본문으로 바로가기

코드 및 인공지능 모델 데이터
DeepECtransformer: EC number prediction for protein sequences with transformer layers

  • Accession
    KDT-135-10000093
  • Submission date
    2026-11-07

Project Detail
Dataset detail - Accession, 프로젝트의 영문 제목, 프로젝트의 국문 제목, 프로젝트의 영문 설명, 프로젝트의 국문 설명
Accession
KAP242748
프로젝트의 영문 제목
DeepECtransformer: Transformer-based functional annotation of enzyme-encoding genes
프로젝트의 국문 제목
DeepECtransformer: 트랜스포머 기반 효소 암호화 유전자의 기능 주석
프로젝트의 영문 설명
This study developed DeepECtransformer, a computational framework that combines a transformer-based neural network with sequence homology searches to predict Enzyme Commission (EC) numbers from protein amino acid sequences and support functional annotation of enzyme-encoding genes. Analysis of self-attention scores revealed that the model highlights functionally relevant regions, including enzyme active sites and cofactor-binding sites. The framework was used to reassess existing enzyme annotations and investigate potential functional differences among protein variants across Escherichia coli strains. In E. coli K-12 MG1655, EC numbers were predicted for 464 y-ome proteins with insufficient functional characterization, and the predicted enzymatic activities of YgfF, YciO, and YjdM were validated through in vitro enzyme assays. Application to 128,100,490 protein sequences from 70,600 genomes provided a large-scale resource of predicted enzyme functions to support functional genomics and metabolic research.
프로젝트의 국문 설명
본 연구에서는 트랜스포머 기반 신경망과 서열 상동성 검색을 결합하여 단백질의 아미노산 서열로부터 효소의 EC 번호를 예측하고, 효소를 암호화하는 유전자의 기능 주석을 지원하는 DeepECtransformer를 개발하였다. 셀프 어텐션 점수 분석을 통해 모델이 효소 활성 부위와 보조인자 결합 부위 등 기능적으로 중요한 영역에 주목함을 확인하였다. 이를 활용하여 기존 효소 기능 주석을 재평가하고, 대장균 균주 간 단백질 변이체의 잠재적 기능 차이를 분석하였다. 대장균 Escherichia coli K-12 MG1655의 기능 규명이 불충분한 y-ome 단백질 464개에 대해 EC 번호를 예측하였으며, 이 가운데 YgfF, YciO, YjdM의 예측된 효소 활성을 시험관 내 효소 활성 실험으로 검증하였다. 나아가 70,600개 유전체의 128,100,490개 단백질 서열에 적용하여 기능 유전체학 및 대사 연구에 활용할 수 있는 대규모 효소 기능 예측 자원을 구축하였다.

Metadata
데이터 타입
Model
제목
DeepECtransformer: EC number prediction for protein sequences with transformer layers
키워드
EC number prediction, enzyme annotation, deep learning, transformer, protein sequence, DIAMOND
제작자
Gi Bae Kim
Version
v_1_0
코드 언어
Python
입력
input.fasta(FASTA, protein sequences in upper-case one-letter amino acid codes with unique sequence IDs, e.g. example/mdh_ecoli.fa)
출력
DeepECv2_result.txt(TSV, columns sequence_ID/prediction/score with one row per EC number when the neural network predicts every sequence, otherwise columns sequence_ID/prediction with multiple EC numbers joined by semicolons and None when no EC number is found), tmp/DL_prediction_result.txt(TSV, columns sequence_ID/prediction/score, neural network result)
데이터셋 정보
UniProtKB/Swiss-Prot and TrEMBL enzymes released April 2018, uniprot dataset 80% split (train), UniProtKB/Swiss-Prot and TrEMBL enzymes released April 2018, uniprot dataset 10% split (validation), UniProtKB/Swiss-Prot and TrEMBL enzymes released April 2018, uniprot dataset 10% split (test), Swiss-Prot enzymes, swissprot dataset (DIAMOND reference database), NEW-392 (test), Price-149 (test)
작업/기능
EC number prediction, enzyme function annotation, multi-label protein sequence classification
프레임워크
PyTorch, Hugging Face Transformers
모델 유형
Transformer, BERT, CNN

readme
# DeepECtransformer ## Procedure **Note**: This source code was developed in Linux, and has been tested in Ubuntu 16.04 with Python 3.6. 1. Clone the repository git clone https://github.com/kaistsystemsbiology/DeepProZyme.git 2. Create and activate virtual environment (takes 5~10 min) conda env create -f environment.yml conda activate deepectransformer 3. To use gpus properly, install the pytorch and cuda for your gpus. This code was tested on pytorch=1.7.0 using cuda version 10.2 ## Example - Run DeepECtransformer (takes < 1 min) python run_deepectransformer.py -i ./example/mdh_ecoli.fa -o ./example/results -g cpu -b 128 -cpu 2 python run_deepectransformer.py -i ./example/mdh_ecoli.fa -o ./example/results -g cuda:3 -b 128 -cpu 2 ## Notes for the K-BDS deposit (October 2026) This deposit includes the tokenizer files; use the unpacked `DeepECtransformer/` folder instead of the clone in step 1. Run the commands from the repository root. ### Example files `example/output/` contains the expected result of the example. The example commands above write their result to `example/results/`, which can be compared with `example/output/`. ### Tokenizer files Transformers 3.5.1 can no longer download the tokenizer of Rostlab/prot_bert_bfd from the Hugging Face Hub, so the tokenizer files were added unmodified to `Rostlab/prot_bert_bfd/`. - Source: https://huggingface.co/Rostlab/prot_bert_bfd (revision 6c5c8a55a52ff08a664dfd584aa1773f125a0487); ProtTrans, Elnaggar et al., IEEE Trans. Pattern Anal. Mach. Intell. 44, 7112-7127 (2022) - License: Academic Free License v3.0 (https://github.com/agemagician/ProtTrans) ### GPU PyTorch 1.7.0 in environment.yml is a CUDA 10.2 build and supports only GPUs with compute capability 3.7 to 7.5 (e.g. Tesla V100, T4, GeForce RTX 20xx). In other environments, run on CPU (`-g cpu`).

File