Where cancer genomes break, and why

1.1 · The posting, factually

1.2 · What the project is really about

Strip the advert down and one sentence remains: cancer genomes are rearranged in recurring patterns, and this project asks why those rearrangements happen in those particular places.

The conventional answer has been selection — a rearrangement persists because it helped the tumour grow. That is certainly part of the story, but it cannot be the whole of it, because some positions break far more often than chance allows even before any selection has had a chance to act. Something about those positions makes the DNA physically fragile.

The group's answer is that fragility is produced by traffic. A chromosome is read by two different machines that both travel along it: the replication machinery copying the DNA, and the transcription machinery copying genes into RNA. Where those two machines meet head-on, the DNA is at risk. Where they travel in the same direction, it is much safer. Break positions are therefore predictable from the local geometry of replication and transcription — and if they are predictable, they are also potentially druggable.

1.3 · Where the project came from

This is not a speculative proposal. It is the next step in a decade-long line of work, and knowing that line is the single most useful piece of interview preparation available to you.

What an RDC actually is

A recurrent DNA break cluster is a genomic interval, typically tens to hundreds of kilobases wide, in which independent cells repeatedly sustain double-strand breaks. They are found inside long, actively transcribed genes that replicate late.

Three features define the vulnerable locus, and all three are measurable:

• Length. The gene is so long that transcribing it once takes longer than the cell cycle allows, so transcription and replication are necessarily simultaneous.
• Late and slow replication. These regions are copied at the end of S phase and are poor in replication origins, so a small number of forks must travel a long way — and a fork that must travel far has more opportunity to meet trouble.
• Transcription that cannot be switched off. Deleting the promoter of an RDC gene abolishes the breaks. In the group's 2026 work, removing the promoter or enhancer of Ctnna2 or Nrg3 eliminated both the nascent transcription and the RDC, without advancing replication timing — which is close to a formal demonstration that the transcription, not merely the lateness, is causal.

In the 2024 paper the 152 RDCs were sorted by the geometry of the forks that converge on them: 87 were inward-moving, of which 63 overlapped a replication-timing transition region; 15 unidirectional; 6 outward-moving; 12 complex; and 32 could not be assigned. That classification is the project's intellectual unit of analysis. When the advert says “map and classify DSBs”, this is the kind of classification it means.

1.4 · Why a laboratory in Jerusalem is on the advert

The Aqeilan laboratory works on WWOX, and WWOX is the reason the collaboration makes sense. It has a double identity that sits precisely on this project's axis.

First, WWOX is the gene that occupies the common fragile site FRA16D. It is long, late-replicating and famously prone to breakage and deletion across many tumour types — in other words, it is itself one of the things that breaks.

Second, WWOX is a participant in the response to breakage. Work from that laboratory reported that WWOX regulates ATM activation and the DNA damage response, and separately that it modulates the ATR-mediated checkpoint; a broader analysis argued that tumour suppressor genes located within common fragile sites are active players in the damage response rather than passive victims of it.

Third, and this is the part that explains the tumour types named in the advert, the same laboratory reported that WWOX promotes osteosarcoma development through upregulation of Myc — a result that complicates the simple tumour-suppressor reading and makes osteosarcoma a natural model for them.

1.5 · Why osteosarcoma, specifically

Osteosarcoma is an unusually good test case, and being able to say why signals that you understand the project rather than just its methods.

Osteosarcoma genomes are among the most structurally chaotic in human cancer. They carry very high numbers of structural variants, frequent chromothripsis, breakage-fusion-bridge cycles and complex amplifications, and comparatively few simple recurrent point-mutation drivers. For a project about break positions this is ideal: there is an enormous amount of rearrangement to explain, and relatively little of it can be explained by selection on individual favourable mutations.

It is also a predominantly adolescent and young-adult cancer with little therapeutic progress in decades, which is what makes the reverse-translational framing in the advert more than a formality.

Breast cancer provides the contrast: a far better characterised genome, well-described replication-stress biology, and established damage-response drug targets. Leukaemia provides a third, mechanistically distinct comparison, because lymphoid cells generate programmed breaks as part of normal development — which is, historically, the biology the LAM-HTGTS method was invented to study.

1.6 · The two candidate profiles

The advert describes two entry routes with different requirements, and it is worth reading carefully which column you are standing in, because it determines what you have to prove.

2.1 · Level zero: the machinery that will later break

DNA is two strands wound together. Each strand runs in a direction, and the two run opposite to each other — written as 5′ to 3′ for one and 3′ to 5′ for the other. That sounds like pedantry and it is not: almost everything in this project is a question about direction, and direction only exists because the two strands are not interchangeable.

A cell that is going to divide passes through four stages, together called the cell cycle. In G1 it grows and prepares. In S phase it copies its entire DNA. In G2 it checks the work. In M phase it divides. Everything this project studies happens in S phase, or is a consequence of something that went wrong there.

How DNA is copied, and why that creates direction

The genome is far too long to copy from one end. Instead, copying starts at thousands of designated points called origins of replication. From each origin, copying proceeds in both directions at once, like a zip opening outwards from its middle. The moving point where the two strands are being separated and copied is a replication fork.

So at any moment in S phase, the genome contains thousands of forks, and each one is travelling either leftward or rightward along the chromosome coordinate. That left-or-right property is replication fork direction, and measuring it genome-wide is one of the four assays in Module 3.

Origins do not all fire at the same time. Some regions are copied in the first minutes of S phase and others at the very end — this is replication timing. Early regions tend to be gene-rich and chemically open; late regions tend to be gene-poor, tightly packed and attached to the nuclear envelope. Between an early region and a late region the timing slopes gradually across a wide interval called a timing transition region, or TTR. A TTR is typically crossed by a single fork travelling a long way with no origin to help it — which, as Figure 1 showed, is exactly where trouble concentrates.

How genes are read, and why that also creates direction

Transcription is the copying of one gene into RNA by the enzyme RNA polymerase. The polymerase uses one of the two DNA strands as its template, and which strand it uses determines which way along the chromosome it travels. A gene is therefore described as being on the plus strand or the minus strand, and that is its transcription direction.

Two directions now exist on the same piece of DNA: the direction the fork is travelling, and the direction the polymerase is travelling. The relationship between those two directions is the subject of this entire PhD.

What happens when DNA breaks, and the three ways cells fix it

A double-strand break (DSB) is both strands severed at the same place. It is the most dangerous form of DNA damage, because unlike a single-strand nick there is no intact strand left to use as a guide. Cells have three main ways of dealing with one, and which one acts determines what scar is left behind.

Non-homologous end joining (NHEJ)

The fast, always-available route: recognise the two ends, trim them, join them. Core components are Ku70/Ku80, DNA-PKcs, and the ligase complex LIG4–XRCC4–XLF. Usually accurate enough, but often gains or loses a few letters. Remember the name XRCC4 — it returns below in a way that explains the group's experimental design.

Homologous recombination (HR)

The accurate route: chew back one strand at each end, then use the intact sister copy of the chromosome as a template to rebuild what was lost. Needs MRE11–RAD50–NBS1 and CtIP for the initial chewing, RPA to coat the exposed single strand, and RAD51 to perform the search. It only works after the DNA has been copied, so it is restricted to S and G2 phase.

Microhomology-mediated end joining (MMEJ, also alt-EJ or TMEJ)

The improviser. It looks for a very short stretch of matching sequence — typically two to twenty letters — near both ends, lines those up, and joins, deleting whatever lay between. Its signature enzyme is DNA polymerase theta, encoded by POLQ. The scar it leaves is characteristic: a deletion with a few letters of microhomology exactly at the junction. The group's 2026 work used a polymerase theta inhibitor and watched junctions shift away from roughly two base pairs of microhomology towards direct joining, which is how you demonstrate that this pathway was responsible.

2.2 · Replication stress and replication–transcription conflicts

Replication stress is the slowing, stalling or collapse of replication forks. It is not one molecular event but a category, and its causes fall into four groups:

• Running out of materials — insufficient nucleotides to build new DNA, which is what happens when a cell is forced into S phase before it is ready.
• Obstacles on the track — chemical damage to the DNA, unusual secondary structures such as G-quadruplexes, tightly bound proteins, and RNA:DNA hybrids (§2.4).
• Too many origins firing, or too few — both are harmful. Excess firing exhausts the nucleotide supply and the pool of protective proteins; too few leaves long stretches dependent on a single distant fork.
• Topological strain — DNA is a twisted double helix, and any machine travelling along it generates over- and under-winding that must be released by topoisomerases.

A stalled fork exposes single-stranded DNA, which is coated by RPA, which recruits the kinase ATR, which activates CHK1. That signalling axis slows the cell cycle, suppresses further origin firing elsewhere and stabilises the fork. It is simultaneously the cell's defence against replication stress and — because cancer cells depend on it far more heavily than normal cells do — the most heavily drugged target in this field (§2.6).

If the fork cannot be rescued it may reverse, be cut, or simply be abandoned, leaving a region unreplicated as the cell enters mitosis. Unreplicated DNA pulled apart during division produces breaks, micronuclei and, over cycles, the large-scale rearrangements seen in tumour genomes.

Oncogene-induced replication stress

This is the reason the whole field exists. Activating an oncogene such as cyclin E (CCNE1) or MYC pushes cells into S phase early and fires too many origins at once. Nucleotide pools are exhausted, forks slow, and breaks accumulate — in the earliest stages of tumour development, before any treatment. The influential synthesis of this idea holds that replication stress is the engine generating the genomic instability on which tumour evolution then feeds, and that this is why loss of p53 is selected for so early and so universally: it removes the response that would otherwise eliminate the damaged cells.

Head-on versus co-directional

Now the central concept. A fork and a transcribing polymerase on the same stretch of DNA can be arranged two ways.

Co-directional: both travelling the same way. The fork catches up with the polymerase from behind. This causes delay and is not harmless, but it is survivable; the fork can displace the polymerase or wait.

Head-on: travelling towards each other. Three things make this much worse. The DNA between them accumulates positive supercoiling with nowhere to go, because both machines are pushing twist into the same shrinking interval. Any RNA:DNA hybrid in that interval is stabilised rather than displaced. And the fork is far more likely to reverse or collapse, producing a break.

Same direction is a traffic jam. Opposite directions is a head-on crash on a single-track line.

The group's 2024 paper turned this into a measurement. Because the break-mapping assay preserves the orientation of each captured broken end, each break can be labelled as pointing towards the centromere or towards the telomere — written Dcen and Dtel. A rightward-moving fork generates centromere-oriented ends; a leftward-moving fork generates telomere-oriented ends. Comparing break orientation against locally measured fork direction and gene direction therefore classifies each break as head-on or co-directional, and the proportions are reported as PHO and PCD. Fifty per cent means no preference. Under replication stress induced with aphidicolin, the reported head-on proportion rose to roughly sixty per cent while the co-directional proportion fell to roughly forty.

2.3 · Breaks, hotspots and translocations

A break hotspot is a genomic interval where breaks are recovered more often than the local background allows. The phrase hides a statistical decision: “more often than background” requires a background model, and the choice of that model determines which hotspots exist. The group's analyses smooth break density over intervals of roughly 25 to 50 kilobases, which is the scale at which this biology operates — a conflict zone is tens of kilobases wide, not tens of base pairs.

A translocation is the joining of two broken ends that came from different places in the genome. It requires three things: a break at locus A, a break at locus B, and the two loci being physically near each other inside the nucleus at that moment. That third requirement is why translocation frequency reports on nuclear architecture as well as on breakage, and why the historical name for this assay family refers to proximity-based break joining.

Not all breaks are accidents. Developing lymphocytes make them on purpose, cutting and rejoining their own antigen-receptor genes using the RAG1/RAG2 enzymes and, later, the enzyme AID. This is the biology the break-mapping method was originally built to study, which is worth knowing because it explains the leukaemia arm of this project: in lymphoid cells you have both programmed and accidental breakage in the same genome, and distinguishing them is a genuine analytical problem.

Fragile sites

Common fragile sites are chromosomal regions that break reproducibly when cells are treated with low-dose aphidicolin. Two are canonical: FRA3B, which contains the gene FHIT, and FRA16D, which contains WWOX — the gene that brings the Jerusalem collaboration into this project. Their properties will now sound familiar: very large transcribed genes, late replication, few origins, and sequences with a tendency to adopt awkward secondary structures. They are the human-cancer counterpart of the recurrent break clusters the group mapped in mouse neural progenitors, which is precisely why the framework should transfer.

A separate class, early replicating fragile sites, breaks under the opposite conditions — in early S phase, at highly transcribed gene-dense regions — which is a useful reminder that late replication is not a universal requirement for fragility.

2.4 · R-loops

An R-loop is a three-stranded structure. Normally, after RNA polymerase has read the DNA, the two DNA strands close behind it and the new RNA leaves. Sometimes the RNA instead folds back and re-anneals to its template strand, forming an RNA:DNA hybrid and leaving the other DNA strand hanging as exposed single-stranded DNA.

They are promoted by G-rich sequence on the displaced strand, by the under-winding left behind a polymerase, by high transcription, and by anything that stops the new RNA being packaged and exported promptly — defective splicing or RNA-processing machinery is a classic cause. They are removed by the enzymes RNase H1 and H2, which digest the RNA within a hybrid, and by helicases such as senataxin that unwind it.

R-loops are not simply damage. They have normal functions at promoters and transcription terminators, in antibody class switching, and at mitochondrial replication origins. The problem is when they persist in the wrong place — above all in the interval between a converging fork and an oncoming polymerase, where they are stabilised rather than displaced and become a physical block that turns a stalled fork into a broken one.

Measurement is by immunoprecipitation with the S9.6 antibody, which recognises RNA:DNA hybrids — the DRIP-seq assay of Module 3.

2.5 · Copy-number alterations and structural variants

A structural variant is any rearrangement of the genome larger than about fifty base pairs. A copy-number alteration is the subset of those that changes how many copies of a region are present. The distinction matters because it maps onto what breaks can do: an inversion or a balanced translocation moves DNA without changing its amount, while a deletion or duplication changes the amount.

Two methods from the group's 2026 paper are worth knowing by name because they answer questions bulk sequencing cannot. Strand-seq is single-cell sequencing that keeps track of which parental strand each read came from, which makes inversions and individual breakpoints directly visible one cell at a time. In that work it showed that 118 of 302 breakpoints in cells with reduced copy number mapped to known break clusters — a per-cell link between the break map and the resulting structural change. And the under-replication index quantifies how much of a region failed to finish replicating, which is the measurement that connects replication stress to the downstream rearrangement.

For this project's tumour types the relevant fact is that osteosarcoma genomes are dominated by exactly the complex classes in the lower half of that table — chromothripsis, breakage–fusion–bridge, extrachromosomal amplification — rather than by simple recurrent point mutations.

2.6 · Reverse-translational targeting of the damage response

Classical translational research runs from laboratory to clinic. Reverse translation runs the other way: it starts from an observation in human patients or patient material, goes back into the laboratory to find the mechanism, and uses that mechanism to propose a treatment. The advert's phrasing — help develop reverse-translational therapy strategies for osteosarcoma and breast cancer — means exactly this. The observation is that these tumour genomes are rearranged in particular patterns; the laboratory work is to find out what produces those patterns; the therapeutic proposal follows from the mechanism.

The logic that makes it a therapy is synthetic lethality: two defects that are each survivable alone but lethal together. If a cancer already depends heavily on one route for coping with replication stress, blocking that route kills it while leaving normal cells, which were not relying on it, comparatively unharmed.

2.7 · The whole vocabulary, collected

3.1 · Words for this module

3.2 · LAM-HTGTS — where the genome breaks, with direction

The measurement problem is severe. A double-strand break is transient, it occurs at an unknown position, and it occurs in a small fraction of cells. You cannot design a primer against a position you do not know.

The solution is indirect and rather elegant. You make one break yourself, at a position you choose — the bait. The cell's repair machinery will occasionally join that bait end to some other broken end elsewhere in the genome — a prey. Because you know the bait's sequence exactly, you can anchor on it and read outward into whatever got attached. Each such molecule reports the position of one genuine break, discovered without knowing in advance where to look.

Why linear amplification

Ordinary PCR needs two primers, one on each side of the target. Here only one side is known. So the first amplification uses a single primer that sits in the bait and carries a biotin tag, and it is run for many cycles of simple extension: each cycle copies once from the genomic template, running from the bait across the junction and out into the unknown prey. The products are single-stranded, each carries biotin at its known end, and they can therefore be captured on streptavidin beads while everything else is washed away. Only then is an adapter ligated to the unknown end, giving the second primer a landing site, and a nested PCR produces a sequenceable library.

You cannot amplify between two points when you only know one of them. So you copy outward from the one you know, grab those copies by their tag, and only afterwards attach a handle to the far end.

1. Induce the bait break A sequence-specific nuclease — Cas9 with a defined guide, or the meganuclease I-SceI at an engineered site — cuts at the chosen position. The group's 2026 work placed baits on chromosomes 6, 8 and 12.
2. Harvest and fragment the genomic DNA Sonication is used, and the randomness is deliberate: because every molecule breaks at a different place, two sequencing reads sharing both the same junction and the same fragment end must be copies of one original molecule, whereas the same junction with different fragment ends represents two independent events. This is the basis of honest counting.
3. Linear amplification from the bait Many cycles of extension with the single biotinylated bait primer.
4. Capture on streptavidin beads Everything without the tag is washed away.
5. Ligate an adapter to the unknown end A single-stranded adapter is joined to the free prey-side end.
6. Remove uncut bait Most copies of the bait locus in the cell population were never cut, or were cut and perfectly rejoined, and they would otherwise flood the library. A blocking digestion removes them. Skipping or under-performing this step is the single most common reason a library returns almost nothing useful.
7. Nested PCR and sequencing A bait-internal primer plus an adapter primer, then paired-end sequencing.
8. Call junctions For each read, find where bait sequence ends and prey begins, align the prey portion, and record its chromosome, position, orientation, fragment end, and the sequence present at the junction itself — which is where any microhomology shows up.

The orientation is the point of the whole exercise. Because the assay reads from a fixed anchor outward, the strand and direction of the prey sequence are preserved, so each recovered end can be classified as centromere-oriented or telomere-oriented. That is what makes the head-on versus co-directional analysis possible at all, and it is why the advert describes the method as direction-resolved.

3.3 · Hi-RepliSeq — when each region is copied, and which way the fork runs

The principle is a stopwatch built from a cell sorter. Newly made DNA is tagged as it is synthesised; cells are then separated according to how far through S phase they have progressed; and the tagged DNA from each group is sequenced. Regions appearing in the earliest group replicate early; regions appearing only in the last group replicate late.

1. Pulse-label new DNA Cells are given BrdU, a modified building block, for a short window. Only DNA made during that window carries the tag.
2. Sort cells by position in S phase Cells are fixed, their total DNA content is stained and measured, and they are sorted by that content — a cell early in S has barely more DNA than a G1 cell, one late in S has nearly twice as much. High-resolution Repli-seq takes sixteen fractions, S1 to S16, rather than the two of the original method, and that resolution is what makes the next steps possible.
3. Pull down the tagged DNA from each fraction An anti-BrdU antibody recovers only newly synthesised DNA, separately for each of the sixteen fractions.
4. Sequence each fraction and build the matrix Sixteen libraries are mapped and counted in genomic bins, giving a matrix of sixteen rows by however many bins the genome is divided into. Displayed as a heatmap, it is the recognisable Repli-seq picture.
5. Extract the features From that matrix: replication timing per bin; initiation zones, where signal first appears in a narrow region and then spreads outward in both directions; termination zones, where two spreading regions converge and replication finishes last; and timing transition regions, the broad slopes between early and late domains.
6. Derive fork direction The DNA between an initiation zone and the nearest termination zone must have been copied by a fork travelling from the former towards the latter. Connecting each initiation zone to its neighbouring termination zones therefore assigns a direction to each interval. The group used a convolutional neural network to assist this calling, and accepted a direction only where both replication-stressed technical replicates agreed.

3.4 · GRO-seq — transcription as it is happening, with a direction

Ordinary RNA sequencing measures the RNA that has accumulated in a cell. For this project that is the wrong quantity twice over. A stable RNA can still be abundant hours after its transcription stopped, so abundance does not report current activity; and many library preparations discard strand information entirely, so the direction — the thing the whole project turns on — is lost.

GRO-seq measures only polymerases that are engaged on DNA at the moment the cells are harvested, and keeps the strand.

1. Isolate nuclei The cell membrane is removed but the nucleus is left intact, with polymerases still sitting on the DNA where they were.
2. Run-on with the on-ramps closed Sarkosyl is added, which prevents any new polymerase from starting and releases paused ones into elongation. Then a modified nucleotide, Br-UTP, is supplied along with the ordinary ones. Every polymerase that was already engaged extends its transcript and incorporates the tag; nothing that was not already engaged contributes at all.
3. Fragment and immunoprecipitate The RNA is fragmented and the tagged portions are pulled down with an anti-BrdU antibody, which recognises the incorporated Br-U.
4. Strand-specific library and sequencing The library must preserve which strand the RNA came from. A non-stranded protocol here destroys the experiment's entire purpose.

The variants worth knowing by name: PRO-seq resolves the polymerase position to roughly a single nucleotide rather than a fragment; TT-seq labels nascent RNA metabolically in living cells without isolating nuclei; NET-seq sequences the RNA still held within the polymerase.

3.5 · DRIP-seq — where RNA is annealed to DNA

The structure being measured is an RNA:DNA hybrid, and the tool is the S9.6 antibody, which binds that structure. The assay is therefore conceptually the simplest of the four: fragment the DNA, pull down with S9.6, sequence what came along.

1. Extract genomic DNA gently Harsh handling destroys the structures you are trying to measure.
2. Fragment Classically a cocktail of restriction enzymes, which cut at fixed positions and so limit resolution to the fragment size; sonication gives finer resolution but risks disrupting hybrids. The choice is a genuine trade-off and should be stated in any methods section.
3. Split the sample and treat one half with RNase H This is the control the assay lives or dies by. RNase H digests the RNA within a hybrid specifically. Signal that disappears was a hybrid; signal that survives was something else.
4. Immunoprecipitate with S9.6, build libraries, sequence The group normalised to fifty million reads per sample for comparability.
5. Call peaks against the RNase H-treated sample Not against input alone.

For strand information there is DRIPc-seq, which converts the recovered RNA to cDNA and so reports which strand the hybridised RNA came from — necessary if you want to ask whether the hybrid lies in the path of an oncoming fork. The group reported that roughly one per cent of the genome carried coding-strand-specific hybrids. Alternative approaches worth naming are R-ChIP, which uses a catalytically dead RNase H1 as a probe inside living cells, and MapR, which fuses a nuclease to RNase H1 and needs no antibody at all.

3.6 · The four side by side

3.7 · The sequence-processing chain

All four assays end as sequencing reads, and three of them then follow a largely common path. LAM-HTGTS is the exception: because its information is a junction within a read rather than a read position, it needs a dedicated junction-calling step rather than a standard alignment-and-coverage workflow.

4.1 · The two deadlines

The application closes on 28 October 2026. Everything before that date has one purpose: to put something in the application that a group leader cannot get from a transcript. Everything after it has a different purpose: to be able to talk about replication stress with someone who has spent ten years on it.

Confusing these two jobs is the standard failure. A curriculum that starts with the cell cycle and reaches multi-omics integration in month four is the right curriculum and the wrong schedule, because by month four the decision has been made. So the sprint comes first, and it is deliberately narrow.

4.2 · The twelve weeks after submission

If the application succeeds you will be asked to discuss this science with people who know it well, probably six to twelve weeks later. The phases below are the curriculum; the hours are realistic for someone working alongside other commitments.

5.1 · File formats, and the coordinate problem

Almost every error made by a beginner in this field is a coordinate error, and the reason is that two conventions coexist and neither announces itself.

BED files count from zero and exclude the end. A feature written as chr1 100 110 covers the eleventh through twentieth bases, and its length is simply 110 − 100 = 10.

GTF, GFF, VCF and SAM count from one and include the end. The same ten bases are written as chr1 101 110, and the length is 110 − 101 + 1 = 10.

One format starts counting at zero, the other at one. Mixing them shifts everything by a single base — which is invisible on a plot, survives every sanity check, and quietly corrupts any analysis that cares about exact positions. For junction-level work, it corrupts the result.

5.2 · A worked integration

The task: given break hotspots, a replication fork-direction track, a stranded transcription track and gene annotation, classify the genome into head-on and co-directional compartments, count breaks in each, and test whether the difference is more than chance allows.

Step 1 — prepare the inputs on the command line

# Everything sorted the same way, or bedtools will lie to you quietly.
sort -k1,1 -k2,2n dsb_junctions.bed > dsb.sorted.bed

# Transcription start sites out of a GTF.
# GTF is 1-based inclusive; BED is 0-based exclusive -> subtract 1 from the start.
# For a minus-strand gene the TSS is at the END coordinate, not the start.
awk 'BEGIN{FS=OFS="\t"} $3=="gene" {
        tss = ($7=="+") ? $4-1 : $5-1
        match($0, /gene_name "[^"]+"/)
        name = substr($0, RSTART+11, RLENGTH-12)
        print $1, tss, tss+1, name, ".", $7
     }' gencode.gtf \
  | sort -k1,1 -k2,2n > tss.bed

# Distance from each break to the nearest TSS (-d appends the distance).
bedtools closest -a dsb.sorted.bed -b tss.bed -d > dsb_to_tss.tsv

# Which breaks sit in late-replicating domains.
bedtools intersect -a dsb.sorted.bed -b rt_late_domains.bed -u > dsb_late.bed

# A background set matched for chromosome and restricted to mappable sequence.
# -chrom keeps each region on its own chromosome; -incl confines it to mappable space.
bedtools shuffle -i dsb.sorted.bed -g hg38.chrom.sizes \
                 -incl mappable_k50.bed -chrom -seed 1 > dsb_shuffled.bed

Step 2 — classify the compartments in R

This is the heart of the project, and it is about thirty lines. The logic is a single truth table: a gene on the plus strand is transcribed towards higher coordinates, so a rightward fork is co-directional with it and a leftward fork is head-on. For a minus-strand gene, the reverse.

library(GenomicRanges)
library(rtracklayer)
library(regioneR)

## ---- inputs -------------------------------------------------------------
dsb   <- import("dsb.sorted.bed")             # break positions
genes <- import("gencode.gtf")
genes <- genes[genes$type == "gene" & width(genes) > 1e5]   # long genes only

fork  <- import("fork_direction.bedGraph")     # score > 0 = rightward
fork$dir <- ifelse(fork$score > 0, "R", "L")

gro   <- import("gro_seq_plus.bigWig")         # measured, not annotated

## ---- build conflict compartments ---------------------------------------
## Intersect gene bodies with fork-direction segments, keeping both labels.
hits <- findOverlaps(genes, fork)

comp <- pintersect(genes[queryHits(hits)], fork[subjectHits(hits)])
comp$tx_dir   <- as.character(strand(genes))[queryHits(hits)]
comp$fork_dir <- fork$dir[subjectHits(hits)]

## The truth table, in one expression.
comp$conflict <- ifelse(
  (comp$tx_dir == "+" & comp$fork_dir == "L") |
  (comp$tx_dir == "-" & comp$fork_dir == "R"),
  "head_on", "co_directional"
)

## ---- count breaks per compartment class --------------------------------
ov  <- findOverlaps(dsb, comp)
tab <- table(comp$conflict[subjectHits(ov)])

p_ho <- tab[["head_on"]] / sum(tab)
p_cd <- tab[["co_directional"]] / sum(tab)

message(sprintf("P_HO = %.3f   P_CD = %.3f", p_ho, p_cd))
## 0.50 / 0.50 would mean geometry carries no information.

Step 3 — the same thing in Python

import pyranges as pr
import numpy as np

dsb   = pr.read_bed("dsb.sorted.bed")
genes = pr.read_gtf("gencode.gtf")
genes = genes[(genes.Feature == "gene")]
genes = genes[(genes.End - genes.Start) > 100_000]

fork  = pr.read_bed("fork_direction.bed")      # Score > 0 = rightward

# join carries both sets of columns into one frame
comp = genes.join(fork, suffix="_fork")
df   = comp.df

df["fork_dir"] = np.where(df["Score_fork"] > 0, "R", "L")
df["conflict"] = np.where(
    ((df["Strand"] == "+") & (df["fork_dir"] == "L")) |
    ((df["Strand"] == "-") & (df["fork_dir"] == "R")),
    "head_on", "co_directional",
)

comp = pr.PyRanges(df)
counts = dsb.join(comp).df["conflict"].value_counts()
p_ho = counts["head_on"] / counts.sum()
print(f"P_HO = {p_ho:.3f}")

Step 4 — the statistic, with a null model that survives scrutiny

Here is where most analyses fail, and it is worth understanding exactly why rather than copying the code.

Suppose you compare your break hotspots against late-replicating domains with a simple two-by-two test and obtain an overwhelming p-value. That number is almost certainly meaningless, because your break hotspots are not a random sample of genomic positions. They are long, they sit in genes, those genes are expressed, expressed long genes replicate late, and all four properties are correlated with one another. A test that ignores this will declare enrichment for whichever of the correlated properties you happen to ask about.

The fix is a null model built by resampling from a universe of regions that share the properties you are not interested in, so that only the property you are testing is allowed to vary.

## Build a universe: mappable, gene-body, length-matched, timing-matched bins.
universe <- import("candidate_regions_matched.bed")

pt <- permTest(
  A                   = dsb,
  B                   = late_domains,
  universe            = universe,
  randomize.function  = resampleRegions,   # draw FROM the universe,
                                           # do not scatter at random
  evaluate.function   = numOverlaps,
  ntimes              = 1000,
  genome              = "hg38",
  mc.set.seed         = FALSE
)

plot(pt)        # observed value against the permuted distribution
pt$numOverlaps$pval

The distinction between randomizeRegions and resampleRegions is the whole point. The first scatters your regions anywhere in the genome, which tests a question nobody asked — your regions are obviously unlike random genomic positions. The second draws them from a pool of plausible alternatives, which tests whether late replication specifically distinguishes the real breaks from comparable regions that did not break.

6.1 · What the conversation will be

Prepare for the first and the second comes largely free, because the hard part of both is being able to say something substantive about a specific piece of science. Below, the questions.

6.2 · Five conceptual questions

6.3 · Three analytical scenarios

6.4 · The three-minute account of your own work

Whichever format applies, you will be asked what you have done. The structure that works is fixed and short:

1. The question, in one sentence Not the field, not the background — the actual question your work addressed.
2. Why it mattered One sentence. What was not known, or what could not be done before.
3. What you did Your methods, in the first person singular where that is accurate. If the work was computational, say what the input and output were.
4. What you found The result, including its size or direction. A number is better than an adjective.
5. What was wrong with it Do not skip this. One genuine limitation, and what you would do differently. Candidates who present flawless work are either not believed or not trusted to notice problems. A stated weakness is the strongest part of the account.
6. The bridge One sentence connecting it to this project. For a thesis on computational drug discovery, the honest bridge is the final chapter of this project — the move from a mechanism to a druggable dependency — not a forced claim of methodological overlap.

6.5 · Questions worth asking them

The questions you ask are assessed, and generic ones are a missed opportunity. Each of these demonstrates something specific about what you have understood.

• “Which bait positions are you planning to use in the human lines, and how are you handling the proximity bias when comparing cancer types?” — shows you understand the assay's central limitation and that comparisons require matched baits.
• “Cancer lines already carry their own repair defects. Does that help break capture, or does it make the backgrounds non-comparable?” — shows you have thought about why the mouse work used an engineered repair-deficient background, and that transferring the approach is not trivial.
• “Is the human cancer-line resource from the preprint the starting point, or would this project generate its own break maps?” — shows you have read beyond the published papers, and asks a real question about what year one looks like.
• “Since replication timing is itself remodelled in cancer, would timing be measured in each line rather than taken from a reference?” — a genuinely sharp methodological question.
• “How is the work split between bench and analysis, and would I be doing both?” — practical, legitimate, and it matters to you.
• “What would make you consider the project successful after three years?” — tells you how they think, and whether the project is scoped.

6.6 · The question you cannot avoid

Somebody will ask why a pharmacist is applying for a genome-instability PhD. There are three ways to answer and only one of them works.

Do not minimise it. “Pharmacy is really quite close to molecular biology” is not true, and claiming it invites a question you cannot answer. Do not over-apologise either — a candidate who treats their own degree as a liability has already conceded the point.

The answer that works is a plain sequence of facts with a direction. You came from drug development, including a thesis on computational approaches to discovering drugs. You became interested in where the targets come from rather than how they are screened, which led you to DNA-damage-response pharmacology, which led you to this project — where the endpoint is a damage-response dependency specific to a tumour type. You taught yourself the genomics you were missing, and here is the analysis you built, which you can talk about in detail. The gap you still have is wet-lab sequencing library experience, and you would expect the first year to be largely about closing it.