What is algorithmic bias in drug discovery?

Algorithmic bias in drug discovery refers to systematic differences in an AI model's performance or predictions across different patient populations, caused by the underlying training data rather than the algorithm's design. It typically originates in genomic databases, clinical trial records or preclinical screening data that were not collected equally across populations, and it propagates into any model trained on that data.

Are drug discovery AI models biased?

Most drug discovery AI models are trained on genomic and clinical datasets that are not equally representative across populations, so some degree of bias is likely present unless specifically tested for and addressed. The clearest documented example is polygenic risk scores, which are measurably more accurate for people of European ancestry than other ancestries because of skewed genomic reference data; the same underlying dynamic can create bias in other model types.

How does training data bias affect AI drug discovery?

Training data bias affects AI drug discovery by causing a model to perform well for populations well-represented in its training data and less reliably, often without any obvious warning sign, for populations that are underrepresented. Since genomic and clinical trial data both currently skew heavily toward European-ancestry and historically well-studied populations, models trained on that data are least likely to be validated for everyone else.

Bias and Fairness in AI Drug Discovery

Highlights in this article

Where bias enters AI drug discovery
Genomic database representativeness
Clinical data bias
What 'algorithmic fairness' means
Steps pharma is taking
Regulatory perspectives
Key takeaways

A model is only ever as representative as the data it learned from, and that is the uncomfortable starting point for any honest conversation about bias in AI drug discovery questions. Genomic databases, clinical trial records, and even the cell lines used in high-throughput screening all carry the imprint of decades of research that was not conducted equally across human populations. Artificial intelligence (AI) does not introduce that imbalance. It learns it, often invisibly, and then reproduces it at a speed and scale no individual chemist or clinician ever could.

This connects to the regulatory guide to AI in drug discovery for how FDA, EMA and ICH are, and in some cases are not, addressing representativeness directly, and to the guide to AI across drug discovery for the wider technical picture. What follows works through where bias actually enters the pipeline, what fairness can reasonably mean in this context, and what is being done about it.

Where bias enters AI drug discovery

Bias rarely enters through a single, identifiable decision. It accumulates across several independent layers of the pipeline, each shaped by its own history. Genomic reference data reflects which populations research funding and infrastructure historically prioritized. Clinical trial data reflects who could access, was recruited into, or was willing to trust a given trial. Even preclinical screening data, the cell lines and model systems used to generate the phenotypic and biochemical readouts that train early discovery models, carries its own narrower kind of sampling bias: a handful of widely used, well-characterized cell lines standing in for the vastly more diverse biology of an actual patient population.

What makes this different from bias in most other applied AI is timing. By the time a discovery-stage model is flagged for a fairness review, the underlying imbalance was usually locked in years or decades earlier, in datasets a modern AI team did not build and often cannot meaningfully rebalance after the fact. That does not make the problem someone else's to solve. It does mean the realistic remedies looked at later in this piece work mostly by adding better data forward, and by testing rigorously for where a model's blind spots actually fall, rather than by fixing history retroactively.

Genomic database representativeness

The single most cited number in this discussion is worth stating precisely: as of 2023, participants of European ancestry made up 86.5% of the cumulative GWAS Catalog, the reference database underlying a large share of genomic association research, while participants labeled African, not including African American or Afro-Caribbean samples, made up just 0.47%. A real-time snapshot of ongoing research collected in September 2024 found the pattern still holding for newly published studies, with every population-labeled group other than Asian (3.96%) still representing less than 1% of new participants.

The downstream consequence is measurable, not theoretical. Research from the Broad Institute found that polygenic risk scores, which estimate an individual's genetic risk for a given condition, are substantially more accurate for people of European ancestry than for people of African ancestry, precisely because the underlying association data so heavily overrepresents European populations. A model trained predominantly on one ancestry's genomic architecture does not fail loudly on other populations; it fails quietly, with lower and less-noticed accuracy.

The counter-effort is real and growing, if still early. The NIH's All of Us Research Program had enrolled more than 860,000 participants as of April 2025, with more than 80% meeting the program's criteria for being historically underrepresented in biomedical research. At the Broad Institute, the PUMAS initiative, Populations Underrepresented in Mental Illness Associations Studies, has used a lower-cost blended genome-exome sequencing method to sequence more than 53,000 samples from African, African American, and Latin American populations specifically, alongside a separately harmonized reference resource combining the 1000 Genomes Project and Human Genome Diversity Project. None of this closes an 86.5-to-0.47 gap quickly, but it is the clearest sign that the field recognizes the gap as a problem worth funding rather than an acceptable status quo.

Clinical data bias

Genomic data is only one input. Clinical trial data, the record of who a drug was actually tested on, carries its own separate and compounding bias, and it feeds directly into the AI models increasingly used to predict trial outcomes, optimize dosing, and flag safety signals. A peer-reviewed analysis of FDA Drug Trials Snapshots data covering 341 pivotal trials supporting 348 drug approvals between 2017 and 2023 found that only 6% of trials enrolled a participant population whose racial and ethnic makeup matched the broader US population; the remaining 94% underrepresented at least one major demographic group. The same analysis found Black and Hispanic enrollment declining over that seven-year window even as Asian enrollment rose, with the shift partly tracking a broader move of pivotal trials to sites outside the United States.

For an AI model, this matters in a very specific way: a model trained or validated on historical trial outcomes learns the response patterns of whichever population was actually enrolled. If that population systematically underrepresents a given group, the model's predictions for that group are, at best, an extrapolation rather than an observation, and extrapolations are exactly where models tend to be both least accurate and least likely to signal their own uncertainty.

What 'algorithmic fairness' means

"Fair" sounds like a single, checkable property. In practice, machine learning researchers work with several distinct, mathematically incompatible definitions of it, and knowing which one is in play changes what a fairness claim actually promises. Demographic parity asks whether a model's positive predictions, a compound flagged as promising, a patient flagged as high-risk, occur at similar rates across groups. Equalized odds asks a narrower question: among people who truly belong in a given outcome category, does the model catch them at similar rates across groups, regardless of the raw prediction rate? Calibration asks whether a predicted probability means the same thing across groups: does a model's "70% likely to respond" actually correspond to a 70% response rate in every population it is applied to, not just on average?

The uncomfortable, well-established result in the fairness literature is that except in narrow special cases, a model cannot simultaneously satisfy all of these definitions when the underlying base rates differ across groups, which they very often do in biology. That is not a reason to give up on fairness as a goal. It is a reason to be specific: a claim that a drug discovery model is "fair" is close to meaningless without saying fair by which definition, and for which specific decision the model is informing.

Steps pharma is taking

Responses so far split into a data-side approach and a practice-side approach, and most serious programs are pursuing both. On the data side, efforts like Broad's PUMAS sequencing initiative and NIH's All of Us program are direct attempts to make the next generation of genomic and clinical reference data less skewed than the last. A related but distinct kind of representativeness gap shows up in preclinical screening: platforms built around very large-scale phenomic and transcriptomic datasets, such as Recursion's screening platform, address a different axis of the same underlying problem by deliberately broadening the range of cell types, perturbations and biological contexts a model is trained across, rather than relying on the narrower set of well-characterized lines that dominated earlier discovery-stage AI work.

Stratified performance testing: reporting model accuracy separately by demographic or ancestry subgroup, rather than only as a single aggregate figure that can hide large subgroup-level gaps.
Explicit uncertainty flagging: building models that signal lower confidence for inputs that fall outside their training population, rather than returning a confident-looking prediction regardless of how well-represented that input actually was.
Diverse reference panel adoption: incorporating newer, more representative resources, such as All of Us or harmonized diverse-genome datasets, alongside legacy databases rather than as a replacement discussed but not implemented.
Demographic data tracking independent of any single regulatory mandate, on the reasoning that the underlying representativeness problem does not disappear if a specific guidance requirement changes.

Regulatory perspectives

The two major regulators covered in the broader regulatory guide currently sit in a genuinely different place on this specific issue. EMA's reflection paper is explicit that AI systems trained on data underrepresenting small populations, including children and people with rare diseases, warrant particular scrutiny for bias, and ties that scrutiny directly to its broader AI validation expectations.

FDA's parallel clinical trial diversity framework, developed independently of its AI-specific guidance and covered in more depth in the companion piece on FDA's AI action plan, has followed a more disrupted path. The Food and Drug Omnibus Reform Act required sponsors to submit Diversity Action Plans, and FDA issued draft guidance on their format in June 2024, with a statutory deadline to finalize that guidance by June 26, 2025. Following a January 2025 executive order addressing federal diversity, equity, and inclusion policy, FDA removed the draft guidance from its website; a subsequent court order required its restoration for public access in February 2025. As of this writing, final guidance has still not been issued, more than a year past its original statutory deadline, and the framework's practical status remains unsettled.

Whatever the outcome of that specific rulemaking, the underlying data representativeness problem this article describes does not depend on it. A model trained predominantly on one ancestry's genomic data or one population's clinical trial history carries that imbalance regardless of which regulatory framework, if any, is actively requiring sponsors to document it.


Key takeaways European-ancestry participants make up 86.5% of the cumulative GWAS Catalog against roughly 0.47% for African ancestry alone, and polygenic risk scores measurably inherit that gap. Only 6% of pivotal FDA trials from 2017 to 2023 enrolled a demographically representative population, and Black and Hispanic enrollment specifically declined over that period. "Algorithmic fairness" is not one property; common technical definitions are mathematically incompatible with each other when group base rates differ, so any fairness claim needs to specify which definition it means. EMA's AI framework explicitly flags small-population data gaps as a bias risk; FDA's separate clinical trial diversity mandate remains in regulatory limbo, independent of that AI-specific question.

Key takeaways

European-ancestry participants make up 86.5% of the cumulative GWAS Catalog against roughly 0.47% for African ancestry alone, and polygenic risk scores measurably inherit that gap.
Only 6% of pivotal FDA trials from 2017 to 2023 enrolled a demographically representative population, and Black and Hispanic enrollment specifically declined over that period.
"Algorithmic fairness" is not one property; common technical definitions are mathematically incompatible with each other when group base rates differ, so any fairness claim needs to specify which definition it means.
EMA's AI framework explicitly flags small-population data gaps as a bias risk; FDA's separate clinical trial diversity mandate remains in regulatory limbo, independent of that AI-specific question.

This article was produced in accordance with Drug Discovery News’ AI Editorial Policies.