should increase the value of mechanistic data: 
– Enable more rigorous systematic reviews 
– Increased precision of estimated effects 
– Identify gaps in the research evidence 
– Reduce selective citation of mechanistic evidence 
– Inform relevance to humans (cross species/model heterogeneity) 
 
Tool for translating basic science into policy & practice 
 
Important for the Continuous Update Project
Overall approach 
Guideline development 
 
Four workshops with experts within our group 
– mixture of presentations with discussion, small group exercises, 
round table discussions 
 
Regular meetings in between workshops 
– Refine the protocol 
– Carry-out searches 
– Determine inclusion/exclusion criteria 
– Investigate QC criteria 
– Consider relevance, publication bias, reporting/display of results 
 
Expertise in: 
- Systematic reviews of epidemiological studies 
- Experimental studies of cancer 
- Bioinformatics 
- Information technology
The Team 
University of Bristol: 
 
PI - Dr Sarah Lewis - Genetic epidemiology/systematic reviews of genetic studies 
 
Co-PI - Prof Richard Martin - Cancer epidemiology/systematic reviews 
 
Dr Mona Jeffreys - Cancer epidemiology/systematic reviews 
 
Dr Mike Gardner - Animal biology/systematic reviews 
 
Prof Jeff Holly - Molecular biology – IGF and cancer 
 
Dr Claire Perks - Molecular biology 
 
Dr Tom Gaunt - Genetic epidemiology/bioinformatics 
 
Prof Jonathan Sterne - Meta-analysis and systematic review methodology 
 
Professor Julian Higgins - Meta-analysis and systematic review methodology 
 
Prof Steve Thomas - Head and neck surgeon 
 
Dr Pauline Emmett - Nutritional epidemiology 
 
Dr Kate Northstone - Nutritional Epidemiology 
 
Cath Borwick – Information specialist 
University of Cambridge: World Cancer Research Fund International 
Dr Suzanne Turner - Animal models Prof Martin Wiseman 
Dr Pierre Hainut (advisor to WCRF Int) 
Dr Panagiota Mitrou, Dr Rachel Thompson 
IARC: 
Dr Sabina Rinaldi - Hormones and cancer
Summary of the process 
Research 
question 
(PEMO)) 
• Identify 
• Appraise individual 
studies 
• Integrate body of 
evidence 
• Mechanism discovery 
(unbiased) 
• Specific mechanisms (targeted) 
• Risk of bias 
• Relevance 
• Within an evidence stream 
(human, animal, in vitro) 
• Across evidence streams 
• Confidence 
conclusion 
 High 
 Moderate 
 Low 
Eligibility 
criteria
Identifying the evidence 
Two step process for searching for mechanisms 
 
Broad automated search to encompass all mechanisms 
– Mechanism discovery 
– Quantitative assessment of mechanism evidence 
– Assists prioritization of mechanisms for review 
– “Hypothesis-free” (to some extent) 
– Identifies efficient starting points for review 
– Identifies potential mechanisms unknown to the reviewer 
 
Targeted search – focus on a particular pathway 
– Apply pre-specified inclusion/exclusion 
– Has the cancer arisen in the animal model rather than being 
transplanted into the animal? 
– Have cell lines been authenticated and results replicated? 
Primary 
search 
Mechanism 
discovery 
Specific 
searches
Automated mechanism quantification and display 
EXPOSURES INTERMEDIATE 
MECHANISMS 
OUTCOMES
Evidence synthesis 
Three step process 
 
Appraisal of the individual studies 
– Risk of bias 
– Present summary data, stratified for presence of 
heterogeneity 
 
Integrate within evidence streams (human / animal) 
– Risk of bias Magnitude of effect 
– Inconsistency Confounding controlled for 
– Imprecision Dose response 
– Publication bias 
– Irrelevance 
 
Integrate across evidence streams 
–
Causal conclusions 
Integrating evidence 
Level of evidence in 
human studies 
High 
Strong 
Moderate 
Weak Modest 
Low 
Inconclusive Weak Modest 
Low Moderate High 
Level of evidence in animal 
studies 
Supportive evidence from in vitro and xenograft models 
underpinning biological plausibility
Milk and prostate cancer 
exemplar 
 
Milk implicated in prostate cancer, but 
– is measured semi-quantitatively 
– is susceptible to confounding 
– large differences between individuals in the same group 
lead to attenuation by measurement errors 
 
Systematic reviews of human observational studies are 
inconclusive 
– Limited-suggestive evidence (World Cancer Research 
Fund International, 2014) 
 
Experimental studies not systematically reviewed
Results of automated mechanism 
quantification (36,000 hits)
Targeted search results 
4946 targeted studies identified 
725 studies retrieved for detailed 
evaluation 
4221 studies excluded after 
review of title and abstract 
(duplicates or clearly 
ineligible) 
Databases searched: 
• Medline 
• Embase 
• BIOSIS 
• CINAHL 
268 studies potentially eligible 
325 studies excluded after 
review of full text 
132 studies awaiting ILL 
23 milk-IGF studies 245 IGF-prostate cancer 
22 human studies extracted (5 
RCT, 3 cohort, 14 cross-sectional) 
50 human studies 
extracted (total: 99) 
7 animal studies 
extracted (total: 8) 
138 Cell line studies
Evidence synthesis 
Milk and IGF in human RCTs by increasing length of follow-up 
Study 
Rich-Edwards 2007 
Hoppe 2004 
Cadogan 1997 
Zhu 2005 
Ben-Shlomo 2005 
Follow-up 
(yrs) 
.02 
.02 
1.5 
2 
25 
% 
Male 
Control 
0 
100 
0 
0 
52.8 
Overall (I-squared = 69.0%, p = 0.012) 
NOTE: Weights are from random effects analysis 
ES (95% CI) 
13.30 (-28.64, 55.24) 
40.20 (-7.43, 87.83) 
69.00 (5.89, 132.11) 
28.60 (-17.66, 74.86) 
-9.50 (-16.75, -2.25) 
20.87 (-8.75, 50.49) 
% 
Weight 
19.59 
17.59 
13.09 
18.06 
31.66 
100.00 
-132 0 132 
Favours control Favours intervention 
Difference in IGF-I (ng/ml) 
IGF and hallmarks of cancer 
(authenticated cell lines)
Future work 
 
Automated mechanisms discovery 
– Validation (not missing important studies) 
– Inter-relationships between mechanisms 
– Develop stand-alone automation software and beta 
testing 
 
Validation of relevance questions 
 
Acceptability 
 
Reliability
For further information 
Sarah Lewis: S.j.lewis@bristol.ac.uk 
Richard Martin: Richard.martin@bristol.ac.uk 
@wcrfint 
facebook.com/wcrfint 
 
 
www.wcrf.org

New methods for reviewing mechanistic evidence

  • 1.
    New methods forreviewing mechanistic evidence Systematic review guidelines for integrating evidence from human, animal and other mechanistic studies which link diet, nutrition and physical activity to cancer Richard Martin School of Social and Community Medicine, University of Bristol
  • 2.
    Why systematically synthesizemechanistic data? Having confidence in the evidence  Extensive mechanistic data link diet with cancer – Biological plausibility; informs interventions  Of 52 observational claims tested in trials published in JAMA, JNCI & NEJM none were confirmed & 10% contradictory  < 10% highly promising basic science discoveries enter clinical use  Evidence distorted (falsely inflated) by – Irrelevance (population, exposure, outcomes) – Weak internal validity (randomisation, allocation concealment, blinding, drop outs, co-morbidities) – Publication bias (98% of animal studies ‘significant’)
  • 3.
    Selective reporting biasin cancer prognostic studies Meta-analysis of association of TP53 status with risk of death at 2 years Kyzas P A et al. JNCI J Natl Cancer Inst 2005;97:1043-1055
  • 4.
    Our aims Todevelop a detailed protocol for comprehensive systematic reviews of mechanistic studies  Systematic reviews – Allow objective appraisal of evidence – Reduce false-positive & false-negative results – Identify sources of bias, improving study quality  This project should increase the value of mechanistic data: – Enable more rigorous systematic reviews – Increased precision of estimated effects – Identify gaps in the research evidence – Reduce selective citation of mechanistic evidence – Inform relevance to humans (cross species/model heterogeneity)  Tool for translating basic science into policy & practice  Important for the Continuous Update Project
  • 5.
    Overall approach Guidelinedevelopment  Four workshops with experts within our group – mixture of presentations with discussion, small group exercises, round table discussions  Regular meetings in between workshops – Refine the protocol – Carry-out searches – Determine inclusion/exclusion criteria – Investigate QC criteria – Consider relevance, publication bias, reporting/display of results  Expertise in: - Systematic reviews of epidemiological studies - Experimental studies of cancer - Bioinformatics - Information technology
  • 6.
    The Team Universityof Bristol:  PI - Dr Sarah Lewis - Genetic epidemiology/systematic reviews of genetic studies  Co-PI - Prof Richard Martin - Cancer epidemiology/systematic reviews  Dr Mona Jeffreys - Cancer epidemiology/systematic reviews  Dr Mike Gardner - Animal biology/systematic reviews  Prof Jeff Holly - Molecular biology – IGF and cancer  Dr Claire Perks - Molecular biology  Dr Tom Gaunt - Genetic epidemiology/bioinformatics  Prof Jonathan Sterne - Meta-analysis and systematic review methodology  Professor Julian Higgins - Meta-analysis and systematic review methodology  Prof Steve Thomas - Head and neck surgeon  Dr Pauline Emmett - Nutritional epidemiology  Dr Kate Northstone - Nutritional Epidemiology  Cath Borwick – Information specialist University of Cambridge: World Cancer Research Fund International Dr Suzanne Turner - Animal models Prof Martin Wiseman Dr Pierre Hainut (advisor to WCRF Int) Dr Panagiota Mitrou, Dr Rachel Thompson IARC: Dr Sabina Rinaldi - Hormones and cancer
  • 7.
    Summary of theprocess Research question (PEMO)) • Identify • Appraise individual studies • Integrate body of evidence • Mechanism discovery (unbiased) • Specific mechanisms (targeted) • Risk of bias • Relevance • Within an evidence stream (human, animal, in vitro) • Across evidence streams • Confidence conclusion  High  Moderate  Low Eligibility criteria
  • 8.
    Identifying the evidence Two step process for searching for mechanisms  Broad automated search to encompass all mechanisms – Mechanism discovery – Quantitative assessment of mechanism evidence – Assists prioritization of mechanisms for review – “Hypothesis-free” (to some extent) – Identifies efficient starting points for review – Identifies potential mechanisms unknown to the reviewer  Targeted search – focus on a particular pathway – Apply pre-specified inclusion/exclusion – Has the cancer arisen in the animal model rather than being transplanted into the animal? – Have cell lines been authenticated and results replicated? Primary search Mechanism discovery Specific searches
  • 9.
    Automated mechanism quantificationand display EXPOSURES INTERMEDIATE MECHANISMS OUTCOMES
  • 10.
    Evidence synthesis Threestep process  Appraisal of the individual studies – Risk of bias – Present summary data, stratified for presence of heterogeneity  Integrate within evidence streams (human / animal) – Risk of bias Magnitude of effect – Inconsistency Confounding controlled for – Imprecision Dose response – Publication bias – Irrelevance  Integrate across evidence streams –
  • 11.
    Causal conclusions Integratingevidence Level of evidence in human studies High Strong Moderate Weak Modest Low Inconclusive Weak Modest Low Moderate High Level of evidence in animal studies Supportive evidence from in vitro and xenograft models underpinning biological plausibility
  • 12.
    Milk and prostatecancer exemplar  Milk implicated in prostate cancer, but – is measured semi-quantitatively – is susceptible to confounding – large differences between individuals in the same group lead to attenuation by measurement errors  Systematic reviews of human observational studies are inconclusive – Limited-suggestive evidence (World Cancer Research Fund International, 2014)  Experimental studies not systematically reviewed
  • 13.
    Results of automatedmechanism quantification (36,000 hits)
  • 14.
    Targeted search results 4946 targeted studies identified 725 studies retrieved for detailed evaluation 4221 studies excluded after review of title and abstract (duplicates or clearly ineligible) Databases searched: • Medline • Embase • BIOSIS • CINAHL 268 studies potentially eligible 325 studies excluded after review of full text 132 studies awaiting ILL 23 milk-IGF studies 245 IGF-prostate cancer 22 human studies extracted (5 RCT, 3 cohort, 14 cross-sectional) 50 human studies extracted (total: 99) 7 animal studies extracted (total: 8) 138 Cell line studies
  • 15.
    Evidence synthesis Milkand IGF in human RCTs by increasing length of follow-up Study Rich-Edwards 2007 Hoppe 2004 Cadogan 1997 Zhu 2005 Ben-Shlomo 2005 Follow-up (yrs) .02 .02 1.5 2 25 % Male Control 0 100 0 0 52.8 Overall (I-squared = 69.0%, p = 0.012) NOTE: Weights are from random effects analysis ES (95% CI) 13.30 (-28.64, 55.24) 40.20 (-7.43, 87.83) 69.00 (5.89, 132.11) 28.60 (-17.66, 74.86) -9.50 (-16.75, -2.25) 20.87 (-8.75, 50.49) % Weight 19.59 17.59 13.09 18.06 31.66 100.00 -132 0 132 Favours control Favours intervention Difference in IGF-I (ng/ml) IGF and hallmarks of cancer (authenticated cell lines)
  • 16.
    Future work  Automated mechanisms discovery – Validation (not missing important studies) – Inter-relationships between mechanisms – Develop stand-alone automation software and beta testing  Validation of relevance questions  Acceptability  Reliability
  • 17.
    For further information Sarah Lewis: S.j.lewis@bristol.ac.uk Richard Martin: Richard.martin@bristol.ac.uk @wcrfint facebook.com/wcrfint   www.wcrf.org