KLSD: A Curated Kinase–Ligand Database Mapping Selectivity Landscapes and Polypharmacology

Abstract

Herein, we present a database, KLSD, which is a curated resource of 787 213 small-molecule kinase inhibitors annotated with 1.8 M quantitative activity records across 428 human kinases, emphasizing selectivity and polypharmacology. Moreover, we introduce a dual-task ensemble that simultaneously regresses pAct and computes selectivity scores. The core is a multibranch residual multilayer perceptron (MLP) whose branches are kinase-specific; this is augmented by SVM, RF, XGBoost, CNN, GCN, GAT, RGCN and VAE-enhanced graph nets. Continuous potency labels replace categorical classes to improve resolution. Benchmarked on the JAK family (JAK1/2/3, TYK2), the ensemble achieves classification accuracies of ≥0.84 for each kinase and 0.98 overall, demonstrating strong generalizability. KLSD and models are freely available at http://ai.njucm.edu.cn:8080.

1. Introduction

Protein kinases play a pivotal role in cellular signal transduction, and their proper activity is essential for maintaining normal cellular functions. However, aberrant kinase activity or overexpression can disrupt phosphorylation processes, leading to a wide range of diseases, including cancer, diabetes, and Alzheimer’s disease. Therefore, monitoring kinase activity and identifying effective kinase inhibitors are not only fundamental to basic biochemical research but also represent critical steps in the development of targeted kinase therapies. Traditional methods for evaluating kinase activity and inhibition-such as electrochemical assays, colorimetric assays, resonance light scattering, surface plasmon resonance (SPR), and fluorescence-based techniques -have been widely used, but they are often expensive and time-consuming. To address these limitations, early computational and bioinformatics approaches were explored to investigate kinase-related functions and interactions, including protein–protein interaction-based models for kinase substrate identification. With the rapid advancement of machine learning, computational approaches for predicting the activity and selectivity of kinase inhibitors have emerged as a research hotspot. Recent advances in deep learning and large language models (LLMs) have further expanded the scope of biomedical data analysis, enabling large-scale knowledge extraction, molecular representation learning, and drug discovery applications.

KLSD, is a curated repository of small-molecule kinase inhibitors (SMIs) paired with 1.8 M quantitative bioactivities spanning 428 human kinases, prioritizing selectivity and polypharmacology. Users may query (i) by compound to retrieve full-panel quantitative activity profiles across all kinases, or (ii) by any kinase pair to retrieve the complete inhibitor set, with each compound explicitly annotated as selective for one kinase or as a multitarget (polypharmacological) agent. KLSD delivers comprehensive, ready-to-use selectivity and polypharmacology data for known kinase inhibitors, streamlining rational ligand design. Yet, as a knowledge base, it is inherently retrospective and limited to known chemotypes, and lacks predictive capabilities for novel chemical matter. In the context of drug discovery and development, accurately predicting the activity and selectivity of kinase inhibitors remains a critical challenge. With advances in computational biology, cheminformatics, and machine learning, this area has seen considerable progress. Supervised learning techniques have been widely employed to predict the bioactivity of chemical compounds, including both the potency and selectivity of kinase inhibitors. These approaches typically rely on large-scale compound–activity datasets and utilize various feature descriptors to represent molecular structures and physicochemical properties. In machine learning models, compounds are generally encoded using molecular descriptors, fingerprints, or other cheminformatics representations to capture their chemical characteristics. For instance, molecular fingerprints such as extended-connectivity fingerprints (ECFPs) are widely used to represent the structural features of compounds. These representations enable the application of diverse machine learning algorithms-including support vector machines (SVM), random forests (RF), gradient boosting decision trees (GBDT), and deep neural networks-to model the complex relationships between molecular features and inhibitory activity. Well-trained models can be used to predict the inhibitory potency of novel compounds against specific kinases, as well as their selectivity across multiple kinase targets. Selectivity prediction is particularly important, as it helps identify compounds with a reduced risk of off-target effects and toxicity. Recent studies have demonstrated the effectiveness of machine learning in this domain. For example, Cao et al. developed a random forest-based model to predict inhibitor activity across multiple kinases, showing strong predictive performance on several datasets. Wen et al. applied deep learning to build a kinase activity prediction model capable of identifying novel inhibitors and guiding rational drug design. While these studies have advanced the field of kinase inhibitor prediction, machine learning methods still face significant challenges. On the one hand, traditional approaches often rely on single-feature representations such as molecular fingerprints or descriptors, failing to effectively integrate 2D/3D physicochemical properties, pharmacophore patterns, and topological structures, which limits their ability to capture the complex mechanisms of kinase inhibition. On the other hand, most methods focus solely on the “inhibitory activity” metric, lacking the capacity for joint modeling of “target selectivity”, thereby falling short of the multidimensional performance evaluation required in early stage drug development. To address these limitations, this study proposes a predictive framework that integrates molecular feature engineering with a dual-path deep learning architecture. By combining multidimensional feature fusion, dual-path collaboration, and small-sample augmentation strategies, the framework aims to achieve accurate joint prediction of both “activity” and “selectivity” of kinase inhibitors, thereby providing technical support for rational drug design.

Among kinase families, Janus kinases (JAKs) hold particular importance due to their central role in multiple signaling pathways, especially those related to immune response and hematopoiesis. The JAK family consists of four members-JAK1, JAK2, JAK3, and TYK2-which transmit extracellular signals by binding to specific cell surface receptors and activating the signal transducer and activator of transcription (STAT) family, ultimately regulating gene expression. These functions position JAKs as critical regulators in both physiological and pathological processes, including cell proliferation, differentiation, apoptosis, and inflammation. Given their essential physiological roles, selectively inhibiting hyperactive JAKs is vital for minimizing off-target effects and improving therapeutic safety. On the other hand, JAK families share high sequence similarity which makes it challenging to obtain selective JAK inhibitors (JAKi). Current JAKi are limited for their selectivity, which can cause off-target effects. This remains a major clinical challenge. Consequently, accurately predicting the activity and selectivity of kinase inhibitors against specific JAK isoforms can guide the rational design of targeted therapies, ensuring effective suppression of disease-relevant pathways while minimizing disruption to normal signaling processes. This study took the JAK family as an example and introduced a dual-task ensemble that simultaneously regresses pAct and computes selectivity scores.

2. Results and Discussion

2.1. Model Comparison and Analysis

To comprehensively evaluate the performance of the proposed model, we compared it against several traditional machine learning and deep learning approaches. Among traditional machine learning methods, we selected k-nearest neighbors (KNN), random forests (RF), XGBoost, and support vector machines (SVM), as these algorithms consistently demonstrate stable performance and computational efficiency on structured data, providing reliable baseline comparisons. , To more effectively capture the intricate structural and semantic information inherent in molecular data, we additionally introduced a convolutional neural network (CNN) based on molecular fingerprints, which excels at extracting both local and global molecular features. , Furthermore, considering that molecular data inherently possesses graph-structured characteristics, we also implemented graph neural network models such as graph convolutional networks (GCN), graph attention networks (GAT), relational graph convolutional networks (RGCN), as well as their variants incorporating variational autoencoders (GCN_VAE, GAT_VAE, RGCN_VAE). − These graph-based models directly handle the molecular graph structure, effectively capturing topological relationships among atoms and chemical bonds, making them particularly suitable for predicting properties of structurally complex molecules. By conducting a comparative analysis with these diverse approaches, we assessed the performance of our proposed dual-objective prediction modelintegrating molecular feature engineering and deep learningfrom multiple perspectives. This comprehensive comparison aims to verify the model’s effectiveness and superiority in predicting molecular bioactivity and selectivity, providing robust theoretical support for its practical application.

The dataset was partitioned into training and testing subsets using an 8:2 ratio, ensuring a consistent distribution of active and inactive compounds between subsets. Model performance on the test set was evaluated using accuracy, precision, recall, F1-score, and ROC AUC metrics, with results summarized in Table , and the overall workflow is illustrated in Figure .

1. Comparative Analysis of Model Results.

dataset	model	acc	pre	recall	F1	ROC AUC
JAK1	XGboost	0.9186	0.9382	0.9656	0.9517	0.9539
	KNN	0.8606	0.8634	0.9885	0.9217	0.9184
	SVM	0.9048	0.9318	0.9552	0.9434	0.9369
	RF	0.9143	0.9232	0.9781	0.9499	0.9516
	CNN	0.8204	0.8204	0.9534	0.9011	0.9198
	RGCN	0.8622	0.8707	0.9825	0.9232	0.9326
	GCN	0.8433	0.8433	0.9882	0.9150	0.8795
	GAT	0.8432	0.8432	0.9986	0.9149	0.8726
	RGCN_VAE	0.8683	0.8735	0.9866	0.9266	0.9170
	GCN_VAE	0.8433	0.8433	0.9881	0.9150	0.8682
	GAT_VAE	0.8432	0.8432	0.9886	0.9149	0.8735
	KLSD	0.9195	0.9604	0.9437	0.9520	0.9613
JAK2	XGBoost	0.8679	0.8904	0.9424	0.9156	0.9122
	KNN	0.7820	0.7940	0.9634	0.8705	0.8190
	SVM	0.8643	0.8859	0.9431	0.9136	0.9024
	RF	0.8620	0.8799	0.9478	0.9126	0.9077
	CNN	0.8095	0.8184	0.9616	0.8864	0.8307
	RGCN	0.8263	0.8291	0.9667	0.8926	0.8655
	GCN	0.7469	0.7469	0.9111	0.8551	0.7943
	GAT	0.7842	0.7864	0.9762	0.8711	0.8119
	RGCN_VAE	0.8162	0.8158	0.9738	0.8878	0.8443
	GCN_VAE	0.7808	0.7813	0.9810	0.8698	0.7999
	GAT_VAE	0.7475	0.7473	0.9703	0.8554	0.8209
	KLSD	0.8673	0.9085	0.9171	0.9127	0.9244
JAK3	XGBoost	0.8574	0.8691	0.9119	0.8900	0.9300
	KNN	0.7545	0.7412	0.9403	0.8284	0.8488
	SVM	0.8650	0.8802	0.9104	0.8951	0.9062
	RF	0.8527	0.8569	0.9209	0.8878	0.9166
	CNN	0.7471	0.7331	0.9686	0.8346	0.8652
	RGCN	0.7769	0.7708	0.9328	0.8441	0.8393
	GCN	0.6827	0.6805	0.9606	0.7966	0.7585
	GAT	0.7148	0.7107	0.9431	0.8105	0.7943
	RGCN_VAE	0.7035	0.6910	0.9796	0.8104	0.8140
	GCN_VAE	0.6468	0.6468	0.9331	0.7856	0.7466
	GAT_VAE	0.7044	0.7000	0.7504	0.8062	0.7500
	KLSD	0.8427	0.8478	0.8850	0.8799	0.9095
TYK2	XGboost	0.8728	0.8919	0.8943	0.8931	0.9282
	KNN	0.8470	0.8568	0.8916	0.8738	0.8847
	SVM	0.8567	0.8743	0.8862	0.8802	0.9109
	RF	0.8519	0.8713	0.8808	0.8760	0.9192
	CNN	0.7624	0.7504	0.9033	0.8203	0.8811
	RGCN	0.7230	0.6882	0.9784	0.8080	0.8855
	GCN	0.6667	0.6448	0.9811	0.7781	0.8299
	GAT	0.7665	0.7742	0.8997	0.8727	0.7861
	RGCN_VAE	0.7907	0.7667	0.9324	0.8415	0.9031
	GCN_VAE	0.7161	0.6902	0.9514	0.8000	0.8254
	GAT_VAE	0.7568	0.7461	0.8973	0.8147	0.8324
	KLSD	0.8857	0.9386	0.8652	0.9004	0.9322

1.

ROC curve of each model.

Traditional machine learning methods, such as XGBoost and random forest (RF), demonstrate robust performance on targets with sufficient sample sizes and balanced data distributionsfor instance, the JAK1 dataset with 5774 samples and the JAK2 dataset with 8438 samples. These models leverage their adaptability to structured features, achieving strong results: as shown in Table , XGBoost attained an accuracy of 0.9186 and an ROC AUC of 0.9539 on the JAK1 dataset, while RF achieved an accuracy of 0.8620 and an ROC AUC of 0.9077 on the JAK2 dataset. However, these models rely heavily on manually crafted low-dimensional descriptors, making it difficult to capture deeper structural information such as three-dimensional conformations and dynamic molecular interactions. As a result, their predictive performance declines on more structurally diverse targets like JAK3 or datasets with limited samples such as TYK2. For example, the F1 scores of XGBoost on JAK3 and TYK2 dropped to 0.8900 and 0.8931, respectively, both inferior to those achieved by KLSD on the corresponding targets.

Graph neural networks (GNNs), including GCN, GAT, and RGCN, inherently model molecular topologies and thus offer advantages in representing chemical structures. As evidenced by Table and Figure , these models excel on large-scale datasets such as JAK1. For instance, RGCN achieved a recall of 0.9825 (close to KLSD’s 0.9437) on JAK1, while GAT attained a recall of 0.9986, demonstrating its sensitivity to key structural motifs such as ATP-binding pockets. Nevertheless, their performance deteriorates on datasets with limited sample sizes and higher structural heterogeneity, such as JAK3 and TYK2. Specifically, GCN achieved an accuracy of only 0.6667 on TYK2, GAT yielded an F1 score of 0.7943 on JAK3, and GCN_VAE reached just 0.7161 in accuracy on TYK2substantially lower than KLSD’s 0.8857 on the same target.

KLSD overcomes the limitations of both traditional machine learning and GNN-based models by integrating multidimensional feature fusion and task-specific architectural design. For example, on the JAK1 dataset, KLSD achieved an F1 score of 0.9520, slightly surpassing that of XGBoost (0.9517). In low-resource settings such as TYK2, KLSD utilizes SMILES randomization for data augmentation and applies a three-block residual network optimized with a mixed MSE + Huber loss function. This strategy yields an F1 score of 0.9004 on TYK2, outperforming RF (0.8760) and SVM (0.8802).

Architecturally, KLSD employs a dual-pathway collaborative framework: a global activity prediction path captures overall molecular bioactivity trends (R ² = 0.9828), while a target-specific path focuses on subtype selectivity. A shared residual MLP layer enhances feature reuse across paths, and specialized output layers accommodate intertarget heterogeneity. The JAK1-specific model achieved the best predictive performance (R ² = 0.779, RMSE = 0.148); following data augmentation, the TYK2 model reached comparable performance to JAK3 (R ² = 0.648, RMSE = 0.184), highlighting KLSD’s effectiveness under data-scarce conditions. Across all targets, performance differences between training and test sets remained within 8%for example, the JAK3 validation set R ² was 0.678 versus 0.648 on the test set (a 4.4% difference). As shown in Figure , KLSD exhibits consistently superior ROC curves across all targets, approaching the top-left corner of the plots more closely than most GNN models, thereby demonstrating better generalization capacity.

2.2. Model Application

The KLSD kinase activity analysis platform adopts a distributed system architecture with front-end and back-end separation, enabling efficient data exchange via RESTful APIs. The front end is developed using the Vue.js framework in conjunction with Node.js to provide a responsive and interactive user interface, while the back end is implemented based on a spring boot architecture with Java as the core programming language, integrating key functional modules including molecular manipulation, feature engineering, and multimodel prediction. − A MySQL relational database is employed for structured data storage and management. Together, these components form a comprehensive computational platform that supports both kinase target comparison and compound activity prediction.

A key feature of the platform is kinase target differential activity analysis. Users can specify two kinase targets along with a range of pAct values to initiate an automated workflow for data retrieval and standardized analysis. The system performs stringent quality control on compound-related data, including outlier removal and missing value imputation. Notably, bioactivity measurements such as K _i, IC₅₀, K _d, and EC₅₀ (collectively referred to as Act) are uniformly converted to nanomolar (nM) units to ensure data consistency and comparability. Compounds explicitly labeled as “inactive” or “not active” are automatically assigned an Act value of zero to serve as baseline reference data. Upon completion of data preprocessing, the platform calculates the differential activity (ΔpAct) between the selected kinase targets and returns the corresponding compound structures, activity values, and classification results. The pAct value, a logarithmic transformation of the Act value, serves as a standardized metric for compound potency, with higher pAct indicating greater inhibitory activity. Figure shows the target difference analysis.

2.

Comparative differential analysis chart of kinase targets.

To expand the functional scope of the platform, KLSD has recently introduced a compound activity prediction module, thereby enabling end-to-end coverage from kinase target differential analysis to molecular activity screening. This extension significantly enhances the platform’s utility as a comprehensive tool for drug discovery support. The system maintains a front-end/back-end decoupled architecture, with data communication handled via RESTful APIs. The front end, built with the Vue.js framework, manages the user interface and interaction logic, while the back end, powered by Python, integrates molecular processing, feature engineering, and multimodel prediction components. Upon receiving SMILES strings from users via the front-end interface and triggering the prediction event, the front-end sends a POST request containing the SMILES input to the back-end using Axios, concurrently invoking an API call to generate molecular structure images. The back-end processes these requests, returning predictive results in JSON format alongside image URLs. The front-end subsequently parses these data, rendering result cards and molecular visualization areas. An overview of the complete system architecture is presented in Figure .

3.

KLSD compound activity prediction system architecture diagram.

The front-end implementation employs responsive design principles, relying on Vue’s reactive data properties (e.g., smilesInput, result) to ensure automatic view updates without manual DOM manipulations. The input module utilizes two-way data binding (v-model) for SMILES input and an event listener (@key up.enter) to trigger predictions upon pressing the Enter key. The card-based layout is developed using CSS Grid, enabling an adaptive grid structure that transitions seamlessly to a single-column format on mobile devices, thus ensuring cross-device compatibility. Loading states are managed through the dynamic use of v-if, controlling overlay layers and rotation animations to enhance user experience. For molecular structure visualization, images are rendered dynamically using the <img/src = “molImage”> element, where molImage refers to the URL returned from the back-end. Placeholders are shown during image loading, and CSS3 animations provide a smooth transition effect when displaying prediction results. Prediction outcomes are extracted from back-end JSON responses, displaying target-specific pIC₅₀ values and inhibitory conclusions. Conditional class bindings visually differentiate inhibitors (blue) from noninhibitors (red), while the confidence in predicted activity is represented through linear-gradient bars generated by normalizing pIC₅₀ values to a scale of [0%, 100%] using the get Confidence Width method.

For back-end implementation, molecular feature engineering and model input preparation were conducted using RDKit, which parses SMILES strings into molecular objects and computes multidimensional feature vectors. Specifically, 1024-bit Morgan fingerprints (radius = 2) and 167-dimensional MACCS fingerprints were generated, optimized for deep learning and traditional machine learning approaches, respectively. Additionally, various 2D and 3D descriptors, along with atomic charge statistics, formed a comprehensive 1040-dimensional feature vector. Molecules were also converted into graph-structured data, capturing node and edge attributes for use with graph neural network models such as GCN, GAT, and RGCN. The predictive architecture integrates both deep learning and traditional machine learning models: deep learning models include an Enhanced Neural Network (EnhancedNN), which processes the 1040-dimensional features via self-attention and multilayer perceptron modules to produce pAct values, employing BatchNorm and Dropout layers to mitigate overfitting. Graph-based models (GCN, GAT, RGCN) capture topological molecular features to output activity probabilities. Traditional machine learning models, including SVM, random forest, and XGBoost, utilize MACCS fingerprints or engineered features for binary classification tasks, whereas the CNN model employs a one-dimensional convolutional architecture based on Morgan fingerprints.

Molecular image generation begins with back-end validation of SMILES strings, followed by hydrogen atom addition and molecular mechanics-based structure optimization to obtain stable 3D conformations. Chem.Draw.MolToImage is then employed to generate standardized PNG images (300 × 200 pixels), applying consistent atomic coloring and thematic backgrounds. Front-end responsive image rendering restricts maximum heights to 300 pixels, preserving stable and consistent interface layouts. The integrated kinase inhibitor activity and selectivity prediction functionalities–combining molecular input, structure visualization, predictive outcome, and confidence-level visualizationare illustrated in Figure .

4.

Kinase inhibitor activity and selectivity prediction function interface.

2.3. Comparison with Existing Kinase Prediction Tools and Databases

KLSD is an integrated kinase informatics platform that consolidates large-scale bioactivity data management, quantitative potency prediction, and selectivity analysis within a single framework. Unlike most existing kinase-related resources, which address these aspects separately, KLSD is designed to support a continuous workflow spanning data collection, modeling, and predictive analysis.

From a data resource perspective, a key strength of KLSD lies in the provision of comprehensive, ready-to-use kinase bioactivity data. KLSD is derived from ChEMBL35, from which we curated 1,131,037 kinase-related activity records. These data are organized into 14 kinase groups, 240 families, and 339 subfamilies according to the KinBase classification system, and are delivered in a preprocessed and quality-controlled format with automated pAct calculation and standardized activity measurements. As a result, KLSD removes the substantial manual effort typically required when working directly with raw ChEMBL data, such as unit harmonization, activity normalization, and data filtering. In contrast, existing kinase databases are generally specialized in scope: KLIFS emphasizes kinase structural information and ligand–binding interactions across 297 kinases; KinaseMD focuses on mutations associated with drug response and resistance; and HKPocket characterizes kinase binding pocket druggability. While these resources are highly valuable for their respective applications, they are primarily designed around specific dimensions of kinase research and generally require additional data processing or external tools to support large-scale activity-based screening and selectivity analysis. In this context, KLSD complements such specialized databases by providing standardized bioactivity data together with integrated analytical functionality, enabling activity-centric kinase studies within a unified framework.

From a predictive tool perspective, existing kinase-related computational methods can be broadly grouped into three categories based on their modeling objectives and input data. The first category comprises kinase–substrate and phosphorylation site prediction tools (e.g., KinasePred, Phosformer-ST, NetPhosK, GPS, MusiteDeep, and KSIMC), which are primarily based on protein sequence features. For example, KSIMC employs a matrix completion approach that integrates kinase and substrate sequence similarities to predict potential kinase–substrate interactions.The second category includes approaches for inferring kinase activity states from omics data, such as KSTAR, KinomePro-DL, and MMCLK. These two classes of methods do not incorporate small-molecule chemical structures and are therefore designed to address biological kinase activity rather than inhibitor potency prediction. The third category, kinase inhibition profiling, is conceptually closer to KLSD. For example, KIPP predicts multikinase inhibition profiles. However, such methods typically output probabilistic inhibition scores rather than standardized quantitative activity measures (e.g., IC₅₀, K _i, or pAct), and they do not provide integrated database functionality or automated selectivity analysis. By contrast, KLSD uniquely integrates curated kinase bioactivity data with standardized quantitative potency prediction and built-in target differential analysis, thereby offering a unified platform that streamlines the entire workflow from data retrieval to potency prediction and selectivity assessment.

3. Conclusion

Currently, kinase database platforms have accumulated extensive kinase-related datasets encompassing structural characteristics, functional mechanisms, regulatory pathways, and biological roles of protein kinases. These data resources serve as fundamental support for research and applications in kinase-related fields. The KLSD platform developed in this study not only provides accurate data queries, in-depth analyses, and interactive visualizations tailored to researchers’ specific needs, but also innovatively integrates a machine learning-based prediction model designed for kinase inhibitors, achieving dual-objective predictions of inhibitor activity and selectivity. Through deep mining and intelligent analysis of kinase data, KLSD effectively identifies underlying patterns and critical molecular features, thereby assisting researchers in obtaining valuable insights more rapidly and accurately, enhancing overall research productivity. Additionally, the platform emphasizes rigorous data quality control and calibration procedures to ensure data accuracy and reliability, thus maximizing its value across biomedical research, drug discovery, and precision medicine applications. Nonetheless, deep learning and artificial intelligence technologies related to kinase research remain promising and expansive areas for further exploration. In the future, we will continue to explore and integrate advanced computational methods to conduct deeper analysis of large-scale kinase datasets. This effort will uncover novel patterns and hidden relationships, predict new kinase functions and biological effects, and continually expand the platform’s functionality and range of applications. These technological advancements will enable researchers to perform more comprehensive and sophisticated scientific inquiries, adapt to evolving research demands, and provide substantial support and innovative resources for the kinase research community.

4. Materials and Methods

4.1. Materials

4.1.1. Data Collection and Preprocessing

Building upon the original KLSD framework, we updated the database to incorporate data from ChEMBL version 35. Comprehensive data preprocessing and standardization were conducted to improve data coverage and quality, thereby enhancing its utility for model development and pharmacological studies. ChEMBL provides several publicly accessible web services for retrieving compound and bioactivity data. Through these services, we retrieved 79 data tables comprising 65,600,363 entries spanning genomics, protein expression, small molecules, biological systems, ontologies, and scientific literature.

Given KLSD’s emphasis on kinase-related information, we selectively extracted kinase-relevant records to construct a dedicated kinase dataset. To ensure data integrity and consistency, we implemented systematic preprocessing, including removal of anomalous entries (e.g., missing values, invalid characters, and numerical outliers), format conversion, and normalization. Compound types and relationships were standardized across datasets to improve data alignment and cross-source compatibility. Additionally, a terminology normalization protocol was established to replace common abbreviations and variants (e.g., “pip”) with standardized nomenclature (e.g., “phosphatidylinositol”) to maintain terminological consistency. Following these procedures, we obtained 787,213 kinase-related records, including 490,555 active and 291,935 inactive compounds. The overall data processing workflow is illustrated in Figure .

5.

Workflow for KLSD data processing and standardization.

In addition, as the predictive module in this study focuses primarily on the JAK family, we extracted data specific to JAK kinases (JAK1, JAK2, JAK3, and TYK2) to construct a dedicated JAK family kinase dataset. This dataset includes not only the molecular structures of JAK family kinase inhibitors (represented in SMILES format), but also a rich set of experimental activity measurementssuch as pIC₅₀ and pK _iwhich quantitatively reflect the binding affinity and inhibitory potency of the compounds against their corresponding kinase targets.

Data preprocessing forms the foundation for building an effective predictive model. To ensure high data quality, we first filtered the original dataset to retain only compounds with the organism labeled as “Homo sapiens”. All entries in the standard_units field were normalized by converting values expressed in micromolar (μM) to nanomolar (nM). Records with missing units or containing nonstandard formats such as “%”, “/min”, or “/s” as well as entries labeled with “% inhibition” were excluded from the analysis. The remaining data were then divided into four subsets corresponding to JAK1, JAK2, JAK3, and TYK2. For compounds with multiple target annotations in the pref_name field (e.g., “Cereblon/JAK1”), we assigned the entry to the first listed target (e.g., JAK1). For compounds with broad or ambiguous target labels such as “Janus Kinase (JAK),” the entries were removed to avoid uncertainty in model training. Subsequently, we performed deduplication and standardization within each subset. For entries with identical pref_name, chembl_id, and canonical_smiles values but differing standard_value, we retained the record containing a valid pchembl_value when available. If both entries had or lacked pchembl_value, their values were averaged. In cases where three or more distinct standard_value records were present, we applied the interquartile range (IQR) method to eliminate outliers before averaging. After this process, all standard_value entries were uniformly converted to a standardized format referred to as pAct.

Additionally, we filtered the dataset based on the type of activity measurement. Only entries where the standard_type corresponded to Activity K _i, pIC₅₀, pK _b, pK _d, or pEC₅₀ were retained. Entries with the standard_comments field marked as “Inhibition” but lacking quantitative values were classified as inactive and assigned a pAct value of zero. Based on established pharmacological criteria, compounds with (p)­IC₅₀ values below 10 nM were considered highly active, and this threshold was used to classify compounds into active and inactive groups. As a result of the above procedures, a total of 22,608 curated records were obtained from the ChEMBL35 database. Among these, JAK2 accounted for the largest proportion (37.3%, n = 8438), followed by JAK1 (25.6%, n = 5774), JAK3 (23.4%, n = 5294), and TYK2 (13.7%, n = 3102). In terms of activity types, pIC₅₀ values were the most common (n = 17,279), followed by pK _i (n = 3982) and pK _d (n = 881). We further analyzed the distribution characteristics of activity values across different JAK subtypes and activity categories (pIC₅₀, pK _i, pK _d, pEC₅₀) to better understand subtype-specific patterns within the dataset. The statistical analysis of the JAK family dataset constructed in this study is illustrated in Figure .

6.

Statistical analysis of the JAK family dataset.

4.1.2. Molecular Descriptors and Fingerprints

Molecular descriptors and fingerprints effectively represent structural features of molecules in drug screening and bioactivity prediction, providing robust inputs for predictive models. In this study, we employed the concise and efficient SMILES notation to represent molecular structures, and performed rigorous data cleaning and standardization procedures to ensure data quality and enhance model accuracy. , Experimental bioactivity metrics, such as pIC₅₀ and pK _i, directly reflect molecular biological activity and serve as critical benchmarks for identifying promising drug candidates. Therefore, SMILES-based molecular fingerprinting was used to extract key structural features of drug-like molecules, facilitating analysis of inhibitory characteristics against JAK family kinases, thereby guiding drug design and optimization.

Specifically, we utilized software packages including RDKit and Open Babel to generate 2D and 3D molecular structures and associated descriptors. Among these, 2D fingerprints such as MACCS and ECFP efficiently capture molecular topology and functional group information, while 3D descriptors characterize molecular spatial properties, including molecular volume, surface area, and polarity. , Energy minimization and conformational optimization of molecular structures were conducted using molecular mechanics and quantum chemical calculations to obtain stable and reliable 3D configurations. Combining both 2D and 3D descriptors further enhanced the precision of bioactivity prediction.

4.2. Method

4.2.1. Construction of Prediction Model

The prediction model for kinase inhibitor activity and selectivity was developed based on a dual-path deep learning framework, designed to simultaneously predict inhibitory potency and selectivity against the Janus kinase (JAK) family membersJAK1, JAK2, JAK3, and TYK2from molecular structural features. The model architecture is illustrated in Figure .

7.

Architecture of the kinase inhibitor activity and selectivity prediction model.

The model input stage begins with a shared core data preprocessing workflow for both the overall activity prediction and the target-specific selectivity modeling tasks, while targeted adjustments are made in the feature engineering and data augmentation stages. Input molecules were represented by SMILES strings, validated and standardized using RDKit. Invalid or duplicate structures were removed, aromatic features were standardized (e.g., normalizing benzene representation to “c1ccccc1”), and canonical SMILES were generated. Subsequently, the overall activity prediction model employed feature vectors composed of 1024-bit Morgan fingerprints (radius = 2) combined with 16 two-dimensional (2D) molecular descriptors, such as molecular weight, Log P, and topological polar surface area (TPSA). In contrast, the subtype-specific selectivity model utilized 2048-bit Morgan fingerprints (radius = 3) enriched with additional pharmacophore descriptors (e.g., number of hydrogen-bond donors/acceptors) and three-dimensional (3D) descriptors including principal moments of inertia (PMI1–3) and radius of gyration. Furthermore, a four-dimensional one-hot vector encoding target subtype was appended to the feature vector. A unified, multistep standardization procedure was adopted for processing activity values. Initially, logarithmically transformed activity values were screened for outliers using an interquartile range (IQR) criterion (Q1-2.5 × IQR to Q3 + 2.5 × IQR). After removing outliers, the data underwent a dynamic Box-Cox transformation (with λ optimized per dataset) and scaling with RobustScaler (5th–95th percentile range), resulting in activity values approximating a normal distribution. The processed data were then stratified and partitioned into training, test, and validation sets at an 8:1:1 ratio, ensuring distributional consistency verified by the Kolmogorov–Smirnov (KS) test (p > 0.05).

After the above standardization and feature engineering processes, the model is able to perform dual-path collaborative prediction based on high-quality data. The overall activity prediction pathway focuses on evaluating the comprehensive inhibitory activity of compounds against the JAK family, while the target-specific selectivity pathway delves into the activity differences among individual JAK family members. These two pathways achieve functional complementarity through distinct network architecture designs. For the overall activity prediction path, the input vector underwent attention-weighting via a 512-dimensional attention layer, followed by nonlinear transformations through three fully connected layers (dimensions: 1536 → 768 → 384) employing GELU activation and a dropout rate of 0.4, ultimately yielding a single predicted activity value. The attention module applied two linear transformations followed by Sigmoid normalization to generate feature importance weights, effectively emphasizing key molecular features. In addressing the imbalance within the TYK2 data subset, the subtype-specific selectivity prediction path utilized SMILES randomization to generate equivalent molecular representations (e.g., transforming “CCO” into “OCC”), achieving a 5-fold data augmentation. Each subtype dataset was independently fed into the model. A shared residual multilayer perceptron (MLP) layer (1024-dimensional, pretrained on JAK1 data) captured cross-target common features, such as hydrophobic interactions within ATP-binding pockets. Target-specific prediction heads were then independently trained: JAK1, JAK2, and JAK3 employed two-layer residual blocks, while TYK2 utilized deeper three-layer residual blocks. Residual block designs included skip connections (x + residual) and LayerNorm layers, facilitating stable training of deeper network architectures. Model optimization was performed using the AdamW optimizer, with learning rates set at 1 × 10^–4 for the attention module and 3 × 10^–4 for the main network in the overall activity model, and at 5 × 10^–4 for general subtype-specific models, with a higher learning rate (1 × 10^–3) for TYK2. Adaptive loss functions enhanced robustness against outliers: Huber loss (δ = 1.0) was applied in the overall activity model, whereas TYK2 employed an adaptive blended loss (MSE + Huber) and the other subtypes employed Smooth L1 loss (δ = 1.0). An early stopping criterion was implemented to prevent overfitting, halting training when validation loss ceased improving for 30 consecutive epochs. Ultimately, the models jointly produced predicted pAct values for each kinase subtype.

Model predictive performance was assessed using the coefficient of determination (R ²) and mean absolute error (MAE). R ² quantifies the global fit of activity predictions, indicating the proportion of variance in experimental data explained by the model, while MAE represents the average absolute difference between predicted and experimental values, directly reflecting prediction accuracy. An R ² value approaching 1 indicates strong predictive capability in capturing activity trends, whereas a lower MAE signifies smaller discrepancies between predictions and actual observations. Collectively, these metrics ensure the model’s practical utility in drug discovery, providing accurate activity predictions and reliable selectivity assessments that effectively prioritize compounds for subsequent experimental validation.

4.2.2. Model Evaluation Metrics

To evaluate the practical utility of the model in drug discovery scenarios, a multidimensional assessment framework was established in this study to systematically validate model performance from perspectives such as overall fitting ability and prediction bias quantification. Candidate molecules identified through postprocessing of multitarget predicted activity values can be prioritized for subsequent experimental validation, significantly enhancing the efficiency of drug discovery efforts.

(1) Standardization of activity values. During the preprocessing of activity values, outliers were identified and removed using an interquartile range (IQR) criterion, followed by dynamic Box-Cox transformation and scaling with RobustScaler. The mathematical formulations are as follows:

IQR calculation:

$$\mathrm{I}\mathrm{Q}\mathrm{R}=Q_3-Q_1$$ 1

range for outlier removal:

$$\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}=[Q_1-2.5\times \mathrm{I}\mathrm{Q}\mathrm{R},Q_3+2.5\times \mathrm{I}\mathrm{Q}\mathrm{R}]$$ 2

Box-Cox transformation:

$$\mathit{y}(\lambda )=\{\begin{array}{ll}\frac{y^{\lambda }-1}{\lambda },& \lambda \ne 0\\ \log (\mathit{y}),& \lambda =0\end{array}$$ 3

where the parameter λ is optimized based on data.

RobustScaler transformation:

$$\mathrm{R}\mathrm{o}\mathrm{b}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{S}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{r}(\mathit{x})=\frac{\mathit{x}-\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}(\mathit{x})}{\mathrm{I}\mathrm{Q}\mathrm{R}(\mathit{x})}$$ 4

where IQR represents the interquartile range and median­(x) is the median of the data.

(2) Model output metrics. The coefficient of determination R ² quantifies the global goodness of fit for predictions:

$$R^2=1-\frac{\sum _{i=1}^n{(y_i-{ŷ}_i)}^2}{\sum _{i=1}^n{(y_i-\mathrm{y̅})}^2}$$ 5

where y _i is the observed value, ŷ _i is the predicted value, and y̅ is the mean observed value.

The mean absolute error (MAE) is calculated as

$$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{n}\sum _{\mathit{i}=1}^n|y_i-{ŷ}_i|$$ 6

The root-mean-square error (RMSE) is expressed as

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}\sum _{\mathit{i}=1}^n{(y_i-{ŷ}_i)}^2}$$ 7

The Huber loss function is defined by

$$L_{\delta }(y,ŷ)=\{\begin{array}{ll}\frac{1}{2}{(y-ŷ)}^2,& |y-ŷ|\le \delta \\ \delta (|y-ŷ|-\frac{1}{2}\delta ),& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}$$ 8

where δ is a threshold typically set to 1.

The Smooth L1 loss function (similar to the Huber loss function) is defined by

$$\mathrm{S}\mathrm{m}\mathrm{o}\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}(y,ŷ)=\{\begin{array}{ll}0.5{(y-ŷ)}^2,& |y-ŷ|<1\\ |y-ŷ|-0.5,& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}$$ 9

(3) Optimizer and training procedure. During training, the AdamW optimizer was utilized with the following parameter update rule:

$${\theta }_t={\theta }_{\mathit{t}-1}-\eta \frac{m_t}{\sqrt{v_t}+\epsilon }$$ 10

where m _t denotes the first moment (mean) estimate of gradients, v _t is the second moment (variance) estimate, η is the learning rate, and ε is a small constant introduced to prevent division by zero.

The KLSD database, available at http://ai.njucm.edu.cn:8080, is maintained by Nanjing University of Chinese Medicine under a time-based access policy to ensure data integrity, confidentiality, and availability. Access is permitted daily from 06:30 to 23:30 Beijing time; external access is restricted outside these hours. KLSD is guaranteed to remain freely available for at least 2 years and supports major browsers on Windows and macOS. The full implementation of the study, including the architecture code of the KLSD platform, scripts for model training, code for baseline models, all raw and processed data, and the associated training, validation, and test datasets are available at the following GitHub repository: https://github.com/ccheng0113ooo-rgb/KLSD. Small-molecule kinase inhibitor datasets, including the JAK family benchmark data, were obtained from ChEMBL (https://www.ebi.ac.uk/chembl/). The processed JAK family datasets are available on GitHub, with the specific download links accessible at https://github.com/ccheng0113ooo-rgb/KLSD/tree/main/finaldata.

∇.

C.C., Y.Y., and H.L. contributed equally to this work. The manuscript was written with contributions from all authors. All of the authors approved submission of the final version of the manuscript.

The authors declare no competing financial interest.