A reproducible benchmark for single-cell multiome peak–gene link scoring, showing that apparent ranking improvements can be driven by promoter proximity rather than regulatory signal.
Docker · Snakemake · R / Seurat / Signac · 12k-cell PBMC multiome · 11 ranking models ·
genome-wide cross-method comparison · Zenodo-archived · reproduced from a clean clone
A containerized Snakemake pipeline takes candidate peak–gene pairs from Signac::LinkPeaks() on
10x PBMC multiome data, reranks the same fixed candidate set under 11 interpretable score modes
(RNA–ATAC coactivity, a genomic distance prior, TF/motif support), and evaluates every mode
against SCENT run genome-wide across 22 autosomes.
One of those modes is a control with no biology in it: rank by distance to the TSS.
The control won at the top of the list. Its top 50 links had a median peak-to-TSS distance of 3.5 bp and a promoter fraction of 1.00, i.e. promoters sitting on their own genes rather than enhancers. Raw top-N support could not distinguish a real result from proximity collapse.
Comparing within distance bins rather than across them, the coactivity scores hold up:
| Method | 0–10 kb | 10–50 kb | 50–100 kb |
|---|---|---|---|
full_lambda_0_1 (coactivity + distance + TF) |
5.06 | 3.15 | 2.64 |
coactivity alone |
4.68 | 2.89 | 3.08 |
linkpeaks (baseline) |
2.60 | 1.78 | 1.94 |
distance_only (control) |
1.59 | 0.92 | 1.12 |
Odds ratio, SCENT-supported links in the top decile vs the rest, within each distance bin.
At approximately fixed distance the reranking scores concentrate supported links roughly twice as strongly as the baseline, while the distance-only control drops to 0.92 — below 1, meaning proximity ranking within a distance bin is worse than arbitrary. So there is real signal beyond distance. It is smaller and narrower than the raw top-N numbers suggest.
SCENT only tested pairs within ±100 kb, so this benchmark says nothing about distal enhancer–gene links. Gene-level enrichment analysis favours the LinkPeaks baseline. Coactivity is co-occurrence, not regulation. One dataset, one tissue.
This is a benchmark and a diagnostic study, not a peak–gene linking method. Full accounting in
Limitations and docs/results_report.md.
Re-run end to end from a clean clone using the published image
(ghcr.io/inkasimo/multiome-link-ranking:v0.1.0). LinkPeaks output, the SCENT tested set and
every top-N supported fraction reproduced exactly. Archived on Zenodo; inputs fetched and
checksum-verified by a single script.
docker pull ghcr.io/inkasimo/multiome-link-ranking:v0.1.0
bash scripts/download_inputs.sh
python3 run_analysis.py run_reranker_with_scentThe image is ~5 GB (the restored R/Bioconductor library dominates), and a full run can take several days, almost all of it the SCENT sweep.
| Document | Contents |
|---|---|
docs/results_report.md |
Full results, what can and cannot be claimed |
docs/method_report.md |
Formal method description with formulas |
docs/input_output_reference.md |
Every input, config field, output column, and interface hazard |
docs/similar_tools.md |
Positioning against 11 existing methods |
docs/future_standalone_v0.md |
Next-phase plan |
| Stage | Frozen benchmark. Complete, documented, not under active development |
| Dataset | 10x Genomics pbmc_unsorted_10k, Cell Ranger ARC 2.0.0, hg38 / GRCh38-2020-A |
| Candidate universe | 5,000 LinkPeaks-derived pairs over 1,390 genes |
| Score modes | 11 committed; 7 compared against SCENT |
| External comparator | SCENT, 22 autosomes, 52,482 tested pairs |
| Conclusion | Sufficient to justify a de novo candidate-generation test. Not sufficient to claim a method |
Linking regulatory peaks to their target genes from paired single-cell RNA + ATAC data is unsolved, and the field has many methods that combine similar ingredients: accessibility– expression correlation, genomic proximity, and TF motif evidence. Most report improvements over a baseline. Few report what happens when a pure proximity ranking is included as a control.
This benchmark was built to answer one narrow question honestly:
Given a fixed candidate set produced by LinkPeaks, does reordering it with interpretable coactivity, distance and TF/motif terms produce a ranking with more external support than the LinkPeaks ordering — and does any advantage survive controlling for promoter proximity?
- Reads a 10x multiome filtered feature-barcode matrix plus ATAC fragments.
- Builds a Seurat object with a Signac chromatin assay; preprocesses RNA (PCA) and ATAC (TF-IDF/LSI); integrates by WNN; clusters.
- Runs
Signac::LinkPeaks()to generate candidate peak–gene pairs (500 kb window), filters to positive-score candidates, and keeps the top 5,000. - Builds one fixed feature table for those candidates: coactivity variants, peak–TSS distance and distance score, peak-level motif and TF/motif-expression scores.
- Reranks the identical candidate set under 11 interpretable score modes — LinkPeaks baseline,
coactivity-only, distance-only, a modified-distance control, coactivity+distance,
coactivity+TF, the full score at three λ values (0.1, 0.2, and 0.3 as the mode named
full), and the modified distance prior at two. - Runs SCENT independently across all 22 autosomes as an external comparator, using its own cis-window candidate set.
- Compares every score mode against SCENT support: top-N supported fraction, rank of supported links, and distance-matched enrichment within distance bins.
- Runs proximal-removal controls, discarding links within 10 / 25 / 50 kb of the TSS and recomputing the comparison.
Steps 7 and 8 are the point. Steps 1–6 exist to make them possible.
The pipeline does not generate its own candidate universe (LinkPeaks defines it, so recall is
bounded by LinkPeaks), perform causal inference, produce cell-type-specific output, or validate
against orthogonal data such as CRISPRi perturbation or eQTLs. It does not replace SCENT,
SCARlink, CREMA, SCENIC+, Pando, LINGER, FigR, ArchR Peak2GeneLinks, Cicero or TRIPOD — see
docs/similar_tools.md.
Evaluated universe: 4,976 pairs, 1,375 genes, 4,087 peaks (the 5,000-pair table restricted to
SCENT-covered chromosomes). SCENT: 52,482 tested rows, 4,758 supporting under
beta > 0 & boot_p <= 0.05.
SCENT-supported fraction, top 200 — scent_validation_topN_support_summary.csv:
| Method | frac | median distance |
|---|---|---|
full (λ = 0.3) |
0.680 | 2,775 bp |
full_lambda_0_1 |
0.605 | 7,855 bp |
full_moddist_lambda_0_1 |
0.600 | 7,855 bp |
coactivity |
0.525 | 10,035 bp |
coactivity_tf |
0.515 | 12,530 bp |
distance_only |
0.510 | 15 bp |
linkpeaks |
0.445 | 13,473 bp |
full_lambda_0_1 (λ = 0.1) is the conservative primary setting. full (λ = 0.3) is an aggressive
distance-prior sensitivity setting; see the distance-matched table before reading its raw support
figure.
At top 50, the distance-only control wins (0.580 vs 0.440), with a median distance of 3.5 bp
and a promoter fraction of 1.00. This table covers the whole 500 kb candidate universe, in which
links beyond 100 kb are counted as unsupported because SCENT never tested them; the
proximal-removal analysis below restricts to the tested window and reverses the distance_only
result.
Distance-matched enrichment, odds ratio, top decile vs rest —
scent_validation_distance_matched_enrichment.csv. Bins align with SCENT's 100 kb window;
everything beyond it is untestable and reported as such:
| Method | 0–10 kb | 10–50 kb | 50–100 kb |
|---|---|---|---|
full_lambda_0_1 |
5.06 | 3.15 | 2.64 |
full (λ = 0.3) |
5.06 | 3.15 | 2.44 |
coactivity_tf |
5.06 | 3.02 | 2.64 |
coactivity |
4.68 | 2.89 | 3.08 |
linkpeaks |
2.60 | 1.78 | 1.94 |
distance_only |
1.59 | 0.92 | 1.12 |
λ = 0.3 is identical to λ = 0.1 in both proximal bins — same odds ratio, same supported counts
— and slightly worse at 50–100 kb. There is no distance bin in which raising the distance prior
improves discrimination. coactivity alone is strongest in the outermost testable bin; the TF
term helps proximally and costs distally.
Proximal removal — scent_min_distance_topN_support_summary.csv and
scent_min_distance_delta_vs_linkpeaks.csv. This analysis is restricted to candidate links within
SCENT's 100 kb window before each threshold is applied, so that untested distal candidates are not
scored as unsupported. Every reranking mode stays ahead of LinkPeaks at all three thresholds and
at N = 50, 100 and 200 — full_lambda_0_1 by +0.02 to +0.135, full (λ = 0.3) by +0.02 to +0.18.
distance_only is the weakest method at every threshold and depth, and below LinkPeaks at
every one (−0.05 to −0.24).
full (λ = 0.3) has the higher raw support at several cells but gains no within-bin advantage
anywhere, so it is reported as a distance-prior sensitivity result and not as a better model.
Three things, in order.
1. The distance-matched result is the real finding. At approximately fixed distance, the reranking scores concentrate SCENT-supported links in their top decile roughly 1.4 to 2 times as strongly as LinkPeaks does, in every bin SCENT could test, while the distance-only control sits at 1.59, 0.92 and 1.12 — and at 0.84 and 0.69 in the finer 10–25 kb and 25–50 kb bins, i.e. below 1. Ranking by proximity within a distance bin is worse than arbitrary. Proximity alone cannot produce the reranking pattern, so the coactivity term carries information beyond distance across the whole 0–100 kb tested range.
2. The raw top-N advantage is partly a proximity effect. The full models' top-100 median
distance is 7.3 kb against LinkPeaks' 16.2 kb, and their promoter fraction is higher. Some of the
top-N gain is bought by ranking closer to promoters. Read
scent_validation_topK_supported_fraction.png and scent_validation_topK_median_distance.png
together; neither is interpretable alone.
3. Nothing here speaks to distal links. The SCENT sweep used a 100 kb window while candidates
extend to 500 kb, so the 100_200kb, 200_500kb and gt500kb bins contain zero supported links
for every method, and the odds ratios reported for them (8.906, 8.906 and 0.333) are
continuity-correction artifacts on empty cells. They must not be quoted. This is the benchmark's
largest limitation and no control within it addresses it.
Full analysis, including the gene-level ORA result that points the other way:
docs/results_report.md.
-
LinkPeaks defines the candidate universe. Recall is bounded by LinkPeaks, and the baseline is also the candidate generator. Nothing here finds links LinkPeaks missed.
-
SCENT is a comparator, not ground truth. It is built from the same two matrices as the scores it evaluates, is itself correlational, and only tested pairs within ±100 kb. Agreement between two correlational methods on shared input is weaker evidence than it looks.
-
Coactivity is co-occurrence, not regulation. A high
mul_weighmeans a peak is accessible in the same cells where a gene is expressed. It does not show the peak regulates the gene, and cannot separate a real regulatory link from two features that happen to be active in the same cell type. -
Coactivity is also sensitive to how often each feature is detected. If a gene and a peak are both frequently detected, they score high together whether or not they are related:
mul_weighcorrelates with the product of the two marginal detection rates at +0.68. LinkPeaks controls for this with a GC- and accessibility-matched background; neither this score nor SCENT does, so part of the apparent advantage over LinkPeaks may be a bias shared with the comparator. The distance controls hold proximity fixed, not detectability. Untested. -
The coactivity statistic is shape-sensitive, and more so for ATAC than RNA. Z-scoring itself assumes nothing about the distribution, but the
max(z, 0)clipping does. RNA counts are zero-inflated yet retain graded values above zero. ATAC counts are near-binary: a peak is usually seen in a cell with zero, one or two fragments, so after clipping the ATAC term is close to a detection indicator carrying almost no magnitude.mul_weightherefore behaves closer to a weighted count of co-detected cells than to a correlation, which is the mechanism behind the detection-rate association above. -
Raw top-N metrics reward promoter collapse. Over the unrestricted 500 kb universe,
distance_onlywins at top-50 with a median distance of 3.5 bp. Inside SCENT's 100 kb tested window the confound is controlled anddistance_onlybecomes the weakest method at every proximal-removal threshold, but no support fraction should be quoted without a distance control beside it. That distance is a strong baseline is an established result, not a finding of this work (seedocs/similar_tools.md). -
Gene-level ORA favours the baseline. LinkPeaks yields 17 enriched GO BP terms against 5 for
full_lambda_0_1. -
Nothing is cell-type-specific. Coactivity pools all cells, TF weights use global mean expression, and SCENT ran with a synthetic
all_cellslabel. A link active in one small population is diluted. -
The TF/motif score is peak-level only. It has no gene or cell-type dependence, so it cannot express TF-to-target specificity. Its contribution is small and inconsistent.
-
Scores are not portable across datasets. The TF/motif score is rescaled to [0, 1] using the minimum and maximum observed in this dataset, so 0.8 means "high relative to the other peaks here", not a fixed quantity. A single outlier peak shifts every other score.
-
λ and α are hand-set, not fitted. No held-out selection was performed. Seven of the eleven committed modes are compared against SCENT;
coactivity_distance,full_lambda_0_2,full_moddist_lambda_0_2anddistance_mod_only_lambda_0_1have no external comparison at all, and any statement about them rests on internal diagnostics only. -
λ = 0.3 is a sensitivity setting, not a better model.
full_lambda_0_1(λ = 0.1) is the conservative primary setting, chosen a priori as a guard against proximity domination.full(λ = 0.3) shows higher raw SCENT support but no within-bin advantage, so its gain is consistent with shifting the ranking into SCENT's tested window rather than discriminating better. -
SCENT results are not bit-reproducible. Its bootstrap p-values are stochastic and its workers are forked, so
seed: 42does not fully determine per-pair output. Re-running changes roughly 0.3% of support calls and shifts distance-matched odds ratios in the third decimal. See the reproduction check under Data availability. -
One dataset, one tissue, one sample. 5,000 pairs over 1,390 genes, drawn from the top of a LinkPeaks ranking rather than sampled at random.
-
The two halves of the pipeline are not perfectly aligned. The reranker used all cells; SCENT used 1,000. TSS conventions and peak ID formats differ. See
docs/input_output_reference.md§9.
- Docker. The full R/Bioconductor stack is inside the image:
ghcr.io/inkasimo/multiome-link-ranking:v0.1.0 - Python ≥ 3.9 with
pyyaml, only if usingrun_analysis.py(pip install -r wrapper-requirements.txt). - Memory: the Seurat/Signac feature-generation step is the peak consumer. The SCENT sweep is the
slowest step and was run with
max_cells: 1000andscent_cores: 4.
docker pull ghcr.io/inkasimo/multiome-link-ranking:v0.1.0Digest for the v0.1.0 release image:
ghcr.io/inkasimo/multiome-link-ranking@sha256:43216d1e5e8f196672f4423ccf4e88c4bab1137ba25c479ef3b1cbb1c1c10d18
Pin by digest for exact reproducibility; the tag is mutable, the digest is not.
Or build locally:
docker build -t ghcr.io/inkasimo/multiome-link-ranking:v0.1.0 -f containers/Dockerfile .Smoke-test:
docker run --rm ghcr.io/inkasimo/multiome-link-ranking:v0.1.0 \
Rscript -e 'library(Seurat); library(Signac); library(TFBSTools); library(JASPAR2022); library(motifmatchr); cat("R stack OK\n")'bash scripts/download_inputs.shDownloads the three 10x files into data/, fetches resources/jaspar/JASPAR2022.sqlite from the
supplementary Zenodo deposit
(10.5281/zenodo.22032568), and verifies everything
against resources/input_manifest.tsv. Use --verify to check without downloading.
| File | Notes |
|---|---|
data/filtered_feature_bc_matrix.h5 |
10x multiome; must carry both Gene Expression and Peaks assays |
data/atac_fragments.tsv.gz |
ATAC fragments |
data/atac_fragments.tsv.gz.tbi |
Tabix index — required, declared as an explicit workflow input |
resources/jaspar/JASPAR2022.sqlite |
JASPAR2022 motif database — see below |
JASPAR2022. Not tracked in Git and not downloadable from upstream: the URL in
scripts/run_linkpeaks_reranker.R is a BiocFileCache key, not a live source, and JASPAR has
since moved to ELIXIR hosting. scripts/run_linkpeaks_reranker.R seeds BiocFileCache with the
local file so the JASPAR2022 package does not attempt a network download at motif-loading time;
without it, feature generation fails in an offline container. It must be present in the
bind-mounted working directory, not only inside the image.
Genome build is hg38, via EnsDb.Hsapiens.v86 and BSgenome.Hsapiens.UCSC.hg38. Motifs are
JASPAR2022 CORE, tax_group=vertebrates, species=9606.
python3 run_analysis.py list_score_modes
python3 run_analysis.py run_all_score_modes --dry-run# heavy step once: builds the fixed feature table
python3 run_analysis.py build_linkpeaks_features
# all score modes from that table
python3 run_analysis.py run_all_score_modes
# SCENT sweep, then validation
python3 run_analysis.py run_scent_pipelinepython3 run_analysis.py <section> [options]
sections:
build_linkpeaks_features build the fixed feature table (heavy)
run_default_score run config default_score_mode
run_score_mode run one mode; requires --mode
run_all_score_modes run every mode in config/ablations.yaml
run_reranker_score_suite feature table + all score modes
run_scent_sweep SCENT producer across configured chromosomes
run_scent_validation SCENT consumer, cross-method comparison
run_scent_validation_min_distance proximal-removal controls (post-processing only)
run_scent_pipeline sweep + validation
run_reranker_with_scent everything
list_score_modes print configured modes
list_scent_run print SCENT producer settings
list_scent_methods print methods in the SCENT comparison
unlock release a stale Snakemake lock
options:
--mode MODE score mode for run_score_mode
--image IMAGE default: ghcr.io/inkasimo/multiome-link-ranking:v0.1.0
--snakefile PATH default: workflow/Snakefile
--configfile PATH default: config/default.yaml
--cores N / --cpus N default: 4
--dry-run
--rerun-incomplete / --no-rerun-incomplete default: on
--rerun-triggers POLICY default: mtime
--extra ... passed through to Snakemake; use last
docker run --rm -it -v "$PWD":/work -w /work \
ghcr.io/inkasimo/multiome-link-ranking:v0.1.0 \
snakemake --snakefile workflow/Snakefile --configfile config/default.yaml --cores 4 all
# including the SCENT sweep and validation
docker run --rm -it -v "$PWD":/work -w /work \
ghcr.io/inkasimo/multiome-link-ranking:v0.1.0 \
snakemake --snakefile workflow/Snakefile --configfile config/default.yaml --cores 4 all_with_scentTargets: all (features + all rankings), all_with_scent (adds the SCENT sweep, the cross-method
validation and the proximal-removal controls).
The proximal-removal controls are a rule of their own and can also be requested directly. They are light post-processing of the validation output and do not re-run SCENT:
python3 run_analysis.py run_scent_validation_min_distanceThresholds and rank depths are version-controlled in
config/scent_validation_min_distance.yaml (min_distances: 10000,25000,50000,
top_n_values: 50,100,200,500, high_fraction: 0.10), so the committed outputs and the
configuration agree.
config/ # user-editable configuration
default.yaml # dataset paths, feature and scoring parameters
ablations.yaml # score-mode definitions
scent_run.yaml # SCENT producer settings
scent_validation.yaml # SCENT consumer settings
containers/Dockerfile # reproducible R + Snakemake runtime
workflow/Snakefile # DAG: features -> rankings -> SCENT sweep -> validation
scripts/
run_linkpeaks_reranker.R # heavy: object build, LinkPeaks, coactivity, distance, motifs
evaluate_rankings.R # light: one score mode -> ranking + diagnostics
run_scent_chr_sweep.R # SCENT producer, per chromosome
benchmark_scent_validation.R# SCENT consumer, cross-method comparison
summarize_scent_validation_min_distance.R # proximal-removal controls
docs/ # method report, results report, I/O reference, positioning
data/ # input data (not versioned)
resources/jaspar/ # local JASPAR2022 SQLite (not versioned)
results/<dataset>/ # outputs
features/ # the fixed candidate universe and feature table
rankings/<mode>/ # one directory per score mode
scent_chr_sweep_<tag>/ # per-chromosome SCENT output
scent_validation/ # cross-method SCENT comparison
scent_validation_min_distance/ # proximal-removal controls
run_analysis.py # Docker + Snakemake wrapper
renv.lock # pinned R dependencies (268 packages)
| Path | Contents |
|---|---|
results/pbmc/features/pbmc_link_features.csv |
The fixed candidate universe. 5,000 pairs × 31 feature columns |
results/pbmc/features/pbmc_baseline_links_full.csv |
15,806 retained LinkPeaks candidates |
results/pbmc/rankings/<mode>/pbmc_<mode>_ranked_links.csv |
Primary ranking output per mode |
results/pbmc/rankings/<mode>/pbmc_<mode>_summary_metrics.csv |
Per-mode metrics |
results/pbmc/rankings/<mode>/pbmc_<mode>_distance_distribution.png |
Distance distribution of top links — the clearest proximity diagnostic |
results/pbmc/scent_validation/scent_validation_topN_support_summary.csv |
Headline comparison table |
results/pbmc/scent_validation/scent_validation_distance_matched_enrichment.csv |
Strongest evidence — within-bin enrichment |
results/pbmc/scent_validation_min_distance/scent_min_distance_delta_vs_linkpeaks.csv |
Proximal-removal controls |
Full column semantics: docs/input_output_reference.md.
The broader scientific goal is a standalone single-cell multiome peak–gene prioritization method that generates its own candidates and produces cell-type-specific, tiered, calibrated output usable for experimental follow-up. That is next-phase work and is not implemented in this repository.
Two findings from this benchmark shape it:
- A de novo cis-window candidate universe already exists here. The 22
results/pbmc/scent_chr_sweep_*/chr*/scent_candidates_chr*.csvfiles hold 117,811 pairs over 9,891 genes and 46,936 peaks, generated by this repository's own code from expressed genes, accessible peaks, same chromosome and TSS ±100 kb — independent of LinkPeaks, overlapping it by only 36.5%. - The bottleneck is the evaluation axis, not the score. Every result here is limited by what SCENT can test. More parameter tuning cannot improve the evidence.
Plan, prerequisites and go/no-go criteria:
docs/future_standalone_v0.md.
Raw input data is not included. Dataset provenance, download URLs, file sizes and SHA256 checksums
are recorded in config/default.yaml.
Large outputs — the eleven *_ranked_links.csv files, per-chromosome SCENT output, and the
combined validation table — are excluded from version control and archived separately at
10.5281/zenodo.22032568 (CC BY 4.0), together with the
JASPAR2022 motif database. scripts/download_inputs.sh fetches the JASPAR file from that deposit
automatically and verifies it against resources/jaspar/JASPAR2022.sqlite.sha256.
Note on tracked results. results/pbmc/scent_validation/ and
results/pbmc/scent_validation_min_distance/ are committed to Git so that the figures referenced
in docs/ render on GitHub. Snakemake .done sentinels are not committed, so a fresh clone
contains results but no sentinels — the workflow will therefore recompute everything and overwrite
the committed files. That is intentional and is how the reproduction check works: after a full run,
git diff shows whether the recomputed results match the committed ones.
Reproduction check. The full pipeline was re-run from a clean clone of this repository using
the published container image (ghcr.io/inkasimo/multiome-link-ranking:v0.1.0). LinkPeaks output
reproduced exactly — 16,824 links before filtering, 15,806 after, 5,000 candidates over 1,390
genes — as did the SCENT tested set (52,482 rows across 22 autosomes) and every top-N supported
fraction reported above. SCENT's bootstrap p-values are stochastic and its workers are forked, so
the seed does not fully determine per-pair results: the re-run gave 4,742 support rows against
4,758 originally, a difference of 16 pairs (0.3%). Distance-matched odds ratios differ in the third
decimal. No conclusion in this repository depends on that variation.
Please cite this repository using CITATION.cff and the archived Zenodo DOI:
10.5281/zenodo.22032459 resolves to the latest version;
10.5281/zenodo.22032460 pins v0.1.0. Supplementary data
has its own DOI, 10.5281/zenodo.22032568.
This workflow also depends on external software and resources. Cite the relevant upstream projects
when using or reusing the workflow: Signac / LinkPeaks, Seurat, SCENT
(immunogenomics/SCENT v1.0.1, commit e80b5ba6b445f972c7fe28fb41e24ef4f5b2e373), JASPAR2022,
motifmatchr, TFBSTools, EnsDb.Hsapiens.v86, BSgenome.Hsapiens.UCSC.hg38, clusterProfiler
and Snakemake.
The supplementary Zenodo deposit includes resources/jaspar/JASPAR2022.sqlite for
reproducibility. The JASPAR database is licensed under Creative Commons Attribution 4.0
International; cite the JASPAR 2022 Nucleic Acids Research database paper and retain attribution
when reusing the file.
Code: MIT (see LICENSE). Documentation (docs/, README.md) and results (results/) are
released under CC BY 4.0.