CoPaLink: Linking Bioinformatics Text and Code

• CoPaLink is an end-to-end system that bridges narrative bioinformatics workflows and executable implementations by linking tool mentions across text and code.
• It integrates advanced named entity recognition and knowledge-base pivot linking to map tools from scientific publications to their computational counterparts.
• Empirical evaluations on gold-standard corpora demonstrate enhanced reproducibility and workflow reuse, despite challenges from upstream entity recognition errors.

Searching arXiv for the cited CoPaLink-related papers to ground the article and clarify the term’s usage. arxiv.search({"query":"id:(Sebe et al., 9 Mar 2026) OR id:(Vatter et al., 2016) OR id:(Ismail et al., 2014)","max_results":10,"sort_by":"relevance"}) CoPaLink is an end-to-end system designed to bridge the gap between narrative descriptions of Bioinformatics workflows in scientific publications and their executable implementations, such as Nextflow workflows, by automatically identifying Bioinformatics tool mentions in both text and code and linking them across modalities. It extends entity linking to an “intermodal” setting in which mentions are aligned either directly, or via pivoting on shared knowledge-base entities. Its stated aims are to enhance reproducibility, facilitate reuse and understanding of published workflows, and support both pre-publication consistency checks and post-publication exploration (Sebe et al., 9 Mar 2026).

1. Conceptual definition and intermodal setting

CoPaLink operationalizes workflow reproducibility at the level of tool mentions. Traditional entity linking associates entity mentions in text with entries in a knowledge base. CoPaLink extends this notion to the case of two different sequence genres—natural language and workflow code—so that a tool described in a paper can be mapped to its counterpart in an executable workflow. In the formulation reported for the system, the linking can be performed directly between modalities or indirectly through shared knowledge-base identifiers (Sebe et al., 9 Mar 2026).

The architecture is organized around three components: Named Entity Recognition in scientific text, Named Entity Recognition in workflow code, and entity linking grounded on Bioinformatics knowledge bases. The design choice is consequential for reproducibility because it makes explicit whether “each step described in a paper corresponds to the actual code executed.” It is also consequential for workflow reuse because it exposes the correspondence between narrative explanation and implementation rather than treating them as separate artifacts.

A plausible implication is that CoPaLink is less a single classifier than a pipeline for structured alignment across heterogeneous representations. The reported evaluation confirms that this decomposition is technically important: high per-component performance does not automatically translate into equally high end-to-end accuracy because the linking stage inherits upstream NER errors.

2. Corpora, annotation protocol, and evaluation splits

CoPaLink relies on three publicly available gold-standard corpora on Zenodo: CPL-Article, CPL-Code, and CPL-Gold-Entity-Link. Annotation was performed in BRAT following detailed guidelines, and the three corpora jointly support text-side extraction, code-side extraction, and cross-modal alignment (Sebe et al., 9 Mar 2026).

Corpus	Source and scope	Annotation facts
CPL-Article	205 PubMed Central articles; 26 “Materials and Methods” sections annotated	36 564 words, 669 mentions, 282 unique tool names, IAA 80%
CPL-Code	~3 000 Nextflow workflows on GitHub; 797 processes manually annotated	78 562 words, 1911 mentions, 421 unique tools, IAA 85%
CPL-Gold-Entity-Link	Fifteen workflows with both text and code annotated	190 true cross-modal links, 42 text-only, 65 code-only, IAA 95%

For CPL-Article, the source articles were retrieved from PubMed Central with the query “(nextflow[Abstract] OR nextflow[Title]) AND github[All Fields]” in January 2026. The subset actually annotated consists of 26 “Materials and Methods” sections extracted with a rule-based heuristic. The evaluation protocol holds out five workflows for test, corresponding to 10 427 words and 310 mentions, while the remaining 21 workflows are randomly split 75%/25% into train and development sets. Five random splits are used to ensure robustness.

For CPL-Code, processes were extracted via BioFlow-Insight. The train, development, and test splits mirror CPL-Article so that both modalities are evaluated on the same five workflows; the test partition contains 173 processes and 515 mentions. CPL-Gold-Entity-Link provides the supervision required to assess intermodal matching rather than only within-modality recognition.

This corpus design constrains the scope of the reported results. The data are centered on Nextflow, and the authors explicitly note that other workflow languages such as Snakemake and WDL are not covered.

3. Named Entity Recognition in text and workflow code

CoPaLink evaluates three families of NER approaches, followed by a vocabulary-expansion mechanism that modifies encoder tokenization. The NER task is to identify spans in text or code that correspond to Bioinformatics tools (Sebe et al., 9 Mar 2026).

The first family is KB-based dictionary lookup. It uses Biotools, Bioconda, Biocontainers, and Bioweb, builds a flat dictionary of tool names and binary commands, and performs exact string match on text or code tokens. This requires no training and serves as a zero-shot baseline.

The second family is decoder-based few-shot NER using LLM prompting. The evaluated models are Llama-3.1-8B-Instruct, Qwen-3-8B, CodeLlama-13B-Instruct-hf, and Qwen-2.5-Coder-14B-Instruct. The prompt structure has three parts: a task description defining “Bioinformatics tool,” 5–10 annotated examples, and the test instance. The examples differ by modality: complete sentences for text and 4-line code snippets for code. Inference is auto-regressive generation of labeled spans.

The third family is encoder-based supervised NER using a BiLSTM-CRF architecture implemented in NLStruct. The architecture is specified as token embeddings $\rightarrow$ BiLSTM $\rightarrow$ CRF for BIO-tagging, with nested entities supported. The pre-trained backbones tested are SciBERT and ModernBERT for text, and CodeBERT, SciBERT, and ModernBERT for code.

CoPaLink further evaluates “Vocabulary Expansion (Early Fusion).” Tool names and binary commands from the knowledge bases are injected into the encoder tokenizer and vocabulary. New embeddings are initialized as the average of constituent tokens, following “fast vocabulary transfer,” and the entire model is then fine-tuned on the annotated corpus. The intended effect is to align the tokenization regime with domain-specific lexical units that standard encoders may otherwise fragment.

4. Intermodal entity linking strategies and formal evaluation

After mention extraction, CoPaLink aligns tool mentions across text and code through four strategies: string-to-string matching, KB-pivot linking, embedding-similarity linking, and decoder-based linking (Sebe et al., 9 Mar 2026).

String-to-string matching includes exact match, Levenshtein distance with thresholds $k=1,2,3$, and common prefix or suffix matching. These methods are purely surface-form based.

KB-pivot linking maps each mention in either modality to all matching knowledge-base entries via exact match on any alternative name from Bioconda, Biocontainers, or Bioweb. A text mention and a code mention are linked if they share any common KB entity. Variants are evaluated for individual KBs and for their fusion, the latter using transitive closure across all dictionaries.

Embedding-similarity linking encodes the tool name, optionally together with a context window, using a pre-trained LLM; the reported implementation uses ModernBERT with injected vocabulary. Cosine similarity between text-side and code-side mention embeddings is thresholded to decide whether to link. Experiments are performed both with and without KB-based candidate filtering and with and without context.

Decoder-based linking uses prompting over aggregated sentences and processes, asking a model to pair each text mention with its code counterpart. The tested models are Qwen-3-8B and Llama-3.1-8B-Instruct, again with and without context.

The reported evaluation for both NER and EL uses Precision, Recall, and F1-measure:

$$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$

For NER, true positives are correctly predicted tool spans. For EL, true positives are correctly linked mention pairs, including correct “no-link” decisions. This inclusion matters because cross-modal alignment is a structured decision problem in which abstaining from an incorrect link is itself evaluated.

5. Empirical results and end-to-end behavior

The empirical results separate component-level performance from end-to-end pipeline accuracy. Across the reported experiments, supervised encoder-based NER and KB-grounded linking are the strongest combination, while few-shot decoder approaches are weaker, especially on code (Sebe et al., 9 Mar 2026).

Task	Best reported configuration	Main result
Text NER	SciBERT + vocab injection	$P=88.7\pm1.7$, $R=80.3\pm3.3$, $F1=84.2\pm1.6$
Code NER	CodeBERT + injection	$P=88.5\pm1.9$, $R=90.1\pm1.4$, $F1=89.2\pm1.1$
Intermodal EL	Bioconda + Bioweb fusion	$\rightarrow$0, $\rightarrow$1, $\rightarrow$2
End-to-end	Best NER + KB-Pivot linking	Joint accuracy $\rightarrow$3

The dictionary baselines show substantial modality dependence. For CPL-Article, the best single KB is Bioconda with $\rightarrow$4, $\rightarrow$5, $\rightarrow$6. For CPL-Code, the best is Biocontainers with $\rightarrow$7, $\rightarrow$8, $\rightarrow$9. The union of all KBs raises recall to 82–94% but lowers precision to 12–22%, producing no F1 gain.

Few-shot decoder-based NER improves over naive lookup in text but not in code. On CPL-Article, Qwen3-8B achieves the best F1 among the prompted models with $k=1,2,3$0 ($k=1,2,3$1, $k=1,2,3$2). On CPL-Code, Qwen2.5-Coder-14B reaches only $k=1,2,3$3 ($k=1,2,3$4, $k=1,2,3$5). The reported conclusion is that few-shot LLMs underperform KB-based and supervised models, especially on code.

Encoder-based NER is the strongest family. On text, SciBERT with vocabulary injection reaches the best score and is reported as statistically significant relative to the compared baselines. On code, CodeBERT with injection achieves a slight, non-significant gain over the corresponding non-injected model.

For linking, the best string-based method is Levenshtein $k=1,2,3$6 with $k=1,2,3$7, $k=1,2,3$8, $k=1,2,3$9, while prefix or suffix matching yields $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$0. KB-pivot linking surpasses these baselines: Bioconda-exact reaches $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$1, Bioweb-exact reaches $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$2, and Bioconda + Bioweb fusion reaches the best overall result with $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$3. Adding other KBs gives no further gain.

Embedding-similarity linking is weaker. Embeddings alone yield $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$4, $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$5, $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$6, and embeddings with KB-candidates improve to $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$7, $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$8, $$P = \frac{TP}{TP + FP},\quad R = \frac{TP}{TP + FN},\quad F1 = 2\cdot\frac{P\,R}{P + R}.$$9. Adding context produces high recall but very low precision, with $P=88.7\pm1.7$0. Decoder-based linking reaches roughly $P=88.7\pm1.7$1–$P=88.7\pm1.7$2 without context; with context, Qwen3-8B reaches $P=88.7\pm1.7$3, $P=88.7\pm1.7$4, $P=88.7\pm1.7$5, and Llama-3.1 reaches $P=88.7\pm1.7$6, $P=88.7\pm1.7$7, $P=88.7\pm1.7$8, still below string-based and KB-based methods.

The end-to-end pipeline combines the best text NER, the best code NER, and KB-Pivot linking using Bioconda + Bioweb. This yields joint accuracy 65.7% on the five test workflows. The authors identify error propagation between NER and EL as the main bottleneck. This suggests that further gains in workflow reproducibility will depend not only on stronger linking algorithms but also on reducing upstream span-detection errors.

6. Knowledge-base integration, implementation stack, and reproducibility assets

Knowledge bases are central to CoPaLink at both the extraction and linking stages. The reported resources are Bioconda, Biocontainers, Biotools, and Bioweb, with different trade-offs in coverage and precision (Sebe et al., 9 Mar 2026).

Bioconda contains approximately 5 000 packages and provides multiple alternative tool names per entry. Biocontainers has approximately 900 containers and is described as having high precision but limited coverage. Biotools and Bioweb provide moderate coverage with synonyms and commands. The integration procedure has four explicit steps: extract all $P=88.7\pm1.7$9 pairs per KB; build a per-KB mapping from name to KB entity IDs; during NER, link any detected span matching a name to that KB ID; and during EL, link two mentions if they share at least one common KB ID, using transitive closure when fusing KBs.

The implementation stack is in Python. NLStruct is used for the BiLSTM-CRF models, BRAT for annotation, Entrezpy for PubMed Central retrieval, and BioFlow-Insight for process extraction. LLM inference and prompt code use HuggingFace transformers with custom prompt templates. The reported hardware is NVIDIA RTX A6000 or Tesla V100 GPUs.

The paper also reports environmental impact per run, expressed as Global Warming Potential in gCO$R=80.3\pm3.3$0e via Green Algorithms. Encoder-based NER is reported at 80 gCO$R=80.3\pm3.3$1e for the base model and 100 gCO$R=80.3\pm3.3$2e with vocabulary injection, whereas decoder-based few-shot inference is reported at approximately 3 gCO$R=80.3\pm3.3$3e. This does not imply superior predictive quality for low-emission methods; it simply documents the computational profile of the tested configurations.

Reproducibility assets are explicitly released. The CoPaLink method and pipelines are available at https://gitlab.liris.cnrs.fr/sharefair/copalink, the experiments and data splits at https://gitlab.liris.cnrs.fr/sharefair/copalink-experiments, and the corpora in BRAT format at DOI 10.5281/zenodo.18526700 for CPL-Article, 10.5281/zenodo.18526760 for CPL-Code, and 10.5281/zenodo.18543814 for CPL-Gold-Entity-Link.

7. Terminological scope and homonymous uses in the supplied literature

The label “CoPaLink” is not unique across the supplied literature. In the 2026 bioinformatics work, it denotes the intermodal workflow-linking system described above. In the supplied summaries, however, the same label is also attached to distinct copula-based constructions from statistical dependence modeling, which are conceptually unrelated to workflow reproducibility (Vatter et al., 2016).

One such use is “Conditional Pair-copula Link,” described as a flexible framework for modeling multivariate densities via pair-copula constructions (vines), in which each bivariate copula parameter is allowed to depend on covariates through a generalized additive link. In that formulation, each copula is reparameterized by Kendall’s $R=80.3\pm3.3$4 or a natural scalar parameter $R=80.3\pm3.3$5, and a link is introduced as

$R=80.3\pm3.3$6

with additive predictor

$R=80.3\pm3.3$7

The framework uses penalized likelihood, proceeds tree by tree in a vine, and is implemented in an R package, gamCopula. Its application domain in the supplied summary is time-varying dependence among intraday foreign-exchange returns.

A second use appears in the supplied summary of a “Two-component copula” for insurance dependence modeling. There, “CoPaLink” denotes a model in which a common macro loss factor $R=80.3\pm3.3$8 is multiplied by company-specific factors $R=80.3\pm3.3$9 satisfying $F1=84.2\pm1.6$0, with total losses

$F1=84.2\pm1.6$1

The resulting margins are Type II Pareto, and the model is described as not Archimedean in general, not extreme-value in general, and asymmetric if $F1=84.2\pm1.6$2 (Ismail et al., 2014).

These homonymous uses are methodologically disjoint. One concerns workflow-tool alignment across papers and executable code; the others concern copula-based dependence structures. A plausible implication is that “CoPaLink” should be disambiguated by domain and citation context, especially in bibliographic databases or survey writing where the label alone would be insufficiently specific.