version: "1.0.0" name: codex-surface-targets description: Use Rosalind Workbench plugins to identify disease-associated cell-surface targets, compare binding sites and known interventions, and select targets for molecule screening or binder design. Build linked target pages with source evidence, structures, and results.

Codex Surface Atlas

Use this skill to find cell-surface targets in a disease state, compare their evidence and binding sites, and select targets for screening or binder design. Start with a disease and cell population, a supplied molecule library, or an existing atlas. Adapt the sources, molecule formats, and tools to the project.

Start with Rosalind Workbench plugins. Use Life Sciences Databases and Life Sciences Literature to retrieve evidence, NGS Analysis Workbench to analyze sequencing data, and Molecular Structure Viewer or Biological Sequence & Alignment Viewer to inspect structures and constructs. NVIDIA BioNeMo Agent Toolkit and Biohub ESM provide modeling workflows; Adaptyv Bio connects selected candidates to experiments. Read plugin recipes for prompts and ways to combine these tools.

Select the installed plugins that fit the project and read their skills before use. The CLI creates the atlas workspace, registers supplied libraries, validates records, and assembles the report. Add local tools or cloud compute where a selected recipe needs preparation, modeling, or analysis.

Read installed local workflows before writing custom intake or report extensions. The CLI already supplies a complete offline replay tutorial, deterministic evidence intake and reconciliation, action comparison, verified sequence/site selection, binder-run import, and assay-return validation. Use these contracts to preserve identities and source artifacts through the workflow.

Public companions include Codex Binder Lane for binder campaigns, BioSymphony Structure Factory for structural campaign planning and execution workflows, and Proteus for structure analysis and rendering. Keep their returned results linked to the atlas's target, site, and run records.

The workflow is:

disease evidence
  -> accessible target census
  -> known interventions
  -> structures, complexes, and accessible sites
  -> action and modality choice
  -> selected screening and design workflows
  -> evidence-linked offline report

Record the disease, population, and intended use

Before gathering evidence, record the disease or subtype, organism, population and cell context, evidence cutoff, purpose, requested molecule formats, and compute budget. Specify which data may be sent to which external services. Read workflow and data contracts before adding records.

Locate surface-atlas on PATH first. If unavailable, check a user-supplied executable path or the project's .venv/bin/surface-atlas on macOS or Linux, or .venv/Scripts/surface-atlas.exe on Windows. Resolve these paths against the project where the CLI was installed. Verify the executable with --version and use it for every CLI command in this workflow. If no executable is available, report the missing CLI and refer to the installation guide. Install packages only when the user requests installation.

Create a local workspace with that executable:

surface-atlas init ATLAS_ID OUTPUT_DIRECTORY \
  --disease "DISEASE OR SUBTYPE" \
  --profile standard --json

The initializer refuses an existing destination and writes the versioned plan, search ledger, and empty collections. Structure tiers set how many targets receive each depth of review; retain the complete searched target census.

Review consequential decisions independently

Read adversarial review and launch bounded independent subagent reviews at the relevant checkpoints: target census and normal-tissue evidence, binding site and construct, screening or design with controls, and final report claims. Use an available equivalent when native subagents are unavailable; adapt the assignments to the user's scope and the materials available. Run independent reviews in parallel when supported.

Give reviewers exact files, record IDs, sources, and the decision to challenge. Request supported findings with the consequence and proposed correction. Check findings against the evidence, correct accepted issues, rerun affected checks, and record unresolved limitations before dependent work. Review does not authorize additional compute or establish scientific proof.

Find targets and record the evidence

Search every source class named in the plan. Record the query, retrieval date, returned count, inclusion rule, exclusion reason, disease and cell context, source identifier, and limitation. Keep every normalized discovery row, including retained, excluded, and unresolved entities. Record an exclusion only when a source supports the reason for excluding that entity.

For retained entities, resolve stable identity, aliases and isoforms, extracellular topology, disease alteration, surface evidence, normal-tissue exposure, shedding and internalization, known ligands and interventions, structure availability, and unresolved measurements. Keep observations, derived prioritization, inferences, and hypotheses in separate fields. Read the evidence model and target reconciliation.

Select binding sites and molecule formats

Record the molecule's target and the effect it is intended to produce. Specify its format, binding site, supporting structure, risks, missing measurements, and proposed analysis or experiment. Internalization may support payload delivery while reducing surface occupancy; explain how it affects the intended action. Use the structure atlas and the molecular-library guide when those views are in scope.

When independent target-site screens answer the same question, preserve each run's source, controls, and confirmation context. Use multiple screening runs before merging their local records.

Any paid compute, private-data transfer, experimental submission, provider license acceptance, or campaign expansion requires an explicit authorization record. Preserve route, model and version, scale, destination, price state, spend ceiling, stop rule, controls, and expected artifacts. A predicted pose or complex remains a computational hypothesis. Use Codex Binder Lane to coordinate protein-binder design and evaluation. Read structure-to-design and compute budget first.

Validate and deliver

Read reader workflow when preparing the report. Open pages with their recorded findings and useful actions. Build comparisons around the user's intended action, and identify the next measurement that could change a target or candidate selection. Name the corresponding Rosalind plugin recipe and its required inputs.

When interactive exploration fits the project, connect sequence regions, binding sites, and structures through verified residue maps. Preserve the reader's selection and required files in copied Codex requests. Have an independent reviewer check filters, residue links, downloads, and requests against the source records. Choose extensions that fit the available data.

Use the package CLI for local validation and report assembly:

surface-atlas validate ATLAS_DIRECTORY --json
surface-atlas report ATLAS_DIRECTORY --output-root REPORT_ROOT \
  --run-id RUN_ID --json

The report is offline-capable and includes local JSON and CSV exports, source links, target pages, structures and artifacts that passed their recorded integrity checks, and a run manifest. Read report presentation and visual delivery for output and accessibility requirements. Validate the report's files and review its claim ceiling; the report does not establish efficacy, safety, or clinical suitability.