AbhijeetKumar2125/da3410-ec-retrieval-v7

DA3410 – Leak CURBER Enzyme Classification (Retrieval v7)

Overview

A retrieval-based enzyme classification system using MMseqs2 sequence similarity search against a reference sequence database.

Method

1. Sequence retrieval: MMseqs2 searches each query against the reference database.
2. Similarity-weighted voting: Retrieved hits contribute votes weighted by alignment score, sequence identity, query and target coverage, and retrieval rank.
3. Adaptive nearest-hit anchoring: The strongest retrieved neighbour receives an adaptive contribution based on identity and coverage.
4. Hierarchical EC voting: Votes are aggregated through EC hierarchy levels (1-, 2-, 3-, and 4-level prefixes). The most supported branch is selected at each level.
5. Valid-label constraint: Predictions are restricted to fully specified EC labels in the supplied vocabulary.

Repository contents

• predict.py – inference and hierarchical voting implementation.
• inputs.py – input loading, validation, and sequence deduplication.
• task.json – task configuration.
• ec_vocabulary.json – valid EC vocabulary.
• full_retrieval_train_labels.tsv – reference target-to-EC annotations.
• mmseqs_db/ – prebuilt MMseqs2 reference database.
• Dockerfile – inference container definition.
• requirements.txt – Python dependencies.

Inference

The evaluation environment runs:

python /app/predict.py --input INPUT --output OUTPUT

The predictor reads the input Parquet file and writes JSONL predictions. Each output record contains the corresponding input ID and a list containing one fully specified EC number.