Variant and sequence effect probes
Evo 2 style workflows can explore sequence context and variant effect signals. Treat output as model evidence that needs comparison with annotation databases, phenotype context, and independent validation.
Research capability map
What NVIDIA BioNeMo skills can add to genome literacy work.
The skills kit is useful after a genetic question has been normalized into a sequence, variant, protein, ligand, or workflow object. It can help generate molecular prediction evidence, but it does not replace identity checks, clinical annotation, confirmatory testing, or expert interpretation.
Use it after normalization, not before.
For DNA work, the first responsibility is still to stabilize the variant: genome build, coordinate, strand, reference allele, alternate allele, transcript or gene context, and the exact prediction target. NVIDIA tools become useful once the object is precise enough to model.
Fold and complex prediction
Boltz-2, AlphaFold-style, OpenFold, and related workflows can predict structures, complexes, and confidence metrics for normalized protein questions. AlphaFold itself is an important reference implementation, but local runs can be too heavy for ordinary hardware.
Docking and affinity hypotheses
DiffDock and Boltz-2 ligand workflows can test ligand poses or protein-ligand hypotheses. These results are screening evidence, not pharmacology or treatment guidance.
Binder and sequence design
Binder-design and protein-generation skills can support research ideation around candidate binders or sequence variants, with strong separation from wet-lab claims.
Molecule generation and ADMET screens
GenMol, MolMIM, ADMET, and cheminformatics utilities can propose or score molecular candidates for research triage, property checks, and prioritization.
Alignment and pipeline acceleration
Parabricks-oriented skills can help with GPU-accelerated genomics pipeline tasks where the input data, reference, and privacy boundary are already controlled.
Research workflow
A safe order for using model skills.
NormalizeResolve build, strand, coordinate, reference, and alternate allele before modeling.
AnnotateCheck identity, clinical assertions, population frequency, gene context, and actionability boundaries.
ModelChoose the BioNeMo skill that matches the object: DNA, protein, ligand, molecule, or pipeline.
ComparePlace model output beside AlphaGenome, curated databases, literature, and explicit uncertainty.
ProtectKeep personal sequence data out of public pages, demos, logs, and casual model probes.
Record the exact object
Log genome build, interval, coordinates, allele orientation, transcript, protein sequence, ligand string, or file hash. Ambiguous input makes confident output meaningless.
Record the service and version
Log NVIDIA hosted NIM versus local NIM, model name, endpoint, parameters, date, and whether an API key or container workflow was used.
Record confidence and limits
Save raw response files, confidence metrics, model warnings, failed assumptions, and a short interpretation boundary for every run.
What this page is not
This is not a promise that any personal genome can be uploaded here, and it is not a diagnostic pathway. Later experiments on private DNA should run in a controlled workspace with explicit consent, local artifact logging, no public exposure, and a decision about whether NVIDIA, Google AlphaGenome, or both are appropriate for the specific prediction target.