RECOVER identifies synergistic drug combinations in vitro through sequential model optimization.

For large libraries of small molecules, exhaustive combinatorial chemical screens become infeasible to perform when considering a range of disease models, assay conditions, and dose ranges. Deep learning models have achieved state-of-the-art results in silico for the prediction of synergy scores. However, databases of drug combinations are biased toward synergistic agents and results do not generalize out of distribution. During 5 rounds of experimentation, we employ sequential model optimization with a deep learning model to select drug combinations increasingly enriched for synergism and active against a cancer cell line-evaluating only ∼5% of the total search space. Moreover, we find that learned drug embeddings (using structural information) begin to reflect biological mechanisms. In silico benchmarking suggests search queries are ∼5-10× enriched for highly synergistic drug combinations by using sequential rounds of evaluation when compared with random selection or ∼3× when using a pretrained model.

A screen for MeCP2-TBL1 interaction inhibitors using a luminescence-based assay.

Understanding the molecular pathology of neurodevelopmental disorders should aid the development of therapies for these conditions. In MeCP2 duplication syndrome (MDS)-a severe autism spectrum disorder-neuronal dysfunction is caused by increased levels of MeCP2. MeCP2 is a nuclear protein that binds to methylated DNA and recruits the nuclear co-repressor (NCoR) complex to chromatin via an interaction with the WD repeat-containing proteins TBL1 and TBLR1. The peptide motif in MeCP2 that binds to TBL1/TBLR1 is essential for the toxicity of excess MeCP2 in animal models of MDS, suggesting that small molecules capable of disrupting this interaction might be useful therapeutically. To facilitate the search for such compounds, we devised a simple and scalable NanoLuc luciferase complementation assay for measuring the interaction of MeCP2 with TBL1/TBLR1. The assay allowed excellent separation between positive and negative controls, and had low signal variance (Z-factor = 0.85). We interrogated compound libraries using this assay in combination with a counter-screen based on luciferase complementation by the two subunits of protein kinase A (PKA). Using this dual screening approach, we identified candidate inhibitors of the interaction between MeCP2 and TBL1/TBLR1. This work demonstrates the feasibility of future screens of large compound collections, which we anticipate will enable the development of small molecule therapeutics to ameliorate MDS.

Identification and optimization of molecular glue compounds that inhibit a noncovalent E2 enzyme-ubiquitin complex.

Pharmacological control of the ubiquitin-proteasome system (UPS) is of intense interest in drug discovery. Here, we report the development of chemical inhibitors of the ubiquitin-conjugating (E2) enzyme CDC34A (also known as UBE2R1), which donates activated ubiquitin to the cullin-RING ligase (CRL) family of ubiquitin ligase (E3) enzymes. A FRET-based interaction assay was used to screen for novel compounds that stabilize the noncovalent complex between CDC34A and ubiquitin, and thereby inhibit the CDC34A catalytic cycle. An isonipecotamide hit compound was elaborated into analogs with ~1000-fold increased potency in stabilizing the CDC34A-ubiquitin complex. These analogs specifically inhibited CDC34A-dependent ubiquitination in vitro and stabilized an E2~ubiquitin thioester reaction intermediate in cells. The x-ray crystal structure of a CDC34A-ubiquitin-inhibitor complex uncovered the basis for analog structure-activity relationships. The development of chemical stabilizers of the CDC34A-ubiquitin complex illustrates a general strategy for de novo discovery of molecular glue compounds that stabilize weak protein interactions.

Crippling life support for SARS-CoV-2 and other viruses through synthetic lethality.

With the rapid global spread of SARS-CoV-2, we have become acutely aware of the inadequacies of our ability to respond to viral epidemics. Although disrupting the viral life cycle is critical for limiting viral spread and disease, it has proven challenging to develop targeted and selective therapeutics. Synthetic lethality offers a promising but largely unexploited strategy against infectious viral disease; as viruses infect cells, they abnormally alter the cell state, unwittingly exposing new vulnerabilities in the infected cell. Therefore, we propose that effective therapies can be developed to selectively target the virally reconfigured host cell networks that accompany altered cellular states to cripple the host cell that has been converted into a virus factory, thus disrupting the viral life cycle.

Imipridone Anticancer Compounds Ectopically Activate the ClpP Protease and Represent a New Scaffold for Antibiotic Development.

Systematic genetic interaction profiles can reveal the mechanisms-of-action of bioactive compounds. The imipridone ONC201, which is currently in cancer clinical trials, has been ascribed a variety of different targets. To investigate the genetic dependencies of imipridone action, we screened a genome-wide clustered regularly interspaced short palindromic repeats (CRISPR) knockout library in the presence of either ONC201 or its more potent analog ONC212. Loss of the mitochondrial matrix protease CLPP or the mitochondrial intermediate peptidase MIPEP conferred strong resistance to both compounds. Biochemical and surrogate genetic assays showed that impridones directly activate CLPP and that MIPEP is necessary for proteolytic maturation of CLPP into a catalytically competent form. Quantitative proteomic analysis of cells treated with ONC212 revealed degradation of many mitochondrial as well as nonmitochondrial proteins. Prompted by the conservation of ClpP from bacteria to humans, we found that the imipridones also activate ClpP from Escherichia coli, Bacillus subtilis, and Staphylococcus aureus in biochemical and genetic assays. ONC212 and acyldepsipeptide-4 (ADEP4), a known activator of bacterial ClpP, caused similar proteome-wide degradation profiles in S. aureus ONC212 suppressed the proliferation of a number of Gram-positive (S. aureus, B. subtilis, and Enterococcus faecium) and Gram-negative species (E. coli and Neisseria gonorrhoeae). Moreover, ONC212 enhanced the ability of rifampin to eradicate antibiotic-tolerant S. aureus persister cells. These results reveal the genetic dependencies of imipridone action in human cells and identify the imipridone scaffold as a new entry point for antibiotic development.

At Long Last, a C-Terminal Bookend for the Ubiquitin Code.

The ubiquitin-proteasome system controls the stability of myriad protein substrates via short sequence motifs called degrons. Studies by Koren et al. (2018) and Lin et al. (2018) have uncovered a broad new class of degrons located at the extreme C terminus of substrates.

Generation of rationally-designed nerve growth factor (NGF) variants with receptor specificity.

Nerve growth factor (NGF) is the prototypic member of the neurotrophin family and binds two receptors, TrkA and the 75 kDa neurotrophin receptor (p75NTR), through which diverse and sometimes opposing effects are mediated. Using the FoldX protein design algorithm, we generated eight NGF variants with different point mutations predicted to have altered binding to TrkA or p75NTR. Of these, the I31R NGF variant exhibited specific binding to p75NTR. The generation of this NGF variant with selective affinity for p75NTR can be used to enhance understanding of neurotrophin receptor imbalance in diseases and identifies a key targetable residue for the development of small molecules to disrupt binding of NGF to TrkA with potential uses in chronic pain.

Evaluating CHARGE syndrome in congenital hypogonadotropic hypogonadism patients harboring CHD7 variants.

PURPOSE: Congenital hypogonadotropic hypogonadism (CHH), a rare genetic disease caused by gonadotropin-releasing hormone deficiency, can also be part of complex syndromes (e.g., CHARGE syndrome). CHD7 mutations were reported in 60% of patients with CHARGE syndrome, and in 6% of CHH patients. However, the definition of CHD7 mutations was variable, and the associated CHARGE signs in CHH were not systematically examined. METHODS: Rare sequencing variants (RSVs) in CHD7 were identified through exome sequencing in 116 CHH probands, and were interpreted according to American College of Medical Genetics and Genomics guidelines. Detailed phenotyping was performed in CHH probands who were positive for CHD7 RSVs, and genotype-phenotype correlations were evaluated. RESULTS: Of the CHH probands, 16% (18/116) were found to harbor heterozygous CHD7 RSVs, and detailed phenotyping was performed in 17 of them. Of CHH patients with pathogenic or likely pathogenic CHD7 variants, 80% (4/5) were found to exhibit multiple CHARGE features, and 3 of these patients were reclassified as having CHARGE syndrome. In contrast, only 8% (1/12) of CHH patients with nonpathogenic CHD7 variants exhibited multiple CHARGE features (P = 0.01). CONCLUSION: Pathogenic or likely pathogenic CHD7 variants rarely cause isolated CHH. Therefore a detailed clinical investigation is indicated to clarify the diagnosis (CHH versus CHARGE) and to optimize clinical management.

Synthetic Genomics: Rewriting the Genome Chromosome by Chromosome.

The Synthetic Yeast Genome Project (Sc2.0) consortium has recently published a collection of seven reports on the design, assembly, and in vivo characterization of five entirely synthetic yeast chromosomes that set the stage for de novo construction of synthetic custom-engineered eukaryotes.

Engineering of weak helper interactions for high-efficiency FRET probes.

Fluorescence resonance energy transfer (FRET)-based detection of protein interactions is limited by the very narrow range of FRET-permitting distances. We show two different strategies for the rational design of weak helper interactions that co-recruit donor and acceptor fluorophores for a more robust detection of bimolecular FRET: (i) in silico design of electrostatically driven encounter complexes and (ii) fusion of tunable domain-peptide interaction modules based on WW or SH3 domains. We tested each strategy for optimization of FRET between (m)Citrine and mCherry, which do not natively interact. Both approaches yielded comparable and large increases in FRET efficiencies with little or no background. Helper-interaction modules can be fused to any pair of fluorescent proteins and could, we found, enhance FRET between mTFP1 and mCherry as well as between mTurquoise2 and mCitrine. We applied enhanced helper-interaction FRET (hiFRET) probes to study the binding between full-length H-Ras and Raf1 as well as the drug-induced interaction between Raf1 and B-Raf.

The design and characterization of receptor-selective APRIL variants.

A proliferation-inducing ligand (APRIL), a member of the TNF ligand superfamily with an important role in humoral immunity, is also implicated in several cancers as a prosurvival factor. APRIL binds two different TNF receptors, B cell maturation antigen (BCMA) and transmembrane activator and cylclophilin ligand interactor (TACI), and also interacts independently with heparan sulfate proteoglycans. Because APRIL shares binding of the TNF receptors with B cell activation factor, separating the precise signaling pathways activated by either ligand in a given context has proven quite difficult. In this study, we have used the protein design algorithm FoldX to successfully generate a BCMA-specific variant of APRIL, APRIL-R206E, and two TACI-selective variants, D132F and D132Y. These APRIL variants show selective activity toward their receptors in several in vitro assays. Moreover, we have used these ligands to show that BCMA and TACI have a distinct role in APRIL-induced B cell stimulation. We conclude that these ligands are useful tools for studying APRIL biology in the context of individual receptor activation.

An improved understanding of TNFL/TNFR interactions using structure-based classifications.

Tumor Necrosis Factor Ligand (TNFL)-Tumor Necrosis Factor Receptor (TNFR) interactions control key cellular processes; however, the molecular basis of the specificity of these interactions remains poorly understood. Using the T-RMSD (tree based on root mean square deviation), a newly developed structure-based sequence clustering method, we have re-analyzed the available structural data to re-interpret the interactions between TNFLs and TNFRs. This improves the classification of both TNFLs and TNFRs, such that the new groups defined here are in much stronger agreement with structural and functional features than existing schemes. Our clustering approach also identifies traces of a convergent evolutionary process for TNFLs and TNFRs, leading us to propose the co-evolution of TNFLs and the third cysteine rich domain (CRD) of large TNFRs.

A novel classification system to predict the pathogenic effects of CHD7 missense variants in CHARGE syndrome.

CHARGE syndrome is characterized by the variable occurrence of multisensory impairment, congenital anomalies, and developmental delay, and is caused by heterozygous mutations in the CHD7 gene. Correct interpretation of CHD7 variants is essential for genetic counseling. This is particularly difficult for missense variants because most variants in the CHD7 gene are private and a functional assay is not yet available. We have therefore developed a novel classification system to predict the pathogenic effects of CHD7 missense variants that can be used in a diagnostic setting. Our classification system combines the results from two computational algorithms (PolyPhen-2 and Align-GVGD) and the prediction of a newly developed structural model of the chromo- and helicase domains of CHD7 with segregation and phenotypic data. The combination of different variables will lead to a more confident prediction of pathogenicity than was previously possible. We have used our system to classify 145 CHD7 missense variants. Our data show that pathogenic missense mutations are mainly present in the middle of the CHD7 gene, whereas benign variants are mainly clustered in the 5' and 3' regions. Finally, we show that CHD7 missense mutations are, in general, associated with a milder phenotype than truncating mutations.

Kinetics in signal transduction pathways involving promiscuous oligomerizing receptors can be determined by receptor specificity: apoptosis induction by TRAIL.

Here we show by computer modeling that kinetics and outcome of signal transduction in case of hetero-oligomerizing receptors of a promiscuous ligand largely depend on the relative amounts of its receptors. Promiscuous ligands can trigger the formation of nonproductive receptor complexes, which slows down the formation of active receptor complexes and thus can block signal transduction. Our model predicts that increasing the receptor specificity of the ligand without changing its binding parameters should result in faster receptor activation and enhanced signaling. We experimentally validated this hypothesis using the cytokine tumor necrosis factor-related apoptosis-inducing ligand (TRAIL) and its four membrane-bound receptors as an example. Bypassing ligand-induced receptor hetero-oligomerization by receptor-selective TRAIL variants enhanced the kinetics of receptor activation and augmented apoptosis. Our results suggest that control of signaling pathways by promiscuous ligands could result in apparent slow biological kinetics and blocking signal transmission. By modulating the relative amount of the different receptors for the ligand, signaling processes like apoptosis can be accelerated or decelerated and even inhibited. It also implies that more effective treatments using protein therapeutics could be achieved simply by altering specificity.

Protein design with fragment databases.

Structure-based computational methods are popular tools for designing proteins and interactions between proteins because they provide the necessary insight and details required for rational engineering. Here, we first argue that large-scale databases of fragments contain a discrete but complete set of building blocks that can be used to design structures. We show that these structural alphabets can be saturated to provide conformational ensembles that sample the native structure space around energetic minima. Second, we show that catalogs of interaction patterns hold the key to overcome the lack of scaffolds when computationally designing protein interactions. Finally, we illustrate the power of database-driven computational protein design methods by recent successful applications and discuss what challenges remain to push this field forward.

Computational design of peptide ligands.

Peptides possess several attractive features when compared to small molecule and protein therapeutics, such as high structural compatibility with target proteins, the ability to disrupt protein-protein interfaces, and small size. Efficient design of high-affinity peptide ligands via rational methods has been a major obstacle to the development of this potential drug class. However, structural insights into the architecture of protein-peptide interfaces have recently culminated in several computational approaches for the rational design of peptides that target proteins. These methods provide a valuable alternative to experimental high-resolution structures of target protein-peptide complexes, bringing closer the dream of in silico designed peptides for therapeutic applications.

Targeting AML through DR4 with a novel variant of rhTRAIL.

Despite progress in the treatment of acute myelogenous leukaemia (AML) the outcome often remains poor. Tumour necrosis factor related apoptosis-inducing ligand (TRAIL) is a promising therapeutic agent in many different types of tumours, but AML cells are relatively insensitive to TRAIL-induced apoptosis. Here we show that TRAIL-induced apoptosis in AML cells is predominantly mediated by death receptor 4 (DR4) and not DR5. Therefore, we constructed a variant of TRAIL (rhTRAIL-C3) that is a strong inducer of DR4-mediated apoptosis. TRAIL-C3 demonstrated much stronger pro-apoptotic activity than wild-type (WT) TRAIL in a panel of AML cell lines as well as in primary AML blasts. The higher pro-apoptotic potential was further enhanced when the TRAIL mutant was used in combination with BMS-345541, a selective inhibitor of inhibitor-κB kinases. It illustrates that combination of this TRAIL variant with chemotherapeutics or other targeted agents can kill AML with high efficacy. This may represent a major advantage over the currently used therapies that have serious toxic side effects. The high efficacy of rhTRAIL-C3 containing therapies may enable the use of lower drug doses to reduce the toxic side effects and improve patient outcome. Our findings suggest that the rational design of TRAIL variants that target DR4 potentiate the death-inducing activity of TRAIL and offer a novel therapeutic strategy for the treatment of AML.

T-RMSD: a fine-grained, structure-based classification method and its application to the functional characterization of TNF receptors.

This study addresses the relation between structural and functional similarity in proteins. We introduce a novel method named tree based on root mean square deviation (T-RMSD), which uses distance RMSD (dRMSD) variations to build fine-grained structure-based classifications of proteins. The main improvement of the T-RMSD over similar methods, such as Dali, is its capacity to produce the equivalent of a bootstrap value for each cluster node. We validated our approach on two domain families studied extensively for their role in many biological and pathological pathways: the small GTPase RAS superfamily and the cysteine-rich domains (CRDs) associated with the tumor necrosis factor receptors (TNFRs) family. Our analysis showed that T-RMSD is able to automatically recover and refine existing classifications. In the case of the small GTPase ARF subfamily, T-RMSD can distinguish GTP- from GDP-bound states, while in the case of CRDs it can identify two new subgroups associated with well defined functional features (ligand binding and formation of ligand pre-assembly complex). We show how hidden Markov models (HMMs) can be built on these new groups and propose a methodology to use these models simultaneously in order to do fine-grained functional genomic annotation without known 3D structures. T-RMSD, an open source freeware incorporated in the T-Coffee package, is available online.

Building blocks for protein interaction devices.

Here, we propose a framework for the design of synthetic protein networks from modular protein-protein or protein-peptide interactions and provide a starter toolkit of protein building blocks. Our proof of concept experiments outline a general work flow for part-based protein systems engineering. We streamlined the iterative BioBrick cloning protocol and assembled 25 synthetic multidomain proteins each from seven standardized DNA fragments. A systematic screen revealed two main factors controlling protein expression in Escherichia coli: obstruction of translation initiation by mRNA secondary structure or toxicity of individual domains. Eventually, 13 proteins were purified for further characterization. Starting from well-established biotechnological tools, two general-purpose interaction input and two readout devices were built and characterized in vitro. Constitutive interaction input was achieved with a pair of synthetic leucine zippers. The second interaction was drug-controlled utilizing the rapamycin-induced binding of FRB(T2098L) to FKBP12. The interaction kinetics of both devices were analyzed by surface plasmon resonance. Readout was based on Förster resonance energy transfer between fluorescent proteins and was quantified for various combinations of input and output devices. Our results demonstrate the feasibility of parts-based protein synthetic biology. Additionally, we identify future challenges and limitations of modular design along with approaches to address them.