Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways

Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways View ORCID Profile Phuong-Nam Nguyen doi: https://doi.org/10.1101/2024.07.23.604746 Phuong-Nam Nguyen 1 Faculty of Computer Science, PHENIKAA University , Yen Nghia, Ha Dong, 12116, Hanoi, Vietnam Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Phuong-Nam Nguyen For correspondence: namphuongnguyen510{at}gmail.com Abstract Full Text Info/History Metrics Supplementary material Preview PDF ABSTRACT Background Contrastive learning has emerged as a pivotal technique in representation learning, particularly for self-supervised and unsupervised tasks. Link prediction, crucial for network analysis, forecasts the formation of connections between nodes. Machine learning enhances link prediction by learning patterns from data, leading to improved performance and scalability. Method In this study, we propose a contrastive learning approach tailored for isomorphic graphs to uncover intrinsic interactions within biological networks. By creating data augmentations through vertex permutations, we train models to learn permutation-invariant representations. Results In this study, we propose a contrastive learning approach tailored for isomorphic graphs to uncover intrinsic interactions within biological networks. By creating data augmentations through vertex permutations, we train models to learn permutation-invariant representations. Our approach was validated using five cancer-targeting biomarkers: ADGRF5, TP53, BRAF, KRAS , and GNAS . Conclusion We discovered new connections between G-coupled receptors ( GPR137B, GPR161 , and GPR27 ) and key path-ways, interactions between cyclin-dependent kinase inhibitors ( CDKN1A and CDK8 ) and specific biomarkers, and identified NFK-BIA as a central node linking all targeting biomarkers. This study highlights the potential of contrastive learning to reveal novel insights into cancer research and therapeutic targets. The implementation of this project is made available at: https://github.com/namnguyen0510/Contrastive-Learning-for-Graph-Based-Biological-Interaction-Discovery . 1 Introduction Contrastive learning has emerged as a powerful technique in the field of representation learning, especially for self-supervised and unsupervised learning tasks. This method leverages contrasting positive and negative sample pairs to learn discriminative features without requiring explicit labels. Contrastive learning aims to learn an embedding space where similar instances are pulled together and dissimilar instances are pushed apart. The foundational work by Hadsell, Chopra, and LeCun (2006) introduced this concept through the Siamese network, which utilized a contrastive loss function to train a neural network on pairs of similar and dissimilar images to learn useful representations for tasks such as face verification 1 . This approach is an emergent training technique in Computer Vision literature 2 – 8 . Link prediction is a crucial task in network analysis, aiming to predict the existence or emergence of links between nodes in a network. Applying machine learning techniques to link prediction has led to considerable improvements in performance. These methods can automatically learn and generalize patterns from data, providing a more robust and scalable solution. We categorize ML-enabled link prediction into two subgroups: (1) feature-based methods and (2) matrix factorization techniques. Feature-based methods involve constructing a feature vector for each pair of nodes and using these vectors as input to a machine learning model 9 , 10 . Matrix factorization techniques have been widely used for link prediction, particularly in recommendation systems. Matrix factorization decomposes the adjacency matrix of a network into latent factors, capturing the underlying structure 11 . Link prediction in networks, especially using machine learning techniques, is highly valuable for predicting new interactions among biological networks. Biological networks, such as protein-protein interaction networks, gene regulatory networks, and metabolic networks, consist of nodes representing biological entities (e.g., proteins, genes) and edges representing interactions between them. Contrastive learning has become a prominent representation learning method, especially in graphs and networks. Contrastive learning frameworks have significantly improved learning robust and meaningful node, edge, and graph representations by leveraging the principles of contrasting positive and negative samples. Deep Graph Infomax (DGI), introduced by 12 , is one of the pioneering works in contrastive learning for graphs. DGI maximizes the mutual information between patch representations (local) and summary representations (global) of graphs. It uses a corruption function to generate negative samples by shuffling node features and a discriminator to differentiate between positive and negative pairs. This approach has shown significant improvements in unsupervised graph representation learning tasks. GraphCL, proposed in 13 , applies data augmentation techniques to generate different views of the same graph. These views serve as positive samples, while other graphs in the batch act as negative samples. The method employs a contrastive loss to maximize the agreement between the augmented views of the same graph. CMRL is introduced in 14 , which leverages multiple graph views, such as node attributes and structural information, to learn node embeddings. By contrasting these different views, CMRL captures complementary information and enhances the quality of learned representations. Despite its usefulness, there still needs to be a research gap in contrastive learning for graph-based datasets. A self-supervised framework is proposed in 15 that integrates contrastive learning with a node classification task. This approach uses contrastive loss to align node embeddings from different augmentations while predicting node labels. Subgraph Contrast introduced 16 , focuses on learning representations by contrasting subgraphs within a larger graph. This method generates subgraphs through random walks and applies a contrastive loss to maximize the similarity between overlapping subgraphs. This article proposes a contrastive learning approach on isomorphic graphs to uncover the intrinsic interactions within biological networks. We create data augmentation by permuting vertices and train the models to learn permutation-invariant representations of these networks. Our numerical results demonstrate the proof of concept using five cancer-targeting biomarkers: ADGRF5 (GPR116), TP53, BRAF, KRAS, and GNAS. Compared to established targeting pathways, these new biomarkers have yet to be extensively studied in the clinical literature. First, we discovered new connections between G-coupled receptors, including GPR137B, GPR161, and GPR27, with the GNAS and BRAF pathways. Second, we identified interactions between cyclin-dependent kinase inhibitors, such as CDKN1A and CDK8, with GNAS and TP53, respectively. Finally, we found that NFKBIA is involved in a pathway linking all the targeting biomarkers: ADGRF5, TP53, BRAF, KRAS, and GNAS. NFKBIA acts as an inhibitor of NF- κ B, which has numerous therapeutic applications. This article is organized as follows: Section 2 presents the proposed framework. Section 3 reports the experimental design and result analysis. Section 4 further investigates the mutual exclusivity of found and targeting biomarkers studied across ten databases. We also provide a comprehensive trend analysis of clinical interests regarding the new biomarkers. Section 5 concludes the work by proposing several research extensions. 2 Method 2.1 Overview In our proposed framework, we diverge from existing works 12 – 16 , by focusing on the intrinsic representations derived from the isomorphism of the original graph. While a comparable approach can be seen in GraphCL, which emphasizes data augmentation through techniques such as node dropping, edge perturbation, attribute masking, and subgraph extraction, our method solely employs creating different isomorphic graphs via permutation. This approach is motivated by two key factors. First, methods like node dropping, attribute masking, and subgraph extraction may overlook essential markers during learning. At the same time, edge perturbation may introduce incorrect interactions by randomly adding or dropping edges. Second, our approach compels the model to learn permutation-invariant representations, thereby uncovering intrinsic interaction patterns. The details of our proposed framework will be discussed in Section 2.2 , with a theoretical interpretation provided in Section 2.3 . 2.2 The Proposed Framework Model Training: Let A n×n be the adjacency matrix associated with a weighted biological network ℊ = ( V, E ), where the node set V = { p 1 , p 2 , …, p n } is the set of proteins with | V | = n and the edge set E is weighted by relevant score between protein pairs. Specifically, we have where e ii = 0 (diagonal elements) and e ij = e ji , i≠j (symmetric graph). An l -layer neural network ℱ Θ with model weight Θ = (Θ 2 , Θ 2 , … Θ l ), for which Θ k has the structure Θ k = ( W k , b k ) ( k ∈ {1, 2, … l }). For each forward pass, we create two versions of A by permutation as data augmentation on the graph, denoted as A ↦ ( A 1 , A 2 ). The embedded representation of A i ( i = { 1, 2 } ) is computed as where * is point-wise multiplication. Clearly, we cannot guarantee H i is a symmetric matrix; thus, we use post-processing the induced representations as follows The first step ensures symmetry, and the second removes self-interaction as in the original matrix A . We aim to maximize the agreement between H 1 and H 2 as in a conventional contrastive learning framework. We use the objective function ℒ:= (MSE, MAE) and the optimization problem is Model Inference: We compute Ĥ = ℱ Θ * ( A ) to predict the interaction between protein nodes. The Softmax function is applied Ĥ ←SoftMax( Ĥ ) to normalized the interaction score to [0, 1] before symmetrizing using Equation 3 . Remarks: We use the point-wise multiplication * to avoid saturated embedding H i and then avoid representation collapsing. Specifically, if the matrix multiplication W · H is used in Equation 2 , this forward pass is attention learning since all other protein interactions are attended in a protein embedding. Consequently, such a message-passing mechanism creates similar embeddings as we increase the number of layers in the model. We want to compute independent latent maps for such interactions and force the model to maximize their agreements. A limitation of our approach is the model complexity, which is T ( n ) = l ·2 n 2 ∈ 𝒪( n 2 ), where l is the number of layers and n is the number of input genes. 2.3 Theoretical Interpretation To create an isomorphic graph using a permutation via its adjacency matrix, consider a graph G with an adjacency matrix A . By applying a permutation matrix P , which reorders the vertices of G , we obtain an isomorphic graph G ′ whose adjacency matrix A ′ is given by A ′ = P T AP . Here, P T is the transpose of the permutation matrix P . This transformation ensures that the structure of G ′ is identical to G , as the permutation merely reorders vertices without altering the graph’s connectivity. We study the interactions through permutations, which the concept of orbits in permutation groups can elucidate. In this context, an orbit of a vertex under a permutation group action consists of all vertices that can be reached by applying the permutations in the group to the vertex. These orbits reveal the symmetry and equivalence classes of vertices under the permutation, providing insights into the structural properties and the equivalence of interactions within the graph. Permutation-invariant representations on isomorphic graphs are crucial as they reveal intrinsic interactions by focusing on the entire graph’s structure rather than the specific subgraph. By ensuring that the representation remains unchanged under any vertex permutation, we can capture the fundamental properties and relationships within the graph. This invariance is vital for learning from graph data, which requires models to recognize and leverage these intrinsic interactions, leading to more robust and generalizable insights. 3 Numerical Results 3.1 Case Study In the numerical demonstration, we aim to discover the pathway between ADGRF5 (GPR116) and iconic drivers of cancers (TP53, BRAF, KRAS, GNAS). Of note, GPR116 is a potential biomarker for the activation of CTLA4 discovered in our previous work 17 . TP53 , often called the “guardian of the genome,” is extensively studied for its role in tumor suppression. Similarly, BRAF and KRAS are key players in cell signaling pathways frequently mutated in cancers, driving substantial research for targeted therapies. BRAF encodes a protein known as B-Raf, part of the MAPK/ERK signaling pathway that regulates cell division, differentiation, and secretion. KRAS is a member of the RAS gene family, encoding a protein that also plays a role in the MAPK/ERK signaling pathway. KRAS acts as a molecular switch, cycling between active and inactive states to regulate cell proliferation and survival. GNAS encodes the G α s protein, a subunit of the heterotrimeric G protein complex involved in transmitting signals from cell surface receptors to intracellular effectors, primarily through the activation of adenylyl cyclase and the subsequent production of cAMP. Mutations in GNAS can lead to constitutive activation of this signaling pathway, contributing to the development and progression of various tumors. ADGRF5 (GPR116), also known as Adhesion G Protein-Coupled Receptor F5, is a member of the adhesion G protein-coupled receptor (GPCR) family. GPCRs are a large family of cell surface receptors that play key roles in various physiological processes and are involved in many diseases. ADGRF5 is predicted to facilitate G protein-coupled receptor activity and is involved in processes like glomerular filtration, pharyngeal arch artery development, and surfactant regulation. It is located in the apical part of the cell and is expressed in the genitourinary system, immune system, liver, lungs, and spinal cord vasculature (Alliance of Genome Resources, April 2022). ADGRF5 plays a crucial role in glucose homeostasis 18 . This biomarker is involved in the release of somatostatin from pancreatic delta cells, as shown in knock-out mouse models. Whole-body deficiency of GPR116 results in decreased beta-cell mass, fewer small islets, and reduced pancreatic insulin content. Despite these changes, glucose homeostasis is maintained through compensatory mechanisms that modulate insulin degradation. This highlights GPR116 ’s significant role in regulating glucose levels. ADGRF5 promotes breast cancer cell migration and metastasis, with its absence impairing cell motility and tumor growth, correlating with increased MMP8 expression and a shift in tumor-associated neutrophils towards an antitumor phenotype 19 . Mechanistically, ADGRF5 enhances RhoA activation and inhibits ERK1/2 activity, reducing C/EBP β activation and leading to heightened MMP8 expression, which attracts antitumor neutrophils and reduces TGF- β -enhanced cell motility, making ADGRF5 a promising therapeutic target for breast cancer. Despite its potential, ADGRF5 has not been studied widely when compared with well-known biomarkers such as TP53, BRAF or KRAS , illustrated in Figure 3 and 4 . We use the String database 20 of protein-protein interaction for this evaluation. In the original graph ( SupMat-S1 ), only KRAS and BRAF are connected by a direct pathway, while BRAF connects with TP53 through HSP90AA1. Besides, there is not sufficient evidence to show the connection between GNAS to KRAS, BRAF, and TP53, yet GNAS could be co-activated with these drivers in some certain cancers 21 , 22 . ADGRF5 and GNAS are connected via RAMP2. The input protein-protein interaction network includes 2,785 proteins and 570,110 edges. 3.2 Experiment Designs We use 2-,4- and 8-layer NNs, trained with three loss functions MSE, MAE, and KL divergence under SGD 23 , Adam 24 and AdamW 25 optimizers. We use a flat learning rate of 0.3 for each experiment for the optimizer and train the model for 500 epochs. SGD and KL loss could be more effective (not shown) for our framework, resulting in minimal reduced or intractable loss values. Only 18 models, ignoring the intractable models trained under KL loss, are used for further evaluations. We consider the NNs produced by SGD to be random classifiers. In contrast, models trained with Adam and AdamW can learn to maximize the agreement between the augmented representation, as shown in Figure 2 . The optimal neural architectures are found at 4 layers, while the lower depth model (2-layer) has much higher contrastive loss or two deep models (2-layer) do not guarantee the performance gain. To improve the robustness of the proposed pipeline, we use two ensemble training techniques, which are mean average and weighted average. Specifically, we use 18 models with the score computed by the model with the least loss value will contribute the most to the final synthesized graph. We compute where  is the synthesized adjacency matrix of the predicted protein-protein interaction network. 3.3 Evaluation of Synthesized Protein-Protein Interaction The analysis of the protein interaction networks reveals distinct structural differences between the original, mean-averaged, and weighted-assembled graphs. The original interaction network was filtered to include only significant links within the 98 th percentile, resulting in a subgraph of 257 proteins and 275 links. This subgraph differs from the mean-averaged graph, which includes 267 proteins and 280 links, highlighting a slight increase in complexity when interactions are averaged. A noteworthy finding is that 38 protein and 18 links are common between both graphs, indicating a core subset of stable interactions. The graphical illustrations are given in SupMat-S2 . Specifically, in the mean-averaged graph, the protein GNAS retains its connection with RAMP2 but loses its direct link with ADGRF5, instead connecting through PDE1A. Three subtypes, PDE1A, PDE1B, and PDE1C, are included in the PDE1 family. PDE1 has potential as a therapeutic target for various diseases, including cardiovascular, pulmonary, metabolic, neurocognitive, renal disorders, cancers, and possibly others 26 . Additionally, GNAS shows a new predicted interaction with GPR161, and KRAS emerges with connections to CTLA4, MRAS, RASA4B, and TP53BP2. TP53 displays novel connections to TP53INP1 and CDK8, with a new indirect link to ADGRF5 via CRBN. CDK8 is a key oncogenic factor, positioning it as a potential tumor biomarker and a promising target for cancer therapy 27 . In the weighted-assembled graph, PDE1A continues to serve as the intermediary between GNAS and ADGRF5, and CRBN remains the bridge for TP53 and ADGRF5. GNAS is also predicted to link to MTOR and CDKN1A, while KRAS exhibits fewer connections, maintaining links only to TP53BP2 and CTLA4. The mTOR pathway is crucial for regulating cell survival, metabolism, growth, and protein synthesis in response to upstream signals in both normal and pathological conditions, particularly in cancer, where it is aberrant signaling, often due to genetic alterations, promotes tumor initiation and progression. Despite extensive use of mTOR inhibitors in therapy, more effective treatments with fewer resistances are needed, with new biomarkers and sequencing technologies enhancing personalized therapy 28 . CDKN1A/p21 is involved in various processes critical for cancer progression, including cell cycle regulation, DNA damage response, apoptosis, senescence, stem cell fate, and EMT. However, its effects can be both pro-tumorigenic and anti-tumorigenic 29 . Both graphs reveal new interactions between BRAF and GPR137B and GPR27, suggesting previously unrecognized pathways. These findings underscore the dynamic nature of protein interaction networks and the importance of different analytical approaches in uncovering novel biological relationships. The GPCR-like lysosomal protein GPR137B controls the localization and activity of Rag and mTORC1 30 . GPR27 is an orphan GPCR that may play a role in neuronal plasticity 31 . We summarize the discovered biomarkers in Table 1 . View this table: View inline View popup Download powerpoint Table 1. Discovered biomarkers by our proposed framework, stratified by targeting biomarkers 3.4 Extraction of Significant Sub-graphs for Investigated Pathway We now extract the most significant sub-graphs that include the targeting biomarkers (ADGRF5, TP53, BRAF, KRAS, GNAS). This is equivalent to applying the longest path problem on both synthesized networks. In the mean average case, we found the pathway: [NFKBIA, ADGRF5, GNB4, GNAS, ZP2, BRAF, VDAC1, KRAS, BPTF, TP53, RBPJ]. In the weighted average case, we found: [NFKBIA, ADGRF5, WEE1, TP53, C16orf89, GNAS, NCOA1, BRAF, VDAC1, KRAS, ROCK1]. The discovered pathways are shown in Figure 1 and summarized in Table 1 . Of note, NFKBIA pre I κ B α (nuclear factor of kappa light polypeptide gene enhancer in B-cells inhibitor alpha) is addressed by both assembled methods. I κ B α can inhibit NF- κ B proteins, holding them in an inactive state within the cytoplasm 32 . NF- κ B appears to be crucial in regulating the expression of immunoregulatory genes involved in critical illness, inflammatory diseases 33 , apoptosis, and cancer. Specifically, NF- κ B is central in controlling the transcription of cytokines, adhesion molecules, and other mediators 34 . Agents inhibiting this pathway, like glucocorticoids and aspirin, can diminish the inflammatory response, while others, such as dominant negative I κ B proteins, enhance the effects of chemotherapy and radiation therapy in cancer treatment 35 . A broad range of common human malignancies display abnormal constitutive expression of NF- κ B, leading to tumorigenesis and cancer survival in various solid tumors, such as pancreatic, lung, cervical, prostate, breast, and gastric carcinomas 36 . Extensive evidence suggests that NF- κ B signaling is primarily involved in the progression of several cancers, which could deepen our understanding of its role in diseases, particularly lung tumorigenesis. Despite the critical role and potential adverse reactions associated with this pathway, various biologics, macromolecules, and small molecules targeting it are available or in trials, with many repurposed drugs also targeting NF- κ B, and small molecules comprising 76% of US FDA approvals in the last decade 37 . Download figure Open in new tab Figure 1. New interaction links discovered by the proposed framework considering targeting pathways (ADGRF5, TP53, BRAF, KRAS, GNAS), left: mean average ensemble, right: weighted ensemble. The score is annotated from low (blue) to high (red). Download figure Open in new tab Figure 2. Results of contrastive learning using 2,4 and 8-layer NNs with MSE and MAE loss under Adam and AdamW optimizers Download figure Open in new tab Figure 3. Trends in clinical literature of well-known cancer targets from 2014 − 2024 Download figure Open in new tab Figure 4. Trends in clinical literature of the discovered targets from 2014 − 2024. Top: less than 100 papers in total, bottom: more than 100 papers in total. 4 Discussion 4.1 Mutual Exclusivity Analysis We investigate the mutual exclusivity of the discovered pathways using the cBioPortal 38 databases from nine projects 39 – 47 with a total of 73,717 samples. We use six queries reported in Table 1 . Mutual exclusivity (ME) in cancer genetics refers to the phenomenon where specific genetic mutations rarely or never occur together within the same tumor, suggesting they may perform overlapping roles in cancer development. Co-occurrence in cancer genetics refers to the phenomenon where specific genetic mutations frequently occur together within the same tumor, indicating potential collaborative roles in cancer progression. From our analysis, the ME pairs are (BRAF, KRAS) and (BRAF, TP53). All other pairs are co-occurrences supported by relatively-significant statistical evidence ( SupMat-S3 ). 4.2 Clinical Interests of The Discovered Biomarkers Figure 3 displays the number of publications for the genes BRAF, KRAS, MTOR , and TP53 from 2014 to 2024. All four genes show a marked increase in publications over this period, with TP53 leading significantly in the research volume. This indicates a strong and growing research interest in these genes, critical in cancer biology. MTOR , central to cell growth and metabolism, also shows significant publication growth, reflecting its importance in cancer and other metabolic disorders. The upward trend underscores the continued and intensified focus on understanding these genes to develop more effective cancer treatments and personalized medicine approaches. The top panel of Figure 4 shows the number of publications related to several genes ADGRF5, C16orf89, GNB4, GPR137B, GPR161, GPR27 , and PDE1A in the same decade. The general trend for most genes shows increased publications over the ten years. This suggests growing interest and research in these specific genes, which could be attributed to advancements in genetic research techniques, increasing recognition of their relevance in various conditions, or the rise of personalized medicine approaches. For ADGRF5 , the trend shows a steady increase in the number of publications starting from around 2015, with a significant spike in the early 2020s. This might indicate discoveries or the emerging importance of this gene in specific medical research areas. For C16orf89 , it displays a more gradual increase in publications, with a notable rise starting around 2020. This gene’s research may be in its earlier stages of gaining attention compared to others. Publications on GNB4 have steadily increased, with a more noticeable rise from around 2018 onwards. This could reflect a growing understanding of its role in disease mechanisms or therapeutic potential. Both statistics of GPR137B and GPR161 show a consistent upward trend in publications, indicating ongoing and possibly expanding research interest. Although publications for GPR27 started lower, a gradual increase suggests emerging research interest. Similar to GPR27 , there is a steady increase in publications on PDE1A , indicating growing research interest over the years. Some of these genes might be implicated in cancer research. For instance, G-protein coupled receptors (GPCRs) like GPR137B and GPR161 are often studied for their roles in cell signaling and cancer. The bottom panel of Figure 4 illustrates the number of publications related to a set of genes BPTF, CDK8, CDKN1A, CRBN, GNAS, MRAS, NCOA1, NFKBIA, RAMP2, RBPJ, ROCK1, TP53BP2, TP53INP1, VDAC1, WEE1, ZP2 from the same period. A clear upward trend is observed across most genes, particularly for CDKN1A, NFKBIA , and TP53INP1 , reflecting a substantial increase in research interest. This surge can be attributed to advancements in genomic research and the growing importance of these genes in personalized medicine and cancer genetics. Notably, genes like TP53INP1 and CDKN1A , crucial in cell cycle regulation and apoptosis, show significant increases, indicating their critical role in cancer research and therapy development. 5 Conclusion In conclusion, we introduced a contrastive learning framework on graphs that preserves the intrinsic features by graph isomorphism. Our findings revealed novel connections between G-coupled receptors (GPR137B, GPR161, and GPR27) with the GNAS and BRAF pathways, interactions between cyclin-dependent kinase inhibitors (CDKN1A and CDK8) with GNAS and TP53, and the involvement of NFKBIA in a pathway linking all the targeted biomarkers. Future work should validate these interactions in clinical settings, explore the therapeutic potential of targeting NFKBIA, and enhance the proposed framework to incorporate additional biomarkers and pathways. Furthermore, integrating deep sequencing technologies and developing innovative therapies with better efficacy and less drug resistance is crucial for advancing personalized cancer treatment. Supplemental Materials SupMat-S1: Original protein-protein interaction networks integrated from String database. SupMat-S2: Synthesized interaction networks. SupMat-S3: Mutual exclusivity analysis from ten projects on cBioportal. Declarations Funding The research was supported by RF-Bio-N.00001 and RF-AI-N.00007 from G.A.I.A QTech LLC, Vietnam. The author would like to thank Phenikaa university for supports during the project. Conflict of interest The authors declare that they have no conflict of interest. Ethics approval and consent to participate The study was conducted in accordance with ethical standards. Consent for publication All authors consent to the publication of this work. Data availability The data supporting the findings of this study are available upon reasonable request. Code availability The implementation of this project is made available at: https://github.com/namnguyen0510/Contrastive-Learning-for-Graph-Based-Biological-Interaction-Discovery . References 1. ↵ Hadsell , R. , Chopra , S. & LeCun , Y. Dimensionality reduction by learning an invariant mapping . 2006 IEEE Comput. Soc. Conf. on Comput. Vis. Pattern Recognit. (CVPR’06) 2 , 1735 – 1742 ( 2006 ). OpenUrl 2. ↵ Hjelm , R. D. et al. Learning deep representations by mutual information estimation and maximization . Int. Conf. on Learn. Represent . ( 2019 ). 3. Wu , Z. , Xiong , Y. , Yu , S. X. & Lin , D. Unsupervised feature learning via non-parametric instance discrimination . In Proceedings of the IEEE conference on computer vision and pattern recognition , 3733 – 3742 ( 2018 ). 4. van den Oord , A. , Li , Y. & Vinyals , O. Representation learning with contrastive predictive coding . arXiv preprint arxiv: 1807.03748 ( 2018 ). 5. Chen , T. et al. Simple and effective unsupervised speech representation learning with large-scale contrastive learning . In Interspeech , 3780 – 3784 ( 2020 ). 6. He , K. , Fan , H. , Wu , Y. , Xie , S. & Girshick , R. Momentum contrast for unsupervised visual representation learning . Proc. IEEE/CVF Conf. on Comput. Vis. Pattern Recognit. 9729–9738 ( 2020 ). 7. Grill , J.-B. et al. Bootstrap your own latent: A new approach to self-supervised learning . Adv. Neural Inf. Process. Syst . 33 , 21271 – 21284 ( 2020 ). OpenUrl 8. ↵ Nguyen , N. & Chang , J. M. Csnas: Contrastive self-supervised learning neural architecture search via sequential model-based optimization . IEEE Trans. on Artif. Intell . 3 , 609 – 624 ( 2021 ). OpenUrl 9. ↵ Liben-Nowell , D. & Kleinberg , J. The link prediction problem for social networks . J. Am. society for information science technology 58 , 1019 – 1031 ( 2007 ). OpenUrl 10. ↵ Al Hasan , M. , Chaoji , V. , Salem , S. & Zaki , M. J. Link prediction using supervised learning . In SDM06: workshop on link analysis, counter-terrorism and security ( 2006 ). 11. ↵ Menon , A. K. & Elkan , C. Link prediction via matrix factorization . In Proceedings of the 2011 European conference on machine learning and knowledge discovery in databases-Volume Part II , 437 – 452 ( 2011 ). 12. ↵ Veličković , P. et al. Deep graph infomax . In International Conference on Learning Representations ( 2019 ). 13. ↵ You , Y. et al. Graph contrastive learning with augmentations . In Advances in Neural Information Processing Systems ( 2020 ). 14. ↵ Hassani , K. & Khasahmadi , A. H. Contrastive multi-view representation learning on graphs . In International conference on machine learning , 4116 – 4126 ( PMLR , 2020 ). 15. ↵ Sun , K. , Lin , Z. & Zhu , Z. Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes . In Proceedings of the AAAI conference on artificial intelligence , vol. 34 , 5892 – 5899 ( 2020 ). OpenUrl 16. ↵ Jiao , Y. et al. Sub-graph contrast for scalable self-supervised graph representation learning . In 2020 IEEE international conference on data mining (ICDM) , 222 – 231 ( IEEE , 2020 ). 17. ↵ Nguyen , P.-N. Biomarker discovery with quantum neural networks: a case-study in ctla4-activation pathways . BMC bioinformatics 25 , 149 ( 2024 ). OpenUrl 18. ↵ Röthe , J. et al. The adhesion gpcr gpr116/adgrf5 has a dual function in pancreatic islets regulating somatostatin release and islet development . Commun. Biol . 7 , 104 ( 2024 ). OpenUrl 19. ↵ Wu , Y. et al. The adhesion-gpcr adgrf5 fuels breast cancer progression by suppressing the mmp8-mediated antitumorigenic effects . Cell Death & Dis . 15 , 455 ( 2024 ). OpenUrl 20. ↵ Szklarczyk , D. et al. String v11: protein–protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets . Nucleic acids research 47 , D607 – D613 ( 2019 ). OpenUrl CrossRef PubMed 21. ↵ Innamorati , G. et al. The curious case of gαs gain-of-function in neoplasia . BMC cancer 18 , 1 – 15 ( 2018 ). OpenUrl CrossRef PubMed 22. ↵ Garcia-Murillas , I. et al. An sirna screen identifies the gnas locus as a driver in 20q amplified breast cancer . Oncogene 33 , 2478 – 2486 ( 2014 ). OpenUrl CrossRef PubMed 23. ↵ Robbins , H. & Monro , S. A stochastic approximation method . The annals mathematical statistics 400 – 407 ( 1951 ). 24. ↵ Kingma , D. P. & Ba , J. Adam: A method for stochastic optimization . arXiv preprint arxiv: 1412.6980 ( 2014 ). 25. ↵ Loshchilov , I. & Hutter , F. Decoupled weight decay regularization . arXiv preprint arxiv: 1711.05101 ( 2017 ). 26. ↵ Samidurai , A. et al. Role of phosphodiesterase 1 in the pathophysiology of diseases and potential therapeutic opportunities . Pharmacol. & therapeutics 226 , 107858 ( 2021 ). OpenUrl CrossRef 27. ↵ Yin , X. et al. Unveiling the impact of cdk8 on tumor progression: mechanisms and therapeutic strategies . Front. Pharmacol . 15 , 1386929 ( 2024 ). OpenUrl 28. ↵ Tian , T. , Li , X. & Zhang , J. mtor signaling in cancer and mtor inhibitors in solid tumor targeting therapy . Int. journal molecular sciences 20 , 755 ( 2019 ). OpenUrl 29. ↵ Manousakis , E. , Miralles , C. M. , Esquerda , M. G. & Wright , R. H. Cdkn1a/p21 in breast cancer: Part of the problem, or part of the solution? Int. J. Mol. Sci . 24 , 17488 ( 2023 ). OpenUrl 30. ↵ Gan , L. et al. The lysosomal gpcr-like protein gpr137b regulates rag and mtorc1 localization and activity . Nat. cell biology 21 , 614 – 626 ( 2019 ). OpenUrl CrossRef 31. ↵ Pillaiyar , T. et al. Structure-activity relationships of agonists for the orphan g protein-coupled receptor gpr27 . Eur. J. Med. Chem . 225 , 113777 ( 2021 ). OpenUrl 32. ↵ Jacobs , M. D. & Harrison , S. C. Structure of an iκbα/nf-κb complex . Cell 95 , 749 – 758 ( 1998 ). OpenUrl CrossRef PubMed Web of Science 33. ↵ Liu , T. , Zhang , L. , Joo , D. & Sun , S.-C. Nf-κb signaling in inflammation . Signal transduction targeted therapy 2 , 1 – 9 ( 2017 ). OpenUrl 34. ↵ Sun , Z. & Andersson , R. Nf-κb activation and inhibition: a review . Shock 18 , 99 – 106 ( 2002 ). OpenUrl CrossRef PubMed Web of Science 35. ↵ Yamamoto , Y. , Gaynor , R. B. et al. Therapeutic potential of inhibition of the nf-κb pathway in the treatment of inflammation and cancer . The J. clinical investigation 107 , 135 – 142 ( 2001 ). OpenUrl 36. ↵ Rasmi , R. R. , Sakthivel , K. M. & Guruvayoorappan , C. Nf-κb inhibitors in treatment and prevention of lung cancer . Biomed. & Pharmacother . 130 , 110569 ( 2020 ). OpenUrl CrossRef PubMed 37. ↵ Ramadass , V. , Vaiyapuri , T. & Tergaonkar , V. Small molecule nf-κb pathway inhibitors in clinic . Int. journal molecular sciences 21 , 5164 ( 2020 ). OpenUrl 38. ↵ Cerami , E. et al. The cbio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data . Cancer discovery 2 , 401 – 404 ( 2012 ). OpenUrl Abstract / FREE Full Text 39. ↵ Zehir , A. et al. Mutational landscape of metastatic cancer revealed from prospective clinical sequencing of 10,000 patients . Nat. medicine 23 , 703 – 713 ( 2017 ). OpenUrl CrossRef PubMed 40. Robinson , D. R. et al. Integrative clinical genomics of metastatic cancer . Nature 548 , 297 – 303 ( 2017 ). OpenUrl CrossRef PubMed 41. Miao , D. et al. Genomic correlates of response to immune checkpoint blockade in microsatellite-stable solid tumors . Nat. genetics 50 , 1271 – 1281 ( 2018 ). OpenUrl CrossRef PubMed 42. Hyman , D. M. et al. Her kinase inhibition in patients with her2-and her3-mutant cancers . Nature 554 , 189 – 194 ( 2018 ). OpenUrl CrossRef PubMed 43. Samstein , R. M. et al. Tumor mutational load predicts survival after immunotherapy across multiple cancer types . Nat. genetics 51 , 202 – 206 ( 2019 ). OpenUrl CrossRef PubMed 44. Rosen , E. Y. et al. Trk fusions are enriched in cancers with uncommon histologies and the absence of canonical driver mutationslandscape and outcome of trk fusion–positive cancers . Clin. Cancer Res . 26 , 1624 – 1632 ( 2020 ). OpenUrl Abstract / FREE Full Text 45. Bolton , K. L. et al. Cancer therapy shapes the fitness landscape of clonal hematopoiesis . Nat. genetics 52 , 1219 – 1226 ( 2020 ). OpenUrl CrossRef PubMed 46. Wu , L. et al. Landscape of somatic alterations in large-scale solid tumors from an asian population . Nat. communications 13 , 1 – 11 ( 2022 ). OpenUrl CrossRef 47. ↵ Nguyen , B. et al. Genomic characterization of metastatic patterns from prospective clinical sequencing of 25,000 patients . Cell 185 , 563 – 575 ( 2022 ). OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted July 24, 2024. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways Phuong-Nam Nguyen bioRxiv 2024.07.23.604746; doi: https://doi.org/10.1101/2024.07.23.604746 Share This Article: Copy Citation Tools Contrastive Learning for Graph-Based Biological Interaction Discovery: Insights from Oncologic Pathways Phuong-Nam Nguyen bioRxiv 2024.07.23.604746; doi: https://doi.org/10.1101/2024.07.23.604746 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioinformatics Subject Areas All Articles Animal Behavior and Cognition (7651) Biochemistry (17746) Bioengineering (13928) Bioinformatics (42066) Biophysics (21499) Cancer Biology (18650) Cell Biology (25579) Clinical Trials (138) Developmental Biology (13409) Ecology (19947) Epidemiology (2067) Evolutionary Biology (24374) Genetics (15633) Genomics (22557) Immunology (17775) Microbiology (40505) Molecular Biology (17217) Neuroscience (88796) Paleontology (667) Pathology (2845) Pharmacology and Toxicology (4836) Physiology (7664) Plant Biology (15179) Scientific Communication and Education (2047) Synthetic Biology (4304) Systems Biology (9839) Zoology (2272)