Update README.md

README.md

@@ -37,7 +37,6 @@ The dataset can be used to train or benchmark other machine learning models for
 This dataset reflects in vitro measurements and may not generalize to in vivo conditions. It should not be used as a sole basis for clinical or industrial enzyme selection without additional experimental validation.
 
 ## Dataset Structure
-
 The repository contains:
 - datasets/ – CSV files for *k*cat, *K*m, and *K*i with train/test splits
 - scripts/ – Preprocessing and utility scripts
@@ -50,6 +49,52 @@ Each entry typically includes:
 
 Data was compiled and curated from public biochemical databases, including BRENDA and SABIO-RK, as described in the CatPred publication. Splits were designed to evaluate generalization to enzyme sequences dissimilar to those seen during training.
 
+## Dataset Splits
+Each kinetic parameter (kcat, km, ki) has two split strategies, described below.
+
+### Split Strategies
+
+**Random splits** divide the data without regard to sequence similarity. These are useful
+for a general baseline but tend to overestimate real-world model performance, since
+training and test enzymes may be closely related.
+
+**Sequence-similarity splits** (`seq_test_sequence_XXcluster`) ensure that test set enzymes
+share less than XX% sequence identity with any enzyme in the training set. This is the
+more rigorous benchmark — a model that performs well here is genuinely generalizing to
+novel enzymes rather than recognizing similar sequences it has effectively seen before.
+
+Five strictness levels are provided:
+| Cluster threshold | Test set stringency |
+| --- | --- |
+| 20% | Hardest — test enzymes are very dissimilar to training data |
+| 40% | Hard |
+| 60% | Moderate |
+| 80% | Easy |
+| 99% | Easiest — nearly any sequence may appear in test |
+
+### File Naming
+
+Each split file is named `{parameter}-{strategy}_{subset}.csv`. The subsets are:
+
+| Subset | Contents | When to use |
+|---|---|---|
+| `train` | Training data only | Model development and hyperparameter tuning |
+| `val` | Validation data only | Monitoring training, early stopping |
+| `test` | Test data only | Final benchmark evaluation |
+| `trainval` | Train + val combined | Retrain final model after hyperparameters are locked in |
+| `trainvaltest` | All data combined | Train a release model once all evaluation is complete |
+
+### Recommended Workflow
+```
+train + val         →   tune architecture and hyperparameters
+trainval + test     →   final benchmark (report results here)
+trainvaltest        →   train the final released model on all available data
+```
+
+This three-stage approach is standard practice in ML: you only touch the test set once,
+and the combined files make it easy to retrain on progressively more data as you move
+from experimentation to deployment.
+
 ## Citation
 
 If you use this dataset, please cite: