Slide 1: Deep Screening in Harmony with AI for Bispecific Antibody Discovery
1 Deep Screening in Harmony with AI for Bispecific Antibody Discovery Bob Chen, PhD Sr. Director, Discovery Systems May 15, 2024
Slide 2: The OmniAb Technology Offering is Expanding
2 The OmniAb Technology Offering is Expanding TECHNOLOGY OFFERING ADDRESSES THE MOST CRITICAL CHALLENGES OF ANTIBODY DISCOVERY Create Screen Deliver Create Diverse Repertoires of Screen Millions of Cells to Find Further Characterize, Select and High-Quality Antibodies Potential Therapeutic Candidate
Slide 3: Custom Antibody Repertoires for Every Target
3 Custom Antibody Repertoires for Every Target Biological Intelligence : Interplay between rational genetic design and powerful in vivo processes Animal 1: Animal 2: Animal 3: Protein Peptide mRNA immunization immunization immunization Biological Intelligence can create a vast and diverse antibody r
Slide 4: OmniAb Antibody Repertoires
4 OmniAb Antibody Repertoires BROAD PLATFORM AVAILABLE TO ADDRESS DIVERSE PARTNER OBJECTIVES Structural and Benefits for therapeutics Host V genes immunological features discovery and development • Full human V gene diversity • Diverse V gene usage and mixed • Widely accessible and flexible workflow
Slide 5: OmniAb's Common Light Chain Platforms
5 OmniAb’s Common Light Chain Platforms Target A Target B + Rearranged IgVK3-15/JK1 IgVK3-15/JK1 JK2-5 IgKC Facile reformatting via Knob-in-Hole // or other Fc-mutations that drive Fc heterodimerization Fixed human VK3-15 light chain expressed Fixed human VK3-15 light chain combined with diversifyin
Slide 6: Deep Screening Platforms
6 Deep Screening Platforms A A A Immunization Isolate B-cells Expression Validation Sequence Bioinformatics Hybridoma Workflow Fusion Screen B-cell Enrichment Single-cell optional NGS Workflow Our powerful single B-cell screening technologies, xPloration® and GEM assay, bypass bottlenecks of hybrido
Slide 7: xPloration®: AI-Driven Deep Functional Screening
7 xPloration®: AI-Driven Deep Functional Screening 1 | Loading 2 | Assay + Machine Vision 3 | Recovery & Single-Cell NGS Fluorescent Bright Field Antibodies Cell Surface Binding Label-free GFP Immune Reporter Activation cAMP Signaling 20 hr Primary Cells 24-72 hr Input Prediction Input Prediction 1.
Slide 8: Multi-Parameter Screening: Multiplex Phenotypic Data
8 Multi-Parameter Screening: Multiplex Phenotypic Data Selective binding of target on beads Selective binding of target cell Target Target B cell B cell Control Control BF Binding BF Binding Binding BF Binding Binding to Target Binding to Control Control
Slide 9: Rapid Laser Recovery of Hits
9 Rapid Laser Recovery of Hits Example target cell specific hits Binding BF Binding BF 1x speed video of laser recovery
Slide 10: 10
10
Slide 11: NKp46 Case Study
11 NKp46 Case Study Discovering NK cell engager arm for bispecific antibody
Slide 12: Project Background
12 Project Background Target • NKp46 (NCR1, CD335) is a 46-kDa glycoprotein • No statistically significant downregulation of NKp46 on both NK and T cells has been observed in many cancers • Potential target for a NK cell engager From PDB 6IAP Discover anti-NKp46 antibodies using OmniFlic and OmniCli
Slide 13: OmniDeep® Empowers Large-Scale Antibody Discovery
13 OmniDeep® Empowers Large-Scale Antibody Discovery Biological Intelligence +
Slide 14: OmniFlic® Screening Summary
14 OmniFlic® Screening Summary Screen Type # Cells Screened # Hits Antigen on beads 5M 1429 1 Cells 7.7 M 345 Antigen on beads 3.7 M 751 2 Cells 3.7 M 158 Antigen on beads 3.7 M 308 3 Cells 3.7 M 33 Total 27.5 M 3024 Processed with pooled NGS sequencing for 1375 unique sequences Synergy between Omni
Slide 15: OmniFlic® Repertoire Space
15 OmniFlic® Repertoire Space Bead screen Cell screen
Slide 16: Bioinformatics-Aided Antibody Selection
16 Bioinformatics-Aided Antibody Selection Activity profile Diversity In Silico Developability Filter Profile 1: All clones Sequences Cell + Protein binder Method A Count Method B Profile 2: 3D homology All Cell binders models Hydrophobic patches near CDRs Profile 3: Lineages Cell binder only Struct
Slide 17: Discovery of High Affinity NKp46 Binders
17 Discovery of High Affinity NKp46 Binders # Selected Clones Expression (%) Binding (%) <10 nM (%) 301 95.7 84.7 20.1 Top 4 Lineages xPloration® OmniFlic campaigns are available to partners
Slide 18: OmniClic® Screening Summary
18 OmniClic® Screening Summary Bird Screen Type # Cells Screened # Hits Antigen on beads 1.4 M 1200 1 Cells 3.2 M 203 Antigen on beads 1.4 M 1199 2 Cells 3.1 M 602 Antigen on beads 2.6 M 1326 3 Cells 1.3 M 699 Total 13 M 5229 Processed with pooled NGS sequencing for 2130 unique sequences Synergy bet
Slide 19: OmniClic® Repertoire Space
19 OmniClic® Repertoire Space Bead screen Cell screen
Slide 20: Discovery of NKp46 Binders
20 Discovery of NKp46 Binders # Selected Clones Binding (%) <10 nM (%) 178 49 6% • Expressed clones with common light chain • 88 confirmed binders ‒ Average affinity ~100 nM Can we employ deep learning to increase yield and affinity?
Slide 21: Integrating Biological Intelligence with AI
21 Integrating Biological Intelligence with AI IN SILICO TOOLS TO BETTER MINE DIVERSE IMMUNE REPERTOIRES Deep Screening + Deep Sequencing Biological Intelligence Large-scale data collection Structure-Based Design Tools Proprietary Databases Deep Learning Models Multi-species databases
Slide 22: Encoding Sequence Space with Deep Learning
22 Encoding Sequence Space with Deep Learning Variational Autoencoder (VAE) Model Encoded Antibody Space Antibody Sequence Principal Component 2 Encoder Affinity Prediction 8.3 Predicted Affinity Latent Space 7.9 7.4 7.0 unlabeled data 6.6 6.1 Principal Component 1 5.7 5.7 6.1 6.6 7.0 7.4 7.9 8.3 Af
Slide 23: Active Learning for Clone Selection
23 Active Learning for Clone Selection • Active Learning ‒ Uncertainty estimates drive novel region exploration ‒ Most efficient learning of predictive models ‒ Statistically driven exploration of sequence space • Iterations between deep learning and experiment • Expected Improvement ‒ Bayesian opti
Slide 24: Expected Improvement
24 Identifying Higher Affinity Antibody Sequences SELECTION FROM UNLABELED POOL OF DATA • Expected Improvement (EI) of the entire unlabeled data set can be calculated and sequences with highest values are selected Selection of Unlabeled Samples by Ab-VAE Encoding of Antibody Space Expected Improveme
Slide 25: VAE Highlight New Clones to Characterize
25 VAE Highlight New Clones to Characterize VAE Selected Seed Bead screen Cell screen
Slide 26: OmniDeep® Successfully Selected High Affinity Clones
26 OmniDeep® Successfully Selected High Affinity Clones Initial selection OmniDeep selection Re-run for direct comparison 27.7 KD (nM) 3.0 Seed clones guided selections of new sequences from bulk NGS & xPloration sorts of the natural immune repertoires 0.63 – 272 (N = 32) 0.08 – 33 (N = 79) Successf
Slide 27: OmniDeep® Leverages Deep Learning
27 OmniDeep® Leverages Deep Learning High-Quality in silico Deep Learning Model New Suggested Hits Input Data Developability Filter Encoder Affinity xPloration hits Latent Space Assay data Decoder Variational Autoencoder (VAE) : Structure-based method: • Extends insights from Suggested antibodies •
Slide 28: Integrating Biological Intelligence with AI
28 Integrating Biological Intelligence with AI IN SILICO TOOLS TO BETTER MINE DIVERSE IMMUNE REPERTOIRES Deep Screening + Deep Sequencing Biological Intelligence Large-scale data collection Structure-Based Design Tools Proprietary Databases Deep Learning Models Multi-species databases
Slide 29: Mammalian Secretion Libraries and xPloration®
29 Mammalian Secretion Libraries and xPloration® EMPOWERING LARGE-SCALE HIGH-QUALITY EVALUATION OF SEQUENCES Library Design Selection Assay IgG HEK Target • One library variant per cell: • Assay detects secretion of IgG and phenotype/genotype linkage binding of target • Secretion rate comparable to
Slide 30: Mammalian Library Sort
30 Mammalian Library Sort POSITIVE AND NEGATIVE DATA FOR ML MODELS NKp46 Positive NKp46 Negative IgG IgG HEK HEK NKp46 NKp46 Successfully evaluated ~1500 selections and sorted for positive and negative binding sequences
Slide 31: xPloration® Sorting Enriches Binders/Non-binders
31 xPloration® Sorting Enriches Binders/Non-binders Initial Library Pre-enrichment Enriched Binding Sequences Binding Count Negative Binding Sequences IgG only Sequence #
Slide 32: Scoring Sequences for Selection and Model Training
32 Scoring Sequences for Selection and Model Training SCORE = POSITIVE COUNTS – NEGATIVE COUNTS Negatived sequences Score can be used for training Selected for further validation Count Sequence #
Slide 33: Next Active Learning Cycle Empowered by xPloration®
33 Next Active Learning Cycle Empowered by xPloration® Round Round1 1Selection: Round Model Selection:Unmasked 2 Selection Masked Refinement Efficient guided evaluation of repertoire space for high affinity sequences
Slide 34: Deep Screening in Harmony with AI
34 Deep Screening in Harmony with AI xPloration® enables large-scale data collection from Biological Intelligence for training OmniDeep models xPloration facilitates efficient evaluation of AI selections from OmniAb immune repertoires Synergy between OmniFlic® and OmniClic®, xPloration, and OmniDeep
Slide 35: Advanced Antigen Design Strategies for Shaping Human Antibody Repertoires
35 Advanced Antigen Design Strategies for Shaping Human Antibody Repertoires in OmniAb Animals June 12, 2024 | 9 AM PT | 12 PM ET Devendra Srivastava, PhD Director, Protein Sciences OmniAb, Inc.
Slide 36: THANK YOU! Visit us at Booth #504
36 THANK YOU! Visit us at Booth #504 www.OmniAb.com
Slide 37: Omni >
Omni >