Sparse dictionary learning recovers pleiotropy from human cell fitness screens

The present and future of the Cancer Dependency Map

Despite tremendous progress in the past decade, the complex and heterogeneous nature of cancer complicates efforts to identify new therapies and therapeutic combinations that achieve durable responses in most patients. Further advances in cancer therapy will rely, in part, on the development of targeted therapeutics matched with the genetic and molecular characteristics of cancer. The Cancer Dependency Map (DepMap) is a large-scale data repository and research platform, aiming to systematically reveal the landscape of cancer vulnerabilities in thousands of genetically and molecularly annotated cancer models. DepMap is used routinely by cancer researchers and translational scientists and has facilitated the identification of several novel and selective therapeutic strategies for multiple cancer types that are being tested in the clinic. However, it is also clear that the current version of DepMap is not yet comprehensive. In this Perspective, we review (1) the impact and current uses of DepMap, (2) the opportunities to enhance DepMap to overcome its current limitations, and (3) the ongoing efforts to further improve and expand DepMap.

Combinatorial prediction of therapeutic perturbations using causally-inspired neural networks

As an alternative to target-driven drug discovery, phenotype-driven approaches identify compounds that counteract the overall disease effects by analyzing phenotypic signatures. Our study introduces a novel approach to this field, aiming to expand the search space for new therapeutic agents. We introduce PDGrapher, a causally-inspired graph neural network (GNN) designed to predict combinatorial perturbagens - sets of therapeutic targets - capable of reversing disease effects. Unlike methods that learn responses to perturbations, PDGrapher solves the inverse problem, which is to infer the perturbagens necessary to achieve a specific response - i.e., directly predicting perturbagens by learning which perturbations elicit a desired response. By encoding gene regulatory networks or protein-protein interactions, PDGrapher can predict unseen chemical or genetic perturbagens, aiding in the discovery of novel drugs or therapeutic targets. Experiments across nine cell lines with chemical perturbations show that PDGrapher successfully predicted effective perturbagens in up to 13.33% additional test samples and ranked therapeutic targets up to 35% higher than the competing methods, and the method shows competitive performance across ten genetic perturbation datasets. A key innovation of PDGrapher is its direct prediction capability, which contrasts with the indirect, computationally intensive models traditionally used in phenotype-driven drug discovery that only predict changes in phenotypes due to perturbations. The direct approach enables PDGrapher to train up to 25 times faster than methods like scGEN and CellOT, representing a considerable leap in efficiency. Our results suggest that PDGrapher can advance phenotype-driven drug discovery, offering a fast and comprehensive approach to identifying therapeutically useful perturbations.

Contextual AI models for single-cell protein biology

Understanding protein function and developing molecular therapies require deciphering the cell types in which proteins act as well as the interactions between proteins. However, modeling protein interactions across biological contexts remains challenging for existing algorithms. Here we introduce PINNACLE, a geometric deep learning approach that generates context-aware protein representations. Leveraging a multiorgan single-cell atlas, PINNACLE learns on contextualized protein interaction networks to produce 394,760 protein representations from 156 cell type contexts across 24 tissues. PINNACLE's embedding space reflects cellular and tissue organization, enabling zero-shot retrieval of the tissue hierarchy. Pretrained protein representations can be adapted for downstream tasks: enhancing 3D structure-based representations for resolving immuno-oncological protein interactions, and investigating drugs' effects across cell types. PINNACLE outperforms state-of-the-art models in nominating therapeutic targets for rheumatoid arthritis and inflammatory bowel diseases and pinpoints cell type contexts with higher predictive capability than context-free models. PINNACLE's ability to adjust its outputs on the basis of the context in which it operates paves the way for large-scale context-specific predictions in biology.

Contextual AI models for single-cell protein biology

Understanding protein function and developing molecular therapies require deciphering the cell types in which proteins act as well as the interactions between proteins. However, modeling protein interactions across biological contexts remains challenging for existing algorithms. Here, we introduce Pinnacle, a geometric deep learning approach that generates context-aware protein representations. Leveraging a multi-organ single-cell atlas, Pinnacle learns on contextualized protein interaction networks to produce 394,760 protein representations from 156 cell type contexts across 24 tissues. Pinnacle's embedding space reflects cellular and tissue organization, enabling zero-shot retrieval of the tissue hierarchy. Pretrained protein representations can be adapted for downstream tasks: enhancing 3D structure-based representations for resolving immuno-oncological protein interactions, and investigating drugs' effects across cell types. Pinnacle outperforms state-of-the-art models in nominating therapeutic targets for rheumatoid arthritis and inflammatory bowel diseases, and pinpoints cell type contexts with higher predictive capability than context-free models. Pinnacle's ability to adjust its outputs based on the context in which it operates paves way for large-scale context-specific predictions in biology.

Structural basis of the Integrator complex assembly and association with transcription factors

Integrator is a multi-subunit protein complex responsible for premature transcription termination of coding and non-coding RNAs. This is achieved via two enzymatic activities, RNA endonuclease and protein phosphatase, acting on the promoter-proximally paused RNA polymerase Ⅱ (RNAPⅡ). Yet, it remains unclear how Integrator assembly and recruitment are regulated and what the functions of many of its core subunits are. Here, we report the structures of two human Integrator sub-complexes: INTS10/13/14/15 and INTS5/8/10/15, and an integrative model of the fully assembled Integrator bound to the RNAPⅡ paused elongating complex (PEC). An in silico protein-protein interaction screen of over 1,500 human transcription factors (TFs) identified ZNF655 as a direct interacting partner of INTS13 within the fully assembled Integrator. We propose a model wherein INTS13 acts as a platform for the recruitment of TFs that could modulate the stability of the Integrator's association at specific loci and regulate transcription attenuation of the target genes.

On knowing a gene: A distributional hypothesis of gene function

As words can have multiple meanings that depend on sentence context, genes can have various functions that depend on the surrounding biological system. This pleiotropic nature of gene function is limited by ontologies, which annotate gene functions without considering biological contexts. We contend that the gene function problem in genetics may be informed by recent technological leaps in natural language processing, in which representations of word semantics can be automatically learned from diverse language contexts. In contrast to efforts to model semantics as "is-a" relationships in the 1990s, modern distributional semantics represents words as vectors in a learned semantic space and fuels current advances in transformer-based models such as large language models and generative pre-trained transformers. A similar shift in thinking of gene functions as distributions over cellular contexts may enable a similar breakthrough in data-driven learning from large biological datasets to inform gene function.

Predicting potential target genes in molecular biology experiments using machine learning and multifaceted data sources

Experimental analysis of functionally related genes is key to understanding biological phenomena. The selection of genes to study is a crucial and challenging step, as it requires extensive knowledge of the literature and diverse biomedical data resources. Although software tools that predict relationships between genes are available to accelerate this process, they do not directly incorporate experiment information derived from the literature. Here, we develop LEXAS, a target gene suggestion system for molecular biology experiments. LEXAS is based on machine learning models trained with diverse information sources, including 24 million experiment descriptions extracted from full-text articles in PubMed Central by using a deep-learning-based natural language processing model. By integrating the extracted experiment contexts with biomedical data sources, LEXAS suggests potential target genes for upcoming experiments, complementing existing tools like STRING, FunCoup, and GOSemSim. A simple web interface enables biologists to consider newly derived gene information while planning experiments.

Evaluation of input data modality choices on functional gene embeddings

Functional gene embeddings, numerical vectors capturing gene function, provide a promising way to integrate functional gene information into machine learning models. These embeddings are learnt by applying self-supervised machine-learning algorithms on various data types including quantitative omics measurements, protein-protein interaction networks and literature. However, downstream evaluations comparing alternative data modalities used to construct functional gene embeddings have been lacking. Here we benchmarked functional gene embeddings obtained from various data modalities for predicting disease-gene lists, cancer drivers, phenotype-gene associations and scores from genome-wide association studies. Off-the-shelf predictors trained on precomputed embeddings matched or outperformed dedicated state-of-the-art predictors, demonstrating their high utility. Embeddings based on literature and protein-protein interactions inferred from low-throughput experiments outperformed embeddings derived from genome-wide experimental data (transcriptomics, deletion screens and protein sequence) when predicting curated gene lists. In contrast, they did not perform better when predicting genome-wide association signals and were biased towards highly-studied genes. These results indicate that embeddings derived from literature and low-throughput experiments appear favourable in many existing benchmarks because they are biased towards well-studied genes and should therefore be considered with caution. Altogether, our study and precomputed embeddings will facilitate the development of machine-learning models in genetics and related fields.

Dimensionality reduction methods for extracting functional networks from large-scale CRISPR screens

CRISPR-Cas9 screens facilitate the discovery of gene functional relationships and phenotype-specific dependencies. The Cancer Dependency Map (DepMap) is the largest compendium of whole-genome CRISPR screens aimed at identifying cancer-specific genetic dependencies across human cell lines. A mitochondria-associated bias has been previously reported to mask signals for genes involved in other functions, and thus, methods for normalizing this dominant signal to improve co-essentiality networks are of interest. In this study, we explore three unsupervised dimensionality reduction methods-autoencoders, robust, and classical principal component analyses (PCA)-for normalizing the DepMap to improve functional networks extracted from these data. We propose a novel "onion" normalization technique to combine several normalized data layers into a single network. Benchmarking analyses reveal that robust PCA combined with onion normalization outperforms existing methods for normalizing the DepMap. Our work demonstrates the value of removing low-dimensional signals from the DepMap before constructing functional gene networks and provides generalizable dimensionality reduction-based normalization tools.

Inferring sparse structure in genotype-phenotype maps

Correlation among multiple phenotypes across related individuals may reflect some pattern of shared genetic architecture: individual genetic loci affect multiple phenotypes (an effect known as pleiotropy), creating observable relationships between phenotypes. A natural hypothesis is that pleiotropic effects reflect a relatively small set of common "core" cellular processes: each genetic locus affects one or a few core processes, and these core processes in turn determine the observed phenotypes. Here, we propose a method to infer such structure in genotype-phenotype data. Our approach, sparse structure discovery (SSD) is based on a penalized matrix decomposition designed to identify latent structure that is low-dimensional (many fewer core processes than phenotypes and genetic loci), locus-sparse (each locus affects few core processes), and/or phenotype-sparse (each phenotype is influenced by few core processes). Our use of sparsity as a guide in the matrix decomposition is motivated by the results of a novel empirical test indicating evidence of sparse structure in several recent genotype-phenotype datasets. First, we use synthetic data to show that our SSD approach can accurately recover core processes if each genetic locus affects few core processes or if each phenotype is affected by few core processes. Next, we apply the method to three datasets spanning adaptive mutations in yeast, genotoxin robustness assay in human cell lines, and genetic loci identified from a yeast cross, and evaluate the biological plausibility of the core process identified. More generally, we propose sparsity as a guiding prior for resolving latent structure in empirical genotype-phenotype maps.

Context-defined cancer co-dependency mapping identifies a functional interplay between PRC2 and MLL-MEN1 complex in lymphoma

Interplay between chromatin-associated complexes and modifications critically contribute to the partitioning of epigenome into stable and functionally distinct domains. Yet there is a lack of systematic identification of chromatin crosstalk mechanisms, limiting our understanding of the dynamic transition between chromatin states during development and disease. Here we perform co-dependency mapping of genes using CRISPR-Cas9-mediated fitness screens in pan-cancer cell lines to quantify gene-gene functional relationships. We identify 145 co-dependency modules and further define the molecular context underlying the essentiality of these modules by incorporating mutational, epigenome, gene expression and drug sensitivity profiles of cell lines. These analyses assign new protein complex composition and function, and predict new functional interactions, including an unexpected co-dependency between two transcriptionally counteracting chromatin complexes - polycomb repressive complex 2 (PRC2) and MLL-MEN1 complex. We show that PRC2-mediated H3K27 tri-methylation regulates the genome-wide distribution of MLL1 and MEN1. In lymphoma cells with EZH2 gain-of-function mutations, the re-localization of MLL-MEN1 complex drives oncogenic gene expression and results in a hypersensitivity to pharmacologic inhibition of MEN1. Together, our findings provide a resource for discovery of trans-regulatory interactions as mechanisms of chromatin regulation and potential targets of synthetic lethality.

A multi-scale map of protein assemblies in the DNA damage response

The DNA damage response (DDR) ensures error-free DNA replication and transcription and is disrupted in numerous diseases. An ongoing challenge is to determine the proteins orchestrating DDR and their organization into complexes, including constitutive interactions and those responding to genomic insult. Here, we use multi-conditional network analysis to systematically map DDR assemblies at multiple scales. Affinity purifications of 21 DDR proteins, with/without genotoxin exposure, are combined with multi-omics data to reveal a hierarchical organization of 605 proteins into 109 assemblies. The map captures canonical repair mechanisms and proposes new DDR-associated proteins extending to stress, transport, and chromatin functions. We find that protein assemblies closely align with genetic dependencies in processing specific genotoxins and that proteins in multiple assemblies typically act in multiple genotoxin responses. Follow-up by DDR functional readouts newly implicates 12 assembly members in double-strand-break repair. The DNA damage response assemblies map is available for interactive visualization and query (ccmi.org/ddram/).

A combinatorial approach to uncover an additional Integrator subunit

RNA polymerase II (RNAPII) controls expression of all protein-coding genes and most noncoding loci in higher eukaryotes. Calibrating RNAPII activity requires an assortment of polymerase-associated factors that are recruited at sites of active transcription. The Integrator complex is one of the most elusive transcriptional regulators in metazoans, deemed to be recruited after initiation to help establish and modulate paused RNAPII. Integrator is known to be composed of 14 subunits that assemble and operate in a modular fashion. We employed proteomics and machine-learning structure prediction (AlphaFold2) to identify an additional Integrator subunit, INTS15. We report that INTS15 assembles primarily with the INTS13/14/10 module and interfaces with the Int-PP2A module. Functional genomics analysis further reveals a role for INTS15 in modulating RNAPII pausing at a subset of genes. Our study shows that omics approaches combined with AlphaFold2-based predictions provide additional insights into the molecular architecture of large and dynamic multiprotein complexes.

Integrator is a global promoter-proximal termination complex

Integrator is a metazoan-specific protein complex capable of inducing termination at all RNAPII-transcribed loci. Integrator recognizes paused, promoter-proximal RNAPII and drives premature termination using dual enzymatic activities: an endonuclease that cleaves nascent RNA and a protein phosphatase that removes stimulatory phosphorylation associated with RNAPII pause release and productive elongation. Recent breakthroughs in structural biology have revealed the overall architecture of Integrator and provided insights into how multiple Integrator modules are coordinated to elicit termination effectively. Furthermore, functional genomics and biochemical studies have unraveled how Integrator-mediated termination impacts protein-coding and noncoding loci. Here, we review the current knowledge about the assembly and activity of Integrator and describe the role of Integrator in gene regulation, highlighting the importance of this complex for human health.

Optimal construction of a functional interaction network from pooled library CRISPR fitness screens

BACKGROUND

Functional interaction networks, where edges connect genes likely to operate in the same biological process or pathway, can be inferred from CRISPR knockout screens in cancer cell lines. Genes with similar knockout fitness profiles across a sufficiently diverse set of cell line screens are likely to be co-functional, and these "coessentiality" networks are increasingly powerful predictors of gene function and biological modularity. While several such networks have been published, most use different algorithms for each step of the network construction process.

RESULTS

In this study, we identify an optimal measure of functional interaction and test all combinations of options at each step-essentiality scoring, sample variance and covariance normalization, and similarity measurement-to identify best practices for generating a functional interaction network from CRISPR knockout data. We show that Bayes Factor and Ceres scores give the best results, that Ceres outperforms the newer Chronos scoring scheme, and that covariance normalization is a critical step in network construction. We further show that Pearson correlation, mathematically identical to ordinary least squares after covariance normalization, can be extended by using partial correlation to detect and amplify signals from "moonlighting" proteins which show context-dependent interaction with different partners.

CONCLUSIONS

We describe a systematic survey of methods for generating coessentiality networks from the Cancer Dependency Map data and provide a partial correlation-based approach for exploring context-dependent interactions.

The phenotypic landscape of essential human genes

Understanding the basis for cellular growth, proliferation, and function requires determining the roles of essential genes in diverse cellular processes, including visualizing their contributions to cellular organization and morphology. Here, we combined pooled CRISPR-Cas9-based functional screening of 5,072 fitness-conferring genes in human HeLa cells with microscopy-based imaging of DNA, the DNA damage response, actin, and microtubules. Analysis of >31 million individual cells identified measurable phenotypes for >90% of gene knockouts, implicating gene targets in specific cellular processes. Clustering of phenotypic similarities based on hundreds of quantitative parameters further revealed co-functional genes across diverse cellular activities, providing predictions for gene functions and associations. By conducting pooled live-cell screening of ∼450,000 cell division events for 239 genes, we additionally identified diverse genes with functional contributions to chromosome segregation. Our work establishes a resource detailing the consequences of disrupting core cellular processes that represents the functional landscape of essential human genes.

Integrator endonuclease drives promoter-proximal termination at all RNA polymerase II-transcribed loci

RNA polymerase II (RNAPII) pausing in early elongation is critical for gene regulation. Paused RNAPII can be released into productive elongation by the kinase P-TEFb or targeted for premature termination by the Integrator complex. Integrator comprises endonuclease and phosphatase activities, driving termination by cleavage of nascent RNA and removal of stimulatory phosphorylation. We generated a degron system for rapid Integrator endonuclease (INTS11) depletion to probe the direct consequences of Integrator-mediated RNA cleavage. Degradation of INTS11 elicits nearly universal increases in active early elongation complexes. However, these RNAPII complexes fail to achieve optimal elongation rates and exhibit persistent Integrator phosphatase activity. Thus, only short transcripts are significantly upregulated following INTS11 loss, including transcription factors, signaling regulators, and non-coding RNAs. We propose a uniform molecular function for INTS11 across all RNAPII-transcribed loci, with differential effects on particular genes, pathways, or RNA biotypes reflective of transcript lengths rather than specificity of Integrator activity.

Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq

A central goal of genetics is to define the relationships between genotypes and phenotypes. High-content phenotypic screens such as Perturb-seq (CRISPR-based screens with single-cell RNA-sequencing readouts) enable massively parallel functional genomic mapping but, to date, have been used at limited scales. Here, we perform genome-scale Perturb-seq targeting all expressed genes with CRISPR interference (CRISPRi) across >2.5 million human cells. We use transcriptional phenotypes to predict the function of poorly characterized genes, uncovering new regulators of ribosome biogenesis (including CCDC86, ZNF236, and SPATA5L1), transcription (C7orf26), and mitochondrial respiration (TMEM242). In addition to assigning gene function, single-cell transcriptional phenotypes allow for in-depth dissection of complex cellular phenomena-from RNA processing to differentiation. We leverage this ability to systematically identify genetic drivers and consequences of aneuploidy and to discover an unanticipated layer of stress-specific regulation of the mitochondrial genome. Our information-rich genotype-phenotype map reveals a multidimensional portrait of gene and cellular function.

Phosphate dysregulation via the XPR1-KIDINS220 protein complex is a therapeutic vulnerability in ovarian cancer

Despite advances in precision medicine, the clinical prospects for patients with ovarian and uterine cancers have not substantially improved. Here, we analyzed genome-scale CRISPR/Cas9 loss-of-function screens across 851 human cancer cell lines and found that frequent overexpression of SLC34A2 – encoding a phosphate importer – is correlated to sensitivity to loss of the phosphate exporter XPR1 in vitro and in vivo. In patient-derived tumor samples, we observed frequent PAX8-dependent overexpression of SLC34A2, XPR1 copy number amplifications, and XPR1 mRNA overexpression. Mechanistically, in SLC34A2-high cancer cell lines, genetic or pharmacologic inhibition of XPR1-dependent phosphate efflux leads to the toxic accumulation of intracellular phosphate. Finally, we show that XPR1 requires the novel partner protein KIDINS220 for proper cellular localization and activity, and that disruption of this protein complex results in acidic vacuolar structures preceding cell death. These data point to the XPR1:KIDINS220 complex and phosphate dysregulation as a therapeutic vulnerability in ovarian cancer.

Golub and colleagues identify the phosphate exporter XPR1 as a therapeutic vulnerability in ovarian and uterine cancers, and show that phosphate efflux inhibition reduces tumor cell viability through accumulation of intracellular phosphate.