The HOX Code in Cancer: Exploring a Master Regulator Family with R2
Introduction
In 1978, Edward Lewis described fruit flies with legs sprouting from their heads where antennae should be. The culprit was a mutated Hox gene — one of a conserved set of master regulators that tell every cell in a developing animal where it is in the body and what it should become. The discovery earned Lewis a Nobel Prize, and the gene family its name.
Decades later, those same genes are drawing attention in oncology — not for growing legs in the wrong place, but for something arguably more dangerous: hijacking the developmental logic of human tumours.
In humans, the HOX family comprises 39 genes arranged in four clusters — HOXA, HOXB, HOXC, and HOXD — spread across chromosomes 7, 17, 12, and 2 respectively. Their arrangement is not accidental. The genes are expressed in the same order along the chromosome as they are along the head-to-tail body axis during development. This principle — known as collinearity — is one of the most striking examples of genome organisation mirroring biological function in all of biology.
In cancer, that collinear logic breaks down. HOX genes are aberrantly activated or silenced across virtually every tumour type, and their misexpression is increasingly linked to stemness, therapy resistance, and poor patient outcomes. Yet because the family is so large and so functionally redundant, studying any single HOX gene in isolation misses the bigger picture.
That is where R2: Genomics Analysis and Visualization Platform comes in. R2 is a web-based platform developed at Amsterdam UMC that lets researchers interrogate gene expression across hundreds of curated cancer datasets — without writing a single line of code. In this post, we will use HOX genes as a showcase: a biologically rich, clinically relevant gene family that reveals what R2's tools can do when you move beyond studying one gene at a time.
The Cluster Architecture: Start With the Genome Browser
Before we look at expression data, it is worth pausing at the genome itself — because the physical organisation of HOX genes is part of what makes them so unusual.
Open R2 and navigate to the Genome Browser. Enter the coordinates for the HOXA cluster on chromosome 7 (roughly chr7:27,100,000–27,280,000 in hg38). What you immediately see is striking: ten genes packed into less than 200 kilobases, all oriented in the same direction, with almost no intervening sequence. Zoom out to include the flanking regions and the contrast is stark — dense cluster, then silence.
Now overlay epigenetic tracks. In stem cells and early progenitors, the HOXA cluster carries a characteristic bivalent chromatin signature: H3K4me3 (a mark of active promoters) sits alongside H3K27me3 (deposited by the Polycomb repressor complex). This co-occupancy keeps genes poised — neither fully on nor fully off — ready to be activated in one direction or the other as cells differentiate. In cancer, this balance is frequently disrupted: whole-cluster silencing through aberrant Polycomb activity, or wholesale activation through loss of repression.
The genome browser makes this visible at a glance. More importantly, it reminds us of something that a table of expression values cannot: the genes in a HOX cluster are not independent units. They share regulatory elements, respond to the same chromatin environment, and are likely co-regulated as a block. That spatial logic should inform how we interpret their expression.
All 39 HOX genes by cluster. Gaps reflect genes lost during vertebrate evolution; HOXA is the most complete cluster, HOXC and HOXD the most reduced at the anterior end.
Cluster | Chromosome | Genes (anterior → posterior) | Count |
|---|---|---|---|
HOXA | 7 | HOXA1, HOXA2, HOXA3, HOXA4, HOXA5, HOXA6, HOXA7, | 11 |
HOXB | 17 | HOXB1, HOXB2, HOXB3, HOXB4, HOXB5, HOXB6, HOXB7, HOXB8, HOXB9, | 10 |
HOXC | 12 |
| 9 |
HOXD | 2 | HOXD1, | 9 |
Who Is Expressed Where? The Family-Wide Heatmap
With the genomic architecture in mind, the next question is simple: across tumour types, which HOX genes are actually expressed?
This is where R2's Heatmap tool earns its keep. Select all 39 HOX genes as a custom gene set, choose a large pan-cancer cohort — TCGA's pan-cancer dataset works beautifully here — and generate the heatmap. Set samples as columns and genes as rows, and allow both to cluster by similarity.
Several things jump out immediately.
First, the cluster structure is preserved in tumours. Genes from the same genomic cluster tend to group together in the expression heatmap, even though the clustering algorithm has no knowledge of chromosomal location. HOXA genes co-cluster; HOXD genes co-cluster. The developmental collinearity echoes, faintly but unmistakably, through cancer transcriptomes.
Second, there is tissue-of-origin specificity. Leukaemia samples light up with HOXA and HOXB expression. Brain tumours show a distinct HOXD-enriched pattern. Epithelial cancers — lung, breast, colon — display yet another configuration. The heatmap is, in effect, a molecular atlas of the HOX code across tumour types, and you have generated it with a handful of clicks.
Third — and this is where it gets clinically interesting — many samples show heterogeneous HOX expression within the same tumour type. Not all gliomas have the same HOX signature. Not all AMLs look alike. That within-type heterogeneity is a hint that HOX expression may be tracking something biologically meaningful beneath the surface of the clinical diagnosis.
Before moving on, pull out one gene to explore individually. HOXA9 is a good choice — it is among the most consistently dysregulated HOX genes across cancer types. Use R2's Gene Expression View to plot HOXA9 across every dataset in the platform. The range is remarkable: near-zero in most normal adult tissues, dramatically elevated in AML, substantial in a subset of gliomas, variable in neuroblastoma. That expression range is not noise. It is a biological signal waiting to be interrogated.
HOX Expression as a Subtype Discriminator: the t-SNE Map
The heatmap showed us heterogeneity within tumour types. The next step is to ask: does that heterogeneity map onto known biology?
Load a well-annotated AML dataset — the TCGA AML cohort, or the Verhaak dataset which includes defined molecular subgroups — and run R2's t-SNE module using all 39 HOX genes as the input feature set. Ask R2 to colour the resulting map by known cytogenetic subtype: NPM1-mutant, RUNX1-RUNX1T1 fusion, MLL-rearranged, and so on.
tSNE using only the HOX genes
The result is one of those moments that justifies the entire approach. The samples separate cleanly into islands. MLL-rearranged AMLs — the cases where the MLL gene is fused to a partner, driving wholesale HOXA cluster activation — cluster tightly on one side of the map. NPM1-mutant cases, which also show elevated HOXA expression but through a different mechanism, sit in a related but distinct neighbourhood. Normal karyotype AMLs scatter across a broader region.
You did not tell the algorithm about cytogenetics. You gave it HOX expression alone. And it recapitulated known molecular subgroups.
This is an important demonstration of what R2's dimensionality reduction tools can do in a teaching or exploratory context. The t-SNE map is not a new discovery — the relationship between HOX expression and AML subtype is well established in the literature. But being able to reproduce it interactively, with your own choice of genes and cohort, makes it a powerful pedagogical tool and a fast sanity check when working with a new dataset.
For a more striking visual, colour the same t-SNE map by HOXA9 expression as a continuous variable rather than by subtype label. The gradient will align almost perfectly with the cluster structure — a vivid illustration of how a single gene tracks the broader epigenetic state of the whole cluster.
Does HOX Expression Matter for Patients? Survival Analysis and the Megasampler
Cluster separation is elegant, but the question oncologists care about is blunter: does it affect survival?
Open R2's Kaplan-Meier module and select a glioma cohort — the combined TCGA GBM+LGG dataset is large enough to give statistical power and spans a wide range of grades and molecular subtypes. Select HOXA9 as the gene of interest and allow R2 to determine the optimal expression cutoff. Run the analysis.
The curve split is dramatic. High HOXA9 expression is associated with markedly worse overall survival — a log-rank p-value well below 0.01 in most glioma cohorts. The median survival difference between the top and bottom quartiles can exceed two years. For a gene that most neuropathologists have never considered, that is a striking finding.
But a single survival curve in a single cohort is not a discovery — it is a hypothesis. This is where R2's Megasampler becomes essential. Keep HOXA9 selected and switch to the Megasampler view. Add ten to fifteen datasets spanning different tumour types: glioma, AML, neuroblastoma, breast cancer, lung adenocarcinoma. R2 will display HOXA9 expression across all of them side by side, in a single plot, normalised to allow comparison.
Two things become immediately clear. First, HOXA9 elevation is not a glioma-specific phenomenon — it recurs across multiple lineages. Second, and more interestingly, the degree of elevation varies enormously: AML sits at the extreme high end, normal brain tissue near zero, and most solid tumours fall somewhere in between. This context is critical for interpreting any single cohort's survival result. High HOXA9 in a glioma may not mean the same thing mechanistically as high HOXA9 in AML, even if the prognostic direction is similar.
For a second survival example that tells a different story, switch to HOXB13 in prostate cancer. Here the relationship flips in some datasets — HOXB13 has been reported as both a tumour suppressor and an oncogene depending on the androgen receptor context. The Megasampler will show you that its expression is highly lineage-restricted in ways that HOXA9 is not. This is a useful teaching moment: within the same gene family, the cancer biology can be qualitatively different.
What Does a HOX Gene Actually Do in a Tumour? Correlating Genes and GO Enrichment
A survival association tells us that HOXA9 expression matters. It does not tell us why. For that, we need to understand what else is happening in tumours where HOXA9 is high.
In the same glioma cohort, open R2's Find Correlated Genes module with HOXA9 as the anchor. R2 will rank all other genes in the dataset by Pearson correlation with HOXA9 expression across samples. Export the top 200 positive correlates and feed them into R2's built-in Gene Ontology enrichment analysis.
The top enriched terms are revealing. You will typically see:
- Anterior/posterior pattern specification — the developmental HOX programme itself, indicating that HOXA9 is co-expressed with other positional identity genes
- Chromatin organisation and histone modification — consistent with the Polycomb/Trithorax axis that governs HOX cluster regulation
- Stem cell maintenance and self-renewal — a recurring theme in HOX-high tumours across lineages
- Response to growth factors and receptor tyrosine kinase signalling — connecting HOX activity to the proliferative programmes that sustain tumour growth
What is striking about this list is how little it looks like a standard cancer gene signature. There is no MYC, no p53, no KRAS. Instead, the HOXA9 transcriptional neighbourhood is populated by developmental genes — transcription factors, chromatin regulators, and niche-signalling components that one would more naturally associate with a foetal tissue than an adult tumour.
This is the central insight of the HOX-in-cancer field, and the correlation analysis in R2 surfaces it in about five minutes: tumours with high HOXA9 are not simply growing faster — they are recapitulating a developmental transcriptional state. They have, in a meaningful sense, regressed towards a less differentiated, more stem-like identity. That is why they are aggressive, why they resist differentiation-inducing therapies, and why HOXA9 makes such a good prognostic marker — it is a proxy for the degree to which a tumour has shed its adult identity.
Run the same analysis on the negative correlates of HOXA9 — the genes that go down when HOXA9 goes up. You will find differentiation markers, lineage-specific transcription factors, and cell-type identity genes. The picture is symmetric: HOX activation and differentiation suppression are two sides of the same coin.
Not All HOX Genes Tell the Same Story: Differential Expression Across the Family
Everything so far has used HOXA9 as the protagonist. But one gene is not a family, and treating all 39 HOX members as interchangeable would be a mistake. The clusters have distinct biological identities, and R2's multi-gene tools let us respect that.
Create a custom gene set in R2 containing all 39 HOX genes — or four separate sets, one per cluster. Then use the Find Differential Expression module to ask a specific question: which HOX genes most distinguish, say, MLL-rearranged AML from NPM1-mutant AML?
The answer is almost entirely HOXA and HOXB genes. HOXC and HOXD members barely move. Within HOXA, it is the posterior members — HOXA7 through HOXA11 — that drive the separation. The anterior HOXA genes (HOXA1–3) are less discriminating. This fine-grained specificity is lost entirely if you collapse the family into a single score.
For a different angle, run the differential expression analysis between two solid tumour subtypes — say, basal-like versus luminal breast cancer. Here the story shifts: HOXB and HOXC genes emerge as the discriminators, and the AML-relevant HOXA posterior genes are nearly absent from both. The HOX code is tissue-specific in cancer just as it is in development.
Finally, try building a HOX cluster signature score using R2's gene set scoring functionality. Create four separate scores — one per cluster — and plot them against each other across a pan-cancer cohort. A scatter plot of HOXA score versus HOXD score will show that most tumour types sit in one of a small number of distinct configurations: HOXA-high/HOXD-low (haematological malignancies), HOXD-high/HOXA-moderate (some brain tumours), or double-low (many epithelial cancers). There is no tumour type that is uniformly high for all four clusters simultaneously — a reminder that the HOX code, even when deregulated, retains a vestige of its developmental specificity.
This is the practical lesson for researchers: when studying HOX genes in your tumour of interest, define your hypothesis at the cluster level, not the family level. R2 makes it straightforward to test cluster-specific patterns, and the biology rewards the specificity.
From Body Plan to Bedside: What HOX Genes Teach Us About R2
We began with a fly with legs on its head, and we have ended with survival curves, molecular subtype maps, and transcriptional signatures that look more like embryo than tumour. That is, in a sentence, the HOX story in cancer: developmental logic persisting — and malfunctioning — in a context where it was never meant to operate.
But this post has also been a demonstration of something else: what it looks like to interrogate a gene family systematically, rather than one gene at a time.
In a single analytical session in R2, we have:
- Visualised the genomic architecture of the HOX clusters in the genome browser — understanding the physical context before we opened any expression data
- Mapped expression across the entire family using a pan-cancer heatmap — seeing cluster-specific patterns emerge without any prior assumptions
- Stratified tumour samples by HOX expression alone using t-SNE — recapitulating known molecular subtypes without providing subtype labels
- Linked expression to outcome using Kaplan-Meier survival analysis — and validated the finding across independent cohorts with the Megasampler
- Deconstructed the biology using correlating gene analysis and GO enrichment — revealing that HOX-high tumours are developmentally rewired, not simply proliferating faster
- Compared across the family using differential expression tools — showing that HOXA, HOXB, HOXC and HOXD have distinct cancer roles that cluster-aware analysis can resolve
None of this required writing code. All of it required biological thinking — about what to ask, which cohorts to trust, and what the patterns mean. R2 is designed to keep the computational friction low so that the biological reasoning can stay front and centre.
If you want to explore further, the R2 platform is freely accessible at r2.amc.nl. Start with any HOX gene that is relevant to your tumour type, follow the correlation network, and see where it leads. The body plan has been hiding in the cancer genome all along — you just need the right tools to find it.
Comments
Post a Comment