Autotune

Autotune generates custom scoring profiles for your target without requiring labeled data (active/decoy sets). Upload your receptor and compound library, and Autotune analyzes the binding pocket to produce an optimized weight profile.

Overview

How It Works

Autotune uses a 4-step process:

1. Upload & Detect - Upload your receptor PDB and compound library SDF. Autotune auto-detects the target class (kinase, protease, GPCR, etc.) from the protein name and binding pocket features.

2. Generate Profile - Based on the detected target class, Autotune interpolates between validated reference profiles using AUC-weighted blending. This produces a custom weight vector tuned for your target's characteristics.

3. Review Weights - Examine the generated scoring weights across all 14 physics terms. A confidence score indicates how well the reference profiles cover your target class. Higher confidence means more reliable predictions.

4. Export / Save - Download the profile as a JSON pack, or save it directly to your account for use in IsoPose and IsoScore.

Scientific Methodology

Profile Interpolation

Autotune does not train a model from scratch. Instead, it leverages a reference library that includes validated demo profiles, dock-mode profiles, and class-level physics recipes. When validated references exist for the detected class, they dominate the blend; class-level recipes provide the untuned fallback path. The key insight is that targets within the same structural class tend to share similar binding physics.

The algorithm works by:

1. Target class detection - Analyzing the protein name, fold classification, and binding pocket properties to assign one of 7 structural classes.
2. Reference profile selection - Retrieving validated references for the detected class (plus class-level fallback recipes for edge cases).
3. AUC-weighted blending - Each reference profile contributes proportionally to its validated AUC. Higher-performing profiles have more influence on the blend. This is a form of performance-weighted ensemble averaging.

The resulting weight vector is a weighted centroid of proven profiles - a conservative starting point that captures the consensus physics for the target class.

Why Unsupervised?

Traditional scoring function optimization (as in the Optimizer) requires labeled data: compounds known to bind (actives) and compounds known not to bind (decoys). For most real-world drug discovery campaigns, this data doesn't exist yet - you have a target and a compound library, but no binding data.

Autotune bridges this gap by transferring knowledge from well-studied targets to new ones. The assumption is that a kinase binding pocket shares common physics (hinge hydrogen bonds, hydrophobic back pocket, gatekeeper interactions) with other kinases, even if the specific residues differ.

Validation

Autotune profiles are assessed using the same metrics as supervised profiles:

• Estimated tier - Based on the AUC distribution of contributing reference profiles
• Confidence score - Reflects the number and quality of reference profiles for the target class
• Cross-validation - When benchmark data is available, Autotune profiles are compared against supervised optimization

In internal testing, Autotune profiles typically achieve 75-85% of the AUC of fully supervised profiles for well-covered target classes (kinases, proteases). For underrepresented classes, performance may be lower.

Target Classes

Autotune recognizes 7 target families:

Class Detection

Class detection uses a cascading approach:

1. Name matching - Common target names and families are matched against a curated dictionary
2. Fold classification - ECOD/SCOP fold annotations, when available, map to target classes
3. Pocket analysis - Pocket volume, polarity ratio, and metal presence serve as fallback features
4. Manual override - Users can always override the auto-detected class

Using Autotune

From the Web Interface

1. Navigate to Workbench → Autotune in the sidebar
2. Upload your receptor PDB file
3. Upload your compound library SDF
4. Review the detected target class (override if needed)
5. Click Generate Profile
6. Review the weight breakdown and confidence score
7. Click Save Profile or Export JSON

Using the Saved Profile

After saving, your custom profile appears in the IsoPose profile selector dropdown under your target class. Select it when submitting docking jobs to use your optimized scoring weights.

Confidence Score

The confidence score (0-100%) indicates how well Autotune's reference library covers your target:

Confidence is computed from: (1) the number of reference profiles for the target class, (2) their AUC distribution, and (3) the similarity of the target to reference proteins.

When to Use Autotune vs. Optimizer vs. Default

Limitations

• Novel target classes: Autotune relies on transfer from known classes. Truly novel binding modes (allosteric sites, protein-protein interfaces not in the reference set) may produce low-confidence results.
• Covalent targets: Profile weights are optimized for non-covalent scoring. Covalent binder recognition requires additional terms not in the current scoring function.
• Small library bias: Very small compound libraries (<50 compounds) provide less statistical power for pocket analysis features.
• Not a replacement for experimental validation: Autotune profiles are computational predictions. Always validate top hits experimentally.

Further Reading

• Scoring Profiles - How profiles weight the energy function
• Benchmark Results - Validation methodology and results
• Optimizer - Supervised optimization with labeled data
• IsoScore Methodology - How the scoring function works