Isomorphic Labs Raises $2.1B as AI Drug Discovery Matures

Isomorphic Labs closed a $2.1 billion funding round on May 13, 2026, led by Thrive Capital — the second-largest biotech fundraise on record, behind only Altos Labs, according to Forbes. The Alphabet-founded company, best known for the AlphaFold protein-structure model, is using the capital to push its AI-driven drug design platform toward clinical candidates. The round follows a $600 million outside funding in 2025 and signals sustained investor conviction in AI’s ability to cut the cost and time of bringing new medicines to market.

Isomorphic Labs and the AlphaFold Foundation

Isomorphic Labs was formed as an Alphabet company in 2021, built directly on the scientific foundation of AlphaFold. CEO Demis Hassabis — who also leads Google DeepMind — was awarded the 2024 Nobel Prize in Chemistry for AlphaFold’s protein-structure prediction work, according to Forbes.

The latest version, AlphaFold 3, released in May 2024, extended the model’s scope beyond proteins to include small molecules, peptides, and antibodies — the full toolkit of modern drug design. Isomorphic has since built on that base with its Isomorphic Labs Drug Design Engine (IsoDDE), which President Max Jaderberg described to Forbes as “like half a dozen AlphaFold breakthroughs” combined into a single platform.

Jaderberg told Forbes the new funding represents “a lot of validation of what we’ve been building out the past four-and-a-half, almost five, years.” The company has not disclosed which drug candidates it plans to advance to clinical trials, nor has it revealed the valuation at which the round was closed.

Tazbentetol Trial Points to Synaptic Repair as a Drug Target

While Isomorphic operates at the molecular modeling layer, a separate clinical signal is illustrating what AI-assisted target identification can unlock at the bedside. Spinogenix reported early Phase 2 data on tazbentetol, a first-in-class investigational synaptic regenerative therapy, at the Schizophrenia International Research Society 2026 Annual Congress.

According to Spinogenix’s press release, the drug produced a placebo-adjusted 6.3-point reduction in the Positive and Negative Syndrome Scale (PANSS) in an add-on trial design. Crucially, patients who discontinued after six weeks retained measurable benefit for multiple days — a durability profile unusual for a psychiatric add-on therapy.

Tazbentetol is believed to modulate fascin-1/F-actin dynamics, promoting the formation of dendritic spines with glutamatergic synapses — in effect, triggering neurons to build new connections. Spinogenix is also studying the compound in Alzheimer’s disease, ALS, glaucoma, and diabetic retinopathy, suggesting the synaptic regeneration mechanism may have broad applicability across neurological and degenerative conditions.

PolitNuggets Benchmark Tests Agentic Fact Discovery

On the research infrastructure side, a new arXiv preprint introduces PolitNuggets (arXiv:2605.14002), a multilingual benchmark designed to evaluate how well large reasoning models (LRMs) find and synthesize obscure, dispersed facts — what the authors call “long-tail” information.

The benchmark constructs political biographies for 400 global elites, covering more than 10,000 political facts across multiple languages. Rather than testing static retrieval, PolitNuggets places models inside agentic frameworks and measures their ability to explore open-ended information spaces — closer to how these systems are actually deployed in research and intelligence workflows.

The authors propose FactNet, an evidence-conditional evaluation protocol that scores three dimensions separately: discovery (did the model find the fact?), fine-grained accuracy (did it get the details right?), and efficiency (how many steps did it take?). Across tested models, the paper finds that current systems struggle most with fine-grained detail accuracy and show wide variance in efficiency. The benchmark also identifies short-context extraction quality, multilingual robustness, and reliable tool use as the three capabilities most predictive of strong agent performance.

Cerebras IPO Validates Decade-Old Hardware Thesis

The commercial momentum behind AI research is also visible in public markets. Cerebras Systems completed a successful IPO this week, generating billions for its investors, including Benchmark, which holds a 9.5% stake in the company, according to TechCrunch.

Benchmark general partner Eric Vishria co-led Cerebras’ $25 million Series A in 2016 — a hardware bet the firm almost declined. Vishria told TechCrunch he was skeptical going into the meeting: “It was five founders and a deck, and it was our first hardware investment in 10 years.” His position shifted by the third slide, when CEO Andrew Feldman argued that GPUs were architecturally mismatched to deep learning workloads and that a purpose-built AI chip was the correct solution.

That pitch predated Google’s 2017 Transformer paper by roughly a year, meaning Cerebras was betting on a hardware architecture shift before the software demand for it had fully materialized. The IPO outcome validates a thesis held for nearly a decade — and underscores that the infrastructure layer of AI, not just the application layer, is generating substantial returns.

What This Means

The week’s developments, read together, describe an AI research-to-application pipeline that is maturing across multiple fronts simultaneously.

Isomorphic’s $2.1 billion raise is the clearest financial signal yet that institutional capital believes AI-native drug design is past proof-of-concept and approaching clinical relevance. The company’s silence on specific pipeline candidates is notable — it suggests either competitive sensitivity or that the platform is still generating candidates rather than advancing them. Investors appear willing to fund the uncertainty.

The tazbentetol Phase 2 data, while from a single early-stage trial, illustrates a mechanism — synaptic regeneration — that AI target identification tools are well-positioned to explore at scale. If the durability signal holds in larger trials, it would represent a qualitatively different approach to schizophrenia treatment than current dopamine-modulating therapies.

PolitNuggets addresses a gap that matters beyond political research: most real-world agentic deployments require models to find facts that are not in any single document. Benchmarks that reward discovery and efficiency separately, rather than just answer accuracy, will produce better-calibrated models for enterprise research tasks.

And Cerebras’ IPO performance is a reminder that hardware investment cycles in AI are long — nearly a decade from Series A to public market — but the eventual returns can be substantial when the architectural thesis proves correct.

FAQ

What is Isomorphic Labs and how does it use AI for drug discovery?

Isomorphic Labs is an Alphabet-founded company that applies AI models to accelerate the drug development process. Its core platform builds on AlphaFold, which predicts the 3D structures of proteins and other biological molecules, and extends into drug design through its IsoDDE engine, which models how candidate compounds interact with biological targets.

What is the PANSS score and why does a 6.3-point reduction matter?

The Positive and Negative Syndrome Scale (PANSS) is a standard clinical instrument for measuring schizophrenia symptom severity, with higher scores indicating worse symptoms. A placebo-adjusted reduction of 6.3 points in an add-on trial design is considered clinically meaningful because add-on trials — where patients are already on background medication — typically show smaller effect sizes than monotherapy studies.

What does the PolitNuggets benchmark measure that existing benchmarks do not?

PolitNuggets evaluates agentic AI systems on their ability to discover and synthesize “long-tail” facts scattered across multiple sources, rather than answering questions from a single provided document. Its FactNet protocol scores discovery, fine-grained accuracy, and efficiency as separate dimensions, giving a more detailed picture of where agentic systems fail than standard accuracy-only benchmarks.