The core difficulty the authors set out to address is the black-box nature of existing predictors. Recent deep learning tools such as deep-Sep demonstrate high sensitivity, but they rely solely on de novo probabilistic classification and offer no mechanism to verify the biological plausibility of their outputs. That limitation makes them prone to high false-positive rates, particularly when sequences share physicochemical features with non-selenoprotein redox families. SelCysML was designed from the ground up to counter this weakness by combining statistical prediction with verifiable homology evidence, so that every call is backed by both a probability score and a functional cross-check against curated reference databases.

To train and evaluate the system, the team compiled a balanced, non-redundant dataset of 30,000 protein sequences, split evenly between 15,000 selenoproteins and 15,000 control proteins. The positive set was drawn from BSepDB, a repository encompassing more than 57,000 selenoprotein entries across 111 families and 16,621 distinct bacterial organisms, spanning 88 clades. Negative controls came from UniProt using two complementary strategies: a multi-domain filtering strategy that retrieved sequences from all three domains of life while strictly excluding any entry annotated with the keyword Selenocysteine, and a bacteria-exclusive control set restricted to the bacterial taxonomic domain. This dual design gave the researchers both a general-purpose model and a bacteria-specific classifier, and the authors note that the reported metrics primarily reflect the model’s ability to distinguish bacterial selenoproteins against a broad-domain background.

Rigorous quality control shaped the dataset before any model training began. The full set was clustered with CD-HIT at a sequence identity cut-off of 0.70 prior to splitting into training and test partitions, guaranteeing that no test sequence shared more than 70 percent identity with a training sequence and preventing data leakage. Non-standard characters were removed, sequences shorter than 40 amino acids were discarded, and negative controls were filtered against high-confidence UniProt and RefSeq proteomes to exclude entries with ambiguous stop codons or anomalous fragment lengths. Notably, selenocysteine residues were normalized to cysteine for feature extraction. Although explicitly encoding selenocysteine would offer higher biochemical fidelity, the authors argue that this U-to-C masking mimics the real-world scenario of automated genomic annotation, in which UGA codons are often prematurely read as stop signals.

The predictive engine rests on two transparent physicochemical descriptors calculated with the protr package in R. Amino Acid Composition yields a 20-dimensional vector representing the frequency of each standard amino acid, while Composition, Transition and Distribution describes how hydrophobicity, charge and polarity are distributed along the sequence. Numerical features were normalized using a median-imputation and scaling strategy. The authors deliberately engineered the tool to operate at the protein sequence level rather than reading nucleotide context such as UGA recoding signals or downstream SECIS elements, which are biologically definitive but often missing from fragmented metagenomic assemblies. That loss of nucleotide context is actively compensated for by the pipeline’s third layer of validation.

Model training followed the tidymodels framework and aggregated four algorithms: XGBoost, Ranger random forests, Kernlab support vector machines and KKNN nearest neighbors. A down-sampling step mitigated class imbalance, and the ensemble combined the probability outputs of the individual algorithms through mean averaging. Performance was validated on an independent test set representing a 30 percent split, with confidence intervals computed over 50 bootstrap iterations, and the classification threshold was optimized to maximize both sensitivity and specificity. On balanced test sets of 2,000 proteins each, the general model reached a sensitivity of 97.3 percent, a specificity of 80.3 percent and a balanced accuracy of 88.8 percent, while the bacteria-specific model achieved a sensitivity of 96.5 percent, a specificity of 87.5 percent, a balanced accuracy of 92.0 percent, an AUC-ROC of 0.977 and an MCC of 0.843, confirming the value of domain-adjusted architectures for prokaryotic screening.

The architecture itself unfolds in three layers. Layer 1 extracts the AAC and CTDC descriptors from raw sequences; Layer 2 runs the ensemble classifier and produces a probability-based verdict of selenoprotein versus standard protein; and Layer 3 performs evidential cross-validation, using a dual-BLAST strategy against local, standalone replicas of BSepDB and UniProt that ship with the pipeline to ensure fully offline execution and reproducibility. This third layer acts as an automated auditor of the machine learning output, assigning each prediction an evidential status of Validated, Potential False Positive or Potential False Negative and calculating an adjusted confidence index. A penalty factor of 0.3 is applied to machine-learning-positive predictions that lack homology in the annotated databases, and such sequences are automatically flagged for expert manual inspection within the tool’s interface.

The authors argue that this design turns a known biological difficulty into a feature rather than a flaw. Their analysis indicates that most false-positive predictions occur in proteins sharing structural motifs with selenoproteins, such as redox-active cysteine-rich domains, and the validation layer flagged examples including chorismate synthase homologs detected in the test set. Because cysteine and selenocysteine are closely related chemically, proteins involved in sulfur-dependent catalysis frequently share conserved scaffolds and active-site environments with bona fide selenoproteins, and standard probabilistic classifiers alone cannot fully resolve this topological convergence. A functional enrichment analysis using STRING-db version 12.0 on the flagged predictions revealed a marked overrepresentation of proteins involved in RNA processing, methyltransferase activity and redox-related enzymatic functions, which the authors interpret as evidence that the model is highly sensitive to structurally analogous catalytic centers of genuine biological interest rather than producing stochastic errors.

Compared with homology-based genomic annotation efforts and with deep-learning-only classifiers, SelCysML occupies a complementary position. By focusing on physicochemical descriptors rather than strictly on sequence similarity, the tool is positioned to identify novel or highly divergent selenoproteins that may escape conventional alignment-based detection, while the BLAST layer converts a binary prediction into a verifiable functional annotation. The authors also acknowledge a potential confounder: because the initial dataset paired multi-domain negative controls with an exclusively bacterial positive class, global compositional descriptors could in principle reflect broad phylogenetic differences rather than intrinsic selenoprotein properties. The bacteria-specific engine, trained with domain-matched negative controls, was built to isolate local physicochemical profiles such as thiol- and selenol-associated amino acid distributions from that background signal, and its superior performance in prokaryotic settings supports that strategy.

SelCysML is delivered as a user-friendly R/Shiny framework, with source code and manuals publicly available on GitHub and the ensemble model bundle and BLAST databases hosted on Zenodo, allowing fully offline and reproducible runs. The authors present the pipeline as a step toward addressing the reproducibility and transparency problems associated with opaque deep-learning classifiers, transforming computational screening into a platform in which ambiguous structural analogs are surfaced for expert review rather than silently misclassified. Future iterations are planned to incorporate U-aware encodings with a 21-amino-acid composition to quantify the specific contribution of explicit selenocysteine residues versus context-based sequence features. For researchers probing bacterial selenoproteomes, the tool offers a route from raw proteomic data to functionally annotated candidates that is both statistically robust and biologically auditable.

Subject of Research: A hybrid machine learning and BLAST pipeline for identifying and functionally annotating bacterial selenoproteins

Article Title: SelCysML: a hybrid ML-BLAST pipeline for the automated identification and functional annotation of bacterial selenoproteins

Article References: Lavín, J. L., & Oguiza, J. A. (2026). SelCysML: a hybrid ML-BLAST pipeline for the automated identification and functional annotation of bacterial selenoproteins. Frontiers in Bioinformatics, 6, Article 2001675. https://doi.org/10.3389/fbinf.2026.2001675

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.2001675

Keywords: selenoproteins, selenocysteine, machine learning, BLAST, bioinformatics, bacteria, protein annotation, proteomics, ensemble models, BSepDB, redox biology, functional genomics