Table 1:

Ablation on the number of recycling rounds and sampling steps for Boltz-1 on the test set. We run the ablation study generating 5 samples and evaluating both the best (oracle) and highest confidence prediction (top-1) out of the 5 for every metric. All models used pre-computed MSAs with up to 4,096 sequences. It is worth noting that the metrics are noisy, so minor inconsistencies (e.g., lack of improvement with increased recycling rounds or diffusion steps) should not be overinterpreted. Moreover, there is a slight difference with the results in Figure 4 due to differences in MSA parameters as well as the set of structures passing all ablations.

		Mean LDDT		DockQ > 0.23		Mean LDDT-PLI		L-RMSD < 2Å
# rec.	# steps	oracle	top-1	oracle	top-1	oracle	top-1	oracle	top-1
3	200	0.729	0.716	0.654	0.625	0.621	0.580	0.581	0.545
0	200	0.698	0.681	0.579	0.544	0.573	0.530	0.582	0.541
1	200	0.718	0.702	0.656	0.635	0.623	0.573	0.588	0.535
2	200	0.726	0.710	0.651	0.632	0.616	0.581	0.587	0.546
3	200	0.729	0.716	0.654	0.625	0.621	0.580	0.581	0.545
6	200	0.732	0.714	0.644	0.635	0.630	0.593	0.595	0.555
8	200	0.733	0.717	0.644	0.633	0.630	0.584	0.588	0.545
10	200	0.735	0.720	0.644	0.631	0.619	0.577	0.575	0.541
3	20	0.720	0.693	0.615	0.592	0.577	0.547	0.550	0.532
3	50	0.727	0.710	0.645	0.627	0.621	0.579	0.586	0.540
3	200	0.729	0.716	0.654	0.625	0.621	0.580	0.581	0.545