ukaikotaro/UNKAI-dataset

UNKAI Protein Pair Dataset

This repository contains protein-pair datasets used for UNKAI, a binary classification model that predicts whether two proteins are associated with the same enzymatic reaction.

Three dataset variants are provided:

• original
• seen_unseen
• strict

Each dataset is divided into training, validation, and test splits.

Data format

Each TSV file contains two columns:

pair    label
A0RQT4_C1EZA3    1

pair

A pair of protein accession IDs represented as:

PROTEIN1_PROTEIN2

label

Binary classification label:

1 = same enzymatic reaction
0 = different enzymatic reaction

Repository structure

original/
├── train.tsv
├── validation.tsv
└── test.tsv

seen_unseen/
├── train.tsv
├── validation.tsv
└── test.tsv

strict/
├── train.tsv
├── validation.tsv
└── test.tsv

Dataset variants

Original

The original dataset uses a pair-level random split.

The positive and negative classes are balanced within each split.

Split	Total	Positive	Negative
Train	140,000	70,000	70,000
Validation	30,000	15,000	15,000
Test	30,000	15,000	15,000

No exact protein-pair duplicates are shared between train, validation, and test.

This split provides the least restrictive evaluation setting among the three released variants.

Seen-unseen

The seen-unseen dataset was constructed using protein clusters.

Protein clusters are first assigned to train, validation, or test groups.

Pairs are then assigned according to the following rules:

Training

Both proteins must belong to clusters assigned to training.

train cluster + train cluster

Validation

Exactly one protein belongs to a training cluster and the other belongs to a validation cluster.

train cluster + validation cluster

Test

Exactly one protein belongs to a training cluster and the other belongs to a test cluster.

train cluster + test cluster

Therefore, each validation/test pair contains one side from a cluster distribution observed during training and one side from an unseen cluster distribution.

The released dataset contains:

Split	Total	Positive	Negative
Train	73,463	37,765	35,698
Validation	10,000	5,000	5,000
Test	10,000	5,000	5,000

No exact protein-pair duplicates are shared between train, validation, and test.

Same-cluster pairs are excluded from the released seen-unseen setting because validation and test pairs necessarily consist of proteins from two differently assigned clusters.

Exact protein accessions from training may appear on the seen side of validation/test pairs by design.

Strict

The strict dataset provides the strongest cluster-separation setting.

Protein clusters are first divided into mutually exclusive train, validation, and test groups.

A protein pair is included in a split only when both proteins belong to clusters assigned to that same split.

Conceptually:

Train:
train cluster + train cluster

Validation:
validation cluster + validation cluster

Test:
test cluster + test cluster

Therefore, clusters used in validation and test are not present in the training cluster set.

The released strict dataset contains:

Split	Total	Positive	Negative
Train	67,699	32,236	35,463
Validation	10,000	5,000	5,000
Test	10,000	5,000	5,000

No exact protein-pair duplicates are shared between train, validation, and test.

Sampling controls

The cluster-based datasets were constructed with additional sampling controls to reduce excessive representation of individual proteins or clusters.

Reverse protein pairs are treated as identical:

A_B == B_A

This is appropriate for UNKAI because its pair representation is based on:

|v1 - v2|

which is symmetric with respect to protein order.

During final dataset selection, the following frequency limits were used:

Constraint	Maximum
Pairs per protein accession	10
Pairs per cluster	80
Pairs per cluster-pair per label	15

During candidate collection, at most 40 examples per cluster-pair per label were retained before final sampling.

Cluster split

Protein clusters were randomly divided using seed 42.

Approximately:

80% training clusters
10% validation clusters
10% test clusters

were assigned before constructing the strict and seen-unseen datasets.

Pair overlap verification

The released TSV files were independently checked for exact pair overlap between splits.

For all three dataset variants:

train ∩ validation = 0
train ∩ test       = 0
validation ∩ test  = 0

Relationship between the datasets

The three datasets are intended to represent increasingly challenging generalization settings.

Original
   |
   | pair-level split
   v
Seen-unseen
   |
   | one unseen cluster side
   v
Strict
   |
   | completely cluster-separated
   v
stronger distribution shift

The strict dataset is therefore intended to provide a more conservative estimate of generalization to proteins from clusters not represented during training.

Models

Pretrained UNKAI checkpoints are available separately:

ukaikotaro/UNKAI

Source code and inference utilities:

https://github.com/ukai3313/UNKAI

Limitations

The labels describe enzymatic-reaction association according to the source data used to construct the protein pairs.

The datasets should not be interpreted as a complete representation of all protein functions or all enzyme reaction relationships.

Results obtained using different splitting strategies are not directly interchangeable. In particular, random pair-level evaluation can be substantially easier than evaluation under cluster separation.

Citation

Citation information for the associated publication will be added here.

License

MIT License.