Page 20 - ebook
P. 20
Deep Y : AI-based biosimilarity assessment platform for
antibody drugs using intact glycoproteins analysis
1,2
3
1,2
1,2
1,2
Dae Sik Cho , Ga Hyeon Kim , Hee Youn Hwang , Myung Jin Oh , and Hyun Joo An *
1. Graduate School of Analytical Science and Technology, Chungnam National University, Daejeon, 34134, Korea
2. Asia-pacific Glycomics Reference Site, Daejeon, 34134, Korea
3. Korea Basic Science Institute, Ochang, 2819, Korea
Overview
A biosimilar is a biological drug that is nearly identical to the original product.
Development of artificial intelligence based model
Glycosylation on antibody drugs typically exhibits a high degree of complexity
Development of a data extraction algorithm
and can be very sensitive to changes in the manufacturing process. This
glycosylation has the greatest impact on maintaining consistent quality and
clinical performance throughout the biosimilar lifecycle. Nevertheless, the
objective criteria for evaluating similarity in terms of glycosylation in biosimilar
development are still ambiguous and difficult. Therefore, we developed DeepY,
a deep learning platform for assessing the biosimilarity of mAbs using intact
glycoprotein analysis by LC-Q-TOF MS. A total of six original and biosimilar were
secured, including Trastuzumab, Infliximab, Adalimumab, and Bevacizumab,
and the glycoproteome forms were analyzed for biosimilarity evaluation. A Development of AI algorithm model for Advancement of AI-based model
total 2,880 spectra and 86,400 glycoform data including test set, training set, biosimilarity assessment Determination of Learning Rate Test set result
1.2 False
and validation set were generated for development of an artificial intelligence Training Fully connected neural network 1 1.33%
Convolution Pooling Convolution Pooling 0.8
algorithm using convolutional and fully connected neural network. Decoy data 0.6
Flattening 0.4
also was created to improve a data extraction algorithm. Deep Y can now 0.2
0
0.1 0.01 0.001 0.0001
determine the biosimilarity of mAbs with about 99% accuracy and is available 1.2 Determination of epoch True
1
on the website (http://Deepy.kr). There is no doubt that DeepY will be a 0.8 98.67%
Feature learning Classification Averaged Accuracy 0.6
groundbreaking scientific assistant for the development of antibody biosimilars. 0.4
Test data Data extraction Model Classification 0.2 1.00 No. 36
0.975
0
Background Molecular weight 1.2 Determination of Node 150 Accuracy 0.950
100
10
50
0.925
1 0.900
0.875
0.8
Importance of biosimilarity Purpose of study Peak intensity 0.6 0.850
0.825
0.4 0.800
Problems Solution Analysis Big Data AI assessment AI Cloud Application Classification : 0.2
0
platform
- Expiration of patents Biosimilarity assessment between original and biosimilar 10 50 100 500 1000
training test Accuracy = 98.67%
- Expanding the development
of biosimilars
Deep Y : AI-based biosimilarity assessment platform
- Complexity & Heterogeneity
- Sensitive change in the process LC/MS analysis Cloud computing architecture Web page (http://Deepy.kr)
Omics Development of Combing cloud Comparability
technology Al data mining computing and assessment &
- Ambiguity & Difficulty of Need for a new approach based big data algorithms AI quality
objective criteria production verification
-AI-based biosimilarity Convolution & by researcher
- Absence of a platform assessment platform MS spectrum fully connected & regulatory Result Files
neural network agencies End Framework Build a Website Analyzed DATA RAW DATA
(JS)
Biosimilarity in glycosylation Sample information User Development Tool Parsing AI-development Machine- Analytics
learning
Front-End
for data
Originator Biosimilar Front – Web UX Design for feature clustering
selection
Cell line Media Tool(JS) . . .
Similar Trastuzumab Herceptin Herzuma Event Run Python
glycosylation Trastuzumab Infliximab
profile? Glycosylation Infliximab Remicade Remsima End Node.js PHP End
Biosimilar Adalimumab Humira Framework Framework
Cell culture Purification Back – Back – http://Deepy.kr
condition method OS ( Linux )
Originator Adalimumab Bevacizumab Bevacizumab Avastin
Result Summary
Generating big data • Even though glycosylation on mAbs is an important factor in comparing
the biosimilarity between original and biosimilar, the objective criteria are
Process of intact glycoprotein analysis by LC-Q-TOF/MS
Trastuzumab Infliximab Adalimumab still ambiguous and difficult.
mAbs UHPLC/Q-TOF MS Deconvolution Training, Test, Validation originator originator originator
data
x10 5 x10 5 x10 5
x10 5 • In this study, we have developed a deep learning model, Deep Y. Big data
4 2.5
4
148063.09 148063.09 1.6 148516.78 148085.97
3.5 3.5
1.4 2
3
3
2.5 1.2 acquired from intact protein analysis consisted of deconvoluted masses
2.5
2 1.5
1
1.5
2
1 0.8 1
1.5
0.5 0.6
0 1 0.4 and their abundances were used as data sets which were divided into
147500 148000 148500 149000 0.5
0.5 0.2
Process of generating decoy data 0 147500148000148500149000 0 148250148500148750149000 0 147800148000148200148400148600 training, validation, and test set, for the development of a deep learning
mAbs Enzymatic digestion Adjustment of Decoy data
antibody glycosylation 5 Decoy data 3 Decoy data Decoy data
4
(beta N-glucosaminidase) x10 4 x10 147245.98 x10 147705.10 x10 model using convolutional and fully connected neural network.
147271.49 1.6 148062.48 147271.59
4 148086.21 6 148517.48 7
1.4 148086.66
x4 5 x4 6 x4
x4 1.2 • The model was mounted on the cloud computing architecture for
2 1 4 5
4
0.8 3
3
0 0.6
2 2 convenient use for mAbs biosimilar developers and evaluators can utilize it
147000 148000 149000 0.4
Counts vs. Deconvolute 1 1
0.2
Decoy Total secured data 0 147000 148000 149000 Counts vs. Deconvoluted Mass (amu) 0 146000 148000 150000 conveniently.
0
147000 148000 149000 150000
(experimental) Bevacizumab Trastuzumab Infliximab
originator
540 6 4 biosimilar biosimilar
Decoy x10 x10 9 x10 5
1.8
2.2
8
spectra 1.6 149202.70 7 148223.65 2 148677.12 Acknowledgement
(in silico) 1560 spectra 1.4 6 1.8
1.6
1.2
5 1.4
1 4 1.2
1080 spectra 0.8 3 0.8 • This research was supported by a grant(19172MFDS169) from Ministry of
1
0.6
2 0.6
0.4
Original 0.2 1 0.4
0.2
0 0 0 Food and Drug Safety in 2021.
147800148200148600149400 149800
147600148000148400148800149200 148400148800149200149600
4 Decoy data 4 Decoy data x10 4 Decoy data
720
780 spectra x10 148387.25 x10 147271.60 2.5 • This work was supported by the National Research Foundation of Korea
spectra
8
149201.81 8
7 7 148086.64
x4 x4 2
6 6
360 5 5 1.5 147705.17 148519.70 Grant funded by the Korean Government (MSIP)(2020, R&D Equipment
spectra 480 4 4 x4
1
spectra 3 3
2 2 0.5 Engineer Education Program, 2014R1A6A9064166).
240 Biosimilar 360 1 0 1 0
0
spectra spectra 146000 148000 150000 147000 148000 149000 147000 148000 149000 150000
Counts vs. Deconvoluted Mass (amu) • We thank LABMS members (Laboratory for Advanced Bio-analytical Mass
180 Decoy data Enzymatic Non-Enzymatic
spectra Spectrometry) for their tremendous help.
Homepage : www.agrs.kr / E-mail: sugar@cnu.ac.kr
