1 Motivation

Combination therapy has long been a cornerstone of clinical practice, with demonstrated benefits across infectious diseases and complex disorders including cancer. Compared with monotherapy, rational drug combinations can enhance therapeutic efficacy, reduce toxicity through dose reduction, and delay or prevent the emergence of drug resistance. However, the efficacy of single-agent therapies is frequently compromised by intrinsic and acquired resistance driven by compensatory signaling pathways, especially for oncology, underscoring the need for systematic approaches to identify synergistic therapeutic strategies beyond empirical screening. Large-scale perturbational transcriptomics initiatives, such as the LINCS L1000 project, have generated extensive public resources comprising hundreds of thousands of chemical and genetic (knockout and knockdown) gene expression signatures. These datasets provide unprecedented opportunities for in silico identification of synergistic interactions, including compounds that can be used in synergistic combinations with a reference compound. Notably, the SynergySeq(http://synergyseq.com/) framework demonstrated the feasibility of leveraging perturbational signatures to identify combinations capable of overcoming resistance to targeted therapies, such as BET inhibitors in glioblastoma¹. Despite these advances, translating heterogeneous perturbational datasets into actionable therapeutic hypotheses remains challenging, and existing tools often lack integrated, user-friendly frameworks accessible to users without specialized bioinformatics expertise. To address these limitations, we developed PerturbSynergy(Figure 1), an open-access predictive web server for the systematic discovery of synergistic drug and molecular target combinations. Building upon the SynergySeq concept, PerturbSynergy integrates large-scale chemical and genetic perturbational transcriptomic data from the LINCS L1000 and LINCS2020 projects to enable dual-mode inference of both drug–drug and drug–target synergy within a unified analytical framework.The algorithm of PerturbSynergy is shown in the figure below (Figure 2). Synergy inference is achieved through computing disease discordance and drug concordance to identify drug combinations that induce a synergistic response in cancer. By allowing users to directly query disease-specific and reference drug signatures against comprehensive perturbational libraries and to interpret results through interactive visualizations, PerturbSynergy provides a practical and scalable platform for discovering synergistic therapeutic strategies tailored to specific biological and clinical contexts.

Note:

Discordance Ratio(DR):The Discordance Ratio assesses how a drug's transcriptional response differs from a disease-specific gene expression signature. It calculates the ratio of drug-induced genes that are in the opposite direction to the disease signature, compared to those that align with it. A high DR indicates that the drug may target pathways opposite to the disease’s transcriptional profile, making it a potentially effective treatment.

Concordance Ratio(CR):The Concordance Ratio measures the similarity in the direction of gene expression changes between the transcriptional responses induced by a drug and the reference gene signature. It is calculated as the ratio of genes that change in the same direction as the reference signature to those that change in the opposite direction. A high CR indicates that the drug's transcriptional response is similar to that of the reference compound, while a low CR suggests divergence in gene expression direction.

Orthogonality Score(OS):The Orthogonality Score (OS) combines CR and DR to quantify the orthogonality (or distinctiveness) of a drug's effect relative to the reference compound and disease signature. It provides a single score that represents the drug's orthogonality, calculated as the distance between its concordance and discordance values. A higher OS indicates that the drug targets pathways not covered by the reference or disease signature, increasing the likelihood of synergy in combination therapy. Therefore, we define OS as the synergy score, which can assess the signature's potential for synergy.

The OS formula:

2 Data Overview

PerturbSynergy incorporates a total of 1080000 gene sets across 272 cell lines(Figure 3) in 21 tissue types and includes two main modules: Synergistic Drugs,Synergistic Targets.

PerturbSynergy consolidates large-scale chemical and genetic (KO/KD) perturbational transcriptomic data from the LINCS L1000 and LINCS2020 projects. In this study, KO and KD perturbations were treated as complementary loss-of-function signatures and processed using an identical analytical workflow. To ensure transparency and reproducibility, we provide a detailed description of each data source and its corresponding input options:

2.1 Chemical perturbation

LINCS L1000:

Source: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE92742;

https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE70138;

Dataset: GSE92742 and GSE70138;

Description: Phase I and Phase II LINCS L1000 perturbational transcriptomic profiles generated from small-molecule compound treatments across diverse human cell lines and Phase I and Phase II data were merged. After removing the completely repeated perturbations of each condition, the matrix comprises 978 (landmark) genes (rows) and 297334 compound-induced transcriptional signatures (columns), including 122933 compounds, 88 cell lines.

LINCS 2020:

Source: https://clue.io/releases/data-dashboard;

Dataset: level5_beta_trt_cp_n720216x12328.gctx;

Description: Newly released CMap LINCS 2020 Level 5 (MODZ) replicate-collapsed z-score gene expression signatures generated from small-molecule (chemical) perturbations (trt_cp). After removing the completely repeated perturbations of each condition, the matrix comprises 978 (landmark) genes (rows) and 646625 compound-induced transcriptional signatures (columns), including 196283 compounds, 263 cell lines.

2.2 Genetic perturbation:

LINCS L1000 KO:

Source: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE70138;

Dataset: GSE70138;

Description: LINCS L1000 Phase II CRISPR/sgRNA-mediated gene knockout (loss-of-function) perturbational profiles, measuring transcriptional responses of human cultured cells to targeted gene knockouts using the L1000 assay. After removing the completely repeated perturbations of each condition, the matrix contains 978 (landmark) genes (rows) and 583 genetic perturbation signatures (columns), including 11 cell lines.

LINCS L1000 KD:

Source: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE92742;

Dataset: GSE92742;

Description: LINCS L1000 Phase I shRNA-mediated gene knockdown (KD) perturbational profiles, capturing transcriptional responses of human cultured cells to targeted gene knockdown using the L1000 assay. After removing the completely repeated perturbations of each condition, the matrix contains 978 (landmark) genes (rows) and 33817 genetic perturbation signatures (columns), including 17 cell lines.

LINCS 2020 KO:

Source: https://clue.io/data/CMap2020#LINCS2020 (CMap/LINCS 2020 resources);

Dataset: level5_beta_trt_xpr_n142901x12328.gctx;

Description: Level 5 (MODZ) replicate-collapsed z-score gene expression signatures generated from CRISPR knockout (KO) perturbations. After removing the completely repeated perturbations of each condition, the matrix contains 978 (landmark) genes (rows) and 55260 genetic perturbation signatures (columns), including 37 cell lines.

LINCS 2020 KD:

Source: https://clue.io/data/CMap2020#LINCS2020 (CMap/LINCS 2020 resources);

Dataset: level5_beta_trt_sh_n238351x12328.gctx;

Description: Level 5 (MODZ) replicate-collapsed z-score gene expression signatures derived from shRNA-mediated gene knockdown (KD) perturbations. After removing the completely repeated perturbations of each condition, the matrix contains 978 (landmark) genes (rows) and 33817 genetic perturbation signatures (columns), including 17 cell lines.

2.3 Method

Chemical perturbation：

Tissue:We used the Phase I and Phase II LINCS L1000 perturbational profiles (GSE92742 and GSE70138, respectively, downloaded from the Gene Expression Omnibus) in this study. Replicate-collapsed differential expression signatures (Level 5 dataset) of the measured (landmark) genes were used throughout the analysis pipeline. LINCS L1000 signatures were accessed using the cmapPy Python library, and Phase I and Phase II datasets were merged. Each perturbational signature is annotated by treatment, cell line, time point, and concentration in the case of chemical compounds, with cell lines further mapped to corresponding tissue types. For the LINCS2020 dataset, the data processing procedure followed the same workflow as described above.

Consensus:Consensus signatures corresponding to identical experimental conditions (treatment, cell line, time point, and concentration in the case of chemical compounds) were aggregated using the MODZ method. For the LINCS2020 dataset, the data processing followed the same workflow as described above.

Genetic perturbation：

Tissue：

LINCS L1000 KO：We used the Phase II LINCS L1000 genetic perturbational profiles (GSE70138) corresponding to gene knockout–related perturbations in this study. Replicate-collapsed differential expression signatures (Level 5 dataset) of the measured (landmark) genes were used in the analysis pipeline. LINCS L1000 signatures were accessed using the cmapPy Python library.

LINCS L1000 KD：Phase I LINCS L1000 perturbational profiles (GSE92742) corresponding to gene knockdown–related perturbations were used. Replicate-collapsed Level 5 differential expression signatures of landmark genes were extracted and processed using the cmapPy library.

LINCS 2020 KO：For genetic knockout perturbations in the LINCS2020 dataset, we used the newly released Connectivity Map LINCS 2020 perturbational profiles obtained from the CLUE platform (https://clue.io/releases/data-dashboard). Gene knockout–associated Level 5 signatures were extracted and processed in a manner analogous to that used for LINCS L1000 data.

LINCS 2020 KD：LINCS 2020 genetic knockdown perturbational signatures were retrieved from the CLUE data release portal. Gene knockdown–associated Level 5 signatures were extracted and processed in a manner analogous to that used for LINCS L1000 data.

Consensus：

Signatures (LINCS L1000 KO/LINCS L1000 KD/LINCS2020 KO/ LINCS 2020 KD) corresponding to identical experimental conditions (gene perturbation, cell line, and time point) were aggregated using the MODZ method to derive consensus knockout/knoc signatures.

Drug annotation:

Drug Repurposing Hub:A curated and annotated collection of FDA-approved drugs, clinical trial drugs, and selected pre-clinical tool compounds.

FDA:FDA-approved drugs.

3 Analysis Workflows

PerturbSynergy features two analytical modules that utilize chemical and genetic perturbation data from the LINCS L1000 and LINCS 2020 projects to analyze synergistic drug interactions, respectively. An executable input & output file can be downloaded here.

×

Synergy Scatter Plot Interpretation Guide

1. Axis Interpretation

• X-axis (Horizontal): Concordance Ratio (CR) - Measures transcriptional similarity to the reference drug.
  Higher values = more similar mechanism of action
• Y-axis (Vertical): Discordance Ratio (DR) - Measures ability to reverse disease gene expression signature.
  Higher values = stronger disease reversal effect

Note: CR ranges from 0 to 1 (0 = completely different, 1 = identical)

---

2. Orthogonality Score (OS)

The Orthogonality Score quantifies synergistic potential:

Formula: OS = √[(1 - CR)² + (DR)²]

• High OS values indicate greater synergistic potential
• Low OS values indicate limited synergy
• Compounds are automatically ranked by OS in the results table

The platform calculates OS to identify drugs that are both:
• Mechanistically different from reference drug (low CR)
• Effective at reversing disease signature (high DR)

---

Synergy Scatter Plot

---

3. How to Use the SynergySeq Platform

1. View the scatter plot: Each point represents a drug from the LINCS L1000 database
2. Identify candidate drugs:
   • Focus on the upper-left quadrant (low CR, high DR)
   • Check the OS ranking in the results table
   • Click on individual points for detailed information

---

Synergy Scatter Plot

---

✅ Ideal Synergistic Combination

• Low CR (mechanistically different)
• High DR (effective disease reversal)
• High OS score

❌ Combinations to Avoid

• High CR, Low DR - Similar mechanism with poor disease effect
• Low CR, Low DR - Different but ineffective

Close

3.1 Prepare input file

Disease signature:Please upload the differentially expressed genes along with their expression directions (upregulated or downregulated), reflecting the differences between the disease state and the control group in your study. You may also include any genes of interest.The column names must be 'Gene' and 'direction'. For example, in PerturbSynergy's demo, the disease signature we used was based on |log₂FC| > 1 and FDR < 0.05. Differential expression analysis, conducted using the limma package on integrated TCGA and GTEx data, identified genes differentially expressed in association with BRCA (breast cancer).

Reference signature:Please upload the gene expression profile following drug perturbation, which represents the differentially expressed genes and their expression directions (upregulated or downregulated) compared to the control group. You may also include any genes of interest.The column name for genes must be 'Gene', and the column name for gene expression direction is generally the name of the drug.For example, in PerturbSynergy's demo, the reference signature we used were perturbation signatures for lapatinib&olaparib derived from the LINCS L1000&2020 database.

3.2 Synergistic Drugs

The Synergistic Drugs module in PerturbSynergy is designed to enable the systematic prediction and analysis of synergistic drug combinations by utilizing large-scale perturbation datasets, such as the LINCS L1000 and LINCS2020 projects. This module allows users to assess the combined effects of drug pairs, identify potential synergistic interactions, and prioritize drug combinations for further investigation.

Follow these steps:

① Upload Disease Signature : a text file describing the differentially expressed genes of the disease state ( Such as in BRCA-related Breast Cancer ). It needs to contain two columns, Gene and direction ( up or down ).

② Upload Reference Signature : A document describing the genetic perturbation of the ' Reference ' that you are concerned about. The document needs to list the genes and their expression status.

③ Users are required to select a dataset for analysis, which integrates over 943,000 perturbational signatures generated by chemical perturbations from large-scale projects such as LINCS L1000. Depending on their specific needs, users can choose between the 'Consensus' dataset, which provides aggregated data, or the more granular 'Tissue' dataset. The latter includes additional layers of information, such as cell line, perturbation time, and drug dose, offering greater resolution for in-depth analysis.

④ Click ' Submit ' to view the visualization results, scores and rankings of drug synergy in the selected data set in ③.

Upon successful completion of the analysis, users will receive visual representations including a scatter plot of the synergy score and a synergy score plot(Figure 5).The output results allow users to adjust the number of top-ranked drugs displayed in the graph(Figure 5A), and the Quick Tips(Figure 5B) provide guidance on how to interpret the Figure 5C and Figure 5D graphs, thereby facilitating the identification of synergistic drug combinations.We also present a multidimensional view that illustrates the distribution of directional data uploaded by users, specifically focusing on landmark genes and the mechanisms of action (MOA) of FDA-approved drugs. When the Consensus data mode is selected, the display emphasizes gene expression patterns and the distribution of drug mechanisms of action. In contrast, selecting the Tissue mode provides a more granular view, highlighting gene signatures, cell lines, and the number of drugs within a specific dataset(Figure 6).

Once an instance is submitted, a results table is generated, sorted by the signature exhibiting the highest score of synergy. Users can filter the results to identify Drug Repurposing Hubs and FDA-approved drugs. In Tissue mode, additional filtering options are available, allowing users to refine results by cell lines, drug concentrations, and time points(Figure 7).

3.3 Synergistic Targets

The Synergistic Target module in PerturbSynergy identifies synergistic target inhibitors through gene knockout and knockdown signatures or target inhibitor input. Synergistic target inhibitors are identified by uploading perturbational signatures from gene knockout signatures or target inhibitor treatments, which reflect the biological effects of gene inhibition. By comparing these signatures with chemical inhibitor data, the platform identifies synergistic drug combinations that may enhance therapeutic efficacy or overcome resistance mechanisms. The system calculates synergy scores, with higher scores indicating greater therapeutic effects from combined target inhibition, suggesting promising drug combinations for further validation. Additionally, the platform allows exploration of synergy across different cell lines, concentrations, and treatment durations, providing a comprehensive approach to discovering novel therapeutic strategies.

The process here is nearly identical to section 3.2, simply follow these steps:

① Upload Disease Signature : a text file describing the differentially expressed genes of the disease state ( such as a certain cancer ). It needs to contain two columns, Gene and direction ( up or down ).

② Upload Reference Signature : A document describing the genetic perturbation(knockout&knockdown) of the ' Reference ' that you are concerned about. The document needs to list the genes and their expression status.

③ Users are required to select a dataset for analysis, which integrates over 137000 perturbational signatures generated by genetics perturbations from large-scale projects such as LINCS L1000. Depending on their specific needs, users can choose between the 'consensus' dataset, which provides aggregated data, or the more granular 'Tissue' dataset. The latter includes additional layers of information, such as cell line, offering greater resolution for in-depth analysis.

④ Click ' Submit ' to view the visualization results, scores and rankings of target synergy in the selected data set in ③.

Upon successful completion of the analysis, users will receive visual representations including a scatter plot of the synergy score and a synergy score plot (Figure 9). The output allows for the adjustment of the number of top-ranked targets displayed in the figure (Figure 9A). Additionally, Quick Tips (Figure 9B) offer guidance on interpreting the graphics in Figures 9C and 9D, which aids in identifying synergistic target-inhibitor combinations. The multidimensional view presented illustrates the distribution of directional data uploaded by the user, as well as the classification of targets. When the 'Consensus' data mode is selected, the display emphasizes the classification of gene expression patterns and targets. Conversely, the 'Tissue' mode provides a more granular view, highlighting the number of gene signatures, cell lines, and knockout/knockdown targets within a specific dataset (Figure 10).

Once an instance is submitted, a results table is generated, sorted by the signature exhibiting the highest score of synergy. Users can filter the results to identify drugable target. In Tissue mode, additional filtering options are available, allowing users to refine results by cell lines(Figure 11).

4 Comparison with Existing Tools:

Compared with existing perturbation analysis tools,PerturbSynergy provides:

① An integrated framework for both drug-drug and drug-target synergy analysis.

② A fully web-based interface requiring no local installation or programming experience.

③ Support for user-defined disease signatures and reference perturbation datasets.

④ Interactive visualization of synergy scores and synergy relationships.

⑤ Downloadable, publication-ready result tables.

We provide a table that highlights the functionalities of PerturbSynergy in comparison with other similar platforms. This comparison helps users understand the unique features and capabilities of PerturbSynergy relative to other available tools in the field.

Figure 12

5 Benchmark & Validation

Predictive performance was evaluated using the independent DrugComb² benchmark dataset. Five clinically representative reference drugs—axitinib, celecoxib, sorafenib, altretamine, and sunitinib—covering the largest numbers of experimentally tested drug pairs in DrugComb were selected for validation.Synergistic drug combinations were defined as positive cases using a Synergy ZIP score threshold greater than 10.For axitinib, there are 31 synergistic and 733 non-synergistic combinations. Celecoxib has 48 synergistic and 1368 non-synergistic combinations. Sorafenib includes 27 synergistic and 772 non-synergistic combinations. Altretamine has 13 synergistic and 478 non-synergistic combinations, while sunitinib shows 17 synergistic and 1328 non-synergistic combinations. The reference drug signature are perturbation signatures for those drugs derived from the LINCS L1000. Across these compounds, PerturbSynergy consistently achieved high precision (≥0.98), strong F1 scores (0.80–0.94), and robust AUC values (0.71–0.76), demonstrating reliable discrimination between synergistic and non-synergistic combinations on unseen data (Fig 13).

Figure 13

6 Method and Validation

The core methodology implemented in PerturbSynergy has been described and validated in apeer-reviewed publication:

Stathias V., Jermakowicz A. M., Maloof M. E., et al.(2018). Drug and disease signature integration identifies synergistic combinations in glioblastoma.Nature Communications,9, 5315. DOI: 10.1038/s41467-018-07659-z.

PerturbSynergy extends this methodology by integrating multi-perturbation synergy analysisand providing an interactive web-based interface for data exploration and visualization.

7 Data Privacy and Data Handling

All data uploaded to PerturbSynergy are treated as strictly confidential.User-submitted disease signatures and reference signatures are processed only for the purpose of the requested analysis and are not shared with other users or third parties.

Uploaded data and intermediate results are stored temporarily on the server and automatically deleted after the analysis is completed.We do not reuse, redistribute, or repurpose user-submitted data for model trainingbenchmarking, or any other secondary use.

PerturbSynergy does not require user registration or login, and no personally identifiable information is collected.

8 Maintenance & Availability

PerturbSynergy is freely accessible to all users, including commercial users, without login.Users may use the web server and generated results for academic or commercial purposes.Please cite our paper when using the server. We are committed to maintaining the server for at least 5 years after publication and will provide a stable URL and versioned updates.

Reference

1 Stathias, V., Jermakowicz, A.M., Maloof, M.E. et al. Drug and disease signature integration identifies synergistic combinations in glioblastoma. Nat Commun 9, 5315 (2018).

2 Bulat Zagidullin, Jehad Aldahdooh, Shuyu Zheng, et al. DrugComb: an integrative cancer drug combination data portal, Nucleic Acids Research, Volume 47, Issue W1, 02 July 2019, Pages W43–W51.