lhqing commited on
Commit
6f0fdb0
·
verified ·
1 Parent(s): 9d75a9b

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +35 -67
README.md CHANGED
@@ -1,82 +1,50 @@
1
  ---
2
- pretty_name: seqforge benchmark
3
  tags:
4
- - genomics
5
- - sequencing
6
  - single-cell
 
7
  - benchmark
8
- - metadata-resolution
9
- viewer: false
10
  ---
11
 
12
- # seqforge-benchmark
13
-
14
- The standard benchmark for [**seqforge**](https://github.com/liuhlab/seqforge) — a compiler that turns
15
- arbitrary FASTQ files plus unstructured metadata into a validated, machine-independent `manifest.yaml`
16
- and a runnable Snakemake pipeline.
17
-
18
- This repo hosts the benchmark's **fingerprint packages** so that seqforge's evaluation can run
19
- **automatically in CI** without needing the full (often multi-terabyte) datasets present. It is the
20
- online home referenced by seqforge milestone **M1**
21
- ([issue #56](https://github.com/liuhlab/seqforge/issues/56)).
22
-
23
- > **Status: early / work in progress.** The layout below is the intended shape; packages land as the
24
- > M1 preflight tooling comes online. Nothing here is final yet.
25
-
26
- ## What is a fingerprint package?
27
-
28
- A fingerprint package is a small, self-contained `<dataset>.fingerprint.tar.gz` that stands in for a
29
- whole dataset. It contains:
30
-
31
- - **Information files** — the paper PDF plus its extracted text and images; any spreadsheets with each
32
- sheet extracted to text.
33
- - **Subsampled FASTQs** — every FASTQ cut to the first N reads, preserving the original directory
34
- structure and filenames exactly.
35
-
36
- seqforge decides what a dataset *is* from a bounded sample of reads (it never reads a whole FASTQ), so
37
- a fingerprint carries enough signal to reproduce the **same manifest** the full dataset would — even
38
- when the original FASTQs are gone. That is what makes a light, portable, CI-friendly benchmark
39
- possible.
40
-
41
- ## What this benchmark is for
42
-
43
- - **Regression safety.** As seqforge grows to new technologies and pipelines, these representative
44
- datasets must keep resolving correctly — and must never *silently* mis-compile. Each case ships a
45
- pre-registered expectation; a verified-but-wrong answer is a hard failure.
46
- - **A validation set, not a test set.** We develop against this corpus. A truly held-out test set for
47
- unbiased accuracy will be a separate, later effort.
48
-
49
- ## Intended layout
50
-
51
- ```
52
- cases/
53
- <dataset-id>/
54
- <dataset-id>.fingerprint.tar.gz # subsampled FASTQs + extracted info files
55
- expected.yaml # pre-registered, checkable predictions
56
- README.md # provenance: source accession, chemistry, why it's here
57
- ```
58
 
59
- Pre-registration discipline: `expected.yaml` records what the answer should be **before** a run —
60
- only a prediction can be wrong. Claims are checkable against the compiled manifest (e.g. per-sample
61
- attributes, resolved chemistry).
 
 
62
 
63
- ## Seed corpus (M1)
64
 
65
- The first packages come from seqforge's real-world compile audits:
 
 
 
 
 
66
 
67
- - 13 *C. elegans* single-cell datasets (seqforge issue #42).
68
- - 1 BD Rhapsody Enhanced ("v2") dataset (seqforge issue #43).
69
 
70
- Coverage will broaden with seqforge itself — additional modalities (ATAC, multiome, ChIP, long-read,
71
- WGBS, feature-barcode) as those pipelines land.
 
 
72
 
73
- ## Provenance and licensing
74
 
75
- Fingerprints are derived from **public** datasets (subsampled reads + published supplementary
76
- material). Each case's `README.md` records its source accession. Private lab data is **never**
77
- uploaded here — those cases stay out-of-band and run only where their data roots are configured.
78
 
79
- ## Links
 
 
 
 
 
 
 
 
80
 
81
- - seqforge: <https://github.com/liuhlab/seqforge>
82
- - Milestone M1 (this benchmark's tracking issue): <https://github.com/liuhlab/seqforge/issues/56>
 
1
  ---
2
+ license: cc-by-4.0
3
  tags:
 
 
4
  - single-cell
5
+ - genomics
6
  - benchmark
7
+ - seqforge
8
+ pretty_name: seqforge validation benchmark
9
  ---
10
 
11
+ # seqforge validation benchmark
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
 
13
+ Byte-light **fingerprint packages** for [`seqforge`](https://github.com/liuhlab/seqforge)'s validation
14
+ benchmark. Each `packages/<accession>.fingerprint.tar.gz` is a head-slice of a real public dataset's
15
+ FASTQs (the first ~2000 reads per file) plus a pin carrying the whole-file identity — enough for
16
+ seqforge to reproduce the same `manifest.yaml`, hash and all, **without the original FASTQs**. This is
17
+ the growing validation set the compiler is developed against; a true held-out test set is separate.
18
 
19
+ ## What's here (and what's not)
20
 
21
+ - **Reads:** head slices only — never a whole FASTQ. Enough to fingerprint the chemistry, far too few
22
+ to reconstruct the biology.
23
+ - **Prose:** extracted **text only** (`info/text/`). The raw papers and their figures are **not**
24
+ redistributed here (copyright); the packages are built with `seqforge preflight --redistributable`
25
+ (or repacked with `seqforge strip-fingerprint`), which is why they are usable for harvest yet carry
26
+ nothing that may not be shared.
27
 
28
+ ## How it is used
 
29
 
30
+ seqforge's eval harness pulls a package by accession (anonymous, pooch-cached — no token) and runs the
31
+ real compiler over the slice, grading the resolved library against a pre-registered `expected.yaml`
32
+ committed in the repo. Sample-level facts are graded from the archive's own BioSample/SRA records, also
33
+ committed in the repo. See `evals/benchmark/` in the seqforge repository.
34
 
35
+ ## Datasets
36
 
37
+ All *C. elegans* (taxid 6239), chosen to span technology × prep × layout × known regressions at
38
+ moderate sample size:
 
39
 
40
+ | accession | chemistry | notes |
41
+ |---|---|---|
42
+ | PRJNA1027859 | 10x 3' v3/v3.1 | single-nucleus neurons (the pilot demo dataset) |
43
+ | GSE274290 | BD Rhapsody WTA | the only BD-original dataset; germline |
44
+ | GSE234962 | 10x 3' v3 | motor neurons |
45
+ | GSE256266 | 10x 3' v3/v3.1 | glia across sexes; irregular filename layout |
46
+ | PRJNA658829 | 10x 3' v3 | whole-organism; equal-length read layout |
47
+ | PRJNA1195922 | 10x 3' v3/v3.1 | male single-nucleus neurons |
48
+ | GSE126954 | 10x 3' v2 | embryogenesis atlas |
49
 
50
+ The set grows as more datasets are curated.