Roche Holding Ltd · SIX:RO
Roche — Investor Day (November 2024)
64 slides from Roche's investor day presentation, Nov 25, 2024
Slide 1: Doing now what patients need next
Doing now what patients need next 1
Slide 2: Forward-looking statements
This presentation contains certain forward-looking statements. These forward-looking statements may be identified by words such as ‘believes’, ‘expects’, ‘anticipates’, ‘projects’, ‘intends’, ‘should’, ‘seeks’, ‘estimates’, ‘future’ or similar expressions or by discussion of, among other things, str
Slide 3: Welcome
Welcome Bruno Eschli Head of Investor Relations 25 November 20243
Slide 4: Agenda
Agenda Welcome Bruno Eschli, Head of Investor Relations Revolutionizing drug R&D with Lab in a Loop Aviv Regev, Head of Genentech Research and Early Development (gRED) Q&A and closing Panel 4
Slide 5: Revolutionizing drug R&D with Lab in a Loop
Revolutionizing drug R&D with Lab in a Loop Aviv Regev Genentech Research and Early Development (gRED), Genentech/Roche 25 November 20245
Slide 6: Drug R&D is a challenging process with a low success rate
Drug R&D is a challenging process with a low success rate 6
Slide 7: Goal: Enhancing R&D productivity
Goal: Enhancing R&D productivity Volume Revenue PTS Effectiveness R&D productivity Costs Cycle time Efficiency 7
Slide 8: Drug discovery phases: Classical approach
Drug discovery phases: Classical approach Target discovery Lead Lead Clinical and assessment Identification optimization candidate 8
Slide 9: Example: Enhancing Research productivity
Example: Enhancing Research productivity Enriching targets from human Develop disease pathobiology genetics, genomics and maps to reveal new targets translational research Earlier portfolio entry to increase Volume Revenue PTS Effectiveness AI-assisted target qualification FIC competitiveness R&D In
Slide 10: LAB* IN A LOOP
DATA LAB* IN A LOOP *or clinic EXPERIMENT SYNTHESIS HUMAN BIOLOGY AI/ML 10
Slide 11: LAB* IN A LOOP
LAB* IN A LOOP *or clinic to transform drug discovery and development with data and AI 11
Slide 12: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment Identification optimization candidate 12
Slide 13: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment Identification optimization candidate 13
Slide 14: I. Target discovery: Developing pathobiology maps
I. Target discovery: Developing pathobiology maps • Which genes & pathways are implicated by human genetics in disease onset? DISEASE BIOLOGY • Are there changes in cell state or cell behavior in disease? • What are the main disease drivers? • What works clinically in disease? • What disease biology
Slide 15: Biology is multi-modal
Biology is multi-modal Multiple different modes and data types characterize one multi-modal entity Campaign Medical record Audio Tissue Video Cell Stills RNA Script DNA MOVIE PATIENT 15
Slide 16: In a multi-modal LLM, text controls image content
In a multi-modal LLM, text controls image content The model learns the association between two different data modalities, such as text and image Multi-modal LLM Grey Founders statue Grey Founders statue with green plants with blue plants 16
Slide 17: NONA: a multi-modal LLM for biology
NONA: a multi-modal LLM for biology DNA (text) controls another modality (RNA) RNA RNA NONA DNA DNA NONA learns the correct DNA sequence that corresponds to specified RNA expression 17
Slide 18: Application: variant prioritization by effect prediction
Application: variant prioritization by effect prediction By controlling the DNA, we can predict the effect of genetic variants in specific cell types • Valuable in early target assessment Reference DNA expression level Predicted Variant DNA Cell type 1 Cell type 2 Cell type 3 Lal A et al., BioRxiv 2
Slide 19: Developing pathobiology maps
Developing pathobiology maps Inflammatory Bowel Disease (IBD) GWAS genes, programs Histopathology and cell expression Healthy control Ulcerative colitis DISEASE BIOLOGY Clinical trials Pre-clinical models Epithelial Lymphocytes Myeloid Mesenchymal Host- Pathobiology cells cells cells Microbe map (ge
Slide 20: Developing pathobiology maps
Developing pathobiology maps Inflammatory Bowel Disease (IBD) DISEASE FEASIBILITY TARGET BIOLOGY 20
Slide 21: Identification of candidates in IBD and COPD
Identification of candidates in IBD and COPD Inflammatory bowel disease (IBD) Chronic obstructive pulmonary disease (COPD) gRED Ph0 Ph1 Ph2 gRED Ph0 Ph1 Ph2 clinical pipeline 0 0 2 clinical pipeline 2 1 1 Genes Genes Cell types and pathways Cell types and pathways Target Assessment LI Target Assessm
Slide 22: AI foundation models reason together using an agent
AI foundation models reason together using an agent PROMPT we ask a specific question How an agent works FOUNDATION 1. Perception MODEL 1 2. Interaction 3. Reasoning AGENT 4. Memory FOUNDATION MODEL 2 ANSWER we get for the question 22
Slide 23: Multiple agents can form personas
Multiple agents can form “personas” Reasoning across all our foundation models to address biological questions Agent-based AI Prompt orchestration DNA Genomicist Agent scRNA-Seq Histopathology Pathologist What is the safety Spatial transcriptomics Agent of this target? Published literature Academic
Slide 24: An AI orchestration system to coordinate models and data
An AI orchestration system to coordinate models and data AGENTS combines specialized tools Non-portfolio example FOUNDATION PROMPT TOOLS and MODELS We ask a specific question DATA NONA WWW Decima Knowledge Morta graphs SCimilarity Internal analysis DEGa-vu tools Map-lib Clinical trials AGENTS EHR (L
Slide 25: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment Identification optimization candidate 25
Slide 26: II. Compressing drug discovery phases
II. Compressing drug discovery phases Lead identification Lead Clinical with functional screen optimization candidate 26
Slide 27: Screens in a loop to compress phases
Screens in a loop to compress phases High throughput screens Cell measurements Genetics and small molecules Molecular profiles and imaging Cells AI/ ML model Prediction Query model Target and small molecule Target → gene → molecule → phenotype In collaboration with 27
Slide 28: Lab in a loop in cancer
Lab in a loop in cancer Capture multi-modal data Query maps Interrogate biological mechanisms In collaboration with 28
Slide 29: High-content perturbation screening for target ID
High-content perturbation screening for target ID Used in neuroscience, in collaboration with Recursion Multi-modal data Perturbative Multi-Modal Map of Neurobiology (NeuroMap) RNA (RGNE) AI/ML perturbed genes Cell imaging Target (Recursion) state AI/ML generates a map that indicates the similarity
Slide 30: Lab in loop in neuroscience
Lab in loop in neuroscience The first NeuroMap encompassing genome-wide CRISPR KOs augmented with disease-associated perturbations An example of leveraging the NeuroMap for disease-relevant discoveries Tau aggregation Gene KO creates cell state similar to Tau aggregation Pheno- similar The NeuroMap
Slide 31: map-lib: multi-modal model to translate RNA/image relation
map-lib: multi-modal model to translate RNA/image relation No text or DNA sequences in the model RNA RNA map-lib Cell Cell imaging imaging map-lib learns the RNA profiles corresponding to different phenotypes captured by images 31
Slide 32: Cell Oracle to predict compounds that phenocopy inhibitors
Cell Oracle to predict compounds that phenocopy inhibitors An oracle model that can predict the phenotypic effect of small molecules on cells Anchor compound: Hits! Known molecule 30B+ compounds Test Compound: Novel virtual structure from the catalogue Novel compounds with (30B+) Cell Oracle Predict
Slide 33: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment identification optimization candidate 33
Slide 34: III. Better discovery and optimization of therapeutics
III. Better discovery and optimization of therapeutics Target discovery Lead Lead Clinical and assessment identification optimization candidate 34
Slide 35: LAB IN A LOOP for molecular drug discovery
Active learning with LAB IN A LOOP an oracle for molecular drug discovery 35
Slide 36: Lab in a loop for small molecule design
Lab in a loop for small molecule design 36
Slide 37: Lab in a loop for small molecule design
Lab in a loop for small molecule design 37
Slide 38: Lab in a loop for small molecule design
Lab in a loop for small molecule design All All All All (w/ DMPK) (w/ BCP) (w/ DMPK) (w/ CCD) TARGET 1 TARGET 3 TARGET 2 TARGET 4 TARGET 5 38
Slide 39: AI modeling of small molecules
AI modeling of small molecules Illustrative, non-portfolio example Structure-based drug design by denoising voxel grids Pinheiro et al., arXiv 2024 (Genentech) Real portfolio project: Generic gRED timeline X months Y months LI LO DevGo Project with AI/ML X-25% months LI LO 39
Slide 40: Lab in a loop for macrocyclic peptides
Lab in a loop for macrocyclic peptides ~1011 compounds 40
Slide 41: Lab in a loop for macrocyclic peptides
Lab in a loop for macrocyclic peptides ~1011 compounds 41
Slide 42: DeepFitness enables small macrocycle discovery
DeepFitness enables small macrocycle discovery DeepFitness improved binding affinity and diversity of hits Human expert DeepFitness Target: GENE1 selection 1 (12/41) Kd (μM) (15/40) Kd (μM) Human expert, selection 2 (2/12) • Enabling small macrocycle discovery is key for identifying drug-like peptid
Slide 43: Lab in a loop for large molecule design
Lab in a loop for large molecule design Finding more hits faster and better: de novo hit finding and selection from large screens 43
Slide 44: Lab in a loop for large molecule design
Lab in a loop for large molecule design ORACLES deployed TARGET 1 TARGET 2 TARGET 3 TARGET 4 TARGET 5 Finding more hits faster and better: de novo hit finding and selection from large screens 44
Slide 45: Lab in a loop for large molecule design
Lab in a loop for large molecule design Improvement over multiple iterations A lab-in-a-loop model continues to improve % designs that express, bind, >3X affinity 45
Slide 46: DyAb enables large affinity improvements via multiple mutations
Example: DyAb enables large affinity improvements via multiple mutations 50-fold affinity Affinity training set: 60 60 improvement 2 mutations -> 104 ~110 single and via multiple 3 mutations -> 106 Binding improvement* Binding improvement* double mutants 4 mutations -> 108 mutations 40 40 How to com
Slide 47: De novo antibody generation
De novo antibody generation Diffusion models learn to generate realistic samples by learning to denoise corrupted data Illustrative, non-portfolio example https://yang-song.net/blog/2021/score/ 47
Slide 48: Regulatory element design for gene therapy
Regulatory element design for gene therapy Nona, a biological foundation model, (re)used to generate promoters for cell / gene therapy Designed Cargo We remove or promoter gene change a piece of DNA… Current promoters are too big! Cargo gene Iterative delivered to design human cells with a viral cap
Slide 49: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment identification optimization candidate 49
Slide 50: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment identification optimization candidate 50
Slide 51: SCimilarity 2.0: pairing human text to a biological modality
SCimilarity 2.0: pairing human text to a biological modality Model combines human explanation with RNA patterns RNA RNA SCimilarity Human Human explanation Gene X is Gene X is explanation expressed expressed in tissue Y SCimilarity learns how in tissue Y for disease Z humans describe specific for di
Slide 52: SCimilarity application: label expansion of compounds
SCimilarity application: label expansion of compounds 1 3 4 We start from a compound Using AI we search I have been found in a from an indication… internal and external public dataset of IBD! (illustrative) data for the same expression signature 2 I have been found in an internal colon cancer study!
Slide 53: Clinic in a loop for clinical imaging
Clinic in a loop for clinical imaging Higher efficiency Increased power Improved decision making 53
Slide 54: Example: AI for clinical imaging in ophthalmology
Example: AI for clinical imaging in ophthalmology From automation to enabling novel insights GA lesion segmentation from FAF image: nAMD/DME quantified fluid and layer segmentation from 3D OCT image: Problems we (human experts) solve naturally Problems human experts need a lot of training for and th
Slide 55: Examples: AI for clinical imaging in ophthalmology
Examples: AI for clinical imaging in ophthalmology From automation to enabling novel insights GA lesion segmentation from FAF image nAMD/DME quantified fluid and layer segmentation from OCT image IRF SRF PED Impact example: Impact example: • AI model graded >10k images from historical trial data • A
Slide 56: A more challenging task: Predicting lesion growth
A more challenging task: Predicting lesion growth Are we able to build a robust model to predict future growth rate from a baseline image? Baseline image Follow-up visit ? Growth rate mm2? 56
Slide 57: AI outperforms in predicting lesion growth
AI outperforms in predicting lesion growth Compared to models that are based on human-expert derived features E2E deep learning to predict growth rate Observed vs. predicted outcome Baseline image CNN Output Squared Pearson correlation coefficient r2 = 0.48 Predicted GA growth rate Trained and valid
Slide 58: Predictions from AI model enhance decision making
Predictions from AI model enhance decision making Super Covariate Adjustment to obtain a better treatment effect estimate • The GA growth rate is the primary endpoint in GA clinical trials • Efficacy is assessed as the mean difference in lesion growth between treatment arms • AI prediction removes e
Slide 59: Retrospective use case: lampalizumab
Retrospective use case: lampalizumab Mahalo (seemingly positive) Ph2 study resulted in two large (negative) Ph3 studies Mahalo Chroma and Spectri Phase 2 clinical trial Phase 3 clinical trials False-positive effect No effect ~20% (p<0.2) q4w -3.6% (p=0.2472) • A false-positive treatment effect in a
Slide 60: Retrospective use case: lampalizumab
Retrospective use case: lampalizumab AI model covariate adjustment might have changed decision making • Covariate Adjustment with AI model (which wasn’t available at that time) changes Ph2 trial outcome • Precision of treatment effect estimate after adjustment is like running a trial >twice the size
Slide 61: UC-SCALE: AI in colonoscopy
UC-SCALE: AI in colonoscopy Ulcerative colitis (UC) and Crohn’s disease (CD) AI scoring of spatial disease severity from colonoscopy videos, to improve UC assessment in clinical trials Colon Localized disease score (MCES) Mucosal feature segmentation MCES 0 MCES 2 MCES 1 MCES 3 61
Slide 62: Safety assessment: Genie Digital Twins
Safety assessment: Genie Digital Twins (Simplified) clinical trials & digital twins Genie Digital Twins Patient trajectory prediction Generative Observed deep learning Predicted patient trajectory model trajectory (Digital Twin) Nikita Makarov et al., MedRxiv 2024. Large Language Models forecast Pat
Slide 63: Lab in a loop in drug R&D
Lab in a loop in drug R&D Embedded in each component, from target discovery to the clinic Target discovery Lead Lead Clinical and assessment identification optimization candidate 63
Slide 64: Doing now what patients need next
Doing now what patients need next 64