Kavlahkaff commited on
Commit
245e1e6
·
verified ·
1 Parent(s): fe04f4b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +221 -0
README.md CHANGED
@@ -1,3 +1,224 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ task_categories:
4
+ - tabular-regression
5
+ tags:
6
+ - hyperparameter-optimization
7
+ - automl
8
+ - bioinformatics
9
+ - multi-omics
10
+ - autoencoder
11
+ - benchmark
12
+ size_categories:
13
+ - 100K<n<1M
14
  ---
15
+
16
+ # BBOmix: Raw HPO Data
17
+
18
+ BBOmix is a large-scale tabular benchmark for hyperparameter optimization (HPO) of
19
+ unsupervised autoencoder (AE) representation learning on real-world biological
20
+ multi-omics data. This dataset contains the **raw, per-run evaluation records**
21
+ underlying the benchmark: 105,000 individual training runs spanning 5 AE
22
+ architectures, 7 dataset-modality tasks, 1,000 sampled hyperparameter
23
+ configurations per architecture, and 3 random seeds.
24
+
25
+ This collection is the unprocessed source data. It was used to build the
26
+ serialized tabular blackboxes used in the [BBOmix paper](https://arxiv.org/abs/2606.05139). If you want to run HPO
27
+ optimizers against the benchmark directly, use [Syne Tune](https://github.com/syne-tune/syne-tune). Use this raw data if you want to
28
+ build your own surrogate/analysis pipeline.
29
+
30
+ **Total size:** ~1 GB (uncompressed JSON)
31
+ **Total files:** 105,000 JSON files (one per training run)
32
+
33
+ The files are now stored as 4 separate ZIP archives based on their architecture: `vanillix.zip`, `varix.zip`, `disentanglix.zip`, and `ontix.zip`.
34
+
35
+ ## Downloading and using the data
36
+
37
+ The raw JSON files are packaged into four zip archives based on their architecture, hosted on Hugging Face at `autoencodix/BBOmix`:
38
+ - `vanillix.zip`
39
+ - `varix.zip`
40
+ - `disentanglix.zip`
41
+ - `ontix.zip`
42
+
43
+ You can download them using the `huggingface_hub` Python library. For example, to download and extract the `varix` data:
44
+
45
+ ```python
46
+ from huggingface_hub import hf_hub_download
47
+ import zipfile
48
+
49
+ # Download the zip file
50
+ file_path = hf_hub_download(repo_id="autoencodix/BBOmix", filename="varix.zip", repo_type="dataset")
51
+
52
+ # Extract to your desired directory
53
+ with zipfile.ZipFile(file_path, 'r') as zip_ref:
54
+ zip_ref.extractall("bbomix_raw/")
55
+ ```
56
+
57
+ Alternatively, you can use the Hugging Face CLI:
58
+
59
+ ```bash
60
+ huggingface-cli download autoencodix/BBOmix varix.zip --repo-type dataset
61
+ unzip varix.zip -d raw_hpo_data/
62
+ ```
63
+
64
+ Once extracted, you will have access to the individual JSON files, which can be loaded and parsed using standard tools (e.g., Python's `json` or `pandas` libraries).
65
+
66
+ ## What's in the benchmark
67
+
68
+ - **5 AE architectures:** Vanillix (standard AE), Varix (β-VAE), Disentanglix
69
+ (β-TCVAE-style disentangled VAE), and two Ontix variants (ontology-constrained VAE, evaluated
70
+ with two ontology variants: `chromosome` and `reactome`)
71
+ - **2 source datasets:** TCGA (The Cancer Genome Atlas, bulk multi-omics,
72
+ pan-cancer) and SCHC (Single-Cell Human Cortex, single-cell multi-omics)
73
+ - **7 dataset-modality tasks:** three TCGA modalities (RNA, EPIGENETIC, DNA) and their
74
+ combinations, plus two SCHC modalities (RNA, EPIGENETIC) and their combination —
75
+ each task always includes the relevant clinical/annotation labels (`CLIN`)
76
+ used for downstream evaluation
77
+ - **1,000 hyperparameter configurations** sampled uniformly at random per
78
+ architecture
79
+ - **3 random seeds** per configuration to account for stochastic training
80
+ variation
81
+ - **300 training epochs** per run, with reconstruction loss logged at every
82
+ epoch to support multi-fidelity / early-stopping HPO methods
83
+
84
+ Each configuration was trained on real biological data and evaluated both by
85
+ its unsupervised reconstruction loss and by downstream supervised task
86
+ performance (logistic regression AUC-ROC on frozen embeddings), enabling
87
+ analysis of how well reconstruction loss serves as a cheap proxy for
88
+ downstream utility — see the paper for details.
89
+
90
+ ## Directory / file naming
91
+
92
+ Each JSON file corresponds to a single training run. Files are named/grouped
93
+ by architecture, dataset, modality combination, seed, and hyperparameter
94
+ configuration index, following the `RUN_ID` field inside the file itself, e.g.:
95
+
96
+ ```
97
+ disentanglix_schc_1_METH_CLIN_hp0.json
98
+ ```
99
+
100
+ reads as: architecture = `disentanglix`, dataset = `schc`, seed = `1`,
101
+ modalities = `METH` + `CLIN`, hyperparameter configuration index = `hp0`.
102
+
103
+ For Ontix runs, the ontology variant (`chromosome` or `reactome`) is also
104
+ encoded in the filename/RUN_ID, since the two ontologies define distinct
105
+ decoder structures and are treated as separate architecture variants.
106
+
107
+ ## JSON schema
108
+
109
+ Every file has the same top-level schema:
110
+
111
+ | Field | Type | Description |
112
+ |---|---|---|
113
+ | `RUN_ID` | string | Unique identifier for this run (see naming convention above). |
114
+ | `ARCHITECTURE` | string | One of `vanillix`, `varix`, `disentanglix`, `ontix`. |
115
+ | `SEED` | int | Random seed used for this run (`0`, `1`, or `2`). |
116
+ | `DATASET` | string | Source dataset: `tcga` or `schc`. |
117
+ | `MODALITIES` | list[string] | Data modalities used as input, e.g. `["METH", "CLIN"]`. `CLIN` denotes the clinical/annotation labels used for downstream evaluation, not a training input modality. |
118
+ | `ONTOLOGY` | string or null | For Ontix runs: `"chromosome"` or `"reactome"`. `null` for all other architectures. |
119
+ | `HYPERPARAMETERS` | dict | The sampled hyperparameter configuration for this run. Keys vary by architecture — see [Hyperparameter search spaces](#hyperparameter-search-spaces) below. |
120
+ | `AVG_ML_TASK_PERFORMANCE` | float | Downstream performance (AUC-ROC), averaged across all downstream tasks defined for this dataset/modality combination. This is the paper's primary "downstream utility" metric. |
121
+ | `PER_TASK_PERFORMANCE` | dict | Downstream AUC-ROC broken down per individual clinical/annotation task (e.g. `sex`, `age_group`, cancer subtype, survival status — task names depend on `DATASET`; see Appendix A of the paper for the full list per dataset). |
122
+ | `VALID_RECON_LOSS` | float | Final validation-set reconstruction loss (or ELBO/total loss for VAE-style architectures) at the last logged epoch. |
123
+ | `loss_per_epoch` | dict[str, float] | Per-epoch training/validation loss curve, keyed by epoch index (`"0"` through `"299"`) as strings. Enables multi-fidelity and early-stopping HPO methods (e.g. ASHA, BOHB, Hyperband) to query intermediate fidelities without retraining. |
124
+ | `RUNTIME_SECONDS` | float | Wall-clock training time for this run in seconds. |
125
+
126
+ ### Hyperparameter search spaces
127
+
128
+ The keys present in `HYPERPARAMETERS` depend on `ARCHITECTURE`:
129
+
130
+ - **Shared across all architectures:** `k_filter` (input filter/feature size),
131
+ `n_layers` (number of encoder layers), `enc_factor` (encoding factor),
132
+ `learning_rate`, `batch_size`, `drop_p` (dropout probability), `weight_decay`,
133
+ `epochs`, `checkpoint_interval`, `loss_reduction`
134
+ - **Vanillix:** shared parameters only, plus `latent_dim`
135
+ - **Varix / Ontix:** adds `beta` (KL-divergence weight)
136
+ - **Disentanglix:** adds `beta_mi` (mutual information weight), `beta_tc`
137
+ (total-correlation weight), `beta_dimKL` (per-dimension KL weight)
138
+ - **Ontix:** `latent_dim` is architecture-defined by the chosen ontology rather
139
+ than sampled (see `ONTOLOGY` field)
140
+
141
+ See Table 1 in the paper for the exact sampling distributions (uniform vs.
142
+ log-uniform vs. categorical) for every hyperparameter.
143
+
144
+ ## Example record
145
+
146
+ ```json
147
+ {
148
+ "RUN_ID": "disentanglix_schc_1_METH_CLIN_hp0",
149
+ "ARCHITECTURE": "disentanglix",
150
+ "SEED": 1,
151
+ "DATASET": "schc",
152
+ "MODALITIES": ["METH", "CLIN"],
153
+ "ONTOLOGY": null,
154
+ "HYPERPARAMETERS": {
155
+ "epochs": 300,
156
+ "checkpoint_interval": 300,
157
+ "loss_reduction": "sum",
158
+ "k_filter": 4096,
159
+ "n_layers": 2,
160
+ "enc_factor": 1,
161
+ "learning_rate": 0.009251283695699343,
162
+ "weight_decay": 9.54041826560807e-05,
163
+ "batch_size": 64,
164
+ "drop_p": 0.6628240927476112,
165
+ "beta_mi": 0.0509093966079994,
166
+ "beta_tc": 288.99999718361437,
167
+ "beta_dimKL": 0.00022271800642020435,
168
+ "latent_dim": 16
169
+ },
170
+ "AVG_ML_TASK_PERFORMANCE": 0.6147933697669551,
171
+ "PER_TASK_PERFORMANCE": {
172
+ "author_cell_type": 0.7042096421209849,
173
+ "age_group": 0.5984282088126611,
174
+ "sex": 0.5417422583672192
175
+ },
176
+ "VALID_RECON_LOSS": 92.06772205753184,
177
+ "loss_per_epoch": {
178
+ "0": 139.4601534783552,
179
+ "1": 132.60988325103102,
180
+ "...": "...",
181
+ "299": 92.06772205753184
182
+ },
183
+ "RUNTIME_SECONDS": 1571.6251
184
+ }
185
+ ```
186
+
187
+ ## Intended use
188
+
189
+ This raw data is intended for:
190
+
191
+ - Building custom tabular/surrogate HPO benchmarks beyond the provided Syne
192
+ Tune blackboxes (e.g., different aggregation of seeds, custom multi-fidelity
193
+ schedules, alternative surrogates for off-grid queries)
194
+ - Studying the relationship between reconstruction loss and downstream
195
+ biological utility across architectures and modalities
196
+ - Hyperparameter importance analysis (e.g., permutation importance, fANOVA)
197
+ using the full configuration-to-performance mapping
198
+ - Meta-learning / transfer-learning research across tasks, architectures, and
199
+ modalities, using the full per-run records rather than aggregated summaries
200
+
201
+ This dataset does **not** contain raw omics measurements (e.g. sequencing
202
+ reads, methylation beta values, mutation calls) — only the evaluation results
203
+ of models trained on that data. The underlying TCGA and SCHC data are public
204
+ and de-identified; see the paper for their original sources.
205
+
206
+ ## Citation
207
+
208
+ If you use this dataset, please cite the BBOmix paper:
209
+
210
+ ```
211
+ @misc{thalebombien2026bbomixtabularbenchmarkhyperparameter,
212
+ title={BBOmix: A Tabular Benchmark for Hyperparameter Optimization of Unsupervised Biological Representation Learning},
213
+ author={Luca Thale-Bombien and Jan Ewald and Ralf König and Aaron Klein},
214
+ year={2026},
215
+ eprint={2606.05139},
216
+ archivePrefix={arXiv},
217
+ primaryClass={cs.LG},
218
+ url={https://arxiv.org/abs/2606.05139},
219
+ }
220
+ ```
221
+
222
+ ## License
223
+
224
+ Released under the Apache 2.0 license.