In discrete tomography sometimes it is necessary to reduce the number of projections used for reconstructing the image. Earlier, it was shown that the choice of projection angles can significantly influence the quality of the reconstructions. In this study, we apply convolutional neural networks to select projections in order to reconstruct the original images from their sinograms with the smallest possible error. The training of neural networks is generally a time-consuming process, but after the network has been trained, the prediction for a previously unseen input is fast. We trained convolutional neural networks using sinograms as input and the desired, algorithmically determined k-best projections as labels in a supervised setting. We achieved a significantly faster projection selection and only a slight increase in the Relative Mean Error (RME).

Object detection is an image analysis task with a wide range of applications, which is difficult to accomplish with traditional programming. Recent breakthroughs in machine learning have made significant progress in this area. However, these algorithms are generally compatible with traditional pixelated images and cannot be directly applied for pointillist datasets generated by single molecule localization microscopy (SMLM) methods. Here, we have improved the averaging method developed for the analysis of SMLM images of sarcomere structures based on a machine learning object detection algorithm. The ordered structure of sarcomeres allows us to determine the location of the proteins more accurately by superimposing SMLM images of identically assembled proteins. However, the area segmentation process required for averaging can be extremely time-consuming and tedious. In this work, we have automated this process. The developed algorithm not only finds the regions of interest, but also classifies the localizations and identifies the true positive ones. For training, we used simulations to generate large amounts of labelled data. After tuning the neural network’s internal parameters, it could find the localizations associated with the structures we were looking for with high accuracy. We validated our results by comparing them with previous manual evaluations. It has also been proven that the simulations can generate data of sufficient quality for training. Our method is suitable for the identification of other types of structures in SMLM data.

No abstract

Transcription factors (TFs) play an essential role in molecular biology by regulating gene expression. The binding sites of TFs can vary by a large amount and the numerous possible binding locations make their detection a challenging issue. Recently, several machine learning approaches using nucleotide sequence data were applied to classify DNA sequences regarding Transcription Factor Binding Sites (TFBS). We propose a novel training strategy without the traditional 1D nucleotide-based DNA sequence representation by instead using a 2D topological matrix of sub-nucleotide chemical functional groups substantially defining the protein binding ability of DNA fragments. We train convolutional neural networks using this novel Functional Group DNA Representation (FGDR) to solve a TFBS classification task. We compare our results with the efficiency of previous nucleotide-based training approaches and show that learning from an FGDR data sequence has several benefits regarding TFBS classification. Moreover, we reason that learning deep neural networks from the FGDR representation produces competitive results while only introducing a pre-processing conversion step. Finally, we show that employing an ensemble of models from the nucleotide and FGDR representations for network training results in higher classification performance than any of the single input approaches.

No abstract