Search plan

The search plan included a scoping phase during which we conducted a string search on Scopus using the following terms: (systematic mapping OR systematic literature review OR systematic review) AND (indirect driver*) AND (biodiversity loss) (S2 Table); however, this search yielded no relevant results. A subsequent search specifically targeting indirect drivers of biodiversity loss, or indirect threats, resulted in fewer than 400 articles, indicating a significant underrepresentation of knowledge regarding drivers of biodiversity loss and their remote sources. For this second preliminary search, we used the following search string: (indirect) AND (threat* OR driver* OR pressure*) AND (biodiversity OR biodiversity loss). Consequently, we opted to examine documents addressing direct drivers of biodiversity loss, assessing whether the authors acknowledged the presence of indirect sources in terms of typology and spatial scale. Because our approach emphasizes direct drivers over indirect ones, we recognize that authors of eligible papers might briefly mention indirect drivers to provide context for a direct driver of biodiversity loss. To facilitate the mapping process, we established a benchmark list of fifteen articles to guide our search and evaluate the comprehensiveness of our findings (S2 File). These fifteen articles were selected to be representative of a broad spectrum of research methodologies, geographic scales, ecoregions, ecological levels of organization, and direct and indirect drivers. We extracted official and unofficial keywords from the titles and abstracts of these articles and organized them into thematic terms. Search strings were then run on OpenAlex and the resulting retrieved documents were compared (S2 Table). OpenAlex is an open-access bibliographic catalogue of scientific articles, authors, institutions, and other bibliographic metadata records including more than 250 million works such as books, datasets, and journal articles [32]. It has been compared with databases behind a paywall (e.g., Scopus, Web of Science) and search engines (e.g., Google Scholar, Semantic Scholar), showing a good degree of article coverage, availability, and metadata records [33,34]. Candidate search strings were tested based on the number of search terms and retrieved documents. Each search attempt was evaluated for comprehensiveness based on its capacity to yield the full list of fifteen benchmark articles (S2 File). Because we will bulk download an extensive literature base and filter it during successive stages, we choose the search string with the maximum combination of relevant broad words, which is the following:

(pressures OR threats OR footprint OR risk OR drivers OR stress OR impacts OR loss) AND (species OR ecosystems OR wildlife OR habitat) AND (hotspot OR map OR spatial OR direct OR indirect OR feedback OR long-distance OR human OR anthropogenic OR actions) AND (biodiversity OR conservation)

We will restrict sources to literature published between 1999 and 2024 published in English, Spanish, French, and Italian due to the authors’ fluency in these languages, which ensures accurate comprehension and analysis. Focusing our systematic literature mapping on the past 25 years will ensure relevance by incorporating current theories, methodologies, and technologies, highlighting contemporary trends and gaps, and making the review process more manageable. Firstly, we will use the IUCN Red List threat classification scheme [12] to categorize existing direct drivers of biodiversity loss. Secondly, we will manually compile relevant information on impacted organisms, taxonomic groups, ecological levels of organization, countries, and ecoregions. Thirdly, we will manually collect data on indirect drivers when explicitly mentioned (i.e., observed or recorded), hypothesized (i.e., modeled on an existing and observed threat), or referred to from another source (i.e., external citation). The outcome to be assessed will be the presence of an explicit term indicating indirect drivers of anthropogenic direct drivers of biodiversity loss. We will classify human-driven indirect factors following Díaz et al. [13]. For our research purposes, we will consider the terms ‘indirect drivers (of biodiversity loss)’ and ‘indirect threats (to biodiversity)’ as synonymous. Similarly, we will also treat as synonyms the terms related to ‘direct drivers (of biodiversity loss)’, namely ‘direct threats (to biodiversity)’, ‘threats’, ‘IUCN threat categories’, ‘sources of stress’, and ‘proximate pressures’. When possible, we will investigate the spatial source of indirect drivers differentiating between those acting at the same spatial scale as direct drivers and those that are not.

We anticipate that data collection will be finalized between September and October 2026. This timeline accounts for the comprehensive search and screening process required for a systematic map. Following the completion of data collection, we expect to analyze the data and prepare the results for publication. The results will likely be available between January and February 2027. This period includes time for data synthesis, interpretation, and manuscript preparation. We acknowledge that unforeseen delays may arise during the systematic mapping process. Potential challenges include slower-than-anticipated full-text screening due to limited team availability, a larger-than-expected volume of search results, and difficulties in obtaining abstracts or full-text articles. To mitigate these risks, we have incorporated buffer periods into the timeline, we will conduct regular team check-ins to monitor progress, and we will reassign tasks or adjust workloads as necessary. In addition, we are using automation tools to support screening and data management, which we anticipate will improve efficiency and help prevent process bottlenecks. All data generated during the systematic mapping process, including the bibliographic database, screening stages, and coding extraction strategy, will be made publicly available via Zenodo, in line with open science principles. Copyrighted materials such as full-text PDFs will not be shared publicly but will be referenced appropriately. We emphasize both rigorous data stewardship and transparent authorship practices. Detailed records of data collection, coding, and extraction strategies will be maintained and made accessible to ensure reproducibility. Sensitive information will be anonymized in accordance with ethical guidelines. Automation tools used in screening and data management will be validated through systematic manual checks, and all updates to the dataset or methodology will be carefully documented to maintain a clear audit trail.

Bibliographic databases and data management

A structured full-text search will be conducted exclusively on OpenAlex [32], focusing on academic literature that includes peer-reviewed research articles and book chapters (i.e., “journal article” and “book chapter” types). Given the extensive nature of the literature base, we will exclude additional sources such as organizational websites or grey literature. The collected data will undergo filtering at various stages to minimize noise in the literature base. First, we will concentrate on primary research published in English, Spanish, French, or Italian, from 1999 to 2024. We will not consider retracted articles, non-article types, articles missing reference lists. duplicate records, and articles without a DOI. Second, we will eliminate all duplicates. Additionally, the initial benchmark list has been expanded to include 182 papers known to be relevant to the authors, which will be utilized to train an automated code for filtering the literature retrieved from OpenAlex (S3 Table) [35]. All 182 papers were manually screened at the title and abstract level and confirmed as eligible for inclusion. This curated set constitutes the positive training dataset used to calibrate and evaluate the automated screening that will be applied to the larger corpus retrieved from OpenAlex. The benchmark articles were selected to be representative of a broad spectrum of research methodologies, geographic scales, taxonomic groups, ecological levels of organization, and direct and indirect drivers of biodiversity loss.

Filtered articles will be downloaded using the PyAlex Python library [36]. In Python (v3.7.12), we will apply the following exclusion criteria: (i) retracted articles, (ii) non-article types (e.g., editorials, datasets, reports, books, letters), (iii) articles missing reference lists, (iv) duplicate records, and (v) articles without a DOI will be excluded. Title and abstract screening will be conducted using Natural Language Processing (NLP), applying Term Frequency–Inverse Document Frequency (TF-IDF) vectorization [37–39] and cosine similarity [40]. First, TF-IDF will be trained on the benchmark article titles and applied to all article titles, using a cosine similarity threshold of 15%. For records lacking abstracts, we will attempt automated extraction using tools such as citationchaser [41] and crossref-commons [42], and, if unsuccessful, we will proceed with manual recovery of available abstracts. A second TF-IDF model will then be trained on benchmark abstracts, using a 13% cosine similarity threshold for screening remaining records. If the reported set thresholds do not fully capture the benchmark articles, they will be adjusted accordingly. Finally, full-text articles will be manually screened for inclusion. All data processing will be conducted using dask v2022.2.0 [43], pandas v1.3.5 [44], numpy v1.21.6 [45], nltk v3.8.1 [46], and scikit-learn v1.0.2 [47]. All scripts and workflows will be made available in the project’s GitHub repository (Fig 1). After the NLP-assisted title and abstract screening, each review will be performed by human reviewers rather than deep-learning models. The rationale is to ensure interpretative accuracy and maintain transparency in classification decisions. While AI-based semantic approaches, such as embeddings or transformer models, can facilitate or fully automate the screening process, these would inevitably introduce complexity and potential bias that we aim to avoid for this study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow for literature search, screening, and data retrieval.

Flow of articles through the screening process, following ROSES reporting standards for systematic evidence synthesis [31] and incorporating machine learning–assisted screening (adapted from Avila-Ortega et al.[35]), including OpenAlex retrieval, metadata filtering, TF-IDF and cosine similarity filtering, title/abstract screening, and full-text selection.

https://doi.org/10.1371/journal.pone.0341928.g001

Article screening and study inclusion criteria

Data extraction

Each article included at the full-text stage will be manually coded using a data collection sheet developed by Ridley et al. [16], which will be modified to align with the current research objectives. Key descriptive information regarding the presence of indirect drivers of threats to biodiversity will be examined and coded. Following Díaz et al. [13], each IUCN threat category will be linked to the main groups of direct drivers of biodiversity loss. The data sheet’s usability and clarity were assessed using the test list of fifteen articles (S2 File). A second reviewer will code 10% of the papers selected randomly. Any discrepancies will be discussed, and decisions regarding the appropriate coding will be made collaboratively by the reviewers. Data will be collected from articles on bibliographies, study design, direct and indirect threat measurement, ecological levels of organization, and the populations investigated. To be coded, the data must be present in the article or supplementary materials. Authors will not be contacted for missing information.

Data synthesis and presentation

This study seeks to synthesize information from eligible studies through a systematic mapping exercise. It will provide a comprehensive database of the reviewed literature, along with coded metadata, as supplementary material. We will derive summary statistics from the extracted data and organize them by ecoregion, ecological levels of organization, study scale, and employed methodological approaches. The study will emphasize the presence of indirect drivers of biodiversity loss and their inherent geographic features. We will include detailed information on the progression of articles through the mapping process, specifying the number of papers sourced and those excluded at each stage. We will produce a geographic map to illustrate the global coverage and density of study areas and the distribution of studies across various spatial scales. We will summarize and visualize key extracted information as well as gaps in knowledge about indirect drivers of biodiversity loss and assess relationships among taxonomy, ecological levels of organization, direct and indirect threats studied, and geographic distribution. Finally, we will propose guidelines for future reporting, insofar as permitted by the scope of the extracted data.