pooper-b8 (407M, 128 kb context)
A ~407M-parameter model that predicts 1,885 ENCODE signal tracks (ATAC-seq, DNase-seq, histone ChIP-seq, RNA-seq; GRCh38) at 1 bp resolution from a 131,072 bp DNA window. It uses an AlphaGenome-style backbone trained from random initialisation (no AlphaGenome pretrained weights).
b8 is a sibling of pooper-b6: per the b6 reproduction notes the two runs differ only in --level-weight (b8 uses 1.0). Architecture, data and schedule are otherwise the same as b6, see the b6 repository for the full recipe and reproduction package. A separate b8 training log was not kept with this checkpoint.
Training supervision comes from the Zipoop teacher shards: one over-fitted small "teacher" network per 32,768 bp window, each compressing that window's 1,885 tracks.
Results
Official-style evaluation: per-track Pearson pooled over all positions of 400 windows. ATAC / DNase / RNA are scored at 1 bp, histone ChIP at 128 bp bins. Checkpoint at step 20,000.
| Assay | Tracks | fold0 valid: Pearson / log1p | fold1 test: Pearson / log1p |
|---|---|---|---|
| ATAC | 200 | 0.716 / 0.683 | 0.728 / 0.690 |
| DNase | 200 | 0.757 / 0.758 | 0.778 / 0.775 |
| RNA-seq | 1,285 | 0.424 / 0.547 | 0.358 / 0.549 |
| ChIP (histone) | 200 | 0.618 / 0.491 | 0.630 / 0.497 |
The validation and test splits follow AlphaGenome's fold0 convention only to keep evaluation comparable. This card makes no claim of matching or beating AlphaGenome: the track sets differ, and no same-track comparison was done.
Files
pooper_genomeb8_407M_128K.pt: checkpoint (1,814,125,498 bytes).
Limitations
- RNA-seq is the weakest assay, with a drop from validation to test.
- Research prototype, not a validated tool for clinical or variant-effect use.
- The model was trained against Zipoop teacher shards; this card does not claim that scores reflect agreement with raw ENCODE signal.
- Single seed; the evaluation uses 400 windows per split.