SiBaSe

SiBaSe (Sidechains, Backbone, Sequence) is a generative model for de novo protein design that produces all three at once. The dominant design pipeline treats them separately: a structure generator such as RFdiffusion emits a bare backbone, an inverse-folding model such as ProteinMPNN writes a sequence for it, and a structure predictor such as ESMFold or AlphaFold 2 checks whether that sequence folds back to the intended shape. Every step in that chain discards sidechains, even though sidechain packing, burial, and non-covalent contacts are what actually drive folding and function.

SiBaSe was built by Cooper Svajda at Texas A&M University and Joshua Yuan at Washington University in St. Louis, and posted to bioRxiv in August 2025. It is deliberately constructed as a controlled experiment rather than a performance play: the architecture mirrors the FrameDiff / FrameFlow / MultiFlow family of backbone flow models closely enough that sidechain inclusion is close to the only variable that changes. The key representational choice is to encode each sidechain as a local rigid frame, the same object used for backbone residues, so that sidechain geometry can be manipulated before the sequence is known and fed through the same Invariant Point Attention machinery.

The headline result is a negative one, and the authors report it directly. Despite having sidechain information available throughout generation, SiBaSe behaves like a backbone-first model: the backbone assumes its fold early while sidechains follow behind it, and conditioning on sidechains alone is markedly less effective than conditioning on backbone.

#Key Features

• Joint co-design: Generates backbone geometry, sidechain placement, and amino acid sequence within a single flow trajectory, rather than chaining separate structure and sequence models.
• Sequence-agnostic sidechain frames: Represents each sidechain as a local reference frame built by Gram-Schmidt orthogonalization from a curated per-residue atom library, so geometry can be modeled before residue identity is decided.
• Controlled architectural comparison: Matches the FrameFamily architecture apart from the sidechain additions, isolating the effect of the extra chemical detail.
• Motif conditioning modes: Supports scaffolding from a motif supplied as backbone-plus-sidechain-plus-sequence, backbone-plus-sequence, or sidechain-plus-sequence, which is how the relative usefulness of each information type is measured.
• Stochastic sampling: Resamples noise at each step rather than following a purely deterministic ODE path, which lets trajectories escape early backbone breaks and clashes.

#Technical Details

SiBaSe is a continuous normalizing flow trained by flow matching, with a Gaussian prior for translations, a uniform prior on SO(3) for rotations, and the continuous-time Markov chain formulation from MultiFlow for the categorical sequence. The network stacks Invariant Point Attention, MLP, transformer-encoder, and geometric update layers with a node dimension of 512, four blocks, eight IPA heads, and four transformer layers, totaling 33.8 million trainable parameters. It omits the pairwise edge representation used by the FrameFamily because sidechain frames roughly double the effective token count. The loss combines the SE(3) frame loss, a sequence cross-entropy term, and a pairwise-distance topology term with separate cutoffs for backbone-backbone (12 A), backbone-sidechain (8 A), and sidechain-sidechain (8 A) pairs. Training used 3,902 monomeric soluble PDB proteins of 75-225 residues for 250 epochs over 28 days on a single T4 GPU.

Across 450 unconditional designs, full co-design succeeded in 1.6% of cases at a 3 A self-consistency RMSD threshold, below MultiFlow, the only comparable co-generative baseline. Re-sequencing the same backbones with ProteinMPNN raised designability to 30% at 3 A and 24% at the stricter 2.5 A cutoff used in FrameFamily papers, with diversity comparable to FrameDiff and above FrameFlow, and slightly improved novelty. Sidechain placement analysis using normalized Earth Mover's Distance showed accurate positioning for small, common, rigid residues and systematic divergence for arginine, methionine, and lysine, with designed sidechains consistently packed closer to the backbone than real ones.

#Applications

SiBaSe is usable as an unconditional generator of novel protein backbones and as a motif scaffolding tool, and its backbone outputs are readily redesigned with ProteinMPNN in the standard pipeline. Its more immediate value is diagnostic: for groups building all-atom generative models, the conditioning experiments and sidechain-placement statistics identify where flow-based co-design loses information, and the sequence confidence trajectories make the model's internal decision dynamics directly inspectable.

#Impact

The paper's contribution is an architectural finding rather than a benchmark win. SiBaSe demonstrates that adding sidechains to a flow-based generator does not make the model use them: because each change in predicted residue identity implies a different sidechain conformation, the model defers sidechain commitment until the backbone stabilizes, a "penalty of uncertainty" the authors argue will generalize to any approach that treats sidechains deterministically during sequence generation. The proposed remedy is representations that carry multiple sidechain configurations in parallel. As a two-author preprint with no released code or weights and no experimental validation, its results are in-silico and await independent replication.