Methods (Leaderboard Construction)
The web application provides a leaderboard of high-impact scientific publications derived from OpenAlex bibliographic data. The system is designed to be topic-agnostic, enabling consistent ranking of research outputs across multiple scientific domains using a unified methodology.
Data sources and cohorts
Bibliographic records are retrieved from OpenAlex, including metadata such as publication dates, citation relationships, and document types. Analyses are organized into two fixed publication cohorts: the Class of 2025 cohort (2023 publications) and the Class of 2026 cohort (2024 publications). Only journal articles are included, and non-article document types are excluded to ensure comparability across cohorts.
Early citation impact metric
Research impact is quantified using an 18-month post-publication citation window, defined as the number of citing works published within 18 months of each paper's publication date. This metric captures early citation impact while controlling for differences in publication age and mitigating long-term accumulation bias.
Top-cited paper selection
To construct the leaderboard efficiently at scale, an exact threshold-expansion ranking strategy is applied. First, an initial high-recall candidate set is retrieved from OpenAlex by selecting papers ranked by lifetime citation counts. Lifetime citations are used as an upper-bound proxy for early citation performance, ensuring that no potentially high-impact papers are excluded at the candidate stage.
Exact 18-month citation counts are then computed for all papers in the candidate pool using citation graph traversal constrained by publication dates. A dynamic threshold corresponding to the desired leaderboard size is derived from this scored subset. The candidate pool is then expanded to include all papers whose lifetime citation counts exceed this threshold, guaranteeing completeness of the final ranked set. The final leaderboard is obtained by sorting all expanded candidates by their 18-month citation counts in descending order.
Topic relevance filtering
To ensure that retrieved papers are substantively relevant to the intended scientific domain, an additional large language model (LLM)–based relevance classifier is applied. For each candidate paper, the model is provided with the title and abstract and asked to determine whether the paper is meaningfully related to the target domain. The classifier outputs a binary decision indicating topical relevance. Only papers classified as relevant are retained for final ranking and presentation in the leaderboard.
Implementation and availability
The full pipeline is implemented in Python. Detailed methodological specifications are provided in the associated bioRxiv preprint. The complete source code, including data retrieval, ranking, and filtering components, is available in the project's GitHub repository https://github.com/pepkio/pri-top-papers.