RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning

Zefeng Lin; Xianyong Fang; Tianfan Fu; Xiaohua Xu

q-bio.BMSep 29, 2026

RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning

Authors: Zefeng Lin, Xianyong Fang, Tianfan Fu, Xiaohua Xu

Organizations: School of Computer Science and Technology, University of Science and Technology of China, Hefei, Anhui, China · School of Computer Science and Technology, Anhui University, Hefei, Anhui, China · State Key Laboratory for Novel Software Technology at Nanjing University, School of Computer Science, Nanjing University, Nanjing, Jiangsu, China

Abstract

RNA design aims to identify sequences that fold into specified secondary structures. Existing methods formulate the task as target-specific search or conditional generation. However, natural RNA evolution proceeds through sequence variation and selection, with compensatory substitutions, whereas these methods do not explicitly model this process. To address this limitation, we propose a two-stage framework comprising RNA Inverse-Folding Flow (RNA-IFlow) and RNA-IFlow-RL. RNA-IFlow uses structure-conditioned Dirichlet Flow Matching to model coordinated variation across the sequence, while RNA-IFlow-RL maps the learned flow to a pairing-preserving finite policy and refines it with thermodynamic feedback. Our framework achieves leading performance on multiple benchmarks, reaching 85.19% Pass@1 on Rfam-27. Further analyses reveal thermodynamic gains, policy dynamics, and robustness across settings. Our work couples coordinated variation with thermodynamic selection, offering a novel paradigm for RNA design.

Figures & tables

Figure 1: Overview of the two-stage framework. (a) illustrates the variation–selection motivation. (b) shows RNA-IFlow supervised training. (c) shows RNA-IFlow-RL training and inference.

Figure 2: Post-training dynamics and sensitivity analysis. (a) shows reward and P@8 changes during post-training. (b,c) show shifts in target-probability and NED distributions. (d,e) show mean P@8

\pm

SD across three seeds for

H

and

G

, together with inference and update times, respectively.

Figure 3: Continuous-flow and finite-policy sequence evolution. (a) shows sequence uncertainty, prediction confidence, and the proportion of legal target pairs over normalized flow time. (b) compares scheduled (blue) and actual (orange) nucleotide updates across policy steps. (c) compares matched RNA-IFlow and RNA-IFlow-RL trajectories; red boxes mark differences from the initial sequence.

Figure 4: This figure compares Anemone designs (214 nt). Scores and recovery use full sequences to evaluate. In the visualization of secondary structure, red bases mark differences from the reference, while green, orange, and dashed gray bonds denote recovered, alternative, and missing pairs.

Appendix figures & tables37 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: This figure compares candidate budgets on Eterna100-v2. Panel (a) shows uMFE coverage, panels (b,c) show the best target probability and NED, and panel (d) shows distinct valid sequences. Curves use nested candidate prefixes.

Figure 6: This figure analyzes training-data composition and coverage. Panel (a) compares equal-size training mixtures, while panel (b) relates composition to quality (left) and fixed-exposure coverage to diversity (right). Unique samples denote distinct target structures.

Figure 7: This figure compares quality and cost under the recorded generation and native-search protocols. Panel (a) shows runtime, panel (b) shows uMFE success, and panels (c,d) show target probability and NED. Boxes show quartiles and diamonds denote means.

Method	Total parameters	Millions	Trained parameters
RNA-Design-LM SL	357921408	$357.92$	-
RNA-Design-LM SL+RL	357921408	$357.92$	-
GoForth	59392004	59.39	-
DRAG	37285	0.04	-
RNAinverse-pf	-	-	-
RNA-IFlow	87548613	$87.55$	86958021

Appendix

Table 3: This table reports model capacity. A dash denotes either no neural parameters for pure search or an unknown historical training mask.

Method	P@1	P@8	MFE@1	MFE@8	Prob.	F1	NED	Div.
RNA-Design-LM SL	$0.2633$	$0.4333$	$0.2700$	$0.4433$	$0.3364$	$0.7735$	$0.1694$	0.9996
RNA-Design-LM SL+RL	0.5133	0.6367	0.5233	0.6567	0.5256	0.9324	0.0652	$0.9783$
GoForth	$0.3933$	$0.5700$	$0.4133$	$0.5833$	$0.3499$	$0.9236$	$0.0804$	$0.9862$
DRAG	$0.3633$	$0.5133$	$0.3633$	$0.5200$	$0.2391$	$0.9062$	$0.1345$	$0.8029$
RNAinverse-pf	$0.4367$	$0.5433$	$0.4533$	$0.5700$	$0.4344$	$0.6544$	$0.3625$	$0.6429$
RNA-IFlow (Ours)	$0.3033$	$0.5200$	$0.3133$	$0.5367$	$0.4089$	$0.8731$	$0.1017$	0.9979

Appendix

Table 4: This table reports complete Eterna100-v2 results for eight returned candidates. Lower NED is better; all other metrics are higher-is-better.

Method	P@1	P@8	MFE@1	MFE@8	Prob.	F1	NED	Div.
RNA-Design-LM SL	$0.2633$	$0.4267$	$0.2700$	$0.4333$	$0.3326$	$0.7669$	$0.1702$	0.9996
RNA-Design-LM SL+RL	0.5033	0.6133	0.5167	0.6333	0.5165	0.9196	0.0691	$0.9812$
GoForth	$0.3933$	$0.5667$	$0.4133$	$0.5733$	$0.3435$	$0.9116$	$0.0848$	$0.9858$
DRAG	$0.3633$	$0.5133$	$0.3633$	$0.5167$	$0.2376$	$0.8960$	$0.1373$	$0.8063$
RNAinverse-pf	$0.4433$	$0.5100$	$0.4567$	$0.5267$	$0.4292$	$0.6442$	$0.3594$	$0.6400$
RNA-IFlow (Ours)	$0.3000$	$0.4867$	$0.3100$	$0.5067$	$0.3988$	$0.8596$	$0.1099$	0.9975

Appendix

Table 5: This table reports complete Eterna100 results for eight returned candidates. Lower NED is better; all other metrics are higher-is-better.

Method	P@1	P@8	MFE@1	MFE@8	Prob.	F1	NED	Div.
RNA-Design-LM SL	$0.5185$	$0.7531$	$0.5185$	$0.7531$	$0.6200$	$0.9814$	$0.0267$	1.0000
RNA-Design-LM SL+RL	0.8148	0.8642	0.8148	0.8642	0.7765	0.9970	0.0058	1.0000
GoForth	$0.6296$	$0.7284$	$0.6296$	$0.7407$	$0.4878$	$0.9930$	$0.0162$	1.0000
DRAG	$0.7160$	$0.8272$	$0.7407$	$0.8272$	$0.4218$	$0.9962$	$0.0468$	$0.5170$
RNAinverse-pf	$0.3827$	$0.4568$	$0.3827$	$0.4568$	$0.4375$	$0.4932$	$0.5076$	$0.4398$
RNA-IFlow (Ours)	$0.6543$	$0.8272$	$0.6667$	$0.8395$	$0.7158$	$0.9944$	$0.0104$	1.0000

Appendix

Table 6: This table reports complete Rfam-27 results for eight returned candidates. Lower NED is better; all other metrics are higher-is-better.

Method	P@1	P@8	MFE@1	MFE@8	Prob.	F1	NED	Div.
RNA-Design-LM SL	$0.5951$	$0.7256$	$0.6069$	$0.7378$	$0.6418$	$0.9651$	$0.0347$	0.9999
RNA-Design-LM SL+RL	0.7199	0.7531	0.7356	0.7657	0.7156	0.9808	0.0154	$0.9774$
GoForth	$0.6252$	$0.7216$	$0.6440$	$0.7347$	$0.5858$	$0.9756$	$0.0274$	$0.9848$
RNA-IFlow (Ours)	$0.6165$	0.7395	$0.6339$	$0.7535$	$0.6890$	$0.9755$	$0.0203$	0.9992
RNA-IFlow-RL (Ours)	0.7120	$0.7312$	0.7299	0.7544	0.7109	0.9783	0.0157	$0.9186$

Appendix

Table 7: This table reports complete RNAsolo-764 results for eight returned candidates. Lower NED is better; all other metrics are higher-is-better.

Method	Full P@8	755 P@8	Full NED	755 NED
RNA-Design-LM SL	$0.7256$	$0.7223$	$0.0347$	$0.0351$
RNA-Design-LM SL+RL	0.7531	0.7501	0.0154	0.0156
GoForth	$0.7216$	$0.7183$	$0.0274$	$0.0274$
RNA-IFlow	0.7395	0.7364	$0.0203$	$0.0205$
RNA-IFlow-RL	$0.7312$	$0.7280$	0.0157	0.0159

Appendix

Table 8: RNAsolo overlap sensitivity. The 755-target subset excludes the nine exact supervised target-structure overlaps without using model outcomes.

Generator	P@1	P@8	Prob.	NED
Native RNA-IFlow	0.3033	0.5200	0.4089	0.1017
RNA-IFlow-no-RL	$0.1233$	$0.2433$	$0.1822$	$0.2399$
RNA-IFlow-RL	0.5400	0.6500	0.5398	0.0534

Appendix

Table 10: This table summarizes the policy interface and post-training results on Eterna100-v2.

Axis	Value	P@1	P@8	Prob.	NED
H	1	$0.4533$	$0.6033$	$0.4821$	$0.0736$
H	2	0.5000	0.6567	$0.5312$	0.0590
H	4	$0.4900$	$0.6367$	$0.5258$	$0.0598$
H	8	0.5167	0.6500	0.5424	0.0558
H	16	0.5000	0.6500	0.5317	$0.0599$
G	2	0.5167	$0.6500$	$0.5312$	$0.0587$

Appendix

Table 11: This table reports complete single-seed policy sensitivity; group size one is the zero-advantage boundary.

Configuration	P@1	P@8	Prob.	NED
$H=2,G=8$	$0.5033\pm 0.0058$	$0.6511\pm 0.0096$	$0.5309\pm 0.0063$	$0.0593\pm 0.0014$
$H=8,G=8$	$0.5122\pm 0.0051$	$0.6578\pm 0.0107$	$0.5391\pm 0.0043$	$0.0554\pm 0.0016$
$H=8,G=4$	$0.5122\pm 0.0102$	$0.6544\pm 0.0051$	$0.5351\pm 0.0023$	$0.0562\pm 0.0008$
$H=8,G=16$	$0.5178\pm 0.0158$	$0.6578\pm 0.0019$	$0.5365\pm 0.0016$	$0.0557\pm 0.0013$

Appendix

Table 12: Values are reported as means and sample standard deviations across three repeated sensitivity runs.

G	Median (s)	Minimum (s)	Maximum (s)
2	25.7918	10.5854	34.0818
4	27.4301	10.4608	35.9115
8	$39.0918$	$11.9554$	$72.1897$
16	$56.9375$	$12.3717$	$302.8546$

Appendix

Table 13: This table reports observed group-size update costs on shared GPUs; minimum and maximum values summarize profiling variability.

Steps	P@1	P@8	MFE@8	Prob.
4	$0.2800$	$0.5033$	$0.5033$	$0.3991$
8	$0.2900$	0.5133	0.5367	$0.4041$
16	$0.2967$	0.5133	0.5367	$0.4056$
32	0.3000	0.5133	0.5333	0.4095
50	0.3033	0.5200	0.5367	0.4089

Appendix

Table 14: This table reports native-flow integration-step sensitivity on 100 Eterna100-v2 targets.

Updates	P@1	P@8	MFE@8	Prob.
1744	$0.5233$	0.6567	0.6867	0.5446
2093	0.5333	0.6533	$0.6767$	$0.5364$
2442	0.5400	$0.6500$	$0.6733$	$0.5398$
2790	$0.5300$	$0.6500$	0.6800	0.5437

Appendix

Table 15: This table reports checkpoint performance during post-training continuation.

Setting	P@1	P@8	Prob.	NED	Div.
Original mixture	0.5133	0.6533	0.5363	0.0555	0.9513
Core-only	$0.4867$	0.6333	$0.5203$	$0.0616$	$0.9279$
Hard-enriched	0.4900	0.6333	0.5248	0.0610	0.9379

Appendix

Table 16: This table compares matched-size training-data compositions.

Setting	P@1	P@8	Prob.	NED	Div.
1395 unique tasks	$0.5000$	0.6500	0.5361	0.0590	$0.9071$
2790 unique tasks	0.5133	$0.6267$	$0.5269$	$0.0596$	0.9175
4143 unique tasks	0.5033	0.6467	0.5367	0.0565	0.9292

Appendix

Table 17: This table reports coverage under a fixed number of target visits.

Setting	P@1	P@8	Prob.	NED	Div.
Lower CE weight	$0.5133$	$0.6433$	$0.5332$	$0.0576$	0.9487
Higher CE weight	0.5167	0.6467	0.5361	$0.0582$	0.9517
Lower KL weight	0.5167	0.6533	0.5408	0.0546	$0.9413$
Higher KL weight	0.5233	$0.6433$	$0.5313$	0.0561	$0.9400$

Appendix

Table 18: This table reports one-axis regularization sensitivity with all other settings fixed.

Setting	P@1	P@8	Prob.	NED
One fresh round	$0.5133$	$0.6633$	$0.5425$	$0.0551$
Two fresh rounds	0.5167	0.6500	0.5424	0.0558
Fixed behavior snapshot, two rounds	0.5233	0.6533	0.5396	0.0557
Three fresh rounds	0.5100	0.6467	0.5396	0.0558
Fixed behavior snapshot, three rounds	0.5233	0.6600	0.5435	0.0554

Appendix

Table 19: This table reports rollout-freshness and reuse controls.

Method	Return time (s)	With scoring (s)	Timed groups
RNA-Design-LM SL	$4.4085$	$4.7356$	600
RNA-Design-LM SL+RL	$4.4085$	$4.6438$	600
GoForth	$1.3518$	$1.6502$	600
DRAG	$14.8637$	$15.0512$	300
RNAinverse-pf	$133.5565$	$133.7930$	300
RNA-IFlow	0.8845	1.1684	600

Appendix

Table 20: This table reports time to return eight candidates and time after common scoring. Neural entries use matched resident-model medians, while search rows retain their native timing scope.

K	Pass@K	MFE@K	Prob.	NED	Pair-F1	Diversity
1	$0.2633$	$0.2700$	$0.2136$	$0.2762$	$0.5899$	1.0000
2	$0.3433$	$0.3500$	$0.2618$	$0.2276$	$0.6662$	1.0000
4	$0.4033$	$0.4100$	$0.3024$	$0.1946$	$0.7313$	1.0000
8	$0.4333$	$0.4433$	$0.3364$	$0.1694$	$0.7735$	0.9996
16	$0.4800$	$0.4967$	$0.3719$	$0.1497$	$0.8187$	$0.9992$
32	$0.5167$	$0.5467$	$0.4064$	$0.1319$	$0.8539$	$0.9985$

Appendix

Table 21: This table reports RNA-Design-LM SL nested-prefix results from the same ordered candidate set, with each metric aggregated separately.

K	Pass@K	MFE@K	Prob.	NED	Pair-F1	Diversity
1	$0.5133$	$0.5233$	$0.4238$	$0.1099$	$0.8319$	1.0000
2	$0.5833$	$0.5900$	$0.4756$	$0.0855$	$0.8847$	0.9950
4	$0.6200$	$0.6333$	$0.5036$	$0.0743$	$0.9145$	$0.9883$
8	$0.6367$	$0.6567$	$0.5256$	$0.0652$	$0.9324$	$0.9783$
16	$0.6500$	$0.6800$	$0.5375$	$0.0582$	$0.9493$	$0.9683$
32	$0.6600$	$0.6967$	$0.5501$	$0.0530$	$0.9543$	$0.9581$

Appendix

Table 22: This table reports RNA-Design-LM SL+RL nested-prefix results from the same ordered candidate set, with each metric aggregated separately.

K	Pass@K	MFE@K	Prob.	NED	Pair-F1	Diversity
1	$0.3933$	$0.4133$	$0.2302$	$0.1357$	$0.8096$	1.0000
2	$0.4633$	$0.4800$	$0.2714$	$0.1101$	$0.8667$	0.9933
4	$0.5433$	$0.5533$	$0.3276$	$0.0895$	$0.9075$	$0.9900$
8	$0.5700$	$0.5833$	$0.3499$	$0.0804$	$0.9236$	$0.9862$
16	$0.5933$	$0.6200$	$0.3731$	$0.0744$	$0.9360$	$0.9831$
32	$0.6100$	$0.6300$	$0.3894$	$0.0693$	$0.9435$	$0.9793$

Appendix

Table 23: This table reports GoForth nested-prefix results from the same ordered candidate set, with each metric aggregated separately.

K	Pass@K	MFE@K	Prob.	NED	Pair-F1	Diversity
1	$0.3033$	$0.3133$	$0.2534$	$0.1880$	$0.6872$	1.0000
2	$0.3733$	$0.3900$	$0.3076$	$0.1477$	$0.7766$	1.0000
4	$0.4600$	$0.4800$	$0.3680$	$0.1224$	$0.8334$	$0.9975$
8	$0.5200$	$0.5367$	$0.4089$	$0.1017$	$0.8731$	0.9979
16	$0.5700$	$0.5967$	$0.4410$	$0.0894$	$0.9020$	$0.9960$
32	$0.5933$	$0.6167$	$0.4650$	$0.0818$	$0.9175$	$0.9931$

Appendix

Table 24: This table reports RNA-IFlow nested-prefix results from the same ordered candidate set, with each metric aggregated separately.

K	Pass@K	MFE@K	Prob.	NED	Pair-F1	Diversity
1	$0.5400$	$0.5433$	$0.4483$	$0.0906$	$0.8527$	1.0000
2	$0.5867$	$0.5900$	$0.4854$	$0.0748$	$0.8906$	0.9850
4	$0.6233$	$0.6300$	$0.5127$	$0.0633$	$0.9190$	$0.9683$
8	$0.6500$	$0.6733$	$0.5398$	$0.0534$	$0.9396$	$0.9363$
16	$0.6633$	$0.6933$	$0.5517$	$0.0499$	$0.9510$	$0.9071$
32	$0.6833$	$0.7133$	$0.5626$	$0.0462$	$0.9597$	$0.8772$

Appendix

Table 25: This table reports RNA-IFlow-RL nested-prefix results from the same ordered candidate set, with each metric aggregated separately.

Method	uMFE	MFE	Prob.	NED	Pair-F1	Run time (s)
FastDesign	0.4900	0.5267	0.3827	$0.3589$	0.6623	33.5269
SAMFEO	0.6667	0.6933	0.5283	0.0579	0.9474	$43.3045$
NEMO	$0.2933$	$0.3167$	$0.1720$	0.2928	$0.6530$	15.9501

Appendix

Table 26: This table reports native-search performance on Eterna100-v2 with at most 60 seconds per target and seed. Time is the observed mean search time, and all 300 groups contribute to the metrics.

Method	Groups	Timeouts	Empty returns	Mean returned
FastDesign	300	93	93	$3.4267$
SAMFEO	300	172	0	$9.9200$
NEMO	300	68	48	$0.8400$

Appendix

Table 27: This table reports search completion and returned-candidate counts under the same native 60-second protocol. Returned counts describe the final candidate sets.

Method	MFE	uMFE	Budget / source
RNA-IFlow-RL	73	68	At most 10,000 candidates; global stop
RNA-Design-LM SL+RL	74	72	Published 10,000-sample evaluation ( Gautam et al., 2026 )
FastDesign	80	78	Published motif/cube search ( Zhou et al., 2026 )
SAMFEO	77	74	Five-run union, probability objective ( Zhou et al., 2023 )
NEMO	79	77	Five-run union reported by Zhou et al. (2023)

Appendix

Table 28: This table reports Eterna100 large-budget coverage as solved targets out of 100. NEMO counts follow the SAMFEO authors’ reproduction.

Benchmark	Method	MFE	uMFE	Protocol
Eterna100-v2	RNA-IFlow-RL	81	76	Maximum 10,000 candidates
Eterna100-v2	RNA-Design-LM SL+RL	81	79	Published 10,000 samples ( Gautam et al., 2026 )
Rfam-27	RNA-IFlow-RL	24	24	Maximum 10,000 candidates
Rfam-27	RNA-Design-LM SL+RL	24	24	Published 10,000 samples ( Gautam et al., 2026 )
RNAsolo-764	RNA-IFlow-RL	591	585	Maximum 10,000 candidates

Appendix

Table 29: This table reports additional large-budget generator coverage under each source’s original stopping and aggregation protocol.

Method	MFE solved	uMFE solved	Source
FastDesign	611	606	Zhou et al. (2026) , Table 2
SAMFEO	608	602	Zhou et al. (2026) , Table 2
NEMO	609	605	Zhou et al. (2026) , Table 2

Appendix

Table 30: This table reports published RNAsolo-764 search coverage from the FastDesign comparison using the source paper’s native search protocol.

Sequence	Prob.	NED	MFE	uMFE	Group MFE	Group uMFE
Reference sequence	$0.1803$	$0.0367$	yes	yes	–	–
RNA-IFlow-RL	0.5178	0.0128	yes	yes	2/8	2/8
RNA-IFlow	0.1257	$0.0773$	no	no	0/8	0/8
RNA-Design-LM SL+RL	$0.0769$	0.0297	no	no	0/8	0/8
GoForth	$0.0019$	$0.1477$	no	no	0/8	0/8
DRAG	$0.0007$	$0.0447$	no	no	0/8	0/8

Appendix

Table 31: This table reports displayed-candidate thermodynamics and group-level hit counts for Anemone. Reference metrics are evaluated independently and do not form a generated

K=8

group.

Figure 8: This figure shows additional first-candidate designs, with a recovered target above and a residual uMFE failure below. Green, orange, and dashed gray bonds denote recovered, alternative, and missing pairs, while nucleotide colors denote A/C/G/U in orange/blue/green/purple.

Target	nt	Model	Prob.	NED	uMFE
Ete_19	104	RNA-IFlow	0.0023	0.0647	no
Ete_19	104	RNA-IFlow-RL	0.9467	0.0022	yes
Ete_50	108	RNA-IFlow	$5.5166\times 10^{-7}$	0.3343	no
Ete_50	108	RNA-IFlow-RL	$0.0006$	0.2935	no

Appendix

Table 32: This table reports additional fixed-first-candidate thermodynamics; small nonzero probabilities are retained in scientific notation.

Figure 9: This figure shows frozen-supervised-flow profiles over 12,613 training-pool targets, each with eight valid candidates. Panels (a,b) report target fractions by empirical uMFE success and mean normalized structural distance, respectively.

Figure 10: This figure compares candidate-level target-probability histograms before and after RL. Gray and blue outlines denote the finite policy without RL and RNA-IFlow-RL, while the vertical axis reports the candidate fraction in each probability bin.

Figure 11: This figure shows training-set length distributions. Bars report sample fractions in each length interval, with green and blue denoting supervised and RL training data, respectively.

Sequence	Matching positions	Recovery (%)
Author/player reference	214/214	100.00
RNA-Design-LM SL+RL	87/214	40.65
GoForth	88/214	41.12
DRAG	101/214	47.20
RNA-IFlow	97/214	45.33
RNA-IFlow-RL	114/214	53.27

Appendix

Table 33: This table reports full-length recovery for the displayed Anemone candidates using the same 214-nucleotide reference throughout.

Explore similar work

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching
Oct 8, 2026Runze Ma, Will Hua, Shuangjia ZhengInverse Design Flow-Based Generative Modeling
Shanghai Jiao Tong University · Independent Researcher
MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design
Oct 9, 2026Jixin Tang, Ji Guo, Jun ZhaoInverse Design Reward Design for RL
Fudan University
RIDER: 3D RNA Inverse Design with Reinforcement Learning-Guided Diffusion
Feb 18, 2026Tianmeng Hu, Yongzheng Cui, Biao Luo +1Inverse Design 3D Diffusion Models
University of Exeter · Central South University

Oct 8, 2026J/K move · Enter open · S save

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching

Runze Ma, Will Hua, Shuangjia Zheng

Shanghai Jiao Tong University · Independent Researcher

Inverse Design Flow-Based Generative Modeling

Open paper AskPDF arXiv

RNA function arises from the coupling of nucleotide sequence and three-dimensional structure, motivating their joint design. Coordinating global folding with nucleotide-level detail remains challenging under limited structural supervision. We introduce La-Ribo, a generative framework for RNA sequence-structure co-design via geometry-latent flow matching. La-Ribo retains a sparse phosphate-sugar--base scaffold and encodes nucleotide identity and local conformation in residue-wise latents. A shared flow network generates both jointly, and an RNA-specific decoder then reconstructs all heavy atoms. To expand supervision, we construct a quality-controlled corpus of 168,561 RNA structures, integrating experimental data with predictions from three folding models, including 10,631 MSA-supported structures generated in this work. La-Ribo improves designability and codesignability over the evaluated baselines across sampling budgets and two refolding models, and the same prior supports scaffold-conditioned inverse folding without additional training.

Abstract

Explore similar work

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching

MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design

RIDER: 3D RNA Inverse Design with Reinforcement Learning-Guided Diffusion