Understanding inherent image features in CNN-based assessment of diabetic retinopathy.

Diabetic retinopathy (DR) is a leading cause of blindness and affects millions of people throughout the world. Early detection and timely checkups are key to reduce the risk of blindness. Automated grading of DR is a cost-effective way to ensure early detection and timely checkups. Deep learning or more specifically convolutional neural network (CNN)-based methods produce state-of-the-art performance in DR detection. Whilst CNN based methods have been proposed, no comparisons have been done between the extracted image features and their clinical relevance. Here we first adopt a CNN visualization strategy to discover the inherent image features involved in the CNN's decision-making process. Then, we critically analyze those features with respect to commonly known pathologies namely microaneurysms, hemorrhages and exudates, and other ocular components. We also critically analyze different CNNs by considering what image features they pick up during learning to predict and justify their clinical relevance. The experiments are executed on publicly available fundus datasets (EyePACS and DIARETDB1) achieving an accuracy of 89 ~ 95% with AUC, sensitivity and specificity of respectively 95 ~ 98%, 74 ~ 86%, and 93 ~ 97%, for disease level grading of DR. Whilst different CNNs produce consistent classification results, the rate of picked-up image features disagreement between models could be as high as 70%.

The burden of disease of three food-associated heavy metals in clusters in the Danish population - Towards targeted public health strategies.

Lifestyle and sociodemographics are likely to influence dietary patterns, and, as a result, human exposure to chemical contaminants in foods and their associated health impact. We aimed to characterize subgroups of the Danish population based on diet and sociodemographic indicators, and identify those bearing a higher disease burden due to exposure to methylmercury (MeHg), cadmium (Cd) and inorganic arsenic (i-As). We collected dietary, lifestyle, and sociodemographic data on the occurrence of chemical contaminants in foods from Danish surveys. We grouped participants according to similarities in diet, lifestyle, and sociodemographics using Self-Organizing Maps (SOM), and estimated disease burden in disability-adjusted life years (DALY). SOM clustering resulted in 12 population groups with distinct characteristics. Exposure to contaminants varied between clusters and was largely driven by intake of fish, seafood and cereal products. Five clusters had an estimated annual burden >20 DALY/100,000. The cluster with the highest burden had a high proportion of women of childbearing age, with most of the burden attributed to MeHg. Individuals belonging to the top three clusters had higher education and physical activity, were mainly non-smokers and lived in urban areas. Our findings may facilitate the development of preventive strategies targeted to the most affected subgroups.

Legacy data sharing to improve drug safety assessment: the eTOX project.

The sharing of legacy preclinical safety data among pharmaceutical companies and its integration with other information sources offers unprecedented opportunities to improve the early assessment of drug safety. Here, we discuss the experience of the eTOX project, which was established through the Innovative Medicines Initiative to explore this possibility.

Cofactory: sequence-based prediction of cofactor specificity of Rossmann folds.

Obtaining optimal cofactor balance to drive production is a challenge in metabolically engineered microbial production strains. To facilitate identification of heterologous enzymes with desirable altered cofactor requirements from native content, we have developed Cofactory, a method for prediction of enzyme cofactor specificity using only primary amino acid sequence information. The algorithm identifies potential cofactor binding Rossmann folds and predicts the specificity for the cofactors FAD(H2), NAD(H), and NADP(H). The Rossmann fold sequence search is carried out using hidden Markov models whereas artificial neural networks are used for specificity prediction. Training was carried out using experimental data from protein-cofactor structure complexes. The overall performance was benchmarked against an independent evaluation set obtaining Matthews correlation coefficients of 0.94, 0.79, and 0.65 for FAD(H2), NAD(H), and NADP(H), respectively. The Cofactory method is made publicly available at http://www.cbs.dtu.dk/services/Cofactory.

Genome-wide assessment of the association of rare and common copy number variations to testicular germ cell cancer.

Testicular germ cell cancer (TGCC) is one of the most heritable forms of cancer. Previous genome-wide association studies have focused on single nucleotide polymorphisms, largely ignoring the influence of copy number variants (CNVs). Here we present a genome-wide study of CNV on a cohort of 212 cases and 437 controls from Denmark, which was genotyped at ∼1.8 million markers, half of which were non-polymorphic copy number markers. No association of common variants were found, whereas analysis of rare variants (present in less than 1% of the samples) initially indicated a single gene with significantly higher accumulation of rare CNVs in cases as compared to controls, at the gene PTPN1 (P = 3.8 × 10(-2), 0.9% of cases and 0% of controls). However, the CNV could not be verified by qPCR in the affected samples. Further, the CNV calling of the array-data was validated by sequencing of the GSTM1 gene, which showed that the CNV frequency was in complete agreement between the two platforms. This study therefore disconfirms the hypothesis that there exists a single CNV locus with a major effect size that predisposes to TGCC. Genome-wide pathway association analysis indicated a weak association of rare CNVs related to cell migration (false-discovery rate = 0.021, 1.8% of cases and 1.1% of controls). Dysregulation during migration of primordial germ cells has previously been suspected to be a part of TGCC development and this set of multiple rare variants may thereby have a minor contribution to an increased susceptibility of TGCCs.

Analysis and prediction of leucine-rich nuclear export signals.

We present a thorough analysis of nuclear export signals and a prediction server, which we have made publicly available. The machine learning prediction method is a significant improvement over the generally used consensus patterns. Nuclear export signals (NESs) are extremely important regulators of the subcellular location of proteins. This regulation has an impact on transcription and other nuclear processes, which are fundamental to the viability of the cell. NESs are studied in relation to cancer, the cell cycle, cell differentiation and other important aspects of molecular biology. Our conclusion from this analysis is that the most important properties of NESs are accessibility and flexibility allowing relevant proteins to interact with the signal. Furthermore, we show that not only the known hydrophobic residues are important in defining a nuclear export signals. We employ both neural networks and hidden Markov models in the prediction algorithm and verify the method on the most recently discovered NESs. The NES predictor (NetNES) is made available for general use at http://www.cbs.dtu.dk/.

Improved prediction of signal peptides: SignalP 3.0.

We describe improvements of the currently most popular method for prediction of classically secreted proteins, SignalP. SignalP consists of two different predictors based on neural network and hidden Markov model algorithms, where both components have been updated. Motivated by the idea that the cleavage site position and the amino acid composition of the signal peptide are correlated, new features have been included as input to the neural network. This addition, combined with a thorough error-correction of a new data set, have improved the performance of the predictor significantly over SignalP version 2. In version 3, correctness of the cleavage site predictions has increased notably for all three organism groups, eukaryotes, Gram-negative and Gram-positive bacteria. The accuracy of cleavage site prediction has increased in the range 6-17% over the previous version, whereas the signal peptide discrimination improvement is mainly due to the elimination of false-positive predictions, as well as the introduction of a new discrimination score for the neural network. The new method has been benchmarked against other available methods. Predictions can be made at the publicly available web server

Definition of supertypes for HLA molecules using clustering of specificity matrices.

Major histocompatibility complex (MHC) proteins are encoded by extremely polymorphic genes and play a crucial role in immunity. However, not all genetically different MHC molecules are functionally different. Sette and Sidney (1999) have defined nine HLA class I supertypes and showed that with only nine main functional binding specificities it is possible to cover the binding properties of almost all known HLA class I molecules. Here we present a comprehensive study of the functional relationship between all HLA molecules with known specificities in a uniform and automated way. We have developed a novel method for clustering sequence motifs. We construct hidden Markov models for HLA class I molecules using a Gibbs sampling procedure and use the similarities among these to define clusters of specificities. These clusters are extensions of the previously suggested ones. We suggest splitting some of the alleles in the A1 supertype into a new A26 supertype, and some of the alleles in the B27 supertype into a new B39 supertype. Furthermore the B8 alleles may define their own supertype. We also use the published specificities for a number of HLA-DR types to define clusters with similar specificities. We report that the previously observed specificities of these class II molecules can be clustered into nine classes, which only partly correspond to the serological classification. We show that classification of HLA molecules may be done in a uniform and automated way. The definition of clusters allows for selection of representative HLA molecules that can cover the HLA specificity space better. This makes it possible to target most of the known HLA alleles with known specificities using only a few peptides, and may be used in construction of vaccines. Supplementary material is available at http://www.cbs.dtu.dk/researchgroups/immunology/supertypes.html.

Reliable prediction of T-cell epitopes using neural networks with novel sequence representations.

In this paper we describe an improved neural network method to predict T-cell class I epitopes. A novel input representation has been developed consisting of a combination of sparse encoding, Blosum encoding, and input derived from hidden Markov models. We demonstrate that the combination of several neural networks derived using different sequence-encoding schemes has a performance superior to neural networks derived using a single sequence-encoding scheme. The new method is shown to have a performance that is substantially higher than that of other methods. By use of mutual information calculations we show that peptides that bind to the HLA A*0204 complex display signal of higher order sequence correlations. Neural networks are ideally suited to integrate such higher order correlations when predicting the binding affinity. It is this feature combined with the use of several neural networks derived from different and novel sequence-encoding schemes and the ability of the neural network to be trained on data consisting of continuous binding affinities that gives the new method an improved performance. The difference in predictive performance between the neural network methods and that of the matrix-driven methods is found to be most significant for peptides that bind strongly to the HLA molecule, confirming that the signal of higher order sequence correlation is most strongly present in high-binding peptides. Finally, we use the method to predict T-cell epitopes for the genome of hepatitis C virus and discuss possible applications of the prediction method to guide the process of rational vaccine design.

Sigma A recognition sites in the Bacillus subtilis genome.

A hidden Markov model of sigma(A) RNA polymerase cofactor recognition sites in Bacillus subtilis, containing either the common or the extended -10 motifs, has been constructed based on experimentally verified sigma(A) recognition sites. This work suggests that more information exists at the initiation site of transcription in both types of promoters than previously thought. When tested on the entire B. subtilis genome, the model predicts that approximately half of the sigma(A) recognition sites are of the extended type. Some of the response-regulator aspartate phosphatases were among the predictions of promoters containing extended sites. The expression of rapA and rapB was confirmed by site-directed mutagenesis to depend on the extended -10 region.