EVA

Overview

EVA (Evolutionary Versatile Architect) is a generative RNA foundation model from the GENTEL Lab at the Chinese Academy of Sciences, posted to bioRxiv in March 2026. EVA is trained on OpenRNA v1, a curated corpus of 114 million full-length RNA sequences spanning all domains of life, and is the first unified architecture targeting the breadth of functional RNA design tasks: transfer RNAs, aptamers, CRISPR guide RNAs, messenger RNAs, and circular RNAs.

Unlike earlier RNA language models such as RNA-FM or Uni-RNA, which are primarily encoders for downstream classification or regression, EVA is generative — it samples RNA sequences from learned distributions conditioned on task-specific context. The model achieves state-of-the-art results on 7 of 9 public RNA design benchmarks, outperforming RfamGen, GenerRNA, and conventional structure-guided design tools.

Key Features

• Long-context generation: Trained at full-length resolution rather than fixed windows, enabling design of long structured RNAs (lncRNAs, circRNAs, full mRNAs).
• Unified architecture for multiple RNA classes: One model handles tRNA, aptamer, gRNA, mRNA, and circRNA design without per-class specialization.
• OpenRNA v1 training corpus: 114M curated full-length RNA sequences, broader and deeper than corpora used in prior RNA FMs.
• State-of-the-art design benchmarks: SOTA on 7 of 9 RNA design tasks evaluated, including tRNA scaffold design and aptamer generation.
• Domain-balanced training: Training corpus spans bacterial, archaeal, and eukaryotic RNAs to avoid the human-centric bias of prior models.

Technical Details

EVA uses a decoder-only transformer architecture trained autoregressively on the OpenRNA v1 corpus. Sequence tokenization operates at the nucleotide level. The training objective is standard next-token prediction; conditional generation is supported via prefix prompting with class tokens or structural constraints. The preprint describes ablations on context length, training data filtering, and class-balancing strategies.

Benchmarks reported in the preprint include tRNA acceptor-stem design, theophylline aptamer generation, CRISPR-Cas9 guide RNA on-target activity, mRNA codon optimization, and circular-RNA scaffolding. EVA outperforms prior task-specific tools on 7 of 9 evaluated benchmarks.

Applications

EVA is suited for synthetic biology and RNA therapeutics groups that need to design functional RNAs without committing to a separate task-specific tool per RNA class. In therapeutic mRNA design, it provides a generative alternative to rule-based codon optimization. In aptamer engineering, it can propose candidate sequences that meet structural and binding constraints. For CRISPR applications, it offers guide-RNA design with predicted on-target activity informed by evolutionary context.

Impact

EVA is the first RNA foundation model to span the breadth of functional RNA design as a generative system rather than as a downstream encoder. By demonstrating SOTA on 7 of 9 tasks within a single unified model, it argues for foundation-model approaches in RNA design analogous to the protein-design trajectory established by ProGen and ESM-3. The 114M-sequence OpenRNA v1 corpus is itself a valuable community resource for further work in RNA foundation modeling.