Abstract
Predicting accurate drug associations, including drug‐target
interactions, drug side effects, and drug‐disease relationships, is
crucial in biomedical research and precision medicine. Recently, the
research community is increasingly adopted graph representation
learning methods to investigate drug associations. However, translating
advancements in graph pre‐training to the domain of drug development
faces significant challenges, particularly in multi‐task learning and
few‐shot scenarios. A unified Multi‐task Graph PrompT (MGPT) learning
model is proposed providing generalizable and robust graph
representations for few‐shot drug association prediction. MGPT
constructs a heterogeneous graph network using different entity pairs
as nodes and utilizes self‐supervised contrastive learning of
sub‐graphs in pre‐training. For downstream tasks, MGPT employs
learnable functional prompts, embedded with task‐specific knowledge, to
enable robust performance across a range of tasks. MGPT demonstrates
the ability of seamless task switching and outperforms competitive
approaches in few‐shot scenarios. MGPT emerges as a robust solution to
the complexities of multi‐task learning and the challenges associated
with limited data in drug development.
Keywords: drug associations, heterogeneous graph network, multi‐task
prompt tuning, self‐supervised contrastive learning
__________________________________________________________________
MGPT is a unified multi‐task graph prompt learning model providing
generalizable and robust graph representations for few‐shot drug
association prediction. MGPT demonstrates the ability of seamless task
switching and outperforms competitive approaches in few‐shot scenarios.
MGPT emerges as a robust solution to the complexities of multi‐task
learning and the challenges associated with limited data in drug
development.
graphic file with name ADVS-12-e06444-g006.jpg
1. Introduction
The diverse associations in pharmacology, such as drug‐target
interactions (DTI), drug‐disease associations, drug side effects, and
drug chemical properties, are essential for advancing drug
development,^[ [32]^1 ^] personalizing medicine, and understanding
treatment efficacy. Identifying drug‐target interactions, for instance,
aids in understanding drug mechanisms, streamlining target
identification, and expediting drug design. Predicting potential side
effects is crucial for drug safety, enabling early detection of adverse
reactions and modification of drug candidates. Accurately predicting
these drug associations can improve drug discovery efficiency, reduce
the costs of failed trials, and advance personalized medical care by
tailoring treatments to individual patient characteristics, thereby
revolutionizing the pharmaceutical industry. The integration of
bioinformatics, pharmacology, and computational modeling has become
increasingly vital in the era of big data and artificial intelligence
for precision medicine and new drug discovery.^[ [33]^2 ^]
The research landscape in drug association predictions has experienced
a paradigm shift, with various methods being used to uncover complex
relationships in the pharmaceutical field. Key methodological trends
include machine learning and graph‐based approaches, such as support
vector machines (SVM), random forests, convolutional neural networks
(CNN), have been widely applied. These techniques leverage diverse
datasets encompassing molecular structures, biological pathways, and
clinical information to discern complex patterns and predict
drug‐target interactions,^[ [34]^3 , [35]^4 , [36]^5 ^] drug side
effects,^[ [37]^6 , [38]^7 ^] and drug disease associations. F.
Napolitano, et al.^[ [39]^8 ^] explored drug repositioning using a
machine‐learning approach, incorporating chemical structure similarity,
target similarity, and gene expression similarity into a support vector
machine (SVM). J. Peng, et al.^[ [40]^9 ^] proposed DTI‐CNN, employing
Jaccard similarity coefficients and a restart random walk model with a
convolutional neural network for DTI prediction. S. Dey, et al.^[
[41]^10 ^] employed a chemical fingerprint algorithm to transform drugs
into graphical structures, compressed through convolution, and used a
fully connected neural network to predict drug‐side effect
associations. H. Luo, et, al.^[ [42]^11 ^] proposed a computational
method based on the assumption that similar drugs are often associated
with similar diseases and vice versa. This approach utilized integrated
similarity metrics and the BiRW (Biased Random Walk) algorithm to
identify potential new indications for a given drug.
In recent years, graph neural network methods have seen extensive
application in bioinformatics due to their ability to learn rich
topological information and biological data.^[ [43]^12 , [44]^13 ^]
Graph convolutional neural networks (GCN) and graph attention networks
(GAT) have been utilized to model relationships between drugs, targets,
and diseases. Y. Luo, et al.^[ [45]^14 ^] introduced the
“guilt‐by‐association” concept, proposing the DTINet model that
captures complex relationships in drug discovery through a
heterogeneous network and network representation algorithms. Zhao
et al.^[ [46]^15 ^] established the GCN‐DTI model, utilizing a graph
convolutional neural network for predictive learning. Unlike most
existing methods, GCN‐DTI doesn't separately construct drug and target
networks; instead, it considers the interactions between drugs and
protein pairs (DPP). Another model, proposed by J. Peng, et al.,^[
[47]^16 ^] is the EEG‐DTI model, which takes a heterogeneous network
input containing various biological entities like drugs, proteins,
diseases, and side effects. It employs a graph convolutional neural
network to achieve end‐to‐end information prediction. B. Hu, et al.^[
[48]^17 ^] introduced a method for predicting drug‐related side effects
using a heterogeneous network that integrates various interaction data.
They represented the correlations between drugs and side effects as a
network graph, with each node's representation synthesized from its
adjacent nodes. P. Xuan, et al.^[ [49]^18 ^] developed heterogeneous
graphs combining drug‐disease associations with medicinal chemical
substructures. It integrated both specific and common topologies along
with pairwise attributes of drugs and side effects. Z. Gao, et al.^[
[50]^19 ^] utilized three views to construct a similarity network
between drugs and diseases. Employing two graph encoders, both local
and global topological structures are concurrently modeled.
Subsequently, a graph contrastive learning approach is applied to
collaboratively train node representations, thereby enhancing the
quality of predictions.
Despite the significant contributions that the above‐mentioned methods
have made to drug association research, they still have some
shortcomings. First, all these methods require a significant amount of
training data. However, in the field of drug development, obtaining
large‐scale annotated data is both expensive and time‐consuming.^[
[51]^20 ^] Second, how do we ensure the accuracy of downstream tasks
with few‐shot learning? Machine learning often thrives on large‐scale
data, and its performance can suffer significantly when confronted with
limited samples.^[ [52]^21 ^] Third, most existing research on
predicting drug association is focused on individual tasks, lacking a
unified framework to integrate multiple tasks.^[ [53]^22 , [54]^23 ,
[55]^24 ^] An effective approach to these challenges is the
“pre‐training and prompt‐tuning” schema, where a model is pre‐trained
on related tasks with abundant data and then prompt‐tuned on a
downstream task of interest. While pre‐training models have been
effective in language and vision domains, as well as in graphs,^[
[56]^25 , [57]^26 , [58]^27 , [59]^28 ^] it remains an open question
how to effectively use pre‐training in drug association research.
To address the aforementioned challenges, we present MGPT (Multi‐task
Graph PrompT), a unified learning framework for few‐shot drug
association prediction that integrates graph‐based representation
learning with prompt‐based task adaptation. MGPT constructs a
heterogeneous graph in which each node denotes a concatenated entity
pair (e.g., drug–protein, drug–disease), encompassing diverse
biomedical entities such as drugs, proteins, and diseases. This graph
is pre‐trained via a self‐supervised contrastive learning strategy to
encode both structural and semantic similarities across entity pairs.
In the subsequent prompt‐tuning phase, a learnable task‐specific prompt
vector is introduced to incorporate prior knowledge captured during
pre‐training. Serving as a semantic anchor, the prompt enables
effective knowledge transfer and rapid adaptation to downstream tasks
under limited supervision. By leveraging the learned representations of
entity pairs, MGPT facilitates few‐shot learning across multiple
drug‐related tasks, including drug‐target interaction prediction,
drug‐side effect association, and drug‐disease relationship inference.
We evaluate MGPT on two benchmark drug association datasets and compare
its performance against a range of competitive baseline models. Results
show that MGPT consistently outperforms existing approaches, achieving
state‐of‐the‐art performance in few‐shot settings. In particular, MGPT
surpasses the strongest baseline, GraphControl, by over 8% in average
accuracy, demonstrating its superior generalization and robustness in
low‐resource scenarios. To further elucidate the model's cross‐task
transferability, we analyze the similarity of learned prompt vectors
using cosine similarity. We observe notably high similarity scores
among pharmacologically related tasks, especially between drug‐side
effect interaction and drug substitution, suggesting that MGPT captures
shared semantic structures critical for effective multi‐task learning.
Collectively, these findings highlight MGPT as a powerful and
generalizable framework for few‐shot learning across diverse drug
association tasks. Given the promising results of MGPT in few‐shot
scenarios, we believe it would be worthwhile to explore its application
in real‐world settings where data is often limited. This could include
domains such as drug discovery, where obtaining large labeled datasets
is often challenging due to the cost and time required for experimental
validation. MGPT's ability to effectively learn from limited data could
potentially speed up the discovery process by enabling more efficient
use of resources and reduced dependence on large datasets.
2. Results
2.1. Overview of MGPT Model
The Multi‐task Graph Prompt (MGPT) Learning model is a cutting‐edge
framework developed for few‐shot drug association prediction. This
process begins with the construction of a heterogeneous graph network,
wherein nodes are entity pairs created by combining different node
types such as proteins, drugs, and diseases. The model then undergoes
self‐supervised contrastive learning to pre‐train these graph nodes
based on their similarities. In specific downstream tasks, MGPT
utilizes a learnable prompt vector. This vector incorporates
pre‐trained knowledge to semantically represent the tasks, aiding in
few‐shot learning across various tasks like predicting drug‐target
interactions, drug side effects, and drug‐disease relationships. The
MGPT Framework stands out as a sophisticated and specialized approach
for consolidating information across diverse tasks in bioinformatics.
2.2. MGPT Outperforms the Baseline Methods Under the Few‐Shot Learning
Condition in all the Downstream Tasks
To evaluate the performance of MGPT in six downstream tasks, including
predicting the interactions of drug‐chemical structure, drug‐drug side
effect, drug‐drug substitute, drug‐target protein, target protein‐gene
ontology, we initially conducted a comparison with seven
state‐of‐the‐art methods in graph representation learning (detailed in
Supporting Information I) on two datasets, detailed experimental
settings are available in the “Methods” section. These methods spanned
supervised learning, unsupervised learning, and a combination of
pre‐training, prompting, and fine‐tuning. Here's an overview of each
approach: Supervised learning based methods trained a Graph Neural
Network (GNN) using nodes and edges in a graph to learn node
representation vectors, which were then directly applied for model
inference. We selected well‐established supervised learning baselines,
including GCN, GAT, and GraphSAGE. Unsupervised learning based methods
didn't rely on labeled node data or predefined outcomes. It enabled the
model to autonomously uncover hidden structures or patterns in the
data, thus acquiring node information for use in downstream tasks. We
included DGI, an approach to contrastive learning that maximized the
mutual information between local and global representations of nodes in
the graph. For specific downstream tasks, the model used the node
information to make predictions, training a Multi‐Layer Perceptron
(MLP) for this purpose. Pre‐training, prompting, and fine‐tuning
methods typically involved training a Graph Neural Network model with a
self‐supervised task. Prompts were used to bridge the gap between
pre‐training and downstream tasks, followed by fine‐tuning the
pre‐trained model on downstream tasks. We considered the GPPT model,
which pre‐trained the Graph Neural Network using a masked edge
prediction task to learn graph node representations. Special tokens
were designed to prompt the model, which was then fine‐tuned for
downstream tasks. In addition, we included GCC, a self‐supervised graph
pre‐training framework that generalizes across diverse graph structures
by discriminating subgraph instances within and across datasets,
enabling the learning of transferable structural representations. We
also evaluated GraphControl, a ControlNet‐inspired deployment module
that conditionally integrates task‐specific information during
fine‐tuning or prompt tuning by aligning input spaces and incorporating
target‐specific attributes, thereby enhancing model adaptability and
accelerating convergence on attributed graphs. Both methods share
conceptual alignment with our framework in their use of pre‐training
for transferable graph representation learning and
task‐specific adaptation.
MGPT outperformed all the baselines across all tasks in the few‐shot
setting (refer to Table [60] 1 and Table [61] 2 ). Specifically, MGPT
showed significant improvements over the baselines, achieving over a 8%
increase in average performance on two datasets compared to the most
competitive baseline method, GraphControl. The results indicated that
MGPT could effectively capture interactions between drugs and other
entities, which is crucial in drug design and modification.
Furthermore, it also proved the effectiveness of MGPT in pre‐training
and prompting, highlighting its ability to learn structural information
and content within graphs, and transfer the knowledge to downstream
tasks through prompting.
Table 1.
The comparison results of MGPT and the baselines in Luo's dataset under
the few‐shot (1% samples) learning scenario (meaning that there were 1%
training samples during the prompt‐tuning).
Method Luo's data
All Drug–protein Drug–disease Drug‐side effect Protein‐disease
GCN 50.83 52.78 50.23 52.00 50.04
GAT 49.97 49.65 50.33 49.96 50.46
GraphSAGE 50.21 49.91 50.18 52.83 49.94
GPPT 49.95 50.02 50.07 50.27 49.97
DGI 51.09 51.61 52.50 53.57 53.75
GCC 54.19 49.84 56.39 58.08 62.83
GraphControl 55.80 53.15 61.54 56.63 61.08
MGPT 65.17 54.95 74.20 72.98 77.12
[62]Open in a new tab
Table 2.
The comparison results of MGPT and the baselines in Zheng's dataset
under the few‐shot (1% samples) learning scenario.
Method Zheng's data
All Drug‐chemical Drug‐side effect Drug‐substituent drug‐target
Target‐go
GCN 50.13 52.83 51.81 52.30 51.53 50.16
GAT 50.19 50.03 49.62 50.09 49.88 50.96
GraphSAGE 50.35 50.02 49.96 49.93 50.05 49.92
GPPT 50.26 53.05 53.06 56.61 51.29 50.16
DGI 51.06 52.87 55.30 52.00 53.66 51.41
GCC 54.58 50.33 51.13 50.49 50.93 54.36
GraphControl 60.52 65.33 65.93 69.30 69.90 59.79
MGPT 70.10 70.06 70.07 77.78 75.44 65.46
[63]Open in a new tab
Interestingly, we observed that two graph pre‐training methods, GPPT
and DGI, performed well in certain tasks. Apart from MGPT, GPPT
achieved state‐of‐the‐art performance on Zheng's dataset, while DGI
excelled on Luo's dataset. This suggests that graph pre‐training
methods can effectively capture relationships between entities and
enhance performance in downstream tasks. However, MGPT still maintained
a lead in various downstream tasks compared to these methods. For GPPT,
its more complex prompting strategy might not have achieved the
expected results in a few‐shot setting, and its fine‐tuning strategy
had limited modifications to the model itself. As for DGI, the lack of
a connection between self‐supervised models and downstream tasks meant
that merely training a new MLP for predictions did not allow the model
to effectively gain prior knowledge in few‐shot scenarios. It is worth
noting that, aside from MGPT, GraphControl achieved the strongest
overall performance by dynamically injecting task‐specific cues during
the fine‐tuning process, thereby effectively alleviating the
transferability‐specificity dilemma. Across both benchmark datasets, it
almost outperformed GPPT and DGI across all evaluated tasks. However,
some supervised models did not perform well in few‐shot scenarios. The
complexity of the model, combined with the few‐shot mode, might not
facilitate effective learning, particularly without appropriate
prompting to transfer the acquired knowledge to downstream tasks.
2.3. Performance of MGPT in Different Settings of Few Shot Learning
In order to explore the performance of MGPT in different downstream
tasks including predicting the interactions of drug‐chemical structure,
drug‐drug side effect, drug‐drug substitute, drug‐target protein,
target protein‐gene ontology, with a small amount of training data, we
conducted an evaluation of the performance of the MGPT under various
few‐shot learning configurations on two established benchmark datasets
(Figure [64] 2a and Figure [65] 3a). We set k% values to 1%, 3%, 5%,
7%, and 10%, meaning that there were k% training samples during the
prompt‐tuning. We observed that the model demonstrated strong
performance in predicting drug substituents and drug target, even with
minimal values of k%, achieving a quasi‐elimination rate exceeding 75%.
It suggested that these two tasks were relatively simple, and
pre‐training can capture sufficient information without extensive
prompt‐tuning for downstream tasks. This observation provided valuable
insights for the effective application of drug pre‐training models in
downstream tasks. The task that was most sensitive to the number of
prompt‐tuning samples was drug‐chemical structure prediction. This
sensitivity indicated that the drug chemical structure task demanded a
high level of accuracy and precision. A larger number of samples could
provide more information, thereby enhancing the model's ability to
capture patterns and nuances in this specific task.
Figure 2.
Figure 2
[66]Open in a new tab
Overall performance of MGPT in Luo's dataset. a) Few‐shot (k% samples)
learning results of MGPT on downstream tasks using Luo's dataset. b)
Heatmap illustrating the similarity of prompt vectors across
variousdownstream tasks in Luo's dataset. c) Ablation study results for
MGPT using Luo's dataset. d) Analysis of various readout strategies in
MGPT using Luo's dataset. e) Visualization of node representations
corresponding to different task types, conducted during pretraining
without the use of prompts using Luo's dataset and pretraining with the
inclusion of prompts using Luo's dataset. f) Comparison of MGPT with
state‐of‐the‐art (SOTA) methods in the few‐shot (k% samples) learning
scenario using Luo's dataset.
Figure 3.
Figure 3
[67]Open in a new tab
Overall performance of MGPT in Zheng's dataset. a) Few‐shot (k%
samples) learning results of MGPT on downstream tasks using Zheng's
dataset. b) Ablation study results for MGPT using Zheng's dataset. c)
Analysis of various readout strategies in MGPT using Zheng's dataset.
d) Visualization of node representations corresponding to different
task types, conducted during pretraining without the use of prompts
using Zheng's dataset and pretraining with the inclusion of prompts
using Zheng's dataset. e) Comparison of MGPT with state‐of‐the‐art
(SOTA) methods in the few‐shot (k% samples) learning scenario using
Zheng's dataset. f) Heatmap illustrating the similarity of prompt
vectors across various downstream tasks in Zheng's dataset. g)
Parameter sensitivity analysis for different hidden dimensions: 8, 16,
32, 64, and 128. h) Parameter sensitivity analysis for different input
dimensions: 16, 32, 64, 128 and 256.
Simultaneously, we conducted evaluations of MGPT and other prominent
methods using two datasets in the different settings of few‐shot
learning scenario, presenting the average prediction results for all
tasks (Figure [68]2f and Figure [69]3e). It can be observed that MGPT
significantly outperformed the other methods comprehensively. With the
increase of the k% value, the performance of most methods improved.
However, MGPT maintained its advantages throughout the range of k%
values. The experimental results also demonstrated that, in general,
MGPT performed better when there were more samples available for
training. However, even with a limited number of samples, MGPT could
still achieve good results.
2.4. MGPT Discovers New Drug Targets
MGPT has demonstrated notable potential in identifying new drug
targets. For example, it predicted an interaction between nortriptyline
(NT) and the multidrug transporter P‐glycoprotein (P‐gp), despite
contradicting the negative ground‐truth label. Remarkably, this
prediction aligns with findings from recent experimental research,^[
[70]^29 ^] which investigated how NT interacts with P‐gp to influence
brain concentrations of psychotropic drugs.
To further support the model's prediction, we conducted molecular
docking using AutoDock Vina,^[ [71]^30 , [72]^31 ^] which yielded a
binding energy of –7.2 kcal/mol for the NT‐P‐gp complex, indicating a
favorable interaction. Structural analysis via the PLIP^[ [73]^32 ^]
revealed that NT forms a stable complex with P‐gp, involving seven
hydrophobic interactions and four hydrogen bonds (Figure [74]6 ). These
interactions suggest a strong binding affinity between the
two molecules.
Figure 6.
Figure 6
[75]Open in a new tab
The interactions between nortriptyline and P‐glycoprotein profiled by
PLIP.^[ [76]^32 ^]
Moreover, in vivo experiments reported in Ref. [[77]29] demonstrated
that pre‐administration of Cyclosporine A, a known P‐gp inhibitor,
significantly increased the brain/blood ratio of NT in rats. This
implies that P‐gp plays a critical role in regulating NT concentration
in the brain, thereby supporting the biological relevance of the
interaction predicted by MGPT.
2.5. Ablation Studies on MGPT
To comprehensively analyze the effectiveness of each component of MGPT,
we conducted two ablation experiments: MGPT without prompt to evaluate
the impact of our prompt strategy, and MGPT without pretrain to
evaluate whether the pre‐training stage effectively acquired prior
knowledge (Figure [78]2c, Figure [79]3b). The following observations
were made: (i) The full model MGPT consistently achieved the highest
performance, showcasing the necessity of the pre‐training and prompt
strategy, regardless of the absence of pre‐training or prompt modules.
(ii) Without pre‐training, the model's performance was substantially
compromised, highlighting the critical nature of unsupervised data
pre‐training and validating the potential of leveraging graph
pre‐training frameworks for drug association predictions.
To better leverage prompts for guiding downstream tasks, we explored
the effects of different readout strategies in MGPT, including SUM,
LINEAR‐MEAN, and FEATURE‐WEIGHTED‐SUM. These strategies showed varying
impacts across tasks. It became evident that FEATURE‐WEIGHTED‐SUM was
effective for all tasks, while the SUM strategy was more suitable for
target‐go prediction. In contrast, the performance of the LINEAR‐MEAN
strategy is relatively weaker compared to the other two strategies. It
suggested the possibility of customizing specific guidance strategies
for downstream tasks or adopting more cost‐effective general strategies
(Figure [80]2d and Figure [81]3c).
We next evaluated the parameter sensitivity in MGPT. The hidden
dimensions were adjusted to 8, 16, 32, 64, and 128 layers
(Figure [82]3g). For Zheng's dataset, encompassing numerous downstream
tasks, optimal performance was achieved with 32 dimensions. In
contrast, Luo's dataset, with fewer downstream tasks, showed better
results with 16 dimensions.
We also conducted a sensitivity analysis of the node input dimensions
(Figure [83]3h). The node dimension set to 32 resulted in the highest
performance for Luo's dataset, while a dimension of 256 was most
effective for Zheng's dataset. Importantly, changes in node dimension
had only a slight effect on accuracy, highlighting the robustness
of MGPT.
2.6. Investigating the Prompt Learned by MGPT
In this section, our objective is to delve into the learning mechanisms
of MGPT, particularly focusing on the role of prompts in both
pre‐training and prompt‐tuning phases for various downstream tasks. Our
approach includes a detailed analysis of how different task types are
represented within the network's node structure, and how this
representation is influenced by the use of prompts.
Our investigation began with the visualization of node representations
corresponding to different task types within the network, conducted in
two phases: pretraining without prompts and pretraining with prompts
(Figure [84]2e for Luo's dataset and Figure [85]3d for Zheng's
dataset). Initially, we visualized node representations learned without
prompts, establishing a baseline. This visualization revealed the
natural segregation or grouping of tasks within the network's
architecture (see Figure [86]2e). The subsequent phase, illustrated in
Figure [87]2e, entailed visualizing nodes after prompt‐based
pre‐training. This step enabled us to assess the impact of prompts on
node representation. Comparison analysis indicated that the nodes
visualized in Figure [88]2e, influenced by prompts, showed a clearer
differentiation between task types. This distinct separation is
significant, suggesting that prompts substantially enhance MGPT's
ability to differentiate various task types. Similar conclusions were
also observed in Zheng's dataset (Figure [89]3d). The improved
distinction in task types facilitated by prompts has significant
implications for MGPT's functionality. It implies that prompts act as a
guiding tool, enabling MGPT to transition more effectively between
downstream tasks. This flexibility is essential for MGPT's application
in diverse real‐world scenarios, where rapid and accurate task
switching is essential. Overall, these results highlight the pivotal
role of prompts in enhancing MGPT's task‐specific learning, solidifying
its utility in a range of downstream tasks. The visual evidence from
our node analysis further emphasizes the importance of prompts in
refining and directing MGPT's learning trajectory.
To better understand the effectiveness of prompt‐tuning in downstream
tasks with pre‐training hints, we utilized t‐SNE visualization to
scrutinize the distribution of sample vectors across various downstream
tasks within two datasets, specifically under the condition of a
few‐shot (1% samples) learning scenario (Figure [90] 4a). We
represented negative nodes with the number 0 and the color red, while
positive nodes were denoted as 1 and colored blue. In Zheng's dataset,
our model effectively identified the decision boundaries in all five
downstream tasks, compared to the initial vector distribution. Notably,
in the drug‐substituent task diagram, the decision boundary was more
apparent than in other tasks, and this clarity corresponded to the
highest predictive accuracy among the downstream tasks. This
observation indicated that using t‐SNE to visualize the representation
of node pairs accurately reflected the performance of the model.
Figure 4.
Figure 4
[91]Open in a new tab
Visualization of the role played by prompt in decision boundaries and
domain adaptation. a) Visualization of the decision boundaries of the
downstream tasks. b) Distribution of the number of drug targets. The
top three drugs with the highest number of targets are Quetiapine, and
Pramipexole. c) Performance of prompt domain adaptation in Luo's
dataset. d) Performance of prompt domain adaptation in Zheng's dataset.
We further evaluated the applicability of MGPT's prompt vectors in
downstream tasks by performing a cosine similarity analysis
(Figure [92]2b and Figure [93]3f). The results revealed some notable
patterns in the similarity of task specific prompt vectors across
different tasks. The results indicate that prompt vectors of the same
type generally exhibit higher similarity. For example, in the Luo's
dataset, the protein‐disease and drug‐disease vectors, as well as the
target‐go and drug‐target vectors in the Zheng's dataset, show this
pattern. The prominence of these similarities was an important insight,
as it reflected the inherent relatedness in the nature of these tasks.
It suggested that these particular task areas shared common features,
which the MGPT model was able to effectively identify and capture. On
the other hand, the prompt vectors for other downstream tasks exhibited
significantly lower levels of similarity. The variation was
particularly significant as it showcased the prompt vectors' capacity
to effectively distinguish between various tasks. This ability to
differentiate is essential when utilizing pre‐trained models, allowing
for the customized application of pre‐trained knowledge. Specific
adaptation to the unique features of each task ensures that the model
can be prompt‐tuned effectively to address the distinct requirements
and characteristics of different tasks.
2.7. Negative Sampling Strategies in Pretraining
During pretraining, we adopt a contrastive learning framework, wherein
the selection of positive and negative samples plays a critical role in
determining the quality of the learned representations. This
section evaluates three commonly used strategies for negative sample
selection in graph‐based contrastive learning:
Strategy 1: Community‐Aware Negative Sampling This approach applies a
community detection algorithm to partition nodes into distinct
communities. Negative samples are selected from nodes outside the
target node's community, thereby reducing the likelihood of selecting
false negatives‐nodes that are semantically similar yet incorrectly
treated as dissimilar.
Strategy 2: Degree‐Based Stratified Negative Sampling Nodes are grouped
into degree‐based strata. Negative samples are drawn either from
different degree tiers or from distant nodes within the same tier. To
further mitigate false negatives, nodes with lower degrees are sampled
with higher probability, as they are less likely to be functionally
similar to the target node.
Strategy 3: Random Negative Sampling This baseline method selects
negative samples randomly from nodes not directly connected to the
target node. While simple and computationally efficient, it lacks
semantic awareness.
As shown in Figure [94] 5a,b, the three strategies exhibit slight
variations in downstream task performance. However, the overall
differences are marginal, likely due to the scale of the graph enabling
the model to learn robust representations regardless of the sampling
strategy. These results suggest that the model exhibits a degree of
robustness to negative sampling design choices.
Figure 5.
Figure 5
[95]Open in a new tab
Comparative analysis of negative sampling strategies in contrastive
learning and enrichment analysis of three typical drug targets. a)
Performance comparison of different negative sampling strategies in
contrastive learning on luo's dataset. b) Performance comparison of
different negative sampling strategies in contrastive learning on
Zheng's dataset. c) GO enrichment analysis of genes linked to drug
Quetiapine. d) KEGG enrichment analysis of genes linked to drug
Quetiapine. e) GO enrichment analysis of genes linked to drug
Pramipexole. f) KEGG enrichment analysis of genes linked to drug
Pramipexole.
2.8. Exploring Domain Adaptation Using Prompts in MGPT
To evaluate the domain adaptation capability of prompts in MGPT, we
conducted experiments where a prompt from one task was used to guide
predictions in different tasks (Figure [96]4c,d). The findings
indicated that, in Luo's dataset, prompts learned from protein‐disease
and drug‐disease interactions exhibited enhanced domain transfer
guidance for various tasks. In Zheng's dataset, the prompt learned from
drug‐substituent interactions demonstrated a more pronounced domain
adaptation ability. Overall, these results confirm that prompts learned
by MGPT can be effectively used for domain adaptation between
downstream tasks, offering valuable insights for domain transfer
learning in drug association analysis.
2.9. Results of Enrichment Analysis of Target Genes
We analyzed the drugs with the top three highest numbers of targets as
predicted by MGPT, focusing on Quetiapine, and Pramipexole of which are
psychotropic drugs (Figure [97]4b). Quetiapine is utilized for managing
bipolar disorder, schizophrenia, and major depressive disorder.
Pramipexole is utilized for managing Parkinson's disease and restless
legs syndrome. It is a dopamine agonist that helps alleviate symptoms
such as stiffness, tremors, muscle spasms, and poor muscle control
associated with Parkinson's disease. The enrichment analysis of GO and
KEGG after integration of the genes linked to the predicted targets
were shown in (Figure [98]5) and Supporting Information Table S4. The
GO enrichment analysis demonstrated that the targets are significantly
related to Neuron‐glial cell signaling, postsynaptic membranes, and
neuronal cell bodies ‐ all crucial to emotional processing
(Figure [99]5c,e). The KEGG pathway enrichment analysis showed that
these targets were notably enriched in pathways including neuroactive
ligand‐receptor interaction, calcium signaling pathway, and the
serotonergic synapse pathway, as detailed in (Figure [100]5d,f).
3. Discussion
This article presents a unified Multi‐task Graph PrompT (MGPT) learning
framework tailored for various few‐shot drug association prediction
tasks. By leveraging a heterogeneous graph network and employing
self‐supervised contrastive learning during pre‐training, MGPT
effectively addresses the challenges of multi‐task learning in the
context of drug development. The integration of a learnable prompt
vector in the prompt‐tuning stage enhances semantic task
representation, enabling efficient few‐shot learning across diverse
tasks such as drug‐target interactions, drug‐side effects, and
drug‐disease relationships. The evaluation against various benchmarks
on two datasets demonstrates the robustness and effectiveness of the
MGPT framework. It exhibits exceptional task‐switching capabilities and
consistently outperforms competitive approaches, achieving an average
improvement of over 8%. Case Study demonstrated that MGPT can identify
previously unexplored associations, such as novel drug targets.
We conducted a comprehensive analysis of the MGPT prompt vector. By
visualizing the node representations for different task types, with and
without the task prompt vector, we discovered that MGPT, through its
pre‐training and prompt‐tuning mechanism, can effectively differentiate
downstream task types, thus promoting the rapid switching among these
tasks. The cosine similarity analysis of prompt vectors further
emphasized the model's proficiency in identifying similarities between
related tasks, as well as its capability to discern differences among
various types of tasks. This property is very important for the
effective application of the pre‐training model to a wide range of
downstream tasks, which can ensure that each task can benefit from the
most relevant and specially customized pre‐training knowledge.
Additionally, the domain adaptation experiment demonstrated that the
prompts learned from MGPT can be effectively utilized for domain
adaptation between downstream tasks, offering valuable insights for
transfer learning in drug association analysis. Finally, by visualizing
the classification boundaries of each downstream task before and after
pre‐training, we discovered that MGPT's pre‐training effectively learns
the inherent semantic information of nodes and uncovers the implicit
relationships within the network structure. It leads to notable
improvements in the few‐shot classification of downstream tasks.
Despite its strong empirical performance, the MGPT framework has
several limitations that warrant further exploration. A primary concern
is its lack of interpretability. Although the integration of prompt
vectors improves task discrimination and downstream performance, the
semantic meaning of these vectors remains opaque, limiting our ability
to understand the model's decision‐making process. To address this,
future work could incorporate explainable AI techniques, such as
attention mechanisms or gradient‐based attribution methods, to shed
light on how prompts influence model behavior. Additionally, MGPT
currently represents pairwise entity relationships as individual nodes
within the graph. While this design simplifies modeling, it may fail to
capture the complexity of higher‐order or multi‐relational biological
interactions, potentially constraining the framework's expressiveness
in more intricate biomedical settings. Overall, while the MGPT
framework shows strong potential for few‐shot drug association
prediction, further refinement and expansion will be essential to fully
realize its applicability to real‐world biomedical challenges,
particularly within the dynamic and data‐scarce landscape of modern
drug discovery.
4. Conclusion
MGPT represents a significant advancement in addressing the critical
challenge of limited data and multi‐task integration in drug
association prediction. By leveraging a unified multi‐task graph
learning framework, MGPT constructs a heterogeneous graph network using
entity pairs and employs self‐supervised contrastive learning to
pre‐train robust graph representations. This novel pre‐training
strategy, combined with learnable functional prompts that incorporate
task‐specific knowledge, enables seamless task switching and achieves
state‐of‐the‐art performance in few‐shot learning scenarios.
Comprehensive experiments across multiple downstream tasks, including
drug‐target interactions, drug‐side effects, and drug‐disease
relationships, demonstrate MGPT's superior ability to generalize and
adapt to diverse tasks with minimal annotated data. Moreover, its
ability to effectively utilize limited samples makes it particularly
valuable in real‐world applications where data scarcity is a common
challenge, such as drug discovery and personalized medicine. By
reducing reliance on large‐scale annotated datasets, MGPT accelerates
the drug development process, enhances prediction accuracy, and
provides actionable insights for precision medicine. This framework not
only advances computational methods in pharmacology but also paves the
way for more efficient and data‐efficient approaches in
biomedical research.
Beyond the current scope, the proposed framework holds strong potential
for extension to other critical areas of drug discovery, such as drug
property classification and toxicity prediction, where labeled data is
typically scarce. Future work may focus on adapting MGPT to incorporate
a broader range of biomedical entities and relational types, or on
integrating domain‐specific knowledge sources to further enhance
predictive performance. Pursuing these directions could significantly
expand the applicability of MGPT and contribute to the development of
more robust and comprehensive computational tools for
pharmacological research.
5. Experimental Section
MGPT Architecture
MGPT consists of three primary components: Heterogeneous Network
Construction, Graph Pre‐training, and Multi‐task Prompt Learning
(Figure [101] 1 ). In the first stage, we build a heterogeneous network
by aggregating data from four distinct tasks into node pairs. These
nodes are linked based on a specific rule: nodes sharing the same
biological substance are connected, creating a network interlinking
drugs and their related entities. The next phase is graph pre‐training,
where we use contrastive learning to analyze node similarities.
Following this, we developed a trainable prompt vector, and with a
limited dataset, directed the pre‐trained model toward specific
downstream tasks. In the final step, we execute drug association
prediction tasks using the learned node representations and the prompt.
Figure 1.
Figure 1
[102]Open in a new tab
An illustrative diagram of MGPT. a) Heterogeneous Network Construction:
Entities such as drugs, proteins, diseases, and side effects are
concatenated into node pairs, laying the groundwork for a heterogeneous
network. b) Graph pre‐training utilizes sub‐graph similarity
contrastive learning, where nodes sharing entities exhibit a higher
scientific basis for similarity. c) Multi‐task Prompt Learning. During
this stage, learnable prompt vectors are introduced for downstream
tasks, serving as parameters for the readout operation, facilitating
the use of diverse aggregation functions on the sub‐graphs specific to
each task.
MGPT Architecture‐Heterogeneous Network Construction
We formally define a heterogeneous graph as
[MATH: G :MATH]
=
[MATH: (V,E,A,R) :MATH]
, characterized by a node type mapping function
[MATH: ϕ:V→A :MATH]
, and an edge type mapping function
[MATH: ψ:E→R :MATH]
. Here,
[MATH: V :MATH]
represents node set,
[MATH: E :MATH]
denotes edge set,
[MATH: A :MATH]
represents node types, and
[MATH: R :MATH]
represents edge types. It's essential to highlight that each node
[MATH: v∈V :MATH]
and each edge
[MATH: e∈E :MATH]
are specifically associated with specific types in
[MATH: A :MATH]
and
[MATH: R :MATH]
, respectively. This signifies that
[MATH: ϕ(v)∈A :MATH]
and
[MATH: ψ(e)∈R :MATH]
. Additionally, heterogeneous graphs consist of various node and edge
types, with the requirement
[MATH: |A|+|R|>2 :MATH]
for their definition.
We first concatenate drugs with other entities such as targets, side
effects, and diseases to form entity pairs, which are treated as nodes
in the heterogeneous graph
[MATH: G :MATH]
, such as v ^(drug, target). These entity pair nodes are undirected and
unweighted, and they are sampled randomly from all possible
associations present in the dataset. A subset of these entity pairs is
selected based on task relevance and data balance considerations, as
detailed in the Experimental Setup section. This construction allows us
to reformulate the downstream association prediction tasks as
classification problems over entity pairs.^[ [103]^3 , [104]^15 ^]
Specifically, our heterogeneous network includes four types of entity
pair nodes, including drug‐protein, drug‐disease, drug‐side effect, and
protein‐disease pairs. When constructing the edges in the heterogeneous
network, given the hypothesis that if two entity pair nodes v ^(a, b)
and v ^(a, c) share any common component entity a, then these two
entity pair nodes are more likely to share similar features, and we
connect them with an edge. Note that, in the process of establishing
connections in the heterogeneous graph, we did not use any
annotated data.
MGPT Architecture‐Graph Pre‐Training
During the pre‐training phase, we design a low‐cost task that doesn't
require biological experiments or annotated data based on the presence
of connections between two nodes. Generally, connected nodes exhibit
higher similarity compared to unconnected nodes, as shown in
Figure [105]1. To transfer the knowledge learned in the pre‐training
stage to downstream node classification, we adopt the graph
pre‐training method based on sub‐graph similarity contrastive learning.
Given the heterogeneous graph
[MATH: G=(V,E,A,R) :MATH]
, our pre‐training objective is to train a function:
[MATH: f:V→Rd
:MATH]
that maps each node
[MATH: v∈V :MATH]
to a d‐dimensional vector representation, where d ≪ |V|. These learned
vectors should encapsulate both node features and structural
information, facilitating downstream tasks, particularly the node
classification problem. A natural idea is to use Graph Neural Networks
(GNN) to learn the function.
Node Representation: Following a strategy that adheres to GNN spatial
message passing, we learn node representations through recursive
aggregation. For instance, at the k‐th layer, the representation of
node v is defined as follows:
[MATH: hvk=AGGREGATE(hvk−1
,{huk−1
:l∈Nv};θk) :MATH]
(1)
where
[MATH: hvk∈
Rd
:MATH]
represents the representation vector of node v in the k‐th layer.
Initially,
[MATH: hv0 :MATH]
contains the features from the original input. Θ, serving as the set of
learnable parameters in GNN, can be expressed as Θ = {θ^1, θ^2, …}.
Exactly, the AGGREGATE function is crucial in incorporating information
from neighboring nodes
[MATH: huk−1
:MATH]
, where
[MATH: u∈Nv :MATH]
, into the representation of the central node v during the aggregation
process. The specific form of this function determines how information
is combined and updated across the node's neighborhoods. Specifically,
we use the AGGREGATE function proposed by GIN:^[ [106]^33 ^]
[MATH: hvk=MLPk(1+εk)·hvk−1
+∑u∈
Nvhuk−1
:MATH]
(2)
where ε is a learnable parameter or a fixed scalar.
Sub‐graph Representation: In a heterogeneous graph
[MATH: G=(V,E,A,R) :MATH]
, we generally define the sub‐graph for node v as follows
[MATH: Sv=(V(Sv),E(Sv)) :MATH]
, where
[MATH: V :MATH]
and
[MATH: E :MATH]
represent sets of nodes and edges, as given by the following formula:
[MATH: V(Sv)={d(u,v)≤δ∣u∈V},E(Sv)={(u,u′)∈E∣u∈V(Sv),u′∈V(Sv)} :MATH]
(3)
where δ represents the scope of the subgraph S [v ]as defined, and d(u,
v) denotes the distance between nodes u and v within the graph. The
subgraph S [v ]aggregates information from node v and its neighbors
within the δ range, facilitating a smoother transition of pre‐training
tasks to downstream tasks.
To compute subgraphs, a readout operation is employed to aggregate the
representations of node v and its neighbors within the subgraph. The
READOUT function is defined as:
[MATH: sv=READOUT({hu:u∈V(Sv)}) :MATH]
(4)
In this specific case, the READOUT function SUM is defined as:
[MATH: sv=(hv+∑u∈V(Sv)hu) :MATH]
(5)
Self‐supervised Contrastive Learning: For an entity pair node v in our
graph
[MATH: G :MATH]
, we choose two nodes, m and n. While node m is directly connected to
node v, there is no direct connection between node n and node v. Our
objective is to reduce the association between the subgraph of node v
and the subgraph of node n while enhancing the association between the
subgraph of node v and the subgraph of node m. Note that when
constructing the heterogeneous graph, edges are only established
between nodes if they share a common entity. Therefore, nodes that
share entities have a higher scientific basis for similarity. To
achieve this, we define our contrastive loss function in the
pre‐training stage as follows:
[MATH: Lpre-train(Θ)=−∑(v,m,n)∈Vpre<
/msub>lnexp(sim(sv,sm)/τ)∑c∈{m,n}exp(sim(sv,sc)/τ),
:MATH]
(6)
where
[MATH: Vpre<
/msub> :MATH]
represents the randomly selected training set, sim is the similarity
between the two sub‐graphs, Θ denotes the GNN parameters, and τ is the
temperature parameter. The optimal
[MATH: Θ=argminΘLpre-train(Θ) :MATH]
obtained at this stage will serve as the weight for the downstream task
model, with the aim of transferring the learned graph knowledge to
downstream tasks.
MGPT Architecture‐Multi‐Task Prompt Tuning
In the new drug development stage, the available samples are typically
limited. Hence, the capability of few‐shot learning is crucial for drug
research and development. Additionally, drawing inspiration from NLP
prompts, we've crafted a highly informative and trainable prompt for
multi‐task drug‐related associations. The objective is to guide the
pre‐trained model toward our downstream tasks, enhancing its ability to
leverage prior knowledge.
Prompt Design: Graph prompts differ significantly from language prompts
due to the following reasons. First, the format of prompts varies: in
natural language processing, prompts take the form of textual
instructions for downstream tasks, while in our tasks, they are
represented graphically. This makes designing graph prompts more
challenging than language prompts as we must consider not only the
content but also the structural information of the graph. Second, while
prompts in natural language processing are often manually crafted,
manual prompt creation in graph processing is impractical.
Current research on graph prompts is limited, with few studies applying
graph prompts to drug association prediction. Furthermore, most
existing graphs are homogeneous, and the design of prompts for
heterogeneous graphs has not been explored. Additionally, there is a
scarcity of research on multi‐task learning. To address these issues,
we propose learnable prompts specifically designed for drug association
prediction at the graph level. These prompts integrate both content and
structural information, employing diverse readout strategies for
different tasks.
To enable flexible adaptation across multiple downstream tasks, we
introduce a learnable prompt vector p that acts as a task‐specific
control signal during the readout phase. In contrast to conventional
feature embeddings, this vector serves as a lightweight, differentiable
instruction that guides the aggregation of subgraph representations in
a task‐aware manner. Conceptually, the prompt vector conditions the
model to attend to task‐relevant features during representation
extraction, thereby facilitating precise and efficient multi‐task
generalization. For instance, the prompt readout operation for a
specific task, such as drug‐target interaction, is as follows:
[MATH: sv(d−p)=READOUT({p(d−p)⊙hu:u∈V(Sv)}), :MATH]
(7)
Here,
[MATH: sv(d−p) :MATH]
signifies the subgraph representation of node v in a particular task,
drug‐protein interaction (d − p). The symbol ⊙ represents element‐wise
multiplication. There is a diverse range of readout options available,
and specific READOUT schemes can be applied based on different node
categories to enhance the model's performance. Specifically, we design
different READOUT strategies, including SUMMATION, AVERAGING, and
FEATURE‐WEIGHTED‐SUM.^[ [107]^34 ^] In this specific case, the READOUT
function FEATURE‐WEIGHTED‐SUM is defined as:^[ [108]^34 ^]
[MATH: sv=(hv+∑u∈V(Sv)hu)·W :MATH]
(8)
This formulation highlights the combination of individual node
representations with a summation operation, followed by a
multiplication with the weight matrix W.
Few‐shot Prompt Tuning: After conducting research, it has been observed
that the success of pre‐training‐prompt tasks in the NLP field can
largely be attributed to the significant similarities between
pre‐training and downstream tasks. These commonalities allow the
knowledge acquired during pre‐training to be effectively utilized in
downstream tasks.
To better adapt the designed prompt to downstream tasks, we freeze the
parameters of the pre‐trained model and fine‐tune the prompt with a
small amount of data, aligning the pre‐trained model more effectively
with the downstream task. In order to identify the shared information
between pre‐training and downstream tasks, we have established two
virtual task bridges,
[MATH: BT :MATH]
and
[MATH: BF :MATH]
, representing positive and negative node instances during the few‐shot
prompt tuning step:
[MATH: BT=1n<
mo>∑vi∈
T<
mi
mathvariant="bold">svi,BF=1n
∑vi
∈F<
mi
mathvariant="bold">svi :MATH]
(9)
where
[MATH: T :MATH]
represents the training dataset of positive node instances, and
[MATH: F :MATH]
represents the training dataset of negative node instances. In the
few‐shot scenarios, each of these sets contains only k% nodes, and each
node consists of two entities with known interactions. Therefore, we
design the following loss function to fine‐tune the downstream tasks:
[MATH: Lprompt(p(z))=−∑(xi,yi)∈V(z)ln
exp(sim(sxi(z),Byi(z))/τ)∑C∈(T,F)exp(sim(sxi(z),BC(z))/τ) :MATH]
(10)
For a specific task z, we define
[MATH: V(z)={(x1,y1),(x2,y2),…} :MATH]
as the training set, where x represents a node, and y represents the
label of x. In drug association prediction tasks, for instance, DTI, x
represents the combination of a drug and a target, and y represents
whether they have interactions.
[MATH: Byi(z) :MATH]
represents virtual task bridges, where y [i ]indicates the positive or
negative label for a specific task z.
[MATH: BC(z) :MATH]
denotes virtual task bridges of both positive and negative samples.
It's important to note that, at this stage, adjustments are made only
to the prompt vector p ^(z). We freeze the parameters from the
pre‐training phase to enhance efficiency, reducing reliance on labeled
data, and making it more suitable for few‐shot learning scenarios.
Experimental Setup
To evaluate the performance of our model in handling downstream
multi‐task scenarios with limited samples, we conduct the pre‐training
process separately on two datasets. During the pre‐training of the
model, we pair the nodes in the dataset based on the inherent
interaction types, selecting 10,000 pairs for each interaction to
construct the heterogeneous network. Note that during the pre‐training
stage, we did not utilize any known interaction information. For
specific tasks, we utilize the same pre‐trained model and specific
prompt vectors. During testing, we employ a particular prompt vector to
perform node classification. In our model, the GNN encoding layer is
set to 3, with a weight decay coefficient of 1e‐4. The dimensionality
of the prompt vector is set as the product of the GNN hidden dimension
and the number of GNN layers, which ensures consistency with the
hierarchical representation structure of the encoder and provides
sufficient capacity for downstream adaptation. In all experiments, we
applied 10‐fold cross‐validation. In terms of computational cost, the
floating‐point operations (FLOPs) required for pre‐training amount to
approximately 1.687 GFLOPs. For each downstream fine‐tuning task
(excluding pre‐training), the model requires approximately 3840 FLOPs,
indicating a relatively low computational burden and enabling efficient
deployment in resource‐constrained environments. For downstream tasks,
we adopt a few‐shot setting, meaning that only k% samples are used for
fine‐tuning the model for each task. Additionally, to balance positive
and negative examples in the fine‐tuning data, we randomly shuffle an
equal number of negative examples each time, i.e., meaning that we use
k/2% positive and k/2% negative samples. We evaluate our model using
the metric of accuracy. For all mentioned baseline methods, we adhere
to using the original parameters and code from the respective papers in
our experiments on the same datasets.
Dataset
Our experiments are conducted on two publicly available datasets that
include a variety of drug association relationships. Luo's data ^[
[109]^14 ^] consists of four entity types: 708 drugs, 1512 proteins,
5603 diseases, and 4192 side effects. The dataset comprises six types
of interactions: drug–protein, drug–drug, drug‐disease, drug‐side
effect, protein‐protein, and protein‐disease interactions (Table [110]
3 ). Zheng's data ^[ [111]^35 ^] includes six entity types: 1,094
drugs, 1,556 target proteins, 738 drug substitutes, 881 chemical
structures, 4,063 drug side effects, and 4,098 gene ontology. The
dataset comprises six types of interactions: drug‐chemical structure,
drug‐drug side effect, drug‐drug substitute, drug‐target protein, and
target protein‐gene ontology interactions (Table [112] 4 ).
Table 3.
Statistics of Luo's dataset.
Luo's data
Node Edges
node node Drug‐ Drug‐ Drug‐ Protein‐
number type protein disease side effect disease
12,015 4 1922 199,214 80,164 1,596,745
[113]Open in a new tab
Table 4.
Statistics of Zheng's dataset.
Zheng's data
Node Edges
node node Drug‐ Drug‐ Drug‐ Drug‐ Target‐
number type chemical side effect substituent target go
12,430 6 133,880 122,792 20,798 10,819 35,980
[114]Open in a new tab
Main Comparison Methods‐GCN
GCN (Graph Convolutional Network)^[ [115]^36 ^] is used for extracting
information from graph data. It primarily utilizes the graph's
adjacency matrix and feature matrix, employing linear transformations
and nonlinear activation functions to obtain low‐dimensional vector
representations for each node, which are then applied in
downstream tasks.
Main Comparison Methods‐GAT
GAT (Graph Attention Network)^[ [116]^37 ^] is a graph neural network
that incorporates attention mechanisms. It utilizes multiple graph
networks to learn low‐dimensional representations of nodes and edges.
The model employs self‐attention mechanisms to identify optimal
mappings and predict relationships between nodes.
Main Comparison Methods‐GraphSAGE
GraphSAGE^[ [117]^34 ^] is an inductive framework for extracting graph
data, which effectively leverages node attribute information to learn
embeddings for newly encountered nodes. Its central concept involves
the acquisition of an aggregation function, such as averaging or
maximum pooling, to merge neighborhood information of nodes into their
own features, thereby obtaining low‐dimensional vector representations
for the nodes.
Main Comparison Methods‐GPPT
GPPT^[ [118]^27 ^] introduces a novel transfer learning paradigm for
the generalization of Graph Neural Networks (GNN). The core concept
involves pretraining a GNN by leveraging a task focused on predicting
masked edges. Following this, a graph prompt function is used to
reframe downstream tasks, aligning them with the pretraining task to
narrow the task gap. The approach introduces methods for generating
task and structure tokens, facilitating the creation of node prompts
for node classification tasks.
Main Comparison Methods‐DGI
DGI^[ [119]^38 ^] framework is designed for unsupervised representation
learning of graph‐structured data, efficiently utilizing node attribute
information to embed newly encountered nodes. At its core, it employs
Graph Convolutional Networks (GCN) to generate local features for nodes
and global features for the entire graph. The training process involves
maximizing the mutual information between these features. This approach
enables the model to learn node vectors that capture both the graph
structure and node attributes, thereby enhancing the performance of
downstream tasks, such as node classification.
Main Comparison Methods‐GCC
GCC^[ [120]^39 ^] is a self‐supervised graph neural network
pre‐training framework designed to learn transferable structural
representations from heterogeneous graphs. Drawing inspiration from
pre‐training paradigms in natural language processing and computer
vision, GCC formulates a contrastive learning objective based on
subgraph instance discrimination, both within and across different
graph structures. This approach enables the model to capture universal
topological patterns and produce graph representations that generalize
effectively to out‐of‐distribution tasks and unseen domains.
Main Comparison Methods‐Graphcontrol
GraphControl^[ [121]^40 ^] is a recently proposed method aimed at
improving domain transferability in pre‐trained graph models by
addressing the “transferability ‐ specificity dilemma”. While
conventional self‐supervised graph pre‐training captures
domain‐invariant structural knowledge, it often neglects task‐ or
domain‐specific node attributes crucial for downstream adaptation.
GraphControl mitigates this limitation by aligning the input spaces of
source and target graphs and conditionally integrating task‐specific
features during fine‐tuning or prompt tuning. This is achieved through
a progressive conditioning mechanism that adaptively injects
target‐specific signals, thereby enabling personalized and
context‐aware knowledge transfer.
Conflict of Interest
The authors declare no conflict of interest.
Author Contributions
G.W. conceived the work, assisted in algorithm design and
implementation. Y.L. and Y.S. contributed to algorithm design,
implementation, and computational experiment analysis. X.Q. contributed
to computational experiment analysis and visualization. X.G.
contributed to algorithm design. Y.L. drafted the initial manuscript.
G.W. and Y.L. contributed to manuscript revision. Y.S. contributed to
data acquisition and processing.
Acknowledgements