MolDAIS: Molecular Descriptors with Actively Identified Subsets

MolDAIS is a powerful tool for efficient molecular property optimization through adaptive learning of sparse subspaces.

Overview

MolDAIS provides a framework for accelerating black-box molecular property optimization by intelligently identifying relevant molecular subspaces. This approach significantly reduces the computational cost of exploring chemical space while maintaining optimization effectiveness.

Installation

pip install MolDAIS

Quick Start Example

More examples are presented in the jupyter notebook

import torch
from MolDAIS.moldais import MolDAIS

# Step 1: Define your molecular search space
smiles_list = [
    "CCO", "CCC", "CCN", "CCCl", "CCCC", "CCCO", "CCCN", "ClCCCl",
    "CCOCC", "CCON", "CCCNCl", "CCCCCl", "CCCCN", "CCCCCO", "CCNCl",
    "CCNO", "ClCCl", "CCOO", "CCNN", "COCO"
]

# Step 2: Define target property values (e.g., logP)
logP_values = torch.tensor([0.5, 1.2, 0.8, 2.1, 3.0, 1.5, 2.3, 2.8, 0.9, 1.4,
                            2.5, 1.9, 1.1, 2.7, 2.6, 1.6, 1.8, 2.2, 1.3, 1.7], 
                            dtype=torch.float32)

# Step 3: Define the optimization problem
problem = MolDAIS.Problem(smiles_search_space=smiles_list,
                          targets=logP_values.unsqueeze(1),
                          experiment_name="LogP_Optimization_Test")
problem.compute_descriptors()

# Step 4: Configure optimizer parameters
optimizer_parameters = MolDAIS.OptimizerParameters(
    sparsity_method='MI',           # Feature selection method (Mutual Information)
    acq_fun='EI',                   # Acquisition function (Expected Improvement)
    num_sparsity_feats=10,          # Number of features to select
    multi_objective=False,          # Single-objective optimization
    constrained=False,              # No constraints
    total_sample_budget=7,          # Total evaluation budget
    initialization_budget=2,        # Initial points for model building
    seed=123                        # Random seed for reproducibility
)

# Step 5: Create and run the optimization
mol_dais = MolDAIS(problem=problem, optimizer_parameters=optimizer_parameters)
mol_dais.configuration.optimize()

# Step 6: Get results
print("Best molecules found:")
print(mol_dais.results.best_molecules)
print("Best property values:")
print(mol_dais.results.best_values)

# Step 7: Visualize optimization progress
mol_dais.configuration.plot_convergence()

Features

• Adaptive Feature Selection: Intelligently identifies relevant molecular descriptors
• Flexible Acquisition Functions: Multiple strategies for exploring chemical space
• Support for Constraints: Handle constrained optimization problems
• Multi-objective Optimization: Optimize multiple molecular properties simultaneously

Planned Features

• Custom Callable Functions: Upcoming support for custom python wrapped functions
• Human-in-the-loop: Upcoming support for interactive optimization

Usage Options

With a SMILES List and Callable Function

# Coming soon - Human-in-the-loop optimization capabilities

Citation

Submitted to RSC Digital Discovery. Formal citation coming soon...

Based on previous work:

@misc{sorourifar2024accelerating,
      title={Accelerating Black-Box Molecular Property Optimization by Adaptively Learning Sparse Subspaces}, 
      author={Farshud Sorourifar and Thomas Banker and Joel A. Paulson},
      year={2024},
      eprint={2401.01398},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM}
}

Contact

For questions, support, or collaboration opportunities:
Farshud Sorourifar - sorourifar.1@osu.edu

License

This project is licensed under the MIT License - see the LICENSE file for details.