Self-mentoring: A new deep learning pipeline to train a self-supervised U-net for few-shot learning of bio-artificial capsule segmentation.

BACKGROUND: Accurate segmentation of microscopic structures such as bio-artificial capsules in microscopy imaging is a prerequisite to the computer-aided understanding of important biomechanical phenomenons. State-of-the-art segmentation performances are achieved by deep neural networks and related data-driven approaches. Training these networks from only a few annotated examples is challenging while producing manually annotated images that provide supervision is tedious. METHOD: Recently, self-supervision, i.e. designing a neural pipeline providing synthetic or indirect supervision, has proved to significantly increase generalization performances of models trained on few shots. The objective of this paper is to introduce one such neural pipeline in the context of micro-capsule image segmentation. Our method leverages the rather simple content of these images so that a trainee network can be mentored by a referee network which has been previously trained on synthetically generated pairs of corrupted/correct region masks. RESULTS: Challenging experimental setups are investigated. They involve from only 3 to 10 annotated images along with moderately large amounts of unannotated images. In a bio-artificial capsule dataset, our approach consistently and drastically improves accuracy. We also show that the learnt referee network is transferable to another Glioblastoma cell dataset and that it can be efficiently coupled with data augmentation strategies. CONCLUSIONS: Experimental results show that very significant accuracy increments are obtained by the proposed pipeline, leading to the conclusion that the self-supervision mechanism introduced in this paper has the potential to replace human annotations.

Genetic engineering of the branched fatty acid metabolic pathway of Bacillus subtilis for the overproduction of surfactin C14 isoform.

Surfactin, a lipopeptide produced by Bacillus subtilis, is one of the most powerful biosurfactants known. This molecule consists of a cyclic heptapeptide linked to a ß-hydroxy fatty acid chain. The isomery and the length of the fatty acid (FA) chain are responsible for the surfactin's activities. In this study, the gene codY, which encode for the global transcriptional regulator and the gene lpdV, located in the bkd operon (lpdV, bkdAA, bkdAB and bkdB genes), which is responsible for the last step of the branched chain amino acid (BCAA) degradation in acyl-CoA were deleted. The influence of these deletions on the quantitative and qualitative surfactin production was analysed. The surfactin production was quantified by RP-HPLC and the surfactin isoforms were characterized using LC-MS-MS and GC-MS analysis. The results obtained in the mutants showed an enhancement of surfactin specific production by a factor of 5.8 for the codY mutant and 1.4 for lpdV mutant. Moreover qualitative analysis of the lpdV mutant reveals that it mainly produced surfactin C14 isoform (2 fold more than the wild type) with linear FA chain. Complete analysis of the extracellular metabolites using 1 H quantitative NMR reveals a reduced production of acetoin in this mutant. This work demonstrates for the first time an original approach to overproduce specifically surfactin with C14 FA chain.

Long-term tracking of budding yeast cells in brightfield microscopy: CellStar and the Evaluation Platform.

With the continuous expansion of single cell biology, the observation of the behaviour of individual cells over extended durations and with high accuracy has become a problem of central importance. Surprisingly, even for yeast cells that have relatively regular shapes, no solution has been proposed that reaches the high quality required for long-term experiments for segmentation and tracking (S&T) based on brightfield images. Here, we present CellStar, a tool chain designed to achieve good performance in long-term experiments. The key features are the use of a new variant of parametrized active rays for segmentation, a neighbourhood-preserving criterion for tracking, and the use of an iterative approach that incrementally improves S&T quality. A graphical user interface enables manual corrections of S&T errors and their use for the automated correction of other, related errors and for parameter learning. We created a benchmark dataset with manually analysed images and compared CellStar with six other tools, showing its high performance, notably in long-term tracking. As a community effort, we set up a website, the Yeast Image Toolkit, with the benchmark and the Evaluation Platform to gather this and additional information provided by others.

Predicting changes of reaction networks with partial kinetic information.

We wish to predict changes of reaction networks with partial kinetic information that lead to target changes of their steady states. The changes may be either increases or decreases of influxes, reaction knockouts, or multiple changes of these two kinds. Our prime applications are knockout prediction tasks for metabolic and regulation networks. In a first step, we propose a formal modeling language for reaction networks with partial kinetic information. The modeling language has a graphical syntax reminiscent to Petri nets. Each reaction in a model comes with a partial description of its kinetics, based on a similarity relation on kinetic functions that we introduce. Such partial descriptions are able to model the regulation of existing metabolic networks for which precise kinetic knowledge is usually not available. In a second step, we develop prediction algorithms that can be applied to any reaction network modeled in our language. These algorithms perform qualitative reasoning based on abstract interpretation, by which the kinetic unknowns are abstracted away. Given a reaction network, abstract interpretation produces a finite domain constraint in a novel class. We show how to solve these finite domain constraints with an existing finite domain constraint solver, and how to interpret the solution sets as predictions of multiple reaction knockouts that lead to a desired change of the steady states. We have implemented the prediction algorithm and integrated it into a prediction tool. This journal article extends the two conference papers John et al. (2013) and Niehren et al. (2015) while adding a new prediction algorithm for multiple gene knockouts. An application to single gene knockout prediction for surfactin overproduction was presented in Coutte et al. (2015). It illustrates the adequacy of the model-based predictions made by our algorithm in the wet lab.

What Population Reveals about Individual Cell Identity: Single-Cell Parameter Estimation of Models of Gene Expression in Yeast.

Significant cell-to-cell heterogeneity is ubiquitously observed in isogenic cell populations. Consequently, parameters of models of intracellular processes, usually fitted to population-averaged data, should rather be fitted to individual cells to obtain a population of models of similar but non-identical individuals. Here, we propose a quantitative modeling framework that attributes specific parameter values to single cells for a standard model of gene expression. We combine high quality single-cell measurements of the response of yeast cells to repeated hyperosmotic shocks and state-of-the-art statistical inference approaches for mixed-effects models to infer multidimensional parameter distributions describing the population, and then derive specific parameters for individual cells. The analysis of single-cell parameters shows that single-cell identity (e.g. gene expression dynamics, cell size, growth rate, mother-daughter relationships) is, at least partially, captured by the parameter values of gene expression models (e.g. rates of transcription, translation and degradation). Our approach shows how to use the rich information contained into longitudinal single-cell data to infer parameters that can faithfully represent single-cell identity.

Modeling leucine's metabolic pathway and knockout prediction improving the production of surfactin, a biosurfactant from Bacillus subtilis.

A Bacillus subtilis mutant strain overexpressing surfactin biosynthetic genes was previously constructed. In order to further increase the production of this biosurfactant, our hypothesis is that the surfactin precursors, especially leucine, must be overproduced. We present a three step approach for leucine overproduction directed by methods from computational biology. Firstly, we develop a new algorithm for gene knockout prediction based on abstract interpretation, which applies to a recent modeling language for reaction networks with partial kinetic information. Secondly, we model the leucine metabolic pathway as a reaction network in this language, and apply the knockout prediction algorithm with the target of leucine overproduction. Out of the 21 reactions corresponding to potential gene knockouts, the prediction algorithm selects 12 reactions. Six knockouts were introduced in B. subtilis 168 derivatives strains to verify their effects on surfactin production. For all generated mutants, the specific surfactin production is increased from 1.6- to 20.9-fold during the exponential growth phase, depending on the medium composition. These results show the effectiveness of the knockout prediction approach based on formal models for metabolic reaction networks with partial kinetic information, and confirms our hypothesis that precursors supply is one of the main parameters to optimize surfactin overproduction.