Pharmacokinetic Property Prediction

Accurately predicting half-life and clearance rate for monomer and multimer peptide / small protein therapeutics from historical data to reduce the need for pre-clinical animal trials. We use multitask Gaussian processes to leverage and factor all high dimension data relationships of interest, and make uncertainty-quantified predictions for novel molecular entities.

Motivation

Pharmacokinetic (PK) properties—particularly half-life and clearance rate—are critical determinants of therapeutic success. However, measuring these properties requires:

• Expensive animal studies
• Long experimental timelines (weeks to months)
• Ethical concerns about animal use
• Resource-intensive analytical methods

Accurate computational prediction can:

• Prioritize candidates before animal studies
• Reduce development costs and timelines
• Minimize animal testing
• Enable rapid iteration in drug design

Challenge

Predicting PK properties for peptide therapeutics is difficult because:

• Complex Structure-Function Relationships: Non-linear relationships between sequence/structure and PK
• Multi-Scale Factors: Properties depend on molecular weight, charge, hydrophobicity, conformational flexibility, and more
• Limited Training Data: Relatively few measured PK values compared to chemical diversity
• Multimer Complexity: Behavior of dimers, trimers, etc. is not simply predictable from monomers

Approach

We employ multitask Gaussian processes (MTGPs) which excel in this setting:

Key Features

1. Multitask Learning: Jointly model half-life and clearance rate, leveraging correlations between properties
2. Uncertainty Quantification: Provide confidence intervals on predictions
3. Data Efficiency: Perform well with limited training data
4. Interpretability: Can identify which molecular features drive PK properties
5. Transfer Learning: Knowledge from monomers informs multimer predictions

Model Architecture

• Input Features: Molecular descriptors, sequence embeddings, structural features, physicochemical properties
• Kernel Design: Custom kernels capturing relevant molecular similarities
• Output: Half-life and clearance predictions with uncertainty estimates
• Validation: Cross-validation on held-out molecules and retrospective analysis

Technical Details

• Gaussian Process Framework: Bayesian non-parametric approach
• Feature Engineering: Combination of physics-based descriptors and learned embeddings
• Multimer Modeling: Explicit representation of quaternary structure
• Hyperparameter Optimization: Bayesian optimization of kernel parameters
• Scalability: Sparse GP approximations for larger datasets

Results

Our models demonstrate:

• High Predictive Accuracy: Strong correlation with experimental values (R² > 0.7)
• Reliable Uncertainty Estimates: Calibrated confidence intervals guide decision-making
• Successful Prospective Predictions: Validated on new molecules not in training set
• Reduced Animal Testing: ~70% reduction in candidates progressed to animal studies

Applications

• Lead Optimization: Rapidly screen and optimize candidates for favorable PK
• Rational Design: Guide modifications to improve half-life or clearance
• Formulation Strategy: Inform dosing regimens and delivery methods
• Portfolio Prioritization: Rank candidates based on predicted PK profiles

Impact

This work enables:

• Faster therapeutic development: Weeks instead of months for PK assessment
• Cost reduction: Fewer expensive animal studies
• Ethical improvement: Reduced animal use in drug development
• Better therapeutics: Design molecules with optimal PK from the start

Future Directions

• Expand to additional PK parameters (bioavailability, volume of distribution)
• Incorporate mechanistic models of metabolism and excretion
• Integration with efficacy and safety predictions for holistic drug design
• Real-time predictions as part of generative design loops