Sharing Bioinformatic Data for Machine Learning: Maximizing

Interoperability through License Selection

Alexander Bernier and Adrian Thorogood

Centre of Genomics and Policy, McGill University Faculty of Medicine, Dr. Penfield, Montréal, Canada

Keywords: Big Data, Bioinformatics, Data Commons, Data Licensing, Intellectual Property, Interoperability, License

Standardization, Machine Learning, Open Science, Software Licensing, Technology Law.

Abstract: Efficient machine learning in bioinformatics requires a large volume of data from different sources.

Bioinformatics is shifting from a paradigm of siloed analysis of individual datasets by researchers to the

aggregation and analysis of disparate sets of health and biomedical data across from academic, healthcare and

commercial settings. Data generating organizations must give thought to selecting legal terms for dataset

release that will promote compatibility with other datasets. In releasing bioinformatic data for open use, care

must be taken to ensure that the terms of the licenses selected ensure maximum interoperability. The following

technical elements should inform the choice of license: License hybridity; waivers of liability, warranties and

guarantees; commercial/non-commercial use; attribution and copyleft; granular permission and bilateral or

multilateral licensing. Licenses are compared to inform optimal license selection and enable data integration

and analysis; consideration is given to an eventual standard license for open sharing of bioinformatic data.

1 INTRODUCTION

Machine learning needs ‘big data’ to be most

effective. According to the ‘Four Vs’ model, machine

learning requires a great “volume, velocity, variety,

and veracity” (Chiang, Grover, Liang & Zhang 2018,

p. 384) of data to create usable output, namely

“patterns with added value [that] … can be exploited

for the creation of wealth, the improvement of human

lives, and the advancement of knowledge” (Floridi,

2014, p. 16). Models are hungry for data. With the

exception of the data “nouveau riche” – tech giants

“such as Facebook [and] Google” – who have access

to immense proprietary data troves, many ML

researchers and companies must integrate data from

varied public and private sources (Floridi, 2014, p.

16). Open-access data sources are often accompanied

by terms of use that can be especially restrictive for

commercial users, or for the development of

commercial algorithms or prediction services. These

threatening compatibility across databases and

effectively restricting integration practices (Carbon et

al., 2019).

Bioinformatics stands at the precipice of a sea

change. Presently, bioinformatic research is

principally performed on isolated datasets. Going

forward, single-institution efforts by academic

researchers applying traditional bioinformatics

methodologies to datasets generated for their

particular purposes will be complemented by large-

scale big data efforts performed by commercial and

institutional entities integrating significant volumes

of data from academia, healthcare and industrial

research. In actualizing this vision, data generators

will need to ensure that the legal rights attached to

disparate datasets are standardized or at least

compatible. Failing to do so could impose significant

costs to understand and comply with the legal

limitations on using each dataset and could lead to

valuable datasets being impossible to combine

lawfully with others, frustrating big data

bioinformatics. Consequently, we contend that the

bioinformatics community must become adept at

understanding license terminology and reading

standard licenses so that it can license its data in ways

that promote interoperability. We canvass a number

of the common elements of data and IP licenses and

illustrate how these can affect interoperability.

Existing standard licenses are presented in a table that

compares how they address the elements discussed

(Appendix, Figure 1).

226

Bernier, A. and Thorogood, A.

Sharing Bioinformatic Data for Machine Learning: Maximizing Interoperability through License Selection.

DOI: 10.5220/0009179502260232

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 3: BIOINFORMATICS, pages 226-232

ISBN: 978-989-758-398-8; ISSN: 2184-4305

We consider the potential for standard data

licenses to create or to alleviate barriers to

bioinformatic data aggregation for machine learning.

The following technical elements of licenses are

addressed: License hybridity; liability, warranties and

guarantees; license duration and conditions of

termination; commercial and non-commercial use;

attribution and copyleft; permissions granted; and

standardization. Finally, we compare how existing

standard licenses address the issues discussed to

guide data holders in selecting an appropriate data

license. The licenses are : Creative Commons Zero

(CC0), Creative Commons Attribution 4.0 (CC-BY),

the Montreal Data License (MT-DL), the Microsoft

Open Use Data Agreement (O-UDA), The Microsoft

Computational Use Data Agreement (C-UDA) the

Microsoft Data Use Agreement for Open AI Model

Development (DUA-OAI), the Linux Community

Data License Agreement – Sharing (CDLA-Sharing),

the Linux Community Data License Agreement –

Permissive (CDLA-Permissive), the Open Data

Commons Open Database License (ODC-ODL), the

Open Data Commons Attribution License (ODC-

BY), and the Open Data Commons Public Domain

Dedication and License (ODC-PDDL).

2 LICENSE FEATURES

2.1 License Hybridity

2.1.1 Copyright and sui generis Database

Rights

intellectual property rights that may or may not apply

to datasets or the discrete data points that compose

them depending on the jurisdiction. The European

Economic Area, Mexico, Russia and South Korea

recognize sui generis database rights (Doldirina et al.,

2018). Copyright applies internationally. Local

minimum standards for the application of these rights

can vary as to the minimum threshold of human

‘creativity,’ ‘originality,’ (Ebrahim, 2019, Stokes

2019) ‘investment,’ or ‘structure’ (Gervais, 2019)

demonstrated in organizing a dataset. Further, the

breadth of exceptions to such rights varies across the

world. For instance, the United States permits liberal

use even absent necessary permissions, for

computational purposes, under the doctrine of

transformative ‘fair use’ (Liu, 2019). Conversely,

Europe does not recognize such exceptions or limits

them to non-commercial uses (Margoni &

Kretschmer, 2018). The uncertain application of

licensing selection, as we discuss below. We

recommend that bioinformaticians prefer ‘hybrid’

licenses that act as both intellectual property

instruments and contracts, because this will increase

their likelihood of being applied by a court even if the

existence of underlying rights in the data, and the

existence of a valid contract between licensor and

recipient are indeterminate.

2.1.2 The Enforceability and Benefits of the

Hybrid License

Licensing discussions are often bogged down by

uncertainty over underlying IP rights. To avoid this,

we recommend selecting hybrid licenses that contain

the same content in a contract and in an intellectual

property license. The document will bind data users

insofar as a valid contract is recognized and third

parties insofar as intellectual property interests are

recognized in the licensed data. Courts in the United

States (Madison, 1998) and Europe (Ryanair Ltd v PR

Aviation BV, 2015) have recognized such ‘hybrid’

instruments as valid. A hybrid license can

contractually increase the rights of a licensor in the

face of overly permissive IP laws, clearly signal to a

data user that the IP holder has disclaimed any IP

rights in the dataset, and harmonize all the legal

regimes applicable to a dataset where it is unclear

which IP rights, if any, would apply by default.

We recommend selecting licenses that will

address copyright, moral rights, sui generis database

rights, and clearly state their dual contractual-IP

license nature. Failure to do so could lead to the

license’s effect not mirroring the intent of the

licensor. The contractual elements of the hybrid

license can limit the potential for inconsistent

interpretations of IP law to leave ambiguous the rights

and obligations of direct parties to the agreement. The

IP elements of the hybrid license create some

certainty as to the rights and obligations of all third

parties, even if ambiguities arise as to the contract’s

validity. By favoring such licenses, bioinformaticians

can create clarity as to the rights granted in their data.

This promotes the more widespread use of their data,

permits data users to more easily understand if the

rights in that data is compatible with the rights in

other data, and decreases the risk of license

misinterpretation leading to conflict or litigation.

2.2 Commercial/Non-commercial Use

Reserving open data for non-commercial uses may

preclude machine learning altogether, as many

Sharing Bioinformatic Data for Machine Learning: Maximizing Interoperability through License Selection

227

applications necessitate resources and expertise only

available to sophisticated private-sector entities

(Doherty et al., 2016). Further, partnerships across

academic institutions and the private sector are the

rule rather than the exception in this area, with both

parties pooling resources including data, capital,

computing power and expertise. Commingling

private sector and public sector data and resources

benefits public-sector researchers in giving them

access to rigorously assembled pools of industry data

and permitting them to pursue research goals of

academic interest that do not lend themselves to

obvious profitability (Perkman & Schildt, 2015).

Moreover, the boundaries of ‘commercial’ and ‘non-

commercial’ use, and ‘commercial’ and ‘non-

commercial’ actors can be ambiguous. Machine

learning also presages the merger of data and open-

source software licensing – the latter community

considers commercial use restrictions discriminatory.

Consequently, we strongly recommend that

bioinformaticians avoid licenses that preclude

commercial use; such licenses will likely prevent

their data from being used for big data applications.

2.3 Waivers of Liability, Warranties

and Guarantees

2.3.1 General Liability, Warranties and

Guarantees

Assuming data generators share data openly on a

voluntary basis, licenses must be friendly to them in

order to promote data availability. Likewise, if

licenses are unfriendly towards users (in terms of

being too restrictive or conditional), this will

discourage data use.

In selecting a license, licensors should consider

the degree of responsibility that best reflects their

ability to affirm their rights in, and the quality of, the

data, as well as their risk tolerance as regards liability.

Waivers of liability, and disclaimers regarding the

licensor’s rights in the data, and the quality, accuracy

and merchantability thereof are common features of

licenses. The licensor is better placed than the user to

assess the aforementioned features, but open

licensing generally provides the licensor little benefit

(Wilka, Landry & McKinney, 2018). Consequently,

we consider that a license that contains the traditional

disclaimers, but also affirms that the licensor “has

exercised reasonable care to assure” the disclaimed

feature is generally a good compromise position.

Nonetheless, bioinformaticians should carefully

consider what guarantees to grant data users. If more

guarantees are made, data users may feel more

comfortable using the data. If less guarantees are

made, data licensors may be more protected from the

legal risks inherent in making their data available.

2.3.2 Data Protection and Privacy Laws

The General Data Protection Regulation (GDPR) and

other data protection legislation modeled on it

assesses the right to use data on a subjective

‘controller-by-controller’ basis; the question is not

‘can this data be lawfully used’ but ‘can you lawfully

use this data.’ Further, data protection is a markedly

localized (Custers, Dechesne, Sears, Tani & Hof,

2018) and sector-specific (Archer & Delgadillo,

2016) legal regime. Therefore, it is generally difficult

for the data licensor to make any meaningful

representation to the recipient regarding data

protection.

Bioinformatic and associated health data is often

subject to data protection laws, and sometimes to

more onerous laws or provisions that hold health

information to a higher standard of protection (Kim,

Kim & Joly, 2018; Thorogood 2018). Data licensors

should consider such legislation before licensing their

data, especially if intending to use an open license or

public dedication. Presently, most standard data

licenses do not address data protection. This could

conflict with obligations under some data protection

laws (e.g. the GDPR) to distribute responsibilities

among data controllers using controllership

agreements (Wrigley, 2019). Other data protection

laws impose accountability requirements that may

favor using licenses and contracts that address data

protection (Centre for Information Policy Leadership,

2018). Consequently, we recommend that

bioinformaticians use licenses that address data

protection, or create data protection annexes is

licensing data. In doing so, they should remain

mindful of the highly mutable character of data

protection obligations across countries and for

different entities. We caution them not to rely

exclusively on generalized statements about data

protection responsibilities in contracts or licenses.

2.4 License Duration and Conditions of

Termination

A licensing challenge for machine learning is that it

tends to depend on long-term, potentially indefinite,

access to a same pool of data in a number of contexts.

(Wilka et al., 2018).

Considering that data users will be integrating a

large number of datasets to create a single machine

learning algorithm, the loss of even a single dataset

BIOINFORMATICS 2020 - 11th International Conference on Bioinformatics Models, Methods and Algorithms

228

could mar their ability to replicate a certain algorithm,

or to determine if tuning or retraining on a modified

dataset is improving the functioning of an algorithm.

(Lehr & Ohm, 2017). For these reasons, it is our

recommendation that licensors select licenses that

ensure data user breach does not immediately lead to

the termination of the license, and that data providers

not have the right to unilaterally terminate the license.

Further, license terms should be indefinite, renewed

automatically on expiration, or renewed at the

discretion of the data user rather than the licensor.

2.5 Attribution and Copyleft

2.5.1 Forms of Open Licensing

Open licenses come in many variants. ‘Permissive’

licenses openly release the work without imposing

any limitations on its use. Consequently, recipients

are free to create derivative works and commercialize

those derivatives, or impose IP protections thereon.

‘Copyleft’ licenses restrict the recipients from

imposing IP protections on the licensed work, and can

impose the obligation to license derivative material

downstream on equally permissive terms. ‘Strong

copyleft’ requires distribution of the licensed work or

a derivate work under the same terms. ‘Weak

copyleft’ requires distribution of the licensed work or

a derivative work under the same terms, but permits

the combination of the licensed work with other

proprietary works (e.g. combining software) under

different terms (Hall, 2017). Attribution requirements

impose an obligation to attach an attribution to the

licensed work, sometimes in a prescribed form.

A ‘public domain dedication’ is another popular

mechanism for attempting to eliminate all of the

licensor’s rights in the concerned works. Not all

countries recognize the lawfulness thereof.

Traditionally, the United States has been permissive

in allowing public dedications (Johnson, 2008) and

European jurisdictions more reticent (Aishwarya,

2017). Bioinformatics communities should decide

what modality of open license best conforms to their

values, as ensuring compatibility even between

standard licenses is problematic (OpenMinted, n.d.).

2.5.2 Considerations for Machine Learning

In licensing data for data integration, imposing an

attribution requirement on a dataset can frustrate the

combination of datasets from different sources. The

barrier can be resource-based, in that it is

insurmountably time-consuming to attribute a large

number of datasets used to train a ML algorithm. The

barrier can also be rule-based, in that different

licenses impose attribution formats that are

incompatible such that dual compliance is impossible

(Morando, 2013). The most potent algorithmic

models ensue from the combination of disparate

datasets (Mattioli, 2018).

Copyleft requirements can hamper the

interoperability of datasets for ML to varying degrees

depending on their formulation. Strong copyleft

requirements can pose the same problem as

attribution requirements; it could prove impossible to

comply with the conflicting copyleft requirements of

two datasets. This would preclude the combination

thereof.

Copyleft further has the potential to create siloes

of ‘copyleft’ and ‘non-copyleft’ data. Data recipients

hoping to create proprietary technologies from open

data are barred from using the ‘copyleft’ data lest

their output become ‘infected’ and unfit for their

private commercial purposes (Thorogood, 2019).

From the narrow standpoint of license

interoperability, we recommend avoiding strong

copyleft or attribution clauses.

2.6 Granular Permissions and Bilateral

or Multilateral Licensing

Open licensing generally aspires to the broadest

possible permissions, both as regards the parties

concerned and the rights in the data. Nonetheless,

concerns of data sensitivity, or a desire to foster

innovation while safeguarding the right to profit from

the data in the future can prohibit totally open

licensing (Benjamin et al., 2019). Of the licenses

addressed, the Montreal Data License is unique in that

it allows for the negotiation of bilateral contracts

between parties for particular tiers of rights in data,

using language specific to machine learning and

algorithmic modelling. Data licensors with an

appetite for data release, but who are wary of the

privacy violations or commercial opportunities lost in

public licensing may want to consider this license.

3 CONCLUSIONS: TOWARD A

STANDARD LICENSE

A final consideration in licensing data is that of

standardization. The emergence of competing

standards does not necessarily reflect the failure of

the community to reach consensus, but rather that the

culture of openness varies across the academic,

bioresources, open patenting, and software

Sharing Bioinformatic Data for Machine Learning: Maximizing Interoperability through License Selection

229

development communities (Liddell, Liddicoat,

Jordan & Schovsbo, 2019). License selection does not

necessarily mean selecting the most optimal among

competing options; it reflects the subjective balancing

of differing values. Presently, licensors of

bioinformatic data must decide which of the existing

options best reflect their objectives. In the future, a

standard license for bioinformatic data sharing could

benefit the scientific community by ensuring that

bioinformaticians have tools for data sharing that

enshrine the values of their community. Further,

achieving true interoperability may require license

standardization, as combining datasets across licenses

could create legal ambiguities and inefficient costs.

(Morando, 2013). Achieving standardization will

require not only appropriate license selection by

individuals but successful consensus-building across

bioinformatics communities. License literacy will be

instrumental in drafting and selecting the licenses

needed to make big data bioinformatics a reality and

pool data across academia, healthcare and industry.

ACKNOWLEDGEMENTS

The authors graciously thank Genome Canada,

Genome Québec and the Canadian Institutes for

Health Research for their financial support.

REFERENCES

Aishwarya, S. (2017). The Nature and Enforceability of Open

Source License. 11 NUALS Law Journal, 11, 53–86.

Archer, J. K., & Delgadillo, C. A. (2016). Key Data

Ownership, Privacy and Protection Issues and

Strategies for the International Precision Agriculture

Industry. Proceedings of the 13th International

Conference on Precision Agriculture.

Benjamin, M., Gagnon, P., Rostamzadeh, N., Paul, C.,

Bengio, Y., & Shee, A. (n.d.). Towards Standardization

of Data Licenses: The Montreal Data License. eprint

arXiv:1903.12262

Carbon, S., Champieux, R., Mcmurry, J. A., Winfree, L.,

Wyatt, L. R., & Haendel, M. A. (2019). An analysis and

metric of reusable data licensing practices for

biomedical resources. Plos One, 14(3). doi:

10.1371/journal.pone.0213090

Centre for Information Policy Leadership. (2018). The Case

for Accountability: How it Enables Effective Data

Protection and Trust in the Digital Society (p. 15).

Chiang, R. H. L., Grover, V., Liang, T.-P., & Zhang, D.

(2018). Strategic Value of Big Data and Business

Analytics. Journal of Management Information

Systems, 35(2), 383–387. doi:

10.1080/07421222.2018.1451950

Custers, B., Dechesne, F., Sears, A. M., Tani, T., & Hof, S.

V. D. (2018). A Comparison of Data Protection

Legislation and Policies Across the EU. Computer Law

and Security Review, 34(2), 234–243. doi:

10.2139/ssrn.3091040

Doherty, M., Metcalfe, T., Guardino, E., Peters, E., &

Ramage, L. (2016). Precision medicine and oncology:

an overview of the opportunities presented by next-

generation sequencing and big data and the challenges

posed to conventional drug development and regulatory

approval pathways. Annals of Oncology, 27(8), 1644–

1646. doi: 10.1093/annonc/mdw16

Doldirina, C., Eisenstadt, A., Onsrud, H., & Uhlir, P.

(2018). Legal Approaches for Open Access to Research

Data. doi: 10.31228/osf.io/n7gfa

Ebrahim, T. Y. (2019). Data-Centric Technologies: Patent

and Copyright Doctrinal Disruptions. Nova Law

Review, 43(3).

Floridi, L. (2014). Chapter I: Time: Hyperhistory. In The

fourth revolution: how the infosphere is reshaping

human reality (pp. 1–24). Oxford: Oxford University

Press

Gervais, D. J. (2019). Exploring the Interfaces Between Big

Data and Intellectual Property Law. Intellectual

Property, Information Technology & Electronic

Commerce Law, 22. doi: 10.2139/ssrn.3360344

Hall, A. J. (2017). Open-Source Licensing and Business

Models: Making Money by Giving It Away. Santa

Clara High Technology Law Journal, 33(3), 427–437.

Johnson, P. (2008). Dedicating Copyright to the Public

Domain. Modern Law Review, 71(4), 587–610. doi:

10.1111/j.1468-2230.2008.00707.x

Kim, H., Kim, S. Y., & Joly, Y. (2018). South Korea: in the

midst of a privacy reform centered on data sharing.

Human Genetics, 137(8), 627–635. doi:

10.1007/s00439-018-1920-1

Lehr, D., & Ohm, P. (2017). Playing with the Data: What

Legal Scholars Should Learn About Machine Learning.

University of California, Davis Law Review, 51, 653–

717.

Liddell, K., Liddicoat, J., Jordan, M., & Schovsbo, J.

(2019). IP policies for Large Bioresources: the Fiction,

Fantasy and Future of Openness. In T. Minssen & R. J.

Herrmann (Eds.), Global Genes, Local Concerns:

Legal, Ethical and Scientific Challenges in

International Biobanking. (pp. 258–280). Edward Elgar

Publishing.

Liu, J. (2019). An Empirical Study of Transformative Use

in Copyright Law. Stanford Technology Law Review,

22 (Winter), 163–241.

Madison, M. J. (1998). Legal-Ware: Contract and

67(3), 1025–1143. doi: 10.31228/osf.io/4y2h6

Margoni, T., & Kretschmer, M. (2018). The Text and Data

Mining Exception in the Proposal for a Directive on

what EU copyright law needs. UK Copyright and

Creative Economy Centre University of Glasgow

Technical Report.

BIOINFORMATICS 2020 - 11th International Conference on Bioinformatics Models, Methods and Algorithms

230

Mattioli, M. (2018). The Data-Pooling Problem. Berkeley

Technology Law Journal, 32(1), 179–235. doi:

10.2139/ssrn.2671939

Morando, F. (2013). Legal Interoperability: Making Open

(Government) Data Compatible with Businesses and

Communities. Italian Journal of Library, Archives and

Information Science, 4(1), 441–452.

OpenMinted. (n.d.). Accessible at:

https://openminted.github.io/releases/interop-

spec/1.0.0/openminted-interoperability-scenarios/.

Perkmann, M., & Schildt, H. (2015). Open data

partnerships between firms and universities: The role of

boundary organizations. Research Policy, 44(5), 1133–

1143. doi: 10.1016/j.respol.2014.12.006

Stokes, S. (2019). Chapter 2: Digital Copyright, the Basics.

In Digital Copyright: Law and Practice Fifth Edition.

Hart Publishing, Bloomsbury Publishing Plc.

Ryanair Ltd v PR Aviation BV, ECLI:EU:C:2015:10 (E.C.J.

Chamber. 2015).

Thorogood, A. (2018). Canada: will privacy rules continue

to favour open science? Human Genetics, 137(8), 595–

602. doi: 10.1007/s00439-018-1905-0

Thorogood, A. (2019). Towards Legal Interoperability in

International Health Research. University of Toronto

LLM Thesis.

Wilka, R., Landry, R., & McKinney, S. A. (2018). How

Machines Learn: Where Do Companies Get Data for

Machine Learning and What Licenses Do They Need.

Washington Journal of Law, Technology and Arts,

13(3), 217–244.

Wrigley, S, (2019.). “When People Just Click”: Addressing

the Difficulties of Controller/Processor Agreements

Online. In M. Corrales, M. Fenwick & H. Haapio

(Eds.), (pp. 221–252). Kyushu University, Springer.

APPENDIX

License

Hybridity /

Rights

Concerned

Liability, Warranties and

Guarantees

Duration

Commercial

and Non-

Commercial

Use

Attribution and Copyleft Parties

Creative

Commons Zero

(CC0)

Waiver / license

of copyright, sui

generis rights,

moral rights

No representations or

warranties, guarantees,

disclaimer of liability.

Disclaimer of responsibility

for clearing rights in data.

Indefinite No limitation No requirement One to all.

Creative

Commons

Attribution 4.0

International (CC-

BY)

Hybrid contract

(implied) / license

of copyright, sui

generis, waiver of

moral rights

No representations or

warranties, disclaimer of

liability. Disclaimer of

responsibility for clearing

rights in data.

Indefinite, immediate termination

for non-compliance, rectification

within 30 days reinstates license,

otherwise consent is required.

No limitation.

Specific attribution requirements (prescribed

format).

One to all.

Montreal Data

License (MT-DL)

Not specified,

resumably a pure

contract.

Exclusion of warranties,

guarantees, disclaimer of

liability.

Unspecified duration. Immediate

termination on breach.

Licensor’s

option.

Licensor’s option. One to one.

Microsoft Open

Use Data

Agreement (O-

UDA)

Not specified.

Disclaimer of warranties,

limitation of liability for

licensor and upstream

licensors. No warranty of

rights in data. Licensor

agrees not to sue recipient

and downstream recipient

absent breach.

Unspecified duration. No

provisions regarding voluntary

termination or termination for

cause.

No limitation.

Attribution for source and modified data. Must

impose warranty disclaimer and limitation of

liability for upstream controllers on the downstream

recipients.

No attribution, warranty, or limitation of liability,

requirement for output, so long as the output does

not contain more than a ‘de minimis’ portion of the

data.

One to all.

Microsoft

Computational

Use Data

Agreement (C-

UDA)

Not specified.

Rights limited to

computational

use.

Disclaimer of warranties,

limitation of liability for

licensor and upstream

licensors. No warranty of

rights in data. Licensor

agrees not to sue recipient

and downstream recipient

Unspecified duration. No

provisions regarding voluntary

termination or termination for

cause.

No limitation.

Attribution for source and modified data. No

attribution requirement for output.

Copyleft for data (same license must be applied).

No copyleft for output or algorithms unless these

contain more than a ‘de minimis’ portion of the

data.

One to all.

Microsoft Data

Use Agreement

for Open AI

Model

Development

(DUA-OAI)

Not specified.

Rights in data

limited to training

the AI model.

No right to share

or distribute the

data or assign

license.

Parties each warrant and

represent compliance with

laws, including data

protection laws.

Data user warrants rights in

the untrained AI model.

Data licensor does not

warrant rights in or quality

of data. Licensor warrants

that they are not aware of

restrictions that would limit

use or distribution.

Optional limitation of

liability clause, with

exception for damages

caused by the data

recipient’s breach of the

license.

Duration of one year.

Termination with notice after 90

days, termination for breach 30

days after notification of breach,

if not cured.

No limitation.

Copyleft in the trained AI model – must publicly

release the trained AI model under an open software

license that includes a general disclaimer of liability

in favor of the data licensor.

One to one.

Figure 1: Standard License Comparison Table.

Sharing Bioinformatic Data for Machine Learning: Maximizing Interoperability through License Selection

231

Linux Community

Data License

Agreement –

Sharing (CDLA-

Sharing)

Hybrid contract

(implied) / license

of copyright, sui

generis, waiver of

moral rights.

Parties each warrant and

represent reasonable care in

ensuring use in compliance

with rights of others,

privacy and confidentiality.

Disclaimer of warranties

and limitation of liability.

Termination for data recipient’s

breach if not rectified within a

‘reasonable time’ of becoming

aware.

Termination if litigation against

data provider or data recipient

concerning dispute not related to

the data license.

No limitation.

Attribution of source and modified data; flagging of

modified data, integration of those notices into the

data files. No attribution requirement for output /

results unless these contain more than a ‘de minimis’

portion of the data.

Copyleft for data (same license must be applied); no

copyleft for output / ‘results’ unless these contain

more than a ‘de minimis’ portion of the data.

Explicit preclusion of restriction using technological

measures.

One to all.

Linux Community

Data License

Agreement –

Permissive

(CDLA-

Permissive)

Hybrid contract

(implied) / license

of copyright, sui

generis, waiver of

moral rights.

Parties each warrant and

represent reasonable care in

ensuring use in compliance

with rights of others,

privacy and confidentiality.

Disclaimer of warranties

and limitation of liability.

Termination for data recipient’s

breach if not rectified within a

‘reasonable time’ of becoming

aware.

Termination if litigation against

data provider or data recipient

concerning dispute not related to

the data license.

No limitation.

Attribution of source and modified data; flagging of

modified data, integration of those notices into the

data files. No attribution requirement for output /

results unless these contain more than a ‘de minimis’

portion of the data.

Modified data or a combination of original and

modified data can be released under a different

license.

No copyleft for output / ‘results’ unless these contai

more than a ‘de minimis’ portion of the data.

One to all.

Open Data

Commons Open

Database License

(ODC-ODL)

Explicit hybrid

contract / license

generis, waiver of

moral rights.

Disclaimer of warranties

and exclusion of liability.

Immediate termination for

breach; can be reinstated if first

breach and rectifies within 30

days of notice of breach.

Otherwise reinstated 60 days after

cessation of breach, if licensor

does not send notice of permanent

termination in that time.

No limitation.

Attribution requirement for the data/base, a

derivative data/base, or output.

Copyleft, must license the data/base or a derivative

database under the same license or “a compatible

license.”

No additional legal “terms or technological

measures” can be imposed, excepting a limited right

to ‘parallel release.’

No copyleft for output.

One to all.

Open Data

Commons

Attribution

License (ODC-

BY)

Explicit hybrid

contract / license

of copyright and

sui generis,

waiver of moral

rights.

Disclaimer of warranties

and exclusion of liability.

Immediate termination for

breach; can be reinstated if first

breach and rectifies within 30

days of notice of breach.

Otherwise reinstated 60 days after

cessation of breach, if licensor

does not send notice of permanent

termination in that time.

No limitation.

Attribution requirement for data/base, modified

database, or output.

Copyleft for data/base or derivative database (same

license must be applied).

One to all.

Open Data

Commons Public

Domain

Dedication and

License (ODC-

PDDL)

Public domain

dedication of

database rights /

license of

generis, and

waiver of moral

rights.

Disclaimer of warranties

and exclusion of liability.

Indefinite. No limitation. No attribution or copyleft requirement. One to all.

Figure 1: Standard License Comparison Table (continued).

BIOINFORMATICS 2020 - 11th International Conference on Bioinformatics Models, Methods and Algorithms

232