A computational pipeline designed to connect zebrafish genes with evidence from corresponding human genes retrieved relevant evidence for 99.33% of 150 benchmark queries, according to an arXiv preprint. Its abstract reports two ranking measures, NDCG@10 at 0.9879 and mean reciprocal rank, or MRR, at 0.99, using an evidence-weighting setting called lambda 0.50. It also reports that 80% of retrieved documents were exclusive to a single query. The paper's stated limitations include the need for biologically verified reference answers.
A cross-species search for evidence
Called Homo-RAG, the framework starts with a target zebrafish gene, uses its corresponding human ortholog as a cross-species bridge, and then retrieves functional and literature evidence before applying retrieval-augmented generation. The retrieval stage combines dense semantic retrieval with sparse lexical BM25 retrieval and can add exact-gene matching when gene identity information is available.
The reconstructed retrieval engine contained approximately 17,224 gene-profile documents, 17,259 UniProt documents and 32,909 PubMed documents. The evaluation used 150 queries, 7,200 retrieved evidence instances and 951 frozen gold evidence pairs, with 672 gold pairs present in the candidate pool. The protocol froze gold-membership labels before ranking and did not supply them to the ranker.
After retrieval, the evidence score gave 50% weight to semantic score, 20% to gene match, 15% to ortholog match, 10% to source reliability and 5% to PubMed support. The paper also varied lambda and reported different results at different settings.
Strong scores, but an unresolved reporting mismatch
The detailed sensitivity results report the highest NDCG@10 at lambda 0.30, with a score of 0.9879. At lambda 0.50, the analysis lists P@10 of 0.8856, R@10 of 0.9343, NDCG@10 of 0.9858 and MRR of 0.990. At lambda 1.00, NDCG@10 is 0.8204 and MRR is 0.903. But the abstract assigns NDCG@10 of 0.9879 to lambda 0.50, leaving the paper's best setting unclear.
At a top-10 cutoff, the preprint reports precision of 0.7981, recall of 0.7430, Hit of 0.9933 and NDCG of 0.8119. A separate progressive comparison gives the hybrid BM25+FAISS system P@10 of 0.38, R@10 of 0.73, NDCG@10 of 0.83 and MRR of 0.87. The Hybrid+3-Hop configuration without ECS is listed at 0.4187, 0.9361, 0.9805 and 0.980 on the same measures, while the full ECS lambda 0.50 row is listed at 0.8856, 0.9361, 0.9879 and 0.990. The differing full-system figures across sections make the results difficult to compare as one clean sequence.
Source-removal tests also produced large differences. With all sources present, the reported P@10 was 0.8856, R@10 was 0.9361, Hit@10 was 0.9933, NDCG@10 was 0.9805 and MRR was 0.980. Without Gene Profile, P@10 was 0.3353; without UniProt, it was 0.3400; and without PubMed, it was 0.1700, with R@10 of 0.7233. These computational comparisons show that the reported scores varied with the source configuration, but do not establish that any one source caused the change.
The remaining test is biological verification
The text-generation stage was compared across five lightweight language models. Phi-3.5-mini-Instruct had the highest reported BERTScore at 0.87, semantic similarity at 0.84, Distinct-1 at 0.19 and Distinct-2 at 0.44. TinyLlama-1.1B-Chat had the lowest corresponding values, at 0.75, 0.72, 0.12 and 0.28. Reported success rates ranged from 0.980 to 1.000, while listed generation latency ranged from 1.8 to 7.5 seconds; Phi had the highest listed latency.
The authors describe several reasons to treat the result as a work in progress. The final evidence score has only one validated evidence-specific feature and uses fixed ranking values, while the tested models are open-source rather than frontier LLMs. They call for richer evidence features, claim-level attribution, cross-source consistency checks and biologically verified reference answers. The paper therefore leaves agreement with verified biological functions as an open requirement.
The manuscript is an arXiv preprint, version 1, dated 26 August 2026. Its inputs came from publicly accessible ZFIN, UniProt and PubMed resources. Processed datasets and data-processing procedures are available from the corresponding author on reasonable request, subject to the original source terms, while the code and implementation are promised for release in a public repository upon manuscript acceptance. The manuscript reports no specific grant and no known competing financial interests or personal relationships; it says ethics approval and consent were not required because the study used public databases and literature and involved no human participants, animals or personal data.
Paper data and sources
Original title: Homo-RAG: Homology-Guided Retrieval-Augmented Generation for Cross-Species Gene Function Prediction
Authors: Azrin Sultana
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text