TERP：基于机器学习的 TCR-肽结合特异性预测工具发布
TERP: Machine Learning-based prediction of TCR-peptide binding specificity
研究团队开发了 TERP（TCR-Epitope Recognition Prediction），用于预测 TCR 与 pMHC 的结合特异性，并上线为免费网页工具。方法上，作者系统比较了 TCR CDR3β 序列与候选肽抗原的不同嵌入方式，以及 CNN、ANN、随机森林、XGBoost、逻辑回归和朴素贝叶斯等多种机器学习方法。使用朴素 TCR 和随机非 cognate 配对两类负样本数据集、多种测试集划分进行基准评测后，随机森林算法与 BLOSUM50 特征表现最好。
来源：Bioinformatics Advances · doi.org
Motivation
The ability to accurately predict TCR-pMHC binding is crucial for the development of cancer immunotherapies, diagnostics, and new vaccines. Tetramer-based single-cell TCR sequencing techniques have provided information on TCR clonotypes and their cognate peptide antigens, which have been catalogued in databases such as VDJdb, IEDB, and McPAS-TCR. Recent advances in AI/ML techniques have opened up the opportunities for development of machine learning based methods for accurate predictions of TCR-pMHC binding.
Results
We have developed TERP (TCR-Epitope Recognition Prediction) based on a systematic analysis using different types of embeddings for CDR3β sequences of TCRs and putative peptide antigens, and a variety of different machine learning methods like CNN, ANN, random forest, XGBoost, logistic regression and naïve Bayes etc. Extensive benchmarking studies using different types of negative datasets (naïve TCRs and random non-cognate pairing) and different splits for test datasets revealed that the Random Forest algorithm and BLOSUM50 features have best predictive performance. TERP had superior performance compared to other published methods for TCR-Split (80:20 random split) and Strict-Split (less frequent peptide-TCR pairs in the test set) settings in both naïve and cross-combination negative sampling. Further, TERP demonstrated generalizability in predicting unseen TCR-peptide pairs, whereas other models failed to do so. TERP has also been implemented as a user-friendly web portal, accessible at http://www.nii.ac.in/TERP. The portal offers various functionalities, including TCR-peptide binding prediction, screening of TCRs for a given epitope, and prediction of epitopes for TCR clonotypes using CDR3β sequences or results from TCRSeq in 10x or immunoseq format.
Availability and implementation
Freely available online at http://www.nii.ac.in/TERP. The website is developed with the Django Python framework and supports all major browsers.