https://github.com/zozo123/t2t-vs-grch38
https://github.com/zozo123/t2t-vs-grch38
Last synced: 11 days ago
JSON representation
- Host: GitHub
- URL: https://github.com/zozo123/t2t-vs-grch38
- Owner: zozo123
- Created: 2026-06-14T16:01:23.000Z (20 days ago)
- Default Branch: main
- Last Pushed: 2026-06-14T16:11:59.000Z (20 days ago)
- Last Synced: 2026-06-14T18:09:16.451Z (20 days ago)
- Language: Shell
- Size: 362 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# What T2T changed in the human reference
**A genome-wide GRCh38 -> T2T-CHM13 delta, computed by Claude Code with [OpenClaw Crabbox](https://github.com/openclaw/crabbox) on [islo.dev](https://islo.dev).**
Live site: https://zozo123.github.io/t2t-vs-grch38/
This is the follow-up to the reference-broadcast demo:
https://zozo123.github.io/genomics-sandboxes/
## Result
T2T-CHM13 closes many gaps in GRCh38. Running the same CpG-island kernel across both references shows:
| metric | value |
|---|---:|
| Newly usable sequence in T2T | 179.6 Mb |
| GRCh38 gap sequence removed | 150.6 Mb |
| GRCh38 candidate CpG islands | 264,816 |
| T2T candidate CpG islands | 305,308 |
| Extra candidate islands | 40,492 |
| Candidate-island density in added sequence | 225.4/Mb |
The added sequence is not one signal:
- chrY and chr9 add large amounts of sequence with little candidate-island gain, consistent with newly resolved heterochromatin and repeat-rich sequence.
- chr13, chr15, chr21, and chr22 gain many candidate islands, consistent with acrocentric/rDNA-rich sequence and with the known tendency of simple CpG-island rules to over-call GC/CpG-rich repeats.
This is not framed as a novel biological discovery. It is a compact, reproducible measurement of how completing the reference changes what a transparent sequence-composition caller sees.
## Harness
The harness is a Claude Code agent driving OpenClaw Crabbox:
```bash
export ISLO_API_KEY=...
./crabbox.sh t2t
```
The harness follows the official Crabbox Islo-provider model:
1. `crabbox run --provider islo --keep --lease-output ...` warms an Islo sandbox with UCSC `hs1.fa.gz` (T2T-CHM13v2.0), splits it by chromosome, and keeps the sandbox.
2. The harness saves that kept sandbox as an Islo snapshot, because the paper is explicitly testing snapshot broadcast.
3. Each chromosome runs through `crabbox run --provider islo --islo-snapshot-name ...`, so Crabbox owns repo sync, guardrails, timing, and run lifecycle while Islo owns sandbox state and streaming exec.
4. The reducer merges T2T per-chromosome JSON against the GRCh38 genome-wide receipts.
If `ISLO_API_KEY` / `CRABBOX_ISLO_API_KEY` is missing, the script prints a warning and falls back to the direct Islo CLI path used to generate the published receipts. The preferred path is OpenClaw Crabbox.
Measured run:
| phase | time |
|---|---:|
| T2T warm-up | 77.6 s |
| snapshot save | 15.0 s |
| 24-way fan-out | 60.0 s |
| snapshot size | 1009.5 MB |
## Files
| file | purpose |
|---|---|
| `index.html`, `styles.css`, `script.js` | static GitHub Pages paper |
| `data/compare.json` | comparison receipts used by the figures |
| `crabbox.sh` | genome-wide OpenClaw Crabbox harness with `run` and `t2t` modes |
| `scripts/t2t_warmup.sh` | T2T warm-up: download, split, index |
| `scripts/compute.py` | per-chromosome MAP kernel |
| `scripts/reduce_compare.py` | GRCh38 vs T2T reducer |
`crabbox run --provider islo` requires `ISLO_API_KEY`. New islo.dev accounts can use coupon `YOSSI150` for 150 free credits.
## Caveats
The CpG-island calls are candidates from a simple sequence-composition rule. The 1987 Gardiner-Garden and Frommer criterion is transparent and useful for this demonstration, but it is known to over-call in some GC-rich repeat contexts. The result should be read as a reference-composition comparison, not as promoter annotation, methylation measurement, or medical inference.