AI in Drug Discovery

The most promising molecules have only minor value if they are difficult to access by AI in drug discovery. Our methods enable access to the largest compound collections with tangible entries: Combinatorial Chemical Spaces.

• 1. The core of the innovation is the AI/ML machinery, which identifies complex patterns from input data. The generated model can then be applied to predict outcomes.
• 2. The method suggests optimized queries that have been fine-tuned for potency or ADME properties. Depending on the workflow and its units, these may be more or less synthetically accessible.
• 3. The search in Chemical Spaces delivers only synthetically accessible compounds with a high degree of similarity to the proposed candidates.

Bigger Is Better!
Chemical Spaces are the largest collections of potential drug candidates, offering a vast selection of compounds that can be screened significantly faster and more efficiently than pre-enumerated libraries given the context of AI in drug discovery. While the concept of Chemical Spaces has existed for more than 20 years, they are now undergoing a true limelight experience.

This transformation is driven by companies specializing in on-demand compound synthesis, making these enormous collections not only theoretically accessible but also practically viable for real-world drug discovery. By leveraging this synergy, researchers can explore billion- to trillion-sized Chemical Spaces with the confidence that promising hits can be synthesized and tested quickly.

Compounds are money. And money should be spend wisely.
The byproduct of the dynamic creation of Chemical Spaces can be the effort-associated mapping of the contained compounds, providing additional factors for informed decision-making.

Furthermore, Chemical Spaces can be trained and improved. Minor tweaks to the used building blocks and reactions increase the aptness of the Space.

The individual atom scores are written out in a CSV format, which can be easily read by the common applications to identify important interaction hotspots within the AI/ML workflow.

HYDE captures the ligand-complex holistically: Important factors such as torsions, inter- and intra-molecular clashes, number of H-bonds, and the estimated affinity of the entire ligand are part of the output.

All modeling approaches require the best starting points to ensure the highest possible precision of predictions. Embedding our tools into your structure-based workflows enhances the performance of your virtual screenings and molecular modeling studies.

Nice-to-know: The --hbond-only option allows that only the calculation and optimization of the H-bond network is performed for every input ligand. No HYDE scoring and pose optimization is executed, e.g., no HYDE scores are available in the output file.

The conformational sampling of the complex can be seamlessly combined with the interpretable feedback from HYDE to enable iterative optimization within ligand design. By adding an additional layer of information (calculated binding affinities, LE/LLE, torsions, clashes), new insights into the binding behavior of the compound (series) can be gained.

Identifying hot spots for high-quality interactions between the (potential) ligand and target structure can serve as the basis for rational drug design, which is implemented through a generative workflow. It is particularly helpful that both aspects, ligand AND target, are analyzed for their contributions to the overall affinity.

To enhance the virtual screening performance and compound selection, Kuhn, Ghogare et al. from Merck and Millipore Sigma embedded Chemical Space navigation with FTrees into the AIDDISON workflow.
FTrees is used for 2D pharmacophore searches, enabling scaffold hopping and efficient ligand-based screening to expand the chemical diversity of related compounds to a proposed candidated. AIDDISON incorporates several Chemical Spaces, Enamine's REAL Space, WuXi's GalaXi, and their own in-house SA-Space, each covering billions of synthetically accessible molecules.

The integration of the search algorithm and the combinatorial Chemical Spaces accelerates searching, filtering, and identifying of promising drug-like molecules while expanding into uncharged areas of the chemical space that would not be covered by other 2D methods.

(Workflow representation modified after Kuhn, Ghogare et al. 2023.)

In order to mine relevant chemistry based on an enumerated set of seed compounds, Kalliokoski et al. integrated SpaceLight and FTrees into their SpaceHASTEN workflow to optimize the performance of a structure-based virtual screening approach.
SpaceHASTEN is designed to handle giga-scale compound collections like Enamine's REAL Space and combines molecular docking and machine learning-guided compounds selection.
Docking scores are predicted with ML models used to prioritize compounds retrieved through similarity searches. An iteratively refined selection aims to maximize hit rates and deliver diverse and synthesizable virtual hits with high docking scores for experimental validation.

This approach enables efficient handling of billion-sized molecule clusters with reduced computational costs compared to brute-force docking.

(Workflow representation modified after Kalliokoski et al. 2023.)