Artificial Intelligence (AI) has the potential to transform biomedical research. By leveraging large datasets, AI may help uncover complex, hidden biological relationships, improve disease prediction, and accelerate drug development. However, the usefulness of “big-data” driven models is inherently constrained by the data used to train them. 

These issues are most salient in the context of female biology. Indeed, females have been critically underrepresented in biomedical research. Worse, studies that include both sexes generally fail to stratify and report results by sex. Atlasing efforts not only failed to include matched numbers and age ranges in females vs. males, but they also overlooked a key female-specific organ: the ovary. Thus, our current understanding of biology is disproportionately derived from observations of male biology. To note, the problem extends beyond the number of represented females: even female-derived datasets frequently lack physiological and life-history information needed to understand female biology. As existing large datasets are now being leveraged to train AI models (e.g., CZI’s “virtual cell”), this historical imbalance represents not only a major blind spot for models to understand human health and disease, but will propagate these issues further. 

How did we get here? 

The historical exclusion of female biology in biomedical research is the product of many past decisions. On the human side of things, underrepresentation of women in clinical trials was initially a byproduct of legitimate concerns about reproductive/developmental toxicity. Specifically, in the 1960s, use of thalidomide to treat morning sickness led to infants born with missing limbs and other severe birth defects. In the wake of this tragedy, additional concerns about reproductive toxicity followed suit, with generalized concerns about prenatal exposures emerging. In response, the U.S. Food and Drug Administration issued guidance in 1977 recommending exclusion of females of “childbearing potential” from early-stage clinical trials, only reversed in the mid-90s. Risk aversion led clinical trials, in practice, to systematically exclude women, regardless of contraceptive use or reproductive status. Although the policy’s intent was to protect women and pregnancies, it backfired, preventing the collection of data about how drugs interact with female biology. Thus, it is hardly surprising that women experience adverse drug reactions at a significantly increased rate to men, since dosing regimens and drug efficacy are not tested against female biology. 

Similar biases have been persistent in preclinical research. Researchers often invoke the notion that data derived from female animals is inherently more “noisy”, due to hormonal cycling, and thus that studying females decreases statistical power and increases cost. This argument also implicitly assumes that biological responses are largely similar across sexes, making the focus on males merely a pragmatic cost-effective decision to maximize discovery. However, the myth that female data is more variable has been largely debunked, invalidating the power rationale for excluding females in the first place. Further, accumulating data shows that biological sex shapes lifelong health, disease risk and drug responses. Ignoring female biology may simplify the interpretation of an experiment, but it also hides crucial health-relevant biology. 

Ovarian aging: a case-study for the female biology blind spot. Ovarian aging represents a striking example of how biased research has hobbled our understanding of female biology. Despite clear evidence linking menopause to increased disease risk, the ovary has historically been reduced to its role in fertility, ignoring its function as a major endocrine organ, producing hormones with systemic targets throughout the body (i.e., bones, brain, arteries). 

In humans, the ovary ages faster than most somatic tissues, with menopause at ~51 years. Women spend decades of their lives in a postreproductive, postmenopausal state. Decreased ovarian function impacts the brain, cardiovascular system, skeleton, metabolism, and immunity. Menopause is associated with accelerated “biological” aging, while later age-at-menopause predicts increased longevity and lower lifelong risk for chronic diseases. 

Yet, ovarian biology remains poorly captured in biomedical datasets. A study may record that an individual is female and 52 years old, without recording menopausal status, age-at-menopause or reproductive history (i.e., pregnancies, lactation). Because age-at-menopause varies substantially between individuals, women of the same chronological age can have different ovarian function. And that crucial information is missing in large datasets generated to study human biology (e.g., GTEx). 

An additional complication to this conundrum: menopause is hard to model preclinically. Few species undergo menopause. In contrast to humans, aged female mice maintain low but persistent estrogen, and commonly used “menopause” models abruptly remove the ovary in young animals, failing to reproduce both transitional effects and interactions with the aging milieu. 

These limitations matter when existing datasets are used to train AI models: no amount of training can teach models about what was never measured or recorded. Algorithms cannot learn from experiments that were never performed. Thus, AI models cannot account for the impact of ovarian aging if ovarian function was never measured or recorded in the underlying training datasets. 

Missing data is not the same as noisy data. It’s crucial, missing biology. The biology that has been studied most extensively – male biology – represents the bulk of currently available training data. Beyond a missed opportunity, missing data risks providing false confidence in the applicability of broad conclusions derived from faulty training data. Building tomorrow’s models on yesterday’s data creates a shaky, unreliable foundation. 

Moving forward: an opportunity 

More than increasing numbers of female participants/subjects in biological datasets, biomedical research needs a paradigm shift prioritizing collection of rich metadata. Studies must be designed to detect sex-specific effects, capture variables relevant to female physiology (not only ovarian function and menopause, but also pregnancy, lactation or contraceptive use). 

We must treat sex differences less like an inconvenient source of noise and more like a focal point of interest. We have an unprecedented opportunity to generate historically missing data at scale - single-cell genomics and spatial profiling can help us resolve intrinsic and systemic changes associated with ovarian aging. Proteomics and metabolomics can capture systemic consequences of altered endocrine function and reproductive status. Longitudinal sampling can determine how ovarian aging relates to organism-level aging trajectories. 

The female data gap is often framed primarily as a representation problem. But it is also an issue of foundational knowledge. Decades of research have left fundamental aspects of female physiology poorly characterized. Together with deliberate generation of sex- and life-history-aware data, AI has the unprecedented power to uncover previously hidden biological relationships. However, the boring but key first step in this revolution must be acknowledged: you can only model what you have measured. To understand female biology, a lot must still be measured.