README.md

---
pretty_name: ProteinBase Interactions
license: mit
language:
  - en
tags:
  - protein-protein-interaction
  - binder-design
  - antibody
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.csv
---

# ProteinBase Interactions

## Overview

Each row represents a single binder-target pair with:

- the ProteinBase binder identifier
- the binder sequence
- the target name
- the target sequence
- an experimental binding-strength label
- a binary classification label
- design-class metadata
- an `antibody` flag indicating whether the binder is an antibody-derived design (`Nanobody` or `scFv`)

## Schema

The dataset contains the following columns:

- `proteinbase_id`: Original ProteinBase binder ID.
- `binder_name`: Binder name uploaded by the original protein designers.
- `target`: Common name of the target protein.
- `target_sequence`: Amino acid sequence of the target protein.
- `binder_sequence`: Amino acid sequence of the binder protein.
- `binding_strength`: Labels designated by Proteinbase. The possible values are `None`, `Weak`, `Medium`, and `Strong`.
- `label`: Binary benchmark label derived from `binding_strength`. `0` corresponds to `None`; `1` corresponds to `Weak`, `Medium`, or `Strong`.
- `value`: KD value for the binder-target pair, when KD measurements exist.
- `max_similarity_to_any_database`: Maximum sequence similarity to any reference database as extracted from the raw ProteinBase evaluations.
- `design_class`: Binder design class (e.g. scFv, miniprotein, peptide)
- `antibody`: `True` for `Nanobody` and `scFv` binders, `False` otherwise.

## Filtering

Rows are retained only if all of the following are true:

- the raw entry contains at least one target
- the binder is marked as expressed (poor expression yield does not indicate non-binding)

Recognized design classes are split into two groups:

- non-antibody binders:
  - `Miniprotein`
  - `Other`
  - `Peptide`
- antibody binders:
  - `Nanobody`
  - `scFv`

Additional similarity filtering is applied only to non-antibody binders:

- non-antibody rows are kept only if `max_similarity_to_any_database <= 50`
- antibody rows are retained regardless of this non-antibody threshold, and are marked by `antibody=True`

## Label Definition

The binary classification label is defined as:

- `label = 0` if `binding_strength == "None"`
- `label = 1` if `binding_strength` is `Weak`, `Medium`, or `Strong`

## Binding Affinity Value

The `value` column stores:

- the minimum KD value across available KD measurements for the corresponding binder-target pair

If no numeric KD is available, `value` is left blank.

## Raw Source Collections

This dataset is currently constructed from the following ProteinBase collections:

- `proteinbase_collection_adaptyv-egfr-competition-round-1.csv`
- `proteinbase_collection_adaptyv-egfr-competition-round-2.csv`
- `proteinbase_collection_adaptyv-x-muni-hackathon-ai-agents-vs-humans.csv`
- `proteinbase_collection_bindcraft1-revalidation.csv`
- `proteinbase_collection_boltzgen-release.csv`
- `proteinbase_collection_boolean-biotech-vhh-competition-2025.csv`
- `proteinbase_collection_cradle-egfr-competition.csv`
- `proteinbase_collection_dsm-round-1.csv`
- `proteinbase_collection_egfr-round1-second-submission.csv`
- `proteinbase_collection_evodiff-validation.csv`
- `proteinbase_collection_evolved-2024-bio-x-ml-team-silica-egfr-nanobodies.csv`
- `proteinbase_collection_evolved-hackathon.csv`
- `proteinbase_collection_gem-x-adaptyv-rbx1-binder-design-competition-results.csv`
- `proteinbase_collection_mog-dfm-spotlight.csv`
- `proteinbase_collection_mosaic-development.csv`
- `proteinbase_collection_mosaic-multispecifics.csv`
- `proteinbase_collection_nipah-binder-competition-results.csv`
- `proteinbase_collection_pd-l1-foldcraft.csv`
- `proteinbase_collection_pro-1-validation.csv`
- `proteinbase_collection_protrl-validation.csv`
- `proteinbase_collection_rfdiffusion-re-validation.csv`