GSK plc · LSE:GSK
GSK — Algorithms for Hope 2021
16 slides from GSK's conference presentation, Jun 15, 2021
Slide 1: Using AI to develop Medicines with a higher probability of success
Algorithms for Hope, 15th June 2021 Using AI to develop Medicines with a higher probability of success Dr Patrick Schwab, GSK Artificial Intelligence and Machine Learning
Slide 2: Introduction
Introduction Hanafi, age 52 Chronic obstructive pulmonary disease (COPD) - Chronic inflammatory lung disease (difficulty breathing, cough, mucus) - Treated once-daily with triple therapy (inhaled glucosteroid, long-acting beta2-agonist, long-acting muscarinic agonist) - 15 and 25% reduction in moder
Slide 3: How can we leverage AI and ML to get more breakthrough medicines to patients faster?
How can we leverage artificial intelligence (AI) and machine learning (ML) to get more breakthrough medicines to patients faster?
Slide 4: Drug Development Process
Drug Development Process Discovery Preclinical Phase I Phase II Phase III Probability 60% 50% 30% 60% of success approved 1 medicine 10+ years Cook, David, et al. "Lessons learned from the fate of AstraZeneca's drug pipeline: a five-dimensional framework." Nature reviews Drug discovery 13.6 (2014):
Slide 5: Eroom's law: Exponential drop in R&D productivity
“Eroom’s law”: Exponential drop in R&D productivity Number of New Molecular Entities (NMEs) Probability of success for a per billion US$ R&D spending new medicine is roughly 5.5%. A single NME on average requires 100s of millions US$ R&D spending. This trend was, for decades, exponentially worsening
Slide 6: Tools Enabling AI in Medicine Development
Tools Enabling AI in Medicine Development CRISPR cell systems Multi-omics interventions 1 Functional Genomics deep learning lab automation 2 Machine Learning and Automation hardware 3 Large-scale Computation 6
Slide 7: Human Genetics – A rich Tool for Medicine Development
Human Genetics – A rich Tool for Medicine Development Causal understanding of the underlying mechanism of disease Higher probability of approval (EU/US) Targets (gene products) with high-quality human genetic evidence are up to three times more likely to succeed in the clinic. Population-scale Bioba
Slide 8: How can we create a map of cellular causal mechanisms?
Given the compelling evidence that a causal understanding of a medicine significantly increases probability of success, how can we create a map of cellular causal mechanisms?
Slide 9: Decoding the Language of the Cell with High Content Imaging
Decoding the Language of the Cell with High Content Imaging culture cells treated control (retinoic acid) apply perturbation untreated (CRISPR, control Molecule, …) cardiomyocytes (heart muscle cells) High content imaging with immunofluorescent staining of specific antigens enables us to understand
Slide 10: Robotic Lab Automation
Robotic Lab Automation Advances in robotic automation technology enable running thousands of petri dish (in-vitro) Experiments per day – generating experimental evidence at scale. Robots … • physically move 384 well plates, • apply perturbations, and • measure relevant cell parameters (e.g., imaging
Slide 11: Using AI to Create a Map of Cellular Phenotypes
Using AI to Create a Map of Cellular Phenotypes unhealthy Small healthy Molecule 1 Deep Learning Large Datasets Small Molecule 2 healthy CRISPR-KO UMAP2 Gene X unhealthy UMAP1 What perturbations move the molecular phenotype towards the healthy state? Effective perturbations could be attractive treat
Slide 12: Uncovering Biological Causal Networks
Uncovering Biological Causal Networks Human Genetics Studies (Genome-wide Association Studies) Gene X Disease Y Causal Evidence Scientific Literature Deep Learning single cell RNA sequencing (scRNAseq) Natural Language Processing 2 ML and Automation Representation Learning 12
Slide 13: Experimental Feedback Loops
Experimental Feedback Loops Experimental Design Generate Hypotheses Run Experiments “Gene X contributes to disease Y” Causal Experimental Maps Data Build Machine Learning Models 3 Large-scale Computation 13
Slide 14: Continuous Improvement to Create a Map of Health and Disease
Continuous Improvement to Create a Map Health and Disease biological understanding high-accuracy model With every experimental iteration, we ask new questions, get new answers and generate a rich and high-resolution understanding of the underlying disease biology. time 14 initial understanding
Slide 15: Conclusion
Conclusion With more than 8’000 potential treatments in preclinical development across the pharmaceutical and biotechnology R&D pipeline size industry1, an increase of the probability of success by just one percent would lead to the discovery of 80 potentially life- changing medicines for patients –
Slide 16: Thank you!
Thank you! Patrick Schwab patrick.x.schwab@gsk.com GlaxoSmithKline plc