dp-export-deep-feature-inputs

dp-export-deep-feature-inputs is the step that prepares a completed segmentation result for DeepProfiler.

It tells DeepProfiler which multichannel Cell Painting image belongs to each field and where the segmented cells are located inside that field.

Purpose

Use this skill when you want:

  • to carry a segmentation result into the deep-feature branch

  • one DeepProfiler-ready image record for each field

  • one coordinate table for the segmented cells in each field

  • a clean handoff from segmentation into deep-feature extraction

Main Outcome

This skill does not calculate learned features yet.

Its job is to tell DeepProfiler two things:

  • which multichannel image belongs to each field

  • where each segmented cell is located in that field

In other words, this is the step that turns a segmentation result into model-readable image fields plus cell positions.

Inputs

This skill reads:

  • a project config such as configs/project_config.demo.json

  • the completed segmentation image table

  • the completed nuclei table

  • the segmentation load-data table that records the Cell Painting channel paths

  • an optional output directory

In the demo setup, these inputs come directly from the completed segmentation result.

Outputs

The main user-facing results are:

  • one DeepProfiler-ready image record for each field

  • one coordinate table for the segmented cells in each field

  • a compact summary of how many fields and cells were prepared

Agent Demo

This page is based on a real local OpenClaw run:

  • turn timestamp: 2026-05-06 18:41 GMT+8

  • model: vibe/gpt-5-mini

Request

I already have a completed segmentation result and now want to move into the DeepProfiler branch. Please prepare the demo image fields and single-cell locations that DeepProfiler will need, then explain what was prepared biologically.

Structured Trace

user_input:
I already have a completed segmentation result and now want to move into the DeepProfiler branch. Please prepare the demo image fields and single-cell locations that DeepProfiler will need, then explain what was prepared biologically.

planner_step_1:
intent:
Load the documented Cell Painting routing guidance before choosing a DeepProfiler skill.
tool_calls:
- read("integrations/openclaw/autodl/workspace/skills/cellpaint-pipeline/SKILL.md")
observed_result:
- the workspace skill file loaded successfully
- routing guidance pointed to `dp-export-deep-feature-inputs` for requests that stop at the DeepProfiler export bundle

planner_step_2:
intent:
Confirm the documented skill contract before running it.
tool_calls:
- exec("cd $REPO_ROOT && cellpainting-skills describe --skill dp-export-deep-feature-inputs")
observed_result:
- the skill description confirmed that this step writes `manifest.json`, `images/field_metadata.csv`, and `locations/`

planner_step_3:
intent:
Run the documented export skill against the demo config and requested output root.
tool_calls:
- exec("cd $REPO_ROOT && /root/autodl-tmp/miniconda3_envs/lyx_env/bin/cellpainting-skills run --config $REPO_ROOT/configs/project_config.demo.json --skill dp-export-deep-feature-inputs --output-dir $REPO_ROOT/demo/workspace/outputs/agent_demo_dp_inputs")
observed_result:
- skill_key = dp-export-deep-feature-inputs
- field_count = 2
- location_file_count = 2
- total_nuclei = 4
- output_root = /root/autodl-tmp/pipeline/CellPainting-Claw/demo/workspace/outputs/agent_demo_dp_inputs

planner_step_4:
intent:
Inspect the exported field manifest and cell-location files so the biological meaning can be explained back to the user.
tool_calls:
- exec("read field_metadata.csv preview")
- exec("read A01 and A02 location CSV previews")
- exec("read manifest.json")
observed_result:
- the export contained two fields:
  - BR00000001 / A01 / site 1
  - BR00000001 / A02 / site 1
- each field linked the five Cell Painting channels: DNA, RNA, ER, AGP, and Mito
- A01/site_1.csv listed 2 nuclei
- A02/site_1.csv listed 2 nuclei

final_answer_snapshot:
The DeepProfiler export completed successfully and prepared 2 image fields plus per-field cell locations for 4 total segmented cells.

Observed Result

The recorded OpenClaw turn exported a complete DeepProfiler input bundle for:

  • 2 image fields

  • 2 location files

  • 4 total segmented cells

The two exported fields were:

  • plate BR00000001, well A01, site 1

  • plate BR00000001, well A02, site 1

Recorded Reply

I exported the DeepProfiler input bundle successfully.

Summary of what was produced:
- fields exported: `2`
- nuclei location files: `2`
- total nuclei listed: `4`

What this means biologically:
- each image field is now paired with the five Cell Painting channels that describe the cell state in that field
- each segmented cell now has a recorded location that DeepProfiler can use to center a single-cell view
- the segmentation result has therefore been translated into a form that the deep-feature model can read directly

Example cell locations from this export:
- `A01/site_1.csv`:
  - cell 1 at `(22, 20)`
  - cell 2 at `(40, 42)`
- `A02/site_1.csv`:
  - cell 1 at `(24, 22)`
  - cell 2 at `(38, 44)`

Next Skills