New generation of NMR analysis assisted by Deep Learning ... · Ann. Meeting BSJ 2016-IPR seminar 2016-1106--25. 03 . New generation of NMR analysis assisted by Deep Learning and

IPR seminar 2016-06-03 Ann. Meeting BSJ 2016-11-25

New generation of NMR analysis assisted by Deep Learning and highly sophisticated Web-

tools developed by PDBj-BMRB

Naohiro Kobayashi

PDBj-BMRB group Institute for Protein Research

Osaka university, Japan

The 55th Annual Meeting of the Biophysical Society of Japan Luncheon meeting

1


NBDC

wwPDB

Project director Haruki Nakamura

PDBj: Database for X-ray and EM structure and experimental data Haruki Nakamura (Professor) Atsushi Nakagawa (Professor) Rei Kinjyo (Associate Prof.) Daron Standley (Prof. at iFREc) 5 annotators, 2 programmers, 2 research fellows

PDBj-BMRB: Database for NMR structure and experimental data Toshimichi Fujiwara (Professor) Chojiro Kojima (Prof. YNU & Osaka Univ.) Naohiro Kobayashi (Sp.Ap. Associate Prof.) Masashi Yokochi (Programmer Annotator) Takeshi Iwata (Sys. Admin. Annotator)

3-year project April 2014-March 2017

Organization of the project of PDBj at Osaka university supported by JST-NBDC

2


We are collecting chemical shift data in collaboration with Univ. Wisconsin-Madison and wwPDB

3


O

C CH3 N

H

C O

C N H

C O

C

H

N H

CH2

CH CH3 CH2

C O

C N H

CH2

N C O

C

H

CH2

CH2

COOH

N C O

C

H

CH2 CH2

C O

C N H

CH2

OH

Plasmid

[15N]-NH4Cl [13C]-Glucose Expression Purification

NMR tube

CH2 CH3 CH2

NH2

C

Nearly all of the atoms can be labeled with 13C, 15N!

Stable isotope labeling of protein made amazing NMR analysis possible!

4


We can observe nearly all of the atom signals in protein… each one of the atom

This is really amazing… as a spectroscopy technique

5


If you prepare 15N-labeled protein…

If the protein has 135 amino acid residues…you can use the 135 signal as 135 atomic probe

How you can imagine, if you can see the signal of each atom in protein

6


If you assign all of the NMR signals….

You can hear the sound of atoms!

7


Domain orientation Structure modeling

What you can do if you assigned all the NMR signals

Backbone atom dynamics

0.8

0.4

0.0

-0.4

NOE

430 450 470 490 500 Residue number

Ligand interaction

Sugase et al., Nature, 2007

8


BioMagResBank (BMRB) Databak of NMR database

9

IPR seminar 2016-06-03

Our activities on BMRB processing (2015)

• 78 entries processed at Osaka (~10%)

• BMRB total 759 entries

0

300

600

900

1200

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

Ent

ry

Year

MadisonOsaka

10


NMR-STAR v3 BMRB/XML BMRBxTool

bmr15400.str

Yokochi et al., J. Biomed. Sem. (2016)

Conversion of BMRB data into machine readable format

11


BMRB/XML

bmr15400.xml

BMRBoTool

BMRB/RDF

bmr15400.rdf

RDF: Resource Description Framework

Toward the more machine readable format

12


BMRB Mutation OMIM dbSNP 2nd SASA% ∆∆Ghydr 4280 L100V 300005 rs28935168 Coil 22.5 3.9 4280 R106W 300005 rs28934907 Strand 8.0 11 4280 R133C 300005 rs28934904 Coil 49.4 7.2 4280 E137G 300005 rs61748392 Helix 41.7 6.6 4280 A140V 300005 rs28934908 Helix 42.0 1.7

High SASA

High DDGhydr

MeCP2: methylated CpG binding protein NMR structure 1QK9: Wakefiled et al., 1999

R133

E137 The mutations,R133C and E137 have been found in patients with X-linked mental retardation.

-- detailed information from OMIM--

Data analysis using BMRB/PDB/OMIM database

13


Multiple database search on the PDBj-BMRB portal

Universal search box Search filter

Search tabs & settings

Search button

14

IPR seminar 2016-06-03 Ann. Meeting BSJ 2016-11-25 15


MagRO A tool for integrated NMR data analysis

16


MagRO: A package of tools for highly automated NMR data analysis

GUIs and automated tools suppot user to analyze NMR signal easily

17


However… None of them can start from peak identification

There are many programs that can assign NMR signals automatically

[Why?]

• So many noises and artifacts in NMR

spectra

• Difficult to get weak important signals • Sometimes signals are overlapped

18


By the way, the technology neural network has been developed

dramatically in this a few years

Developments of techniques for deep layered parceptron

Developments of Graphic Processing Unit (GPU)

Many layered and connected neural networks is now called: Deep Learning (DNN)

19


What is Convolutional Neural Network (CNN)?

Neocognitron: Fukushima et al. (1979)

LeNet-5: LeCun et al. (1988), DQN Hassabis et al. (2015)

Convolution

Filter size:5x5

Stlide:1

40x40 Conv1 Pool1

Max-pooling Window： 2x2 Stlide:2

Conv2

Convolution

Filter size:5x5

Stlide:1

Pool2

Stlide:2

…

Out-put

Hidden leyer

128 dim

Noise or Signal?

Max-pooling Window： 2x2

10x10x32 tensor

20


CNN+FNN->ZNN: fully automated noise/signal recognition

CNN: like a “eye-sight” of human, can recognize 2D NMR image but the view field is narrow

FNN: collects the other information, helps CNN ZNN: integrates CNN, FNN noise/signal recognition

ZNN

Noise or Signal?

FNN CNN

weak

strong

wiggle close lone

noisy

border

side-chain

spec-type

8->6->1 : NN

10->8->1 : NN

1600dim 5-leyer

21


Id_Robot

Annealing sequential assignment

Assignments

Matrix AASS

1HN-15Na clusters, Ca, Cb, CO chemical shifts

Matrix ART

3D-NMR spectra: HNCO, HNCACB, CBCA(CO)NH

Matrix CCON

Peak identification: CNN/FNN->ZNN

Perfectly automated assignment system using Id_Robot + Anneal_Robot

Anneal_Robot Random mutation 20cycles 40stages (Total:800 annealings)

22


Performance of new version

Backbone signal assignment (117 a.a.) manual: 6-8 hours semi-automated (previous version): 30-60min

new version with AI: 50sec (4-CPUs) 30-60 times faster!

…this is only working for spectra with very good quality still required improvements on peak recognition

23


Acknowledgement

Takashi Nagata, Masato Katahira Kenji Sugase

Kyoto University：

Wisconsin University: Eldon L. Ulrich, John L. Markley

Osaka University: Toshimichi Fujiwara, Chojiro Kojima (YNU, Osaka Univ.) Masashi Yokochi, Takeshi Iwata

Tokushima Bunri University: Yoshikazu hattori

Kengo Tsuda RIKEN Yokohama：

Yutaka Muto, Kanako Kuwasako Musashino University：

24

Documents

New generation of NMR analysis assisted by Deep Learning ... · Ann. Meeting BSJ 2016-IPR seminar 2016-1106--25. 03 . New generation of NMR analysis assisted by Deep Learning and