Machine learning classifiers predict key genomic and evolutionary traits across the kingdoms of life

MICROPHERRET: MICRObial PHEnotypic tRait ClassifieR using Machine lEarning Techniques

BACKGROUND

In recent years, there has been a rapid increase in the number of microbial genomes reconstructed through shotgun sequencing, and obtained by newly developed approaches including metagenomic binning and single-cell sequencing. However, our ability to functionally characterize these genomes by experimental assays is orders of magnitude less efficient. Consequently, there is a pressing need for the development of swift and automated strategies for the functional classification of microbial genomes.

RESULTS

The present work leverages a suite of supervised machine learning algorithms to establish a range of 86 metabolic and other ecological functions, such as methanotrophy and plastic degradation, starting from widely obtainable microbial genome annotations. Tests performed on independent datasets demonstrated robust performance across complete, fragmented, and incomplete genomes above a 70% completeness level for most of the considered functions. Application of the algorithms to the Biogas Microbiome database yielded predictions broadly consistent with current biological knowledge and correctly detecting functionally-related nuances of archaeal genomes. Finally, a case study focused on acetoclastic methanogenesis demonstrated how the developed machine learning models can be refined or expanded with models describing novel functions of interest.

CONCLUSIONS

The resulting tool, MICROPHERRET, incorporates a total of 86 models, one for each tested functional class, and can be applied to high-quality microbial genomes as well as to low-quality genomes derived from metagenomics and single-cell sequencing. MICROPHERRET can thus aid in understanding the functional role of newly generated genomes within their micro-ecological context.

Identifying biomarkers related to motor function in chronic stroke: A fNIRS and TMS study

BACKGROUND

Upper limb motor impairment commonly occurs after stroke, impairing quality of life. Brain network reorganization likely differs between subgroups with differing impairment severity. This study explored differences in functional connectivity (FC) and corticospinal tract (CST) integrity between patients with mild/moderate versus severe hemiplegia poststroke to clarify the neural correlates underlying motor deficits.

METHOD

Sixty chronic stroke patients with upper limb motor impairment were categorized into mild/moderate and severe groups based on Fugl-Meyer scores. Resting-state FC was assessed using functional near-infrared spectroscopy (fNIRS) to compare connectivity patterns between groups across motor regions. CST integrity was evaluated by inducing motor evoked potentials (MEP) via transcranial magnetic stimulation.

RESULTS

Compared to the mild/moderate group, the severe group exhibited heightened premotor cortex-primary motor cortex (PMC-M1) connectivity (t = 4.56, p < 0.01). Absence of MEP was also more frequent in the severe group (χ2 = 12.31, p = 0.01). Bayesian models effectively distinguished subgroups and identified the PMC-M1 connection as highly contributory (accuracy = 91.30%, area under the receiver operating characteristic curve [AUC] = 0.86).

CONCLUSION

Distinct patterns of connectivity and corticospinal integrity exist between stroke subgroups with differing impairments. Strengthened connectivity potentially indicates recruitment of additional motor resources to compensate for damage. These findings elucidate the neural correlates underlying motor deficits poststroke and could guide personalized, network-based therapies targeting predictive biomarkers to improve rehabilitation outcomes.

In Silico Analysis of the Phylogenetic and Physiological Characteristics of Sphingobium indicum B90A: A Hexachlorocyclohexane-Degrading Bacterium

The study focuses on the in silico genomic characterization of Sphingobium indicum B90A, revealing a wealth of genes involved in stress response, carbon monoxide oxidation, β-carotene biosynthesis, heavy metal resistance, and aromatic compound degradation, suggesting its potential as a bioremediation agent. Furthermore, genomic adaptations among nine Sphingomonad strains were explored, highlighting shared core genes via pangenome analysis, including those related to the shikimate pathway and heavy metal resistance. The majority of genes associated with aromatic compound degradation, heavy metal resistance, and stress response were found within genomic islands across all strains. Sphingobium indicum UT26S exhibited the highest number of genomic islands, while Sphingopyxis alaskensis RB2256 had the maximum fraction of its genome covered by genomic islands. The distribution of lin genes varied among the strains, indicating diverse genetic responses to environmental pressures. Additionally, in silico evidence of horizontal gene transfer (HGT) between plasmids pSRL3 and pISP3 of the Sphingobium and Sphingomonas genera, respectively, has been provided. The manuscript offers novel insights into strain B90A, highlighting its role in horizontal gene transfer and refining evolutionary relationships among Sphingomonad strains. The discovery of stress response genes and the czcABCD operon emphasizes the potential of Sphingomonads in consortia development, supported by genomic island analysis.

Flipping the script: Understanding riboswitches from an alternative perspective

Riboswitches are broadly distributed regulatory elements most frequently found in the 5'-leader sequence of bacterial mRNAs that regulate gene expression in response to the binding of a small molecule effector. The occupancy status of the ligand-binding aptamer domain manipulates downstream information in the message that instructs the expression machinery. Currently, there are over 55 validated riboswitch classes, where each class is defined based on the identity of the ligand it binds and/or sequence and structure conservation patterns within the aptamer domain. This classification reflects an "aptamer-centric" perspective that dominates our understanding of riboswitches. In this review, we propose a conceptual framework that groups riboswitches based on the mechanism by which RNA manipulates information directly instructing the expression machinery. This scheme does not replace the established aptamer domain-based classification of riboswitches but rather serves to facilitate hypothesis-driven investigation of riboswitch regulatory mechanisms. Based on current bioinformatic, structural, and biochemical studies of a broad spectrum of riboswitches, we propose three major mechanistic groups: (1) "direct occlusion", (2) "interdomain docking", and (3) "strand exchange". We discuss the defining features of each group, present representative examples of riboswitches from each group, and illustrate how these RNAs couple small molecule binding to gene regulation. While mechanistic studies of the occlusion and docking groups have yielded compelling models for how these riboswitches function, much less is known about strand exchange processes. To conclude, we outline the limitations of our mechanism-based conceptual framework and discuss how critical information within riboswitch expression platforms can inform gene regulation.

cdsBERT - Extending Protein Language Models with Codon Awareness

Recent advancements in Protein Language Models (pLMs) have enabled high-throughput analysis of proteins through primary sequence alone. At the same time, newfound evidence illustrates that codon usage bias is remarkably predictive and can even change the final structure of a protein. Here, we explore these findings by extending the traditional vocabulary of pLMs from amino acids to codons to encapsulate more information inside CoDing Sequences (CDS). We build upon traditional transfer learning techniques with a novel pipeline of token embedding matrix seeding, masked language modeling, and student-teacher knowledge distillation, called MELD. This transformed the pretrained ProtBERT into cdsBERT; a pLM with a codon vocabulary trained on a massive corpus of CDS. Interestingly, cdsBERT variants produced a highly biochemically relevant latent space, outperforming their amino acid-based counterparts on enzyme commission number prediction. Further analysis revealed that synonymous codon token embeddings moved distinctly in the embedding space, showcasing unique additions of information across broad phylogeny inside these traditionally "silent" mutations. This embedding movement correlated significantly with average usage bias across phylogeny. Future fine-tuned organism-specific codon pLMs may potentially have a more significant increase in codon usage fidelity. This work enables an exciting potential in using the codon vocabulary to improve current state-of-the-art structure and function prediction that necessitates the creation of a codon pLM foundation model alongside the addition of high-quality CDS to large-scale protein databases.