Protein function prediction through multi-instance multi-label transfer learning

Haifeng HU; Mao ZHENG; Weijian WU; Jun WANG; Jiansheng WU

doi:10.1360/N112017-00090

SCIENTIA SINICA Informationis

Download PDF

SCIENTIA SINICA Informationis, Volume 47 , Issue 11 : 1538-1550(2017) https://doi.org/10.1360/N112017-00090

Protein function prediction through multi-instance multi-label transfer learning

Haifeng HU ^1,2, Mao ZHENG ¹, Weijian WU ³, Jun WANG ⁴, Jiansheng WU ^4,*

More info

Affiliations：

1. School of Telecommunication and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China

2. A Project Funded by the Priority Academic Program Development of Jiangsu Higer Education Institutions, Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology, Nanjing University of Information Science Technology, Nanjing 210044, China

3. College of Computer and Information, Hohai University, Nanjing 211100, China

4. School of Geography and Biological Information, Nanjing University of Posts and Telecommunications, Nanjing 210023, China

* Corresponding author (jansen@njupt.edu.cn)

ReceivedMay 14, 2017
AcceptedJun 30, 2017
PublishedNov 13, 2017

Previous Table of Contents Next

Rating

Abstract

Funded by

国家自然科学基金(61571233,61271082)

国家重点基础研究发展计划(973)(2011CB302903)

江苏省高校自然科学研究重大项目(14KJA510003)

江苏省重点研发计划(BE2015700)

南京信息工程大学PAPD与CICA- EET

transfer learning

protein function prediction

multi-instance multi-label learning

new sequencing-completed species

sample reweighting

References

[1] Jiang J, Zhai C X. Instance weighting for domain adaptation in NLP. In: Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, Prague, 2007. 264--271. Google Scholar

Click to view Download high-quality image

Figure 1
Multi-instance multi-label transfer learning framework(TR-MIML), including the re-weighting data samples from source domain stageand the classification model construction stage
Click to view Download high-quality image

Figure 2
(Color online) Effect on protein function prediction of Rattus norvegicus by transfer learning using five species with different phylogenetic relationship

Algorithm 1 Pseudo code of TR-MIML learning framework

$\widehat{h}~=~$TR-MIML${(}{~D}^{~T}{~,}~{~D}^{~S}{~)}~$

Input: $D^T:$ Target domain dataset; $D^S:$ Source domain dataset.

Output: $\widehat{h}:$ Classifier

Steps:

1.for $~X_i^S~$ in $D^S$ do

2.$f_i^S~=~{\rm~miFV}(~{X_i^S~}~)$;

3.end for

4.for $~X_i^T$ in $D^T$ do

5.$f_i^T~=~{\rm~miFV}(~{X_i^T}~)$;

6.end for

7.Compute $\beta~$ by solving (2);

8.Learn the classifier $\widehat{h}$ by solving (5).

Table 1 Experimental dataset statistics

Species	Proteins	GO terms	Domains per protein	GO terms per protein
			(Mean$\pm$std.)	(Mean$\pm$std.)
Geobacter sulfurreducens	379	320	3.20$\pm$1.21	3.14$\pm$3.33
Azotobacter vinelandii	407	340	3.07$\pm$1.16	4.00$\pm$6.97
Mus musculus	11676	3065	2.76$\pm$1.84	44.64$\pm$50.27
Rattus norvegicus	5991	2600	2.53$\pm$1.71	39.51$\pm$44.78
Homo sapiens	13773	3311	2.98$\pm$4.30	55.81$\pm$126.63
Arabidopsis thaliana	8986	1811	2.02$\pm$1.46	27.68$\pm$70.37
Saccharomyces cerevisiae	3509	1566	1.86$\pm$1.36	15.89$\pm$11.52

Table 2 Performance comparison on Geobacter sulfurreducens (target domain) by multiplemulti-instance multi-label learning methods with or without transferlearning. Results demonstrate that transfer learning can improve theperformance of protein function prediction on this task

Source domain	Method	AP ($~\uparrow~)$	CV ($~\downarrow~)$	HL ($~\downarrow~)$	OE ($~\downarrow~)$	RL ($~\downarrow~)$
	TR-MIMLfast	0.58$\pm~$0.02	4.04$\pm~$0.07	0.15$\pm~$0.00	0.55$\pm~$0.04	0.32$\pm~$0.01
	MIMLfast	0.44$\pm~$0.02	4.76$\pm~$0.03	0.20$\pm~$0.01	0.71$\pm~$0.03	0.43$\pm~$0.01
Mus musculus	TR-MIMLNN	0.56$\pm~$0.01	4.28$\pm~$0.20	0.21$\pm~$0.01	0.58$\pm~$0.02	0.36$\pm~$0.03
	MIMLNN	0.54$\pm~$0.01	4.51$\pm~$0.19	0.24$\pm~$0.01	0.58$\pm~$0.01	0.37$\pm~$0.01
	TR-MIMLSVM	0.53$\pm~$0.02	4.16$\pm~$0.02	0.18$\pm~$0.0	0.61$\pm~$0.01	0.37$\pm~$0.09
	MIMLSVM	0.44$\pm~$0.01	4.62$\pm~$0.05	0.19$\pm~$0.02	0.67$\pm~$0.01	0.40$\pm~$0.01
	TR-MIMLfast	0.56$\pm~$0.02	4.02$\pm~$0.17	0.16$\pm~$0.00	0.58$\pm~$0.03	0.33$\pm~$0.02
	MIMLfast	0.43$\pm~$0.05	5.22$\pm~$0.07	0.23$\pm~$0.03	0.75$\pm~$0.08	0.42$\pm~$0.06
Rattus norvegicus	TR-MIMLNN	0.53$\pm~$0.01	4.11$\pm~$0.06	0.16$\pm~$0.02	0.58$\pm~$0.01	0.38$\pm~$0.01
	MIMLNN	0.48$\pm~$0.01	4.74$\pm~$0.09	0.19$\pm~$0.00	0.66$\pm~$0.03	0.41$\pm~$0.01
	TR-MIMLSVM	0.53$\pm~$0.03	4.29$\pm~$0.03	0.17$\pm~$0.01	0.60$\pm~$0.03	0.39$\pm~$0.01
	MIMLSVM	0.52$\pm~$0.01	5.10$\pm~$0.08	0.17$\pm~$0.01	0.66$\pm~$0.01	0.35$\pm~$0.02
	TR-MIMLfast	0.53$\pm~$0.04	4.42$\pm~$0.14	0.17$\pm~$0.00	0.62$\pm~$0.05	0.35$\pm~$0.01
	MIMLfast	0.50$\pm~$0.02	4.50$\pm~$0.14	0.22$\pm~$0.02	0.66$\pm~$0.04	0.35$\pm~$0.02
Saccharomyces cerevisiae	TR-MIMLNN	0.53$\pm~$0.01	4.58$\pm~$0.09	0.14$\pm~$0.04	0.61$\pm~$0.09	0.37$\pm~$0.01
	MIMLNN	0.52$\pm~$0.02	4.72$\pm~$0.11	0.16$\pm~$0.01	0.62$\pm~$0.02	0.41$\pm~$0.02
	TR-MIMLSVM	0.54$\pm~$0.02	4.61$\pm~$0.07	0.17$\pm~$0.01	0.57$\pm~$0.01	0.40$\pm~$0.03
	MIMLSVM	0.53$\pm~$0.01	4.72$\pm~$0.04	0.18$\pm~$0.01	0.60$\pm~$0.01	0.43$\pm~$0.01

a) 粗体表示有迁移学习的结果要显著好于无迁移学习的结果(基于置信度为95%的配对样本t检验).

Table 3 Performance comparison on Azotobacter vinelandii (target domain) by multiplemulti-instance multi-label learning methods with or without transferlearning. Results demonstrate that transfer learning can improve theperformance of protein function prediction on this task

Source domain	Method	AP ($~\uparrow~)$	CV ($~\downarrow~)$	HL ($~\downarrow~)$	OE ($~\downarrow~)$	RL ($~\downarrow~)$
	TR-MIMLfast	0.55$\pm~$0.00	4.30$\pm~$0.40	0.15$\pm~$0.00	0.58$\pm~$0.01	0.34$\pm~$0.02
	MIMLfast	0.49$\pm~$0.03	4.55$\pm~$0.15	0.21$\pm~$0.02	0.69$\pm~$0.05	0.38$\pm~$0.02
Mus musculus	TR-MIMLNN	0.52$\pm~$0.00	4.35$\pm~$0.0	0.22$\pm~$0.03	0.68$\pm~$0.02	0.39$\pm~$0.01
	MIMLNN	0.48$\pm~$0.01	4.79$\pm~$0.03	0.27$\pm~$0.01	0.65$\pm~$0.00	0.41$\pm~$0.00
	TR-MIMLSVM	0.50$\pm~$0.0	4.55$\pm~$0.12	0.22$\pm~$0.04	0.63$\pm~$0.01	0.41$\pm~$0.02
	MIMLSVM	0.47$\pm~$0.01	4.67$\pm~$0.12	0.28$\pm~$0.01	0.64$\pm~$0.02	0.44$\pm~$0.02
	TR-MIMLfast	0.54$\pm~$0.02	4.64$\pm~$0.23	0.17$\pm~$0.00	0.64$\pm~$0.02	0.39$\pm~$0.04
	MIMLfast	0.49$\pm~$0.02	5.03$\pm~$0.25	0.25$\pm~$0.00	0.67$\pm~$0.03	0.40$\pm~$0.02
Rattus norvegicus	TR-MIMLNN	0.50$\pm~$0.03	4.93$\pm~$0.20	0.20$\pm~$0.01	0.66$\pm~$0.02	0.40$\pm~$0.01
	MIMLNN	0.46$\pm~$0.01	5.36$\pm~$0.13	0.23$\pm~$0.01	0.71$\pm~$0.02	0.46$\pm~$0.02
	TR-MIMLSVM	0.52$\pm~$0.01	4.73$\pm~$0.05	0.21$\pm~$0.02	0.65$\pm~$0.01	0.42$\pm~$0.01
	MIMLSVM	0.49$\pm~$0.00	4.82$\pm~$0.03	0.25$\pm~$0.01	0.70$\pm~$0.01	0.46$\pm~$0.01
	TR-MIMLfast	0.62$\pm~$0.02	4.52$\pm~$0.10	0.18$\pm~$0.00	0.53$\pm~$0.02	0.37$\pm~$0.02
	MIMLfast	0.55$\pm~$0.03	4.87$\pm~$0.28	0.24$\pm~$0.02	0.60$\pm~$0.03	0.35$\pm~$0.02
Saccharomyces cerevisiae	TR-MIMLNN	0.59$\pm~$0.01	5.02$\pm~$0.10	0.19$\pm~$0.04	0.60$\pm~$0.01	0.43$\pm~$0.02
	MIMLNN	0.51$\pm~$0.00	5.49$\pm~$0.10	0.19$\pm~$0.01	0.64$\pm~$0.00	0.46$\pm~$0.00
	TR-MIMLSVM	0.52$\pm~$0.01	4.77$\pm~$0.03	0.19$\pm~$0.01	0.62$\pm~$0.03	0.40$\pm~$0.01
	MIMLSVM	0.49$\pm~$0.01	4.75$\pm~$0.00	0.20$\pm~$0.01	0.66$\pm~$0.01	0.46$\pm~$0.02

a) 每个评价指标上最好的结果用粗体表示.

Table 4 Performance comparison on Geobacter sulfurreducens (target domain) by AdaBoost andLogistic Regression (LR) learning methods with or without transfer learning.Results demonstrate that transfer learning can improve the performance ofprotein function prediction on this task

Source domain	Method	AP ($~\uparrow~)$	CV ($~\downarrow~)$	HL ($~\downarrow~)$	OE ($~\downarrow~)$	RL ($~\downarrow~)$
Mus musculus	TrAdaBoost	0.58$\pm~$0.01	4.10$\pm~$0.03	0.23$\pm~$0.01	0.61$\pm~$0.01	0.33$\pm~$0.01
	AdaBoost	0.47$\pm~$0.01	4.51$\pm~$0.04	0.28$\pm~$0.01	0.69$\pm~$0.01	0.41$\pm~$0.01
	DALR	0.44$\pm~$0.02	4.24$\pm~$0.02	0.26$\pm~$0.00	0.66$\pm~$0.00	0.34$\pm~$0.02
	LR	0.35$\pm~$0.03	4.32$\pm~$0.05	0.27$\pm~$0.02	0.82$\pm~$0.01	0.45$\pm~$0.02
Rattus norvegicus	TrAdaBoost	0.45$\pm~$0.03	4.23$\pm~$0.02	0.27$\pm~$0.01	0.65$\pm~$0.02	0.33$\pm~$0.01
	AdaBoost	0.36$\pm~$0.02	4.31$\pm~$0.03	0.28$\pm~$0.03	0.83$\pm~$0.03	0.45$\pm~$0.02
	DALR	0.45$\pm~$0.01	4.11$\pm~$0.03	0.9$\pm~$0.01	0.78$\pm~$0.01	0.46$\pm~$0.04
	LR	0.32$\pm~$0.01	4.41$\pm~$0.05	0.30$\pm~$0.01	0.88$\pm~$0.00	0.54$\pm~$0.01
Saccharomyces cerevisiae	TrAdaBoost	0.44$\pm~$0.01	4.11$\pm~$0.02	0.28$\pm~$0.01	0.77$\pm~$0.01	0.46$\pm~$0.04
	AdaBoost	0.32$\pm~$0.01	4.41$\pm~$0.05	0.30$\pm~$0.01	0.88$\pm~$0.01	0.54$\pm~$0.01
	DALR	0.56$\pm~$0.08	4.35$\pm~$0.01	0.28$\pm~$0.02	0.61$\pm~$0.00	0.32$\pm~$0.01
	LR	0.47$\pm~$0.02	4.48$\pm~$0.00	0.33$\pm~$0.01	0.67$\pm~$0.01	0.34$\pm~$0.00

a) 每个评价指标上最好的结果用粗体表示.

Table 5 Performance comparison on Azotobacter vinelandii (target domain) by AdaBoost andLogistic Regression (LR) learning methods with or without transfer learning.Results demonstrate that transfer learning can improve the performance ofprotein function prediction on this task

Source domain	Method	AP ($~\uparrow~)$	CV ($~\downarrow~)$	HL ($~\downarrow~)$	OE ($~\downarrow~)$	RL ($~\downarrow~)$
Mus musculus	TrAdaBoost	0.52$\pm~$0.00	4.07$\pm~$0.01	0.28$\pm~$0.00	0.55$\pm~$0.00	0.40$\pm~$0.00
	AdaBoost	0.50$\pm~$0.00	4.77$\pm~$0.03	0.36$\pm~$0.00	0.69$\pm~$0.00	0.42$\pm~$0.00
	DALR	0.53$\pm~$0.00	4.47$\pm~$0.02	0.26$\pm~$0.01	0.55$\pm~$0.00	0.38$\pm~$0.00
	LR	0.48$\pm~$0.01	4.79$\pm~$0.05	0.31$\pm~$0.02	0.67$\pm~$0.03	0.41$\pm~$0.02
Rattus norvegicus	TrAdaBoost	0.57$\pm~$0.03	4.25$\pm~$0.00	0.27$\pm~$0.01	0.62$\pm~$0.01	0.33$\pm~$0.01
	AdaBoost	0.46$\pm~$0.02	4.49$\pm~$0.01	0.34$\pm~$0.02	0.66$\pm~$0.02	0.35$\pm~$0.02
	DALR	0.53$\pm~$0.01	4.15$\pm~$0.01	0.17$\pm~$0.00	0.60$\pm~$0.01	0.36$\pm~$0.09
	LR	0.44$\pm~$0.00	4.61$\pm~$0.05	0.19$\pm~$0.01	0.67$\pm~$0.00	0.39$\pm~$0.00
Saccharomyces cerevisiae	TrAdaBoost	0.49$\pm~$0.01	4.43$\pm~$0.01	0.30$\pm~$0.00	0.60$\pm~$0.01	0.41$\pm~$0.04
	AdaBoost	0.45$\pm~$0.01	4.69$\pm~$0.09	0.37$\pm~$0.01	0.66$\pm~$0.01	0.48$\pm~$0.01
	DALR	0.44$\pm~$0.02	4.24$\pm~$0.02	0.26$\pm~$0.00	0.66$\pm~$0.00	0.34$\pm~$0.01
	LR	0.35$\pm~$0.01	4.33$\pm~$0.07	0.28$\pm~$0.00	0.84$\pm~$0.01	0.44$\pm~$0.01

a) 每个评价指标上最好的结果用粗体表示.

0

Citations
0

Altmetric
Article metrics

Email
Export

Protein function prediction through multi-instance multi-label transfer learning

Abstract

Funded by

References

0

0

Interesting search

Top 5Related Articles