Biology Is Not Just Data
Why Context Is the Bottleneck for AI Biology
By Li Lei on July 13, 2026
When I first worked with large-scale genomics data, I remember staring at long tables of significant genes, variants, and enriched pathways.
The results looked impressive.
But the real question was harder:
Which signals were biologically real?
Which were artifacts?
Which were statistically significant but biologically weak?
Which were worth following up on in an experiment?
That experience still shapes how I think about AI biology.
At first glance, genomics looks like a perfect field for artificial intelligence. Biology is producing enormous amounts of data: genome sequences, variants, RNA-seq, ATAC-seq, single-cell profiles, methylomes, proteomics, metabolomics, phenotypes, images, environmental records, and clinical measurements. Over the past decade, deep learning has become increasingly important for genome analysis and interpretation, including regulatory genomics, variant effect prediction, and functional genomics [1,2].
It is tempting to think that biology is simply waiting for better models.
More data.
Bigger models.
More compute.
Better prediction.
But the longer I work at the intersection of genomics, computational biology, and AI, the more I believe this view is incomplete.
Biology is not just a data problem.
Biology is a context problem.
A gene is not just a gene.
A variant is not just a variant.
A cell is not just a cell.
A phenotype is not just a label.
A model prediction is not automatically a discovery.
Biological meaning depends on context: tissue, cell type, developmental stage, genotype, environment, population history, evolutionary constraint, experimental design, assay technology, and validation evidence.
This is why I believe one of the biggest bottlenecks in AI biology is not only model architecture or compute.
It is context.
Imagine an AI system ranks a gene as a top candidate for drought tolerance in barley or sorghum.
At first, this sounds exciting. The model has scanned a large biological space and found something worth attention. Maybe the gene is near a trait-associated variant. Maybe it is differentially expressed under drought. Maybe it appears in a regulatory network. Maybe the literature suggests a connection to stress response.
But that ranking is not yet a discovery.
It is the beginning of a biological investigation.
A biologist will immediately ask:
Is this gene expressed in the relevant tissue — root, leaf, seed, or reproductive organ?
Is it active at the relevant developmental stage?
Is the response specific to drought, or is it a general stress response?
Is the association driven by population structure or linkage disequilibrium?
Is there chromatin or regulatory evidence?
Is the gene conserved across grasses?
Does perturbing the gene actually change the phenotype?
Would improving drought tolerance create a tradeoff with growth, yield, or reproduction?
These questions are not secondary details.
They are the biology.
This is where AI biology becomes difficult. A model can learn patterns from data, but biological meaning depends on where, when, how, and under what conditions those patterns appear.
In biology, signals are often real, but their meaning is conditional.
That conditionality is exactly what makes biological interpretation hard. It is also what makes AI biology hard.
The same signal can mean different things
A gene highly expressed in leaves under drought stress may be important for water response. But the same gene expressed in roots, seeds, or reproductive tissues may tell a very different story.
A transcription factor may regulate stress response in one developmental stage and growth in another.
A variant may look associated with a trait, but the signal may be driven by population structure, linkage disequilibrium, or environmental stratification.
A chromatin region may be accessible, but that does not automatically mean it is an active enhancer.
A gene may be differentially expressed, but that does not automatically mean it is causal.
A pathway may be enriched, but that does not tell us which mechanism is driving the phenotype.
This is why context is not a detail added after analysis.
Context is part of the biological signal.
If the data does not carry enough context, an AI model may learn shortcuts instead of biology. It may learn species-specific bias, batch effects, tissue composition, sequencing depth, publication bias, or experimental artifacts. It may perform well on a benchmark but fail when applied to a new genotype, tissue, environment, disease cohort, or crop field.
In biology, a prediction is useful only when we understand its boundary.
When we say “context,” it can sound vague. But in biology, context is not abstract. It has real layers.
I think of biological context as at least seven connected layers.
At the molecular level, we ask: what is happening in the genome, epigenome, transcriptome, proteome, or metabolome?
A DNA variant may change a protein sequence. Or it may affect a regulatory element. Or it may do nothing obvious by itself but matter in combination with other variants. A gene expression signal may reflect transcriptional regulation, RNA stability, cell composition, or technical noise.
A sequence is never just a string of letters. It exists in chromatin, in a genome, in a regulatory architecture, and in an evolutionary history.
This is why models such as Enformer are exciting. They show that sequence-based models can improve gene expression prediction by incorporating long-range regulatory information, but they also remind us that regulatory interpretation depends on genomic and cellular context [3].
2. Cellular context
The same gene can behave differently across cell types. This is one of the major lessons from single-cell and single-nucleus biology.
A signal that disappears in bulk RNA-seq may become clear at the cell-type level. A gene may be active only in a rare cell population. A regulatory program may be specific to a transient developmental state. A disease or stress response may not affect all cells equally.
Large efforts such as the Human Cell Atlas were built around this idea: to understand human biology, we need to define cell types by molecular profiles and connect them with location, morphology, and function [4].
AI models that ignore cell type and cell state may average away the biology we actually care about.
Biology is spatial and tissue-specific.
A drought response in roots is not the same as a drought response in leaves. A regulatory element active in one tissue may be silent in another. A variant affecting seed development may have little relevance to vegetative growth. A disease marker in blood may not reflect what is happening in the target tissue.
This matters deeply for both plant biology and human biology.
When context is missing, a model may connect the right entities in the wrong place.
Genes act inside organisms, not in isolation.
The same variant may have different effects depending on genetic background. The same pathway may behave differently across accessions, cultivars, strains, populations, or species. A phenotype may result from many small effects across the genome rather than one obvious causal gene.
This is something population genetics teaches us very well: variation is structured. History matters. Background matters. Selection matters. What appears as a clean association may be entangled with demography, ancestry, and environment.
AI systems that ignore organismal and population context may confuse correlation with cause.
Biology is responsive.
Temperature, drought, light, nutrients, pathogens, microbiome, management practice, diet, treatment, and climate can all change biological meaning. A gene that matters under stress may not matter under control conditions. A genotype that performs well in one environment may fail in another.
In agriculture, this is obvious. A trait is rarely just genetic. It is genotype-by-environment interaction.
In medicine, it is also true. Disease risk, treatment response, and immune state are shaped by environment, history, and exposure.
AI biology must learn not only biological entities, but biological conditions.
Biology has history.
Some genes are deeply conserved. Others are lineage-specific. Some regulatory elements evolve quickly. Others remain constrained across millions of years. A pathway discovered in Arabidopsis may or may not translate directly to maize, sorghum, rice, or Brachypodium. A signal from mouse may or may not translate to human.
Evolutionary context helps us ask whether a pattern is likely to be functional, conserved, diverged, or species-specific.
Without evolutionary thinking, AI may treat all biological observations as flat examples. But biology is not flat. It is shaped by descent, divergence, constraint, and adaptation.
7. Experimental context
Finally, every biological dataset comes from an experiment.
How was the sample collected?
What protocol was used?
What sequencing platform?
What batch?
What controls?
What replicates?
What normalization?
What quality filters?
What reference genome?
What annotation version?
These details are not technical decoration. They determine what conclusions can be trusted.
This is also why the FAIR principles — making data findable, accessible, interoperable, and reusable — are so important for data-intensive science. They emphasize that data, metadata, tools, and workflows should be reusable by both humans and machines [5].
A beautiful machine learning model trained on poorly annotated, biased, or inconsistent data may produce confident predictions with weak biological value.
In AI biology, metadata is not paperwork.
Metadata is part of intelligence.
I am not against bigger models.
Foundation models have already shown remarkable potential in protein biology, genomics, chemistry, and medicine. AlphaFold demonstrated the power of deep learning for protein structure prediction and changed how many scientists think about structural biology [6]. Models such as Enformer show how deep learning can extract regulatory information from DNA sequences at scale [3].
These are real breakthroughs.
But bigger models alone do not solve the context problem.
A model trained only on sequence can learn sequence patterns, but unless tissue-specific activity is represented in the training data or system design, it may not distinguish where a regulatory element is active.
A model trained on expression can learn cell states, but unless treatment, genotype, and sampling conditions are represented, it may not understand why those states appear.
A model trained on literature can learn published knowledge, but it may also inherit publication bias, outdated assumptions, and unsupported claims.
The issue is not whether models are powerful.
They are.
The issue is whether the system around the model carries enough biological context to make the output meaningful.
This is where many AI biology efforts become fragile. They focus heavily on model architecture but underinvest in data provenance, metadata, knowledge representation, benchmark design, biological validation, and feedback loops.
In other words, they build impressive engines but give them incomplete maps.
From prediction to meaning
AI is very good at prediction. But science is not only prediction.
Science also asks:
Why does this happen?
Under what conditions does it happen?
Is the relationship causal?
Can it generalize?
Can we intervene?
What evidence would change our mind?
A prediction becomes biologically valuable only when it helps us move toward understanding, decision, or action.
This is why AI biology needs scientists who understand both models and mechanisms. The model can point us toward a pattern. But the scientist must decide whether the pattern is meaningful, testable, and worth pursuing.
For the drought-tolerance gene example, the model’s ranking may be the first signal. But the biological work begins afterward.
We need to ask whether the gene is expressed in the right tissue, active under the right condition, supported by regulatory evidence, robust across genotypes, conserved across species, and testable through perturbation or validation.
That is how a prediction becomes a hypothesis.
And only through evidence does a hypothesis become discovery.
Context-free AI can be fast, fluent, and impressive.
It can summarize papers.
It can rank genes.
It can generate hypotheses.
It can produce beautiful explanations.
But if it does not represent biological context, it can also mislead.
It may overgeneralize from one species to another.
It may treat correlation as causation.
It may ignore tissue specificity.
It may miss population structure.
It may combine evidence from incompatible assays.
It may present a weak hypothesis with strong language.
It may sound confident because language models are good at fluency, not because the biology is correct.
This is especially risky because biological AI outputs often look plausible.
A wrong gene ranking may not look obviously wrong.
A weak mechanistic explanation may sound elegant.
A false literature connection may appear reasonable.
A biased model may perform well on an internal benchmark.
The danger is not only that AI can be wrong.
The danger is that AI can be wrong in a biologically plausible way.
That is why context-aware evaluation is essential.
We should not only ask whether a model performs well on average. We should ask where it performs well, where it fails, and why.
Does it generalize across species?
Across tissues?
Across environments?
Across populations?
Across experimental platforms?
Across time?
A model that performs well only in familiar contexts may still be useful. But we must know its boundary.
In biology, knowing the boundary of a model may be as important as knowing its prediction.
Context engineering: making biological meaning computable
If prompt engineering was one of the first popular skills of the AI era, context engineering may become one of the most important skills in AI biology.
By context engineering, I mean the work of making biological context computable.
This is where knowledge graphs, ontologies, metadata standards, multimodal integration, retrieval systems, and validation loops become important.
A knowledge graph can connect genes, variants, traits, tissues, pathways, publications, experiments, species, environmental conditions, and evidence types. Biomedical knowledge graphs are often used to represent biological and clinical concepts as nodes and relationships, making them useful for data integration, machine learning, and reasoning [7].
Graph representation learning extends this idea by learning from network structure and node relationships. In biomedicine, such approaches have been applied across molecular networks, disease relationships, patient data, drug discovery, and other connected biological systems [8].
This does not make the system perfect. But it gives the model a richer map.
Instead of asking AI to reason from isolated data points, we can ask it to reason over structured biological relationships.
This is especially important for AI agents in science.
If an agent is expected to answer biological questions, design analyses, prioritize candidates, or propose experiments, it needs more than a language model. It needs access to reliable data, structured knowledge, metadata, literature, tools, evaluation criteria, and human feedback.
An AI agent without biological context is like a brilliant student with no lab notebook, no experimental design, and no memory of what was actually measured.
It may be clever, but it is not yet a scientist.
What context-aware AI biology might look like
A context-aware AI biology system would not simply take a gene list and return a polished interpretation.
It would ask better questions.
What species is this from?
Which tissue?
Which developmental stage?
Which treatment?
Which genotype?
Which assay?
Which controls?
Which reference annotation?
Which evidence is direct and which is inferred?
Which claims come from literature and which come from data?
Which results are robust across datasets?
Which hypotheses are testable?
It would not only produce an answer.
It would show the evidence path.
It would connect a prediction to the data, the metadata, the prior knowledge, the uncertainty, and the next experiment.
This is the kind of AI biology I find exciting.
Not AI that simply gives us more answers.
AI that helps us ask better questions.
In the AI era, biologists should not underestimate their role.
It is easy to feel that machine learning experts are driving the future and biologists are only data providers. But that would be a mistake.
Biologists understand what the question means.
Biologists understand whether the data is appropriate.
Biologists understand what the model might be missing.
Biologists understand what evidence is convincing.
Biologists understand whether a prediction is experimentally actionable.
This expertise is not secondary.
It is central.
The future of AI biology needs biologists who can define meaningful problems, identify relevant context, design better evaluation, and interpret model outputs with scientific discipline.
This does not mean every biologist must become a machine learning engineer. But more biologists should learn enough AI to participate in system design, not only tool usage.
They should be able to say:
This dataset is not enough.
This benchmark is misleading.
This prediction is interesting but not actionable.
This context is missing.
This result needs validation.
This hypothesis is worth testing.
That is scientific leadership in the AI era.
Genomics taught me humility.
When I first learned to analyze large biological datasets, I was fascinated by the power of scale. Suddenly, we could look across the genome, across populations, across tissues, across conditions. We could move from one gene at a time to thousands of genes at once.
But scale also made the problem harder.
More data did not automatically create more understanding. Sometimes it created more ambiguity. More correlations. More candidates. More possible explanations.
The real work was to connect signals back to biology.
Which signal is robust?
Which one is confounded?
Which one generalizes?
Which one has mechanism?
Which one can be validated?
Which one matters?
AI biology is entering the same stage.
We are surrounded by new models, new tools, new predictions, and new promises. The opportunity is real. But the central question remains deeply scientific:
What does it mean?
Biology is data with meaning
I believe AI will transform biology. It will help us read genomes, design proteins, interpret cells, mine literature, integrate knowledge, generate hypotheses, and accelerate discovery.
But the most important progress will not come from treating biology as a giant spreadsheet.
It will come from building systems that understand relationships, constraints, mechanisms, uncertainty, and context.
Biology is not just data.
Biology is data with history.
Data with structure.
Data with environment.
Data with measurement noise.
Data with evolutionary memory.
Data with experimental limits.
Data with meaning.
The future of AI biology will depend on how well we teach machines to work with that meaning.
And that work cannot be done by AI alone.
It requires biologists who understand data, computational scientists who understand models, engineers who understand systems, and teams that respect the complexity of life.
In the first essay of this series, I wrote that AI will not replace biologists, but biologists who understand AI systems will replace those who only use AI tools.
In this second essay, I want to go one step deeper:
Biologists who understand context will be essential to the future of AI biology.
Because in biology, context is not background.
Context is the bottleneck.
And perhaps, context is also the key.
If we want AI to become truly useful for biology, the question is not only how large the model is.
The question is:
What context does the system need in order to make a biological answer trustworthy?
That is the conversation I hope more biologists will join.
[1] Zou, J., Huss, M., Abid, A., Mohammadi, P., Torkamani, A. & Telenti, A. A primer on deep learning in genomics. Nature Genetics 51, 12–18 (2019). doi:10.1038/s41588-018-0295-5.
[2] Eraslan, G., Avsec, Ž., Gagneur, J. & Theis, F. J. Deep learning: new computational modelling techniques for genomics. Nature Reviews Genetics 20, 389–403 (2019). doi:10.1038/s41576-019-0122-6.
[3] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nature Methods 18, 1196–1203 (2021). doi:10.1038/s41592-021-01252-x.
[4] Regev, A. et al. The Human Cell Atlas. eLife 6, e27041 (2017). doi:10.7554/eLife.27041.
[5] Wilkinson, M. D. et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 3, 160018 (2016). doi:10.1038/sdata.2016.18.
[6] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2.
[7] Nicholson, D. N. & Greene, C. S. Constructing knowledge graphs and their biomedical applications. Computational and Structural Biotechnology Journal 18, 1414–1428 (2020). doi:10.1016/j.csbj.2020.05.017.
[8] Li, M. M., Huang, K. & Zitnik, M. Graph representation learning in biomedicine and healthcare. Nature Biomedical Engineering 6, 1353–1369 (2022). doi:10.1038/s41551-022-00942-x.
No comments:
Post a Comment