Flexible Analog Search with Kernel PCA Embedded Molecule Vectors

Stefano Rensi, Russ B. AltmanComputational 2017 · DOI

TL;DR

This paper introduces a novel method for "flexible analog search" using kernel PCA to embed molecules into a continuous vector space. This approach allows for approximate matched molecular pair (MMP) searches with tunable levels of abstraction, overcoming limitations of existing methods that are either too rigid or too context-dependent.

What Problem Does This Paper Address?

Drug discovery relies heavily on optimizing lead compounds by making small structural changes to improve activity and ADMET properties. Matched Molecular Pairs (MMPs) are a key concept, representing pairs of molecules differing by a single, well-defined transformation. However, existing MMP search algorithms struggle with "abstraction" – they either miss important, slightly inexact MMPs or include irrelevant ones due to overly strict or overly loose definitions of context. This paper aims to bridge this gap by enabling flexible control over how much contextual information is considered during the search.

Methods

The core of the method involves representing molecules as vectors and then embedding these vectors into a continuous space using Kernel Principal Component Analysis (KPCA) with a Tanimoto similarity kernel. This KPCA embedding creates continuous vectors where dot products approximate Tanimoto similarities, allowing for non-linear dimensionality reduction.

The search itself is framed as a supervised query. A key innovation is the "analog score," which measures the similarity between two relationships (e.g., a:b and c:d). This score allows for finding molecules d that complete an analogy a:b::c:d.

Crucially, the method allows for "uncoupled feature sets." This means the "relation" (the transformation between two molecules, b-a) and the "target" (the molecule being queried, c) can be represented using different numbers of principal components (dimensions). This provides fine-grained control over the level of abstraction: a low-dimensional relation vector captures a more abstract transformation, while a high-dimensional target vector captures more specific molecular details.

Datasets

The paper uses two primary datasets for illustration and benchmarking:

  • Small molecule FDA-approved drugs: From Drugbank [22].
  • Biologically important metabolites: From KEGGBRITE ontology [23].

These datasets were filtered for molecular weight (100-700 Da) and non-human native classes.

For benchmarking approximate context-independent MMP search, a large dataset of 72,251 compounds from ChEMBL was used, grouped into "Easy," "Intermediate," and "Difficult" superclasses based on size and diversity.

Key Insights and Ideas

  • Abstraction is the bottleneck: The paper clearly articulates that the rigidity of existing MMP search methods in handling abstraction is a major hurdle in finding relevant analogs.
  • KPCA for continuous representation: Embedding molecules into a continuous space via KPCA provides a powerful way to represent molecular relationships and allows for flexible manipulation of these representations.
  • Analog score for relationship similarity: The proposed analog score offers a novel way to quantify the similarity between chemical transformations, going beyond simple pairwise molecule similarity.
  • Uncoupled feature sets unlock flexibility: The ability to independently control the dimensionality of the "relation" and "target" vectors is the "secret sauce" that allows for tunable abstraction, enabling searches that are neither fully context-dependent nor fully context-independent.
  • Feature selection is paramount: The performance of the method is highly sensitive to how feature sets are chosen, highlighting the importance of application-specific tuning.

Key Results

  • Nucleobase/nucleoside analogy: The analog score successfully ranked correct purine/pyrimidine and hydrogen bond donor/acceptor correspondences for nucleobases.
  • Flexible search performance: The uncoupled feature selection strategy (low-dimensional relation, high-dimensional target) significantly improved search performance for queries across purine/pyrimidine contexts, while maintaining good performance for hydrogen bond donor/acceptor contexts.
  • Approximation of context-independent search: The proposed method, with an adaptive feature selection heuristic, was able to accurately approximate context-independent MMP search results obtained from fragment indexing, demonstrating robustness across different dataset sizes and diversity levels.
  • Fingerprint impact: The choice of 2D fingerprint representation (ECFP, APFP, RDFP, TTFP) had a measurable impact on search accuracy, with some fingerprints performing better than others.

Limitations and Future Work

  • Feature selection sensitivity: The authors acknowledge that the method is sensitive to feature selection, and finding optimal parameters is challenging and application-dependent. They suggest that the "secret sauce" of uncoupled feature sets is key, but the strategy for choosing these sets needs further exploration beyond grid search.
  • Non-smoothness of rankings: The non-smoothness of search rankings in some figures suggests that contextual information might not be neatly organized in consecutive principal components, implying that simple grid searches might not be universally sufficient.
  • Future directions: The paper suggests that kernel embedding can be applied to any symbolic representation supporting similarity computation, opening possibilities for searching relationships beyond structural aspects, such as binding affinity profiles or other non-structural chemical properties. Further work could explore alternative embedding techniques (e.g., neural network-based graph convolutions) and more sophisticated feature selection strategies.

Summary updated

Abstract

Studying analog series to find structural transformations that enhance the activity and ADME properties of lead compounds is an important part of drug development. Matched molecular pair (MMP) search is a powerful tool for analog analysis that imitates researchers' ability to select pairs of compounds that differ only by small well-defined transformations. Abstraction is a challenge for existing MMP search algorithms, which can result in the omission of relevant, inexact MMPs, and inclusion of irrelevant, contextually dissimilar MMPs. In this work, we present a new method for MMP search that returns approximate results and enables flexible control over abstraction of contextual information. We illustrate the concepts and mechanics of our method with a series of exemplar MMP queries, and then benchmark search accuracy using MMPs found by fragment indexing. We show that we can search for MMPs in a context dependent manner, and accurately approximate context independent fragment index based MMP search over a range of fingerprint and dataset conditions. Our method can be used to search for pairwise correspondences among analog sets and bolster MMP datasets where data is missing or incomplete.
BibTeX
@inproceedings{rensi2017flexible,
  author = {Stefano Rensi and Russ B. Altman},
  title = {{Flexible Analog Search with Kernel PCA Embedded Molecule Vectors}},
  booktitle = {Computational},
  year = {2017},
  doi = {10.1016/j.csbj.2017.03.003},
}
Figure