Appearance
Target Discovery
Watch the Target Discovery workspace walkthrough.
The 5-minute-51-second walkthrough uses edited actual browser captures of the frontend, source review and recorded foundation/evidence checks. Its public data extract, local response/store adapters and synthetic proposal controls are identified. Test screens identify the foundation and evidence revisions; the ranking PR records its own exact revision and pre-commit checks. It does not demonstrate live backend ingestion, experimental efficacy or maintainer approval.
Open View → Exploration → Target Discovery, or choose Layouts → Target Discovery to place the workspace beside the map, Heatmap Explorer, Knowledge Synthesizer and Reasoning Cards. The workspace connects Data → Hypotheses → Ranking → Proposal. Existing analysis and reasoning tools can also be opened from its tool list.
1. Inspect the data
Import an H5AD through the existing H5AD source workflow. The panel reads the map's ingestion dataset_id with the existing authenticated H5AD client. It does not upload, alter, preprocess or train on the source.
Read the reported cell and feature totals, choose a cell-type annotation column, and inspect available QC columns, representations, embedding shapes and alignment. Distinct annotation totals and sample labels are not cell-type cohort counts. Column presence does not establish that QC passed; unreported information stays unknown. Use the existing analysis panels to inspect preprocessing decisions. Legacy sources with absent or null shape/axis metadata display Not reported; missing metadata does not imply that the dataset is empty or invalid.
Search the source gene catalog, select 1–200 map cells and inspect one gene's stored X measurements. The table preserves the returned cell-to-value association, including reordered responses and missing values. These may be processed or scaled values rather than raw counts. Expression is descriptive; it does not establish a causal disease role.
Start hypothesis with this evidence records the dataset, feature, exact cell IDs and corresponding original source-cell IDs, source URI, descriptive summary and a contextual evidence stance. A map without an H5AD source can still hold explicitly cited external research hypotheses; no single-cell summary or measurements are fabricated for it.
2. Interpret evidence
Record a gene/protein, cell type, cell state, disease context, intervention direction and modality, proposed mechanism, expected benefit, uncertainty, liabilities and the next experiment's comparison and readout. State unknowns explicitly when the evidence is insufficient.
Attach source links and concise summaries as supporting, contradictory or contextual evidence. Keep counter-evidence visible. Selected map records can also be attached with their exact IDs; a selection does not by itself establish a biological claim. Select source cells restores the exact loaded source selection and opens Inspector; missing cells or a different map/dataset leave the selection unchanged. HTTP(S) and Mantis evidence URIs are accepted; executable, credential-bearing and local-file URIs are rejected. Map-level references without a dataset ID apply to their linked map; references with an explicit dataset ID must match the current source.
Text edits save immediately to the session, preserving spaces and newlines. Assessment form changes apply only with Save assessment. Every numeric assessment needs a rationale and links to existing evidence. Removing a source resets every assessment citing it to unknown so a score never silently outlives its evidence.
3. Rank research priorities
The transparent v1 rubric uses four equally weighted, user-assessed criteria:
| Criterion | What to justify with linked evidence |
|---|---|
| Causal support | Proposed causal disease role in this context |
| Intervention | Plausible intervention and modality |
| Benefit | Expected benefit in the stated cell and disease context |
| Safety | Liability and safety evidence, including unresolved risks |
Scores run from 0 (evidence argues against) through 1 (weak), 2 (mixed), 3 (supportive), to 4 (strong). Unknown remains unassessed. Priority is sum / 16 × 100; unknown contributes zero without being presented as an observed zero. The list reports assessment coverage, contradictory source count and missing proposal fields. Ties sort by target, then hypothesis ID.
Expand a row to inspect each criterion's rationale and every attached source, including uncited counter-evidence, intervention direction, mechanism, uncertainty and liabilities. The score is an authored research prioritization aid, not clinical efficacy, a probability or automatic causal inference. No model is trained and no disease treatment is recommended.
Six-field proposal and backups
The proposal view exports JSON or Markdown with exactly these six fields:
- Target and complete cell/disease context.
- Intervention direction and modality.
- Mechanism and expected benefit.
- Evidence, including contradictory and contextual sources with provenance.
- Uncertainty and safety liabilities.
- A testable next experiment.
An incomplete proposal cannot be exported as complete. The export also includes the map, dataset, hypothesis ID and transparent assessment rationale. The JSON preview exposes the same content as the JSON download. Both JSON and Markdown preserve attached cell IDs and dataset provenance. Missing scientific evidence can be acknowledged explicitly; a structurally complete proposal is not equivalent to experimentally validated science.
Drafts belong to the current map instance and dataset session. They are not saved to the server or shared with collaborators. Export a workspace JSON backup before leaving, then restore it into the same map and dataset. Restoring replaces session drafts only after confirmation and validates the complete file before applying it. A concurrent edit invalidates a pending import. Separate maps and datasets never share draft stores.
Bounds keep UI and agent responses inspectable: at most 100 hypotheses, 50 sources per hypothesis, 200 attached cells, 32,000 formatted characters per hypothesis and 300,000 formatted characters per workspace backup. Imports are limited to 1 MB. Link full source documents instead of pasting them into summaries.
Agent controls
The frontend Drive manifest advertises five standard-level commands. They use the same store, validation, ranking and export as the user interface:
| Command | Purpose |
|---|---|
target_discovery.open | Open/focus a workspace bound to a loaded map |
target_discovery.get | Page ranking summaries (1–10 rows); pass id for one full hypothesis |
target_discovery.load | Refresh the existing imported H5AD summary |
target_discovery.edit | Create, save, remove or navigate with source/revision guards |
target_discovery.export | Return one complete JSON or Markdown proposal |
For example, an agent reads the workspace, then creates a fully shaped unknown draft with { action: "create", mapId: <get.mapId>, datasetId: <get.datasetId>, revision: <get.revision> } as the edit command's operation. To save, supply a full hypothesis and the latest revision. Navigate with the exact dataset ID, stage and optional hypothesis ID. Every edit operation includes the exact map ID too. Export with dataset_id: <get.datasetId> (omit it for unlinked maps) and format: "json" or "markdown". Pass map_id when controlling a map other than the active view. Export also accepts datasetId for consistency with edit operations; if both spellings are supplied, their values must agree. Reopening an existing panel preserves layout undo/redo. An unbound user-opened panel is reused only for the active map; other maps receive their own bound panel. Stale writes, cross-dataset replay and expired edit commands change nothing. The manifest must be supported by the connected Drive service; this frontend change does not deploy that service or invent server-side analysis endpoints.
Verification data
Tests include a bounded extract of the public Scanpy PBMC3k processed example. The 2,638-cell, 1,838-feature H5AD is pinned to the hash published in Scanpy's dataset registry. Its fixture retains actual annotations, stored embeddings, eight cell IDs and stored X values for four available genes. Its searchable catalog contains only those four genes; the manifest retains the full source dimensions. The local verification API translates those source records into the existing client response shapes; it is not a live MantisAPI ingestion. Other proposal and score examples are explicitly synthetic, not biological conclusions.
Developers can reproduce the extract with scripts/target-discovery/extract-pbmc3k-fixture.py and a local h5py environment. Application dependencies do not change. Regression tests cover scoring, formatting, unsafe sources, backup bounds, source ownership, asynchronous races, agent replay, real-data expression attribution and the connected proposal flow.