Zipoop 32k × 2,847 RNA 教师(021)
Per-window teacher networks that compress 2,847 ENCODE RNA-seq tracks (GRCh38) over 32,768 bp windows at 1 bp resolution. 89,436 windows, stored as one <chrom>.npz per chromosome (25 files, ~126 GB).
This is the 021 batch: total RNA-seq 1,285, polyA plus 1,434, polyA minus 128. Each window has its own teacher network; nothing is shared across windows. The artifact is meant to be decoded (or fine-tuned further) segment by segment, not to predict signal from DNA.
What a window contains
| member | meaning |
|---|---|
dna |
(32768,) uint8, 0–3 = A/C/G/T, 4 = N — the model input |
parameter_0…143 + parameter_names |
trained network weights (144 arrays, float16) |
model_config |
width 48, rank 48, 2 transformer layers, 4 heads, 7 pooling levels, max pooling, sinusoidal positions |
mean, power, gain, offset |
per-track target scaling and calibration |
fft_indices, fft_packed, fft_scale |
sparse FFT base: top-54 rfft bins per track, float16 with one float32 scale per track |
final_gain, final_offset |
per-track affine calibration to real units |
valid_v2, valid_encoding |
missing-data mask, wire format track-rle-lzma-v2 |
metadata |
the 2,847 ENCODE track records |
format, hybrid_format, storage_encoding, byte_shuffle |
format tags and byte-shuffle index |
164 members per window. Reconstruction is signal = irfft(sparse FFT) + DNA-conditioned residual, then × final_gain + final_offset; missing positions follow the mask, and RNA-seq tracks are 0 and valid where the bigWig did not record.
Quality of this batch
Measured on the production receipts over all 89,436 trained windows (per-track 1 bp Pearson against the real ENCODE signal):
mean Pearson 0.9727 median 0.9746 p05 0.9429 min 0.7256
primary 52.5% secondary 42.4% no_signal 5.1% (40.3% hit the 1,000-step cap)
A 0.90 mean Pearson is the "usable" label; secondary means the step cap was reached before the 0.95 target. Pearson-optimal weights do not guarantee absolute amplitude — that is what the per-track affine calibration is for.
What it is not
Not a genome-wide predictor and not a model that generalises across windows. The teacher is fitted independently on each window; transplanting one window's archive onto unseen windows was measured on the level-2 (128 kb) batch at per-track Pearson 0.051 as-is and 0.061 with the sparse coefficients zeroed, NMSE ≈ 0.99, versus 0.818 on the window it was trained on. Treat each window as its own codec.
Reading it
zipoop 0.2.x reads the pooper-dna-v4 / fft-plus-dna-residual-v2 members but does not apply the per-track calibration offsets in the way the training code does; the reference reader is the pooper package:
import io, json, numpy as np
from pooper.teacher_batch import open_shard # logical arrays, byte-shuffle undone
from pooper.hybrid import archive_decoder # applies final_gain/offset
with open_shard("/path/to/chr1.npz::chr1_0_32768") as z:
dna, valid = np.asarray(z["dna"]), np.asarray(z["valid_v2"])
sig = archive_decoder("/path/to/chr1_0_32768.npz").decode(missing_nan=False) # (32768, 2847)
If you only have this merged file, extract one window's members into a standalone npz first (member names are <seg>/<key>; strip the <seg>/ prefix), then decode. Because the archival arrays are byte-shuffled, the standalone archive must be written back with dtype/shape restored from the byte_shuffle index, fft_coefficients = fft_packed * fft_scale[..., 0], format=pooper-dna-v3, hybrid_format=fft-plus-dna-residual-v1.
Layout
<chrom>.npz— zip64/LZMA with members<seg>/<key>(164 per window), windows in genome order; window names<chrom>_<start>_<end>.MANIFEST.tsv— chrom, segments, members, bytes, sha256 per file.tracks.tsv— the 2,847 tracks (index, ENCODE accession, modality, assay, output type, replicates, experiment).
License and attribution
This dataset is released under CC BY 4.0. You may use, copy, modify and redistribute it, including commercially and including for training or distilling models, provided you give credit:
Zipoop teacher shards 021, by zatumout (2026), derived from ENCODE data. Licensed under CC BY 4.0.
The shards are derived from public ENCODE signal tracks. The ENCODE Consortium asks users of ENCODE data to cite the Consortium's most recent integrative publication, reference the ENCODE accession numbers (listed per track in the shard metadata and in the panel file) and acknowledge the ENCODE Consortium and the production laboratories (see https://www.encodeproject.org/help/citing-encode).
- Downloads last month
- 67