This article is sponsored by CDD Vault and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page.
Drug discovery is one of the slowest, costliest processes in enterprise R&D. Developing a single FDA-approved therapy typically takes 10 to 15 years, according to workshop proceedings published by the National Academies of Sciences, Engineering, and Medicine, and roughly 90 percent of drug candidates that enter development fail before ever reaching patients, according to research published in JAMA, often after years of investment in a single research direction.
Much of that cost and delay traces back to how research data is managed. Across biotech and pharma, discovery data is often distributed across multiple platforms, formats, and organizations, creating challenges for data sharing, reproducibility, and scientific reuse that researchers from the University of Pennsylvania have identified as persistent barriers to efficient biomedical research. In practice, that fragmentation appears as assay data stored in separate systems, limited interoperability between research platforms, and teams working from inconsistent data sources.
Similar data integration and standardization challenges are described in a data science roadmap authored by a working group from the Structural Genomics Consortium and published in Nature Communications. The authors argue that centralized architectures, standardized vocabularies, and better-connected research workflows are essential for generating AI-ready datasets. As organizations expand the use of AI in research and development, the quality, governance, traceability, and accessibility of scientific data become increasingly important.
Researchers from the University of Maryland, Baltimore County, and the University of Illinois Chicago, writing in a review of FDA workshop perspectives on AI in drug development, note that successful AI adoption depends on robust data management practices and fit-for-purpose data to support reliable model development and regulatory confidence.
Emerj recently hosted conversations on the pressing question facing nearly every R&D organization: what does it take to move AI from isolated pilot wins to organization-wide adoption in drug discovery? In an internal podcast interview, Barry Bunin, CEO and President of CDD Vault, drew on two decades of data-unification successes and failures across the pharma and biotech industries. In a companion webinar, Xiong Liu, Director of Data Science and AI at Novartis, and Mitchell Buckley, Application Scientist at CDD Vault, examined the same friction from inside a large enterprise R&D organization.
This article examines the core insights these leaders shared for organizations working to scale AI across drug discovery programs:
- Unified discovery data for scalable model training: Give models access to complete, governed discovery context so they avoid accuracy failures caused by fragmented biology, chemistry, and computational data.
- Metadata rigor to establish reproducible outputs: Apply shared ontologies and consistent annotation to ensure AI outputs can be audited, reproduced, and approved across scientific and regulatory stakeholders.
- Foundation‑first sequencing to enable compounding adoption: Build the discovery data foundation once to let every new use case scale without re‑engineering pipelines or resetting governance.
- Culture and governance as the gate to organization‑wide AI scale: Split risk governance from value governance and align scientific and computational groups around shared trust so AI initiatives earn a credible green light.
Listen to the full conversations from the series below:
Episode 1: Driving the Transformation of Drug Discovery Through AI‑Ready Data Foundations – with Barry Bunin of CDD
Guest: Barry Bunin, CEO and President at CDD Vault
Expertise: Drug Discovery Informatics, AI/ML Platform Strategy, Preclinical Data Science, Scientific Software Leadership
Brief Recognition: Bunin founded CDD in 2004 after serving as an entrepreneur-in-residence at Eli Lilly, and has since grown CDD Vault into a platform used across pharma and biotech for collaborative drug discovery data management. He holds a PhD in chemistry, is named on a patent tied to the FDA-approved cancer therapy Kyprolis, and co-authored Behind the Code: The Human Side of Collaborative Drug Discovery.
Webinar on-demand: Building AI-Ready Foundations for Drug Discovery
Guest: Xiong Liu, Director of Data Science and AI at Novartis
Expertise: Data Science & AI, Drug Discovery, Clinical Trial Analytics, Biomedical Informatics
Brief Recognition: Dr. Xiong (Sean) Liu is a data science and AI leader with more than a decade of pharmaceutical R&D experience at Novartis and Eli Lilly. At Novartis, he leads data science and AI initiatives in Biomedical Research spanning drug discovery and clinical trials, following his role as a founding member of the company’s global AI Innovation Lab. Previously, at Eli Lilly, he led data science and NLP initiatives supporting drug discovery, clinical development, patient safety, and outcomes research. Earlier in his career, he served as a Principal Investigator at Intelligent Automation, Inc., where he led 10 government-sponsored projects and secured multi-million-dollar SBIR funding. He completed a Ph.D. in Information Science at the University of Pittsburgh and a postdoctorate in Bioinformatics at Johns Hopkins University School of Medicine.
Guest: Mitchell Buckley, Application Scientist at CDD Vault
Expertise: Drug Discovery, Medicinal Chemistry, Biomedical Informatics, Scientific Partnerships
Brief Recognition: Mitchell Buckley is Head of Partnerships and Technical Marketing at Collaborative Drug Discovery, bringing a background in drug discovery research, biotech, and scientific strategy. He previously co-founded and served as CTO of Modulate Bio and was Director of Strategy and Operations at Nucleate, working across academic, venture, and industry partnerships. Earlier, he conducted drug discovery research in neurodegeneration at Brigham and Women’s Hospital and received the 2020 American Chemical Society Division of Inorganic Chemistry Undergraduate Research Award. He holds a BS in Biochemistry and Molecular Biology from the University of Massachusetts Amherst.
Unified Discovery Data for Scalable Model Training
Mitchell Buckley identifies fragmented data infrastructure as the first obstacle any drug discovery organization has to clear before AI delivers value. When databases across teams, programs, or sites don’t communicate with each other, there is no way to pool the data that machine learning models need to be trained and validated on.
Barry Bunin traces the same friction back further, to the divide between the experimentalists generating lab results and the computational scientists modeling them. Without a system built around how each group naturally works, organizations lose what he calls the economics of specialization: the efficiency gained when biologists, chemists, and data scientists build directly on each other’s work.
That divide has historically bred mutual skepticism, with modelers overselling their predictions and experimentalists left holding a multi-year synthesis project when one doesn’t pan out. As Bunin puts it, “there’s been a lot of mistrust and hype and misunderstanding in the past” between computational and experimental teams, and closing it is as much organizational as technical.
Buckley frames the core obstacle organizations must resolve before AI can deliver reliable value in drug discovery:
“The primary issue is different data silos. If databases aren’t talking to each other, there isn’t an opportunity to pool that data together and feed it into these models. A secondary issue: even a well-organized database creates problems if the data isn’t properly annotated, both for reproducibility and because that metadata is critical context for machine learning and AI applications.”
— Mitchell Buckley, Application Scientist at CDD Vault
Liu expands this same problem to the enterprise level, emphasizing that technical inconsistency is only half the challenge and that the harder work is resolving ownership across teams. Data arrives from a range of internal platforms and external partners, each with its own format, and even a technically sound data lake becomes difficult to govern or scale when teams disagree about who owns which dataset.
He points to organizations consolidating departmental data into shared data lakes and layering governance and semantic structure on top as an early, practical step, though resolving ownership between teams is often the harder part of that work. According to Liu, the tool for R&D leaders here is a sequencing test, not a technology purchase. Before funding a new AI use case, teams should be able to answer a basic question: can the data this model needs be pooled and queried across the systems that hold it, and is there a clear owner for each dataset? If the answer is no, the immediate priority is not a better model, it is resolving who owns each dataset and how those datasets connect.
Liu’s before‑and‑after example shows how this plays out in practice. In a siloed organization, a researcher chasing a lead compound has to manually track down results scattered across separate platforms held by different teams, cross‑referencing spreadsheets by hand before a model ever sees the data, often duplicating experiments another team already ran. In an organization that has closed this gap, that same researcher queries one connected environment and the relevant data across programs surfaces directly, cutting the manual reconciliation that precedes most modeling effort and reducing the duplicated experimentation silos tend to produce.
Metadata Rigor to Establish Reproducible Outputs
Solving data silos only gets an organization halfway there. Buckley’s second point is that raw access to pooled data isn’t enough if that data isn’t consistently annotated: two labs can store technically similar results, but if one team labels a compound’s activity differently from another, or omits metadata that explains how a result was generated, the combined dataset becomes unreliable for both reproducibility and downstream model training.
According to Mitchell, this is why annotation quality functions as trust infrastructure rather than a documentation afterthought. A model trained on inconsistently annotated data will produce outputs that even the scientists who built it cannot fully explain or defend to reviewers, regulators, or budget owners.
Buckley describes the discipline this requires in three habits:
- Capturing complete and correctly structured experimental data so models see the full range of tested conditions rather than only favorable results.
- Keeping experiments reproducible through consistent annotation and metadata practices.
- Maintaining shared ontologies and uniform data formats so results generated in one lab mean the same thing when read by another team or model.
He frames this discipline as the primary value driver for research operations teams trying to get AI systems to work reliably, more so than any single modeling technique.
Buckley also notes that this challenge doesn’t discriminate by company size. Large pharmaceutical organizations and early-stage biotechs alike are actively trying to eliminate data silos and put stronger data management practices in place, and the ones already in the industry tend to know they need this groundwork before they can responsibly scale AI further.
Across the conversation, Buckley and Liu surface an actionable rule of thumb for enterprise teams: treat metadata standards as a gate, not a cleanup task. Before a dataset is allowed to feed a model, it should pass a consistency check against a shared ontology, the same way code passes a review before merging. Teams that skip this step often discover the gap only after a model produces results nobody can trace back to a defensible data source.
Liu argues that before this discipline is in place, a promising result from one program typically stays trapped with the team that generated it, because no one else can confidently interpret how it was produced. After it’s in place, that same result becomes usable context for other programs and models across the organization, extending the value of every experiment rather than confining it to a single use.
Foundation‑First Sequencing to Enable Compounding Adoption
“My suggestion is build the foundation once and scale by adoption. The foundation means data foundations, the semantic layer, AI-ready data. Adoption means showing that it’s delivering promise for business decision-making. Once you have that information, you usually get the green light to scale your AI systems.”
— Xiong Liu, Director of Data Science and AI at Novartis
Liu offered this central recommendation when asked what he wanted enterprise leaders to take away from their conversation.
He further illustrates the concept with an example from genomics. Rather than building a separate data pipeline for each disease area, an organization pulls related data types, such as single-cell omics data from multiple programs, into one location and applies a consistent semantic layer.
Once that foundation exists, individual teams can pull data by their specific context, whether that’s a disease area or cell type, without rebuilding the underlying plumbing each time. Liu also points to closed-loop learning as a downstream benefit. Once discovery data, lab results, and model outputs sit on a shared foundation, findings from the lab can feed back into the models that generated the original leads, creating an iterative cycle instead of a one-way handoff.
Bunin describes the same compounding effect on the platform side. Once a scientist defines a data structure for one experiment, that structure carries forward automatically for every future upload instead of being rebuilt by hand, what he calls “same song, second verse.” He ties this back to a foundational habit shared by the organizations he’s watched succeed over two decades:
“The first thing is centering on a source of truth for your data, having everybody able to see and use the data. The better organizations will have multiple departments, or even multiple organizations, working as one, so no time is lost. You treat your partners as intelligent as the scientists you’re working with directly. That’s the first foundational thing.”
— Barry Bunin, CEO and President at CDD Vault
Buckley reaches a similar conclusion from the operations side. The biggest advantage goes to organizations that build AI infrastructure deliberately from the start, rather than retrofitting it onto a stack never designed to support it. Early-stage biotechs, he notes, often have an edge here, since less legacy infrastructure gives them more agility to build for AI rather than work around it.
The guests converge on a single sequencing principle that determines whether AI adoption compounds or stalls. Resist the instinct to fund the most exciting or highest-profile use case first. Instead, fund the shared data foundation, semantic layer, and governance structure every future use case will depend on, and treat the first funded use case as a proof point rather than an end in itself.
Before this sequencing discipline, every new AI initiative rebuilds data plumbing a previous team already solved for a different program. After it, new use cases plug into an existing foundation, and each additional application becomes faster and cheaper to stand up than the last.
Culture and Governance as the Gate to Organization‑Wide AI Scale
Even a well-built data foundation stalls without governance that can move a proposal from pilot to enterprise deployment. Liu breaks this into two distinct layers that leaders often conflate, to the detriment of their proposals.
“For larger-scale AI systems, we need governance at different layers. One is basic risk and IT governance, where the data has to be secured and de-identified. Then there is scientific and functional governance: are we building the right AI, and is it actually working? Those are the questions we have to answer for leadership and budget owners before they can give us the green light.”
— Xiong Liu, Director of Data Science and AI at Novartis
Liu frames governance as two distinct layers that determine whether an AI proposal earns a credible green light:
- Risk and IT governance: Addresses whether data is secured, de‑identified, and compliant.
- Scientific and functional governance: Determines whether the proposed AI application is appropriate and whether it produces results the organization can act on.
Liu notes that proposals often stall because teams present only the technical layer without the value sizing that answers the business question, leaving decision‑makers without enough information to move forward confidently.
Buckley connects this back to trust, noting that teams that get sustained value from AI are the ones whose leadership and scientists trust the underlying data and models, not just the ones running the most sophisticated algorithms. Humans stay in the loop by design, he adds, and that trust, more than any technical benchmark, determines whether a team scales an initiative or keeps re-litigating the same pilot.
Bunin locates that trust one level up, in leadership behavior. Breaking down silos between departments, or between an organization and its outside partners, is cultural work that no governance framework alone can substitute for.
The guests point to a practical tool for leaders preparing an AI proposal: separate the two governance questions explicitly before presenting to leadership. Show that the data is secure and compliant, and, independently, show the specific numbers and value sizing that justify the scientific bet. Then pair that proposal with the cultural work of getting disciplines to trust each other’s inputs, since a technically sound governance framework still stalls if the teams behind the data don’t yet trust one another.
Proposals that answer both questions with distinct evidence, backed by leadership willing to bridge disciplines, tend to move through gated decisions faster than proposals that treat governance as a single checkbox.
