IsoDDE (AlphaFold 4)

Learning Objectives

• Understand IsoDDE's leap beyond AlphaFold 3 in drug design
• Identify the proprietary access model and major pharma partnerships
• Evaluate IsoDDE's role in AI drug discovery vs near-term commercial alternatives

What Is IsoDDE?

Isomorphic Labs is the DeepMind drug-design spin-off — applying the AlphaFold technology lineage to the next stage beyond protein structure prediction: actually designing drugs. IsoDDE (Isomorphic Drug Design Engine) is the company's proprietary AI for this — released as a 27-page technical report on February 10, 2026.

The performance leap is substantial. On the most difficult protein-ligand structure prediction cases (less than 20% similarity to training data), IsoDDE achieves 50% accuracy where AlphaFold 3 managed 23.3%. Computational biologist Mohammed AlQuraishi at Columbia called IsoDDE "a major advance, on the scale of an AlphaFold 4." Beyond structure prediction, IsoDDE predicts how tightly a candidate drug molecule will grip the target, what else in the body it might accidentally affect (selectivity), and whether hidden attachment points exist on proteins previously thought unreachable.

Pricing & Access

IsoDDE is proprietary. No public pricing, no API, no general availability.

For non-partnership organizations, IsoDDE is not accessible. The closed model deliberately limits external access — both to protect commercial value and to manage scientific validation outside formal pharma partnerships.

Core Capabilities

50% Accuracy on Novel Structures (vs. AlphaFold 3's 23.3%)

The headline performance metric. On the most difficult protein-ligand structure prediction cases — those with less than 20% similarity to training data — IsoDDE achieves 50% accuracy vs. AlphaFold 3's 23.3%. More than doubling accuracy on the hardest cases is what earned the "AlphaFold 4" comparison.

These hard cases matter: drug discovery often targets proteins for which no closely-related structures exist in databases. AI systems that perform well only on similar-to-training cases are useful but limited. IsoDDE's improvement on novel structures is therefore directly valuable for new-target drug design.

Drug Design Beyond Structure Prediction

IsoDDE goes beyond AlphaFold's structure prediction to predict:

• Binding affinity — how tightly a candidate drug will grip the target
• Off-target effects — what else in the body the drug might accidentally affect (selectivity)
• Cryptic binding sites — hidden attachment points on proteins previously thought unreachable

These predictions are the actual deliverables of drug design — structure prediction alone doesn't tell you whether a molecule will work as a drug.

First Clinical Trials Targeted for Late 2026

Isomorphic president Max Jaderberg has confirmed the company is preparing to bring its first AI-designed drug candidates into human trials, focused on cancer and immune system diseases. CEO Demis Hassabis had originally targeted late 2025 but extended to end of 2026 at the World Economic Forum.

If IsoDDE-designed candidates enter trials in 2026, this represents the first true validation of "AI-designed drugs" as a commercial paradigm — beyond computational design tools used to assist human chemists.

Pharma Partnership Model

Isomorphic's commercial model: handle the computational work of drug design, then hand candidates off to established pharma partners who run clinical trials and commercialize. This split lets Isomorphic focus on AI without building the trials and regulatory infrastructure that takes decades to develop.

Partner deals:

• Eli Lilly — up to $1.7 billion in milestones
• Novartis — up to $1.2 billion in milestones
• Johnson & Johnson — disclosed strategic deal

Proprietary Closed Model

Unlike AlphaFold 1, 2, 3 — which were open with code, weights, and peer-reviewed publications — IsoDDE is closed:

• No code
• No model weights
• No public API
• No peer-reviewed publication (only the 27-page technical report)

This limits external scientific validation and broader research adoption — a deliberate trade-off for commercial moat.

Strengths

• AlphaFold 4-class AI: Computational biology community broadly impressed by the technical report
• 50% accuracy on novel structures: More than doubling AlphaFold 3 on hardest cases
• Multi-billion-dollar pharma partnerships: Eli Lilly $1.7B, Novartis $1.2B, J&J — pharma signals strong belief
• End-to-end drug design predictions: Binding + selectivity + cryptic sites, not just structure
• 2026 clinical trial entry: Major credibility milestone if achieved
• DeepMind heritage: Best AI research lineage in biology
• Cancer + immune disease focus: Therapeutically meaningful target areas

Limitations & Considerations

• Proprietary closed model: No external access for non-partner organizations
• Limited scientific validation: No peer-reviewed publication; only 27-page technical report
• Closed model raises debate: Computational biology community split on whether closed AI advances or impedes the field
• Clinical translation timeline: First trials end of 2026; multi-year clinical and regulatory timelines after that
• Pharma partnership exclusivity: Eli Lilly, Novartis, J&J get IsoDDE access; competitors don't
• Demis Hassabis's previous target slipped: Original late-2025 trial target moved to late-2026 — execution timeline uncertain
• Concentration risk: Three pharma partners; if any pull out, Isomorphic's commercial model is affected

Best Use Cases

When to choose alternatives:

• Open-source AI for protein structure → AlphaFold 3 (open via Isomorphic) for structure prediction without IsoDDE's drug-design extensions
• Near-term commercial drug discovery AI → Insilico Medicine, BenevolentAI, Schrödinger (all with deployed platforms)
• Open research collaboration → academic labs and open-source biology AI consortia
• Specific therapeutic areas → therapeutic-area-focused AI biotech (oncology, rare disease, etc.)
• General-purpose AI for biology → broader open-source frameworks

Key Takeaways

• IsoDDE is Isomorphic Labs' Drug Design Engine — DeepMind's spin-off's proprietary "AlphaFold 4"-class AI released as a 27-page technical report on February 10, 2026
• 50 percent accuracy on novel protein-ligand structures (less than 20 percent similarity to training data) where AlphaFold 3 managed 23.3 percent — major leap on the hardest cases
• Predicts not just structures but binding affinity, off-target effects, and cryptic binding sites — actual drug-design deliverables
• Proprietary closed model: no code, no weights, no API, no peer-reviewed publication; access exclusively through partnerships with Eli Lilly ($1.7B milestones), Novartis ($1.2B), and Johnson & Johnson
• First AI-designed drug candidates targeted for clinical trials by end of 2026; for non-partnership access to AI drug discovery, Insilico Medicine, BenevolentAI, and Schrödinger offer commercial alternatives