zERExtractor: Enzyme Data Extraction

• zERExtractor is an automated, extensible multimodal platform that extracts enzyme kinetics data from varied scientific sources.
• It integrates deep learning, advanced OCR, and large language models to transform heterogeneous tables, figures, and text into structured records.
• The system bridges the gap between raw enzyme literature and curated databases, enhancing data reliability for enzyme modeling.

zERExtractor is an automated, extensible multimodal platform for extracting enzyme-catalyzed reaction and activity data from scientific literature. It is presented as infrastructure for converting enzyme kinetics information scattered across tables, figures, molecular diagrams, and free text into structured machine-readable records suitable for databases and machine learning. The system combines domain-adapted deep learning, advanced OCR, semantic entity recognition, prompt-driven LLM modules, and human expert correction to recover kinetic parameters such as $k_\text{cat}$ and $K_m$, enzyme identity, substrate SMILES, experimental conditions, molecular diagrams, and related biochemical fields from heterogeneous document formats (Zhou et al., 30 Jul 2025).

1. Scientific Rationale and Scope

The system is motivated by what the authors describe as a major literature-to-database gap in enzymology. Only a small portion of experimentally reported enzyme kinetic data is curated into resources such as BRENDA and SABIO-RK, while much of the enzyme–substrate–condition evidence remains embedded in publications. zERExtractor is framed as a mechanism for recovering this inaccessible layer of biochemical evidence and converting it into structured records for downstream enzyme modeling, enzyme engineering, and biochemical knowledge discovery (Zhou et al., 30 Jul 2025).

The paper places particular emphasis on enzyme kinetics literature, where key information is distributed across multiple modalities rather than expressed in a single normalized representation. Parameters such as $k_\text{cat}$ and $K_m$, together with enzyme identity, sequence, substrate structure, mutation context, and assay conditions, are described as foundational inputs both for biochemical databases and for AI models that predict turnover numbers, specificity, catalytic behavior, and enzyme–substrate relationships. This suggests that zERExtractor is intended not merely as a document-mining utility, but as a bridge between primary literature and structured enzyme informatics.

Its initial benchmark domain is P450-related enzymology. That concentration gives the system a concrete biochemical target, but it also makes clear that the paper’s strongest empirical claims are anchored in a specific enzyme-family literature rather than in a universal survey of enzymology.

2. System Architecture and Document Workflow

zERExtractor is described as a unified but modular pipeline with a plug-and-play or plugin-ready design. The major components can be swapped, augmented, or upgraded as better models become available, including LLMs. The system is therefore not presented as a monolithic extractor. Instead, it integrates document parsing, OCR-based table understanding, molecular image recognition, semantic entity extraction, relation extraction, schema alignment, normalization, and human correction within a single workflow (Zhou et al., 30 Jul 2025).

The processing begins with raw PDF ingestion. Documents are parsed and segmented into tables, figures, and plain text. Table localization is performed with PP-YOLOv2, while structure-aware OCR and layout reconstruction are handled by SLANet within PaddleOCR PP-Structure. The table extractor uses both textual content and visual layout cues, including cell bounding boxes and spatial relationships, to reconstruct tables containing multi-line headers, merged cells, and nested parameter groupings. The reconstructed output is converted into JSON, with rows corresponding to biological reaction entries.

A subsequent semantic interpretation layer aligns OCR-recovered columns to a predefined schema. This stage uses schema alignment and fuzzy header matching so that heterogeneous biochemical headers can be mapped to standard fields even when exact header names are absent. Quantitative fields are normalized into value–unit pairs. The paper also describes prompt-driven LLM modules for document understanding, schema alignment in difficult or noisy tables, semantic interpretation, and ambiguous field resolution. Figure descriptions further indicate an alternative LLM-based pathway in which prompt-based document parsing, optionally with zero-shot or few-shot adaptation and even LoRA-style tuning, can directly extract structured biochemical entries from unstructured text or semi-structured tables; however, the paper does not report exact prompts, model identities, or LoRA hyperparameters.

For figures and reaction schematics, zERExtractor introduces UniMolRec, a dedicated module for molecular structure interpretation. UniMolRec converts molecular depictions into canonical SMILES and is described as an ensemble of multiple specialized deep learning models designed to improve robustness across publication styles, noise conditions, and image formats. These recognized molecules are then linked to biochemical roles such as substrate, product, and cofactor through a relation extraction stage that uses visual-semantic alignment and rule-guided decoding. The pipeline ultimately integrates table-derived and figure-derived outputs into schema-constrained structured JSON records.

3. Extraction Schema and Multimodal Outputs

The extraction target is broader than numerical kinetics alone. The paper states that zERExtractor extracts kinetic parameters, relative activities, yield or conversion percentages, enzyme names, enzyme IDs, substrate names, product names, substrate SMILES, mutations, source organism, pH, reaction metadata, molecular diagrams, and reaction roles such as substrate, product, enzyme, and cofactor (Zhou et al., 30 Jul 2025).

The explicit schema reported in the paper is as follows.

Field	Type
Enzyme name	string
Substrate	string
Enzyme ID	string
Product	string
pH	float
$K_\text{m}$	object containing value and unit
$k_\text{cat}$	object containing value and unit
Mutations	list of strings or string
Source Organism	string
Relative Activity	object containing value and unit

The paper specifies that Enzyme ID includes UniProt, GenBank, or PDB identifiers. Quantitative fields are extracted as value–unit pairs, with examples including $\mu M$, $\text{s}^{-1}$, and $\%$. Relative Activity also covers yield or conversion percentages.

Two scope details are noteworthy. First, enzyme sequences are claimed in the abstract and introduction, and “sequences” are also included in the list of annotated biological fields, but sequence-specific schema fields are not listed in the core schema table. Second, the system performs biochemical role assignment at the reaction level, linking a substrate SMILES to the corresponding enzyme, product, and kinetic parameters under a specified condition set. The paper does not define a formal ontology, BIO tagging scheme, or relation label inventory beyond the biochemical role names described in prose.

4. Annotation Platform, Expert Supervision, and Active Learning

Human expert correction is a core part of zERExtractor rather than an after-the-fact quality check. The paper describes an interactive annotation and validation platform with three coordinated views: a PDF and table annotation panel for in-context inspection, a chemical structure and reaction relationship interface for molecule recognition and linkage, and a unified structured output panel in which fields such as enzyme name, SMILES, mutations, $K_m$, $K_m$0, and relative activity are aggregated. Annotators can modify, delete, or reprocess extracted entries, and both pre- and post-annotation data are retained for evaluation and retraining (Zhou et al., 30 Jul 2025).

Annotation was performed by ten researchers with biological backgrounds under a unified schema-aligned labeling guideline. The paper also reports a dual-phase review process for tables, with initial annotation followed by cross-checking by independent annotators. In an earlier quality-control stage, approximately 20% of extracted tables, specifically 54 tables, were manually reviewed by two graduate-level annotators with expertise in enzymology. After cleaning and validation, 175 high-quality tables were retained, yielding 612 structured reaction entries for downstream modeling and analysis. The paper presents these numbers as an earlier curation subset rather than the full released benchmark.

The authors repeatedly characterize the workflow as active learning, but the description is conceptual rather than algorithmically explicit. AI-assisted annotation produces draft labels; experts validate and correct them; corrected data are returned to the training loop; and the system adapts to new data sources through iterative refinement. No formal acquisition function, uncertainty score, entropy criterion, or ranking rule is provided. The paper does report that corrected data were used to retrain SLANet for better structural and semantic field alignment in domain tables and to fine-tune an LLM for reaction-level inference and ambiguous entity resolution.

5. Benchmark Dataset and Empirical Results

The released benchmark is centered on 270 P450-related publications from 2000 to 2024. These documents were curated from open-access or peer-reviewed enzymology literature and filtered with keywords such as “P450” and “enzyme kinetic parameter,” with an early-stage requirement that documents contain at least one experimental table. The benchmark is described as including over 1,000 fully annotated tables and approximately 5,000 biological fields, although one section elsewhere refers to over 800 tables for the table extraction module; the counts are not fully reconciled within the paper (Zhou et al., 30 Jul 2025).

The training resources reported for individual modules are more detailed than the overall benchmark description. For table extraction, SLANet is first pretrained on PubTabNet, whose size is given as 500,777 training images, 9,115 validation images, and 9,138 testing images, and then fine-tuned on a biochemical domain dataset automatically annotated by the system. For molecular image recognition, the training data include approximately 1 million synthetic molecular images rendered with Indigo from PubChem molecules, approximately 680,000 molecular images from USPTO patents, and an in-house set of exactly 1,355 manually labeled molecular reaction images from scientific literature PDFs.

The headline performance numbers are concentrated in three areas.

Component	Benchmark or dataset	Reported result
Table recognition	PubTabNet	Acc 89.9%
Molecular image recognition	UOB	99.1%
Relation extraction	Manually annotated reaction benchmark	accuracy 94.2%

The paper also reports a comparison on PubTabNet in which TableMaster achieves 77.90%, LGPMA 65.74%, SLANet 86.0%, and zERExtractor 89.9%, corresponding to a 3.9% gain over SLANet. For molecular image recognition, reported accuracies are 98.9% on Indigo, 98.1% on ChemDraw, 95.7% on CLEF, 88.7% on Staker, 75.8% on ACS, 94.2% on USPTO, 68.8% on JPO, and 99.1% on UOB. The authors state that relation-role assignment accuracy reaches 94.2% for roles including substrate, product, and cofactor. During refinement with human-annotated data, table field extraction accuracy improves by 3.9% and molecular relation recognition by 3.0%.

Several benchmarking details remain unresolved in the paper’s presentation. One section refers to a manually annotated benchmark of 270 full-text articles containing 100 enzyme-related tables for evaluation of table extraction, but its relation to the larger 800+/1,000+ table corpus is not fully clarified. The paper also mentions additional molecular-recognition metrics such as InChIKey Accuracy, Perfect Match ratio, Validity, and Tanimoto Similarity, yet only the accuracy table is numerically reported in the described results.

6. Position within Enzyme Informatics, Reproducibility, and Limitations

The paper identifies three main contributions: a unified multimodal extraction system combining domain-adapted deep learning and LLMs; an expert-guided active-learning and annotation loop for continuous improvement; and a large expert-annotated benchmark dataset for enzyme informatics (Zhou et al., 30 Jul 2025). In technical terms, zERExtractor is best understood as a hybrid multimodal information extraction platform tailored to enzymology papers. It combines layout analysis, structure-aware OCR, schema-aligned table parsing, optical chemical structure recognition, biochemical role extraction, LLM-assisted semantic interpretation, and human correction inside an iterative retraining cycle. Its intended output is not isolated snippets, but normalized structured reaction records.

A common misconception would be to interpret zERExtractor as an OCR system specialized for biochemical tables. The paper instead presents it as a broader reaction-record construction framework. Table extraction is only one layer; the overall objective is relation-level integration across text, tables, and figures so that the system can determine which $K_m$1 and $K_m$2 belong to which enzyme–substrate pair under which conditions.

The paper also leaves substantial reproducibility gaps. The benchmark is concentrated on P450 literature, so generalization to other enzyme families remains open. Robustness to highly heterogeneous layouts appears incomplete, as indicated by the need for domain-specific fine-tuning, fuzzy alignment, and expert correction. LLM integration is operationally underspecified: the exact models, prompt templates, shot counts, data formats, optimization settings, and decoding parameters are not reported. The mathematical treatment is also informal. Apart from domain variables and units such as $K_m$3, $K_m$4, $K_m$5, $K_m$6, and $K_m$7, the paper provides no formal equations for losses, active-learning scores, or evaluation metrics. It likewise offers no broad ablation study and no detailed error analysis beyond qualitative remarks about heterogeneous table formats, irregular layouts, nested headers, and ambiguous biological field expressions.

Within those limits, the paper’s central claim is clear: zERExtractor is intended as the missing bridge between large volumes of uncurated enzyme kinetics literature and the structured corpora required by modern AI models of enzyme function. Its modularity suggests an architecture designed to evolve as OCR engines, recognition models, and LLM components improve, but the present empirical record is most concrete for P450-focused multimodal extraction rather than for fully general enzymology.