CESAR: LLM screening plus citation chaining automates drug-discovery literature reviews
- CESAR is an open-source workflow that builds a literature corpus by chasing references and citing papers through OpenAlex, then has LLMs screen abstracts, calibrated against human-annotated sets.
- Tested on a broad field (total synthesis) and a narrow target (MAGL), retrospectively against curated references; the corpus recovered most known papers and cheaper models came close to costly ones.
- An open-weight model run locally matched cloud results while also extracting selectivity and compound names, allowing confidential screening. Preprint, not peer reviewed.
💡 Points Based on the abstract
Recovering most known papers is a recall figure measured against curated lists that were probably built from the same literature and citation links, so it says little about what the chain misses. Before trusting it on your target, hand-screen a random sample of the rejected abstracts to estimate false negatives, since that is where a landscape quietly goes wrong.