Search

Supervised machine learning is widely used in biology and deserves more scrutiny. We present a set of community-wide recommendations (DOME) aiming to help establish standards of supervised machine learning validation in biology. Formulated as questions, the DOME recommendations improve the assessment and reproducibility of papers when included as supplementary material. The work of the Machine Learning Focus Group was funded by ELIXIR, the research infrastructure for life-science data. IW was funded by the A*STAR Career Development Award (project no. C210112057) from the Agency for Science, Technology and Research (A*STAR), Singapore. D.F. was supported by Estonian Research Council grants (PRG1095, PSG59 and ERA-NET TRANSCAN-2 (BioEndoCar)); Project No 2014-2020.4.01.16-0271, ELIXIR and the European Regional Development Fund through EXCITE Center of Excellence. S.C.E.T. has received funding from the European Union’s Horizon 2020 research and innovation programme under Marie Skłodowska-Curie Grant agreements No. 778247 and No. 823886, and Italian Ministry of University and Research PRIN 2017 grant 2017483NH8. Peer Reviewed "Article signat per 8 autors més 28 autors/es de l' ELIXIR Machine Learning Focus Group: Emidio Capriotti, Rita Casadio, Salvador Capella-Gutierrez, Davide Cirillo, Alessio Del Conte, Alexandros C. Dimopoulos, Victoria Dominguez Del Angel, Joaquin Dopazo, Piero Fariselli, José Maria Fernández, Florian Huber, Anna Kreshuk, Tom Lenaerts, Pier Luigi Martelli, Arcadi Navarro, Pilib Ó Broin, Janet Piñero, Damiano Piovesan, Martin Reczko, Francesco Ronzano, Venkata Satagopam, Castrense Savojardo, Vojtech Spiwok, Marco Antonio Tangaro, Giacomo Tartari, David Salgado, Alfonso Valencia & Federico Zambelli"

We revise a major important problem in bioinformatics: how to annotate protein sequences in the genomic era and all the solutions that have been described by implementing tools based on labelling methods. In this paper we mainly focus on our own work and the theoretical methods that are popular in the field of biosequence analysis in modern molecular biology. We will also review a recent application from our group that largely improves on the topology prediction of disulfide bonds in proteins from Eukaryotic organisms.

Motivation: Alternative splicing has been suggested as a key mechanism for increasing the the functional landscape of the human genome. In order to detect structural and functional features of alternative protein variants originated from the same gene, a pipeline for annotating all the alternative splicing variants included in the ASPicDB database has been implemented, by integrating different state of the art tools for similarity search and for the prediction of structural and functional features of a protein starting from its residue sequence. Methods: For each one of the 254,195 protein variants coming from 17,142 human genes a first layer of annotation consists in searching with BLAST for similar sequences annotated in the SwissProt data base (rel. 53.0) or endowed with a resolved three dimensional structure in the PDB data base (rel. Apr 09). Moreover, remote homology searches are performed by mapping on the sequence the structural and functional domains collected in the PFAM database (rel. 23.0). To this aim we adopted the pfam_scan.pl program, downloaded from the PFAM ftp site (ftp://ftp.sanger.ac.uk/pub/databases/Pfam/Tools/). The second layer of annotation results form a decision tree integrating several predictive tools developed by the Bologna Biocomputing Group. As a first step, N-terminal signal peptides and C-terminal GPI-anchor propeptides are predicted with Spep (Fariselli et al., 2003) and PredGPI (Pierleoni et al., 2008), respectively. When present, they are cleaved from the protein sequence and the presence and localization of alpha-helical transmembrane domains is predicted with ENSEMBLE (Martelli et al., 2003). Secondary structure and cysteine bonding state are predicted with SecPred (Jacoboni et al., 2000) and CysPres (Martelli et al., 2004), respectively. The subcellular localization of globular proteins is predicted with BaCelLo (Pierleoni et al., 2006). Results: As a result of the first annotation layer, 228,737 protein variants share similarity with a SwissProt sequence with an E-value lower than 10-5. Out of these proteins, 129,828 share also similarity with a sequence included in the PDB database. In these cases, the transfer of the functional and structural annotations by similarity is feasible, at least for the aligned regions. 350,870 PFAM domains are also mapped onto 159,538 protein variants with an E-value lower than 10-5. The second layer of annotation discriminates 28,991 and 1,679 sequences endowed with signal peptides and GPI-anchor propeptides, respectively. 41,594 variants are predicted as transmembrane and, among globular proteins, BaCelLo classifies 69,251 sequences as nuclear, 90,267 as cytoplasmic, 19,514 as mitochondrial and 31,996 as secreted. The structural and functional annotations of the proteins encoded by the transcript variants were added in the ASPicDB database (Castrignanò et al. 2008) and can be browsed by means of a graphical search interface, that also allows to retrieve all the genes whose splicing variants encode proteins with specific structural or functional properties (e.g. PFAM or TM domain, protein type, etc.) (Fig. 1A) or showing differences in specific features (Fig. 1B) . Availability: The ASPicDB, supplemented with the annotations for the protein coding transcripts, is available at http://www.caspur.it/ASPicDB/ Availability: http://www.caspur.it/ASPicDB/

In this paper we evaluate the performance of machine learning methods in the task of predicting the bonding state of cysteines starting from protein sequences. This task is very important and is the first step for the identification of disulfide bonds in proteins. We score the performance of three different approaches, such as: 1) Hidden Support Vector Machines (HSVM) which integrates the SVM predictions with a Hidden Markov Model; 2) SVM-HMM which discriminatively trains models that are isomorphic to a kth-order hidden Markov model; 3) Grammatical-Restrained Hidden Conditional Random Fields (GRHCRF) that we recently introduced. To evaluate the present (and future) methods we built a new non-redundant dataset. We report the performance using indices based on per-cysteine and per-protein scores. Furthermore, we evaluate two different encoding schemes based on sequence profile and position specific scoring matrix (PSSM) as computed with the PSI-BLAST program. The sequence profile consists of a matrix that reports the different residue frequencies for each sequence position as internally scored by PSI-BLAST program, while the PSSM modulates such frequencies with respect to the reference substitution score matrix (BLOSUM62). The evaluation is carried out with different dimensions of the local cysteine environment and using different Markov models. Our results show that when the evolutionary information is encoded with PSSM all the methods perform better than with sequence profile. Finally, among the different methods it appears that GRHCRFs performs slightly better than the others achieving a per protein accuracy of 87% with a correlation coefficient of 0.73. Considering that our dataset does not contain trivial protein cases (only one cysteine per protein) the accuracy achieved is among the best performing reported so far.

Genome annotation is one of the most important issues in the genomic era. The exponential grow rate of newly sequenced genomes and proteomes urges the development of fast and reliable annotation methods, suited to exploit all the information available in curated data bases of protein sequences and structures. To this aim we developed BAR, the Bologna Annotation Resource that is now updated (available at http://microserf.biocomp.unibo.it/bar/). The basic notion is that sequences with high identity value to a counterpart can inherit the same function/s and structure, if available. What is totally new in our analysis is to cluster sequences with the constraint that sequence identity should be equal or higher than 40% on at least 90% of the pairwise alignment length. By this sequences are clustered in sets that can be annotated in terms of function and structure depending on the annotation level of the sequences within the cluster. Our method starts with on all-against-all alignment of all the sequences in a GRID environment. The alignments are then regarded as an undirected graph and after the clustering procedure that constrains both the sequence identity value and the alignment length, all the connected nodes (proteins) collapse into a single group (cluster). A cluster that incorporates a UniProt entry inherits its annotations (GO terms that are statistically validated, PDB structures, SCOP classifications, Pfam families, if available). Clusters can contain distantly related proteins that by this can be annotated with high confidence. Ultimately the method analyses a total of over 12 million protein sequences taken from 988 genomes and UniProt release 13. In this version HMM models of those clusters that contain PDB templates are also provided to the end-user for computing structural models of distantly related sequences.

The International Laboratory of BioInformatics (LIBI - www.libi.it) aims at setting up of an advanced Bioinformatics and Computational Biology Laboratory, focusing on the central activities of basic and applied research in modern Biology and Biotechnologies. The main activities include the construction and the maintenance of general genomic, proteomic and transcriptomic databases (e.g. ENSEMBL) as well as specialized databases developed by LIBI partners (e.g. MitoRes, UTRdb, UTRSite, ASPicDB, CSTdb, etc.); the design and implementation of new algorithms and software for the analysis of genomes and their expression products and for the prediction of the structure of proteins; the development of a Grid Problem Solving Environment in order to exploit the distributed computational infrastructure.

We describe a modeling procedure that allowed computing the three dimensional low-resolution folding (3D) of a set of very important human membrane proteins, mitochondrial porins, or voltage-dependent anion selective channels (VDACs) whose structure is not yet known with atomic resolution. Our procedure is a somewhat modified threading procedure that takes advantage of aligning the predicted membrane domains starting from the protein sequence with a suited bacterial template of undetectable sequence homology with the target and of the rather conserved architecture of the membrane protein type known as beta-barrel. By this it was possible to overcome the absence of homologous sequences with known atomic structure in the PDB data base and compute a set of models that are good enough to fit most of the existing experimental data. Our results add to the structural data base, and allow the possibility to have computed models carefully built to design experiments of site directed mutagenesis and/or adopt the models for interpreting phenomelogical-biophysical characterization of transmembrane transport. The repository of VDAC models is available at http://gpcr2.biocomp.unibo.it/~vdac/.

The annotation of the subcellular localization is a major step in protein functional annotation. This is particularly important in eukaryotic cells, which contain several subcellular compartments enclosed by membranes hosting different relevant functions. Methods are available for the ‘ab initio’ prediction of the subcellular localization of globular proteins, relying both on the explicit search of sorting signals in the residue sequence, or on the capture of intrinsic features coded in the aminoacidic composition. We developed BaCelLo, an SVM-based predictor that outperforms other available methods when the major localizations are discriminated: extracytoplasmic space, cytoplasm, nucleus, mitochondrion and chloroplast. A common way for targeting globular proteins to the membrane surface is the attachment with a lipid anchor. The most common and studied lipid anchor modification is the glycosylphosphatidylinositol (GPI) linkage to the C-terminal residue, upon the cleavage of a 20-30 residue long peptide. We developed GPIPE, a method based on HMMs and SVMs that is able to accurately discriminate GPI-anchored proteins in a proteome and to determine the position of the cleavage site.

In this work we present an integrated platform for large-scale eukaryotic genome annotation based on the prediction of subcellular localization, GPI-anchor prediction and membrane protein discrimination into inner and outer classes. Large scale proteomic projects have determined a huge number of aminoacidic sequences whose functions are, in the largest part, still unknown. In eukaryotes compartmentalization plays a major role in intracellular biochemical pathways. However the determination of subcellular localization with experimental high-throughput procedures is a difficult task and computational procedures are needed. We developed BaCelLo (1), a predictor for five classes of subcellular localizations (secretory pathway, cytoplasm, nucleus, mitochondrion and chloroplast) that is based on different SVMs organized in a decision tree. The system exploits the information derived from the aminoacidic sequence and from the evolutionary information contained in alignment profiles. It analyzes the whole sequence composition and the compositions of both the N- and C-termini. The training set is curated in order to avoid redundancy. For the first time a balancing procedure is introduced in order to mitigate the effect of biased training sets. Three kingdom-specific predictors are implemented: for animals, plants and fungi, respectively. When distributing the proteins from animals and fungi into four classes, accuracy of BaCelLo reach 74% and 76%, respectively; a score of 67% is obtained when proteins from plants are distributed into five classes. BaCelLo outperforms the other presently available methods for the same task and gives more balanced accuracy and coverage values for each class. BaCelLo is also described in Nature Protocols, in the Bioinformatics section (2) BaCelLo can be accessed at http://www.biocomp.unibo.it/bacello/. BaCelLo is currently under integration in a workflow which will allow GO functional integration, prediction of GPI-anchors and discrimination between inner and outer membrane proteins. The workflow will be tested on large-scale genome annotation. With a suite of machine learning based methods, developed in house (BaCelLo, SpepLip (3) and ENSEMBLE (4)), we presently built eSLDB (eukaryotic Subcellular Localization DataBase) (5) an online database collecting the annotations of subcellular localization of eukaryotic proteomes. So far five proteomes have been processed and stored: Homo sapiens, Mus musculus, Caenorhabditis elegans, Saccharomyces cerevisiae and Arabidospis thaliana. For each sequence, the database lists localization obtained adopting three different approaches: 1) experimentally determined (when available); 2) homology based (when possible); 3) predicted. All the data are available at the website and can be searched by sequence, by protein code and/or by protein description. Furthermore a more complex search can be performed combining different search fields and keys. All the data contained in the database can be freely downloaded in flat file format. The Database is available at: http://gpcr.biocomp.unibo.it/esldb/.

La Bioinformatica è una disciplina "nuova" tanto che la parola stessa è comparsa nella letteratura scientifica alla fine degli anni ottanta. Essa cerca di "ricavare informazione" dalla montagna di dati biologici che si stanno accumulando con la serie di ambiziosi e grandiosi progetti di ricerca in corso di realizzazione (vedi il progetto Genoma e quelli sulla Genomica Strutturale e la Proteomica). L'espressione "ricavare informazione" è volutamente generica e generale perché l'informazione a cui si fa riferimento può coprire, ed in effetti copre, molte se non tutte le aree di ricerca delle scienze della vita. La Bioinformatica necessita di un approccio multidisciplinare in grado di sfruttare le competenze ed esperienze di biologi molecolari, biochimici, genetisti, informatici, chimici, fisici ed ingegneri.

Mediterranean civilizations and the Atlantic bluefin tuna (Thunnus thynnus) have shared a dramatic history spanning millennia. Today, traditional fishing techniques are fading into the past and industrial fisheries command the seas. Analyzing tuna bones recovered from archeological excavations throughout Europe, we aim to reconstruct the species’ adaptive responses to fisheries pressure and climate change. Using novel molecular techniques, DNA has been extracted from bluefin tuna vertebrae excavated from late iron age and ancient roman settlements in coastal Iberia (4th-­-2nd century BC) and Byzantine-­-era Constantinople (4th-­-15th century AD), as well as vertebrae from the Massimo Sella archive located at the University of Bologna (Ionian, Tyrrhenian and Adriatic Seas, early 20th century). The species identity of all samples has been verified using newly developed mini-­- barcodes targeting mitochondrial and nuclear markers. Samples have been quantified and analyzed using multiple technique (qPCR, Agilent BioAnalyzer), in order to correlate Atlantic bluefin tuna DNA content with samples’ age and diagenetic processes. Suitable samples were selected for subsequent high-­- throughput sequencing on Ion Torrent platform to assess temporal genomic changes induced by environmental and natural history drivers. All historical samples, along with modern ones collected from the same geographic areas, have been genotyped with a dedicated SNP panel, which is composed of 76 SNPs with high similarity to a wide variety of annotated genes and additional 20 SNPs that provide significant discrimination between modern populations. These efforts will provide new insights into the spatial dynamics, population structure and genetic evolution of the species over the past 2000 years.