Domain-wide Mapping of Peer-reviewed Literature for Genetic Developmental Disorders using Machine Learning and Gene2Phenotype

Genetically determined developmental disorders (GDD) are rare, heterogeneous conditions for which clinical diagnosis increasingly depends on genomic variant prioritisation and rapid synthesis of genotype-phenotype evidence scattered across the literature. Manual curation of this evidence is labour-intensive, difficult to scale and to keep up to date. We present a domain-wide, automated pipeline that identifies PubMed abstracts describing human case reports/series and maps them to molecularly defined diseases in the Gene2Phenotype database (G2P). The natural language processing system combines a fine-tuned BERT classifier (LitDD BERT) to detect GDD-relevant abstracts, a fine-tuned cross-encoder (LitDD Crossencoder) to propose disease candidates, and a constrained large language model to adjudicate final mappings. Trained on 13,738 annotated title-abstract pairs spanning 231 genes and 354 G2P entries, LitDD BERT achieved an F1 of 0.89 (precision 0.83, recall 0.94). The cross-encoder reached 0.99 top-5 recall, and the full ensemble attained precision 0.89, recall 0.82, and F1 0.85 on a held-out test set. Applied to the entirety of PubMed, the pipeline identified approximately 69,000 manuscripts which could be mapped to G2P diseases. Against independent manually curated sets, it retrieved about 70% of manuscripts at both micro and macro levels, indicating generalization across diverse disease mechanisms, differing curation standards and inheritance patterns. The resulting corpus is accessible through G2P and is being integrated into routine biocuration. This work delivers scalable, updateable literature surveillance for GDD, enabling faster evidence review in a diagnostic setting, supporting bioinformatic pipelines, and laying a foundation for downstream case-level extraction and standardized phenotype integration.