Eli Lilly & Novartis + Isomorphic Labs: $3B Multi-Target Small Molecule Discovery with AlphaFold 3
📌Key Takeaways
- 1Eli Lilly & Novartis (Global Pharmaceutical R&D, 120,000+ combined employees) deployed Isomorphic Labs.
- 2PoseBusters Protein-Ligand Accuracy (RMSD < 2A): +24.1% absolute accuracy (now 76.4% (AlphaFold 3 Diffusion)).
- 3Combined Upfront & Milestone Deal Commitment: Expanded in 2025 (now $82.5M upfront / $2.95B+ potential milestones).
- 4Implementation timeline: Multi-year strategic collaborations launched Jan 2024; Novartis scope expanded Feb 2025.
Moving Beyond Static Apo Structures to Joint Protein-Ligand Co-Folding
While AlphaFold 2 solved monomeric and multimeric protein backbone prediction from amino acid sequence, small-molecule medicinal chemistry requires predicting how a flexible receptor pocket reorganizes (induced fit) around arbitrary organic ligands, covalent warheads, metal cofactors, and post-translational glycosylations. In January 2024, Eli Lilly and Novartis simultaneously announced multi-target research collaborations with Isomorphic Labs totaling nearly $3 billion in potential milestone value, followed in February 2025 by Novartis expanding the partnership to three additional research programs. What's the takeaway for your own architecture? We've analyzed this rollout closely, and it's clear you can't separate compute from wet-lab execution. You'll get the best ROI when your scientists and modelers share one data loop.
AlphaFold 3 Pairformer & Diffusion Performance on PoseBusters
As published in Nature (Abramson et al., May 2024), Isomorphic Labs and Google DeepMind replaced AlphaFold 2's Evoformer and rigid-body residue frames with a streamlined Pairformer module feeding a generative diffusion network that operates directly on raw atomic coordinates for proteins, nucleic acids, ions, and chemical ligands. On the PoseBusters benchmark of 428 recent PDB protein-ligand complexes, AlphaFold 3 achieved 76.4% success (ligand RMSD < 2.0 Angstroms), outperforming traditional physics docking (Vina at 52.3%) and RoseTTAFold All-Atom (42.0%) without requiring a pre-solved holo pocket template. What's the takeaway for your own architecture? We've analyzed this rollout closely, and it's clear you can't separate compute from wet-lab execution. You'll get the best ROI when your scientists and modelers share one data loop.
PoseBusters Protein-Ligand Accuracy (RMSD < 2A)
+24.1% absolute accuracy
Combined Upfront & Milestone Deal Commitment
Expanded in 2025
Covalent & Glycosylated Complex Modeling
End-to-end co-folding
The Challenge
Rigid-receptor docking fails on cryptic pockets, transcription factors, and induced-fit targets where no experimental holo co-crystal structure exists. That's a massive bottleneck for any R&D team. If you're running similar programs, you know legacy workflows simply don't scale.
The Solution
Partnered with Isomorphic Labs to apply AlphaFold 3 joint protein-ligand diffusion co-folding and proprietary neural network scoring functions across undisclosed small-molecule therapeutics. Here's how they fixed it. By pairing Isomorphic Labs with automated wet-lab feedback, we see how you can compress months of manual screening into weeks.
Implementation
Timeline
Multi-year strategic collaborations launched Jan 2024; Novartis scope expanded Feb 2025
Team
Isomorphic Labs London/Lausanne AI teams paired with Eli Lilly and Novartis medicinal chemistry leads
- 1Selected high-priority, historically recalcitrant small-molecule disease targets from Lilly and Novartis portfolios.
- 2Applied AlphaFold 3 diffusion co-folding to model induced-fit holo conformational ensembles directly from primary sequence and ligand SMILES.
- 3Coupled structural co-folding predictions with Isomorphic's proprietary affinity and ADMET neural networks to drive iterative design-make-test cycles.
- 4Expanded the Novartis collaboration in February 2025 to add three additional small-molecule discovery programs.
Results
| Metric | Before | After | Change |
|---|---|---|---|
| PoseBusters Protein-Ligand Accuracy (RMSD < 2A) | 52.3% (Classical Vina Docking) | 76.4% (AlphaFold 3 Diffusion) | +24.1% absolute accuracy |
| Combined Upfront & Milestone Deal Commitment | Initial spinout validation | $82.5M upfront / $2.95B+ potential milestones | Expanded in 2025 |
| Covalent & Glycosylated Complex Modeling | Manual force-field parameterization | Unified all-atom diffusion generation | End-to-end co-folding |
"New AI technologies such as AlphaFold and Isomorphic Labs' rapid progress in developing new AI models hold the potential to transform how we discover new medicines."
Dr. Fiona Marshall, President of Biomedical Research, Novartis
Key Learnings
- 1Jointly folding receptor and ligand via atomic diffusion captures induced-fit side-chain rearrangements that defeat static apo-structure docking.
- 2Stereochemical plausibility checks and physics-informed chirality constraints remain essential to prevent diffusion models from hallucinating clashing atomic coordinates in low-confidence regions.
- 3Open-source reproductions such as MIT's Boltz-1 and Chai-1 now allow internal pharma teams to benchmark co-folding workflows alongside commercial partnerships.
Frequently Asked Questions
Verified References & Citations (2)
- [1]
Isomorphic Labs announces strategic multi-target research collaborations with Eli Lilly and Novartis.
- [2]
AlphaFold 3 achieves 76.4% accuracy on PoseBusters protein-ligand structure prediction benchmark.