From healthy breast to lncRNA drivers: two papers, two ends of my thesis

I paired these two papers for a reason: they sit at the two ends of my own project. One maps what the healthy breast is made of, cell type by cell type. The other hunts for non-coding driver mutations at population scale, which is the core of my thesis. Read together, they bracket the exact gap I’m trying to fill: you need a good description of normal before you can say how a tumour diverged, and you need a powered way to find the mutations that pushed it there.

Paper 1, a single-cell reference for the healthy breast

Breast cancer is heterogeneous, and mutations alone don’t explain that heterogeneity. The cell a tumour starts from, and the state of the tissue around it, matter too. So before you can model divergence, you need normal mapped properly.

Reed and colleagues built the Human Breast Cell Atlas from 55 donors, women who’d had a reduction mammoplasty, or a risk-reducing mastectomy because they carried a BRCA1/2 mutation. After QC they kept 801,360 single cells, grouped into 41 subclusters across the epithelial, stromal and immune compartments. That scale is the point: it’s a reference, not a snapshot.

Three findings stuck with me:

• Age and parity leave a fingerprint. Using only the 22 average-risk donors (to avoid BRCA confounding) and a differential-abundance test (Milo), they show age shifts the luminal-progenitor subclusters and drives “lineage infidelity”, cells losing their canonical markers. Several epithelial changes go in opposite directions for age versus parity, which is striking because age and parity also have opposing effects on breast cancer risk.
• BRCA carriers show an immune shift, but an exhausted one. High-risk tissue has more immune cells (a ~2-fold enrichment of a pro-inflammatory CD8 TC1 subset in BRCA1 carriers), yet those cells carry an exhaustion signature: PD1, PDL1, LAG3, TIGIT, HAVCR2.
• The headline: that exhaustion signature appears in tissue with no tumour in it, validated by immunohistochemistry. Immune escape may begin during initiation, in histologically normal tissue.

Paper 2, finding driver lncRNAs in 12,631 genomes

This one is much closer to the centre of my thesis. Over 90% of somatic mutations in cancer fall outside protein-coding genes, and long non-coding RNAs are a large, poorly understood class of those non-coding genes. The problem is power: finding driver lncRNAs needs enormous cohorts.

The Johnson lab analysed 12,631 whole tumour genomes from the 100,000 Genomes Project and found 121 driver lncRNAs under positive selection across 19 cancer types, roughly two-thirds of tumours carry at least one. Earlier efforts like PCAWG had fewer than 3,000 genomes and were underpowered; here, 13 cancer types are powered to detect lncRNAs mutated 5% above background.

The method matters for me because I want to reuse it. They test for two signatures of positive selection, excess mutational burden and functional impact in lncRNA exons, correcting for local trinucleotide mutation rate, gene length and mutational signatures. The novelty is engineering: a rebuilt pipeline (ExInAtorPS) that runs ~276× faster, finishing the whole pan-cancer analysis in ~31 minutes. And they validate before trusting it, recovering known coding drivers at 67–80% precision, an empirical FDR of 0.6%, and 4.6-fold enrichment for known cancer lncRNAs.

Breast cancer had the most driver lncRNAs of any type, 27 of them. The most recurrent is SNHG14 (mutated in ~35% of tumours), then NEAT1 and MALAT1. Recurrence at the same positions is one of the cleanest signatures of selection.

What the mutations actually hit

An element-level pipeline (ExInAtorEL) shows the mutations concentrate in conserved RNA secondary structures, RNA-binding-protein (RBP) sites, and conserved sequence elements. Then the part most relevant to me: a convolutional neural network (Pysster) predicts how each mutation changes RBP binding affinity. Most mutations reduce binding, and they cluster in the core recognition motif.

The proposed mechanism is elegant: in healthy cells, RBPs bind and repress these oncogenic lncRNAs; driver mutations disrupt that interaction and release the lncRNA. Their worked example is MALAT1 and FTO (an m6A demethylase), a breast-tumour mutation reduces FTO binding, MALAT1 rises, and m6A on MALAT1 goes up. A somatic genome change wired straight into the epitranscriptome.

The deeper idea: a large enough tumour-genome cohort can be used like a high-resolution mutagenesis screen that maps the functional sites inside a non-coding gene.

Why I read them together

The atlas tells me what the healthy breast is; the lncRNA screen shows how to find what breaks it, at scale, in the non-coding genome. My contribution is the join: a breast-specific driver search that’s aware of the subtype-resolved regulatory regions I’m mapping in Objective 1, so a candidate non-coding mutation isn’t just “recurrent”, it’s recurrent in an element that’s actually active in the right cell state. That’s the difference between a statistical hit and a mechanistic hypothesis.