In Roudnicky et al. (PNAS 2020), we reported how we identified inducers of the endothelial cell barrier, among others by using the SPARK (Small-molecule PAthway Research Kit) library and genome-edited endothelial cells derived from human pluripotent stem-cells.

Some biological background

If you have not heard of endothelial cells, they form the inner surface of blood vessels and lymphatic vessels as a cell monolayer. Endothelial cells are involved in many biological processes, including blood clotting, inflammation, formation and growth of new blood vessels (known as angiogenesis), and regulation of blood pressure.

A speciality of endothelial cells is that they can form tight barriers, known as endothelial barriers, that control the flow of substances, fluid, and immune cell into and out of a tissue. In particular, the blood-retinal barrier (often abbreviated as BRB) and blood-brain barrier (BBB) contain endothelial cells which let selected molecules but not others pass through. In addition to usual functions of endothelial cells, they support and protect cells in the central nervous system. Therefore, it may not surprise you that impaired integrity of these barriers and malfunction of endothelial cells can cause many diseases or contribute to disease progression.

A phenotypic screening to identify pathways regulating endothelial cells

How can we develop drugs to improve the integrity of endothelial barriers? The first hurdle is to identify genes and biological pathways that regulate the barrier forming function of endothelial cells, and second is to find out against which targets can we develop drugs to modulate their function. To address both questions at the same time, we can screen many compounds with annotated target and pathway-regulation profiles with a endothelial-cell-based phenotypic assay. That is exactly what we did.

Why did we setup a phenotypic screening? When the biological system that we want to model is complex (for instance endothelial cells in our case) and when we want to identify novel regulators of the biological system, cell-based phenotypic screening is probably an better approach than a biochemical assay which focuses on one particular target that we assume is important for the system. Compared with biochemical assays, a robust and well-designed cell-based assay can reveal effects of pharmaceutical intervention in cells, which better models the situation where drugs enter in our body and reach the cells. At the same time, cell-based phenotypic screening offers a higher throughput than in vivo animal models. Human cell-based assays, in addition, may potentially address human biology with reduced dependency on animal models.

In this study, we used stem-cell-derived endothelial cells as the cell system to model endothelial-cell biology. These cells, even without genome editing, already show endothelial-cell-like expression profiles and functions, which was revealed among others by our analysis with the BioQC algorithm and a comprehensive survey of the FANTOM5 database (see details in the supplementary information). Furthermore, we used the genome-editing technique to modify the cells so that they report the abundance of Claudin-5, the protein encoded by gene CLDN5. The edited CLDN5 gene is coupled with a green fluorescent protein (GFP) reporter, so that the cells with higher expression of CLDN5 protein emit brighter green fluorescent when exposed to light in the blue to ultraviolet range. We used the reporter as the readout of our phenotypic screening, because CLDN5 has been found to stabilizes endothelial cells, and we believe that its expression is correlated with and mechanistically linked with the protective function of endothelial cells. Using flow cytometry, we could identify drug-like molecules that induce CLDN5 expression, which, if further validated, hopefully improve the protective function of endothelial cells in blood-retinal and blood-brain barriers.

But what chemical molecules do we want to screen with and test on these cells? There are millions of compounds that have been synthesized and reported in public and proprietary databases, and many more compounds can be at least theoretically constructed - the potential number of drug-like molecules far exceeds the total number of atoms in the universe. How can we select a compact set of molecules, which is nevertheless comprehensive with regard to their biological spaces and pathway regulation patterns, in order to identify genes and pathways that are putative therapeutic starting points?

The SPARK library

To address this challenge, we combined data, prior knowledge, and data analysis to construct the SPARK library. The workflow of construction and the benchmark results can be found in the Figure 2 of the paper. In short, we integrated high-quality compound-activity-target data from ChEMBL, a public database, and Roche’s proprietary database. By using Gini Index, a metric to quantify the specificity of compound against targets (which we also used in our BioQC software), and by using Affinity propagation (AP), an unsupervised machine-learning algorithm, we identified compounds that are specific, potent, and representative with regard to their biological targets. We further manually curated the list and annotated the compounds with biological pathways. The result is SPARK, a pathway-annotated chemogenomic library with about two-thousand small molecules.

The screening and RepSox, an interesting hit

Now we have a cell system ready for screening and a chemogenomic-library with small molecules associated with target genes and pathways, we can use them to identify small molecules and targets that modulate Claudin-5, our molecular phenotypic marker of endothelial cell integrity. The screening may sound less glamorous than building a genome-edited reporter cell system or constructing a new chemogenomic library, but in fact it is at least equally important as other activities, because its quality directly determines what we can get for further study.

Thanks to detail-oriented work by the team, we identified 62 compounds that activated CLDN5. Interestingly, several hits that we found are TGF-beta pathway inhibitors. While it has been long suspected that the TGF-beta pathway is involved in modulating endothelial cell integrity, few results are available to reveal the causal link between the two, and little is known about the biological effect and the therapeutic potential of modulating the TGF-beta pathway in retinal endothelial cells with small molecules.

Further work revealed that RepSox, a TGF-beta pathway inhibitor, elevated resistance of endothelial barriers in primary retinal and brain endothelial cells. It also reduces para-cellular permeability and prevented barrier breakdown induced by vascular endothelial growth factor A (VEGFA) in vitro. In animal models, it altered vascular patterning in the mouse retina during development.

Further studies of RepSox to reveal its activities against kinases (kinome profiling), gene expression profiles (transcriptome profiling), and protein expression and post-translation modulation profiles (proteome profiling) revealed that RepSox inhibits several biological pathways besides TGF-beta, especially downstream pathways of VEGFA and inflammation gene networks. At the same time, it activates Notch and Wnt pathways, which are believed to stabilize vasculature and to promote the establishment of endothelial barriers, while inducing individual tight junctions and transporters. In short, RepSox modulates endothelial cells by targeting multiple pathways.

My learnings

Put together, the results suggest that inhibiting multiple pathways may be an effective strategy for the development of better endothelial-cell-barrier models and new therapeutics that induce endothelial-cell barriers. From a drug-discovery perspective, the study suggests that it is feasible to discover target genes and pathways by screening relevant in vitro cell systems with pathway-annotated small-molecule libraries such as SPARK. Personally, I was particularly excited by learning more about the endothelial biology in a team that include experts in stem-cell technology, genome editing, phenotypic screening, omics, and data analysis.

A few studies in the recent years have described new chemogenomic library sets, in particular those from the BROAD institute (The Drug Repurposing Hub), Cancer Research UK (The Probe Miner), Novartis (Systematic Chemogenomic Library Assembly), and other ongoing efforts like Open Access Chemogenomics Library and Chemical Probes for the Druggable Genome. In our manuscript, we also share with the community how we constructed the SPARK library and how its application helped us gain new insights in endothelial cell biology. While chemogenomic libraries with pathway annotation are not silver bullets to solve all problems in drug discovery, if used wisely and appropriately, they can help identify new and sometimes unexpected targets and pathways.

Finally, our work complements a recent publication by Ietswaart et al. on EBioMedicine, which is co-authored by colleagues at Novartis Institute of Biomedical Research and academic partners, which uses compound-target-activity data and adverse drug reaction (ADR) data to associate drug targets with adverse effects. Though the scope, data source, and methodologies of both studies are distinct, they both suggest that in vitro pharmacology data can be used to address safety and efficacy questions of future drugs. We can probably learn how to develop new drugs more effectively by understanding our established drugs better.