Immunogenicity risk is a critical challenge in therapeutic antibody development. Traditional antibody humanness scoring tools typically output a single score for an entire sequence and rely heavily on framework region (FR) information, which can obscure potential non-human signals in complementarity-determining regions (CDRs). Cyagen Biosciences has published a preprint on bioRxiv entitled Multi-resolution humanness scoring of antibodies via a region-isolated language model. The study introduces IgScope, a region-isolated antibody language model that evaluates humanness at three resolutions—chain, region, and amino-acid residue—addressing the limitations of conventional whole-sequence scoring and providing a new computational tool for AI-generated antibody auditing, antibody humanization, and quality control of fully human antibody mouse platforms.

DOI: https://doi.org/10.64898/2026.10.23.761395
Industry challenge: A single overall score can mask potential antibody immunogenicity risks
In antibody drug development, humanness scoring is a key early-stage quality-control step. It evaluates the similarity between an antibody sequence and naturally occurring human antibodies, helps anticipate the risk of anti-drug antibody (ADA) formation, and reduces potential clinical immunogenicity concerns. Leading tools such as OASis, T20, AbNatiV, HumAb, and AntiBERTy generally output a single score for the entire antibody chain. Framework regions account for approximately three quarters of the antibody variable domain and are relatively conserved. By contrast, complementarity-determining regions (CDRs), which determine antigen binding, are central to antibody humanization grafting and AI-generated antibody design. Conventional models can learn shortcuts by relying on conserved framework signals to assign high humanness scores while failing to identify hidden non-human sequence segments within CDR loops.
Key risk: T-cell immunogenic epitopes are recognized as local short peptides rather than as an entire antibody sequence. Even when the overall chain score is excellent, a non-human segment in a local CDR can still present an ADA risk. This issue is especially important for AI-generated antibodies: many generated antibodies meet whole-chain humanness thresholds, while their CDR sequences can deviate substantially from natural human repertoires, a defect that traditional tools may not detect.
Technical breakthrough: IgScope uses a region-isolated language model for three-resolution humanness assessment
IgScope is a 23-million-parameter antibody Transformer encoder trained from scratch. Its training dataset contains 119 million human, mouse, and camelid heavy-chain, light-chain, and VHH nanobody sequences. The model introduces a region-isolated attention module and incorporates chimeric-sequence augmentation at a 30% proportion, preventing shortcut learning in which framework information is used to infer CDR humanness.
Two core architecture innovations
Module A, region-isolated encoder: The attention mechanism allows sequence interactions only within the same IMGT-defined region. CDR representations are derived entirely from their own amino acids without interference from framework sequences, ensuring that regional scores are genuinely locally independent.
Module B, global encoder: Full-sequence attention is enabled for whole-chain species classification and residue-level assessment. Two evolutionary auxiliary tasks—ancestral germline gene prediction (AGP) and mutation process prediction (MPP)—provide fine-grained residue-level signals.
A single forward pass produces three levels of output:
Chain level: Human probability for the entire heavy chain, light chain, or VHH.
Region level: Humanness scores for the seven IMGT-defined regions: FR1-CDR1-FR2-CDR2-FR3-CDR3-FR4.
Residue level: Human probability for every amino-acid position, enabling localization of non-human mutations.

Figure 1. Overall architecture of IgScope.
Robust validation: Multiple experiments demonstrate model reliability
1. Locality validation: CDR scoring is independent of framework sequences
When all framework-region tokens were masked, IgScope showed almost no decrease in CDR-recognition accuracy (Figure 2a and Figure 3c). In contrast, the species-classification AUC-ROC of the leading antibody model AntiBERTy dropped sharply to 0.672 after framework masking (Figure 3c), indicating that its prediction relied largely on framework signals. In a reversion-mutation test that introduced mouse residues back into a human framework, IgScope achieved 93.4% residue-recognition accuracy and accurately localized individual non-human amino-acid positions embedded in a human framework (Figure 2b).

Figure 2. Multi-resolution diagnostic experiments.
2. Leading overall performance on external datasets
On the fully independent VDJServer external test set, IgScope achieved an AUC-ROC of 0.9997 for species classification, significantly outperforming mainstream tools including Humatch, AntiBERTy, OASis, and T20 (Figure 3b). Among 69,899 synthetic chimeric-grafted antibodies, IgScope achieved a grafting-detection ROC-AUC of 0.9938 and accurately identified grafted molecules whose FR and CDR source species were inconsistent (Figure 3d–f).

Figure 3. Comparison with mainstream antibody humanness scoring tools.
3. Strong correlation with clinical anti-drug antibody (ADA) incidence
In a dataset of 217 therapeutic antibodies with real clinical ADA data, the IgScope chain-level score showed a Pearson correlation of r=-0.567 with clinical ADA incidence, placing it in the same tier as the leading models in the field (Figure 4b). Using ADA>10% as a binary threshold, IgScope achieved an ROC-AUC of 0.807 and outperformed four immunogenicity-prediction tools based on HLA-II epitopes (Figure 4e). Residue-level indicators, including the longest non-human segment and the proportion of non-human residues, were also highly correlated with ADA, supporting the role of local non-human segments as important drivers of immunogenicity rather than relying only on whole-chain similarity (Extended Data Table 2).

Figure 4. Correlation between IgScope scores and clinical ADA incidence.

Extended Data Table 2. Relationship between IgScope scores and clinical ADA incidence.
Two core application scenarios address practical antibody-development challenges
Scenario 1: Auditing de novo AI-generated antibodies and identifying high-score traps
The research team evaluated 3,359 antibody sequences generated by four mainstream antibody-generation models: BoltzGen, IgGM, ODesign, and RFantibody. Whole-chain humanness scores for AI-generated antibodies reached 0.966–0.988 (Figure 5d), making them nearly indistinguishable from naturally human antibodies. However, all four models produced CDR scores significantly lower than those of natural human CDRs (P<1e-30; Figure 5a–c). In other words, using only conventional chain-level scoring could allow many AI-designed antibodies with potential non-human CDR risks to pass, whereas IgScope regional scoring can identify this defect and add a critical audit checkpoint to AI antibody design.

Figure 5. Regional humanness characteristics of de novo designed antibodies.
Scenario 2: Fine-grained quality control for fully human antibody mouse platforms
Cyagen Biosciences’ self-developed HUGO-Ab fully human antibody mouse platform products—including HUGO-Mab fully human monoclonal antibody mice, HUGO-Light fully human common-light-chain antibody mice, and HUGO-Nano chVHH camelized human nanobody mice—are important antibody-discovery tools. Their variable domains are designed to be human overall. To maintain VHH nanobody stability, however, the FR2 region intentionally retains four camelid characteristic amino acids, known as the FERF quartet (Figure 6e).
For humanized mouse sequences, both whole-chain scores and scores across the seven FR regions were close to those of natural human sequences, except for the relatively low FR4 score of HIS-Mouse light chains. These results were significantly better than those of de novo AI-generated antibody sequences (Figure 6a–c).
For HUGO-Nano chVHH camelized human nanobody mouse sequences, whole-chain scores remained close to fully human scores, while conventional tools completely failed to detect the four intentionally retained non-human positions (Figure 6c).
IgScope regional scoring accurately localized this signal. In HUGO-Nano chVHH camelized human nanobody mouse sequences, only the FR2 score decreased significantly, while all other regions maintained high scores. When the four characteristic residues were specifically masked, the FR2 score immediately returned to the human level (Figure 6d).
These results demonstrate that IgScope can support fine-grained quality control of sequences produced by humanized mouse platforms, distinguish genuine humanness defects from non-human positions intentionally retained for engineering purposes, and provide a computational quality-control benchmark for antibody sequences generated by the HUGO-Ab fully human antibody mouse series.

Figure 6. Humanized variable domains obtained from a humanized mouse platform.
Model limitations and outlook
The paper also describes the boundaries of the model:
If a CDR motif is completely identical between human and mouse germline genes, the region-local model has a theoretical upper limit for identification (Figure 2e–f).
The current model processes a single antibody sequence and does not consider heavy- and light-chain pairing, the VH–VL interface, or three-dimensional structural information.
Clinical ADA endpoints are affected by multiple confounding factors, including dose, dosing regimen, and detection method. Sequence prediction is a computational hypothesis and still requires in vitro and in vivo validation.
In the future, IgScope will be integrated into AbSeek® intelligent antibody computational platform. It will support humanization optimization and risk auditing of AI-generated antibodies, while real sequencing data from the HUGO-Ab fully human antibody mouse platform platform will help connect AI computational prediction with antibodies produced in vivo. This closed loop can support early developability screening of therapeutic antibodies and help reduce the risk of late-stage clinical failure caused by immunogenicity. If you are interested in IgScope multi-resolution antibody humanness analysis, please Contact us to discuss technical support and collaboration.
Publication information
Title: Multi-resolution humanness scoring of antibodies via a region-isolated language model


