Biological object
Protein, DNA, RNA, molecule, cell, tissue, image and spatial neighborhood are different scientific objects. BioAtlas keeps the modality and scale explicit.
Choose models by biology, representation and evidence—not by hype.
BioAtlas separates the biological object from its computational representation: sequence tokens, structural alphabets, molecular graphs, all-atom coordinates, gene-expression vectors, spatial graphs, tissue images and other forms are tracked explicitly alongside architecture, objective, output and validation.
Model ontology
64 model families · 12 primary domainsOne category is not enough to describe biology.
Representation
Track how the object becomes computation: sequence tokens, graphs, pair features, 3D coordinates, expression vectors, embeddings, image patches or spatial graphs.
Architecture & objective
Separate Transformer, diffusion, GNN, state-space, physics and hybrid architectures from the objective used to learn or generate the representation.
Output & validation
Predictions, embeddings, structures and generated candidates are linked to task-specific evidence rather than being collapsed into a universal model score.
Scientific task hubs
5 curated topic hubsStart from the problem you need to solve.
Protein Structure Prediction AI Models
Compare AI models for protein and biomolecular structure prediction using source-linked capabilities, access, benchmarks, versions and known limitations.
Explore evidence →AI Molecular Docking and Pose Prediction Models
Explore AI docking and protein–ligand pose models with benchmark context, access terms, input/output contracts and evidence boundaries.
Explore evidence →Protein Design AI Models
Compare generative protein and binder-design models, including evidence, access, experimental validation, workflows and limitations.
Explore evidence →Genomic Foundation Models
Explore AI foundation models for DNA, RNA and genomic sequence understanding with evidence, versions, access and scientific limitations.
Explore evidence →Virtual Cell and Perturbation AI Models
Compare virtual-cell, single-cell and perturbation-prediction models by inputs, outputs, benchmark context and evidence limitations.
Explore evidence →Evidence rules
Model capability is always conditional.
Task-specific evaluation
Structure accuracy, docking pose accuracy, affinity ranking, representation transfer, generation quality and perturbation prediction are different metrics and should not be collapsed into one readiness score.
Version-aware records
A model family can change architecture, scope, representation, access terms and benchmark behavior between releases. Unknown version histories remain pending rather than receiving inferred dates.
Entity-aware records
A model, a checkpoint, a platform and a company are different entities. BioAtlas now labels them explicitly instead of presenting every item as the same kind of model.
Evidence boundaries
Developer-reported, preprint, peer-reviewed, independently reproduced and experimentally validated evidence can be tracked separately, with source links and caveats retained.