Deep Learning (DL) has recently enabled unprecedented advances in one of the grand challenges in computational biology: the half-century-old problem of protein structure prediction. In this paper we discuss recent advances, limitations, and future perspectives of DL on five broad areas: protein structure prediction, protein function prediction, genome engineering, systems biology and data integration, and phylogenetic inference. We discuss each application area and cover the main bottlenecks of DL approaches, such as training data, problem scope, and the ability to leverage existing DL architectures in new contexts. To conclude, we provide a summary of the subject-specific and general challenges for DL across the biosciences.

Received ...; reviews returned ...; accepted ...Abstract.---The multispecies coalescent (MSC) has emerged as a powerful and desirable framework for species tree inference in phylogenomic studies. Under this framework, the data for each locus is assumed to consist of orthologous, single-copy genes, and heterogeneity across loci is assumed to be due to incomplete lineage sorting (ILS). These assumptions have led biologists that use ILS-aware inference methods, whether based directly on the MSC or proven to be statistically consistent under it (collectively referred to here as MSC-based methods), to exclude all loci that are present in more than a single copy in any of the studied genomes. Furthermore, such analyses entail orthology assignment to avoid the potential of hidden paralogy in the data. The question we seek to answer in this study is: What happens if one runs such species tree inference methods on data where paralogy is present, in addition to or without ILS being present? Through simulation studies and analyses of two biological data sets, we show that running such methods on data with paralogs provide very accurate results, either by treating all gene copies within a family as alleles from multiple individuals or by randomly selecting one copy per species. Our results have significant implications for the use of MSC-based phylogenomic analyses, demonstrating that they do not have to be restricted to single-copy loci, thus greatly increasing the amount of data that can be used. [Multispecies coalescent; incomplete lineage sorting; gene duplication and loss; orthology; paralogy.]

Many recent phylogenetic methods have focused on accurately inferring species trees when there is gene tree discordance due to incomplete lineage sorting (ILS). For almost all of these methods, and for phylogenetic methods in general, the data for each locus is assumed to consist of orthologous, single-copy sequences. Loci that are present in more than a single copy in any of the studied genomes are excluded from the data. These steps greatly reduce the number of loci available for analysis. The question we seek to answer in this study is: What happens if one runs such species tree inference methods on data where paralogy is present, in addition to or without ILS being present? Through simulation studies and analyses of two large biological data sets, we show that running such methods on data with paralogs can still provide accurate results. We use multiple different methods, some of which are based directly on the multispecies coalescent (MSC) model, and some of which have been proven to be statistically consistent under it. We also treat the paralogous loci in multiple ways: from explicitly denoting them as paralogs, to randomly selecting one copy per species. In all cases the inferred species trees are as accurate as equivalent analyses using single-copy orthologs. Our results have significant implications for the use of ILS-aware phylogenomic analyses, demonstrating that they do not have to be restricted to single-copy loci. This will greatly increase the amount of data that can be used for phylogenetic inference.

Phylogenetic networks extend trees to enable simultaneous modeling of both vertical and horizontal evolutionary processes. PhyloNet is a software package that has been under constant development for over 10 years and includes a wide array of functionalities for inferring and analyzing phylogenetic networks. These functionalities differ in terms of the input data they require, the criteria and models they employ, and the types of information they allow to infer about the networks beyond their topologies. Furthermore, PhyloNet includes functionalities for simulating synthetic data on phylogenetic networks, quantifying the topological differences between phylogenetic networks, and evaluating evolutionary hypotheses given in the form of phylogenetic networks.In this paper, we use a simulated data set to illustrate the use of several of PhyloNet's functionalities and make recommendations on how to analyze data sets and interpret the results when using these functionalities. All inference methods that we illustrate are incomplete lineage sorting (ILS) aware; that is, they account for the potential of ILS in the data while inferring the phylogenetic network. While the models do not include gene duplication and loss, we discuss how the methods can be used to analyze data in the presence of polyploidy.The concept of species is irrelevant for the computational analyses enabled by PhyloNet in that species-individuals mappings are user-defined. Consequently, none of the functionalities in PhyloNet deals with the task of species delimitation. In this sense, the data being analyzed could come from different individuals within a single species, in which case population structure along with potential gene flow is inferred (assuming the data has sufficient signal), or from different individuals sampled from different species, in which case the species phylogeny is being inferred. * nakhleh@rice.edu P(D|Ψ) = -1443 P(D|Ψ) = -1439 P(D|Ψ) = -1454 P(D|Ψ) = -1435 P(D|Ψ) = -1435 P(D|Ψ) = -1439 P(D|Ψ) = -1435 P(D|Ψ) = -1439 P(D|Ψ) = -1434 P(D|Ψ) = -1443 P(D|Ψ) = -1439 P(D|Ψ) = -1454 P(D|Ψ) = -14547 posterior distribution over all the possible trees. As Fig 3(b) shows, the random walk visits the different trees, but "circulates" among three topologies more than the other 12 topologies. The posteriors of the trees visited are plotted as in Fig. 3(c), and then the marginal probability distribution on tree topologies is summarized from the samples, as in Fig. 3(d). While one could return the tree (D,((A,E),(B,C))) as the one having the highest marginal probability (the yellow bar is the highest in Fig. 3(d)), the power of Bayesian analysis is that the totality of the results shown in the figure provide more information than just the optimal point. For example, the results in Fig. 3(d) show that the confidence level does not exceed 40% for any of the possible 15 trees. Furthermore, it shows that two trees have comparable probabilities, while a third one is close to them. Back to network inference, any search strategy begins with one or more s...

No abstract