README.md

---
license: mit
task_categories:
- tabular-regression
- reinforcement-learning
tags:
- protein-language-model
- biology
- enzymes
- gh114
- saprot
- dpo
pretty_name: GH114 Enzyme Activity (FLOP Benchmark)
size_categories:
- n<1k
base_model: NorseDrunkenSailor/Qwen_smol_GH114
---

# GH114 Enzyme Activity Dataset (from FLOP Benchmark)

## Dataset Description
This dataset contains protein sequences, SaProt structure-aware sequence, and activity labels for 55 **Glycoside Hydrolase Family 114 (GH114)** enzymes.

It is specifically formatted for aligning the **[NorseDrunkenSailor/Qwen_smol_GH114](https://huggingface.co/NorseDrunkenSailor/Qwen_smol_GH114)** base model. This model is a generative Causal Language Model (CLM) adapted from Qwen2 and trained on SaProt 3Di-aware sequences. The goal is to use **Direct Preference Optimization (DPO)** or other **Reinforcement Learning** techniques to align the model towards correctly ranking highly active GH114 variants across Cross-Validation (CV) folds. 
The CV folds are specifically curated in FLOP (https://www.biorxiv.org/content/10.1101/2023.06.21.545880v2.full.pdf) to prevent train/test leakage. This is considered a archetypal wildtype-discovery task. The median sequence identity is 0.46. The three CV folds contain 20, 18, and 17 sequences respectively.

- **Curated:** NorseDrunkenSailor
- **Original Source (Paper):** [FLOP: Tasks for Fitness Landscapes Of Protein wildtypes](https://github.com/petergroth/FLOP) and (https://www.biorxiv.org/content/10.1101/2023.06.21.545880v2.full.pdf) (Groth et al., 2024)
- **Original Source (Data):** Patent [WO/2019/228448, NOVOZYMES A/S](https://patentscope.wipo.int/search/en/detail.jsf?docId=WO2019228448)

## Biological Context & Assay
**Family GH114** (endo-alpha-1,4-polygalactosaminidases) are enzymes capable of degrading **PEL**, an exopolysaccharide that forms a critical structural component of biofilms in bacteria like *Pseudomonas aeruginosa*.
The 'target_reg' column represents the activity of the enzyme in concentration normalised units. 

**Note:** While seemingly small in size (N=55), this dataset represents high-quality ground truth derived from purified enzymes, intended for testing few-shot learning and low-data regime alignment capabilities. This is in fact a luxury and relatively large wild-type discovery dataset to start with. These are rare.

### The Assay
The `target_reg` column represents the **hydrolytic activity** of the enzyme against the PEL polysaccharide. 
- **Method:** The data comes from assays performed on **purified, concentration-normalized** natural enzymes (as opposed to crude lysates), ensuring the signal reflects true catalytic efficiency rather than expression levels.
- **Goal:** Disruption of PEL compromises biofilm integrity. High-activity variants are sought for industrial applications to remove biofilms (cleaning and detergent compositions).

## Data Fields
| Column | Type | Description |
| :--- | :--- | :--- |
| `name` | string | Sequence ID (SEQ ID from Patent/Paper). |
| `target_reg` | float | **Hydrolytic Activity**. Higher values indicate better degradation of PEL. |
| `seq` | string | **SaProt-formatted Sequence**. Input for the `Qwen_smol_GH114` tokenizer. Contains both amino acid and 3Di structural tokens. |
| `part_0` | int | Indicator (0/1) for FLOP Cross-Validation Fold 0. |
| `part_1` | int | Indicator (0/1) for FLOP Cross-Validation Fold 1. |
| `part_2` | int | Indicator (0/1) for FLOP Cross-Validation Fold 2. |

## Usage for Alignment (DPO/RL)
This dataset is designed to train a reward model or directly align the generator. The `seq` column allows the model to "read" the structure-aware sequence, while `target_reg` provides the ground truth signal for fitness.

```python
from datasets import load_dataset
dataset = load_dataset("NorseDrunkenSailor/GH114_activity_DPO")

# Example: Select high-activity variants for DPO positive pairs
high_activity = dataset.filter(lambda x: x['target_reg'] > 0.8)
```

Cross-Validation Strategy
This dataset uses the homology-stratified splitting defined in the FLOP paper.
Do NOT random split. Use part_0, part_1, part_2 columns.

Splits were generated using GraphPart to ensure that sequences in the test set share <55% identity with the training set, testing the model's ability to generalize to new subfamilies rather than memorizing homologs.

Citation

@article{groth2024flop,
  title={FLOP: Tasks for Fitness Landscapes Of Protein wildtypes},
  author={Groth, Peter M{\o}rch and Michael, Richard and Salomon, Jesper and Tian, Pengfei and Boomsma, Wouter},
  journal={bioRxiv},
  year={2024},
  doi={2023.06.21.545880v2}
}

@patent{WO2019228448,
  title={Polypeptides},
  author={Novozymes A/S},
  year={2019},
  number={WO/2019/228448}
}