TABLE 4.

Comparison between Coordinates to Design (C → D) and reversible learning (C ↔ D) on CASP 7–12. The letters in the brackets represent the data considered by the loss function. P denotes PSSM, O denotes one-hot encoding of the sequence and P/O implies both data are utilized. We report the KL-divergence scores of the PSSM, the sequence or both. The accuracy of the sequence is depicted in brackets where applicable.

Dataset	C → D (P)	C → D (O)	C → D (P/O)	C ↔ D(P)	C ↔ D(P/O)
CASP 7	1.08	1.79 (16.7%)	1.02/1.69 (23.2%)	0.98	0.96/1.69 (23.1%)
CASP 8	0.87	1.77 (14.2%)	0.89/1.67 (21.1%)	0.87	0.88/1.64 (21.5%)
CASP 9	0.90	1.74 (13.9%)	0.90/1.61 (21.3%)	0.92	0.91/1.62 (21.0%)
CASP 10	0.83	1.76 (16.1%)	0.84/1.65 (21.7%)	0.88	0.83/1.69 (21.2%)
CASP 11	0.98	1.82 (14.3%)	1.00/1.75 (20.2%)	0.91	0.93/1.73 (20.7%)
CASP 12	0.97	1.85 (14.0%)	0.99/1.75 (20.3%)	0.94	0.91/1.74 (20.3%)