Documentation

Login
Docs / Lead Optimization

Lead Optimization

Lead Optimization is the customer-facing LP+ workspace for staging analog pairs, choosing target context, reviewing chemistry deltas, and saving provenance-rich reports. LP+ Benchmarks is the separate evidence surface for JACS2015, Merck, showcase packs, and historical validation results.

What The Split Means

Surface Purpose Honest status
Lead Optimization Chemist workflow, pair staging, profile intent, saved reports Production workspace path in progress
LP+ Benchmarks Published benchmark packs, showcase/reference runs, validation evidence Benchmark surface, not the module

This split is deliberate. Benchmark evidence should stay visible, but it should not masquerade as the chemist workflow for arbitrary molecule edits.

What You Can Do Today

  • Stage related molecule pairs directly in the module landing
  • Upload a pair CSV with reference and analog SMILES
  • Choose the target pack context that the pair set should inherit
  • Capture the intended profile path:
    • default pack recipe
    • calibrated live path
    • native IsoPose handoff
    • benchmark reference only
  • Review scaffold continuity and chemistry deltas such as:
    • DeltaMW
    • DeltalogP
    • DeltaTPSA
    • DeltaHBA
    • DeltaHBD
  • Save or export the pair queue while the full live module path is still being completed

Current Target Refinement Snapshot

  • CDK2 is now customer-visible as Calibrated live with browser-validated directional signal around rho = 0.861
  • JNK1 is now customer-visible as Calibrated live with browser-validated directional signal around rho = 0.775
  • TYK2 remains visible, but should stay qualified as beta because the calibrated-live signal is weaker (rho = 0.311) and the dataset is only 12 edges

This matters because the UI should distinguish:

  • Showcase
  • Raw live physics
  • Calibrated live physics
  • Beta calibrated live

Current Limitation

The module does not yet promise truthful LP+ DeltaDeltaG for arbitrary chemist edits from plain SMILES alone.

Why:

  • real LP+ scoring needs receptor context
  • it also needs 3D ligand poses or a docking/preparation path
  • benchmark packs already contain that context
  • arbitrary chemist edits still need the SMILES-to-3D module path to be fully wired

That means the workspace is honest about two states:

  • staged for module scoring
  • predicted DeltaDeltaG available

If a pair does not yet have a validated LP+ value, it should stay labeled as awaiting live module scoring.

Minimum Chemist Workflow

  1. Pick a target pack context.
  2. Add a reference compound and a modified analog.
  3. Record the intended change, such as:
    • add methyl
    • swap heteroatom
    • extend chain
    • add halogen
  4. Review whether the pair appears to stay in the same scaffold series.
  5. Inspect chemistry deltas before committing to a live LP+ run or export.
  6. When the live path is available, run LP+ and save the resulting report with provenance.

Pair Upload Format

The module pair uploader accepts a CSV or tab-separated file with these columns:

Column Required Meaning
ref_name optional Name of the starting compound
ref_smiles yes Reference compound SMILES
analog_name optional Name of the modified analog
analog_smiles yes Analog SMILES
change_note optional Human-readable edit description
predicted_ddg optional LP+ DeltaDeltaG if already computed elsewhere

If predicted_ddg is omitted, the pair is still useful and should remain in an awaiting-scoring state.

Interpreting LP+ Correlation

The question chemists usually ask is: "What does rho = 0.5 actually mean for my project?"

rho range Practical meaning How to use it
>= 0.70 Strong directional signal Good for prioritizing close analogs and deciding synthesis order
0.50 - 0.69 Real but imperfect signal Good for triage and edge direction, not precise potency sizing
0.35 - 0.49 Weak exploratory signal Use for hypothesis generation only
< 0.35 Not decision-grade Do not rely on it for medicinal chemistry decisions

rho ~= 0.5 usually means the target is useful for "better vs worse" calls, but not yet strong enough to trust the exact predicted magnitude of every edit.

Raw Live vs Calibrated Live vs Showcase

The LP+ product now distinguishes these modes explicitly:

Mode Meaning
Raw live physics Live LP+ scoring from the current physics recipe
Calibrated live physics Live LP+ scoring plus a calibrated model layer on top of the live feature outputs
Showcase / precomputed Pre-calculated benchmark predictions, kept separate from the module

These should never be conflated in customer copy or saved provenance.

Saved Provenance

Every saved LP+ report should preserve enough information to answer:

  • which target pack was used
  • whether the run was benchmark or module
  • whether the path was raw live, calibrated live, native handoff, or showcase
  • which pose source was used
  • which profile or profile intent was selected

That provenance matters just as much as the final correlation number.

On this page
What The Split Means What You Can Do Today Current Target Refinement Snapshot Current Limitation Minimum Chemist Workflow Pair Upload Format Interpreting LP+ Correlation Raw Live vs Calibrated Live vs Showcase Saved Provenance Related Docs