[Paper Review] LIMO: Latent Inceptionism for Targeted Molecule Generation

Peter Eckmann, Kunyang Sun+4 more|PubMed|Jun 17, 2022

Computational Drug Discovery Methods36 references17 citations

TL;DR

LIMO introduces a latent space optimization framework using variational autoencoders and sequential neural property predictors to accelerate de novo molecule generation with high binding affinity. It achieves nanomolar binding affinity predictions (e.g., $K_D = 6 \cdot 10^{-14}$ M) against human estrogen receptor, outperforming state-of-the-art methods by 6–12× in speed while maintaining high diversity and validity.

ABSTRACT

Generation of drug-like molecules with high binding affinity to target proteins remains a difficult and resource-intensive task in drug discovery. Existing approaches primarily employ reinforcement learning, Markov sampling, or deep generative models guided by Gaussian processes, which can be prohibitively slow when generating molecules with high binding affinity calculated by computationally-expensive physics-based methods. We present Latent Inceptionism on Molecules (LIMO), which significantly accelerates molecule generation with an inceptionism-like technique. LIMO employs a variational autoencoder-generated latent space and property prediction by two neural networks in sequence to enable faster gradient-based reverse-optimization of molecular properties. Comprehensive experiments show that LIMO performs competitively on benchmark tasks and markedly outperforms state-of-the-art techniques on the novel task of generating drug-like compounds with high binding affinity, reaching nanomolar range against two protein targets. We corroborate these docking-based results with more accurate molecular dynamics-based calculations of absolute binding free energy and show that one of our generated drug-like compounds has a predicted <i>K</i> _<i>D</i> (a measure of binding affinity) of 6 · 10^-14 M against the human estrogen receptor, well beyond the affinities of typical early-stage drug candidates and most FDA-approved drugs to their respective targets. Code is available at https://github.com/Rose-STL-Lab/LIMO.

Motivation & Objective

• Accelerate the generation of drug-like molecules with high binding affinity to target proteins, a major bottleneck in drug discovery.
• Overcome the computational inefficiency of existing methods that rely on slow, physics-based property evaluations like molecular dynamics or docking.
• Enable fast, gradient-based optimization in a learned latent space to generate diverse, high-affinity compounds efficiently.
• Support substructure-constrained optimization for lead optimization, allowing modification of existing pharmacophores while preserving synthetic feasibility.
• Achieve superior performance on high-affinity molecule generation compared to reinforcement learning and sampling-based baselines.

Proposed method

• Leverages a variational autoencoder (VAE) to map SMILES strings into a continuous, disentangled latent space for molecular representation.
• Employs two sequential neural networks to predict molecular properties, enabling efficient reverse-optimization via gradient-based methods.
• Applies an inceptionism-like reverse optimization technique in the latent space to iteratively refine molecules toward desired properties.
• Uses docking software (AutoDock-GPU) and molecular dynamics-based absolute binding free energy (ABFE) calculations to validate binding affinity predictions.
• Integrates substructure constraints by fixing parts of the molecular scaffold during optimization, enabling targeted lead enhancement.
• Employs a two-stage optimization: first, latent space refinement via gradients; second, decoding to generate valid, synthesizable molecules.

Experimental results

Research questions

• RQ1 Can latent space optimization via gradient-based reverse-optimization significantly accelerate the generation of high-affinity drug-like molecules compared to reinforcement learning or sampling-based methods?
• RQ2 To what extent can LIMO generate molecules with binding affinities in the nanomolar range, particularly when validated with high-accuracy ABFE calculations?
• RQ3 How effective is LIMO in substructure-constrained optimization for lead compound enhancement, preserving key pharmacophores while improving affinity?
• RQ4 How well do docking-predicted affinities correlate with more accurate ABFE results for LIMO-generated compounds?
• RQ5 Can LIMO generate diverse, high-quality candidates at scale, enabling efficient downstream filtering for experimental testing?

Key findings

• LIMO generates molecules with predicted $K_D = 6 \cdot 10^{-14}$ M against the human estrogen receptor, surpassing the affinity of FDA-approved drugs like tamoxifen and raloxifene.
• LIMO achieves 6–8× faster generation than reinforcement learning baselines and 12× faster than sampling-based methods, while maintaining or exceeding performance on benchmark tasks.
• The method produces high-affinity compounds in the nanomolar range, with one compound showing $K_D = 6 \cdot 10^{-5}$ nM (i.e., $6 \cdot 10^{-14}$ M) via ABFE calculations.
• LIMO-generated compounds exhibit favorable drug-likeness metrics and form stable, energetically favorable interactions in protein binding pockets, as shown by docking pose analysis.
• The framework maintains high molecular diversity (as measured in Table 2 and 6), increasing the likelihood of identifying viable candidates after downstream filtering.
• Despite docking predictions not perfectly correlating with ABFE results, LIMO’s ability to rapidly generate hundreds of diverse, high-affinity candidates increases the probability of identifying true binders.