Search perturbation-induced morphology with biological language
Perturb-LM is currently a text-to-morphology benchmark. It validates leakage-aware retrieval over Cell Painting profiles while keeping image-level retrieval as a longer-term direction.
Research question
Can frozen biomedical language representations retrieve perturbation-induced cellular morphology better than strong identifier-stripped lexical controls?
mAP 95% query-bootstrap CI: 0.2445 to 0.2582. This is a lexical control, not a learned-model result.
Why leakage matters
Treatment names, target sequences, plates, wells, and replicates can make retrieval look stronger than it is. Perturb-LM treats those shortcuts as measurable failure modes.
Naive evaluation
Can reward exact treatment identifiers, target sequences, replicate structure, and acquisition artifacts.
Leakage-aware evaluation
Removes identifiers, reports evaluable queries, and tests held-out and plate/well-aware conditions.
How the system works
Biological language query
Researcher describes a phenotype or mechanism.
Frozen text representation
The real biomedical encoder evaluation is pending.
Lightweight alignment
A small projection is the next controlled experiment.
Morphology-profile retrieval
Profile-space ranking before any image-level claim.
Perturbation-level evaluation
Metrics aggregate at the perturbation level.
Validated foundation
QC, leakage checks, deterministic queries, train-only preprocessing, and reproducible smoke workflows are in place.
Lexical benchmark
Identifier-stripped TF-IDF establishes the primary lexical control that future models must beat.
Pending learned model
Frozen text embeddings and a lightweight projection are the next experiment, not a completed result.
Current limitations
- The real biomedical text-alignment experiment has not started.
- Split-specific learned-model results remain pending.
- Held-out-batch evaluation is unavailable.
- The additional profile plate changes the feature schema and is only a compatibility investigation.
- The current benchmark is text-to-morphology-profile retrieval, not validated text-to-image retrieval.