{"paper_id":"0a58130a-0fef-4b1c-a89c-9807cdbe8a2b","body_text":"1 \nInvestigating the impacts of sidechains on de-novo protein design. \nCooper Svajda\n1\n,  Joshua Yuan\n2 \nCorresponding Author Email: cooper.svajda@ag.tamu.edu \n1 – Department of Plant Pathology and Microbiology, Texas A&M University \n2 – Department of Energy, Environmental and Chemical Engineering, Washington University in \nSt. Louis \nDe novo protein design aims to create novel protein structures and sequences, often to \nenable specific functions. Most current generative models operate on simplified backbone-only \nrepresentations. However, in vivo and in vitro protein folding and function are largely mediated by \namino acid sidechains. Given this biochemical relevance, we ask: what happens when sidechain \nand sequence information are introduced into a protein generative model? To address this, we \ndeveloped SiBaSe, a novel generative model that simultaneously co-designs sidechains, backbones, \nand sequence. SiBaSe achieves design performance approaching that of state-of-the-art backbone-\nonly models. However, analysis of design patterns reveals that, despite access to sidechain data, the \nmodel behaves similarly to a backbone-based model. This appears to arise from uncertainties in \nthe simultaneous modeling of sequence and sidechains that are inherent to flow-based \narchitectures and offers new insight into architectural limitations and opportunities for improving \ngenerative protein design. \n1 Introduction  \n Protein design is an emerging field with the potential to transform nearly every area of \nbiology. By designing proteins de novo for specific functions, it becomes possible to bypass the \nconstraints of natural evolution and create novel, non-natural phenotypes. Recent advances in this \nfield have been driven by machine learning-based tools that automate and accelerate the design \nprocess. \n The standard pipeline (Yeh et al. 2023, Lauko et al. 2024) begins with a structure \ngeneration tool , which produces a protein backbone lacking sequence or sidechain information \n(e.g., Watson et al. 2023, Yim et al. 2023a, Lin et al. 2023). These backbone structures can be \nguided toward specific functions through the inclusion of structural or functional motifs. Next, a \nsequence design model , most commonly ProteinMPNN (Dauparas et al. 2022), is used to infer \nan amino acid sequence intended to fold into the target backbone. Finally, the designed sequence \nis evaluated using a structure prediction model , such as AlphaFold2 or ESMFold, to assess \nwhether the predicted fold matches the original design (Jumper et al. 2021, Lin et al. 2023). The \nagreement between the designed and predicted structures, often referred to as designability , is \nused to determine the in-silico success of the design. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 2 \n This design pipeline has enabled the creation of novel and functional proteins that would \nhave been unattainable only a few years ago, including de novo luciferases, serine hydrolases, and \nhigh-affinity binders with micro- to nanomolar affinities (Yeh et al. 2023, Lauko et al. 2023, \nKrishna et al. 2024). Remarkably, these successes have been achieved using models that, at a \nfundamental level, do not directly manipulate one of the most critical determinants of protein \nstructure: the amino acid sidechains. According to Anfinsen’s Dogma, a protein’s sequence alone \ndetermines its folded structure, and the distinguishing chemical features of a sequence arise from \nthe unique compositions and order of its sidechains (Anfinsen, 1973). Decades of biochemical \nresearch supports this, showing that folding is driven by sidechain interactions; whether through \nmodulating the backbone’s electronic structure immediately after translation (Spassov et al. 2007) \nor later via sidechain packing, burial, and non-covalent interactions (Dill 1990, Dill et al. 2008, \nFarber & Mittermaier 2008). This biochemical importance stands in contrast to the simplifications \nof backbone-only models, prompting the central question of this study: what happens when \nsidechain information is introduced into a protein generative model?  \n To investigate this question, we developed a novel protein generative model called SiBaSe \n(Sidechains, Backbone, Sequence). Incorporating sidechains into the design process necessitated \nthe inclusion of sequence information, as an amino acid’s identity is inherently defined by its \nsidechain. To isolate the effect of sidechain inclusion, we designed SiBaSe to be architecturally \nsimilar to a recent family of backbone models: FrameDiff, FrameFlow, and MultiFlow (Yim et al. \n2023a, Yim et al. 2023b, Campbell et al. 2024), henceforth the FrameFamily. By maintaining \narchitectural similarity with these models, we aimed to control for spurious variables and directly \nassess the impact of incorporating sidechain-level detail.  \n A key challenge in the co-generation of sidechains and backbones is the inherent \nsequence dependence of sidechains: their geometry and identity are tightly coupled. In de novo \nprotein generation, however, the sequence is initially unknown, requiring a sidechain \nrepresentation that is independent of residue identity. SiBaSe addresses this by representing \nsidechains using the same local-frame based system employed for the backbone. This choice \nenables the model to manipulate sidechain geometries without requiring prior knowledge of the \nsequence. Additionally, it facilitates seamless architectural integration, minimizes memory usage, \nand supports the use of the rotationally invariant Invariant Point Attention (IPA) module (Jumper \net al. 2021). \n To evaluate the impact of sidechain inclusion on generative performance, we compared \nSiBaSe to other models in the FrameFamily across three key metrics: designability , diversity , and \nnovelty . To further probe the relative contribution of each data type (backbone, sidechain, and \nsequence) we conducted a partial conditioning  experiment, selectively providing subsets of \nstructural information during generation. Finally, we performed a qualitative assessment of model \nbehavior, including sequence confidence , sidechain placement , and overall design patterns, to \nbetter understand how sidechain information influenced the generative process. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 3 \n 2 Results  \n 2.1 Unconditional Generation  \n The first test of SiBaSe was the unconstrained design of 450 proteins ranging from 75 to \n225 amino acids in length. Each of these generated proteins possessed a SiBaSe generated \nbackbone structure, sequence and sidechain placements. The designed sequence of each protein \nwas then provided to ESMFold (Lin et al. 2023), whose predictions were compared to the original \ndesigns to calculate designability (backbone-RMSD). Each of the outputs with designability <3 Å \nwere then compared to their nearest structural neighbor in the PDB, using the program FoldSeek, \nwhich returned a pdbTM value for each design (VanKempen et al. 2023). SiBaSe was capable of \ndesigning protein structures and sequences which achieved designability backbone-RMSD values \nas low as 2.42 Å (Fig 1). Despite this, out of the 450 designs only 7 had designability scores of less \nthan 3 Å (1.6%) This designability success rate is significantly lower than peer backbone-only \nmodels. It is important to note, the likelihood of successfully designing sequences and structures \nwhich fold at under 3 Å of accuracy by random chance is infinitesimal. Any level of success at this \ntask indicates the model had learnt a true mapping in protein sequence-structure space. \n \nFigure 1: Unconditional Co -Generation Sampling of SiBaSe:  Above are three proteins whose \nstructure, sequence and sidechains were generated using SiBaSe. The saturated color is the SiBaSe \noutput while the faded-colored and grey structures are the superimposed ESMFold and \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 4 \nAlphaFold2 predictions respectively. The alignments are shown as scRMSD and scRMSD-AF2 \nrespectively. pdbTM and nnSeq  record the TM score and sequence identity to the nearest \nneighbor in the PDB as determined by FoldSeek. The insets below show zoomed-in fragments \nwith sidechain designs/predictions displayed in same color scheme as above. The sequence and \ncorresponding fragment are color coded for amino acid identification. \n In order to assess the quality of the protein structures themselves, and not the quality of \nsequence-structure pairs, the same 450 designs were each given 3 new sequences designed by \nProteinMPNN (Dauparas et al. 2022). These sequences were then subjected to the same \nESMFold reconstructions for designability assessment as before (Fig 2). Additionally, all designs \nwith designability under 3 Å (the ‘success subset’) had a nearest-neighbor found in the PDB using \nFoldSeek with which to calculate pdbTM. The ‘success subset’ was then clustered via TM score \nfollowing established conventions (Herbert 2008), with the number of clusters divided by total \nproteins in the subset to arrive at a diversity score.   \n \nFigure 2: MPNN Redesigns of SiBaSe Designed Backbones:  Above are 6 designs from SiBaSe \nwhose sequences were redesigned using ProteinMPNN. The sequences were folded using \nESMFold and structures aligned and visualized as in Figure 1. The scRMSD and pdbTM \ndisplayed derived from same approach described in Figure 1. \n SiBaSe demonstrated generative behavior broadly consistent with other models in the \nFrameFamily across multiple key metrics (Table 1). Its generated structures were comparably The \naverage diversity of its generated structures was comparable to that of Frame Diff and exceeded the \ndiversity observed in FrameFlow. In the subset of successful designs, SiBaSe achieved slightly \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 5 \nimproved novelty relative to peer models. When evaluating designability, co-design success was \nobserved in 1.6% of cases, a lower rate than backbone-only models, including the co-generative \nMultiFlow. However, when sidechains were excluded and structures were re-sequenced using \nProteinMPNN, success rates improved substantially. At a 3 Å cutoff, designability reached 30%, \nand at the more stringent 2.5 Å cutoff used in prior FrameFamily publications, SiBaSe achieved a \n24% success rate. Together, these results suggest that while co-design introduces challenges for \nstructural fidelity, SiBaSe shares key generative characteristics with established backbone-focused \nmodels. \n \nTable 1: Comparison of SiBaSe to FrameFamily:  SDE – stochastic sampling strategy, ODE – \ndeterministic/flow sampling strategy, CoDesign – simultaneous design of structure and sequence, * \n- proportion < 3 Å vs <2.5 Å for other designability values, **- value of 1 not reported due to \noverinflation from small number of total proteins, *** - no published values \n In nearly all cases, sequences redesigned with ProteinMPNN achieved higher designability \nthan the original SiBaSe-generated sequences (Fig. 3A). While SiBaSe was capable of producing \nsuccessful designs, a gradual decline in designability was observed with increasing protein length \n(Fig. 3B). Although novelty showed a modest negative correlation with designability, the model was \nstill able to generate proteins that were both highly novel and structurally accurate (Fig. 3C). In \nterms of secondary structure composition, SiBaSe exhibited no clear bias, generating proteins with \ndiverse secondary structural composition (Fig. 3D). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 6 \n \nFigure 3: Analysis of SiBaSe Designs: Graphical representation of SiBaSe designs illustrating \ndesignability versus sequence origin, length and novelty. An additional panel illustrates secondary \nstructure diversity of designs. Panel A plots designability RMSD for SiBaSe designed sequence (x-\naxis) versus MPNN redesign (y-axis). Panel B plots designability RMSD (bester performer, y-axis) \nagainst protein length (x-axis). Panel C plots designability RMSD (y-axis) against novelty (TM, x-\naxis). Panel D plots proportion of secondary structure; y-axis is beta-sheet proportion, x-axis is \nalpha-helix proportion. \n 2.2 Conditional Generation  \n In an attempt to investigate the relative impact or importance of backbone and sidechain \ngeometric information during design, an experiment was conducted using conditioning. Briefly, \nconditioning involves providing the model with a fragment(s) of protein structure and/or sequence \ncalled the ‘motif’ and tasking it with generating a protein containing it, the ‘scaffold’ (Fig 4). The \nreplacement method of conditioning, while facing key limitations, was more than adequate for this \nexperiment and required no specialized training for the conditioning task (Trippe et al. 2022).  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 7 \n To evaluate how different types of structural information influence design, we tested \nSiBaSe under three distinct conditioning modes. In Mode 1, the conditioning motif included \nbackbone, sidechains, and sequence. Mode 2 used backbone and sequence, while Mode 3 \nprovided sidechains and sequence without the backbone (Fig 4A). Motifs were sampled from the \nProtein Data Bank (PDB) and selected to present increasing levels of difficulty: ranging from large, \ncontiguous secondary structure elements to small, non-contiguous fragments. For each mode–\nmotif combination, 30 designs were generated (Fig 4b). Each SiBaSe-designed sequence, along \nwith three ProteinMPNN redesigns (with the motif sequence fixed), was submitted to ESMFold for \nstructure prediction. Design–prediction pairs were evaluated using whole-protein designability \nscores as before, and additionally, motif-only RMSD  was computed in an all-atom manner to \nassess local structural fidelity. \n \nFigure 4: Conditioning Modes and Motifs: Above is a graphical representation of the 3 \nconditioning modes used in this work, the concept of conditioning, and the 8 motifs used in this \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 8 \nwork. Panel A shows three conditioning modes: 1 – backbone/sidechain/sequence, 2 – \nbackbone/sequence, 3 – sidechain/sequence. Panel B illustrates the process of conditioning, motif \nin red scaffold in grey. Panel C shows the 8 motifs used, the PDB ID of origin for each motif is \nshown next to the motifs. \n Designability scores, both for whole-protein structures and motif-level all-atom RMSD, \ndeclined with increasing motif complexity and across conditioning modes, with Modes 1 and 2 \nconsistently outperforming Mode 3 (Fig. 5A). Because RMSD is sensitive to the size of the \nstructure being evaluated, motif-level RMSDs were normalized to allow fair comparison across \nmotifs of varying sizes. This normalization accentuated the observed trend, reinforcing the \ncorrelation between reduced motif size, or lack of backbones, and lower design accuracy (Fig. 5B). \nThese results suggest that the model more effectively integrates backbone information than \nsidechain information during the generative process. Moreover, relatively small amounts of \nbackbone input were sufficient to guide design outcomes, whereas equivalent sidechain \ninformation had a more limited effect (Fig. 5B). \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 9 \n \nFigure 5: Conditional Sampling Evaluation : Above is a graphical representation, via boxplots, of \nthe designability of motif-conditioned outputs of SiBaSe. Panel A shows the 8 motifs (x-axis) in \norder of increasing ‘difficulty’ from left to right. Each motif has 6 boxplots (left to right) \ncorresponding to whole protein RMSD, modes 1, 2 and 3, then motif-all-atom RMSD, modes 1, \n2, 3. Panel B shows the normalized motif-all-atom RMSD for the 8 motifs, 3 boxplots correspond \n(left to right) modes 1, 2, 3. \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 10 \n 2.3 Sidechain Placement  \n Sidechain placement was analyzed by comparing 450 unconditioned SiBaSe designs to \n450 randomly sampled proteins from the training set. Amino acids were separated by type, and \ntheir backbone frames were aligned to assess sidechain translation. The central atom of each \nsidechain frame was plotted, and normalized Earth Mover’s Distance (nEMD) was computed to \nquantify similarity between real and designed point clouds (Fig. 6A). Placement accuracy varied by \namino acid type: smaller, more common, and less flexible residues (e.g., Leucine, Asparagine, \nValine, Glutamate) showed tighter cluster overlap and lower nEMD values, while larger, flexible, \nand rarer residues (e.g., Arginine, Methionine, Lysine) diverged more. A key observation is that \nthe placement of designed sidechains was slightly closer to the backbone as compared to the real \nprotein distributions for most of the amino acid types (exceptions being Valine and Leucine). \nSerine and Isoleucine were notable exceptions to this pattern with poor performance despite their \nsize and frequency in the training data. \n To assess orientation, the same central atoms were centered at the origin and the terminal \nframe atoms were projected onto a unit sphere (Fig. 6B). Orientation accuracy similarly depended \non residue type, with narrow, well-overlapping distributions for Tyrosine, Tryptophan, Valine, \nIsoleucine, and Threonine, and broader divergence for Arginine, Methionine, and Lysine. \nInterestingly, Serine and Isoleucine again showed strong orientation overlap despite translational \ninaccuracy. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 11 \n \nFigure 6: Sidechain Placement Visualization:  Above is an illustration of the sidechain frame \nplacements in SiBaSe designs versus real proteins. Panel A demonstrates the sidechain placement \n(translation) of real (blue) and designed (red) amino acids relative to the amino acid backbone \n(green). The normalized earth movers distance (nEMD) value is adjacent the amino acid name. \nPanel B shows the orientations of amino acids sidechains in a unit-sphere-like representation. Real \nsidechain frame terminal atoms are shown in blue, designs shown in red.  \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 12 \n 2.4 Sequence Generation Dynamics  \n Following the continuous-time Markov chain framework described in MultiFlow \n(Campbell et al. 2024), SiBaSe outputs a rate matrix that reflects the model’s confidence across all \n20 amino acid types at each position during generation. This enables direct visualization of \nsequence dynamics throughout the design process (Fig. 7). Initially, with the sequence tensor fully \nmasked, the model exhibits low confidence across all positions, as indicated by low color \nsaturation. Around timestep t = 0.33, confidence begins to increase and the model starts converging \non a prospective sequence. However, even after this point, many positions continue to undergo \nrapid turnover in predicted amino acid identity. The most confident residue at a given position \noften shifts significantly in both biochemical properties and size. One illustrative example \n(highlighted by a red arrow in Fig 7) shows a single position transitioning from alanine to \nphenylalanine, cysteine, glutamate, lysine, and finally leucine over the course of a single trajectory. \n \nFigure 7: Amino Acid Confidence Matrix : The above matrix visualizes the amino acid sequence, \nand confidence, of a SiBaSe generated protein during the design process. X-axis corresponds to \nthe amino acid sequence of designed protein in N-to-C order (left to right). Y-axis corresponds to \ngenerative timestep, top of graph 𝑡 = 0, bottom of graph 𝑡 = 1. Color corresponds to amino acid \ntype. Saturation indicates confidence, lower saturation indicating lower confidence.  \n 2.4 Intra-Generation Structural Dynamics  \n As a final assessment of model behavior during the generative process, a single design \ntrajectory was sampled, and the model’s outputs were visualized at each timestep (Fig. 8). Protein \nbackbone frames were rendered in purple and sidechain frames in green. In the final structure, \nsidechain frames were well-distributed, filling the space around and between backbone elements. \nEarlier in the trajectory (t < 0.75), however, sidechain frames remained closely clustered around \ntheir respective backbone frames. The backbone appeared to extend and adopt a coarse fold first, \nwith sidechain frames remaining tightly coupled to their backbones. This visualization highlights a \ntemporal separation in structure formation, with backbone geometry evolving earlier in the \ntrajectory and sidechain placement following. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 13 \n \nFigure 8: Visualization of Protein Generative Trajectory within SiBaSe: Above are 6 snapshots \nof the SiBaSe protein ‘prediction’, or output, for 6 timesteps during a single design trajectory. \nTimestep value of snapshot to the top left of each image. Backbone frames are colored purple and \nsidechain frames colored green. \n 3 Discussion  \n Sidechains play an indispensable role in protein folding and function (Dill 1990, Dill et al. \n2008, Spassov et al. 2007, Farber & Mittermaier 2008), so their inclusion in a generative model \nwas expected to produce significant deviations in behavior from backbone-only structural models. \nHowever, the most immediate and striking result was that SiBaSe, despite incorporating sidechain \ninformation, exhibited only modest differences from the backbone-focused FrameFamily models \nin terms of designability, novelty, and diversity. The co-generation success rate of 1.6% (at a <3 Å \nthreshold) was notably lower than that of MultiFlow, the only other model capable of joint \nsequence and structure generation. Nonetheless, even modest success in the co-design task \ndemonstrates that SiBaSe was able to learn a non-trivial mapping between structure and sequence \nspaces. When structure alone was evaluated, by remaking sequence with ProteinMPNN, SiBaSe \nachieved a substantially higher success rate of 30%, revealing a stark contrast between structure-\nonly and co-generation performance. This divergence in structure-only vs cogeneration success \npointed toward a deeper phenomenon that was further explored through targeted analyses. \n In order to assess the relative contributions of the various kinds of information, \nconditioning was used as a method for controllable data-injection. During the design process, the \npositions/sequence of the motif inform the surrounding designed protein, and as such the models \nability to utilize this information serves as a practical proxy for the importance and interpretability \nof the injected data. As expected, designability was highest for larger and structurally simpler \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 14 \nmotifs. However, as motif size decreased, the performance of sidechain-only (Mode 3) outputs \ndeclined more rapidly than that of backbone-only (Mode 2) outputs. Visual inspection of selected \ndesigns (Supp. Fig. 3) revealed that small sidechain-only motifs were often disregarded entirely, \nwith final sidechain frames appearing up to 10 Å away from their corresponding backbone \npositions. This suggests that the model had more difficulty extracting actionable structural cues \nfrom sidechains alone, particularly in sparse or fragmentary contexts. \n Generative models generate from distributions they have learned, so visualizing the \nplacements of sidechains (tantamount to rotamer assignment) was expected to yield clues about the \nlearning that took place in sidechain space. In terms of translational accuracy, the model was \ngenerally able to assign plausible sidechain positions relative to the backbone, with performance \ncorrelated to amino acid size, frequency, and flexibility. However, the predicted sidechains tended \nto remain closer to the backbone than those in real proteins, suggesting a conservative placement \nstrategy. Notably, Serine and Isoleucine exhibited unexpectedly poor translational accuracy despite \ntheir high frequency and simple rotameric profiles. In contrast, orientation accuracy was more \nconsistent across residue types. This pattern may reflect differences in learning complexity, as the \nspace of orientations (the 𝑆𝑂(3) manifold) is more constrained and potentially easier to model \nthan the open-ended Cartesian space of translations. Overall, the model showed substantial \ndivergence in sidechain placement accuracy between amino acid types and had an overall pattern \nof timid translational placement. \n Sequence design in SiBaSe, following the continuous-time Markov chain approach of \nFrameFlow, proceeds via iterative updates of amino acid type probability distributions. A key \nadvantage of this framework is that it enables inspection of the model’s internal decision-making \ndynamics during generation. Early in the design process, the model exhibited high uncertainty \nacross the sequence, with low confidence in amino acid identity for the first third to half of the \ntrajectory. Even after convergence began, substantial stochasticity remained, both in confidence \nlevels and in the identity of the most probable amino acid at each position. In many cases, the \nmost confident residue type at a given site shifted drastically between chemically and structurally \ndistinct amino acids, suggesting ongoing competition between diverse sequence solutions. \n These patterns come into focus with the last data point, the visualization of a generative \ntrajectory. The models’ behavior was backbone-centric, in spite of the sidechain information \npresent. For the majority of the trajectory, the backbone was assuming a nascent-then-final shape \nwhile the sidechains were ‘dragged’ along by their respective backbones. Rather than any apparent \nsharing of burden between backbones and sidechains with respect to structure resolution, the \nmodel appears to have learned to be a backbone-first model with sidechains serving a supporting \nrole at best; and potentially a drag on the model’s design freedom at worst.  \n In answering why, we postulate that there is a ‘penalty of uncertainty’ which is a result of \nthe interplay between sequence and sidechain geometry. Throughout the design trajectory, \nsequence predictions remained volatile, and sidechains were often positioned conservatively, closer \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 15 \nto the backbone than observed in real proteins. While changes in sequence do not necessarily \nperturb backbone geometry, each shift in amino acid identity implies a distinct sidechain \nconformation, often requiring substantial spatial rearrangement. This geometric sensitivity \nintroduces inertia into the design process: the model defers confident sidechain placement until \nthe backbone stabilizes. As a result, sidechains are treated as secondary elements, trailing behind \nbackbone formation rather than serving as active guides. \n As a result, SiBaSe operates as a backbone-first model, with sidechains integrated late in \nthe process and contributing minimally to early structural decisions. This highlights a key \narchitectural limitation: without explicit mechanisms to resolve or mitigate the uncertainty from \nsidechain-sequence interdependence, the generative process defaults to the simpler, and more \nstable, backbone space. \n 4 Conclusions  \n SiBaSe was developed to explore the effect of incorporating biochemically rich sidechain \ninformation into generative protein design, and it was able to achieve near-peer performance in \nseveral metrics. Experimentation with the model revealed that despite having access to sidechain \ndata, the model ultimately adopted a backbone-centric strategy. This suggests that the presence of \nsidechains, when represented in a singular format, introduces uncertainty that the model is \nreluctant to resolve. This behavior is likely generalizable to other computational approaches that \ntreat sidechains deterministically during sequence generation. To meaningfully leverage the \ninformational richness of sidechains, future models may need to adopt representations that \naccommodate multiple evolving sidechain configurations in parallel. Such flexibility could enable \nmore effective exploration of the intertwined sequence–structure landscape and help overcome the \narchitectural inertia imposed by sidechain uncertainty.  \n 5 Materials and Methods  \n 5.1 Flow Models  \n SiBaSe was designed to emulate the behavior of FrameFlow/MultiFlow, and as such it is of \nthe Continuous Normalizing Flow (CNF) family of models (Yim et al. 2023a, Cambell et al. 2024). \nThese models learn a reversible mapping between two distributions, the data distribution, 𝑝!, and \nsome prior distribution (often a form of noise), 𝑝\", by integrating an ordinary differential equation \n(ODE) over a learned vector field, , 𝑣#. \n \t\n𝑑𝐳(𝑡)\n𝑑𝑡 = 𝑣#(𝐳(𝑡), 𝑡), 𝐳(0) ∼ 𝑝\", 𝐳(𝑇) ∼ 𝑝! \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 16 \n Flow matching is an approach to training in which, rather than solving the entirety of the \ndifferential equation, the model learns an approximation of the full vector field, 𝑢$(𝑥 ∣ 𝑥!), in a \nsupervised fashion (Lipman et al. 2022).  \n𝐿FM = 𝔼$,&!,&[‖𝑣$(𝑥) − 𝑢$(𝑥 ∣ 𝑥!)‖'] \n Progress along the vector field is quantified by a continuous variable, the timestep (𝑡), in \nthe range of [0, 1]. The noise priors exist at 𝑡 = 0, and true data at 𝑡 = 1. Intermediate levels of \nnoise in the range [0, 1] are arrived at via a linear interpolation along the vector connecting the two \npoints in either Euclidian space for coordinates, or the geodesic on the 𝑆𝑂(3) manifold for \norientations (Supp. Fig. 1). The selected prior distributions are a three-dimensional gaussian \ndistribution, 𝑁(0,1) ∈ \t ℝ( , and the uniform distribution in 𝑆𝑂(3), 𝑈(𝑆𝑂(3)), for coordinate and \norientation data respectively. Sequence, as categorical data, will be described in section 5.3.  \n 5.2 Frame Representation of Backbones and Sidechains  \n The backbones and sidechains of amino acids are represented as frames, mathematical \nobjects with a coordinate and rotation matrix which describes a local reference frame. Frames are \nconstructed, using the Gram-Schmidt orthogonalization process, from 3 points in space. \nBackbones are readily convertible in this format as they consist of a repeating pattern of N-C-Cα \natoms with consistent internal bond geometry.  \n Representing sidechains as frames has several benefits, namely the sequence-agnosticism \nof the frame representation and ease of incorporating these additional frames with those of the \nbackbones. In order to convert the sidechains into a frame representation it was necessary to create \na library of atom types for each amino acid type which met 3 criteria: \n -The atoms needed to possess the biochemically relevant atoms (ex. Charged atoms) \n -The atoms needed to be part of a rigid group for un-orthogonalization \n -The atoms needed to be distal to the backbone to capture full sidechain size \n The atoms which were determined best to meet the listed criteria are shown in Table 2. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 17 \n \nTable 2: Library of Amino Acid Sidechain Atoms for Frame Construction. The table above \nshows the library of atoms that would be used, for each amino acid type, to construct its sidechain \nframes. For amino acids with ‘NA’, the sidechain is either non-existent or exists as a rigid group \nwith no rotamers. In these cases, the backbone frame was duplicated to serve as a placeholder. \n 5.3 Categorical Data  \n Categorical data is not readily usable in its native format within a flow model. As such, \nSiBaSe take the continuous-time Markov chain approach from MultiFlow (Campbell et al. 2024). \nIn this approach, the sequence of the protein is represented as a one-hot encoded tensor for all \namino acid types and masking tensor. The model predicts a rate matrix, or confidence matrix, of \neach amino acid type for each position. Further description below. \n 5.4 SiBaSe Architecture, Training and Sampling  \n SiBaSe’s training dataset consisted of 3902 proteins from the Protein Data Bank (PDB) \nthat were monomeric, soluble proteins ranging from 75-225 amino acids in length, post-screened \nto remove proteins with missing loops/fragments. For each training step, SiBaSe would sample a \nprotein from the training set, a timestep value 𝑡, and matching-size noise tensors from the \n𝑁(0,1) ∈ \t ℝ( and 𝑈(𝑆𝑂(3)) distributions (for coordinate and rotations respectively). The timestep \nwas scaled for rotations so that any 𝑡 < 0.1 → 𝑡 = 0. Finally, an independent timestep was \nsampled for the sequence.  \n The geometric data was interpolated to the selected noise levels following the \nmathematical approach in FramFlow (Yim et al. 2023b). The sequence timestep was then used to \nguide the random masking of sequence information such that the proportion of masked positions \napproximated (1 − 𝑡) as in MultiFlow. The final noised protein was passed to the model to \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 18 \ngenerate a prediction for the original un-noised protein, \t𝑥!\n∧ = 𝜑(𝑥$, 𝑡). The model’s loss was \ntripartite, with the 𝑆𝐸(3) loss of the FrameFamily (ℒ*+( ), a cross-entropy loss on sequence \n(ℒ,-. ), and pairwise distance topology loss (ℒ$/0/1/23 ), being scaled by coefficients and then \ncombined to produce the final loss, ℒ$/$41. The topology loss was a MSE loss on pairwise \ndistances, with separate distance cutoffs for backbone-backbone (12 Å), backbone-sidechain (8 Å) \nand sidechain-sidechain (8 Å) components derived from the true structure distograms.  \n \nℒ5//67 = 1\n𝑁 E (\nnodes\n|𝑐!\n∧ − 𝑐!|')\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tℒ6/$ = 1\n𝑁 E (\nnodes\n|log\n8\n^\n∆\n− log8∆ |') \nℒ*+( = \t ℒ6/$ + \t ℒ5//67  \nℒ,-. = − 1\n𝑁 E (\n9/7-,\n𝑆(!)log\t 𝑆\n^\n(!)) \n𝕀=> = \t M1,\t\t\t\t\t\t𝑑=>\n$6?- \t ≤ 𝑐𝑢𝑡𝑜𝑓𝑓\n0,\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒       ℒ@@ = \t\n!\nA$$\n∑ 𝕀=>\n@@\n=,> X𝑑=>\n06-7 − 𝑑=>\n$6?- Y\n'\n \nℒ*B = \t\n!\nA%&\n∑ 𝕀=>\n*B\n=,> X𝑑=>\n06-7 − 𝑑=>\n$6?- Y\n'\n    ℒ@* = \t\n!\nA$%\n∑ 𝕀=>\n@*\n=,> X𝑑=>\n06-7 − 𝑑=>\n$6?- Y\n'\n \nℒ$/0/1/23 =ℒ@@ + ℒ*B + ℒ@* \nℒ$/$41 = 1\n1 − 𝑡$_514D0\t\".G\nMℒ*+( '' + \t ℒ*+( ()\n2 \t + \t ℒ,-.\n3 + \t ℒ$/0/\n3 \t[ \n SiBaSe was trained for 28 days (250 epochs) on a single T4 GPU using the Adam \nOptimizer (Kingman and Ba, 2014), with a learning rate of 1e-4, β1=0.9 and β2=0.999. Each \ntraining set sample was drawn singly and then duplicated along the batch dimension, with \nindependent timesteps and noise between batches, in a size-dependent manner to avoid memory \nwaste on padding. After each of the final 80 epochs, the model generated 30 proteins for \ndesignability quantification with ESMFold. The checkpoints of the top 4 epochs (lowest \ndesignability score) were retained and the checkpoint from epoch 239 was selected as the vinal \nversion (Supp. Fig. 4). \n To sample SiBaSe, the model begins by drawing from the prior described noise \ndistributions (𝑁(0,1) ∈ \t ℝ(  and 𝑈(𝑆𝑂(3))) to generate starting tensors of the desired protein \nlength. A corresponding fully-masked sequence tensor is generated. The model generates a \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 19 \nprediction for the true protein, \t𝑥!\n∧, at each step. Initially, \t𝑥!\n∧ was simply interpolated back to the \nnext timestep using the ODE approach of FrameFlow/MultiFlow. It was discovered that the model \nwould settle on sub-optimal solutions (backbone breaks and clashes) earth in design and not be \ncapable of escaping. To circumvent this, a stochastic approach was adopted. In the new approach, \neach time the model generates a prediction \t𝑥!\n∧, new noise tensors are sampled, and the prediction \nis noised using these new tensors to the next timestep input. In doing this, the model was capable \nof escaping early suboptimal solutions. \n Following FrameDiff/FrameFlow/MultiFlow, SiBaSe is constructed from repeating layers \nof invariant point attention (IPA), MLP, Transformer-encoder and Geometric Update layers \n(Supp. Fig. 2). The main difference between SiBaSe and the models in the FrameFamily is the \nabsence of edge representation. This is due to memory constraints that were created after the \naddition of sidechain frames, which effectively double the protein size. The model \nhyperparameters are as follows: \n Global Parameters:  node dim = 512 L = 4 \n IPA Parameters:  heads = 8  query points = 8 value points = 12 \n Transformer Parameters: heads = 8 layers = 4 \n SiBaSe has a total of 33.8 million trainable parameters.  \n \n 6 References  \n1. Yeh A, Norn C, Kipnis Y, Tischer D, Pellock S, Evans D, Ma P, Lee G, Zhang J, \nAnishchenko I, Coventry B, Cao L, Dauparas J, Halabiya S, DeWitt M, Carter L, Houk K, \nBaker D. 2023. De novo design of luciferases using deep learning. Nature 614:774–780. \ndoi:10.1038/s41586-023-05696-3. \n2. Lauko A, Pellock S, Anischanka I, Sumida K, Juergens D, Ahern W, Shida A, Hunt A, Kalvet \nI, Norn C, Humphreys I, Jamieson C, Kang A, Brackenbrough E, Bera A, Baker D. 2024. \nComputational design of serine hydrolases. Science. doi:10.1126/science.adu2454. \n3. Watson J, Juergens D, Bennet N, Trippe B, Yim J, Eisenach H, Ahern W, Borst A, Ragotte \nR, Milles L, Wicky B, Hanikel N, Pellock S, Courbet A, Sheffler W, Want J, Venkatesh P, \nSappington I, Torres S, Baker D. 2023. De novo design of protein structures and function with \nRFDiffusion. Nature 620:1089–1100. doi:10.1038/s41586-023-06415-8. \n4. Yim J, Trippe B, Bortoli V, Mathieu E, Doucet A, Barzilay R, Jaakkola T. 2023. SE(3) \ndiffusion model with application to protein backbone generation. arXiv. \ndoi:10.48550/arXiv.2302.02277. \n5. Dauparas J, Anishchenko I, Bennet N, Bai H, Ragotte J, Milles L, Wicky B, Courbet A, De \nHaas R, Bethel N, Huddy T, Pellock S, Tischer D, Chan F, Koepnick B, Nguyen H, Kang A, \nSankaran A, Bera A, King N, Baker D. 2022. Robust deep learning-based protein sequence \ndesign using ProteinMPNN. Science 378(6615): 49-56. doi:10.1123/science.add2187. \n6. Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, \nBates R, Zidek A, Potapenko A, Bridgland A, Meyer C, Kohl S, Ballard A, Cowie A, Romera-\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 20 \nParedes B, Nikolov S, Hassabis D. 2021. Highly accurate protein structure prediction with \nAlphaFold. Nature 596:583–589. doi:10.1038/s41586-021-03819-2. \n7. Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, Smetanin N, Verkuil R, Kabeli O, Shmueli Y, \nCosta ADS, Fazel-Zarandi M, Sercu T, Candido S, Rives A. 2023. Evolutionary-scale \nprediction of atomic-level protein structure with a language model. Science 379(6637):1123–\n1130. doi:10.1126/science.ade2574. \n8. Krishna R, Wang J, Ahern W, Sturmfels P, Venkatesh P, Kalvet I, Lee G, Morey-Burrows F, \nAnishchenko I, Humphreys I, McHugh R, Vaffeados D, Li X, Sutherland G, Hitchcock A, \nHunter C, Kang A, Brackenbrough E, Bera A, Baker D. 2024. Generalized biomolecular \nmodeling and design with RoseTTAFold all-atom. Science 384(6693). \ndoi:10.1126/science.adl2528. \n9. Afinsen C. 1973. Principles that govern the folding of protein chains. Science 181(4096):223–\n230. doi:10.1126/science.181.4096.223. \n10. Spassov Z, Yan L, Flook P. 2007. The dominant role of side-chain backbone interactions in \nstructural realization of amino acid code. ChiRotor: A side-chain prediction algorithm based \non side-chain backbone interactions. Protein Science 16(3):494–506. \ndoi:10.1110/ps.062447107. \n11. Dill K. 1990. Dominant forces in protein folding. Biochemistry 29(31):7133–7155. \ndoi:10.1021/bi00483a001. \n12. Dill K, Ozkan S, Shell M, Weikl T. 2008. The protein folding problem. Annu Rev Biophys \n37:289–316. doi:10.1146/annurev.biophys.37.092707.153558. \n13. Farber P, Mittermaier A. 2008. Side chain burial and hydrophobic core packing in protein \nfolding transition states. Protein Science 17(4):644–651. doi:10.1110/ps.073105408. \n14. Yim J, Campbell A, Foong A, Gastegger M, Jimenez-Luna J, Lewis S, Satorras V, Veeling B, \nBarzilay R, Jaakkola T, Noe F. 2023. Fast protein backbone generation with SE(3) flow \nmatching. arXiv. doi:10.48550/arXiv.2310.05297. \n15. Campbell A, Yim J, Barzilay R, Rainforth T, Jaakkola T. 2024. Generative Flows on Discrete \nState-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design. arXiv. \ndoi:10.48550/arXiv.2402.04997. \n16. Van Kempen M, Kim SS, Tumescheit C, Mirdita M, Lee J, Gilchrist CLM, Söding J, \nSteinegger M. 2023. Fast and accurate protein structure search with Foldseek. Nature \nBiotechnology 42(2):243–246. doi:10.1038/s41587-023-01773-0. \n17. Herbert A, Sterberg MJE. 2008. MaxCluster: a tool for protein structure comparison and \nclustering. https://www.sbg.bio.ic.ac.uk/maxcluster/ \n18. Trippe B, Yim J, Tischer D, Baker D, Broderick T, Barzilay R, Jaakkola T. 2022. Diffusion \nprobabilistic modeling of protein backbones in 3d for the motif-scaffolding problem. arXiv. \ndoi:10.48550/arXiv.2206.04119. \n19. Lipman Y, Chen RTQ, Ben-Hamu H, Nickel M, Le M. 2022. Flow matching for generative \nmodeling. arXiv (Cornell University). doi:10.48550/arxiv.2210.02747. \n20. Kingma D, Ba J. 2014. Adam: A method for stochastic optimization. arXiv. \ndoi:10.48550/arXiv.1412.6980. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint \n\n 21 \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted August 12, 2025. ; https://doi.org/10.1101/2025.08.08.669410doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}