Aqueous solubility is the most important physicochemical property for agrochemical and drug candidates and a prerequisite for uptake, distribution, transport, and finally the bioavailability in living species. Prediction of solubility is not straightforward, with the two main problems being (i) that the free energy of solvation is comprised of the lattice energy and the dissolution of the solute from vacuum, and (ii) the dependence of the solvation free energy on the charge state of the molecule, i.e. its pH-dependence.

We approach the first problem by machine learning models for the prediction of the energy needed to reach an artificial-liquid-state, the super-cooled liquid, based on QM-derived COSMO-RS and Morgan descriptors, and the calculation of the free energy of solvation by COSMO-RS itself. Results are presented for carefully selected datasets from the Bayer compound collection.

The second problem, we address with the first-ever direct machine learning models for pH-dependent solubility in buffered water. For this, we combined almost 300000 data points from 11 solubility assays performed over 24 years and over one million data points from lipophilicity and melting point experiments. Data were split into three pH-classes − acidic, neutral and basic − , representing the conditions of stomach and intestinal tract for animals and humans, and phloem and xylem for plants. We find that multi-task neural networks using ECFP-6 fingerprints outperform baseline models with root mean square errors of 0.56 log units and Spearman rank correlations of 0.83 obteined over the three classes, allowing for the prediction of compound pH profiles.

Halogen bonds can be utilized as highly directional interactions in medicinal chemistry and drug discovery. More than a decade ago, we aimed to highlight their principles and applications as a perspective for the community. However, their use in molecular design demands a reasonably precise description of their energy-geometry-dependence. How such QM-derived scoring functions can be generated from machine learning approaches will be discussed.

An alternative route to identifying unconventional binding modes and innovative chemistry involving halogen bonds, is the use of halogen-enriched fragment libraries (HEFLibs). Examples showcasing the applicability of this strategy will be discussed for a variety of drug targets.

As an orthogonal principle to the established (hetero)aryl halides, CF₂X (X = Cl, Br, I) moieties provide the opportunity to release halogen bonding (XB) from its two-dimensional confinement. From monitoring the PDB, we identified the allosteric Bcr-Abl inhibitor asciminib as a first example for halogen bonding (XB) through a CF2Cl group. Challenges and opportunities of CF₂X groups will be analyzed and discussed from a multitude of different perspectives. CF₂X moieties can be valuable functions for fragments, but have been used as well to design unconventional amino acids for incorporation into therapeutic peptides or peptidomimetics.

A multitude of methods exist for binding site comparison in the scientific literature. All of them have been tested & validated versus various datasets. However, as of today, there is still no reference benchmark & study in the field despite several attempts made in the past.

In this talk I'll explain my perspective on issues I identified when reviewing these benchmark datasets. I'll cover common misconceptions around binding site comparison methods & outline several areas of improvement that need to happen in the field to address these issues. I'll give an outline on actions that we are taking to validate our own pocket comparison methodology and make use of the inherent knowledge contained in structural repositories like 3decision & PSILO.

Virtual screening is a rapid scouting method for finding hits in large compound collections, and typically involves integration of several computational methods at scale. Additionally, multiple criteria are used for nominating the hits for experimental testing, often in screen-test-rescreen cycles. In this talk, I will discuss the virtual screening workflow we implemented for NSP13 target as part of the CACHE challenge, and how we integrated different tools in a modular fashion for consistency and scalability. The talk will cover aspects of molecular docking, protein-ligand interaction analysis, and molecular dynamics simulation.

Sometimes it’s the little things that take the most time out of your day, preparing proteins, submitting jobs, assessing docking results, sharing data, etc. Thankfully some of those tasks can take care of themselves through automation. One way to achieve this is through the use of KNIME. In this talk I would like to introduce KNIME, the KNIME MOE nodes, and some of the ways KNIME can be used to offload some of the more repetitive tasks. With real-life examples from our lab that show how it can link to a wide range of existing tools and resources.

Though many existing covalent docking tools are successful, they have several limitations. They require time-consuming tasks such as pre-processing of ligands (standardization, filtering based on warheads type), setting up of reactions, protein preparation, complex formation, etc... They can only be used to dock a small library of compounds due to the high CPU time needed for performing covalent computations and scoring, rendering them unsuitable for performing large-scale virtual screening. We developed a toolkit for performing automated covalent docking in a fast and effective manner by combining GOLD, MOE, KNIME, and Python programs. In the first step, a KNIME workflow was developed to perform ligand preparation in a manner compatible with the covalent docking protocol of GOLD. The protein structures were prepared using the MOE QuickPrep module. Subsequently, a Python program was written to perform the covalent docking of ligands, by invoking the GOLD docking engine, in a parallelized fashion. Finally, the protein-ligand complexes were generated for the docking poses by calling a submodule of the Python program. We applied this toolkit retrospectively on six targets (NUDT7, OTUB2, EGFR, Cathepsin K, XPO1, and HCV NS3 protease) with known crystal structures and sets of active and inactive covalent inhibitors and assessed the potential of four GOLD docking scoring functions (GoldScore, ASP, ChemScore, and PLP). For most of the targets, the virtual screening metrics were satisfactory for the different scoring functions, demonstrating the program's usefulness in performing prospective virtual screening.

The KNIME workflow “Evotec_Covalent_Processing_forGOLD.knwf” for the preparation of the ligands is available in the KNIME Hub https://hub.knime.com/emilie_pihan/spaces. The Python program is available at https://gitlab.com/seb-buch/covalent_docking_helper.

De novo design is not a new concept in drug discovery, but until recently has made little quantifiable impact. Recent advances in artificial intelligence and novel deep-learning models alongside a commensurate increase in compute power have enabled the creation of new de novo methods capable of learning how to build drug-like molecules with good success rates. Further, these methods can be directed either towards diversity or towards achieving specific design goals.

In this talk I will introduce REINVENT[1], a promising method capable of building diverse compounds as SMILES notation that has been used to great effect on drug discovery projects at AstraZeneca. While the standard prior has been trained on a large and diverse set of drug-like molecules, transfer learning can be used to focus compound generation into specific areas of chemical space. Reward functions of various forms can be applied to each generation of compound design and through reinforcement learning the output will converge upon an optimal set of solutions. We tailor this approach to many tasks, including scaffold-hopping, hit-expansion and lead optimisation.

I will also introduce some tailored applications of REINVENT, created to specifically address the challenges and workflows of drug discovery. LibInvent[2] is a tool to generate de novo libraries, keeping one synthon as a constant and generating suitable reagents for a given reaction, followed by assessment of the final molecule using the reward function. Unlike standard reagent enumeration, this is not limited to an input set of reagents, but will stay within established project chemistry for straightforward follow-up. LinkInvent[3] has a similar concept but is directed at replacing a central part of the molecule, leaving components at either end untouched. This tool is particularly applicable to scaffold-hopping.

Finally, I will reflect on the cultural impact of AI within drug discovery. It has become increasingly apparent that there is an important role for human input into these systems and in reducing these ideas to practise. For all the promise and hype of artificial intelligence, it seems a partnership of AI and human inputs are the most efficient solution to rapidly progress drug design.

---

1. REINVENT 2.0: An AI Tool for De Novo Drug Design. Blaschke, Arús-Pous, Chen, et al., J. Chem. Inf. Model., 2020 60 (12), 5918-5922. DOI: 10.1021/acs.jcim.0c00915
2. LibINVENT: Reaction-based Generative Scaffold Decoration for in Silico Library Design. Fialková, Zhao, Papadopoulos, et al., J. Chem. Inf. Model., 2022 62 (9), 2046-2063. DOI: 10.1021/acs.jcim.1c00469
3. Link-INVENT: generative linker design with reinforcement learning. Guo, Knuth, Margreitter, et al., Digital Discovery, 2023 2, 392-408. DOI: 10.1039/D2DD00115B

RAS is the most frequently mutated oncogene in cancer, with KRAS G12C mutations most commonly found in lung adenocarcinoma, colorectal cancer, and other solid tumor malignancies. Covalent inhibitors of KRAS^G12C have shown antitumor activity against advanced/metastatic KRAS G12C-mutated cancers. Here we report the identification of a new class of pyrazole based KRAS^G12C inhibitors discovered by structure based de-novo design. The compounds bind to the KRAS^G12C switch II pocket with a novel binding mode, exploiting unique interactions with the GDP-bound form of the KRAS^G12C protein. We describe the hit to lead and lead optimization by structure-based design of the novel chemical series leading to the discovery of JDQ443, a novel, potent and selective, orally bioavailable KRAS^G12C covalent inhibitor.

cGMP-dependent protein kinase I-α (PKG1α) is a target for pulmonary arterial hypertension due to its role in the regulation of smooth muscle function. While most work has focused on regulation of cGMP turnover, we recently described several small molecule tool compounds which were capable of activating PKG1α via a cGMP independent pathway. Selected molecules were crystallized in the presence of PKG1α and were found to bind to an allosteric site proximal to the low-affinity nucleotide binding domain. These molecules act to displace the regulatory helix and cause activation of PKG1α representing a new mechanism for the activation and control of this critical therapeutic path. The described structures are vital to understanding the function and control of this key regulatory pathway.