[Paper Review] ChemoVerse: Manifold traversal of latent spaces for novel molecule discovery

Harshdeep Singh, Nicholas McCarthy+2 more|arXiv (Cornell University)|Sep 29, 2020

Machine Learning in Materials Science6 references4 citations

TL;DR

This paper proposes ChemoVerse, a model-agnostic framework for directed exploration of latent spaces in molecular generative models using Riemannian manifold interpolation guided by domain-specific heuristics. By combining grammar variational autoencoders with k-d trees and Yen’s algorithm for pathfinding, it enables efficient, valid molecule generation with improved drug-likeness and target-specific properties, outperforming linear interpolation in both novelty and structural validity.

ABSTRACT

In order to design a more potent and effective chemical entity, it is essential to identify molecular structures with the desired chemical properties. Recent advances in generative models using neural networks and machine learning are being widely used by many emerging startups and researchers in this domain to design virtual libraries of drug-like compounds. Although these models can help a scientist to produce novel molecular structures rapidly, the challenge still exists in the intelligent exploration of the latent spaces of generative models, thereby reducing the randomness in the generative procedure. In this work we present a manifold traversal with heuristic search to explore the latent chemical space. Different heuristics and scores such as the Tanimoto coefficient, synthetic accessibility, binding activity, and QED drug-likeness can be incorporated to increase the validity and proximity for desired molecular properties of the generated molecules. For evaluating the manifold traversal exploration, we produce the latent chemical space using various generative models such as grammar variational autoencoders (with and without attention) as they deal with the randomized generation and validity of compounds. With this novel traversal method, we are able to find more unseen compounds and more specific regions to mine in the latent space. Finally, these components are brought together in a simple platform allowing users to perform search, visualization and selection of novel generated compounds.

Motivation & Objective

• To address the challenge of random and inefficient exploration in high-dimensional, sparse latent spaces of generative models for molecular design.
• To enable intelligent, directed traversal of latent chemical spaces to discover novel, valid molecules with specific desired properties.
• To integrate domain-specific heuristics—such as fingerprint similarity, synthetic accessibility, drug-likeness, and binding activity—into the exploration process.
• To develop a scalable, model-agnostic platform for virtual screening that supports search, visualization, and selection of generated compounds.

Proposed method

• Employed grammar variational autoencoders (GVAEs), including an attention-augmented variant, to generate high-quality, valid latent representations of SMILES-encoded molecules.
• Calculated Jacobian distances between latent points to model local geometric distortions and preserve structural similarity in the latent space.
• Constructed a k-dimensional tree (k-d tree) using Jacobian distances as edge weights to enable efficient, logarithmic-time search in the latent manifold.
• Integrated multiple domain heuristics—Tanimoto fingerprint similarity, synthetic accessibility (SA) score, QED drug-likeness, and binding activity—into edge weights for guided pathfinding.
• Applied Yen’s algorithm with A* search to compute the shortest path between source and destination points in the k-d tree, prioritizing paths with desired molecular properties.
• Performed equidistant interpolation along the computed paths and decoded latent vectors using the generator to produce novel, valid molecules.

Experimental results

Research questions

• RQ1 Can Riemannian manifold interpolation with heuristic-guided pathfinding improve the discovery of novel, valid molecules in sparse latent spaces compared to random or linear sampling?
• RQ2 How effective is the integration of multiple molecular property heuristics (e.g., Tanimoto similarity, SA score, QED, binding activity) in guiding directed traversal toward desired chemical regions?
• RQ3 To what extent does manifold traversal outperform linear interpolation in terms of structural validity and diversity of generated compounds?
• RQ4 Can the framework identify and explore previously uncharted regions of latent chemical space with high potential for novel drug-like compounds?

Key findings

• Manifold traversal with fingerprint similarity and Yen’s algorithm generated 156 valid, novel compounds between latent regions of diabetes and lung cancer drugs, compared to only 3 valid compounds from linear interpolation with 100 points.
• The attention-augmented grammar VAE achieved a low rate of invalid molecule generation, demonstrating improved structural validity in latent space sampling.
• The k-d tree with Jacobian-based distances enabled efficient logarithmic-time search, significantly improving runtime performance over brute-force methods.
• Heuristic-weighted pathfinding allowed for targeted exploration of latent space regions rich in molecules with desired properties such as high drug-likeness and binding activity.
• Perturbing source and destination points enabled discovery of multiple distinct paths, increasing the diversity of novel compound generation.
• The platform successfully visualized and bounded regions of interest in the latent space, enabling focused mining for compounds with specific structural and functional characteristics.