Biology Beyond Human Intuition

Why understanding life may require learning to see it differently

Living systems are extraordinarily complex. They operate across molecular, cellular, tissue, organismal, and ecological scales, with interactions that are nonlinear, emergent, context-dependent, dynamic, and hierarchical. A perturbation at one scale can propagate through countless interactions to produce consequences at another, and yet the same molecular intervention can yield dramatically different outcomes depending on the cellular environment, physiological state, genetic background, or broader biological context.

Humans have made tremendous strides in understanding facets of biology. We have discovered the structure of DNA, sequenced entire genomes, developed vaccines and transformative therapeutics, and even learned how to manipulate cellular identity and function. These advances have fundamentally changed our understanding of life and our ability to intervene in it. Yet despite this progress, much of biology remains beyond our understanding. Unlike mathematics and many areas of physics and chemistry, where relatively compact principles can often explain and predict a wide range of phenomena, biology remains difficult to predict even when we understand many of its individual components. We can identify molecular mechanisms, but often cannot anticipate how they will behave in different contexts or what their consequences will be at the level of an entire organism.

It's reasonable to speculate that human intuition may be poorly suited to understanding biology. We struggle to reason about high-dimensional systems because our cognitive capacity is unable to mentally represent thousands of interacting variables, nonlinear dependencies, feedback loops, and emergent phenomena simultaneously.

Thus, our solution has been to make biology more intelligible. To simplify it.

We organize biological systems into genes, proteins, pathways, cell types, tissues, organs, and diseases. We construct diagrams that describe interactions between selected components. We classify continuous and heterogeneous biological states into discrete categories. We develop mechanistic explanations that isolate a manageable number of variables from the vast complexity in which they operate.

These abstractions have been enormously productive. They have enabled modern molecular biology, genetics, medicine, and biotechnology. But their success does not establish that they are the most informative representations of living systems.

In fact, our desire to make things interpretable or simple may be one of our greatest limiting factors.

Our mental models shape the biology we observe

Our mental models and desire to comprehend biology shape every facet of how we study it.

Scientific measurement is not a neutral window into nature. What we choose to measure is informed by what we believe matters. We select molecular markers, define cell types, quantify pathway activity, and construct disease classifications based on existing biological knowledge and the technologies available to us. This bias is even further amplified by what we chose to report or release in the scientific manuscripts and products.

Even technologies that generate enormous amounts of data are typically designed and interpreted through human-defined biological concepts. We favor measurements that produce recognizable, interpretable quantities, often reducing high-dimensional observations into a handful of familiar markers or categories.

Michael Bronstein and Luca Naef have recently argued for a shift from traditional "white-box data," designed to be directly interpretable by humans, toward "black-box data," optimized for machine learning rather than human intuition.

Interpretation often comes at the cost of information. A cell, for example, does not necessarily occupy a discrete biological state corresponding to one of our established cell-type labels. Cells can exist in metastable states that depend on combinations of molecular, spatial, temporal, and environmental features that do not align neatly with those categories. Likewise, a disease may not be adequately represented by a diagnostic label, and an organism's physiological state cannot be fully reduced to a collection of clinical measurements.

The issue is not simply that these abstractions are wrong. It is that they are invariably incomplete, and we rarely know what information is lost when we constrain them.

More importantly, our decisions about what to measure and what matters determine which aspects of biology become visible to subsequent scientific inquiry. Our measurements are shaped by our models of biology, and our models of biology are shaped by our measurements. This creates a feedback loop in which existing conceptual frameworks influence the evidence available to revise them.

Our mental models also shape the experiments we perform

The same constraints extend to experimental biology. We typically design experiments by identifying a hypothesis, selecting variables we believe are relevant, manipulating one or more of those variables, and measuring outcomes we expect to be informative.

Experimental design is often organized around familiar entities and established relationships. Consequently, the experiments we perform are not merely constrained by technological feasibility or cost. They are also constrained by what we can conceive of testing.

This has important consequences for the diversity of biological observations we generate. Much of experimental biology relies on a relatively small number of well-characterized cell lines, model organisms, assays, and experimental conditions. We repeatedly measure the same readouts, using the same technologies, in the same biological contexts. There are good reasons for this. Standardization improves reproducibility, facilitates comparisons across studies, and allows results to accumulate within established scientific frameworks.

But these practices also constrain the biological variation represented in our data. An experiment that is highly reproducible within a particular cell line or model organism does not necessarily tell us how the same biological process behaves in another cell type, species, physiological state, or environmental context. This distinction becomes especially important for machine learning. Models trained on narrow, highly standardized experimental conditions may become exceptionally good at predicting outcomes within those conditions while struggling to generalize beyond them.

For models intended to learn general principles of biology, diversity of context may be as important as, or even more important than, repeated observations within a limited set of conditions.

Our mental models shape how we intervene

These limitations become especially consequential when we attempt to change biological outcomes. Modern drug discovery is built around a sequence of abstractions and proxies. We identify a disease, nominate a target, develop an intervention, evaluate molecular or cellular effects, test preclinical models, and ultimately assess outcomes in humans.

At each stage, we rely on simplified representations of a much more complex biological process. Target engagement serves as a proxy for functional activity. Cellular phenotypes serve as proxies for tissue responses. Animal models serve as proxies for human physiology. Biomarkers serve as proxies for clinical outcomes. But uncertainty accumulates as we move between translational stages.

These approximations have been necessary, but their limitations contribute to the high failure rates in drug development. A molecular intervention can engage its intended target, produce the expected cellular response, and even demonstrate efficacy in an animal model, yet fail to produce a meaningful clinical benefit.

LLMs have inherited many of our abstractions

Artificial intelligence is beginning to transform biological research. But much of its current development inherits the same conceptual constraints. Large language models trained on scientific literature learn representations of biology through human-generated descriptions. They absorb an enormous amount of scientific knowledge, including relationships that no individual researcher could remember or synthesize. But they learned from what we decided to write down. Scientific language is itself a compressed representation of biological reality, filtered through our hypotheses, experiments, terminology, and publication practices.

A language model can become exceptionally proficient at manipulating biological concepts without necessarily possessing a representation of the underlying biological systems from which those concepts arose. This distinction matters. Reasoning more effectively within an existing conceptual framework is not the same as discovering a better framework, which may be what we are actually lacking.

There is an important distinction between perceiving a system and reasoning about it. Reasoning allows us to manipulate representations, infer relationships, construct explanations, and plan actions. But the quality of those operations depends on what the representations preserve. If critical aspects of a system are absent from its representation, even sophisticated reasoning may fail to recover them. For biological AI, this suggests that progress in reasoning alone may not be sufficient. We also need advances in how machines represent biological reality.

The problem is also not limited to language models. Many biological AI systems are trained to predict human-defined labels, measurements, or tasks. Their architectures and objectives often reflect the same decomposition of biology into molecular, cellular, and clinical domains.

Even biological foundation models trained directly on experimental data can inherit these limitations if they are optimized primarily to reproduce existing measurements or perform tasks defined by our current understanding. By not acknowledging this head-on, we risk building increasingly powerful AI systems that operate within the same simplified representations that have constrained biological science for decades.

What if machines could learn to perceive biology differently?

Rather than beginning with a predetermined set of biological concepts and teaching machines to reason about them, we could train models to discover useful representations from biological observations themselves. The objective would not be to reproduce our existing classifications of genes, pathways, cell types, or diseases. It would be to learn a representation of biological systems that preserves the information necessary to predict their behavior.

This creates an opportunity to exploit a defining characteristic of biology: structure. The same underlying system can be observed through many different measurements and at many different scales. No single measurement captures the whole system. Yet these observations are not independent. They are connected through the physical processes that generate biological function.

Importantly, this does not mean biological measurements can be understood without context, or that all useful biological information can be compressed into a single universal description. It means that we should allow the structure of biological observations, rather than only our preexisting concepts, to help determine how those observations are represented.