A decade in the non-coding genome (2015–2025)

This one was a review rather than a single paper, I wanted to step back and trace how non-coding cancer genomics actually evolved over ten years, because that arc is the justification for my PhD. It’s also personal: the story starts with my supervisor’s work, and it ends right where I’m trying to push it. Here’s the decade, and the through-line I took from it.

The premise: 98% of the genome, mostly ignored

Only ~1.5–2% of the genome is protein-coding. The other ~98% is non-coding, promoters, enhancers, silencers and insulators (CTCF sites), UTRs, introns, non-coding RNAs (lncRNAs, miRNAs), and vast intergenic stretches. Mutations out there can change gene expression without touching a protein: shifting transcription-factor binding, chromatin accessibility or RNA processing, enough to activate an oncogene (the textbook case is the TERT promoter) or silence a tumour suppressor.

Ten years of non-coding driver discovery: from the first hotspot scans (2015) to pan-cancer paucity (2020), genome-wide methods (2024) and the functional-validation era (2025).

2015–2017, hotspots and the first scans

The early era was about finding recurrence. Piraino & Furney’s work laid out why non-coding mutations matter and built unbiased genome-wide scans: tile the genome into 50 bp windows and score each on recurrence (mutations above local background) and conservation (evolutionary constraint), then filter out low-mappability and hypermutated regions. The deliverable was a unified ranking of candidate driver regions, coding and non-coding together. This is, essentially, the ancestor of the hotspot-calling I do in my Objective 2.

The first scans tiled the genome into 50 bp windows, scoring each on recurrence and conservation to rank candidate driver regions.

2018–2021, the pan-cancer verdict, and a hard truth

Then came scale: PCAWG / Rheinbay et al. (2020) searched 2,658 whole genomes, integrating 13 driver-discovery algorithms with Brown’s method to combine dependent p-values, and filtered aggressively (discarding 46% of significant hits as artefacts of APOBEC/UV/AID mutational processes).

The verdict was sobering: beyond TERT, non-coding point-mutation drivers were “surprisingly limited”, an estimated 96 promoter driver mutations (73 of them TERT) versus >1,475 in coding sequence.

Across 2,658 genomes, aggressive filtering of mutational-process artefacts left almost nothing standing beyond the TERT promoter, the field's sobering "paucity" verdict.

But the honest reading, which Elliott & Larsson’s review captures, is that paucity is entangled with detectability:

GC bias. Promoters are GC-rich, GC-rich regions get poor WGS coverage, so real drivers get missed. Up to a third of PCAWG tumours had <20% sensitivity at the TERT hotspots themselves.
Targeted depth rescues signal. Deep-sequencing the FOXA1 promoter across 360 breast tumours recovered recurrent mutations invisible at standard depth (sensitivity was ~1% in the TCGA breast WGS), though even then they sat at 2.9%, and FOXA1’s high genome-wide indel rate hints some 3′UTR indels are passengers, not drivers. A good lesson in not over-reading recurrence.
Structural variants matter too. PCAWG’s recurrent-breakpoint methods found a recurrent focal microdeletion at BRD4 that lowers expression of an otherwise-amplified gene in breast and ovarian tumours, the first evidence of a microdeletion limiting an amplified gene.

The line I keep quoting (Elliott & Larsson): mutation-rate models can never perfectly reflect reality, so selection p-values are "simply scores that reflect… whether a signal is interesting, given today's knowledge." Every non-coding hotspot list is provisional. That humility is a design requirement, not a disclaimer.

2022–2024, going genome-wide and integrative

The methods matured. Dietlein et al. (2024) ran a genome-wide sliding window over 61.2 million mutations from 3,949 patients, combining three significance tests on tiled 1/10/100 kb intervals: an epigenomic comparison (more mutations than the histone-ChIP signal predicts), an inter-tumour comparison (cancer-type-specific accumulation as a proxy for missing epigenomic data), and positional clustering. It’s the clearest statement that you detect non-coding drivers by integrating mutation patterns with regulatory context, not by recurrence alone.

Dietlein integrates three signals, epigenomic, inter-tumour and positional clustering, into a single genome-wide driver call.

2025, the functional-validation era

The frontier moved from “is this region recurrently mutated?” to “what does the mutation do?”, exactly the ultraconserved-elements study (Bayraktar 2025), which paired burden tests with CRISPR screens and showed mutated ncUCEs acting as real enhancers and silencers. Statistics → mechanism → bench.

The through-line I took for my PhD

Two things crystallised for me writing this:

The classical toolkit is almost entirely statistical. Look at the standard tables of non-coding selection tools, LARVA, OncodriveFML, ExInAtor and the rest, and none are really machine-learning or deep-learning models. The whole field’s detection layer is mutation-rate modelling. That’s the gap I want to work in: bringing learned regulatory representations (the kind AlphaGenome and MPRA models now provide) into driver prioritisation, not just counting.
“Paucity” was partly a measurement artefact. GC bias, coverage, and the wrong unit of analysis hid signal. A breast-specific, regulatory-map-aware search, with contact-supported enhancer–gene links and proper, conservatively-calibrated nulls, is a genuinely different experiment from the pan-cancer scans that concluded the cupboard was bare.

A decade in, the non-coding genome isn’t “done”, it’s where the most interesting, least-finished questions in cancer genomics still live. Which is the whole reason I’m here.