Benchmarking causal reasoning algorithms for gene expression-based compound mechanism of action analysis

BACKGROUND

Elucidating compound mechanism of action (MoA) is beneficial to drug discovery, but in practice often represents a significant challenge. Causal Reasoning approaches aim to address this situation by inferring dysregulated signalling proteins using transcriptomics data and biological networks; however, a comprehensive benchmarking of such approaches has not yet been reported. Here we benchmarked four causal reasoning algorithms (SigNet, CausalR, CausalR ScanR and CARNIVAL) with four networks (the smaller Omnipath network vs. 3 larger MetaBase™ networks), using LINCS L1000 and CMap microarray data, and assessed to what extent each factor dictated the successful recovery of direct targets and compound-associated signalling pathways in a benchmark dataset comprising 269 compounds. We additionally examined impact on performance in terms of the functions and roles of protein targets and their connectivity bias in the prior knowledge networks.

RESULTS

According to statistical analysis (negative binomial model), the combination of algorithm and network most significantly dictated the performance of causal reasoning algorithms, with the SigNet recovering the greatest number of direct targets. With respect to the recovery of signalling pathways, CARNIVAL with the Omnipath network was able to recover the most informative pathways containing compound targets, based on the Reactome pathway hierarchy. Additionally, CARNIVAL, SigNet and CausalR ScanR all outperformed baseline gene expression pathway enrichment results. We found no significant difference in performance between L1000 data or microarray data, even when limited to just 978 'landmark' genes. Notably, all causal reasoning algorithms also outperformed pathway recovery based on input DEGs, despite these often being used for pathway enrichment. Causal reasoning methods performance was somewhat correlated with connectivity and biological role of the targets.

CONCLUSIONS

Overall, we conclude that causal reasoning performs well at recovering signalling proteins related to compound MoA upstream from gene expression changes by leveraging prior knowledge networks, and that the choice of network and algorithm has a profound impact on the performance of causal reasoning algorithms. Based on the analyses presented here this is true for both microarray-based gene expression data as well as those based on the L1000 platform.

Small-molecule screen reveals pathways that regulate C4 secretion in stem cell-derived astrocytes

In the brain, the complement system plays a crucial role in the immune response and in synaptic elimination during normal development and disease. Here, we sought to identify pathways that modulate the production of complement component 4 (C4), recently associated with an increased risk of schizophrenia. To design a disease-relevant assay, we first developed a rapid and robust 3D protocol capable of producing large numbers of astrocytes from pluripotent cells. Transcriptional profiling of these astrocytes confirmed the homogeneity of this population of dorsal fetal-like astrocytes. Using a novel ELISA-based small-molecule screen, we identified epigenetic regulators, as well as inhibitors of intracellular signaling pathways, able to modulate C4 secretion from astrocytes. We then built a connectivity map to predict and validate additional key regulatory pathways, including one involving c-Jun-kinase. This work provides a foundation for developing therapies for CNS diseases involving the complement cascade.

Morphology and gene expression profiling provide complementary information for mapping cell state

Morphological and gene expression profiling can cost-effectively capture thousands of features in thousands of samples across perturbations by disease, mutation, or drug treatments, but it is unclear to what extent the two modalities capture overlapping versus complementary information. Here, using both the L1000 and Cell Painting assays to profile gene expression and cell morphology, respectively, we perturb human A549 lung cancer cells with 1,327 small molecules from the Drug Repurposing Hub across six doses, providing a data resource including dose-response data from both assays. The two assays capture both shared and complementary information for mapping cell state. Cell Painting profiles from compound perturbations are more reproducible and show more diversity but measure fewer distinct groups of features. Applying unsupervised and supervised methods to predict compound mechanisms of action (MOAs) and gene targets, we find that the two assays not only provide a partially shared but also a complementary view of drug mechanisms. Given the numerous applications of profiling in biology, our analyses provide guidance for planning experiments that profile cells for detecting distinct cell types, disease phenotypes, and response to chemical or genetic perturbations.

In silico Drug Screening Approach Using L1000-Based Connectivity Map and Its Application to COVID-19

Conventional drug screening methods search for a limited number of small molecules that directly interact with the target protein. This process can be slow, cumbersome and has driven the need for developing new drug screening approaches to counter rapidly emerging diseases such as COVID-19. We propose a pipeline for drug repurposing combining in silico drug candidate identification followed by in vitro characterization of these candidates. We first identified a gene target of interest, the entry receptor for the SARS-CoV-2 virus, angiotensin converting enzyme 2 (ACE2). Next, we employed a gene expression profile database, L1000-based Connectivity Map to query gene expression patterns in lung epithelial cells, which act as the primary site of SARS-CoV-2 infection. Using gene expression profiles from 5 different lung epithelial cell lines, we computationally identified 17 small molecules that were predicted to decrease ACE2 expression. We further performed a streamlined validation in the normal human epithelial cell line BEAS-2B to demonstrate that these compounds can indeed decrease ACE2 surface expression and to profile cell health and viability upon drug treatment. This proposed pipeline combining in silico drug compound identification and in vitro expression and viability characterization in relevant cell types can aid in the repurposing of FDA-approved drugs to combat rapidly emerging diseases.

High-Throughput Strategies for the Discovery of Anticancer Drugs by Targeting Transcriptional Reprogramming

Transcriptional reprogramming contributes to the progression and recurrence of cancer. However, the poorly elucidated mechanisms of transcriptional reprogramming in tumors make the development of effective drugs difficult, and gene expression signature is helpful for connecting genetic information and pharmacologic treatment. So far, there are two gene-expression signature-based high-throughput drug discovery approaches: L1000, which measures the mRNA transcript abundance of 978 "landmark" genes, and high-throughput sequencing-based high-throughput screening (HTS²); they are suitable for anticancer drug discovery by targeting transcriptional reprogramming. L1000 uses ligation-mediated amplification and hybridization to Luminex beads and highlights gene expression changes by detecting bead colors and fluorescence intensity of phycoerythrin signal. HTS² takes advantage of RNA-mediated oligonucleotide annealing, selection, and ligation, high throughput sequencing, to quantify gene expression changes by directly measuring gene sequences. This article summarizes technological principles and applications of L1000 and HTS², and discusses their advantages and limitations in anticancer drug discovery.

Modeling drug mechanism of action with large scale gene-expression profiles using GPAR, an artificial intelligence platform

Background

Querying drug-induced gene expression profiles with machine learning method is an effective way for revealing drug mechanism of actions (MOAs), which is strongly supported by the growth of large scale and high-throughput gene expression databases. However, due to the lack of code-free and user friendly applications, it is not easy for biologists and pharmacologists to model MOAs with state-of-art deep learning approach.

Results

In this work, a newly developed online collaborative tool, Genetic profile-activity relationship (GPAR) was built to help modeling and predicting MOAs easily via deep learning. The users can use GPAR to customize their training sets to train self-defined MOA prediction models, to evaluate the model performances and to make further predictions automatically. Cross-validation tests show GPAR outperforms Gene set enrichment analysis in predicting MOAs.

Conclusion

GPAR can serve as a better approach in MOAs prediction, which may facilitate researchers to generate more reliable MOA hypothesis.

Integrating adverse outcome pathways (AOPs) and high throughput in vitro assays for better risk evaluations, a study with drug-induced liver injury (DILI)

The emergence of high throughput in vitro assays has the potential to significantly improve toxicological evaluations and lead to more efficient, accurate, and less animal-intensive testing. However, directly using all available in vitro assays in a model is usually impractical and inefficient. On the other hand, mechanistic knowledge has always been critical for toxicological evaluations and should not be ignored even with the increasing availability of data. In this paper, we illus­trate a promising approach to integrating mechanistic knowledge with multiple data sources for in vitro assays, using drug-induced liver injury (DILI) as an example. The adverse outcome pathway (AOP) framework was used as a source for mechanistic knowledge and as a selection tool for high throughput predictors. Our results confirm the value of AOPs as a knowledge source for understanding adverse events and show that the majority of drugs classified as most-DILI-concern were mapped to AOPs related to liver toxicity found in AOPwiki. AOPs were also used effectively to select a subset of assays from the Tox21 and L1000 projects as the predictors in predictive modeling of DILI risk. Together with previously published drug properties for daily dose, lipophilicity, and reactive metabolite formation, these assay endpoints were used to build a penalized logistic regression model for assessing DILI risk. This model obtained an accuracy of 0.91, thus confirming the potential power of integrating mechanistic knowledge with high throughput assays for toxicological evalu­ations. The results also provide a roadmap for data integration that could be used for other adverse effects.

Drug Effect Prediction by Integrating L1000 Genomic and Proteomic Big Data

The library of integrated Network-Based Cellular Signatures (LINCS) project aims to create a network-based understanding of biology by cataloging changes in gene expression and signal transduction. Gene expression and proteomic data in LINCS L1000 are cataloged for human cancer cells treated with compounds and genetic reagents. For understanding the related cell pathways and facilitating drug discovery, we developed binary linear programming (BLP) to infer cell-specific pathways and identify compounds' effects using L1000 gene expression and phosphoproteomics data. A generic pathway map for the MCF7 breast cancer cell line was built. Within them, BLP extracted the cell-specific pathways, which reliably predicted the compounds' effects. In this way, the potential drug effects are revealed by our models.

Drug Signature Detection Based on L1000 Genomic and Proteomic Big Data

The library of integrated Network-Based Cellular Signatures (LINCS) project aims to create a network-based understanding of biology by cataloging changes in gene expression and signal transduction. L1000 big datasets provide gene expression profiles induced by over 10,000 compounds, shRNAs, and kinase inhibitors using L1000 platform. We developed a systematic compound signature discovery pipeline named csNMF, which covers from raw L1000 data processing to drug screening and mechanism generation. The discovered compound signatures of breast cancer were consistent with the LINCS KINOMEscan data and were clinically relevant. In this way, the potential mechanisms of compounds' efficacy are elucidated by our computational model.

Transcriptomic Data Mining and Repurposing for Computational Drug Discovery

Conventional drug discovery in general is costly and time-consuming with extremely low success and relatively high attrition rates. The disparity between high cost of drug discovery and vast unmet medical needs resulted in advent of an increasing number of computational approaches that can "connect" disease with a candidate therapeutic. This includes computational drug repurposing or repositioning wherein the goal is to discover a new indication for an approved drug. Computational drug discovery approaches that are commonly used are similarity-based wherein network analysis or machine learning-based methods are used. One such approach is matching gene expression signatures from disease to those from small molecules, commonly referred to as connectivity mapping. In this chapter, we will focus on how publicly available existing transcriptomic data from diseases can be reused to identify novel candidate therapeutics and drug repositioning candidates. To elucidate these, we will present two case studies: (1) using transcriptional signature similarity or positive correlation to identify novel small molecules that are similar to an approved drug and (2) identifying candidate therapeutics via reciprocal connectivity or negative correlation between transcriptional signatures from a disease and small molecule.

Drug Repurposing Using Deep Embeddings of Gene Expression Profiles

Computational drug repositioning requires assessment of the functional similarities among compounds. Here, we report a new method for measuring compound functional similarity based on gene expression data. This approach takes advantage of deep neural networks to learn an embedding that substantially denoises expression data, making replicates of the same compound more similar. Our method uses unlabeled data in the sense that it only requires compounds to be labeled by identity rather than detailed pharmacological information, which is often unavailable and costly to obtain. Similarity in the learned embedding space accurately predicted pharmacological similarities despite the lack of any such labels during training and achieved substantially improved performance in comparison with previous similarity measures applied to gene expression measurements. Our method could identify drugs with shared therapeutic and biological targets even when the compounds were structurally dissimilar, thereby revealing previously unreported functional relationships between compounds. Thus, our approach provides an improved engine for drug repurposing based on expression data, which we have made available through the online tool DeepCodex ( http://deepcodex.org ).

From gene networks to drugs: systems pharmacology approaches for AUD

The alcohol research field has amassed an impressive number of gene expression datasets spanning key brain areas for addiction, species (humans as well as multiple animal models), and stages in the addiction cycle (binge/intoxication, withdrawal/negative effect, and preoccupation/anticipation). These data have improved our understanding of the molecular adaptations that eventually lead to dysregulation of brain function and the chronic, relapsing disorder of addiction. Identification of new medications to treat alcohol use disorder (AUD) will likely benefit from the integration of genetic, genomic, and behavioral information included in these important datasets. Systems pharmacology considers drug effects as the outcome of the complex network of interactions a drug has rather than a single drug-molecule interaction. Computational strategies based on this principle that integrate gene expression signatures of pharmaceuticals and disease states have shown promise for identifying treatments that ameliorate disease symptoms (called in silico gene mapping or connectivity mapping). In this review, we suggest that gene expression profiling for in silico mapping is critical to improve drug repurposing and discovery for AUD and other psychiatric illnesses. We highlight studies that successfully apply gene mapping computational approaches to identify or repurpose pharmaceutical treatments for psychiatric illnesses. Furthermore, we address important challenges that must be overcome to maximize the potential of these strategies to translate to the clinic and improve healthcare outcomes.

Multi-target drug repositioning by bipartite block-wise sparse multi-task learning

Background

Finding potential drug targets is a crucial step in drug discovery and development. Recently, resources such as the Library of Integrated Network-Based Cellular Signatures (LINCS) L1000 database provide gene expression profiles induced by various chemical and genetic perturbations and thereby make it possible to analyze the relationship between compounds and gene targets at a genome-wide scale. Current approaches for comparing the expression profiles are based on pairwise connectivity mapping analysis. However, this method makes the simple assumption that the effect of a drug treatment is similar to knocking down its single target gene. Since many compounds can bind multiple targets, the pairwise mapping ignores the combined effects of multiple targets, and therefore fails to detect many potential targets of the compounds.

Results

We propose an algorithm to find sets of gene knock-downs that induce gene expression changes similar to a drug treatment. Assuming that the effects of gene knock-downs are additive, we propose a novel bipartite block-wise sparse multi-task learning model with super-graph structure (BBSS-MTL) for multi-target drug repositioning that overcomes the restrictive assumptions of connectivity mapping analysis.

Conclusions

The proposed method BBSS-MTL is more accurate for predicting potential drug targets than the simple pairwise connectivity mapping analysis on five datasets generated from different cancer cell lines.

Availability

The code can be obtained at http://gr.xjtu.edu.cn/web/liminli/codes.

A Library of Phosphoproteomic and Chromatin Signatures for Characterizing Cellular Responses to Drug Perturbations

Although the value of proteomics has been demonstrated, cost and scale are typically prohibitive, and gene expression profiling remains dominant for characterizing cellular responses to perturbations. However, high-throughput sentinel assays provide an opportunity for proteomics to contribute at a meaningful scale. We present a systematic library resource (90 drugs × 6 cell lines) of proteomic signatures that measure changes in the reduced-representation phosphoproteome (P100) and changes in epigenetic marks on histones (GCP). A majority of these drugs elicited reproducible signatures, but notable cell line- and assay-specific differences were observed. Using the "connectivity" framework, we compared signatures across cell types and integrated data across assays, including a transcriptional assay (L1000). Consistent connectivity among cell types revealed cellular responses that transcended lineage, and consistent connectivity among assays revealed unexpected associations between drugs. We further leveraged the resource against public data to formulate hypotheses for treatment of multiple myeloma and acute lymphocytic leukemia. This resource is publicly available at https://clue.io/proteomics.

The Library of Integrated Network-Based Cellular Signatures NIH Program: System-Level Cataloging of Human Cells Response to Perturbations

The Library of Integrated Network-Based Cellular Signatures (LINCS) is an NIH Common Fund program that catalogs how human cells globally respond to chemical, genetic, and disease perturbations. Resources generated by LINCS include experimental and computational methods, visualization tools, molecular and imaging data, and signatures. By assembling an integrated picture of the range of responses of human cells exposed to many perturbations, the LINCS program aims to better understand human disease and to advance the development of new therapies. Perturbations under study include drugs, genetic perturbations, tissue micro-environments, antibodies, and disease-causing mutations. Responses to perturbations are measured by transcript profiling, mass spectrometry, cell imaging, and biochemical methods, among other assays. The LINCS program focuses on cellular physiology shared among tissues and cell types relevant to an array of diseases, including cancer, heart disease, and neurodegenerative disorders. This Perspective describes LINCS technologies, datasets, tools, and approaches to data accessibility and reusability.

Large-Scale Gene Expression Profiling Platform for Identification of Context-Dependent Drug Responses in Multicellular Tumor Spheroids

Cancer cell lines grown as two-dimensional (2D) cultures have been an essential model for studying cancer biology and anticancer drug discovery. However, 2D cancer cell cultures have major limitations, as they do not closely mimic the heterogeneity and tissue context of in vivo tumors. Developing three-dimensional (3D) cell cultures, such as multicellular tumor spheroids, has the potential to address some of these limitations. Here, we combined a high-throughput gene expression profiling method with a tumor spheroid-based drug-screening assay to identify context-dependent treatment responses. As a proof of concept, we examined drug responses of quiescent cancer cells to oxidative phosphorylation (OXPHOS) inhibitors. Use of multicellular tumor spheroids led to discovery that the mevalonate pathway is upregulated in quiescent cells during OXPHOS inhibition, and that OXPHOS inhibitors and mevalonate pathway inhibitors were synergistically toxic to quiescent spheroids. This work illustrates how 3D cellular models yield functional and mechanistic insights not accessible via 2D cultures.