Drug-induced liver injury (DILI) presents a significant challenge due to its complexity, small datasets, and severe class imbalance. While unsupervised pretraining is a common approach to learn molecular representations for downstream tasks, it often lacks insights into how molecules interact with biological systems. We therefore introduce VitroBERT, a bidirectional encoder representations from transformers (BERT) model pretrained on large-scale in vitro assay profiles to generate biologically informed molecular embeddings. When leveraged to predict in vivo DILI endpoints, these embeddings delivered up to a 29% improvement in biochemistry-related tasks and a 16% gain in histopathology endpoints compared to unsupervised pretraining (MolBERT). However, no significant improvement was observed in clinical tasks. Furthermore, to address the critical issue of class imbalance, we evaluated multiple loss functions-including BCE, weighted BCE, Focal loss, and weighted Focal loss-and identified weighted Focal loss as the most effective. Our findings demonstrate the potential of integrating biological context into molecular models and highlight the importance of selecting appropriate loss functions in improving model performance of highly imbalanced DILI-related tasks.

The advent of agentic AI systems is leading to significant transformations across scientific and technological domains. Advances in large language models (LLMs), reasoning capabilities, and integration with external tools have ushered in a new era where agentic AI systems can autonomously perform computational tasks that were traditionally carried out by humans. Computer-aided drug design (CADD)─a multifaceted process encompassing complex, interdependent tasks─stands to benefit profoundly from these advancements. However, one of the key challenges in enabling agentic systems to autonomously take over tasks in CADD is constructing models for property estimation that match the quality and reliability of those developed by human experts. As this is not currently straightforward, this capability represents a major bottleneck for fully realizing the potential of autonomous pipelines in drug discovery. We present here MolAgent, a system-agnostic agentic AI framework designed for high-fidelity modeling of molecular properties in early-stage drug discovery. MolAgent autonomously implements expert-level pipelines for both classification and regression, empowering agentic systems to efficiently construct and deploy models. With integrated automated feature engineering, robust model selection, advanced ensemble methodologies, and comprehensive validation frameworks, MolAgent ensures optimal accuracy and model robustness. The platform seamlessly accepts 2D and 3D structural data for ligands and receptors and harmonizes traditional molecular descriptors with advanced deep learning features extracted from pretrained 2D and 3D encoders. Ultimately, the platform's fully automated, end-to-end workflow is designed for seamless agentic execution. Adherence to the Model Context Protocol (MCP) guarantees interoperability with diverse agentic AI infrastructures, ensuring flexible integration into complex, future discovery pipelines.

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 contributionWe 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.

As training volume increases predictive model quality, leveraging existing external data sources holds the promise of time- and cost-efficiency. In a drug discovery setting, pharmaceutical companies all own substantial but confidential datasets. The MELLODDY project develops a privacy-preserving federated machine learning solution and deploys it at an unprecedented scale (more than 100,000 tasks across ten major pharmaceutical companies), while ensuring the security and privacy of each partner’s sensitive data. Each partner builds models that benefit from a shared representation, for their own private assays. Established predictive performance metrics such as AUC ROC or AUC PR are constrained to unseen labelled chemical space. However, they cannot gauge performance gains in unlabelled chemical space. Federated learning indirectly extends labelled space, but in a privacy-preserving context, a partner cannot use this label extension for performance assessment. Metrics that estimate uncertainty on a prediction can be calculated even where no label is known. Practically, the chemical space covered with predictions of sufficient confidence, reflects the applicability domain of a model. After establishing a link to established performance metrics, we propose the efficiency from the conformal prediction framework (‘conformal efficiency’) as a proxy to the applicability domain size. A documented extension of the applicability domain would qualify as a tangible benefit from federated learning. In interim assessments, MELLODDY partners report a median increase in conformal efficiency of the federated over the single-partner model of 5.5% (with increases up to 9.7%). Subject to distributional conditions, that efficiency increase can be directly interpreted as the expected increase in conformal i.e. high confidence predictions. In conclusion, we present the first evidence that privacy-preserving federated machine learning across massive drug-discovery datasets from ten pharma partners indeed extends the applicability domain of property prediction models.

We here introduce the Aquamarine (AQM) dataset, an extensive quantum-mechanical (QM) dataset that contains the structural and electronic information of 59,783 low-and high-energy conformers of 1,653 molecules with a total number of atoms ranging from 2 to 92 (mean:50.9), and containing up to 54 (mean:28.2) non-hydrogen atoms. To gain insights into the solvent effects as well as collective dispersion interactions for drug-like molecules, we have performed QM calculations supplemented with a treatment of many-body dispersion (MBD) interactions of structures and properties in the gas phase and implicit water. Thus, AQM contains over 40 global (molecular) and local (atom-in-a-molecule) physicochemical properties (including ground-state and response properties) per conformer computed at the tightly converged PBE0+MBD level of theory for gas-phase molecules, whereas PBE0+MBD with the modified Poisson-Boltzmann (MPB) model of water was used for solvated molecules. By addressing both molecule-solvent and dispersion interactions, AQM dataset can serve as a challenging benchmark for state-of-the-art machine learning methods for property modeling and \textit{de novo} generation of large (solvated) molecules with pharmaceutical and biological relevance.

In drug discovery, different data modalities (chemical structure, cell biology, quantum mechanics, etc.) are abundant, and their integration can help with understanding aspects of chemistry, biology, and their interactions. Within cell biology, cell painting (CP) and transcriptomics RNA-Seq (TX) screens are powerful tools in early drug discovery, as they are complementary views of the biological effect of compounds on a population of cells post-treatment. While multimodal learning of chemical structure-cell painting, or different omics data has been experimented; a cell painting-bulk transcriptomics multimodal model is still unexplored. In this work, we benchmark two representation learning methods: contrastive learning and bimodal autoencoder. We use the setting of cross modality learning where representation learning is performed with two modalities (CP and TX), but only cell painting is available for new compounds embeddings generation and downstream task. This is because for new compounds, we would only have CP data and not TX, due to high data generation cost of the RNA-Seq screen. We show that in the absence of TX features for new compounds, using learned embeddings like those obtained from Constrastive Learning enhances performance of CP features on tasks where TX features excels but CP features does not. Additionally, we observed that learned representation improves cluster quality for clustering of CP replicates and different mechanisms of action (MoA), as well as improves performance on several subsets of bioactivity tasks grouped by protein target families.

Human dose prediction (HDP) is a useful tool for compound optimization in preclinical drug discovery. We describe here our exclusively in silico HDP strategy to triage compound designs for synthesis and experimental profiling. Our goal is a model that provides a preliminary estimate of the dose for a given exposure target based on chemical structure. First, we construct machine learning models to estimate rat pharmacokinetics, which are subsequently allometrically scaled to estimate human pharmacokinetics. Second, we establish a 10 nM free concentration target for early HDP where potency data are not yet available. Finally, we assess the uncertainty associated with each model and propagate these into the final estimate, providing us with actionable guidance on the level of accuracy of these estimates. We find that this strategy can reduce preparation of compounds with poor properties relative to an unstructured approach, but extensive experimental testing remains required.

Drug-induced liver injury (DILI) presents a multifaceted challenge, influenced by interconnected biological mechanisms. Current DILI datasets are characterized by small sizes and high imbalance, posing difficulties in learning robust representations and accurate modeling. To address these challenges, we trained a multi-modal multi-task model integrating preclinical histopathologies, biochemistry (blood markers), and clinical DILI-related adverse drug reactions (ADRs). Leveraging pre-trained BERT models, we extracted representations covering a broad chemical space, facilitating robust learning in both frozen and fine-tuned settings. To address imbalanced data, we explored weighted Binary Cross-Entropy (w-BCE) and weighted Focal Loss (w-FL). Our results demonstrate that the frozen BERT model consistently enhances performance across all metrics and modalities with weighted loss functions compared to their non-weighted counterparts. However, the efficacy of fine-tuning BERT varies across modalities, yielding inconclusive results. In summary, the incorporation of BERT features with weighted loss functions demonstrates advantages, while the efficacy of fine-tuning remains uncertain.

AbstractImage-based models that use features extracted from cell microscopy images can estimate the activity of small molecules in various biological assays. Typically, models are trained on images stained by an optimized protocol (e.g. Cell Painting) after exposure to a fairly high small molecule concentration (referred to as ’image concentration’) of $$10\; \upmu {\text{M}}$$ 10 μ M or higher. Low concentration images (e.g. $$0.16$$ 0.16 μM, $$0.8$$ 0.8 μM, $$4$$ 4 μM) tend to yield models with worse performance. In this work, we nevertheless report a practical use for low image concentration data. We propose the combination of well-performing models trained at higher image concentrations, with lower image concentration for inference to identify more potent compounds. We show that this approach improves on the conventional method (directly training a high-potency model) in 65$$\%$$ % of assays investigated in terms of AUC-ROC, and 75$$\%$$ % of assays in terms of RIPtoP-corrected AUC-PR.

We here introduce the Aquamarine (AQM) dataset, an extensive quantum-mechanical (QM) dataset that contains the structural and electronic information of 59,783 low-and high-energy conformers of 1,653 molecules with a total number of atoms ranging from 2 to 92 (mean: 50.9), and containing up to 54 (mean: 28.2) non-hydrogen atoms. To gain insights into the solvent effects as well as collective dispersion interactions for drug-like molecules, we have performed QM calculations supplemented with a treatment of many-body dispersion (MBD) interactions of structures and properties in the gas phase and implicit water. Thus, AQM contains over 40 global and local physicochemical properties (including ground-state and response properties) per conformer computed at the tightly converged PBE0+MBD level of theory for gas-phase molecules, whereas PBE0+MBD with the modified Poisson-Boltzmann (MPB) model of water was used for solvated molecules. By addressing both molecule-solvent and dispersion interactions, AQM dataset can serve as a challenging benchmark for state-of-the-art machine learning methods for property modeling and de novo generation of large (solvated) molecules with pharmaceutical and biological relevance.

In a drug discovery setting, pharmaceutical companies own substantial but confidential datasets. The MELLODDY project developed a privacy-preserving federated machine learning solution and deployed it at an unprecedented scale. Each partner built models for their own private assays that benefitted from a shared representation. Established predictive performance metrics such as AUC ROC or AUC PR are constrained to unseen labeled chemical space and cannot gage performance gains in unlabeled chemical space. Federated learning indirectly extends labeled space, but in a privacy-preserving context, a partner cannot use this label extension for performance assessment. Metrics that estimate uncertainty on a prediction can be calculated even where no label is known. Practically, the chemical space covered with predictions above an uncertainty threshold, reflects the applicability domain of a model. After establishing a link to established performance metrics, we propose the efficiency from the conformal prediction framework (‘conformal efficiency’) as a proxy to the applicability domain size. A documented extension of the applicability domain would qualify as a tangible benefit from federated learning. In interim assessments, MELLODDY partners reported a median increase in conformal efficiency of the federated over the single-partner model of 5.5% (with increases up to 9.7%). Subject to distributional conditions, that efficiency increase can be directly interpreted as the expected increase in conformal i.e. low uncertainty predictions. In conclusion, we present the first indication that privacy-preserving federated machine learning across massive drug-discovery datasets from ten pharma partners indeed extends the applicability domain of property prediction models.

The search for chemical hit material is a lengthy and increasingly expensive drug discovery process. To improve it, ligand-based quantitative structure–activity relationship models have been broadly applied to optimize primary and secondary compound properties. Although these models can be deployed as early as the stage of molecule design, they have a limited applicability domain—if the structures of interest differ substantially from the chemical space on which the model was trained, a reliable prediction will not be possible. Image-informed ligand-based models partly solve this shortcoming by focusing on the phenotype of a cell caused by small molecules, rather than on their structure. While this enables chemical diversity expansion, it limits the application to compounds physically available and imaged. Here, we employ an active learning approach to capitalize on both of these methods’ strengths and boost the model performance of a mitochondrial toxicity assay (Glu/Gal). Specifically, we used a phenotypic Cell Painting screen to build a chemistry-independent model and adopted the results as the main factor in selecting compounds for experimental testing. With the additional Glu/Gal annotation for selected compounds we were able to dramatically improve the chemistry-informed ligand-based model with respect to the increased recognition of compounds from a 10% broader chemical space.

Drug-induced liver injury (DILI) is an important safety concern and a major reason to remove a drug from the market. Advancements in recent machine learning methods have led to a wide range of in silico models for DILI predictive methods based on molecule chemical structures (fingerprints). Existing publicly available DILI data sets used for model building are based on the interpretation of drug labels or patient case reports, resulting in a typical binary clinical DILI annotation. We developed a novel phenotype-based annotation to process hepatotoxicity information extracted from repeated dose in vivo preclinical toxicology studies using INHAND annotation to provide a more informative and reliable data set for machine learning algorithms. This work resulted in a data set of 430 unique compounds covering diverse liver pathology findings which were utilized to develop multiple DILI prediction models trained on the publicly available data (TG-GATEs) using the compound's fingerprint. We demonstrate that the TG-GATEs compounds DILI labels can be predicted well and how the differences between TG-GATEs and the external test compounds (Johnson & Johnson) impact the model generalization performance.

Federated multipartner machine learning has been touted as an appealing and efficient method to increase the effective training data volume and thereby the predictivity of models, particularly when the generation of training data is resource-intensive. In the landmark MELLODDY project, indeed, each of ten pharmaceutical companies realized aggregated improvements on its own classification or regression models through federated learning. To this end, they leveraged a novel implementation extending multitask learning across partners, on a platform audited for privacy and security. The experiments involved an unprecedented cross-pharma data set of 2.6+ billion confidential experimental activity data points, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. Appropriate complementary metrics were developed to evaluate the predictive performance in the federated setting. In addition to predictive performance increases in labeled space, the results point toward an extended applicability domain in federated learning. Increases in collective training data volume, including by means of auxiliary data resulting from single concentration high-throughput and imaging assays, continued to boost predictive performance, albeit with a saturating return. Markedly higher improvements were observed for the pharmacokinetics and safety panel assay-based task subsets.

Image-based profiling has emerged as a powerful technology for various steps in basic biological and pharmaceutical discovery, but the community has lacked a large, public reference set of data from chemical and genetic perturbations. Here we present data generated by the Joint Undertaking for Morphological Profiling (JUMP)-Cell Painting Consortium, a collaboration between 10 pharmaceutical companies, six supporting technology companies, and two non-profit partners. When completed, the dataset will contain images and profiles from the Cell Painting assay for over 116,750 unique compounds, over-expression of 12,602 genes, and knockout of 7,975 genes using CRISPR-Cas9, all in human osteosarcoma cells (U2OS). The dataset is estimated to be 115 TB in size and capturing 1.6 billion cells and their single-cell profiles. File quality control and upload is underway and will be completed over the coming months at the Cell Painting Gallery: https://registry.opendata.aws/cellpainting-gallery . A portal to visualize a subset of the data is available at https://phenaid.ardigen.com/jumpcpexplorer/ .

To apply federated learning to drug discovery we developed a novel platform in the context of European Innovative Medicines Initiative (IMI) project MELLODDY (grant n°831472), which was comprised of 10 pharmaceutical companies, academic research labs, large industrial companies and startups. The MELLODDY platform was the first industry-scale platform to enable the creation of a global federated model for drug discovery without sharing the confidential data sets of the individual partners. The federated model was trained on the platform by aggregating the gradients of all contributing partners in a cryptographic, secure way following each training iteration. The platform was deployed on an Amazon Web Services (AWS) multi-account architecture running Kubernetes clusters in private subnets. Organisationally, the roles of the different partners were codified as different rights and permissions on the platform and administrated in a decentralized way. The MELLODDY platform generated new scientific discoveries which are described in a companion paper.

In this paper, we compare the most popular Atom-to-Atom Mapping (AAM) tools: ChemAxon,[1] Indigo,[2] RDTool,[3] NameRXN (NextMove),[4] and RXNMapper[5] which implement different AAM algorithms. An open-source RDTool program was optimized, and its modified version ("new RDTool") was considered together with several consensus mapping strategies. The Condensed Graph of Reaction approach was used to calculate chemical distances and develop the "AAM fixer" algorithm for an automatized correction of erroneous mapping. The benchmarking calculations were performed on a Golden dataset containing 1851 manually mapped and curated reactions. The best performing RXNMapper program together with the AMM Fixer was applied to map the USPTO database. The Golden dataset, mapped USPTO and optimized RDTool are available in the GitHub repository https://github.com/Laboratoire-de-Chemoinformatique.

Background: Accurate prediction of absorption, distribution, metabolism and excretion (ADME) properties can facilitate the identification of promising drug candidates. Methodology & Results: The authors present the Janssen generic Target Product Profile (gTPP) model, which predicts 18 early ADME properties, employs a graph convolutional neural network algorithm and was trained on between 1000-10,000 internal data points per predicted parameter. gTPP demonstrated stronger predictive power than pretrained commercial ADME models and automatic model builders. Through a novel logging method, the authors report gTPP usage for more than 200 Janssen drug discovery scientists. Conclusion: The investigators successfully enabled the rapid and systematic implementation of predictive ML tools across a drug discovery pipeline in all therapeutic areas. This experience provides useful guidance for other large-scale AI/ML deployment efforts.

The quality of experimental data for chemical reactions is a critical consideration for any reaction-driven study. However, the curation of reaction data has not been extensively discussed in the literature so far. Here, we suggest a 4 steps protocol that includes the curation of individual structures (reactants and products), chemical transformations, reaction conditions and endpoints. Its implementation in Python3 using CGRTools toolkit has been used to clean three popular reaction databases Reaxys, USPTO and Pistachio. The curated USPTO database is available in the GitHub repository (Laboratoire-de-Chemoinformatique/Reaction_Data_Cleaning).