LIMO

LIMO generates drug-like molecules optimized for high binding affinity to target proteins using Latent Inceptionism and latent-space optimization.

Key Features:

• Latent Inceptionism: Implements Latent Inceptionism to navigate molecular latent spaces for targeted molecule generation.
• Variational Autoencoder (VAE): Uses a variational autoencoder to create a latent-space representation of molecules.
• Sequential property-predictor networks: Employs two sequential neural networks that predict molecular properties from latent vectors.
• Gradient-based reverse-optimization: Enables gradient-based reverse-optimization in latent space to adjust generated molecules toward desired properties.
• Binding affinity outcomes: Produces drug-like compounds achieving nanomolar-range binding affinities against two protein targets.
• High-affinity example: Generates a compound with a predicted dissociation constant (K_D) of 6 × 10^-14 M against the human estrogen receptor.
• Molecular dynamics validation: Validates binding affinities using molecular dynamics–based calculations of absolute binding free energy.
• Comparison to traditional approaches: Addresses computational expense associated with reinforcement learning, Markov sampling, and Gaussian process–guided deep generative models that rely on physics-based binding affinity calculations.
• Benchmark performance: Demonstrates competitive results on established benchmark tasks for molecule generation.

Scientific Applications:

• Targeted molecule generation for drug discovery: Generates candidate drug-like molecules optimized for binding to specified protein targets.
• Binding affinity optimization: Optimizes molecular structures in latent space to improve predicted binding affinity metrics, including K_D.
• Validation and benchmarking: Supports validation of predicted affinities via molecular dynamics–based absolute binding free energy calculations and benchmarking on standard tasks.

Methodology:

Constructs a molecular latent space with a variational autoencoder, navigates that space using two sequential neural networks that predict molecular properties, applies gradient-based reverse-optimization in latent space, and validates affinities with molecular dynamics–based absolute binding free energy calculations.