Causal discovery in biology

How causal discovery helps biologists turn measurements into testable hypotheses: signalling and gene regulation, which experiment to run next, how to prepare omics and cytometry tables, and the pitfalls to avoid.

Why biologists use causal discovery

Modern biology measures many things at once: the whole transcriptome, dozens of proteins in single cells, hundreds of metabolites, the microbes in a sample. Correlation networks built from such data are dense: almost everything moves with something. The question a lab can act on is narrower: which molecule regulates which, and what happens if we block it?

Causal discovery reads the measurements and proposes a directed graph: which variables act directly on which, and which relationships are only indirect. It does not replace experiments. It decides which experiments are worth running, which saves time, reagents and animals.

What it helps with

• Signalling pathways. From measurements of phosphorylated proteins in many single cells, propose which kinase acts on which. A widely studied public example is the flow-cytometry dataset of Sachs and colleagues (Science, 2005), with eleven signalling proteins measured in many individual human immune cells; UCMT Labs uses it in its own examples.
• Gene regulation. From expression data, propose which transcription factors drive which genes, and separate direct targets from genes further down the cascade.
• Metabolism and the microbiome. Propose which metabolites or species drive others, rather than only co-occurring.
• Biomarkers and clinical measures. Separate markers that sit on the path to an outcome from markers that only move with it, before choosing a drug target or a test.

From graph to experiment

An open link, and the experiment that settles it

A discovered graph usually contains three kinds of links:

• Directed links the data supports clearly: candidates for confirmation.
• Undirected links, where the data cannot tell which way the effect runs: the most informative places for an experiment. Inhibiting or knocking down one side and measuring the other settles the direction.
• Links that contradict the literature: either a data problem (batch, mislabelling) or a finding worth a closer look.

Rank the links by how much they matter to your question and how cheap they are to test, and test the top of the list first.

A practical workflow

From measurement to the next experiment

1. Measure. One row per sample or per cell, one column per gene, protein or metabolite.
2. Choose the variables. Discovery works best on tens of variables, not on a whole genome. Start from a pathway, a set of differentially expressed genes, or the markers your question is about.
3. Clean. See the cleaning checklist below.
4. Discover and compare. Run the suggested method and one from another family. For data where unmeasured factors (cell state, cell cycle, tissue composition) are likely, include FCI, which marks where a hidden cause may act.
5. Read with knowledge. Check the graph against known pathways; the agreements build trust, the disagreements are the interesting part.
6. Test. Plan perturbations for the top-ranked open or surprising links.

Cleaning checklist

• Log-transform counts and intensities when relationships look multiplicative.
• Deal with batches: correct them, or include the batch as a column so it is not mistaken for biology.
• Average technical replicates; keep biological replicates as rows.
• Remove constant columns and duplicates (UCMT Labs' data check flags them).
• Remove direct identifiers of people; keep only what the analysis needs.

How to lay out your table

Leave the sample ID out of the discovery. A condition or batch column, coded as a number, can stay in: the graph then shows what it affects.

Pitfalls

• Feedback loops. Biology is full of them; most discovery methods assume none. A loop may appear as an undirected link or a single direction. Time-course data and perturbations help.
• Pooled conditions. Mixing very different conditions in one table can create links that hold in none of them. Run per condition as well as pooled.
• Few samples, many variables. With fifty samples, only strong effects are found. Reduce the variables before adding methods.
• Measured proxies. An antibody measures a phosphorylation state, not activity; the graph is about what you measured.

What causal discovery does not do

It does not prove mechanisms, and it does not replace perturbation experiments or the literature. Its value is in focus: from a hundred possible next experiments to the few most likely to tell you something. UCMT Labs gives you the data check, several methods to compare on the same table, and a what-if view to estimate what blocking a node might change, all without code.