Google Just Built the Best Drug Design AI in History. You Can't Use It.

The Best AI You'll Never Touch

On February 10, 2026, Isomorphic Labs released a 27 page technical report describing IsoDDE, the Isomorphic Labs Drug Design Engine. Mohammed AlQuraishi, a computational biologist at Columbia University, called it "a major advance, on the scale of an AlphaFold 4." The benchmarks back that up. On the most difficult protein ligand structure prediction cases, those with less than 20% similarity to training data, IsoDDE achieved 50% accuracy where AlphaFold 3 managed 23.3%. More than double the performance of the system that won the 2024 Nobel Prize in Chemistry.

IsoDDE runs protein structure prediction, ligand binding modeling, affinity estimation, antibody interaction prediction, and binding pocket discovery as a single integrated engine rather than separate tasks. It identifies novel binding pockets from amino acid sequence alone, in seconds, without specifying ligand identities. In one demonstration, it correctly predicted both the known thalidomide binding pocket and a novel cryptic allosteric site on cereblon, a protein that had been studied for 15 years before experimentalists discovered that second pocket in 2026. The system approaches the accuracy of experimental fragment soaking techniques while running orders of magnitude faster.

Unlike AlphaFold 1, 2, and 3, IsoDDE is proprietary. No code, no weights, no API, no peer reviewed publication. Nature reported that the technical paper offers "scant insight into how to achieve similar results." Scientists working on open source alternatives are, as AlQuraishi put it, "left guessing." The system is exclusively available through partnerships with Eli Lilly (up to $1.7 billion in milestones), Novartis (up to $1.2 billion), and Johnson & Johnson. Combined potential deal value across Lilly and Novartis alone: nearly $3 billion in biobucks, excluding royalties.

Meanwhile, on the same day this briefing publishes, Insilico Medicine and Liquid AI announced a partnership producing LFM2 2.6B, a 2.6 billion parameter model that outperforms Google's own TxGemma 27B, a model ten times its size, on 13 of 22 pharmacokinetics and toxicology tasks. It runs entirely on private infrastructure. In January, Eli Lilly and NVIDIA announced a $1 billion coinnovation lab in South San Francisco. AstraZeneca acquired Modella AI. GSK licensed Noetik's cancer foundation models for $50 million and committed $1.2 billion to an AI powered biologics facility in Pennsylvania.

The divide between open and proprietary AI in drug discovery is widening, and it will define who has access to the best ai in research tools for the rest of this decade.

What IsoDDE Actually Does

IsoDDE is architecturally different from its predecessors. AlphaFold 3, Chai 1, Boltz 1/2, ESMFold, RoseTTAFold, and OpenFold each handle specific prediction tasks. IsoDDE runs all of them together: structure prediction, docking, affinity estimation, and pocket discovery feed into each other. According to the technical report, this joint architecture enables generalization to unseen chemistry, the exact capability where previous molecular modeling software and molecular docking software have consistently failed and the benchmark that matters most for real world in silico drug design.

On antibody antigen docking, IsoDDE achieves a 2.3x improvement over AlphaFold 3 on high fidelity predictions and a 19.8x improvement over Boltz 2 on low homology test sets across 334 structures. Its binding affinity predictions surpass physics based free energy perturbation (FEP) methods on the FEP+, OpenFE, and CASP16 benchmarks, without requiring experimental crystal structures and at a fraction of the computational cost. Traditional FEP calculations take hours to days on cluster hardware and cost $100 or more per molecule; IsoDDE runs the same predictions in seconds.

The Cereblon Case

For 15 years, researchers believed cereblon had one principal druggable pocket. When Dippon et al. experimentally discovered a novel allosteric, cryptic binding site in 2026, IsoDDE independently predicted its location using only the protein's amino acid sequence, with no specified ligands and no prior structural data on the novel site. No publicly available molecular docking software or ai protein folding system can do blind pocket identification from sequence.

The system is entirely focused on the molecular design phase of in silico drug design. It says nothing about dissolution, excipient compatibility, stability, bioavailability, or manufacturing. Every candidate it produces still needs to be formulated into a drug product, and that process sits outside its scope entirely.

The Open vs. Proprietary War

The history of ai protein folding has been a history of openness. AlphaFold 2 was published in Nature and its code was released. The AlphaFold Protein Structure Database expanded to 214 million predicted structures, covering nearly all cataloged proteins known to science. When AlphaFold 3 withheld its code for six months in 2024, over 1,000 scientists signed an open letter protesting, and the code was eventually released under a noncommercial license. The entire field of deep learning drug discovery grew on the assumption that foundational models would be shared.

IsoDDE breaks that assumption. Isomorphic Labs, valued at $2.5 billion after a $600 million Series A in March 2025, is a commercial entity whose revenue depends on milestone payments from pharma partnerships. Forty percent of its Series A went to compute infrastructure expansion. The company has 17 active drug development programs across oncology, immunology, and cardiovascular disease, with the first AI designed cancer drug set to enter Phase 1 clinical trials by end of 2026. The incentive is simple: monetize performance, protect methodology.

The proprietary camp goes well past Isomorphic. Eli Lilly and NVIDIA are building a $1 billion coinnovation lab in South San Francisco, using NVIDIA's BioNeMo platform, Vera Rubin architecture, and Clara foundation models. The lab will run 24/7 AI assisted experimentation where AI agents direct robotic lab systems to perform physical experiments around the clock, with results feeding back into AI models in real time. AstraZeneca's acquisition of Modella AI in January 2026, its first AI company acquisition, puts generative and agentic AI into oncology R&D. GSK licensed Noetik's foundation models for non small cell lung cancer and colorectal cancer for $50 million, described as "among the first and largest transactions monetizing a biological foundation model as a scalable enterprise asset." Xaira Therapeutics launched in 2024 with $1 billion in venture capital and the backing of Nobel laureate David Baker.

The open source ecosystem remains productive. Boltz 2, developed by MIT and Recursion Pharmaceuticals, is the first deep learning drug discovery model to approach the accuracy of physics based FEP methods while running 1,000 times faster. It performs protein ligand affinity predictions in approximately 18 seconds on a single consumer GPU, under an MIT license. Chai Discovery, backed by $225 million including participation from OpenAI, achieved a 77% success rate on the PoseBusters benchmark, slightly exceeding AlphaFold 3's 76%. ESMFold from Meta AI runs an order of magnitude faster than AlphaFold 2 without requiring multiple sequence alignments. RoseTTAFold, from David Baker's lab, computes structures in as little as 10 minutes on a single gaming computer.

The distance between the two camps is growing. Gabriele Corso, formerly of MIT and now at Boltz, framed IsoDDE as "a new baseline to match, but also to pass." Diego del Alamo at Takeda Pharmaceuticals pointed to the real issue: "We don't know how impactful that extra data is," referring to the proprietary training data that open source teams cannot access.

The regulatory dimension makes this worse. The FDA/EMA joint AI principles, published January 14, 2026, require data governance, model explainability, and transparent communication about AI capabilities and limitations. EMA's Annex 22, expected to be enforced on the same timeline as the EU AI Act in August 2026, explicitly requires that AI models used in GMP applications be explainable and that training data meet GxP standards for accuracy, integrity, and traceability. A proprietary model that cannot be independently inspected or reproduced creates a compliance problem that benchmark performance cannot solve.

As we analyzed in our FDA/EMA AI Principles briefing , the regulatory trajectory points in one direction: transparency and data lineage are baseline requirements for ai in medicine.

The Lightweight Model Revolution

On March 3, 2026, Insilico Medicine and Liquid AI announced a partnership that takes a different path from both the proprietary megalab model and the open source academic approach. Their product, LFM2 2.6B MMAI, is a 2.6 billion parameter model trained on approximately 120 billion tokens across more than 200 pharmaceutical tasks, built on Liquid AI's foundation model architecture inspired by dynamical systems and signal processing.

The model punches well above its weight class. On 13 of 22 pharmacokinetics and toxicology tasks, LFM2 2.6B outperformed TxGemma 27B, a model with ten times the parameters. It achieved best in class results on 3 tasks compared to specialist models. On multi parameter molecular optimization, it reached up to 98.8% success rate on MuMO Instruct benchmarks. On affinity prediction, it produced better correlation scores than GPT 5.1, Claude Opus 4.5, and Grok 4.1 on Insilico's internal benchmark covering 2.5 million measurements across 689 protein targets.

The architecture matters. Liquid AI's models, originating from MIT research inspired by roundworm neural circuits, can have as few as 20,000 parameters and fewer than 20 neurons. The LFM2 2.6B runs entirely on private infrastructure without cloud services. For pharmaceutical companies operating under data sovereignty requirements, on premise deployment is a regulatory necessity. Every prediction carries a full audit trail. Every training datum is traceable.

Insilico Medicine brings clinical validation to the partnership. Its rentosertib (ISM001 055) Phase IIa results, published in Nature Medicine in June 2025, showed a mean FVC improvement of +98.4 mL in the 60 mg treatment group versus a 20.3 mL decline in placebo across 71 patients. That is the first clinical validation of ai powered drug discovery published in Nature Medicine. The company raised $293 million in its Hong Kong IPO in December 2025, oversubscribed 1,427 times in the public offering. Its MMAI Gym for Science, launched in January 2026, transforms general purpose language models into pharmaceutical grade engines, demonstrating up to 10x performance gains on drug discovery benchmarks and pushing one model (Qwen3 14B) from failing roughly 70% of medicinal chemistry benchmarks to solving 95% or more within two weeks.

Efficient architecture design, not just scale, makes foundation models practical for sciences.

Ramin Hasani, CEO, Liquid AI

The argument is simple: domain specific, lightweight, privately deployable bioinformatics tools can outperform billion parameter general purpose models on pharmaceutical tasks, at a fraction of the cost and with full regulatory traceability. For the hundreds of pharmaceutical companies that will never have access to IsoDDE's proprietary predictions or a billion dollar AI facility, the lightweight model path is the realistic one.

The cost numbers tell the story. Traditional physics based FEP calculations cost $100 or more per molecule and require cluster hardware. Boltz 2 runs predictions in 18 seconds on a consumer GPU. The LFM2 2.6B runs on private servers. The price of entry into ai powered drug discovery and molecular modeling software keeps falling, even as the ceiling of what proprietary systems can do keeps rising. For organizations pursuing ai in research without billion dollar budgets, these bioinformatics tools are the practical option.

What This Means for Formulation Science

All three approaches, IsoDDE's proprietary engine, lightweight open models, and billion dollar corporate AI labs, focus exclusively on drug discovery. None of them touch formulation. As discovery AI improves, the formulation gap widens.

IsoDDE doubles accuracy on protein ligand binding prediction and identifies novel pockets in seconds. It will produce better molecular candidates faster. Every one of those candidates still needs to be formulated into a drug product that dissolves, remains stable, scales in manufacturing, and delivers the active ingredient to the right place in the patient's body. The clinical failure rate sits at approximately 90% and has not moved despite a decade of AI investment in discovery. AI discovered compounds show 80 to 90% Phase I success, approximately 40% Phase II success; those numbers are indistinguishable from the industry baseline.

The arithmetic is simple. If AI compresses early discovery from 4.5 years to 12 to 18 months but formulation development, stability testing, and manufacturing scale up still consume the same years they always have, faster discovery increases the relative burden of formulation. More candidates arriving faster at the formulation stage means more pressure on a process that is still largely manual and empirical. With 3.6 million possible formulation combinations for a single compound, the design space is too large for trial and error.

The solubility numbers have not changed. Only 8% of novel drug candidates exhibit both excellent permeability and solubility. Between 40% and 70% of all new hits have solubility too low for complete gastrointestinal absorption. Approximately 40% of new chemical entities are insoluble in water. Over a third of US Pharmacopeia listed pharmaceuticals are poorly soluble. The majority fall into BCS Class II: low solubility, high permeability. AI designed molecules frequently have high melting points and poor solvent solubility, characteristics that compound the formulation challenge.

No discovery AI addresses dissolution behavior, excipient compatibility, physical and chemical stability, photostability, bioavailability enhancement strategies (solid dispersions, nanosuspensions, lipid systems), dosage form design, or manufacturing scale up. The underlying architecture does not matter; whether it is deep learning drug discovery or traditional molecular docking software, the blind spot is the same. As we observed in our digital twin disconnect analysis , 79% of pharma companies virtualize their equipment, but fewer than 17% virtualize their formulations.

Turning those designs into bioavailable, stable, scalable, patient-ready products still depends on formulation science and process know-how, not algorithms.

• IsoDDE predicts where a molecule binds. It cannot predict whether it will dissolve.
• LFM2 2.6B optimizes molecular properties. It cannot predict excipient compatibility.
• The Lilly NVIDIA lab creates digital twins of manufacturing lines. It does not address the formulation decisions that determine what those lines produce.
• Over 173 AI designed programs are in clinical development. The overall clinical failure rate remains approximately 90%.

Eroom's Law, the observation that drug development costs double roughly every nine years despite technological improvements, will not reverse by accelerating one phase while leaving the rest alone. The $2.23 billion average cost per new drug, per Deloitte's 2024 analysis, reflects a system where $7.7 billion was spent on terminated candidates in a single year and Phase III cycle times increased 12%. R&D return on investment was 5.9%, dropping to 3.8% excluding GLP 1 drugs. Average time from Phase I through regulatory filing now exceeds 100 months. The application of ai in medicine has to extend beyond molecule design, and ai in healthcare has to reach formulation, manufacturing, and quality for these economics to change.

DeepC Builds What They Don't

DeepC occupies the space that IsoDDE, LFM2 2.6B, and the Lilly NVIDIA lab do not touch: formulation intelligence. Over $1.6 billion in proprietary discovery AI deals were announced in January and February 2026 alone. The ai powered drug discovery market is projected to grow from $19.89 billion to $160.49 billion by 2035. Formulation development, where drugs actually fail, receives a fraction of that attention and investment.

DeepC is an AI coscientist for formulation scientists, built for the work that happens after a molecule is found. Its tools address the specific problems that discovery AI creates but cannot solve.

Purpose Built Research Tools

• Formulation Agent: Excipient selection, compatibility assessment, and formulation design using pharmaceutical machine learning models trained on real world formulation data. Reduces the traditional 3 to 6 month empirical screening process to a prioritized, evidence based shortlist.
• SMILES2SPEC: Predicts LC-MS/MS spectra from molecular structures, allowing formulation scientists to anticipate analytical challenges before running expensive experiments. Computational chemistry software built for the formulation workflow.
• MiCQ: ML powered Critical Quality Attribute imputation using KNN, MICE, MissForest, and CatBoost to address incomplete CQA datasets. Turns a persistent formulation development blocker into a solvable problem.
• Elute: Extracts structured data from over 50 document formats, including CMC documents, batch records, stability reports, analytical certificates, and regulatory filings. Pharmaceutical document intelligence at scale.

Grounded in Authoritative Data

DeepC's validation advantage comes from its data foundation. The platform is grounded in authoritative regulatory data: the FDA Inactive Ingredient Database (all FDA approved excipients with maximum potency levels), FAERS (individual case safety reports for excipient related signal detection), DailyMed (154,834+ drug labeling records with complete formulation data), ClinicalTrials.gov, and PubMed. Every recommendation carries data lineage from source to output. Every prediction is traceable to its regulatory foundation.

That traceability separates DeepC from every molecular modeling software platform and in silico drug design tool that stops at the molecule. The FDA/EMA joint principles (Principles 6, 7, and 10) require data governance, model explainability, and plain language communication about AI capabilities and limitations. The FDA's January 2025 draft guidance establishes a seven step credibility assessment framework for AI models used in regulatory submissions. EMA's Annex 22 requires that AI models in GMP applications be explainable and that training data meet GxP standards for accuracy, integrity, and traceability. DeepC was built to meet those specifications. Proprietary black box APIs, however accurate their predictions, face a compliance gap that will widen as these frameworks take binding effect.

Over 530 companies focus on AI drug discovery. Over $11 billion in venture capital flowed into AI ML drug discovery in 2025 alone, across 348 financing rounds. The formulation development side of the pipeline, where drugs actually fail, receives a fraction of that attention and investment. DeepC does not compete with Isomorphic Labs, Recursion, or Insilico Medicine for the molecule design market. It builds the platform their molecules will need when they leave the computational domain and enter the physical world of formulation, manufacturing, and regulatory submission.

Each small molecule discovery is like a work of art. If we can make that an engineering problem, versus this sort of discovery, this artisanal drug-making problem, think of the impact on human life.

David Ricks, CEO, Eli Lilly

What Ricks calls the "artisanal drug making problem" is the formulation challenge. Discovery is becoming engineering. Formulation has not made that transition. DeepC is built to close that gap.