The first subsection describes the development of the EnzChemRED, which is the main focus of this paper. The subsections that follow describe the development of a prototype end-to-end NLP pipeline for enzyme functions that makes use of EnzChemRED for fine-tuning and benchmarking. The last two subsections describe the combination of methods to create the end-to-end pipeline, as well as methods to process and visualize the output.EnzChemRED developmentSelection of abstracts for curationTo build EnzChemRED we selected papers curated in UniProtKB/Swiss-Prot that describe enzyme functions. We queried the UniProt SPARQL endpoint (https://sparql.uniprot.org/) to identify papers that provided experimental evidence used to link protein sequence records from UniProtKB/Swiss-Prot to reactions from Rhea (specifically those reactions that involve only small molecules, excluding papers linked to Rhea reactions that involve proteins and other macromolecules) (Fig. 1). UniProtKB/Swiss-Prot uses evidence tags and the Evidence and Conclusions Ontology (ECO)45 to denote provenance and evidence for functional annotations; our SPARQL query selected papers linked to Rhea annotations in UniProtKB/Swiss-Prot with evidence tags with experimental evidence codes from ECO, such as “ECO:0000269”, which denotes “experimental evidence used in manual assertion”. We also narrowed the selection of abstracts to those including mentions of at least one pair of reactants found in Rhea, and to those having a score of at least 0.9 according to our LitSuggest46 model for abstracts relevant to enzyme function (see Section Literature triage). This LitSuggest score threshold is exceeded by 99% of abstracts of papers curated in Rhea. We selected 1,210 abstracts and divided these into 11 groups of 110 abstracts for curation by our team of expert curators.Fig. 1SPARQL query used to identify papers for abstract curation in EnzChemRED.Curation of chemical and protein mentionsCuration of abstracts in EnzChemRED was performed using the collaborative curation tool TeamTat (www.teamtat.org)47 following a protocol based on that developed for the curation of the BioRED dataset (available at https://ftp.ncbi.nlm.nih.gov/pub/lu/BioRED/), with modifications described below. We curated five types of entity in EnzChemRED: Chemical, Protein, Domain, MutantEnzyme, and Coreference, which are described below, with examples shown in Fig. 2.
Chemical: a mention of a chemical entity – including chemical structures, chemical classes, and chemical groups. Where possible we normalized chemical mentions to identifiers from ChEBI, which provides chemical structure information, and which is used to describe chemical entities in both Rhea and UniProtKB. Where no ChEBI identifier was available we used MeSH. A small number of chemical mentions have no mapping to either resource.
Protein: a mention of a protein, or family of proteins, normalized to UniProtKB accession numbers (UniProtKB ACs). As in BioRED, we included gene names in our annotation, which we also normalized to UniProtKB ACs.
Domain: a mention of a protein domain, normalized to the UniProtKB ACs of the protein in which the domain occurs.
MutantEnzyme: a mention of a mutant protein, normalized to the UniProtKB accession number of the wild-type protein.
Coreference: not a mention per se, but rather a reference to a protein or chemical mention that appears elsewhere in the abstract. In the following example, “It” is a coreference to a specific protein mention found in the preceding sentence: “ABC1 is a hydrolase. It catalyses the hydrolysis of phospholipids.”. Coreferences were normalized to the chemical or protein identifier for the mention being referenced.
Fig. 2Entity curation in EnzChemRED. We curated all Chemical and Protein mentions, but not of Domain, MutantEnzyme and Coreference (denoted by ‘*’), for which curation focused on those mentions that participate in conversions. The latter are not included in our evaluations.We curated all Chemical and Protein mentions found in abstracts, irrespective of whether those mentions were part of descriptions of enzymatic reactions or not. We did not systematically curate Domain, MutantEnzyme and Co-reference mentions, but focused on those that participate in enzymatic reactions. For this reason, we did not consider these three types of mentions in our evaluations of NER, NEN, and RE performance, but these could serve as valuable annotations for the development of a more extensive dataset in the future.Curation of chemical conversionsWe based our schema for the curation of relations relevant for enzyme functions in EnzChemRED on that developed in BioRED, but with three major alterations.First, we defined two additional relation types in EnzChemRED. BioRED captures chemical reactions by using the relation “Conversion” to link pairs of reactants. In EnzChemRED, we also added “Indirect_conversion” and “Non_conversion”, giving three possible relations (Table 2) that are defined as follows:
Conversion: links two chemicals that, according to the text, may participate on opposite sides of a reaction equation – such as one substrate and one product.
Indirect_conversion: links two chemicals that, according to the text, can interconvert, but not directly – such as conversions involving the first and last chemical in a pathway. While these kinds of relations will not give rise to enzyme function annotations, they constitute a significant but minor fraction of relations in the EnzChemRED dataset (see Section Dataset statistics and inter-annotator agreement).
Non_conversion: links two chemicals that, according to the text, were experimentally tested but did not interconvert at all (at least under the experimental conditions used). These are the rarest type of relation in EnzChemRED.
Table 2 Examples of curated relations in EnzChemRED. Chemical mentions and protein mentions are denoted by the numbered subscripts “c” and “p” respectively.Second, while BioRED features only binary pairs, in EnzChemRED we also introduced ternary tuples, which allow us to link mentions of enzymes to the “Conversions” they catalyze. We assign each enzyme the role of “Converter”.Third, we modified the granularity of annotations in BioRED. While BioRED provides document-level relation pairs, EnzChemRED provides relations annotated at the level of individual mentions and sentences.Curation workflowFig. 3 outlines the curation workflow for EnzChemRED; we describe the main steps below.Fig. 3Curation workflow for the EnzChemRED corpus. A total of 1,210 abstracts were curated by 11 experts in three phases. These 1,210 abstracts were then used to train an interim BioREx model, which was then run on all EnzChemRED abstracts; abstracts with putative FP and FN predictions by the model were then analyzed again in phase 4, and, where necessary, re-curated.Pre-Annotation of chemicals and proteins: We used PubTator48,49 to pre-tag chemical and gene/protein mentions in the 1,210 abstracts of EnzChemRED prior to their curation. PubTator assigns MeSH IDs for chemical mentions and Entrez IDs for gene and protein mentions, which we converted to ChEBI IDs and UniProtKB ACs using MeSH-to-ChEBI and Entrez-to-UniProtKB mapping tables. ChEBI provides multiple distinct identifiers for different protonation states of a given chemical compound, so we mapped all ChEBI IDs to those of the major protonation state at pH 7.3 – the form used in UniProtKB and Rhea – using a mapping file created for this purpose by Rhea (the file “chebi_pH7_3_mapping.tsv”, which is available at https://www.rhea-db.org/help/download).Curation by human experts: Curation was performed using TeamTat (Fig. 4) by a team of 11 professional curators with expertise in biochemistry and the curation of UniProtKB/Swiss-Prot and Rhea. Curators were required to review all PubTator tagging results for gene/protein and chemical mentions (both text spans and IDs) and correct them and add missing protein and chemical annotations and identifiers as necessary. Curators were allowed to use external information sources, including the full text of the article, as well as knowledge resources such as UniProtKB, Rhea, ChEBI, MeSH, and PubChem50, when curating chemical and protein mentions. Following curation of all protein and chemical mentions, curators were then required to link chemical mentions that participate in relations of the type “Conversion”, “Indirect_conversion”, or “Non_conversion”, thereby creating binary (chemical-chemical) pairs, as well as mentions of enzymes that catalyze those conversions (“Converter”) if applicable, creating ternary [enzyme-(chemical-chemical)] tuples. Curators were prohibited from using external information, such as the full text of the publication or prior knowledge of the chemistry or enzymes involved in the reactions, when annotating relations of any type, or linking converters to conversions. Put another way, all the evidence needed to create a conversion, and to link a converter to it, had to be contained in the abstract itself, either within one sentence, or across multiple sentences.Fig. 4Annotation of EnzChemRED abstracts using TeamTat. In the abstract shown (from ref. 76) the mentions of “Cysteine dioxygenase” and “CDO” refer to the enzyme (UniProt:P60334) that is responsible for the conversion of L-cysteine (CHEBI:35235) and cysteine sulfinic acid (CHEBI:61085). The inset shows details of the curated relation within the TeamTat tool, including the type of relation (“Conversion”), the text spans that define the participants in that relation, and their offsets, the unique identifiers from UniProtKB and ChEBI that were used to tag those text spans, and the assignment of the role “Converter” to the text spans of the protein mentions.We divided the curation process into four phases.
1.
Phase 1. We provided each curator with 10 abstracts for familiarization with the curation workflow and guidelines. Following curation, the abstracts were frozen, and a copy was made, which was reviewed and corrected by a second curator. The curation team then met to discuss curation issues, revise guidelines, and finalize the set of abstracts from phase 1. The output from phase 1 consisted of 110 curated abstracts, each reviewed and where necessary revised, and a set of revised guidelines.
2.
Phase 2. We provided each curator with 50 additional abstracts for curation. Curated abstracts were frozen, and a copy was made, which was reviewed and corrected by a second curator. The curation team then met to discuss curation issues, revise guidelines, and finalize the set of abstracts from phase 2. The output from phase 2 consisted of a further 550 curated abstracts, each reviewed and where necessary revised, and a set of revised guidelines.
3.
Phase 3. We provided each curator with 50 additional abstracts for curation. These were not reviewed after curation. The output from phase 3 consisted of a further 550 curated abstracts that had not been reviewed.
4.
Phase 4. We performed a round of “model guided” re-curation of abstracts, using a “preliminary” BioREx model (see Section Relation extraction) to identify potential curation errors, such as missed chemical conversions. We trained this model using the set of 1,210 abstracts curated in phases 1–3 and used it to perform RE on the entire dataset of 1,210 abstracts. We identified potential false positive (FP) or false negative (FN) predictions in 575 of the 1,210 abstracts. Each potential FP or FN prediction identified by the preliminary BioREx model was then examined by two curators, who were free to compare and discuss their interpretations of the models’ predictions. In some cases, the potential FP and FN errors from the model were deemed to be correct and were re-curated as TP or TN as appropriate, and the curation guidelines were updated if needed. We used this Phase 4 EnzChemRED dataset to train our final models for NER (see Section Named entity recognition) and RE (see Section Relation extraction).
Overview of the end-to-end pipelineFig. 5 shows the four main steps of our end-to-end NLP pipeline for enzyme function extraction, which are:
1.
Literature triage, to identify relevant papers about enzyme functions.
2.
Named entity recognition (NER), to tag chemical and protein mentions.
3.
Named entity normalization (NEN), to link chemical and protein mentions to stable unique database identifiers.
4.
Relation extraction (RE), to extract information about chemical conversions and the enzymes that catalyze them.
Fig. 5Overview of the end-to-end pipeline.The following sections describe the methods used in each of the steps, the combination of methods to create the end-to-end pipeline, and methods to process and visualize the output from it.Literature triageThe goal of the literature triage step is to identify relevant abstracts, and reduce the number of irrelevant abstracts for processing during the subsequent steps of NER, NEN, and RE. For literature triage we used LitSuggest (https://www.ncbi.nlm.nih.gov/research/litsuggest/)46, a web-based machine-learning framework for literature recommendations. LitSuggest frames literature recommendation as a document classification task and addresses that task using a stacking ensemble learning method. LitSuggest uses a variety of fields from each publication, including the journal name, publication type, title, abstract, registry numbers (for substance identifiers and names), and user-submitted keywords. Text of the fields is concatenated and converted into a bag-of-words representation, which serves as inputs/features for a diverse array of classifiers for text-mining available through the scikit-learn library. The outputs from the individual classifiers are fed into a logistic regression model, which ensemble and obtain the optimal classification. In addition to literature triage, we also used LitSuggest to confirm the relevance of abstracts selected for curation in EnzChemRED (see Section Selection of abstracts for curation).Positive training examples for LitSuggest consisted of abstracts from papers that provided experimental evidence used to annotate enzymes in UniProtKB/Swiss-Prot with Rhea reactions (dataset created November 7th, 2020). As with the EnzChemRED dataset, we defined papers that provided experimental evidence as those linked to an evidence tag with an experimental evidence code from the Evidence and Conclusions Ontology (ECO), such as “ECO:0000269”, and excluded abstracts of papers linked to Rhea reactions that involve proteins and other macromolecules, such as DNA. This exclusion criterion strongly reduced the propensity of LitSuggest models to retrieve papers about signalling pathways, where proteins are modified by enzymes.We trained and tested LitSuggest models using a set of 9,055 positive abstracts split into 5 sets of 7,244 positive abstracts for training and 1,811 positive abstracts for testing, with 14,488 negative abstracts for training selected at random from PubMed using the LitSuggest curation interface. LitSuggest models provide a score of 0-1 for each abstract, with scores above 0.5 denoting that the abstract is relevant (belongs to the same class as the positive training data). The five LitSuggest models had a mean recall of 98% when tasked with classifying the 1,811 abstracts left out. To test precision and recall using more realistic ratios of relevant and irrelevant literature, we also performed “spike-in” tests that mixed 250 relevant papers (from the set of 1,811 abstracts left out) with a set of 100,000 abstracts selected from PubMed using NCBI e-utils (so a ratio of 400:1 irrelevant to relevant abstracts). At a score threshold of 0.8, our best performing LitSuggest model had precision of 90.1%, sensitivity of 94.8%, and F1 score of 92.4% in these “spike-in” tests. Swiss-Prot curators now use this LitSuggest model as a tool to triage literature for curation of protein sequences in UniProtKB/Swiss-Prot with reactions from Rhea on a weekly basis. It is available at https://www.ncbi.nlm.nih.gov/research/litsuggest/project/5fa57e75bf71b3730469a83b. We also used this model in the first step of our end-to-end pipeline for enzyme function extraction from PubMed abstracts.Named entity recognition (NER)The goal of Named entity recognition (NER) is to identify chemical and protein mentions in text. The NER task is framed as one of sequence labelling; text is represented as a sequence of tokens \(x=({x}_{1},{x}_{2},\ldots ,{x}_{n})\), where n denotes the length of the text, and the goal is to classify the sequence of tokens x into a corresponding label sequence \(y=({y}_{1},{y}_{2},\ldots ,{y}_{n})\), where \({y}_{i}\in Y\). Y is a label set for the model, and each label represents the entity type and position in the text. A common scheme for label sets in biomedical NER is the IOB2 tagging scheme, where IOB stands for “inside, outside, beginning”. In IOB2 the label set Y consists of the “O” label, which denotes a text chunk that is “outside” the tagging schema (so not a chemical or protein entity in the case of EnzChemRED), and entity type labels, with the prefixes “B-”, denoting the beginning of the entity token, and “I-”, denoting the inside of the entity token. So, under the IOB2 schema, a mention of a “membrane fatty acid” would be labelled as “membrane [O] fatty [B-Chemical] acid [B-Chemical]”.For the EnzChemRED task we used AIONER51 (https://github.com/ncbi/AIONER), an all-in-one scheme-based biomedical NER tool that integrates multiple datasets into a single task by adding task-oriented tagging labels, allowing the model to learn the synonyms present in texts covering multiple subjects. AIONER performs optimally on 14 BioNER benchmark datasets, such as BioRED, BC5CDR52, GNormPlus/GNorm253, NLM-Gene54, and NLM-Chem55. It has been shown to achieve competitive performance with Multi-Task Learning (MTL) methods for multiple NER tasks, while being more stable and requiring fewer model modifications. AIONER replaces the “O” (outside) label with a specific form of O-label for each NER dataset, such as “O-Gene” for the task of gene finding in the case of GNormPlus/GNorm2 and NLM-Gene. For EnzChemRED we use “O-Reaction” to signify the NER task, giving five labels for our dataset, namely “O-Reaction”, “B-Gene”, “I-Gene”, “B-Chemical”, and “I-Chemical”. So, under the IOB2 schema used by AIONER to integrate EnzChemRED, our mention of a “membrane fatty acid” would be labelled as “membrane [O-Reaction] fatty [B-Chemical] acid [I-Chemical]”.We illustrate the training process for the AIONER model using EnzChemRED in Fig. 6. We employed spaCy (https://spacy.io/) for sentence detection and tokenization, allowing us to convert our entire biochemical reaction text dataset for AIONER. Once the dataset was converted to the AIONER representations of input and label sequence, we optimized our model using the fine-tuning script provided by AIONER on GitHub. A similar procedure was followed during the testing phase without inputting the label sequence. We tested four different pre-trained language models (PLMs) for NER, namely Bioformer56, PubMedBERT28, AIONER-Bioformer51, and AIONER-PubMedBERT51. We performed 10-fold cross validation for each model using EnzChemRED, fine-tuning the PLMs using the training set partition and evaluating them on the test set partition.Fig. 6Overview of the fine-tuning process for AIONER on EnzChemRED.Named entity normalization (NEN)Named entity normalization (NEN) takes the entities identified during NER a step further. It aims to determine the exact meaning of each mention in context by mapping it to a unique identifier from a knowledgebase or ontology such as UniProtKB (for proteins) or ChEBI (for chemical entities). This process helps clarify and standardize the entities detected in text and is an essential step in transforming natural language into a structured knowledgebase.NEN can be formulated as follows: given a named entity e in context and a lexicon L (essentially a list of IDs and their corresponding synonyms, where an ID can have multiple synonyms), the goal is to find the unique ID in L for e. For EnzChemRED we used the Multiple Terminology Candidate Resolution (MTCR) pipeline to map chemical mentions in abstracts to ChEBI and MeSH IDs. MTCR is a structured approach for linking entities in the biomedical domain, including chemical terminologies; a similar process, referred to as sieve-based entity linking, has been described by D’Souza and Ng57. There are three main steps in the MTCR pipeline: abbreviation resolution, candidate lookup, and post-processing.
1.
During the abbreviation resolution phase, the pipeline identifies pairs of short and long forms in each document using the Ab3P Abbreviation Resolution tool58. Short forms are then expanded into long forms before looking up (so “TPP” to “triphenyl phosphate”).
2.
The candidate lookup step involves finding the candidate IDs of e in lexicon L. The process starts with a precise lookup and proceeds to higher recall queries, stopping once a match is found. The steps are as follows: (1) search for e in lexicon L; (2) lower the case of e and strip non-alphanumeric characters from it, creating ep, then search for ep in lexicon L; (3) stem ep, and search for stemmed forms of ep in lexicon L; (4) search for ep in Lall and map to L, where “Lall” refers to all lexicons; (5) Search for stemmed e in Lall and map to L. Mappings can take place in two ways: ‘single’, where terms are mapped directly to the target using cross references in the two lexicons, and ‘pivot’, where terms are mapped directly to the target through identifiers shared with another lexicon, such as the International Chemical Identifier (InChI) (or it’s hashed form, the InChIKey) (https://www.inchi-trust.org)59, the SMILES (Simplified Molecular-Input Line-Entry System) (http://opensmiles.org), a linear notation for chemical structures, the Chemical Abstracts Service (CAS) number, or others.
3.
Post-processing removes annotations for mentions identified as non-chemical, which is particularly relevant when the target terminology is broad, such as MeSH. Ambiguous mentions (those with two or more potential target identifiers) are resolved using unambiguous mentions.
The MTCR pipeline has been benchmarked for the BioCreative VII NLM Chem task60 with MeSH as the target terminology. BlueBERT27 showed higher NER performance, but MTCR demonstrated outstanding NEN performance, with a precision of 81.5% and a recall of 76.4%, with only 29% of teams outperforming MTCR in NEN.Relation extraction (RE)We frame the problem of extracting pairs of chemical reactants as one of relation classification. There are two tasks, binary pair classification, for (chemical-chemical) relations that link reaction participants, and ternary tuple classification, for (protein-(chemical-chemical)) relations, that link chemical reactants, and the enzymes that catalyze their conversion.
For the task of binary pair classification for (chemical-chemical) relations, given a chemical mention pair (c1, c2) and the corresponding sentence s, the objective is to classify the relation type r of the chemical pair (c1, c2).
For the task of ternary tuple classification for (protein-(chemical-chemical)) relations, given a protein mention p, a chemical mention pair (c1, c2), and the corresponding sentence s, the objective is to classify the relation type r of the ternary tuple (p (c1, c2)).
Valid relation types for binary pairs and ternary tuples are “Conversion”, “Indirect_conversion”, and “Non_conversion”, which are curated (see Section Curation of chemical conversions), and “None”, which is assigned automatically during evaluation. For binary pairs, “None” is assigned to all pairs of chemical mentions that are not curated using one of the three valid relation types. For ternary tuples, “None” is assigned to all ternary tuples that include pairs of chemical mentions not curated using one of the three valid relation types, and to all ternary tuples that include pairs of chemicals curated with a valid relation but that also include a protein mention that was not linked to them by a curator (i.e. it is not the enzyme responsible).We performed relation classification using PubMedBERT and BioREx61, which is a PubMedBERT model trained on the BioRED dataset and eight other common biomedical RE benchmark datasets (PubMedBERT is essentially the same model but without this additional training step). BioREx offers a reliable and effective approach to chemical reaction extraction and has shown consistently high performance for relation classification across seven different entity pairs. In PubMedBERT and BioREx, each input sequence \(x=({x}_{1},{x}_{2},\ldots ,{x}_{n})\) is prefixed with a special [CLS] token xCLS. This token is processed through the neural network layers of the model along with the sequence, and the output corresponding to the xCLS token is a high-dimensional vector that aggregates, or summarizes, the contextual information from the entire sequence. We denote the output embedding of the [CLS] token as hCLS.To adapt the [CLS] vector hCLS for relation classification, it is passed through a linear neural network layer, which aims to map the high-dimensional hCLS vector into a lower-dimensional vector space suitable for the classification task. In our case, this is a four-dimensional vector, with each dimension corresponding to one relation type label – Conversion, Indirect_conversion, Non_conversion, or None. Mathematically, we can express the operation of the linear layer as:$$r=W\times {h}_{{CLS}}+b$$where W is the weight matrix, b is the bias vector, and r is the output vector. The length of r matches the number of classes in our task, which is four.Fig. 7 illustrates the process of fine-tuning on EnzChemRED using BioREx. EnzChemRED is annotated at both mention and sentence levels, with locations specified, unlike BioRED, which uses document-level annotation, and where relations are given in the format of ID pairs without specifying the exact locations of the entity mentions involved. We therefore adjust the fine-tuning procedure used in BioRED, replacing the [Corpus] tag with a [Reaction] tag, and using individual sentences as input rather than full documents.Fig. 7An illustration of the specific input representation for the EnzChemRED dataset and the fine-tuning process of the BioREx model.Fig. 8 illustrates an example input representation of a ternary tuple. The classification of ternary tuples follows similar rules to that for binary pairs. We insert additional boundary tags, “<P>” and “</P>”, to denote the enzyme in the input instance, but otherwise follow the same procedure as for binary pair RE. As with binary pairs, ternary tuples are annotated at both sentence and mention levels, such that if the same enzyme appears more than once in a sentence, each occurrence would be treated as a different instance.Fig. 8An example of the input representation for a ternary tuple in EnzChemRED.End-to-end pipelineWe combined the best performing methods for NER (AIONER-PubMedBERT, fine-tuned using EnzChemRED) and RE (BioREx, fine-tuned using EnzChemRED) with MTCR for chemical NEN to create a prototype end-to-end pipeline for enzyme function extraction from text. We applied this pipeline to EnzChemRED abstracts for cross validation purposes, and to relevant PubMed abstracts (up to December 2023) identified using the LitSuggest model described in the Section Literature triage to map enzyme functions in literature. The latter necessitated comparison of chemical pairs extracted from PubMed abstracts to pairs of chemical reactants from Rhea, which was accomplished as follows. To create the set of Rhea pairs for comparison, we extracted pairs of chemical reactants from Rhea using a heuristic procedure in which we removed the top 100 most frequently occurring compounds in Rhea reactions such as water, oxygen, and protons, and then enumerated all possible pairs of the remaining compounds within each Rhea reaction. We also removed pairs of identical ChEBI IDs from the Rhea set, which in Rhea can occur as part of transport reactions. To prepare the chemical pairs extracted from PubMed abstracts for comparison to Rhea, we first normalized their ChEBI IDs to those representing the major microspecies at pH7.3, the form used in Rhea, removed pairs that include any of the top 100 most frequently occurring compounds in Rhea reactions, and removed pairs where both members had the same ChEBI ID. This can occur due to errors in NER and NEN, where erroneous text spans can cause distinct but related chemical names to be mapped to the same identifier. After processing we compared the degree of overlap in the two sets (chemical pairs from Rhea reactions and PubMed abstracts) using their ChEBI IDs.Visualization of chemical pairs from PubMed abstracts and RheaTo visualize chemical pairs from PubMed abstracts and Rhea we used the Tree Map (TMAP) algorithm to create TMAP trees using code from http://tmap.gdb.tools as described62, clustering chemical pairs in TMAP trees according to their Differential Reaction Fingerprint (DRFP), calculated according to the method of Probst63. We used the degree of atom conservation between the members of each chemical pair to filter the output of our end-to-end NLP pipeline. To calculate atom conservation, we first converted molecular structures into graphs by replacing all bond types with single bonds. This ensures a standardized representation of molecular structures, simplifying subsequent analyses. We then computed the Maximum Common Substructure (MCS) using the rdkit.Chem.rdFMCS.FindMCS function (from the open-source cheminformatics toolkit RDKit, at www.rdkit.org) with a permissive ring fusion parameter. The MCS represents the largest common atomic framework shared by the two molecules (after conversion into a graph of atoms linked by single bonds). The atom conservation is the average of the percentage of common atoms, as given by:$$ \% \,\mathrm{atom\; conservation}=1/2\times ({n}_{{MCS}}/{n}_{L}+{n}_{{MCS}}/{n}_{R})\times 100$$where nMCS is the number of atoms in the maximum common substructure, nL is the number of atoms in the molecule on the left side of the pair, and nR is the number of atoms in the molecule on the right side of the pair. This metric provides a standardized measure of structural similarity, facilitating the comparison of chemical compounds in each pair.