In the drug discovery process, where experiments can be costly and time-consuming, computational models that predict drug-target interactions are valuable tools to accelerate the development of new therapeutic agents. Estimating the uncertainty inherent in these neural network predictions provides valuable information that facilitates optimal decision-making when risk assessment is crucial. However, such models can be poorly calibrated, which results in unreliable uncertainty estimates that do not reflect the true predictive uncertainty. In this study, we compare different metrics, including accuracy and calibration scores, used for model hyperparameter tuning to investigate which model selection strategy achieves well-calibrated models. Furthermore, we propose to use a computationally efficient Bayesian uncertainty estimation method named HMC Bayesian Last Layer (HBLL), which generates Hamiltonian Monte Carlo (HMC) trajectories to obtain samples for the parameters of a Bayesian logistic regression fitted to the hidden layer of the baseline neural network. We report that this approach improves model calibration and achieves the performance of common uncertainty quantification methods by combining the benefits of uncertainty estimation and probability calibration methods. Finally, we show that combining post hoc calibration method with well-performing uncertainty quantification approaches can boost model accuracy and calibration.
Molecular optimization is a crucial step in drug development, involving structural modifications to improve the desired properties of drug candidates. Although many deep-learning-based molecular optimization algorithms have been proposed and may perform well on benchmarks, they usually do not pay sufficient attention to the synthesizability of molecules, resulting in optimized compounds difficult to be synthesized. To address this issue, we first developed a general pipeline capable of constructing functional reaction template library specific to any property where a predictive model can be built. Based on these functional templates, we introduced Syn-MolOpt, a synthesis planning-oriented molecular optimization method. During optimization, functional reaction templates steer the process towards specific properties by effectively transforming relevant structural fragments. In four diverse tasks, including two toxicity-related (GSK3β-Mutagenicity and GSK3β-hERG) and two metabolism-related (GSK3β-CYP3A4 and GSK3β-CYP2C19) multi-property molecular optimizations, Syn-MolOpt outperformed three benchmark models (Modof, HierG2G, and SynNet), highlighting its efficacy and adaptability. Additionally, visualization of the synthetic routes for molecules optimized by Syn-MolOpt confirms the effectiveness of functional reaction templates in molecular optimization. Notably, Syn-MolOpt’s robust performance in scenarios with limited scoring accuracy demonstrates its potential for real-world molecular optimization applications. By considering both optimization and synthesizability, Syn-MolOpt promises to be a valuable tool in molecular optimization.
Scientific contribution Syn-MolOpt takes into account both molecular optimization and synthesis, allowing for the design of property-specific functional reaction template libraries for the properties to be optimized, and providing reference synthesis routes for the optimized compounds while optimizing the targeted properties. Syn-MolOpt’s universal workflow makes it suitable for various types of molecular optimization tasks.
Computer-aided drug design has the potential to significantly reduce the astronomical costs of drug development, and molecular docking plays a prominent role in this process. Molecular docking is an in silico technique that predicts the bound 3D conformations of two molecules, a necessary step for other structure-based methods. Here, we describe version 1.3 of the open-source molecular docking software Gnina. This release updates the underlying deep learning framework to PyTorch, resulting in more computationally efficient docking and paving the way for seamless integration of other deep learning methods into the docking pipeline. We retrained our CNN scoring functions on the updated CrossDocked2020 v1.3 dataset and introduce knowledge-distilled CNN scoring functions to facilitate high-throughput virtual screening with Gnina. Furthermore, we add functionality for covalent docking, where an atom of the ligand is covalently bound to an atom of the receptor. This update expands the scope of docking with Gnina and further positions Gnina as a user-friendly, open-source molecular docking framework. Gnina is available at https://github.com/gnina/gnina.
Scientific contributions: GNINA 1.3 is an open source a molecular docking tool with enhanced support for covalent docking and updated deep learning models for more effective docking and screening.
We evaluate the impact of pretraining Graph Transformer architectures on atom-level quantum-mechanical features for the modeling of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of drug-like compounds. We compare this pretraining strategy with two others: one based on molecular quantum properties (specifically the HOMO-LUMO gap) and one using a self-supervised atom masking technique. After fine-tuning on Therapeutic Data Commons ADMET datasets, we evaluate the performance improvement in the different models observing that models pretrained with atomic quantum mechanical properties produce in general better results. We then analyze the latent representations and observe that the supervised strategies preserve the pretraining information after fine-tuning and that different pretrainings produce different trends in latent expressivity across layers. Furthermore, we find that models pretrained on atomic quantum mechanical properties capture more low-frequency Laplacian eigenmodes of the input graph via the attention weights and produce better representations of atomic environments within the molecule. Application of the analysis to a much larger non-public dataset for microsomal clearance illustrates generalizability of the studied indicators. In this case the performances of the models are in accordance with the representation analysis and highlight, especially for the case of masking pretraining and atom-level quantum property pretraining, how model types with similar performance on public benchmarks can have different performances on large scale pharmaceutical data.
Scientific contribution
We systematically compared three different data type/methodologies for pretraining molecular Graphormer with the purpose of modeling ADMET properties as downstream tasks. The learned representations from differently pretrained models were analyzed in addition to comparison of downstream task performances that have been typically reported in similar works. Such examination methodologies, including a newly introduced analysis of Graphormer’s Attention Rollout Matrix, can guide pretraining strategy selection, as corroborated by a performance evaluation on a larger internal dataset.
Retrosynthesis consists of recursively breaking down a target molecule to produce a synthesis route composed of readily accessible building blocks. In recent years, computer-aided synthesis planning methods have allowed a greater exploration of potential synthesis routes, combining state-of-the-art machine-learning methods with chemical knowledge. However, these methods are generally developed to produce individual routes from a singular product to a set of proposed building blocks and are not designed to leverage potential shared paths between targets. These methods do not necessarily encompass real-world use cases in medicinal chemistry, where one seeks to synthesize sets of target compounds in a library mode, looking for maximal convergence into a shared retrosynthetic path going via advanced key intermediate compounds. Using a graph-based processing pipeline, we explore Johnson & Johnson Electronic Laboratory Notebooks (J&J ELN) and publicly available datasets to identify complex routes with multiple target molecules sharing common intermediates, producing convergent synthesis routes. We find that over 70% of all reactions are involved in convergent synthesis, covering over 80% of all projects in the case of J&J ELN data.
Scientific contribution
We introduce a novel planning approach to develop convergent synthesis routes, which can search multiple products and intermediates simultaneously guided by state-of-the-art machine learning single-step retrosynthesis models, enhancing the overall efficiency and practical applicability of retrosynthetic planning. We evaluate the multi-step synthesis planning approach using the extracted convergent routes and observe that solvability is generally high across those routes, being able to identify a convergent route for over 80% of the test routes and showing an individual compound solvability of over 90%. We find that by using a convergent search approach, we can synthesize almost 30% more compounds simultaneously for J&J ELN as compared to using an individual search, while providing an increased use of common intermediates.
In this study, we propose a neural network- based approach to analyze IR spectra and detect the presence of functional groups. Our neural network architecture is based on the concept of learning split representations. We demonstrate that our method achieves favorable validation performance using the NIST dataset. Furthermore, by incorporating additional data from the open-access research data repository Chemotion, we show that our model improves the classification performance for nitriles and amides.
Scientific contribution: Our method exclusively uses IR data as input for a neural network, making its performance, unlike other well-performing models, independent of additional data types obtained from analytical measurements. Furthermore, our proposed method leverages a deep learning model that outperforms previous approaches, achieving F1 scores above 0.7 to identify 17 functional groups. By incorporating real-world data from various laboratories, we demonstrate how open-access, specialized research data repositories can serve as yet unexplored, valuable benchmark datasets for future machine learning research.
Machine learning is quickly becoming integral to drug discovery pipelines, particularly quantitative structure-activity relationship (QSAR) and absorption, distribution, metabolism, and excretion (ADME) tasks. Graph Convolutional Network (GCN) models have proven especially promising due to their inherent ability to model molecular structures using graph-based representations. However, maximizing the potential of such models in practice is challenging, as companies prioritize data privacy and security over collaboration initiatives to improve model performance and robustness. kMoL is an open-source machine learning library with integrated federated learning capabilities developed to address such challenges. Its key features include state-of-the-art model architectures, Bayesian optimization, explainability, and federated learning mechanisms. It demonstrates extensive customization possibilities, advanced security features, straightforward implementation of user-specific models, and high adaptability to custom datasets without additional programming requirements. kMoL is evaluated through locally trained benchmark settings and distributed federated learning experiments using various datasets to assess the features and flexibility of the library, as well as the ability to facilitate fast and practical experimentation. Additionally, results of these experiments provide further insights into the performance trade-offs associated with federated learning strategies, presenting valuable guidance for deploying machine learning models in a privacy-preserving manner within drug discovery pipelines. kMoL is available on GitHub at https://github.com/elix-tech/kmol.
Scientific contribution The primary scientific contribution of this research project is the introduction and evaluation of kMoL, an open-source machine learning library with integrated federated learning capabilities. By demonstrating advanced customization and security capabilities without additional programming requirements, kMoL represents an accessible yet secure open-source platform for collaborative drug discovery projects. Additionally, the experiment results provide further insights into the performance trade-offs associated with federated learning strategies, presenting valuable guidance for deploying machine learning models in a privacy-preserving manner within drug discovery pipelines.
With the cost/yield-ratio of drug development becoming increasingly unfavourable, recent work has explored machine learning to accelerate early stages of the development process. Given the current success of deep generative models across domains, we here investigated their application to the property-based proposal of new small molecules for drug development. Specifically, we trained a latent diffusion model—DrugDiff—paired with predictor guidance to generate novel compounds with a variety of desired molecular properties. The architecture was designed to be highly flexible and easily adaptable to future scenarios. Our experiments showed successful generation of unique, diverse and novel small molecules with targeted properties. The code is available at https://github.com/MarieOestreich/DrugDiff.
This work expands the use of generative modelling in the field of drug development from previously introduced models for proteins and RNA to the here presented application to small molecules. With small molecules making up the majority of drugs, but simultaneously being difficult to model due to their elaborate chemical rules, this work tackles a new level of difficulty in comparison to sequence-based molecule generation as is the case for proteins and RNA. Additionally, the demonstrated framework is highly flexible, allowing easy addition or removal of considered molecular properties without the need to retrain the model, making it highly adaptable to diverse research settings and it shows compelling performance for a wide variety of targeted molecular properties.
In recent years, the integration of machine learning techniques into chemical reaction product prediction has opened new avenues for understanding and predicting the behaviour of chemical substances. The necessity for such predictive methods stems from the growing regulatory and social awareness of the environmental consequences associated with the persistence and accumulation of chemical residues. Traditional biodegradation prediction methods rely on expert knowledge to perform predictions. However, creating this expert knowledge is becoming increasingly prohibitive due to the complexity and diversity of newer datasets, leaving existing methods unable to perform predictions on these datasets. We formulate the product prediction problem as a sequence-to-sequence generation task and take inspiration from natural language processing and other reaction prediction tasks. In doing so, we reduce the need for the expensive manual creation of expert-based rules.