Kinase inhibitors are central to targeted therapies, yet the disparity between large activity-focused databases (≈ 224, 000 unique kinase-ligand pairs in ChEMBL) and much smaller, structural datasets (≈ 6, 738 unique structures in KLIFS) limits the performance of structure-based machine learning (ML) for kinase-inhibitor binding affinity predictions. Although docking and other computational methods can generate large numbers of in silico structures, their use is typically restricted to poses within 2−4 Å of an experimental reference, discarding much of the simulated data. In this work, we introduce multi-objective Kinodata3D, a multi-objective E(3)invariant graph neural network (GNN) that leverages kinase–ligand structures across a broad range of structural quality. The model jointly predicts binding affinity, activity uncertainty, and structural quality, using the predicted reliability of each complex to weight its contribution to the activity loss. This allows the model to learn from a wide range of in silico structures without treating all of them as equally informative. Compared with an equivalent model without structure-quality prediction, multiobjective Kinodata3D reduces activity-prediction error and assigns lower uncertainty to structurally reliable complexes. Incorporating low-quality poses helps the model identify structurally unreliable complexes and assign them higher uncertainty, while prioritizing affinity learning from high-quality structures. By explicitly modeling structural quality, multi-objective Kinodata3D makes affinity predictions more interpretable by linking prediction confidence to the reliability of the input structure. These results show that in silico structures spanning a broad range of pose qualities can be used reliably for binding affinity prediction when structural uncertainty is modeled directly. Because the framework is agnostic to the structuregeneration method, it provides a general strategy for reducing structural data scarcity and delivering more accurate, better-calibrated, and more interpretable predictions for structure-based ML in drug discovery. multi-objective Kinodata3D is available at https://github.com/openkinome/multi-objective-kinodata-3D

Structure-based virtual screening (SBVS) is a cornerstone of computer-aided drug design, yet its success depends on selecting a combination of docking tools, scoring function (SF), and ranking strategies. MolDockLab addresses this challenge with an automated, data-driven framework that optimizes SBVS workflows for a protein target, balancing predictive performance and computational efficiency. It systematically explores combinations of five docking engines, 15 SF, and three consensus ranking strategies using a calibration set of ≈ 200 compounds with known bioactivity, and applies the best-correlating workflow to the larger screening library. Final hit selection from the top 1% integrates protein-ligand interaction profiler (PLIP)-derived interaction fingerprints, structural-diversity assessment, and expert visual inspection. In a retrospective evaluation on the epidermal growth factor receptor (EGFR), the chosen pipeline achieved a Spearman correlation of 0.36 and an enrichment factor (EF) at 10% of 1.57, consistent with calibration. Prospectively, for the energy coupling factor transporters (ECF-T)-a challenging transmembrane target with a cryptic binding site and no co-crystallized ligand-the pipeline reached a correlation of 0.45 and enrichment of 3.13. Post-processing enabled in vitro confirmation of two chemically novel inhibitors rivaling the most potent ECF-T inhibitors reported to date.

Accurately modeling interactions between small molecules and proteins using machine learning (ML) remains a central challenge in modern drug discovery. In particular, achieving and reliably assessing generalization in ML-based mod- els for binding affinity prediction has proven difficult. Focusing on the protein kinase domain, we investigate both the predictive performance and biophysical plausibility of graph neural networks (GNNs) trained to predict binding affinity, with and without access to docked protein–ligand complexes. To this end, we introduce a model-agnostic explainable AI framework that interprets model pre- dictions by attributing them to specific residues within the ATP-binding pocket of kinases. On the basis of discrete interaction fingerprints (IFPs), we employ robust statistical methods and feature selection to derive reference interaction profiles that serve as ground-truth explanations. Our analysis identifies 131 dis- tinct residues across 29 unique kinases that play key roles in linking IFP and binding affinity data. This enables a complementary benchmark, spanning nearly 10,000 docked complexes, to quantitatively assess model quality not only in terms of predictive accuracy but also biophysical alignment. The benchmark quantifies to what extent the model’s predictive mechanism aligns with meaningful bio- physical mechanisms—specifically, whether it recognizes protein regions actively involved in ligand binding. In a pilot study, we find that models incorporating 3D structural features exhibit moderate biophysical alignment, whereas baseline models without 3D information show no significant alignment, underscoring the structural nature of molecular binding, and importance thereof for training GNN models.

Structure-based virtual screening (SBVS) is a cornerstone of modern computer-aided drug design, yet its success is highly dependent on selecting an appropriate combination of docking tools, scoring functions (SFs), and ranking strategies from the vast array of available options. MolDockLab addresses this challenge by providing an automated, data-driven framework that optimizes SBVS workflows for a given protein target, for balanced predictive performance and computational efficiency, ultimately improving hit identification.MolDockLab systematically explores all combinations of five docking engines, fifteen SFs, and three consensus ranking strategies using a calibration set of approximately 200 compounds with known bioactivity. The SBVS workflow showing the highest correlation with experimental data is then applied to the larger screening library. Final hit selection from the top 1% integrates protein–ligand interaction profiler (PLIP)-derived protein–ligand interaction fingerprints, structural diversity assessment, and expert manual inspection.MolDockLab was validated through both a retrospective and a prospective case study. In the retrospective evaluation on epidermal growth factor receptor (EGFR), the chosen pipeline achieved a Spearman correlation of 0.35 and an enrichment factor at 10% (EF₁₀%) of 1.51, both consistent with the calibration set, used for pipeline selection. In the prospective application to the energy coupling factor transporters (ECF-T), a challenging transmembrane target featuring a dynamic, membrane-embedded cryptic binding site and no co-crystallized ligand, the optimal pipeline on the calibration set reached a Spearman correlation of 0.45 and an EF₁₀% of 3.13. Subsequent post-processing enabled the identification and in vitro confirmation of two chemically novel active compounds, which their potencies rival those of the best ECF-T inhibitors reported to date.Application to a challenging case study demonstrates that MolDockLab effectively enhances hit identification by tailoring SBVS workflows to individual targets. While the framework improves robustness and reproducibility, its performance remains dependent on the availability of an adequate calibration set for pipeline selection.

Recent advances in machine learning (ML) are reshaping drug discovery. Structure-based ML methods use physically-inspired models to predict binding affinities from protein:ligand complexes. These methods promise to enable the integration of data for many related targets, which addresses issues related to data scarcity for single targets and could enable generalizable predictions for a broad range of targets, including mutations. In this work, we report our experiences in building KinoML, a novel framework for ML in target-based small molecule drug discovery with an emphasis on structure-enabled methods. KinoML focuses currently on kinases as the relative structural conservation of this protein superfamily, particularly in the kinase domain, means it is possible to leverage data from the entire superfamily to make structure-informed predictions about binding affinities, selectivities, and drug resistance. Some key lessons learned in building KinoML include the importance of reproducible data collection and deposition, the harmonization of molecular data and featurization, and the selection of the right data format to ensure reusability and reproducibility of ML models. As a result, KinoML allows users to easily achieve three tasks: accessing and curating molecular data; featurizing this data with representations suitable for ML applications; and running reproducible ML experiments that require access to ligand, protein, and assay information to predict ligand affinity. Despite KinoML focusing on kinases, this framework can be applied to other proteins. The lessons reported here can help guide the development of platforms for structure-enabled ML in other areas of drug discovery.

The effectiveness of machine learning (ML) in drug discovery hinges on evaluation and modeling approaches that align with how compounds are tested and compared in real experimental contexts. We observe that experimental data in public repositories like ChEMBL naturally clusters by assay origin, while retaining significant overlap between training and test sets even when using common splitting strategies. This clustering effect is notable as simply predicting the mean activity value from a training assay yields surprisingly strong performance on test compounds from the same assay. To address these observations, we propose a paradigm for ML in drug discovery that respects the inherent structure of aggregated experimental data. We implement this approach through: (1) data splitting that allocates entire assays to either training or test sets, (2) evaluation metrics that assess ranking performance within individual assays rather than absolute prediction accuracy across heterogeneous experiments, and (3) set-based ranking models trained specifically on compound sets drawn from the same assay rather than random sets. Evaluating our approach on three datasets derived from ChEMBL, we demonstrate that ranking models trained on intra-assay sets consistently outperform both traditional IC50 prediction and ranking models trained on arbitrary compound sets. This performance advantage is most pronounced when data is curated minimally, suggesting that our approach effectively mitigates inconsistencies between experiments. Our findings indicate that ML methods for drug discovery should prioritize intra-assay ranking capability over absolute value prediction when working with aggregated experimental data.

Recent advances in machine learning (ML) are reshaping drug discovery. Structure-based ML methods use physically-inspired models to predict binding affinities from protein:ligand complexes. These methods promise to enable the integration of data for many related targets, which addresses issues related to data scarcity for single targets and could enable generalizable predictions for a broad range of targets, including mutants. In this work, we report our experiences in building KinoML, a novel framework for ML in target-based small molecule drug discovery with an emphasis on structure-enabled methods. KinoML focuses currently on kinases as the relative structural conservation of this protein superfamily, particularly in the kinase domain, means it is possible to leverage data from the entire superfamily to make structure-informed predictions about binding affinities, selectivities, and drug resistance. Some key lessons learned in building KinoML include: the importance of reproducible data collection and deposition, the harmonization of molecular data and featurization, and the choice of the right data format to ensure reusability and reproducibility of ML models. As a result, KinoML allows users to easily achieve three tasks: accessing and curating molecular data; featurizing this data with representations suitable for ML applications; and running reproducible ML experiments that require access to ligand, protein, and assay information to predict ligand affinity. Despite KinoML focusing on kinases, this framework can be applied to other proteins. The lessons reported here can help guide the development of platforms for structure-enabled ML in other areas of drug discovery.

Herein, we present the first application of target-directed dynamic combinatorial chemistry (tdDCC) to the whole complex of the highly dynamic transmembrane, energy-coupling factor (ECF) transporter ECF-PanT in Streptococcus pneumoniae. In addition, we successfully employed the tdDCC technique as a hit-identification and -optimization strategy that led to the identification of optimized ECF inhibitors with improved activity. We characterized the best compounds regarding cytotoxicity and performed computational modeling studies on the crystal structure of ECF-PanT to rationalize their binding mode. Notably, docking studies showed that the acylhydrazone linker is able to maintain the crucial interactions.

Drug discovery pipelines nowadays rely on machine learning models to explore and evaluate large chemical spaces. While including 3D structural information is considered beneficial, structural models are hindered by the availability of protein-ligand complex structures. Exemplified for kinase drug discovery, we address this issue by generating kinase-ligand complex data using template docking for the kinase compound subset of available ChEMBL assay data. To evaluate the benefit of the created complex data, we use it to train a structure-based E(3)-invariant graph neural network (GNN). Our evaluation shows that binding affinities can be predicted with significantly higher precision by models that take synthetic binding poses into account compared to ligand or DTI models only.

Problem settingStochastic dynamical systems in which local interactions give rise to complex emerging phenomena are ubiquitous in nature and society. This work explores the problem of inferring the unknown interaction structure (represented as a graph) of such a system from measurements of its constituent agents or individual components (represented as nodes). We consider a setting where the underlying dynamical model is unknown and where different measurements (i.e., snapshots) may be independent (e.g., may stem from different experiments).MethodOur method is based on the observation that the temporal stochastic evolution manifests itself in local patterns. We show that we can exploit these patterns to infer the underlying graph by formulating a masked reconstruction task. Therefore, we propose GINA (Graph Inference Network Architecture), a machine learning approach to simultaneously learn the latent interaction graph and, conditioned on the interaction graph, the prediction of the (masked) state of a node based only on adjacent vertices. Our method is based on the hypothesis that the ground truth interaction graph-among all other potential graphs-allows us to predict the state of a node, given the states of its neighbors, with the highest accuracy.ResultsWe test this hypothesis and demonstrate GINA's effectiveness on a wide range of interaction graphs and dynamical processes. We find that our paradigm allows to reconstruct the ground truth interaction graph in many cases and that GINA outperforms statistical and machine learning baseline on independent snapshots as well as on time series data.

TeachOpenCADD is a free online platform that offers solutions to common 16 computer-aided drug design (CADD) tasks using Python programming and open-source data and 17 packages. The material is presented through interactive Jupyter notebooks, accommodating 18 users from various backgrounds and programming levels. 19 Due to the tremendous impact of deep learning (DL) methods in drug design, the 20 TeachOpenCADD platform has been expanded to include an introduction to molecular DL tasks. 21 This edition provides an overview of DL and its application in drug design, highlighting the usage 22 of diverse molecular representations in this field. The platform introduces various neural 23 network architectures, including graph neural networks (GNNs), equivariant graph neural 24 networks (EGNNs), and recurrent neural networks (RNNs). It demonstrates how to use these 25 architectures for developing predictive models for molecular property and activity prediction, 26 exemplified by the Quantum Machine 9 (QM9), ChEMBL, and Kinase Inhibitor BioActivity (KiBA) 27 data sets. The DL edition covers methods for evaluating the performance of neural networks 28 using uncertainty estimation. Furthermore, it introduces an application of GNNs for 29 protein-ligand interaction predictions, incorporating protein structure and ligand information. 30 The TeachOpenCADD platform is continuously updated with new content and is open to 31 contributions, bug reports, and questions from the community through its GitHub repository 32 (github.com/volkamerlab/teachopencadd). It can be used for self-study, classroom instruction, and 33 research applications, accommodating users from beginners to advanced levels. 34

TeachOpenCADD is a free online platform that offers solutions to common computer-aided drug design (CADD) tasks using Python programming and open-source data and packages. The material is presented through interactive Jupyter notebooks, accommodating users from various backgrounds and programming levels. Due to the tremendous impact of deep learning (DL) methods in drug design, the TeachOpenCADD platform has been expanded to include an introduction to molecular DL tasks. This edition provides an overview of DL and its application in drug design, highlighting the usage of diverse molecular representations in this field. The platform introduces various neural network architectures, including graph neural networks (GNNs), equivariant graph neural networks (EGNNs), and recurrent neural networks (RNNs). It demonstrates how to use these architectures for developing predictive models for molecular property and activity prediction, exemplified by the Quantum Machine 9 (QM9), ChEMBL, and Kinase Inhibitor BioActivity (KiBA) data sets. The DL edition covers methods for evaluating the performance of neural networks using uncertainty estimation. Furthermore, it introduces an application of GNNs for protein-ligand interaction predictions, incorporating protein structure and ligand information. The TeachOpenCADD platform is continuously updated with new content and is open to contributions, bug reports, and questions from the community through its GitHub repository (https://github.com/volkamerlab/teachopencadd). It can be used for self-study, classroom instruction, and research applications, accommodating users from beginners to advanced levels.

To understand the long-run behavior of Markov population models, the computation of the stationary distribution is often a crucial part. We propose a truncation-based approximation that employs a state-space lumping scheme, aggregating states in a grid structure. The resulting approximate stationary distribution is used to iteratively refine relevant and truncate irrelevant parts of the state-space. This way, the algorithm learns a well-justified finite-state projection tailored to the stationary behavior. We demonstrate the method's applicability to a wide range of non-linear problems with complex stationary behaviors.

Dynamical systems in which local interactions among agents give rise to complex emerging phenomena are ubiquitous in nature and society. This work explores the problem of inferring the unknown interaction structure (represented as a graph) of such a system from measurements of its constituent agents or individual components (represented as nodes). We consider a setting where the underlying dynamical model is unknown and where different measurements (i.e., snapshots) may be independent (e.g., may stem from different experiments). We propose GINA (Graph Inference Network Architecture), a graph neural network (GNN) to simultaneously learn the latent interaction graph and, conditioned on the interaction graph, the prediction of a node's observable state based on adjacent vertices. GINA is based on the hypothesis that the ground truth interaction graph -- among all other potential graphs -- allows to predict the state of a node, given the states of its neighbors, with the highest accuracy. We test this hypothesis and demonstrate GINA's effectiveness on a wide range of interaction graphs and dynamical processes.

AbstractIn the recent COVID-19 pandemic, mathematical modeling constitutes an important tool to evaluate the prospective effectiveness of non-pharmaceutical interventions (NPIs) and to guide policy-making. Most research is, however, centered around characterizing the epidemic based on point estimates like the average infectiousness or the average number of contacts.In this work, we use stochastic simulations to investigate the consequences of a population’s heterogeneity regarding connectivity and individual viral load levels.Therefore, we translate a COVID-19 ODE model to a stochastic multi-agent system. We use contact networks to model complex interaction structures and a probabilistic infection rate to model individual viral load variation.We observe a large dependency of the dispersion and dynamical evolution on the population’s heterogeneity that is not adequately captured by point estimates, for instance, used in ODE models. In particular, models that assume the same clinical and transmission parameters may lead to different conclusions, depending on different types of heterogeneity in the population. For instance, the existence of hubs in the contact network leads to an initial increase of dispersion and the effective reproduction number, but to a lower herd immunity threshold (HIT) compared to homogeneous populations or a population where the heterogeneity stems solely from individual infectivity variations.Author summaryComputational modeling can support decision-making in the face of pandemics like COVID-19. Models help to understand transmission data and predict important epidemiological properties (e.g.,When will herd immunity be reached?). They can also examine the effectiveness of certain measures, and—to a limited extent—extrapolate the dynamics under specific assumptions. In all these cases, the heterogeneity of the population plays an important role. For instance, it is known that connectivity differences in (and among) age groups influence the dynamics of epidemic propagation. Here we focus on two types of differences among individuals: their social interactions and on how infectious they are. We show that only considering population averages (e.g.,What is the average number of contacts of an individual?) may lead to misleading conclusions, because the individual differences (such as those related to the epidemic(over-)dispersion) play an important role in shaping the epidemic dynamics. Many commonly used model classes, such as SEIR-type ODE compartmental models, ignore differences within a population to a large extent. This omission bears the potential of misleading conclusions.

Monte Carlo estimation in plays a crucial role in stochastic reaction networks. However, reducing the statistical uncertainty of the corresponding estimators requires sampling a large number of trajectories. We propose control variates based on the statistical moments of the process to reduce the estimators' variances. We develop an algorithm that selects an efficient subset of infinitely many control variates. To this end, the algorithm uses resampling and a redundancy-aware greedy selection. We demonstrate the efficiency of our approach in several case studies.

In the recent COVID-19 pandemic, mathematical modeling constitutes an important tool to evaluate the prospective effectiveness of non-pharmaceutical interventions (NPIs) and to guide policy-making. Most research is, however, centered around characterizing the epidemic based on point estimates like the average infectiousness or the average number of contacts. In this work, we use stochastic simulations to investigate the consequences of a population’s heterogeneity regarding connectivity and individual viral load levels. Therefore, we translate a COVID-19 ODE model to a stochastic multi-agent system. We use contact networks to model complex interaction structures and a probabilistic infection rate to model individual viral load variation. We observe a large dependency of the dispersion and dynamical evolution on the population’s heterogeneity that is not adequately captured by point estimates, for instance, used in ODE models. In particular, models that assume the same clinical and transmission parameters may lead to different conclusions, depending on different types of heterogeneity in the population. For instance, the existence of hubs in the contact network leads to an initial increase of dispersion and the effective reproduction number, but to a lower herd immunity threshold (HIT) compared to homogeneous populations or a population where the heterogeneity stems solely from individual infectivity variations.

Human mobility is the fuel of global pandemics. In this simulation study, we analyze how mobility restrictions mitigate epidemic processes and how this mitigation is influenced by the epidemic's degree of dispersion. We find that (even imperfect) mobility restrictions are generally efficient in mitigating epidemic spreading. Notably, the effectiveness strongly depends on the dispersion of the offspring distribution associated with the epidemic. We also find that mobility restrictions are useful even when the pathogen is already prevalent in the whole population. However, also a delayed implementation is more efficient in the presence of overdispersion. Conclusively, this means that strategies based on mobility restrictions, like green zones, are easier to implement when the transmission dynamics admits overdispersion (e.g., in the case of COVID-19). To study these relationships at an appropriate level of abstraction, we propose a spatial branching process model combining the flexibility of stochastic branching processes with an agent-based approach allowing a conceptualization of locality, saturation, and interaction structure.