Connecting proteins with drug-like compounds: Open source drug discovery workflows with BindingDB and KNIME

George Nicola, Michael R. Berthold, Michael P. Hedrick, Michael K. GilsonDatabase 2015 · DOI

TL;DR

This paper presents a set of open-source workflows built with KNIME that integrate the BindingDB database to predict potential drug targets for compounds. By leveraging chemical similarity, these workflows can generate hypotheses about drug mechanisms and potential side effects, offering a powerful, accessible tool for drug discovery research.

What Problem Does This Paper Address?

The explosion of publicly available protein-small molecule interaction data, exemplified by the BindingDB database, presents a massive opportunity for drug discovery. However, effectively mining and integrating this data, especially with broader biomolecular analyses, often requires custom scripting and significant bioinformatics expertise. This creates a barrier for many researchers. The paper addresses this by demonstrating how user-friendly, graphical workflow tools like KNIME can be used to harness the power of BindingDB, making complex data analysis and hypothesis generation more accessible.

Methods

The core of this work lies in developing KNIME workflows that interact with BindingDB. Two primary approaches are presented:

  • Direct Web Service Access: Workflows that query BindingDB's RESTful web services in real-time. This ensures up-to-date data but can be slower due to network latency.
  • Local Data Distillation: Workflows that utilize a local, processed version of the BindingDB dataset. This significantly speeds up analysis and enhances privacy but requires initial data preparation.

A key focus is the implementation of "Find My Compound's Targets" (FMCT) workflows. These workflows are based on the principle of chemical similarity: if compound A is similar to compound B, and B binds to a certain protein, then A is likely to bind to that same protein. The paper explores various chemical similarity metrics, including:

  • RDKit Fingerprint Similarity: A fast method using chemical fingerprints.
  • Average of Multiple Fingerprint Metrics: Combining several fingerprint-based similarities.
  • Maximum Common Substructure (MCS) Similarity: A more computationally intensive method that identifies shared structural fragments.
  • Hybrid MCS: Using MCS to re-rank hits from faster fingerprint methods.

The paper also details methods for building and optimizing local BindingDB datasets for faster FMCT analysis, including clustering compounds to create a reduced reference set.

Datasets

The primary dataset used is BindingDB, a large, public database of protein-ligand binding data. The paper utilizes both the live BindingDB web services and a locally processed version of the database. For evaluation, the authors use data from seven existing drugs with previously identified therapeutic targets, as reported in a prior study using the Similarity Ensemble Approach (SEA) (Gregori-Puigjané et al., 2012).

Key Insights and Ideas

  • Democratizing Drug Discovery Data: KNIME workflows can bridge the gap between raw data (like BindingDB) and actionable insights, making advanced analysis accessible to a wider range of researchers.
  • Chemical Similarity as a Predictive Engine: The core idea that chemical similarity can predict protein targets is powerful and, when implemented with robust workflows, can uncover novel hypotheses.
  • Complementary Nature of Similarity Metrics: Different similarity metrics (fingerprint-based vs. MCS) offer distinct advantages. Fingerprint methods are faster and often align well with web-based tools like SEA, while MCS can reveal less obvious structural similarities that might be missed by other methods, albeit at a higher computational cost.
  • The Value of Local Data Processing: For large-scale analysis or when privacy is a concern, processing BindingDB data locally within KNIME offers significant benefits in speed and control.
  • Hypothesizing Side Effects: The FMCT approach isn't just for finding primary targets; it can also flag potential off-target interactions, which are crucial for understanding drug side effects and repurposing opportunities.

Key Results

The study successfully implemented and tested several KNIME workflows for interacting with BindingDB and performing target prediction.

  • Successful Replication and Novelty: When applied to seven drugs with ill-defined targets, the FMCT workflows replicated some previously confirmed results from SEA and also identified new, plausible high-quality hits.
  • Performance Comparison: The fast RDKit fingerprint method performed comparably to more complex methods, while MCS occasionally provided unique and interesting predictions.
  • Speed vs. Accuracy Trade-off: Using a reduced, clustered dataset for FMCT significantly accelerated analysis but led to a substantial decrease in the recovery of validated targets.
  • Workflow Availability: The described workflows are made available for download, enabling other researchers to use and adapt them.

Limitations and Future Work

  • Computational Cost of MCS: The pure MCS similarity metric is computationally expensive, limiting its use for large-scale screening. The paper explores hybrid approaches to mitigate this.
  • Reduced Dataset Performance: While clustering speeds up analysis, it sacrifices accuracy and the ability to discover less obvious hits. Future work could explore optimizing clustering parameters or using different clustering methods.
  • Anecdotal Nature of Validation: The validation of FMCT predictions against SEA results is described as "anecdotal rather than statistical," highlighting the need for further experimental validation.
  • Future Integration: The authors envision broader applications, including integrating FMCT with other chemical similarity metrics (e.g., shape-based, pharmacophore models), natural language processing for literature mining, and machine learning techniques. They also suggest that KNIME's extensibility can accommodate new nodes for advanced analyses like ligand-protein docking.

Summary updated

Abstract

Today's large, public databases of protein-small molecule interaction data are creating important new opportunities for data mining and integration. At the same time, new graphical user interface-based workflow tools offer facile alternatives to custom scripting for informatics and data analysis. Here, we illustrate how the large protein-ligand database BindingDB may be incorporated into KNIME workflows as a step toward the integration of pharmacological data with broader biomolecular analyses. Thus, we describe a collection of KNIME workflows that access BindingDB data via RESTful webservices and, for more intensive queries, via a local distillation of the full BindingDB dataset. We focus in particular on the KNIME implementation of knowledge-based tools to generate informed hypotheses regarding protein targets of bioactive compounds, based on notions of chemical similarity. A number of variants of this basic approach are tested for seven existing drugs with relatively ill-defined therapeutic targets, leading to replication of some previously confirmed results and discovery of new, high-quality hits. Implications for future development are discussed. Database URL: www.bindingdb.org.
BibTeX
@inproceedings{nicola2015connecting,
  author = {George Nicola and Michael R. Berthold and Michael P. Hedrick and Michael K. Gilson},
  title = {{Connecting proteins with drug-like compounds: Open source drug discovery workflows with BindingDB and KNIME}},
  booktitle = {Database},
  year = {2015},
  doi = {10.1093/database/bav087},
}
Figure