基于深度学习识别RiPPs前体肽及裂解位点

doi:10.12211/2096-8280.2022-016

摘要/Abstract

摘要：

得益于基因测序技术的快速发展，基因组测序数据呈现爆炸式增长，核糖体合成和翻译后修饰肽（RiPPs）是近十年逐渐进入人们视野的一大类肽类天然产物。这类化合物在自然界中分布极其广泛，具有丰富的结构多样性和生物活性多样性，是天然药物的重要来源。RiPPs的发现主要依赖低通量生物实验，传统方法精确但成本高昂，随着新型计算机技术的更新迭代，包括antiSMASH、RiPP-PRISM等在内的生物信息学工具能够极大加速RiPPs挖掘进程，但依然无法突破基于同源性方法（例如搜索保守的生物合成酶）的限制——无法有效识别具有不同生物合成机制的新型RiPPs。在这里，本文首次基于自然语言处理预训练模型BERT，提出四种可以完全依赖序列数据识别RiPPs而非基于同源性及基因组上下文信息的深度学习模型，通过对各模型进行验证分析和对比，最终确定在RiPPs识别赛道上表现卓越的最佳模型BERiPPs（bidirectional language model for enhancing the performance of identification of RiPPs precursor peptides）。BERiPPs能够在不考虑基因组背景的情况下以无偏见的方式识别RiPPs前体肽，并可通过条件随机场生成对前导肽裂解位点的预测，为高通量挖掘全新RiPPs提供了思路，并在一定程度下揭示了前体肽和修饰酶间的生物学底层关系。

关键词: 深度学习, RiPPs, 前体肽, 预训练模型, 天然产物挖掘

Abstract:

Genome sequencing data showed explosive growth attributed to the rapid development of DNA sequencing technology. Ribosomally synthesized and post-translationally modified peptides are a kind of natural peptide product that gradually came into people's view in the last decade. These compounds are widely distributed in nature, diverse in structure and bioactivity, and are important sources of natural drugs. The discovery of RiPPs mainly relies on low-throughput biological experiments, which are accurate but costly. With the development of new information technologies, bioinformatics tools such as antiSMASH and RIPP-Prism can greatly accelerate the process of RiPPs mining. However, methods based on gene homology, such as searching for conserved biosynthetic enzymes, are still unable to effectively identify novel RiPPs with different biosynthetic mechanisms. Here, for the first time, based on the natural language processing pre-training model BERT, four deep learning models that can fully rely on sequence data to identify RiPPs instead of homology and genomic context information are proposed and trained on the same RiPPs dataset. Through verification and comparison of these models, the best model BERiPPs performs well on the RiPPs identification track and is as accurate as the homology-based method. BERiPPs can identify RiPPs precursor peptides and RiPPs classes in an unbiased manner regardless of the genomic background, extending the range of novel RiPPs captured by approximately 60% compared to homology-based approaches. By combining BERiPPs with a conditional random field, the prediction of the cleavage site of the leader peptide can be indirectly generated with high accuracy by the recognition of each amino acid label in the sequence. The deep learning based on the pre-training model provides the possibility for high-throughput mining of novel RiPPs in a manner different from that of the gene context-dependent methods and reveals the underlying biological relationship between precursor peptides and modified enzymes.

Key words: deep learning, RiPPs, precursor peptides, pre-training model, natural products mining

中图分类号:

Q819

吕靖伟, 邓子新, 张琪, 丁伟. 基于深度学习识别RiPPs前体肽及裂解位点[J]. 合成生物学, 2022, 3(6): 1262-1276.

Jingwei LYU, Zixin DENG, Qi ZHANG, Wei DING. Identification of RiPPs precursor peptides and cleavage sites based on deep learning[J]. Synthetic Biology Journal, 2022, 3(6): 1262-1276.

图/表 10

图1 RiPPs的生物合成途径（图中各部分序列比例仅作图示用，不代表其实际长度）

Fig. 1 Biosynthetic pathways of RiPPs(The proportion of each part of the sequence in the figure is for illustration only and does not represent its actual length)

图2 根据ORF长度识别RiPPs前体肽

Fig. 2 Identification of RiPPs precursor peptides based on ORF length

图3 基于BIO规则的序列标注示例

Fig. 3 Sample sequence annotations based on BIO rules

图4 模型输入表示及结构

Fig. 4 Model input representation and structure

图5 BERT预训练层的初始化

Fig. 5 Initialization of pre-trained layers of BERT

表1 不同算法在RiPPs前体肽识别任务上的效果对比

Tab. 1 Comparison of different algorithms in identification of RiPPs precursor peptides

Model	Precision	Recall	F1
BERT (top 1 layers initialized)	0.9056	0.8962	0.9009
BERT (top 2 layers initialized)	0.8938	0.9031	0.8985
BERT (fully initialized)	0.8710	0.8408	0.8556
BERT (pre-trained)	0.9031	0.9031	0.9031
BERT-CNN	0.9123	0.8997	0.9059
BERT-DPCNN	0.9126	0.9031	0.9078
BERT-RCNN	0.9127	0.8685	0.8901
BERiPPs (BERT-BiLSTM)	0.9331	0.9170	0.9250

图6 各模型对特定RiPPs家族识别结果对比

Fig. 6 Comparison of identification results of various RiPPs families by different models

图7 不同训练方式下的BERiPPs和DeepRiPP在预测RiPPs类别上的结果对比（因测试集中各类RiPPs样本数量不同，故将混淆矩阵中的数值进行归一化处理，再根据四舍五入原则精确到小数点后一位）

Fig. 7 Comparison of prediction results of RiPPs classes between BERiPPs under different training methods and DeepRiPP(Due to the different number of various RiPPs samples in the test set, the values in the confusion matrix are normalized and then accurate to one decimal point according to the rounding principle.)

表2 不同损失函数下的部分RiPPs类别预测结果对比

Tab. 2 Comparison of prediction results of partial RiPPs classes under different loss functions

Class	Loss Function	Precision	Recall	F1
Autoinducing peptides	Focal Loss	1.0000	0.9583	0.9787
Autoinducing peptides	Cross Entropy Loss	0.9231	1.0000	0.9600
Thiopeptides	Focal Loss	0.8947	0.8947	0.8947
Thiopeptides	Cross Entropy Loss	0.8500	0.8947	0.8718
Lasso peptides	Focal Loss	0.8873	0.9130	0.9000
Lasso peptides	Cross Entropy Loss	0.8889	0.9275	0.9078
Class Ⅲ_Ⅳ Lanthipeptides	Focal Loss	0.9483	0.9649	0.9565
Class Ⅲ_Ⅳ Lanthipeptides	Cross Entropy Loss	0.9655	0.9825	0.9739

图8 BERiPPs与RiPPMiner及antiSMASH对比

Fig. 8 Comparison of BERiPPs with RiPPMiner and antiSMASH

参考文献 37

1	MARTENS E, DEMAIN A L. The antibiotic resistance crisis, with a focus on the United States[J]. The Journal of Antibiotics, 2017, 70(5): 520-526.
2	HUTCHINGS M I, TRUMAN A W, WILKINSON B. Antibiotics: past, present and future[J]. Current Opinion in Microbiology, 2019, 51: 72-80.
3	HUDSON G A, MITCHELL D A. RiPP antibiotics: Biosynthesis and engineering potential[J]. Current Opinion in Microbiology, 2018, 45: 61-69.
4	WANG F T, WEI W Q, ZHAO J F, et al. Genome mining and biosynthesis study of a type B linaridin reveals a highly versatile α-N-methyltransferase[J]. CCS Chemistry, 2021, 3(3): 1049-1057.
5	SKINNIDER M A, JOHNSTON C W, EDGAR R E, et al. Genomic charting of ribosomally synthesized natural product chemical space facilitates targeted mining[J]. Proceedings of the National Academy of Sciences of the United States of America, 2016, 113(42): E6343-E6351.
6	ARNISON P G, BIBB M J, BIERBAUM G, et al. Ribosomally synthesized and post-translationally modified peptide natural products: overview and recommendations for a universal nomenclature[J]. Natural Product Reports, 2013, 30(1): 108-160.
7	YU Y, ZHANG Q, VAN DER DONK W A. Insights into the evolution of lanthipeptide biosynthesis[J]. Protein Science, 2013, 22(11): 1478-1489.
8	ZHONG Z, HE B B, LI J, et al. Challenges and advances in genome mining of ribosomally synthesized and post-translationally modified peptides (RiPPs)[J]. Synthetic and Systems Biotechnology, 2020, 5(3): 155-172.
9	BLIN K, SHAW S, STEINKE K, et al. antiSMASH 5.0: updates to the secondary metabolite genome mining pipeline[J]. Nucleic Acids Research, 2019, 47(W1): W81-W87.
10	HETRICK K J, VAN DER DONK W A. Ribosomally synthesized and post-translationally modified peptide natural product discovery in the genomic era[J]. Current Opinion in Chemical Biology, 2017, 38: 36-44.
11	HYATT D, CHEN G L, LOCASCIO P F, et al. Prodigal: prokaryotic gene recognition and translation initiation site identification[J]. BMC Bioinformatics, 2010, 11: 119.
12	DELCHER A L, BRATKE K A, POWERS E C, et al. Identifying bacterial genes and endosymbiont DNA with Glimmer[J]. Bioinformatics, 2007, 23(6): 673-679.
13	VAN HEEL A J, DE JONG A, MONTALBÁN-LÓPEZ M, et al. BAGEL3: automated identification of genes encoding bacteriocins and (non-) bactericidal posttranslationally modified peptides[J]. Nucleic Acids Research, 2013, 41(W1): W448-W453.
14	TIETZ J I, SCHWALEN C J, PATEL P S, et al. A new genome-mining tool redefines the lasso peptide biosynthetic landscape[J]. Nature Chemical Biology, 2017, 13(5): 470-478.
15	MERWIN N J, MOUSA W K, DEJONG C A, et al. DeepRiPP integrates multiomics data to automate discovery of novel ribosomally synthesized natural products[J]. Proceedings of the National Academy of Sciences of the United States of America, 2020, 117(1): 371-380.
16	AGRAWAL P, KHATER S, GUPTA M, et al. RiPPMiner: a bioinformatics resource for deciphering chemical structures of RiPPs based on prediction of cleavage and cross-links[J]. Nucleic Acids Research, 2017, 45(W1): W80-W88.
17	DE LOS SANTOS E L C. NeuRiPP: Neural network identification of RiPP precursor peptides[J]. Scientific Reports, 2019, 9: 13406.
18	SHIN H C, ROTH H R, GAO M C, et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning[J]. IEEE Transactions on Medical Imaging, 2016, 35(5): 1285-1298.
19	SUNDERMEYER M, SCHLÜTER R, NEY H. LSTM neural networks for language modeling[C]// 13th Annual conference of the International Speech Communication Association 2012 (INTERSPEECH 2012). Portland, OR, USA: International Speech Communications Association, 2012:194-197.
20	DEVLIN J, CHANG M W, LEE K, et al. BERT: pre-training of Deep Bidirectional Transformers for Language Understanding[C]//Proceedings of NAACL-HLT. 2019: 4171-4186.
21	TENNEY I, DAS D, PAVLICK E. BERT rediscovers the classical NLP pipeline[C]//Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence, Italy. Stroudsburg, PA, USA: Association for Computational Linguistics, 2019: 4593-4601.
22	CHO K, VAN MERRIENBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar. Stroudsburg, PA, USA: Association for Computational Linguistics, 2014: 1724-1734.
23	Bahdanau D, Cho K H, Bengio Y. Neural machine translation by jointly learning to align and translate[C]//3rd International Conference on Learning Representations, ICLR 2015. 2015.
24	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Advances in neural information processing systems. 2017: 5998-6008.
25	SHERSTINSKY A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network[J]. Physica D: Nonlinear Phenomena, 2020, 404: 132306.
26	WANG Q, LI B, XIAO T, et al. Learning deep transformer models for machine translation[C]//Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence, Italy. Stroudsburg, PA, USA: Association for Computational Linguistics, 2019: 1810-1822.
27	SÖDING J. Protein homology detection by HMM-HMM comparison[J]. Bioinformatics, 2005, 21(7): 951-960.
28	LIU L Y, REN X, SHANG J B, et al. Efficient contextualized representation: Language model pruning for sequence labeling[C]//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Belgium. Stroudsburg, PA, USA: Association for Computational Linguistics, 2018: 1215-1225.
29	Huang Z, Xu W, Yu K. Bidirectional LSTM-CRF models for sequence tagging[EB/OL]. arXiv preprint: 2015, arXiv:1508.01991.
30	ZHAO H S, JIA J Y, KOLTUN V. Exploring self-attention for image recognition[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020, 10073-10082.
31	Dodge J, Ilharco G, Schwartz R, et al. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping[EB/OL]. arXiv preprint: 2020, arXiv:2002.06305.
32	YADAV S, SHUKLA S. Analysis of k-fold cross-validation over hold-out validation on colossal datasets for quality classification[C]//2016 IEEE 6th International Conference on Advanced Computing. Bhimavaram, India: IEEE, 2016: 78-83.
33	ZHANG Z L, SABUNCU M. Generalized cross entropy loss for training deep neural networks with noisy labels[J]. Montréal: NeurIPS, 2018, 31.
34	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//2017 IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 2999-3007.
35	LI B, SHER D, KELLY L, et al. Catalytic promiscuity in the biosynthesis of cyclic peptide secondary metabolites in planktonic marine cyanobacteria[J]. Proceedings of the National Academy of Sciences of the United States of America, 2010, 107(23): 10430-10435.
36	JUMPER J, EVANS R, PRITZEL A, et al. Highly accurate protein structure prediction with AlphaFold[J]. Nature, 2021, 596(7873): 583-589.
37	XIE Q Z, LUONG M T, HOVY E, et al. Self-training with noisy student improves ImageNet classification[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE,2020 : 10684-10695.

[1]	朱景勇, 李钧翔, 李旭辉, 张瑾, 毋文静. 深度学习在基于序列的蛋白质互作预测中的应用进展[J]. 合成生物学, 2024, 5(1): 88-106.
[2]	康里奇, 谈攀, 洪亮. 人工智能时代下的酶工程[J]. 合成生物学, 2023, 4(3): 524-534.
[3]	陈志航, 季梦麟, 戚逸飞. 人工智能蛋白质结构设计算法研究进展[J]. 合成生物学, 2023, 4(3): 464-487.
[4]	宋益东, 袁乾沐, 杨跃东. 深度学习在蛋白质功能预测中的应用[J]. 合成生物学, 2023, 4(3): 488-506.
[5]	王也, 王昊晨, 晏明皓, 胡冠华, 汪小我. 生物分子序列的人工智能设计[J]. 合成生物学, 2021, 2(1): 1-14.