Documentation

Login
Docs / Tools / IsoPose Lab

IsoPose Lab

IsoPose Lab is the validation and diagnostics workspace for IsoPose bundles. It is where you compare live WebGPU docking against each target's locked benchmark expectations, inspect search-vs-rescore behavior, and run multi-target validation campaigns.

The guide is available at both /docs/tools/isopose-lab and /docs/isopose-lab.

Demo Or Run Mode Vs Lab Mode

Surface Best for What it emphasizes
Demo / Run mode Single-target walkthroughs and user-facing runs Guided setup, polished results, simpler controls
IsoPose Lab Validation, parity tuning, and benchmark diagnostics Expected AUCs, run-mode controls, dock-vs-rescore comparisons, campaign execution

Use Demo / Run mode when you want to show how a target behaves for a normal user flow. Use IsoPose Lab when you need to answer questions like:

  • Is the live WebGPU run still matching the locked benchmark?
  • Did a target regress in pose search, rescoring, or both?
  • Which subset, speed preset, or run mode explains the difference?
  • Does a whole campaign still pass across multiple targets?

Core Metrics

dockAuc

dockAuc measures the ranking quality of the raw docking stage. It answers: "If I rank ligands directly from the docking/search output, how well do actives separate from decoys?"

cascadeAuc / rescoreAuc

The post-docking metric is the rescored ranking after the physics/profile pipeline runs on the selected poses. In code and artifacts this often appears as rescoreAuc. In discussion it is also called cascadeAuc because it comes from the cascade rescoring stage.

How to interpret them together

  • dockAuc low and rescoreAuc low: the search stage is probably failing, or the bundle itself needs attention.
  • dockAuc acceptable but rescoreAuc lower: the rescore/profile stage is the likely regression point.
  • rescoreAuc higher than dockAuc: expected for targets where the scoring profile adds useful discrimination after docking.
  • dockAuc higher than rescoreAuc: keep the dock ranking only when the bundle notes explicitly say the dock ordering is the retained gold path.

The Rescore Pipeline

IsoPose Lab is a two-stage workflow:

  1. Pose search Raw docking generates poses and a dock ranking.
  2. Feature extraction The selected poses are converted into physics features such as burial, clashes, hydrogen-bond quality, water terms, and related target-specific signals.
  3. Profile scoring The target's locked scoring profile weights those features into the final ranking.
  4. Benchmark comparison The live result is compared against the target's published AUC, floor AUC, and validation source metadata.

That separation is why Lab mode is useful for debugging. It makes it obvious whether a regression came from the search stage, the profile stage, or the benchmark expectation itself.

Validation Campaigns

Campaigns batch the same validation workflow across multiple targets.

Use campaigns when you want to:

  • rerun a fixed subset after changing scoring logic
  • compare a preset across a full target class
  • sanity-check a registry update before promoting it

The campaign panel tracks:

  • selected subset
  • target filter or custom target list
  • per-target status
  • per-target dockAuc
  • pass/fail verdict
  • aggregate mean AUC and total runtime

Because campaigns reuse the same target metadata the single-target flow uses, they are the quickest way to spot whether a change is isolated or systemic.

Target Card Anatomy

Each target entry in Lab mode is backed by the registry plus the bundle manifest. The important fields are:

Field Meaning
Display name / target ID Human label plus the stable registry identifier
Status Registry state such as GOLD_LOCKED
Validated AUC The expected benchmark AUC the live run is compared against
Floor AUC Minimum acceptable benchmark threshold for that bundle
AUC source Where the locked benchmark came from in bundle metadata
Scoring mode The ranking path the bundle expects, such as geometric, hbq-optimized, or a target-specific mode
Scoring profile The named profile or manifest-derived gold profile behind the rescore stage
Run mode / baseline variant The execution path the target expects for parity, such as auto, native, or router_v7

Read the card like this:

  1. Confirm the target is runnable and GOLD_LOCKED.
  2. Note the validated AUC and floor AUC before starting a run.
  3. Check whether the selected run mode matches the target's expected path.
  4. Compare live dockAuc and post-rescore AUC against the validated AUC after the run completes.

When To Use Lab Mode

Use IsoPose Lab when you need:

  • target-by-target parity checks
  • campaign sweeps across many bundles
  • confidence that a scoring or routing change did not silently drift a benchmark
  • a fast explanation of whether a discrepancy comes from docking, rescoring, or metadata

For normal product walkthroughs and one-off runs, Demo / Run mode is still the simpler entry point. Lab mode is where you go when you need answers rather than just output.

On this page
Demo Or Run Mode Vs Lab Mode Core Metrics dockAuc cascadeAuc / rescoreAuc How to interpret them together The Rescore Pipeline Validation Campaigns Target Card Anatomy When To Use Lab Mode Related Pages