We are excited to announce the latest Cresset KNIME nodes release (3.0.0) which introduces the new ‘Spark in Flare’ functionality within Flare for the following nodes:

• Spark Database Search
• Spark Fragment Selector
• Generate Spark Database

This release of the Cresset KNIME nodes ensures full compatibility with the latest KNIME release (Version 5.4) and are fully integrated into Cresset’s drug design solutions. You have the option to run nodes within stand-alone Spark, or using Spark within Flare, where you can seamlessly integrate Spark experiments into your workflows with just a click of a button. This allows you to make informed decisions about molecular modifications and improve your workflows for drug discovery and SAR analysis without needing a separate Spark installation.

In this article, we showcase the use of ‘Spark Fragment Selector’ and ‘Spark Database Search’ with a practical example, demonstrating how they can provide impactful insights in structure-activity relationship studies.

Figure 1. Overview of the workflow presented in this article. Execute complex workflows including many steps with the help of the new Cresset KNIME nodes

To illustrate the utility of these new nodes, a set of inhibitors of the aldose reductase (ALR2) enzyme, will be used.^1-4 For this dataset, comprehensive structure-activity relationship (SAR) data is available, as these inhibitors have been extensively studied due to their potential as effective treatments for diabetic complications.^1-4 The goal is to identify regions within the molecules where structural modifications could potentially improve biological activity. Activity Atlas™, Field QSAR and Spark will be used to gain insights into the molecular features that correlate with potency, explore ways to enhance activity, and perform intelligent molecular modifications.

To begin with, an Activity Atlas model is built using the training dataset, which provides a visual overview of the relationship between the 3D electrostatic, shape and hydrophobicity of the aligned ligands and biological activity. This model helps highlight key areas where modifications can lead to improved potency. One significant observation from the model is the presence of a green region at the left-hand side of the molecular structures (Figure 2A), indicating a favorable shape feature in that area: this suggests that introducing bulkier R-groups in this region could potentially lead to increased activity.

Spark can assist in modifying the R-group, but it requires selecting a single molecule from the training dataset. The molecule we decided to use stands out due to its significant activity (8.3 pKi) that makes it a good candidate for further development. This highly active molecule will also be used as a reference point for assessing the impact of subsequent modifications. The fluorine atom highlighted in Figure 2B also extends into the area we want to explore, making it a good attachment point for the Spark R-group exploration.

Figure 2. A) The Activity Atlas result: a smooth green surface as the one highlighted here means that bulkier substitutions may lead to increased biological activity, B) The Spark Fragment Selector window allows you to select the part of the molecule to be replaced

To explore potential bioisosteric substitutions, the ‘Spark Fragment Selector’ and ‘Spark Database Search’ KNIME nodes were used to run a Spark R-group replacement experiment. The former node offers a Spark-like interface to select the R-group to be replaced, whereas the latter enables control over experiment settings, such as the choice of databases to search, scoring method, and search constraints. In this specific experiment, the ChEMBL Common database was used, and all the other options were kept as defaults. At the end of the experiment, a diverse set of molecular structures was generated. Some of these molecules could potentially exhibit higher potency than the reference.

Once the new set of molecules is generated, the next step is to assess their predicted activity. To achieve this, a Field QSAR model was developed using the initial training set, with the Spark molecules as the prediction set. This model establishes a correlation between the three-dimensional electrostatic and steric fields (calculated with the Cresset XED force field⁵) of the training set molecules and their biological activity. Predictive 3D-QSAR models allow for the prediction of new compounds’ activity based on their molecular electrostatic and steric features.

For Field QSAR models in Flare, the ‘Distance to model’ metric provides an assessment of whether the molecules of your prediction set lie ‘Inside’, ‘Close’ or ‘Outside’ of the descriptors space of the molecules in the training set, i.e. whether they have features which are not present in the training set. It provides a clear assessment of how well a predicted molecule aligns with the training data, improving confidence in QSAR predictions. Using this metric, the molecules that lie ‘Outside’ the Domain of Applicability of the model are filtered out with the help of the ‘Row filter’ node, so that only the molecules with more reliable predictions are taken into account. This reduces our dataset to 361 molecules from the 500 initial new molecular designs.

Figure 3. The ‘Dist To Model’ metric informs us whether the new molecules lie ‘Inside’, ‘Close, or ’Outside’ of the descriptors space of the molecules in the training set, i.e. whether they have features which are not present in the training set compounds

Molecules outside this Domain of Applicability were filtered out as their predicted activities may not be reliable.

An Activity Atlas model was then rebuilt, including the 361 molecules from the Spark experiment with their predicted activity, to further assess the impact of the modifications. This updated model provides an enhanced understanding of how the structural changes may affect biological activity. One of the most interesting observations from the final model is that bulkier substituents indeed seem to lead to an increased predicted activity. The smooth green surface that is still present in the same position indicates that molecules with bulkier substituents may demonstrate larger predicted activity.

Indeed, sorting the molecules in descending order of predicted activity reveals that the top molecules, exhibiting predicted activity of values ~9 pKi, all have a bulky R-group in this region, adding weight to the hypothesis that bulkier modifications may enhance potency.

Figure 4. Results of the Activity Atlas experiment on the Spark-generated molecules. The predicted activity indicates that indeed bulkier R-groups substitutions may result into active molecules

The newly released KNIME nodes provide a seamless and powerful way to conduct bio-isosteric replacements using Spark directly from Flare. This update enables researchers to rapidly explore structural modifications using a data-driven approach, predict the impact of modifications through robust QSAR modeling, and validate hypotheses using Activity Atlas insights all within a single workflow.

This example demonstrates how these tools can be used to efficiently optimize lead molecules and drive structure-based drug discovery. With these new capabilities, Cresset KNIME nodes continue to empower scientists with cutting-edge computational chemistry workflows, paving the way for further advancements in molecular modeling and drug discovery.

To try the new KNIME nodes please navigate to our website for instructions on how to download and install or update the Cresset KNIME nodes, or contact Cresset support.

References

1. Mylari, B. L.; Armento, S. J.; Beebe, D. A.; Conn, E. L.; Coutcher, J. B.; Dina, M. S.; O’Gorman, M. T.; Linhares, M. C.; Martin, W. H.; Oates, P. J.; Tess, D. A.; Withbroe, G. J.; Zembrowski, W. J. A Novel Series of Non-Carboxylic Acid, Non-Hydantoin Inhibitors of Aldose Reductase with Potent Oral Activity in Diabetic Rat Models: 6-(5-Chloro-3-Methylbenzofuran-2-Sulfonyl)-2 H -Pyridazin3-One and Congeners. J. Med. Chem. 2005, 48 (20), 6326–6339
2. Mylari, B. L.; Beyer, T. A.; Scott, P. J.; Aldinger, C. E.; Dee, M. F.; Siegel, T. W.; Zembrowski, W. J. Potent, Orally Active Aldose Reductase Inhibitors Related to Zopolrestat: Surrogates for Benzothiazole Side Chain. J. Med. Chem. 1992, 35 (3), 457–465
3. Mylari, B. L.; Larson, E. R.; Beyer, T. A.; Zembrowski, W. J.; Aldinger, C. E.; Dee, M. F.; Siegel, T. W.; Singleton, D. H. Novel, Potent Aldose Reductase Inhibitors: 3,4-Dihydro-4-Oxo-3-[[5(Trifluoromethyl)-2-Benzothiazolyl]Methyl]-1-Phthalazineacetic Acid (Zopolrestat) and Congeners. J. Med. Chem. 1991, 34 (1), 108–122
4. Mylari, B. L.; Armento, S. J.; Beebe, D. A.; Conn, E. L.; Coutcher, J. B.; Dina, M. S.; O’Gorman, M. T.; Linhares, M. C.; Martin, W. H.; Oates, P. J.; Tess, D. A.; Withbroe, G. J.; Zembrowski, W. J. A Highly Selective, Non-Hydantoin, Non-Carboxylic Acid Inhibitor of Aldose Reductase with Potent Oral Activity in Diabetic Rat Models: 6-(5-Chloro-3-Methylbenzofuran- 2-Sulfonyl)-2- H Pyridazin-3-One. J. Med. Chem. 2003, 46 (12), 2283–2286
5. J. G. Vinter, Extended Electron Distributions Applied to the Molecular Mechanics of some Intermolecular Interactions, J. Comput.-Aided Mol. Des. 1994, 8, 653-668