Visible Machine Learning for Biomedicine.

A major ambition of artificial intelligence lies in translating patient data to successful therapies. Machine learning models face particular challenges in biomedicine, however, including handling of extreme data heterogeneity and lack of mechanistic insight into predictions. Here, we argue for "visible" approaches that guide model structure with experimental biology.

A multi-scale map of cell structure fusing protein images and interactions.

The cell is a multi-scale structure with modular organization across at least four orders of magnitude1. Two central approaches for mapping this structure-protein fluorescent imaging and protein biophysical association-each generate extensive datasets, but of distinct qualities and resolutions that are typically treated separately2,3. Here we integrate immunofluorescence images in the Human Protein Atlas4 with affinity purifications in BioPlex5 to create a unified hierarchical map of human cell architecture. Integration is achieved by configuring each approach as a general measure of protein distance, then calibrating the two measures using machine learning. The map, known as the multi-scale integrated cell (MuSIC 1.0), resolves 69 subcellular systems, of which approximately half are to our knowledge undocumented. Accordingly, we perform 134 additional affinity purifications and validate subunit associations for the majority of systems. The map reveals a pre-ribosomal RNA processing assembly and accessory factors, which we show govern rRNA maturation, and functional roles for SRRM1 and FAM120C in chromatin and RPS3A in splicing. By integration across scales, MuSIC increases the resolution of imaging while giving protein interactions a spatial dimension, paving the way to incorporate diverse types of data in proteome-wide cell maps.

Methylome-wide Analysis of Chronic HIV Infection Reveals Five-Year Increase in Biological Age and Epigenetic Targeting of HLA.

HIV-infected individuals are living longer on antiretroviral therapy, but many patients display signs that in some ways resemble premature aging. To investigate and quantify the impact of chronic HIV infection on aging, we report a global analysis of the whole-blood DNA methylomes of 137 HIV+ individuals under sustained therapy along with 44 matched HIV- individuals. First, we develop and validate epigenetic models of aging that are independent of blood cell composition. Using these models, we find that both chronic and recent HIV infection lead to an average aging advancement of 4.9 years, increasing expected mortality risk by 19%. In addition, sustained infection results in global deregulation of the methylome across >80,000 CpGs and specific hypomethylation of the region encoding the human leukocyte antigen locus (HLA). We find that decreased HLA methylation is predictive of lower CD4 / CD8 T cell ratio, linking molecular aging, epigenetic regulation, and disease progression.

Genetic dissection of complex traits using hierarchical biological knowledge.

Despite the growing constellation of genetic loci linked to common traits, these loci have yet to account for most heritable variation, and most act through poorly understood mechanisms. Recent machine learning (ML) systems have used hierarchical biological knowledge to associate genetic mutations with phenotypic outcomes, yielding substantial predictive power and mechanistic insight. Here, we use an ontology-guided ML system to map single nucleotide variants (SNVs) focusing on 6 classic phenotypic traits in natural yeast populations. The 29 identified loci are largely novel and account for ~17% of the phenotypic variance, versus <3% for standard genetic analysis. Representative results show that sensitivity to hydroxyurea is linked to SNVs in two alternative purine biosynthesis pathways, and that sensitivity to copper arises through failure to detoxify reactive oxygen species in fatty acid metabolism. This work demonstrates a knowledge-based approach to amplifying and interpreting signals in population genetic studies.

Combinatorial CRISPR-Cas9 screens for de novo mapping of genetic interactions.

We developed a systematic approach to map human genetic networks by combinatorial CRISPR-Cas9 perturbations coupled to robust analysis of growth kinetics. We targeted all pairs of 73 cancer genes with dual guide RNAs in three cell lines, comprising 141,912 tests of interaction. Numerous therapeutically relevant interactions were identified, and these patterns replicated with combinatorial drugs at 75% precision. From these results, we anticipate that cellular context will be critical to synthetic-lethal therapies.

The condition-dependent transcriptional landscape of Burkholderia pseudomallei.

Burkholderia pseudomallei (Bp), the causative agent of the often-deadly infectious disease melioidosis, contains one of the largest prokaryotic genomes sequenced to date, at 7.2 Mb with two large circular chromosomes (1 and 2). To comprehensively delineate the Bp transcriptome, we integrated whole-genome tiling array expression data of Bp exposed to >80 diverse physical, chemical, and biological conditions. Our results provide direct experimental support for the strand-specific expression of 5,467 Sanger protein-coding genes, 1,041 operons, and 766 non-coding RNAs. A large proportion of these transcripts displayed condition-dependent expression, consistent with them playing functional roles. The two Bp chromosomes exhibited dramatically different transcriptional landscapes--Chr 1 genes were highly and constitutively expressed, while Chr 2 genes exhibited mosaic expression where distinct subsets were expressed in a strongly condition-dependent manner. We identified dozens of cis-regulatory motifs associated with specific condition-dependent expression programs, and used the condition compendium to elucidate key biological processes associated with two complex pathogen phenotypes--quorum sensing and in vivo infection. Our results demonstrate the utility of a Bp condition-compendium as a community resource for biological discovery. Moreover, the observation that significant portions of the Bp virulence machinery can be activated by specific in vitro cues provides insights into Bp's capacity as an "accidental pathogen", where genetic pathways used by the bacterium to survive in environmental niches may have also facilitated its ability to colonize human hosts.

De novo generation of multi-target compounds using deep generative chemistry.

Polypharmacology drugs-compounds that inhibit multiple proteins-have many applications but are difficult to design. To address this challenge we have developed POLYGON, an approach to polypharmacology based on generative reinforcement learning. POLYGON embeds chemical space and iteratively samples it to generate new molecular structures; these are rewarded by the predicted ability to inhibit each of two protein targets and by drug-likeness and ease-of-synthesis. In binding data for >100,000 compounds, POLYGON correctly recognizes polypharmacology interactions with 82.5% accuracy. We subsequently generate de-novo compounds targeting ten pairs of proteins with documented co-dependency. Docking analysis indicates that top structures bind their two targets with low free energies and similar 3D orientations to canonical single-protein inhibitors. We synthesize 32 compounds targeting MEK1 and mTOR, with most yielding >50% reduction in each protein activity and in cell viability when dosed at 1-10 µM. These results support the potential of generative modeling for polypharmacology.

Growth inhibition of pathogenic bacteria by sulfonylurea herbicides.

Emerging resistance to current antibiotics raises the need for new microbial drug targets. We show that targeting branched-chain amino acid (BCAA) biosynthesis using sulfonylurea herbicides, which inhibit the BCAA biosynthetic enzyme acetohydroxyacid synthase (AHAS), can exert bacteriostatic effects on several pathogenic bacteria, including Burkholderia pseudomallei, Pseudomonas aeruginosa, and Acinetobacter baumannii. Our results suggest that targeting biosynthetic enzymes like AHAS, which are lacking in humans, could represent a promising antimicrobial drug strategy.

Modulation of Caenorhabditis elegans infection sensitivity by the LIN-7 cell junction protein.

In Caenorhabditis elegans, the LIN-2/7/10 protein complex regulates the activity of signalling proteins. We found that inhibiting lin-7 function, and also lin-2 and lin-10, resulted in enhanced C. elegans survival after infection by Burkholderia spp., implicating a novel role for these genes in modulating infection outcomes. Genetic experiments suggested that this infection phenotype is likely caused by modulation of the DAF-2 insulin/IGF-1 signalling pathway. Supporting these observations, yeast two-hybrid assays confirmed that the LIN-2 PDZ domain can physically bind to the DAF-2 C-terminus. Loss of lin-7 activity also altered DAF-16 nuclear localization kinetics, indicating an additional contribution by hsf-1. Unexpectedly, silencing lin-7 in the hypodermis, but not the intestine, was protective against infection, implicating the hypodermis as a key tissue in this phenomenon. Finally, consistent with lin-7 acting as a general host infection factor, lin-7 mutants also exhibited enhanced survival upon infection by two other Gram-negative pathogens, Pseudomonas and Salmonella spp.

Mapping the common gene networks that underlie related diseases.

A longstanding goal of biomedicine is to understand how alterations in molecular and cellular networks give rise to the spectrum of human diseases. For diseases with shared etiology, understanding the common causes allows for improved diagnosis of each disease, development of new therapies and more comprehensive identification of disease genes. Accordingly, this protocol describes how to evaluate the extent to which two diseases, each characterized by a set of mapped genes, are colocalized in a reference gene interaction network. This procedure uses network propagation to measure the network 'distance' between gene sets. For colocalized diseases, the network can be further analyzed to extract common gene communities at progressive granularities. In particular, we show how to: (1) obtain input gene sets and a reference gene interaction network; (2) identify common subnetworks of genes that encompass or are in close proximity to all gene sets; (3) use multiscale community detection to identify systems and pathways represented by each common subnetwork to generate a network colocalized systems map; (4) validate identified genes and systems using a mouse variant database; and (5) visualize and further investigate select genes, interactions and systems for relevance to phenotype(s) of interest. We demonstrate the utility of this approach by identifying shared biological mechanisms underlying autism and congenital heart disease. However, this protocol is general and can be applied to any gene sets attributed to diseases or other phenotypes with suspected joint association. A typical NetColoc run takes less than an hour. Software and documentation are available at https://github.com/ucsd-ccbb/NetColoc .

Multimodal perturbation analyses of cyclin-dependent kinases reveal a network of synthetic lethalities associated with cell-cycle regulation and transcriptional regulation.

Cell-cycle control is accomplished by cyclin-dependent kinases (CDKs), motivating extensive research into CDK targeting small-molecule drugs as cancer therapeutics. Here we use combinatorial CRISPR/Cas9 perturbations to uncover an extensive network of functional interdependencies among CDKs and related factors, identifying 43 synthetic-lethal and 12 synergistic interactions. We dissect CDK perturbations using single-cell RNAseq, for which we develop a novel computational framework to precisely quantify cell-cycle effects and diverse cell states orchestrated by specific CDKs. While pairwise disruption of CDK4/6 is synthetic-lethal, only CDK6 is required for normal cell-cycle progression and transcriptional activation. Multiple CDKs (CDK1/7/9/12) are synthetic-lethal in combination with PRMT5, independent of cell-cycle control. In-depth analysis of mRNA expression and splicing patterns provides multiple lines of evidence that the CDK-PRMT5 dependency is due to aberrant transcriptional regulation resulting in premature termination. These inter-dependencies translate to drug-drug synergies, with therapeutic implications in cancer and other diseases.

Genome-wide association studies of human and rat BMI converge on synapse, epigenome, and hormone signaling networks.

A vexing observation in genome-wide association studies (GWASs) is that parallel analyses in different species may not identify orthologous genes. Here, we demonstrate that cross-species translation of GWASs can be greatly improved by an analysis of co-localization within molecular networks. Using body mass index (BMI) as an example, we show that the genes associated with BMI in humans lack significant agreement with those identified in rats. However, the networks interconnecting these genes show substantial overlap, highlighting common mechanisms including synaptic signaling, epigenetic modification, and hormonal regulation. Genetic perturbations within these networks cause abnormal BMI phenotypes in mice, too, supporting their broad conservation across mammals. Other mechanisms appear species specific, including carbohydrate biosynthesis (humans) and glycerolipid metabolism (rodents). Finally, network co-localization also identifies cross-species convergence for height/body length. This study advances a general paradigm for determining whether and how phenotypes measured in model species recapitulate human biology.

Endogenous factors enhance HIV infection of tissue naive CD4 T cells by stimulating high molecular mass APOBEC3G complex formation.

Human immunodeficiency virus (HIV) can infect resting CD4 T cells residing in lymphoid tissues but not those circulating in peripheral blood. The molecular mechanisms producing this difference remain unknown. We explored the potential role of the tissue microenvironment and its influence on the action of the antiviral factor APOBEC3G (A3G) in regulating permissivity to HIV infection. We found that endogenous IL-2 and -15 play a key role in rendering resident naive CD4 T cells susceptible to HIV infection. Infection of memory CD4 T cells also requires endogenous soluble factors, but not IL-2 or -15. A3G is found in a high molecular mass complex in HIV infection-permissive, tissue-resident naive CD4 T cells but resides in a low molecular mass form in nonpermissive, blood-derived naive CD4 T cells. Upon treatment with endogenous soluble factors, these cells become permissive for HIV infection, as low molecular mass A3G is induced to assemble into high molecular mass complexes. These findings suggest that in lymphoid tissues, endogenous soluble factors, likely including IL-2 and -15 and others, stimulate the formation of high molecular mass A3G complexes in tissue-resident naive CD4 T cells, thereby relieving the potent postentry restriction block for HIV infection conferred by low molecular mass A3G.

A genomic survey of positive selection in Burkholderia pseudomallei provides insights into the evolution of accidental virulence.

Certain environmental microorganisms can cause severe human infections, even in the absence of an obvious requirement for transition through an animal host for replication ("accidental virulence"). To understand this process, we compared eleven isolate genomes of Burkholderia pseudomallei (Bp), a tropical soil microbe and causative agent of the human and animal disease melioidosis. We found evidence for the existence of several new genes in the Bp reference genome, identifying 282 novel genes supported by at least two independent lines of supporting evidence (mRNA transcripts, database homologs, and presence of ribosomal binding sites) and 81 novel genes supported by all three lines. Within the Bp core genome, 211 genes exhibited significant levels of positive selection (4.5%), distributed across many cellular pathways including carbohydrate and secondary metabolism. Functional experiments revealed that certain positively selected genes might enhance mammalian virulence by interacting with host cellular pathways or utilizing host nutrients. Evolutionary modifications improving Bp environmental fitness may thus have indirectly facilitated the ability of Bp to colonize and survive in mammalian hosts. These findings improve our understanding of the pathogenesis of melioidosis, and establish Bp as a model system for studying the genetics of accidental virulence.

Cellular APOBEC3G restricts HIV-1 infection in resting CD4+ T cells.

In contrast to activated CD4+ T cells, resting human CD4+ T cells circulating in blood are highly resistant to infection with human immunodeficiency virus (HIV). Whether the inability of HIV to infect these resting CD4+ T cells is due to the lack of a key factor, or alternatively reflects the presence of an efficient mechanism for defence against HIV, is not clear. Here we show that the anti-retroviral deoxycytidine deaminase APOBEC3G strongly protects unstimulated peripheral blood CD4+ T cells against HIV-1 infection. In activated CD4+ T cells, cytoplasmic APOBEC3G resides in an enzymatically inactive, high-molecular-mass (HMM) ribonucleoprotein complex that converts to an enzymatically active low-molecular-mass (LMM) form after treatment with RNase. In contrast, LMM APOBEC3G predominates in unstimulated CD4+ T cells, where HIV-1 replication is blocked and reverse transcription is impaired. Mitogen activation induces the recruitment of LMM APOBEC3G into the HMM complex, and this correlates with a sharp increase in permissivity for HIV infection in these stimulated cells. Notably, when APOBEC3G-specific small interfering RNAs are introduced into unstimulated CD4+ T cells, the early replication block encountered by HIV-1 is greatly relieved. Thus, LMM APOBEC3G functions as a potent post-entry restriction factor for HIV-1 in unstimulated CD4+ T cells. Surprisingly, sequencing of the reverse transcripts slowly formed in unstimulated CD4+ T cells reveals only low levels of dG dA hypermutation, raising the possibility that the APOBEC3G-restricting activity may not be strictly dependent on deoxycytidine deamination

A protein interaction landscape of breast cancer.

Cancers have been associated with a diverse array of genomic alterations. To help mechanistically understand such alterations in breast-invasive carcinoma, we applied affinity purification­mass spectrometry to delineate comprehensive biophysical interaction networks for 40 frequently altered breast cancer (BC) proteins, with and without relevant mutations, across three human breast cell lines. These networks identify cancer-specific protein-protein interactions (PPIs), interconnected and enriched for common and rare cancer mutations, that are substantially rewired by the introduction of key BC mutations. Our analysis identified BPIFA1 and SCGB2A1 as PIK3CA-interacting proteins, which repress PI3K-AKT signaling, and uncovered USP28 and UBE2N as functionally relevant interactors of BRCA1. We also show that the protein phosphatase 1 regulatory subunit spinophilin interacts with and regulates dephosphorylation of BRCA1 to promote DNA double-strand break repair. Thus, PPI landscapes provide a powerful framework for mechanistically interpreting disease genomic data and can identify valuable therapeutic targets.

Interpretation of cancer mutations using a multiscale map of protein systems.

A major goal of cancer research is to understand how mutations distributed across diverse genes affect common cellular systems, including multiprotein complexes and assemblies. Two challenges­how to comprehensively map such systems and how to identify which are under mutational selection­have hindered this understanding. Accordingly, we created a comprehensive map of cancer protein systems integrating both new and published multi-omic interaction data at multiple scales of analysis. We then developed a unified statistical model that pinpoints 395 specific systems under mutational selection across 13 cancer types. This map, called NeST (Nested Systems in Tumors), incorporates canonical processes and notable discoveries, including a PIK3CA-actomyosin complex that inhibits phosphatidylinositol 3-kinase signaling and recurrent mutations in collagen complexes that promote tumor proliferation. These systems can be used as clinical biomarkers and implicate a total of 548 genes in cancer evolution and progression. This work shows how disparate tumor mutations converge on protein assemblies at different scales.

A protein network map of head and neck cancer reveals PIK3CA mutant drug sensitivity.

We outline a framework for elucidating tumor genetic complexity through multidimensional protein-protein interaction maps and apply it to enhancing our understanding of head and neck squamous cell carcinoma. This network uncovers 771 interactions from cancer and noncancerous cell states, including WT and mutant protein isoforms. Prioritization of cancer-enriched interactions reveals a previously unidentified association of the fibroblast growth factor receptor tyrosine kinase 3 with Daple, a guanine-nucleotide exchange factor, resulting in activation of Gαi- and p21-activated protein kinase 1/2 to promote cancer cell migration. Additionally, we observe mutation-enriched interactions between the human epidermal growth factor receptor 3 (HER3) receptor tyrosine kinase and PIK3CA (the alpha catalytic subunit of phosphatidylinositol 3-kinase) that can inform the response to HER3 inhibition in vivo. We anticipate that the application of this framework will be valuable for translating genetic alterations into a molecular and clinical understanding of the underlying biology of many disease areas.

Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells.

Most drugs entering clinical trials fail, often related to an incomplete understanding of the mechanisms governing drug response. Machine learning techniques hold immense promise for better drug response predictions, but most have not reached clinical practice due to their lack of interpretability and their focus on monotherapies. We address these challenges by developing DrugCell, an interpretable deep learning model of human cancer cells trained on the responses of 1,235 tumor cell lines to 684 drugs. Tumor genotypes induce states in cellular subsystems that are integrated with drug structure to predict response to therapy and, simultaneously, learn biological mechanisms underlying the drug response. DrugCell predictions are accurate in cell lines and also stratify clinical outcomes. Analysis of DrugCell mechanisms leads directly to the design of synergistic drug combinations, which we validate systematically by combinatorial CRISPR, drug-drug screening in vitro, and patient-derived xenografts. DrugCell provides a blueprint for constructing interpretable models for predictive medicine.

Quantitative Translation of Dog-to-Human Aging by Conserved Remodeling of the DNA Methylome.

All mammals progress through similar physiological stages throughout life, from early development to puberty, aging, and death. Yet, the extent to which this conserved physiology reflects underlying genomic events is unclear. Here, we map the common methylation changes experienced by mammalian genomes as they age, focusing on comparison of humans with dogs, an emerging model of aging. Using oligo-capture sequencing, we characterize methylomes of 104 Labrador retrievers spanning a 16-year age range, achieving >150× coverage within mammalian syntenic blocks. Comparison with human methylomes reveals a nonlinear relationship that translates dog-to-human years and aligns the timing of major physiological milestones between the two species, with extension to mice. Conserved changes center on developmental gene networks, which are sufficient to translate age and the effects of anti-aging interventions across multiple mammals. These results establish methylation not only as a diagnostic age readout but also as a cross-species translator of physiological aging milestones.