You measure gene activity and a phenotype. They are correlated. Easy. Then you delete the gene and expect the same toolbox to work. That’s where the trouble starts.
Step 1: What the data actually tells you
Observational data tells you one thing only: when gene A varies, the phenotype tends to vary with it. A model can learn that association and become very good at predicting the phenotype given observed gene A.
Observational data does not tell you:
- whether gene A causes the phenotype,
- whether the phenotype regulates gene A, or
- whether a hidden process drives both.
Correlation is symmetric. Cause is not.
Step 2: Why the deletion question is different
Deleting a gene is not asking: “What phenotype value usually shows up when gene A is low?” It is asking: “What would the phenotype be if I force gene A to zero?”
Observation conditions on the world as it is.
Intervention changes how the world works.
A regression model is a conditioning machine. A gene deletion is a rewiring operation. If you treat rewiring like conditioning, you are answering a different question than the one you asked.
Step 3: The hidden fork in the road
Here is the core reason causality matters. Two different causal worlds can generate the same observational correlations:
World A: Gene A → Phenotype
World B: Phenotype → Gene A
In World A, deleting gene A should change the phenotype. Gene A is upstream.
In World B, deleting gene A may do little or nothing to the phenotype. Gene A is downstream.
Observational data alone cannot distinguish these worlds. If you refuse to represent causal direction, you are forced to give the same prediction in both cases.
That prediction is not “conservative”. It is logically unsupported. The only defensible answer becomes: I do not know.
Step 4: What a causal model actually is
A causal model is not just a graph. It is a generative recipe for how variables are produced. A common framework is the Functional Causal Model, also called a Structural Equation Model.
Each variable is computed from its direct causes (its “parents”) plus a noise term that absorbs everything else you did not model.
A = fA(ParentsOfA, UA)
P = fP(ParentsOfP, UP)
The graph states which variables feed into which. The functions state how they feed in. The noise captures unknowns and variation.
Now the intervention becomes clean and mechanical:
Delete gene A means replace its equation with a constant:
do(A = 0)
You cut the incoming arrows into A, fix A to zero, and keep the rest of the system unchanged. Then you compute the new phenotype distribution implied by the remaining equations.
Only at this point does “best prediction after deletion” become a well-defined object:
it is the expected phenotype (or full distribution) under the intervention do(A=0).
Why this matters beyond genes
This is not a biology quirk. It is a universal structure problem. The same confusion shows up when people try to:
- estimate the effect of an advert from click data,
- infer policy impact from historical trends,
- personalise recommendations and assume behaviour stays stable,
- deploy ML models and act surprised when the world changes in response.
If your question contains “what happens if we do X”, you are in the intervention business. Correlation alone cannot run that business.
The payoff of causal thinking
With a causal model you can:
- compute intervention distributions from observational data when assumptions hold,
- reason about hidden variables instead of pretending they do not exist,
- combine observational and interventional datasets coherently,
- separate mechanism from coincidence,
- build models that survive deployment because they understand the knobs.
The core lesson
Prediction answers: “What tends to happen?”
Causality answers: “What would happen if I act?”
Deleting a gene is an action. Correlation has no opinion on actions. Causality is not decorative theory. It is the minimum structure required to make the question answerable.
