Discovering anticancer drug target combinations via network-informed signaling-based approach

Abstract

Background

Oncologists deciding on cancer treatments must make difficult decisions as to which prescription and implementation strategies would best suit each patient. Much is still unknown about combinations of prescription drugs as there are many to choose from. At the outset, the oncologist reckons with at least two established facts: (i) patients receiving successive single molecules treatments are likely to experience drug resistance, and (ii), to select optimal drug combinations requires to pick the ‘best’ protein drug target combinations. Intuitively, target selection should precede drug selection, implying that well-informed strategies would opt to first consider drug targets — not drugs — combinations. Nowadays, drug combinations that oncologists consider are empirical and limited. They are restricted primarily by observations and praxis, that is, scant clinical experience with their application.

Methods

Here we develop a strategy for selecting optimal drug target combinations following nature. We use protein-protein interaction networks and shortest paths to discover communication pathways in cells based on interaction network topology. Our strategy mimics cancer signaling in drug resistance, which commonly harnesses pathways parallel to those blocked by drugs, thereby bypassing them.

Results

We select key communication nodes as combination drug targets inferred from topological features of networks. We test our network-informed signaling-based approach to discover anticancer drug target combinations on available clinical data, patient-derived breast and colorectal cancers. Alpelisib + LJM716 and alpelisib + cetuximab + encorafenib combinations diminish tumors in breast and colorectal cancers, respectively.

Conclusions

Our network-based approach discovers optimal protein co-target combinations to counter resistance, selecting co-targets from alternative pathways and their connectors.

Plain Language Summary

Cancer treatment often involves drug combinations, but choosing the right ones is difficult. Oncologists know that using one drug at a time can result in resistance. However, current drug combinations are mostly based on limited clinical experience. This study suggests that choosing the right targets for treatment is key to success and introduces a new method that helps identify the best combinations, based on analysis of cancer cells’ adaptation to treatment. We tested our method using real data from breast and colorectal cancer patients. Our strategy can help oncologists design smarter drug combinations to overcome resistance.

Yavuz et al. present a network-based strategy to identify optimal protein co-targets for cancer therapy by mimicking resistance mechanisms. Their approach uncovers effective drug target combinations that can be used in clinics across alternative signaling pathways in breast and colorectal cancer models.

Introduction

Precision oncology is increasingly transitioning towards drug combinations^1–4. However, determining the optimal combinations is challenging. Single drug targeting a single protein (monotherapy) frequently succumbs because of inherent or emergent resistance^5–7. The critical challenge in cancer treatment lies in completely eradicating tumor cells before they can develop and propagate resistant mutations^3,8. Sequential single-drug regimens are unlikely to overcome this hurdle, as they merely repeat the cycle of resistance development. In contrast, by simultaneously administering multiple therapeutic agents, oncologists can dramatically reduce the probability of individual cancer cells developing simultaneous resistance to multiple drugs^2,3. The multi-targeted intervention essentially creates a more formidable therapeutic barrier against the cancer’s adaptive potential.

Targeting pathways with small-molecule inhibitors has proven a compelling strategy. Over 80 drugs targeting kinases are FDA-approved, several approved in 2023⁹. However, cancer cells often develop resistance through alternative growth pathways, with pre-existing drug-resistant subclones proliferating rapidly. One example is PI3K/AKT/mTOR, with around 30-40% of breast cancer cases featuring mutations in PIK3CA^10–15. The combination of alpelisib, a pan-PI3K inhibitor^9,16,17, with hormone therapy has demonstrated effectiveness in metastatic breast cancers that are hormone receptor-positive and human epidermal growth factor (HER2) negative^14,18,19. Trastuzumab targets patients with HER2-positive breast cancer; however, they may relapse due to activating mutations in PIK3CA in 20–30% of the cases. Co-administration of trastuzumab and inhibitors targeting the PI3K/AKT/mTOR signaling axis, including PIK3CA, enhances response^20,21. In hepatocellular carcinoma, dual inhibition of mTOR and SHP2 shows promising synergistic effects, preventing Receptor Tyrosine Kinase (RTK)-mediated resistance to mTOR inhibition²². Understanding resistance mechanisms to drugs like RAF inhibitors is critical for developing more effective RAS/RAF/MEK/ERK signaling inhibitors^23,24. Using machine learning and experimental validation, researchers identified the long non-coding RNA RERE-AS1 as a key biomarker that enhances breast cancer patients’ response to ribociclib by suppressing tumor growth through MEK/ERK pathway regulation²⁵.

Methods proposed to optimize combinations include a Graph Convolutional Network (GCN)²⁶. Prioritization was also based on semantic relationships between drug and disease²⁷, and on pathway crosstalk²⁸. Large-scale dual-drug combinations in breast, colorectal, and pancreatic cancer cell lines, revealed the rarity and highly contingent nature of drug synergy⁴. Combinations of FDA-approved cancer drugs, tested against the NCI-60 cohort²⁹ were deposited at the NCI-ALMANAC repository. In vivo investigations of murine xenograft models validated the superiority of these drug pairings relative to monotherapy. NCI-ComboMATCH augmented the efficacy by genomically directed combinatorial interventions³⁰. AstraZeneca’s extensive drug combination dataset from a DREAM Challenge demonstrates promising computational strategies for predicting synergistic cancer drug pairs while also revealing specific genomic interactions that explain drug synergy ³¹. A recent study ³² on computational drug repurposing reviewed approaches, evaluated in silico resources, and presented case studies demonstrating practical applications. REMEDi4ALL showcased phenotypic data of synergistic drug combinations with enhanced efficacy and safety ³². Combi-Seq uses DNA barcoding to efficiently screen hundreds of drug combinations in microfluidic droplets, providing insights into both cell viability and cellular signaling pathways in response to drug combinations³³.

While firsthand knowledge and experience can point to a successful combination, more inclusive successful cancer treatments can be expected to benefit from systematically addressing drug-resistant clones and activating mutations that culminate in offering a learned choice. To identify the mutations, we compile co-existing, tissue-specific mutations, in the same and in different pathways³⁴. Under the premise that not only the protein pairs harboring co-existing mutations play a critical role in amplifying oncogenic signaling but also proteins on paths connecting them, we construct protein-pair specific subnetworks and identified proteins that serve as bridges between them. The oncogenic subset is comprised of RTKs and transcription factors (TFs), including EGFR, ERBB2, ERBB2, and MYC. Co-targeting ESR1/PIK3CA and BRAF/PIK3CA are comparable to patient-derived xenografts (PDXs), responding better to combination drugs. Co-targeting the ESR1/PIK3CA subnetwork pathways, a marker of breast cancer metastasis, with alpelisib-LJM716 combination diminishes the tumor. Co-targeting BRAF/PIK3CA in colorectal cancer with alpelisib, cetuximab, and encorafenib (PIK3CA, EGFR, and BRAF inhibitors, respectively) reveals context-dependent tumor growth inhibition in xenografts, with efficacy modulated by protein subnetwork mutation and expression profiles, suggesting that learning from, and being guided by, nature could be an informed approach to discovery. Here, we refer to protein products of genes (in italics), following the source of our data, TCGA and AACR GENIE.

Methods

Data collection and preprocessing

Somatic mutation profiles were obtained from publicly available large-scale cancer genomics resources, including TCGA³⁵ and AACR Project GENIE³⁶. Standard preprocessing steps were applied, including removal of low-confidence variants with low variant allele frequency and potential germline events, and prioritization of primary tumor samples where multiple tumor records existed.

To identify significant co-existing mutations, we considered mutations present in multiple non-hypermutated tumors and generated pairwise combinations across different proteins. Statistical significance of co-occurrence was assessed using Fisher’s Exact Test, followed by multiple testing correction. Mutation pairs meeting significance thresholds and frequency criteria were retained for downstream analyses and classified as drivers or passengers based on established cancer mutation catalogs.

Protein–protein interaction data were integrated from the HIPPIE³⁷ database, retaining high-confidence interactions after filtering. For pathway analysis, we used the KEGG_2019_Human dataset³⁸, focusing on curated signaling pathways.

All datasets were used in their publicly released processed formats, and basic quality control, filtering, and standardization were applied consistently.

The further details of data collection and preprocessing protocols are described in the Methods sections of references^34,39. All relevant datasets and resources used in this study are listed in Supplementary Tables 1-2, along with their corresponding references. This study did not require institutional review board (IRB) or ethics committee approval because it used only publicly available, de-identified patient data.

Calculation of the shortest paths between protein pairs

We used 3424 different gene double mutations obtained with the above methodology and deposited in our recent study ³⁴. 46 of these doublets are metastatic breast cancer markers and accumulated on 16 protein pairs. We calculated shortest paths by using PathLinker^40–42. PathLinker is a graph-theoretic algorithm designed for reconstructing interactions that efficiently identifies multiple short paths connecting any set of sources to any set of targets within a protein-protein interaction (PPI) network. We downloaded the PathLinker algorithm from https://github.com/Murali-group/PathLinker.

We used the first component of each protein pair as source and the second as the target node with the parameter k = 200 to the PathLinker algorithm to compute the k shortest simple paths between a source and target node⁴². A simple path is a path with no repeated nodes. PathLinker computed 200 simple shortest paths for protein pairs harboring co-existing mutations except for the 21 pairs where at least one of the constituents is not found in the HIPPIE PPI network. The lengths of the shortest paths from source to target node varies from one to five. The corresponding data is available in Supplementary Data 2.

To evaluate the robustness of the default k = 200 parameter in PathLinker, subnetworks were generated using k = 200, k = 300, and k = 400, and compared. Jaccard similarity coefficients were calculated for the node sets on the shortest paths between these k values, with the mean Jaccard index ranging from 0.72 to 0.74, indicating strong overlap between k = 200 and the other subnetworks. Pathway enrichment analysis was performed using the Enrichr⁴³ tool (KEGG 2019 Human library), and 28 of the top 30 significantly enriched pathways (FDR < 0.05) were shared across k = 200, k = 300, and k = 400 subnetworks (Jaccard similarity = 0.88). These pathways included key signaling pathways such as MAPK, PI3K/AKT, and apoptosis, suggesting biological consistency across k values.

Further examination showed that increasing k beyond 200 substantially increased computational cost due to the larger number of paths and nodes included. Although these larger subnetworks were structurally more complex, they did not offer proportional gains in biological insight. Consequently, k = 200 was retained as the default parameter for all primary analyses.

We selected the Human Integrated Protein-Protein Interaction rEference (HIPPIE) as our primary interactome due to its stringent confidence scoring, high reliability, and compatibility with edge-weighted reconstruction algorithms³⁷. HIPPIE’s scoring system, based on experimental PPI data, has been shown to perform best in minimizing false positives³⁷. While HIPPIE offers clear advantages for reconstructing signaling pathways, we acknowledge that the choice of PPI network can influence both the topology of the reconstructed subnetwork and algorithmic outcomes. For example, BioGRID⁴⁴, with its dense coverage derived from manual curation, may increase recall but also introduce more false positives due to the absence of a unified confidence scoring system. STRING⁴⁵, depending on its configuration, can vary widely in coverage and bias; filtered versions reduce literature bias, while unfiltered networks—encompassing co-expression and text-mining data—may increase noise. As shown in prior benchmarking studies⁴⁶, network reconstruction algorithms are sensitive to such structural differences, including node degree distribution, edge density, and biases toward well-studied proteins.

Pathway enrichment analysis

After computing 200 shortest paths for co-mutated protein pairs, we lumped all the nodes on the shortest paths together for each pair and dubbed such proteins as ‘shortest path proteins.’ The number of nodes forming the shortest paths varies from 81 to 166.

Then by using these shortest path proteins, we conducted gene set enrichment analysis by GSEApy (https://github.com/zqfang/GSEApy) to find out prominent pathways through which the signal flows between two nodes. GSEApy represents an implementation of Gene Set Enrichment Analysis (GSEA) in Python and Rust, and functions as a wrapper for the Enrichr tool. GSEA is a computational framework that assesses whether a predefined gene set demonstrates statistically significant and concordant alterations between two distinct biological states. We used the parameter gene_sets = ‘KEGG_2019_Human’. In the enrichment results, we only focused on the “signaling pathways.”

After obtaining pathway enrichment results from GSEApy, we kept the significant signaling pathways in which the shortest path proteins belong to. There are 1248 pairs whose shortest paths are enriched in at least one signaling pathway. We also calculated the fraction of shortest path proteins in each signaling pathway among all shortest path nodes. This would give us the signaling pathway(s) that the majority of the shortest path proteins belong to. So, for each pair, we selected the signaling pathway(s) that contain more than 20% of the nodes in the shortest paths. For the pairs with the shortest path proteins passing through more than one signaling pathway, we labelled the pathways as similar if the Jaccard Index was 0.2; otherwise, we labeled them as different pathways.

Subsequent pathway enrichment analysis of the aggregated shortest path proteins revealed 170 protein pairs; each linked to at least one pathway in which the shortest path proteins showed significant enrichment (Supplementary Data 1, 2). For each of these protein pairs, we curated a seed gene set by selecting the shortest path proteins associated with the enriched pathways. These seed gene sets were then utilized in the PageRank algorithm to facilitate subnetwork reconstruction.

Protein pair specific subnetwork construction from shortest path proteins as seeds in enriched pathways

We applied the PageRank algorithm⁴⁷ from the NetworkX implementation (https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.link_analysis.pagerank_alg.pagerank.html) as follows: Upon identifying the pathways with which over 20% of the shortest path proteins for each protein pair are linked, we employed these genes and their corresponding protein pair components (excluding those in enriched pathways) as initial seeds for the PageRank algorithm. The initial scores were assigned as 2 for the protein pair components and 1 for the remaining seed genes, with alpha set to 0.90. To select the subnetwork genes, we established the threshold as the minimum (M) of the scores of Gene 1 and Gene 2 (M = minimum[pagerank_score[Gene1],pagerank_score[Gene2]]). For the protein pairs with a total number of selected subnetwork genes is less than or equal to 50, we used a less stringent threshold to 0.2*M not to lose interactions in the PPI network. The preprocessed HIPPIE network, devoid of self-edges, was utilized for this process.

The general formula for the PageRank algorithm is as follows:

$$PageRank\left(x_i\right)=\frac{\left(1-d\right)}{N}+d{\sum }_{x_j\in G\left(x_i\right)}\frac{PageRank\left(x_j\right)}{L\left(x_j\right)}$$ 1

where $x_i$ is a webpage and $x_j$ is a webpage with an outgoing link to $x_i$.

The PageRank algorithm ranks web pages by evaluating the quantity and quality of their inbound and outbound hyperlinks^47,48. It employs a dampening factor, denoted as ‘d’, which ranges between 0 and 1 (empirically calibrated to 0.90 in this study). To assess the robustness of our network analysis to parameter selection, we evaluated the sensitivity of gene rankings to the PageRank damping factor α. For each gene pair, we computed subnetworks using α values in (0.60, 0.70,0.80, 0.85, 0.90, 0.95) and calculated Spearman rank correlations (two-sided) between PageRank scores across all α combinations. These correlations were consistently high (minimum = 0.976, maximum = 0.999, median=0.996), indicating that top-ranked genes remained remarkably stable across the tested α range. Given this stability, we selected α = 0.90 as an optimal value that balances subnetwork size with ranking consistency. This factor models the likelihood that a hypothetical user, after navigating through a series of hyperlinks, will randomly access an unlinked page. The algorithm incorporates the cardinality of outbound links from each node, denoted as ‘L’. Essentially, the PageRank algorithm processes an initial set of nodes and produces a ranked list of other nodes in the network, ordered by their relevance to the input set. It uses a recursive computation that spreads importance scores through the network, adjusted by a damping factor and normalized by each node’s out-degree.

Betweenness and eigenvector centrality metrics

Betweenness centrality of a node v is the number of shortest paths passing through that node which can be calculated by the following formula:

$$BetweennessCentrality\left(v\right)={\sum }_{s\ne v\ne t}\frac{{\sigma }_{st}\left(v\right)}{{\sigma }_{st}}$$ 2

where ${\sigma }_{st}$ is the total number of shortest paths connecting node s to node t and ${\sigma }_{st}\left(v\right)$ is the total number of shortest paths connecting node $s$ to node $t$ passing through $v$ (https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.centrality.betweenness_centrality.html)⁴⁹.

To calculate eigenvector centrality, we use the following formula:

$$\lambda v_i={\sum }_{j\to i}{v}_j$$ 3

where $\lambda$ is the eigenvalue of maximum modulus that is positive, $v_i$ is the corresponding node (https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.centrality.eigenvector_centrality.html).

Hierarchical clustering of the connector proteins

We performed hierarchical clustering on a group of genes selected by betweenness centrality metrics. To perform hierarchical clustering, first, we divided protein pairs into three groups according to the number of their enriched signaling pathways of the shortest path proteins: the shortest path proteins are enriched in one pathway, two pathways, and three or more (multiple) pathways. We constructed a classified dataset with proteins as the indices and the signaling pathways as the column names. A cell will have a value of 1 if the corresponding gene is a member of the signaling pathway in the corresponding column, 0 otherwise. Then, for the protein pairs, we identified the proteins that have betweenness centrality greater than Q3 of their corresponding subnetwork. These are dubbed as the “connector proteins” of the corresponding subnetwork. For each group formed according to pathway enrichment results of the shortest path proteins, we lumped together the connector proteins and then got the intersection of these connector proteins in the three groups. The resulting set of connector proteins in the intersection of the three groups is the one that is important in cellular network communication.

Then, we created a dendrogram illustrating the relationships between the proteins by evaluating their shared pathways, employing the x-axis to depict the Hamming distance metric that quantifies dissimilarity between clusters; smaller distances indicate greater similarity. Greater distances on the dendrogram indicate infrequent co-occurrence of genes within the same signaling pathway, reflecting lower connectivity and functional association. Genes with a hamming distance <0.1 can be considered as having similar functions as their corresponding pathways overlap.

Initially, we obtained a catalogue of 46 signaling pathways sourced from KEGG, encompassing a total of 2197 genes distributed among these pathways. Upon analysis of gene-pathway associations, we found genes affiliated with anywhere from one to twelve pathways. Subsequently, we constructed a binary matrix sized at 2197 × 46, where gene names served as row indices and the 46 pathways as columns. Each cell in the matrix was assigned a value of 1 if the corresponding gene belonged to the pathway specified by the column, and 0 otherwise. Then we calculated the betweenness centrality of the nodes (Eq. 2) in each gene-pair specific subnetwork and selected the ones with centrality score >Q3 as connector proteins and clustered them with hierarchical clustering with ‘complete’ parameter.

The Hamming distance, a fundamental metric in information theory and coding theory, serves as a quantitative measure for comparing two binary data strings of equivalent length⁵⁰. This distance is defined as the cardinality of the set of positions at which the corresponding symbols in two strings are discordant. In the context of our genomic analysis, each gene can be conceptualized as a binary vector of dimension 46 (corresponding to the number of signaling pathways under consideration), where each element is assigned a value of unity if the gene is a constituent of the pathway represented by that dimension, and zero otherwise.

The Hamming distance between two vectors, denoted as H(a,b) where a and b are the vectors in question, can be computed through the application of the XOR operation, symbolized as a ⊕ b. The result vector from this operation is then subjected to a summation of its elements, which is equivalent to enumerating the non-zero elements. This process can be formally expressed as:

$$H\left(a,b\right)={\sum }_{i=1}^n{a}_i\oplus b_i$$ 4

The result of the Hamming distance will be the sum of different elements in a series of elements over which the distance is compared, where n represents the dimensionality of the vectors (in our case, 46).

This metric provides a robust means of quantifying the dissimilarity between genes based on their pathway affiliations, allowing for the construction of a distance matrix that serves as the foundation for subsequent hierarchical clustering analyses.

Drug and drug target information

We obtained the data for drugs and their target proteins from the Connectivity Map (CMAP)⁵¹.

Selection of PDXs for pre-clinical validation

We get the list of PDX models⁵² that have at least one mutated target protein, where these proteins are also among the shortest path proteins of the corresponding protein pair. We enlisted the seed genes set, including source and target nodes; the PDXs should contain at least two drug targets among the seed genes, including the source and target nodes. Then, we computed the set of PDX models that have at least two mutations in seed genes for a corresponding given protein pair and also at least one mutation on the protein pair. We also evaluated the volume growth information of the untreated tumor, with single treatment and combination therapy.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Overview of the data and workflow

Our aim is to identify proteins to co-target with combination drugs that intervene in cancer cell proliferation. Optimal target proteins may belong to redundant, parallel, or compensatory pathways acting in cellular feedback mechanisms maintaining homeostasis (Fig. 1a, b)^3,53–55.

Fig. 1. Illustration of the rationale and conceptual model of inter-pathway communication.

a Targeted cancer therapies are frequently thwarted by tumors’ adaptive mechanisms, which enable cellular survival through alternative signaling pathways. By mapping intricate molecular interactions and understanding how cancer cells circumvent initial therapeutic blockades, we propose a strategic approach to combination drug design that anticipates and neutralizes potential resistance mechanisms. This methodology aims to discover optimal proteins to co-target in signaling pathways (same, redundant or parallel) to tackle therapy resistance emerging in monotherapies. Our pipeline transforms the current empirical approach to drug combination selection into a predictive, network-informed strategy that can proactively identify therapeutic interventions capable of simultaneously disrupting multiple cancer cell survival routes. b Parallel pathways operate independently but converge on similar phenotypic outcomes through separate molecular mechanisms. Redundant pathways exhibit functional overlap, allowing cellular processes to persist despite inhibition of one pathway. Compensatory pathways are activated in response to inhibition of a primary pathway, enabling adaptive resistance. Understanding these mechanistic distinctions informs rational combination therapy design.

Redundant pathways involve functionally overlapping mechanisms where multiple pathways perform similar roles, allowing cellular functions to persist despite the inhibition of one. These often involve homologous proteins or members of protein families, for instance, ERK1/2 and ERK5. PI3K and AKT are same-pathway inhibition, making single-pathway targeting potentially insufficient. Parallel pathways are independent signaling routes that converge on similar phenotypic outcomes but operate via distinct molecular cascades. They become activated in response to the inhibition of a primary pathway or an altered environment. For example, EGFR blockade can trigger activation of MET or IGF1R, preserving signaling through e.g., MAPK and/or PI3K/AKT/mTOR. MAPK and PI3K pathways drive proliferation but through different mechanisms, providing an example of compensatory pathways. Both pathways are required for the cell cycle⁵⁶.

These distinctions highlight the resilience of cancer signaling networks and underscore the rationale for combination drugs. Targeting multiple components across these relationships can block escape routes and prevent network rewiring. Effective implementation requires mapping of tumor-specific pathway dependencies shaped by lineage, genomics, and the microenvironment. Ideally, all potential drug combinations—double, triple, multiple—should be experimentally evaluated against diverse tumor genomes. However, this is impractical due to their sheer number. Computational methods can significantly narrow the scope of candidates to a manageable scale.

To determine optimal combinations of drug targets, we delineate mediators that orchestrate intracellular signaling cascades. For this, we used protein pairs harboring tissue-specific co-existing mutations, taken from our recent pan-cancer genomes compilation (details are provided in Supplementary Information and tables of³⁴, including the co-existing mutations). We call a protein pair A | B if there is a co-existing mutation pair where the component mutations are on proteins A and B. We construct subnetworks—smaller, interconnected subsets of proteins within a larger protein-protein interaction network—derived from proteins on the shortest paths to discern connector proteins. We define “connector proteins” as communication bridges in the cellular protein network. In our definition, their interactors are on the shortest paths to facilitate efficient signal transduction, essential for rapid cellular responses^57,58. Combination drugs targeting them can block oncogenic signals in drug resistance. To engineer the oncogenic signaling landscape, we integrate upstream and downstream regulators in the shortest paths in the subnetworks where the shortest paths connect protein pairs with co-existing mutations. Figure 2 illustrates our pipeline, and the Methods section describes the compilation of co-existing mutations³⁴, the computation of the shortest paths in the integrated HIPPIE protein-protein interaction (PPI) network³⁷ by deploying the Path Linker algorithm⁴², the pathway enrichment analysis using Enrichr⁴³ on a KEGG 2019-derived dataset³⁸, and the gene-pair specific subnetworks using the PageRank algorithm⁴⁷. We then computationally validate our results using patient-derived xenograft models⁵² for breast cancer and colorectal cancer-specific subnetworks. The supplemental materials provide further details of our calculations and the data, including the shortest paths for 170 protein pairs (Supplementary Data 1) harboring co-existing mutations³⁴, and both components have pathway information, and the pathway enrichment analysis results (Supplementary Data 1-3).

Fig. 2. Methodological pipeline to identify critical proteins mediating cancer-promoting communication.

Our algorithm starts with selecting protein pairs harboring co-existing mutations such as ESR1/PIK3CA, BRAF/PIK3CA, etc. Then for each protein pair, we compute the shortest paths connecting them in the HIPPIE³⁷ protein-protein interaction (PPI) network using the Path Linker⁴² algorithm. We aggregate proteins on the shortest paths, and here, conduct pathway enrichment analysis among 46 signaling pathways from KEGG³⁸. We follow with pathway enrichment analysis with Enrichr⁴³ focusing on signaling pathways, highlighting those where at least 20% of the shortest path proteins are involved. Using the obtained shortest path proteins in these enriched signaling pathways as seeds for the PageRank⁴⁷ algorithm, we construct protein-pair specific subnetworks. For each such subnetwork, we calculate subnetwork centrality scores, identifying those with high betweenness centrality as key connector proteins (bridges). We propose the highest-scoring nodes as potential targets for combination drugs, as they are critical in mediating cancer-promoting cellular communication.

For each of the 170 co-mutated proteins (Supplementary Data 1) with both components belonging to at least one signaling pathway, we compose a seed protein set by selecting the proteins on the shortest paths in the enriched pathway(s) to feed into the Page Rank algorithm to reconstruct subnetworks (Methods). We then calculate and select co-targeting candidate connector proteins (Supplementary Data 4). Subnetworks constructed from protein pairs in the ‘Single’ pathway category exhibit greater distances compared to ‘Double’ and ‘Multiple’ pathway categories (Supplementary Fig. 1). This suggests that proteins in a single pathway are more isolated, connected primarily via a few upstream/downstream proteins, whereas genes in double and multiple pathways exhibit closer proximities, indicating potential crosstalk. In the ensuing sections, we demonstrate the application of our analytical pipeline to the ESR1/PIK3CA and BRAF/PIK3CA protein pairs, specific to breast and colorectal cancer, respectively.

Supplementary Fig. 2 presents protein pairs where the shortest path proteins traverse two and three signaling pathways, respectively. In Supplementary Fig. 2a, protein pairs in the PI3K/AKT pathway combine with either MAPK, Hippo, ErbB, or the Chemokine pathways. The shortest path proteins connecting metastatic markers³⁴ are predominantly enriched in PI3K/AKT and MAPK pathways (Fig. 3), except for ESR1/PIK3CA, enriched in PI3K/AKT and Chemokine pathways. We identified 91 protein pairs where more than 20% of the shortest path proteins exclusively traverse a single pathway. A bubble plot (Supplementary Fig. 3) visually represents 61 protein pairs, emphasizing those where at least one component is a drug target according to the Connectivity Map (CMAP) data⁵¹. Remarkably, a substantial portion of the shortest path proteins associated with these pairs navigate through the PI3K/AKT pathway, followed prominently by the MAPK, and thyroid pathways. This underscores the pivotal role of the PI3K/AKT pathway in oncogenic signaling, positioning it as a prominent reservoir for identifying multiple drug targets.

Fig. 3. Schematic representation of MAPK and PI3K/AKT signaling pathways.

To clarify the main signaling mechanisms in breast and colorectal cancers, we begin with a simple illustration of MAPK and PI3K/AKT pathways featuring ErbB3 and PI3Kα co-targeted with LJM716 and alpelisib in breast cancer. The combination therapy can block cell growth and transcription of target genes. RTKs such as ErbB3, EGFR, ErbB2, or ErbB4 on the cell surface, along with the downstream nodes, depict the signaling cascade. Estrogen receptor (ERα) is shown. PI3Kα phosphorylates signaling lipid PIP₂ to PIP₃. The ErbB family of RTKs includes EGFR (ErbB1), ErbB2 (Her2), ErbB3, and ErbB4. ErbB2 and ErbB3 frequently form heterodimers with potent signaling capabilities through the PI3K/AKT pathway. For colorectal cancer, the combination therapy targeting EGFR, RAF, and PI3K with cetuximab + encorafenib + alpelisib is shown. In general, if oncogenic signaling is elevated in these pathways through mutations or overexpression, an upstream RTK (EGFR, ErbB2, ErbB3, ErbB4) can be combined with downstream oncogenes RAS, RAF, MEK, or ERK and PI3K, AKT from the compensatory pathways MAPK and PI3K/AKT, respectively.

Deciphering upstream and downstream interactions within pathways, including crosstalks, necessitates discerning their functional similarity in terms of their constituent proteins. We computed the similarity using the Jaccard index (Supplementary Fig. 4). High pathway protein similarity suggests the potential for targeting drug interventions upstream and downstream (in the same pathway or in redundant pathways), while dissimilar pathways highlight proteins relating to inter-pathway (parallel or compensatory pathways) crosstalk^3,59. In the context of cellular pathways in identical cell types, pathways sharing identical proteins or proteins from analogous families are termed ‘redundant,’ whereas those utilizing distinct proteins are classified as ‘parallel.’ Compensatory pathways are the ones that are in crosstalk³.

Understanding how cells communicate is key to identifying the critical drug targets in oncogenic signaling networks. Combination drugs guided by signaling pathways have the potential to inhibit proliferative signaling cascades and detect feedback mechanisms through connector proteins in subnetworks. By examining the pathways through which signals flow and evaluating their similarities, potential drug target combinations can be inferred. We dub important connector proteins “co-targeting candidates”.

Connector proteins as co-targeting candidates

We hierarchically clustered a group of proteins selected as co-targeting candidates due to their pivotal roles identified by betweenness centrality metrics. Betweenness centrality measures the degree to which a node (protein) functions as a bridge in a network. Clustering organized these proteins based on their signaling pathways. This clustering approach facilitates the selection of targets from each cluster for combination drugs, capitalizing on synergistic interactions inherent to these protein networks. Targeting proteins central to essential communication pathways enhances the efficacy of combination drugs compared to single agents^2–4,60.

Figure 4a presents a dendrogram of co-targeting candidates, using the y-axis to represent Hamming distance—a metric quantifying dissimilarity between clusters. Smaller distances on this dendrogram indicate greater similarity between gene clusters. A maximum distance of 0.8 provides insights into cluster divergence, suggesting a balance between intra-cluster homogeneity. Larger distances reveal that genes are less likely to co-occur within the same signaling pathway, indicating lower connectivity and functional association. Such separation suggests these genes may participate in distinct biological pathways, highlighting their diverse cellular roles.

Fig. 4. Selection of co-targeting candidate nodes from the reconstructed subnetworks.

a The dendrogram delineates co-targeting candidate nodes. The y-axis denotes Hamming distance metric that quantifies the dissimilarity between clusters of proteins based on the signaling pathways they belong to. It measures the number of corresponding nodes in the pathways which are different in the two clusters. Smaller distances indicate higher similarity. Larger distances imply lower similarity. Lower similarity between pathways implies infrequent co-occurrence of genes, suggesting distinct biological pathways or networks in cellular processes. Genes with a Hamming distance of less than 0.1 are deemed functionally similar due to overlapping pathways. PIK3CA/B and PIK3R1/2/3, with a distance of zero, belong to the same cluster and are in close proximity to AKT1. This cluster is connected to the KRAS and RAF cluster, which is also near the MAPK1 and MAP2K1 cluster. Genes positioned farther apart are less likely to share interactions and functions. For the oncogenes there are seven RTK co-targeting candidates including EGFR, ERBB2, ERBB3, FGFR2, FGFR3, IGF1R and NTRK1. As to the TFs among OGs, CTNNB1, ESR1, MDM2, MYC and STAT3 are the co-targeting candidates. We see that PIK3CA/B, PIK3R1/2/3 (distance 0) belong to the same cluster and this cluster is in proximity to AKT1. It connects the clusters of KRAS and RAF which are also proximal to the cluster composed of MAPK1 and MAP2K1. b The co-targeting node candidates from (a) are categorized as oncogenes and tumor suppressor genes in boxes. The red and blue fonts are transcription factors (TFs) and the receptor tyrosine kinases (RTKs), respectively. The dataset encompasses 53 co-targeting candidates, including 27 oncogenes and 11 tumor suppressor genes. Among the oncogenes, seven RTKs are identified as co-targeting candidates: EGFR, ERBB2, ERBB3, FGFR2, FGFR3, IGF1R, and NTRK1. TFs within the oncogenes include CTNNB1, ESR1, MDM2, MYC, and STAT3.

Proteins exhibiting a Hamming distance less than 0.1 are considered to have similar functions, reflecting substantial pathway overlap. Figure 4b categorizes these co-targeting candidates into oncogenes (OGs), tumor suppressor genes (TSGs), and their associated TFs. The comprehensive dataset comprises 53 co-targeting candidates, including 27 OGs and 11 TSGs. Proteins positioned farther apart in the network demonstrate reduced likelihood of sharing regulatory or interaction networks, thereby highlighting their distinct functional contexts. Strategically combining proteins from different clusters could potentially enhance therapeutic approaches by broadening treatment efficacy and targeting multiple cellular mechanisms.

Within the PI3K/AKT pathway, several oncogene co-targeting candidates play critical roles. PIK3CA and PIK3CB encode the p110α and p110β catalytic subunits of PI3K, respectively, which are upstream activators of AKT1, a key effector in the PI3K/AKT signaling cascade⁶¹. This pathway regulates cellular processes such as growth, metabolism, and survival. Additionally, GSK3B, a downstream target of AKT, is involved in glycogen metabolism and cell survival, while CDK4, regulated by PI3K/AKT signaling, plays a crucial role in cell cycle progression. PTPN11 (SHP2) also interacts with PI3K signaling, further linking RTK activation to downstream effectors in the PI3K/AKT pathway (Fig. 3)⁶². Another pathway that features prominently among the oncogenes identified in the co-targeting set is MAPK. KRAS, BRAF, ARAF, and RAF1 are key upstream regulators in this pathway. KRAS acts as an upstream regulator of RAF kinases, including BRAF and RAF1, which subsequently activate MEK1 (MAP2K1). MEK1 phosphorylates and activates ERK (MAPK1), culminating in the regulation of cellular proliferation and differentiation^63,64. MAPK is crucial for transmitting signals from the cell surface to the nucleus, transducing gene expression for cell fate decisions (Fig. 3). In the JAK/STAT pathway, JAK2 and JAK3 encode non-receptor tyrosine kinases involved in cytokine signaling and STAT activation. STAT3, a transcription factor activated by JAK kinases, plays an important role in promoting cell growth and survival, and its constitutive activation is associated with malignancies^28,65. Additionally, several oncogenes are involved in other pathways that interface with RTK signaling. SRC and FYN are members of the Src family of tyrosine kinases, which are involved in various signaling processes, including those related to focal adhesion and RTK signaling⁶⁶.

We next searched the Gene Ontology (GO) biological processes, seeking co-targeting candidates enriched in the Molecular Signatures Database (MSigDB)⁶⁷. The analysis reveals that these co-targeting candidates exhibit significant enrichment in a variety of GO biological processes, notably phosphorylation, apoptosis, and regulation of transferases. These processes are integral to cellular communication and regulatory networks, highlighting the biological significance of the clustered genes. The functional similarities observed across these GO processes suggest that inhibiting these pathways should be considered in therapeutic intervention, as they play pivotal roles in maintaining cellular homeostasis and signal transduction.

Combining RTK inhibitors with inhibitors targeting downstream oncogenes from these pathways aims to achieve a more robust suppression⁶⁸. One example is co-administering EGFR and PI3K inhibitors, potentially leading to more robust antitumor activity^69,70. Proteins with a dendrogram distance less than 0.1 suggest functional homology, indicating pathway redundancy. Proteins separated by distances of 0.1 or greater represent distinct functional clusters with divergent molecular roles. Our strategic approach leverages molecular distance to design targeted combination drugs exploiting parallel and compensatory signaling pathways. The critical framework systematically maps co-targeting candidates across molecular networks, identifying upstream and downstream proteins, RTKs, oncogenes, and TFs. This context-dependent mapping enables precise inference of signaling mechanisms, facilitating more nuanced therapeutic interventions.

To reinforce the translational value of the predicted protein combinations, we cross-referenced a subset of top-ranked target pairs with ongoing or completed clinical trials listed on ClinicalTrials.gov. The clinical relevance supports our findings. For example, the combination of PIK3CA and estrogen receptor (ER) is currently being evaluated in a phase III trial of alpelisib (BYL719), a selective PI3Kα inhibitor, in combination with endocrine therapy such as fulvestrant in hormone receptor-positive, HER2-negative, PIK3CA-mutant metastatic breast cancer (ID: NCT05646862). Fulvestrant, a selective estrogen receptor degrader, acts by antagonizing and downregulating ER, thereby synergizing with PI3K inhibition. In line with this, ESR1/PIK3CA co-mutations were observed in 90 breast cancer tumors in our dataset, further supporting the biological plausibility of targeting this axis, given that ESR1 encodes the estrogen receptor (ER) and drives ER-dependent transcription. Similarly, capivasertib, a pan-AKT inhibitor targeting AKT1/2/3, is being tested in combination with fulvestrant in the ReDiscover-2 trial (ID: NCT06982521) among PI3Kα-mutated, HR + /HER2- advanced breast cancer patients, aligning with the predicted therapeutic relevance of the PI3K/AKT pathway. This study compares RLY-2608 and fulvestrant combination versus capivasertib and fulvestrant. Additionally, a triple combination of encorafenib (a BRAF^V600E inhibitor), cetuximab (an EGFR inhibitor), and alpelisib is being evaluated in BRAF-mutant metastatic colorectal cancer, supporting the feasibility of co-targeting BRAF, EGFR, and PIK3CA (ID: NCT01719380). Furthermore, alpelisib is being investigated with tucatinib, a HER2-selective kinase inhibitor, in HER2-positive, PIK3CA-mutant metastatic breast cancer, reinforcing the therapeutic rationale of inhibiting the same pathway (ID: NCT05230810). Notably, EGFR and PIK3CA co-mutations were identified in multiple lung and brain tumors in our dataset, where AKT1 emerging as a downstream effector, suggesting that EGFR + PIK3CA ± AKT1 combinations merit further clinical investigation. Finally, KRAS and PIK3CA co-mutations were highly enriched in uterine and colorectal tumors, indicating strong tumor-type specificity and underscoring their potential as context-dependent combination therapy targets. Collectively, these findings substantiate the translational potential of our strategies and highlight specific combinations that align with, or extend beyond, current clinical efforts.

The co-existing mutations network reveals how mutations across proteins contribute to similar phenotypic outcomes through shared biological pathways, highlighting the advantage of considering protein networks over isolated proteins. This method reduces the likelihood of therapeutic resistance, as targeting multiple network nodes diminishes the chance of cancer cell survival. Additionally, addressing interconnected proteins allows combination drugs achieving greater specificity, minimizing off-target effects and preserving cellular functions.

ESR1/PIK3CA subnetwork decodes combination drug targets for breast cancer

Above, we suggested that the gene sets (encoded proteins) in the Fig. 4a dendrogram are important components of cellular communication, and their combinations could be evaluated for combination drugs. To evaluate them as co-targeting candidates, we identified proteins in the ESR1/PIK3CA-specific subnetwork. We reconstructed subnetwork specific to metastatic breast cancer by identifying the shortest path proteins that interconnect ESR1 and PIK3CA, which are pivotal markers of breast cancer metastasis³⁴. Some of the shortest path proteins and drug targets in the reconstructed subnetwork are depicted in Fig. 5 and the full subnetwork is in Supplementary Fig. 5. For the ESR1/PIK3CA co-mutated pair, we identified 38 proteins that were enriched in the PI3K/AKT and chemokine pathways where 24 proteins were implicated in the chemokine pathway, 30 in the PI3K/AKT signaling pathway, with 16 proteins common to both.

Fig. 5. ESR1 | PIK3CA subnetwork reconstructed with PageRank algorithm in breast cancer.

In protein-protein interaction (PPI) networks, a ‘subnetwork’ refers to a smaller set of interconnected proteins that work together to regulate specific cellular processes. ‘Subnetwork reconstruction’ refers to the process of identifying and mapping a smaller, functionally relevant group of protein interactions within a larger PPI network. This allows focusing on specific interactions that may play a critical role in cellular functions. This group contributes to the list of co-targeting candidates depending on their betweenness centrality score. We constructed an ESR1/PIK3CA specific subnetwork using 38 identified shortest path proteins within the chemokine and PI3K/AKT signaling pathways. In protein networks, ‘shortest paths’ refer to the most direct routes (i.e., sequences of interactions) between two proteins within the network. These paths represent the smallest number of interactions (steps) required to connect the two nodes. Key interactions are likely to influence biological processes. Applying the PageRank algorithm, we delineated the ESR1/PIK3CA specific subnetwork. The resulting subnetwork with 65 nodes provides the drug targets listed in CMAP and RTKs. Key drug targets like EGFR, ERBB2/3/4 (encoding ErbB2/3/4), PDGFRB (encoding PDGFRβ), and IGF1R (encoding IGF1) are pivotal nodes in the dendrogram, which proposes them as co-targeting candidates. In the reconstructed subnetwork of ESR1/PIK3CA, seed genes have purple borders, blue nodes are regular genes, and TFs are V-shaped. RTKs are pink. Inhibitors are in rectangles. The genes in chemokine, PI3K/AKT, and in both pathways are shaded yellow, pink and light blue, respectively.

Then, we reconstructed the ESR1/PIK3CA specific subnetwork using these 38 proteins as seeds. These subnetworks were designed to elucidate tissue-specific interactions within the HIPPIE PPI network and identify key connector proteins and co-targeting candidates. The resulting subnetwork contains 65 proteins with the drug target genes listed in the CMAP. RTKs and shortest path edges are highlighted. The drug target RTKs are EGFR, ERBB2/3/4 (encoding ErbB2/3/4), PDGFRB (PDGFRβ), IGF1R (IGF1) and the seed gene CSF1R (CSF-1) belong to the PI3K/AKT pathway. PIK3CA/B/D/G (PI3Kα/β/δ/γ) and MAP2K1 (MEK1) belong to PIK3K/AKT and chemokine pathways, and SRC (Src) and PRKDC (DNA-PK) belong to the chemokine signaling pathway. The transcription factor ESR1 is not included in these enriched pathways.

We see that the highlighted drug targets in the ESR1/PIK3CA subnetwork have nodes in the Fig. 4a dendrogram, implicating their importance in cellular communication. We also see that ErbB2 can be targeted with the drugs afatinib, lapatinib, neratinib, vandetanib and trastuzumab. ErbB3 can be targeted with vandetanib and the monoclonal antibody LJM716, and EGFR by several drugs, including gefitinib, erlotinib, osimertinib, etc. PI3Kα can be targeted with alpelisib and RLY-2608, an allosteric mutant-selective inhibitor of PI3Kα⁷¹. The clustering in Fig. 4a along with the breast cancer-specific ESR1/PIK3CA subnetwork in Fig. 5 could be a good strategy for planning combinations targeting one or more RTKs from the oncogenes’ subnetwork (including TFs). In general, if a tumor has activating mutations in at least one of the protein-pair constituents, and seed proteins accompanied by sufficient transcriptional activity in the pathways of the seed proteins, we can assume an increased oncogenic activity. Either direct inhibition of the connector proteins or indirect inhibition by targeting their partners in protein complexes could provide a venue for combination drugs.

Co-targeting PI3Kα and ErbB3 in HER2+ breast cancer

In clinical practice, breast cancer is categorized into three therapeutic groups: ER-positive, HER2 (or ErbB2)-amplified, and triple-negative breast cancers. HER2-positive (HER2+) breast cancer, marked by the overexpression of HER2, represents a biologically distinct subset characterized by aggressive tumor behavior and poor prognosis^72,73. To delineate the signaling mechanism in HER2+ breast cancer, we first map some of the key components in the subnetwork in Fig. 5 to the PI3K/AKT pathway (Fig. 3). The subnetwork contains the four members of the ErbB family of transmembrane RTKs: EGFR (ErbB1), ErbB2, ErbB3, and ErbB4. ErbB2 and ErbB3 are closely related to EGFR/ErbB1. ErbB2 and ErbB3 frequently form heterodimers, which are highly active in signaling because ErbB2 has potent kinase activity while ErbB3 has an impaired kinase domain but provides strong docking sites for downstream signaling molecules like PI3K^74,75. We provide further details of signal transduction in HER2+ breast cancer in the Supplementary Information.

To identify the nodes pivotal for communication within the subnetwork, we leveraged the networks topological characteristics (Methods). We propose proteins exhibiting betweenness centrality values exceeding Q3 as key connector proteins (Fig. 6a). This analysis revealed several key proteins: ESR1 (encoding ERα), PIK3CA (encoding the protein PI3Kα), ERBB2 (encoding the protein ErbB2), IGF1R (encoding IGF-1), PIK3R2, PIK3R3, APP (amyloid precursor protein), HCK (HCK Proto-Oncogene, Src Family Tyrosine Kinase), FYN (a membrane-associated tyrosine kinase involved in cell growth regulation and interacting with the p85 subunit of phosphatidylinositol 3-kinase and fyn-binding protein), EGFR, and GRB2 (which binds to the epidermal growth factor receptor and contains one SH2 domain and two SH3 domains).

Fig. 6. Betweenness centrality and PDX models.

a Scatter plot showing the betweenness centrality (x-axis) and eigenvector centrality (y-axis) of the ESR1/PIK3CA subnetwork in breast cancer. Betweenness centrality quantifies the extent to which a node serves as a bridge along the shortest paths between pairs of nodes in a network. Nodes with higher betweenness centrality play crucial roles in efficient communication within the network. Eigenvector centrality measures the influence of a node in a network based on its connections to other central nodes. Nodes with higher eigenvector centrality are not only well-connected but are also connected to other nodes that themselves are highly connected, indicating their influence on the network structure. In the ESR1/PIK3CA subnetwork, nodes positioned towards the upper right quadrant of the scatter plot exhibit both high betweenness and eigenvector centrality, and are co-targeting candidates (red). Nodes that are not potential targets are shown with a blue dot. Node sizes are proportional to the degree centrality; nodes with higher degree centrality have more connections, making them potentially more influential in the network. b Expression scores of upregulated or downregulated pathways in the PDX models with at least one mutation either in ESR1 or PIK3CA and one of the subnetwork genes. There are six such PDX models. All PDXs have high activity in Estrogen, HIF-1, and Oxytocin pathways.

These proteins effectively regulate the flow of communication, as messages typically pass through them. Eigenvector centrality, another metric in network analysis, assesses a node’s importance based not only on its direct connections but also on the centrality of its connected nodes. Nodes with high eigenvector centrality are well-connected and linked to other highly central nodes within the network. This metric posits that connections to more central nodes enhance a node’s own centrality score. In terms of communication dynamics, nodes with high eigenvector centrality are influential hubs, extending their impact beyond their immediate vicinity to the entire network.

HER2+ breast cancer can be targeted by combinations of RTKs, EGFR (ERBB1), ERBB2, ERBB3 or ERBB4 along with a PI3K isoform, PIK3CA/B/D/G. In case of ESR1 mutations or overexpression, ESR1 can be targeted instead of the RTKs.

For validation, we employed patient-derived xenograft (PDX) models from a publication providing daily drug response rates of PDXs, with volumetric measurements recorded over diverse temporal intervals⁵². We selected breast cancer (BRCA) models featuring at least two mutations: one in the protein pair components and the other(s) in the seed genes (shortest path proteins). Additionally, to facilitate comparative analysis, we chose models that had temporal volume data for untreated tumors and those treated with combination drugs as well as the individual components of these combinations. This selection resulted in six BRCA models with volume changes recorded over various time intervals for the untreated cases, those treated with BYL719 (alpelisib, targeting PIK3CA), LJM716 (an anti-HER3 antibody), and their combination BYL719 + LJM716. We provide the expression values of BRCA PDX seed proteins and drug targets expression values to determine the breast cancer subtype of the models. As both ErbB2 (HER2) and ErbB3 (HER3) levels are relatively higher, we can evaluate these models of HER2+ type (Supplementary Fig. 6). In Fig. 6b, we present pathway expression scores for pathways with relatively high expression (Methods). This reveals which pathways exhibit higher expression across all models; notably, the expression scores for the Oxytocin, HIF-1, and Estrogen signaling pathways are elevated. Furthermore, for models X-1828 and X-2524, the scores for Hippo, Thyroid, and Rap1 pathways are also elevated. Regarding the mutation profiles of these models, X-3077 and X-3078 harbor BRCA, PIK3CA (harbors H1047R driver mutation), and TP53 mutations among the seed genes, while X-1828 contains TP53 and PIK3CA (harbors G914R mutation) mutations.

The growth patterns of models X-3077 and X-1828 are illustrated in Fig. 7a, b and X-3078 in Supplementary Fig. 7. All three graphs indicate that untreated tumor growth is highly aggressive. In model X-3077, treatment with LJM716 decelerates tumor growth; however, resistance emerges after day 0. Similarly, in model X-3078, resistance to LJM716 becomes evident after day 10. For both X-3077 and X-3078, BYL719 treatment nearly halts tumor growth. Conversely, in model X-1828, the tumor continues to grow despite an initial shrinkage with BYL719 treatment. Under the combination BYL719 + LJM716, X-3077 experiences cessation and subsequent shrinkage of tumor growth. Model X-3078 initially responds well to the combination drugs for the first 40 days but begins to grow afterward, though the growth rate remains lower compared to the untreated case. Model X-1828 shows responsiveness to the combination drugs during the first 30 days and after 70 days, with a minor tumor growth observed between days 30 and 70. Similarly, PDX models with the metastatic marker mutations of CDH1 and PIK3CA respond better to BYL719 + LJM716 and tumor growth slows down dramatically (Supplementary Figs. 8–10).

Fig. 7. Volumes of BRCA PDXs with and without treatment.

a Tumor growth rates of PDX model X-3077 when untreated and when treated with BYL719 (alpelisib), LJM716 and BYL719 + LJM716. BYL719 is a pan-PI3K inhibitor, and LJM716 is an anti-HER3 monoclonal antibody. The volume of the PDX either shrinks or grows dramatically slower compared to untreated and monotherapy cases. X-3077 has mutations in the seed proteins BRCA, PIK3CA, and TP53. b Tumor volumes of PDX model X-1828 under untreated conditions and following treatment with BYL719, LJM716, and their combination. Similar to X-3077, the combination treatment markedly suppressed tumor growth compared to controls. X-1828 carries mutations in PIK3CA and TP53. Each curve represents an individual xenograft measurement per condition.

Combination drugs work better: Our pursuit of combination drugs is bolstered by the growth rates of patient-derived xenografts (PDXs) under both untreated conditions and various monotherapy and combination drug regimens (Supplementary Fig. 11). We selectively included 223 models with documented volume data from day 0 and observable growth trajectories in untreated scenarios. We conducted a rigorous analysis of the daily growth rates for these models when administered 24 combination drugs and 17 monotherapies, which are constituent elements of the combination regimens.

We evaluated the efficacy of LJM716, a HER3-neutralizing antibody, as monotherapy and in conjunction with BYL719, a p110α-specific inhibitor, in HER2-overexpressing breast and gastric carcinomas. LJM716 effectively diminished HER2-HER3 and HER3-p85 dimerization and phosphorylation of HER3 and AKT, inhibiting tumor progression in xenograft models. The combination of LJM716 with lapatinib and trastuzumab significantly enhanced survival outcomes. Synergistic interaction between LJM716 and BYL719 resulted in greater inhibition of cell proliferation and AKT phosphorylation in HER2+ and PIK3CA mutant cell lines, surpassing monotherapies. In trastuzumab-resistant HER2+/PIK3CA mutant MDA453 xenografts, combination drugs achieved complete tumor regression, in contrast to partial inhibition observed with individual agents. Long-term treatment showed that only 14% of mice receiving the LJM716, BYL719, and trastuzumab combination experienced tumor recurrence, a significantly lower rate compared to other treatment groups. These findings suggest that dual inhibition of HER2 signaling with LJM716 and BYL719 represents an effective approach for HER2-overexpressing tumors⁷⁶.

Combination drugs for colorectal cancer through BRAF | PIK3CA specific subnetwork

Colorectal cancer (CRC) represents a heterogeneous malignancy characterized by diverse genetic alterations that drive tumorigenesis and progression. Among these aberrations, co-existing mutations in BRAF (encoding B-Raf) and PIK3CA have garnered significant attention due to their implications for prognosis and therapeutic response. B-Raf, a component of the MAPK signaling pathway, and PI3Kα, a pivotal regulator of the PI3K/AKT pathway, play critical roles in cell proliferation, survival, and differentiation (Fig. 3). We aim to identify key nodes in the communication network by setting our starting point at B-Raf and PI3Kα, two proteins harboring co-existing mutations specific to CRC³⁴.

To this end, we identified genes in the BRAF/PIK3CA specific subnetwork similar to the one for breast cancer, where the resulting subnetwork contains 65 proteins (Fig. 8 and Supplementary Fig. 12) with the drug target genes listed in the CMAP, RTKs, and shortest path edges are highlighted. In the subnetwork, the drug target RTKs are CSF1R (encoding CSF-1), EGFR (EGFR), ERBB2 (ErbB2), FLT4 (FLT4), IGF1R (IGF1), PDGFRA/B (PDGFRα/β). Downstream in these pathways, we see MAPK pathway members BRAF (B-Raf), RAF1 (Raf1), GNAS (GSα) and MAP2K1/2 (MEK1and MEK2). From the PI3K/AKT pathway, we see PI3K isoforms PIK3CA/B/D/G (PI3Kα/β/δ/γ) and two other genes, PRKDC (encoding PKCδ) and KDR (Kdr).

Fig. 8. BRAF/PIK3CA subnetwork in colorectal cancer reconstructed with PageRank algorithm.

BRAF/PIK3CA subnetwork nodes in the MAPK (shaded in yellow) and PI3K/AKT (shaded in blue) pathways delineate a functionally relevant cluster of protein interactions derived from the shortest path proteins linking BRAF and PIK3CA. Shortest paths refer to the most efficient routes (sequences) of interactions between two nodes within the network. These paths represent the minimal number of interaction steps required to connect the two nodes. The subnetwork of proteins on the shortest paths encompasses RTKs such as EGFR, PDGFRα, IGF1, and ErbB2, alongside downstream B-Raf and PI3Kα, within the compensatory MAPK and PI3K/AKT cascades. The network suggests that co-targeting these RTKs, in conjunction with at least one of the identified oncogenes, offers robust therapeutic strategies as they are the key bridging nodes. These key nodes with high betweenness centrality listed among the co-targeting candidates. In this subnetwork, seed genes are marked with purple borders, regular genes as blue nodes, and TFs are V-shaped. RTKs are in pink. Inhibitors targeting specific nodes within this subnetwork are annotated within rectangular labels.

We see that the highlighted drug targets in the BRAF | PIK3CA subnetwork have nodes in the dendrogram in Fig. 4a implicating their importance in cellular communication. We elaborate the details of the subnetwork reconstruction and the selection of the co-targeting candidates in the Supplementary Information.

For CRC, the RTKs CSF1R (encoding CSF-1), EGFR, ERBB2, FLT4, IGF1R, PDGFRA/B can be targeted in combination with MAPK pathway members SRC (encoding Src), RAF1, BRAF, MAP2K1 (MEK1), MAP2K2 (encoding MEK2) and PI3K/AKT members KDR, PIK3CA/B/D/G or PRKDC (PKCδ). These genes have higher betweenness and edge centrality values (Fig. 9a) and can thwart the oncogenic signal flow in CRC tumors that have mutations in BRAF or PIK3CA and at least one of the seed genes.

Fig. 9. Co-targeting candidates in BRAF/PIK3CA subnetwork and PDXs.

a Scatter plot showing the betweenness centrality (x-axis) and eigenvector centrality (y-axis) of the BRAF | PIK3CA subnetwork in colorectal cancer. Co-targeting nodes (red) on the upper right quadrant of the scatter plot have both high betweenness and eigenvector centrality. These nodes govern important pathways and serve as vital connectors linking sections of the network. The remaining network nodes (blue dots) are not potential targets. Node sizes correspond to their degree centrality, indicating their level of connectivity within the network. b Pathway expression scores of PDXs with mutations in BRAF/PIK3CA subnetwork components show prominent expression in Estrogen, HIF-1, HIPPO, and Oxytocin signaling pathways.

Supplementary Fig. 13 and Fig. 9b show the expression levels of some proteins and pathway scores in these PDXs, respectively. AKT1, ERBB2, ERBB3, and MAP2K2 expression levels are higher and above 0.5. As to the expression scores, Thyroid, Rap1, Oxytocin, Hippo, HIF-1, and Estrogen pathways display prominent activity. In Fig. 10a we see that the CRC PDX model X-2145 diminishes after the combination of BYL719 + cetuximab + encorafenib targeting PIK3CA (or PIK3C/B/D/G), EGF,R and BRAF, respectively. The model has mutations on seed genes BRAF, ERBB3, and GNAI2. CRC models X-3792 and X-3205 also respond better to the BYL719 + cetuximab + encorafenib treatment compared to monotherapies of encorafenib and cetuximab (Fig. 10b and Supplementary Fig. 14). X-3792 has mutations on seed genes CRK, ERBB3, GNAI2, IGF1R, PIK3CA, and PAK2. X-3205 has mutations on BRAF, MAP2K3, and PIK3CA. For this model, signaling could be stronger as both co-mutated genes have mutations on X-3205; therefore, after a small increase followed by a steady state for 20 days, the tumor exhibits a higher growth state. Among the patient-derived xenografts (PDXs) with PIK3CA mutations, those harboring specific driver mutations demonstrated a preferential response to alpelisib monotherapy compared to combination drug treatments. This differential response may be attributed to oncogene addiction, particularly evident in xenografts X-3792 (E545D mutation) and X-3205 (R88Q mutation), which exhibit characteristic driver mutation profiles³⁹.

Fig. 10. Tumor growth rates of CRC PDXs.

a, b Tumor growth rates of the colorectal cancer PDXs with mutations in the BRAF/PIK3CA subnetwork. The growth curves are without treatment, treated with single drugs BYL719 (PI3K inhibitor), cetuximab (EGFR inhibitor), encorafenib (BRAF inhibitor), and the combination treatment BYL719+cetuximab+encorafenib. In the models, combination therapy significantly reduces tumor growth rates. The x-axis is the treatment days. The PDX X-2145 has mutations in the seed proteins BRAF, ERBB3, and GNAI2 where X-3792 has mutations in CRK, ERBB3, GNAI2, IGF1R, PIK3CA, and PAK2. Each curve represents measurements from a single xenograft per condition.

This type of combination drugs falls into targeting parallel pathways³. Both MAPK and PI3K/AKT pathways receive signals through an RTK and transduce it downstream (Fig. 3). Single targeting of these molecules often leads to adaptive resistance mechanisms. BRAF^V600E mutations, present in a subset of colorectal cancers, activate the MAPK, driving proliferation and survival. However, inhibition of BRAF alone can result in compensatory activation of EGFR and PI3K/AKT pathways, diminishing treatment response. Similarly, single drug targeting EGFR or PIK3CA can lead to feedback activation of the MAPK pathway or other compensatory survival signals.

Discussion

Here, we develop a signaling-based method to discover optimal proteins for the oncologist to co-target with drug combinations. Our innovative signaling-based approach harnesses nature, mimicking the cancer’s prolife tactics. A fundamental cancer tenet for evading anti-drugs is circumventing them. Since the drugs commonly block pathways, cancer often maneuvers around the blockade through emerging mutations, and (or) altering expression levels, that allow bypassing it. Our signaling- and mutations-informed method aims to capture cancer’s ploy.

To accomplish this aim, we identified potential targets for combination drugs using a network-based approach. Candidates were selected from pivotal subnetworks nodes, characterized by high betweenness centrality, serving as essential communication hubs. As such, they often serve as homeostatic guardians, making them ideal targets for combination drugs aimed at disrupting oncogenic signaling. By analyzing our compiled coexisting mutations³⁴ and their associated subnetworks, we elucidated the principal communication pathways, employing genes on the shortest paths as seeds. This methodology enabled us to determine key oncogenic signaling nodes and essential connector proteins that reveal potential rewiring routes.

We cluster proteins with high betweenness centrality into three categories with respect to the signaling pathways they belong to: same or redundant, parallel and compensatory pathways^3,59. The oncogenes include RTKs—EGFR, ERBB2, ERBB3, FGFR2, FGFR3, IGF1R, and NTRK1—and transcription factors—CTNNB1, ESR1, MDM2, MYC, and STAT3. We applied our results to breast cancer and colorectal cancer subnetworks, derived from ESR1/PIK3CA and BRAF/PIK3CA protein pairs, which harbor co-existing mutations specific to these tissues, respectively³⁴.

The ESR1/PIK3CA subnetwork underscores the potential efficacy of inhibiting the PI3K/AKT pathway by targeting an RTK and a downstream kinase such as PIK3CA/B/D/G. Concurrently targeting the compensatory pathways MAPK and PI3K/AKT by inhibiting EGFR and the downstream oncogenes BRAF and PIK3CA is also effective. These co-targeting proteins from different clusters in the dendrogram shown in Fig. 4a, which have a Hamming distance greater than 0.1, can inform combination drugs. This approach involves configuring an upstream RTK with downstream oncogene(s) and/or transcription factors while considering pathway organization and crosstalk. RTKs complexes, as in the case of inhibiting HER3 in HER2+ breast cancer, can guide alternative combinations, as targeting one heterodimer component can inhibit ligand-induced signaling.

Biological networks, particularly PPI networks, can help in blocking oncogenic signaling in cancer cells^77–82. By integrating large-scale omics datasets and computational models, context-specific network alterations that drive cancer can be discovered, offering potential therapeutic targets. One prominent algorithm is PageRank, which can identify subnetworks by scoring nodes (genes, proteins, metabolites) based on their network connectivity and importance. Higher PageRank scores establish central elements, revealing critical subnetworks. This method helps pinpoint key components and interactions within complex biological systems^47,80,83,84. Subnetwork reconstruction and topological analysis reveal critical regulators and functional units⁸⁰. Integration of omics data enables identification of key network features such as hubs and modules^80,82. Centrality measures and community detection algorithms elucidate essential proteins and hierarchical organizations in cancer signaling and facilitate the discovery of driver mutations and drug targets by mapping context-dependent relationships between oncogenes and their effectors^81,82.

The topological characteristics of PPI networks provide vital insights into the key components of these systems^85–87. Hubs (proteins with a high degree) and bottlenecks (proteins with a high betweenness) are promising therapeutic targets due to their roles in maintaining network stability and functions⁸⁸. Beyond centrality metrics, functional enrichment analysis and experimental validation are crucial for verifying the potential of identified proteins as drug targets. Node annotations, biological pathways, and gene expression data are crucial for target selection. REFLECT, an advanced machine learning tool was developed to investigate the concept that recurrent co-alteration signatures may be targeted with tailored combination drugs to enhance preclinical and clinical outcomes¹. The significance of network-based methods in identifying combination drugs is underscored by tools like REFLECT to integrate biological data and generate optimal therapeutic strategies. Leveraging the complexity and connectivity within PPI networks, these approaches can uncover critical nodes and pathways, toward highly targeted combination treatments⁸⁹. Through data mining and network analysis, synergistic multidrug combinations were identified, including optimized 3- and 4-drug regimens tailored for ERα+/HER2-/PI3Kα-mutant breast cancer subtypes⁹⁰. These combinations target ERα along with PI3Kα, p21, and PARP1, demonstrating efficacy in tamoxifen-resistant models and patient-derived organoids, suggesting potential for overcoming current treatment limitations. The Pathway Ensemble Tool (PET) identifies pathways and gene combinations associated with poor prognosis, offering potential targets for combination drugs⁹¹. A useful tool that enables sophisticated tissue-specific pathway analysis is SignaLink3, which represents a comprehensive, manually curated signaling network database that integrates protein-protein interactions, regulatory information, and subcellular localization data across humans and three model organisms, providing an unprecedented 700,000 interactions⁹².

The mounting number of new targeted therapies has exponentially expanded the therapeutic combinations search space, calling for an effective solution to this challenge^2,3,93. Selecting optimal small molecule combinations among this vast array renders comprehensive clinical testing impractical. Combinatorial kinase inhibitors aim to disrupt specific resistance signaling, inhibiting multiple targets simultaneously^69,94. Combination strategies entail blocking pathways upstream and downstream, exploiting synthetic lethality, and concerted targeting of multiple pathways^2,3,95.

Network-based, signaling-primed combination drugs represent a transformative paradigm in precision oncology and personalized medicine. It addresses the commonly recognized limitations of monotherapies. It offsets compensatory mechanisms, and by temporally rotating the combinations, it lessens drug resistance. High-throughput omics data harnessed by sophisticated computational algorithms is a recipe for therapeutic regimens to the heterogeneous nature of cancers. Within this framework, signaling-learned network-informed combination drugs establish a first step in designing a combinatorial strategy that precisely targets tumor-specific molecular alterations. This approach also expedites drug discovery through the repurposing of existing compounds and the identification of novel synergies.

We acknowledge that incorporating extensively studied cancer subtypes, such as HER2-positive breast cancer or BRAF/PIK3CA-mutant colorectal cancer, may introduce bias toward these well-characterized contexts. This could potentially lead to overfitting, where predicted therapeutic combinations are disproportionately influenced by cancers with abundant molecular and clinical data. Such bias might limit the generalizability of our findings to less-studied cancer types or molecular subtypes. Future work should focus on systematically validating the identified combinations across a broader spectrum of tumor types with varying levels of molecular annotation to ensure the robustness and clinical applicability of our approach across diverse oncological contexts.

While our study prioritizes therapeutic combinations based on molecular and cellular signaling interactions, effective cancer therapy must also account for the tissue context and the influence of the tumor microenvironment, including interactions with immune cells and stromal components. These factors can impact drug efficacy and resistance, and incorporating such spatial and immunological dimensions remains an important avenue for future research. Additionally, although our computational approach identifies biologically plausible combinations, the safety profiles of these combinations—including potential on- and off-target toxicities—must be rigorously evaluated in preclinical models. The lack of absolute specificity and selectivity of many targeted agents poses challenges in clinical implementation, underlining the need for careful toxicological screening alongside efficacy studies.

Here we launch the first concept-based signaling mechanism-informed strategy, which takes up the question of “which signaling pathway and protein to select to mitigate the patient’s expected drug resistance”³ deescalating the massive number of possibilities facing the physician and fitting the solution to the patient status. Applications of our strategy are validated by existing patient-based xenograft models. However, translating these into clinical practice necessitates further rigorous validation in preclinical models and meticulously conducted clinical trials.

Author contributions

B.R.Y., H.J., and R.N. conceived and designed the study. B.R.Y. did the data curation and visualization, analyzed the results, and drafted the manuscript. B.R.Y., H.J., and R.N. validated, reviewed, and edited the manuscript. R.N. supervised the project.

Peer review

Peer review information

Communications Medicine thanks Dominique Levêque and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Data availability

This paper analyzes existing, publicly available data. The accession numbers for the datasets are listed in Supplementary Table 1. The following publicly available datasets and resources were used in this study: The Cancer Genome Atlas (TCGA) data (https://portal.gdc.cancer.gov). Data from AACR Project GENIE were accessed through Synapse (https://www.synapse.org/genie). Pathway annotations were retrieved from KEGG and the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb). Protein-protein interaction data were sourced from the HIPPIE PPI network (https://cbdm-01.zdv.uni-mainz.de/~mschaefer/hippie/). Transcriptional regulatory interactions were obtained from TRRUST.

Code availability

The code for the analyses can be accessed from the following repository: https://github.com/bengiruken/CombinationTherapyTargets, https://zenodo.org/records/15801594⁹⁶. The software and the algorithms used are listed in Supplementary Table 2.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01150-9.