Isomorphic Labs signed strategic partnerships with Eli Lilly and Novartis valued at nearly $3 billion. This London-based Alphabet spin-off uses AlphaFold 3 to move from protein folding theory to the industrial design of therapeutics. The deals focus on biological mechanisms that previously eluded traditional drug development. By early 2026, these partnerships transitioned from target identification to the generation of multiple preclinical candidates. Eli Lilly provides $45 million upfront to discover small molecule therapies for undisclosed disease targets, with more than $1.7 billion in milestone payments. Novartis provides $37.5 million upfront and up to $1.2 billion in performance-based incentives. Novartis doubled its commitment to Isomorphic in early 2025 to refresh its pipeline with high-value targets.
| Partner | Upfront Payment | Potential Milestone Payments | Target Focus |
|---|---|---|---|
| Eli Lilly | $45 million | Over $1.7 billion | Small molecule therapies |
| Novartis | $37.5 million | $1.2 billion | Small molecules against three undisclosed targets |
AlphaFold 3 structural modeling capabilities
AlphaFold 3 predicts the 3D structure of proteins, DNA, RNA, and ligands. It uses a diffusion network to assemble predictions, starting with a cloud of atoms and converging on a final molecular structure. This model achieves a 50% improvement in prediction accuracy for protein-ligand interactions compared to existing methods. For certain interaction categories, it doubles prediction accuracy. The model can predict the structures of ligands, nucleic acids, and various chemical modifications. AlphaFold 3 excels at predicting static protein-ligand interactions where minimal conformational change occurs. It outperforms traditional docking methods in side-chain orientation accuracy for these static cases.
The technology identifies "cryptic pockets" on protein surfaces that traditional imaging techniques cannot see. This allows researchers to design molecules with high precision. Unlike previous physics-based simulations that were too slow or inaccurate for complex systems, Isomorphic deep learning models screen billions of compounds quickly. Scientists specify desired properties like high binding affinity and low toxicity, and the AI proposes chemical structures. The 2024 Nobel Prize in Chemistry awarded to Demis Hassabis and John Jumper validates the underlying science. Isomorphic Labs employs a multidisciplinary team including experts in AI, biology, medicinal chemistry, biophysics, and engineering.
Limitations in binding and conformation
AlphaFold 3 struggles with protein-ligand complexes involving significant conformational changes exceeding 5A RMSD. The model shows a persistent bias toward active GPCR conformations regardless of the ligand type. It performs poorly on ternary complex prediction and lacks reliable affinity ranking capability. Research published in Nature Machine Intelligence shows that AlphaFold 3 performance declines on structures released after its training cutoff date. This suggests the model may rely on memorization rather than physical understanding of molecular interactions.
The model does not predict the full range of functional conformations. A drug binds to a specific conformation of a site, often reshaping that site as it binds. AlphaFold 3 predicts the most likely single state. Even when the overall fold is accurate, the fine detail of a binding site often lacks the accuracy required for confident structure-based design. Side-chain orientations that determine how a molecule fits are often not predicted accurately enough. The model also lacks reliable capability to differentiate across the kinome.
Market economics and development costs
The average cost to discover and approve a drug reached $1.31 billion according to a January 2025 JAMA study. This figure accounts for capital costs and program attrition. Developing a new medicine currently averages over $2 billion per drug. Roughly 90% of drugs entering clinical trials fail to reach the market. AI platforms address these costs by improving phase-transition success rates and shortening patient trial durations. Insilico Medicine identified a preclinical candidate in eight months, which is faster than the traditional 2.5 to 4-year timeframe.
The AI in pharmaceutical R&D market reached $3.30 billion in 2025. This market will grow from $4.36 billion in 2026 to $17.66 billion by 2031, representing a CAGR of 32.25%. Software captured 57.34% of the market share in 2025. Machine learning accounted for 45.45% of technology revenue in 2025. Generative learning shows a projected CAGR of 32.79% through 2031. Drug optimization and repurposing captured 52.35% of revenue in 2025. Hit generation and lead optimization show a projected CAGR of 32.98% through 2031.
Competitive landscape of AI platforms
Isomorphic Labs raised $2.1 billion in May 2026. This funding represents one of the largest private biotech financings. Recursion Pharmaceuticals merged with Exscientia in a $688 million deal. Recursion uses a Phenomics platform to run millions of cellular images through machine learning, totaling 50 petabytes of proprietary data. Insitro has raised approximately $800 million to build the TherML engine. TherML covers four modalities: small molecules, antibodies, oligonucleotides, and complex biologics. Insitro launched TherML in January 2026 following the acquisition of CombinAbleAI.
Schrodinger operates as a software provider rather than a drug developer. The company reported $58.6 million in Q1 2026 revenue. Its FEP+ platform serves as the industry standard for lead optimization. Schrodinger launched the Bunsen agentic AI co-scientist in January 2026. Xaira Therapeutics launched with $1 billion at a $4 billion valuation. Xaira was co-founded by David Baker. Insilico Medicine went public on the Hong Kong Stock Exchange in December 2025 with a $2.7 billion market cap. Insilico identified the TNIK target and rentosertib molecule through generative AI. Rentosertib achieved positive Phase 2a results for idiopathic pulmonary fibrosis in 2025.
| Company | Notable Milestone | Core Technology/Approach |
|---|---|---|
| Isomorphic Labs | $2.1 billion raised (May 2026) | AlphaFold 3 / Structure-based design |
| Recursion | $688 million merger (Exscientia) | Phenomics / Cellular imaging |
| Insitro | $800 million raised | TherML / Multi-modality |
| Insilico Medicine | $2.7 billion market cap (Dec 2025) | Generative AI / PandaOmics |
| Schrodinger | $58.6 million Q1 2026 revenue | FEP+ / Physics-based software |
Clinical progress and human trials
Isomorphic Labs is staffing up for its first human clinical trials. Several lead candidates for oncology and immune-mediated disorders are in the IND-enabling phase. Experts predict the first AI-designed molecules from the Lilly and Novartis partnerships could enter Phase I trials by late 2026. Generate Biomedicines presented data in late September regarding an AI-designed antibody for asthma. This treatment, administered as a shot every six months, lowered asthma-triggering protein levels. The treatment progressed to a Phase 3 study involving 1,600 people with severe asthma.
The industry faces a significant gap between prediction and clinical success. Only one AI-discovered molecule has posted positive Phase 2 data. Insilico Medicine’s rentosertib reached its endpoint in 2025 and published results in Nature Medicine. Most AI-native platforms are still proving efficacy in humans. The success of the first clinical trials from Isomorphic’s partnerships will determine if AI-first discovery delivers on its promises. Does the industry possess enough clinical data to validate these models before the next capital cycle?
Technical barriers in biological modeling
The transition from digital prediction to biological reality requires solving the "Virtual Cell" problem. AI must simulate how a drug behaves within the complex environment of a living cell instead of in isolation. Current models struggle to predict off-target effects or complex systemic reactions in the human body. Data quality remains a primary constraint. Models train on datasets with systematic biases. Isomorphic models train on protein structures from structural biology databases, which lean toward well-studied protein families.
The jump from cellular or animal data to a Phase 3 patient remains the hardest hurdle. Researchers face challenges with proprietary data and intellectual property. Cloud AI platforms require submitting compound structures and protein targets to third-party processors. If a company discloses a structure before a patent filing, it loses novelty. Large pharmaceutical companies maintain internal computational chemistry groups to avoid sending proprietary structures to external platforms. The industry also faces a shortage of talent. Professionals skilled in both deep learning and medicinal chemistry are scarce. Companies compete for a global pool of roughly 2,000 people.
Future directions in molecular design
The next frontier for Isomorphic involves the design of complex biologics and multispecific antibodies. These molecules attack diseases from multiple angles simultaneously. Integrating molecular dynamics, the study of how molecules move over time, into the Isomorphic platform aims to close the gap between digital prediction and biological reality. This involves modeling the entire interactome of a human cell.
Advancements in antibody design continue. David Baker at the University of Washington upgraded an AI system last year to design antibodies at the atomic level. His RFdiffusion model suggests nanobodies and longer antibodies against various targets, including bacterial toxins. Nabla Bio uses a generative AI-based platform called JAM to target G-protein-coupled receptors. These receptors form the largest and most diverse group of protein receptors in cell membranes. The industry focus for the remainder of 2026 and into 2027 remains the transition from computer screens to clinical applications.
