발표자: 박혜진(신테카바이오) 발표일: 2018.1. 최근 과학 기술의 발전에 따라 환자 개개인에 최적화된 진단 및 치료를 제공하는 정밀의학의 시대가 도래했다. 정밀의학의 핵심은 유전체 등 빅데이터의 처리와 AI를 활용한 분석으로 진단에 유효한 정보를 확보하는 데 있다. 정밀의학이 가장 활발하게 적용되는 암 진단과 치료에 있어 분자 수준(인산화효소 돌연변이)과 세포 수준(암세포주)의 유전체 정보를 활용한 약물 활성 예측은 약물 저항성 관련 진단과 신약 개발에 도움을 줄 것으로 예상된다. 딥러닝 기법을 사용한 분자 수준(인산화효소 돌연변이)의 데이터를 활용한 약물 활성 예측 모형은 항암표적치료제와 표적 단백질의 상호작용을 기존의 모형보다 더 높은 정확도로 예측할 수 있다. 또한 세포 수준(암세포주) 데이터를 활용한 약물 활성 예측 모형은 환자의 약물 저항성 예측 뿐 아니라 신약 개발 이나 신약 재창출(drug repositioning) 연구에도 유용하게 사용될 수 있다. 발표 논문: https://www.nature.com/articles/s41598-018-27214-6

1. ResiScan & CDRscan: Deep learning based
Cancer drug response prediction
2018.01.17. 신테카바이오 용인 인실리코 의학센터 박혜진
1

2. Genome Integration & In silico Drug Company
Data
Yesterday Today Tomorrow
Symptom-based
Intuitional
medicine
Cohort-based
Evidence-based
medicine
Algorithm-based
Precision
medicine
Actions
• Larger hospitals
( More Patients)
• Better Intuitional
Clinical Practices
• Big Data & AI Hospitals
• Application of Precision
Medical Practice
• Specialized
Hospitals
• More Patients with
the Similar Diseases
• Cohort & Evidence
based Practice
Hospital
Paradigm Shift in Medicine

3. Genome Integration & In silico Drug Company
Drug Discovery Technology
3
penicillins
sulphonamides
aspirin
psychotropics
NSAIDS
H2-antagonists
beta blockers
lipid lowerers
ACE-inhibitors
Biotech drugs
chronic
degenerative
disease associated
with ageing,
inflammation,
cancer
drugs against
targets identified
from disease genes
1900 20301950 1960 1970 1980 1990 2000 2010 2020 2040
NewTherapeuticCycles
1st generation 2nd generation 3rd generation
natural products
and derivatives
serendipity
receptors
enzyme
genetic engineering
cell pharmacology/
molecular biology
genomics/ proteomics

4. Genome Integration & In silico Drug Company
Global Pharma Investment Trend in Precision Medicine (USD$ investment amount)
National Human Genome Research Institute. The Cost of Sequencing a Human Genome. Accessed September 13, 2016 at
http://www.genome.gov/sequencingcosts.
Data provided by: Concert Genetics. Available at concertgenetics.com. *Methodological notes: Concert Genetics began
publishing the first reliable data on the number of genetic testing products available in January of 2016. PMC has
published a list of 127 genetic tests commonly associated with the 132 personalized medicines listed in the Appendix of
this document at http:// www.personalizedmedicinecoalition.org/Education/Tests.
Coming of Age

5. Genome Integration & In silico Drug Company
AI 기반 신약 개발 업체
5
Company 국가 기술
투자금액
($M)
Atomwise USA Deep-learning screening from molecular structure data 6.35
BenevolentAI UK Deep-learning and natural language processing of research literature 100
BERG USA Deep-learning screening of biomarkers from patient data
Cloud Pharmaceuticals USA
Quantum Molecular Design process deployed on cloud-based high performance comput
ing platforms
1.35
Envisagenics USA Machine learning algorithm based on known splicing functional outcomes 0.1
Excientia USA Bispecific compounds via Bayesian models of ligand activity from drug discovery
Globavir Bosciences USA Deep-learning screening from drug and disease databases 0.15
GNS Healcare USA Bayesian probabilistic inference for investgating efficacy
Insilco Medicine USA Deep-learning screening from drug and disease databases 10
NuMedii USA Big data analysis using broad human clinical and molecular network data 5.5
Numerate USA Deep-learning from phenotypic data 17.42
Reursion Pharmaceuticals USA Cellular phenotyping via image analyais 15.35
twoXAR USA Deep-learning screening from literature and assay data 4.3
Verge Genomics USA Network algorithms on human genomic data 4
스탠다임 한국 Deep Learning, Drug Repositioning
신테카바이오 한국 Deep Learning, Genome Analysis, Virtual Screening

6. Genome Integration & In silico Drug Company
Cancer Drug
6
Chemotherapy Targeted Therapy Immunotherapy
• Drugs that effect cells
that are doubling
• Not very specific
• Cytotoxic
• Drugs that inhibit a more
specific target in cells
• Many are oral agents
• Mixture of cytostatic and
cytotoxic
• Drugs that effect immune
system
• Antibody
• Cell
• Vaccine

7. Genome Integration & In silico Drug Company
May 23 2017
Keytruda (Pembrolizumab) PD-L1 receptor existing list of approved indications:
Melanoma, lung cancer, Head & neck cancer, Lymphoma, Bladder cancer. More studies are currently
underway
FDA granted accelerated approval for the first time to use for solid tumor with a specific genetic feature
as indication.
First Accelerated FDA Approval
“PD-L1 발현율이 50% 이상이며, EGFR 및 ALK변이가
없는 진행형 비소세포폐암 환자”의 1차 치료제로 적응 증을 확대승인

8. Genome Integration & In silico Drug Company
Cancer Drug Resistance
8

9. Genome Integration & In silico Drug Company
Strategy
9
Cancer types
Drug Response
Drug Structure
CDRscan
ResiScan
Genetic variants
Genetic variants
Kinase Mutants

10. ResiScan: 분자 기반 약물 반응 예측
10

11. Genome Integration & In silico Drug Company
전통적인 In silico Model (SAR)
11
End-Points DataDescriptors
SAR (Structure-Activity-Relationship)  Finding F(X)

12. Genome Integration & In silico Drug Company
Proteochemometrics(PCM) model
12
A C D E F S T V W Y∙∙∙
∙∙∙0 0 0 0 0 0 0 0 0 1
∙∙∙0 0 0 0 0 0 0 0 0 1
∙∙∙0 0 0 0 1 0 0 0 0 0
∙∙∙0 0 0 0 0 0 0 0 0 1
20 Amino Acid one hot encording
Target Protein Drug Compound
Molecular Fingerprint
y = F(x)Activity

13. Genome Integration & In silico Drug Company
System 개요
13

14. Genome Integration & In silico Drug Company
Data structure
14

15. Genome Integration & In silico Drug Company
Target-Drug Data Information
Kinase mutant Drug
Number of Data 96 (wildtype:21개) 183
Descriptor Amino acid one-
hot encoding
CDK fingerprint
Number of
Feature
9900 (495*20) 3072
total 15702 (Data) * 12972 (Feature)
➢ Duong-Ly KC, et.al. Kinase Inhibitor Profiling Reveals Unexpected
Opportunities to Inhibit Disease-Associated Mutant Kinases. Cell Rep.
2016 Feb 2;14(4):772-81.

16. Genome Integration & In silico Drug Company
Virtual docking
16
DrugsBinding site in proteins
Virtual Docking
DrugsBinding site in proteins
기존 Model Virtual Docking Model
Machine learning
Obs.
Pre.

17. Genome Integration & In silico Drug Company
Result: Regression
17
• 기존 머신러닝(SVM, Random Forest) 방법과 비교
• SVM과 RF의 경우, Feature Selection 과정을 거침
SVM RF CNN Virtual docking
R2 0.7799 0.8179 0.8354 0.8741
Scatter plot

18. Genome Integration & In silico Drug Company
Result: Classification
18
Feature
Accuracy Specificity AUC
Protein Drug
DNN
(1D, earlystop)
9900 3072 0.901 0.967 0.940
SVM 499 2355 0.893 0.929 0.919
RF 499 2355 0.895 0.942 0.903
• 기존 머신러닝(SVM, Random Forest) 방법과 비교
• SVM과 RF의 경우, Feature Selection 과정을 거침

19. Genome Integration & In silico Drug Company
항암 약물 내성 기전
19

20. CDRscan: 변이 기반 항암 약물 반응 예측
20

21. Genome Integration & In silico Drug Company
데이터 개요

22. Genome Integration & In silico Drug Company22
Data Distribution
Cancer Type Gene 134
TCGA Desc Cell Line Gene Mutation Instance
1 ACC Adrenocortical carcinoma 1 6 6 197
2 ALL Acute lymphoblastic leukemia 26 118 585 5,728
3 BLCA Bladder Urothelial Carcinoma 19 72 150 3,693
4 BRCA Breast invasive carcinoma 51 108 367 10,037
5 CESC Cervical squamous cell carcinoma and endocervical adenocarcinoma 14 55 100 2,811
6 CLL Chronic Lymphocytic Leukemia 3 5 5 651
7 COAD_READ Colon adenocarcinoma and Rectum adenocarcinoma 51 128 1,568 9,523
8 DLBC Lymphoid Neoplasm Diffuse Large B-cell Lymphoma 35 95 268 7,100
9 ESCA Esophageal carcinoma 35 90 256 6,930
10 GBM Glioblastoma multiforme 36 80 203 7,291
11 HNSC Head and Neck squamous cell carcinoma 42 83 224 8,020
12 KIRC Kidney renal clear cell carcinoma 29 69 145 5,374
13 LAML Acute Myeloid Leukemia 28 96 243 6,143
14 LCML Chronic Myelogenous Leukemia 10 64 108 2,268
15 LGG Brain Lower Grade Glioma 17 53 100 3,759
16 LIHC Liver hepatocellular carcinoma 17 47 84 3,231
17 LUAD Lung adenocarcinoma 63 109 559 12,194
18 LUSC Lung squamous cell carcinoma 15 58 113 2,910
19 MB Medulloblastoma 4 15 17 848
20 MESO Mesothelioma 19 54 88 3,543
21 MM Multiple Myeloma 18 63 104 3,665
22 NB Neuroblastoma 32 77 149 6,908
23 OV Ovarian serous cystadenocarcinoma 34 93 291 5,953
24 PAAD Pancreatic adenocarcinoma 30 65 151 5,709
25 PRAD Prostate adenocarcinoma 6 77 129 1,135
26 SCLC Small Cell Lung Cancer 66 108 484 12,536
27 SKCM Skin Cutaneous Melanoma 55 117 521 10,858
28 STAD Stomach adenocarcinoma 24 94 272 4,372
29 THCA Thyroid carcinoma 16 57 91 3,190
30 UCEC Uterine Corpus Endometrial Carcinoma 9 108 348 1,843
31 unable to classify
992 134 8,105 153,746

23. Genome Integration & In silico Drug Company23
Mutation / Sample
0 20 40 60 80 100 120 140
#271
#99
#84
#75
#57
#48
#40
#35
#31
#27
#23
#19
#15
#11
#7
#3
number of sample
numberofmutation
cell-line 992
gene 134
mutation 8,105

24. Genome Integration & In silico Drug Company
Model 개요
24

25. Genome Integration & In silico Drug Company
a b
Observed ln(IC50)
-5 0 5 10
Predictedln(IC50)
10
5
0
-5
master fc shallow tanh unified
Models
8
6
4
2
0
-2
-4
-6
-8
Error
observedln(IC50)-predictedln(IC50)
master
fully connected
shallow
tanh
unified
예측 결과

26. Genome Integration & In silico Drug Company26
Cancer type별 예측 결과

27. Genome Integration & In silico Drug Company
향후 계획
27
항암약물 반응 예측 모델
세포주 유전체 변이,
발현 데이터
(CCLP,CCLE..)
약물 관련 유전체 변
이, 발현, 경로 데이터
(CCLP,CCLE,L1000…)
세포주-약물 반응
데이터
(GDSC, L1000…)
항암약물반응 바이오마커
공개 임상 유전체
데이터
(TCGA,…)
환자 유전체 데이터
(병원,…)

28. Genome Integration & In silico Drug Company
Conclusion
28
➢ 최근 유전체 분석 등 과학 기술의 발전에 따라 환자 개개인에 최적화된 진단
및 치료를 제공하는 정밀의학의 시대가 도래했다.
➢ 분자 수준(kinase mutant)과 세포 수준(cancer cell line)의 약물 활성 예측은 약
물 저항성 관련 진단과 신약 개발에 도움을 줄 것으로 예상됨.
➢ Kinase mutant에 대한 약물 활성 예측 모형은 virtual docking 기법을 사용하
여 더 높은 정확도를 가짐.
➢ 암세포주 데이터를 활용한 약물 활성 예측 모형은 약물 저항성 뿐 아니라 신약
개발 단계나 신약 재창출(drug repositioning) 단계에도 사용될 수 있음.
➢ 향후 임상 유전체, 유전체 발현 데이터 등의 데이터를 통합하여 예측 모형과 바
이오 마커를 개발함으로써 실제 진단이나 신약 개발에 유용한 툴로 완성도를 높
일 계획임.

29. 29
감사합니다.