AI is useful in drug discovery when it is paired with strong biology, clean data, and realistic goals. The fastest gains come from workflow acceleration, especially coding, documentation, and metadata capture. The biggest failures happen when teams force complex AI onto messy or undersized datasets. Diamond Age helps teams choose the simplest method that answers the scientific question and avoids expensive detours.
Takeaways from Our Webinar on AI in Drug Discovery
This article is based on insights from a Diamond Age-hosted webinar on AI in drug discovery. We brought together scientists and data leaders to talk candidly about what AI is delivering today, where teams are getting stuck, and how to avoid the most common and expensive failure modes.
The conclusions here reflect Diamond Age’s biology-first point of view, informed by real client work and reinforced by the discussion. Rather than recap the entire panel, we focus on the themes that matter most for teams doing real discovery work.
Diamond Age Contributors
-
- Eleanor Howe, Founder & CEO, Diamond Age Data Science
-
- Michael DeRan, Consultant, Diamond Age Data Science
Guest Participants
-
- Chris Frew, CEO, BioBuzz (Moderator)
-
- Jefferson Parker, Senior Manager, Genpact
-
- Alex Jackson, Founder, NextGears
Why AI in Drug Discovery Still Needs a Reality Check
If you work in drug discovery, you have probably been told more than once that you need to use AI. Sometimes that directive comes with a clear goal. More often, it is simply a mandate.
That gap between expectation and execution is where many AI projects begin to struggle.
The Hype Has Quieted, and That Is a Good Thing
A few years ago, AI in drug discovery was framed as a shortcut to breakthroughs. Fully automated pipelines, autonomous hypothesis generation, and cures just around the corner. From our perspective, that framing created unrealistic expectations.
What we see now is a more useful reality. The most consistent gains come from tools that help scientists work faster at tasks they already understand, such as writing and debugging code, standardizing documentation, and making data usable across teams.
“AI is going to help biologists do biology way better. But it still requires biologists.”
–Eleanor Howe
Why Most AI Projects Fail Before Modeling Starts
The most common failure mode we encounter has nothing to do with algorithms. It appears earlier, when experiments are captured inconsistently, metadata is missing, or key context exists only in notebooks or memory rather than in shared systems.
“A lot of organizations generate incredible amounts of data, but without consistency or context. That makes it very hard to turn into something actionable.”
–Jefferson Parker
Diamond Age is often brought in after previous AI efforts have stalled. In many cases, increasingly sophisticated approaches were tried, but no one stopped to assess whether the data itself was suitable for modeling. Cleaning up workflows, integrating ELN and LIMS systems, and standardizing metadata frequently unlock more value than any other model ever could.
“Eighty percent of the work is still data preparation. That has not changed.”
–Michael DeRan
Where AI Is Actually Delivering Value Today
There are areas where AI is clearly delivering value. Diamond Age consistently sees impact in protein structure and sequence modeling, high-content imaging analysis, and integrating large, multimodal datasets across omics platforms.
“The projects that move the needle are not the flashy ones. They are the practical ones that fix how teams actually work today.”
–Chris Frew
In these contexts, AI does not replace scientific judgment. It narrows the search space, surfaces patterns faster, and helps teams decide what is worth testing next, when the underlying data is solid.
“The idea of the undruggable target is becoming less true, but only if you apply the tools carefully.”
–Eleanor Howe
When AI Is the Wrong Tool for the Job
One of the most valuable parts of Diamond Age’s work is helping teams recognize when not to use AI. Not every dataset calls for deep learning, and forcing it often creates more problems than it solves.
When sample sizes are small or experimental design is still evolving, simpler approaches such as linear models, random forests, or careful exploratory analysis are often faster, more robust, and easier to defend.
“Most of the time, AI is not the answer, and that is okay.”
–Eleanor Howe
Why Domain Expertise Still Matters
Large language models generate plausible outputs, but they do not know when they are wrong. Without domain expertise, it is easy to mistake confident language for correct insight.
“AI tools do not think critically. Without experts validating the outputs, it is easy to mistake plausible-looking results for real insight.”
–Alex Jackson
Teams that get the most out of AI involve biologists, statisticians, and data scientists at the beginning, when experiments are being designed, not after the data is already locked in.
“The expert is as important as they have always been. AI just makes that more obvious.”
–Eleanor Howe
Diamond Age Takeaways
AI tends to reward preparation, not ambition. Teams that invest in data quality and experimental design consistently outperform those chasing the newest tools. Critical thinking matters more than ever because while AI accelerates analysis, it also accelerates mistakes if outputs are not questioned. Many of the biggest wins come from foundational improvements such as better documentation, cleaner pipelines, and faster iteration.
Knowing when not to use AI is part of a sound strategy. Used well, AI amplifies good science. Used poorly, it magnifies existing problems.
Talk to Diamond Age
If you’re trying to figure out where AI fits into your discovery workflow or where it doesn’t, Diamond Age can help you pressure-test your approach before you commit time and budget. We focus on biology-first thinking, defensible analytics, and practical progress.
Author
Eleanor Howe is the Founder & CEO of Diamond Age Data Science and a computational biologist specializing in machine learning and transcriptional profiling. Diamond Age works with life science teams to make their data usable, their analytics defensible, and their discovery workflows more efficient.
FAQ: AI in Drug Discovery
Q: Where do most teams get stuck when trying to use AI in drug discovery?
A: Most teams get stuck before any modeling happens. The biggest blockers are inconsistent experimental documentation, missing or unclear metadata, and data spread across disconnected systems. Without a solid foundation, AI tools can’t reliably compare results across experiments, time points, or teams.
Q: What are the most common (and expensive) AI failure modes?
A: The most expensive mistakes usually come from forcing complex AI methods onto datasets that aren’t ready for them. That includes small sample sizes, uncontrolled variability, lack of validation, and overconfidence in outputs that haven’t been checked against biological reality. These failures often send teams down false paths that take months and significant budget to unwind.
Q: Why do AI projects fail even when the technology itself is sound?
A: Because the problem is rarely the model. It’s usually experimental design, data quality, or a mismatch between the scientific question and the method being used. AI can’t compensate for unclear hypotheses or poorly captured data.
Q: How can teams avoid wasting money on the wrong AI approach?
A: Start by being very clear about the question you’re trying to answer. Then choose the simplest method that can answer it. Invest early in consistent data capture and documentation, and involve data scientists and statisticians at the experiment design stage, not after the data is already generated.
Q: When is AI not the right tool for the job?
A: AI is often the wrong choice when datasets are small, highly variable, or still evolving. In those cases, traditional statistical approaches or basic machine learning are usually faster, more interpretable, and easier to defend. Using AI just because it was requested by leadership is a common and costly mistake.
Q: What’s the fastest, lowest-risk way to get value from AI?
A: Use AI to remove friction from existing workflows first. Coding assistance, report generation, literature summarization (with citations), and metadata cleanup deliver immediate productivity gains without introducing major scientific or regulatory risk.
Q: Why does domain expertise still matter so much when using AI?
A: AI tools generate outputs that sound confident, not outputs that are necessarily correct. Domain experts are essential for validating results, spotting subtle errors, and ensuring conclusions make biological and clinical sense. Without expert oversight, it’s easy to mistake plausible-looking results for real insight.
Q: Do teams need to build custom AI systems to stay competitive?
A: Usually not. Most organizations get more value from using existing tools well than from building new ones. Competitive advantage typically comes from strong biology, good data discipline, and thoughtful application of the right methods, not from owning custom AI infrastructure.