Update README.md

README.md

@@ -19,73 +19,17 @@ used to model base-resolution binding signal and to discover sequence motifs via
 TF-MoDISco.
 
 ## Model details
-
-- **Architecture:** BPNet (`bpnet-lite` implementation), ? filters,
-  ? dilated convolutional layers, with bias tracks for both the
-  profile and total-count heads.
 - **Genome:** mm10
-- **Input window:** 2114 bp sequence -> **output window:**
-  1000 bp profile prediction
-- **Training data:** SOX2 ChIP-seq, MACS2 narrow peaks as positive loci,
-  with a matched negative set.
-  - Peaks (loci): `/scratch/prj/stem_cells_pituitary/Bence/atacseq_chip_results/26/bwa/merged_library/macs2/narrow_peak/SOX2_REP1.mLb.clN_peaks.bed`
-  - Signal (BAM): `exp26_SOX2.bw`
-  - Negatives: `/scratch/prj/stem_cells_pituitary/Bence/atacseq_chip_results/26/bwa/merged_library/macs2/narrow_peak/SOX2_negatives_bpnet.bed`
-- **Train/validation split (by chromosome):**
-  - Validation: ``
-  - Training: all other chromosomes (see `exp26_SOX2.bpnet.fit.json` for the
-    exact list)
-- **Training hyperparameters:** lr=None, batch_size=None,
-  max_epochs=None, early_stopping=None epochs,
-  max_jitter=None, reverse_complement_augmentation=None
 - **Checkpoint selection:** `exp26_SOX2.torch` is the best-validation-loss checkpoint saved
   during training by `bpnet-lite`'s `fit` routine (as opposed to a `.final.torch` checkpoint,
   which reflects the last training epoch and is not included here).
 
 ## Performance
 
-Reported metrics from `exp26_SOX2.performance.tsv`:
-
-```
-profile_mnll	profile_jsd	profile_pearson	profile_spearman	count_pearson	count_spearman	count_mse
-472.9859924316406	0.44859686493873596	0.20588907599449158	0.06264603137969971	0.39160212874412537	0.3859022855758667	0.6483629941940308
-```
+Reported metrics found in `exp26_SOX2.performance.tsv`.
 
 See `exp26_SOX2.evaluate.json` for the full evaluation output.
 
-## How to load
-
-This model was trained with `bpnet-lite` and is saved as a raw PyTorch `state_dict`-style
-`.torch` checkpoint (not a `transformers`-compatible format). To load it, install
-`bpnet-lite` and use its `BPNet` class directly:
-
-```bash
-pip install bpnet-lite
-```
-
-```python
-import torch
-from bpnetlite import BPNet
-
-# Re-create the architecture with the same hyperparameters used in training
-# (see exp26_SOX2.bpnet.fit.json for the full config)
-model = BPNet(
-    n_filters=64,
-    n_layers=8,
-    profile_output_bias=True,
-    count_output_bias=True,
-)
-
-state_dict = torch.load("exp26_SOX2.torch", map_location="cpu")
-model.load_state_dict(state_dict)
-model.eval()
-```
-
-> Double check the exact `BPNet(...)` constructor signature against the `bpnet-lite`
-> version you have installed, since argument names/defaults can change between releases.
-> Pin the same `bpnet-lite` version that produced this model if you need exact
-> reproducibility — see `exp26_SOX2.log` for environment details if recorded.
-
 ## Files
 
 | File | Description |
@@ -96,16 +40,3 @@ model.eval()
 | `exp26_SOX2.performance.tsv` | Reported performance metrics |
 | `exp26_SOX2.evaluate.json` | Full evaluation output |
 | `exp26_SOX2.log` | Training log |
-
-## Citation
-
-If you use this model, please cite the original BPNet paper and the `bpnet-lite`
-software:
-
-- Avsec et al., "Base-resolution models of transcription-factor binding reveal soft
-  motif syntax", *Nature Genetics*, 2021.
-- `bpnet-lite`: https://github.com/jmschrei/bpnet-lite
-
-## Repo
-
-`ge8rgia/BPNet_ChIP-seq_SOX2`