The problem the Canadian team set out to solve is deceptively simple to state and notoriously hard to crack. When a protein’s structure has been solved experimentally, domain boundaries can be traced by eye or by algorithm through the folded chain. But structural determination remains slow, expensive, and unevenly distributed across the tree of life. For everything else, scientists lean on conserved domain annotations from resources such as Pfam, InterPro, and the NCBI Conserved Domain Database, or they resort to iterative truncation screening, building a panel of constructs and testing which ones express, fold, and behave well. Each of these strategies has blind spots. Annotation databases capture only domains that have already been characterized in other proteins, leaving novel or fast-evolving domains invisible. Truncation screening is empirical, laborious, and often burns months of laboratory time on constructs that fail. Seq2DomML was designed to complement these approaches by learning, from data, what a domain boundary looks like in the raw language of sequence.

The foundation of the framework is the Evolutionary Classification of Domains, or ECOD, a curated database that organizes protein domains by their evolutionary and structural relationships. The researchers converted ECOD’s structural domain annotations into binary residue-level labels: each amino acid in a bacterial protein sequence was marked as belonging to an annotated domain, or alternatively as part of a domain boundary, an inter-domain linker, or a terminal flanking region at the ends of the chain. This transformation turns a structural biology question into a supervised sequence labeling problem, the same class of task that powers speech recognition and named-entity extraction in language technology. Every residue becomes a token to be classified, and the model must learn the contextual cues that signal whether a given position sits inside a folded unit or in the flexible connective tissue between them.

Feeding that problem required an unusually rich encoding of each residue. The team initially described every amino acid with 2,389 features spanning amino acid identity, positional information along the chain, physicochemical properties, and interaction-related descriptors. That breadth is a deliberate hedge: no single property cleanly distinguishes a boundary residue from an interior one. Hydrophobicity, conservation patterns, local sequence composition, and position relative to the termini all carry partial signals, and a boundary might be legible only in their combination. But a feature space of nearly 2,400 dimensions also brings the classic curse of dimensionality, inviting overfitting and burying informative signals in noise. The researchers therefore invested heavily in feature selection, and this is where Seq2DomML acquires its most distinctive technical character.

Feature reduction proceeded in two stages. First, correlation-based pruning removed redundant descriptors that carried overlapping information. Then came the evolutionary step: a multi-objective genetic algorithm that treated the feature set itself as a genome to be optimized. In the initialization stage, the algorithm randomly assigned numbers of active features across a population and generated 100 unique binary feature vectors, each one a candidate recipe of which descriptors the model should see. Pairs of parents were then matched at random and combined through two-point crossover to produce two offspring per couple, with bit-flip mutation applied to the offspring to preserve diversity across the population. Each generation, every parent and offspring was trained and evaluated on three objectives simultaneously: the macro F1-score, the prediction rate for the minority boundary class, and the recall for that class. A non-dominated sorting algorithm, borrowed from multi-objective optimization theory, ranked individuals by their best achievable trade-offs across the three objectives, and the top individuals, equal in number to the original population, survived into the next generation. Crossover, mutation, and selection repeated for 50 generations, gradually sculpting a compact feature set that balanced overall accuracy against sensitivity to the rare and difficult boundary residues.

The classifier at the heart of the framework is a bidirectional long short-term memory network, a recurrent architecture well suited to sequential data. Bidirectionality matters here for a reason rooted in protein biology: the signal that a residue sits at a domain boundary depends on context from both sides of the chain, the tail end of one folded unit and the beginning of the next. An LSTM reading the sequence only left to right would encounter boundary evidence after the fact; a bidirectional model integrates information flowing in both directions, allowing the hidden state at each position to reflect the entire local sequence environment. The LSTM’s gating mechanisms, which learn to retain or forget information over long stretches of input, help the network bridge the dozens or hundreds of residues that may separate a boundary from the nearest obvious sequence landmark.

Evaluation was designed to guard against the most common pitfall in protein machine learning: leakage of evolutionary memory. The researchers removed from the held-out test set any sequences sharing at least 35 percent global sequence identity with the training set, ensuring the model could not simply recognize close relatives of proteins it had already seen. Against this stringent benchmark, Seq2DomML was compared with established annotation tools including InterPro, Pfam, and NCBI CDD, and it outperformed them on the classification metrics. It achieved a macro F1-score of 0.734, a figure that weights performance across all residue classes equally and therefore reflects genuine skill on the underrepresented boundary and linker categories, alongside a within-domain F1-score of 0.924, showing near-expert reliability at recognizing residues inside annotated domains. Its domain count mean absolute error of 0.952 was the second-lowest among the evaluated methods, trailing only Pfam.

Perhaps the most telling statistic is the mean domain Intersection over Union, a metric borrowed from computer vision that measures the spatial overlap between predicted and reference regions. Seq2DomML posted the highest mean IoU of the field, 0.846, indicating the strongest spatial agreement with the ECOD-derived domain regions among all methods tested. In practical terms, this means the model does not merely tally the right number of domains; it draws their edges in roughly the right places. For an experimentalist deciding where to cut a protein to isolate a single domain for structural studies, edge placement is everything. A prediction that captures the correct domain count but misplaces a boundary by thirty residues can yield a construct that aggregates, misfolds, or refuses to crystallize. The IoU result suggests Seq2DomML’s boundaries are tight enough to be actionable.

The authors are careful to position the tool as a complement rather than a replacement for the annotation infrastructure that has served biology for decades. Conserved domain databases rest on curated evolutionary evidence and remain authoritative where homologs exist. Seq2DomML’s value emerges precisely where that evidence runs out: in bacterial proteins with no experimentally determined structure, no recognizable domain homologs, and no reliable annotation to guide construct design. In those cases, the model’s residue-level partitions can help prioritize candidate splice regions, flagging the most promising places to truncate a protein before a single clone is made. Because it requires only sequence as input, it scales to the enormous and rapidly growing output of bacterial genome sequencing projects, where the majority of encoded proteins still await structural or functional characterization.

The broader significance of the work lies in its demonstration that structurally defined partitions are recoverable from sequence-derived features alone. The multi-objective evolutionary feature selection adds a methodological lesson for the field: when the class of interest is rare and the feature space is vast, optimizing a single accuracy metric can quietly sacrifice exactly the predictions that matter most. By forcing the feature set to balance macro F1, boundary prediction rate, and boundary recall simultaneously, the Wilfrid Laurier team built a model that stays honest about its hardest task. As structural biology pipelines increasingly depend on computational triage to decide which proteins and which fragments are worth experimental effort, frameworks like Seq2DomML point toward a future in which the chapter breaks of every protein, however novel, can be read straight from its sequence.

Subject of Research: Machine learning prediction of protein domain boundaries from bacterial amino acid sequences

Article Title: Seq2DomML: Protein domain boundary prediction from bacterial protein sequences using a machine learning framework

Article References: Monks, A., Asilian Bidgoli, A., & Suits, M. D. L. (2026). Seq2DomML: Protein domain boundary prediction from bacterial protein sequences using a machine learning framework. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06666-y

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06666-y

Keywords: protein domains, domain boundary prediction, machine learning, LSTM, feature selection, ECOD, bioinformatics, bacterial proteins, construct design, sequence annotation, Pfam, InterPro

The study was led by Professor Rachel Kolodny and PhD candidate Guy Yanai of the University of Haifa, together with Professor Nir Ben-Tal and graduate student Gabriel Axel of Tel Aviv University, and Specially Appointed Associate Professor Liam M. Longo of ELSI. Kolodny also spent five months as a visiting researcher at ELSI, developing methods to analyze the new model. Their creation, dubbed CLSS for Contrastive Learning Sequence-Structure, is a protein language model designed to overcome a stubborn problem that has limited previous computational approaches: the awkward relationship between what a protein’s sequence says and what its structure actually does.

Scientists have long organized proteins into hierarchical groups based on relatedness, much like the genus and species categories biologists use to classify organisms. These curated systems, such as the widely used ECOD and CATH databases, distill decades of expert knowledge. But with artificial intelligence now capable of generating ’embeddings’—numerical representations in which proteins with similar properties receive nearby coordinates, like postal codes on a map—researchers can visualize relationships across millions of proteins at once, producing what the team calls a protein world map. The catch is that sequence and structure do not map neatly onto each other. Unrelated sequences can fold into similar shapes, while even identical sequences can sometimes adopt wildly different structures.

Most existing protein language models treat sequence and structure as separate worlds, processing one or the other independently. Even hybrid models that incorporate both kinds of data rarely place the sequence and the structure of the same protein at the same location on a global map, leaving researchers with two conflicting atlases of protein space. CLSS was engineered specifically to resolve this discordance. Using a machine learning strategy known as contrastive learning, the model is trained on pairs of protein sequences and their corresponding structures, learning to pull matching sequence-structure pairs together in the embedding space while pushing unrelated pairs apart.

The result is a single shared map in which a protein occupies essentially the same location whether the model is given its sequence or its structure. When benchmarked against other state-of-the-art protein language models, CLSS succeeded in producing a cohesive unified representation, something its predecessors could not achieve. Remarkably, the model’s maps closely reproduced the relationships recorded in the expert-curated ECOD and CATH classification systems, even though those classifications were never shown to the model during training. In direct classification tests, CLSS also performed strongly, demonstrating that merging sequence and structure information yields genuinely more informative protein representations.

Perhaps the most exciting feature of CLSS is its ability to handle fragments. Most protein language models require a complete sequence or structure to generate a meaningful embedding, but CLSS showed that short sequence fragments can in many cases be positioned meaningfully alongside full-length proteins and structures. This capability matters enormously for evolutionary studies, because small pieces of proteins have been repeatedly reused and rearranged throughout the history of life. Some fragments may even have served as the primordial building blocks from which the earliest protein domains were assembled, meaning that similar fragments appearing in otherwise unrelated proteins can hint at ancient evolutionary connections.

The maps produced by CLSS also revealed sweeping patterns across protein space that were previously difficult to see. When the researchers overlaid biological properties onto the maps, proteins associated with organic cofactors turned out to cluster in particular regions, while metal-binding proteins were scattered more broadly. Such patterns illustrate how global protein maps can serve not only as classification tools but as instruments for exploring the interplay between sequence, structure, function, and deep evolutionary history, potentially exposing large-scale patterns invisible to conventional pairwise comparison methods.

‘This gives us a way to look at the protein universe through sequence and structure at the same time, rather than treating them as separate worlds,’ said Longo. ‘What is particularly exciting for us is the possibility of using these maps to uncover large-scale evolutionary patterns that are difficult to recognise using conventional approaches.’ The team ultimately envisions unified sequence-structure representations opening new frontiers in database searches, protein engineering, and the reconstruction of evolutionary trajectories—offering a fresh window onto how the staggering diversity of proteins found in life today emerged over nearly four billion years of evolution.

Subject of Research: A contrastive-learning protein language model that unifies protein sequence and structure representations to map the protein universe

Article Title: Uniting sequence and structure to map the protein universe

Article References: Uniting sequence and structure to map the protein universe. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: protein language model, CLSS, protein evolution, contrastive learning, protein structure, amino acid sequence, ECOD, CATH, protein embeddings, evolutionary biochemistry, protein classification, Institute of Science Tokyo