Bringing a single FDA-approved therapy to market typically takes 10 to 15 years, according to workshop proceedings from the National Academies of Sciences, Engineering, and Medicine. Research published in JAMA puts the failure rate at roughly 90 percent for candidates that enter development, often after years of investment in one research direction. Few industries face costs and delays on that scale.
Much of the drag comes from how discovery data is managed. Researchers at the University of Pennsylvania have identified data spread across platforms, formats, and organizations as a persistent barrier to sharing, reproducibility, and reuse. In practice, that means assay data in separate systems, platforms that don’t interoperate, and teams working from inconsistent sources.
A data science roadmap from a Structural Genomics Consortium working group, published in Nature Communications, reaches a similar conclusion. Its authors argue that centralized architectures, standardized vocabularies, and connected workflows are essential to producing AI-ready datasets. Researchers from the University of Maryland, Baltimore County, and the University of Illinois Chicago add, in a review of FDA workshop perspectives, that AI adoption depends on robust data management and fit-for-purpose data if models are to earn regulatory confidence.
So what does it take to move AI from isolated pilot wins to organization-wide adoption? Emerj explored that question with three practitioners. Barry Bunin, CEO and President of CDD Vault, founded CDD in 2004 after serving as an entrepreneur-in-residence at Eli Lilly. He holds a PhD in chemistry, is named on a patent tied to the cancer therapy Kyprolis, and co-authored Behind the Code: The Human Side of Collaborative Drug Discovery. Xiong Liu, Director of Data Science and AI at Novartis, has more than a decade of pharma R&D experience at Novartis and Eli Lilly and was a founding member of Novartis’s global AI Innovation Lab. Mitchell Buckley, Application Scientist at CDD Vault, is also Head of Partnerships and Technical Marketing at Collaborative Drug Discovery and previously co-founded and served as CTO of Modulate Bio.
Their advice comes down to four lessons.
Lesson 1: Pool the Data Before You Train the Model
Buckley is blunt about where to start. Fragmented infrastructure is the first obstacle, because databases that don’t talk to each other leave nothing to pool for training and validating models.
“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
Bunin traces the problem to a divide between experimentalists generating lab results and computational scientists modeling them. When no system fits how each group works, organizations lose what he calls the economics of specialization, the gains that come when biologists, chemists, and data scientists build directly on one another’s work. The divide has a history: modelers overselling predictions, and experimentalists stuck with a multi-year synthesis project when a prediction fails. In Bunin’s words, “there’s been a lot of mistrust and hype and misunderstanding in the past.” Closing that gap is as much organizational as technical.
Liu adds that the enterprise version of the problem is ownership. Data comes from internal platforms and external partners in different formats, and even a well-built data lake is hard to govern when teams disagree about who owns what. His test for R&D leaders is a question to ask before funding any new AI use case: can the data this model needs be pooled and queried across the systems that hold it, and does each dataset have a clear owner? If not, the priority is resolving ownership and connections, not buying a better model.
The payoff is easy to picture. In a siloed organization, a researcher chasing a lead compound hunts through platforms held by different teams, reconciles spreadsheets by hand, and may repeat experiments another team already ran. In a connected one, the same researcher queries a single environment and relevant data across programs surfaces directly.
Lesson 2: Treat Metadata as Trust Infrastructure
Access to pooled data isn’t enough if the data is annotated inconsistently. Two labs can store similar results, but if one labels a compound’s activity differently or omits how a result was generated, the combined dataset becomes unreliable for reproducibility and model training alike. A model built on such data produces outputs that even its creators struggle to defend to reviewers, regulators, or budget owners.
Buckley describes the discipline in three habits: capturing complete, correctly structured experimental data so models see the full range of tested conditions and not just favorable results; keeping experiments reproducible through consistent annotation; and maintaining shared ontologies and uniform formats so a result means the same thing in every lab. He calls this the primary value driver for research operations teams, more so than any single modeling technique. It also isn’t a big-company problem. Large pharma and early-stage biotechs alike are working to eliminate silos.
Buckley and Liu offer a practical rule: treat metadata standards as a gate, not a cleanup task. A dataset should pass a consistency check against a shared ontology before it feeds a model, much as code is reviewed before merging. Teams that skip the check often discover the gap only when a model produces results no one can trace to a defensible source. Done well, the benefits compound. A promising result stops being trapped with the team that generated it and becomes usable context across programs.
Lesson 3: Build the Foundation First
“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 illustrates this with genomics. Rather than building a pipeline for each disease area, an organization brings related data, such as single-cell omics from multiple programs, into one location under a consistent semantic layer. Teams then pull data by disease area or cell type without rebuilding the plumbing. A shared foundation also enables closed-loop learning, where lab findings feed back into the models that generated the original leads.
Bunin sees the same compounding on the platform side: once a scientist defines a data structure for one experiment, it carries forward to every future upload, or as he puts it, “same song, second verse.” He names the habit shared by the organizations he has watched succeed:
“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 notes that early-stage biotechs often have an edge, since less legacy infrastructure means they can build for AI instead of working around old systems. The shared principle is to resist funding the most exciting use case first. Fund the data foundation, semantic layer, and governance every future use case will need, and treat the first project as a proof point. Without that sequencing, each new initiative rebuilds plumbing a previous team already solved. With it, each new application is faster and cheaper than the last.
Lesson 4: Earn the Green Light with Governance and Culture
A solid foundation still stalls without governance that can carry a proposal from pilot to enterprise deployment. Liu says leaders often conflate two layers, 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
Risk and IT governance covers security, de-identification, and compliance. Scientific and functional governance asks whether the application is appropriate and produces results the organization can act on. Proposals often stall because teams present only the technical layer and skip the value sizing that answers the business question.
Culture matters as much as process. Buckley observes that sustained value comes from teams whose leaders and scientists trust the data and models, not simply those running the most sophisticated algorithms, and that humans stay in the loop by design. Bunin places that trust in leadership behavior: breaking down silos between departments and outside partners is cultural work no framework can replace.
The practical takeaway is to separate the two governance questions before presenting to leadership. Show that the data is secure and compliant, then show the numbers and value sizing that justify the scientific bet, and pair both with real effort to build trust across disciplines. Proposals that answer both questions with distinct evidence tend to move through gated decisions faster than those that treat governance as a single checkbox.
The Bottom Line
Scaling AI in drug discovery is less about the next model than about the groundwork beneath it: connected data, rigorous metadata, a foundation built once, and governance and culture that let leaders say yes with confidence.