Predictive medicine based on analysis and
simulation of time-series clinical data
Yasushi Okuno
Kyoto University, Graduate School of Medicine RIKEN Life Systems Research Center
RIKEN Advanced Institute for Computational Science Foundation for Biomedical Research and Innovation
KUHP Oncology Center
Time-series information acquisition
Time-series sample acquisition
(Blood, DNA, etc)
Collaborating
institutions Multi-omics
(Genome,Proteome)
Novel biomarkers and pharmaceutical targets Feedback to the clinical practice for
personalized medicine
Universities, Pharmas Feedback to the new targets for drug discovery
Medical simulation and optimal treatment prediction Data accumulation Patient reported outcome(PRO) Clinical information
Personalized medicine and drug discovery
from time-series accumulation of biomedical big data
Kyoto University Hospital:
Biobank & Informatics for Cancer (BIC) Project
Medical big data
• More data than parseable by humans
• Complex data correlations not resolvable
by simple statistics.
Our Aim: Data-driven Predictive medicine and Personalized medicine
AI
• Pattern extraction from data
• Machine learning of complex patterns
• Prediction of events from patterns
Supercomputing
Big
data
analysis
and
Simulation
for
medicine
• Prediction of physical state, therapeutic efficiency, side effects, and outcomes.
• Identification of markers for early discovery of conditions.
Criticality of medical big data
Learn from patients ⇒ Give back to patients
Clinical
trial
Standard
of
care
Clinical
practice
Target population for therapy evaluation
with restricted therapy conditions
Real
patients
Differences in distributionData‐driven EBM
More effective, safer care
Answering unmet medical needs
• Rare cases (hundreds, not thousands)
• Mainly adults; children, pregnant mothers excluded • Simultaneous drug/condition constraints reduce
target population; potentially biased results • Long‐term treatment effect unclear
• Insufficient comparative treatment evaluation • Evaluation by specialists of a disease, objective?
Drug
development
Approval
Pharmacovigilence
(Post‐marketing tracking)
P atien t number Examination date
Plots of rate of neutrophils for KUHP Oncology patient
Oncology prediction problem:
• Prognosis prediction: tracking patient outcomes is critical due to possible metastasis
and drug resistance.
• Side effects prediction: monitoring patient symptom is critical because of active
therapy with anti‐cancer drugs.
Medical big data and simulation
using real-world data from Kyoto Univ. Hospital
【Data for analysis】 Kyoto U. Hospital chemotherapy recipients • 5285 patients • 2863 recorded deaths • Survival span after initial exam / start of therapy max 5930.0 days, average 817.5 days
Time-series analysis using conventional clinical statistics
P atien t numberCox regression analysis
Survival following first exam (days)
Sur viv al ra te
Computation using groups split at threshold of NLR = 4
5285 patients
2863 recorded deaths Survival span data
max 5930.0 days average 817.5
Examination date
Time‐course data of KUHP patient neutrophil‐to‐lymphocyte (NLR) ratio NLR is reported to be a predictor of prognosis (low NLR: good outcome)
Reverse calculation: number of days prior to death P atien t number 2863 deaths
Time-series analysis by heatmap plot
P atien t number Examination date 5285 patients 2863 recorded deaths Survival span data
max 5930.0 days average 817.5
Selection of patients with recorded death and alignment using date of death
Time‐course data of KUHP patient neutrophil‐to‐lymphocyte (NLR) ratio NLR is reported to be a predictor of prognosis (low NLR: good outcome)
NLR average
NLR changes dramatically rise 1 year
prior to death
Reverse: #days prior to death
NLR likely better predictor of remaining life than prognosis
Time-series analysis by heatmap plot
Reverse: #days prior to death
P atien t number 2863 deaths Time‐course data of KUHP patient neutrophil‐to‐lymphocyte (NLR) ratio NLR is reported to be a predictor of prognosis (low NLR: good outcome)
NLR average
NLR changes dramatically rise 1 year
prior to death
Reverse: #days prior to death
NLR likely better predictor of remaining life than prognosis
Time-series analysis by heatmap plot
Reverse: #days prior to death
P atien t number 2863 deaths Time‐course data of KUHP patient neutrophil‐to‐lymphocyte (NLR) ratio NLR is reported to be a predictor of prognosis (low NLR: good outcome)
‐ Cases of long lifespan despite high CRP value
‐ Cannot feed group statistics back to individuals
P atien t number Examination date
Plots of rate of neutrophils for KUHP Oncology patient
Oncology prediction problem:
• Prognosis prediction: tracking patient outcomes is critical due to possible metastasis
and drug resistance.
• Side effects prediction: monitoring patient symptom is critical because of active
therapy with anti‐cancer drugs.
Medical big data and simulation
using real-world data from Kyoto Univ. Hospital
【Data for analysis】 Kyoto U. Hospital chemotherapy recipients • 5285 patients • 2863 recorded deaths • Survival span after initial exam / start of therapy max 5930.0 days, average 817.5 days
Distribution of prediction error
Example of
large error Example of
small error
Groups to individuals: neutrophil change simulations
Example of
Cause
Ef
fe
ct
Impulse response method:
Estimate cause‐effect between
variables by observing change in each
variable after perturbation of a
specific single variable
• Monocytes strongly impact neutrophils.
• RBC, ALB, CL, ALP are also effectors.
Cause-effect estimation
between lab data
KUHP Cancer Center
Personal Genome Data
Determination of
treatments and side-effects from individual genotypes. Genotype-driven therapy selection from existing guidelines and research literature.
Clinical sequence
Molecular Simulations for Personalized Medicine in Drug Therapy
Personalized Drug Therapy
In oncology, genome dynamical change due to drug resistance via mutation
Mechanisms for resistance acquisition not fully understood.
Mutation
Non-small cell lung cancer therapy and resistance
Crizotinib
Alectinib
Ceritinib
L1196 G1202 S1206
T1151
Crizotinib G1269
E1167 I/T1171
CH5424802
αC helix
Wild type vs I1171T mutant
V/L1180
CH5424802 Wild type vs V1180L mutant
6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 ‐19 ‐18 ‐17 ‐16 ‐15 ‐14 ‐13 6 6.5 7 7.5 8 8.5 9 9.5 10 ‐18 ‐16 ‐14 ‐12 ‐10 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 ‐18 ‐16 ‐14 ‐12 ‐10 wt wt I1171T wt V1180L Alectinib Ceritinib Crizotinib wild I1171T F1174I F1174V V1180L V1185L L1196M L1198F G1202R G1269A F1174V Experimental pIC50 Experimental pIC50 Calculated ∆G (kcal/mol) Experimental pIC50 Calculated ∆G (kcal/mol) Calculated ∆G (kcal/mol)
Computed free energies vs.
Understanding mechanisms for reduction of Alectinib binding affinity
due to mutations in ALK
wild-CH5424802 (green) ‐77.74 ‐60.41 ‐17.33 I1171T-CH5424802 (cyan) ‐75.06 ‐58.04 ‐17.02 V1180L-CH5424802(magenta) ‐74.55 ‐60.01 ‐14.54 coulomb vdw (kcal/mol) ∆G
Binding free energies using the MP-CAFEE method
E1167 I/T1171 Alectinib αC helix H.B. broken L1196 V/L1180 E1167 αC helix Alectinib
KUHP Cancer Center
Personal Genome Data
Determination of
treatments and side-effects from individual genotypes. Genotype-driven therapy selection from existing guidelines and research literature.
Clinical sequence
Molecular Simulations for Personalized Medicine in Drug Therapy
Personalized Drug Therapy
In oncology, genome dynamical change due to drug resistance via mutation
Mechanisms for resistance acquisition not fully understood.
•Predicting response to drug therapy and side effects using simulation
•Uncovering mechanisms explaining side effects and drug resistance
Acknowledgements
• The KBDD Consortium
• RIKEN Adv. Inst. Comput. Sci.
• Res. Org. for Info. Sci. Tech (RIST)
• Foundation Biomed. Res. Innov.
• Osaka Univ. Cybermedia Center
• Biogrid Center Kansai
• Post-K, Priority issues program
• CREST “Big Data Applications”
• Mitsui Knowledge Industry Co. Ltd.
• Chugai Pharmaceutical Co. Ltd.
• COE program from Kobe and Hyogo
Kyoto University Grad. Sch. Medicine
• Profs. Manabu Muto,
• Assoc. Prof. Shigemi Matsumoto
• Assoc. Prof. Masashi Kanai
• All the members of Okuno Labo.
All the members of Priority Issue 1
All the members of KBDD Consortium
Special thanks
Japn. Foudation for Cancer Res.
• Dr. Ryohei Katayama
RIKEN / AICS
• Dr. Makoto Taiji
• Dr. Mitsugu Araki
• Kei Taneishi
Foundation Biomed. Res. Innov.
• Dr. Tasuku Honjyo
• Dr. Yoichi Nabeshima
• Dr. Hiroaki Iwata