Google DeepMind released AlphaFold 4 on Wednesday — the first version of the structural-biology model with reliable protein-protein interaction predictions across the proteome. On the same day, sister company Isomorphic Labs disclosed that its first AI-discovered drug candidate has entered Phase 2 clinical trials. The dual announcements land together by design: DeepMind is presenting the research advance and Isomorphic is presenting the first translational outcome that depends on the underlying model lineage.

The structural-biology research community has been anticipating this release since AlphaFold 3 in 2024 demonstrated reliable single-protein structure prediction but only partial coverage of protein-protein interactions. Interactions — how one protein binds another — are the substrate for drug discovery, signal transduction, and most biological processes that interest pharmaceutical companies. AlphaFold 4 closes that gap.

What's actually new in AlphaFold 4

The technical advance, summarised from the paper that accompanied the release in Nature, is a substantially-improved capacity to predict protein-protein interaction structures across the full diversity of the human proteome. AlphaFold 3 worked well for protein-protein pairs where the structural similarity to known reference cases was high; AlphaFold 4 extends that reliability to novel pairs without close structural precedent in the training set.

The improvement comes from three architectural changes. First, the training set has been expanded to include large amounts of cryo-EM data that was unavailable when AlphaFold 3 was trained. Second, the model now produces an explicit confidence score for each predicted residue contact, allowing downstream users to filter to high-confidence predictions for drug-discovery decisions. Third, the iterative refinement step that powered AlphaFold 3's accuracy gains has been generalised to protein-complex prediction, producing measurably better structural geometry for the bound complexes.

Isomorphic Labs Phase 2 milestone

The translational news is the more commercially-significant of the two announcements. Isomorphic Labs is the DeepMind spinoff founded in 2021 to apply AI to drug discovery; it has been publicly building a pipeline since 2022 with partnerships at Novartis, Eli Lilly and others. The Phase 2 candidate — code-named IL-2026-01 — targets a protein-protein interaction implicated in a specific autoimmune disease that Isomorphic has not yet publicly disclosed.

What is publicly disclosed is the path from discovery to Phase 2: candidate identified using an internal Isomorphic-extended variant of AlphaFold 3 in mid-2023; medicinal chemistry optimisation through 2024; Phase 1 safety trials in late 2024 and 2025; Phase 2 efficacy trials starting May 2026. The timeline — roughly three years from AI-driven discovery to Phase 2 — is significantly faster than the industry median of seven to eight years for drug candidates of equivalent complexity.

What the Phase 2 milestone does and does not prove

The Phase 2 milestone is significant because it is the first concrete evidence in a regulator-monitored clinical context that AlphaFold-lineage drug discovery produces candidates that survive past the safety stage. Phase 1 trials primarily test for toxicity at varying doses in healthy volunteers; Phase 2 tests for efficacy in patient populations. The transition from Phase 1 to Phase 2 is a non-trivial filter; roughly half of drug candidates fail at the Phase 1 stage. IL-2026-01 surviving Phase 1 is positive evidence that AI-discovered candidates are not systematically toxic relative to traditionally-discovered candidates.

What Phase 2 entry does NOT prove is that the candidate works. Roughly two-thirds of candidates that enter Phase 2 fail to demonstrate efficacy. The drug industry's empirical rate of Phase 2 success is about 30%. The proper evidentiary milestone for AI-driven drug discovery is Phase 3 entry — or even better, approval — and Isomorphic is still many years from that bar with this candidate.

What the pharma industry takes from this

For pharma companies, the AlphaFold 4 + IL-2026-01 announcements together signal two things. First, the structural-biology tooling now reliably handles the protein-protein interactions that underpin a substantial fraction of disease biology, removing a previous bottleneck on AI-driven discovery. Second, the discovery-to-Phase 2 timeline of three years is plausibly achievable with AI-augmented workflows, compared to a historical baseline that ran twice as long.

The implication for pharmaceutical R&D budgets is significant. Pharma R&D historically allocates a substantial fraction of pre-clinical spend to the discovery and optimisation phases that AI-driven approaches compress. If the time-and-cost compression that Isomorphic is demonstrating proves generalisable, it represents an opportunity for the industry to either reduce R&D budgets (which Wall Street would reward) or to maintain budgets while substantially increasing the size of the pipeline (which the industry's own science argues is the better strategic choice).

What's free, what's not

AlphaFold 4 is being released under the same model as AlphaFold 3 — academic and non-commercial research use is free, with structures viewable through the AlphaFold Protein Structure Database. Commercial users — pharmaceutical companies, biotech startups, drug-discovery service providers — must license access. The pricing has not been publicly disclosed but is understood to be in the seven-figures-per-year range for major pharma customers.

The licence structure is the operational point. AlphaFold has become the de facto standard structural-biology tool in industry; the licensing economics meaningfully fund DeepMind's continued research while creating a moat against open-source replication of the full-quality model.

Academic versus industry use diverges

Academic structural biologists are likely to adopt AlphaFold 4 quickly. The free access through the structure database, the consistency with prior AlphaFold releases, and the improvement in protein-complex prediction all align with what academic labs already use AlphaFold for. The transition cost from AlphaFold 3 to AlphaFold 4 is essentially zero for academic users.

Industrial users face a more complex calculation. The licensed full-model access has real value for production drug-discovery pipelines, but the seven-figures-per-year licensing cost has driven several large pharma companies to invest in in-house alternative models — IBM Research's protein-folding work, the Schrödinger ChemFOLD release, RoseTTAFold from the University of Washington's spin-out, and several proprietary efforts that have not been publicly disclosed. The competitive landscape for industrial protein-structure prediction is more contested in 2026 than it was in 2024.

Isomorphic Labs's success with IL-2026-01 is the strongest single argument for the AlphaFold-licensed-access path. If the candidate progresses through Phase 2 successfully — and especially if Isomorphic's pipeline produces further Phase 2 candidates at similar pace through 2027 — the AlphaFold licensing fee becomes manifestly worth paying for any pharma company that wants to compete at AI-driven discovery speed. If IL-2026-01 fails, the calculus is more contested.

What AlphaFold 4 doesn't fix

Structural biology is one input to drug discovery, not the whole pipeline. AlphaFold 4 substantially improves the protein-structure prediction step, but the subsequent steps — finding small molecules that bind a target protein's site, optimising those molecules for ADMET properties (Absorption, Distribution, Metabolism, Excretion, Toxicity), running pre-clinical animal models, navigating regulatory approval — remain expensive, time-consuming and largely unrelieved by improvements in structure prediction alone.

The bottleneck after AlphaFold-class structure improvements becomes the wet-lab validation and clinical-trial cost. AI-driven structural prediction can identify more candidate molecules per unit time, but each candidate still requires a Phase 1, 2 and 3 trial sequence that takes years and tens of millions of dollars. AlphaFold 4 does not change that math. The cost-reduction argument for AI in drug discovery is real but specific to the candidate-generation phase; the larger industry economics are dominated by post-candidate-generation costs.

That nuance matters for pharma-industry strategic decisions. The AlphaFold-class structural improvements increase the number of plausibly-druggable targets a research organisation can pursue per unit budget. They do not significantly reduce the per-candidate cost of carrying a target through approval. The pharma R&D budget question becomes one of pipeline width versus pipeline depth — which is a strategic choice rather than a cost-optimisation one.

Computational requirements and accessibility

AlphaFold 4 is more computationally demanding than its predecessor, particularly for the protein-protein interaction predictions that are its headline capability. Running the model at full quality requires GPU clusters that are accessible to large academic centres and pharma companies but largely out of reach for small biotechs and individual researchers. DeepMind's approach to mitigating this gap is two-fold: the AlphaFold Protein Structure Database continues to provide free access to pre-computed structures for all known protein sequences, and a "lighter" inference profile in AlphaFold 4 trades some accuracy for substantially-reduced GPU requirements.

The lighter inference profile is the more interesting of the two from an accessibility standpoint. It allows research-grade prediction quality on a single high-end consumer GPU — the kind of hardware a well-equipped university lab can afford. The trade-off in accuracy is meaningful for protein-protein interaction predictions but modest for single-protein structure prediction. The dual-tier approach is becoming a pattern across AI-tools-for-science releases.

Implications for academic biology training

AlphaFold 4's release continues a multi-year shift in how computational biology is taught. The 2020s have seen structural-biology curricula increasingly incorporate AI tooling alongside traditional experimental methods. AlphaFold 4 will accelerate that shift: graduate students entering structural biology in 2026 will treat AI-driven structure prediction as a foundational tool the way the 2010s generation treated X-ray crystallography software.

The training implication runs deeper than just learning to use the tool. The conceptual interpretation of AI-predicted structures — when to trust them, when to validate experimentally, how to interpret confidence scores — is becoming part of the core competency for the next generation of structural biologists. Several major US universities have already introduced or planned curriculum changes that reflect this shift. The European Molecular Biology Laboratory has published a working paper proposing standardised curriculum modules for AI-driven structural biology that anticipates AlphaFold-class tools as the default starting point.

The broader DeepMind AI-for-science portfolio

AlphaFold is one of several DeepMind AI-for-science initiatives. AlphaGeometry tackles formal mathematical reasoning. AlphaProteo focuses on protein design rather than just structure prediction. GraphCast handles weather forecasting. AlphaMissense provides clinical-variant interpretation. The portfolio represents a coherent strategy: apply frontier-model capability to specific scientific domains where the rewards justify substantial investment.

The portfolio's collective progress over the past 36 months has been DeepMind's strongest argument for the value of fundamental AI research that is not directly product-monetisable. The Isomorphic Labs commercial path provides revenue capture from one branch of the portfolio; the AlphaFold Protein Structure Database provides public-good distribution from another; the AlphaGeometry and AlphaMissense releases provide research-community impact even without direct revenue. The portfolio strategy is becoming a model that other AI labs are studying — particularly Anthropic, which has signalled interest in similar science-oriented frontier deployment.

Sources

Google DeepMind's announcement and the Nature paper provide the technical detail. Isomorphic Labs' press release covers the Phase 2 milestone. STAT News carried the integrated lede on the pharmaceutical-industry implications.