RILLIE

RNA aptamers are usually evolved by SELEX, which selects binders in a test tube over many rounds of amplification. A sequence optimised for binding in vitro often folds badly inside a cell, so aptamers that look excellent on the bench underperform as live-cell reporters. RILLIE — "RNA In Silico Evolution via LLM and Inverse folding" — replaces those selection rounds with a scoring pass over pretrained models, arguing that an RNA language model trained on natural sequences has already absorbed the cellular folding pressure SELEX omits.

RILLIE is an orchestration over frozen foundation models rather than a newly pretrained backbone. It scores a candidate twice: once with an RNA language model, for plausibility under natural evolutionary patterns, and once with an inverse-folding model conditioned on the wild-type tertiary structure, for compatibility with the fold that makes the aptamer work. Both log-likelihoods are exponentiated, normalised to [0, 1] and multiplied — a product-of-experts combination — and top-ranked sequences go straight to synthesis. Nothing is fine-tuned on the target, so the pipeline runs unchanged on a new aptamer family.

The work comes from Yanjie Huang, Bibi Zhang and colleagues, with Shuangjia Zheng of the GENTEL Lab at Shanghai Jiao Tong University and Yi Yang at East China University of Science and Technology corresponding, plus a co-author at MIT. It was posted to bioRxiv in April 2025 under CC-BY-NC and has not been peer reviewed.

#Key Features

• Benchmark-driven component choice: Nine nucleotide foundation models were scored on six non-coding RNA deep mutational scanning datasets before any was used for design, and the two best were carried forward.
• Complementary sequence and structure signals: Spearman correlation between the language model's and the inverse-folding model's scores is low, so the two capture different facets of the sequence-function map; combining them raises high-fitness prediction precision above either alone.
• Two-phase library construction: Every position is first scanned for single mutations that improve both scores; the highest-scoring sites (11 for Broccoli) are retained, and their mutations combined into libraries of three to seven changes, re-ranked by the composite score.
• Multi-round evolution without retraining: Wet-lab results from one round drop harmful mutations and widen the mutation budget for the next, escaping local optima while model weights stay fixed.
• Wet-lab validation in cells: Designs were assayed for fluorescence and affinity in vitro, then imaged by confocal microscopy and quantified by flow cytometry in HEK293T cells.

#Technical Details

The benchmark covered four RNA language models — AIDO.RNA, RiNALMo, RNA-FM and RNA-MSM — four DNA language models — Evo, Nucleotide Transformer, GENA-LM and GROVER — and the inverse-folding model RhoDesign, scored by Spearman correlation between model likelihood and measured fitness on tRNA, fluorescent aptamer and glmS ribozyme mutational scans. AIDO.RNA at 1.6B parameters and RiNALMo at 651M were most accurate; RhoDesign came close on only 37.1M parameters; no DNA language model reached significant correlation. The pipeline therefore pairs AIDO.RNA with RhoDesign, taking tertiary structures from AlphaFold 3 and secondary structure from RhoFold and DSSR.

For the Broccoli aptamer, which binds DFHBI-1T through a G-quadruplex, 20 designs carrying three to seven mutations were synthesised: 13 were fluorescent and all 13 bound more tightly than wild type, the B2 variant gaining roughly 55% fluorescence and twofold affinity, and eight were brighter than wild type in HEK293T cells. For the already heavily optimised Pepper aptamer, a first round of 20 designs with four to six mutations produced 15 brighter variants with HBC-530 and five with a near-threefold affinity gain for HBC-620. A second round widened the budget to 6-14 mutations and still gained 2.6-fold brightness, at sequence identity as low as 70% over the 49-nucleotide aptamer.

#Applications

RILLIE fits labs that engineer functional RNAs and can afford to synthesise tens, not thousands, of candidates: fluorescent aptamer reporters for live-cell imaging, and more broadly aptamers, riboswitches and ribozymes with a wild-type structure to condition on. Because the scoring is zero-shot it applies to targets with no mutational scanning data of their own — the situation that blocks supervised fitness predictors — and it slots into an existing directed-evolution workflow, each batch of assay results reshaping the next library rather than retraining a model.

#Impact

RILLIE is evidence that pretrained nucleotide models can drive RNA engineering end to end, with variants validated in cells rather than only in silico, at 20 sequences per round — a scale any molecular biology lab can run. Its benchmark also delivers a useful negative result: DNA language models, despite reported transfer to non-coding RNA, did not predict RNA fitness here. The design claims rest on two fluorescent aptamer families, so generalisation to ribozymes and riboswitches is supported by correlation on scanning data rather than design campaigns. The repository carries an MIT licence file, though its README separately asserts a non-commercial restriction and points at a licence document that is absent from the tree, so the terms are not settled. The inverse-folding checkpoint the pipeline depends on is distributed through a Google Drive link referenced in the repository rather than a model hub, making the weights less durable than the code.