Machine Learning-driven Fragment-based Discovery of CIB1-directed Anti-Tumor Agents by FRASE-bot.

Chemical probes are an indispensable tool for translating biological discoveries into new therapies, though are increasingly difficult to identify. Novel therapeutic targets are often hard-to-drug proteins, such as messengers or transcription factors. Computational strategies arise as a promising solution to expedite drug discovery for unconventional therapeutic targets. FRASE-bot exploits big data and machine learning (ML) to distill 3D information relevant to the target protein from thousands of protein-ligand complexes to seed it with ligand fragments. The seeded fragments can then inform either (i) de novo design of 3D ligand structures or (ii) ultra-large-scale virtual screening of commercially available compounds. Here, FRASE-bot was applied to identify ligands for Calcium and Integrin Binding protein 1 (CIB1), a promising but ligand-orphan drug target implicated in triple negative breast cancer. The signaling function of CIB1 relies on protein-protein interactions and its structure does not feature any natural ligand-binding pocket. FRASE-based virtual screening identified the first small-molecule CIB1 ligand (with binding confirmed in a TR-FRET assay) showing specific cell-killing activity in CIB1-dependent cancer cells, but not in CIB1-depleted cells.

Enzymatic Macrolactamization of mRNA Display Libraries for Inhibitor Selection.

mRNA display is a powerful, high-throughput technology for discovering novel, peptide ligands for protein targets. A number of methods have been used to expand the chemical diversity of mRNA display libraries beyond the 20 canonical amino acids, including genetic code reprogramming and biorthogonal chemistries. To date, however, there have been few reports using enzymes as biocompatible reagents for diversifying mRNA display libraries. Here, we report the evaluation and implementation of the common industrial enzyme, microbial transglutaminase (mTG), as a versatile biocatalyst for cyclization of mRNA display peptide libraries via lysine-to-glutamine isopeptide bonds. We establish two separate display-based assays to validate the compatibility of mTG with mRNA-linked peptide substrates. These assays indicate that mTG has a high degree of substrate tolerance and low single round bias. To demonstrate the potential benefits of mTG-mediated cyclization in ligand discovery, high diversity mTG-modified libraries were employed in two separate affinity selections: (1) one against the calcium and integrin binding protein, CIB1, and (2) the second against the immune checkpoint protein and emerging therapeutic target, B7-H3. Both selections resulted in the identification of potent, cyclic, low nanomolar binders, and subsequent structure-activity studies demonstrate the importance of the cyclization to the observed activity. Notably, cyclization in the CIB1 binder stabilizes an α-helical conformation, while the B7-H3 inhibitor employs two bridges, one mTG-derived lactam and a second disulfide to achieve its potency. Together, these results demonstrate potential benefits of enzyme-based biocatalysts in mRNA display ligand selections and establish a framework for employing mTG in mRNA display.

Combining pharmacophore models derived from DNA-encoded chemical libraries with structure-based exploration to predict Tankyrase 1 inhibitors.

DNA-encoded chemical libraries (DECLs) interrogate the interactions of a target of interest with vast numbers of molecules. DECLs hence provide abundant information about the chemical ligand space for therapeutic targets, and there is considerable interest in methods for exploiting DECL screening data to predict novel ligands. Here we introduce one such approach and demonstrate its feasibility using the cancer-related poly-(ADP-ribose)transferase tankyrase 1 (TNKS1) as a model target. First, DECL affinity selections resulted in structurally diverse TNKS1 inhibitors with high potency including compound 2 with an IC50 value of 0.8 nM. Additionally, TNKS1 hits from four DECLs were translated into pharmacophore models, which were exploited in combination with docking-based screening to identify TNKS1 ligand candidates in databases of commercially available compounds. This computational strategy afforded TNKS1 inhibitors that are outside the chemical space covered by the DECLs and yielded the drug-like lead compound 12 with an IC50 value of 22 nM. The study further provided insights in the reliability of screening data and the effect of library design on hit compounds. In particular, the study revealed that while in general DECL screening data are in good agreement with off-DNA ligand binding, unpredictable interactions of the DNA-attachment linker with the target protein contribute to the noise in the affinity selection data.

Scalable estimator of the diversity for de novo molecular generation resulting in a more robust QM dataset (OD9) and a more efficient molecular optimization.

Chemical diversity is one of the key term when dealing with machine learning and molecular generation. This is particularly true for quantum chemical datasets. The composition of which should be done meticulously since the calculation is highly time demanding. Previously we have seen that the most known quantum chemical dataset QM9 lacks chemical diversity. As a consequence, ML models trained on QM9 showed generalizability shortcomings. In this paper we would like to present (i) a fast and generic method to evaluate chemical diversity, (ii) a new quantum chemical dataset of 435k molecules, OD9, that includes QM9 and new molecules generated with a diversity objective, (iii) an analysis of the diversity impact on unconstrained and goal-directed molecular generation on the example of QED optimization. Our innovative approach makes it possible to individually estimate the impact of a solution to the diversity of a set, allowing for effective incremental evaluation. In the first application, we will see how the diversity constraint allows us to generate more than a million of molecules that would efficiently complete the reference datasets. The compounds were calculated with DFT thanks to a collaborative effort through the QuChemPedIA@home BOINC project. With regard to goal-directed molecular generation, getting a high QED score is not complicated, but adding a little diversity can cut the number of calls to the evaluation function by a factor of ten.

EvoMol: a flexible and interpretable evolutionary algorithm for unbiased de novo molecular generation.

The objective of this work is to design a molecular generator capable of exploring known as well as unfamiliar areas of the chemical space. Our method must be flexible to adapt to very different problems. Therefore, it has to be able to work with or without the influence of prior data and knowledge. Moreover, regardless of the success, it should be as interpretable as possible to allow for diagnosis and improvement. We propose here a new open source generation method using an evolutionary algorithm to sequentially build molecular graphs. It is independent of starting data and can generate totally unseen compounds. To be able to search a large part of the chemical space, we define an original set of 7 generic mutations close to the atomic level. Our method achieves excellent performances and even records on the QED, penalised logP, SAscore, CLscore as well as the set of goal-directed functions defined in GuacaMol. To demonstrate its flexibility, we tackle a very different objective issued from the organic molecular materials domain. We show that EvoMol can generate sets of optimised molecules having high energy HOMO or low energy LUMO, starting only from methane. We can also set constraints on a synthesizability score and structural features. Finally, the interpretability of EvoMol allows for the visualisation of its exploration process as a chemically relevant tree.

Predictive Models for Kinetic Parameters of Cycloaddition Reactions.

This paper reports SVR (Support Vector Regression) and GTM (Generative Topographic Mapping) modeling of three kinetic properties of cycloaddition reactions: rate constant (logk), activation energy (Ea) and pre-exponential factor (logA). A data set of 1849 reactions, comprising (4+2), (3+2) and (2+2) cycloadditions (CA) were studied in different solvents and at different temperatures. The reactions were encoded by the ISIDA fragment descriptors generated for Condensed Graph of Reaction (CGR). For a given reaction, a CGR condenses structures of all the reactants and products into one single molecular graph, described both by conventional chemical bonds and "dynamical" bonds characterizing chemical transformations. Different scenarios of logk assessment were exploited: direct modeling, application of the Arrhenius equation and temperature-scaled GTM landscapes. The logk models with optimal cross-validated statistics (Q2 =0.78-0.94 RMSE=0.45-0.86) have been challenged to predict rates for the external test set of 200 reactions, comprising both reactions that were not present in the training set, and training set transformations performed under different reaction conditions. The models are freely available on our web-server: http://cimm.kpfu.ru/models.

Dataset's chemical diversity limits the generalizability of machine learning predictions.

The QM9 dataset has become the golden standard for Machine Learning (ML) predictions of various chemical properties. QM9 is based on the GDB, which is a combinatorial exploration of the chemical space. ML molecular predictions have been recently published with an accuracy on par with Density Functional Theory calculations. Such ML models need to be tested and generalized on real data. PC9, a new QM9 equivalent dataset (only H, C, N, O and F and up to 9 "heavy" atoms) of the PubChemQC project is presented in this article. A statistical study of bonding distances and chemical functions shows that this new dataset encompasses more chemical diversity. Kernel Ridge Regression, Elastic Net and the Neural Network model provided by SchNet have been used on both datasets. The overall accuracy in energy prediction is higher for the QM9 subset. However, a model trained on PC9 shows a stronger ability to predict energies of the other dataset.

Visualization and Analysis of Complex Reaction Data: The Case of Tautomeric Equilibria.

Generative Topographic Mapping (GTM) approach was successfully used to visualize, analyze and model the equilibrium constants (KT ) of tautomeric transformations as a function of both structure and experimental conditions. The modeling set contained 695 entries corresponding to 350 unique transformations of 10 tautomeric types, for which KT values were measured in different solvents and at different temperatures. Two types of GTM-based classification models were trained: first, a "structural" approach focused on separating tautomeric classes, irrespective of reaction conditions, then a "general" approach accounting for both structure and conditions. In both cases, the cross-validated Balanced Accuracy was close to 1 and the clusters, assembling equilibria of particular classes, were well separated in 2-dimentional GTM latent space. Data points corresponding to similar transformations measured under different experimental conditions, are well separated on the maps. Additionally, GTM-driven regression models were found to have their predictive performance dependent on different scenarios of the selection of local fragment descriptors involving special marked atoms (proton donors or acceptors). The application of local descriptors significantly improves the model performance in 5-fold cross-validation: RMSE=0.63 and 0.82 logKT units with and without local descriptors, respectively. This trend was as well observed for SVR calculations, performed for the comparison purposes.

Predictive Models for Halogen-bond Basicity of Binding Sites of Polyfunctional Molecules.

Halogen bonding (XB) strength assesses the ability of an electron-enriched group to be involved in complexes with polarizable electrophilic halogenated or diatomic halogen molecules. Here, we report QSPR models of XB of particular relevance for an efficient screening of large sets of compounds. The basicity is described by pKBI2 , the decimal logarithm of the experimental 1 : 1 (B : I2 ) complexation constant K of organic compounds (B) with diiodine (I2 ) as a reference halogen-bond donor in alkanes at 298 K. Modeling involved ISIDA fragment descriptors, using SVM and MLR methods on a set of 598 organic compounds. Developed models were then challenged to make predictions for an external test set of 11 polyfunctional compounds for which unambiguous assignment of the measured effective complexation constant to specific groups out of the putative acceptor sites is not granted. At this stage, developed models were used to predict pKBI2 of all putative acceptor sites, followed by an estimation of the predicted effective complexation constant using the ChemEqui program. The best consensus models perform well both in cross-validation (root mean squared error RMSE=0.39-0.47 logKBI2 units) and external predictions (RMSE=0.49). The SVM models are implemented on our website (http://infochim.u-strasbg.fr/webserv/VSEngine.html) together with the estimation of their applicability domain and an automatic detection of potential halogen-bond acceptor atoms.

Predictive Models for the Free Energy of Hydrogen Bonded Complexes with Single and Cooperative Hydrogen Bonds.

In this work, we report QSPR modeling of the free energy ΔG of 1 : 1 hydrogen bond complexes of different H-bond acceptors and donors. The modeling was performed on a large and structurally diverse set of 3373 complexes featuring a single hydrogen bond, for which ΔG was measured at 298 K in CCl4 . The models were prepared using Support Vector Machine and Multiple Linear Regression, with ISIDA fragment descriptors. The marked atoms strategy was applied at fragmentation stage, in order to capture the location of H-bond donor and acceptor centers. Different strategies of model validation have been suggested, including the targeted omission of individual H-bond acceptors and donors from the training set, in order to check whether the predictive ability of the model is not limited to the interpolation of H-bond strength between two already encountered partners. Successfully cross-validating individual models were combined into a consensus model, and challenged to predict external test sets of 629 and 12 complexes, in which donor and acceptor formed single and cooperative H-bonds, respectively. In all cases, SVM models outperform MLR. The SVM consensus model performs well both in 3-fold cross-validation (RMSE=1.50 kJ/mol), and on the external test sets containing complexes with single (RMSE=3.20 kJ/mol) and cooperative H-bonds (RMSE=1.63 kJ/mol).