Use Germinal Online

Germinal: De Novo Antibody Design for Epitope-Targeted Nanobodies and scFvs

Germinal is a generative pipeline for de novo antibody design that designs the complementarity-determining regions (CDRs) of nanobodies and single-chain variable fragments (scFvs) against a chosen epitope on a protein target. The authors state that the full computational pipeline and the experimental screening protocol are open-source and freely available.

Designing antibodies from scratch is hard because binding depends on a handful of flexible CDR loops, and a design has to look like a real, well-behaved antibody as well as bind. Germinal addresses this by combining a structure-based binding signal with an antibody-specific language model, so that designs are both confident binders and antibody-like. The authors report that this enables low-throughput testing, with tens of designs per target rather than thousands.

How Germinal Designs Antibodies: Structure Prediction Meets an Antibody Language Model

Germinal designs by gradient-based optimization rather than by sampling from a trained generator. Two signals compete and are balanced during design.

1. Binding confidence: gradients from AlphaFold-Multimer (AF-M) push the antibody toward a confident, well-formed interface with the target.

2. Antibody-likeness: gradients from IgLM, an antibody-specific language model, keep sequences realistic. The authors note a trade-off between AF-M confidence (iPAE) and IgLM likelihood, and stronger IgLM guidance improved predicted developability and humanness scores.

3. Staged optimization: design runs in three phases (logits, softmax with temperature annealing, and semi-greedy). A framework can be fixed as a template while the CDRs are designed freely.

4. Custom losses: a paratope loss keeps binding in the CDRs rather than framework residues, and helix and strand losses push CDRs toward flexible loop conformations.

5. Sequence cleanup and filtering: AbMPNN redesigns non-contacting CDR residues for stability, then AlphaFold 3 (a different predictor from the design stage) and PyRosetta scores are used to filter and rank designs.

Experimental Results: Nanobody Binders Against Four Targets

The authors tested designs against four protein targets using a split-luciferase screen followed by biolayer interferometry (BLI). Most designs had CDR sequences below 50% identity to sequences in SAbDab or OAS.

Overall BLI success was 4 to 22% per target, which the authors describe as likely conservative, and each target yielded between 2 and 11 sub-micromolar binders. For PD-L1, alanine substitutions at I54, Y56 and E58 each weakened binding by more than 10-fold, consistent with the intended epitope. scFv designs passed filters at rates comparable to nanobodies.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with Germinal on Tamarind Bio

• Nanobody discovery: generate candidate binders against a protein target without immunization or large library screening.

• Epitope-focused design: direct designs toward a chosen region of an antigen, as the authors did for PD-L1.

• scFv design: design scFv binders, which the authors report pass filters at rates comparable to nanobodies.

• Small experimental campaigns: shortlist tens of ranked designs for validation instead of thousands.

How to Use Germinal on Tamarind Bio

1. Open the tool: log in to tamarind.bio and open Germinal.

2. Provide the target: supply the antigen structure. The authors used AlphaFold 3 predictions as input structures for their targets.

3. Choose the epitope: specify the target residues the antibody should bind, if exposed.

4. Select the format: choose nanobody or scFv and the framework template, if exposed.

5. Set sampling: choose the number of design trajectories, if exposed. The authors ran roughly 700 to 1,600 trajectories per target that passed initial filters.

6. Run the job: submit the design run and wait for completion.

7. Download and rank: download the designed sequences and structures, and review the confidence and biophysical scores.

8. Validate: test a shortlist experimentally, for example with a binding assay followed by BLI.

Things to Keep in Mind

• Performance depends on AlphaFold-Multimer prediction accuracy and confidence.

• Iterative backpropagation and structure prediction are computationally intensive and require extensive sampling.

• Memory limits restrict use to smaller targets; large proteins may need truncation to the region of interest.

• Only protein epitopes are supported, not glycans, small molecules or other non-protein antigens.

• Affinity maturation is not addressed, and polyspecificity needs broader testing.

Source: Germinal preprint on bioRxiv, DOI 10.1101/2025.09.19.677421 (https://www.biorxiv.org/content/10.1101/2025.09.19.677421v1). The authors state the pipeline is open source; no repository URL was found on the page read.