Extended Data Fig. 1: Training ablations reveal determinants of RFdiffusion success.

A–C) RFdiffusion can generate high quality large unconditional monomers. Designs are routinely accurately recapitulated by AF2 (see also Fig. 2c), with high confidence (A) for proteins up to approximately 400 amino acids in length. B) Further orthogonal validation of designs by ESMFold. C) Recapitulation of the design structure is often better with ESMFold compared with AF2. For each backbone, the best of 8 ProteinMPNN sequences is plotted, with points therefore paired by backbone rather than sequence. D) Comparing RFdiffusion trained with MSE loss on Cα atoms and N-Cα-C backbone frames (Methods 2.5), rather than with FAPE loss^8,17. The MSE loss is not invariant to the global coordinate frame, unlike FAPE loss, and is required for good performance at unconditional generation (left, two-proportion z-test of in silico success rate, n = 400 designs per condition, z = 4.1, p = 4.1e-5). For motif scaffolding problems, where the ‘motif’ provides a means to align the global coordinate frame between timesteps, FAPE loss performs approximately as well as MSE loss, suggesting the L2 nature of MSE loss (as opposed to the L1 loss in FAPE) is not empirically critical for performance. E) Allowing the model to condition on its X₀ prediction at the previous timestep (see Supplementary Methods 2.4) improves designs. Designs with self-conditioning (pink) have improved recapitulation by AF2 (left) and better AF2 confidence in the prediction (right). Two-proportion z-test of in silico success rate, n = 800 designs per condition z = 11.4, p = 6.1e-30. F) RFdiffusion leverages the representations learned during RF pre-training. RFdiffusion fine-tuned from pre-trained RF (pink) comprehensively outperforms a model trained for an equivalent amount of time, from untrained weights (gray). For context, sequences generated by ProteinMPNN on these output backbones are little better than sampling ProteinMPNN sequences from random Gaussian-sampled coordinates (white). Two-proportion z-test of in silico success rate, pre-training vs without pre-training (or vs random noise; both have zero success rate), n = 800 designs per condition, z = 23.0, p = 3.1e-117. Note that the data in pink in D–F is the same data, reproduced in each plot for clarity. G) The median (by AF2 r.m.s.d. vs design) 300 amino acid unconditional sample highlighting the importance of self-conditioning and pre-training. Without pre-training (at least when trained with equivalent compute), RFdiffusion outputs bear little resemblance to proteins (gray, left). Without self-conditioning, outputs show characteristic protein secondary structures, but lack core-packing and ideality (gray, middle). With pre-training and self-conditioning, proteins are diverse and well-packed (pink, right). H) Greater coherence during unconditional denoising may partly explain the effect of self-conditioning. Successive X₀ predictions are more similar when the model can self-condition (lower r.m.s.d. between X₀ predictions, pink curve). Data are aggregated from unconditional design trajectories of 100, 200 and 300 residues. I) During the reverse (generation) process, the noise added at each step can be scaled (reduced). Reducing the noise scale improves the in silico design success rates (left, middle; two-proportion z-test of in silico success rate, n = 800 designs per condition, 0 vs 0.5: z = 1.7, p = 0.09, 0 vs 1: z = 6.5, p = 6.8e-11; 0.5 vs 1: z = 4.8, p = 1.4e-6). This comes at the expense of diversity, with the number of unique clusters at a TM-score cutoff of 0.6 reduced when noise is reduced (right). Note throughout this figure the 6EXZ_long benchmarking problem is abbreviated to 6EXZ for brevity. Boxplots represent median±IQR; tails: min/max excluding outliers (±1.5xIQR).