version: "1.0.1" name: gtm-density-landscape description: Build or reuse a GTM map, create density landscapes, inspect compound distributions, and prepare report-ready density artifacts. metadata: title: GTM density landscape status: stable version: 2.0.0 depends_on: [] profiles:

• gtm-analysis

permissions:

• compute:execute
• artifact:read
• artifact:write

input_artifacts:

• name: clean_dataset_path

kind: dataset required: true

• name: gtm_model_path

kind: model required: false output_artifacts:

• name: gtm_model_path

kind: model required: true

• name: projected_dataset_path

kind: dataset required: true

• name: density_matrix_path

kind: analysis-table required: false

• name: density_plot_path

kind: visualization required: true tags:

• gtm
• chemography
• density-landscape
• compound-distribution

keywords:

• gtm
• density
• density map
• density landscape
• compound distribution
• densest nodes

required_tools:

• chemspace_plan_analysis
• gtm_save_density_plot
• gtm_get_density_summary

optional_tools:

• gtm_optimization
• gtm_save_model_and_data
• gtm_load_density_matrix
• gtm_load_model_only
• gtm_load_and_prep_data
• gtm_sample_dense_nodes
• gtm_project_data
• report_save_rich

example_prompts:

• Show the GTM density landscape for the current clean dataset and summarize the densest nodes.

GTM Density Landscape

Use this skill for GTM density maps, compound distribution analysis, dense-node inspection, and report-ready density artifacts.

Procedure

1. Call chemspace_plan_analysis with the user's request and available session summary.
2. If preflight returns needs_clarification=true, ask the returned clarification questions before calling mutating GTM tools.
3. Resolve the active clean dataset from session memory or from an explicit user path. Prefer clean_dataset_path; dataset_path is only a legacy clean-data alias.
4. Respect map mode when available. default_map means project onto the pretrained default map unless the user explicitly asks to build/train a new map. new_map or missing means use the session-local GTM behavior.
5. Reuse an existing GTM map from session state when available and appropriate. Load it with gtm_load_model_only and prepare the active dataset with gtm_load_and_prep_data.
6. Build a new GTM only when no suitable map exists or the user asks for one. Default to a low optimization strategy unless the user explicitly requests medium/high/thorough search. If building, call gtm_optimization, then persist the model and projected data with gtm_save_model_and_data.
7. If the user supplies new data or generated candidates for an existing map, materialize the candidate set if needed and use gtm_project_data before density analysis.
8. For density visualizations in MCP mode, call gtm_save_density_plot; gtm_load_density_matrix returns density tables but does not itself write plots.
9. Summarize dense regions with gtm_get_density_summary. Sample compounds from dense nodes with gtm_sample_dense_nodes only when sampling is part of the request.
10. Do not describe dense nodes as active or potent unless measured activity values from a loaded table or tool output support that claim.
11. Save or reference density plot artifacts when preparing a report. Put density figures directly after density analysis in report sections.

Expected Outputs

• Fitted or loaded GTM model.
• Projected source dataset.
• Density matrix or density summary table.
• Static density plot artifact.
• Summary of dense regions and compound distributions suitable for downstream report generation.

Details

• Density table columns are x, y, nodes, filtered_density. Report global stats (max / min / mean / median density) and identify the top-5 densest and top-5 sparsest nodes.
• Neighborhood-preservation table columns are x, y, nodes, density, neighborhood score; report preservation quality and flag well-preserved vs poorly-preserved regions.
• Describe spatial patterns in compass/quadrant terms (e.g. "dense band across the center", "sparse south-west corner").
• Close with a 3-bullet executive summary of the density structure.
• Optimization strategy (when building): gtm_optimization(strategy=...) levels are low (fast heuristic, ~9 combinations), medium (extended grid, up to 144 combinations), high (Bayesian, ~50 trials). Always low by default; for datasets >5,000 molecules always low and tell the user medium/high are available.