Unsupervised classification to improve the quality of a bird song recording dataset

Open audio databases such as Xeno-Canto are widely used to build datasets to explore bird song repertoire or to train models for automatic bird sound classification by deep learning algorithms. However, such databases suffer from the fact that bird sounds are weakly labelled: a species name is attributed to each audio recording without timestamps that provide the temporal localization of the bird song of interest. Manual annotations can solve this issue, but they are time consuming, expert-dependent, and cannot run on large datasets. Another solution consists in using a labelling function that automatically segments audio recordings before assigning a label to each segmented audio sample. Although labelling functions were introduced to expedite strong label assignment, their classification performance remains mostly unknown. To address this issue and reduce label noise (wrong label assignment) in large bird song datasets, we introduce a data-centric novel labelling function composed of three successive steps: 1) time-frequency sound unit segmentation, 2) feature computation for each sound unit, and 3) classification of each sound unit as bird song or noise with either an unsupervised DBSCAN algorithm or the supervised BirdNET neural network. The labelling function was optimized, validated, and tested on the songs of 44 West-Palearctic common bird species. We first showed that the segmentation of bird songs alone aggregated from 10% to 83% of label noise depending on the species. We also demonstrated that our labelling function was able to significantly reduce the initial label noise present in the dataset by up to a factor of three. Finally, we discuss different opportunities to design suitable labelling functions to build high-quality animal vocalizations with minimum expert annotation effort.     

Segmentation of moving images by the human visual system

 New segments appearing in an image sequence or spontaneously accelerated segments are band limited by the visual system due to a nonperfect tracking of these segments by eye movements. In spite of this band limitation and acceleration of segments, a coarse segmentation (initial segmentation phase) can be performed by the visual system. This is interesting for the development of purely automatic segmentation algorithms for multimedia applications. In this paper the segmentation of the visual system is modelled and used in an automatic coarse initial segmentation. A suitable model for motion processing based on a spectral representation is presented and applied to the segmentation of synthetic and real image sequences with band limited and accelerated moving foreground and background segments. Received: 1 August 1995/Accepted in revised form: 25 February 1997     

Effective pre-processing of long term noisy audio recordings: An aid to clinical monitoring

Nowadays, great attention is devoted to minimizing the discomfort caused by connection of patients to sensors for long-term monitoring of physiological parameters. Hence, the need for contact-less monitoring systems is increasingly recognized in clinical investigation. To this aim, audio signals recorded by ambient microphones are an appealing and increasing field of research: in the biomedical field, application of contact-less audio recording of long duration may concern obstructive apnoea syndrome, preterm newborns in Intensive Care Units, daily monitoring in occupational dysphonia, speech therapy, Parkinson and Alzheimer disease, monitoring of psychiatric and autistic subjects, etc. However, a significant amount of ambient noise is inevitably included in the records.Especially in the case of recordings that take a long time, manual extraction of clinically useful information from a whole record is a time-consuming operator-dependent task, the length of a whole recording (even several hours) being prohibitive both for perceptual analysis made by listening to it and for visual inspection of signal patterns. Moreover, objective measures of signal characteristics may serve clinicians as a common ground for diagnosis. Hence, automatic methods are needed to speed up and objectify the analysis task.The present work describes a new, automatic, fast and reliable method for extracting “voiced candidates” from audio recordings of long duration for both clinical and home applications.To demonstrate its effectiveness, the method is compared to existing software tools commonly used in biomedical applications using synthetic signals.     

An alternative approach to multiple genome comparison

Genome comparison is now a crucial step for genome annotation and identification of regulatory motifs. Genome comparison aims for instance at finding genomic regions either specific to or in one-to-one correspondence between individuals/strains/species. It serves e.g. to pre-annotate a new genome by automatically transferring annotations from a known one. However, efficiency, flexibility and objectives of current methods do not suit the whole spectrum of applications, genome sizes and organizations. Innovative approaches are still needed. Hence, we propose an alternative way of comparing multiple genomes based on segmentation by similarity. In this framework, rather than being formulated as a complex optimization problem, genome comparison is seen as a segmentation question for which a single optimal solution can be found in almost linear time. We apply our method to analyse three strains of a virulent pathogenic bacteria, Ehrlichia ruminantium, and identify 92 new genes. We also find out that a substantial number of genes thought to be strain specific have potential orthologs in the other strains. Our solution is implemented in an efficient program, qod, equipped with a user-friendly interface, and enables the automatic transfer of annotations between compared genomes or contigs (Video in Supplementary Data). Because it somehow disregards the relative order of genomic blocks, qod can handle unfinished genomes, which due to the difficulty of sequencing completion may become an interesting characteristic for the future. Availabilty: http://www.atgc-montpellier.fr/qod.     

Automated Analysis and Reannotation of Subcellular Locations in Confocal Images from the Human Protein Atlas

The Human Protein Atlas contains immunofluorescence images showing subcellular locations for thousands of proteins. These are currently annotated by visual inspection. In this paper, we describe automated approaches to analyze the images and their use to improve annotation. We began by training classifiers to recognize the annotated patterns. By ranking proteins according to the confidence of the classifier, we generated a list of proteins that were strong candidates for reexamination. In parallel, we applied hierarchical clustering to group proteins and identified proteins whose annotations were inconsistent with the remainder of the proteins in their cluster. These proteins were reexamined by the original annotators, and a significant fraction had their annotations changed. The results demonstrate that automated approaches can provide an important complement to visual annotation.     

Automatic EEG arousal detection for sleep apnea syndrome

Electroencephalographic (EEG) arousals are seen in EEG recordings as an awakening response of the human brain. Sleep apnea is a serious sleep disorder. Severe sleep apnea brings about EEG arousals and sleep for patients with sleep apnea syndrome (SAS) is thus frequently interrupted. The number of respiratory-related arousals during the whole night on PSG recordings is directly related to the quality of sleep. Detecting EEG arousals in the PSG record is thus a significant task for clinical diagnosis in sleep medicine. In this paper, a method for automatic detection of EEG arousals in SAS patients was proposed. To effectively detect respiratory-related arousals, threshold values were determined according to pathological events as sleep apnea and electromyogram (EMG). If resumption of ventilation (end of the apnea interval) was detected, much lower thresholds were adopted for detecting EEG arousals, including relatively doubtful arousals. Conversely, threshold was maintained high when pathological events were undetected. The proposed method was applied to polysomnographic (PSG) records of eight patients with SAS and accuracy of EEG arousal detection was verified by comparative visual inspection. Effectiveness of the proposed method in clinical diagnosis was also investigated.     

Evaluation of annotation strategies using an entire genome sequence

MOTIVATION: Genome-wide functional annotation either by manual or automatic means has raised considerable concerns regarding the accuracy of assignments and the reproducibility of methodologies. In addition, a performance evaluation of automated systems that attempt to tackle sequence analyses rapidly and reproducibly is generally missing. In order to quantify the accuracy and reproducibility of function assignments on a genome-wide scale, we have re-annotated the entire genome sequence of Chlamydia trachomatis (serovar D), in a collaborative manner. RESULTS: We have encoded all annotations in a structured format to allow further comparison and data exchange and have used a scale that records the different levels of potential annotation errors according to their propensity to propagate in the database due to transitive function assignments. We conclude that genome annotation may entail a considerable amount of errors, ranging from simple typographical errors to complex sequence analysis problems. The most surprising result of this comparative study is that automatic systems might perform as well as the teams of experts annotating genome sequences.     

A segmentation approach to long duration surface EMG recordings.

The purpose of this study was to develop an automatic segmentation method in order to identify postural surface EMG segments in long-duration recordings. Surface EMG signals were collected from the cervical erector spinae (CES), erector spinae (ES), external oblique (EO), and tibialis anterior (TA) muscles of 11 subjects using a bipolar electrode configuration. Subjects remained seated in a car seat over the 150-min data-collection period. The modified dynamic cumulative sum (MDCS) algorithm was used to automatically segment the surface EMG signals. Signals were rejected by comparison with an exponential mathematical model of the spectrum of a surface EMG signal. The average power ratio computed between two successive retained segments was used to classify segments as postural or surface EMG. The presence of a negative slope of a regression line fitted to the median frequency values of postural surface EMG segments was taken as an indication of fatigue. Alpha level was set at 0.05. The overall classification error rate was 8%, and could be performed in 25 min for a 150-min signal using a custom-built software program written in C (Borland Software Corporation, CA, USA). This error rate could be enhanced by concentrating on the rejection method, which caused most of the misclassification (6%). Furthermore, the elimination of non-postural surface EMG segments by the use of a segmentation approach enabled muscular fatigue to be identified in signals that contained no evidence of fatigue when analysed using traditional methods.     

A Framework for Annotating Human Genome in Disease Context

Identification of gene-disease association is crucial to understanding disease mechanism. A rapid increase in biomedical literatures, led by advances of genome-scale technologies, poses challenge for manually-curated-based annotation databases to characterize gene-disease associations effectively and timely. We propose an automatic method-The Disease Ontology Annotation Framework (DOAF) to provide a comprehensive annotation of the human genome using the computable Disease Ontology (DO), the NCBO Annotator service and NCBI Gene Reference Into Function (GeneRIF). DOAF can keep the resulting knowledgebase current by periodically executing automatic pipeline to re-annotate the human genome using the latest DO and GeneRIF releases at any frequency such as daily or monthly. Further, DOAF provides a computable and programmable environment which enables large-scale and integrative analysis by working with external analytic software or online service platforms. A user-friendly web interface (doa.nubic.northwestern.edu) is implemented to allow users to efficiently query, download, and view disease annotations and the underlying evidences.     

Functional annotation prediction: all for one and one for all

In an era of rapid genome sequencing and high-throughput technology, automatic function prediction for a novel sequence is of utter importance in bioinformatics. While automatic annotation methods based on local alignment searches can be simple and straightforward, they suffer from several drawbacks, including relatively low sensitivity and assignment of incorrect annotations that are not associated with the region of similarity. ProtoNet is a hierarchical organization of the protein sequences in the UniProt database. Although the hierarchy is constructed in an unsupervised automatic manner, it has been shown to be coherent with several biological data sources. We extend the ProtoNet system in order to assign functional annotations automatically. By leveraging on the scaffold of the hierarchical classification, the method is able to overcome some frequent annotation pitfalls.     

Automatic detection of false annotations via binary property clustering

Background  

Computational protein annotation methods occasionally introduce errors. False-positive (FP) errors are annotations that are mistakenly associated with a protein. Such false annotations introduce errors that may spread into databases through similarity with other proteins. Generally, methods used to minimize the chance for FPs result in decreased sensitivity or low throughput. We present a novel protein-clustering method that enables automatic separation of FP from true hits. The method quantifies the biological similarity between pairs of proteins by examining each protein's annotations, and then proceeds by clustering sets of proteins that received similar annotation into biological groups.     

Combined evidence annotation of transposable elements in genome sequences

Transposable elements (TEs) are mobile, repetitive sequences that make up significant fractions of metazoan genomes. Despite their near ubiquity and importance in genome and chromosome biology, most efforts to annotate TEs in genome sequences rely on the results of a single computational program, RepeatMasker. In contrast, recent advances in gene annotation indicate that high-quality gene models can be produced from combining multiple independent sources of computational evidence. To elevate the quality of TE annotations to a level comparable to that of gene models, we have developed a combined evidence-model TE annotation pipeline, analogous to systems used for gene annotation, by integrating results from multiple homology-based and de novo TE identification methods. As proof of principle, we have annotated "TE models" in Drosophila melanogaster Release 4 genomic sequences using the combined computational evidence derived from RepeatMasker, BLASTER, TBLASTX, all-by-all BLASTN, RECON, TE-HMM and the previous Release 3.1 annotation. Our system is designed for use with the Apollo genome annotation tool, allowing automatic results to be curated manually to produce reliable annotations. The euchromatic TE fraction of D. melanogaster is now estimated at 5.3% (cf. 3.86% in Release 3.1), and we found a substantially higher number of TEs (n = 6,013) than previously identified (n = 1,572). Most of the new TEs derive from small fragments of a few hundred nucleotides long and highly abundant families not previously annotated (e.g., INE-1). We also estimated that 518 TE copies (8.6%) are inserted into at least one other TE, forming a nest of elements. The pipeline allows rapid and thorough annotation of even the most complex TE models, including highly deleted and/or nested elements such as those often found in heterochromatic sequences. Our pipeline can be easily adapted to other genome sequences, such as those of the D. melanogaster heterochromatin or other species in the genus Drosophila.     

The time course of segmentation and cue-selectivity in the human visual cortex

Texture discontinuities are a fundamental cue by which the visual system segments objects from their background. The neural mechanisms supporting texture-based segmentation are therefore critical to visual perception and cognition. In the present experiment we employ an EEG source-imaging approach in order to study the time course of texture-based segmentation in the human brain. Visual Evoked Potentials were recorded to four types of stimuli in which periodic temporal modulation of a central 3° figure region could either support figure-ground segmentation, or have identical local texture modulations but not produce changes in global image segmentation. The image discontinuities were defined either by orientation or phase differences across image regions. Evoked responses to these four stimuli were analyzed both at the scalp and on the cortical surface in retinotopic and functional regions-of-interest (ROIs) defined separately using fMRI on a subject-by-subject basis. Texture segmentation (tsVEP: segmenting versus non-segmenting) and cue-specific (csVEP: orientation versus phase) responses exhibited distinctive patterns of activity. Alternations between uniform and segmented images produced highly asymmetric responses that were larger after transitions from the uniform to the segmented state. Texture modulations that signaled the appearance of a figure evoked a pattern of increased activity starting at ～143 ms that was larger in V1 and LOC ROIs, relative to identical modulations that didn't signal figure-ground segmentation. This segmentation-related activity occurred after an initial response phase that did not depend on the global segmentation structure of the image. The two cue types evoked similar tsVEPs up to 230 ms when they differed in the V4 and LOC ROIs. The evolution of the response proceeded largely in the feed-forward direction, with only weak evidence for feedback-related activity.     

Haptic Exploratory Behavior During Object Discrimination: A Novel Automatic Annotation Method

In order to acquire information concerning the geometry and material of handheld objects, people tend to execute stereotypical hand movement patterns called haptic Exploratory Procedures (EPs). Manual annotation of haptic exploration trials with these EPs is a laborious task that is affected by subjectivity, attentional lapses, and viewing angle limitations. In this paper we propose an automatic EP annotation method based on position and orientation data from motion tracking sensors placed on both hands and inside a stimulus. A set of kinematic variables is computed from these data and compared to sets of predefined criteria for each of four EPs. Whenever all criteria for a specific EP are met, it is assumed that that particular hand movement pattern was performed. This method is applied to data from an experiment where blindfolded participants haptically discriminated between objects differing in hardness, roughness, volume, and weight. In order to validate the method, its output is compared to manual annotation based on video recordings of the same trials. Although mean pairwise agreement is less between human-automatic pairs than between human-human pairs (55.7% vs 74.5%), the proposed method performs much better than random annotation (2.4%). Furthermore, each EP is linked to a specific object property for which it is optimal (e.g., Lateral Motion for roughness). We found that the percentage of trials where the expected EP was found does not differ between manual and automatic annotation. For now, this method cannot yet completely replace a manual annotation procedure. However, it could be used as a starting point that can be supplemented by manual annotation.     

A semantic analysis of the annotations of the human genome

The correct interpretation of any biological experiment depends in an essential way on the accuracy and consistency of the existing annotation databases. Such databases are ubiquitous and used by all life scientists in most experiments. However, it is well known that such databases are incomplete and many annotations may also be incorrect. In this paper we describe a technique that can be used to analyze the semantic content of such annotation databases. Our approach is able to extract implicit semantic relationships between genes and functions. This ability allows us to discover novel functions for known genes. This approach is able to identify missing and inaccurate annotations in existing annotation databases, and thus help improve their accuracy. We used our technique to analyze the current annotations of the human genome. From this body of annotations, we were able to predict 212 additional gene-function assignments. A subsequent literature search found that 138 of these gene-functions assignments are supported by existing peer-reviewed papers. An additional 23 assignments have been confirmed in the meantime by the addition of the respective annotations in later releases of the Gene Ontology database. Overall, the 161 confirmed assignments represent 75.95% of the proposed gene-function assignments. Only one of our predictions (0.4%) was contradicted by the existing literature. We could not find any relevant articles for 50 of our predictions (23.58%). The method is independent of the organism and can be used to analyze and improve the quality of the data of any public or private annotation database.