From Correlation to Causation
Can AI Discover Biological Mechanisms?
In genomics, we are remarkably good at finding associations.
A variant is associated with disease.
A gene is correlated with drought tolerance.
A transcription factor is connected to hundreds of other genes.
An AI model predicts cellular responses with impressive accuracy.
We discover patterns everywhere.
But biology asks a harder question:
What actually causes what?
If a gene is highly expressed during drought stress, does it help the plant survive? Or is its expression simply a response to damage?
If a genetic variant is associated with disease, is it the causal variant? Does it regulate the nearby gene, a distant gene, or something else entirely?
If an AI model predicts that perturbing a gene will change cellular state, does that mean it has learned the underlying biological mechanism?
These are not merely questions about prediction.
They are questions about causation.
And I believe this distinction may define one of the next frontiers of AI biology.
In the first essay of this series, I argued that biologists need to become AI system thinkers.
In the second, I argued that biological context is essential for meaningful AI.
In the third, I explored why benchmark performance is not the same as biological discovery.
In the fourth, I asked whether AI could move beyond prediction and participate in the experimental cycle itself.
Now comes a deeper question:
Can AI move beyond recognizing biological patterns to discovering why those patterns exist?
Because ultimately, science is not satisfied with knowing what happens.
We want to understand why it happens — and what would happen if we changed the system.
1. Correlation is powerful. But it is not causation.
Consider a simple example.
Imagine that we analyze thousands of plants under drought stress and discover that Gene X is strongly correlated with drought tolerance.
An AI model ranks Gene X among the most important genes.
The statistics are convincing.
The association replicates.
The model achieves impressive predictive performance.
Should we conclude that Gene X causes drought tolerance?
Not yet.
There are several possible explanations.
Possibility 1: Gene X contributes to drought tolerance.
Changing its activity could alter how the plant responds to water limitation.
Possibility 2: Gene X responds to drought stress.
Its expression increases because the plant is stressed, but changing the gene may not improve tolerance.
Possibility 3: Gene X is associated with another causal factor.
Perhaps genetic linkage, environmental differences, or developmental state explains the correlation.
All three scenarios could produce a strong statistical association.
But they represent very different biology.
Figure 1. The same association can have different causal explanations. Gene activity may influence a phenotype, a shared factor may affect both gene expression and phenotype, or expression may be a consequence of an altered biological state.
A predictive model may distinguish drought-tolerant plants from drought-sensitive plants without identifying the mechanism responsible for the difference.
That does not make the model useless.
Prediction can be extremely valuable.
But predictive usefulness and causal understanding are different scientific achievements.
This distinction becomes especially important as foundation models learn increasingly complex relationships from large biological datasets.
A model can capture correlations extraordinarily well.
But whether those correlations represent causal biological relationships is another question.
2. From genetic association to biological mechanism
Population genetics offers a particularly clear example of this challenge.
Genome-wide association studies have transformed our ability to discover genetic loci associated with complex traits.
We can now identify thousands of variants associated with disease, crop performance, environmental adaptation, and many other phenotypes.
But identifying a locus is not the same as identifying the causal gene.
And identifying the causal gene is not the same as understanding its mechanism.
Consider a hypothetical GWAS signal associated with drought tolerance.
The associated marker may be in linkage disequilibrium with the functional variant.
The functional variant may lie in a noncoding regulatory region.
That region may affect a gene hundreds of kilobases away.
Its regulatory effect may be tissue-specific.
And the resulting phenotype may depend on genotype, developmental stage, and environment.
How do we move from the association to a credible mechanism?
We need several layers of evidence.
Fine-mapping can prioritize candidate variants.
Expression QTL analysis can reveal genetic effects on gene expression.
Colocalization can assess whether molecular and trait associations are consistent with a shared causal variant.
Regulatory genomics can provide evidence about candidate enhancers and target genes.
Functional perturbation can test whether changing a candidate gene or regulatory element alters a biological outcome.
But none of these methods alone necessarily establishes the entire causal pathway.
Recent reviews of causal-gene prioritization and Mendelian randomization emphasize precisely this issue: even sophisticated genetic approaches depend on assumptions that must be examined, and causal claims are strengthened by combining statistical inference with molecular and experimental evidence [1,2].
The challenge is not simply finding another association.
It is connecting the pieces into an explanation.
Figure 2. Moving from genetic association to mechanism requires multiple layers of evidence: candidate prioritization, intervention, mechanistic measurements, and replication. Each stage strengthens the hypothesis but may also reveal alternative explanations.
In other words:
A genetic association is a clue.
A causal effect is stronger evidence.
A mechanism is an explanation of how that effect occurs.
3. Observation tells us what is associated. Intervention asks what changes.
There is a useful conceptual distinction in causal inference.
Observational analysis asks:
What happens when X and Y vary together?
Interventional analysis asks:
What happens to Y when we deliberately change X?
Counterfactual reasoning asks:
What would have happened to Y if X had been different, all else appropriately accounted for?
These questions are related, but they require different assumptions and evidence.
For example, suppose a transcription factor is correlated with the expression of several hundred genes.
We might infer a regulatory network.
But does the transcription factor actually regulate those genes?
Or are they responding together to another upstream signal?
Now imagine that we knock down the transcription factor and measure the transcriptome.
If certain genes change expression reproducibly, we obtain stronger evidence that the transcription factor influences their expression under the tested conditions.
This is why perturbation technologies are transforming systems biology.
CRISPR screens, Perturb-seq, reporter assays, targeted genetic manipulations, and other experimental approaches allow us to move beyond observing correlations toward testing interventions.
In 2025, Jiang and colleagues demonstrated scalable single-cell perturbation screening across more than 1,500 perturbations, six cell lines, and five signaling contexts. Their work illustrates how experimentally measured perturbation responses can help reconstruct molecular pathway signatures rather than relying entirely on observational expression relationships [3].
The transition is important.
Instead of inferring relationships only from which genes move together, we begin asking which changes follow when a specific regulator is experimentally altered.
But even here, caution is necessary.
A perturbation may have off-target effects.
Knocking out a gene may trigger compensation.
The measured effect may depend on cell type or experimental conditions.
And a downstream response does not automatically indicate a direct molecular interaction.
Intervention provides stronger causal evidence, but intervention alone does not guarantee a complete mechanistic explanation.
That distinction matters.
4. From causal effects to mechanisms
This is where I think the scientific challenge becomes particularly interesting.
Suppose we knock out Gene X.
The phenotype changes.
We now have evidence that Gene X contributes causally to the phenotype under the experimental conditions.
But do we understand the mechanism?
Not necessarily.
We still need to ask:
Which molecular process changed?
Which downstream genes or proteins mediate the effect?
Is the effect direct or indirect?
At what developmental stage does it occur?
Does it depend on genetic background?
Can restoring Gene X rescue the phenotype?
Can we intervene at an intermediate step and reproduce or block the outcome?
A mechanism explains not only that an intervention changes an outcome, but how the influence propagates through the biological system.
This is the distinction highlighted in a 2026 Nature Reviews Genetics perspective by Maizels and Briscoe [4].
The authors argue that gene regulatory networks are increasingly represented as large statistical correlation graphs. These networks may be visually impressive, but they do not necessarily explain molecular causation.
They advocate integrating biologically grounded model structure, molecular constraints, and experimental perturbations to move from descriptive relationships toward mechanistic explanations.
I think this is a crucial point.
A network containing thousands of connections is not automatically an explanation.
A highly accurate model is not automatically a mechanism.
And a model that reproduces observed data may still fail to explain how the system responds when disturbed.
To understand biology, we need more than a map of relationships. We need testable explanations of how changes propagate through those relationships.
5. A real example: connecting genes, cellular programs, and traits
One of the most interesting recent studies addressing this challenge was published in Nature in 2026.
Ota and colleagues developed an approach called Causal modelling of gene effects from regulators to programs to traits [5].
The problem they addressed is familiar to anyone working in statistical genetics.
GWAS and rare-variant studies identify many gene–trait relationships, but the molecular mechanisms linking those genes to organismal phenotypes often remain poorly understood.
The researchers integrated two complementary types of evidence:
Genetic evidence from gene-level loss-of-function burden analyses, providing estimates of how gene disruption relates to human traits.
Perturbation evidence from Perturb-seq, providing information about how experimentally changing genes affects cellular transcriptional programs.
By combining these data, they inferred regulatory relationships connecting genes, downstream programs, and blood traits.
Their results demonstrated how genetic evidence and experimental perturbation can be integrated to generate structured, testable hypotheses about biological mechanisms.
Importantly, the inferred causal structure remains dependent on model assumptions and the available experimental evidence. It is not equivalent to proving every molecular interaction in the system.
Nevertheless, the study illustrates a direction I find compelling:
From variants to genes.
From genes to cellular programs.
From cellular programs to phenotypes.
That is much closer to the biological explanation we ultimately want.
It also shows why the future of AI biology may require combining statistical genetics, molecular biology, perturbation experiments, and computational modeling rather than treating any one data type as sufficient.
6. Can foundation models learn causal biology?
This brings us back to AI.
Biological foundation models are becoming increasingly powerful.
They can learn representations from DNA sequences, proteins, single-cell transcriptomes, and other molecular data.
Some predict gene expression.
Others predict variant effects.
Some generate new biological sequences.
Others attempt to predict what happens when genes are perturbed.
But there is an important question:
Does a model that accurately predicts an intervention understand the causal mechanism?
Not necessarily.
A model might learn enough statistical structure to predict perturbation responses without learning an explicit, biologically faithful mechanism.
Conversely, a mechanistically structured model may offer a useful causal explanation while remaining incomplete or less accurate on some prediction tasks.
This makes evaluating causal claims especially difficult.
In 2025, Tejada-Lapuerta and colleagues discussed how causal machine learning could help single-cell genomics move toward predicting responses under new interventions and environments [6].
Their perspective highlights the importance of distinguishing observational patterns from intervention effects and incorporating causal assumptions into modeling.
A 2026 bioRxiv preprint, DoFormer: Causal Transformer for Gene Perturbation, takes a more explicitly intervention-aware approach. Its authors propose adapting a causal intervention operator within a transformer architecture and report improved prediction of unseen perturbation responses relative to their evaluated baselines [7].
These are promising directions.
But a model's use of causal terminology or intervention-aware architecture is not sufficient evidence that it has discovered a genuine biological mechanism.
We still need independent experimental tests.
In fact, a 2026 study examining regulatory signals inside scGPT made an important distinction: an intervention on a model's internal representation can establish a causal effect within the model's computation without necessarily proving a causal relationship in a living cell [8].
This distinction is subtle but fundamental.
Causality inside a neural network is not automatically causality inside biology.
7. What does it mean to predict a perturbation correctly?
Even evaluating perturbation models is more difficult than it first appears.
Suppose an AI system predicts gene-expression changes following CRISPR knockout.
We compare its predictions with experimental measurements.
The correlation is high.
Does this mean the model has learned cellular causality?
Not necessarily.
Some genes barely change after perturbation.
Many transcriptomic measurements share common baseline structure.
A model can achieve apparently strong aggregate performance while missing the most important perturbation-specific effects.
In 2025, Ahlmann-Eltze and colleagues reported that several deep-learning approaches did not consistently outperform simple linear baselines for predicting genetic perturbation effects [9].
However, a subsequent 2026 Nature Biotechnology study by Miller and colleagues showed that common evaluation metrics can be poorly calibrated for this problem. Using more informative metrics and positive-control baselines, they found that deep-learning perturbation models can outperform uninformative baselines under certain evaluation settings [10].
Together, these studies reveal an important lesson.
The question is not simply whether AI works or fails.
We need to ask:
Which intervention is being predicted?
What biological effect are we measuring?
Does the metric capture the signal we care about?
Does the result generalize to an unseen perturbation?
Is the model recovering a meaningful biological response?
This is why causal AI needs rigorous experimental design and evaluation.
A high prediction score does not establish mechanism.
And a poor benchmark score does not necessarily mean a model contains no useful biological information.
The evaluation must match the scientific claim.
8. Counterfactuals: the next frontier of biological reasoning
To me, one of the most interesting capabilities scientific AI could develop is counterfactual reasoning.
Biologists regularly ask questions such as:
What would happen if this gene were knocked out?
What if it were overexpressed?
What if the same variant occurred in another genetic background?
What if the plant encountered drought earlier in development?
What if we inhibited one pathway while activating another?
These questions go beyond describing observed data.
They ask us to reason about alternative biological worlds.
But counterfactual reasoning is difficult because we generally cannot observe both outcomes of the same intervention in the same biological unit under identical conditions.
We must estimate the unobserved alternative using experimental comparisons, causal models, and assumptions.
This is where combining biological knowledge with AI may become especially valuable.
Imagine a future AI system that can compare several competing mechanistic explanations and predict how each would respond to a carefully chosen intervention.
It might tell us:
Under Mechanism A, knocking out Gene X should reduce expression of downstream Gene Y.
Under Mechanism B, knocking out Gene X should have little effect on Gene Y but alter another pathway.
Therefore, measuring Gene Y after perturbation would help distinguish the hypotheses.
That would move beyond generic prediction.
It would support causal hypothesis discrimination.
And it connects naturally to Essay 4, where I argued that the most valuable experiment is not necessarily the one most likely to succeed.
It may be the one that most effectively reduces uncertainty.
9. Biology is contextual — and causality is too
There is one more complication.
Even when a causal relationship is established, it may not be universal.
A gene may causally influence a phenotype in one tissue but not another.
A regulatory pathway may operate differently across developmental stages.
A genetic perturbation may have one effect in a reference genotype and another in a different genetic background.
The same intervention may succeed under controlled conditions but fail under environmental stress.
This is particularly important in plant biology, where genotype-by-environment interactions can strongly influence phenotypes.
It is equally relevant in medicine, where disease state, cell type, genetic background, and treatment history influence biological responses.
A recent 2026 Nature Reviews Genetics review examines how genetic effects propagate across molecular, cellular, tissue, and organismal levels, emphasizing that integration across these scales remains a major challenge [11].
This leads to an important principle:
A causal effect is always estimated under some set of conditions.
One experiment does not automatically establish a universal biological law.
And one mechanistic explanation may need to be revised when the context changes.
This is why the themes of my first four essays are connected.
We need systems thinking because biology contains interacting components.
We need context because biological meaning depends on conditions.
We need evaluation because model performance can be misleading.
We need experiments because observation alone cannot resolve many causal questions.
And we need causal reasoning because the ultimate objective is not simply prediction.
It is understanding.
10. What should biologists contribute to causal AI?
As AI becomes more powerful, I believe the role of biological expertise becomes even more important.
A machine-learning researcher may build a sophisticated model of gene regulation.
But a biologist may recognize that a proposed relationship is inconsistent with developmental timing.
A computational model may identify a highly connected regulatory gene.
But a population geneticist may recognize that the association is influenced by linkage disequilibrium or population structure.
An AI system may propose a gene perturbation.
But an experimental biologist may recognize that the resulting phenotype could arise from compensation, off-target effects, or an inappropriate assay.
These are not minor technical details.
They determine what we can legitimately conclude.
Future-ready biologists should understand enough causal inference to question assumptions, enough AI to evaluate model claims, and enough experimental biology to recognize what evidence would distinguish competing explanations.
We do not all need to become experts in structural causal models.
But we should understand the difference between:
association and intervention,
prediction and explanation,
causal effect and mechanism,
model interpretation and biological validation.
That is scientific judgment in the AI era.
11. From pattern recognition to mechanistic discovery
So, can AI discover biological mechanisms?
I believe the answer is increasingly yes, but only with the right evidence.
AI can help identify patterns humans might miss.
It can integrate genetic and molecular evidence.
It can prioritize candidate regulators.
It can help model perturbation responses.
It can propose mechanistic hypotheses.
It can help identify experiments that distinguish among competing explanations.
And, as discussed in Essay 4, it may increasingly participate in closed-loop experimental discovery.
But generating a plausible mechanism is not the same as establishing one.
Mechanistic claims must survive interventions, independent measurements, alternative explanations, and relevant biological contexts.
The exciting possibility is not that AI will produce perfect causal explanations from data alone.
It is that AI may help scientists move more efficiently through the cycle:
Observation → Hypothesis → Intervention → Evidence → Mechanistic understanding
The scientist's role remains central, not because humans must always be correct, but because scientific discovery requires disciplined judgment about what evidence means.
The future of AI biology is not just prediction
The first generation of AI biology largely asked:
Can we predict biological outcomes?
A second generation increasingly asks:
Can AI generate useful hypotheses and guide experiments?
The next frontier may ask:
Can AI explain why biological systems behave as they do — and predict what would happen if we changed them?
That is a more ambitious challenge.
A strong correlation can guide our attention.
A successful prediction can help prioritize our work.
A controlled intervention can establish stronger evidence of a causal effect.
But a mechanism explains how the effect arises, through which processes, and under what conditions.
These are different achievements.
And each requires different kinds of evidence.
I believe the future of AI biology will not be defined only by models that predict more accurately.
It will be defined by systems that help us move from patterns to causes, from causes to mechanisms, and from mechanisms to new biological understanding.
Because biology is not simply about knowing which variables move together.
It is about understanding how living systems work.
And what happens when we change them.
Prediction tells us what may happen.
Causality asks what would happen if we intervened.
Mechanism explains how and why.
That is the journey from correlation to causation.
And perhaps it is also the journey from artificial intelligence toward a deeper form of scientific intelligence.
References and Further Reading
[1] Methods for Prioritizing Causal Genes in Molecular Studies of Human Disease: The State of the Art. Genetic Epidemiology (2026). https://doi.org/10.1002/gepi.70037
[2] Sanderson, E. et al. Challenges and future directions for Mendelian randomization. Nature Genetics 58, 984–994 (2026). https://doi.org/10.1038/s41588-026-02546-6
[3] Jiang, L. et al. Systematic reconstruction of molecular pathway signatures using scalable single-cell perturbation screens. Nature Cell Biology (2025). https://doi.org/10.1038/s41556-025-01622-z
[4] Maizels, R. J. & Briscoe, J. Gene regulatory networks: from correlative models to causal explanations. Nature Reviews Genetics 27, 485–498 (2026). https://doi.org/10.1038/s41576-026-00939-1
[5] Ota, M., Spence, J. P., Zeng, T. et al. Causal modelling of gene effects from regulators to programs to traits. Nature 650, 399–408 (2026). https://doi.org/10.1038/s41586-025-09866-3
[6] Tejada-Lapuerta, A. et al. Causal machine learning for single-cell genomics. Nature Genetics 57, 797–808 (2025). https://doi.org/10.1038/s41588-025-02124-2
[7] Karbalayghareh, A., Paull, E. & Califano, A. DoFormer: Causal Transformer for Gene Perturbation. bioRxiv (2026). https://doi.org/10.64898/2026.05.02.722054. Preprint.
[8] Kendiukhov, I. Causal intervention validation of gene regulatory signals in scGPT. Journal of Biomedical Informatics 180, 105080 (2026). https://doi.org/10.1016/j.jbi.2026.105080
[9] Ahlmann-Eltze, C., Huber, W. & Anders, S. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 22, 1657–1661 (2025). https://doi.org/10.1038/s41592-025-02772-6
[10] Miller, H. E. et al. Deep learning perturbation models can outperform baselines on calibrated metrics. Nature Biotechnology (2026). https://doi.org/10.1038/s41587-026-03307-w
[11] Laman Trip, D. S. & Beltrao, P. Human genetics across levels of biological organization. Nature Reviews Genetics (2026). https://doi.org/10.1038/s41576-026-00991-x
Note on preprints
Reference [7] is a bioRxiv preprint and has not yet undergone peer review. I include it because intervention-aware foundation models are an active area of research. Its findings should be interpreted as preliminary evidence rather than established consensus.
