A Generative Neuro‐Symbolic AI for Protein Sequence Design Authors: Marianne Defresne, Delphine Dessaux, Samuel Buchet, Lucie Barthe, Liza Ammar‐Khodja, Bessam Azizi, Valentin Durante, Gianluca Cioci, Simon de Givry, Alain Roussel, Luis F Garcia‐Alles, Thomas Schiex, Sophie Barbe Published: 2026 ABSTRACT ABSTRACTDeep learning has revolutionized computational protein design, enabling the generation of sequences that fold onto target backbones with unprecedented accuracy. However, state‐of‐the‐art inverse folding tools largely rely on auto‐regressive sampling. While powerful, this paradigm is increasingly recognized for its inability to “think ahead”, a crucial capacity to reliably create the complex, long‐range inter‐residue dependencies essential for most biological functions. To overcome these fundamental limitations, we introduced EffieDes, a generative neuro‐symbolic AI framework that synergizes the predictive capabilities of deep learning with the logical precision of automated reasoning. EffieDes leverages deep learning to encode the target backbone's fitness landscape into Effie—a fully decomposable probabilistic graphical model (Potts model). This landscape can then be rigorously explored by an automated reasoning prover to identify sequences that simultaneously satisfy complex design constraints and optimize backbone fitness. We validated this neuro‐symbolic approach through the design of orthogonal sequence pairs that adopt identical folds but exhibit selective self‐assembly, as well as the design of a de novo selective nanobody with nanomolar affinity for an immune‐evasive SARS‐CoV‐2 variant. EffieDes provides a robust architecture for precisely dissecting learned fitness landscapes, offering a new path toward proteins with highly optimized performances and sophisticated functional objectives.Keywords: bacterial microcompartments, generative AI, inverse folding, nanobodies, neuro‐symbolic AI, protein design, protein engineering Introduction Proteins serve as the fundamental molecular machinery of life, orchestrating the vast array of biochemical processes essential for any organism's survival. However, the repertoire of natural proteins represents only a negligible sliver of the theoretically possible sequence space. Protein design seeks to navigate this uncharted territory, employing strategies that range from the strategic modification of existing scaffolds to the de novo construction of entirely novel sequences. While rational design [1] and directed evolution [2] have historically dominated the field, they are increasingly complemented—or superseded—by computational frameworks [3, 4, 5, 6]. By leveraging burgeoning structural and genomic databases, these digital approaches facilitate sequence‐to‐structure mapping while circumventing the bottlenecks of exhaustive experimental screening. Central to this evolution is the rise of deep learning (DL), which has framed protein structure prediction and design as inverse operations of a fundamental biological puzzle, ushering in a new era of predictive and generative molecular biology. In the path from sequence to structure and eventually function, generative protein language models (PLMs) [7, 8, 9] are trained to sample sequences mimicking natural distributions. Although PLMs benefit from the vastness of known sequence space and can generate novel folds, they remain difficult to steer toward specific functional niches unexplored by nature. While PLMs have successfully produced functional enzymes [10] or GFP sequences [11], these designs typically adopt natural folds and retain significant (>30%) sequence similarity to known homologs. Consequently, the design of proteins with radically new functions or architectures most often relies on a structure‐based approach [3, 12]. Under this paradigm, a backbone supporting the desired function is first generated [13] and subsequently “inverse‐folded” to identify compatible sequences. Recently, inverse folding has been transformed by deep learning models such as GVP [14] and ProteinMPNN [15], among others [16, 17, 18, 19]. These models leverage structural data [20] to learn potent structure‐sequence mappings that frequently outperform physics‐based methods, both in silico and experimentally [15]. Nevertheless, inverse folding remains a formidable computational puzzle [21]. Atomic interactions form an intricate network connecting side‐chains to the backbone and to one another; a valid sequence must satisfy both local backbone compatibility and global self‐consistency. To manage this complexity, most neural architectures avoid predicting all amino acids simultaneously. Even models capable of parallel prediction, like CARBonAra [16], eventually resort to iterative refinement. This strategy is exemplified by auto‐regressive models such as ProteinMPNN [15], which fix amino acid identities sequentially, conditioned on previously designed positions. This approach is computationally efficient and grounded in the chain rule of probability. For a sequence of n amino acids s1,…,sn given a backbone B, it is expressed as Once all conditional probabilities P(si|s0.75 and pLDDT>75. AlphaFold2 is not specialized for single‐sequence structure prediction, which may explain its poorer performance here. For both models, a clear correlation was observed between pLDDT and TM scores, except for AlphaFold on the most challenging case of k=5. Although reducing the amino acid repertoire expectedly complicates the design landscape, ESMFold continued to discern high‐quality, foldable sequences even for k=5 (Figure 2h). By leveraging automated reasoning to optimize Potts models under complex constraints, EffieDes can steer sampling into fitness regions inaccessible to auto‐regressive methods. This bypasses the need for retraining—a process that is unfeasible here precisely because no training data exists for the specific design objective. This makes EffieDes a powerful tool for exploring a “dark” sequence space that remains out of reach of purely data‐dependent architectures. EffieDes can Design Symmetrical Multi‐Component Assemblies Protein assemblies are crucial for many biological processes [73, 74], and the ability to design specific protein–protein interactions opens up new opportunities for progress in a wide range of fields, from fundamental science to health and biotechnology [75, 76, 77, 78, 79, 80, 81]. A significant challenge in this area lies in precisely controlling the interactions between protein components to achieve the desired assembly. This challenge is even more critical when designing symmetrical assemblies composed of sequence‐distinct proteins sharing a common fold. To evaluate the ability of EffieDes to handle such complex settings, we applied it to the redesign of bacterial microcompartment shell proteins (BMC‐H), which natively form homo‐hexamers [82, 83, 84, 85] (Figure S6), into hetero‐hexamers composed of two distinct subunits (e.g., A and B, Figure 3a). Such a redesign is attractive from a biotechnological perspective, as hetero‐assemblies could expand the functional versatility of BMC scaffolds by potentially enabling asymmetric catalytic arrangements, tunable permeability, or modular surface display [86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98]. However, achieving such controlled hetero‐assemblies requires the simultaneous enforcement of symmetry constraints and a multi‐state design strategy [99, 100], in which the heteromeric state (AB, positive) is favored while the homomeric states (AA and BB, negative) are avoided (Figure 3a). Design of protein interfaces in self‐assembling heteromeric BMC‐H proteins. (a) Representation of the negative design strategy for BMC‐H proteins showing the heterohexameric positive state (AB) and the two homohexameric negative states (AA and BB). (b) Comparison of the pre‐minimization score differences between positive and negative hexameric states for mutant BMC‐Hs designed with EffieDes (top row) or ProteinMPNN (bottom row), using the small designable region. Δ¯A and Δ¯B are represented on the X and Y axes and each dot represents a design. The color scale indicates the Δ¯WT scores, with blue corresponding to heterohexamers (AB) that are likely weakly stable. Regardless of the scoring approach used, EffieDes designs exhibit better Δ¯A and Δ¯B, as highlighted by the averaged ΔA¯ and ΔB¯ (black dots) and their γ values. (c) Experimental screening of the associations between designed pairs of BMC‐H monomers A and B using the tripartite GFP technology. Maximal fluorescence (Fmax) values are given as a percentage of the WT RMM value (dark gray). The dashed black line highlights maximal signals emitted by negative controls: non‐associating pairs (buckwheat trypsin inhibitor or BWI), aggregating proteins (CcmK3 from Synechocystis sp. PCC 6803), or cell auto‐fluorescence (Empty: cells transformed with a pET26b empty vector). A 3D symmetrical hexamer design template was built from the crystal structure of the BMC‐H shell protein RMM (PDB ID = 5L38) [82], and two designable interfacial regions (small and large) were defined (see Computational and Experimental section and Figure S7). Both EffieDes and ProteinMPNN [15] were tasked with generating sequence pairs A and B that satisfy the negative design objective of selective heteromeric interaction (see Computational and Experimental section and Supporting Information). A total of 377 pairs (AB) of sequences were designed using EffieDes and 495 using ProteinMPNN. To quantitatively evaluate how effectively each method favored the hetero‐hexamer over the homo‐hexamers, all sequences were cross‐scored by both Effie and ProteinMPNN, and the normalized differences in score of the positive and negative states were computed (Δ¯A=EAB−EAAσ and Δ¯B=EAB−EBBσ, where σ represents the standard deviation of all ΔA and ΔB values; see Computational and Experimental section). Similarly, the stability of the heterohexamer AB relative to the wild type (WT) homomeric template was assessed by computing Δ¯WT=EAB−EWTσ. All the pairs of sequences designed by EffieDes showed a better score for the positive state AB than for the negative states AA or BB, as highlighted by the negative Δ¯A and Δ¯B Effie scores. In contrast, ProteinMPNN was unable to completely account for the negative design objective, as only ∼ 85% of its designs showed negative Δ¯A and Δ¯B ProteinMPNN scores (Figure 3b and Figures S8–S11). This trend was further captured by the composite metric γ=−ΔA¯−ΔB¯ (with ΔA¯ and ΔB¯ the averaged values, see Computational and Experimental Section), which quantifies progress along the intended optimization direction. Indeed, γEffieDes was greater than γProtMPNN regardless of the scoring method used for evaluation (Table S2). Strikingly, ProteinMPNN gave better ΔA¯ and ΔB¯ to EffieDes designs than to its own designs (Figure S11). This underlines the incapacity of auto‐regressive models to sample high‐quality sequences when strong conditioning requires the sampling of conditional distributions that drift far away from those specified in the fundamental chain rule. As a decomposable function, Effie also provides valuable insights into how mutations affect monomer interactions by separating intra‐ and inter‐chain scores. The ΔAinter and ΔBinter score differences (Figures S12 and S13) further highlighted the limitations of ProteinMPNN in solving this negative design problem, as its γinter was strictly negative (Table S3). The stability of all designed hexamers was investigated by constructing and minimizing the structures of all positive and negative states using the Rosetta beta_nov16 score function [101] (Figures S9–S15). Candidates from either approach were then selected for experimental testing based on their favorable Effie or ProteinMPNN score differences ΔA, ΔB, and ΔWT post‐minimization, as well as low root‐mean square deviation (RMSD) of the positive state AB backbone to the WT hexamer (see Computational and Experimental Section, Supplementary Discussion and Tables S4–S7). The tripartite‐green fluorescent protein (tGFP) technology [102, 103] was used to detect AB interactions in vivo upon co‐expression in Escherichia coli. This method relies on a GFP reporter split into three nonfluorescent portions, two of them being attached, respectively, to A or B. The reconstitution of the complete GFP in an environment containing the third missing GFP component, with concomitant fluorescence emission, requires the association of the two protein partners (A and B) (Figure S17). The experimental success rate of EffieDes was markedly higher than that of ProteinMPNN, with 86% (12/14) of designs exhibiting fluorescence levels well above negative controls compared to just 20% (2/10) for ProteinMPNN (Figure 3c). Furthermore, while nearly half of the EffieDes designs (6/14; 43%) exceeded 50% of the WT RMM positive control's fluorescence, this level of performance was not reached by any ProteinMPNN‐designed sequence (0/10). Analysis of protein contents by SDS‐PAGE corroborated the superior efficiency of EffieDes, as ProteinMPNN designs more frequently resulted in aggregation of a fraction of the expressed mutant proteins (see Supplementary Discussion and Figure S16). Among all candidates, the “Duo4” design from EffieDes showed the highest expression and interaction between AB protein partners. This was supported by fluorescence measurements after comparison of protein expression levels (Figure 3c) and by co‐purification experiments (see Supplementary Discussion and Figures S17 and S18). Moreover, the formation of Duo4 hexamers was confirmed by size‐exclusion chromatography (SEC) experiments, as well as their stability (see Supplementary Discussion and Figure S19). EffieDes thus successfully designed new BMC‐H proteins that self‐assemble in vivo. Overall, in silico and experimental results indicated that EffieDes can better address complex negative design problems involving symmetric objects than the auto‐regressive ProteinMPNN method for new associations of pairs of mutant BMC‐H proteins. EffieDes can Design Nanobodies Using De Novo Backbones Nanobodies, also known as VHHs (single‐domain antibody fragments derived from the variable domain of heavy‐chain‐only antibodies found in camelids), are increasingly used in biotechnology and therapeutics due to their ease of production, high stability, and ability to access cryptic epitopes. Engineering nanobodies using computational design methods can offer a fast and scalable alternative to conventional discovery approaches, such as animal immunization and the screening of immune or synthetic libraries. However, the structure‐based design of nanobodies remains challenging, as it requires an accurate modeling of their complementarity‐determining regions (CDRs), highly variable and functionally critical loops that govern antigen‐binding specificity and affinity [104]. While generative AI approaches, such as RFdiffusion [13, 105], can, in principle, de novo generate CDR backbones tailored to an antigen surface, the challenge remains to identify sequences that effectively encode these pre‐organized geometries while simultaneously optimizing the interfacial interactions for affinity and target specificity. This task is particularly demanding for highly flexible CDR loops, where small sequence changes can profoundly reshape the conformational and energetic landscape of the binding interface. Here, we evaluated whether the global optimization capabilities of EffieDes could design functional sequences for novel CDR backbones targeting an immune‐evasive virus. We centered our study on MR17, a synthetic nanobody (sybody) originally developed against the ancestral Wuhan strain of SARS‐CoV‐2, which binds to the receptor‐binding domain (RBD) of the spike protein and blocks its interaction with the human angiotensin‐converting enzyme 2 (ACE2) receptor [106]. It retains binding to the Delta variant RBD, but fails to interact with the Omicron XBB.1.16 variant, which carries 11 mutations at the MR17 binding interface compared to the Wuhan strain (Figures S27 and S28). As MR17 targets a functionally critical epitope, it provides a relevant structural template to assess whether EffieDes‐based sequence design on de novo CDR backbones can rapidly retarget a nanobody toward an emerging immune‐evasive variant. To achieve this, we generated novel CDR backbones for MR17 [106] in complex with the SARS‐CoV‐2 XBB.1.16 RBD by partial diffusion using RFdiffusion [13] and subsequently designed sequences by either ProteinMPNN sampling or EffieDes optimization, using global diversity constraints [48, 51] to ensure a systematic exploration of the highest‐fitness regions of the landscape. A total of 5600 sequences were sampled with ProteinMPNN and, independently, 90 optimized with EffieDes. These designs were evaluated primarily based on predicted ΔG binding affinities computed with Rosetta's InterfaceAnalyzer and the beta_nov16 score function [101, 107] (Table S10), and with MM/GBSA and MM/PBSA binding energy estimates [108, 109, 110, 111, 112] from molecular dynamics (MD) simulations (see Supplementary Methods). Additional structural criteria, such as interfacial hydrogen bond propensity and interface size, were also considered throughout the MD trajectories (Tables S11 and S12). A limited set of nine nanobody candidates—eight from ProteinMPNN and one from EffieDes—was selected for experimental testing (Figure S30 and Table S13). Of the nine designs tested, only the EffieDes‐derived nanobody (NbRM‐E1, Figure 4a) exhibited binding to the SARS‐CoV‐2 XBB.1.16 RBD in biolayer interferometry (BLI) experiments (Figure S21). While this restricted experimental validation cannot provide a quantitative benchmark of overall success rates, it provides a proof‐of‐concept that EffieDes can identify a functional solution, where ProteinMPNN fails under the same screening pipeline. Notably, NbRM‐E1 binds to its target with a higher affinity than MR17 to the Wuhan RBD [106]: KD for MR17 was =83.7nM (koff=9.710−3s−1) versus 64nM for NbRM‐E1 (koff=4.7410−3s−1; Figure S24). Furthermore, while MR17 interacts with the RBD of the SARS‐CoV‐2 Delta variant, NbRM‐E1 selectively binds to the XBB.1.16 RBD, with no detectable interaction with the Delta RBD (Figures S25 and S26). Design of nanobodies binding the RBD of SARS‐CoV‐2 XBB.1.16. (a) 3D structural model of nanobody NbRM‐E1 (cyan) in complex with the SARS‐CoV‐2 XBB.1.16 RBD (green). Interfacial residues are represented as lines; on NbRM‐E1, the 7 native designable residues are colored in orange and the 29 residues distinguishing NbRM‐E1 from MR17 are shown as dark brown sticks. (b) 3D model showing the binding sites of each protein partner (NbRM‐E1: cyan, ACE2: pink, and VHH‐72: orange) on XBB.1.16 RBD (green), generated by aligning the backbone of the XBB.1.16 RBD with that of the RBD in complex with ACE2 (XBB.1.5, PDB ID: 8VKP) [113] or with VHH‐72 (SARS‐CoV‐1, PDB ID: 6WAQ) [114]. (c) BLI competition assays showing the inhibition of ACE2 binding when NbRM‐E1 is interacting with XBB.1.16 RBD (dark brown) and the binding of VHH‐72 nanobody to the complex NbRM‐E1/RBD (green), which supports the predicted binding site of NbRM‐E1. BLI‐based competition assays (see Computational and Experimental Section) showed that the NbRM‐E1 nanobody effectively blocked the interaction between the XBB.1.16 RBD and its ACE2 receptor. Furthermore, NbRM‐E1 allowed for the simultaneous binding of VHH‐72—a broad neutralizing nanobody targeting a conserved, non‐overlapping epitope [114, 115] (Figure 4b,c). This dual observation—ACE2 blockade coupled with a lack of competition with VHH‐72—provided a robust functional mapping that aligns with the binding interface proposed in our structural model (Figure 4a,b). EffieDes thus proved its ability to design a functional sequence on a de novo CDR backbone that enables binding to the SARS‐CoV‐2 XBB.1.16 RBD, thereby successfully retargeting it toward an immune‐evasive viral variant. The combination of backbone generation and EffieDes sequence design thus provides a rapid and effective strategy for engineering nanobodies in response to emerging viral threats. Discussion The rapid ascent of deep learning for inverse folding has been largely driven by the efficiency of auto‐regressive architectures. However, our findings suggest that the reliance on sequential, token‐by‐token decoding—effectively a “Next‐Token Prediction” approach—imposes a structural ceiling on design complexity. In inverse folding, where long‐range side‐chain coordination is paramount, the inability of these models to “think ahead” leads to an accumulation of local errors. By replacing sequential sampling with the global optimization of a backbone‐dependent Potts model, EffieDes restores the ability to account for the intricate, long‐range inter‐residue interactions inherent in protein folding. This shift from a “greedy” local walk to a global Maximum a Posteriori inference allows for the exploration of fitness landscapes that are mathematically unreachable by current autoregressive standards. The most significant advantage of the neuro‐symbolic framework is its capacity for “hard” conditioning without the need for specialized retraining. Our results on the BMC‐H hexameric platform provided a clear demonstration of this: while ProteinMPNN struggled to balance the competing requirements of symmetry and negative design, EffieDes identified viable sequences by treating these requirements as formal logical invariants. The fact that ProteinMPNN, during in silico cross‐evaluation, favored sequences designed by EffieDes over its own suggests that the failure of autoregressive models in these settings is not a lack of “knowledge” within the weights but a failure of the sampling process itself under strong conditioning. This “sampling breakdown” occurs because, as constraints are added, the effective search space narrows to a region that the sequential chain rule can no longer accurately approximate. Furthermore, while the limited number of experimentally tested candidates frames the nanobody design primarily as a proof‐of‐concept, the successful identification of the NbRM‐E1 nanobody against an immune‐evasive SARS‐CoV‐2 variant highlights a second critical strength of EffieDes: data efficiency. By utilizing the inductive bias of a Potts model—which naturally captures the pairwise nature of physical interactions—EffieDes demonstrated an ability to generalize to de novo CDR loop structures that exist outside the training distribution. This suggests that neuro‐symbolic architectures may be less prone to the “memorization” pitfalls of pure deep learning, offering a more robust path for designing proteins for which evolutionary data is sparse or nonexistent. The nanomolar affinity and ACE2‐blocking capability of NbRM‐E1 confirmed that this theoretical advantage translates directly into functional, high‐performance molecular tools. To further expand the reach of EffieDes, several architectural evolutions can be envisioned. While current inverse folding tools are beginning to incorporate non‐protein ligands [16, 22], their authors acknowledged the limitations of pure DL methods, particularly in low‐data regimes, suggesting their hybridization with physics‐based methods. Effie score functions would be a fitting candidate as they can already be combined with physics‐based pairwise score functions for ligands, directly or using a new all‐atom version of EffieNN. Finally, while the current pairwise Potts representation is highly effective, extending EffieNN to capture many‐body interactions could further refine the resolution of the fitness landscape, paving the way for the design of increasingly sophisticated molecular machines. Computational and Experimental Section Protein Structure Representation and Dataset The input protein backbone B is defined by the 3D coordinates of N, Cα, C, and O atoms for n residues, such that B∈Rn×4×3. To enforce the SE(3) (global rotation and translation) invariance of protein structures [116] EffieNN encodes backbone geometry using invariant pairwise features between residue pairs (i,j): Interatomic distances: for each residue pair (i,j), all pairwise distances between the backbone atoms (N, Cα, C, O, and virtual Cβ) of residues i and j are computed [15]. The resulting 25 distances are encoded using 16 Gaussian radial basis functions with centers evenly spaced between 0 and 20 Å. Interatomic distances: for each residue pair (i,j), all pairwise distances between the backbone atoms (N, Cα, C, O, and virtual Cβ) of residues i and j are computed [15]. The resulting 25 distances are encoded using 16 Gaussian radial basis functions with centers evenly spaced between 0 and 20 Å. Positional encoding: a 16‐dimensional sinusoidal embedding is used to encode the number of residues between positions i and j in the sequence (|j−i|) [117]. Positional encoding: a 16‐dimensional sinusoidal embedding is used to encode the number of residues between positions i and j in the sequence (|j−i|) [117]. Chain identity: for multi‐chain proteins, a Boolean feature indicates whether residues i and j belong to the same chain, to learn interfacial properties [118]. Chain identity: for multi‐chain proteins, a Boolean feature indicates whether residues i and j belong to the same chain, to learn interfacial properties [118]. The final representation of the protein backbone is a matrix of pairwise features with shape (n,n,25×16+16+1). Residues with missing atomic coordinates in experimentally resolved structures are not predicted and excluded from all metrics. EffieNN was trained on two datasets extracted from the Protein Data Bank [20]. The first single‐chain dataset, assembled by Ingraham, Garg, Barzilay, and Jaakkola [61], is composed of single‐chain protein structures. Training, validation, and test sets (respectively, 17 000, 600, and 1200 proteins) were split based on a sequence similarity cutoff of 40% and CATH 4.2 classification to ensure that similar proteins are in the same set. This dataset has been widely used to train and test inverse folding architectures, allowing for fair comparison across methods. The second multi‐chain dataset, used for ProteinMPNN training, is composed of protein assemblies clustered into training (23 358), validation (1464), and testing (1539) sets. Each assembly can be associated with multiple sequences and/or conformations, one of which is randomly sampled at each training epoch. This dataset overcomes a key limitation of the single‐chain dataset, which can fragment biologically relevant complexes into isolated chains and overlook interfacial properties. Only soluble proteins shorter than 1000 amino acids were retained. Neural Architecture and Training EffieNN comprises two main components, as illustrated in Figure 1. The first component extracts information from the local environment of each residue, defined by its 128 nearest neighbors. For proteins smaller than 128 amino acids, zero‐padding was applied. The environment is processed using blocks composed of a gated Multi‐Layer Perceptron (gMLP) [60] and an MLP with residual connections [119] (ResMLP). Inspired by Graph Neural Network architectures, the same gMLP+ResMLP block is repeated 6 times, with the output of the previous iteration added to the input. The final output is averaged over the neighbors to produce an embedding of the target residue's environment. The second component, a single ResMLP, predicts pairwise score matrices Ei,j. For all pairs of neighboring residues (the k=128 nearest neighbors), the concatenated environment vectors of each residue in the pair, along with the distance and positional features, are fed to this ResMLP to produce a 20×20 matrix. The set of all such pairwise matrices defines the generated Effie score function, conditioned on the input backbone B. An ablation study and detailed hyperparameters are available in the Supporting Information. EffieNN training is self‐supervised to maximize the probability of the sequences observed in the training set given their associated backbone. The exact loglikelihood criteria being intractable for Potts models, it is usual to use Besag's pseudo‐loglikelihood (PLL) [57]. We relied instead on the E‐PLL [59], which avoids the known incapacity of the PLL to estimate high‐energy terms [58] by randomly masking a fraction of pairwise terms during training. To accommodate variable protein sizes, the E‐PLL eliminates 10% of the pairwise terms connecting a residue to its neighbors during training. A L1 regularization of 10−4 was applied to the predicted score matrix to encourage sparsity and simplify discrete optimization. Following preliminary training that showed constant zero functions estimation for pairs of distant vertices, and consistently with physics‐based score functions, no function is predicted for pairs of residues separated by more than 15 Å (cutoffs in recent Rosetta's score functions range from 6 to 12 Å [120]). The absence of detectable long‐range correlations does not mean that they do not exist, as they may just vary based on various protein properties (function,...) that are not available to EffieNN. EffieNN was trained using the Adam optimizer, with a weight decay of 10−3, an initial learning rate of 5.10−4, and a learning rate divided by 10 when the validation loss plateaued (with patience 0) until it reached 10−8. Each protein in the training set was treated as a batch of n residues (of varying size). This variation becomes important in the multi‐chain dataset, where proteins have from 50 to 1000 amino acids: the learning rate was set proportionally to the square root of the batch size [121]. In Silico Evaluation of EffieDes In our architecture, EffieNN is optimized by the automated reasoning prover toulbar2 [39]. Three algorithms can be used: exact optimization with the Hybrid Best‐First Search algorithm [122], low‐rank convex relaxation optimization LR‐BCD [66] for efficient solving, and a custom bi‐objective solver developed for multi‐state design. Exact optimization was used for benchmarking against Rosetta on small proteins and for the redesign of RNA‐polymerase, using toulbar2's Python interface, PyToulbar2 (version 0.0.0.2). For large proteins (e.g., NSR comparisons on the full single‐chain dataset), optimization was performed using LR‐BCD with 50 rounding runs, 5 iterations, and option −4 for the rank. The sequence with the lowest predicted score among the 50 produced was retained. LocalColabFold [123] (version 1.5.2) with 3 recycles was used to assess the quality of the designed proteins through forward folding. RNA Polymerase Redesign The crystal structure of a previously designed double‐ψ β‐barrel (DPBB), composed of seven amino acid types [71] (PDB ID: 7DXZ) was used as a starting point. The symmetry axis between the two DPBB monomers was determined using AnAnaS [124], and the backbone was idealized to enforce perfect symmetry with Rosetta. Each monomer consists of 42 amino acid residues. Four Effie score functions were predicted by four versions of EffieNN trained on the single‐chain or multi‐chain dataset, with or without added backbone noise of 0.2 Å during training and optimized exactly with toulbar2. To constrain the number k of amino acid types allowed in the design, an N‐value constraint [72] was encoded as a generalized linear constraint [49] in toulbar2, with auxiliary Boolean variables indicating whether each amino acid type is used. Without any constraint, the optimal sequence uses 12 different types of amino acids among the 20 available. Constrained design was performed for k∈5,6,7 amino acid types by enumerating all sequences within 10 units of the optimal score. The 100 best‐scoring sequences (per k and per Effie model) were retained for downstream analysis; for k=5, all valid sequences found (fewer than 100) were kept. Implementation of Multi‐State Design Approaches for the Redesign of BMC‐H Proteins We used EffieNN trained on the multi‐chain protein dataset, with 0.14 Å std‐dev Gaussian noise added to the atomic coordinates during training. EffieNN independently processed both the positive and negative states, and the resulting decomposable score function was projected on monomers based on the symmetrical properties of each state. The approximate Pareto front was produced for two designable regions of, respectively, 29 and 39 residues, as described in Supporting Information. This produced 450 (resp. 778) solutions for the small (resp. large) region (Figure S4). Following sequence clustering, the sequence minimizing the absolute value of the difference in Effie scores between the two negative states, AA and BB, was selected in each cluster, producing 210 and 204 unique sequences for evaluation. To favor optimization of the positive state, only sequences for which |λ−|<λ+ were kept, with λ− the weight applied on the negative states and λ+ the one applied on the positive state, resulting in a final count of 189 (resp. 188) sequences. ProteinMPNN was used with its default parameters: model v_48_020, trained with 0.20 Å Gaussian noise, and a sampling temperature T=0.1. The positive state (AB) was assigned a weight λ+=1. As with EffieDes, to give more importance to the positive state, weights λ− ranging from −0.9 to −0.1 by increments of 0.1 were tested for the negative states (AA and BB), using the same weight for both states. Symmetrical chains in all states were tied using the homooligomer option of the make_pos_neg_tied_positions_dict.py script of ProteinMPNN. Combinations with a negative weight between −0.45 and −0.25 were determined to be the most suitable for our design objective and were explored with a finer resolution of 0.01 (see Supporting Information). For each weight combination, 10 designs were generated, for a total of 280 pairs of sequences for each designable region. Duplicate sequences were removed, leading to a total of 236 (resp. 259) unique designs. Construction of Conformations and In Silico Assessments of the Redesign of BMC‐H Proteins After identification of the symmetry axis of the x‐ray structure of the BMC‐H shell protein MSM0272 forming the Rhodococcus and Mycobacterium microcompartment (RMM) from Mycobacterium smegmatis MC2 155 (PDB ID = 5L38) [82] using AnAnaS [125], Rosetta cyclic C6 symmetry definition [126] was used to build a symmetrical hexamer conformation, with a 60° rotation between each asymmetric subunit and a distance of 17.1 Å between the center of mass of the subunits and the center of the cyclic hexamer. This hexamer was then minimized using Rosetta ref2015 Cartesian scoring function [120] with the same symmetry constraints. The resulting conformation is denoted as the wild‐type (WT) template and was used for the design of heteromers AB (Figure 3a) and as a reference for design evaluation. A small (29 residues) and a larger (39 residues) designable region were defined at the interface between the subunits (Figure S7). Mutant conformations of the three different states considered in our multi‐state designs (one heterohexamer AB and two homohexamers AA and BB) were built by mapping the designed sequences on the WT template conformation using our in‐house Custozyme+ tool, side chain packing all residues using Rosetta beta_nov16 scoring function [101]. These mutant conformations were then minimized using the Rosetta minimize executable with the multi‐step L‐BFGS algorithm (lbfgs_armijo_nonmonotone), a minimization tolerance of 0.001, and beta_nov16 scoring function. No symmetry constraints were applied, allowing for large conformational changes in the hexameric conformations (Figures S12 and S13). Conformations pre‐ and post‐minimization were scored using both Effie and ProteinMPNN and the scores differences between the positive (AB) and each negative (AA or BB) hexameric state ΔA=EAB−EAA and ΔB=EAB−EBB were computed, as well as those between the positive state and the wild‐type homomeric RMM template ΔWT=EAB−EWT. To account for possible discrepancies in the score units from Effie and ProteinMPNN, score differences were normalized by the standard deviations over all ΔA and ΔB values from Effie scores or from ProteinMPNN scores (σEffie=171.134, σProtMPNN=64.1453). For a design scored with Effie, this defines Δ¯AEffie=ΔAEffieσEffie. The averaged values ΔA¯ and ΔB¯ were calculated for each approach (e.g., for each design i, ΔA¯=1n∑i=1nΔ¯Ai), as well as the scalar product γ=u⃗·v0⃗ with u⃗=ΔA¯,ΔB¯ and v0⃗=(−1,−1). Similarly, the Effie inter‐chain score differences were calculated for all designs, and normalized by the standard deviation on all ΔAinter and ΔBinter values (σinter=148.985), as well as the averaged values and the scalar product γinter. Production and Characterization of the Designed BMC‐H Proteins Full‐length DNA fragments coding for either Duo‐GFP10 or Duo‐GFP11 were synthesized by Twist Bioscience (sequences provided in Table S8). Each pair of fragments was then Gibson‐assembled within a NdeI/SalI‐treated pET26b‐based vector‐1 (KanR) that also coded for an His6‐tagged GFP1‐9 under an independent T7‐transcription control. Gibson assembly reactions were carried out with the NEBuilder Hifi DNA Assembly Master Mix (E2621, NEB) following supplier instructions. Reactions were transformed in TOP10 cells and plasmids recovered by standard purification procedures [103]. Final vectors were verified by restriction and sequencing (Eurofins). Protein expression studies were performed after the transformation of chemically‐competent BL21(DE3) E. coli cells with the plasmids. A preculture of each strain in Luria–Bertani broth (LB, 100 μL) media supplemented with kanamycin (40μgmL−1 final conc.: LBK media) was grown for 6 to 8 h at and 200 rpm orbital‐shaking. Then, 2μL were seeded into LBK (200 μL) supplemented with IPTG (10μM), previously dispensed in a 96‐well black plate with glass flat bottom (Greiner, reference 655892). Cultures were continued in a CLARIOstar Plus (BMG Labtech), at with continuous shaking. Optical density (at 600nm) and GFP fluorescence (excitation wavelength at 470±15 nm, emission at 515±20 nm) were recorded approximately every 10 min, for 15 h. Data were fitted to the following sigmoidal function using PRISM 6 (GraphPad) y=Fbasal+Fmax−Fbasal1+10((log10(halfFmax)−x)×slopefactor) where halfFmax is the OD for which F=Fmax−Fbasal2. For cases with too weak fluorescence readings, the maximum fluorescence values Fmax were extracted manually (typically the value read after 12–14 h of culture). Plotted averages and standard deviations were derived from at least two independent experiments, each one including several replicates per case (typically 3 clones). The data presented were normalized with respect to the value of the RMM pair, which served as positive control. Further experimental characterizations are detailed in Supplementary Methods. Design of Nanobodies Using De Novo Backbones A structural model of the MR17 synthetic nanobody in interaction with the SARS‐CoV‐2 XBB.1.16 RBD was constructed by aligning the crystal structure of the MR17‐RBD Wuhan complex (PDB 7C8W) [106] with the cryo‐EM structure of the XBB.1.5 RBD‐ACE2 complex (PDB 8VKP) [113] using PyMOL [127]. The K478 residue in the XBB.1.5 RBD was mutated to arginine using Custozyme+ to model the XBB.1.16 RBD. Missing residues in the MR17 structure and steric clashes at the MR17‐RBD interface were resolved with MODELLER (v10.5) [128]. The resulting MR17‐XBB.1.16 RBD complex was used as input to RFdiffusion [13] to generate multiple novel CDR backbones through its denoising neural network. Partial diffusion (github.com/RosettaCommons/RFdiffusion) was used in combination with the radius of gyration (ROG) potential (partial_T∈{1,3,5,7,9,11,13,15}, guide_scale∈{0.25,0.4,0.55,0.7,0.85,1.0,1.15} with a cubic guide decay). The noising/denoising process was applied solely to the sybody structure and parts of its sequence were masked during the generation process (28 residues corresponding to the 3 CDRs of MR17). In this setting, an input structure is partially noised to serve as the initial input instead of starting with a completely random structure. Ten structures were generated for each combination of the given parameters by varying the random seed, resulting in a total of 560 structures. ProteinMPNN was applied to these 560 backbones to design 10 sequences per structure, targeting a designable region of 36 residues, including the three CDRs and additional interface residues (Figure S26). Sequence design used a sampling temperature of T=0.1. Conformations of nanobody–RBD complexes were constructed using Custozyme+, relaxed using Rosetta FastRelax protocol [129] with the beta_nov16 score function [101]. Disulfide bonds between relevant cysteines were fixed (4 disulfide bonds on the RBD and 1 on the nanobody). Interactions of the top 100 complexes, ranked by Rosetta InterfaceAnalyzer ΔG [107], were further analyzed. From these, 14 nanobody designs showing the best Rosetta's binding scores with the RBD, as well as having the best propensity to form hydrogen bonds and a high sequence diversity, were chosen for further characterization. 50 ns MD simulations were performed, using AMBER18 [130] with the all‐atom ff14SB AMBER force field [131] under near‐physiological conditions (T=310K, P=1bar, protonation states of titrable residues assigned at pH 6.5 using the pdb2pqr server [132], 150 mM NaCl concentration with the number of ions computed using the SPLIT method [133]). Details on MD simulations are provided in the Supporting Information. The three most stable complexes along MD simulations were then used as templates for EffieDes sequence design, using EffieNN (trained on multi‐chain protein dataset with added std‐dev 0.2 Å Gaussian noise on atomic coordinates) and the exact optimization algorithm of toulbar2. For each template, a total of 30 sequences were designed with, as above, a designable region containing 36 residues, and diversity constraints set to 2 amino acids between sequences. Six complexes were selected for MD simulations based on the same criteria as above: Rosetta binding scores with the RBD, propensity to form interfacial hydrogen bonds, and sequence diversity. From both ProteinMPNN and EffieDes pipelines, the 9 nanobodies (Table S13) exhibiting the strongest interactions with RBD XBB.1.16, based on MM/GBSA and MM/PBSA calculations [108, 109, 110, 111, 112] over 50 ns MD trajectories, as well as showing the highest numbers of interfacial contacts and hydrogen bonds, were selected for experimental characterization. These nanobodies were named NbRM‐P or NbRM‐E (for Nanobodies generated with RFdiffusion from the MR17 structure with a sequence predicted by ProteinMPNN or EffieDes). Experimental Characterization of Nanobodies Interaction With SARS‐CoV‐2 RBDs RBDs of the SARS‐CoV‐2 spike protein (strains XBB.1.16 and Delta) were expressed in EXPI‐293F cells (A14527) (Thermofisher), in fusion with a N‐ter signal peptide and a C‐ter His‐tag, using the eukaryotic pYD11 expression plasmid under CMV promoter control. Genes encoding nanobodies were cloned into pET24 expression vector, enabling expression of a C‐terminal His‐tagged protein. Plasmids were transformed into E. coli BL21(DE3) cells. For each nanobody, a 100 mL overnight preculture was prepared from a single colony in 2YT (or LB) medium supplemented with 50μgmL−1 kanamycin and incubated at with shaking at 200 rpm. Small‐scale cultures of 12 mL LB medium were inoculated and incubated at . When the optical density at 600 nm reached 0.6, protein expression was induced with 1 mM IPTG. Cultures were incubated for an additional 5 h at under agitation before harvesting. Cells were then collected by centrifugation (5000 × g, 10 min, ) and pellets were resuspended in 2 mL of lysis buffer (50 mM Tris‐HCl pH 8.0, 300 mM NaCl, 1mgmL−1 lysozyme, 5μgmL−1 DNase I, 5 mM MgCl2). Cell disruption was achieved by three freeze–thaw cycles (liquid nitrogen followed by water bath). Lysates were clarified by centrifugation (15 000 g, 15 min, ), and the soluble fractions were collected for western blot and BLI analyses (see Supporting Information for more details and Figures S20 and S21). Nanobodies and RBD proteins were both purified by affinity of their His‐tag by IMAC on a nickel‐charged resin, followed by size‐exclusion chromatography using a Superdex 75 16/60 column (see Supporting Information and Figures S22 and S23 for production and purification details). Biolayer interferometry (BLI) was used to measure nanobodies interaction with the RBD. It is a label‐free optical biosensing technique that measures biomolecular interactions in real time by monitoring wavelength shifts in the reflected light spectrum of a sensor surface during analyte binding. BLI experiments were performed on an Octet system (Octet RED96) (FortéBio, CA). A black bottom 96‐well microplate (Greiner Bio‐One # 655209) was filled with 200 µL of solution (nanobodies in PBS buffer) and agitated at 1000 rpm, at . Tips were hydrated in PBS buffer for 20 min at room temperature prior to further experiments. Biotinylated SARS‐CoV‐2 RBDs (5μg.mL−1) were loaded on streptavidin SA (18‐0009) biosensors (Pall ForteBio) for 1 min. After a baseline step in assay buffer (PBS [pH 7.4], 0.1% bovine serum albumin, 0.02% Tween 20) and a quenching step in 5μg.mL−1 biocytin, RBD‐loaded sensors were dipped into known concentrations of NbRM‐E1 nanobody (or total cell extract for interaction screening) for an association phase during 200–300 s. The sensors were then dipped in assay buffer for a dissociation step during 500 s. The association and dissociation curves were globally fitted to a 1:1 binding model. KD was determined by fitting binding curves with the “association then dissociation” equation in the FortéBio Data analysis software version 7.1, resulting in K . Conflicts of Interest S. Barbe and T. Schiex serve as scientific advisors and shareholders of the protein design startup Amineo. These roles have not influenced the design, analysis, or interpretation of the work presented here. All authors declare that there are no other competing interests. Supporting information Supporting File: advs76464‐sup‐0001‐SuppMat.pdf.