Albumin Binding Classification
Human serum albumin binding-site classification and binary binding prediction
Executive Summary
The Spirographic AI albumin binding classifier predicts whether a drug binds human serum albumin (HSA) at clinically significant levels (≥50% bound) using structural analysis of the drug's molecular features. The system classifies drugs into four albumin binding sites (Sudlow's Site I, Site II, Subdomain IB, and Surface) and applies physicochemical plausibility filtering.
Validated on 616 drugs with experimentally measured albumin binding percentages, the classifier achieves:
Positive class: ≥50% albumin bound. The 616-drug dataset includes 440 binders and 176 non-binders.
Performance Metrics
| Metric | Value | 95% CI (Wilson) |
|---|---|---|
| Sensitivity (Recall) | 95.9% | 93.7 - 97.4% |
| Specificity | 85.8% | 79.8 - 90.2% |
| Accuracy | 93.0% | 90.7 - 94.8% |
| PPV (Precision) | 94.4% | 91.9 - 96.2% |
| NPV | 89.3% | 83.8 - 93.2% |
| Youden's J | 0.817 | |
| F1 Score | 0.952 |
Confusion Matrix
| Predicted Binder | Predicted Non-Binder | Total | |
|---|---|---|---|
| Actual Binder | 422 (TP) | 18 (FN) | 440 |
| Actual Non-Binder | 25 (FP) | 151 (TN) | 176 |
| Total | 447 | 169 | 616 |
Methods
Drug Set
616 drugs were selected based on the following inclusion criteria:
- Experimentally measured albumin binding percentage available
- Valid SMILES structure available
- Binding threshold: ≥50% albumin bound = positive (binder)
Binders (≥50%): 440 drugs (71.4%). Non-binders (<50%): 176 drugs (28.6%).
Classification Architecture
The classifier operates as a four-site cascade:
- Site I (Sudlow's Site I / Warfarin site): Large, flexible pocket in subdomain IIA.
- Site II (Sudlow's Site II / Diazepam site): Smaller, rigid, stereoselective pocket in subdomain IIIA.
- Subdomain IB: Fatty acid binding pocket.
- Surface (Electrostatic): Charged surface binding.
Input
Drug structures are processed directly from SMILES notation. No external chemistry libraries are used, no ML, the entire platform is chemistry based.
Misclassification Analysis
Review the individual false negatives and false positives in the expandable sections below.
False negatives · 18 drugs
Drugs with ≥50% albumin binding that the classifier predicts as non-binders:
False positives · 25 drugs
Drugs with <50% albumin binding that the classifier predicts as binders:
Observations: 17 of 25 FP drugs have binding between 35-47% — near the 50% classification threshold. The mean binding percentage of FP drugs is 35.5%, indicating these are moderate binders rather than true non-binders. The classifier errs toward over-prediction of binding, which is the safer direction for drug safety screening (failing to identify a binder has greater clinical consequence than over-calling a non-binder).
Context: Literature Comparison
Albumin binding prediction remains an active area of computational pharmacology. Published methods include:
Published method comparison
| Method | Dataset Size | Metric | Performance | Reference |
|---|---|---|---|---|
| Spirographic AI (this work) | 616 | Accuracy | 93.0% | This report |
| Random Forest (QSAR) | 1300 | Accuracy | 88% | Zhu et al., J Chem Inf Model, 2013 |
| SVM + molecular descriptors | 816 | Accuracy | 85-89% | Votano et al., J Med Chem, 2006 |
| Deep learning (GNN) | 1400 | AUC | 0.88 | Li et al., Brief Bioinform, 2022 |
| Consensus QSAR | 967 | Accuracy | 87% | Ghafourian & Amin, J Pharm Sci, 2013 |
| Pharmacokinetic PBPK | Varies | R-squared | 0.70-0.85 | Jones et al., CPT, 2015 |
Spirographic AI accuracy (93.0%) is shown alongside published accuracy figures. AUC and R-squared are different metrics and should not be compared numerically with accuracy. Cross-study results are not direct head-to-head comparisons because datasets, class distributions, endpoints, and validation methods differ. Spirographic AI Youden’s J = 0.817 is reported separately in the Performance Metrics section.
Key Differentiators
- This classifier uses no training data or machine learning — it operates from first-principles structural analysis
- It provides binding site classification (Site I, II, IB, Surface) in addition to Y/N prediction
- The 95.9% sensitivity ensures minimal missed binders, critical for drug interaction screening
- Validation set of 616 drugs spans all major therapeutic classes
Limitations
- Binary threshold: The 50% cutoff is clinically standard but loses nuance for drugs near the boundary (35-65% range). Binding tier prediction (MILD/MODERATE/STRONG) partially addresses this.
- Static binding: The classifier predicts equilibrium binding and does not account for competitive displacement, saturation kinetics, or disease-state alterations (hypoalbuminemia, uremia).
- Single-compound SMILES: Multi-component formulations (salts, co-crystals) use the largest SMILES fragment.
- Experimental reference variability: Published albumin binding percentages vary by method (equilibrium dialysis, ultrafiltration, chromatography) and conditions (pH, temperature, drug concentration).
Conclusion
The Spirographic AI albumin binding classifier achieves J=0.817 on 616 drugs with 95.9% sensitivity and 85.8% specificity, using first-principles structural analysis without training data. The system correctly identifies albumin binders across all major drug classes while providing binding site localization. Performance is competitive with published machine learning approaches while offering mechanistic interpretability through site-specific classification.