How Data Scientists Work Together With Domain Experts in Scientific Collaborations: To Find The Right Answer Or To Ask The Right Question?

1 INTRODUCTION

• Large, complex scientific datasets exceed the analytic capability of a local team working within a single expertise domain.The paper connects this need for new research practices to open science’s advocacy for transparency, accessibility, reproducibility, and interdisciplinary collaboration.Source: §1 · p. 2
• Interdisciplinary collaborations have produced successes such as AlphaFold, but those examples coexist with reported collaboration tensions and challenges.Source: §1 · p. 2
• Data-centric collaborations face tensions linked to interdisciplinary work, differing motivations and cultures, remote teams, data-centric practices, and limited technology or infrastructure support.The paper also notes tensions between domain experts’ and data volunteers’ expectations in a civic data hackathon.Source: §1 · p. 2
• Biomedical research is an active open-science domain because projects often curate and integrate large datasets, while analytic or interpretive mistakes may have catastrophic consequences for human lives.Source: §1 · p. 2
• The study asks what challenges data scientists and biomedical scientists encounter across common ground, coupling of work, collaboration readiness, and technology readiness.The authors use semi-structured interviews with 22 participants and apply the Olson framework as an analytic lens, not as a framework to test.Source: §1 · p. 3

2.1 The Olsons’ Framework and Remote Scientific Collaborations

• The Olson framework describes coupling of work, common ground, collaboration readiness, and technology readiness; this study applies it to interdisciplinary open-science projects and examines possible mismatches with its high-common-ground best practices.Source: §2.1 · p. 3

2.2 Teams and Infrastructures in Data-Centric Open Science Projects

• Prior data-centric collaboration research examined selected workflow steps or particular tools, while this paper seeks a comprehensive account spanning both small and large teams.Open-science teams can take varied forms, including geographically distributed and self-organizing communities.Source: §2.2 · pp. 4–5

2.3 Interdisciplinary Collaboration Teams

• Interdisciplinary teams distribute domain knowledge and task understanding across experts, making knowledge sharing and integration important but costly.The literature describes a trade-off between sharing unique perspectives and the costs of pooling, exchanging, and integrating information.Source: §2.3 · p. 5

2.3.1 Third Space and Hybridity.

• Third-space theory describes a hybrid space at disciplinary boundaries where participants can preserve their own perspectives while negotiating and integrating differences.The authors propose that data scientists and biomedical scientists also construct such a space, potentially making existing support techniques relevant.Source: §2.3.1 · pp. 5–6

2.3.2 Common Ground: Content and Process.

• Content common ground concerns shared understanding of the work, whereas process common ground concerns shared and continually updated rules and practices for conducting it.The authors argue that biomedical collaborations distinguish domain content expertise from analytics process expertise, and that the two forms of common ground may develop differently over time.Source: §2.3.2 · p. 6

3.1 Participants

• The study interviewed 22 U.S.-based informants reporting on 26 projects, including data scientists and biomedical scientists across small teams and crowdsourcing collaborations.Participants averaged five years of experience in their expert domains; 16 identified primarily as data scientists and six as biomedical scientists.Source: §3.1 · p. 7

3.2 Semi-structured Interview

• Researchers conducted 19 face-to-face and three remote semi-structured interviews over three months in summer 2017, asking about collaboration experiences and recent projects.They also invited participants to provide project artifacts such as meeting notes, analysis results, and publications.Source: §3.2 · p. 7

3.3 Data Analysis and Verification

• Two coders first used deductive coding for Olson’s dimensions and common-ground subdimensions, then inductively developed themes and axial codes through iterative discussion and cross-checking.They added codes until theoretical saturation and revisited the framework and transcripts to resolve disagreements.Source: §3.3 · p. 7

4 FINDINGS

• The findings are organized in this order: coupling of work, collaboration readiness, technology readiness, and common ground.Source: §4 · p. 9

4.1.1 Common Workflow.

• The ideal workflow moves from biomedical data and questions through data-science translation and modeling to evaluation, but is iterative and can revise the research question.Informants also described brainstorming and rapidly testing prototypes together.Source: Figure 1 · p. 9, §4.1.1 · p. 9
• Small-group teams are managed by a team researcher, while large crowdsourcing projects use organizers or project managers to structure and monitor subteams.Source: §4.1.2 · p. 9

4.1.2 Team Structure and Coupling of Work.

• Small teams rely on timely communication for closely coupled work; crowdsourcing managers divide questions into subproblems and regulate coupling through clarification and participant support.Examples include forum feedback, webinars, and expert advisory boards.Source: §4.1.2 · pp. 9–10

4.2 Collaboration Readiness

• Reciprocal skills and resources initially attract collaborators, but motivation changes; small teams can become dependent on a few core members, frustrating them and reducing engagement.Source: §4.2.1 · p. 10

4.2.1 Challenges of Maintaining Motivation.

• Small teams risk management overload, while large crowdsourcing subteams avoid it in the short term but often struggle to sustain motivation beyond short projects and limited publication opportunities.These projects commonly last three to four months, and informants reported difficulty finding collaborators afterward.Source: §4.2.1 · p. 10

4.2.2 Reward Attribution and Over-Competing with Other Teams.

• Leaderboard rankings based on one metric can prioritize higher scores over biomedical meaning or new questions, encouraging repetitive solutions and limiting discovery.Source: §4.2.2 · pp. 10–11

4.3 Technology Readiness

• Teams use co-editing, communication, co-creation and version-control, data and code repositories, and expertise systems; technology-readiness challenges vary with collaborators’ backgrounds and over time.Source: §4.3 · p. 11

4.3.1 Information Needs and Tool Preferences.

• Biomedical scientists prioritize transparency and interpretability, while data scientists prioritize model performance, generalizability, and efficiency; their distinct needs shape tool preferences.Biomedical scientists wanted documentation of processing and alternatives, while data scientists sought reusable models and examples.Source: §4.3.1 · p. 12
• Teams often choose familiar overlapping tools instead of advanced options because persuasion and training take time; one challenge’s unified-tool requirement coincided with lower participation.Source: §4.3.1 · p. 12

4.3.2 Fragmented Information.

• Information scattered across tools and systems is especially difficult to track and synthesize in small teams without a dedicated management role.Informants described using separate local coding tools, documents, slides, email, and Slack.Source: §4.3.2 · p. 12

4.4.1 Formulating the Initial BMS and DS Research Questions.

• At project outset, small teams often converge on paired biomedical and data-science questions, with each discipline believing it is answering its own version.Biomedical scientists define a domain-specific question that is translated into a computable data-science question.Source: §4.4.1 · pp. 12–13
• Large crowdsourcing projects use expert panels and dry runs to check that proposed questions are meaningful, feasible, and manageable within the available time.Participants reported that planning such collaborations could take months.Source: §4.4.1 · p. 13

4.4.2 New Research Questions Emerge During the Project Process.

• As projects unfold, questions can shift toward better or more impactful directions, but unexpressed changes break common ground and can confuse teams.The shift may move collaborators from answering an initial question toward exploring what question to ask.Source: §4.4.2 · p. 13

4.4.3 Obscure Data.

• Biomedical data’s ambiguity, bias, missing context, and experiential structure knowledge complicate method evaluation and result interpretation, contributing to frustration and distrust.Relevant clinical context may be absent from metadata, while data structures vary in form, jargon, and routine across projects.Source: §4.4.3 · p. 14

4.4.4 Ask The Right Question.

• Biomedical scientists may seek alternative research questions but find the translation into data-science questions opaque, while prediction tasks offer mature methods and standardized loss functions.Informants described alternative hypotheses as important in biomedical research.Source: §4.4.4 · p. 14
• Biomedical scientists seek meaningful, intervention-relevant discovery, while data scientists may prioritize novel methods, predictive performance, and standardized evaluation.Biomedical problems often lack clear predictive outcomes, and this prediction-oriented tendency appears more severe in small teams.Source: §4.4.4 · p. 15
• In large crowdsourcing collaborations, an ensemble of winning teams’ models can aggregate different approaches’ focuses and merits to support new insight.Source: §4.4.4 · p. 15

5.1 Successful Collaborations In Scientific Discovery

• Small teams need timely coordination, while large crowdsourcing projects need loose coupling; tight coupling under proper management can also strengthen common ground and collaboration readiness.Many projects were non-divisible and highly iterative, making modular work and routine coordination across locations difficult.Source: §5.1.1 · p. 15

5.1.1 Coupling of Work, Collaboration Readiness, and Technology Readiness.

• Different research agendas can make data scientists’ methods and performance priorities misalign with biomedical scientists’ interest in interpretable, intervention-relevant results.The discussion recommends weighing both sides’ interests in team organization and notes that a broker can translate stakeholders’ goals; it expects this role to smooth collaboration.Source: §5.1.1 · p. 16
• A shared “third space” lets biomedical and data scientists develop hybrid language, negotiate shared tools, and revise the initial problem as new questions emerge.Collaboration readiness also involves organizational dependencies and relationships between teams over time, while tool choices often favor familiar options and interoperability matters as needs evolve.Source: §5.1.1 · pp. 15–16, §5.1.2 · p. 16

5.1.2 Ever-Evolving Common Ground and Better Scientific Discovery.

• Scientific discovery teams benefit from continually revising content common ground while building process common ground, rather than settling on one question or solution.Increasing process common ground can support breaking and updating content common ground; without it, teams may become confused or settle prematurely on less optimal questions or solutions.Source: §5.1.2 · pp. 16–17

5.2 Principles for Technology Design

• The discussion proposes tools that support divergent-to-convergent communication and distinguish evolving content knowledge from process protocols and their histories.A shared team view can coexist with role-specific details; suggested supports include brainstorming, summarizing, evaluating ideas, and tracking changes in shared objects.Source: §5.2 · p. 17

5.3 Project Management Guideline

• For small teams, management training, shared vocabulary and data documentation, and interoperable project-management tools may complement technical solutions.The discussion links these suggestions to difficulties managing fragmented, repetitive information and maintaining collaboration readiness without a specialized project manager.Source: §5.3 · p. 18

5.4 AI as a Partner in the Future of Data-Centric Scientific Discovery

• The discussion suggests AutoML could partner with human teams by helping biomedical scientists explore ways to formulate questions, while noting that technical and non-technical challenges remain.The study describes gaps in translating biomedical questions into data-science questions and interpreting results, alongside a need for iterative data-science attempts to inspire inquiry.Source: §5.4 · p. 18

• Snowball sampling may introduce selection bias because informants reached through the researchers’ social network may be unusually active in open science and report more positive experiences.All informants were U.S.-based, which may not represent cultural differences or varied geographical distances in open science collaborations.Source: §6 · p. 19
• Interview self-reports are subjective, single-sided, and potentially oversimplified, limiting detail about dynamic interdisciplinary interactions.The authors call for measures of content and process common ground, observation of real team interactions, and longitudinal case studies across the research pipeline.Source: §6 · p. 19
• Because the study focuses on biomedical research, it remains unknown how collaboration challenges vary in domains such as physics, geology, and psychology.Source: §6 · p. 19

• The study reports collaboration challenges from interviews with 22 data scientists and biomedical scientists and offers a systematic account of stakeholder practices.Source: §7 · p. 19
• It distinguishes content common ground from process common ground and finds scientific collaborations require breaking the former while accumulating the latter.The conclusion contrasts this pattern with decision-making or problem-solving scenarios where one decision or solution is the final product.Source: §7 · p. 19
• The authors say the findings inform practices for interdisciplinary collaboration and offer actionable design suggestions for data-analytic tools and cloud-sharing platforms.Source: §7 · p. 19