SCEDetector

SCEDetector screens single-cell Strand-seq data for candidate sister chromatid exchanges (SCEs), inversions, and recurrent translocation-like breakpoints. It works from fixed 200 kb genomic bins labeled with Watson/Crick strand states (CC, WC, WW) and uses read depth plus structural-variant (SV) calls to remove common artifacts.

Important. Every call is a screening candidate, not a definitive biological event. The label Translocation only means that the same breakpoint coordinate recurs across cells; the tool does not identify translocation partners or prove that a rearrangement occurred.

---

Biological background (read this first)

What Strand-seq measures

In Strand-seq, DNA is labeled so that, after replication, each sister chromatid can be read as mostly Watson (W) or mostly Crick (C) template. For every cell and every 200 kb bin the pipeline reports:

Class	Meaning in a diploid autosome
WW	Both homologs / both sisters read as Watson-dominated (homozygous Watson)
CC	Both read as Crick-dominated (homozygous Crick)
WC	One Watson-like and one Crick-like contribution (heterozygous / mixed)

The raw file also carries per-bin depths c and w (Crick and Watson read counts). Those depths are essential: the same letter pattern can be a true exchange, a duplication, a deletion, or noise next to the centromere.

What an SCE looks like

A sister chromatid exchange swaps template strands between sisters. In Strand-seq this appears as a single, clean change of state along a chromosome arm, for example:

WW WWW WWW | WC WC WC WC     →  breakpoint = start of the WC run
CC CC | WC WC | WW WW WW     →  not a single SCE (two switches; handled separately)

Valid single-switch SCE transitions used here:

From	To
WC	WW or CC
WW	WC
CC	WC

Direct WW ↔ CC without a WC middle is not treated as a normal autosomal SCE (except on male chrX, which is hemizygous: only WW ↔ CC is allowed, and WC is discarded).

Why inversions and “two SCEs” look similar

A heterozygous inversion can flip the strand state of one interval and then restore the original flank state, producing an A–B–A sandwich:

WW ── WC ── WW     or     CC ── WC ── CC

The same pattern can also be two genuine SCEs a few megabases apart. SCEDetector cannot always tell them apart from one cell alone, so it uses:

1. Recurrence across cells at both breakpoints, and/or
2. A single inversion SV call that covers most of the sandwich

before labeling Inversion; otherwise it reports two SCE rows.

Why many filters exist

Strand-seq chromosomes are full of regions that look like state switches but are not SCEs:

Artifact	Why it fools a naïve switch caller
Centromeres / heterochromatin	Sparse or None bins; state stubs at edges
Deletions	Homozygous-looking dips inside WC
Duplications	Extra WC-like signal with elevated total depth
Inversions / complex SVs	Large masked holes; tiny state scraps on either side
Acrocentric p-arms	Masking leaves a short homozygous stub before a long WC

The pipeline therefore masks unreliable bins, drops depth/SV-supported artifacts, then calls switches on what remains.

---

Pipeline at a glance

For each QC-passed cell × chromosome (chrY is used only to infer sex):

[1] Bin masking
      Drop None / centromere / SV / invalid classes
        ↓
[2] Artifact removal (depth + SV geometry)
      (a)  Duplication-like WC
      (b)  Short stubs at SV hole edges
      (b′) Sparse remnants after large SV holes
      (c)  Lopsided A–WC–A WC islands
      (c′) Short A–WC–A WC scraps beside large SV-tiled gaps
      (d)  Short / deletion-backed / long-shallow WC–A–WC homozygous islands
      (e)  Shallow homozygous tips (deletion-like)
      (f)  Short leading homozygous stubs on acrocentrics
        ↓
[3] Double-switch extraction
      A–B–A → Inversion or two SCEs
      WW→WC→CC / CC→WC→WW → always two SCEs
        ↓
[4] Single-switch calling
      Exactly one valid switch per arm → SCE
      (+ centromere-crossing exception)
        ↓
[5] Breakpoint refinement
      Move stitched breakpoints to the true switch inside an SV hole
        ↓
[6] Recurrence relabeling
      Shared SCE-like breakpoints across cells → Translocation

---

Requirements

• Python 3.9+
• pandas, openpyxl

pip install -r requirements.txt

Usage

python3 SCEDetector.py \
  -i fastq0022_HPNE_M.txt.raw.gz \
  --sv lenient_filterFALSE.tsv \
  --qc StrandPhaseR_final_output.txt \
  -o SCE_detected.xlsx

---

Inputs

Argument	File	Role
-i	*.txt.raw.gz	200 kb bins: chrom, start, end, sample, cell, class, c, w
--sv	lenient_filterFALSE.tsv	Per-cell SV intervals (deletion, duplication, inversion, complex, …)
--qc	StrandPhaseR_final_output.txt	QC-passed cells only these are analyzed
--mask	HGSVC.200000.txt (default)	Low-mappability / None-bin mask
--species	human → chromosome_arm_positions_grch38.txt	Centromere barriers [p.End, q.Start)

Sex. Inferred from chrY read depth. On male chrX (hemizygous), WC bins are removed and only WW ↔ CC transitions are allowed. Deletion SVs on male chrX are not masked (hemizygous deletions would otherwise erase the only informative state).

---

Algorithm (step by step)

Step 1 — Bin masking

Each 200 kb bin is dropped if any of the following applies.

1.1 None / low-mappability mask (HGSVC.200000.txt)

Rule	Detail
Explicit None	Always dropped
Short good islands	≤5 bins of good fully flanked by None → treated as None (mappability speckles)
Edge absorption	A large None run (≥10 bins) also absorbs up to 5 flanking good bins on each side so heterochromatin edges do not leave tiny state stubs

1.2 Centromere barrier

Bins overlapping [p.End, q.Start) from the arm table are dropped.
If a chromosome has no arm-table entry, contiguous None stretches ≥1 Mb (merged across gaps ≤1 Mb) are used as a fallback barrier.

1.3 SV intervals

SV class	Masked?
Deletion / duplication / inverted duplication (del_*, dup_*, idup_*)	Always (except male-chrX deletions — kept)
Inversion / complex	Masked by default; kept (not masked) only if tip-linked

Tip-linked means any of:

• the SV reaches a chromosome tip within ±1 Mb, or
• it starts within 2 Mb of pter / ends within 2 Mb of qter, or
• it abuts another tip-linked SV (including tip-linked CN events).

Why tip-linked inversions are kept. A tip inversion often is the biological event that should remain visible as a state change. Masking it would erase the signal. Interior inversions / complex events are usually masked so their messy interiors do not invent extra SCE breakpoints; their effect is reconsidered later when resolving A–B–A sandwiches.

1.4 Invalid strand class

Keep only CC / WC / WW (male chrX: only CC / WW).

Surviving bins are merged into state runs. Example:

WW WW WW WC WC CC CC  →  WW | WC | CC

---

Step 2 — Artifact removal

After masking, some runs still look like SCE switches but are better explained by copy-number change or mask geometry. Depth terms used below:

Symbol	Meaning
c, w	Median Crick / Watson depth on the full contiguous raw WC run (including bins the None mask would skip)
dom	Dominant-strand depth of flanking homozygous run(s): WW → w, CC → c. Average of both sides when present; floored by the cell-wide homozygous median when flanks are missing or too shallow. Flanks may be taken across an SV/None gap when nothing abuts the WC edge.
balance	min(c,w) / max(c,w) — near 1 means strands are even
Expected coverage	Per-bin across-cell median of relative depth × this cell’s genome-wide median. Stops mappability-hot bands (deep in every cell) from looking like gains

---

2a. Drop duplication-like WC

Biology. An unmarked (or incompletely called) duplication often keeps one strand near full homozygous depth and adds signal on the other strand. True copy-neutral WC should sit near half depth on each strand and ~1× the homozygous flank in total.

Drop if either path fires.

Path A — flank asymmetry (all three required):

1. max(c,w) / dom ≥ 0.85 — one strand still near homozygous depth
2. min(c,w) / dom ≥ 0.35 — the other strand is clearly present
3. (c+w) / dom ≥ 1.35 — total depth is elevated vs the flank

Path B — coverage ceiling (alone enough):

• (c+w) > expected coverage × 1.50

	True WC (keep)	Duplication-like WC (drop)
Strands	both ≈ 0.5 × dom	one ≈ dom, other ≈ 0.5 × dom
Total	≈ dom	≈ 1.5 × dom

---

2b. Drop short stubs at SV hole edges

Biology. After a large SV is removed, 1–2 bins of a foreign state often cling to the hole edge and create a false A–B–A sandwich.

Rule. Any run of ≤2 bins (≤400 kb) that touches an SV skip interval is dropped so the true flanks can merge.

---

2b′. Drop sparse remnants after large SV holes

Biology. _merge_runs reports genomic start/end across skipped bins. A scrap on either side of a del/dup/inv block can therefore look megabases long even though almost no kept sequence remains — enough to seed a false SCE.

Drop a WC or homozygous run when all hold:

1. The run touches an SV skip interval
2. Kept bin length ≤ 2 Mb (WC) or ≤ 0.8 Mb (WW/CC)
3. Genomic hole (span − kept) is strictly larger than kept length

Solid short flanks with kept ≈ span are retained.

How 2b / 2b′ / 2c′ differ (easy to confuse — they all touch SV holes):

	2b	2b′	2c′
Target	Any class	WC or WW/CC run	A–WC–A only (WW-WC-WW / CC-WC-CC)
Length idea	≤ 2 bins (0.4 Mb) — shortest of the three	Small kept vs large hole (WC ≤2 Mb / homo ≤0.8 Mb)	WC ≤ 1 Mb (can be longer than a 2b stub)
Geometry	Touches SV	Touches SV and hole > kept	Flank gap ≥ 10 Mb almost fully SV-tiled
Role	Dust on the hole edge	Run that looks long only because it spans a hole	Fake dual-SCE from a short WC island after a huge SV desert

So “short” in 2c′ means short relative to a real ABA WC island / dual-SCE spacing, not shorter than 2b. A 0.6–1.0 Mb WC between identical flanks across an ~18 Mb SV gap is a 2c′ case that 2b misses (>2 bins).

---

2c. Drop lopsided A–WC–A WC islands

Biology. A WC island between two identical homozygous runs (WW–WC–WW / CC–WC–CC) is the classic “two SCE or one inversion” pattern. If the island is really a homozygous stretch with background on the minor strand (lopsided, near-flank total depth), treating it as two exchanges is wrong.

Keep the island if any escape rule holds; otherwise drop it (flanks merge). Single WC transitions (not sandwiched) are never touched here.

#	Keep when	Intuition
1	balance ≥ 0.75	Strands are even enough for real WC
2	length ≥ 10 Mb and balance ≥ 0.50	Long islands get a relaxed balance bar
3	(c+w) < 0.55 × dom	Depth-depleted islands are a different artifact — handled elsewhere
4	Flank-dominant strand inside the island is ≤ 0.47 × the deeper flank and balance ≥ 0.35	Clean halving (copy-neutral exchange)
5	Both flanks exist, length ≤ 5 Mb, balance ≥ 0.55, island total lies between the two flanks, dominant strand still > 0.55 × deeper flank, minor strand ≥ 0.45 × shallower flank	True exchange next to uneven flanks

---

2c′. Drop short A–WC–A WC scraps beside large SV-tiled gaps

Biology. After a multi-megabase SV/complex mask, a short WC remnant (≤1 Mb) can remain between identical homozygous flanks and invent a nonsense dual SCE a few hundred kb apart. Depth-depleted islands escape rule 2c (keep when (c+w) < 0.55 × dom), so a separate ABA + gap + SV-tiling rule is needed. This is not “shorter than 2b”: 2b already removes ≤2-bin edge stubs; 2c′ targets short WC islands next to a huge SV-tiled hole (including 3–5 bin scraps that 2b leaves).

Drop WW–WC–WW / CC–WC–CC when all hold:

1. WC length ≤ 1 Mb
2. At least one flank gap ≥ 10 Mb
3. That gap is almost fully tiled by SV skip (uncovered ≤ 1 Mb)

Depth-carved holes without SV tiling (for example inversion sandwiches where WC was trimmed by other filters) are kept, so true ABA inversion calls are not destroyed. Overlap with 2b is possible when the WC scrap is also ≤2 bins; 2b runs earlier and usually clears those first.

---

2d. Drop short / deletion-backed WC–A–WC homozygous islands

Biology. Mirror of 2c: a short WW/CC island inside WC (WC–CC–WC / WC–WW–WC) looks like two SCEs. Two exchanges only a few Mb apart are rare; a real island would stay roughly depth-neutral. A shallow island overlapping a deletion call is usually lost sequence, not two exchanges.

Short island — drop when all hold:

1. Island length < 5 Mb
2. Island (c+w) < deeper flanking WC × 0.80
3. Island (c+w) < shallower flanking WC (depleted vs both sides)

Deletion-backed island (no length cap) — drop when either holds:

1. Deletion SV covers ≥ 20% of the island and island (c+w) < deeper WC × 0.80
2. Deletion SV covers ≥ 15% and island (c+w) < deeper WC × 0.40 (already deeply depleted)

Long shallow island without deletion support — drop when all hold:

1. Island length ≥ 5 Mb
2. Deletion SV covers < 15% of the island (weak/no del call; otherwise the deletion-backed path applies)
3. Island (c+w) < deeper WC × 0.40
4. Island (c+w) < shallower WC

This closes the gap where a long unmarked deletion-like island is too long for the short path and has no del_* call for the deletion-backed path. The depth bar matches the soft deletion-backed tier so only deeply depleted islands are removed.

Depth-neutral islands, or islands deeper than one WC flank, are kept.

---

2e. Drop shallow homozygous tips (deletion-like)

Biology. A terminal WW/CC tip next to inward WC can be a real tip SCE, or a deletion/noise stub that seeds a false WW→WC→CC two-step. Male chrX is skipped.

A. Deletion-backed (exactly two runs on the chromosome) — drop when all hold:

1. Tip abuts the WC flank (no masked gap)
2. Tip (c+w) < WC flank × 0.70
3. Deletion SV calls cover ≥ 10% of the tip in total
4. At least one deletion lies within 2 Mb of the WC↔homo junction

B. Short shallow tip (any run count) — drop when both hold:

1. Tip length ≤ 15 Mb
2. Tip (c+w) ≤ WC flank × 0.70

Path B catches compact terminal homozygous blocks even when no deletion call sits on the tip.

C. Long shallow tip with heavy deletion cover (any run count) — suppress tip-junction SCE when all hold:

1. Tip length > 15 Mb
2. Tip abuts the WC flank
3. Tip (c+w) ≤ WC flank × 0.50
4. Deletion SV calls cover ≥ 50% of the tip

Path C does not remove the tip run (that would collapse multi-switch chromosomes and rewrite unrelated proximal calls). It only suppresses the WC↔homo tip SCE. Targets arm-scale hemizygous tips where the first del_* starts far past the junction (e.g. T3 A10 chr4 @142.2 Mb).

---

2f. Drop short leading homozygous stubs on acrocentrics

Biology. On acrocentric chromosomes (chr13/14/15/21/22), p-arm and centromere masking often leaves a few bins of WW/CC abutting a long WC on the q-arm. That stub looks like a tip SCE but is almost always mask-edge noise.

Rule. Leading WW/CC ≤ 1 Mb that abuts a following WC → drop the stub.
Stubs separated from WC by a masked gap are kept.

---

Step 3 — Double-switch extraction and resolution

Double switches are removed from the filtered run list and resolved on their own.

3.1 A–B–A sandwich (WW–WC–WW, CC–WC–CC, …)

Return to the original flank class → either one inversion or two SCEs.

Priority	Condition	Result
1	Both breakpoints shared within ±10 kb by ≥5% of QC cells (same sample·chrom; ≥2 cells)	Inversion (+ shared %)
2	Exactly one inversion SV call lies inside the sandwich and covers ≥ 40% of it	Inversion (shared % empty = SV-backed)
3	Otherwise	Two SCE rows

Crossing the centromere alone does not force Inversion.

Long sandwiches that merely contain several scattered inversion calls stay as two SCEs (rule 2 requires exactly one contained call).

3.2 Two-step opposite (WW→WC→CC / CC→WC→WW)

Always two SCEs. A centromere inside the final homozygous run does not convert this into an inversion. (Not used on male chrX, where WC is absent.)

3.3 Other patterns

• Matching flanks but not a valid SCE sandwich (e.g. CC–WW–CC): clear the middle, emit nothing
• Other A–B–C middles: clear the middle

---

Step 4 — Single-switch SCE

After double switches are stripped, each chromosome is split at the centromere into arms. For each arm:

• exactly two state runs, and
• a valid SCE transition

→ one breakpoint = SCE (the start of the second run, after refinement in Step 5).

Centromere-crossing exception. If the kept flanks on either side of the barrier form a valid SCE transition and each flank is ≥ 5 Mb, emit one cross-centromere SCE. Prefer the raw switch inside the barrier when there is exactly one; otherwise use the start of the right flank.

Arms that still have multiple state changes after filtering are left without a single-SCE call (their double switches were already handled in Step 3).

---

Step 5 — Breakpoint refinement across SV holes

After stitching, the default breakpoint is the start of the first surviving bin after a removed gap. If that gap (or the breakpoint) touches an SV, the tool peeks at raw bins inside the gap without the SV mask:

Peek result	Action
Exactly one left → right switch	Move the coordinate there
None or more than one	Keep the default (far / right edge)

Sandwich breakpoints use the same correction, but only when SV intervals tile the gap completely. Without that guard, a chromosome-wide sandwich peek can wander into the centromere or a None desert and land tens of Mb away.

Why this matters. Without correction, a stitched breakpoint snaps to the far edge of a removed SV. That coordinate can collide with a real breakpoint in another cell and fake a recurrent subclone (Translocation).

---

Step 6 — Recurrence → Translocation

Among collected SCE-like breakpoints (not already resolved as SV-backed-only logic for inversions):

1. Compare only within the same sample + chrom
2. Count distinct cells with a breakpoint within ±10 kb
3. If that count is ≥ 5% of QC cells (and at least 2 cells) → relabel as Translocation and store the shared-cell percent
4. Rows already labeled Inversion are never relabeled as translocation

Recurrence is evidence of a shared breakpoint (clonal SV, recurrent fragile site, or mapping artifact). It is not proof of a translocation partner.

---

Output

Excel table with one row per candidate breakpoint (two rows for dual SCEs; one row with start/end for inversions).

Column	Meaning
Sample	Sample name
Cell_ID	Cell barcode / ID
chr	Chromosome
start	Breakpoint (left end for inversions)
end	Right breakpoint for Inversion; empty for SCE / Translocation
Event	SCE / Translocation / Inversion
Shared_cell_percent	Shared-cell % for Translocation and recurrent Inversion; empty for ordinary SCE and for SV-backed Inversion

How to read empty Shared_cell_percent on an Inversion

Shared_cell_percent	Meaning
Filled	Called from cross-cell recurrence of both sandwich ends
Empty	Called from the single inversion SV covering ≥40% of the sandwich

---

Command-line arguments

Only file / species paths are exposed on the CLI. Filter thresholds are fixed in SCEDetector.py to the values documented above (tuned against reference outputs such as SCE_detected.xlsx).

Argument	Default	Description
-i / --input	(required)	Raw 200 kb bins
--sv	(required)	SV intervals
--qc	(required)	QC-passed cells
--mask	HGSVC.200000.txt	None-bin mask
--species	human	Arm / centromere table
-o / --output	SCE_detected.xlsx	Output path

---

Scope and limitations

Included	Not included
Single valid switch per arm after filtering	Full generative model of multiple crossovers
A–B–A / two-step double-switch resolution	Haplotype-resolved validation of every call
Depth- and SV-aware artifact filters	Partner chromosome for translocation candidates
Breakpoint recurrence → translocation candidates	Proof that a recurrent site is a true rearrangement

Practical caveats:

• Calls remain candidates; always inspect depth and SV context for important events.
• Large duplications can erase a true nearby SCE if the whole WC block is dropped as duplication-like.
• True dual SCEs with WC ≤1 Mb beside a large SV-tiled hole are intentionally suppressed (rule 2c′).
• Thresholds were tuned on HPNE Strand-seq cohorts; new chemistries or bin sizes may need retuning.

---

License

Use freely for research and analysis.

Contributor

• Hyunjin Cho

hyun-jin891