What is Reproducibility? The R* brouhaha (and how Research Objects can help)

What is Reproducibility?

The R* brouhaha(and how Research Objects can help)

Professor Carole GobleThe University of Manchester, UKSoftware Sustainability Institute, UKELIXIR-UK, FAIRDOM Association e.V.carole.goble@manchester.ac.ukFirst International Workshop on Reproducible Open Science @ TPDL, 9 Sept 2016, Hannover, Germany

Acknowledgements• Dagstuhl Seminar 16041 , January 2016

– http://www.dagstuhl.de/en/program/calendar/semhp/?semnr=16041• ATI Symposium Reproducibility, Sustainability and Preservation , April

2016– https://turing.ac.uk/events/reproducibility-sustainability-and-preservation/– https://osf.io/bcef5/files/

• C Titus Brown• Juliana Freire• David De Roure• Stian Soiland-Reyes• Barend Mons• Tim Clark• Daniel Garijo• Norman Morrison

“When I use a word," Humpty Dumpty said in rather a scornful tone, "it means just what I choose it to mean - neither more nor less.”

Carroll, Through the Looking Glass

re-compute

replicatererun

repeat

re-examine

repurpose

recreate

reuse

restorereconstruct review

regeneraterevise

recycle

redo

robustness tolerance

verification compliance validation assurance

remix

Reproducibility of Reproducibility Research

Computational Science

http://tpeterka.github.io/maui-project/From: The Future of Scientific Workflows, Report of DOE Workshop 2015, http://science.energy.gov/~/media/ascr/pdf/programdocuments/docs/workflows_final_report.pd

1. Observational, experimental

2. Theoretical3. Simulation4. Data

intensive

BioSTIF

Computational Science

Scientific publications goals: (i) announce a result(ii) convince readers its correct.

Papers in experimental science should describe the results and provide a clear enough protocol to allow successful repetition and extension.

Papers in computational science should describe the results and provide the complete software development environment, data and set of instructions which generated the figures.

Virtual Witnessing*

*Leviathan and the Air-Pump: Hobbes, Boyle, and the Experimental Life (1985) Shapin and Schaffer.

Jill Mesirov

David Donoho

Datasets, Data collectionsStandard operating proceduresSoftware, algorithmsConfigurations, Tools and apps, services

Slide

share

Github

figsh

are

Commun

ityDB

Arxiv.o

rg

Pubm

ed

Docke

rim

age

Codes, code librariesWorkflows, scriptsSystem software Infrastructure Compilers, hardware

Systems of SystemsHeterogeneous hybrid patchwork of tools and service evolving over time

10 “Simple” Rules for Reproducible Computational

Research: RACE1. For Every Result, Keep Track of How It

Was Produced2. Avoid Manual Data Manipulation Steps3. Archive the Exact Versions of All

External Programs Used4. Version Control All Custom Scripts5. Record All Intermediate Results, When

Possible in Standardized Formats6. For Analyses That Include Randomness,

Note Underlying Random Seeds7. Always Store Raw Data behind Plots8. Generate Hierarchical Analysis Output,

Allowing Layers of Increasing Detail to Be Inspected

9. Connect Textual Statements to Underlying Results

10.Provide Public Access to Scripts, Runs, and Results

Sandve GK, Nekrutenko A, Taylor J, Hovig E (2013) Ten Simple Rules for Reproducible Computational Research. PLoS Comput Biol 9(10): e1003285. doi:10.1371/journal.pcbi.1003285

Record Everything

Automate Everything

Contain Everything

ExposeEverything

Preparation painindependent testing trials and

tribulations

[Norman Morrison]

replication hostility no funding, time, recognition, place to publishresource intensive access to the complete environment

Lab Analogy: Witnessing “Datascopes”

Input Data

Software

Output Data

ConfigParameters

Methodstechniques, algorithms, spec. of the steps, models

Materialsdatasets, parameters, algorithm seedsExperim

ent

Instrumentscodes, services, scripts, underlying libraries, workflows, , ref resourcesLaboratory

sw and hw infrastructure, systems software, integrative platformscomputational environment

Setup

“Micro” Reproducibility

“Macro” Reproducibility

Fixivity

Validate

Verify

Trust

Repeat, Replicate, Robust

[C Titus Brown]

https://2016-oslo-repeatability.readthedocs.org/en/latest/repeatability-discussion.html

Why the differences?

Reproduce, Trust

“an experiment is reproducible until another

laboratory tries to repeat it” Alexander Kohn

Repeatability:“Sameness”Same result1 Lab1 experiment

Reproducibility:“Similarity”Similar result> 1 Lab> 1 experiment

Validate

Verify

Method Reproducibility

the provision of enough detail about study procedures and data so the same procedures could, in theory or in actuality, be exactly repeated.

Result Reproducibility (aka replicability)

obtaining the same results from the conduct of an independent study whose procedures are as closely matched to the original experiment as possible

What does research reproducibility mean? Steven N. Goodman, Daniele Fanelli, John P. A. Ioannidis Science Translational Medicine 8 (341), 341ps12. [doi: 10.1126/scitranslmed.aaf5027] http://stm.sciencemag.org/content/scitransmed/8/341/341ps12.full.pdf

ProductivityTrack differences

Validate

Verify

reviewers want additional workstatistician wants more runsanalysis needs to be repeatedpost-doc leaves, student arrivesnew/revised datasetsupdated/new versions of algorithms/codessample was contaminatedbetter kit - longer simulationsnew partners, new projects

Personal & Lab

Productivity

Public GoodReproducibili

ty

“Datascope” Lab Analogy

Methodstechniques, algorithms, spec. of the steps, models

Materialsdatasets, parameters, algorithm seedsExperim

ent

Instrumentscodes, services, scripts, underlying libraries, workflows, ref datasets

Laboratorysw and hw infrastructure, systems software, integrative platformscomputational environment

Setup

“Datascope” Lab Analogy

Methodstechniques, algorithms, spec. of the steps, models

Materialsdatasets, parameters, algorithm seedsExperim

ent

Instrumentscodes, services, scripts, underlying libraries, workflows, ref datasets

Laboratorysw and hw infrastructure, systems software, integrative platformscomputational environment

Setup Form

Function

“Datascope” Practicalities

Methodstechniques, algorithms, spec. of the steps, models

Materialsdatasets, parameters, algorithm seeds

Experim

ent

Instrumentscodes, services, scripts, underlying libraries, workflows, ref datasets

Laboratorysw and hw infrastructure, systems software, integrative platformscomputational environment

Setup

Living DependenciesScience, methods, datasetsquestions stay, answers change

breakage, labs decay, services and techniques come and go, new instruments, updated datasets, services, codes, hardware

One offs, streams,stochastics, sensitivities,scale, non-portable datablack boxes

supercomputer accessnon-portable softwarelicensing restrictionsunreliable resourcesblack boxescomplexity

T1 T2

evolving ref datasets,new simulation codes

EnvironmentArchived vs Active

Contained vs DistributedRegimented vs Free-for-

allWho owns the dependencies?

Dependencies -> Manage

Black boxes -> Expose

Dynamics -> FixityReliability

Replicate harder than Reproduce?

Repeating the experiment or the set up?

Container Conundrum Results will Vary

Replicability WindowAll experiments become less replicable over

timePrepare to repair

Levels of Computational Reproducibility

Coverage: how much of an experiment is reproducible

Orig

inal

Exp

erim

ent S

imila

r Exp

erim

ent Di

ffere

nt E

xper

imen

tPo

rtabi

lity

Depth: how much of an experiment is available

Binaries + Data

Source Code / Workflow+ Data

Binaries + Data + Dependencies

Source Code / Workflow+ Data + Dependencies

Virtual MachineBinaries + Data + Dependencies

Virtual MachineSource Code / Workflow+ Data + Dependencies

Figures + Data

[Freire, 2014]

Minimum: data and source code available under terms that permit inspection and execution.

Measuring Information Gain from Reproducibility

Research goal

Method/Alg.

Platform/Exec Env

Data Parameters

Input data

Actors

Information Gain Consistency

Robustness/S

ensitivity

Generality

Portability/Adoption 1

Portability/Adoption 2

Independent validation

Repurposability

Implementation/Code

No changeChangeDon’t care

https://linkingresearch.wordpress.com/2016/02/21/dagstuhl-seminar-report-reproducibility-of-data-oriented-experiments-in-e-scienc/

http://www.dagstuhl.de/16041

How? Preserve by Reporting, Reproduce by Reading

Archived Record

Description Zoostandards, common metadata

How? Preserve by Maintaining, Repairing, ContainingReproduce by Running, Emulating, Reconstructing

Active Instrument Byte level Buildability Zoo

provenance

portability, preservation

robustness, versioning

access descriptionstandards

common APIslicensing, identifiers

standards,common metadata

change variation sensitivity

discrepancy handling

packaging, containers

FAIR RACE Reproducibility Dimensions

dependenciessteps

Research ObjectStandards-based metadata framework for logically and physically bundling resources with context,

http://researchobject.org

Bigger on the inside than the outsideexternal referencing

Manifest Constructi

on

Aggregates link things

togetherAnnotations

about things & their

relationships

Container

Research Object Standards-based metadata framework for logically and physically bundling resources with context, http://researchobject.org

Packaging content & links: Zip files, BagIt, Docker

images

Catalogues & Commons Platforms: FAIRDOM

Manifest Descripti

onDependencies

what else is needed

Versioning its evolution

Checklists what should be there

Provenance

where it came from

Identificationlocate things

regardless whereid

Systems Biology Commons• Link data,

models and SOPs

• Standards• Span resources• Snapshot +

DOIs• Bundle and

export• Logical bundles

Belhajjame et al (2015) Using a suite of ontologies for preserving workflow-centric research objects, J Web Semantics doi:10.1016/j.websem.2015.01.003

application/vnd.wf4ever.robundle+zip

Workflow Research Objects exchange, portability and maintenance

*https://2016-oslo-repeatability.readthedocs.org/en/latest/overview-and-agenda.html

Asthma Research e-Lab

Dataset building and releasing

Standardised packing of Systems Biology models

European Space Agency RO Library

Large dataset management for life science workflows

LHC ATLAS experiments

Notre Dame U Rostock

Encyclopedia of DNA Elements

PeptideAtlas

Words matter.

Reproducibility is not a end. Its a means to an end.Beware reproducibility

zealots.

50 Shades of Reproducibility.form vs function

A conundrum: big co-operative data-driven science makes

reproducibility desirable but also means

dependency and change are to be expected.

Lab analogy for computational

science

Bonus Slides