Audris Mockus links 69.8M edges between research papers and code repositories

A new cross-corpus analysis from Audris Mockus directly links the world of scientific papers to the world of public software repositories for the first time at global scale. By connecting the World of Code (a near-complete mirror of version-control history) to Semantic Scholar and OpenAlex, the researcher built a typed graph of 69.8 million edges spanning eight relation types—paper-to-software mentions, software-to-paper citations, software dependencies, authorship, affiliation, and identity bridges. The study anchors on 18,247 curated science repositories and asks two reciprocal questions: what is the impact of science on software, and what is the impact of software on science?

The results reveal two distinct strata. Literature's reach into software is dominated by a reproducibility and packaging layer (nf-core, Nextflow, Bioconda) and sequence-analysis tools, reflecting scientists' need to make computational methods discoverable and reusable. In contrast, software's reach back into science is proxied by an invisible machine-learning and data-science infrastructure tier—PyTorch, seaborn, NLTK—that powers modern research without always being explicitly cited. Critically, direct paper-names-software links are too sparse to rank: a human-curated gold benchmark found none of its 65 in-scope cases. Using dependency reuse as a proxy yields only a weak coupling to citation count and stars (Spearman rho=0.36). Most cautionary, the sign of the reuse–citation correlation flips depending on how one pairs a repository with citation data—through papers naming it (n=137, rho=0.05, CI straddling zero) versus DOIs a repo declares (n=1,067, rho=0.13, CI [0.07,0.19]). The study refrains from strong claims, underscoring that measurement choices can reverse headline correlations.