Germinal: Open-Source Nanobody Design With Epitope Evidence and Some Unresolved Gaps

Stanford and Arc Institute’s pipeline designs nanobodies against specified epitopes, testing only 43–101 candidates per target.

A Stanford/Arc Institute team released Germinal in Sep 2025, an open-source generative framework for de novo nanobody design. The approach combines AlphaFold-Multimer (AF-M) for structural confidence with IgLM, an antibody-specific language model, for sequence naturalness. CDRs are designed from scratch on a fixed framework, filtered with AF3 and PyRosetta-derived biophysical and biochemical scores, and validated experimentally.

The results across four diverse targets: 4–22% BLI-verified success rates from libraries of just 43–101 nanobodies per antigen. Best affinities reached 140 nM (BHRF1), 170 nM (PD-L1), 190 nM (IL20), and 560 nM (IL3). All nanomolar. All from small design pools.

I’m going to walk through what they did, then go through the validation gaps I found.

How the Model Works

Germinal works in three stages: design, sequence optimization, and filtering.

In the design stage, gradients from AF-M and IgLM are merged at each iteration to co-optimize structure confidence and antibody-likeness. Three custom loss functions steer CDRs away from alpha-helices and beta-strands (which AF-M otherwise favors) and enforce that binding contacts occur through CDR residues rather than the framework. The design progresses through logit optimization, softmax annealing, and semi-greedy discrete selection (see appendix for more info). Each design iteration takes 2–8 minutes on an H100 GPU (per their repo README), and full campaigns require 200–500 H100-GPU hours per target.

In the sequence optimization stage, AbMPNN redesigns non-contact CDR residues to improve stability while preserving the predicted binding interface.

In the filtering stage, AF3 provides an independent structural assessment (different from the AF-M used during design), and PyRosetta scores provide biophysical metrics. Only designs passing stringent thresholds on ipTM, pLDDT, PAE, CDR interface percentage, shape complementarity, hydrogen bonds, and several other criteria advance to experimental testing.

Using AF3 as a separate validator from the AF-M used during optimization reduces the risk of the pipeline finding sequences that game a single model’s confidence metrics.

Germinal pipeline overview and experimental results. (A) The pipeline takes a target structure and antibody framework as input, designs CDRs by merging AF-M and IgLM gradients through three phases (logits, softmax with temperature annealing, semi-greedy selection), then optimizes non-contact CDR residues with AbMPNN and filters with AF3. (B) AF-M and IgLM define separate optima in sequence space; joint optimization steers toward sequences that score well on both objectives. (C) AF3-predicted structures of the best binders for each target, with CDRs in pink and target binding domains in blue. Best KD values by BLI are annotated. (D) Screening results per target. Of 43–101 designs tested per library, 4–22% were verified as binders by BLI. “Max. Id.” is the highest pairwise CDR sequence identity between any design and the PDB, confirming the designs are novel.

The funnel from computation to validated binders looks like this:

The total number of trajectories attempted before hallucination filtering isn’t reported. The README in the GitHub repo says they typically run “until there are around 1,000 passing designs,” which at ~1 pass per GPU hour implies thousands of total attempts per target. The dropoff from hallucination to experimental selection is steep: only the top 3–7% of passing designs were chosen for testing based on the metrics listed in the table below.

Target Diversity

They tested against four soluble proteins:

• PD-L1: 101 designs tested, 7 BLI-verified binders (7%), best KD = 170 nM.
• IL3: a cytokine with no previously reported de novo binders. 46 designs, 2 BLI-verified binders (4%), best KD = 560 nM.
• IL20: another cytokine without prior de novo binders. 43 designs, 4 BLI-verified binders (9%), best KD = 190 nM (KD1 from a 2:1 heterogeneous ligand fit).
• BHRF1: an Epstein-Barr virus Bcl-2-like protein. 52 designs, 11 BLI-verified binders (22%), best KD = 140 nM.

All input target structures were AF3 predictions rather than crystal structures, which demonstrates the pipeline can work without experimentally solved target structures.

Epitope Targeting Evidence

For PD-L1, they introduced alanine substitutions at three predicted interface residues (I54, Y56, E58). Each substitution reduced binding by >10-fold for some designs and had little effect on others. This pattern makes sense if different nanobodies contact different residues within the epitope. A nanobody that relies heavily on Y56 for binding loses affinity when Y56 is mutated to alanine, but a nanobody that primarily contacts I54 is unaffected by that same mutation. If the mutations were simply destabilizing PD-L1 as a whole, all nanobodies would lose binding equally. This differential sensitivity pattern is consistent with the designed binders contacting the intended epitope rather than binding somewhere else entirely.

However, the alanine scan was only performed on PD-L1. There’s no epitope validation for IL3, IL20, or BHRF1. Given that IL3 and IL20 are new targets where no prior binder data exists, epitope mapping on at least one additional target would strengthen the claim that Germinal’s epitope targeting generalizes.

Screening Assay and Controls

The experimental workflow uses a split-luciferase (NanoBiT) assay as the initial screen. Binder-LgBiT fusions expressed in HEK293 cells are tested for binding by adding antigen-SmBiT fusions. They included both positive controls (known binders to the same target, when available) and negative controls (binders against unrelated antigens).

The assay shows standard curves with R² > 0.97 for both expression (HiBiT) and binding (SmBiT).

The dropout between luciferase hits and BLI-verified binders is substantial. For PD-L1: 25 candidates passed luciferase filters, but only 7 showed detectable binding on BLI. For BHRF1: 20 passed luciferase, 11 confirmed on BLI.

Strengths

• Full open-source release with pipeline code, hyperparameter configs for PD-L1 and IL3, all designed sequences, and Docker/Singularity containerization.
• Alanine scanning on PD-L1 provides good epitope targeting evidence.
• Split-luciferase screening assay is practical, well-characterized with standard curves, and accessible to labs without specialized high-throughput equipment.
• Both positive and negative controls included in each screening plate.
• Luciferase-to-BLI dropout rates transparently reported.

Gaps

Gap #1: No Polyreactivity or Polyspecificity Testing

The paper’s specificity testing is limited to a single off-target protein per screen (IL2 for PD-L1 binders; PD-L1 for the other three targets). For the BLI specificity check (Figure S12), they tested only two binders per target against one off-target antigen at a single concentration (2000 nM).

Testing against one off-target protein doesn’t tell you whether a binder is polyreactive. Polyreactive antibodies bind many unrelated targets through non-specific hydrophobic or electrostatic interactions. You can pass a single-antigen specificity check and still be polyreactive. The standard assays for this are baculovirus particle (BVP) ELISA, polyspecificity reagent (PSR) binding, heparin chromatography, or at minimum a panel of unrelated antigens. None of these were performed.

The authors do acknowledge this directly: “more comprehensive poly-specificity measurements will be required to rigorously confirm antigen specificity.”

How nature handles polyreactivity: During B cell development, antibodies pass through multiple tolerance checkpoints that actively remove polyreactive clones from the repertoire. In the bone marrow, immature B cells expressing polyreactive BCRs undergo receptor editing or clonal deletion. Those that escape face additional selection in the periphery, where chronically self-reactive cells are anergized or excluded from germinal centers. De novo designed CDRs skip all of these checkpoints. There is no selection pressure against non-specific hydrophobic or electrostatic interactions, which means polyreactivity has to be assessed experimentally after the fact.

Gap #2: Affinities Are 10–100x Weaker Than Therapeutic Requirements

Best affinities range from 140 nM to 560 nM across the four targets. Therapeutic antibodies typically need single-digit nanomolar or better to be clinically viable. These affinities are starting points that would require extensive affinity maturation.

The paper frames these as strong results (“nanomolar binding affinities” in the abstract), but context is important here. The PD-L1 positive control (KN035) has a KD of 3 nM. The best Germinal PD-L1 binder is 57-fold weaker at 170 nM. For IL3, the best binder is 560 nM, which is closer to micromolar territory.

The authors do note that protein language models could enable affinity maturation as a future step.

Gap #3: No Experimental Developability Data

The paper presents computational developability metrics (Therapeutic Nanobody Profiler - TNP scores, OASis humanness scores, surface hydrophobicity, charge distribution) and shows that IgLM guidance improves these metrics. But there’s no experimental measurement of any developability parameter.

At a minimum, I’d want to see the following to assess whether these designs are viable starting points, with extended assays (marked with *) as bonuses:

• Thermal stability by DSF or DSC
• Analytical SEC for aggregation and oligomeric state
• Expression yields across the design set
• Polyreactivity panel (BVP, PSR, or heparin chromatography)
• Self-interaction as viscosity proxy (e.g., via AC-SINS)*
• Polydispersity with DLS*

The framework choice (hNbBCII) is well-characterized and likely does confer stability, but designing new CDRs onto a stable framework can still introduce aggregation-prone motifs, affect thermal stability, or create developability liabilities.

Even a basic DSF + DLS screen on the 24 BLI-verified binders would have provided useful data. The purified proteins were already in hand. They could have also run PAIA’s plate-based developability kits to get some good assessments quickly with little material:

• Hydrophobicity (HIC assay)
• Polyreactivity against heparin
• Polyreactivity against ovalbumin

Gap #4: IL20 Kinetics Require a 2:1 Binding Model, and BLI Loading Levels Are Not Reported

The IL20 sensorgrams required fitting to a 2:1 heterogeneous ligand model rather than the standard 1:1 kinetics used for the other targets. The paper explains that 1:1 fits showed systematic residuals during dissociation, and the 2:1 model minimized residuals.

Biphasic kinetics on BLI can arise from several sources. Analyte aggregation at the sensor surface has been shown to produce biphasic association and dissociation that worsens with longer association times and higher analyte concentrations (Sherer and Cho, ACS Omega 2025). High ligand loading density can introduce mass transport limitation and analyte rebinding artifacts (Kamat & Rafique, Anal. Biochem. 2017).

Conformational heterogeneity of the immobilized ligand or the analyte is another possibility. The nanobody analytes are monovalent, so bivalent binding can be ruled out. But the paper doesn’t investigate which of the remaining causes is responsible.

The paper never reports actual sensor loading levels for any target. The methods describe loading biotinylated target at 200 nM “for 300 s or until a 1 nm BLI shift was reached,” but the achieved loading levels are not disclosed. High ligand density on streptavidin sensors is one of the most common causes of biphasic kinetics in BLI, because it creates heterogeneous binding populations and can introduce mass transport limitation and analyte rebinding artifacts. Without knowing the loading levels, we can’t assess whether the IL20 biphasic behavior (and potentially the IL3 poor fits) reflects an assay artifact or something intrinsic to the interaction.

If they had tested at lower loading densities and the biphasic behavior went away, that points to an assay artifact. If it persisted, the two kinetic populations are likely intrinsic to the interaction. Flipping the assay orientation (immobilizing the nanobody, flowing IL20 as analyte) would also help. Neither experiment was done.

Gap #5: IL3 BLI Fits Below Quality Threshold

For IL3, both binders showed R² = 0.93 on the kinetic fits. The paper acknowledges this falls below their own quality threshold of R² > 0.95 for the other targets. To compensate, they averaged KD values across replicate experiments and tested against both C-terminal and N-terminal AviTag versions of IL3.

Averaging replicates of a poor-fit model doesn’t improve the underlying fit quality. It reduces random error but doesn’t address systematic deviation between the model and the data. If the 1:1 binding model doesn’t describe the sensorgrams well, the extracted KD is unreliable regardless of how many times you repeat the experiment. The same loading-level concerns from IL20 apply here: without knowing the ligand density on the sensors, we can’t distinguish poor fits caused by the interaction itself from those caused by assay artifacts.

The reported IL3 KD values (560 nM and 2.5 µM) are approximate at best. The weaker of the two binders is already in the micromolar range.

Gap #6: Configs in the Repo Don’t Fully Match the Paper

The GitHub repo provides exact hyperparameter configs for PD-L1 and IL3, but not for IL20 or BHRF1. Anyone wanting to reproduce the IL20 or BHRF1 results would need to guess the parameters, and the README notes that “final sampling runs in the preprint all used slightly modified parameters” from the defaults.

There’s also a species conditioning detail that isn’t discussed in the paper. The PD-L1 config conditions IgLM on [CAMEL] species while IL3 uses [HUMAN]. Since the hNbBCII-FGLA framework originates from a camelid scaffold, [CAMEL] conditioning isn't unreasonable. But the paper doesn't mention that species conditioning was varied per target. The methods section gives <HUMAN> as the example species token, implying uniform conditioning.

Gap #7: scFv Results Built on a Framework That Failed Grafting Controls

The paper reports computational scFv design results (Figures S5, S6) with metric distributions comparable to nanobodies. None were tested experimentally.

The framework validation (Figure S1) shows why. They tested two scFv frameworks: trastuzumab and FW1.4. FW1.4 didn’t express. Trastuzumab expressed but lost all binding activity when grafted with known functional CDRs from F5111. They chose trastuzumab anyway and used it for all computational scFv designs.

The question is why include scFv results at all. They had no working scFv framework, they didn’t screen additional frameworks to find one that could support CDR grafting, and they didn’t test any of the computational designs. Computational metrics on a framework that can’t support known binders don’t add information.

Gap #8: Expression Failures Not Analyzed

Looking at Figure 1D, substantial fractions of designs failed to express: 23/101 for PD-L1 (23%), 10/46 for IL3 (22%), 13/43 for IL20 (30%), and 21/52 for BHRF1 (40%).

The paper’s differential analysis (Figure S7A) identified a few metrics correlated with expression (CDR SAP score, pLDDT), but doesn’t examine why expression failure varies so much across targets.

Since all designs use the same framework, expression failures should be driven by CDR properties. A systematic analysis of what distinguishes expressed from non-expressed designs (CDR hydrophobicity, aggregation-prone motifs, unusual amino acid compositions) would improve future design cycles.

Gap #9: No Audit of False Negatives from the Luciferase Screen

The luciferase filter removes a large fraction of designs before they ever reach BLI. For PD-L1, only 25 out of 78 expressing designs were sent to BLI. For BHRF1, 20 out of 31. The reported success rates (4–22%) are calculated from total designs tested per target, but the luciferase triage determines which of those designs get a chance at BLI confirmation. The authors themselves note the assay reflects “apparent association” and is influenced by transfection efficiency, cell debris, and other artifacts.

Taking 10–20 luciferase-failing designs per target and running them on BLI anyway would quantify the false negative rate. If even a few show binding, the luciferase filter is discarding functional binders and the conditional success rate (28–55% among designs that reached BLI) better reflects the pipeline’s design capability.

Gap #10: No Experimental Ablation of Pipeline Components

Germinal introduces paratope loss, alpha-helix loss, beta-strand loss, IgLM guidance, AbMPNN CDR redesign, and three different gradient merging strategies (weighted sum, PCGrad, MGDA). The paper shows computational effects of some of these (IgLM guidance improves predicted developability scores, secondary structure losses shift CDR conformations), but none of these ablations are connected to experimental outcomes.

Running even a small ablation on one target (e.g., designs with vs without IgLM guidance, tested through to BLI) would show whether the computational improvements translate to experimental differences. As it stands, some of these components could be computationally measurable but experimentally irrelevant.

Gap #11: Target Truncation and Presentation Format Not Validated

PD-L1 was truncated to the V-like domain (residues 19–127) for computational efficiency. All targets were presented as soluble ectodomains with fusion tags (SmBiT for luciferase, AviTag + His for BLI). Binding was never confirmed against full-length targets or in a cell-surface context.

For PD-L1, a type I transmembrane protein, cell-surface binding by flow cytometry on PD-L1-expressing cells would confirm the binders work on the native target in its membrane context. Domain truncation can remove steric constraints that block access to the epitope in the full-length protein, and fusion tags near the epitope can interfere with or artificially enhance binding. The IL3 AviTag orientation issue (requiring both C-term and N-term versions) already shows tag placement affects the binding readout.

Gap #12: No Run-to-Run Reproducibility Data

The paper reports results from single design campaigns per target. The README notes that “in silico success rates vary wildly across targets,” and hallucination-based design is stochastic. Running the same target, same epitope, same parameters a second time could yield a different hit rate.

Even two independent runs on PD-L1 (their most data-rich target) carried through to BLI would show whether the reported hit rates are reproducible or target-specific outliers.

Too many acronyms in one post? **Check out my biotech abbreviation cheat sheet** and feel free to suggest additions.

Follow me on **Substack, Medium, Bluesky, and LinkedIn **for more posts on drug discovery, assay development, and screening workflows.

Appendix: Germinal’s Three Design Phases

Germinal designs CDR sequences through three phases that progressively shift from continuous exploration to discrete sequence selection.

Logit optimization. Instead of starting with a fixed amino acid sequence, the model assigns a continuous score (logit) to each of the 20 possible amino acids at every CDR position. These scores are updated iteratively using gradients from both AF-M and IgLM. Because the scores are continuous, the optimization can make smooth adjustments rather than swapping one amino acid for another in discrete steps.

Softmax annealing. The logits are converted to probability distributions over amino acids at each position using the softmax function. Early in optimization, the temperature parameter is high, so probabilities are spread across multiple amino acids at each position. As the temperature decreases, the distribution sharpens until one amino acid dominates at each position. This is analogous to simulated annealing in protein structure prediction: high temperature allows broad conformational sampling, low temperature locks in the final state.

Semi-greedy discrete selection. In the final phase, the model switches from continuous probabilities to actual amino acid identities. At each iteration, it samples 5 candidate sequences from the current distributions, scores each with the combined AF-M + IgLM objective, and keeps the best. This is “semi-greedy” because it evaluates multiple samples rather than deterministically picking the highest-probability residue at each position, reducing the risk of converging on a local optimum.

For more info, here are some additional reading materials: • Softmax function and temperature: https://en.wikipedia.org/wiki/Softmax_function • Simulated annealing: https://en.wikipedia.org/wiki/Simulated_annealing • Greedy algorithms: https://en.wikipedia.org/wiki/Greedy_algorithm

Views expressed are my own and do not represent my employer.