Data_Preparation

LAB2_project - Data Preparation

This stage of the project focuses on generating non-redundant and unbiased datasets that can be used for training and benchmarking machine learning models. The process involves clustering, representative selection, TSV filtering, and final dataset splitting.

---

Workflow Details

1. Clustering of protein sequences

Positive and negative datasets were clustered independently using MMseqs2 in order to reduce redundancy and ensure that no highly similar proteins appear in both training and benchmarking sets.

Main outputs:

• pos_cluster-results_cluster.tsv / neg_cluster-results_cluster.tsv
  Tables mapping each input sequence to its cluster representative.
• pos-cluster-results_rep_seq.fasta / neg-cluster-results_rep_seq.fasta
  FASTA files containing only the representative sequences (one per cluster).

---

2. Extraction of representative IDs

Representative identifiers were retrieved directly from the FASTA headers of the MMseqs2 output.

Generated files:

• rep_positive.ids → list of representative IDs for the positive dataset
• rep_negative.ids → list of representative IDs for the negative dataset

---

3. Filtering of original TSV files

The script get_tsv.py takes the representative ID lists (.ids) and the cluster mapping tables (*_cluster.tsv) as input. It produces reduced TSV files that contain only the representative entries.

Outputs:

• pos_cluster_results.tsv → positive dataset, reduced to representatives
• neg_cluster_results.tsv → negative dataset, reduced to representatives

These files serve as the non-redundant reference datasets for downstream analysis.

---

4. Splitting into training and benchmarking sets

The script get_sets.py takes the representative ID lists (.ids) as input and generates:

• Training set (80%)
  Used for model training and hyperparameter tuning.
  Within this set, each sequence is also assigned to one of five cross-validation folds, preserving the positive/negative ratio.

• Benchmarking set (20%)
  Held out and never used during model training, providing an unbiased evaluation of generalization performance.

Outputs in Cross_Validation/:

• pos_train.tsv and neg_train.tsv → training data for positive and negative classes, including fold assignments
• pos_bench.tsv and neg_bench.tsv → benchmarking data for positive and negative classes

---

Reproducibility

All data preparation steps (clustering, representative selection, TSV filtering, and train/benchmark splitting)
are automated in the script clustering.sh:

bash clustering.sh

---

About MMseqs2

MMseqs2 (Many-against-Many sequence searching) was used in this project to:

• Cluster sequences based on ≥30% identity and ≥40% alignment coverage
• Select a single representative per cluster
• Prevent redundancy and overlap between subsets

These thresholds are specifically chosen to minimize the risk of data leakage, which occurs if similar sequences are present in both training and benchmarking datasets.

Official repository: soedinglab/MMseqs2

---

Current Results

After clustering and representative selection:

Dataset	Total sequences	Representative sequences	File containing representatives
Positive	2,949	1093	POS CLUSTER TSV
Negative	20,615	8934	NEG CLUSTER TSV

After splitting with get_sets.py:

Subset	Positive	Negative	Files generated
Training (80%)	874	7147	POS TRAIN TSV, NEG TRAIN TSV
Benchmarking (20%)	219	1787	POS BENCH TSV, NEG BENCH TSV

Each sequence in the training set is also annotated with a cross-validation fold index (1–5).

---

Next step → The non-redundant training and benchmarking sets generated in this stage
are used as input for the Data_Analysis module, where exploratory data analysis (EDA) is performed to assess dataset quality, visualize protein length distributions, and identify conserved sequence motifs before proceeding to feature extraction and modeling.