Simply Mycoplasma: The bacterium Mycoplasma pneumoniae , a human pathogen, has a genome of reduced size and is one of the simplest organisms that can reproduce outside of host cells. As such, it represents an excellent model organism in which to attempt a systems-level understanding of its biological organization. Now three papers provide a comprehensive and quantitative analysis of the proteome, the metabolic network, and the transcriptome of M. pneumoniae (see the Perspective by Ochman and Raghavan ). Anticipating what might be possible in the future for more complex organisms, Kühner et al. (p. 1235 ) combine analysis of protein interactions by mass spectrometry with extensive structural information on M. pneumoniae proteins to reveal how proteins work together as molecular machines and map their organization within the cell by electron tomography. The manageable genome size of M. pneumoniae allowed Yus et al. (p. 1263 ) to map the metabolic network of the organism manually and validate it experimentally. Analysis of the network aided development of a minimal medium in which the bacterium could be cultured. Finally, G‡ell et al. (p. 1268 ) applied state-of-the-art sequencing techniques to reveal that this “simple” organism makes extensive use of noncoding RNAs and has exon- and intron-like structure within transcriptional operons that allows complex gene regulation resembling that of eukaryotes.

Information leakage is an increasingly important topic in machine learning research for biomedical applications. When information leakage happens during a model’s training, it risks memorizing the training data instead of learning generalizable properties. This can lead to inflated performance metrics that do not reflect the actual performance at inference time. We present DataSAIL, a versatile Python package to facilitate leakage-reduced data splitting to enable realistic evaluation of machine learning models for biological data that are intended to be applied in out-of-distribution scenarios. DataSAIL is based on formulating the problem to find leakage-reduced data splits as a combinatorial optimization problem. We prove that this problem is NP-hard and provide a scalable heuristic based on clustering and integer linear programming. Finally, we empirically demonstrate DataSAIL’s impact on evaluating biomedical machine learning models.

The increasing volume of genomic data opens new possibilities for analysis of protein function. We introduce a method for automated selection of residues that determine the functional specificity of proteins with a common general function (the specificity‐determining positions [SDP] prediction method). Such residues are assumed to be conserved within groups of orthologs (that may be assumed to have the same specificity) and to vary between paralogs. Thus, considering a multiple sequence alignment of a protein family divided into orthologous groups, one can select positions where the distribution of amino acids correlates with this division. Unlike previously published techniques, the introduced method directly takes into account nonuniformity of amino acid substitution frequencies. In addition, it does not require setting arbitrary thresholds. Instead, a formal procedure for threshold selection using the Bernoulli estimator is implemented. We tested the SDP prediction method on the LacI family of bacterial transcription factors and a sample of bacterial water and glycerol transporters belonging to the major intrinsic protein (MIP) family. In both cases, the comparison with available experimental and structural data strongly supported our predictions.

SDPpred (Specificity Determining Position prediction) is a tool for prediction of residues in protein sequences that determine the proteins' functional specificity. It is designed for analysis of protein families whose members have biochemically similar but not identical interaction partners (e.g. different substrates for a family of transporters). SDPpred predicts residues that could be responsible for the proteins' choice of their correct interaction partners. The input of SDPpred is a multiple alignment of a protein family divided into a number of specificity groups, within which the interaction partner is believed to be the same. SDPpred does not require information about the secondary or three-dimensional structure of proteins. It produces a set of the alignment positions (specificity determining positions) that determine differences in functional specificity. SDPpred is available at http://math.genebee.msu.ru/~psn/.