Table 1:

Ablation study and aggregated benchmark results for gRNAde. We report metrics averaged over 100 test sets samples and standard deviations across 3 consistent random seeds. The percentages reported in brackets for the 3D self-consistency scores are the percentage of designed samples within the ‘designability’ threshold values (scRMSD≤2Å, scTM≥0.45, scGDT≥0.5).

Split	Max. #states	Model	GNN	Max. train length	Perplexity (↓)	Native seq. recovery (↑)	Self-consistency metrics
							2D – EternaFold scMCC (↑)	scRMSD (↓)	3D – RhoFold scTM-score (↑)	scGDT_TS (↑)
Single-state split	1	AR	Equiv	500	1.77±0.07	0.438±0.01	0.624±0.07	13.01±1.18 (0.5%)	0.21±0.0 (14.3%)	0.22±0.0 (12.7%)
	1	AR	Equiv	1000	1.73±0.08	0.453±0.01	0.648±0.01	13.10±0.58 (1.0%)	0.20±0.0 (10.8%)	0.21±0.0 (10.6%)
	1	AR	Equiv	2500	1.41±0.01	0.493±0.01	0.633±0.03	11.76±0.91 (1.4%)	0.27±0.0 (28.8%)	0.27±0.0 (28.0%)
	1	AR	Equiv	5000	1.29±0.02	0.530±0.01	0.585±0.03	11.70±0.56 (1.3%)	0.26±0.0 (24.8%)	0.25±0.0 (20.1%)
	1	AR	Inv	5000	1.32±0.04	0.549±0.00	0.612±0.02	11.50±0.64 (1.9%)	0.28±0.0 (32.1%)	0.28±0.0 (26.2%)
	1	NAR	Inv	5000	1.54±0.04	0.571±0.00	0.430±0.02	14.26±0.51 (1.3%)	0.19±0.0 (15.9%)	0.18±0.0 (12.7%)
	1	NAR	Equiv	5000	1.46±0.06	0.584±0.00	0.473±0.02	13.04±0.88 (1.3%)	0.23±0.0 (24.0%)	0.22±0.0 (17.9%)
	3	AR	Equiv	5000	1.23±0.05	0.539±0.01	0.620±0.01	11.47±1.05 (2.5%)	0.28±0.0 (31.4%)	0.28±0.0 (27.2%)
	5	AR	Equiv	5000	1.25±0.01	0.539±0.02	0.596±0.03	11.90±1.00 (2.9%)	0.27±0.0 (31.6%)	0.26±0.0 (26.4%)
	Groundtruth sequence prediction baseline:				-	1.000±0.00	0.686±0.00	5.23±0.07 (27.9%)	0.56±0.0 (68.7%)	0.55±0.0 (68.7%)
	Random sequence prediction baseline:				-	0.251±0.00	0.012±0.00	24.40±0.34 (0.0%)	0.04±0.0 (0.0%)	0.02±0.0 (0.0%)
	ViennaRNA 2D-only baseline:				-	0.259±0.00	0.611±0.00	20.34±0.10 (0.0%)	0.07±0.0 (0.6%)	0.07±0.0 (1.1%)
Multi-state split	1	AR	Equiv	500	1.87±0.06	0.445±0.01	0.603±0.03	13.08±0.20 (3.5%)	0.10±0.0 (1.2%)	0.25±0.0 (20.7%)
	1	AR	Equiv	1000	1.84±0.01	0.447±0.01	0.580±0.01	13.02±0.56 (2.3%)	0.09±0.0 (0.9%)	0.25±0.0 (20.4%)
	1	AR	Equiv	2500	1.73±0.04	0.480±0.02	0.567±0.01	12.83±0.05 (3.4%)	0.10±0.0 (1.9%)	0.26±0.0 (21.2%)
	1	AR	Equiv	5000	1.68±0.03	0.455±0.01	0.569±0.02	12.88±0.20 (4.1%)	0.11±0.0 (1.6%)	0.26±0.0 (22.6%)
	1	AR	Inv	5000	1.72±0.01	0.463±0.01	0.559±0.03	13.09±0.27 (4.1%)	0.10±0.0 (2.2%)	0.27±0.0 (23.0%)
	1	NAR	Inv	5000	2.01±0.04	0.457±0.01	0.461±0.01	14.06±0.23 (3.2%)	0.08±0.0 (1.7%)	0.23±0.0 (16.5%)
	1	NAR	Equiv	5000	1.89±0.06	0.432±0.01	0.423±0.01	13.63±0.27 (3.6%)	0.09±0.0 (1.2%)	0.24±0.0 (18.3%)
	3	AR	Equiv	5000	1.60±0.03	0.467±0.03	0.561±0.03	13.31±0.38 (3.4%)	0.10±0.0 (2.6%)	0.24±0.0 (19.0%)
	5	AR	Equiv	5000	1.55±0.04	0.473±0.01	0.549±0.03	13.48±0.79 (3.3%)	0.10±0.0 (3.0%)	0.24±0.0 (20.2%)
	Groundtruth sequence prediction baseline:				-	1.000±0.00	0.570±0.01	9.78±0.13 (10.3%)	0.16±0.0 (11.7%)	0.36±0.0 (36.7%)
	Random sequence prediction baseline:				-	0.249±0.00	0.128±0.00	21.15±0.21 (0.9%)	0.02±0.0 (0.0%)	0.09±0.0 (3.3%)
	ViennaRNA 2D-only baseline:				-	0.258±0.00	0.601±0.00	15.47±0.20 (2.4%)	0.05±0.0 (0.2%)	0.19±0.0 (15.2%)
All data	1	AR	Equiv	5000	1.23±0.01	0.733±0.00	0.627±0.02	8.10±0.28 (20.7%)	0.42±0.0 (46.1%)	0.41±0.0 (43.0%)
	2	AR	Equiv	5000	1.21±0.01	0.783±0.01	0.629±0.03	8.40±0.09 (19.1%)	0.42±0.0 (47.8%)	0.41±0.0 (41.7%)
	3	AR	Equiv	5000	1.19±0.01	0.787±0.01	0.606±0.02	7.88±0.68 (20.5%)	0.43±0.0 (47.4%)	0.42±0.0 (44.0%)
	5	AR	Equiv	5000	1.15±0.01	0.811±0.01	0.617±0.02	7.51±0.30 (20.7%)	0.45±0.0 (50.2%)	0.44±0.0 (46.7%)