How to build and validate 3D-QSAR models - Insights from Xu et al., 2020

Dataset Preparation

A high-quality dataset is the cornerstone of any QSAR study.

1. Collect experimental bioactivity data — typically IC₅₀ values from reliable sources.
2. Normalize the activity scale — convert IC₅₀ to pIC₅₀:
   $$ \text{pIC}_{50} = -\log_{10}(\text{IC}_{50} \, \text{in molar}) $$
3. Data partitioning:
   • Training set: ~75–80% of compounds for model development.
   • Test set: ~20–25% of compounds for external validation.

A balanced chemical space between sets is critical to avoid extrapolation during prediction.

Molecular Docking & Alignment

Accurate alignment of molecules is pivotal in 3D-QSAR since descriptor calculation is spatially dependent.

• Protein structure selection: Obtain the target’s crystallographic structure from the Protein Data Bank (PDB).
• Docking:
  • Use software such as SYBYL, AutoDock, or other docking engines.
  • Perform redocking to verify reliability — RMSD ≤ 2.0 Å is generally considered acceptable.
• Alignment strategy: Select the best-scoring pose as the reference for aligning all ligands.

Descriptor Calculation

Two classical methods dominate 3D-QSAR descriptor generation:

Comparative Molecular Field Analysis (CoMFA)

• Fields: Steric + Electrostatic
• Grid spacing: 2.0 Å
• Probe: sp³ carbon atom with +1 charge
• Energy cutoff: 30 kcal/mol

Comparative Molecular Similarity Indices Analysis (CoMSIA)

• Fields: Steric, Electrostatic, Hydrophobic, H-bond Donor, H-bond Acceptor
• Attenuation factor: 0.3 — controls the exponential distance dependence.

Model Building with Partial Least Squares (PLS)

PLS regression is the workhorse for 3D-QSAR, capable of handling collinear and noisy descriptors.

1. Internal validation:
   • Perform Leave-One-Out Cross-Validation (LOO-CV).
   • Record:
     • q²: cross-validated \( R^2 \), internal predictive power.
     • ONC: Optimal Number of Components.
2. Final model construction (using ONC):
   • r²: Goodness-of-fit to training set.
   • SEE: Standard Error of Estimate.
   • F-statistic: Statistical significance of the regression.

Model Validation

Validation ensures the model is predictive, robust, and not the result of chance correlations.

External Validation

• Predict pIC₅₀ for the test set.
• Calculate r²ₚᵣₑd for predictive performance.

Tropsha’s Criteria

A set of diagnostic metrics assessing external predictivity:

• \( r^2_0 \), \( k \), \( k' \), \( r_m^2 \), \( \Delta r_m^2 \)
• Good predictive models: \( k \approx 1 \), \( r_m^2 > 0.5 \)

Y-Randomization

• Randomize biological activities and rebuild the model.
• A valid QSAR will show low q² and r² for randomized trials.

Contour Map Interpretation

CoMFA and CoMSIA produce contour maps that visually indicate where modifications may enhance or reduce activity.

• Steric maps: Green = bulk-favorable; Yellow = bulk-unfavorable.
• Electrostatic maps: Blue = electropositive-favorable; Red = electronegative-favorable.
• Hydrophobic & H-bond maps: Guide lipophilicity and hydrogen bonding optimization.

These maps serve as structure–activity roadmaps for rational ligand design.

Summary of Key Metrics

Final Note:
A robust 3D-QSAR workflow is not solely about generating good statistical values — it is about building interpretable models that reliably guide the design of novel, potent ligands.

References

• Xu, Y., He, Z., Liu, H., Chen, Y., Gao, Y., Zhang, S., … & Yang, M. (2020). 3D-QSAR, molecular docking, and molecular dynamics simulation study of thieno [3, 2-b] pyrrole-5-carboxamide derivatives as LSD1 inhibitors. RSC advances, 10(12), 6927-6943.