The Data Problem at the Heart of AI-Powered Drug Discovery

As AI rapidly becomes embedded in drug discovery, perhaps the most important question isn’t how powerful the next model will be but it’s what data those models are actually learning from.

Anthropic’s recent expansion into drug development is the latest sign that artificial intelligence is moving beyond general-purpose reasoning and into applied science. It reflects real momentum in the field and growing confidence that AI can meaningfully reshape how new medicines are discovered. But if the goal is to improve clinical success rates—not simply accelerate discovery—we should be asking whether the biological data underpinning these systems is capable of teaching them what human biology actually looks like.

For years, the industry’s focus has been on building increasingly sophisticated algorithms. Larger models, more compute, and better architectures have driven remarkable advances in protein structure prediction, target identification, molecular design, and other areas of biomedical research. Those innovations deserve the attention they have received.

What has received far less attention is the biological foundation beneath them.

AI is exceptionally good at finding patterns. But it cannot distinguish between biology that is merely measurable and biology that is truly predictive of what happens in patients. The ceiling for AI-powered drug discovery will ultimately be determined not only by the intelligence of the models, but by the fidelity of the human data used to train, validate, and benchmark them. If we want AI to deliver better medicines rather than simply faster hypotheses, building high-quality human biological data infrastructure may prove to be just as important as building the next generation of AI itself.

AI Can Only Learn From the Biology We Give It

Every AI model is constrained by the information it learns from. Drug development presents a particularly difficult challenge because biology is extraordinarily complex. Human disease is influenced by genetics, age, sex, comorbidities, immune responses, environmental exposures, and thousands of interacting molecular pathways that remain only partially understood.

Historically, much of the experimental data used throughout drug development has originated from animal studies, immortalized cell lines, simplified in vitro systems, and computational simulations. These tools have advanced biomedical science tremendously and remain indispensable across many stages of research. However, decades of experience have also demonstrated that they cannot fully replicate human physiology.

Approximately 90% of drug candidates entering clinical trials ultimately fail, with lack of efficacy and unexpected safety findings representing major contributors. While many factors contribute to these failures, one recurring theme is that preclinical models often struggle to predict what actually happens in patients.

If AI systems are trained primarily on data generated from models that themselves have known translational limitations, it should not be surprising if those limitations persist. AI can recognize patterns remarkably well but it cannot invent biological truth that was never present in its training data.

More Data Is Not Necessarily Better Data

The AI community often speaks about scaling laws: the observation that larger datasets frequently improve model performance. In biology, however, quantity and quality are not interchangeable. Millions of data points collected from experimental systems that poorly reflect human physiology may offer less predictive value than smaller datasets generated directly from clinically relevant human biology. This distinction becomes especially important in drug development, where the goal is not simply prediction but prediction that translates into successful patient outcomes.

High-fidelity human biological data differs from traditional laboratory datasets in several important ways. It captures biological function rather than static snapshots. It preserves relationships between tissues, cellular architecture, perfusion, metabolism, and pharmacology. It incorporates patient history, clinical characteristics, molecular measurements, and functional responses within the same biological system. Most importantly, it measures biology that actually exists in humans.

As AI becomes increasingly capable of extracting subtle patterns from complex data, the quality of those patterns becomes inseparable from the quality of the underlying biology.

The Next Competitive Advantage May Be Human Data Infrastructure

Over the past several years, enormous investments have gone into foundation models, computational infrastructure, and specialized AI architectures. Far less attention has been paid to the infrastructure required to systematically generate high-quality human biological data at scale.

Human biological specimens are inherently limited and require integration with some form of donation infrastructure. Standardization remains challenging. Experimental protocols must be reproducible. Metadata must be comprehensive. Multiomic measurements, imaging, functional assays, and clinical context all need to be integrated into datasets that are sufficiently consistent for computational learning.

None of this resembles traditional software engineering.Instead, it requires coordinated biological operations capable of producing reproducible, regulator-ready datasets over many years. Just as cloud infrastructure became essential for modern software development, human biological infrastructure may become foundational for the next generation of AI-driven therapeutics.The organizations investing in that infrastructure today may ultimately shape what tomorrow’s AI models are capable of learning.

Regulators Are Moving in This Direction

Importantly, this discussion extends beyond academic curiosity. Regulators themselves are increasingly emphasizing human-relevant evidence. The FDA has expanded its support for New Approach Methodologies (NAMs), outlining pathways through which computational models, organ-on-chip technologies, advanced in vitro systems, and other human-relevant approaches can contribute to regulatory decision-making. 

This shift recognizes an important reality. As computational methods become more sophisticated, confidence in their predictions depends upon confidence in the biological evidence supporting them. If better AI requires better validation, and better validation requires better human data, then better AI must require better human data that lies at the source of each model.

The Next Decade Will Be Defined by Data Quality

The pharmaceutical industry has experienced multiple technological revolutions from high-throughput screening to next-generation sequencing. Each expanded the volume of available data but the application of AI in this field represents something different.

Its success depends not only on producing more data but on producing data that more faithfully represents human biology. As foundation models become increasingly accessible, algorithmic advantages alone may become less durable. Access to differentiated, high-quality human biological data may instead emerge as one of the defining competitive advantages across the pharmaceutical ecosystem. This is particularly true if the industry’s ultimate objective is improving clinical success rather than simply accelerating discovery.

Anthropic’s entrance into drug development reflects the remarkable progress AI has made over the past several years. It also highlights the next question the industry must answer. If AI truly has the potential to transform medicine, what biological foundation will those models stand upon?

The future of AI-powered drug discovery will undoubtedly depend on better algorithms. But it may depend even more on building the human data infrastructure capable of teaching those algorithms what human biology actually looks like.