Efficient generation of epitope-targeted antibodies with Germinal

Abstract

Obtaining antibodies against specific protein targets is a widely important yet experimentally laborious process. Meanwhile, computational methods for antibody design have been limited by low success rates that require resource-intensive screening. Here, we introduce Germinal, a broadly enabling generative pipeline that designs antibodies against specific epitopes with nanomolar binding affinities while requiring only low-n experimental testing. Our method co-optimizes antibody structure and sequence by integrating a structure predictor with an antibody-specific protein language model to perform de novo design of functional complementarity-determining regions (CDRs) onto a user-specified structural framework. When tested against four diverse protein targets, Germinal designed functional antibodies across all targets and binder formats, testing only 43–101 designs for each antigen. Validated designs also exhibited robust expression in mammalian cells and high sequence and structural novelty. We provide open-source code and full computational and experimental protocols to facilitate wide adoption.

Introduction

Antibodies play a central role in adaptive immunity by binding a remarkable diversity of molecular epitopes on target antigens, often with high specificity [1, 2]. Together with well-characterized biochemical properties and favorable therapeutic profiles, these capabilities have made antibodies widely adopted as general-purpose and high specificity binders in biomedicine, biotechnology, and basic research [3, 4].

Antibody generation against arbitrary antigens traditionally requires animal immunization or large library screening campaigns [5, 6]. However, these experimental approaches are constrained by fundamental limitations. First, these processes are laborious and expensive, and do not always yield successful binders [7]. Second, once binders are identified, their molecular and structural characterization can be difficult [8, 9]. Finally, these methods offer no control over which specific region of the antigen the antibody recognizes, known as the epitope, making it difficult to direct binding to functionally important sites or specific target conformations.

Machine learning has enabled atomically accurate structure prediction of protein monomers and multimeric complexes with methods such as AlphaFold [10, 11]. Subsequent efforts leverage these structure predictors to design protein sequences with desired structural properties, including binding interactions. For example, methods such as RFdiffusion have fine-tuned structure prediction models with diffusion objectives to generate backbone coordinates of potential binders [12, 13]. Other approaches have used gradient-based methods that invert structure prediction to generate proteins and protein binders [14, 15, 16, 17]. Recently, BindCraft inverted AF-M to achieve high experimental success rates for de novo miniproteins against arbitrary targets [18, 11].

Despite numerous preliminary efforts [19, 20, 21], robust de novo antibody design against specific epitopes with high success rates remains challenging, due to the hyper-variability of complementarity determining regions (CDRs) and the highly constrained nature of antibody sequence space. For example, prior efforts to design novel CDRs screened thousands of designs to yield a handful of binders, often with activity in the micromolar range [19]. Such resource-intensive screening campaigns are inaccessible to most laboratories. Generative modeling that reduces experimental validation to tens rather than thousands of designs would democratize access across molecular biology.

Here, we present Germinal, a generative pipeline for de novo antibody design (Figure 1A). Germinal achieves epitope-targeted, de novo CDR design with success rates that enable low-n experimental testing. Germinal biases designs toward a specified antibody framework while freely designing the CDRs, preserving robust expression in human cells and the therapeutic developability profiles of known framework regions (FRs). We combine backpropagation of AlphaFold-Multimer (AF-M; [11]) with an antibody-specific protein language model (IgLM; [22]) to produce designs with realistic CDR sequences (Figure 1B). Custom loss functions ensure correct spatial positioning to favor CDR over framework binding contacts. Lead antibody designs were identified using a split-luciferase assay or a preliminary surface plasmon resonance (SPR; Methods) screen, then validated by bio-layer interferometry (BLI) for robust identification of high-affinity binders.

Figure 1. Efficient antibody generation via joint optimization of AlphaFold-Multimer and IgLM with Germinal.

(A) Overview of the Germinal pipeline. A target and antibody framework are provided as structural templates and sequence inputs, while CDRs are freely designed. Gradients from AlphaFold-Multimer (AF-M) and an antibody-specific language model (IgLM) are combined during the three design phases. (B) Structure and sequence joint optimization landscape. AF-M (structure/complex compatibility) and IgLM (antibody sequence prior) objectives define distinct optima. Their joint objective steers optimization toward sequences that are both structurally confident and antibody-like. (C) Representative predicted complexes across four targets. Designed nanobodies or scFvs bound to specified epitopes on PD-L1, IL3, IL20, and BHRF1 are shown in pink, with CDRs highlighted in magenta. Domains used for design are blue; white domains were excluded from the design target structure. Best dissociation constants $\left(K_{\text{D}}\right)$ measured by BLI are annotated. Structures are AF3-predicted. Asterisks (*) denote binders derived from an initial Germinal design. (D) Screening results by target for designed nanobodies (left) and scFvs (right). Stacked bars summarize per-library results (43 to 101 designs per target). Designs are grouped by expression and binding profile: low expression (gray), expressed but non-binding in split-luciferase assay (blue), split-luciferase binders (pink), and BLI-verified binders (magenta). scFvs were screened directly in fragment antibody (Fab) format by SPR/BLI, so luciferase-related categories are omitted. “Max. Id.” denotes the highest pairwise CDR sequence identity of any design in the library to the PDB. For scFvs, Max. Id. was calculated independently for variable heavy (VH) and variable light (VL) chains and averaged. AF-M: AlphaFold-Multimer; IgLM: antibody-specific language model; CDR: complementarity-determining region; BLI: bio-layer interferometry; SPR: surface plasmon resonance; $K_{\text{D}}$: dissociation constant.

Across four distinct antigens (Figure 1C), we identified validated binders by screening only tens of de novo nanobody or single-chain variable fragment (scFv) designs per target (Figure 1D). BLI showed that validated designs of both formats achieved nanomolar-to-low-micromolar dissociation constants $\left(K_{\text{D}}\right)$ for all antigens tested. Cryo-EM and alanine mutagenesis at computationally specified hotspots confirmed that most designs engage their intended epitopes with atomic-level accuracy. All designs exhibited low polyreactivity, comparable to nanobody and antibody controls with previously determined specificity.

To promote broad accessibility, we make our full computational pipeline open-source and freely available to the scientific community, alongside our experimental protocol for initial binder screening. We anticipate that Germinal will greatly reduce the need for the extensive experimental infrastructure developed specifically for antibody discovery. By supporting precise epitope targeting while maintaining favorable antibody properties, Germinal opens new possibilities for the development of antibody molecules for molecular biology and therapeutic design.

Results

Design of antibodies via dual-objective structure and sequence optimization

Computational design of de novo antibodies remains challenging due to their specific structural constraints, the high variability of their binding regions (CDRs), and a lack of antibody training data for structure prediction tools [23, 24]. In natural antibodies, the conserved $\beta$-sheet framework regions serve as a rigid scaffold that positions flexible CDR loops for optimal antigen recognition [25]. Although secondary structures do occur in CDRs, flexible loops (where flexible refers to the paucity of secondary structures rather than conformational change upon target binding [26, 27, 28]) are more prevalent in natural antibodies [27, 29]. Binder design methods built on these structure predictors, however, are biased toward protein–protein interfaces dominated by secondary structures rather than the more flexible loop conformations characteristic of CDRs [12, 18, 30].

Antibody-specific language models offer a complementary source of information. Trained on vast repositories of antibody sequences [31], these models learn the distribution of natural antibody sequences and capture implicit structural information from sequence data alone [22, 32, 33, 34]. Consistent with this, recent work has shown that general structure-based models alone struggle to redesign functional CDRs, but perform substantially better when complemented by antibody-specific language models [35]. We therefore hypothesized that integrating structure prediction with antibody-specific language models could overcome the twin challenges of data scarcity and structural bias, unlocking efficient de novo antibody design. Germinal realizes this integration by combining a protein structure predictor, AF-M [11], with an antibody-specific protein language model, IgLM [22] (Figure 1A), merging their gradients to co-optimize both fold geometry and antibody naturalness in a joint optimization landscape (Figure 1B). Germinal’s generation pipeline consists of three main stages: design, sequence optimization, and filtering.

In the design stage, Germinal uses the gradients of AF-M to sample sequences that, with high confidence, bind desired epitopes in the predicted structures. Germinal also uses the gradients of IgLM to bias sampling toward sequences with high probability under the language model, reflecting their likelihood of resembling naturally occurring antibodies. Given AF-M’s direct conditioning on input sequences and the intrinsic coupling of sequence and structure, the antibody language model also implicitly influences the resulting structure. During this stage, Germinal also allows users to explicitly bias samples toward a desired antibody framework sequence (e.g., a framework with a favorable developability profile; Figure S1), as well as augmenting structure guidance by providing the framework’s structure as a template to AF-M (Methods).

Sampled sequences with high-quality predicted structures are then carried forward to a sequence optimization stage. Similar to other hallucination methods, we use AbMPNN [36, 37]), a structure-conditioned sequence design model fine-tuned on antibodies, to redesign CDR residues that are not in direct contact with the antigen. This serves two purposes: improving binder stability by optimizing non-interface CDR residues while preserving key paratope–epitope contacts, and generating multiple sequence variants per design structure, effectively expanding the diversity of the candidate pool without requiring additional computational trajectories (Methods; Table S1).

Finally, designs are passed through a filtering stage (Tables S2–S3), where we use a different structure predictor from the one used during the design stage to provide a separate assessment of design quality; in particular, we used AlphaFold 3 (AF3) due to its superior accuracy on antibody–antigen complexes [38, 24]. We filter and rank (Tables S4–S5) designs based on AF3 confidence scores alongside PyRosetta-derived biophysical and biochemical scores [39], yielding a final set of candidates for experimental testing (Methods).

We initially found that naive application of AF-M guidance towards antibody binder design resulted in generations with paratopes (the region of the antibody that contacts the target) composed of framework residues or enriched with secondary-structures. To address this, Germinal incorporates three custom loss functions during the design stage (Figure 2A) that guide generations toward natural antibody binding conformations in contrast to those favored by naive AF-M. A paratope-specific loss ensures that binding occurs primarily through the designed CDRs (Figure 2B). Additionally, Germinal incorporates two secondary-structure losses, an $\alpha$-helix loss and $\beta$-strand loss. These losses encourage the generated CDRs to adopt loop conformations, instead of the secondary-structure–rich conformations from vanilla AF-M guidance (visualized in Figure 2C,D). Together, these losses guide designs toward binding interfaces dominated by CDRs [40, 41, 42] while promoting their conformational flexibility, a feature that has been linked to enhanced antigen specificity and the capacity to generate broadly neutralizing binders in natural antibodies [43, 44].

Figure 2. Germinal’s design stage steers generations toward natural antibody-like properties through structural and sequence guidance.

(A) Schematic of Germinal’s design stage. Structure and confidence losses are integrated with IgLM sequence guidance through gradient merging, and the resulting gradient is used to update the sequence PSSM. A straight-through estimator converts the PSSM into a one-hot encoded sequence for input to IgLM, whereas AF-M operates directly on the PSSM. (B) Paratope loss ensures binding occurs primarily through CDR regions rather than framework residues. (C-D) Secondary structure losses prevent CDRs from being dominated by $\alpha$-helices and $\beta$-strands. Ternary kernel density plots display the secondary structure composition of the interface for the designs. Selected points represent the median interface fraction for designs generated at each weight. (E) Pareto frontier between AF-M structural confidence, shown as iPAE, and IgLM sequence likelihood reveals competing optimization objectives. (F) Comparison of developability properties from the Therapeutic Nanobody Profiler (TNP) between sequences generated with (weight of 1.0) and without (weight of 0) IgLM guidance. Designs within the amber region are defined as low risk, between amber and red, medium risk, and above red, high risk. (G) Comparison of the OASis humanness score amongst sequences generated with varying weighting of IgLM guidance. Shown in gray and pink are the median humanness scores of all nanobodies, and all camelid nanobodies, respectively, in the SAbDab-nano database.

While combining AF-M structure guidance with IgLM sequence guidance is central to our approach, we found that the preference an antibody language model has for a given sequence (IgLM log-likelihood) and the predicted binding confidence (AF-M interface predicted aligned error; iPAE) are competing objectives. This behavior is expected given that antibody-antigen complexes have been challenging to model [27, 24], as well as the fact that backpropagating through AF-M can lead to adversarial designs with unrealistic poses that improve structure confidence. This trade-off yields a Pareto frontier, underscoring the need for joint optimization (Figure 2E; S2). As the strength of IgLM guidance increased, we observed improvements in the therapeutic developability and safety profiles for generated sequences as predicted by the Therapeutic Nanobody Profiler (TNP) [45] (Figure 2F; S3) and Therapeutic Antibody Profiler (TAP) [46] (Figure S4). Additionally, IgLM guidance improved the ”humanness” and reduced the immunogenicity risk of generated sequences, as predicted by the OASis tool [47], which calculates the similarity of a given sequence against a large database of human antibodies [31] (Figure 2G). In total, Germinal enables controllable, multi-objective generation of antibody-like sequences.

Using Germinal to target diverse antigens

We next sought to apply Germinal to four diverse targets: Protein Death Ligand 1 (PD-L1), an immune checkpoint ligand expressed on both tumor and immune cells, which is both a clinically validated therapeutic target and a frequent target of de novo binder design efforts; interleukin-3 (IL3) and interleukin-20 (IL20), cytokines involved in immune signaling for which no de novo binders have been reported to date; and BHRF1, a viral Bcl-2–like anti-apoptotic protein from the Epstein–Barr virus, as a representative example of a non-human, pathogen-derived target.

After sampling, we observed that 1,584 nanobody trajectories for PD-L1, 1,379 trajectories for IL3, 739 trajectories for IL20, and 1,456 trajectories for BHRF1 had passed the pipeline filters with promising computational metrics (Figure S5). We then ranked these designs for experimental screening based on a comprehensive set of metrics (Table S4; Methods), ultimately selecting the top 101 nanobody designs for PD-L1, 46 designs for IL3, 43 designs for IL20, and 52 designs for BHRF1 for downstream validation. These designs exhibited low CDR sequence identity (<55%; median of ~30%; Figure S6) to any existing sequence in the PDB [48] or OAS [31]. Additionally, the tested designs demonstrated structural novelty when compared against experimentally resolved complexes in the PDB (Figure S7; Table S6; Methods), with low iAlign interface similarity scores (IS<0.47) indicating that the predicted interfaces do not resemble those of any known complexes. Furthermore, the designs selected for experimental testing exhibited substantial structural and sequence diversity within each library (Tables S7, S8).

We also used Germinal to design scFvs against two of these targets, PD-L1 and IL3. The resulting designs passed our filtering criteria at rates comparable to nanobody designs and achieved similarly favorable values across key evaluation metrics (Figure S5, S8–S9). Notably, several of the metrics identified as strong predictors of nanobody binding success were based on interface contact area (e.g., LIA, number of hydrogen bonds, and number of interface residues; Figure S10), and were thus naturally elevated in scFv designs due to the presence of two binding chains. Therefore, using the information available from the nanobody design rounds, we updated our filters for scFvs to calibrate for the interface area (Table S3) and sampled designs for PD-L1 and IL3. These runs yielded 470 PD-L1 and 324 IL3 designs that passed the pipeline filters, of which 48 designs each were selected for experimental validation following stringent filtering and ranking (Table S5). Moreover, consistent with our nanobody results, these scFv designs exhibited both CDR sequence and structural novelty. CDR sequence identities to any sequence in the PDB or OAS remained at ~40% (<52%; Figure S6), and binder–target interfaces differed substantially from known complexes in the PDB (IS ≤ 0.2; Figure S7; Table S6) as well as from other designs within the generated set (Tables S7, S8).

It is worth noting that all input antigen structures to the pipeline are AF3 predictions rather than experimentally determined structures. This demonstrates that Germinal can generate functional binders using predicted rather than experimentally resolved antigen structures, highlighting its potential applicability to targets for which such data are not available (Supplementary Discussion). However, for all four targets tested experimentally, the AF3 predictions used as input were themselves validated against known experimental structures $(\text{C}\alpha \text{RMSD}<1Å)$. As with other hallucination-based methods, Germinal assumes accurate antigen structure conditioning, and design quality will therefore depend on the fidelity of the input model.

Binding affinity characterization of generated antibodies

To experimentally validate nanobody designs produced by Germinal, we employed a split-luciferase assay based on the NanoBiT system (Figure S11) [49] to enable efficient scaling and fast iteration across initial design–validation rounds. This assay served as a smaller-scale, simpler alternative to binding-identification assays such as display-based approaches used by other groups [21, 50]. This system is built on the large luciferase subunit (LgBiT), which we fused to our binders, and the small luciferase subunit (SmBiT) that we fused to the antigen. LgBiT can be complemented by either a high-affinity peptide (HiBiT, $K_{\text{D}}=0.7\text{nM}$) or the very low-affinity peptide (SmBiT, $K_{\text{D}}=190\mu \text{M}$). Addition of HiBiT enables direct quantification of LgBiT-tagged binder expression, while addition of an antigen–SmBiT fusion protein allows detection of potential binders. For all four targets, designed nanobodies were fused to LgBiT and encoded in plasmids that were transiently transfected into HEK293 cells. Alongside our designs, we included a validated nanobody against the same antigen (when available) as a positive control, and a nanobody targeting an unrelated antigen as a negative control (Methods). To identify promising candidates, binding measurements (SmBiT–LgBiT) were normalized to the corresponding expression measurements (HiBiT–LgBiT), providing a relative measure of binding potential (Methods).

Candidates meeting both expression and binding thresholds were advanced to BLI. As a result, we tested 25 of 101 PD-L1 nanobodies, 11 of 46 IL3 nanobodies, 11 of 43 IL20 nanobodies, and 20 of 52 BHRF1 nanobodies for binding to their respective antigens. We observed detectable binding for 7 of 25 PD-L1 designs, 2 of 11 IL3 designs, 4 of 11 IL20 designs, and 5 of 20 BHRF1 designs (Figure 3A–D; S12). We obtained nanomolar binders for all four antigens tested.

Figure 3. Biophysical characterization of Germinal-designed binders.

(A–D) For each target, a Germinal-predicted structure of an experimentally validated nanobody (pink cartoon) in complex with the antigen (blue surface) is shown alongside the corresponding BLI sensorgram: PD-L1 (A, E11, $K_{\text{D}}=170\text{nM}$); IL3 (B, D2_Ser, $K_{\text{D}}=280\text{nM}$); IL20 (C, H5, $K_{\text{D}}=190\text{nM}$); BHRF1 (D, A5, $K_{\text{D}}=1.2\mu \text{M}$). Heat maps of BLI-derived $K_{\text{D}}$ values (M) for all nanobody hits for each target are also shown. (E, F) Germinal-predicted structures of designed scFv hits against PD-L1 (E) and IL3 (F), shown from two orientations (~180° apart), with heat maps of SPR- and BLI-derived $K_{\text{D}}$ values and representative SPR and BLI sensorgrams for the highest-affinity hit in each panel: PD-L1 scFv H5 and IL3 scFv F4. For all sensorgrams, binding curves are colored by analyte concentration, kinetic fits are shown as black dashed lines, and the derived $K_{\text{D}}$ is shown in bold. All BLI binding were fit with a global 1:1 model unless otherwise specified. (G–H) A polyspecificity particle assay was performed to measure the binding of Germinal-designed nanobodies (G) or scFvs (H) to a polyspecificity reagent (PSR). PSR binding scores are normalized to the polyspecific controls F02’ for nanobodies and Ixekizumab for scFvs. Each design was tested in triplicate (n=3 technical replicates). Dashed vertical lines delineate different targets. n.d., not determined.

Given the success rate of the initial nanobody designs and the consistency of in silico predictions between Germinal-designed nanobodies and scFvs, 48 anti-PD-L1 and 48 anti-IL3 scFv designs were reformatted as fragment antibodies (Fabs) and subjected to a preliminary surface plasmon resonance (SPR) affinity screen to identify hits for subsequent biophysical characterization (Figure S13; Methods). Fabs were chosen as the screening format due to their closer resemblance to therapeutic IgGs, their clinical relevance, and more predictable retention of binding affinity upon IgG conversion [51]. From this screen, we identified 4 potential PD-L1 binders and 1 IL3 binder.

These hits were subsequently characterized by BLI, both to confirm binding and to establish a baseline for downstream characterization. Of the SPR-validated designs, 3 of 4 anti-PD-L1 designs and the 1 anti-IL3 design had measurable binding activity by BLI (Figures 3E–F; S14–S15) and advanced to downstream biochemical characterization, while the remaining anti-PD-L1 design produced inconclusive measurements when tested with BLI. We hypothesize that this is due to discrepancies between SPR and BLI that arise from differences in experimental setup, and thus proceeded only with designs validated by both SPR and BLI. More broadly, despite applying computational filters retrospectively identified from the nanobody screens (Figure S10), experimental success rates for scFvs remained comparable to those of the nanobody campaigns, suggesting that these filters may not generalize across antibody formats.

Rescue and optimization of low-expressing designs

After observing low expression levels for a subset of Germinal-generated designs, we explored targeted expression rescue strategies. For the anti-IL3 scFv F4, we evaluated two additional designs that ranked just below the selection cutoff. The first was a parent sequence representing the original Germinal design prior to AbMPNN redesign (F4_Parent; differing at 6 residues from the hit); the second was a sister sequence sharing the same Germinal seed but carrying a different sampled AbMPNN sequence (F4_Sister; differing at 2 residues). Both variants expressed at higher levels than the original F4 (Figure S16), while retaining comparable or improved binding affinities, and were thus carried forward for downstream characterization (Figure 3F).

For the anti-BHRF1 nanobody C4, which consistently yielded insufficient protein for biochemical characterization after purification, we mutated its framework to match the Legobody framework [52] (to yield C4_Lego; 94.53% sequence similarity), which is nearly identical to those used in synthetic nanobody libraries engineered for high thermal stability [53, 54]. To test whether the legobody framework mutations affect nanobody binding activity, we applied the same re-scaffolding to the BHRF1 nanobody D3 (D3_Lego), a design that expressed robustly and showed confirmed binding. D3_Lego maintained high expression yields and showed improved binding affinity relative to D3, establishing that transfer to the Legobody framework is well-tolerated. Indeed, C4_Lego showed improved expression yields in Expi293F cells compared to the original C4 and was found to bind BHRF1 with 42 nM affinity (Figure S17F; Figure 3D). Therefore, we proceeded with C4_Lego for all downstream experiments.

Separately, the anti-IL3 nanobody D2 contained a solvent-exposed cysteine in CDR3 that likely promoted disulfide-mediated dimerization, as observed by non-reducing gel electrophoresis (Figures S17F–G). We hypothesized that this dimerization contributed to D2’s low expression yield and poor-quality BLI measurements (Figure S12C). Mutating this residue to serine (D2_Ser) restored monomeric expression, with improved purity by gel electrophoresis and well-resolved BLI binding curves (Figures 3B; S12B). D2_Ser was therefore used in all subsequent experiments.

Polyreactivity

We evaluated non-specific binding of our designs using a flow cytometry-based polyspecificity particle assay, in which His- and FLAG-tagged binders are captured on anti-His magnetic beads and assessed for binding to biotinylated solubilized Expi293F membrane proteins (polyspecificity reagent or PSR) [55, 56]. Detection of biotin with streptavidin and FLAG-tag with an anti-FLAG M2 antibody enabled distinction of polyreactive binders from target-specific species, as verified with previously validated positive controls (F02’ for nanobodies and Ixekizumab for scFvs) and non-polyreactive negative controls (A02’ for nanobodies and Elotuzumab for scFvs) (Figure S18) [57, 55]. We found that all Germinal-generated designs exhibited low polyreactivity (<4% of the polyreactive positive control), comparable to the negative control samples (Figure 3G–H; S19–S20). This favorable profile was retained in the derivative sequences C4_Lego and F4_Sister. In contrast, D2_Ser showed markedly elevated polyreactivity relative to the original D2 design (Figure 3G; Figure S19), suggesting that although mutation of a CDR cysteine to serine can rescue monomeric expression and target binding, it may also introduce non-specific binding interactions. This result underscores the importance of comprehensive developability assessment beyond binding affinity alone.

Epitope specificity of generated antibodies

To experimentally validate whether Germinal can indeed design epitope-specific antibodies, we determined the structure of a representative anti-PD-L1 scFv (H5) in complex with PD-L1 at 3.9 Å resolution via cryoEM. Superposition of the experimental structure with the Germinal-predicted model demonstrated close overall agreement, with a global $\text{C}\alpha$ RMSD of 1.25 Å (Figure 4A). The predicted model also showed a reasonable local fit to the experimental cryo-EM density across all six CDR loops (Figure 4B,C), indicating that the predicted backbone geometry is consistent with the experimental map. Comparison at the binding interface confirmed that contacts to the targeted hotspot residues (I37, Y39, E41) were maintained as designed (Figure 4D), demonstrating that Germinal can successfully produce designs that bind at the intended epitope.

Figure 4. Cryo-EM and mutational mapping of epitopes targeted by Germinal-designed binders.

(A) Cryo-EM structure of the Germinal-designed anti-PD-L1 scFv H5 (white) overlaid with the AF3-predicted structure (pink), in complex with PD-L1 subdomain I (blue). (B) Per-CDR local comparison of experimental and predicted loop conformations for V_L (top) and V_H (bottom), with key interface residues shown as sticks and labeled. (C) Experimental cryo-EM density map of the H5 scFv:PD-L1 complex with fitted atomic models overlaid, colored by chain. (D) Close-up of the binding interface comparing designed (magenta) and experimental (white) residue conformations, with hotspot residues highlighted in green. (E) $K_{\text{D}}$ values (M) for three anti-PD-L1 scFv hits against PD-L1 alanine variants I37A and Y39A. (F) Germinal-predicted binding poses for representative binders across all four targets (pink ribbons on blue surfaces), with zoom panels showing binder contacts (≤3.5 Å) at alanine-substituted hotspot residues (green sticks, labeled). (G) $K_{\text{D}}$ matrices from alanine mutagenesis at hotspot residues for all four targets (PD-L1: I37A, Y39A, E41A; IL3: L26A, T103A, F104A; IL20: R59A, R63A, R99A; BHRF1: V85A, I89A, R92A). For BHRF1 A5 (denoted by an asterisk (*)), the corresponding hotspot mutations were E76A, H77A, and Y122A. Binding below the limit of detection was assigned $K_{\text{D}}>1.0\times {10}^{-4}\text{M}$. n.d., not determined. All residue numbers refer to the mature protein sequence after signal peptide cleavage (see Methods for details).

To further assess whether binders across all four targets engaged their intended epitopes, we measured the effects of alanine substitutions at predicted hotspot residues. For PD-L1, hotspot substitutions I37A and Y39A reduced binding affinity across all three designed scFvs (Figures 4E, S14), consistent with the direct contacts observed at these residues in the cryo-EM structure of H5. We performed hotspot mutagenesis across all targets and binder formats, substituting predicted hotspot residues to an alanine: I37, Y39, and E41 for PD-L1; L26, T103, and F104 for IL3; R59, R63, and R99 for IL20; and V85, I89, and R92 for BHRF1 (Figures 4F, G) (residue numbers refer to the mature protein sequence after signal peptide cleavage; Methods). Substitution of one or more hotspot residues reduced binding affinity by at least two-fold relative to wild-type antigen, with at least one substitution abrogating detectable binding entirely in 17 out of 26 designs (Figure 4E, G; Figures S14–S15, S21–S24). Binders with distinct binding modes were differentially affected by individual substitutions, with no single mutation abolishing binding across all designs (Figure 4F (left), G (left)). To further confirm that the observed affinity reductions reflected disruption of the binder–antigen interface rather than global destabilisation of antigen folding, monoclonal antibodies with proprietary epitopes (for IL3 and IL20) or binders with overlapping but distinct epitopes (for PD-L1 and BHRF1) were included as positive controls; in each case, control binders retained near wild-type affinity for at least one alanine variant (Figure 4G), confirming the structural integrity of the antigen variants. Together, these results demonstrate that Germinal-designed binders engage their computationally specified epitopes with the predicted contacts.

Discussion

We present an end-to-end pipeline for de novo antibody design that achieves nanomolar-to-low-micromolar binders with low-n experimental validation. A scalable split-luciferase assay screens and filters designs to identify lead nanobodies for BLI validation. This pipeline yielded nanobodies against four diverse soluble protein targets, including IL3, for which, to our knowledge, no antibody or non-natural binder has been reported in the literature or deposited in the PDB aside from its cognate receptor— a target unlikely to benefit from memorization of known binder conformations. We extended the methodology to de novo scFv antibody fragments, obtaining binders for PD-L1 and IL3. We make our complete methodology and code publicly available.

Our method differs from previous protein language model approaches that guide directed evolution of antibodies for improved binding affinity, stability, or other therapeutic properties [56, 58, 59]. These efforts assume at least weak initial function (e.g., affinity maturation of a weak binder), whereas Germinal designs CDRs de novo without any starting binder; protein language models could nonetheless enable affinity maturation of Germinal designs. Concurrent industry-led efforts [60, 61, 62] report double-digit success rates but have not publicly released code, weights, or detailed methods. Open-source pipelines [50, 63], by contrast, provide greater methodological transparency. mBER, like Germinal, leverages backpropagation-based hallucination with partial structural conditioning for nanobody design, while BoltzGen is a diffusion-based all-atom generative model for nanobodies and other binder modalities. Germinal differs by co-optimizing structural confidence and antibody-likeness while controlling binding pose through explicit paratope- and structure-based loss functions, and by coupling design to cryo-EM structural validation of a binder–antigen complex, epitope-directed alanine mutagenesis, and polyreactivity profiling. Bennett et al. separately demonstrated de novo antibody design with atomic-level accuracy using RFDiffusion, but validation required screening tens of thousands of candidates to identify tens of binders [19]. Our reanalysis of their reported AF3-filtered data (Supplementary Discussion) places their success rate at ~1.2–1.5%, several-fold below Germinal’s. Although we did not use ipSAE [64] for scoring, retrospective analysis showed all but one validated binder exceeded the suggested min_ipSAE > 0.5 threshold (majority above 0.6; Figures S25; Table S9–S10) [65]; however, many non-binders also exceeded these thresholds, highlighting the difficulty of identifying computational scores that reliably predict binding across formats.

Beyond the initial design pipeline, we note that computational de novo antibody design benefits from rational optimization strategies rooted in traditional protein engineering and biochemistry. We employed three such approaches: directed framework mutations (Legobody re-scaffolding) to rescue low-expressing candidates (BHRF1 C4_Lego), cysteine substitution to recover monomeric expression (IL3 D2_Ser), and recovery of sister and parent variants from design trajectories (IL3 scFv F4). These optional, modular steps expand the pool of functional binders or improve initial hits.

Several aspects of the Germinal model and sampling pipeline also present clear opportunities for future development. First, performance is constrained by AF-M’s confidence and accuracy. Additionally, structure prediction and backpropagation at each iteration require extensive inference-time sampling, making the pipeline computationally intensive (see Computational resources in Methods). Improvements in optimization algorithms and integration with other generative approaches could reduce this burden. Memory constraints also limit the method to smaller protein targets, though most epitopes can be successfully targeted by truncating large proteins to the domain of interest. To support wider adoption, future work could substitute restrictively-licensed components (i.e., IgLM, AF3) with open-source alternatives (e.g., AbLang, ProteniX) [32, 66].

Like other methods [50, 67, 61], Germinal is restricted to favorable protein epitopes, limiting success rates for hard-to-target surfaces (Supplementary Discussion). We note that our results demonstrate the ability to target specific epitopes rather than any arbitrary ones, and thus performance will vary depending on the biophysical properties of the selected surface. Similarly, because AF-M only models canonical amino acids, important target classes such as glycans, small molecules, and non-protein antigens remain beyond its scope. Although Germinal can produce high-confidence designs against predicted antigen structures, our experimental validation used targets with existing experimental structures. For antigens like MCF2, which lacks an experimental structure (Supplementary Discussion), systematic evaluation of how design success depends on antigen model quality remains an open question. Finally, although Germinal’s loss functions steer designs toward CDR loop-dominated poses, the optimization yields a distribution of interface conformations, with a subset engaging the target through secondary-structure-rich contacts. Some of these designs exhibited stronger binding affinities but higher polyspecificity signal (Figure 3G–H). Although small sample sizes preclude firm conclusions, this warrants caution when prioritizing designs with secondary-structure-rich interfaces, particularly those involving partial framework regions.

Computational design of epitope-targeted antibodies with high success rates opens transformative possibilities across biotechnology and medicine. Germinal could unlock targeting of previously inaccessible epitopes, including intracellular domains, conformational states, and conserved regions that fail to elicit natural immune responses. Rapid antibody design could compress therapeutic antibody identification from weeks to days, improving response to novel pathogens or evolving disease targets. As our generative pipeline becomes become more efficient, systematic campaigns could produce antibodies against entire proteomes, expanding our repertoire of affinity reagents and therapeutic molecules.

Methods

Germinal algorithm

Germinal performs gradient-based optimization on the joint landscape defined by AF-M and IgLM objectives, merging their gradients into a single update direction. Following previous work [18], design proceeds through three phases (logit, softmax, and semi-greedy optimization) that progressively anneal continuous logits into one-hot sequence representations (Figure 2A). Non-interface CDR residues are then redesigned during sequence optimization using an antibody-finetuned sequence design model. Finally, designs undergo filtering and ranking to select top candidates for experimental testing.

Input preparation

To run the design pipeline, we first defined the antigen structure and the target epitope. To reduce computational cost, multi-domain antigens were restricted, whenever possible, to the independently folding domain containing the target epitope. For PD-L1, we used the V-like domain, while for IL3, IL20, and BHRF1, we used the full mature protein sequence excluding signal peptides. Antigens were predicted using AF3 [38] and validated to fold with high confidence and low RMSD when aligned to the experimental structure (RMSD < 1 Å), if available. This procedure was also applied to new antibody frameworks.

Epitope residues on the target were specified in an input settings file, as were CDR lengths for the antibody framework (identified via AbNumber (version 0.4.2) using IMGT numbering) [68, 69]. These lengths, together with user-defined starting points, were used to calculate CDR positions. For antigens with known binders (e.g., natural receptors), the epitope corresponding to the reference biologically relevant interaction was designated as the preferred epitope and the one to be consistently tested first. This choice was motivated by: 1) targeting natural interaction interfaces is often of direct therapeutic interest, and 2) in preliminary unconstrained runs, these regions emerged most frequently among successful designs. However, for some targets the known epitope was not the most productive and an alternative was used instead. See Supplementary Discussion for a retrospective analysis of epitope-dependent design success rates across targets.

Initialization

Before the design stage, the antibody framework and antigen structures are merged into a unified PDB containing two chains, which serves as the starting complex. The sequence position-specific scoring matrix (PSSM) is initialized by setting framework positions to their one-hot encodings and CDR positions to one of three different initializations depending on user choice: (i) uniform random noise, (ii) Gumbel noise centered around the template framework sequence, or (iii) zeros. This flexibility allows users to control the trade-off between broad sampling and guided starting points. Gumbel initialization was adopted as the default based on empirical performance. This initialization, together with the starting conditions and random seed, determines the diversity of the design trajectories.

During AF-M initialization, the antigen and antibody framework structures are also supplied as templates to AF-M. Importantly, only the framework regions (FRs) of the antibody structure are used as templates, with all CDR residues masked in the input feature matrices (e.g., atom positions, side chains, sequence). This ensures that AF-M provides structural guidance based only on stable framework features without biasing the optimization of CDR loops to the natural CDRs of the framework structure.

Sequence bias

To bias sequences toward a chosen framework, we use a bias weight to modulate between framework preservation and exploration. This term is applied to framework region (FR) residues only (i.e., all residues other than CDRs). Specifically, at every iteration, this bias is added to the logits corresponding to the original framework sequence before the current binder is used as input to AF-M and IgLM, essentially shifting the probability distribution to favor the original framework sequence. Unlike hard-fixing, this approach still permits mutations in FR positions if gradients strongly favor them. Based on parameter sweeps (Figure S26), we selected a value of 10, which minimized framework mutations while allowing enough flexibility for confidence optimization, and hence successful trajectories.

Design stage phases

• Logits: In this phase, the sequence is represented as continuous logits for each position and amino acid type. Gradient-based optimization is performed directly on these logits using the combined gradients from AF-M (structural confidence) and IgLM (sequence naturalness). The logits allow for smooth optimization in continuous space while maintaining differentiability.

• Softmax: Here, the sequence logits of the binder are normalized to sequence probabilities using the softmax function. Gradients are similarly merged and applied to update this representation of the sequence.

• Semi-greedy optimization: In this final phase, the softmax probabilities from the previous stage are used to sample mutations, and mutations with the best loss are fixed.

Optimization strategy

Gradients from AF-M and IgLM are merged at each iteration of the logit and softmax phases using either a weighted sum, Multiple Gradient Descent Algorithm (MGDA) [70] or a weighted version of PCGrad [71] (see Gradient merging). In the semi-greedy phase, optimization shifts from continuous gradients to discrete sequence updates. At each of the 10 semi-greedy iterations, five candidate sequences are sampled, and the one that best balances AF-M structural confidence with IgLM sequence likelihood—calculated as a weighted sum of the AF-M losses and IgLM log-likelihood—is accepted.

For all three stages, the IgLM weight is adjusted according to a four-step schedule. During the logits phase, the IgLM weight is linearly annealed from $v_1$ to $v_2$. In the softmax phase, the weight is set to $v_3$. Finally, in the semi-greedy phase, the weight is set to $v_4$. This annealing schedule empirically provided more robust performance than fixed IgLM weights, and is the default setting in Germinal.

Additionally, because Germinal does not include a discrete one-hot optimization phase, we check for sequence convergence at the end of the softmax phase. Specifically, for each CDR position we: (i) compute the maximum probability across all 20 amino acids and (ii) average these values across all CDR positions. The resulting value represents the average positional certainty in the most likely amino acid per position. To ensure that the model is confident enough in the sequence to justify discretization, only sequences exceeding an empirically determined threshold (0.1) for this metric are advanced to the semi-greedy stage.

At the end of each phase, the current best design is evaluated against AF-M confidence metrics (pLDDT > 0.8, ipTM > 0.68 and iPAE < 0.35) and designs advance to the next phase only if they satisfy these thresholds. At the end of the design stage, remaining sequences are co-folded with AF3 for independent evaluation prior to Sequence optimization. Designs must satisfy strict AF3 confidence thresholds (ipTM > 0.75, PAE < 8, pTM > 0.85, and pLDDT > 0.85), exhibit a CDR interface percentage greater than 60%, and be predicted to bind the specified epitope to proceed.

Importantly, the final AF-M loss used in Germinal is computed as a weighted sum of multiple structural and interface metrics. A full list of the loss terms and their weightings, as well as biases added during optimization are shown below (see Loss and bias terms). Additionally, during each iteration, AF-M is run using three recycles, with the final recycle taken as the prediction for each iteration.

Gradient merging

Central to Germinal is the gradient merging algorithm, which reconciles updates from AF-M and IgLM objectives. Before merging is applied, gradients from both models are normalized to their respective ${\ell }_2$-norms. This ensures optimization of the two objectives is not biased due to magnitude differences between the two gradients. The resulting gradient is then re-normalized to have unit ${\ell }_2$-norm again and scaled by the effective sequence length (the number of non-zero positions in the AF-M gradient) as originally done by ColabDesign. Three strategies were explored used for gradient merging:

• Weighted Sum: gradients are merged by a direct weighted sum where the AF-M weight is fixed to 1, and the IgLM weight $\lambda$ is set by a user-defined parameter.

  $$\nabla \text{Germinal}=\nabla \text{AF-M}+\lambda ·\nabla \text{IgLM}$$

• Weighted PCGrad: As described in [71], PCGrad mitigates destructive interference between gradients by resolving conflicts in direction. Specifically, one gradient is projected onto the orthogonal complement of the other if the dot product is negative, thereby preventing cancellation of progress along critical dimensions. In our implementation, the previously normalized vectors are scaled by a user-defined weight before applying PCGrad.

• Multiple Gradient Descent Algorithm: MGDA [70] formulates gradient merging as a multi-objective optimization problem, where the goal is to compute a Pareto-optimal descent direction. Concretely, MGDA solves for the convex combination of AF-M and IgLM gradients that minimizes the squared norm of the merged update:

  $$\nabla \text{Germinal}=\alpha ·\nabla \text{AFM}+(1-a)·\nabla \text{IgLM}$$

  with $\alpha \in [0,1]$ chosen to balance objectives without privileging one a priori. This is achieved by solving a small quadratic program at each step. In practice, MGDA is more computationally intensive than PCGrad or weighted sums, but it provides a principled way to balance objectives when their contributions vary dynamically during optimization.

Loss and bias terms

Below, we list all biases, weights, and custom loss terms utilized in Germinal.

We use the following biases:

• Sequence bias: Higher values enforce greater similarity to the framework sequence.

• IgLM scale: Scaling factor applied to normalized IgLM gradients relative to AF-M gradients.

We use the following ColabDesign loss functions (weights):

• pLDDT loss: Emphasizes high per-residue confidence in AF-M predictions.

• ipLDDT loss: Emphasizes high per-residue confidence at the predicted interface.

• iPAE loss: Penalizes high alignment error between predicted and target structures at the interface.

• PAE loss: Penalizes high alignment error of the predicted complex.

• ipTM loss: Favors predictions with high predicted TM-score, encouraging correct geometry at the interface.

• pTM loss: Favors predictions with high predicted TM-score, encouraging correct overall geometry.

• Inter-protein contacts: Ensures that the designed antibody has enough contacts to the antigen chain within a certain distance. The number of contacts and distance can be specified by the user. If a hotspot is provided, the loss favors contacts with the specified positions.

• Intra-protein contacts: Encourages each residue in the designed sequence to form a sufficient number of contacts with other residues within the same protein. The user can specify both the minimum number of contacts required and the distance threshold used to define a contact.

We use the following custom loss functions:

• $\alpha$-helical loss: This loss is a variant of the helix loss from [18], restricted to CDR positions. We discourage $\alpha$-helical geometry within the CDRs by penalizing $\{i,i+3\}$ pairs whose predicted distances fall in the 2–6.2 Å range characteristic of $\alpha$-helices. Using AlphaFold’s distogram, we compute a binary cross-entropy penalty per residue pair that is large when probability mass lies inside this window and small otherwise. We also apply a positional mask to restrict evaluation to CDR residues, averaging the resulting penalties. Minimizing this loss reduces helix-like contacts in the CDRs and biases designs toward flexible, loop-like conformations.

• $\beta$-strand loss: Analogous to the helix term, this loss suppresses $\beta$-strand–like geometry within the selected CDR positions by penalizing $\{i,i+3\}$ pairs whose predicted distances fall in a $\beta$-like window (9.75–11.5 Å). Similar to the helix loss, we compute a binary cross-entropy penalty per pair with respect to this window using the distogram. To restrict this loss to CDR residues, we apply a positional mask that includes all CDR including the position directly before and directly after each CDR. We sort the probability of each pair constituting a $\beta$-strand and the top three most likely pairs were averaged as the final $\beta$-strand loss. Minimizing this loss reduces $\beta$-strands in the targeted regions, complementing the $\alpha$-helix loss to favor loop-like conformations.

• Paratope loss: We developed a paratope loss to achieve two goals: (1) encourage the CDR residue to bind to the specified epitope and (2) discourage the framework binding to the target. For the first goal, we define a loss ${ℒ}_{\text{CDR}}$, which operates on two sets of residues: target_id, which includes all specified hotspot residues on the target (or all residues in the target if no hotspot is specified), and pos_id, which include all CDR residues. For each pair of residues between target_id and pos_id, we calculate a categorical cross-entropy (CCE) loss using the distogram distance probabilities and a user-defined distance threshold. We then average the top $k$ closest predicted contacts for each epitope residue, across all residues in target_id to yield a final scalar loss. For the second goal, we calculate the same loss with different sets of residues. Here, we set target_id to all target residues and define fw_id as the set of framework residues on the antibody. Following the same loss calculation yields ${ℒ}_{\text{framework}}$ (Figure 2B). We then define the final paratope loss as:

  $${ℒ}_{\text{paratope}}={ℒ}_{\text{CDR}}*\frac{{ℒ}_{\text{CDR}}}{{ℒ}_{\text{framework}}-\lambda }$$

  where $\lambda$ is a user-defined parameter and is referred to as an “offset” value in the codebase. Note that both ${ℒ}_{\text{CDR}}$ and ${ℒ}_{\text{framework}}$ will decrease when the corresponding region gains more contacts with its target. Thus, the combined term ${ℒ}_{\text{paratope}}$, which is proportional to ${ℒ}_{\text{CDR}}$ and inversely proportional to ${ℒ}_{\text{framework}}$, increases when framework contacts are abundant and decreases when CDR residues are predicted to bind to the specified epitope, steering contacts toward the intended CDR pose via minimization of the loss.

Sequence optimization

Design candidates that cleared initial filters are passed through AbMPNN [36, 37], a variant of ProteinMPNN fine-tuned on OAS [31] and SAbDab [72]. AbMPNN was used to generate 40 sequence variants per design, of which the top 4 (by AbMPNN log-likelihood) were retained for re-folding with AF3 and final filtering. Optimization was restricted to CDR residues with all heavy atoms >3 Å from the antigen, preserving key paratope–epitope contacts while allowing AbMPNN to improve the stability of the surrounding loop scaffold. This is consistent with the improvement in binder_score, a PyRosetta energy term for which lower values indicate greater predicted stability, observed across AbMPNN-redesigned sequences relative to their parent designs (Table S1). Beyond stability, this step also serves as a practical means of diversifying the candidate pool, as the AbMPNN-redesigned sequences exhibit only modest changes in other key scores (Figure S27). Since each successful trajectory effectively seeds multiple distinct yet high-quality candidate sequences, it increases the yield of testable designs without additional sampling and can de-risk a given structural proposal, as different sequences may exhibit distinct developability properties that increase the likelihood that at least one will have a favorable profile. This process also implicitly acts as a self-consistency check, as designs from the same seed frequently scored highly and proved functional when tested experimentally. In future iterations of the pipeline, this self-consistency behavior could serve as an additional ranking criterion.

Finally, while the sequence optimization step may not be strictly required in all cases — for example, IL3 scFv F4 was experimentally successful both with and without AbMPNN optimization (Figure S15) — broader experimental validation will be needed to determine how generally it can be omitted.

IgLM Implementation

Slight modifications were made to the original implementation of IgLM [22] in order to incorporate the PSSM sequence representation during design. All inputs to IgLM are one-hot amino acid sequences; at each iteration, we convert the raw PSSM to a probability distribution using a softmax operation with a temperature of 0.6, then use a straight-through estimator to obtain discrete sequence representations. We condition each sequence (split into VH and VL for scFvs) with a chain token (<HEAVY> or <LIGHT>) and a species token (e.g., <HUMAN>) [73]. Loss is calculated by taking the mean negative log-likelihood over the entire sequence, which is then backpropagated to the raw PSSM to yield the update gradient for the iteration.

Design filtering

For filtering, designs were evaluated against a set of stringent thresholds derived from structural prediction confidence scores and PyRosetta-derived metrics [39]. Only sequences passing all thresholds advanced to the next stage. Seeking to utilize an independent structure prediction model from AF-M, and given its superior accuracy in modeling antibody-antigen complexes [24], we selected AlphaFold 3 (AF3) [38] as the final co-folding model. All AF3 predictions were performed without structural templates and only using MSAs for the target. The metrics and thresholds applied for filtering are detailed in Table S2.

Design ranking

After filtering, we proceed to an optional, yet recommended, ranking step, to reduce the number of designs to be experimentally validated. Ranking is conducted using a Borda count–based ranking algorithm, which integrates multiple structural and sequence-based metrics, each weighted by relative importance (Table S4).

Hard threshold filtering followed by Borda ranking ensures the top high-confidence, well-structured, and framework-preserving candidates are chosen. Retrospective analysis further highlighted the importance of several metrics (Figure S10).

Computational resources

All sampling was performed on NVIDIA H100 GPUs. Although exact runtimes vary heavily by antigen, typical sampling runs require 200–500 H100-GPU hours to yield 200–400 designs that pass pipeline filters.

Structural novelty analysis

To determine the structural novelty of designed binder–target interfaces, we first performed a broad structural search using Foldseek-Multimer (version 8.ef4e960) [74], which enables fast and sensitive comparison of large protein complexes. Specifically, we used easy-multimersearch to query every BLI- or SPR-verified hit against the PDB for each target:

foldseek easy-multimersearch path/to/hits pdb result path/to/tmpFolder

PDB structures deposited after the AlphaFold-Multimer training cutoff date were excluded. To quantify interface-level structural similarity against all Foldseek hits, we additionally computed Interface Similarity (IS) scores using iAlign v1.1 [75], which performs structure-based comparison of protein–protein interfaces. For each verified hit, iAlign was run against all multi-chain Foldseek matches (with the same post-cutoff exclusions applied).

ialign.pl <design>.pdb AB <match>.pdb <chains> -a 0 -mini 10

For each campaign, the match with the highest IS-score was selected and the design complex was aligned to the matched PDB structure by superposition of the target chain for visualization.

Sequence novelty analysis

To determine the sequence novelty of the designs, the MMseqs2 (version 15.6f452) toolkit [76] was used to assess the sequence similarity to all entries in the PDB [48] and Observable Antibody Space (OAS) [31]. The MMseqs2 command used is:

mmseqs easy-search <query_fasta> <target_db> <outfile> <tmp_dir> -s 3.0
--format-output “query,target,fident,qaln,taln,qstart,qend,tstart,tend,evalue”

The pairwise alignments are post-processed to compute CDR-specific identity by extracting the alignment positions corresponding to the IMGT-derived [69] CDR positions and computing the number of matching positions divided by total CDR positions.

Structural and sequence diversity analysis

To assess the structural diversity of designed binder–target interfaces within each campaign, we computed the Interface Similarity (IS) scores using iAlign [75] between all pairs of designs within each campaign. Specifically, for each campaign, every design complex was compared against every other on chains A and B (target and binder), and the pairwise IS-scores were recorded. This was done for all tested designs as well as verified hits. Similarly, to assess the sequence diversity of designed binders, we computed CDR sequence identity between all pairs of designs within each campaign, and between all verified hits within each campaign.

Therapeutic Nanobody Profiler analysis

To evaluate designs with the Therapeutic Nanobody Profiler (TNP), the Github repository found at github. com/oxpig/TNP was utilized. Since there is no Github to our knowledge for the Therapeutic Antibody Profiler (TAP), the web server was utilized (https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred/tap). The thresholds for green, amber, and red (shown as blue, purple and pink, respectively, in Figure S3–S4) risk levels are derived from the automatic assignment from the profiler. Humanness scores were calculated using the promb (version 1.0.2) package (https://github.com/MSDLLCpapers/promb) following methodology proposed in BioPhi.

Experimental characterization

Framework selection

To identify a robust scaffold for CDR grafting, we evaluated both nanobody and scFv frameworks for their expression and tolerance to CDR replacement. For nanobodies, we screened three frameworks: the humanized nanobody hNbBCII-FGLA, its camelid predecessor cAbBCII, and an alpaca nanobody that has been used in prior in-house experiments. These scaffolds were chosen due to their known stability and successful use as universal grafting frameworks [73]. For scFvs, we tested an anti-IL2 scFv from our collection (Fab-F5111) [77], an scFv version of trastuzumab [78], and a previously reported scFv framework (FW1.4) [79]. All nanobody scaffolds were grafted with PD-L1-nb [80] CDRs (PDB: 5JDS), and all scFvs were grafted with F5111 CDRs to assess the effect of framework on binding and expression activities. Among the nanobody grafts, only hNbBCII and cAbBCII expressed robustly and retained binding activities (Figure S1). We selected hNbBCII for downstream work based on its prior successful use [19]. Among the scFv grafts, only the trastuzumab scaffold yielded appreciable expression, but none of the scFv grafts showed a significant binding signal. We retained trastuzumab given its previous successful use in other computational antibody design pipelines [19].

Notably, CDR lengths were not modified for any design and were instead kept identical to those of the original framework. This decision was motivated by test experiments in which most designs with non-native CDR lengths failed to express, and by previous work showing that grafting CDR sequences onto a mismatched framework substantially reduces both thermostability and binding affinity [81, 82]. In nanobodies, the lengths of CDR1 and CDR2 are highly conserved across structures deposited in the PDB [29]. Whilst the length of CDR3 is variable, its length carries a structural consequence: whether or not CDR3 makes intramolecular contacts with the framework depends on both CDR3 length and framework identity [81, 82].

Plasmids and cloning

Constructs were cloned by PCR methods using either NEBNext High-Fidelity Polymerase (New England Biolabs; #M0544) followed by Gibson assembly of fragments (New England Biolabs; #E2621S), or Phusion Flash High-Fidelity Master Mix (Thermo Fisher; #F548L) followed by In-Fusion fragment assembly (Takara Bio). In all cases, constructs were assembled in a three-piece In-Fusion strategy, in which the mammalian expression vector backbone was split into two fragments by PCR with pUC-ori primers, and homology regions were designed to overlap with the synthetic DNA fragments. The coding regions of all synthetic DNA fragments were codon-optimized for human expression prior to synthesis. DNA fragments were designed such that only the mature protein-coding sequence (except for IL3, IL20 and PD-L1 was replaced in the plasmid, while all other plasmid elements, including signal peptides, tags, promoters, and terminators, were left unchanged.

DNA sequences of designed binders and targets (PD-L1, IL3, IL20, and BHRF1 (Figure 1C)) were ordered from Twist Biosciences with Gibson cloning adaptors for insertion into a pCMV mammalian expression vector (see Data S1). Target proteins were cloned with their native N-terminal signal peptide, with the exception of BHRF1, which was cloned with the IL2 signal peptide (MYRMQLLSCIALSLALVTN). All designed binders were cloned with the IL2 signal peptide. For split-luciferase assays, all target proteins were fused to a C-terminal SmBit peptide and His tag, while all designed binders were fused to a C-terminal LgBit protein. For BLI experiments, target proteins were cloned with a C-terminal AviTag (GLNDIFEAQKIEWHE) and His tag, except for IL3, which was cloned in both C-terminal and N-terminal AviTag formats due to the proximity of its termini to the target epitope; designed binders were fused to a C-terminal His tag only. For polyspecificity assays, nanobodies were cloned with an N-terminal FLAG tag (DYKDDDDK) and a C-terminal His tag, while Fabs were cloned with a C-terminal FLAG tag on the light chain and a C-terminal His tag on the heavy chain. For Cryo-EM experiments, target antigens were cloned with a C-terminal His tag only, and PD-L1 Fab H5 was cloned with a C-terminal His tag on the heavy chain only. Plasmids were propagated in Turbo Competent Cells (New England Biolabs; #C2984) for high-yield transformation and outgrowth, and purified using standard plasmid preparation protocols. All constructs were validated by whole-plasmid Nanopore sequencing. A full list of tested binder sequences is provided in Data S2.

Expi293 protein expression

All target proteins for split-luciferase assays and lead designed binders were expressed in Expi293F cells (Thermo Fisher; #A14635). For BLI experiments, target proteins possessing an AviTag were expressed in Expi293F in the presence of a BirA enzyme, enabling biotinylation of targets in vivo. Cells were maintained in a 2:1 mixture of FreeStyle 293 Expression Medium (Thermo Fisher; #12338018):Expi293 Expression Medium (Thermo Fisher; #A1435102) and grown at 37 °C with 8% (v/v) CO2 and shaking at 125 rpm. After ≥ 3 passages, cells were diluted to 2 × 106 cells/mL and grown for 20–24 h, when they were diluted to 3 × 106 cells/mL for transfection in a final volume of 30–50 mL (designed binders) or 100–250 mL (targets). Transfection mixtures were prepared with the following components per mL of expression culture: $0.5\mu \text{g}$ midi-prepped DNA, $1.3\mu \text{L}$ FectoPRO transfection reagent (Polyplus; #101000007), and $100\mu \text{L}$ expression media. After incubating for 10 min, the transfection cocktails were added to expression cultures. Cultures were immediately supplemented with glucose (0.45% (w/v) final concentration) and valproic acid (2 mM final concentration), and grown for 4–5 days post-transfection, when cell viability fell below 50%. Secreted protein was harvested from cells by centrifugation at 4000 × g at 4 °C, followed by filtration through a $0.22\mu \text{M}$ syringe filter (Thermo Fisher; 723–2520). 10× phosphate buffered saline (PBS; 1.37 M NaCl, 27 mM KCl, 100 mM Na2HPO4, 18 mM KH2PO4, pH 7.4) and 1.63 M NaCl were added to filtered supernatant each to a final dilution of 1:10 and, if necessary, supernatants were pH-adjusted to pH 7.6 using 1 M Tris-HCl pH 9.0.

Immobilized metal ion affinity chromatography purification of His-tagged proteins

All His-tagged proteins were purified using HisPur Cobalt resin (Thermo Scientific; #89965). Cobalt resin was washed twice in binding buffer (high-salt phosphate buffered saline [HS-PBS; 300 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.8 mM KH2PO4, pH 7.4], 10 mM imidazole) before adding to clarified supernatants and incubating for 1 h at 4 °C with end-over-end rotation. Purification mixtures were transferred to chromatography columns for gravity flow purification. Resin was first washed with binding buffer, followed by wash buffer (HS-PBS, 30 mM imidazole), before proteins were eluted with elution buffer (HS-PBS, 200 mM imidazole). Eluants were immediately buffer exchanged into PBS pH 7.4 and concentrated using centrifugal concentrators with 5 or 10 kDa molecular-weight cut-off (Millipore; #UFC901024), by four rounds of centrifugation (4,000–15,000 × g, 10 min, 4 °C) followed by resuspension in PBS pH 7.4. Protein concentrations were determined from A₂₈₀ using their extinction coefficients predicted from ExPASy ProtParam. Protein purity was assessed by SDS-PAGE (Figure S16–S17), flash-frozen in liquid nitrogen, and stored at −80 °C.

Size-exclusion chromatography purification of antigens

For antigens whose Co-NTA eluates contained visible contaminants by SDS-PAGE, preparations were further purified by size-exclusion chromatography on an AKTA pure FPLC system (Cytiva; using UNICORN^™ 7 software) equipped with a Superdex 200 Increase 10/300 GL column (Cytiva) pre-equilibrated in degassed PBS pH 7.4. A 0.5 mL sample was injected per run and antigen purity and monodispersity within peak fractions were confirmed by SDS-PAGE (Figure S28A). Fractions containing target species were pooled (Figure S28B), concentrated, flash-frozen in liquid nitrogen, and stored at −80 °C.

HEK293T protein expression

Transient transfections were performed on all of the binder plasmids using jetOPTIMUS DNA transfection Reagent (Avantor) in 24-well plates, with each plate incorporating internal controls. When available, validated binders (PD-L1 binder, KN035 [PDB: 5JDS]; BHRF1 binder, GDM_BHRF1_35 [67]) were included as positive controls. Binders not associated with a given target (a PD-L1 binder, KN035, for IL20 and BHRF1; a $\text{TNF}\alpha$ binder [PDB: 5M2J] for PD-L1 and IL3) were included as negative controls. This layout ensured that every plate contained both experimental and control conditions for consistent data normalization and comparison. Each binder plasmid was transfected in two biological replicates, with each replicate consisting of 450 ng of binder plasmid and 5 ng of CMV-driven mCherry plasmid (as co-transfection marker). A total of 990 ng binder plasmid DNA was added to $110\mu \text{L}$ of transfection buffer with co-transfection marker mixed. $0.88\mu \text{L}$ of transfection reagent was then added, well mixed, and incubated for 9 minutes. Following incubation, $50\mu \text{L}$ of the transfection complex was added dropwise to each well of Human Embryonic Kidney (HEK) 293 cells (ATCC CRL-1573) at 80–90% confluency. The cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM), supplemented with 10% fetal bovine serum (Fisher Scientific; #FB1299910), 1 mM sodium pyruvate (EMD Millipore; #TMS-005-C), 1× penicillin–streptomycin (Genesee catalog, #25–512), 2 mM l-glutamine (Genesee; #25–509) and 1× MEM non-essential amino acids (Genesee; #25–536), under standard conditions (37 °C, 5% CO₂). 6 to 12 hours after transfection, $300\mu \text{L}$ of DMEM was removed from each well and replaced with $200\mu \text{L}$ of fresh DMEM media to reduce reagent effects and concentrate final protein concentration. All split-luciferase assays were carried out at least 36 hours after changing the media.

SDS-polyacrylamide gel electrophoresis (SDS-PAGE)

Protein samples were resolved by SDS-PAGE on 4–20% SurePAGE precast gels (GenScript, #M00657) using an XCell SureLock Mini-cell electrophoresis apparatus (Thermo Fisher Scientific, #EI0001). Samples were prepared by addition of 4× Laemmli sample buffer (62.5 mM Tris-HCl pH 6.8, 2% (w/v) SDS, 10% (v/v) glycerol, 0.002% (w/v) bromophenol blue), followed by heat denaturation at 99 °C for 3 min. Separation was performed at 200 V in 1× MES-SDS Running Buffer (GenScript, #M00677). Molecular weight was estimated by co-electrophoresis with either SeeBlue Plus2 Pre-stained Protein Standard (Invitrogen, #LC5925) or PageRuler Plus Prestained Protein Ladder (Thermo Fisher Scientific, #26619). Gels were stained with InstantBlue Coomassie (Abcam, #ab119211) under gentle agitation for up to 1 hour, then destained in Milli-Q water for at least 1 hour prior to imaging on a GelDoc Go system (Bio-Rad). Band analysis was performed using ImageLab v6.1.0 (Bio-Rad). Uncropped versions of all gels presented in this manuscript are provided in (Figure S33).

HiBit-LgBit NanoLuc complementation reporter assay for binder expression

Binder expression was quantified by luminescence of the Hibit-LgBit complementation system (Nano-Glo HiBiT Lytic Detection System; Promega) measured with plate reader (Tecan Infinite^® M Plex, multimode microplate reader; #30190085 using the Tecan I-control version 3.9.1.0.). 200 μL of media were first collected from each well in 24-well plates and split equally between 2 wells in 96-well plates. During all experiments, media from the first biological replicate well were distributed to the first two technical replicates of split-luciferase assay, and media from the second biological replicate added to the third technical replicate. Each split-luciferase assay reaction was prepared in a total volume of $100\mu \text{L}$, consisting of $49\mu \text{L}$ Nano-Glo^® buffer, $44.9\mu \text{L}$ phosphate-buffered saline (PBS, pH 7.4), $1\mu \text{L}$ Nano-Glo^® substrate, $0.1\mu \text{L}$ HiBit (20 nM) and $5\mu \text{L}$ conditioned media. First luminescence reading was recorded immediately after addition of samples and second measurement was performed after 4 minutes of incubation. Longer incubation time leads to signal saturation, causing the plate reader to end the assay.

To validate that the luminescence signal provides a reliable and quantitative measurement, we leveraged a known PD-L1 binder, KN035 [80] (PDB: 5JDS, $K_{\text{D}}=3\text{nM}$), at varying concentrations. We measured luminescence directly from the supernatants after addition of HiBiT (expression) and confirmed a strong linear relationship between luminescence and KN035 concentration upon addition of 20 nM of HiBiT to quantify expression (Figure S11B top). Encouragingly, measured expression levels from the HiBiT assay (Figure S11C left) correlated well with protein yields obtained from independent recombinant expression of nanobodies in Expi293F cells (Figure S29).

SmBit-LgBit NanoLuc complementation reporter assay

Binder-target interaction and specificity was quantified by luminescence of the SmBiT–LgBiT complementation system (Nano-Glo Live Cell Assay System, Promega) on the plate reader (Tecan Infinite^® M Plex, multimode microplate reader; #30190085). Media from the same well used in the Hibit-LgBit assay were used to maintain consistency. Each reaction was prepared in a total volume of $100\mu \text{L}$, consisting of $49\mu \text{L}$ Nano-Glo^® buffer, $39\mu \text{L}$ phosphate-buffered saline (PBS, pH 7.4), $1\mu \text{L}$ Nano-Glo substrate, 20 nM target protein (from $2\mu \text{M}$ stock solution), and $10\mu \text{L}$ conditioned media. To assess specificity, parallel reactions were assembled with the same volume of conditioned media but substituting the target protein with 20 nM off-target protein (IL2 for PD-L1 binders; PD-L1 for BHRF1, IL20, and IL3 binders; all from $2\mu \text{M}$ stock solutions). First luminescence was recorded right after all the binders were added and second measurement was performed after 8 minutes of incubation.

To verify that binding activity could be represented by SmBiT–LgBiT signals normalized by HiBiT–LgBiT (expression) signals, purified PD-L1 KN035 binder at different concentrations (0.01 nM, 0.1 nM, 1 nM, 10 nM, 100 nM) was tested using the same experimental setup as the split-luciferase assays used for all binders with the exception of the $5\mu \text{L}$ of media, which were replaced by direct addition of purified protein at the specified dilutions. HiBiT–LgBiT measurements were not collected at 100 nM due to signal saturation. Both the HiBiT–LgBiT and SmBiT–LgBiT assays exhibited log-linear responses across the tested concentration range (Figure S11B), supporting the use of SmBiT–LgBiT normalized by HiBiT–LgBiT as a valid measure of binding activity.

Identification of nanobody binder leads for further characterization

We selected candidates for further characterization if they met two criteria: (i) expression luminescence was above 150,000 RLU (Figure S11C left) and (ii) we observed an increase in normalized binding luminescence between two time points (Figure S11C right; S11D; S30; Methods).

For each split-luciferase assay condition, three technical replicates were averaged to obtain a representative luminescence value for both expression (HiBiT–LgBiT assay) and binding (SmBiT–LgBiT assay). In the initial screens of PD-L1 and IL3, binders were first defined based on binding activity, using a threshold of at least a five-fold increase in normalized binding signal relative to the negative control. A set of PD-L1 nanobodies with different expected activities was also tested on BLI to gain insights into both their expression and binding behaviors. Based on these initial tests, only samples with HiBiT–LgBiT luminescence values above 150,000 arbitrary luminescence units (ALU) were considered expressed. Binding activity for downstream validation was subsequently determined using more refined criteria following our initial PD-L1 experiments: we defined a binder as either (i) an increase in the binding ratio (on-target/off-target) between T1 and T2 greater than 0.5, or (ii) an increase in the on-target binding ratio (on-target T2 / on-target T1) greater than 1.05.

Bio-layer interferometry (BLI)

Target proteins with AviTag and SmBiT removed and binder proteins with LgBiT removed, were used for BLI analysis. All reactions were run on an Octet RED96 at 30 °C, and samples were run in PBS pH 7.4 supplemented with 0.2% BSA and 0.05% Tween 20 (assay buffer). For each of the binder proteins, a titration of five concentrations was prepared by 2-fold serial dilution from the highest concentration tested. Lead nanobody designs were assayed for binding to biotinylated target proteins immobilized onto Octet Streptavidin Biosensors (Sartorius; 18–5019). Sensors were pre-incubated in an assay buffer for 10 minutes, before incubation in wells containing the assay buffer to determine the baseline signal. Sensors were then immersed in wells containing biotinylated target protein (200 nM in assay buffer) for 300 s or until a 1 nm (for PD-L1, IL3, and IL20) or 0.6 nm (BHRF1) BLI shift was reached. Loaded sensors were again washed and baselined in wells containing assay buffer alone, before immersion in nanobody protein solutions for association. Dissociation was monitored by transferring the sensors back into assay buffer alone. For each binder titration, a reference channel was prepared by loading sensors with target protein, but associating in assay buffer alone, to account for any change in signal due to dissociation of target protein from the biosensor tips. Association and dissociation binding curves were fit in Octet System Data Analysis Software version 9.0.0.15 using a global fit 1:1 monovalent model to determine apparent $K_{\text{D}}\text{s}$, with signals baseline-corrected by subtraction of the reference channel. Binding curves for monoclonal IgG control antibodies against IL3 and IL20 were fit to 1:2 bivalent binding models. BLI hits were identified if the design displayed clear association kinetics (BLI shift > 0.07 nm for the highest tested concentration). $R^2$ values for each derived apparent $K_{\text{D}}$ are found in Data S3.

Surface Plasmon Resonance (SPR)

Affinity quantification of designed scFvs by SPR were carried out by the contract research organization Adaptyv Bio. SPR experiments were conducted using a Carterra LSA XT instrument, and at least two independent binding experiments (n = 2) were performed per design. Briefly, designed scFvs were reformatted into fragment antibody (Fab) format, by fusing the designed variable heavy (VH) domain to human IgG1 CH1 (Uniprot P01857, positions 1–103), and fusing the designed variable light (VL) domain to human Ig $\kappa$ light chain (Uniprot P01834, positions 1–107). Fab designs were expressed with C-terminal Twin-Strep tags in a prokaryotic in vitro translation system and normalized post-expression using an affinity-based quantification assay. SPR sensor chip surfaces were functionalized by covalently attaching Strep-Tactin XT to a carboxymethylated surface using EDC/NHS coupling, before twin-Strep-tagged Fabs were printed onto the chip using 96-channel bidirectional flow (750 s capture, 600 s equilibration). A PD-L1 dilution series (10–1000 nM, half-log dilutions) was prepared using running buffer (10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.05% Tween-20, pH 7.4) and injected onto the chip following a single-cycle kinetics format. Each cycle included 60 s baseline, 300 s association, and 600 s dissociation phases. Surfaces were regenerated between cycles with glycine-HCl, pH 1.5. $R^2$ values were not provided for SPR kinetic fits.

Sensorgrams were processed in Adaptyv Fitting software with baseline and reference subtraction. Data were globally fit to a 1:1 Langmuir model to extract $k_{\text{on}}$, $k_{\text{off}}$, and $K_{\text{D}}$ values. Fabs were classed as hits if they exhibited clearly measurable interaction signal with successful model fits and a calculated $K_D\le 10\mu \text{M}$. Binders producing association signals >300% above negative controls but lacking reliable fits were manually classified based on response magnitude. The raw data for SPR experiments and corresponding kinetic fits for each replicate for all validated binders can be found in Data S3.

Epitope-targeted alanine mutagenesis

Designed epitopes were interrogated by screening binding activity against alanine substitution variants at selected hotspot residues (PD-L1: I37A, Y39A, E41A; IL3: L26A, T103A, F104A; IL20: R59A, R63A, R99A; BHRF1: V85A, I89A, R92A). Residue numbering for all mutants corresponds to the mature protein sequence (i.e., after signal peptide cleavage) and thus differs from full-length UniProt numbering by the length of the signal peptide: +17 for PD-L1, +28 for IL3, +25 for IL20, and +1 for BHRF1. Variant antigens were expressed, purified, and assayed for binding against designed binders by BLI as described for their wild-type counterparts. All BLI data were collected and analyzed using Octet System Data Analysis Software (v9.0.0.15), with signals baseline-corrected by reference channel subtraction. When available, a positive control antibody targeting a different epitope was used to validate structural correctness and folding of the protein. Anti-PD-L1 nanobody KN035, anti-BHRF1 de novo protein binder GDM BHRF1 35, AntiIL3 monoclonal antibody EPR7964 (Abcam, cat. ab167159, 1034174–12) and anti-IL20 monoclonal antibody EPR20756–1 (Abcam, cat. ab244722, 1113252–8) were used.

Polyspecificity reagent preparation

Polyspecificity reagent (PSR) was prepared from Expi293F cells as described previously [55, 57], with minor modifications. Briefly, 10⁹ cells were pelleted at 550×g for 3 min, washed once in ice-cold PBS supplemented with 1 mg/mL BSA (Sigma-Aldrich, A4737; PBS-B), and resuspended in ice-cold Buffer B (50 mM HEPES, 150 mM NaCl, 2 mM CaCl₂, 5 mM KCl, 5 mM MgCl₂, 10% v/v glycerol, pH 7.2) supplemented with protease inhibitor cocktail. Cells were homogenized (3×30 s) and sonicated (3×30 s) on ice. The lysate was centrifuged at 40,000 × g for 1 h at 4 °C; the supernatant (soluble cytosolic protein fraction) was retained, and the membrane-enriched pellet was resuspended in Buffer B with protease inhibitor using a Dounce homogenizer (30 strokes). Protein concentration was determined using the DC Protein Assay Kit I (Bio-Rad, 5000111). The membrane fraction was diluted to approximately 1 mg/mL in solubilization buffer (50 mM HEPES, 150 mM NaCl, 2 mM CaCl₂, 5 mM KCl, 5 mM MgCl₂, 1% n-dodecyl-$\beta$-D-maltopyranoside) with protease inhibitor and rotated end-over-end overnight at 4 °C. The suspension was centrifuged at 40,000×g for 1 h at 4 °C to yield the soluble membrane protein (SMP) fraction in the supernatant. Protein concentration was re-determined and adjusted to 0.7 mg/mL. Biotinylation was performed by addition of 20 μL of freshly prepared 10 mM EZ-Link Sulfo-NHS-SS-Biotin (Thermo Fisher Scientific, 21331) per mg SMP, with incubation at 25 °C for 45 min with gentle agitation. Reactions were quenched by addition of 50 μL of 25 mM Tris-HCl pH 8.0 per mL of reaction. Finally, biotinylated PSR was dialyzed into PBS (pH 7.4) overnight at 4 °C using 6–8 kDa MWCO tubing, aliquoted, and stored at −80 °C.

Polyspecificity assay

Non-specific binding of purified His- and FLAG-tagged hit binders was assessed using a flow cytometry-based polyspecificity particle assay adapted from Makowski et al. [55] and Hie et al. [56]. Magnetic beads coated with an NTA-Cobal-based ligand (His-Tag Isolation and Pulldown Dynabeads; Thermo Fisher Scientific, 10103D) were washed twice in flow buffer (PBS, pH 7.4 supplemented with 1 mg/mL BSA, 4 °C) and diluted to 540 μg/mL. His- and FLAG-tagged nanobodies or Fabs were diluted to 5 μg/mL or 20 μg/mL, respectively, in flow buffer. Beads (30 μL) were incubated with 85 μL of binder solution in a 96-well plate overnight at 4 °C with rocking. Bead-bound binders were washed twice with flow buffer using a magnetic plate stand, then incubated with 50 μL of 0.1 mg/mL biotinylated PSR for 20–30 min at 4 °C with rocking, followed by one wash in flow buffer. Beads were subsequently incubated for 15 min at 4 °C with streptavidin-APC (BioLegend, 405207; 1:1000), anti-FLAG M2 mouse monoclonal antibody (Sigma-Aldrich, F3165, 0000421805; 1:1000), and goat anti-mouse secondary antibody conjugated to Alexa Fluor Plus 488 (Thermo Fisher Scientific, PIA32723TR, 3263235; 1:1000). All detection antibody dilutions were prepared in flow buffer. Beads were washed once more, resuspended in 200 μL flow buffer, and analyzed on an Attune NxT flow cytometer (Invitrogen) using the Attune Cytometric Software. APC fluorescence reported PSR binding as a measure of polyspecificity, and Alexa Fluor 488 fluorescence confirmed binder loading on beads. Flow cytometry data were analyzed in FlowJo v10.10.1 (BD Biosciences).

Expression and purification of anti-Fab nanobody

The anti-Fab nanobody was expressed and purified as previously described [83]. Briefly, sequences encoding the anti-Fab nanobody were cloned into a pET26b+ vector containing an N-terminal polyhistidine tag and transformed into BL21(DE3) E. coli. Cultures were grown overnight at 37 °C with shaking at 250 rpm in Luria–Bertani (LB) medium supplemented with 50 μg/mL kanamycin, then used to inoculate large-scale cultures grown under the same conditions to OD₆₀₀ 0.6–0.8, at which point expression was induced with 1 mM IPTG for 3–4 h. Cells were harvested by centrifugation at 4,000 × g for 15 min and pellets stored at −80 °C or immediately used for purification.

For purification, cell pellets were resuspended in purification buffer (20 mM HEPES, pH 7.4, 150 mM NaCl, 20 mM imidazole, protease inhibitor cocktail) and lysed by sonication. Lysates were clarified by centrifugation at 40,000×g for 45 min and incubated with HisPur Ni-NTA resin (Thermo Fisher Scientific) pre-equilibrated with purification buffer for 1 h at 4 °C. Resin was loaded onto a gravity-flow polypropylene column, washed extensively with purification buffer, and bound protein was eluted with 300 mM imidazole. The eluate was concentrated and further purified by size-exclusion chromatography using a Superdex 200 Increase 10/300 column (Cytiva) pre-equilibrated with imidazole-free purification buffer. Peak fractions corresponding to the anti-Fab nanobody were collected, concentrated, and flash-frozen in liquid nitrogen for storage at −80 °C.

Complex assembly

Purified anti-Fab nanobody and anti-PD-L1 Fab were mixed in a 3:1 molar ratio and incubated on ice for 15 min. The mixture was concentrated and subjected to size-exclusion chromatography as described above. Peak fractions corresponding to the nanobody–Fab complex were collected and concentrated (Figure S31). Purified PD-L1 was then mixed with the nanobody–Fab complex in a 3:1 molar ratio, incubated on ice for 15 min, and further purified by size-exclusion chromatography. Peak fractions corresponding to the ternary antigen–binder complex were collected and concentrated to approximately 1.0 mg/mL for cryo-EM studies.

Cryo-EM sample preparation and data collection

3 μL of purified complex was applied to glow-discharged holey carbon grids (Quantifoil R1.2/1.3 Au, 300 mesh) and vitrified in liquid ethane using a Vitrobot Mark IV instrument (Thermo Fisher Scientific). Grids were imaged on a 300 kV Titan Krios G2 transmission electron microscope equipped with a Selectris imaging filter and Falcon 4i direct electron detector (Thermo Fisher Scientific). Movies were acquired using EPU software (Thermo Fisher Scientific) at a nominal magnification of 130,000×, corresponding to a physical pixel size of 0.98 Å/px, with a defocus range of −1.0 to −2.5 μm. A total electron dose of 50 e⁻/Ų was fractionated across 40 frames.

Cryo-EM data processing

A total of 2,613 movies were collected for the PD-L1 binder complex. All processing was performed in cryoSPARC v4.7.1 [84]. Movies were preprocessed using patch motion correction and patch CTF estimation. Particles with diameters of 120–140 Å were initially picked using the blob picker and extracted at 4× binning. Two rounds of 2D classification were performed to remove low-quality particles, followed by ab initio 3D reconstruction and heterogeneous refinement, yielding classes consistent with a Fab–nanobody–PD-L1 complex. A randomized particle subset was used to train a Topaz model [85] for additional picking, followed by further 2D classification at 4× binning. Particles in high-quality classes were re-extracted without binning and subjected to iterative heterogeneous and non-uniform refinements. The resulting stack was merged with the blob-picked pool, duplicates removed, and subjected to further heterogeneous and non-uniform refinements, yielding an initial ~4.1 Å reconstruction with clear Fab, nanobody, and PD-L1 density.

To improve resolution, an additional round of template-based and Topaz picking was performed in parallel. Particles from each method were processed through 2D classification and heterogeneous refinement, then merged with duplicates removed. The final stack of 94,644 particles was subjected to non-uniform refinement, yielding a ~3.9 Å reconstruction of the full nanobody–Fab–PD-L1 complex. Masked local refinement at the Fab–PD-L1 interface yielded a final reconstruction at 3.93 Å. Local resolution was estimated in cryoSPARC, with all resolutions reported based on the gold-standard Fourier shell correlation (FSC) criterion of 0.143. A schematic depicting this process, along with a representative cryo-EM micrograph and selected 2D class averages are shown in (Figure S32).

Model building, refinement, and structural analysis

The predicted structure of Fab-bound PD-L1 was fitted into experimental cryo-EM density maps using UCSF ChimeraX [86]. Fitted structures were manually rebuilt in Coot [87], refined using Phenix [88], and validated using MolProbity [89]. Final maps and models were visualized in UCSF ChimeraX v1.5.

Data Availability

The cryo-EM map and atomic model of the PD-L1 scFv H5 complex have been deposited in the Electron Microscopy Data Bank (EMDB accession EMD-77181) and the Protein Data Bank (PDB accession 35TL), respectively. We have also uploaded raw data to Zenodo (doi: 10.5281/zenodo.20221722). All data, including the sequences of all tested designs and template plasmid maps used in this study, are available in the main text or as supplementary data.Supplementary data files are provided as follows:

• Supplementary Data 1: Template plasmid maps used in this work.

• Supplementary Data 2: Designed sequences per target and their corresponding names.

• Supplementary Data 3: Derived $K_{\text{D}}\text{s}$ and associated $R^2$ values from BLI experiments.

Code Availability

We make code for running the pipeline available at https://github.com/SantiagoMille/germinal.