Released model

Try AssayFormer

Describe a CRISPR screen, add the genes already tested, and rank the next batch with the AssayFormer model.

Describe the screen

Public example

Phenotype Cell line Cell type Perturbation Organism Condition Additional context (optional)

Add observed results

Values mean Hit threshold A hit is

Observed genes

No observations pasted; cold-start ranking is supported.

Next batch size Exclude DepMap common-essential genes

Predicted next genes

Rank	Gene	Acquisition score	Batch inclusion	Known result	Hits / public screens	Essential	Reactome pathway	CORUM complex
Submit a screen to generate a ranking.

No ranking yet. Choose a public example and click Rank the next batch.

Results are ordered by AssayFormer's acquisition score. The batch-inclusion percentage estimates how often a gene enters the next batch under the stochastic Gumbel-top-k policy; it is not a probability that the gene is a hit.

What is running? AssayFormer ranks the released 21,147-gene evaluation universe, conditioning on the assay description and the hit/non-hit feedback above. Observed genes are removed from the candidate set. Symbols outside the checkpoint vocabulary are reported and ignored rather than mapped to an unknown-gene embedding. The public examples run locally in your browser using the ONNX version of the paper checkpoint.

All five examples are public BioGRID ORCS screens distributed through AssayBench. Each preloads two completed AssayFormer batches as feedback. Loading example source…

To run your own screen.

Public examples use released text embeddings and run in your browser. The first ranking downloads the 20 MB model; configure an embedding endpoint to enter a different screen.

Submitted descriptions and observations are not retained by this application, although the hosting platform may have its own logging policy.