Using federated data sources and Varian Learning Portal framework to train a neural network model for automatic organ segmentation.

PURPOSE: In this study we trained a deep neural network model for female pelvis organ segmentation using data from several sites without any personal data sharing. The goal was to assess its prediction power compared with the model trained in a centralized manner. METHODS: Varian Learning Portal (VLP) is a distributed machine learning (ML) infrastructure enabling privacy-preserving research across hospitals from different regions or countries, within the framework of a trusted consortium. Such a framework is relevant in the case when there is a high level of trust among the participating sites, but there are legal restrictions which do not allow the actual data sharing between them. We trained an organ segmentation model for the female pelvic region using the synchronous data distributed framework provided by the VLP. RESULTS: The prediction performance of the model trained using the federated framework offered by VLP was on the same level as the performance of the model trained in a centralized manner where all training data was pulled together in one centre. CONCLUSIONS: VLP infrastructure can be used for GPU-based training of a deep neural network for organ segmentation for the female pelvic region. This organ segmentation instance is particularly difficult due to the high variation in the organs' shape and size. Being able to train the model using data from several clinics can help, for instance, by exposing the model to a larger range of data variations. VLP framework enables such a distributed training approach without sharing protected health information.

A graph-theoretical approach for motif discovery in protein sequences.

Motif recognition is a challenging problem in bioinformatics due to the diversity of protein motifs. Many existing algorithms identify motifs of a given length, thus being either not applicable or not efficient when searching simultaneously for motifs of various lengths. Searching for gapped motifs, although very important, is a highly time-consuming task due to the combinatorial explosion of possible combinations implied by the consideration of long gaps. We introduce a new graph theoretical approach to identify motifs of various lengths, both with and without gaps. We compare our approach with two widely used methods: MEME and GLAM2 analyzing both the quality of the results and the required computational time. Our method provides results of a slightly higher level of quality than MEME but at a much faster rate, i.e., one eighth of MEME's query time. By using similarity indexing, we drop the query times down to an average of approximately one sixth of the ones required by GLAM2, while achieving a slightly higher level of quality of the results. More precisely, for sequence collections smaller than 50000 bytes GLAM2 is 13 times slower, while being at least as fast as our method on larger ones. The source code of our C++ implementation is freely available in GitHub: https://github.com/hirvolt1/debruijn-motif.

Synthetic Transcription Amplifier System for Orthogonal Control of Gene Expression in Saccharomyces cerevisiae.

This work describes the development and characterization of a modular synthetic expression system that provides a broad range of adjustable and predictable expression levels in S. cerevisiae. The system works as a fixed-gain transcription amplifier, where the input signal is transferred via a synthetic transcription factor (sTF) onto a synthetic promoter, containing a defined core promoter, generating a transcription output signal. The system activation is based on the bacterial LexA-DNA-binding domain, a set of modified, modular LexA-binding sites and a selection of transcription activation domains. We show both experimentally and computationally that the tuning of the system is achieved through the selection of three separate modules, each of which enables an adjustable output signal: 1) the transcription-activation domain of the sTF, 2) the binding-site modules in the output promoter, and 3) the core promoter modules which define the transcription initiation site in the output promoter. The system has a novel bidirectional architecture that enables generation of compact, yet versatile expression modules for multiple genes with highly diversified expression levels ranging from negligible to very strong using one synthetic transcription factor. In contrast to most existing modular gene expression regulation systems, the present system is independent from externally added compounds. Furthermore, the established system was minimally affected by the several tested growth conditions. These features suggest that it can be highly useful in large scale biotechnology applications.

Biclustering methods: biological relevance and application in gene expression analysis.

DNA microarray technologies are used extensively to profile the expression levels of thousands of genes under various conditions, yielding extremely large data-matrices. Thus, analyzing this information and extracting biologically relevant knowledge becomes a considerable challenge. A classical approach for tackling this challenge is to use clustering (also known as one-way clustering) methods where genes (or respectively samples) are grouped together based on the similarity of their expression profiles across the set of all samples (or respectively genes). An alternative approach is to develop biclustering methods to identify local patterns in the data. These methods extract subgroups of genes that are co-expressed across only a subset of samples and may feature important biological or medical implications. In this study we evaluate 13 biclustering and 2 clustering (k-means and hierarchical) methods. We use several approaches to compare their performance on two real gene expression data sets. For this purpose we apply four evaluation measures in our analysis: (1) we examine how well the considered (bi)clustering methods differentiate various sample types; (2) we evaluate how well the groups of genes discovered by the (bi)clustering methods are annotated with similar Gene Ontology categories; (3) we evaluate the capability of the methods to differentiate genes that are known to be specific to the particular sample types we study and (4) we compare the running time of the algorithms. In the end, we conclude that as long as the samples are well defined and annotated, the contamination of the samples is limited, and the samples are well replicated, biclustering methods such as Plaid and SAMBA are useful for discovering relevant subsets of genes and samples.

Quantitative analysis of the self-assembly strategies of intermediate filaments from tetrameric vimentin.

In vitro assembly of intermediate filaments from tetrameric vimentin consists of a very rapid phase of tetramers laterally associating into unit-length filaments and a slow phase of filament elongation. We focus in this paper on a systematic quantitative investigation of two molecular models for filament assembly, recently proposed in (Kirmse et al. J. Biol. Chem. 282, 52 (2007), 18563-18572), through mathematical modeling, model fitting, and model validation. We analyze the quantitative contribution of each filament elongation strategy: with tetramers, with unit-length filaments, with longer filaments, or combinations thereof. In each case, we discuss the numerical fitting of the model with respect to one set of data, and its separate validation with respect to a second, different set of data. We introduce a high-resolution model for vimentin filament self-assembly, able to capture the detailed dynamics of filaments of arbitrary length. This provides much more predictive power for the model, in comparison to previous models where only the mean length of all filaments in the solution could be analyzed. We show how kinetic observations on low-resolution models can be extrapolated to the high-resolution model and used for lowering its complexity.