[Paper Review] Normalized rotation shape descriptors and lossy compression of molecular shape

Jarek Duda|arXiv (Cornell University)|Sep 30, 2015

Computational Drug Discovery Methods4 references3 citations

TL;DR

This paper introduces a framework for normalized rotation shape descriptors using spherical harmonics (PCA-SH) and specialized models like Bent Deformed Cylinder (BDC) and Bent Cylindrical Harmonics (BCH) to enable lossy compression and accurate shape representation of molecular ligands. By combining principal component analysis with orthogonal, rotation-invariant descriptors, the method achieves high representativeness, decodability, and continuity, enabling reconstruction of molecular shapes and improved virtual screening performance.

ABSTRACT

There is a common need to search of molecular databases for compounds resembling some shape, what suggests having similar biological activity while searching for new drugs. The large size of the databases requires fast methods for such initial screening, for example based on feature vectors constructed to fulfill the requirement that similar molecules should correspond to close vectors. Ultrafast Shape Recognition (USR) is a popular approach of this type. It uses vectors of 12 real number as 3 first moments of distances from 4 emphasized points. These coordinates might contain unnecessary correlations and does not allow to reconstruct the approximated shape. In contrast, spherical harmonic (SH) decomposition uses orthogonal coordinates, suggesting their independence and so lager informational content of the feature vector. There is usually considered rotationally invariant SH descriptors, what means discarding of some essential information. This article discusses framework for descriptors with normalized rotation, for example by using principal component analysis (PCA-SH). As one of the most interesting are ligands which have to slide into a protein, we will introduce descriptors optimized for such flat elongated shapes. Bent deformed cylinder (BDC) describes the molecule as a cylinder which was first bent, then deformed such that its cross-sections became ellipses of evolving shape. Legendre polynomials are used to describe the central axis of such bent cylinder. Additional polynomials are used to define evolution of such elliptic cross-section along the main axis. There will be also discussed bent cylindrical harmonics (BCH), which uses cross-sections described by cylindrical harmonics instead of ellipses. All these normalized rotation descriptors allow to reconstruct (decode) the approximated representation of the shape, hence can be also used for lossy compression purposes.

Motivation & Objective

• To develop a shape descriptor framework that enables lossy compression of molecular conformations while preserving key structural features for virtual screening.
• To address limitations in existing methods like USR by improving representativeness, continuity, and decodability through orthogonal, rotation-normalized descriptors.
• To design specialized descriptors—BDC and BCH—for elongated, flat, or bent ligand shapes common in drug discovery.
• To enable reconstruction of approximated molecular shapes from descriptor coefficients, supporting both similarity search and compression.
• To optimize descriptors using PCA and polynomial fitting for features like electron density and atomic mass along the molecular spine.

Proposed method

• Uses spherical harmonics (SH) with principal component analysis (PCA) to generate rotation-invariant, orthogonal descriptors that maximize information content and minimize correlation.
• Proposes the Bent Deformed Cylinder (BDC) model to represent elongated ligands by bending a cylinder and deforming its elliptical cross-sections along a central spine.
• Employs Legendre polynomials to describe the central spine of the BDC and additional polynomials to model the evolution of elliptical cross-sections.
• Introduces Bent Cylindrical Harmonics (BCH) as an alternative to BDC, using cylindrical harmonics instead of ellipses to describe cross-sectional shape evolution.
• Applies polynomial fitting to properties like electron electronegativity and cumulative atomic mass along the spine to extract additional descriptor coordinates.
• Uses PCA on the full descriptor vector to ensure coefficient independence and optimize for selectiveness and information density.

Experimental results

Research questions

• RQ1 Can normalized rotation descriptors based on spherical harmonics and PCA improve shape representation over existing methods like USR in terms of continuity and decodability?
• RQ2 How well can the BDC and BCH models capture the shape of flat, elongated ligands relevant to protein binding?
• RQ3 To what extent can additional features like electron electronegativity and cumulative mass be encoded into the descriptor for improved biological relevance?
• RQ4 Can the descriptor framework be optimized for specific applications using reference similarity metrics and MSE minimization?
• RQ5 How can conformational mobility be captured by averaging coefficients across conformations and including variance terms?

Key findings

• The PCA-normalized spherical harmonic descriptor achieves better orthogonality and reduced correlation among coefficients compared to standard USR, enhancing information density.
• The BDC model successfully captures the bending and cross-sectional evolution of ligands like epinephrine, outperforming standard SH in shape reconstruction fidelity.
• Descriptor coefficients from polynomial fitting of electron electronegativity and cumulative mass provide quantitative, biologically relevant features for similarity assessment.
• The framework supports lossy compression by allowing reconstruction of approximated molecular shapes from descriptor coefficients, with potential for rate-distortion optimization.
• The method enables clustering and virtual screening in a low-dimensional feature space while preserving essential structural and electronic properties.
• The inclusion of variance in coefficients across conformational ensembles allows modeling of molecular flexibility, crucial for ligand-protein binding.