EVA | Yanjie Huang 黄䶮杰

EVA model architecture diagram — Model architecture: long-context MoE Transformer with CLM and GLM generation paradigms.

EVA application concept for RNA modeling and design — Application concept: EVA as a model for RNA understanding, generation, optimization, and design.

EVA, short for Evolutionary Versatile Architect, is a generative RNA foundation model built for full-length RNA modeling and controllable RNA design. It learns from OpenRNA v1, a curated atlas of 114 million full-length RNA sequences across domains of life, and uses a 1.4B-parameter decoder-only Transformer with a Mixture-of-Experts backbone and an 8,192-token context window.

The model unifies RNA sequence scoring and generation in one framework. Causal language modeling supports de novo design and continuation, while general language modeling supports span infilling and domain redesign. EVA can condition generation on RNA type and taxonomic lineage, making it useful for designing mRNA, tRNA, rRNA, lncRNA, miRNA, circRNA, viral RNA, and other RNA classes.

What It Enables

Full-length RNA sequence modeling without aggressive truncation.
Zero-shot RNA fitness prediction and log-likelihood scoring.
Controllable generation by RNA type, species, or lineage.
De novo RNA design, domain redesign, and in-silico directed evolution.
Sequence-to-structure-aware optimization for practical RNA engineering tasks.

My Role

I developed MoE architectures with MegaBlocks and the first pretraining and mid-training code, achieving 30% lower perplexity than the dense baseline. I co-curated and released OpenRNA v1 across more than 15 RNA types, benchmarked zero-shot fitness and gene-essentiality prediction, and used sparse autoencoders to identify structure/function features (Spearman > 0.9) and steer generation.

My work connects model development with experimental validation: I designed and optimized functional RNAs and CRISPR-Cas systems, with the best designed aptamers improving both fluorescence and binding affinity by 2x over the positive control. This work informs my interest in lab-in-the-loop learning through interaction with the physical world.

Release

The public EVA repository releases training, finetuning, inference, generation, scoring, and directed-evolution code, with model and data access documented for reproducible use.