Eterna is Solved

TL;DR

The paper tackles inverse RNA folding with the goal of reliably designing sequences that fold into target structures, targeting the Eterna100 benchmark. It introduces Montparnasse, a framework that combines MOGRLS (a simplified GREED-RNA style local search), Progressive Narrowing (PN), and the multi objective, limited repetition renderer MOGNRPALR based on generalized NRPA. Experimental results show MOGNRPALR vastly outperforms prior methods, solving all Eterna100 v1 problems within one day using 200 parallel runs, with particular gains on hard instances like Problems 90, 99, and 100. These findings demonstrate the viability of multi objective, prior-informed search strategies for efficient, high-coverage RNA design, potentially accelerating designs in synthetic biology and nanotechnology.

Abstract

RNA design consists of discovering a nucleotide sequence that folds into a target secondary structure. It is useful for synthetic biology, medicine, and nanotechnology. We propose Montparnasse, a Multi Objective Generalized Nested Rollout Policy Adaptation with Limited Repetition (MOGNRPALR) RNA design algorithm. It solves the Eterna benchmark.

Figures (6)

• Figure 1: Problems 90, 99 and 100 from Eterna100 v1.
• Figure 2: Comparison of the distributions of BPD on problem 99 for GREED-RNA and MOGRLS for increasing numbers of evaluations. After 270 000 evaluations which corresponds to one day of computation, MOGRLS has solved the problem 19 times out of 200 runs while GREED-RNA has solved the problem 2 times out of 200 runs (see subfigure (c)). Subfigure (d) gives the evolution of the average BPD with the number of evaluations for problem 99. The averages are calculated using 200 runs. The rightmost value corresponds to one day of computation for one run.
• Figure 3: Comparison of the distributions of BPD on problem 99 for PN and MOGRLS for increasing numbers of evaluations. After 270 000 evaluations which corresponds to one day of computation, PN has solved the problem 28 times out of 200 runs while MOGRLS has solved the problem 19 times out of 200 runs (see subfigure (c)). Subfigure (d) gives the evolution of the average BPD with the number of evaluations for problem 99. The averages are calculated using 200 runs. The rightmost value corresponds to one day of computation for one run.
• Figure 4: Comparison of the distributions of BPD on problem 99 for PN and MOGNRPALR for increasing numbers of evaluations. After 270 000 evaluations which corresponds to one day of computation, MOGNRPALR has solved the problem 120 times out of 200 runs while PN has solved the problem 28 times out of 200 runs (see subfigure (c)). Subfigure (d) gives the evolution of the average BPD with the number of evaluations for problem 99 for all algorithms. The averages are calculated using 200 runs. The rightmost value corresponds to one day of computation for one run.
• Figure 5: Comparison of the BPD on problem 90 for GREED-RNA and MOGNRPALR for increasing numbers of evaluations. 220 000 evaluations by one process takes one day. GREED-RNA is stuck and does not solve the problem while MOGNRPALR progresses and solves the problem 6 times out of 200 runs.