Phenotypic Screening of Chemical Libraries Enriched by Molecular Docking to Multiple Targets Selected from Glioblastoma Genomic Data

Article abstract of DOI:10.1021/acschembio.0c00078

Like most solid tumors, glioblastoma multiforme (GBM) harbors multiple overexpressed and mutated genes that affect several signaling pathways. Suppressing tumor growth of solid tumors like GBM without toxicity may be achieved by small molecules that selec

Full text of DOI:10.1021/acschembio.0c00078

Subscriber access provided by Karolinska Institutet, University Library
Article
Phenotypic Screening of Chemical Libraries Enriched by Molecular
Docking to Multiple Targets Selected from Glioblastoma Genomic Data
David Xu, Donghui Zhou, Khuchtumur Bum-Erdene, Barbara J Bailey, Kamakshi
Sishtla, Sheng Liu, Jun Wan, Uma K. Aryal, Jonathan A Lee, Clark D Wells,
Melissa L. Fishel, Timothy W Corson, Karen Pollok, and Samy O. Meroueh
ACS Chem. Biol., Just Accepted Manuscript • DOI: 10.1021/acschembio.0c00078 • Publication Date (Web): 03 Apr 2020
Downloaded from pubs.acs.org on April 5, 2020
Just Accepted
“Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted
online prior to technical editing, formatting for publication and author proofing. The American Chemical
Society provides “Just Accepted” as a service to the research community to expedite the dissemination
of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in
full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully
peer reviewed, but should not be considered the official version of record. They are citable by the
Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore,
the “Just Accepted” Web site may not include all articles that will be published in the journal. After
a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web
site and published as an ASAP article. Note that technical editing may introduce minor changes
to the manuscript text and/or graphics which could affect content, and all legal disclaimers and
ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or
consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W.,
Washington, DC 20036
Published by American Chemical Society. Copyright © American Chemical Society.
However, no copyright claim is made to original U.S. Government works, or works
produced by employees of any Commonwealth realm Crown government in the course
of their duties.

Page 1 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Fig. 1
3
34x178mm (300 x 300 DPI)
ACS Paragon Plus Environment

ACS Chemical Biology
Page 2 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Fig. 2
3
30x172mm (300 x 300 DPI)
ACS Paragon Plus Environment

Page 3 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Fig. 3
2
91x172mm (300 x 300 DPI)
ACS Paragon Plus Environment

ACS Chemical Biology
Page 4 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Fig. 4
228x182mm (300 x 300 DPI)
ACS Paragon Plus Environment

Page 5 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Fig. 5
3
37x173mm (300 x 300 DPI)
ACS Paragon Plus Environment

ACS Chemical Biology
Page 6 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
TOC
3
38x190mm (300 x 300 DPI)
ACS Paragon Plus Environment

Page 7 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
Phenotypic Screening of Chemical Libraries Enriched by Molecular Docking
to Multiple Targets Selected from Glioblastoma Genomic Data
David Xu2,3†, Donghui Zhou , Khuchtumur Bum-Erdene , Barbara J. Bailey4,5,6,
1†
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
8
2,7
2,7
9
1
Kamakshi Sishtla , Sheng Liu , Jun Wan , Uma K. Aryal , Jonathan A. Lee ,
Clark D. Wells , Melissa L. Fishel4,5,6, Timothy W. Corson1,5,6,8, Karen E.
1
Pollok4,5,6*, and Samy O. Meroueh1,2,6
*
1Department of Biochemistry and Molecular Biology, Indiana University School of Medicine, Indianapolis, IN 46202
2Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN 46202
3Department of BioHealth Informatics, Indiana University School of Informatics and Computing, Indianapolis, IN 46202
4Department of Pediatrics, Herman B. Wells Center for Pediatric Research, Indiana University School of Medicine, Indianapolis, IN
46202
5Department of Pharmacology and Toxicology, Indiana University School of Medicine, Indianapolis, IN 46202
6Indiana University Melvin and Bren Simon Cancer Center, Indiana University School of Medicine, Indianapolis, IN 46202
7Department of Medical and Molecular Genetics, Indiana University School of Medicine, Indianapolis, IN 46202
8Eugene and Marilyn Glick Eye Institute, Department of Ophthalmology, Indiana University School of Medicine, Indianapolis, IN 46202
9Purdue Proteomics Facility, Bindley Bioscience Center, Purdue University, West Lafayette, IN 47907
*
Corresponding authors
Authors contributed equally
†
Corresponding Authors: Samy Meroueh
Department of Biochemistry and Molecular Biology
Indiana University School of Medicine
35 Barnhill Drive, MS4021
6
Indianapolis, IN 46202
Tel: (317) 274-8315
E-mail: smeroueh@iu.edu
Karen Pollok
Department of Pediatrics
Indiana University School of Medicine
1044 W Walnut St, Indianapolis, IN 46202, R4 302
Indianapolis, IN 46202
Tel: (317) 274-8891
E-mail: kpollok@iu.edu
ACS Paragon Plus Environment

ACS Chemical Biology
Page 8 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
ABSTRACT
Like most solid tumors, glioblastoma multiforme (GBM) harbors multiple overexpressed and
mutated genes that affect several signaling pathways. Suppressing tumor growth of solid tumors
like GBM without toxicity may be achieved by small molecules that selectively modulate a
collection of targets across different signaling pathways, also known as selective
polypharmacology. Phenotypic screening can be an effective method to uncover such
compounds, but the lack of approaches to create focused libraries tailored to tumor targets has
limited its impact. Here, we create rational libraries for phenotypic screening by structure-based
molecular docking chemical libraries to GBM-specific targets identified using the tumor’s RNA-
seq and mutation data along with cellular protein-protein interaction data. Screening of this
enriched library of 50 candidates led to several active compounds. Among them 1 (IPR-2025),
which (i) inhibited cell viability of low-passage patient-derived GBM spheroids with single-digit
micromolar IC s that are substantially better than standard-of-care temozolomide; (ii) blocked
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
50
tube-formation of endothelial cells in Matrigel with sub-micromolar IC s; and (iii) had no effect on
50
+
primary hematopoietic CD34 progenitor spheroids or astrocyte cell viability. RNA sequencing
RNA-seq) provided potential mechanism of action of 1 and mass spectrometry-based thermal
(
proteome profiling (TPP) confirmed that the compound engages multiple targets. The ability of 1
to inhibit GBM phenotypes without affecting normal cell viability suggests that our screening
approach may hold promise for generating lead compounds with selective polypharmacology for
the development of treatments of incurable diseases like GBM.
ACS Paragon Plus Environment

Page 9 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
INTRODUCTION
Like most solid incurable tumors, glioblastoma multiforme (GBM) exhibit multiple
hallmarks of cancer as delineated by Hanahan and Weinberg(1): Self-sufficiency in growth
signals, insensitivity to growth inhibitory signals, evasion from programmed cell death (apoptosis),
ability to undergo limitless cycles of cell growth, sustained ability to be supplied by blood
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
(angiogenesis), and aggressive invasion of the brain parenchyma. Large-scale sequencing
studies of human tumors such as The Cancer Genome Atlas (TCGA) project have revealed that
the complex phenotypes that define cancer are driven by a large number of somatic mutations
that occur in proteins across the cellular network(2). Whole genome sequencing studies that have
profiled the molecular signatures of various cancers, such as in ovarian(3), colorectal(4),
breast(5), renal(6), lung(7-9), pancreatic(10,11), and brain(12,13), have further confirmed that the
complex phenotypes are driven by multiple targets spanning interconnected signaling pathways
across the human protein-protein interaction network.
Suppressing the growth and metastasis of solid tumors harboring tens of mutations
without causing toxicity will require small molecules that selectively modulate multiple targets and
signaling pathways. The identification of such compounds is a significant challenge, as most
drugs that target multiple proteins are discovered in a retrospective manner. In some cases, drugs
are developed to inhibit one target, but later, they are found to have additional targets(14).
Metformin is one example of a drug that was developed to treat diabetes, but was later found to
have beneficial effects in cancer likely due to off-targets of the drug(15). Drugs that were
discovered either from natural products or phenotypic screens also often work through
polypharmacology. One example is aspirin, which not only inhibits cyclooxygenases but also
neutrophil activation and inducible nitric acid synthesis(16).
There has been a resurgence of interest in phenotypic screening in cancer drug
discovery(17). Between 1999 and 2008, over half of FDA-approved first-in-class small-molecule
drugs were discovered through phenotypic screening(18). The increased interest in phenotypic
screening is due in part to the lack of effective treatment options for incurable tumors such as
GBM, which remains the most aggressive brain tumor and responds poorly to standard-of-care
therapy that includes surgery, irradiation, and temozolomide. Standard-of-care therapies for GBM
have been essentially unchanged for decades with a median survival of only 14-16 months and
a five-year survival rate of 3–5%(19,20). Ineffective tumor cell killing is largely due to intra-tumoral
genetic instability which allows these malignancies to modulate cell survival pathways,
angiogenesis, and invasion(21,22). In addition, the highly immunosuppressive GBM
ACS Paragon Plus Environment

ACS Chemical Biology
Page 10 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
microenvironment complicates therapeutic approaches that minimize tumor burden and promote
host immunity(23-25). Moreover, investigations to date indicate therapies that combine TMZ with
immunotherapy-based approaches can either promote or deplete immunity(26-29). Phenotypic
screening can be an effective strategy for the development of small molecules to perturb the
function of proteins that drive tumor growth and invasion. Despite the increased interest in
phenotypic screening in cancer drug discovery, the main limitations of the approach include (i)
the lack of methods to tailor library selection to the tumor genome, (ii) cellular assays that do not
accurately represent a tumor, (iii) overreliance on immortalized cell lines, (iv) targeting a single
protein when tumors are driven by multiple proteins, and (v) confining compound screening to
one phenotype.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
To date, most phenotypic screens are carried out on well-annotated tool compound
libraries that include FDA-approved drugs. These are known as chemogenomic libraries, and they
are used to uncover new biology for targets associated with these compounds or for drug-
repurposing purposes(30-33). However, existing approved drugs and tool compounds act on less
than 5% of targets in the human genome(34). The lack of target diversity in chemogenomic
libraries presents an opportunity for the development of new chemical libraries or for the
enrichment of existing libraries. The creation of diverse libraries for high-throughput screening is
a major challenge considering the vastness of chemical space. Just among commercially-
available compounds, there are now at least 400 million small organic compounds that can be
purchased(35,36). In addition, specialized libraries designed using diversity-oriented synthesis
(DOS)(37) and de novo combinatorial libraries such as SCUBIDOO(38) offer additional avenues
to screen unexplored chemical space.
Traditional two-dimensional monolayer assays utilizing cancer cell lines have been the
most practical method to phenotypically screen these large libraries. However, these screening
campaigns have yielded compounds that fail to model compound efficacy and cytotoxicity in more
disease-relevant assays(39). Traditional two-dimensional assays do not accurately capture the
three-dimensional microenvironment of tumors, thus leading to toxic compounds that generally
tend to block microtubule dynamics or lead to DNA modification(40,41). The use of immortalized
cell lines to predict efficacy is now also recognized to be inadequate(42). There are now many
examples of small molecules that are efficacious in traditional in vitro and in vivo models yet fail
to show clinical efficacy(43). As a result, there has been intense interest in the development of
more sophisticated three-dimensional assays. Cancer cells grown in three-dimensional spheroids
are now widely used to investigate the effects of small molecules on tumor growth and other
ACS Paragon Plus Environment

Page 11 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
endpoints such as invasion and remodeling of the tumor matrix. More sophisticated assays, such
as spheroid and organoids, have been developed to better represent the tumor and its
microenvironment.
Here we follow a rational approach to create chemical libraries that are used for phenotypic
screening to generate small molecules with selective polypharmacology that inhibit GBM growth
without affecting non-transformed normal cell lines. Our approach combines catalogs of
differentially expressed molecular targets identified by tumor genomic profiles along with cellular
protein-protein interaction data to select a collection of targets with druggable binding pockets.
Approximately 9000 in-house compounds are docked to each of these targets. Small molecules
that are predicted to simultaneously bind to multiple proteins are selected for phenotypic
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
screening using three-dimensional spheroids of patient-derived GBM cells. Compounds were
tested in non-transformed primary normal cell lines in (i) a three-dimensional assay using CD34+
progenitor cells and (ii) a two-dimensional assay using astrocytes. The effect of the compound on
angiogenesis was also tested using a tube formation assay with brain endothelial cells. To
uncover potential mechanisms of action, the compound was selected for RNA sequencing of
compound-treated and untreated cells. Thermal proteome profiling was performed to identify
potential targets. Cellular thermal shift assays using antibodies were used to confirm the binding
of the compound to targets that emerged from the thermal proteome profiling study.
RESULTS
Target Selection, Virtual Screening, and Rank-Ordering of Chemical Library. A
weakness of current implementations of phenotypic screening is the lack of rational approaches
in the creation of chemical libraries. Here, we propose a strategy that uses the tumor’s genomic
profile to enrich chemical libraries for phenotypic screening (Fig. 1a). The process begins with
the identification of druggable pockets on a large number of protein structures obtained from the
Protein Data Bank (PDB)(44). In previous work, we searched for druggable binding sites on
proteins implicated in a range of cancers and classified them in the context of functional
importance(45). Druggable binding sites were classified based on whether they occurred at a
catalytic site (ENZ), a protein-protein interaction interface (PPI), or an allosteric site (OTH).
Using our approach, druggable binding sites were identified for GBM. Gene expression
profiles were collected for GBM patients from TCGA. A total of 169 GBM tumors and 5 normal
samples have been characterized using RNA sequencing platforms. The data were used to
perform differential expression analysis to identify genes that are overexpressed in GBM (p <
ACS Paragon Plus Environment

ACS Chemical Biology
Page 12 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
.001, FDR < 0.01, and log fold change (log FC) > 1) (Fig. 1b). In addition, a set of somatic
2
2
mutations was retrieved from GBM patients at TCGA and identified for 158 of the 169 tumor
samples. In total, 755 genes with somatic mutations were overexpressed in GBM patient samples.
The set of 755 genes were subsequently filtered based on whether their protein products
are involved in protein-protein interactions. Two large-scale protein-protein interaction networks
of the human proteome were recently described by Rolland and co-workers(46). The first network
is from literature curation of seven widely-used protein-protein interaction databases. The second
network is based on systematically mapping human binary protein-protein interactions. The two
protein-protein interaction datasets from both the literature-curated and experimentally-
determined networks were combined to form a large-scale protein-protein interaction network
consisting of approximately 8000 proteins and 27000 interactions. The protein products of the
genes implicated in GBM were mapped onto this protein-protein interaction network to construct
a GBM subnetwork (Fig. 1c). Among the 755 previously-identified genes implicated in GBM, 390
had at least one interaction in the network. In total, 117 of the 390 proteins had at least one
druggable binding site (Fig. 1d and Table S1).
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
To identify small molecules that inhibit phenotypes associated with GBM, an in-house
library of approximately 9000 compounds was docked to the set of 316 druggable binding sites
on proteins in the GBM subnetwork. The Support Vector Machine-Knowledge Based (SVR-
KB)(47) scoring method was used to predict the binding affinities of each pair of protein-
compound interactions. The number of druggable binding sites with affinities better than a given
SVR-KB cutoff was used to rank-order compounds (Fig. 1e). In this work, an SVR-KB cutoff
corresponding to a computational binding affinity of 10 nM was used. Approximately 55 percent
of compounds were predicted to bind to five or less binding sites, and 20 percent of the docked
compounds were predicted to bind to none of the pockets in the GBM subnetwork (Fig. 1f). Less
than 4 percent of the compounds were predicted to bind to at least 30 of the 316 binding sites.
Small molecules with the highest number of predicted GBM targets were selected for further
testing. A total of 154 compounds were selected for the first screen that were predicted to target
between 38 and 86 binding sites on GBM-specific proteins. The mean number of predicted targets
is 47 for the 154 compounds. The top 154 compounds were clustered with hierarchical clustering
using chemical similarity. For each cluster, the compound corresponding to the cluster center was
selected for phenotypic testing (Table S2).
Exploring Compounds in Patient-Derived GBM Spheroids. A total of 47 compounds
selected from the library of 9000 compounds were tested in a three-dimensional spheroid viability
ACS Paragon Plus Environment

Page 13 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
assay(48,49) using the patient-derived primary glioma cell line GBM43(50) at 25 µM (Fig. 2a).
Most compounds showed little to no effect on GBM43 cell viability, except for 1 (IPR-2025), which
inhibited cell viability by 90% (Fig. 2b). The compound was resynthesized and its stability was
tested following incubation in buffer (Fig. 2c). The compound was stable in buffer. A follow-up
concentration-dependent study of 1 revealed an IC of 3.7 ± 0.1 µM in GBM43 (Fig. 2d and e).
50
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Compound 1 was tested in two additional glioblastoma spheroid models derived from different
patients with recurrent GBM, namely GBM10 and SJ-GBM2. GBM10(50) is derived from an adult
patient, while SJ-GBM2(51) is derived from a pediatric patient. The IC50 of 1 was 3.1 ± 1.5 µM
and 2.9 ± 1.9 µM in GBM10 and SJ-GBM2, respectively. The effect of 1 on normal and non-
+
transformed cell viability was explored using CD34 progenitor cells and astrocytes (Fig. 2d). Cell
+
viability studies for CD34 cells were performed using a colony formation assay, while a
monolayer MTT assay was used for astrocytes. Compound 1 had no effect on the cell viability of
+
CD34 or astrocyte cells up to 100 µM. For the sake of comparison, the standard-of-care
temozolomide inhibits GBM43 with an IC of 244 ± 24 µM(52). However, GBM43 is known to be
50
moderately resistant to temozolomide(53). An alternative treatment option is the
chemotherapeutic CCNU (lomustine), which inhibited GBM43 and GBM10 cell viability weakly,
showing about 50% inhibition at 100 µM (Fig. 2f). It is important to note that the astrocyte assay
is a 2-D assay, which is different than the 3-D assay used for the GBM cell viability studies, thus
direct comparison of the exposure times between the two assays is not possible. Cells grown in
2-D are generally more sensitive to chemicals compared with cells grown in 3-D. In 2-D, cells are
not in contact with each other, but they make contacts with the plastic surface, and adopt a
flattened morphology. Cells in a 3-D environment, on the other hand, mimic the morphology of
cells found in tumors due to the fact that cells form contacts with other cells, triggering distinct
signaling pathways that makes them less sensitive to anti-tumor agents. In addition, cells at the
center of spheroids experience hypoxia, which further enhances their survival, making them less
prone to cell killing by compounds. Furthermore, in the 3-D environment found in spheroids,
compounds must be able to diffuse to access cells that are at the center of these spheroids, which
is often another mechanism of resistance, especially for compounds that are less likely to diffuse
into these spheroids due to their physico-chemical properties.
The activity of 1 was also assessed in three pancreatic ductal adenocarcinoma patient-
derived spheroid models(54-56): Pa02C, Panc10.05 (Pa16C), and Panc198 (Pa20C). Pa02C is
from liver metastasis of pancreatic cancer, while Panc10.05 and Panc198 are from the primary
pancreatic tumor. Compound 1 showed no activity in Pa02C and Panc198 and weak activity in
Panc10.05 with an IC of 26.0 ± 5.9 µM IC .
50
50
ACS Paragon Plus Environment

ACS Chemical Biology
Page 14 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
Structure-Activity Relationship (SAR) of 1 (IPR-2025). A set of 5 analogs from our
internal library were identified from commercial libraries with high Tanimoto similarity to 1 (Table
S3): 2 (GFI-027), 3 (IPR-1909), 4 (RAG-021), 19 (IPR-2024), and 20 (KLM-017). These 5
compounds were tested in a concentration-dependent manner in GBM43 (Fig. S1). Only 19
inhibited in a concentration-dependent manner like 1, although the compound only reaches 60%
inhibition at 100 µM in GBM10 and has an approximate IC50 of 33 µM and no effect in GBM43.
An analog-by-catalog approach was followed to identify another 15 derivatives. These
compounds were tested in GBM43 and GBM10 in a concentration-dependent manner (Fig. S1).
Substitution of the fused tricyclic moiety in 14 (IPR-3440), 20 (KLM-17), and 21 (IPR-3442),
among others, led to loss of activity. Among the derivatives that share the fused tricyclic moiety
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
of the parent, two also feature the sulfonyl group on R . 5 (IPR-3474) substitutes the toluene
1
group on R with a biphenyl and lacks the carbonyl group on the pyrrolidinone group, while 6 (IPR-
3476) replaces the pyrrolidinone group with an amine linker. These two compounds suggest the
1
importance of the carbonyl group on the pyrrolidinone in R . Similarly, compounds that lack the
1
sulfonyl group on R , like 3, also lacked activity.
To further assess the importance of the sulfonyl moiety, three additional compounds, 22
IPR-3502), 23 (IPR-3503), and 24 (IPR-3504), were synthesized that lack key moieties in R1
(
(
Fig. 3). The removal of the carbonyl group on the pyrrolidinone, sulfonyl linker, or both the
sulfonyl linker and the methylbenzene group in 22, 23, and 24, respectively, led to loss of activity
in both GBM43 and GBM10. Three additional derivatives were synthesized to add halogen groups
3
to 1. Substitution of the methyl group in R with a trifluoromethyl in 25 (IPR-3593) resulted in
nearly ten-fold decrease in IC across each of the GBM spheroids compared to 1. Chlorine atoms
50
were added to two separate positions of the fused ring structure in 26 (IPR-3594) and 27 (IPR-
595), resulting in no activity in either compound across the GBM spheroids. Among the 21
3
analogs of 1 that were tested across each of the three GBM spheroids, only the trifluoromethyl 25
resulted in measurable IC s.
50
Compound 1 Inhibits Tube-Formation in Matrigel. Since the seminal work of
Folkman(57), uncontrolled angiogenesis (the process of new blood vessel growth from existing
ones) has become an established hallmark of cancer(1). Solid tumors require a dedicated blood
supply once they reach a certain, limiting size, and antiangiogenic agents can block tumor growth
by starving the tumor of oxygen and nutrients. A so-called “angiogenic switch” is an integral part
of tumor development: varied tumor types secrete vascular endothelial growth factor (VEGF) and
other proangiogenic stimuli and downregulate antiangiogenic proteins. There have also been
ACS Paragon Plus Environment

Page 15 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
some promising clinical results with angiogenesis inhibitors in improving progression-free survival
in both primary and recurrent GBM(58,59). This cancer is also characterized by microvascular
proliferation(60) and high levels of VEGF(61). Given the ample evidence of the importance for
angiogenesis in the biology of these tumors, there is significant need to identify novel compounds
with specific antiangiogenic activity. The parent compound 1 was tested for its ability to inhibit
tube formation of brain microvascular endothelial cells (Fig. 2g). Compound 1 inhibited tube
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
formation with approximate 0.1 µM IC (Fig. 2h). The effect of compound 1 on BMECs revealed
50
an IG50 3.5 µM after a 48-hour exposure (Fig. S2). In comparison, pronounced effects on tube
formation were seen at doses well below this, i.e. 100 nM and 1 µM (Fig. 2h). Moreover, the tube
formation assay was conducted over a short term (8 hours), and as seen in the images in Fig.
2g, there were no obvious signs of cell death at these concentrations.
Structural Analysis and RNA Sequencing to Uncover Compound 1 Mechanism of
Action. The binding modes of 1 to each of the targets predicted by the SVR-KB scoring function
were examined in detail. Each target was classified using the structural context of the binding site
and functional context of the protein (Table 1). Generally, the binding modes of 1 were at allosteric
sites outside the active site on enzymes. Three allosteric sites were adjacent to known protein-
protein interaction interfaces on NCF2, NEDD4, and PYGL. Similarly, 1 was predicted to bind to
allosteric sites adjacent to the active sites on KIF11 and TAP1. The compound was predicted to
bind at pockets at the protein-protein interfaces at the PLK1 polo box and at the homodimers of
NCF1 and PNP. The compound was also predicted to bind to the β-propeller structures of CDC20,
GNB1, ITGA5, and RACK1.
RNA sequencing was performed on untreated GBM43 cells and GBM43 cells treated with
10 µM 1 to validate the predicted targets of 1 and uncover a potential mechanism of action (Fig.
4a). Differential expression analysis (p < 0.001, FDR < 0.01, |log FC| > 1) revealed a set of 15
2
overexpressed and 20 underexpressed genes in GBM43 cells treated with 1 (Fig. 4b). The sets
of overexpressed and underexpressed were separately analyzed for overrepresented GO terms.
No significantly overrepresented terms were found among either set of differentially expressed
genes. The set of differentially expressed genes (DEGs) were compared with those known to be
causally implicated in cancer using the Cancer Gene Census (CGC)(62). Two DEGs were
previously identified as oncogenes. The transcription factor GATA2 has been shown to promote
GBM progression through the EGFR/ERK/Elk-1 pathway(63). Similarly, the kinase KDR
(VEGFR2) is a known oncogene in lung, blood, and skin cancers and plays a key role in regulating
angiogenesis(64).
ACS Paragon Plus Environment

ACS Chemical Biology
Page 16 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
The expression profile of cells treated with 1 was compared to the gene signatures of
previously characterized compounds and gene knockdowns using the LINCS L1000 platform. The
L1000 platform is an extension of the Connectivity Map (CMap) project(65), which used
similarities in gene expression signatures to discover shared mechanisms of action between small
molecule and genetic perturbations. The L1000 platform expands on CMap and uses a panel of
approximately 1,000 landmark genes to characterize the molecular profiles of over 19,000
compound and 5,000 gene perturbations across nine cell lines(66). To compare the gene
expression profile of 1 to the existing signatures in L1000, a signature query was generated using
the sets of overexpressed and underexpressed genes. The similarity between the signature query
is compared with all other signatures in L1000 using a connectivity score (τ), which corresponds
to the percentage of all reference gene sets that are more similar than the observed gene
signature. The most similar 30 gene and small molecule perturbations of 1 are shown in Fig. 4c.
The most similar gene signatures are gene knockdowns of GNA15 and OGFOD1. The G protein
α-subunit GNA15 is part of the heterotrimeric G protein complex consisting of α-subunit and βγ
complex, which mediate downstream signaling of G protein-coupled receptors (GPCRs)(67).
While no α-subunits were among the targets predicted to bind to the compound, both the G protein
β-subunit GNB1 and β-like subunit RACK1 were predicted targets of 1. RACK1 acts as a
scaffolding protein in transcription regulation of the transcription factor EIF4E(68). Gene silencing
and knockdown studies of RACK1 have resulted in promotion of apoptosis and inhibition of cell
proliferation, migration, and invasion in glioma(69,70). The oxygenase OGFOD1 belongs to a
family of transcriptional factor and chromatin regulators and has been found to inhibit cell
proliferation in breast cancer cells(71). Also among the knockdown gene signatures most similar
to the gene signature of 1 in L1000 are the tumor suppressors AES(72) and TP53BP2(73).
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Thermal Proteome Profiling to Identify Potential Targets of 1 (IPR-2025). Once a
compound is identified in a phenotypic screen, the challenge is to identify its targets. Several
methods have been used in the past for target identification, such as biochemical, genomic, or
computational approaches(74-77). Computational methods make use of statistical and machine
learning methods to link a novel compound to existing compounds by identifying common features
between the two. Previously developed algorithms have used similarities in compound
scaffolds(78,79), protein structure similarity(80), side effects(81), and bioactivities(82) to infer a
compound’s target. A more recent mass spectrometry-based method is thermal proteome
profiling, which allows for systematic identification of a compound’s direct targets by analyzing
shifts in melting temperature of the targets in cells(83,84). As a compound binds to its protein
ACS Paragon Plus Environment

Page 17 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
target, the protein becomes more resistant to heat-induced unfolding and denaturation. This in
turn increases the melting temperature of the protein.
Thermal proteome profiling was used to identify potential targets of 1. GBM43 cells were
treated with 10 µM of 1 at six temperatures between 37 and 60 °C. Following heating, soluble
proteins were extracted in phosphate-buffered saline (PBS) buffer, quantified and digested into
peptides using trypsin. Digested peptides were analyzed by LC-MS/MS and quantified using
label-free precursor-ion (MS1) intensity. The iBAQ values (intensities normalized by the size of
the proteins) acquired after data analysis using MaxQuant(85) were used to fit thermal melting
curves and to determine the shift in melting temperature between untreated and compound-
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
treated proteins (∆T ). We acquired quantitative iBAQ data for over 1700 proteins across all
m
temperatures at 1% FDR, of which the melting curves of 129 proteins were determined from the
thermal profiling (Fig. 5a). A cutoff of 3 °C was used to identify proteins with significant thermal
shifts when treated with 1. A total of 12 proteins were identified with ∆T ≥ 3 °C (Fig. 5b and S3).
m
Interestingly, RACK1 was both predicted to be a target of 1 (Table 1) as well as showed a
significant thermal shift in TPP. In contrast, a set of 7 proteins were identified to be destabilized
when exposed to 1, resulting in a change in melting temperature ∆T ≤ -3 °C (Fig. 5c and S4).
m
The largest observed destabilization was in DSTN, a component of the cytoskeleton.
Integrated Analysis of Computational, RNA-seq, and TPP Data for Potential
Mechanisms of Action. To uncover the potential mechanism of action of 1, proteins predicted to
bind 1 by SVR-KB, and proteins implicated by compound-specific changes by RNA-seq profiling,
L1000 comparisons, and thermal proteome profiling were integrated into a protein network (Fig.
5d). Proteins implicated from each of the four sources were connected based on structural and
non-structural evidence using the STRING database(86). STRING incorporates protein-protein
interactions from both direct physical interactions as well as indirect functional associations.
Interactions come from a variety of sources, including experimental interactions from protein-
protein databases, pathway knowledge from manually curated databases, co-expression studies,
and homology. Multiple interconnected modules are formed in the protein subnetwork. The largest
module features proteins associated with the cell cycle (PPP1CB, PLK1, EXO1, and CDC20) and
metabolism of compounds with nucleobases (KIF11, EXO1, TOP2A, and CENPE). Within this
cluster, several targets were predicted to bind to 1, including EXO1, TOP2A, CENPE, and KIF11.
Three proteins in this cluster showed positive thermal shifts, DCTN1, KIF5B, and PPP1CB, as
well as a negative thermal shift in ARCN1. Both DCTN1 and KIF5B act as motor proteins: DCTN1
is a subunit of dynactin, which binds to dynein and acts as cytoskeletal motors in cellular
ACS Paragon Plus Environment

ACS Chemical Biology
Page 18 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
transport(87), while KIF5B is a motor protein involved in mitosis and meiosis, and acts as a
catalytic subunit of the tumor suppressors NF1 and NF2(88). Interestingly, upregulation of the
dynactin mediator DYNC1I1 is observed in the RNA-seq analysis. A second interconnected
module is formed by the G protein β-subunit and β-subunit-like proteins GNB1 and RACK1,
respectively. The prediction of RACK1 as a direct target of 1 is supported by direct evidence from
the thermal proteome profiling, as well as a similar stability shift in the RNA-binding protein
HDLBP. Similarly, upregulation of the GPCR PTH2R and downregulation of the potassium
channel KCNJ10, which belong to similar regulation pathways as GNB1, was also observed in
the analysis of RNA-seq derived from treated versus non-treated GBM spheroids.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
To confirm direct binding to RACK1, the effects of 1 and the inactive analog 22 were
examined in a concentration-dependent manner with an antibody-based cellular thermal shift
assay (CETSA) (Fig. 5e; top panel). Similar to thermal proteome profiling, GBM43 cells are
treated with varying concentrations of the compound and then heated to 45 °C. Direct binding of
the compound to the protein will result in protein stabilization and an increase in melting
temperature. In GBM43, there is an increase in RACK1 abundance at 50 and 100 µM in cells
treated with 1, suggesting that RACK1 may be among the targets of 1. To rule out non-specific
binding of the compound, the assay was also performed on GAPDH, an enzyme involved in
glycolysis, resulting in no difference in protein abundance with increase concentration of
compound. When GBM43 were treated with the inactive analog 22, the concentration of the
compound did not affect protein abundance compared to the DMSO control (Fig. 5e; bottom
panel).
DISCUSSION
Tumors such as GBM exhibit multiple phenotypes that include uncontrolled growth,
angiogenesis, invasion, and immune evasion. These phenotypes are driven by perturbations in
genes and their protein products working in concert across multiple signaling pathways. The
multiple targets involved in promoting tumor growth and metastasis pose a major challenge for
the development of small-molecule therapeutic agents to treat these tumors. To date, the most
common strategy in cancer drug discovery is to develop small molecules that modulate the
function of a single target. This approach has not yielded therapeutic agents that are efficacious
for complex tumors such as GBM. At the other extreme, phenotypic screening has been used to
uncover novel anti-cancer agents(17). This strategy has led to several approved drugs, including
eribulin in breast cancer, nelarabine in T-cell lympoblastic leukemia and lymphoma, and
trametinib in metastatic melanoma(89). However, the limited diversity of chemical libraries, the
ACS Paragon Plus Environment

Page 19 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
use of immortalized cell lines, and the reliance on two-dimensional cellular assays has mostly led
to cytotoxic cell cycle inhibitors that have not yielded efficacy in patients(90).
A data-driven approach that combines genomic, structural, and protein-protein interaction
data to enrich chemical libraries using computational docking has the potential to overcome the
limitations of phenotypic screening for cancer drug discovery. Here, in a first proof-of-concept
study, we propose a screening strategy that takes integrates vast orthogonal datasets including
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
(i) tumor genomic data from patient-derived GBM samples available at TCGA, (ii) three-
dimensional structures of human proteins that enable the identification of druggable pockets in
proteins implicated in GBM, and (iii) the large number of cellular protein-protein interactions that
have been mapped over the past decade using yeast-two-hybrid and other methods. Specifically,
our approach initially utilizes expression data from TCGA to select all genes that are
overexpressed and mutated in GBM. The PDB is subsequently mined to retrieve available protein
structures that are encoded by these genes. Druggable pockets within these structures are
identified and used for structure-based screening to identify potential small-molecule inhibitors.
Proteins that are known to be involved in protein-protein interactions and possess druggable
pockets are used for structure-based docking of an in-house library of 9,000 compounds. The
resulting protein-compound complexes were ranked to select 50 small molecules that bind to the
highest number of GBM-specific proteins. The resulting compounds were tested in a cell viability
assay utilizing patient-derived GBM cells grown in a physiologically relevant three-dimensional
format. Our strategy has a significant advantage over standard phenotypic screening, namely that
large libraries containing thousands of compounds can be enriched to select a small collection of
candidates that can be tested in more sophisticated assays and multiple phenotypes. Two assays
are used, spheroid growth and tube formation assay in Matrigel. We also strictly use patient-
derived low-passage cell lines to overcome the limitations of existing phenotypic screens that are
done on established cell lines. In addition to using GBM cancer cell lines, we also employ low
passage pancreatic ductal adenocarcinoma cell lines as well as normal non-transformed cells
+
such as CD34 progenitor cells and astrocytes.
Compound collections were initially screened at a single concentration in duplicate using
GBM43 cells grown in 3D spheroids. In the set of compounds that were predicted to maximally
bind to GBM-specific targets, one compound, namely 1 (IPR-2025), showed substantial inhibition
of cell viability. A concentration-dependence study revealed that 1 (IPR-2025) showed high
potency in all three patient-derived GBM cell lines with IC s < 4 µM. The compound was tested
50
for its effect on tube formation in Matrigel, an indication of their potential usefulness in blocking
ACS Paragon Plus Environment

ACS Chemical Biology
Page 20 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
angiogenesis and a key feature of GBM tumors, inhibiting with sub-micromolar IC . The
50
+
compound was tested in both normal non-transformed human CD34 progenitor cells and primary
+
brain astrocytes. Significantly, 1 had no effect on colony formation of CD34 cells or astrocyte cell
viability suggesting that there is a therapeutic window. To put the activity in perspective, standard-
of-care temozolomide inhibits GBM43 growth with an approximate IC of 250 µM, consistent with
50
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
the high therapeutic doses used in GBM.
RNA-seq was initially employed to uncover a potential mechanism of action for 1 in
GBM43. Compound 1 exhibited high selectivity as relatively few genes were affected by this
compound. The L1000 platform was employed to identify gene knockdowns that led to a similar
effect on gene expression. However, L1000 utilizes adherent cell cultures and established cell
lines in cancers other than GBM. In contrast, GBM43 is a low-passage patient-derived cell line
grown as a spheroid model and is expected to have a different underlying gene expression profile.
Interestingly, the expression profile of 1 was most similar to the knockdown of G protein GNA15,
which belongs to the same family as the predicted target GNB1 and structural homolog RACK1.
Considering the GBM-specific activity of 1 (IPR-2025) and its lack of cytotoxicity in normal
non-transformed cell lines, thermal proteome profiling was used to identify potential targets of this
compound. TPP uncovered up to 11 potential targets of 1, including RACK1 and KIF5B. It is
interesting that among them, one target (RACK1) was also among the targets of 1 that emerged
from the structure-based docking and ranking analysis. Analysis of the predicted binding mode
reveals that the compound may bind within the central tunnel of the β-propeller structure. Follow-
up CETSA analysis seems to support the fact that 1 may bind to RACK1 directly in cell culture. It
is important to note that the TPP list is likely not a comprehensive list of all direct targets of 1, and
it is also likely that the 20 top targets identified in TPP may not be all targets of 1. Furthermore,
RACK1 is a scaffold protein that is engaged in multiple protein-protein interactions by creating a
large multi-protein complex. Even if we were able to confirm RACK1 direct binding, it would not
establish that the compound inhibits RACK1 protein-protein interactions. Also, it is worth noting
that the compound was designed to selectively bind to multiple targets. It is possible that the
biological response that is observed as a result of treatment with compound is the result of binding
of the compound to other targets. Further mechanistic studies beyond the scope of this
manuscript will have to be performed to establish the exact mechanism of this compound.
Several derivatives of compound 1 either from commercial sources or from our own
synthesis were tested for inhibition of cancer cell viability. Ultimately, identification of all targets
will be required to better explain the observed SAR. However, structure-activity analyses can be
ACS Paragon Plus Environment

Page 21 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
performed using the chemical structure of compounds and their biological activity. Compounds
2-21 were obtained from commercial sources through a substructure search. They may share
structure with 1, but they have key differences that makes them in active. For example,
compounds 7-18 and 20-21 lack the dibenzofuran group of 1. This group appears to be critical
for biological activity. In fact, slight modifications at this moiety, such as introduction of chlorine
atoms in compounds 26 and 27 led to complete loss of activity. While compounds 2-6 and
compound 19 have the dibenzofuran moiety, they are substantially different in structure and lack
the critical N-acylsulfonamide core. Compounds 22-24 were designed to explore the importance
of the N-acylsulfonamide on compound 1. N-acylsulfonamides are known to react with
nucleophiles such as serine residues. The N-acylsulfonamide moiety of compound 1 could
possibly react with serine, threonine, or tyrosine nucleophiles as illustrated in Fig. S5. Replacing
the N-acylsulfonamide carbonyl with methylene to generate 22 resulted in complete elimination
of compound activity. A similar observation was made when the sulfone was replaced with
methylene in compound 23. We wondered if the product compound (Fig. S5) following
nucleophilic attack of the N-acylsulfonamide is behind the biological activity of compound 1.
Compound 24 showed no activity, suggesting that a combination of binding and covalent
modification is likely behind the activity of compound 1. However, the fact that subtle
modifications to other parts of the compound, such as introduction of chlorine atoms in 26 and
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
27, led to substantial loss in activity, suggests that the biological activity of compound 1 is a
combination of binding and covalent modification.
In summary, we implement a multi-target screening approach that integrates cancer
genomics with the druggable protein interactome to identify therapeutic candidates of GBM. The
strategy of selecting compounds that were predicted to either target the highest number of GBM-
specific proteins yielded a candidate compound. We pursued the compound in RNA-seq to
discover potential mechanisms of action. Thermal proteome profiling of 1 revealed possible
targets, including the computational predicted target RACK1. The ability of 1 to selectively target
GBM phenotypes without affecting normal cell viability makes it suitable as a lead compound to
uncover new targets in GBM and to develop potential therapeutic agents with multi-targeting
properties.
MATERIALS AND METHODS
Cell Culture. The SJ-GBM2 cell line was obtained from the Children’s Oncology Group.
The GBM10 and GBM43 xenografts were a kind gift from Dr. Jann Sarkaria (Mayo Clinic), and
tumors were expanded by passage in the flank of NOD/SCID γnull mice. To generate GBM10 and
ACS Paragon Plus Environment

ACS Chemical Biology
Page 22 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
GBM43 cell lines, tumors were harvested, disaggregated, and maintained in 2.5% FBS for 14
days on Matrigel-coated plates (BD Biosciences) to remove murine fibroblasts as previously
described (50). Cell lines were propagated in DMEM with 10% FBS medium with glutamine
(Cellgro, Manassas, VA) for no more than 7 passages in 5% CO at 37 °C. Cell line identity was
2
confirmed by DNA fingerprint analysis (IDEXX BioResearch) for species and baseline short-
tandem repeat analysis testing. The pancreatic cancer cells used in these studies were low-
passage patient-derived PDAC cell lines (fewer than 35 passages from patient tumors) and that
were received from Dr. Anirban Maitra (The Johns Hopkins University) and maintained as
previously described (55,91). Pa02C, Panc10.05, and Panc198 were cultured in DMEM medium
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
and supplemented with 10% FBS in 5% CO at 37 °C for fewer than 10 passages. Cell line identity
2
was confirmed by DNA fingerprint analysis (IDEXX BioResearch) for species and baseline short-
tandem repeat analysis testing and cells were routinely checked for mycoplasma.
GBM and PDAC 3D models. GBM10, GBM43, and SJ-GBM2 spheroids were generated
4
by plating early-passage cells at 2.5 × 10 cells per well in 96-well ultralow attachment plates
(Corning Inc.) in DMEM/F12 (1:1; GIBCO) supplemented with 2% B27 supplement (GIBCO), 20
ng/mL epidermal growth factor (EGF), and 20 ng/mL fibroblast growth factor (FGF) (Peprotech)
for 2 days. The spheroids were then treated with compounds and growth analyzed by Alamar
blue staining at day 5 following compound exposure as previously described by our group (52).
Astrocyte cell proliferation was determined by The CellTiter 96® AQueous Non-Radioactive Cell
Proliferation Assay (MTS) (Promega) performed in 96-well plates. Primary human fetal astrocytes
were purchased from ABM (cat #T4033) and cultured in progenitor astrocyte growth media
3
defined by Barris and co-workers (92). Cells were seeded at 8 × 10 /well in DMEM medium with
0% FBS and 1% 100X Penicillin/Streptomycin. Cell numbers were determined after 3 days of
1
incubation with DMSO or compounds at indicated concentrations. 20 μL of MTS solution was
added to the well. After 1 hour incubation at 37 °C, the absorbance was measured at 490 nm.
Experiments were done in triplicate.
PDAC cells were grown as tumor spheroids as described previously using pancreatic
cancer cells stably transduced with TdTomato (red) and the intensity of the spheroids over time
was quantitated as described in our previous studies (56,91,94). Briefly, PDAC cells (1,000
cell/well) were seeded in ultralow adherence 96-well plates (Corning Inc.) in media containing 3%
reduced growth factor Matrigel (Corning Inc.) and 5% FBS. Spheroids were fed or treated on days
4
and 8. Red fluorescence intensity was measured at day 12 following plating using the Thermo
ACS Paragon Plus Environment

Page 23 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
ArrayScan (Thermo Fisher Scientific). Percent viability for drug-treated wells was quantitated
compared to media control.
Half maximal inhibitory concentration (IC ) of cell viability curves were determined using
50
a four-parameter logistic model (93). IC50 values represent relative IC s, that is, the minimum
50
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
parameter of the model is the lowest cell viability observed. Cell viability curves were determined
using the non-linear least-squares method as implemented in the SciPy package in Python.
Colony-forming Unit (CFU) Assay. Briefly, purified human umbilical cord blood CD34+
cells were obtained from the Angio BioCore (IUSCC) and the IC50 values determined by dose
+
response curves. CD34 cells were incubated with compounds and plated in the presence of the
3
compounds at 2 × 10 cells per 35 mm dish in MethoCult H4434 (Stemcell Technologies) as
previously described (95). The number of clonogenic progenitor colonies was enumerated under
a light microscope at day 14 post-compound exposure.
TCGA GBM Gene Expression and Somatic Mutations. RNA-sequencing and mutation
data from GBM were identified from The Cancer Genome Atlas (TCGA) project(12,96) and
collected from the National Cancer Institute’s Genomic Data Commons (GDC) data portal(97). As
part of the GDC pipeline, 169 tumors and 5 normal samples were previously aligned against the
GRCh38 genome assembly using STAR 2-pass (98) and quantified using HTSeq (99). Level 3
HTSeq fragment counts were collected from all 174 samples and used these as read counts for
differential expression analysis. Differentially expressed genes were identified using the
edgeR(100) package and the quasi-likelihood F-test pipeline (101) for hypothesis testing. A
counts-per-million (CPM) threshold corresponding to a library count of 10 (CPM ≈ 0.15) was used
to filter out genes with low reads. Overexpressed genes were defined as those with p < 0.001,
FDR < 0.01, and log fold change (log FC) > 1. Ensembl Gene identifiers from the differential
2
2
expression analysis were mapped to UniProt(102) identifiers using BioMart(103). Open-access
mutation annotation format (MAF) files for GBM were also collected from the GDC data portal.
The GDC pipeline used the MuTect2(104) pipeline for mutation calling. As part of the GDC’s
workflow, low quality and potentially germline variants were removed. Somatic mutations were
mapped to their corresponding proteins using the UniProt ID.
Protein-Protein Interaction Network. Rolland and associates (46) recently constructed
two large-scale protein-protein interaction networks. The first is based on literature curation and
benchmarking of seven protein-protein interaction databases, producing a network of
approximately 5500 proteins and 12000 interactions. The second is based on systematic
ACS Paragon Plus Environment

ACS Chemical Biology
Page 24 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
experimental testing of pairwise combinations of approximately 13000 genes and describes a
network of approximately 4200 proteins and 14000 interactions. Entrez genes were mapped to
UniProt identifiers using UniProt’s mapping tool (http://www.uniprot.org/mapping/) and these two
networks were combined together using the NetworkX (105) module in Python. The resulting
protein-protein interaction network consists of approximately 27000 interactions across 8000
proteins and was visualized using Cytoscape (106) (v3.3.0).
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Druggable Binding Sites. In a previous work, a set of druggable binding pockets were
identified in the human proteome and explored in the context of functional importance, signaling
pathways, and the human interactome(45). In brief, a set of 4124 proteins were identified in the
human proteome that have been solved with crystallography. Druggable binding pockets were
found on these proteins using the SiteMap(107,108) module in Schrödinger (Schrödinger LLC,
New York, NY). Up to 10 binding sites were mapped for each structure. Each binding site was
evaluated by its ability to bind a ligand (SiteScore) and its druggability (DrugScore). Only binding
sites with SiteScore and DrugScore above 0.8 were kept. In total, 5498 binding sites on 2607
proteins were found. Each binding site identified by SiteMap was visually inspected and manually
annotated to determine its functional role in the protein. If an active site residue was in contact
with the SiteMap spheres, or if a catalytic molecule or inhibitor occupied the space of the spheres,
the binding site was labeled ‘enzyme’ (ENZ). If the binding site was at a protein-protein interaction
(PPI) interface on the original structure or on any of the aligned structures, the binding site was
labeled ‘PPI’. Otherwise, if the binding site was neither enzyme nor part of the interaction
interface, it was labeled ‘other’ (OTH).
Virtual Screening GBM Network. The set of druggable binding sites on proteins which
were (i) overexpressed in GBM, (ii) featured somatic mutations in GBM, and (iii) were part of the
PPI interactome were selected. Specifically, we focused on binding sites that were classified as
either PPI or OTH and omitted those with solely ENZ classifications. In total, 316 binding sites
were identified. The set of binding pockets were docked against an internal chemical library of
small molecules. These compounds are from previous screening campaigns primarily focused on
targeting tight protein-protein interactions. Compounds possessing pan-assay interference
compound (PAINS)(109) or rapid elimination of swill (REOS)(110) moieties were discarded. This
resulted in a compound library of 9075 small molecules and their enantiomers. Compounds were
docked against 316 binding sites on proteins implicated in GBM using AutoDock Vina(111)
(v1.1.2). The average coordinates of the SiteMap spheres were used to identify the centroid of
the binding site for docking. Each binding site was represented as a box with dimensions of 21 Å
ACS Paragon Plus Environment

Page 25 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
×
21 Å × 21 Å. All other parameters were set to default values. Docked poses were rescored using
the previously developed Support Vector Machine-Knowledge Based (SVR-KB)(47) scoring
function. In this scoring function, knowledge-based pair-potentials of co-crystal complexes are
used to train a regression model to predict the binding affinity of docked complexes. To rank-order
compounds, the number of pockets with binding affinities better than a given SVR-KB cutoff for
that compound were counted. Here, an SVR-KB cutoff corresponding to a predicted binding
affinity of 10 nM was used. Compounds were rank-ordered using the number of binding sites
predicted to bind to the compound with an affinity better than the SVR-KB cutoff. The top 154
compounds were retrieved and clustered using the Canvas module in Schrödinger. A hashed
binary fingerprint corresponding to atom triplets of Daylight invariant atom types were generated
for the selected compounds. Compounds were hierarchically clustered using their atom triplet
fingerprints and average linkage clustering to 50 clusters. The Tanimoto similarity between a pair
of fingerprints was used as the distance metric. Compounds corresponding to the centroid of each
cluster were selected for experimental validation.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Tube Formation Assay. The Matrigel based tube formation assay was performed as
previously described (112). Briefly, 50 µL Matrigel (Corning, Corning, NY) was allowed to solidify
in a 96 well black, clear bottom plate at 37 °C for 20 min. Primary brain microvascular endothelial
cells (BMECs; Cell Systems, Seattle, WA) were added to the solid Matrigel at 15,000 cells per
well in 100 µL endothelial growth medium-2 (EGM-2, Lonza, Walkersville, MD) and dosed with
appropriate concentrations of compound with 1 µL DMSO per well. Tube formation was observed
every 2 hours by brightfield microscopy and images were taken after 8 hours of tube formation.
Six images per treatment were analyzed with the Angiogenesis Analyzer plugin for ImageJ(113),
and BMEC total tubule length for treated cells was normalized to DMSO-treated samples.
Statistical analysis using one-way ANOVA with Dunnett’s post hoc test to compare treatments
with DMSO control was completed using GraphPad Prism (v7.0, GraphPad Software, La Jolla,
CA).
RNA-Seq of Compound-Treated Cells. GBM43 cells were treated with 10 µM 1 (IPR-
2025). RNA-seq analysis was conducted in triplicates for the untreated control and 1. After
compound treatment, GBM43 cells were collected and rinsed with 1× PBS. The pellet was first
homogenized in RLT lysis buffer plus β-mercaptoethanol and total RNA was extracted with
RNeasy Mini Kit (Qiagen) in combination with QIAshredder (Qiagen). The RNA quality was
assessed by A260/A280 ratio (NanoDrop) and stored at -80 °C.
ACS Paragon Plus Environment

ACS Chemical Biology
Page 26 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
Total RNA was evaluated for its quantity and quality using the Agilent Bioanalyzer 2100
system. For RNA quality, a RIN number of 7 or higher was desired. A total of 500 ng RNA was
used. The cDNA library was prepared through mRNA purification and enrichment, RNA
fragmentation, cDNA synthesis, ligation of index adaptors, and amplification, following the TruSeq
Stranded mRNA Sample Preparation Guide, RS-122-9004DOC, Part# 15031047 Rev. E
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
(Illumina, Inc.). Each resulting indexed library was quantified, and its quality accessed by Qubit
and Agilent Bioanalyzer, and multiple libraries pooled in equal molarity. Five µL of 2 nM pooled
libraries per lane was then denatured, neutralized, and applied to the cBot for flow cell deposition
and cluster amplification, before loading to HiSeq 4000 for sequencing (Illumina, Inc.).
The RNA-seq aligner from STAR(98) (v2.5) was used to map RNA-seq reads to the human
reference genome (hg38), with the following parameter: ‘--outSAMmapqUnique 60’. Uniquely
mapped sequencing reads were assigned to genes using featureCounts(114) (from subread
v1.5.1) with the following parameters: ‘-s 2 -Q 10’. The genes were filtered for further analysis if
their count per million (CPM) of reads was less than 0.5 in more than 4 samples. The method of
trimmed mean of M values (TMM) was adopted for gene expression normalization cross all
samples, followed by differential expression (DE) analysis between different conditions using
edgeR(100,115) (v3.20.8).
Differentially expressed genes (DEG) were determined using p-value and false discovery
rate (FDR) cutoffs of 0.001 and 0.01, respectively. A fold change (|log FC|) cutoff of 1 was used.
2
The functional analysis was performed on overexpressed and underexpressed DEGs separately
with a cutoff of FDR < 0.05 to identify significantly overrepresented Gene Ontology (GO) and/or
KEGG pathways using DAVID(116,117) and PANTHER(118).
Thermal Proteome Profiling. GBM43 cells were trypsinized, washed with PBS, and
suspended in PBS supplemented with a protease inhibitor cocktail (Sigma‐Aldrich). Suspended
cells were treated with 10 µM 1 (IPR-2025) or DMSO for 1 hour at 37 °C, then the treated cells
6
were separated into fractions, each with 5 × 10 cells for proteome thermal profiling. Fractions
were heated for 3 min to the following temperatures: (i) 37 °C, (ii) 40 °C, (iii) 45 °C, (iv) 50 °C, (v)
55 °C, and (vi) 60 °C, then incubated for 3 min at room temperature and frozen in liquid nitrogen.
Samples were lysed with three freeze-thaw cycles using liquid nitrogen and 25 °C water bath.
Cell lysates were centrifuged at 100,000 × g for 20 min at 4 °C to separate protein aggregates
from soluble proteins. Supernatants were collected for LC-MS/MS.
ACS Paragon Plus Environment

Page 27 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
Protein concentration in each sample was determined by bicinchoninic acid (BCA) assay
with BSA as a standard. About 50 µg protein from each sample was denatured by adding 50 µL
of 8M urea, reduced by incubating with 10 mM dithiothreitol (DTT) at 50 °C for 45ꢀmin, and
cysteine alkylated with 20ꢀmM iodoacetamide (IAA) in the dark for 45 min at room temperature,
followed by incubation with 5 mM DTT for 20 min at 37 °C to scavenge residual IAA. Proteins
were digested using sequencing grade trypsin and Lys-C mix from Promega at a 1:25 (w/w)
enzyme-to-protein ratio at 37 °C overnight. The digested peptides were cleaned using C18 silica
micro spin columns (The Nest Group Inc.) and peptides were eluted using 80% acetonitrile
containing 0.1% formic acid (FA). The samples were vacuum dried and re-suspended in 3%
acetonitrile and 0.1% formic acid. The peptide concentration was determined by BCA assay with
BSA as a standard. Peptides concentration was adjusted to 0.2 µg/µL, soluble and insoluble
samples were mixed together and 5 µL was used for LC-MS/MS analysis as described below.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Samples were analyzed by reverse-phase LC-ESI-MS/MS system using the Dionex
UltiMate 3000 RSLCnano System coupled to the Q Exactive High Field (HF) Hybrid Quadrupole-
Orbitrap MS and a Nano-electrospray Flex ion source (Thermo Fisher Scientific). Peptides were
loaded onto a trap column (300 µm ID  5 mm) packed with 5 µm 100 Å PepMap C18 medium,
and then separated on a reverse phase column (15 cm long × 75 µm ID) packed with 3 µm 100
Å PepMap C18 silica (Thermo Fisher Scientific). All the MS measurements were performed in the
positive ion mode, using 120 min LC gradient as previously described(119). The mass
spectrometer was operated using standard data-dependent mode. MS data were acquired with a
Top20 data-dependent MS/MS scan method. The full scan MS spectra were collected in the 400-
1,600 m/z range with a maximum injection time of 100 milliseconds, a resolution of 120,000 at
6
2
00 m/z, spray voltage of 2 and AGC target of 3 × 10 . Fragmentation of precursor ions was
performed by high-energy C-trap dissociation (HCD) with the normalized collision energy of 27
eV. MS/MS scans were acquired at a resolution of 15,000 at m/z 200 with an ion-target value of
5
1
× 10 and a maximum injection of 20 milliseconds. The dynamic exclusion was set at 15 s to
avoid repeated scanning of identical peptides. Instrument was calibrated at the start of each batch
run and then in every 72 hours using calibration mix solution (Thermo Scientific). The performance
of the instrument was also evaluated routinely using complex E. coli digest purchased from
Sigma.
Raw LC-MS/MS data were analyzed using MaxQuant(85) (v1.6.0.16) against the UniProt
human protein database containing 40,707 proteins. The database search was performed with
the precursor mass tolerance set to 10 ppm and MS/MS fragment ions tolerance was set to 20
ACS Paragon Plus Environment

ACS Chemical Biology
Page 28 of 59
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
ppm. Database search was performed with enzyme specificity for trypsin and LysC, allowing up
to two missed cleavages. Oxidation of methionine and N-terminal acetylation were defined as a
variable modification, and carbamidomethylation of cysteine was defined as a fixed modification
for database searches. The ‘unique plus razor peptides’ were used for peptide quantitation. The
false discovery rate (FDR) of both peptides and proteins identification was set at 0.01. Data were
searched with ‘match between runs’ option. In the case of identified peptides that are shared
between two proteins, these were combined and reported as one protein group. Proteins
matching to the reverse database were filtered out.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Protein abundances were quantified using iBAQ(120) and transformed into relative
concentrations relative to the lowest temperature (37 °C). Protein fold changes were normalized
and melting curves were fitted following the protocol described by Savitski(83), Franken(84), and
co-workers. The previously described normalization protocol(83) was adapted to fit the range of
temperatures in our experiments. Melting curves were selected from both the vehicle and
treatment groups where the fold change at 55 °C was between 0.4 and 0.6 and the fold change
at 60 °C was less than 0.4. Normalization was then applied as previously described(83). Following
normalization, melting curves were fitted into the sigmoidal function(84):
1
― 푝
푓(푇) =
+ 푝
푎
―
(푇 ― 푏)
1
+ 푒
where 푇 is the temperature, 푎 and 푏 are constants, and 푝 is the plateau as 푇→∞. The
melting point 푇 is found using:
푚
푎
푇 =
푚
1
― 푦
푏 ― ln
(푦 ― 푝
)
at 푓
푇푚 = 푦 = 0.5. The slope of the melting curve 푠푙표푝푒 = 푓′(푇푖푛푓푙푒푐푡푖표푛) is found at the
( )
′
′
(
)
= 0. Melting curves were determined using the non-
푖푛푓푙푒푐푡푖표푛
inflection point 푇푖푛푓푙푒푐푡푖표푛 where 푓 푇
linear least-squares method as implemented in the SciPy package in Python. Proteins with poorly
fitted curves were excluded if the parameters of the curve fit any of the following criteria: (i) vehicle
2
or treatment curve with 푅 < 0.8, (ii) vehicle 푝푙푎푡푒푎푢 > 0.3, (iii) the number of points between
(푝푙푎푡푒푎푢 + 0.1, 0.9) < 1, and (iv) vehicle or treatment 푠푙표푝푒 > ―0.06. In total, 129 proteins were
identified that passed all quality control filters. The shift in melting point Δ푇 was determined as:
푚
treatment
vehicle
Δ푇 = 푇
― 푇
푚
푚
푚
ACS Paragon Plus Environment

Page 29 of 59
ACS Chemical Biology
1
2
3
4
5
6
7
8
9
A cutoff of Δ푇 ≥ 3.0 was used to identify proteins with significant thermal stability following
푚
compound treatment.
Gene Set Analysis. The STRING (Search Tool for the Retrieval of Interacting
Genes/Proteins) database(86) incorporates protein-protein interactions from both direct physical
interactions as well as indirect functional associations. STRING was used to integrate gene sets
identified from protein target prediction using SVR-KB, differentially expressed genes identified
from RNA sequencing, most similar gene signatures from L1000, and genes identified by thermal
proteome profiling. A concatenated list of all HUGO gene symbols was used as input.
Connections coming from text mining sources were excluded. Edges with confidence score < 0.7
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
(
high confidence) and genes with no edges in the subnetwork were excluded. The resulting
subnetworks were visualized using Cytoscape(106).
General Chemistry. Preparative thin layer chromatography (PTLC) was performed on 20
×
20 cm plates (500-micron thick silica gel). Silica gel chromatography was performed on a
1
Biotage Horizon flash chromatography system. H NMR spectra were recorded on a Bruker
Ascend TM 400 spectrometer at 400 MHz at 298 K, and the chemical shifts are given in parts per
million (ppm) referenced to the residual proton signal of the deuterated solvents: CHCl at δ =
3
7.26 ppm and CH OH or CH OD at δ = 3.30 ppm. LC-MS spectra were taken on an Agilent
3
3
Technologies 1260 Infinity or 6120 Quadrupole spectrometer. The mobile phase for the LC was
acetontrile (A) and water (B) with 0.01% formic acid, and the eluent gradient was from 5-95% A
in 6.0 min, 60-95% A in 5.0 min, 80-100% A in 5.0 min and 85-100% A in 10 min using a SBC18
5
0 mm × 4.6 mm × 2.7 μm capillary column. Mass spectra (MS) were measured by electrospray
ion-mass spectroscopy (ESI). All temperatures are in °C unless otherwise noted.
O
O
S
NH2
O
O
N
O
O
O
S
N
O
S
N
H
O
O
O
O
TFA
O
NaH, THF
O
N
O
O
O
O
NH
O
POCl , pyridine
3
HO
O
1a
1b
1
tert-butyl 5-oxo-1-tosylpyrrolidine-2-carboxylate (1a). To a solution of tert-butyl 5-
oxopyrrolidine-2-carboxylate (530 mg, 2.86 mmol, 1.0 eq) in anhydrous THF (10 mL) at 0 °C, NaH
(
343 mg, 8.58 mmol, 3.0 eq) was added. The mixture was stirred for 15 min, and 4-
ACS Paragon Plus Environment