Original title: Affinage: genome-scale mechanistic gene annotation from the published literature
Matteo Di Bernardo, Iain M. Cheeseman · July 2026
Biologists need to know how genes work, but this knowledge is scattered across thousands of research papers. Human-curated databases often lag years behind, and using AI on demand to read papers for every gene is costly and hard to reproduce. Researchers created Affinage to solve this. It is an AI pipeline that reads the primary literature once per gene and saves the results as a structured summary. It works in two steps. First, a reading pass pulls out direct experimental evidence. Next, a synthesis pass thinks through only those extracted findings. Affinage annotated 19,293 human protein-coding genes. It supplied mechanisms for thousands of genes where UniProt records were empty or short stubs. In a head-to-head comparison scored by an AI judge, Affinage beat the curated database on 99.1% of genes. It also identified the 10% of the proteome that still lacks mechanistic characterization.
Imagine a librarian asked to summarize every book in a massive library. But the librarian must only include facts that are directly proven by experiments inside each book, ignoring opinions or guesses. Now imagine that librarian has to do this for 19,293 different collections of books, one per gene, and also identify which collections contain no proven facts at all. In this paper, Affinage is that librarian: it reads all the research papers for each gene, extracts only the direct experimental evidence, and outputs a structured annotation.
Biologists need to know what each gene does, but that knowledge is scattered across countless papers and often out of date. Affinage reads all that literature and produces a reliable, structured summary for almost every human gene, instantly. It also flags which genes remain truly unknown, helping focus research on the biggest gaps.
The abstract notes that evaluation was done using a cross-family LLM judge, not human experts, so the 99.1% figure reflects agreement with that judge rather than absolute ground truth. Affinage is limited to human protein-coding genes and to direct experimental evidence from primary literature, so it may miss insights from indirect evidence, reviews, or computational predictions. The pipeline also identifies but does not fill the 10% of uncharacterized genes.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Affinage: genome-scale mechanistic gene annotation from the published literature — Matteo Di Bernardo, Iain M. Cheeseman
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.