developer
OpenFold3 Preview entered the model lineage
Extended the OpenFold lineage from AF2 reproduction toward a trainable AF3-class biomolecular model.
Open update record →Releases, milestones and builders.
Follow the biological AI model landscape over time and inspect permanent, source-linked update records.
Canonical release records
View organizations →developer
Stack entered the model lineage
Extended the State lineage toward broader in-context generalization rather than treating State and Stack as one static model.
Open update record →developer
Boltz-2 entered the model lineage
Added a learned affinity capability to the open Boltz structure lineage.
Open update record →developer
BoltzGen entered the model lineage
Extended Boltz from structure/affinity prediction into generative design.
Open update record →developer
Chai-2 entered the model lineage
Expanded the Chai lineage from multimodal complex prediction into generative therapeutic design.
Open update record →developer
Evo 2 entered the model lineage
Scaled the Evo lineage toward genome-scale reasoning and design.
Open update record →peer-reviewed
RFantibody entered the model lineage
Extended RFdiffusion-style generative structural modelling into antibody design.
Open update record →developer
RFdiffusion2 entered the model lineage
Moved the RFdiffusion lineage from backbone-first generation toward atom-level functional design.
Open update record →developer
State entered the model lineage
Established Arc's virtual-cell perturbation prediction lineage.
Open update record →peer-reviewed
AlphaFold 3 entered the model lineage
Replaced the AF2 structure module with a diffusion-based architecture and broadened supported complex types.
Open update record →preprint
Boltz-1 entered the model lineage
Released code and weights under commercially usable terms.
Open update record →preprint
Chai-1 entered the model lineage
Established Chai Discovery's first openly accessible complex-prediction model.
Open update record →peer-reviewed
ESM3 entered the model lineage
Unified sequence, structure and functional annotations in one generative model family.
Open update record →peer-reviewed
Evo entered the model lineage
Introduced a large genomic foundation model trained at nucleotide resolution.
Open update record →peer-reviewed
RoseTTAFold All-Atom entered the model lineage
Extended the RoseTTAFold family beyond protein-only representations.
Open update record →code
LigandMPNN entered the model lineage
Extended the ProteinMPNN concept to functional sites and molecular partners.
Open update record →peer-reviewed
RFdiffusion entered the model lineage
Converted a structure-prediction lineage into a generative protein-design system.
Open update record →peer-reviewed
OpenFold (AF2 reproduction) entered the model lineage
Made AlphaFold-style training and research reproducible in an open PyTorch implementation.
Open update record →peer-reviewed
ProteinMPNN entered the model lineage
Made sequence design a practical downstream step after backbone generation.
Open update record →preprint
AlphaFold-Multimer entered the model lineage
Extended the AlphaFold 2 family from monomers to protein assemblies.
Open update record →peer-reviewed
RoseTTAFold entered the model lineage
Introduced joint reasoning across one-, two- and three-dimensional representations.
Open update record →peer-reviewed
AlphaFold 2 entered the model lineage
Established the modern deep-learning structure-prediction baseline.
Open update record →Broader model timeline
Virtual Cells & Single-Cell
AIDO Cell 1.0
Developer-reported Virtual Cell Benchmark 1.0 spans 31 metrics across five task families; independent reproduction remains pending.
Source ↗Spatial & Tissue Systems
AIDO.Tissue
Public AIDO.Tissue checkpoints and downstream examples are available.
Source ↗Tissue, Pathology & Imaging
GenBio-PathFM
Reported evaluations include THUNDER, HEST and PathoROB; cross-paper comparability remains task-specific.
Source ↗Antibodies & Biologics
RFantibody
Experimentally evaluated de novo antibody designs have been reported.
Source ↗Virtual Cells & Single-Cell
State / Stack
Trained on 270M+ cells — largest single-cell perturbation model at release.
Source ↗Virtual Cells & Single-Cell
TranscriptFormer
Separated from the CZI platform record in BioAtlas ontology v2.
Source ↗Structure Prediction
Boltz-1 / Boltz-2
BoltzGen extends the line toward universal binder design.
Source ↗Virtual Cells & Single-Cell
Cell2Sentence (C2S-Scale)
Lets general LLMs 'speak' single-cell biology.
Source ↗Virtual Cells & Single-Cell
CZI Virtual Cells (rBio, TranscriptFormer)
Acquired the EvolutionaryScale (ESM3) team in 2025.
Source ↗Leading organizations
Google DeepMind
Structure Prediction · Protein & Binder Design · Genomics, DNA & RNA
5 atlas recordsInstitute for Protein Design, UW
Structure Prediction · Protein & Binder Design · Antibodies & Biologics
4 atlas recordsGenBio AI
Virtual Cells & Single-Cell · Genomics, DNA & RNA · RNA Models & Design · Protein Foundation & Representation · Structure Prediction · Spatial & Tissue Systems · Tissue, Pathology & Imaging
3 atlas recordsNVIDIA
Platforms, Data & Infra · Small-Molecule & Chemistry · Structure Prediction · Protein Foundation & Representation · Genomics, DNA & RNA
3 atlas recordsChan Zuckerberg Initiative
Virtual Cells & Single-Cell · Protein Foundation & Representation
2 atlas recordsGenerate Biomedicines
Protein & Binder Design · AI-Native Discovery Cos.
2 atlas recordsMeta AI (FAIR)
Structure Prediction · Protein Foundation & Representation
2 atlas recordsMIT (Barzilay & Jaakkola labs)
Structure Prediction · Small-Molecule & Chemistry
2 atlas recordsRecursion Pharmaceuticals
Platforms, Data & Infra · AI-Native Discovery Cos.
2 atlas recordsAbsci
Antibodies & Biologics
1 atlas recordAcademic RNA foundation-model collaborators
RNA Models & Design
1 atlas recordArc Institute
Virtual Cells & Single-Cell
1 atlas record