AI Biology Timeline: When Models Reached the Wet Lab

AlphaFold made protein structure feel solved too early. The real AI biology timeline starts there, then gets more interesting when models touch small molecules, wet labs, clinical endpoints, and restricted frontier-model access.

The AI biology timeline is the shift from structure prediction (AlphaFold 2), to interaction prediction (AlphaFold 3), to generative design and lab-in-the-loop execution. The recommended action in 2026 is to evaluate AI drug discovery systems by closed-loop evidence: wet-lab throughput, clinical endpoints, licensing constraints, and reproducible benchmarks.

TL;DR

AI biology breakthroughs moved through five phases: predict structures, generate designs, automate experiments, validate in patients, and govern frontier models.

In AI drug discovery 2026, the stronger teams connect models to experiments and regulatory-grade evidence. Better benchmark scores still matter, but they sit below clinical signal, assay throughput, model access rights, and the ability to trace why a candidate moved forward.

A practical operating rule follows from that: use frontier models for hypothesis generation, use open models for reproducibility and control, and treat every closed model as a governed dependency.

Key Takeaways

• AlphaFold 2 turned structure prediction into infrastructure after the Nature paper on July 15, 2021, while AlphaFold 3 expanded the problem to biomolecular interactions in Nature on May 8, 2024.

• AlphaFold 3 drug discovery has a licensing split: code is on GitHub, while trained model weights remain under custom non-commercial terms via Google’s request form.

• Insilico’s rentosertib and Recursion’s REC-4881 are the public clinical evidence points to watch, because they test whether AI-native discovery survives contact with patients.

• Open biological foundation models accelerated after 2024, with Chai-1, Boltz, Protenix, RFdiffusion3, and ESMFold2/ESMC giving teams alternatives to closed frontier systems.

• Frontier model gatekeeping is now part of technical architecture. Access terms, prohibited uses, and audit requirements belong in the model-selection table beside accuracy and latency.

What Is The AI Biology Timeline Practitioners Should Use?

The useful timeline is a capability timeline, rather than a press-release chronology.

The first phase was structure prediction. AlphaFold 2 and RoseTTAFold made 3D protein structure a routine input to downstream work.

The second phase was generative design. ProteinMPNN, RFdiffusion, Chroma, and ESM-3 shifted the field from predicting known biology toward proposing new proteins and functions.

The third phase was interaction modeling. AlphaFold 3, RoseTTAFold All-Atom, Chai-1, Boltz, and Protenix moved the target from protein monomers to complexes involving ligands, DNA, RNA, ions, and other molecules.

The fourth phase was closed-loop execution. Recursion, BioNeMo, Insilico, and autonomous-lab partnerships turned “AI biology” into a systems problem involving data factories, compute, wet labs, and clinical operations.

The fifth phase is governance. By June 2026, model restrictions and biological-design safety boundaries had become first-class product constraints.

Call this the Loop Ladder: structure, design, experiment, clinic, governance. A team climbs only as high as the weakest rung.

2021: Why Did AlphaFold Change The Starting Line?

AlphaFold 2’s canonical moment was the 2021 Nature paper, published after its CASP14 performance showed atomic-accuracy structure prediction competitive with experimental structures for many proteins.

The immediate product move mattered as much as the paper. The AlphaFold Protein Structure Database, jointly run by DeepMind and EMBL-EBI, launched in July 2021 and later expanded to more than 200 million predicted structures; the public database is accessible at alphafold.com.

That changed the default workflow for biologists and drug discovery teams. Before AlphaFold, missing structure was often a program blocker or a costly experimental detour. After AlphaFold, predicted structure became a starting artifact.

The open ecosystem responded fast. RoseTTAFold’s RosettaCommons repository and the Institute for Protein Design’s July 2021 announcement brought a 3-track network into public use, with the Science publication following in August 2021.

OpenFold then attacked reproducibility directly. The OpenFold repository framed the project as an open retraining of AlphaFold 2, which mattered for teams that needed inspectable training, inference, and multimer behavior.

This first wave made one idea obvious: structure prediction had become a substrate. The next question was what to build on top of it.

2022-2023: How Did Prediction Become Design?

ProteinMPNN made inverse folding practical. The Baker lab’s sequence-design model, cited by the Institute for Protein Design in its LigandMPNN follow-up, became a standard way to propose sequences for a desired backbone.

ESM-2 and ESMFold made a parallel bet: protein language models could learn enough from sequence scale to predict structure quickly. Meta’s ESM work appeared first as a bioRxiv preprint in July 2022, with the Science paper following in March 2023.

RFdiffusion then made de novo backbone generation feel real to the broader field. The RosettaCommons RFdiffusion repository and IPD’s July 2023 Nature coverage marked the transition from “predict this protein” to “design a protein that does this job.”

Chroma, announced by Generate Biomedicines in December 2022 through BusinessWire, pushed the same frontier from a company-building angle. Its product framing around programmable protein generation showed where investor and platform narratives were heading.

This period also exposed a permanent split in AI biology. Academic and open-source labs moved fast in public. Commercial drug hunters treated the same capabilities as proprietary workflow pieces inside discovery engines.

That split still defines AI drug discovery 2026.

2024: What Did AlphaFold 3 Actually Add To Drug Discovery?

AlphaFold 3’s important contribution was interaction modeling. The Nature paper by Abramson et al., published May 8, 2024, extended prediction to biomolecular complexes including protein-small-molecule, protein-DNA, protein-RNA, and other molecular interactions.

That sounds like a direct line to drug discovery, but the practical effect is narrower and more useful. AlphaFold 3 can improve the quality of structural hypotheses around binding and molecular context. It does not replace medicinal chemistry, ADMET work, synthesis planning, assays, or clinical judgment.

The release mechanics became a governance lesson. Google DeepMind released AlphaFold 3 inference code under Apache 2.0 in November 2024, while the trained weights request form says model parameters are available for non-commercial use under custom terms.

That licensing split matters for founders. A prototype can run on one access regime, while a regulated commercial product may need another.

Isomorphic Labs is the commercial expression of that boundary. The Alphabet subsidiary announced simultaneous Lilly and Novartis partnerships on January 7, 2024, according to Fierce Biotech: $45 million upfront plus up to $1.7 billion in Lilly milestones, and $37.5 million upfront plus up to $1.2 billion in Novartis milestones.

The chart shows why “nearly $3 billion” deal headlines need interpretation. The biobucks are useful market signal, while the realized upfront cash was $82.5 million across the two deals.

2024-2026: Why Did The Field Move Toward Closed Loops?

A structure model gives a hypothesis. A closed loop gives a learning system.

Recursion is the cleanest public example. On July 12, 2023, Recursion announced a $50 million NVIDIA equity investment and collaboration. By May 13, 2024, NVIDIA said Recursion’s BioHive-2 used 504 H100 GPUs, NVIDIA Quantum-2 InfiniBand, DGX SuperPOD architecture, and delivered 2 exaflops of AI performance in an NVIDIA write-up.

The same NVIDIA article gave the operational numbers that matter: more than 2 million wet-lab experiments per week, more than 50 petabytes of proprietary biological and chemical data, and 3 trillion searchable relationships.

That is the shape of serious AI biology infrastructure. The model sits inside an experiment factory, rather than serving as a detached predictor.

NVIDIA generalized that thesis through BioNeMo. BioNeMo began as a molecular AI framework at GTC 2022, then expanded at GTC 2024 with BioNeMo NIM microservices and foundation models for genomics, single-cell RNA, and docking, according to NVIDIA’s March 2024 BioNeMo announcement.

By January 2026, NVIDIA described BioNeMo as an open development platform for lab-in-the-loop workflows in a J.P. Morgan Healthcare Conference announcement. The release named Eli Lilly, Thermo Fisher, Chai Discovery, Basecamp Research, and Boltz among ecosystem participants, and introduced Clara open models RNAPro and ReaSyn v2.

BioNeMo’s release notes show the infrastructure cadence. As of mid-2026, the research set identified BioNeMo Framework v2.7 as the current stable generation, with subpackages spanning protein language modeling, evolutionary-scale modeling, molecular contrastive learning, and nucleic-acid design.

What Clinical Evidence Actually Landed?

Clinical evidence is where AI biology becomes harder to narrate and easier to evaluate.

Insilico’s rentosertib is the flagship AI-native small-molecule case in the research set. The company says TNIK was identified through PandaOmics and inhibitor candidates were generated with Chemistry42; Insilico describes that target-design path on its TNIK program page.

The Phase IIa trial in idiopathic pulmonary fibrosis enrolled 71 patients across 22 sites in China, with placebo and three rentosertib dosing arms over 12 weeks, according to GEN’s coverage. Insilico announced positive topline results on November 12, 2024, in its Phase IIa release, and then announced Nature Medicine publication and ATS 2025 presentation on June 3, 2025.

Recursion’s clinical picture is more mixed, which makes it more useful. Its REC-994 Phase 2 SYCAMORE trial for cerebral cavernous malformation met the primary safety endpoint in September 2024, but BioPharma Dive reported investor concern about efficacy.

Then REC-4881 delivered a stronger platform-validation story. On December 8, 2025, Recursion announced positive Phase 1b/2 TUPELO results in familial adenomatous polyposis: 75% of evaluable patients showed polyp burden reduction, with a 43% median reduction at 12 weeks. The company called it the first clinical validation of Recursion OS.

This table is the practical antidote to platform hype. If a system claims to discover drugs, ask which patient endpoint moved, in which trial, under which protocol.

What Changed In Open Models By June 2026?

Open biological foundation models made the AlphaFold 3 licensing debate less binary.

Chai-1 appeared in October 2024 with a bioRxiv-linked repository for multimolecular interaction prediction. Boltz-1 followed in November 2024 as an AlphaFold-3-class open model under MIT license in the Boltz repository. ByteDance’s Protenix arrived in January 2025 as an open AlphaFold 3 reproduction effort.

The Baker lab kept pushing design forward. RFdiffusion2 was released through RosettaCommons in September 2025, and IPD announced RFdiffusion3 on December 3, 2025 with atom-level diffusion and ten-fold faster performance over RFdiffusion2. RosettaCommons later described RFdiffusion3 availability in foundry and its atom-level modeling details in a December 22, 2025 note.

The ESM lineage also changed hands. EvolutionaryScale released ESM-3 on June 25, 2024, then ESM Cambrian on December 4, 2024. By June 2026, Chan Zuckerberg Biohub’s ESM repository described ESMC as a new scaling frontier relative to ESM-2, while Hugging Face hosted ESMFold2 and Northwestern reported the Biohub release of a protein-biology world model on June 1, 2026.

Open models now set a floor. Closed models still may lead on some frontier capabilities, but commercial teams can demand clearer evidence because public baselines exist.

Why Are Frontier Biology Models Being Gatekept?

Biology model restrictions are not a side issue. They shape who can build, publish, reproduce, and commercialize.

AlphaFold 3 is the best documented case in this research set. The code license allows open inspection of the inference pipeline, while the weights terms constrain commercial use and downstream model training. For a startup, that difference can decide whether the model belongs in production or only in research.

ESM-3 created another cautionary moment. EvolutionaryScale’s June 2024 release was framed as a frontier multimodal generative protein model, but the research set notes license controversy around restricted weights before later open releases such as ESM Cambrian.

The same standard should be applied to search terms such as Fable 5 biology restrictions or Mythos 5 drug design. If primary model cards, allowed-use policies, safety evaluations, and licensing terms are missing, treat the model as an unqualified dependency.

Frontier-model governance in biology has three practical layers.

First, access governance: who can use the model, for what purpose, and under which license.

Second, design governance: which biological outputs are filtered, blocked, logged, or routed to human review.

Third, evidence governance: which predictions can be reproduced by another team using documented inputs, model versions, and evaluation data.

A useful model-selection record should look more like an audit artifact than a leaderboard screenshot.

biology_model_review:
  model_name: "candidate model"
  as_of_date: "2026-06-24"
  task: "structure prediction | ligand design | protein design | assay triage"
  source_type: "open weights | API | custom license | internal model"
  allowed_use:
    commercial: true
    derivative_training: false
    publication: "check terms"
  evidence:
    benchmark: "name and version"
    wet_lab_validation: "assay, date, sample size"
    clinical_signal: "trial, endpoint, status"
  governance:
    prohibited_outputs: "documented policy link"
    audit_logs: true
    human_review_required: true

This is dull operational work. It is also where expensive mistakes get prevented.

Where Does Isomorphic Fit After AlphaFold 3?

Isomorphic Labs sits at the intersection of frontier model access, pharma partnerships, and unanswered clinical translation.

The company publicly launched as an Alphabet subsidiary in November 2021. It then turned AlphaFold’s scientific prestige into commercial deal flow, starting with the January 2024 Lilly and Novartis collaborations and extending to a cross-modality, multi-target Johnson & Johnson collaboration announced January 20, 2026.

Funding followed the same curve. Forbes reported Isomorphic’s $2.1 billion Series B in May 2026, led by Thrive Capital with Alphabet participating.

The missing public artifact is a human trial. As of June 24, 2026, the research set found no public first-in-human or Phase I entry for an Isomorphic-discovered candidate. Public company guidance pointed to first clinical trials by the end of 2026.

That makes Isomorphic a high-conviction platform story with clinical evidence still pending. The correct benchmark for 2026 is whether its AF3-derived drug design engine produces development candidates that enter and survive clinical testing.

What This Means For You

For senior AI engineers and founders, the AI biology timeline points to a build strategy.

Use AlphaFold-style models to reduce uncertainty around structure and interaction hypotheses. Use open models when reproducibility, cost control, or derivative training rights matter. Use closed frontier systems when their accuracy delta is material and the license allows your actual use case.

Design your platform around feedback latency. A weaker model connected to fast assays can improve faster than a stronger model trapped in a slow validation cycle.

Treat clinical claims as a separate evidence class. A platform publication, a docking benchmark, an animal model, a Phase I safety result, and a Phase II efficacy endpoint answer different questions.

Most of all, stop evaluating AI drug discovery as a single-model contest. The 2026 stack is a loop.

Action Checklist For AI Drug Discovery 2026

• Map every model to its dated version, license, allowed use, and weight-access terms.

• Separate prediction tasks from design tasks, then assign the right evaluation to each.

• Require at least one wet-lab validation path before promoting a model output into a program decision.

• Track clinical evidence by endpoint, trial phase, sample size, and sponsor claim.

• Compare closed frontier systems against open baselines such as Boltz, Protenix, RFdiffusion, and ESM-family releases.

• Build an audit log for biological design prompts, model outputs, filters, human approvals, and downstream experiments.

• Avoid treating partnership value as proof of platform validity. Upfront cash, milestones, and clinical endpoints belong in different columns.

• Watch the next Isomorphic and Recursion clinical milestones. Those readouts will say more about the field than another static benchmark.

Sources

• Highly accurate protein structure prediction with AlphaFold
• Accurate structure prediction of biomolecular interactions with AlphaFold 3
• AlphaFold Protein Structure Database
• google-deepmind/alphafold3
• AlphaFold 3 model parameters request form
• Isomorphic Labs, Lilly, and Novartis deal coverage
• Isomorphic Labs Johnson & Johnson collaboration
• Isomorphic Labs $2.1 billion fundraise
• Insilico rentosertib Phase IIa topline results
• Insilico Nature Medicine publication announcement
• Recursion and NVIDIA collaboration
• Recursion BioHive-2 NVIDIA supercomputer
• Recursion REC-4881 TUPELO Phase 1b/2 results
• NVIDIA BioNeMo drug discovery platform expansion
• BioNeMo Framework release notes
• RFdiffusion3 announcement
• Boltz repository
• Biohub ESM repository