Molecular-level protein semantic learning via structure-aware coarse-grained language modeling.

{"sci_version": "1.2", "schema": "https://corpus.bntrg.co.uk/sci/1.0/sci.schema.json", "id": "pmid:41352986", "kind": "science-record", "revision": 3, "content_hash": "sha256:e2ceb65e1052fee85d6d3ae38bac086ed904fd017e2ec606975b45ae7b43faff", "created_at": "2026-11-01T20:31:22Z", "updated_at": "2026-11-01T22:15:07Z", "generator": {"name": "corpus-assay-engine", "run_id": "reading-sci-2026-11-01"}, "paper": {"state": "complete", "title": "Molecular-level protein semantic learning via structure-aware coarse-grained language modeling.", "doi": "10.1093/bioinformatics/btaf654", "pmid": "41352986", "authors": ["Jun Zhang", "Xueer Weng", "Tiantian Zhu", "Yumeng Liu", "Zexuan Zhu"], "journal": "Bioinformatics (Oxford, England)", "year": 2026, "abstract": "MOTIVATION: Protein language models (PLMs) have emerged as pivotal tools for protein representation, enabling significant advances in structure-function prediction and computational biology. However, current PLMs predominantly rely on fine-grained amino acid sequences as input, treating individual residues as tokens. While this approach facilitates semantic learning at the residue level, it struggles to capture molecular-level semantics, particularly for large proteins, where sequence truncation and inefficient local pattern extraction hinder holistic understanding. The spatial structure of a protein determines its function. Despite the critical role of protein function analysis, coarse-grained protein language frameworks that bridge sequence and structural semantics remain underdeveloped.\nRESULTS: To fill this gap, we introduce a novel structure-aware coarse-grained protein language that discretizes proteins into local structural patterns derived from their secondary structures. By constructing a vocabulary of these patterns as \"words,\" we represent proteins as compact, structure-aware \"sentences\" significantly shorter than raw amino acid sequences. We benchmark the proposed coarse-grained language against three state-of-the-art fine-grained protein languages and a classical language modeling method in natural language processing, using two architectures: a lightweight Doc2Vec model and a Transformer-based BERT model, and evaluating performance across diverse downstream tasks, including function prediction, enzyme classification, and interaction identification. The proposed method achieves stable performance across three tasks, especially for long proteins. These results demonstrate that the proposed coarse-grained protein language preserves critical structural and functional semantics and improves molecular-level analysis, offering a promising direction for decoding higher-order biological insights.\nAVAILABILITY AND IMPLEMENTATION: The data and source code of the proposed method are available at GitHub (https://github.com/bug-0x3f/coarse-grained-protein-language) and Zenodo (DOI: 10.5281/zenodo.17674298).", "claims": [{"id": "claim-method", "statement": "The relationship between SCG’s improvement over Sa on large proteins and the average protein length of the task test dataset (Pearson’s r = 0.9884).", "kind": "methodological", "target": null, "evidence": [{"asset": "source:pmid:41352986", "locator": "sections[11] \"3.2 The influence of length on protein modeling\" paragraph 4", "quote": "The relationship between SCG’s improvement over Sa on large proteins and the average protein length of the task test dataset (Pearson’s r = 0.9884)."}], "state": "complete"}, {"id": "claim-data-access", "statement": "All the datasets used in this study and the source code of the proposed method are available at GitHub (https://github.com/bug-0x3f/coarse-grained-protein-language) and Zenodo (DOI: 10.5281/zenodo.17674298).", "kind": "methodological", "target": null, "evidence": [{"asset": "source:pmid:41352986", "locator": "sections[20] \"Data availability\" paragraph 2", "quote": "All the datasets used in this study and the source code of the proposed method are available at GitHub (https://github.com/bug-0x3f/coarse-grained-protein-language) and Zenodo (DOI: 10.5281/zenodo.17674298)."}], "state": "complete"}, {"id": "claim-target-result", "statement": "In the pretraining dataset, SCG sentences are on average 7.28 times shorter than fine-grained protein-language sequences, and substantially fewer SCG sequences require truncation.", "kind": "quantitative", "target": {"value": 7.28, "unit": "fold change", "tolerance": null, "source": "Statistical analysis of the pre-training dataset revealed that fine-grained sequences are, on average, 7.28 times longer than SCG sentences."}, "evidence": [{"asset": "source:pmid:41352986", "locator": "sections[11] \"3.2 The influence of length on protein modeling\" paragraph 2", "quote": "Statistical analysis of the pre-training dataset revealed that fine-grained sequences are, on average, 7.28 times longer than SCG sentences."}, {"asset": "source:pmid:41352986", "locator": "sections[11] \"3.2 The influence of length on protein modeling\" paragraph 2", "quote": "Under this configuration, only 5 SCG sequences required truncation, compared to 1,367 truncated sequences in fine-grained languages."}], "state": "complete"}], "gaps": []}, "experiment": {"state": "partial", "objective": "In the pretraining dataset, SCG sentences are on average 7.28 times shorter than fine-grained protein-language sequences, and substantially fewer SCG sequences require truncation.", "steps": [{"id": "read_published_estimate", "order": 1, "name": "Read the published estimate from the source", "kind": "other", "state": "complete", "blocked_on": null, "description": "The claim, method and evidence span published with the source article.", "duration_hours": null, "parameters": {}, "inputs": ["source:pmid:41352986"], "outputs": ["the published estimate and its evidence"], "instruments": [], "reagents": []}, {"id": "obtain_source_data", "order": 2, "name": "Obtain the source data from the named repository", "kind": "data-preparation", "state": "blocked", "blocked_on": "data:accession_absent", "description": "The GEO series accession resolves and its sample accessions are listed. This lane did not download the dataset or read its licence.", "duration_hours": null, "parameters": {}, "inputs": [], "outputs": ["the source data with its repository provenance"], "instruments": [], "reagents": []}, {"id": "run_analysis", "order": 3, "name": "Run the published analysis", "kind": "analysis", "state": "complete", "blocked_on": null, "description": "Apply the paper's statistical method to the resolved series and compare the result with the published target.", "duration_hours": null, "parameters": {}, "inputs": [], "outputs": ["the replicated result and its comparison"], "instruments": [], "reagents": []}], "materials": [], "instruments": [], "reagents": [], "conditions": {}, "gaps": ["step:obtain_source_data blocked on data:accession_absent"]}, "data": {"state": "partial", "assets": [{"id": "source:pmid:41352986", "role": "manuscript", "format": "xml", "media_type": "application/xml", "uri": "https://doi.org/10.1093/bioinformatics/btaf654", "sha256": "sha256:309559096a60155d0fbf40e136db8876bf49734e8f366d64c85a094100cb2381", "bytes": 46248, "access": "open", "state": "resolved", "reader": null, "reader_status": "reader_required", "identifier": "10.1093/bioinformatics/btaf654", "version": null, "note": "Full article fetched from PMC; the checksum covers the rendered section body, not the original XML bytes.", "structure": {"pmcid": "PMC12758601", "source": "europe_pmc_fulltext", "coverage": "whole_body", "text_hash_scope": "rendered_section_body"}}], "gaps": []}, "analysis": {"state": "partial", "methods": [{"id": "method-1", "family": "correlation_analysis", "specific": "pearson", "label": "Correlation analysis", "dsl_operations": ["pearson_r"], "missing_operations": [], "expressible": true, "parameters": {}, "data_shapes": ["paired_numeric_columns"], "evidence": {"score": 1.0, "band": "verbatim", "quote": "The relationship between SCG’s improvement over Sa on large proteins and the average protein length of the task test dataset (Pearson’s r = 0.9884)."}}, {"id": "method-2", "family": "image_quantification", "specific": "segmentation", "label": "Image analysis / quantification", "dsl_operations": ["segment_objects", "quantify_intensity"], "missing_operations": [], "expressible": true, "parameters": {}, "data_shapes": ["image_stack"], "evidence": {"score": 0.8, "band": "high", "quote": "Segmentation"}}], "targets": [{"id": "target-published-estimate", "claim_id": "claim-target-result", "operation": "quantify_intensity", "expected_value": 7.28, "unit": "fold change", "tolerance": null, "state": "complete", "note": "The published estimate as the reader quoted it from the paper. No tolerance is published because the paper states none."}], "gaps": ["method.reader_differs_from_extractor"]}, "capability": {"state": "complete", "requirements": [{"capability_id": "capability:computational_analysis", "label": "Computational analysis platform", "assay_families": ["meta_analysis", "machine_learning", "survival_analysis", "statistical_regression", "pca", "molecular_docking", "network_analysis", "differential_expression", "gwas"], "availability_state": "owned", "tier_unlocked": "R1", "candidate": "Automated Assays Laravel application and simulated adapters"}], "gaps": []}, "readiness": {"state": "blocked", "tier": "R4", "blocking_axis": "data", "blocked_on": "data:accession_absent", "axes": {"method": {"tier": "R1", "expressible": true, "required_operations": ["pearson_r"], "missing_operations": [], "basis": "The paper names the method it used and the corpus's method extractor maps it onto the DSL catalogue: pearson."}, "capability": {"tier": "R1", "capability_id": "capability:computational_analysis", "availability_state": "owned", "basis": "The computational analysis is owned. The physical bench the paper used is not, so a physical validation cannot be run here."}, "data": {"tier": "R4", "reachable": false, "accessions": [], "resolved_accessions": [], "statements": ["All the datasets used in this study and the source code of the proposed method are available at GitHub (https://github.com/bug-0x3f/coarse-grained-protein-language) and Zenodo (DOI: 10.5281/zenodo.17674298)."], "basis": "The paper names no accession to its source data, so the data has to come from the authors."}}, "can_run": {"dsl_operations": ["pearson_r"], "assets": ["source:pmid:41352986"], "steps": ["run_analysis"]}, "must_build": {"dsl_operations": [], "converters": [], "readers": [], "capabilities": [], "data_gaps": ["data:geo:None"]}, "gaps": ["data:accession_absent"]}, "review": {"state": "partial", "machine_critique": {"kind": "extraction_review", "reader": {"model": "MiniMax-M3.1-Flash", "provider": "minimax", "provider_family": "minimax", "read_at": "2026-11-01T20:07:21Z", "independence": "single_reader"}, "evidence_check": {"method": "verbatim", "availability": "verbatim"}, "calls": 5, "usage": {"prompt_tokens": 11515, "completion_tokens": 1929, "cached_tokens": 128}}, "human_critique": null, "stages": [{"id": "extraction", "state": "complete", "reader": "MiniMax-M3.1-Flash", "provider": "minimax"}], "gaps": ["review.independent_reader", "review.machine_critique", "review.human_critique"]}, "execution": {"state": "absent", "status": "not_run", "mode": null, "runs": [], "gaps": ["execution.not_run"]}, "provenance": {"entities": [{"id": "pmid:41352986", "type": "paper", "label": "pmid:41352986"}, {"id": "source:pmid:41352986", "type": "manuscript", "label": "PMID 41352986 full text"}, {"id": "data:geo:None", "type": "dataset", "label": "GEO None"}], "activities": [{"id": "activity-reading-sci-2026-11-01", "type": "classification", "used": [{"id": "pmid:41352986", "type": "paper", "label": "pmid:41352986"}, {"id": "source:pmid:41352986", "type": "manuscript", "label": "PMID 41352986 full text"}, {"id": "data:geo:None", "type": "dataset", "label": "GEO None"}], "agent": "agent:corpus-assay-engine"}, {"id": "activity-read-41352986", "type": "extraction", "used": ["source:pmid:41352986"], "agent": "agent:minimax-reader"}, {"id": "activity-resolve-None", "type": "accession_resolution", "used": ["data:geo:None"], "agent": "agent:corpus-assay-engine"}], "agents": [{"id": "agent:corpus-assay-engine", "type": "software", "label": "Corpus assay engine"}, {"id": "agent:minimax-reader", "type": "software", "label": "MiniMax literature reader"}], "source_run_id": "reading-sci-2026-11-01", "dsl_catalogue_fingerprint": "343fb30756ba6ec3", "capability_table_fingerprint": "2cb3e42f047d2ecf", "snapshot_id": "ba34b49490d9bfaf"}, "lineage": {"parent_id": "pmid:41352986", "parent_hash": "sha256:0d9c9793ac58091e434f1245b2dfa2b0d4cbdc512be4a95f1365cf543b95dc74", "fork_of": null, "generation": 2}, "gaps": ["execution.not_run", "method.reader_differs_from_extractor", "review.human_critique", "review.independent_reader", "review.machine_critique", "step:obtain_source_data blocked on data:accession_absent"], "assays": [{"role": "analysis", "definition_id": "correlation_analysis", "definition_version": "f2132e96ae6c7a4e", "definition_digest": null, "required_operations": ["pearson_r"], "support_state": "supported", "source_hashes": {"dsl_catalogue": "f2132e96ae6c7a4e"}, "confidence": 1.0, "evidence": ["The relationship between SCG’s improvement over Sa on large proteins and the average protein length of the task test dataset (Pearson’s r = 0.9884)."], "order_basis": "analysis_only", "id": "assay:1"}]}