Compositional Deep Probabilistic Models of DNA Encoded Libraries

Benson Chen, Mohammad Sultan, Theofanis KaraletsosJournal of Chemical Information and Modeling 2023 · DOI · OpenReview

TL;DR

This paper introduces DEL-Compose, a novel deep probabilistic model that leverages the compositional nature of DNA-Encoded Libraries (DELs) to analyze complex experimental data. By decomposing molecules into their building blocks (synthons) and modeling their interactions, DEL-Compose offers a more interpretable and robust tool for identifying potent drug candidates compared to existing methods.

What Problem Does This Paper Address?

DNA-Encoded Libraries (DELs) are a powerful technology for discovering small molecules with desired properties, such as binding to a specific protein target. However, the selection experiments that generate DEL data are inherently noisy and complex, making it challenging to extract meaningful signals. Traditional methods often struggle to handle this complexity, potentially masking important insights. This paper addresses the need for a computational tool that can effectively model DEL data, account for experimental noise, and provide interpretable insights into molecular binding properties.

Methods

The core of DEL-Compose is a generative probabilistic model that treats molecules as compositions of synthons. It learns embeddings for individual synthons (mono-synthon), pairs of synthons (di-synthon), and triplets of synthons (tri-synthon) using neural networks. These embeddings are then combined to represent the entire molecule. The model explicitly incorporates experimental factors like pre-selection library read-out and replicate-level noise into its observation model, using a zero-inflated Poisson (ZIP) distribution to handle the count data. This compositional approach avoids the need to enumerate all possible full-molecule structures, which can be computationally prohibitive.

Datasets

The authors evaluated DEL-Compose on publicly available DEL data from two well-studied protein targets: Carbonic Anhydrase IX (CA-IX) and Horseradish Peroxidase (HRP). This dataset consists of tri-synthon libraries with a significant number of unique synthons at each position, generating a large chemical space. The data includes experimental conditions with and without the target protein, as well as multiple replicates.

Key Insights and Ideas

  • Compositionality is Key: The paper strongly advocates for treating molecules as compositions of synthons, rather than as monolithic entities. This compositional view unlocks more efficient learning and better interpretability.
  • Modeling Noise is Crucial: DEL data is notoriously noisy. DEL-Compose's explicit modeling of experimental factors like pre-selection and replicate variability significantly improves its performance and robustness.
  • Interpretability Through Latent Variables: The model's learned latent variables and attention mechanisms provide valuable insights into which synthons and their combinations are most important for binding, offering a chemist-friendly interpretation.
  • Factorized Models Excel in Low-Resource Settings: The factorized nature of DEL-Compose allows it to perform competitively, and often better, than models using full molecule representations, especially when training data is limited.

Key Results

  • DEL-Compose achieved state-of-the-art performance on both CA-IX and HRP datasets, outperforming baselines in terms of negative log-likelihood (NLL) and multi-class precision-recall AUC (PR-AUC).
  • The model demonstrated strong generalization capabilities, particularly on synthon splits, indicating its ability to learn robust molecular representations.
  • DEL-Compose successfully identified and ranked important synthons and pharmacophores for both targets, aligning with known biological activity.
  • The model's ability to distinguish between true binding and off-target effects (e.g., high noise in the matrix condition) was highlighted.
  • In low-resource regimes, DEL-Compose consistently outperformed models using full molecule representations in terms of PR-AUC.

Limitations and Future Work

  • Limited to Single Synthon Substructures: The current analysis is primarily focused on cases where important chemical substructures are localized within a single synthon. For more complex targets where substructures span multiple synthons, additional tools or regularizations might be needed.
  • Explicit Chemical Structure of Di-Synthons Not Considered: The model does not explicitly consider the chemical structure of di-synthons. The authors suggest adding regularizers to predict di-synthon and tri-synthon structures to capture higher-order dependencies.
  • Reliance on Substructure Knowledge for Splitting: The data splitting strategies relied on prior knowledge of pharmacophores. When such knowledge is unavailable, alternative splitting methods like scaffold splits or clustering based on molecular fingerprints would be necessary.
  • Future Work: The authors suggest exploring additional regularizers for higher-order synthon structures and developing strategies for splitting data when substructure knowledge is absent. They also envision using the learned representations for out-of-distribution predictions and molecule generation.

Summary updated

Abstract

DNA-Encoded Library (DEL) has proven to be a powerful tool that utilizes combinatorially constructed small molecules to facilitate highly efficient screening assays. These selection experiments, involving multiple stages of washing, elution, and identification of potent binders via unique DNA barcodes, often generate complex data. This complexity can potentially mask the underlying signals, necessitating the application of computational tools such as machine learning to uncover valuable insights. We introduce a compositional deep probabilistic model of DEL data, DEL-Compose, which decomposes molecular representations into their mono-synthon, di-synthon, and tri-synthon building blocks and capitalizes on the inherent hierarchical structure of these molecules by modeling latent reactions between embedded synthons. Additionally, we investigate methods to improve the observation models for DEL count data such as integrating covariate factors to more effectively account for data noise. Across two popular public benchmark datasets (CA-IX and HRP), our model demonstrates strong performance compared to count baselines, enriches the correct pharmacophores, and offers valuable insights via its intrinsic interpretable structure, thereby providing a robust tool for the analysis of DEL data.
BibTeX
@article{chen2023compositional,
  author = {Benson Chen and Mohammad Sultan and Theofanis Karaletsos},
  title = {{Compositional Deep Probabilistic Models of DNA Encoded Libraries}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2023},
  doi = {10.1021/acs.jcim.3c01699},
}
Figure