---
dataset_info:
  features:
  - name: seqs
    dtype: string
  - name: labels
    dtype: int64
  splits:
  - name: train
    num_bytes: 72141633
    num_examples: 154336
  - name: valid
    num_bytes: 2367747
    num_examples: 5000
  - name: test
    num_bytes: 2284093
    num_examples: 5000
  download_size: 75804271
  dataset_size: 76793473
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: valid
    path: data/valid-*
  - split: test
    path: data/test-*
---

Dataset of proteins that homodimerize (1) and randomly selected proteins from StringDB (0). Data splits are random, but the balance is roughly 90:10. The main assumption is that most proteins do not dimerize with themselves.

Homodimers are selected from [here](https://seq2fun.dcmb.med.umich.edu/HomodimerDB/download.cgi) via the 80% identity clustering.

```
@misc {SynthyraHomodimerBenchmark,
	author       = { Hallee, L. and Bichara, D. and Gleghorn, J, P. },
	title        = { Synthyra Homodimer Benchmark },
	year         = 2024,
	url          = { https://huggingface.co/datasets/Synthyra/homodimer_benchmark },
	doi          = { 10.57967/hf/3798 },
	publisher    = { Hugging Face }
}
```