Stanford Team Releases Minerva, a Genome Language Model Uncovering Novel Reverse Transcriptase ncRNA Systems

David Li and Garyk Brixi, working under Brian Hie's guidance at Stanford, released Minerva, a scientific discovery framework driven by a genome language model, with the preprint "Coevolutionary mining of prokaryotic non-coding elements with a genome language model" now public on bioRxiv. Using the method, the team discovered a previously unknown class of structurally shared, diverse ncRNA arrays in the UG27 reverse transcriptase system, demonstrating a path for AI to systematically discover genetic elements and mine the "dark genome" at scale.

Confirmed

• Minerva is a genome-language-model-guided discovery method by David Li and Garyk Brixi under Brian Hie's supervision, with acknowledgments to advisor Michael Grinstein and collaboration from the Michael Fischbach lab; multiple posts note David Li was previously at Arc Institute and is now at EvolutionaryScale.
• Per Anshul Kundaje, the core idea is training a language model on massive genome sequences and performing "coevolutionary mining": compressing the genomic corpus into the model, then using it to generate scalable functional and structural hypotheses about the dark genome; per m18, the method points to hundreds of new candidate elements per bacterial genome.
• Per m18 and m20, Minerva builds on the gLM2 genome language model, combining categorical Jacobian methods with protein contact heads.
• Per m5–m7, the model predicts RNA base pairing, repeats, and other interactions from sequence alone, motivated by the many undiscovered bacterial non-coding RNAs and structural elements; it also surfaced new candidate RNAs and repeat elements.
• Core finding: through large-scale coevolutionary mining across prokaryotic genomes, Minerva found that the UG27 reverse transcriptase system encodes diverse variable ncRNA arrays with shared structures, each ncRNA templating a short DNA hairpin—revealing previously unknown interactions in non-coding DNA and a novel reverse-transcriptase-associated system.

Why it matters

• According to Brian Hie, the work shows AI can not only accelerate biological design but also systematically discover new genetic elements and reverse transcriptase mechanisms, offering a reusable framework for genome data mining. Multiple community voices hailed it as a significant advance in understanding genomic "dark matter" and another case of language models penetrating fundamental biology.

2026-09-23 ~ 2026-09-24 · 24 related posts

• [source] Genome language models uncover new class of reverse-transcriptase mechanisms — BrianHie · 2026-09-23
• Minerva paper details: ncRNA arrays template short DNA hairpins — BrianHie · 2026-09-23
• [source] Minerva: Hie lab's data-mining framework for scientific discovery — BrianHie · 2026-09-23
• [source] Minerva framework preprint is out, says Brian Hie — BrianHie · 2026-09-23
• Minerva: genome language models uncover diverse structured ncRNA arrays in biology — BrianHie · 2026-09-23
• Minerva: genome language models uncover co-evolutionary systems across prokaryotes — BrianHie · 2026-09-23
• Minerva: genome language models uncover variable ncRNA arrays in UG27 reverse transcriptase systems — BrianHie · 2026-09-23
• Minerva mines the dark genome with genome language models — anshulkundaje · 2026-09-23
• Genome Language Model Minerva Uncovers a New Class of Reverse Transcriptase — JosephJacks_ · 2026-09-23
• Minerva: Genome Language Models Uncover Hidden Elements in Non-Coding DNA — BrianHie · 2026-09-23
• Minerva genome-scale model builds on gLM2 with categorical Jacobian and protein contact heads — BrianHie · 2026-09-23
• Arc Institute unveils Minerva, a genome language model for predicting RNA interactions — BrianHie · 2026-09-24
• Stanford's Minerva uses genome language models for biological discovery — BrianHie · 2026-09-24
• Genome language model uncovers hidden interactions in bacterial non-coding DNA — KevinKaichuang · 2026-09-24
• Stanford team mines bacterial non-coding elements with a genome language model, finds new RNAs — KevinKaichuang · 2026-09-24

9 near-duplicate retellings: BrianHie · BrianHie · BrianHie · BrianHie · BrianHie · BrianHie · BrianHie · anshulkundaje · BrianHie