Optimizer Workflow

The LatticeZero Optimizer finds the scoring weight combination that best separates known drugs from known non-drugs for a specific target. This guide explains how the optimizer works and when to use it.

:::note This guide covers the supervised optimizer (requires actives + decoys). For unsupervised profile generation, see Autotune. :::

What Is a Scoring Profile?

A scoring profile is a set of weights that tell the scoring engine how much to care about each physics term for a specific target. The default "equal weight" profile gives AUC ~0.50 (random discrimination). An optimized profile can achieve AUC 0.70–0.98 for the same target.

Why target-specific weights matter:

• ACE (metalloprotease): weights metal coordination heavily (ZBG score)
• CDK2 (kinase): weights hinge hydrogen bonds and depth heavily
• MCL1 (PPI): weights burial and inside-fraction heavily
• ESR1 (nuclear receptor): weights aromatic burial and depth heavily

When to Use the Optimizer

Use the Optimizer when you have:

• A prepared target (receptor + pocket definition)
• Known actives - SMILES or SDF of confirmed binders (≥20 recommended)
• Known decoys - SMILES or SDF of non-binders (≥200 recommended, ideally 30× actives)

The DEKOIS2 benchmark sets provide these for 27 validated targets. For custom targets, use ChEMBL actives with DUD-E or DEKOIS2-style decoys.

The Optimization Pipeline

Step 1 - Input Data

You need a scores CSV with physics term columns for all compounds. The typical pipeline:

1. Dock all actives + decoys with IsoPose (or use pre-docked poses)
2. Export the scores CSV (includes all 14 physics terms + geometric features)
3. Merge with labels (1 = active, 0 = decoy)

Step 2 - Optimization Method

The optimizer uses differential evolution (SciPy DE) to search the weight space:

• Initialization - warm-start from target-class priors (not random)
• Objective - maximize mean holdout AUC across 5 cross-validation folds
• Search - DE tries thousands of weight combinations: maxiter=500, Sobol initialization
• Regularization - L2 penalty prevents extreme weights
• Polish - L-BFGS-B local optimizer fine-tunes after DE convergence

Step 3 - Validation

Results are only accepted if they pass validation:

Y-scramble is mandatory: the optimizer runs again with randomized labels. If it still finds high AUC, the signal is fake (likely due to property biases in the dataset - avoid using MW, TPSA, logP as features).

:::warning Physics features only. Never include molecular descriptors (MW, logP, TPSA, HBD, HBA) in the feature set - they inflate AUC via DEKOIS2 property-matching artifacts without any binding information. :::

Step 4 - Tier Assignment

Feature Importance

After optimization, the profile shows which features drive discrimination:

• Large positive weights - terms that are significantly better in actives
• Large negative weights - terms that are significantly worse in actives (penalize if high)
• Near-zero weights - terms with no discriminative power for this target

This is scientifically informative: high E_hbq weight confirms hydrogen bonds are critical for the target. High inside_frac weight confirms tight burial is key. High ZBG weight confirms zinc coordination is essential.

Re-optimization Triggers

Consider re-optimizing a profile when:

1. Engine update - new physics terms added or existing terms recalibrated
2. More data - additional actives or decoys available
3. Demo drift - health dashboard shows demo AUC significantly above validated AUC (re-opt candidate)
4. New target class - novel target type not well-represented by existing priors

The Platform Health dashboard (admin only) tracks demo vs. validated AUC for all 18 benchmark targets and flags re-optimization candidates automatically.

Validated Profile Library

LatticeZero ships with 27 pre-validated profiles across 5 target classes:

See Benchmark Results for full DEKOIS2 validation numbers.