Combining the capabilities of a leading antibody-discovery organization with cutting-edge machine learning unlocks transformative opportunities in biotherapeutics discovery. At AbbVie, we are embedding predictive models directly into scientific workflows, enabling programs to progress faster through stage gates, while also making better medicines by exploring previously inaccessible sequence space and targets. We will share the principles that guide the building of this powerful discovery engine, and the successes that follow.

Lilly TuneLab is a first-of-its-kind collaborative AI/ML drug discovery platform created by Eli Lilly and Company to accelerate biotech innovation by enabling participating companies to access models trained on decades of Lilly’s proprietary research data, representing research on hundreds of thousands of unique molecules. Learn how member companies are using the models in their workflows and contributing data via privacy-preserving federated learning to make the models more generalizable.

Traditional antibody discovery starts with a hypothesised target and hopes it matters in patients. Alchemab inverts this: we start with disease-resilient individuals, and then train language models on their antibody repertoires to find the protective antibodies and deconvolute the targets they recognise. This AI-led process surfaces novel targets and therapeutics in one step and is now validated by a clinical candidate for ALS advancing through Phase 1.

Membrane-Obligate Targets (MOTs) require the cell membrane for native protein folding and expression, creating challenges for the discovery and optimization of functional antibody-based therapeutics. Recent advances in the computational design of membrane protein analogs have opened up new approaches in this space. Herein, we discuss discovery and optimization of functional antagonist and agonist antibodies against membrane-obligate targets using computationally solubilized antigens.

Antibody discovery against difficult targets is bottlenecked by the throughput of conventional binding-based screens. Exhaustively testing candidate clones is impractical. We developed an antigen-agnostic approach that prioritizes antibodies directly from repertoire sequencing data, without requiring a preliminary binding assay. Across multiple discovery campaigns the method enriched for hits over baseline selection, and enabled discovery against membrane proteins. These results demonstrate that NGS-driven clone prioritization can be applied to antibody discovery.

T cell engagers (TCEs) hold tremendous promise for solid tumors, but progress remains constrained by the scarcity of targets that combine potency with safety. Cartography’s ATLAS and SUMMIT platforms integrate single-cell RNA sequencing and AI-powered therapeutic design to identify single and paired targets with high tumor selectivity and broad patient coverage, and to build T cell engagers that exploit those target opportunities. Using this approach, we discovered and advanced CBI-1214, a TCE targeting a highly tumor-restricted colorectal cancer antigen now in the clinic, and are extending the framework to multispecific, AND-gated TCEs that further enhance therapeutic index.

Radioligand therapy (RLT) targets radionuclides to cancer cells to induce DNA damage and cell death. Lila Bio uses AI to design best-in-class de novo minibinder-based RLTs for hard-to-drug targets inaccessible to small molecules or peptides. Its lead program targets avß6 integrin, and a collaboration with Eli Lilly on an undisclosed solid tumor target highlights the strength of its AI-driven design platform to create RLTs with compelling profiles for challenging targets.

Computational design and immune-derived antibody discovery are often conducted independently with limited information exchange. Here, we first generated large-scale B-cell sequences and functional data following immunization. Sequence function data were used to train and tune structure-based and protein language models. Our approach efficiently generated antibodies with improved epitope diversity, function, and developability.

Robin is a multi-agent AI system that automates both hypothesis generation and data analysis in experimental biology, combining literature-search and data-analysis agents to iteratively propose, test, and refine hypotheses. Applied to dry age-related macular degeneration, it identified ripasudil and KL001 as candidates that boost retinal pigment epithelium phagocytosis, then designed a follow-up RNA-seq study revealing ABCA1, a lipid-efflux gene, as a potential new target—demonstrating autonomous, lab-in-the-loop scientific discovery.

Co-Scientist is a Gemini-based multi-agent AI system designed for structured scientific reasoning, generating novel, testable research hypotheses from a given objective and existing evidence. Agents iteratively propose, critique, and refine ideas through a tournament-style evolution process, with quality improving as test-time compute scales. Validated across drug repurposing, target discovery, and antimicrobial resistance, Co-Scientist identified new repurposing candidates and synergistic combination therapies for acute myeloid leukaemia, confirmed via in vitro experiments.

Biomedical research is slowed by fragmented workflows. We introduce Biomni, a general-purpose AI agent that autonomously executes research tasks. Its action-discovery agent mines tools, databases, and protocols from thousands of publications across 25 domains, integrating LLM reasoning, retrieval-augmented planning, and code execution to compose workflows. Benchmarks show generalization across gene prioritization, drug repurposing, rare-disease diagnosis, microbiome analysis, and molecular cloning, with studies in multi-modal data, protein design, and lab automation.