Cell reprogramming design by transfer learning of functional transcriptional networks.

Recent developments in synthetic biology, next-generation sequencing, and machine learning provide an unprecedented opportunity to rationally design new disease treatments based on measured responses to gene perturbations and drugs to reprogram cells. The main challenges to seizing this opportunity are the incomplete knowledge of the cellular network and the combinatorial explosion of possible interventions, both of which are insurmountable by experiments. To address these challenges, we develop a transfer learning approach to control cell behavior that is pre-trained on transcriptomic data associated with human cell fates, thereby generating a model of the network dynamics that can be transferred to specific reprogramming goals. The approach combines transcriptional responses to gene perturbations to minimize the difference between a given pair of initial and target transcriptional states. We demonstrate our approach's versatility by applying it to a microarray dataset comprising >9,000 microarrays across 54 cell types and 227 unique perturbations, and an RNASeq dataset consisting of >10,000 sequencing runs across 36 cell types and 138 perturbations. Our approach reproduces known reprogramming protocols with an AUROC of 0.91 while innovating over existing methods by pre-training an adaptable model that can be tailored to specific reprogramming transitions. We show that the number of gene perturbations required to steer from one fate to another increases with decreasing developmental relatedness and that fewer genes are needed to progress along developmental paths than to regress. These findings establish a proof-of-concept for our approach to computationally design control strategies and provide insights into how gene regulatory networks govern phenotype.

Cell reprogramming design by transfer learning of functional transcriptional networks.

Recent developments in synthetic biology, next-generation sequencing, and machine learning provide an unprecedented opportunity to rationally design new disease treatments based on measured responses to gene perturbations and drugs to reprogram cell behavior. The main challenges to seizing this opportunity are the incomplete knowledge of the cellular network and the combinatorial explosion of possible interventions, both of which are insurmountable by experiments. To address these challenges, we develop a transfer learning approach to control cell behavior that is pre-trained on transcriptomic data associated with human cell fates to generate a model of the functional network dynamics that can be transferred to specific reprogramming goals. The approach additively combines transcriptional responses to gene perturbations (single-gene knockdowns and overexpressions) to minimize the transcriptional difference between a given pair of initial and target states. We demonstrate the flexibility of our approach by applying it to a microarray dataset comprising over 9,000 microarrays across 54 cell types and 227 unique perturbations, and an RNASeq dataset consisting of over 10,000 sequencing runs across 36 cell types and 138 perturbations. Our approach reproduces known reprogramming protocols with an average AUROC of 0.91 while innovating over existing methods by pre-training an adaptable model that can be tailored to specific reprogramming transitions. We show that the number of gene perturbations required to steer from one fate to another increases as the developmental relatedness decreases. We also show that fewer genes are needed to progress along developmental paths than to regress. Together, these findings establish a proof-of-concept for our approach to computationally design control strategies and demonstrate their ability to provide insights into the dynamics of gene regulatory networks.

Assessment of community efforts to advance network-based prediction of protein-protein interactions.

Comprehensive understanding of the human protein-protein interaction (PPI) network, aka the human interactome, can provide important insights into the molecular mechanisms of complex biological processes and diseases. Despite the remarkable experimental efforts undertaken to date to determine the structure of the human interactome, many PPIs remain unmapped. Computational approaches, especially network-based methods, can facilitate the identification of previously uncharacterized PPIs. Many such methods have been proposed. Yet, a systematic evaluation of existing network-based methods in predicting PPIs is still lacking. Here, we report community efforts initiated by the International Network Medicine Consortium to benchmark the ability of 26 representative network-based methods to predict PPIs across six different interactomes of four different organisms: A. thaliana, C. elegans, S. cerevisiae, and H. sapiens. Through extensive computational and experimental validations, we found that advanced similarity-based methods, which leverage the underlying network characteristics of PPIs, show superior performance over other general link prediction methods in the interactomes we considered.

Computational Inference of Synaptic Polarities in Neuronal Networks.

Synaptic polarity, that is, whether synapses are inhibitory (-) or excitatory (+), is challenging to map, despite being a key to understand brain function. Here, synaptic polarity is inferred computationally considering three experimental scenarios, depending on the nature of available input data, using the Caenorhabditis elegans connectome as an example. First, the inputs consist of detailed neurotransmitter (NT) and receptor (R) gene expression, integrated through the connectome model (CM). The CM formulates the problem through a wiring rule network that summarizes how NT-R pairs govern synaptic polarity, and resolves 356 synaptic polarities in addition to the 1752 known polarities. Second, known synaptic polarities are considered as an input, in addition to the NT and R gene expression data, but without wiring rules. These data train the spatial connectome model, which infers the polarity of 81% of the CM-resolved connections at >95$>95$ % precision, while also inferring 147 of the remaining unknown polarities. Last, without known expression or wiring rules, polarities are inferred through a network sign prediction problem. As an illustration of high performance in this case, the generalized CM is introduced. These results address imminent challenges in unveiling large-scale synaptic polarities, an essential step toward more realistic brain models.

Extreme Antagonism Arising from Gene-Environment Interactions.

Antagonistic interactions in biological systems, which occur when one perturbation blunts the effect of another, are typically interpreted as evidence that the two perturbations impact the same cellular pathway or function. Yet, this interpretation ignores extreme antagonistic interactions wherein an otherwise deleterious perturbation compensates for the function lost because of a prior perturbation. Here, we report on gene-environment interactions involving genetic mutations that are deleterious in a permissive environment but beneficial in a specific environment that restricts growth. These extreme antagonistic interactions constitute gene-environment analogs of synthetic rescues previously observed for gene-gene interactions. Our approach uses two independent adaptive evolution steps to address the lack of experimental methods to systematically identify such extreme interactions. We apply the approach to Escherichia coli by successively adapting it to defined glucose media without and with the antibiotic rifampicin. The approach identified multiple mutations that are beneficial in the presence of rifampicin and deleterious in its absence. The analysis of transcription shows that the antagonistic adaptive mutations repress a stringent response-like transcriptional program, whereas nonantagonistic mutations have an opposite transcriptional profile. Our approach represents a step toward the systematic characterization of extreme antagonistic gene-drug interactions, which can be used to identify targets to select against antibiotic resistance.

Distinguishing cell phenotype using cell epigenotype.

The relationship between microscopic observations and macroscopic behavior is a fundamental open question in biophysical systems. Here, we develop a unified approach that-in contrast with existing methods-predicts cell type from macromolecular data even when accounting for the scale of human tissue diversity and limitations in the available data. We achieve these benefits by applying a k-nearest-neighbors algorithm after projecting our data onto the eigenvectors of the correlation matrix inferred from many observations of gene expression or chromatin conformation. Our approach identifies variations in epigenotype that affect cell type, thereby supporting the cell-type attractor hypothesis and representing the first step toward model-independent control strategies in biological systems.

Predicting growth rate from gene expression.

Growth rate is one of the most important and most complex phenotypic characteristics of unicellular microorganisms, which determines the genetic mutations that dominate at the population level, and ultimately whether the population will survive. Translating changes at the genetic level to their growth-rate consequences remains a subject of intense interest, since such a mapping could rationally direct experiments to optimize antibiotic efficacy or bioreactor productivity. In this work, we directly map transcriptional profiles to growth rates by gathering published gene-expression data from Escherichia coli and Saccharomyces cerevisiae with corresponding growth-rate measurements. Using a machine-learning technique called k-nearest-neighbors regression, we build a model which predicts growth rate from gene expression. By exploiting the correlated nature of gene expression and sparsifying the model, we capture 81% of the variance in growth rate of the E. coli dataset, while reducing the number of features from >4,000 to 9. In S. cerevisiae, we account for 89% of the variance in growth rate, while reducing from >5,500 dimensions to 18. Such a model provides a basis for selecting successful strategies from among the combinatorial number of experimental possibilities when attempting to optimize complex phenotypic traits like growth rate.

Correction: Experimental evolution of diverse Escherichia coli metabolic mutants identifies genetic loci for convergent adaptation of growth rate.

[This corrects the article DOI: 10.1371/journal.pgen.1007284.].

Experimental evolution of diverse Escherichia coli metabolic mutants identifies genetic loci for convergent adaptation of growth rate.

Cell growth is determined by substrate availability and the cell's metabolic capacity to assimilate substrates into building blocks. Metabolic genes that determine growth rate may interact synergistically or antagonistically, and can accelerate or slow growth, depending on genetic background and environmental conditions. We evolved a diverse set of Escherichia coli single-gene deletion mutants with a spectrum of growth rates and identified mutations that generally increase growth rate. Despite the metabolic differences between parent strains, mutations that enhanced growth largely mapped to core transcription machinery, including the ß and ß' subunits of RNA polymerase (RNAP) and the transcription elongation factor, NusA. The structural segments of RNAP that determine enhanced growth have been previously implicated in antibiotic resistance and in the control of transcription elongation and pausing. We further developed a computational framework to characterize how the transcriptional changes that occur upon acquisition of these mutations affect growth rate across strains. Our experimental and computational results provide evidence for cases in which RNAP mutations shift the competitive balance between active transcription and gene silencing. This study demonstrates that mutations in specific regions of RNAP are a convergent adaptive solution that can enhance the growth rate of cells from distinct metabolic states.