Molecular Representation Research Articles

A deep learning framework for predicting molecular property based on multi-type features fusion

Extracting expressive molecular features is essential for molecular property prediction. Sequence-based representation is a common representation of molecules, which ignores the structure information of molecules. While molecular graph representation has a weak ability in expressing the 3D structure. In this article, we try to make use of the advantages of different type representations simultaneously for molecular property prediction. Thus, we propose a fusion model named DLF-MFF, which integrates the multi-type molecular features. Specifically, we first extract four different types of features from molecular fingerprints, 2D molecular graph, 3D molecular graph and molecular image. Then, in order to learn molecular features individually, we use four essential deep learning frameworks, which correspond to four distinct molecular representations. The final molecular representation is created by integrating the four feature vectors and feeding them into prediction layer to predict molecular property. We compare DLF-MFF with 7 state-of-the-art methods on 6 benchmark datasets consisting of multiple molecular properties, the experimental results show that DLF-MFF achieves state-of-the-art performance on 6 benchmark datasets. Moreover, DLF-MFF is applied to identify potential anti-SARS-CoV-2 inhibitor from 2500 drugs. We predict probability of each drug being inferred as a 3CL protease inhibitor and also calculate the binding affinity scores between each drug and 3CL protease. The results show that DLF-MFF product better performance in the identification of anti-SARS-CoV-2 inhibitor. This work is expected to offer novel research perspectives for accurate prediction of molecular properties and provide valuable insights into drug repurposing for COVID-19.

Can Pretrained Models Really Learn Better Molecular Representations for AI-Aided Drug Discovery?

Self-supervised pretrained models are gaining increasingly more popularity in AI-aided drug discovery, leading to more and more pretrained models with the promise that they can extract better feature representations for molecules. Yet, the quality of learned representations has not been fully explored. In this work, inspired by the two phenomena of Activity Cliffs (ACs) and Scaffold Hopping (SH) in traditional Quantitative Structure-Activity Relationship analysis, we propose a method named Representation-Property Relationship Analysis (RePRA) to evaluate the quality of the representations extracted by the pretrained model and visualize the relationship between the representations and properties. The concepts of ACs and SH are generalized from the structure-activity context to the representation-property context, and the underlying principles of RePRA are analyzed theoretically. Two scores are designed to measure the generalized ACs and SH detected by RePRA, and therefore, the quality of representations can be evaluated. In experiments, representations of molecules from 10 target tasks generated by 7 pretrained models are analyzed. The results indicate that the state-of-the-art pretrained models can overcome some shortcomings of canonical Extended-Connectivity FingerPrints, while the correlation between the basis of the representation space and specific molecular substructures are not explicit. Thus, some representations could be even worse than the canonical fingerprints. Our method enables researchers to evaluate the quality of molecular representations generated by their proposed self-supervised pretrained models. And our findings can guide the community to develop better pretraining techniques to regularize the occurrence of ACs and SH.

Molecular toxicity of nitrobenzene derivatives to tetrahymena pyriformis based on SMILES descriptors using Monte Carlo, docking, and MD simulations

It is challenging to model the toxicity of nitroaromatic compounds due to limited experimental data. Nitrobenzene derivatives are commonly used in industry and can lead to environmental contamination. Extensive research, including several QSPR studies, has been conducted to understand their toxicity. Predictive QSPR models can help improve chemical safety, but their limitations must be considered, and the molecular factors affecting toxicity should be carefully investigated. The latest QSPR methods, molecular modeling techniques, machine learning algorithms, and computational chemistry tools are essential for developing accurate and robust models. In this work, we used these methods to study a series of fifty compounds derived from nitrobenzene. The Monte Carlo approach was used for QSPR modeling by applying the SMILES molecular structure representation and optimal molecular descriptors. The correlation ideality index (CII) and correlation contradiction index (CCI) were further introduced as validation parameters to estimate the developed models' predictive ability. The statistical quality of the CII models was better than those without CII. The best QSPR model with the following statistical parameters (Split-3): (R2 = 0.968, CCC = 0.984, IIC = 0.861, CII = 0.979, Q2 = 0.954, QF12 = 0.946, QF22 = 0.938, QF32 = 0.947, Rm2 = 0.878, RMSE = 0.187, MAE = 0.151, FTraining = 390, FInvisible = 218, FCalibration = 240, RTest2 = 0.905) was selected to generate the studied promoters with increasing and decreasing activity.

Is fragment-based graph a better graph-based molecular representation for drug design? A comparison study of graph-based models

Graph Neural Networks (GNNs) have gained significant traction in various sectors of AI-driven drug design. Over recent years, the integration of fragmentation concepts into GNNs has emerged as a potent strategy to augment the efficacy of molecular generative models. Nonetheless, challenges such as symmetry breaking and potential misrepresentation of intricate cycles and undefined functional groups raise questions about the superiority of fragment-based graph representation over traditional methods. In our research, we undertook a rigorous evaluation, contrasting the predictive prowess of eight models—developed using deep learning algorithms—across 12 benchmark datasets that span a range of properties. These models encompass established methods like GCN, AttentiveFP, and D-MPNN, as well as innovative fragment-based representation techniques. Our results indicate that fragment-based methodologies, notably PharmHGT, significantly improve model performance and interpretability, particularly in scenarios characterized by limited data availability. However, in situations with extensive training, fragment-based molecular graph representations may not necessarily eclipse traditional methods. In summation, we posit that the integration of fragmentation, as an avant-garde technique in drug design, harbors considerable promise for the future of AI-enhanced drug design.

Development of QSAR model based on Monte Carlo optimization for predicting GABAA receptor binding of newly emerging benzodiazepines.

Benzodiazepines and their derivatives belong to a category of new psychoactive substances that have been introduced into the continually expanding illicit market. However, there is a notable absence of available pharmacological data for these substances. To gain a deeper understanding of their pharmacology, we employed the Monte Carlo optimization conformation-independent method as a tool for developing QSAR models. These models were built using optimal molecular descriptors derived from both SMILES notation and molecular graph representations. The resulting QSAR model demonstrated robustness and a high degree of predictability, proving to be very reliable. Moreover, we were able to identify specific molecular fragments that exerted both positive and negative effects on binding activity. This discovery paves the way for the swift prediction of binding activity for emerging benzodiazepines, offering a faster and more cost-effective alternative to traditional in vitro/in vivo analyses.

Influence of Model Structures on Predictors of Protein Stability Changes from Single-Point Mutations.

Missense variation in genomes can affect protein structure stability and, in turn, the cell physiology behavior. Predicting the impact of those variations is relevant, and the best-performing computational tools exploit the protein structure information. However, most of the current protein sequence variants are unresolved, and comparative or ab initio tools can provide a structure. Here, we evaluate the impact of model structures, compared to experimental structures, on the predictors of protein stability changes upon single-point mutations, where no significant changes are expected between the original and the mutated structures. We show that there are substantial differences among the computational tools. Methods that rely on coarse-grained representation are less sensitive to the underlying protein structures. In contrast, tools that exploit more detailed molecular representations are sensible to structures generated from comparative modeling, even on single-residue substitutions.

MARS Plus: An Improved Molecular Design Tool for Complex Compounds Involving Ionic, Stereo, and Cis-Trans Isomeric Structures.

MARS (Molecular Assembling and Representation Suite) (Hsu et al. J. Chem. Inf. Model. 2019, 59, 3703-3713) is a toolbox for the molecular design of organic molecules. MARS uses integer arrays to represent the elements and connectivity between elements of a molecule. It provides a collection of operations to manipulate the elemental composition and connectivity of a molecule (or a pair of molecules), enabling the creation of novel chemical compounds. In this work, the original MARS is extended to handle complex molecular structures, including geometric (cis-trans) isomers, stereo isomers, cyclic compounds, and ionic species. The extended version of MARS, referred to as MARS+, has a more comprehensive coverage of the chemical space and therefore can explore molecules with a greater chemical and physical diversity. Compared to other molecular design tools, MARS+ is designed to perform all possible manipulations on a given molecule or a pair of molecules. Molecular structure manipulation can be conducted in either a controlled or a random fashion. Furthermore, every structure manipulation has a counterpart so that the operation can be reversed. Nearly any possible chemical structure can be generated with MARS+ via a combination of molecular operations. The capabilities of MARS+ are examined by the design of new ionic liquids (ILs). The results show that MARS+ is a useful tool for computer-aided molecular design (CAMD) and molecular structure enumeration.

Shaping the Water-Harvesting Behavior of Metal-Organic Frameworks Aided by Fine-Tuned GPT Models.

We construct a data set of metal-organic framework (MOF) linkers and employ a fine-tuned GPT assistant to propose MOF linker designs by mutating and modifying the existing linker structures. This strategy allows the GPT model to learn the intricate language of chemistry in molecular representations, thereby achieving an enhanced accuracy in generating linker structures compared with its base models. Aiming to highlight the significance of linker design strategies in advancing the discovery of water-harvesting MOFs, we conducted a systematic MOF variant expansion upon state-of-the-art MOF-303 utilizing a multidimensional approach that integrates linker extension with multivariate tuning strategies. We synthesized a series of isoreticular aluminum MOFs, termed Long-Arm MOFs (LAMOF-1 to LAMOF-10), featuring linkers that bear various combinations of heteroatoms in their five-membered ring moiety, replacing pyrazole with either thiophene, furan, or thiazole rings or a combination of two. Beyond their consistent and robust architecture, as demonstrated by permanent porosity and thermal stability, the LAMOF series offers a generalizable synthesis strategy. Importantly, these 10 LAMOFs establish new benchmarks for water uptake (up to 0.64 g g-1) and operational humidity ranges (between 13 and 53%), thereby expanding the diversity of water-harvesting MOFs.

Molecular data representation based on gene embeddings for cancer drug response prediction

Cancer drug response prediction is a crucial task in precision medicine, but existing models have limitations in effectively representing molecular profiles of cancer cells. Specifically, when these models represent molecular omics data such as gene expression, they employ a one-hot encoding-based approach, where a fixed gene set is selected for all samples and omics data values are assigned to specific positions in a vector. However, this approach restricts the utilization of embedding-vector-based methods, such as attention-based models, and limits the flexibility of gene selection. To address these issues, our study proposes gene embedding-based fully connected neural networks (GEN) that utilizes gene embedding vectors as input data for cancer drug response prediction. The GEN allows for the use of embedding-vector-based architectures and different gene sets for each sample, providing enhanced flexibility. To validate the efficacy of GEN, we conducted experiments on three cancer drug response datasets. Our results demonstrate that GEN outperforms other recently developed methods in cancer drug prediction tasks and offers improved gene representation capabilities. All source codes are available at https://github.com/DMCB-GIST/GEN/.

Dataset Design for Building Models of Chemical Reactivity.

Models can codify our understanding of chemical reactivity and serve a useful purpose in the development of new synthetic processes via, for example, evaluating hypothetical reaction conditions or in silico substrate tolerance. Perhaps the most determining factor is the composition of the training data and whether it is sufficient to train a model that can make accurate predictions over the full domain of interest. Here, we discuss the design of reaction datasets in ways that are conducive to data-driven modeling, emphasizing the idea that training set diversity and model generalizability rely on the choice of molecular or reaction representation. We additionally discuss the experimental constraints associated with generating common types of chemistry datasets and how these considerations should influence dataset design and model building.

Modelling local and general quantum mechanical properties with attention-based pooling

Atom-centred neural networks represent the state-of-the-art for approximating the quantum chemical properties of molecules, such as internal energies. While the design of machine learning architectures that respect chemical principles has continued to advance, the final atom pooling operation that is necessary to convert from atomic to molecular representations in most models remains relatively undeveloped. The most common choices, sum and average pooling, compute molecular representations that are naturally a good fit for many physical properties, while satisfying properties such as permutation invariance which are desirable from a geometric deep learning perspective. However, there are growing concerns that such simplistic functions might have limited representational power, while also being suboptimal for physical properties that are highly localised or intensive. Based on recent advances in graph representation learning, we investigate the use of a learnable pooling function that leverages an attention mechanism to model interactions between atom representations. The proposed pooling operation is a drop-in replacement requiring no changes to any of the other architectural components. Using SchNet and DimeNet++ as starting models, we demonstrate consistent uplifts in performance compared to sum and mean pooling and a recent physics-aware pooling operation designed specifically for orbital energies, on several datasets, properties, and levels of theory, with up to 85% improvements depending on the specific task.

On the Development of Descriptor-Based Machine Learning Models for Thermodynamic Properties: Part 1—From Data Collection to Model Construction: Understanding of the Methods and Their Effects

In the present work, a multi-angle approach is adopted to develop two ML-QSPR models for the prediction of the enthalpy of formation and the entropy of molecules, in their ideal gas state. The molecules were represented by high-dimensional vectors of structural and physico-chemical characteristics (i.e., descriptors). In this sense, an overview is provided of the possible methods that can be employed at each step of the ML-QSPR procedure (i.e., data preprocessing, dimensionality reduction and model construction) and an attempt is made to increase the understanding of the effects related to a given choice or method on the model performance, interpretability and applicability domain. At the same time, the well-known OECD principles for the validation of (Q)SAR models are also considered and addressed. The employed data set is a good representation of two common problems in ML-QSPR modeling, namely the high-dimensional descriptor-based representation and the high chemical diversity of the molecules. This diversity effectively impacts the subsequent applicability of the developed models to a new molecule. The data set complexity is addressed through customized data preprocessing techniques and genetic algorithms. The former improves the data quality while limiting the loss of information, while the latter allows for the automatic identification of the most important descriptors, in accordance with a physical interpretation. The best performances are obtained with Lasso linear models (MAE test = 25.2 kJ/mol for the enthalpy and 17.9 J/mol/K for the entropy). Finally, the overall developed procedure is also tested on various enthalpy and entropy related data sets from the literature to check its applicability to other problems and competing performances are obtained, highlighting that different methods and molecular representations can lead to good performances.

Differential CD38 Expression on Daratumumab-Naïve Vs. Resistant Multiple Myeloma - a Direct Stochastic Optical Reconstruction Microscopy ( dSTORM) Analysis in 53 Patients

Background: The anti-CD38 antibody Daratumumab (Dara) has become the new standard of care for the treatment of newly-diagnosed and relapsed/ refractory multiple myeloma (MM). At the same time, Dara resistance remains poorly understood. Potential resistance mechanisms have been proposed and range from intrinsic downregulation in CD38 dim MM cells to genetic alterations and splicing variants which may affect CD38 epitope expression (Adamia, ASH 2022). Here, we used super-resolution microscopy for high-precision quantification of CD38 in MM and piloted this approach in a large cohort of 53 MM patients with vs. without Dara resistance. Methods: Direct stochastic optical reconstruction microscopy ( dSTORM) is a novel imaging modality based on photoswitching of certain fluorophores coupled to antibodies for molecular representation of surface molecules at unprecedented levels of resolution (<20 nM). To correlate CD38 expression data with clinical outcome, we divided our cohort based on treatment status into two groups: Dara-naïve (n = 31) vs. Dara-resistant (n = 22) patients. Imaging data was complemented by phenotypic assessment using a well-established multi-epitope anti-CD38 antibody (Cytognos), as per standard-of-care. Results: Mean CD38 receptor density as measured with Dara via dSTORM was significantly higher in Dara-naïve (of 48.81 ± 29.65 receptors/µm²) vs. Dara-resistant patients (15.52 ± 3.9 receptors/µm², p = 0.016). This finding was corroborated by multi-epitope phenotypic analysis which revealed mean CD38 receptor expression levels of 32.09 ± 13.14 receptors/µm² in Dara-naïve vs. 21.65 ± 17.73 receptors/µm² in Dara-resistant patients ( p = 0.05). Lower CD38 expression via dSTORM vs. multi-epitope anti-CD38 staining was consistently observed in our cohort and may reflect ongoing blockage of the receptor binding site by Dara in cases with shorter time to last Dara application (median 53 days, range 1-189 days). As there was no correlation with the interval between Dara application and sampling however, splice-site mutations affecting CD38 receptor accessibility as proposed by Adamia et al. may depict a more relevant mechanism of resistance in our Dara-resistant cohort. In a next step, we therefore investigated alternative epitope binding in Dara-resistant patients. To this end, we investigated antibody binding by dSTORM imaging for isatuximab (Isa), an alternative CD38-directed monoclonal antibody that binds to a location in the C-terminal portion of the CD38 extracellular domain which is entirely different from the Dara epitope. Interestingly, there was a trend for increased detection rate with Isa for mean CD38 receptor expression in Dara-resistant patients (22.2 ± 4.6 receptors/µm²) vs. Dara (15.5 ± 3.9 receptors/µm², p = 0.1) and on par with CD38 receptor expression. No data was available to assess Dara binding in Isa pre-exposed patients. Conclusion: Innovative microscopy using dSTORM allows for high-resolution epitope quantification at unprecedented granularity. While this finding enables visual proof of differential epitope binding of Dara vs. Isa on CD38 in a clinically feasible setting, correlation studies to determine the clinical relevance of this observation remain ongoing.

DeepFPlearn +: enhancing toxicity prediction across the chemical universe using graph neural networks.

Sophisticated approaches for the in silico prediction of toxicity are required to support the risk assessment of chemicals. The number of chemicals on the global chemical market and the speed of chemical innovation stand in massive contrast to the capacity for regularizing chemical use. We recently proved our ready-to-use application deepFPlearn as a suitable approach for this task. Here, we present its extension deepFPlearn+ incorporating (i) a graph neural network to feed our AI with a more sophisticated molecular structure representation and (ii) alternative train-test splitting strategies that involve scaffold structures and the molecular weights of chemicals. We show that the GNNs outperform the previous model substantially and that our models can generalize on unseen data even with a more robust and challenging test set. Therefore, we highly recommend the application of deepFPlearn+ on the chemical inventory to prioritize chemicals for experimental testing or any chemical subset of interest in monitoring studies. The software is compatible with python 3.6 or higher, and the source code can be found on our GitHub repository: https://github.com/yigbt/deepFPlearn. The data underlying this article are available in Zenodo, and can be accessed with the link below: https://zenodo.org/record/8146252. Detailed installation guides via Docker, Singularity, and Conda are provided within the repository for operability across all operating systems.

BatmanNet: bi-branch masked graph transformer autoencoder for molecular representation.

Although substantial efforts have been made using graph neural networks (GNNs) for artificial intelligence (AI)-driven drug discovery, effective molecular representation learning remains an open challenge, especially in the case of insufficient labeled molecules. Recent studies suggest that big GNN models pre-trained by self-supervised learning on unlabeled datasets enable better transfer performance in downstream molecular property prediction tasks. However, the approaches in these studies require multiple complex self-supervised tasks and large-scale datasets , which are time-consuming, computationally expensive and difficult to pre-train end-to-end. Here, we design a simple yet effective self-supervised strategy to simultaneously learn local and global information about molecules, and further propose a novel bi-branch masked graph transformer autoencoder (BatmanNet) to learn molecular representations. BatmanNet features two tailored complementary and asymmetric graph autoencoders to reconstruct the missing nodes and edges, respectively, from a masked molecular graph. With this design, BatmanNet can effectively capture the underlying structure and semantic information of molecules, thus improving the performance of molecular representation. BatmanNet achieves state-of-the-art results for multiple drug discovery tasks, including molecular properties prediction, drug-drug interaction and drug-target interaction, on 13 benchmark datasets, demonstrating its great potential and superiority in molecular representation learning.

From intuition to AI: evolution of small molecule representations in drug discovery.

Within drug discovery, the goal of AI scientists and cheminformaticians is to help identify molecular starting points that will develop into safe and efficacious drugs while reducing costs, time and failure rates. To achieve this goal, it is crucial to represent molecules in a digital format that makes them machine-readable and facilitates the accurate prediction of properties that drive decision-making. Over the years, molecular representations have evolved from intuitive and human-readable formats to bespoke numerical descriptors and fingerprints, and now to learned representations that capture patterns and salient features across vast chemical spaces. Among these, sequence-based and graph-based representations of small molecules have become highly popular. However, each approach has strengths and weaknesses across dimensions such as generality, computational cost, inversibility for generative applications and interpretability, which can be critical in informing practitioners' decisions. As the drug discovery landscape evolves, opportunities for innovation continue to emerge. These include the creation of molecular representations for high-value, low-data regimes, the distillation of broader biological and chemical knowledge into novel learned representations and the modeling of up-and-coming therapeutic modalities.

FormulationAI: a novel web-based platform for drug formulation design driven by artificial intelligence.

Today, pharmaceutical industry faces great pressure to employ more efficient and systematic ways in drug discovery and development process. However, conventional formulation studies still strongly rely on personal experiences by trial-and-error experiments, resulting in a labor-consuming, tedious and costly pipeline. Thus, it is highly required to develop intelligent and efficient methods for formulation development to keep pace with the progress of the pharmaceutical industry. Here, we developed a comprehensive web-based platform (FormulationAI) for in silico formulation design. First, the most comprehensive datasets of six widely used drug formulation systems in the pharmaceutical industry were collected over 10years, including cyclodextrin formulation, solid dispersion, phospholipid complex, nanocrystals, self-emulsifying and liposome systems. Then, intelligent prediction and evaluation of 16 important properties from the six systems were investigated and implemented by systematic study and comparison of different AI algorithms and molecular representations. Finally, an efficient prediction platform was established and validated, which enables the formulation design just by inputting basic information of drugs and excipients. FormulationAI is the first freely available comprehensive web-based platform, which provides a powerful solution to assist the formulation design in pharmaceutical industry. It is available at https://formulationai.computpharm.org/.

GAABind: a geometry-aware attention-based network for accurate protein-ligand binding pose and binding affinity prediction.

Protein-ligand interactions are increasingly profiled at high-throughput, playing a vital role in lead compound discovery and drug optimization. Accurate prediction of binding pose and binding affinity constitutes a pivotal challenge in advancing our computational understanding of protein-ligand interactions. However, inherent limitations still exist, including high computational cost for conformational search sampling in traditional molecular docking tools, and the unsatisfactory molecular representation learning and intermolecular interaction modeling in deep learning-based methods. Here we propose a geometry-aware attention-based deep learning model, GAABind, which effectively predicts the pocket-ligand binding pose and binding affinity within a multi-task learning framework. Specifically, GAABind comprehensively captures the geometric and topological properties of both binding pockets and ligands, and employs expressive molecular representation learning to model intramolecular interactions. Moreover, GAABind proficiently learns the intermolecular many-body interactions and simulates the dynamic conformational adaptations of the ligand during its interaction with the protein through meticulously designed networks. We trained GAABind on the PDBbindv2020 and evaluated it on the CASF2016 dataset; the results indicate that GAABind achieves state-of-the-art performance in binding pose prediction and shows comparable binding affinity prediction performance. Notably, GAABind achieves a success rate of 82.8% in binding pose prediction, and the Pearson correlation between predicted and experimental binding affinities reaches up to 0.803. Additionally, we assessed GAABind's performance on the severe acute respiratory syndrome coronavirus 2 main protease cross-docking dataset. In this evaluation, GAABind demonstrates a notable success rate of 76.5% in binding pose prediction and achieves the highest Pearson correlation coefficient in binding affinity prediction compared with all baseline methods.

CrossFuse-XGBoost: accurate prediction of the maximum recommended daily dose through multi-feature fusion, cross-validation screening and extreme gradient boosting.

In the drug development process, approximately 30% of failures are attributed to drug safety issues. In particular, the first-in-human (FIH) trial of a new drug represents one of the highest safety risks, and initial dose selection is crucial for ensuring safety in clinical trials. With traditional dose estimation methods, which extrapolate data from animals to humans, catastrophic events have occurred during Phase I clinical trials due to interspecies differences in compound sensitivity and unknown molecular mechanisms. To address this issue, this study proposes a CrossFuse-extreme gradient boosting (XGBoost) method that can directly predict the maximum recommended daily dose of a compound based on existing human research data, providing a reference for FIH dose selection. This method not only integrates multiple features, including molecular representations, physicochemical properties and compound-protein interactions, but also improves feature selection based on cross-validation. The results demonstrate that the CrossFuse-XGBoost method not only improves prediction accuracy compared to that of existing local weighted methods [k-nearest neighbor (k-NN) and variable k-NN (v-NN)] but also solves the low prediction coverage issue of v-NN, achieving full coverage of the external validation set and enabling more reliable predictions. Furthermore, this study offers a high level of interpretability by identifying the importance of different features in model construction. The 241 features with the most significant impact on the maximum recommended daily dose were selected, providing references for optimizing the structure of new compounds and guiding experimental research. The datasets and source code are freely available at https://github.com/cqmu-lq/CrossFuse-XGBoost.

A knowledge-guided pre-training framework for improving molecular representation learning

Learning effective molecular feature representation to facilitate molecular property prediction is of great significance for drug discovery. Recently, there has been a surge of interest in pre-training graph neural networks (GNNs) via self-supervised learning techniques to overcome the challenge of data scarcity in molecular property prediction. However, current self-supervised learning-based methods suffer from two main obstacles: the lack of a well-defined self-supervised learning strategy and the limited capacity of GNNs. Here, we propose Knowledge-guided Pre-training of Graph Transformer (KPGT), a self-supervised learning framework to alleviate the aforementioned issues and provide generalizable and robust molecular representations. The KPGT framework integrates a graph transformer specifically designed for molecular graphs and a knowledge-guided pre-training strategy, to fully capture both structural and semantic knowledge of molecules. Through extensive computational tests on 63 datasets, KPGT exhibits superior performance in predicting molecular properties across various domains. Moreover, the practical applicability of KPGT in drug discovery has been validated by identifying potential inhibitors of two antitumor targets: hematopoietic progenitor kinase 1 (HPK1) and fibroblast growth factor receptor 1 (FGFR1). Overall, KPGT can provide a powerful and useful tool for advancing the artificial intelligence (AI)-aided drug discovery process.