Artificial Intelligence in Translational Precision Medicine - PowerPoint PPT Presentation

Artificial Intelligence in Translational Precision Medicine ACOSIS-2019 Marrakech, Morocco Nov 20-22 nd , 2019 Peter J. Tonellato, PhD Professor of Bioinformatics Director of Center for Biomedical Informatics Health management and Informatics School of Medicine, University of Missouri Columbia, Missouri, USA

I have entered Morocco one less time than I have left Morocco. Conceived and Born in Casablanca, so... Bidaoui

Translational Precision Cancer Medicine • “Precision Medicine” with digital molecular profiling • Quantification of Life in the Era of Precision Medicine • CBMI Programs • PGx and Clinical Avatars • DCP – NSCLC - BC • Molecular Tumor Board • AI and Cancer

Translational Precision Cancer Medicine • “Precision Medicine” with digital molecular profiling • Quantification of Life in the Era of Precision Medicine • CBMI Cancer Programs • Two Tier I Proposals • DCP – NSCLC - BC • Molecular Tumor Board • AI and Cancer

Precision Medicine (21 st Century) NIH and US academic healthcare complexes have turned attention to data intensive, evidence-based, patient centric “Precision Medicine” – accounting for individual patient genetics , lifestyle and environment . Era of digitized patho/physiology - “big” data using emerging digital sequencers; high definition 3/4-D imaging, … Seek a translational approach capable of restoring personalized medicine while leveraging ‘big’ data and analytics with objectives: • Leverage experience of healthcare practice • Capture value of digitized evidence before & after patient interaction • Increase quality of (< 30 minute) face time • Improve individual patient outcome • Cost neutral • Iterative active learning manner

Personalized Medicine (Pre-WWII) • Physicians Education – MD no CME • Training in local Family practice by Lead Physician • Experience gained over decades of Family Practice on multiple-generation families Personalized • Average practice < 1000 patients • Average face-to-face time > 30 minutes

De-Personalized Medicine (Post-WWII) • Physicians Education – MD at Academic-Medical Centers followed by Residencies; Fellowships; and additional Specialty training, CMEs and highly technical workshops • Specialized experience gained over decades of referred patients (far less personalized) • Data and Evidence driven using early technologies (imaging, blood analyzers,…) De-Personalized • Average Practice > 3000 patients • Average face-to-face time < 30 minutes

Accelerate De-Personalization with Data-Driven Medicine • In Era of molecular testing (genome, transcriptome, epigenome) individual data and information at Terabyte levels • AI, deep learning and related anonymous analytical methods contain inherent risks far beyond technical weaknesses in approach, methods, sensitivity and specificity • No pedagogical approach to introduce data, evidence, predictive measures to healthcare providers • Specialization increases with data-driven approaches thus accelerating factors driving de-personalization Uber De-Personalized • Terabytes of data and information • Information and predictions inconclusive or contradictory to experience • Average Practice > ??,000 patients • Average face-to-face time << 30 minutes

Translational Research Clinical Enterprise Research Enterprise 1. Round holes arise in clinical setting 2. Square Pegs derived from basic research 3. Translation emerges from Commercial R&D and Regulatory Approval process followed by clinical implementation

Clouded Translational Medicine LPM Translation Insilico Translational Medicine Simulations and Predictions

Translational Precision Cancer Medicine • “Precision Medicine” with digital molecular profiling • Quantification of Life in the Era of Precision Medicine • CBMI Programs • PGx and Clinical Avatars • DCP – NSCLC - BC • Molecular Tumor Board • AI and Cancer

PGx and PM Paradox • Precision (either individual or sub-population) • Multi-factor inclusion criteria (age, gender, genotype,…) • Coupled to multiple (some ~50) warfarin dosing algorithms and protocols • => Explodingly large clinical trials

US Mixed Population Statistics

Clinical Avatars Human Avatar PHI First Name: Animal PHI First Name: Ozzy Last Name: House Last Name: Osborne Physical Height: 6’ 6” Physical Height: 6’ Weight: 180 Weight: 160 Genetic CYP2C9: *3/*3 Genetic CYP2C9: *1/*1 VKORC1: A/B VKORC1: A/A

Phenomenological modeling provides iterative method to accurate representations Short and broad Tall and skinny Ken from Toy Story 3

CA are statistical representations of actual populations Bayesian Model Simulation Framework Clinical avatar records – used as input to the clinical trial simulation framework

Clinical Avatars (Model data set structure) Variable(s) Parameters Age 18 to 24 (21.1%), 25 to 44 (30.3%), 45 to 64 (21.9%), 65 to 94 (26.7%) Gender Male {< 18 (51.26%), 18 to 24 (51.11%), 25 to 44 (50.06%), 45 to 64 (48.65%), 65 and over (41.18%)}; Female {< 18 (48.74%), 18 to 24 (48.89%), 25 to 44 (49.94%), 45 to 64 (51.35%), 65 and over (58.82%)} Race White (75.1%), African American (12.3%), Native American (0.9%), Asian (3.6%), Pacific Islander (0.1%), Other (5.5%), Unknown (2.5%) Height Mean: 69.2”, St.D: 6.6”, Min : 56.0”, Max: 82.4” Weight Mean: 189.8 lb, St.D: 59.1 lb, Min: 71.6 lb, Max : 308.0 lb Smoker White - 20%; African American - 21%; Native American - 35%; Asian / Pacific Islander - 11%; Other - 23% Amiodarone Y - 55%, N - 45% DVT Y - 26.8% N - 73.2% VKORC1 A/A - 65%, A/B - 20%, B/B - 15% CYP2C9 *1/*1 - 64.3%, *1/*2 - 18%, *1/*3 - 11.7% , *2/*2 - 2% , *2/*3 - 2.1% , *3/*3 - 0.25% The clinical avatar population and the resulting variables and statistical distributions.

Methodology • Preprocess data set (errors, clean-up, imputation) • Split “Cleaned” Data into Training and Test Data Sets • Iterative Bayesian Network Modeling: • Select random sample of Data for use as training set • Domain Knowledge integrated into Neural Network model (TETRAD) • Conduct Search • Initialize Search: FCI algorithm – test for latent variables • Test Additional Search Algorithms • Randomize Training Data -> Conduct Search • Use predictive/search metrics to define 3 “best” BNMs • Compare edges/non-edges in 3 “best” fit BNMs • Perform Markov Blanket validation • Compare/Revise Domain Knowledge • Continue until Domain Knowledge fully incorporated • Compare Domain Knowledge, Predictive and Test Metrics across BNMs – select “Optimal” BNM.

BNMs PC Train Data Imputation Data JPC Searches Searches Tetrad Random Test Test Sampling ~70% Data Data PCL FCI FCI GES CPC JCPC Knowledge 1 Training Knowledge 2 Validation Metrics BN1 BN1 i Knowledge i Markov * * Blankets BN2 i BN2 * * BN3 BN3 i * * BN* Test Data Validation Literature/Experts Testing/Predictive Metrics

GENERATED DAGS DAG 1 DAG 5 DAG 2 DAG 6 DAG 4 DAG 3 21

Parameter U.S. Base Actual PharmGKB Warfarin 18 Simulated Warfarin P-Value* (n=5700) (n = 20,000) Age 1 0.75 <18 27.6% (1572) 0.18% (10) 0.13% (26) 18 – 24 7.4% (420) 1.3% (75) 1.2% (235) 25 – 44 30.9% (1763) 9.9% (559) 9.8% (1,957) 45 – 64 21.5% (1227) 36% (2,040) 36.4% (7,282) 65 – 94 2.6% (718) 52.5% (2,974) 52.5% (10,500) Gender by age 1 0.89 <18 M: 49.9% (784), F: 50.1% (788) M: 30% (3), F: 70% (7) M: 34.6% (9), F: 65.4% (17) 18 – 24 M: 48.8% (205), F: 51.2% (215) M: 42.7% (32), F: 57.3% (43) M: 47.2% (111), F: 52.7% (124) 25 – 44 M: 50% (882), F: 50% (881) M: 49.9% (279), F: 50.1% (280) M: 50.6% (990), F: 49.4% (967) 45 – 64 M: 48.4% (594), F: 51.6% (633) M: 60% (1225), F: 40% (815) M: 59.4% (4,324), F: 40.6% (2,958) 65 – 94 M: 41.4% (310), F: 58.6% (438) M: 59.3% (1,855), F: 40.7% (1,272) M: 59.4% (6,353), F: 40.6% (4,344) Race 1 0.51 White 75.1% (4,282) 54.8% (3,122) 54.2% (10,835) African American 12% (684) 8.1% (462) 7.9% (1,583) Native American 0.8% (45) 0% (0) 0% Asian 3.6% (206) 28.7% (1,634) 29.7% (5,936) Pacific Islander 0.09% (5) 0% (0) 0% Other 5.9% (336) 0% (0) 0% Unknown 2.5% (142) 8.4% (482) 8.2% (1,646)

Actual PharmGKB Warfarin 18 Parameter U.S. Base Simulated Warfarin P-Value* (n=5700) (n = 20,000) Weight 5 (lbs) 7.7e-31 Mean 176.55 ± 30.9 171.58 ± 48.2 173.51 ± 27.87 Min 92 66 92 Max 273 524 290 Smoker 4 2.4e-11 White 20.3% (868) 14.4% (324) 14.3% (1,552) African American 21.2% (145) 20.8% (91) 20.9% (332) Native American 40% (18) 0% (0) 0% (0) Asian/Pacific 11.7% (24) 6.4% (18) 5.7% (340) Islander 22.6% (76) 6.5% (16) 5.7% (94) Other/Unknown DVT 6,7,8,9,19,11 Yes 26.8% (1,527) 16.4% (817) 16% (3,203) No 73.2% (4,173) 83.6% (4,191) 84% (16,797) VKORC1 7,10,14,15,16,17 0.83 A/A 15.5% (884) 52.2% (1,245) 52% (10,404) A/B 46.7% (2,661) 25.8% (614) 26.3% (5,261) B/B 37.8% (2,155) 22.0% (525) 21.7% (4,335) CYP2C9 7,10,14,15,16 0.81 *1/*1 64.3% (3,666) 74.9% (4,155) 75.4% (15, 079) *1/*2 18.8% (1,071) 13.4% (742) 13.4% (2,676) *1/*3 12.6% (718) 9% (501) 8.8% (1,756) *2/*2 1.9% (109) 1% (58) 1% (194) *2/*3 2% (114) 1.3% (72) 1.1% (227) *3/*3 0.39% (22) 0.4% (22) 0.3% (68)