A unified framework combining linear and 3D molecular features for robust drug-protein interaction prediction

Summary

We develop PointDPI to predict drug-protein interactions by simultaneously exploiting their linear and 3D structural features. By aligning the features, PointDPI stereoscopically recognizes molecular properties and reduces reliance on 3D structures. Local topological relationships among molecules are further preserved for avoiding distortion. PointDPI predicts key regulatory sites based on the model’s gradient. We demonstrate improved performance over several state-of-the-art (SOTA) methods, including increased accuracy in dealing with unseen molecules. Four predicted drug-protein interactions (DPIs) are experimentally validated at both mRNA and protein levels, highlighting the therapeutic potential of adenosine in inflammatory diseases, ondansetron and etodolac in neurological diseases, and neuroprotective action for dopamine.

Graphical abstract

Highlights

• •PointDPI predicts drug-protein interactions using combined linear and 3D features
• •Aligns molecular features stereoscopically, reducing dependence on full 3D structures
• •Preserves local topology to avoid distortion, improving prediction accuracy
• •Validated predictions reveal therapeutic potential for several drugs

Motivation

Drug-protein interaction (DPI) prediction is widely studied, but current methods that rely on either linear or 3D molecular features face reliability issues. Although combining multiple feature types improves molecular representation, nonlinear embeddings can distort relationships between molecules. To address this, we develop PointDPI, which integrates both linear and 3D structure information of molecules for DPI prediction.

Sun et al. develop PointDPI to predict drug-protein interactions (DPIs) by integrating linear and 3D molecular structures. PointDPI preserves inter-molecular relationships and predicts key regulatory sites, outperforming several state-of-the-art methods.

Introduction

Drugs intervene in diseases by regulating the activity or expression of relevant target proteins.¹ Therefore, identifying drug-protein interactions (DPIs) is crucial for drug discovery.^2,3 Nowadays, numerous studies have begun to explore potential DPIs using in silico methods.⁴ These methods analyze molecular properties by employing multiple modal features of drugs and proteins, thereby providing empirical insights into the interaction patterns between molecules.^5,6,7,8 By providing reliable drug-protein candidates for subsequent wet-lab experimental validation, these methods can reduce the workload of biological validations and enhance the efficiency of drug discovery.^9,10

Early prediction methods infer DPIs predominantly based on the “guilt-by-association” hypothesis, which suggests that drugs with similar properties are likely to interact with the same protein, and vice versa. Many classical machine learning techniques have been employed, such as matrix factorization,¹¹ random forest,^12,13 support vector machines,¹⁴ linear regression,^15,16 etc. Due to the molecular-level similarity comparisons, these methods face challenges in capturing key substructures of molecules and exploring specific binding sites between drugs and targets.

Another category of studies based on the “guilt-by-association” hypothesis constructs networks using the relationships between entities such as drugs, targets, and diseases.^17,18,19 These methods can integrate the functional information of drugs and targets during the prediction process, thereby significantly improving the prediction performance. However, such methods require a large amount of prior knowledge and labeled data to construct the network, resulting in poor scalability.²⁰

With the abundance of drug- and protein-related data and the increasing number of known DPIs, deep-learning-based methods gradually become mainstream.²¹ These methods predict DPIs by analyzing the atomic composition, physicochemical, or evolutionary properties of molecules.^22,23 A common practice of deep learning models is to encode the linear features of molecules, such as the Simplified Molecular Input Line Entry System (SMILES) notation²⁴ for drugs and the amino acid sequences of proteins, to predict the regulatory relationships between molecules.^25,26 Nevertheless, linear features are insufficient in describing the functional groups of molecules and the positional relationships between atoms. Some studies switch to exploiting atomic maps^27,28,29 or skeletal formula^30,31 to represent molecules. These methods effectively capture the connectivity (chemical bonds) and clustering (functional groups) among atoms in drugs, thereby improving prediction performance. Unfortunately, proteins lack representation formats similar to the skeletal formula of drugs. In addition, both skeletal formula and atomic graph simplify and flatten the spatial structures of molecules, hence fail to capture the relative spatial positioning of atoms, such as interatomic distances and conformations, which significantly influence the chemical properties of the molecules.³² In comparison, the 3D structures of molecules provide more information for deciphering molecular properties. Recent studies demonstrate that incorporating the 3D structures of molecules in DPI prediction not only improves the performance of models but also aids in analyzing the active substructures involved in molecular regulation.^33,34 However, due to the difficulties associated with obtaining complete and pure protein structures, the applicability of these 3D structure-based methods is restricted.^35,36 Although some AI models^37,38 have made significant breakthroughs in protein structure prediction, the accuracy is still constrained by the availability of known structures. If the target protein lacks sufficient related structures, the predicted structure may not be reliable.³⁹ These unreliable structures may mislead models and limit the accuracy of prediction.

Integrating multiple modal features is patently beneficial to comprehensively defining molecules. However, some modal features are phenotypic, such as toxic side effects and clinical manifestation of drugs, functions, association with diseases of proteins, etc. These kinds of phenotypic information are generally incomplete and hence may adversely affect models. Moreover, the nonlinear embedding of features during deep learning process, especially with multi-modal embedding and alignment, may distort the inherent relationships among molecules and incur misidentification of DPIs.

In this study, we develop PointDPI, a DPI prediction method that leverages the multi-modal structures determined uniquely by the chemical compositions of molecules. PointDPI incorporates both linear features (drug SMILES and protein amino acid sequences) and 3D structural features (induced by the point cloud) of molecules as inputs of the model. The unique point cloud module enables PointDPI to capture the specific chemical compositions of drugs and proteins while learning the relative position relationships between atoms in the molecules. By further aligning the embeddings learned from the linear features with those learned from the 3D structures, PointDPI induces high-level structural features from the different aspects of chemical compositions to comprehensively appreciate the properties of molecules. When the 3D structures of drugs or proteins are unreliable or unavailable, PointDPI can still work by utilizing the well-aligned linear features. This is also useful when working with large-scale datasets. For maintaining the inherent relationships among molecules in the chemical space and avoiding distortion in the embedding space, we propose to preserve the local topological relationships among drugs and among proteins during deep learning. Ablation experiments demonstrate that integrating 3D structures of proteins and drugs and preserving local topological relationships among molecules significantly improves the accuracy of DPI prediction. PointDPI outperforms several state-of-the-art (SOTA) methods on three public datasets (DrugBank, BindingDB, and Luo’s dataset). Cold-start experiments conducted on new drugs and new proteins confirm the scalability and portability of PointDPI. Case studies showcase PointDPI’s capability to predict interpretable DPIs. By analyzing the gradient of the model, PointDPI provides the key substructures involved in molecular regulation. The results of in vitro experiments on SH-SY5Y cell line demonstrate that PointDPI successfully identified four drug-protein pairs that exhibited clear interactions. These drugs exhibited significant regulatory effects on the expression of their respective candidate proteins at both the mRNA and protein levels. Notably, further downstream signaling pathway analysis indicated that adenosine increased intracellular cAMP content by inhibiting both the expression and enzymatic activity of PDE4B, resulting in reduction of pro-inflammatory cytokines, thus exerting anti-inflammatory effects. This highlights the potential role of adenosine in the control of inflammatory diseases. By inhibiting NMDA2B receptor expression and alleviating excitotoxicity, dopamine promoted BDNF synthesis and modulated associated signaling pathways. This evidently indicates that dopamine may offer protective effects in neurodegenerative and psychiatric diseases. Ondansetron showed to inhibit adenylate cyclase (AC) activity and suppress $C{a}^{2+}$ influx through kappa opioid receptor (KOR) mediation, highlighting its modulatory effects in ion homeostasis-related disorders. The reduction of RXR-$\gamma$ expression by etodolac treatment suggests its regulatory effects in related neuronal diseases, particularly in neuroimmune-related disorders

Results

Framework of PointDPI

For representing drugs and proteins, we introduce both the linear and structural features determined by molecular compositions. The linear feature of a drug is represented by its SMILES, a 1D sequence that describes its atomic and chemical bond information. For a protein, we utilize its amino acid sequence as its linear feature. To explore the interaction between drugs and proteins at structural level, we introduce 3D point cloud data for both drugs and proteins. The point cloud data comprise the 3D coordinates of all the atoms in drugs and proteins.

Let $\mathcal{R}=\left\{r_1,r_2,\dots r_u\right\}$ and $\mathcal{P}=\left\{p_1,p_2,\dots ,p_v\right\}$ represent the sets of $u$ drugs and $v$ proteins in the dataset, respectively. PointDPI is trained to predict the interaction possibility $s_{r_i,p_j}\in \left[0,1\right]$ between any drug $r_i\in \mathcal{R}$ and protein $p_j\in \mathcal{P}$ based on their linear and structural features. Table S1 provides all the notations used in this paper.

The framework of PointDPI is shown in Figure 1. For a given drug-protein pair ($r_i$, $p_j$), the representation learning module (Figure 1A) employs distinct coding layers to learn the linear and structural embeddings for the drug and protein, respectively. The entity alignment module (Figure 1B) is assembled to learn high-level features for comprehensively appreciating the molecular properties and reducing reliance on the 3D structures of molecules. To achieve this, PointDPI minimizes the distances between the linear and structural embeddings of drug $r_i$ and protein $p_j$, aligning the corresponding entities across different data sources. For maintaining the inherent relationships among molecules, we compute the similarities between all pairs of drugs and between all pairs of proteins based on their chemical compositions. With a similarity threshold, we obtain the local topological relationships of drug $r_i$ and protein $p_j$, which involve only some of their neighbors. In the process of aligning different modal features, PointDPI also minimizes the differences between the inherent similarities and the embedding similarities computed between the molecules and their neighbors. For facilitating comparison, we refer to such PoinDPI as PointDPI-tp. The aligned drug and protein embeddings are subsequently concatenated and serve as the input for the predicting layer (Figure 1C) to calculate the interaction propensity between the drug and the protein. In the prediction process, by analyzing the gradients of the model, we calculate the importance of all sites of the molecules to find the possible regulatory sites. For more comprehensive details of the PointDPI framework, please refer to the STAR Methods section.

Figure 1.

PointDPI pipeline

(A) The linear and structural features of drugs and proteins are input into respective coding layers to obtain the embeddings.

(B) By minimizing the differences between the linear and structural embeddings, the alignment vectors of drugs and proteins are generated. Meanwhile, the local topological relationships among drugs and among proteins are preserved to maintain the inherent relationships among molecules.

(C) The interaction score is calculated by the prediction layer based on the alignment vectors. In addition, by calculating the gradient of the features, we analyze the molecules’ key sites that involved the regulation process.

(D) The details of the point cloud coding layer. According to the chemical bonds between atoms, an atom connection graph is established. Similarly, based on the 3D coordinates of the atoms, an atom distance graph is constructed. A graph convolutional layer is equipped to learn the atom embeddings by aggregating the attributes of neighboring atoms in both graphs. A 1D convolutional layer is thus used to induce the embeddings.

PointDPI effectively utilize multi-modal molecular structures to improve DPI prediction

We evaluated the performance of PointDPI and the compared methods on three public datasets: DrugBank,⁴⁰ BindingDB-IBM,⁴¹ and Luo’s dataset.¹⁷ The details of these datasets can be seen in the Datasets section. The area under the receiver operating characteristic curve (ROC-AUC) and the area under the precision-recall curve (PR-AUC) were used as the two major performance metrics. Additionally, we reported the accuracy, sensitivity, and specificity that yielded the best F1 score.

To establish the superiority of PointDPI over existing DPI methods, we compared it with six SOTA methods: BridgeDPI,⁴² GraphDTA,²⁷ MolTrans,⁴³ DrugBAN,²⁸ PSICHIC,⁴⁴ and BINDTI.⁴⁵ GraphDTA is a prediction method originally designed for calculating the binding affinity between molecules. To adapt GraphDTA to binary classification, we followed the procedure outlined in Kexin⁴³ and Huan Yee⁴⁶ to incorporate a Sigmoid function in its final fully connected layer and optimized it using cross-entropy loss. We evaluated it solely using the linear features of the molecules, since that one aim of PointDPI is to eliminate the reliance on the 3D structural features of molecules.

The results in Table 1 show that PointDPI outperforms other methods in terms of ROC-AUC, PR-AUC, and sensitivity on all datasets. It is also compatible in accuracy and specificity.

Table 1.

The performance of PointDPI and four SOTA methods

Dataset	Method	ROC-AUC	PR-AUC	Accuracy	Sensitivity	Specificity
DrugBank	BridgeDPI	0.7754	0.7958	0.6679	0.8068	0.5289
	GraphDTA	0.8246	0.7949	0.7776	0.7894	0.7658
	MolTrans	0.8281	0.8323	0.7200	0.8616	0.5774
	DrugBAN	0.8254	0.8323	0.7116	0.8835∗	0.5396
	PSICHIC	0.8307	0.8413	0.7050	0.8102	0.5799
	BINDTI	0.8352∗	0.8418∗	0.7251	0.8792	0.5711
	PointDPI	0.8471	0.8529	0.7351∗	0.8904	0.5977∗
BindingDB	BridgeDPI	0.9144	0.9243	0.8676	0.8580	0.8766
	GraphDTA	0.9375∗	0.9008	0.9005	0.9053∗	0.8972
	MolTrans	0.8990	0.8684	0.8201	0.8676	0.7865
	DrugBAN	0.9374	0.9415∗	0.8554	0.7595	0.9355
	PSICHIC	0.9071	0.9153	0.8405	0.7922	0.8859
	BINDTI	0.8884	0.8505	0.8464	0.8359	0.8568
	PointDPI	0.9421	0.9501	0.8730∗	0.9094	0.9022∗
Luo’s	BridgeDPI	0.8369	0.8504	0.7754	0.7749	0.7759
	GraphDTA	0.8631	0.8686	0.8221	0.7796	0.8645
	MolTrans	0.8812	0.8642	0.8270	0.8322	0.8218
	DrugBAN	0.8468	0.8611	0.7840	0.7947	0.7734
	PSICHIC	0.8909∗	0.8855∗	0.8402∗	0.8018	0.8026
	BINDTI	0.8854	0.8638	0.7926	0.8437∗	0.8177
	PointDPI	0.8951	0.8911	0.8403	0.8496	0.8307∗

To facilitate comparison and analysis, the maximum two values of each metric are bolded and marked with an asterisk, respectively. Since DrugBank and Luo’s datasets do not provide a standard separation of training and test sets, we conducted a five-fold cross-validation on the two datasets and reported the average performance of each method.

To validate if molecular structural features can effectively improve the prediction of the model, we conducted an ablation experiment. In the experiment, we predicted DPIs only with the linear feature or the structural features of molecules. To highlight the importance of the reliability of structural features, only pairs of molecules with reliable structures (confidence score >45) were selected to train the model. The results in Figure 2A obviously demonstrate that the structural features of molecules provide more valuable information for DPI prediction compared to the linear features. Integrating multi-modal features of molecules can further improve the performance of the model.

Figure 2.

The experimental results on the DrugBank dataset

(A) The ROC curves and PR curves of the ablation experiment demonstrate that considering multi-modal features of the molecule simultaneously is beneficial to improving the accuracy of prediction.

(B) Alignment strategy can effectively release the negative effects of protein structure confidence.

(C) t-SNE visualizes the distribution of the drug-protein pairs’ embeddings.

(D) The ROC-AUC and the PR-AUC of 5-fold cross-validation. $\mu$ denotes the average performance of 5-fold.

(E) The convergence process of the alignment loss of drugs (left) and proteins (right).

(F) The density distribution of prediction score for the positive and negative samples.

(G) The cold-start experimental results of drugs (left) and proteins (right).

It is known that, in most cases, experimentally determining molecular 3D structures is challenging and the 3D structures predicted by AI models suffer from suboptimal confidence levels. As such, we utilize entity alignment strategy in this study to mitigate the negative effects of unreliable structures. As shown in Figure 2B, the aligned model exhibits a higher true positive rate (TPR) compared to the pure structure-based model. The advantage is particularly evident when dealing with molecules with low-confidence structures. The aligned model achieved TPR improvements of 0.15 and 0.22 for structural confidences <45 and <35, respectively.

To visually demonstrate the classification capability of PointDPI, we applied t-SNE (Figure 2C). Here, the representation of each drug-protein pair was set to their embedding before the last fully connected layer, and the coordinates of each pair were calculated by the t-SNE algorithm. It can be observed that positive (known DPIs) and negative (unknown DPIs) samples in the dataset are separated into two distinct clusters (the right cluster and the bottom cluster). The cluster in the top-left region contains many negative samples that may be the unobserved DPIs. Eight candidate DPIs predicted by PointDPI have been validated by published literature or the subsequent wet-lab experiments (see Table S3 for details).

To assess the stability of the model, we conducted 5-fold cross-validation. The known DPIs and unknown DPIs were randomly divided into five equal groups. Each group of known DPIs, along with an equal number of unknown DPIs, was used to test the model in turn. The remaining known DPIs and the same number of unknown DPIs were used for training. As depicted in Figure 2D, PointDPI performs more stably and consistently outperforms other SOTA methods in terms of both ROC-AUC and PR-AUC.

Additionally, to evaluate the performance of each method on individual drugs, we calculated the ROC-AUC and PR-AUC for each method on every drug in the dataset. The experimental results (Table S4) of the Wilcoxon test confirm that PointDPI is significantly superior to other SOTA methods in terms of both ROC-AUC and PR-AUC at the significant level $p<$ 0.05.

PointDPI is capable of eliminating the dependence of the model on the 3D structures of molecules through the entity alignment strategy. The convergence of the mean squared error (MSE) loss between the two types of features for drugs (left) and proteins (right) indicates that PointDPI bridges the gap between different embeddings (Figure 2E). This capability provides the possibility for PointDPI to replace the structural features with linear features when dealing with molecules for which the structures are unavailable or unreliable.

Figure 2F illustrates the density distribution of predicted scores for the positive (left) and negative samples (right). The predicted scores for most positive samples are greater than 0.5, which aligns with the label data. The predicted scores for negative samples exhibit a right-tailing phenomenon. It also aligns with the label data and adheres to a fundamental law of biology that a drug interacts with only a limited number of proteins.

To assess the performance of the prediction methods in dealing with new drugs or proteins (i.e., compounds of unknown use or proteins of unknown pharmacological significance), cold-start experiments were conducted using the drugs and proteins in the DrugBank dataset. In this part, we set the ${\lambda }_s$ in Equation 10 as 1, to preserving the local topological relationships among molecules. This operation helps improve the robustness and transferability of the model, and we use “ ${\text{PointDPI}}_{tp}$” in Figure 2G to mark this method. For drugs, 80% of the drugs in the dataset were selected. The DPIs related to these drugs were used for training, and the remaining DPIs were used for testing, ensuring that there was no overlap between the drugs in the training and test sets. Similarly, cold-start experiments were also conducted for proteins. The performance of each method is shown in Figure 2G. These cold-start experiments demonstrate that PointDPI is more powerful and robust in predicting potential drug-protein interactions for new drugs and proteins. The experimental results of each method on Luo’s dataset and the BindingDB dataset can be found in Figures S1–S3. In addition, the results shown in Figure 2G also demonstrate that maintaining inter-molecular topological relationships significantly benefited DPI prediction.

The predicted results of PointDPI are interpretable

We further analyzed four predicted drug-protein pairs to demonstrate the interpretability of PointDPI. In the deep learning model, the gradient corresponding to a feature reflects its importance to the prediction result. The greater the gradient corresponding to a molecular substructure, the greater the influence of that substructure on the prediction results. In Figures 3A–3D, we mark the key substructures of the candidate pairs according to the gradients of the model. By comparing the candidates with some known DPIs, we explain the rationale behind PointDPI’s decision. For instance, PointDPI predicted that there existed regulatory relationship between adenosine and PDE4B due to the presence of the red-highlighted substructure in adenosine and the residues 503–510 in PDE4B. Upon examining the dataset, we discover that PDE4B is a known target of adenosine phosphate, which shares a highly similar chemical structure with adenosine (Tanimoto coefficient = 0.874). Therefore, it is credible that adenosine and PDE4B may exhibit similar interaction patterns. Another example is the predicted pair of etodolac and RXRG (Figure 3D). RXRA, a protein belonging to the same family as RXRG, exhibits a regulatory relationship with etodolac. Considering the high homology between RXRG and RXRA, it is also reasonable for PointDPI to suggest that etodolac may also have a direct or indirect interaction with RXRG at the highlighted sites.

Figure 3.

Case studies on the DrugBank dataset

(A) PDE4B is a known target of adenosine phosphate, and adenosine and adenosine phosphate have a similar substructure at the predicted interaction region.

(B) For potential interaction with peptide fragments in GRIN2B, we find that DOPO, a known target of dopamine, also has similar peptide fragments.

(C and D) The proteins OPRM and RXRA, which are highly homologous to OPRK1 and RXRG, respectively, are known targets of ondansetron and etodolac.

(E) The known targets (the branches marked with red) and predicted candidates (the branches marked with blue) of Zinc. Heat indicates the sequence similarity between a target and its left neighbor calculated by the Smith-Waterman algorithm. Blue bars represent the predicted interaction score between Zinc and its candidate proteins. Red bars represent the known interaction between Zinc and its known targets.

Figure 3E presents the complete prediction results for Zinc (DrugBank id: DB01593) and the reference molecules associated with these candidates. Zinc has 49 known targets (marked with red branches) and obtained 38 candidate targets (marked with blue branches). As depicted in the figure, most candidate proteins of Zinc can match with a reference protein that bears a high similarity to it.

Further experimental validation of the four predicted DPIs

First, the expression of PDE4B in SH-SY5Y cell line following adenosine treatment was detected to validate the interaction between adenosine and PDE4B predicted by PointDPI. The results showed that, in comparison to the control group, treatment with adenosine significantly reduced the expression of PDE4B at both mRNA (${}^{\ast }p<0.01$, Figure 4A) and protein (Figure 4B, ${}^{\ast }p<0.05$ for Figure 4C) levels. The anti-inflammatory effects of selective PDE4B inhibitors have attracted attention. To investigate the potential mechanism by which adenosine exerts its anti-inflammatory effects through the regulation of PDE4B, we established an inflammatory cell model by stimulating RAW264.7 cells with lipopolysaccharide (LPS). The experimental results demonstrated that adenosine significantly attenuated the LPS-induced elevation of tumor necrosis factor alpha (TNF-α) (${}^{\#\#}p<0.01$, Figure 4D) and interleukin-6 (IL-6) (${}^{\#\#}p<0.01$, Figure 4E). Additionally, adenosine significantly reversed the LPS-induced increase in PDE4B mRNA expression (${}^{\#\#}p<0.01$, Figure 4F) and the decrease in cAMP levels (${}^{\#\#}p<0.01$, Figure 4G).

Figure 4.

Experimental validation of adenosine and PDE4B, dopamine and NMDA2B, ondansetron and KOR, and etodolac and RXR-$\gamma$

(A) Effect of adenosine treatment for 24 h on the expression of PDE4B gene in SH-SY5Y cells (n = 6).

(B) Representative bands of PDE4B in SH-SY5Y cells.

(C) Relative PDE4B expression levels in SH-SY5Y cells after 24 h of adenosine treatment (n = 3).

(D) Effect of adenosine treatment for 24 h on TNF-$\alpha$ concentration of RAW264.7 cells after LPS stimulation (n = 6).

(E) Effect of adenosine treatment for 24 h on IL-6 concentration of RAW264.7 cells after LPS stimulation (n = 6).

(F) Effect of adenosine treatment for 24 h on the expression of PDE4B gene of RAW264.7 cells after LPS stimulation (n = 6).

(G) Effect of adenosine treatment for 24 h on the intracellular cAMP content of RAW264.7 cells after stimulation (n = 3).

(H) Effect of dopamine treatment for 24 h on the expression of GRIN2B gene in SH-SY5Y cells (n = 7).

(I) Representative bands of NMDA2B in SH-SY5Y cells.

(J) Relative NMDA2B expression levels in SH-SY5Y cells after dopamine treatment (n = 3).

(K) Effect of dopamine treatment for 24 h on the expression of BDNF gene in SH-SY5Y cells (n = 7).

(L) Effect of dopamine treatment for 24 h on the expression of BDNF protein in SH-SY5Y cells (n = 6).

(M) Effect of ondansetron treatment for 24 h on the expression of OPRK1 gene in SH-SY5Y cells (n = 6).

(N) Representative bands of KOR in SH-SY5Y cells.

(O) Relative KOR expression levels in SH-SY5Y cells after ondansetron treatment (n = 3).

(P) Effect of ondansetron treatment for 24 h on the intracellular cAMP content of SH-SY5Y cells (n = 5).

(Q) Effect of ondansetron treatment for 24 h on the intracellular calcium content of SH-SY5Y cells (n = 6).

(R) Effect of etodolac treatment for 24 h on the expression of RXRG gene mRNA in SH-SY5Y cells (n = 6).

(S) Representative bands of RXR-$\gamma$ in SH-SY5Y cells.

(T) Relative RXR-$\gamma$ expression levels in SH-SY5Y cells after etodolac treatment (n = 3). ${}^{\ast }p<0.05$, ${}^{\ast }p<0.01$ versus the control group; ${}^{\#\#}p<0.01$ versus the LPS group.

Next, we validated another drug-protein pair predicted by PointDPI, dopamine, and the ionotropic glutamate receptor, NMDA2B. The results showed that, in comparison to the control group, treatment with dopamine significantly reduced the expression of NMDA 2B at both mRNA (${}^{\ast }p<0.05$, Figure 4H) and protein (Figure 4I, ${}^{\ast \ast }p<0.01$ for Figure 4J) levels in SH-SY5Y cell line. To further investigate the relationship between dopamine and the NMDA2B receptor, we assessed the impact of dopamine on intracellular BDNF levels. The results demonstrated that dopamine enhanced BDNF expression at both mRNA (${}^{\ast \ast }p<0.01$ for Figure 4K) and protein levels (${}^{\ast \ast }p<0.01$ for Figure 4L).

Furthermore, the drug-target pair predicted by PointDPI, ondansetron and KOR was confirmed. The results showed that, in comparison to the control group, treatment with ondansetron significantly increased the expression of KOR at both mRNA (${}^{\ast }p<0.05$, Figure 4M) and protein (Figure 4N, ${}^{\ast }p<0.05$ for Figure 4O) levels in SH-SY5Y cell line. To investigate the changes in downstream signaling molecules induced by the activation of KOR, we further assessed intracellular cAMP levels to evaluate the activity of intracellular adenylate cyclase and simultaneously measured intracellular $C{a}^{2+}$ levels. The results suggested that ondansetron significantly decreased the cAMP levels (${}^{\ast \ast }p<0.01$, Figure 4P), indicating inhibition of adenylate cyclase (AC) activity. Additionally, it significantly reduced intracellular calcium ion concentration (${}^{\ast \ast }p<0.01$, Figure 4Q).

Finally, the predicted drug-target pair, etodolac and retinoid X receptor gamma (RXR-$\gamma$), was verified. The results showed that, in comparison to the control group, treatment with etodolac significantly reduced the expression of RXR-$\gamma$ at both mRNA (${}^{\ast }p<0.05$, Figure 4R) and protein (Figure 4S, ${}^{\ast \ast }p<0.01$ for Figure 4T) levels in SH-SY5Y cell.

Wet-lab experimental results

Adenosine is known to serve as an important energy synthesis substance and ubiquitous extracellular signaling molecule.^47,48 As a signaling molecule, adenosine exerts activating effect in regulation of the cardiovascular and respiratory function, applying to the treatment of cardiovascular and respiratory diseases including supraventricular tachycardia, asthma, etc.⁴⁹ PointDPI has predicted the interaction between adenosine and PDE4B. PDE4B is a member of PDE family and serves as hydrolase of the second messenger cAMP, thereby terminating the biochemical effects conducted by the second messenger.⁵⁰ In recent years, there has been a notable focus on the anti-inflammatory effects of selective PDE4B inhibitors.⁵¹ Our results suggest that adenosine may exert potential anti-inflammatory effects by inhibiting the expression of PDE4B, providing preliminary evidence for a potential association between adenosine and PDE4B (Figures 4A–4C). cAMP serves a central role in the resolution of inflammation by promoting granulocyte apoptosis, efferocytosis/phagocytosis, macrophage reprogramming, as well as the biosynthesis of pro-resolving mediator.⁵² Experimental results exploring the underlying mechanism indicate that adenosine increased intracellular cAMP content by inhibiting both the expression and enzymatic activity of PDE4B, resulting in reduction of pro-inflammatory cytokines, thus exerting anti-inflammatory effects (Figures 4D–4G). This further supports the association between adenosine and PDE4B from the perspective of downstream signaling pathways and also highlights the potential role of adenosine in the control of inflammatory diseases, extending beyond the treatment of cardiovascular and respiratory diseases.

Dopamine (DA) is known to be one of the catecholamine neurotransmitters in the brain,⁵³ involving in movement regulation and contributing to development of Parkinson disease and schizophrenia via D1, D2, D3, D4, and D5 receptor subtypes.⁵⁴ PointDPI has predicted the interaction between dopamine and ionotropic glutamate receptor, NMDA2B. NMDA2B is a subtype of NMDA receptors (NMDARs), which are ionotropic glutamate receptors essential for learning, memory, and synaptic development.⁵⁵ However, overactivation of NMDARs triggers a series of enzyme cascades, inducing excitotoxicity, and ultimately results in cell death,⁵⁶ contributing to the development of neurodegenerative diseases including Parkinson disease,⁵⁷ Alzheimer disease,⁵⁸ and stroke.⁵⁹ Our results suggest that dopamine may exert a potential protective role in neurological disorders by inhibiting excitotoxicity through suppression of NMDA2B activation (Figures 4H–4J). Inhibiting NMDARs triggers a cascade of events, including the inhibition of calcium-calmodulin-dependent eEF2 kinase, leading to enhanced protein synthesis of BDNF, which is the pharmacological mechanism of ketamine, an antidepressant.⁶⁰ Further studies found that the acute inhibition of eEF2K activity induces rapid synaptic scaling in the hippocampus.⁶¹ Brain-derived neurotrophic factor (BDNF), a neuropeptide involved in synaptogenesis and synaptic plasticity, has therapeutic potential in neurodegenerative and psychiatric disorders.^62,63 Our data suggest that dopamine, by inhibiting NMDA2B receptor expression and alleviating excitotoxicity, promoted BDNF synthesis (Figures 4K and 4L) and modulated associated signaling pathways, offering protective effects in neurodegenerative and psychiatric diseases. This finding further supports the interaction between dopamine and NMDA2B receptors. More importantly, this provides potential therapeutic insights for neurodegenerative and psychiatric disorders.

Ondansetron, one kind of 5-hydroxytryptamine type 3 receptors (5-HT3R) antagonist, has been used clinically to treat nausea and vomiting via antagonizing the binding of serotonin to 5-HT3R.⁶⁴ PointDPI has predicted the interaction between ondansetron and KOR. KOR, one kind of opioid receptors, participates in a series of physiological and pathological processes including pain,⁶⁵ oligodendrocyte differentiation and myelination,⁶⁶ adult neurogenesis,⁶⁷ and ischemic stroke.⁶⁸ Our experiments revealed that ondansetron may act as an agonist of the KOR, thereby providing preliminary evidence for the interaction between ondansetron and KOR (Figures 4M–4O). Furthermore, activation of opioid receptors induces the dissociation of inhibitory $G\alpha$ and $G\beta \gamma$ subunits. The $G\alpha$ subunits inhibit adenylate cyclase activity, while $G\beta \gamma$ subunits reduce neurotransmitter release by presynaptically blocking voltage-gated calcium channel opening and enhance membrane hyperpolarization by postsynaptically activating G-protein-gated inwardly rectifying potassium channels.⁶⁹ Our results showed that ondansetron inhibited adenylate cyclase (AC) activity and suppressed $C{a}^{2+}$ influx through KOR mediation (Figures 4P and 4Q), further indicating the interaction between ondansetron and KORs. This finding highlights the modulatory effects of ondansetron in ion-homeostasis-related disorders, extending beyond its use as an antiemetic.

Etodolac, one kind of nonsteroidal anti-inflammatory drugs, is mainly used in postoperative pain and rheumatic diseases through selective inhibition of cyclooxygenase enzymes-2, which initiate the formation of prostaglandin.⁷⁰ PointDPI has predicted the interaction between etodolac and retinoid X receptor gamma (RXR-$\gamma$). RXR-$\gamma$ is one isoform of RXR, which serves as ligand of retinoic acid (RA). The binding of RA to RXR or retinoic acid receptors, followed by their combination with a regulatory DNA element, can lead to cascades⁷¹ and thus plays vital roles in pathological processes including cocaine-exposure-induced neuronal dysfunction,⁷² cerebral ischemia/reperfusion associated with thromboinflammation,⁷³ and multiple sclerosis involved in neuroinflammatory networks.⁷⁴ Therefore, the reduction of RXR-$\gamma$ expression by etodolac treatment (Figures 4R–4T) suggests its regulatory effects in related neuronal diseases, particularly in neuroimmune-related disorders.

Discussion

In this paper, we propose PointDPI, a DPI prediction method that utilizes linear and 3D structural features of molecules. PointDPI not only captures the specific chemical composition of drugs and proteins but also learns the relative position relationship between atoms in the molecules. For comprehensively defining molecules’ properties, PointDPI learns high-level molecular features by aligning their linear and structural features, which mitigates its dependence on structural features as well, allowing it to handle a broader range of molecules with unreliable structures. For avoiding distortion of the inherent inter-molecular relationships, the mechanism of preserving local topological relationships among molecules is introduced during the nonlinear embedding process.

Experimental results on three public datasets demonstrate that PointDPI outperforms several SOTA DPI prediction methods steadily. Cold-start experiments further demonstrate that PointDPI is effective in mining potential DPIs for new drugs and proteins. Case studies showcase PointDPI’s capability to predict interpretable DPIs. By analyzing the gradient of the model, PointDPI provides the key substructures involved in molecular regulation. This interpretability enhances the design and development of targeted drug discovery.

The in vitro experiments validated four predicted DPIs—adenosine and PDE4B, dopamine and NMDA 2B, ondansetron and KOR, and etodolac and RXR-$\gamma$. Interestingly, the outcome showed that adenosine had therapeutic potential in inflammatory diseases, while ondansetron and etodolac exhibited promising roles in neurological diseases, extending beyond their current applications. In addition, a new mechanism of neuroprotective action for dopamine was discovered. These findings provide valuable insights for drug development targeting inflammatory and neurological diseases. Moreover, our experimental investigation highlights PointDPI as a valuable deep-neural-network-based framework for predicting drug-protein interactions, offering valuable insights for drug developers.

Limitations of the study

PointDPI requires processing 3D point cloud data of molecules during training, which demands high computing power resources (approximately 40 G of GPU memory), making it impossible to train on lightweight and portable devices. Additionally, PointDPI has not yet taken into account the 2D structures of molecules, such as the skeletal formula and atom graph of drugs. In future work, more diverse molecular structures can be introduced for training. PointDPI is expected to be applied in the relationship prediction of other biological networks besides DPI prediction, such as lncRNA and proteins, miRNA and proteins, etc.

Resource availability

Lead contact

Requests for further information and resources should be directed to the lead contact, Chang Sun (sunchangcn@mail.nankai.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •No new datasets were generated in this study.
• •The source code of PointDPI code has been deposited at Zenodo and is publicly available at https://doi.org/10.5281/zenodo.17442501 and https://github.com/JackieSun818/PointDPI as of the date of publication.
• •Any additional information required to reanalyze the data reported in thispaper is available from the lead contact upon request.

Acknowledgments

This work was partially supported by the National Natural Science Foundation of China (6227226), the National Key R&D Programs of China (2021YFC2100800, 2021YFC2100801, 2020YFA0908700, and 2020YFA0908702), the Tianjin Municipal Health Commission Scientific Research Project of Traditional Chinese Medicine and Integrated Chinese-Western Medicine (no. 2023187), the Tianjin Binhai New Area Health Commission Science and Technology Project (no. 2023BWKQ001) and the CAAI-MindSpore Open Fund.

Author contributions

C.S., J.W., and J.L. conceived the research project. C.S., M.L., Y. Li, and R.T. developed the computational program. Z.Q. and Y.W. designed the wet-lab experiments and analyzed the wet-lab experimental data. Z.Q. performed the wet-lab experiments. Y. Liu supervised and managed the wet-lab experimentation. C.S. and M.L. verified the prediction results. C.S., Z.Q., Y.W., Y. Liu, J.W., and J.L. wrote the manuscript with support from all authors.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-PDE4B antibody	Abcam (Cambridge, UK)	RRID: AB2927553
Anti-KOR antibody	Abcam (Cambridge, UK)	RRID: AB_2868439
HRP-conjugated anti-rabbit antibody	Abcam (Cambridge, UK)	RRID: AB_955447
Anti-NMDA 2B antibody	CST (Boston, USA)	RRID: AB_2112463
Anti-RXR gamma antibody	CST (Boston, USA)	RRID: AB_10698608
Anti-β-actin antibody	CST (Boston, USA)	RRID: AB_2223172
Chemicals, peptides, and recombinant proteins
Adenosine	Sigma-Aldrich (St. Louis, MO, USA)	A4036
Etodolac	TCI (Shanghai, China)	E0858
Ondansetron	Aladdin (Shanghai, China)	O129694
Dopamine	Macklin (Shanghai, China)	A849306
Critical commercial assays
TransZol Up Plus RNA Kit	TransGen Biotech (Beijing, China)	ER501-01
HiScript ® II Q RT SuperMix for qPCR Kit	Vazyme (Nanjing, China)	R223
2 × Universal SYBR Green Fast qPCR mix kit	ABclonal (Wuhan, China)	RK21203
Whole Protein Extraction Kit	Sangon Biotech (ShangHai, China)	C510003
cAMP ELISA Kit	Elabscience Biotechnology Co., Ltd. (Wuhan, Hubei, China)	E-EL-0056
Mouse TNF-α ELISA Kit	Elabscience Biotechnology Co., Ltd. (Wuhan, Hubei, China)	E-EL-M3063
Mouse IL-6 ELISA Kit	Elabscience Biotechnology Co., Ltd. (Wuhan, Hubei, China)	E-EL-M0044
Human BDNF ELISA Kit	Jianglai Biotechnology Co., Ltd. (Shanghai, China)	JL11683
Calcium Colorimetric Assay Kit	Beyotime Biotechnology (Shanghai, China)	S1063S
Deposited data
DrugBank dataset	DrugBank	https://go.drugbank.com/
BindingDB-IBM dataset	Zheng et al.41	https://doi.org/10.1038/s42256-020-0152-y
Luo’s dataset	Luo et al.17	https://doi.org/10.1038/s41467-017-00680-8
Experimental models: Cell lines
SH-SY5Y (human neuroblastoma cell line)	Procell Biotechnology Co., Ltd. (Wuhan, Hubei, China)	CL-0208
RAW264.7 (murine macrophage cell line)	Procell Biotechnology Co., Ltd. (Wuhan, Hubei, China)	CL-0190
Oligonucleotides
PDE4B RT-qPCR primers for SH-SY5Y cells (Forward: ATCTCACGCTTTGGAGTCAAC; Reverse: TTAAGACCCCATTTGTTCAGG)	This paper	Gene: PDE4B; RefSeq: NM_002600.4
GRIN2B RT-qPCR primers for SH-SY5Y cells (Forward: CATCACCTTCATCTGCGAACACC; Reverse: GCAGGATGTTGGAGTGTGTGTTG)	This paper	Gene: GRIN2B; RefSeq: NM_000834.5
RXRG RT-qPCR primers for SH-SY5Y cells (Forward: GGACGATAAGGAAGGACC; Reverse: GGGGAATACGCTTGGC)	This paper	Gene: RXRG; RefSeq: NM_006917.5
OPRK1 RT-qPCR primers for SH-SY5Y cells (Forward: GACTTCCGCACACCCTTGAA; Reverse: ATCGACGTCTTCCCTGACTT)	This paper	Gene: OPRK1; RefSeq: NM_000912.5
BNDF RT-qPCR primers for SH-SY5Y cells (Forward: CGTGACAGCATGAGCAGAGA; Reverse: ACATGCAGTGTTTCCCCCAA)	This paper	Gene: BDNF; RefSeq: NM_170735.6
β-actin RT-qPCR primers for SH-SY5Y cells (Forward: TGAGCGCGGCTACAGCTT; Reverse: TCCTTAATGTCACGCACGATTT)	This paper	Gene: ACTB; RefSeq: NM_001101.5
PDE4B RT-qPCR primers for RAW264.7 cells (Forward: GGTCAGTGCTGCTGAGAGTT; Reverse: CATGTGGGGGCAAATGTTGG)	This paper	Gene: PDE4B; RefSeq: NM_001177980.2
β-actin RT-qPCR primers for RAW264.7 cells (Forward: ATCATTGCTCCTCCTGAGCG; Reverse: CAGCTCAGTAACAGTCCGCC)	This paper	Gene: ACTB; RefSeq: NM_007393.5
Software and algorithms
PointDPI algorithm	This paper	https://doi.org/10.5281/zenodo.17442500
Comparsion method (BridgeDPI)	Wu et al.42	https://doi.org/10.1093/bioinformatics/btac155
Comparsion method (GraphDTA)	Nguyen et al.27	https://doi.org/10.1093/bioinformatics/btaa921
Comparsion method (MolTrans)	Huang et al.43	https://doi.org/10.1093/bioinformatics/btaa880
Comparsion method (DrugBAN)	Bai et al.28	https://doi.org/10.1038/s42256-022-00605-1
Comparsion method (PSICHIC)	Koh et al.44	https://doi.org/10.1038/s42256-024-00847-1
Comparsion method (BINDTI)	Peng6	https://doi.org/10.1109/JBHI.2024.3375025
SPSS 20.0	IBM	N/A
GraphPad Prism 9.4.0	GraphPad Software	N/A
OriginPro 9.5	OriginLab	N/A
ImageJ	Media Cybernetics	https://imagej.nih.gov/ij/
Other
Tanon 5500 Chemical Imaging Luminescence System	Tanon (Shanghai, China)	N/A

Experimental model and study participant details

Cell culture and treatment

The SH-SY5Y cells and RAW 264.7 cells were cultured in a humidified 5% $C{O}_2$ incubator at $37°C$, with maintaining in Dulbecco’s Modified Eagle Medium with 10 $\%$ (v/v) FBS. To validate the interaction between adenosine and PDE4B, SH-SY5Y cells were exposed to adenosine $\left(100\mu M\right)$ for 24h, while cells remained untreated in the control group. Furthermore, the control group, LPS treatment group ($0.1\mu g/mL$ LPS, 6h) and adenosine for LPS treatment group ($100\mu M$ adenosine for 24h following LBS treatment) in RAW 264.7 cells were set up to investigate the potential mechanism of adenosine in anti-inflammatory. In addition, SH-SY5Y cells were respectively exposed to dopamine $\left(100\mu M\right)$, ondansetron $\left(50\mu M\right)$, or etodolac $\left(100\mu M\right)$ for 24h to verify the other three pairs of drug-protein interactions.

Method details

This section outlines the details of each part in PointDPI.

Datasets

The BindingDB-IBM dataset comprises 46,590 drugs, 691 proteins, and 54,194 DPIs obtained from the BindingDB database. The DrugBank dataset involves 6876 drugs, 4628 proteins, and 18,785 DPIs reported by the DrugBank database (Version 5.0). Luo et al. proposed their dataset in 2018, which contains 708 drugs, 1512 proteins, and 1923 known interactions between them. For the BindingDB-IBM dataset, 48,771 drug-protein pairs are designated for training, while 5,423 drug-protein pairs are allocated for testing. Since the DrugBank and Luo’s datasets do not provide a standard separation of training and test sets, we conducted a 5-fold cross-validation on these two datasets, and reported the average performance of the methods. For the drugs in the dataset, we generated their point cloud data with the Rdkit package⁷⁵ based on their SMILES. For the proteins, we obtained their PDB files from the AlpfaFold2 database.^37,38

SMILES coding layer

The SMILES of a drug is a string that encodes information about its atoms and chemical bonds. We use a $d$-dimensional vector to represent each character in the SMILES. Consequently, a drug can be represented by a matrix $R_l\in {\mathbb{R}}^{a\times d}$, where $a$ is the length of the SMILES and $R_l\left(i,j\right)\in \left[0,1\right]$. To learn the linear embedding of a drug, we combine a convolutional neural network (CNN) and a fully connected neural network (FCN). The CNN is responsible for capturing local patterns and features within the drug’s SMILES sequence, while the FCN processes the extracted features and produces the final linear embedding representation of the drug. This process is as follows:

$$\begin{array}{l}{F}_r^{\left(l+1\right)}=\sigma \left(CNN\left(W_c^{\left(l\right)},b_c^{\left(l\right)},F_r^{\left(l\right)}\right)\right)\text{,}\\ h_r^l=\sigma \left(r_l{W}_r^l+b_r^l\right)\text{,}\end{array}$$ (Equation 1)

where $F_r^{\left(l+1\right)}$ represents the feature map output by the CNN. Initially, $F_r^{\left(0\right)}$ is equal to $R_l$, the input matrix representing the drug’s SMILES. The non-linear activation function $\sigma$ is applied to the feature map $F_r^{\left(l+1\right)}$. $W_c^{\left(l\right)}$ and $b_c^{\left(l\right)}$ denote the weight matrix (filter) and bias vector, respectively, of the $l$-th CNN layer. These parameters contribute to the convolutional operations performed by the CNN in extracting local patterns and features from the drug’s SMILES sequence. The vector $r_l$ is obtained by flattening the feature map $F_r^{\left(l+1\right)}$ into a one-dimensional vector. This vector is then passed as input to the fully connected layer characterized by the weight matrix $W_r^l$ and the bias vector $b_r^l$ to produce the final linear embedding representation of the drug, denoted as $h_r^l$. Please refer to Table S2 for the meaning and specific configuration of each hyperparameter.

Protein sequence coding layer

Each protein is represented by an amino acid sequence $s=\left(a_1,a_2,\dots ,a_n\right)$, where each character $a_i$ corresponds to one of the 23 amino acids. By leveraging ProtTrans,⁷⁶ a $d$-dimension vector $p_l$ is generated for each protein based on its amino acid sequence and then fed into a fully connected layer to obtain the embedding $h_p^l$ of the linear feature of protein as follows:

$$h_p^l=\sigma \left(p_l{W}_p^l+b_p^l\right)\text{,}$$ (Equation 2)

where $W_r^l$ is the weight matrix, and $b_r^l$ the bias of the fully connected layer.

Point cloud coding layer

This unique point cloud coding layer in PointDPI enables the utilization of the 3D structure of molecules. For each drug, the connective matrix $R_c\in {\mathbb{R}}^{m\times m}$ is constructed, where $R_c\left(i,j\right)=1$ indicates a chemical bond between the $i$-th and $j$-th atoms. Additionally, the distance matrix $R_d\in {\mathbb{R}}^{m\times m}$ is created by calculating the Euclidean distance between atom pairs. Similarly, for proteins, the connective matrix $P_c\in {\mathbb{R}}^{n\times n}$ and the distance matrix $P_d\in {\mathbb{R}}^{n\times n}$ are constructed based on the connectivity information at the amino acid level and the 3D coordinates of the $C\alpha$ atom of each residue, respectively.

We take a protein as an example to illustrate how this layer works. To capture the impact of both the connectivity and distance between amino acid residues on protein folding and properties, the point cloud coding layer employs a graph convolutional neural network (GCN). The process is as follows:

$$\begin{array}{l}{H}_c^{l+1}=\sigma \left(D_c^{-\frac{1}{2}}\left(P_c+I\right)D_c^{-\frac{1}{2}}{W}_c^l{H}^l+b_c^l\right)\text{,}\\ H_d^{l+1}=\sigma \left(D_d^{-\frac{1}{2}}\left(P_d+I\right)D_d^{-\frac{1}{2}}{W}_d^l{H}^l+b_d^l\right)\text{,}\\ H^{l+1}=H_c^{l+1}\oplus H_d^{l+1}\text{,}\end{array}$$ (Equation 3)

where $H_c^{l+1}$ and $H_d^{l+1}$ are the outputs of the $l$-th GCN layer, $D$ is the degree matrix corresponding to $P$, and $I$ is the identity matrix. $H^l$ is the input of the $l$-th GCN layer, $H^0\in {\mathbb{R}}^{n\times d}$ is the attribute matrix of the protein. Each row in $H^0$ is a $d$-dimensional vector that represents a specific type of amino acid. $H^{l+1}$ integrates the connectivity and distance information between amino acid residues. In addition, a CNN is utilized to learn the amino acid fragment information, which can be defined as follows:

$$\begin{array}{l}{F}_p^{\left(l+1\right)}=\sigma \left(CNN\left(W_c^{\left(l\right)},b_c^{\left(l\right)},F_p^{\left(l\right)}\right)\right)\text{,}\\ h_p^s=\sigma \left(p_s{W}_p^s+b_p^s\right)\text{,}\end{array}$$ (Equation 4)

where $F_p^{\left(l+1\right)}$ is the feature map output by the $l$-th layer of the CNN, $F_p^0$ is the output of the GCN layer. $W_c^{\left(l\right)},b_c^{\left(l\right)}$ are the weight matrix (filter) and bias vector of the $l$-th CNN layer, $p_s$ is the vector obtained by flattening the feature map $F_p^{\left(l+1\right)}$, $W_p^s,b_p^s$ represent the weight matrix and bias of the fully connected layer, and $h_p^s$ is the final structural embedding of the protein. The structural embedding $h_r^s$ of a drug can be obtained similarly.

Entity alignment

Both the linear and structural features of a drug (or protein) are employed in this paper to decipher the same molecule from the different aspects of the molecular composition, rather than from various incomplete phenotypic information, for comprehensively defining the molecule as a whole. To ensure accurate characterization of the same molecule, its linear and structural embeddings obtained through nonlinear learning should evolve to be consistent at a high level. In PointDPI, we achieve alignment by minimizing the MSE loss between molecular linear embedding and structural embedding. The alignment loss is defined as follows:

$$los{s}_a={\Vert h_r^l-h_r^s\Vert }_2^2+{\Vert h_p^l-h_p^s\Vert }_2^2\text{.}$$ (Equation 5)

In an ideal scenario, the above losses converge. In this case, PointDPI can still utilize linear embeddings as substitute for structural embeddings when the 3D structures of molecules are unreliable or unavailable. This enables PointDPI to work without reliance on 3D structure data, thereby mitigating its limitations and enhancing its generalization capabilities.

Local topological relationship preservation

Existing methods devote every effort to predicting the interaction propensity between a pair of drug and protein while overlook the negative impact of nonlinear embedding on relationships between molecules. For avoiding nonlinear distortion, we propose to preserve the local topological relationships among molecules. To this end, we calculate, ahead of embedding, the Tanimoto coefficients $S^r\in {\mathbb{R}}^{u\times u}$ between all drug pairs, and the Smith-Waterman similarities $S^p\in {\mathbb{R}}^{v\times v}$ between all protein pairs. With a similarity threshold $\theta$, we derive the local topological relationship ${\mathcal{T}}_{r_i}={\left\{r_l\right\}}_{r_l\in \mathcal{R},S_{il}^r\le \theta }$ for drug $r_i$ based on $S^r$, and the local topological relationship ${\mathcal{T}}_{p_j}={\left\{p_k\right\}}_{p_k\in \mathcal{P},S_{jk}^p\le \theta }$ for protein $p_j$ based on $S^p$. When aligning the linear and structural features, we calculate the cosine distance between the embeddings of drug $r_i$ (protein $p_j$) and its neighbor $r_l$ $\left(p_k\right)$ as their embedding similarity $\widehat{S_{il}^r}$ $\left(\widehat{S_{jk}^p}\right)$. The preservation of local topological relationship is achieved by minimizing the following Sammon stresses:

$$los{s}_s=\sum _{r_l\in {\mathcal{T}}_{r_i}}\frac{{\left(\widehat{S_{il}^r}-S_{il}^r\right)}^2}{{\left(1-S_{il}^r\right)}^2}+\sum _{p_k\in {\mathcal{T}}_{p_j}}\frac{{\left(\widehat{S_{jk}^p}-S_{jk}^p\right)}^2}{{\left(1-S_{jk}^p\right)}^2}\text{.}$$ (Equation 6)

DPI prediction module

To integrate the features of the drug and protein, PointDPI constructs the embedding of a drug-protein pair $h_{rp}$ as follows:

$$h_{rp}=h_r\oplus h_p\text{,}$$ (Equation 7)

where $h_r$ and $h_p$ represent the embeddings of drug $r$ and protein $p$ after alignment, respectively. $\oplus$ indicates the concatenate operation. The interaction possibility of drug-protein pair, $s_{rp}$, is then calculated by a fully connected layer.

$$s_{rp}=\sigma \left(h_{rp}{W}^l+b^l\right)\text{,}$$ (Equation 8)

where $W^l,b^l$ are the weight matrix and bias of the fully connected layer, respectively.

The prediction loss of PointDPI is computed as the cross entropy:

$$los{s}_p=-\sum _{r\in \mathcal{R}}\sum _{p\in \mathcal{P}}\left[y_{rp}\log s_{rp}+\left(1-y_{rp}\right)\log \left(1-s_{rp}\right)\right]\text{,}$$ (Equation 9)

where $y_{rp}$ is the real label that indicates whether drug $r$ can interact with protein $p$. For aligning the linear and structural embeddings and preserving local topological relationships among molecules, we can simultaneously minimize the alignment loss and Sammon stresses defined in Equations 5 and 6. The final loss of PointDPI can be defined as follows:

$$loss=los{s}_p+{\lambda }_{a}los{s}_a+{\lambda }_{s}los{s}_s\text{.}$$ (Equation 10)

${\lambda }_a$ and ${\lambda }_s$ serve as parameters controlling the weight of the loss terms. As an end-to-end model, all parameters of PointDPI are simultaneously trained and optimized through this loss function. The corresponding settings of hyperparameters in PointDPI are shown in Table S2.

Quantitative real-time PCR

Total RNA was extracted from SH-SY5Y cells, 1pg RNA was reverse transcribed for cDNA. Primer design of related genes was carried out by the software SnapGene Viewer, and the primer-BLAST verification was performed on the NCBI website. The primer sequences of related genes are listed in key resources table. RT-PCR reactions and analysis were performed by a Realplex2 real-time PCR instrument. $\beta$-actin mRNA was used as an endogenous control. The $2^{-\Delta \Delta Ct}$ method was used to quantify the mRNA expression of the target gene.

Western blotting

Whole protein from each group was extracted from SH-SY5Y cells and its concentration was determined by BCA method. Equal amounts of total protein $\left(60\mu g\right)$ were resolved in SDS-PAGE. The PVDF membranes were blocked and incubated with primary antibodies targeting PDE4B (1:2000), NMDA 2B (1:1000), RXRG (1:1000), KOR (1:1000) and $\beta$-actin (1:1000) overnight at $4°C$ and then incubated with the HRP-conjugated anti-rabbit antibody (1:2000). Protein expression was detected by Tanon 5500 Chemical Imaging Luminescence System and intensities of bands were quantitated by ImageJ.

ELISA

The supernatant of the cell lysate from each group was collected for cAMP and BDNF measurement. Additionally, the supernatant of the cell culture medium was collected to measure the TNF-$\alpha$ and IL-6 content. Following steps in the respective instruction manuals, the levels of cAMP, BDNF, TNF-$\alpha$, and IL-6 were respectively measured. The standard wells, blank wells, and sample wells were set up. Then, biotinylated antibody working solution to each well was added and incubated at 37°C. After incubation, the plate was washed. Next, the enzyme conjugate working solution to each well was added and incubated at 37°C, followed by washing the plate again. The substrate solution (TMB) was added to each well and incubated in the dark at 37°C. Finally, stop solution was added to each well to terminate the reaction. The optical density (OD) was measured at 450 nm. The concentration of cAMP, BDNF, TNF-$\alpha$, and IL-6 for each sample were respectively calculated based on standard curve.

Intracellular calcium ion concentration detection

The supernatant of the cell lysate from each group was collected, and the intracellular calcium ion level was measured following the instructions. Standard or sample was added to each well of 96-well plate, followed by the addition of detection reagent. The plate was incubated at room temperature protecting from light for 5 min. The absorbance at 575 nm was measured using microplate reader. The calcium ion level in each group was calculated based on standard curve.

Quantification and statistical analysis

Statistical analyses were performed in Python with Numpy and Scikit-learn package. SPSS20.0, GraphPad Prism 9.4.0, and Origin pro 9.5 were also used for statistical analysis. The experimental results are expressed as mean $\pm$ standard error of the mean (mean $\pm$ SEM). For comparisons among multiple groups, one-way analysis of variance (ANOVA) was used, while comparisons between two groups were performed using t test, with significance at the level of $p<0.05$ and extreme significance at the level of $p<0.01$.

Published: February 12, 2026