| name | crossmodal-discovery |
| description | Use for cross-modal target discovery from an unmet clinical need, combining literature, genetics, druggability, clinical trials, single-cell, and spatial evidence in one auditable context. |
Cross-modal discovery
Use this skill when the user asks for a new target, a translational opportunity, an unmet-need landscape, or a conclusion that must combine several BioCSSwitch packs.
Required workflow
- Call
crossmodal_plan_unmet_need with a concrete disease and unmet need. Preserve its context as the only shared evidence state.
- Execute plan steps in stage order. Resolve
$outputs... references from actual earlier tool results and fan out {target} steps only over targets present in the context.
- After every external pack call, pass the matching plan step and real result to
crossmodal_reduce_evidence. Never summarize a result into an unsupported claim before reduction.
- For measured single-cell or spatial results, add explicit claim-level records with
crossmodal_integrate_observations. A generated analysis recipe is provenance, not biological evidence.
- Call
crossmodal_synthesize. If it reports conflicts, call kg_generate_hypotheses before recommending a decisive next experiment.
- Present target rank, evidence coverage, corroborating sources, conflicts, missing modalities, and the next discriminating data collection step together.
Evidence contract
supports, contradicts, and neutral must be explicit. No result or a failed search means unknown, not refuted.
- No trial record means trial saturation is unknown. Only a completed zero-hit search may support a low-saturation interpretation.
- Preserve PMID, DOI, NCT, Ensembl, ChEMBL, and source-tool provenance exactly as returned. Never invent identifiers.
- Cross-modal corroboration requires independent sources; duplicate database rows are not replication.
- Rankings prioritize investigation only. They do not establish mechanism, efficacy, safety, or clinical utility.
- For PHI or patient-level inputs, use
bio-privacy first and do not place identifiers in an evidence context.
Minimum final table
For every shortlisted target report: score, assessed modalities, missing modalities, strongest evidence, explicit conflict status, trial-saturation basis, and a falsifiable next experiment.