Design oligonucleotide probes for single molecule RNA FISH (smFISH) and HCR v3 experiments.
This repo is a fork of https://github.com/arjunrajlaboratory/ProbeDesign from the Raj Lab.
ProbeDesign creates oligonucleotide probes for smFISH and HCR v3 split-initiator FISH. It selects optimal probe sequences based on:
- Thermodynamic properties — targets specific Gibbs free energy for RNA-DNA hybridization (Sugimoto 1995)
- Sequence masking — avoids cross-hybridization to pseudogenes, repeats, and repetitive genomic regions
- Optimal spacing — uses dynamic programming to find the best probe placements
- Mixed-length probes — variable-length oligos (e.g. 18-22 bp) to maximise probe count on GC-variable sequences
- Low-complexity filtering — excludes probes with homopolymer or dinucleotide repeats (L-mask)
- HCR v3 split-initiator pairs — two 25-nt halves per pair, 13 amplifiers (B1–B17)
Completed
- ✅ Create streamlit GUI application
- ✅ Batch mode for CLI and GUI
- ✅ Implement homopolymer/dinucleotide filter
- ✅ Implement automated mixed-length probe design compatible with dynamic programming
- ✅ Allow bowtie result output
- ✅ Implement HCRv3 probe design algorithms from the RShiny scripts
- ✅ Host test streamlit application on Glasgow VM
- ✅ Implement Isoform Explorer — browse constitutive sequences and download FASTA for probe design (Ensembl v113, Hsap/Mmus/Dmel)
- ✅ IDT 96-well plate order export — per-run button, IDT Plate tab to combine probe sets,
probedesign idtplate
In Development / Planned
- Test local RepeatMasker installation
- Implement hairpin structure & inter-probe duplex formation filters
- Quick Start
- Installation
- Web App (Streamlit)
- Command-Line Interface
- Masking Options
- Output Files
- Troubleshooting
- Project Structure
- Further Reading
# 1. Install a package manager (mamba or conda)
# See Section 2 for details
# 2. Clone and set up
git clone https://github.com/JeffLeeLab/smFISHProbeDesign.git
cd smFISHProbeDesign
chmod +x setup_all.sh
./setup_all.sh # ~18 min first run (downloads genome indices)
# 3. Launch the web app
mamba activate probedesign # or: conda activate probedesign
streamlit run streamlit_app/app.py
# 4. Or use the CLI
probedesign design input.fa --probes 48 --pseudogene-mask --genome-mask # smFISH
probedesign hcrdesign input.fa --pairs 30 -a B1 --pseudogene-mask --genome-mask # HCR| OS | Supported | Notes |
|---|---|---|
| macOS (Apple Silicon / Intel) | ✅ | |
| Linux (x86_64 / aarch64) | ✅ | |
| Windows | wsl --install in PowerShell, then use Ubuntu terminal |
You need one of mamba or conda. The setup script auto-detects whichever is available.
Option A — Miniforge (recommended):
- Download from github.com/conda-forge/miniforge/releases/latest
- Run:
bash ~/Downloads/Miniforge3-*.sh - Restart terminal, verify:
conda --version
git clone https://github.com/JeffLeeLab/smFISHProbeDesign.git
cd smFISHProbeDesign
chmod +x setup_all.sh
./setup_all.shThe script will:
- Create a
probedesignconda environment (Python 3.11 + Bowtie 1.3.1) - Install the
probedesignpackage and the web-app dependencies (Streamlit, pandas, pyarrow, matplotlib, openpyxl, TransQuant) - Build pseudogene alignment indices (~1 min)
- Download genome alignment indices (~12 min, ~6.7 GB)
- Run validation tests
To skip the 6.7 GB genome download: ./setup_all.sh --skip-genome
# Create environment from environment.yml
mamba env create -f environment.yml # or: conda env create -f environment.yml
mamba activate probedesign
pip install -e .
# Update an existing environment
mamba env update -n probedesign -f environment.yml --prune
# Remove and recreate from scratch
mamba env remove -n probedesign
mamba env create -f environment.ymlWithout environment.yml:
mamba create -n probedesign \
-c conda-forge -c bioconda --channel-priority strict \
python=3.11 "click>=8.0" bowtie=1.3.1 "pytest>=7.0" -y
mamba activate probedesign
pip install -e .
pip install "streamlit>=1.48" "pandas>=1.5" "pyarrow>=14.0" "matplotlib>=3.8" "openpyxl>=3.1"
pip install "git+https://github.com/JeffLeeLab/TransQuant-probe-weight.git@v0.2.0"python -m pytest tests/ # HCR unit, smFISH regression, run-header and IDT plate teststests/test_smfish_regression.py reproduces the reference designs in tests/regression/ (CDKN1A, KRT19, mouseMITF). The KRT19 and mouseMITF tests need bowtie and the human/mouse indices, and are skipped without them. setup_all.sh also checks CDKN1A and KRT19 against the same references. tests/test_provenance.py covers the # run header in _seq.txt / _HCR_seq.txt, including that its command reproduces the design.
The primary interface for probe design is a Streamlit web application with five tabs in the top navigation bar: smFISH, HCR, Isoform Explorer, TransQuant and IDT Plate.
mamba activate probedesign # or conda
streamlit run streamlit_app/app.py
# Opens http://localhost:8501- Upload or paste a FASTA file
- Adjust parameters in the sidebar (probes, oligo length, masking)
- Click Design Probes
- Download
_oligos.txt,_seq.txt, the bowtie hit tables (when masking is on), and raw bowtie files (when Save raw bowtie output is ticked) - Download IDT plate (.xlsx) gives an IDT 96-well plate order for this probe set alone. To put several sets on one order, use the IDT Plate tab
- Switch Mode to Batch
- Provide a folder of FASTA files (or upload multiple files)
- Click Run Batch, then download all results as a ZIP plus
batch_summary.tsv. ItsWarningscolumn lists any mask that failed for a gene. - Optional: set an output directory under Auto-save outputs to also write results to disk on the server
| Parameter | Default | Description |
|---|---|---|
| Mode | Single sequence | Single sequence or Batch |
| Species | human | Genome/pseudogene database |
| Number of probes | 48 | Target probe count (no upper limit) |
| Oligo length mode | Fixed (20 bp) | Fixed length or mixed range (e.g. 18-22 bp); 10–60 bp |
| Spacer length | 3 bp | Minimum gap between probes (no upper limit) |
| Target Gibbs FE | −23.0 kcal/mol | Optimal binding free energy |
| Allowable Gibbs range | −26 to −20 | Probes outside this are excluded |
| Homopolymer / Dinucleotide repeats | 5 / 3 | Low-complexity (L-mask) thresholds, 3–10 |
| Pseudogene mask | on | Filter probes matching pseudogenes |
| Genome mask | on | Filter repetitive genome regions |
| Bowtie index directory | <repo>/bowtie_indexes |
Shown when a bowtie mask is on |
| Save raw bowtie output | off | Add raw alignment files to the downloads |
| RepeatMasker | None | None, Auto (requires extra setup), or provide a masked FASTA |
Select Mixed range under "Oligo length mode" to use variable-length probes (e.g. 18-22 bp). This maximises probe count on sequences with variable GC content — GC-rich regions get shorter probes, AT-rich regions get longer probes, all within the same Gibbs FE target range.
The HCR tab designs HCR v3 split-initiator probe pairs. Each pair is two 25-nt probes separated by a 2-nt gap on the target, with half-initiator sequences attached. The two halves hybridise independently, so ΔG targets apply per 25-nt half. Single and batch modes work as for smFISH. Downloads are _HCR_oligos.txt, _HCR_seq.txt, _HCR_hits.txt and an IDT plate order (.xlsx, full oligos with initiators); batch mode adds a ZIP and hcr_batch_summary.tsv.
| Parameter | Default | Description |
|---|---|---|
| Amplifier | B1 | B1–B5, B7, B9–B11, B13–B15, B17 (B7+ use a WW degenerate spacer) |
| Number of pairs | 30 | Target pair count |
| Pair spacing | 3 nt | Minimum gap between consecutive 52-nt pair blocks |
| Target ΔG per half | −31.0 kcal/mol | Optimal binding free energy per half |
| Strict ΔG range | −35 to −27 | Range each half must satisfy |
| Asymmetric ΔG leniency | off | Let one half go down to the Lenient ΔG floor (−42) |
| Homopolymer / Dinucleotide repeats | 5 / 3 | Low-complexity (L-mask) thresholds |
| Pseudogene / Genome mask | on / on | Bowtie off-target filters |
| Asymmetric bowtie | off | Only require one half to be free of off-target hits |
| Resolve WW spacer | blank | Replace WW with explicit bases (e.g. AA) for B7+ |
The Isoform Explorer tab (top navigation bar) lets you inspect the isoform architecture of any gene and identify whether a constitutive sequence exists — a region shared by all isoforms — for use as probe-design input.
Workflow:
- Select species (Human, Mouse, or Drosophila) and search for a gene by name or Ensembl ID.
- The isoform structure plot shows all transcripts coloured by Transcript Support Level (TSL) for mammals. Isoforms excluded by the adaptive TSL filter are de-emphasised.
- If a constitutive sequence exists, copy it from the sequence box or download the FASTA.
- Paste the FASTA into the smFISH or HCR designer — the N markers at exon–exon junctions prevent probes from spanning junctions.
Data: Ensembl v113 · Hsap 55,184 genes · Mmus 54,705 genes · Dmel 16,797 genes. Mammalian data uses an adaptive TSL filter (TSL ≤ 2 where available, otherwise the best available level). The FASTA header records the per-gene TSL cutoff applied.
The TransQuant tab (top navigation bar) embeds the TransQuant probe-weight tool (external transquant_w package). Given the full genomic locus of a gene and a probe set, it computes the probe weight factor W (corrects transcription-site brightness for where probes bind relative to Pol II), the gene length L, and the probe localisation profile N — following the TransQuant methodology (Bahar Halpern & Itzkovitz, Methods 2016). It is fully self-contained — no inputs or results are shared with the smFISH / HCR / Isoform Explorer / IDT Plate tabs.
A standalone, browser-only version (no install needed) is hosted at jeffleelab.github.io/TransQuant-probe-weight.
The embedded version is pinned to a release tag (currently v0.2.0). To update it, bump the @<tag> in all three places it appears — environment.yml, setup_all.sh and the install hint in streamlit_app/pages/transquant.py — then run pip install --upgrade "git+https://github.com/JeffLeeLab/TransQuant-probe-weight.git@<new-tag>" inside the probedesign environment (or re-run ./setup_all.sh).
The IDT Plate tab combines probe sets from several runs into one IDT 96-well plate order (the Excel file IDT's plate upload accepts).
- Upload
_oligos.txt/_HCR_oligos.txtfiles, or a batch results ZIP. smFISH and HCR sets can share a plate. - In the table, set the Order sets go onto plates, untick sets to leave out, and check the Gene column. Gene is the probe-set name without its region (
Eaat1-cmm→Eaat1,TDP-43-CDS-DmelCO→TDP-43); edit it if the region wasn't stripped correctly. - Set Minimum oligos per plate in the sidebar: 24 for standard 96-well plates (default), 48 for HOTPlates.
- Check the plate maps and download
IDT_plate-ordering_<YYYYMMDD>_<genes>.xlsx(genes joined by_; a long list ends in_+N).
Layout rules:
- Wells fill down each column: A1–H1, then A2 …
- Each probe set starts on a new column.
- A set is never split across plates; one that doesn't fit moves to the next plate. Only a set larger than 96 oligos is split, over whole plates.
- Each plate is a worksheet named after its genes (Excel's 31-character limit; extra genes show as
+N). - Plates are filled in order with the fewest plates that each meet the minimum, earlier plates first. If no layout meets it (e.g. a single 16-oligo set), the app warns.
The same rules apply to the single-run Download IDT plate buttons and to probedesign idtplate.
Usage: probedesign design [OPTIONS] INPUT_FILE
Design probes for a target sequence.
INPUT_FILE is a FASTA file containing the target sequence.
Options:
-n, --probes INTEGER Number of probes to design [default: 48]
-l, --oligo-length TEXT Oligo length: single int (e.g. 20) or range
for mixed lengths (e.g. 18-22; range bounds
10-60) [default: 20]
-s, --spacer-length INTEGER Minimum gap between probes [default: 3]
-g, --target-gibbs FLOAT Target Gibbs free energy in kcal/mol [default: -23]
--allowable-gibbs TEXT Allowable Gibbs FE range as min,max [default: -26,-20]
-o, --output TEXT Output file prefix (default: derived from input filename)
-q, --quiet Suppress output to stdout
--species [human|mouse|elegans|drosophila|rat]
Species for masking databases [default: human]
--pseudogene-mask / --no-pseudogene-mask
Mask regions that align to pseudogenes [default: off]
--genome-mask / --no-genome-mask
Mask repetitive genomic regions [default: off]
--index-dir DIRECTORY Directory containing bowtie indexes
[default: <repo>/bowtie_indexes]
--repeatmask-file PATH FASTA file with N's marking repeat regions (manual masking)
--repeatmask / --no-repeatmask Run RepeatMasker automatically to mask repeats [default: off]
--hp-threshold INTEGER Homopolymer repeat threshold for L-mask (probes with
single-nucleotide runs >= this length are excluded)
[default: 5; range: 3-10]
--di-threshold INTEGER Dinucleotide repeat threshold for L-mask (probes with
dinucleotide motifs repeated >= this many times are
excluded) [default: 3; range: 3-10]
--save-bowtie-raw Save raw bowtie alignment output (produces large files)
--help Show this message and exit.
Designs HCR v3 split-initiator probe pairs. ΔG values are per 25-nt half. Don't use design -l 52 for HCR.
Usage: probedesign hcrdesign [OPTIONS] INPUT_FILE
Options:
-p, --pairs INTEGER Number of probe pairs to design [default: 30]
-a, --amplifier [B1|B2|B3|B4|B5|B7|B9|B10|B11|B13|B14|B15|B17]
HCR amplifier [default: B1]
-g, --target-gibbs FLOAT Target ΔG per 25-nt half [default: -31.0]
--allowable-gibbs TEXT Strict ΔG range as min,max [default: -35,-27]
--asymmetric-gibbs One half strict, one lenient
--lenient-gibbs-min FLOAT Floor ΔG for the lenient half [default: -42.0]
--asymmetric-bowtie Require only one half to be bowtie-clean
--pair-spacing INTEGER Minimum gap between 52-nt pair blocks [default: 3]
-s, --species [human|mouse|elegans|drosophila|rat]
Species for masking databases [default: human]
--pseudogene-mask / --no-pseudogene-mask [default: off]
--genome-mask / --no-genome-mask [default: off]
--index-dir DIRECTORY [default: <repo>/bowtie_indexes]
--repeatmask-file PATH FASTA file with N's marking repeat regions
--hp-threshold INTEGER Homopolymer threshold [default: 5; range: 3-10]
--di-threshold INTEGER Dinucleotide threshold [default: 3; range: 3-10]
--resolve-spacer TEXT Replace WW spacer with explicit bases (e.g. AA)
-o, --output TEXT Output file prefix
--save-bowtie-raw Save raw bowtie alignment output
-q, --quiet Suppress output to stdout
Note that -s means --species here but --spacer-length in design.
Check the Gibbs FE distribution of a sequence before designing probes — useful for choosing
appropriate --target-gibbs and --allowable-gibbs values.
Usage: probedesign analyze [OPTIONS] INPUT_FILE
Options:
-l, --oligo-length INTEGER Length of oligonucleotide to analyze [default: 20]
--help Show this message and exit.
Build an IDT 96-well plate order (.xlsx) from oligo files of one or more runs, with the same layout rules as the IDT Plate tab. Files go onto plates in the order given.
Usage: probedesign idtplate [OPTIONS] OLIGO_FILES...
Options:
-o, --output-dir DIRECTORY Directory for the plate file (default: current directory)
--min-oligos INTEGER RANGE IDT minimum oligos per plate: 24 for standard 96-well
plates, 48 for HOTPlates (default: 24)
OLIGO_FILES are _oligos.txt / _HCR_oligos.txt files or batch-result ZIPs. Needs openpyxl (in environment.yml; or pip install -e ".[idt]").
# Check Gibbs FE distribution before designing
probedesign analyze input.fa
# Basic: 32 probes, no masking
probedesign design input.fa --probes 32
# Recommended: pseudogene + genome masking
probedesign design input.fa --probes 48 --pseudogene-mask --genome-mask
# Mixed-length probes (18-22 bp) — maximises count on GC-variable sequences
probedesign design input.fa --probes 48 -l 18-22
# Repeat masking from a pre-masked file (N's mark repeats)
probedesign design input.fa --probes 32 --repeatmask-file input_masked.fa
# Mouse gene, all masking, custom output prefix
probedesign design my_gene.fa --probes 48 --species mouse \
--pseudogene-mask --genome-mask -o MyGene_probes
# HCR: 30 pairs for amplifier B1 with masking
probedesign hcrdesign input.fa --pairs 30 -a B1 --pseudogene-mask --genome-mask
# HCR: relax ΔG on one half when a target is hard to cover
probedesign hcrdesign input.fa -p 25 -a B2 --asymmetric-gibbs
# HCR: B7+ amplifier with WW spacer resolved to AA
probedesign hcrdesign input.fa -p 30 -a B9 --resolve-spacer AA
# IDT plate order for two probe sets (HOTPlates: 48-oligo minimum)
probedesign idtplate Eaat1-cmm_oligos.txt Eaat2-cmm_oligos.txt --min-oligos 48Masking removes probe candidates that would hybridize non-specifically.
| Method | Flag | Use case |
|---|---|---|
| Auto-detect | (none) | FASTA already has Ns at repeats |
| Manual file | --repeatmask-file PATH |
Pre-masked sequence file |
| Automatic | --repeatmask |
Runs RepeatMasker locally (see docs/REPEATMASKER.md) |
16-mer exact alignment against curated pseudogene databases. Indices built by setup_all.sh.
Three bowtie searches (12/14/16-mers) against the whole genome with tiered hit-count thresholds.
Probes containing homopolymer runs (e.g. AAAAA) or dinucleotide repeats (e.g. ACACAC) are excluded. Thresholds are adjustable via --hp-threshold (default 5) and --di-threshold (default 3).
| Species | --species |
Pseudogene | Genome |
|---|---|---|---|
| Human | human |
✅ | ✅ |
| Mouse | mouse |
✅ | ✅ |
| Drosophila | drosophila |
✅ | ✅ |
| C. elegans | elegans |
Manual | Manual |
| Rat | rat |
Manual | Manual |
setup_all.sh builds or downloads indices for human, mouse and drosophila. For elegans or rat, build the indices yourself into bowtie_indexes/, named celegansPseudo / celegans and ratPseudo / rat.
Tab-separated, one probe per line. Order sequences from the last two columns when placing synthesis orders.
| Column | Content |
|---|---|
| 1 — Index | Probe number (1-based) |
| 2 — Start | Start position in target sequence (1-based, nucleotides) |
| 3 — GC% | GC content |
| 4 — Tm | Melting temperature (°C) |
| 5 — Gibbs | ΔG (kcal/mol) |
| 6 — Sequence | Probe sequence (5'→3') |
| 7 — Name | Probe label |
Visual sequence map (110 chars per line) showing probe placement and masking. Useful for checking masking behaviour.
The file starts with a run header: lines beginning with #. It records:
- the equivalent
probedesigncommand; - the input and every design parameter;
- the version and git commit;
- for each mask, whether it was applied, which bowtie index was used, and a short fingerprint of that index (so runs on different machines can be compared);
- the probes found and the percentage covered by each mask.
A mask that was requested but failed (bowtie missing, a broken index) reads FAILED, not applied. grep -v '^#' strips the header and leaves the sequence map unchanged.
Each block shows:
- Sequence line: Original target sequence
- R line (when repeat masking used):
R= masked position, base = unmasked - P line (when pseudogene mask used):
P= masked - B line (when genome mask used):
B= masked - F line (always):
F= thermodynamically excluded, base = valid probe start - L line (always):
L= probe start excluded by the low-complexity filter - Probe complement line: reverse-complement sequence placed below the target
- Probe label line:
Prb# N,Pos X,FE Y,GC Zfor each probe
| File | Content |
|---|---|
<name>_bowtie_pseudogene.txt |
Per-position 16-mer hit counts from pseudogene search |
<name>_bowtie_genome.txt |
Per-position hit counts from genome search (12/14/16-mers) |
<name>_bowtie_*_raw.txt |
Raw bowtie alignment output (only with --save-bowtie-raw) |
| File | Content |
|---|---|
<name>_HCR_oligos.txt |
TSV with header: pair, half (P1/P2), start, GC%, Tm, Gibbs, strict (yes/no), full_oligo, name. full_oligo includes the half-initiator and spacer. Names are <name>_HCR<amp>_NN (odd = P1, even = P2) |
<name>_HCR_seq.txt |
# run header (as for _seq.txt), then the sequence map with mask lines and pair annotations |
<name>_HCR_hits.txt |
Per-pair, per-half bowtie hit counts with strict/lenient annotation |
<name>_HCR_bowtie_*_raw.txt |
Raw bowtie output (only with --save-bowtie-raw; CLI only) |
Excel workbook for IDT's 96-well plate upload. One worksheet per plate, named after its genes, with columns Well Position, Name, Sequence; empty wells are left out. Written by the Download IDT plate buttons, the IDT Plate tab and probedesign idtplate.
| Problem | Solution |
|---|---|
mamba / conda: command not found |
Restart terminal after installing; re-run installer if needed |
bowtie: command not found |
Activate environment: mamba activate probedesign |
Warning: Pseudogene/Genome masking failed: Could not find bowtie executable |
Bowtie isn't on PATH, so the design ran without that mask. Activate the environment and re-run |
| Wrong bowtie (file manager, not aligner) | Only install via bioconda, not homebrew |
| Setup fails at pseudogene indices | Check data/pseudogeneDBs/*.fasta exist |
| Genome download interrupted | Re-run ./setup_all.sh. Species already downloaded are skipped; the interrupted one restarts |
streamlit: command not found |
Activate environment first |
| Port 8501 in use | streamlit run streamlit_app/app.py --server.port 8502 |
IDT plate export needs openpyxl |
Re-run ./setup_all.sh, or pip install openpyxl inside the probedesign environment |
| Windows | Use WSL2 Ubuntu terminal, not PowerShell |
smFISHProbeDesign/
├── src/probedesign/ # Python package
│ ├── cli.py # Click CLI: design, hcrdesign, analyze, idtplate
│ ├── core.py # smFISH algorithm (badness, DP, mixed-length)
│ ├── hcr.py # HCR split-initiator pair design
│ ├── thermodynamics.py # Gibbs FE and Tm (Sugimoto 1995)
│ ├── masking.py # Bowtie masking (pseudogene, genome), RepeatMasker, low-complexity
│ ├── sequence.py # Sequence utilities
│ ├── fasta.py # FASTA I/O
│ ├── provenance.py # Run header for _seq.txt / _HCR_seq.txt
│ ├── output.py # smFISH output files
│ ├── hcr_output.py # HCR output files
│ └── idt.py # IDT 96-well plate order (.xlsx)
├── streamlit_app/ # Streamlit web app
│ ├── app.py # Entry point + top navigation
│ ├── pages/ # smfish.py, hcr.py, isoform_explorer.py, transquant.py, idt_plate.py
│ ├── utils.py # Backend helpers (run/batch wrappers, validation)
│ └── isoform_data.py # Isoform Explorer data access + plot
├── data/
│ ├── isoform_explorer/ # Isoform Explorer data (Ensembl v113, Parquet)
│ └── pseudogeneDBs/ # Pseudogene FASTA files (indices built from these)
├── docs/ # Documentation (see Further Reading)
├── tests/ # pytest suite (HCR unit, smFISH regression, run header, IDT plate)
│ └── regression/ # Regression inputs + reference outputs
├── legacy/ # Archived upstream MATLAB / Python 2 code (not used)
├── bowtie_indexes/ # Genome/pseudogene indices (gitignored)
├── setup_all.sh # Automated setup script
├── environment.yml # Conda environment specification
├── pyproject.toml # Python package config
├── CLAUDE.md # AI coding context
└── README.md # This file
| Document | Description |
|---|---|
| docs/probe_design_principles.md | Thermodynamic model, masking algorithm, DP optimization |
| docs/summary.md | Development notes, Bowtie 1/2 index compatibility |
| docs/BOWTIE.md | Bowtie installation and index setup |
| docs/REPEATMASKER.md | RepeatMasker installation (~56 GB database) |
| docs/MIXED_LENGTH_PLAN.md | Mixed-length probe design plan and analysis |
| docs/hcr_designer_plan.md | HCR split-initiator designer plan |
| docs/nucleotide_repeat_plan.md | Low-complexity (L-mask) filter plan |
| docs/isoform_explorer/ | Isoform Explorer plan and data-build scripts |
The *_plan.md files are design plans for features that have already shipped. They're kept for background, so where they disagree with the code, the code wins.