Prediction of antibody non-specificity using protein language models and biophysical parameters

ABSTRACT

The development of therapeutic antibodies requires optimizing target binding affinity and pharmacodynamics, while ensuring high developability potential, including minimizing non-specific binding. In this study, we address this problem by predicting antibody non-specificity by two complementary approaches: (1) antibody sequence embeddings by protein language models (PLMs) and (2) a comprehensive set of sequence-based biophysical descriptors. We benchmark the PLM embeddings against interpretable, sequence-derived biophysical descriptors and use fragment-specific models (variable heavy (VH) and variable light (VL) chains, concatenated regions, and individual complementary-determining regions (CDRs) to identify region-level determinants of non-specificity in functional antibodies that bind defined targets. These models were trained on previously published human and mouse antibody data and tested on three public datasets. We show that non-specificity is best predicted from the VH domain and heavy-chain CDRs. These region-level analyses show that heavy-chain features, especially H-CDR3, dominate the predictive signal. The top performing PLM, a VH domain-based Evolutionary Scale Modeling 1 v LogisticReg model, resulted in 10-fold cross-validation accuracy of up to 71%. While predictive accuracy is comparable to classical machine-learning baselines, PLM-based embeddings provide sequence-context representations that yield consistent region-level attribution and provide complementarity to the classical models by enhancing the reliability of predicted non‑specificity scores when used in combination. Our biophysical descriptor-based analysis identified the isoelectric point as a key driver of non-specificity, consistent with previous reports. Our findings highlight the importance of biophysical properties in predicting antibody non-specificity and highlight the potential of PLMs for the development of antibody-based therapeutics. These conclusions are robust to alternative class definitions and are not driven by VH-VL mutational bias. We illustrate the generalizability and practical use of the PLM approach by extending it to therapeutic antibodies and nanobodies, providing a tool for early-stage developability assessment, and we make the models publicly available.

Introduction

Monoclonal antibodies (mAbs) continue to be one of the leading drug modalities in the pharmaceutical industry. The success of mAbs for therapeutic application is the result of advances in in vivo and in vitro discovery platforms, which have enabled fast generation of high-affinity binders toward a highly diverse set of targets.^1,2 Recently, de novo design has gained increasing interest in the field as a third-generation discovery approach with the potential of substantially accelerating drug discovery and development timelines.^3,4 Moreover, with advances in mAb engineering, there has also been an increased interest in the development of mAbs with ultra-high target affinity (pM to fM),⁵ context-dependent target binding (e.g., pH-dependent^6,7 or ligand-induced target binding⁸), and various multi-specific functionalities.⁹ To reach optimal binding affinity and potency, mAb hits identified during discovery are often subjected to comprehensive screening campaigns using display platforms (libraries of 10³ to 10¹⁰ variants) and recombinant well-plate variant generation workflows (libraries of 10² to 10⁴ variants).¹⁰

When selecting an antibody lead candidate for development toward clinical testing, optimal target binding affinity and/or pharmacodynamics are key selection parameters. In addition, in the past decade there has been an increased focus on the importance of progressing antibodies that also possess a good developability potential to clinical stage. Antibody developability requires the intersection of multiple disciplines, in which diverse parameters such as expression levels, immunogenicity, processability, and formulation feasibility are addressed, to ensure optimal potential for successful clinical development of lead candidates. Non-specific binding, i.e., weak non-covalent interactions with off-target molecules or interfaces, has emerged as one of the key developability parameters to increase the chance for clinical success.^11,12 Specifically, several studies have reported that high tendency for non-specific binding can translate into faster in vivo clearance, thereby compromising pharmacokinetics.^13–20 Most recently, developability assays measuring non-specific interaction, self-association, and FcRn binding interaction have been shown to correlate most strongly with clearance.²⁰ Furthermore, there is an inherent risk that non-specific interactions can translate into undesirable side-effects.²¹ Non-specific binding is not a rare phenomenon, as recent reports suggest the presence of a trade-off between affinity and specificity. Thus, optimization of affinity and potency comes with an inherent risk of compromising target specificity.^22–25

Given the high level of interest in measuring non-specific interactions, several in vitro screening assays are available for this purpose. A commonly used assay is the enzyme-linked immunosorbent assay (ELISA) with a panel of common antigens, typically insulin, DNA, albumin, cardiolipin, and lipopolysaccharide (LPS).^12,25 Initially, these biomolecules were studied as model antigens for autoimmune responses and diseases. For example, insulin is a self-antigen for autoantibodies associated with type 1 diabetes.^26,27 In addition, DNA, albumin, and cardiolipin are self-antigens for autoantibodies associated with diseases such as systemic lupus erythematosus^28,29 and anti-phospholipid syndrome,³⁰ and LPS is an antigen for immune responses to bacterial infections.³¹ Moreover, ELISA is widely used in immunology, where non-specific antibodies are often referred to as poly-reactive. Such antibodies are characterized as having low-affinity binding to multiple distinct antigens, including self-antigens, and they have been widely studied for targets such as HIV and influenza viruses, as they can be broadly neutralizing.^32,33 While this feature is beneficial for immunity to infectious diseases, to potentially confer broad protection against viruses, it is a highly undesirable feature for therapeutic mAbs. Other common non-specificity assays include baculovirus particle (BVP) ELISA,³⁴ poly-specific reagent (PSR),³⁵ and cross-interaction chromatography with ligands such as heparin¹⁵ and human IgG from serum.¹⁴

In addition to these in vitro assays, in silico methods have been gaining interest, and several tools have been reported recently for prediction of non-specific interactions.^36–39 The development and implementation of predictive computational methods for the prediction and re-design of mAb non-specificity at an early stage is of great interest, as it facilitates the generation of safe and efficacious lead candidates with high developability potential. Several studies reported the identification of non-specificity by in silico approaches. Short, linear sequence motifs (e.g., GG, RR, VG, VV, YY, WW and WxW, where × can be any amino acid) enriched in non-specific antibodies, as reported by Kelly et al.²³ have been utilized to create synthetic antibody libraries free from such motifs in the CDRs.⁴⁰ Moreover, artificial intelligence/machine learning (AI/ML) models that classify non-specific antibodies, leveraging experimental data and sequence-based information, have been reported. Boughter et al.³⁶ developed a classifier to identify non-specific antibodies based on experimental data acquired from ELISA with a panel of common antigens. Harvey et al.³⁷ developed a one-hot LogisticReg model based on a naïve Nb library assessed by the PSR assay.

In this study, we developed ML models to estimate the non-specificity of antibodies (Figure 1A). Commonly used sequence-derived biophysical properties were tested alongside protein language models (PLMs) to embed antibody sequences. PLMs have emerged as powerful tools for extracting informative features from raw protein sequences by leveraging patterns learned from massive sequence databases.^41–44 However, much of the reported success in antibody specificity prediction relies on supervised fine-tuning, antibody-tailored representations, and/or antigen-specific datasets, e.g., data arising from a single well-defined antigen, which is distinct from zero-shot application of a general PLM.^45–47 In addition, PLM-based supervised models have been developed to predict assay-based measures of polyreactivity.⁴⁸ Most recently, Amini et al.⁴⁹ applied PLMs to predict antibody developability liabilities and demonstrated that PLM-derived features can capture non-specificity measured by the PSR assay across a historic dataset spanning 33 therapeutic antibody programs. Finally, recent benchmarking studies highlight limitations of sequence-only likelihoods and indicate that incorporating structural context – and, for binding tasks, explicit Ab-Ag information – can improve predictive performance.^50,51

Figure 1.

Performance evaluation of machine learning models for predicting antibody non-specificity. (A) schematic workflow of the study. Publicly available datasets containing antibody sequences were used. These sequences were annotated and embedded using sequence-based biophysical descriptors and protein language models (PLMs). Different ML models were trained and evaluated using k-fold cross-validation, sensitivity-specificity analysis, and external datasets. (B) histogram showing the distribution of antibody sequences based on the number of flags in the Boughter dataset. Sequences are categorized as influenza-reactive (blue), HIV non-reactive (red), and HIV-reactive (light blue). (C) Bar plots showing the validation performance (k-fold CV and Leave-one- family-out) for a top performing ML algorithm (LogisticReg) and PLMs across various validation schemes (k-fold CV and Leave-one-family-out). (D) Bar plot of 10-fold CV accuracy for different antibody sequences embedded by top performing language model (ESM 1 v mean-mode). (E) histogram of predicted probabilities of antibody non-specificity using the VH-based logistic regression (ESM 1 v) model for the Jain dataset. Antibodies are classified into non-specific (dark blue), mildly non-specific (light blue), and specific (red) categories. (F) boxplot comparing the predicted non-specificity probabilities for antibodies across different datasets using VH-based logistic regression (ESM 1 v). The boxplot displays the median, interquartile range, and outliers, with significant differences indicated by Spearman’s correlation coefficient (SCC) and p-values (*** indicate p-value < 0.001).

Despite these advances, several gaps remain. First, most PLM-based studies for antibody property prediction have focused on antigen-specific binding or developability parameters other than non-specificity. Second, a systematic comparison of PLM embeddings with interpretable biophysical descriptors for non-specificity prediction is lacking. Third, the relative contribution of different antibody regions to non-specificity prediction using PLMs has not been comprehensively evaluated for antibodies raised against a specific target. A previous study⁵² reported on non-specificity ML models, but the training data was limited to single-chain variable (scFv) domains from naïve repertoires that were explicitly enriched for binding to poly-reactivity proxy reagents, such as ovalbumin, insulin, soluble cytosolic proteins, and soluble membrane proteins, with no target specificity. In this study, we address these gaps by systematically benchmarking state‑of‑the‑art PLMs, e.g., Evolutionary Scale Modelling (ESM), against sequence‑based biophysical descriptors for the prediction of antibody non‑specificity, while explicitly dissecting the contribution of individual antibody regions to model performance. ESM models, such as ESM 1 v, encode sequences into high-dimensional embeddings that reflect residue context, conservation, and evolutionary information, which are all factors known to influence antibody behavior. These features make ESM models well suited for predicting complex properties like non-specificity, which can arise from subtle sequence-dependent effects not easily captured by traditional descriptors. By applying ESM models to antibody variable regions, we aim to harness its representation power to identify sequence signatures of non-specific binding and improve early-stage developability assessment. Besides testing which encoding provided the best prediction performance, an additional objective of this study was to determine which regions of the antibody sequence contribute most strongly to non‑specificity. Furthermore, to gain mechanistic and biophysical insight, we analyzed a comprehensive set of 68 sequence‑derived biophysical descriptors alongside the PLM‑based models. Together, these complementary approaches enable the prediction of non‑specific interactions and provide interpretable guidance for the design and selection of antibody candidates with improved specificity and developability.

Results and discussion

Public antibody data

Four different datasets were retrieved from public sources; (1) a curated dataset of > 1000 mouse IgA, influenza-reactive and HIV-reactive antibodies with their respective non-specificity flag from ELISA with a panel of common antigens³⁶ (Figure S1A-C), (2) 137 clinical-stage IgG1-formatted antibodies with their respective non-specificity flag from ELISA with a panel of common antigens¹² (Figure S1D), (3) 398 antibodies, originating from naïve, IgG memory and long-lived plasma cells, with their respective poly-specific reagent score⁵³ (Figure S1E), and (4) 140 000 nanobody (Nb) clones assessed by the PSR assay from a naïve Nb library.³⁷ These four datasets are referred to as the Boughter, Jain, Shehata, and Harvey datasets, respectively. As therapeutic antibodies are engineered to be closely related to human antibodies to avoid immunogenic responses, it is important to exploit human antibody data for development of optimal ML models. As the Boughter dataset partly consists of mouse IgA antibodies, the sequence similarity of these mouse antibodies was compared to human antibodies to ensure that sequence differences within the dataset are not too large. The mouse IgA antibodies appear to differ mostly in the H/L-CDRs (Figure S1B). Another notable difference within the Boughter dataset is that the mouse IgA antibodies have a slightly shorter H-CDR3 loop relative to the human antibodies (Figure S1C).

Ideally, training data sets for classification ML-Models should be balanced when it comes to positive and negative data points. Our evaluation of the machine learning models for predicting antibody non-specificity is shown in Figure 1. An outline of the schematic workflow of the study is depicted in Figure 1A. The distribution of non-specificity for the three datasets is visualized in Figure 1B. The Boughter and Jain datasets are relatively balanced in terms of specific (zero flags), mildly non-specific (1–3 flags) and non-specific (>4 flags) antibodies, while the Shehata dataset is unbalanced, with 7 of 398 antibodies characterized as non-specific only. In this study, the most balanced dataset (i.e., Boughter) was selected for training of ML models, while the remining three (i.e., Jain, Shehata, and Harvey, which consists exclusively of VHH sequences) were used for testing.

Protein language models enable the representation of antibody non-specificity

Following the study original study,³⁶ the Boughter dataset was first parsed into two groups: specific group (0 flags) and non-specific group (>3 flags), leaving out the mildly non-specific antibodies (1–3 flags) (Figure 1A). The amino acid sequences of the parsed dataset were then annotated in the CDRs, and various fragments of the antibody sequences were embedded into vectors representing their physico-chemical and structural properties (i.e., ESM 1b, ESM 1 v, ESM 2, Protbert bfd, AntiBERTy, and AbLang2). This procedure resulted into the training of 16 different antibody fragment-specific binary classification models was trained (see Methods section). The performance of all the generated models from 10-fold cross-validation (CV) can be seen in Figure S2-S7. Overall, all the PLMs performed well with 66–71% 10-fold CV accuracy, including the antibody-specific ones AntiBERTy and AbLang2 (Figure 1C). These deep learning models were trained on large datasets of protein sequences in the million-to-billion range, encoding protein sequences into vectors for representation of their physiological properties, remote homology, and secondary/tertiary structure. PLMs were originally developed for the prediction of protein contacts and structure.^54,55 Going forward, the ESM 1 v was selected as the embedder of choice for this study based on its overall performance and generalizability. Across the 16 antibody fragment input types, ESM 1 v consistently achieved the highest or near‑highest accuracy and showed the most stable balance between sensitivity and specificity (see Figures S2–S7 for per‑model comparisons). ESM 1 v is also a general‑purpose PLM trained on a very large and diverse protein sequences, which supports its applicability beyond conventional antibody formats. The antibody‑specialized PLM, AbLang2, was released during the course of this work and was included in our benchmarking; AbLang2 is competitive and results in accuracies comparable to ESM 1 v for several VH‑based inputs. While AbLang2 is a viable alternative for antibody‑focused applications, we retained ESM‑1 v as our embedder of choice due to its generalizability.

To select a primary ML algorithm, we systematically evaluated all combinations of embeddings (PLMs and biophysical descriptors; descriptor-based models are described below) with multiple classifiers rather than choosing one a priori. Across both 10‑fold CV and Leave‑One‑Family‑Out evaluation, logistic regression performed comparably to, or better than, Random Forest, XGBoost, and Gaussian Process models for most fragment inputs (see Figures S2-S8 for full comparisons of PLM‑ and descriptor‑based models). We selected logistic regression because it: (1) delivered robust performance across validation schemes, (2) provides interpretable coefficients that help identify biophysical drivers of non‑specificity, and (3) is less prone to overfitting on moderate‑sized, high‑dimensional datasets such as ESM 1 v embeddings. Using the same classifier for both PLM‑ and descriptor‑based models also removes classifier choice as a confounder, enabling a fair and direct comparison of input representations.

The highest PLM-based predictability is achieved by encoding the VH domain

An overview of the different antibody fragment-specific models based on the ESM 1 v embedder is shown in Figure 1D. The highest predictability (71% 10-fold CV accuracy of non-specificity) was obtained for the models trained on the VH and H-CDRs sequences. These results suggest that the non-specificity primarily originates from the VH domain, with main contributions from the H-CDR loops. When looking at the models based on the individual H-CDR loops, the order of low-to-high predictability of non-specificity follows H-CDR2, H-CDR1, and H-CDR3 (Figure 1D). H-CDR3 has the highest predictability of non-specificity among all the H-CDR loops. The importance of the H-CDR3 loop for non-specificity is in agreement with Guthmiller et al.³³ who showed by using molecular dynamics simulation that the flexibility of the H-CDR3 loop plays an important role in the non-specific behavior of antibodies. Moreover, Chen et al.⁵² reported on AI/ML models to predict antibody non-specificity from human naïve scFv repertoires that were explicitly enriched for binding to so‑called non‑specificity ligands, i.e., poly-reactivity “proxy” reagents, and that have no known target binding. In this setting, the antibody pool is deliberately biased toward ligand reactivity and does not have known target binding. Accordingly, the resulting models primarily learn sequence features associated with binding to these proxy ligands rather than the broader, developability‑relevant non‑specificity that emerges in target‑directed discovery and optimization. This paradigm differs from approaches that start from antibodies raised or selected against a defined target and then quantify their general off‑target binding propensity, an approach taken in the original study of Boughter et al.³⁶ and our study, and the distinction should be considered when comparing reported model performance and interpreting what aspect of “non‑specificity” is being captured. Nevertheless, our results align with the central conclusion of Chen et al.⁵² that non-specificity originates predominantly from the heavy chain, particularly the VH and H-CDRs, while we demonstrate this explicitly in the context of antibodies that have been raised toward a target.

Accuracy of around 70% was consistently observed across 3, 5, and 10-Fold CV for the top performing models (Figure 1C), and similar performance obtained for sensitivity and specificity. Moreover, when looking at the predictability of one antibody family to another, the overall accuracy was consistently above >60% for the Leave-One Family-Out validations. However, when comparing sensitivity and specificity, classifiers trained on human antibodies performed poorly when tested on mouse antibodies. This is not surprising, as mouse and human antibodies have notable sequence differences, such as mouse IgA having a shorter H-CDR3 and larger sequence differences in the CDRs relative to human antibodies (Figure S1B). Moreover, classifiers trained on mouse IgA and HIV-reactive antibodies perform well across all evaluation metrics (accuracy, sensitivity, and specificity) when tested on influenza-reactive antibodies, while classifiers trained on mouse IgA and influenza-reactive antibodies seem to be better in predicting non-specific HIV-reactive antibodies than specific ones.

To investigate whether germline-level sequence differences could explain the higher performance of VH‑ and H-CDRs-based descriptors, we computed V‑region identity and amino‑acid mutations from the closest human germline for both VH and VL across the Boughter, Jain, and Shehata datasets (Figure S20A-D). Across the different datasets, VH carried a modestly higher mutational load than VL (median VH-VL difference 3–4 residues; mean VH-VL difference 2–4), and these VH-VL differences were statistically significant (paired t‑test and Wilcoxon tests; Figure S20E, G and I). However, this small VH bias does not explain the classifier results: within VH, higher mutation counts did not appear to consistently associate with non‑specificity defined by non-specificity ELISA or PSR assays (Figure S20E, G and I). We also observed modest enrichment of specific germline V‑genes, such as VH1‑69, among non‑specific antibodies in all three datasets (Figure S20F, H, and J), consistent with the Boughter et al. study,³⁶ although germline usage alone remains insufficient to distinguish specific from non‑specific antibodies. Furthermore, the modest enrichment of VH mutations aligns well with affinity maturation, where VH, particularly the H‑CDR loops and most notably H‑CDR3, undergoes more extensive diversification than VL. This pattern underscores the dominant role of VH in antigen recognition and provides a biological explanation for the higher mutational load typically observed in VH relative to VL. This observation aligns with Akbar et al.,⁵⁶ who reported that H‑CDRs contribute more substantially to antigen binding than L‑CDRs, further supporting the asymmetric mutational patterns observed. Taken together, these results indicate that while VH accumulates slightly more somatic variation (particularly in H‑CDRs and H‑CDR3), the higher predictive power of VH‑based descriptors is unlikely to be explained solely to mutation counts. Instead, it likely reflects sequence‑ and gene‑specific determinants within the VH sequence.

Classification probability of non-specificity against non-specificity ELISA flags mimics regression behavior

A prediction probability of non-specificity was computed in addition to the binary output from the binary classification models. When comparing model-predicted probabilities of non-specificity for the held‑out test antibodies from the Boughter dataset (samples from the 10‑fold cross‑validation), the conditional distributions of predicted probability for the predefined specificity classes (specific, mildly non‑specific, non‑specific) are ordered on average (Figure 1E). That is, mean predicted probability increases from the specific to the non‑specific class; however, substantial overlap between these distributions indicates imperfect separation and limited discriminative power. The premise that non-specificity is not a binary property is exemplified by the overlaps between the distributions. This illustrates that the prediction probability can be used beyond the binary output to assess antibodies of varying degree of non-specificity.

When comparing those to non-specificity ELISA flags, a significant regression-like behavior (Spearman’s correlation coefficient (SCC) 0.43) was observed for one of the top performing classifiers, ESM 1 v mean-mode VH-based LogisticReg model (Figure 1F) . The prediction probability for non-specificity followed an uptrend when compared to the non-specificity ELISA flags. An exception to this trend was the antibodies with seven non-specificity ELISA flags, as those were exclusively mouse IgA antibodies (differences discussed above).

The isoelectric point: a key biophysical driver of non-specificity

To gain insight into the biophysical origins of antibody non-specificity, a set of 68 sequence descriptors (Table S1) was computed for the parsed Boughter dataset. These descriptors encompass a wide range of biophysical properties derived from the antibody sequence, including theoretical isoelectric point (pI), secondary structure propensity, and hydrophobicity. To assess the presence of redundancy among the descriptors, we constructed a Spearman’s correlation matrix, which revealed that several descriptors, such as hydrophobicity descriptors, exhibited strong correlation among each other (SCC > 0.5, Figure S9), thus indicating redundancy. For consistency and interpretability, we used Logistic regression for the descriptor‑based models as for the PLM-based models. Using the same classifier isolates differences due to the input representation (PLM vs descriptors) for a fair comparison. Empirically, Logistic regression performed comparably to more complex classifiers (Random Forest, XGBoost, and Gaussian Process) across most settings (Figure S8). It also yields directly interpretable coefficients for descriptor analysis and is less prone to overfitting on our moderate‑sized, high‑dimensional datasets. All the descriptors were ranked according to the absolute logistic regression coefficients (Table S2), whereafter top 25 descriptors were selected, and used for training of a VH-based LogisticReg model. The in-depth analysis of the importance of the 25 descriptors is shown in four different plots in Figure 2A:

• The first plot shows the LogisticReg coefficients, indicating the relative importance of each descriptor. Notably, Disorder_Propensity_DisProt, Aggrescan_a4v, and theoretical pI show significant positive coefficients, suggesting that they are strong drivers of non-specificity.

• The second plot displays the permutation importance, highlighting the change in accuracy when each descriptor is permuted. Descriptors like theoretical pI, bulkiness, Hplc_Hfba_retention, and Polarity_Zimmerman demonstrate substantial decrease in accuracy upon permutation.

• The third plot illustrates the accuracy of models based on single descriptors to underscore their individual predictive power. Theoretical pI resulted in the highest accuracy compared to the other descriptors, confirming its critical role in the prediction of non-specificity.

• The fourth plot shows the Leave-One-Feature-Out accuracy, revealing how the exclusion of each descriptor affects the overall model performance. Most of the descriptors result in a minimal drop in accuracy, indicating that the model performance remains unaffected when a certain descriptor is left out. This can be explained by the remaining redundancy among the 25 descriptors, e.g., theoretical pI appears to be negatively correlated with Polarity_Zimmerman, according to the Spearman’s correlation matrix shown in Figure 2B.

Figure 2.

Analysis of descriptor importance and model performance for VH-Based logistic regression. (A) analysis of descriptor importance using various metrics for the VH-based logistic regression model; (first panel) logistic regression coefficients indicating the relative importance of different features, (second panel) permutation importance showing the change in 10-fold CV accuracy when each descriptor is permuted, (third panel) 10-fold CV accuracy of models based on single descriptors, and (fourth panel) 10-fold CV accuracy of Leave-one-feature-out models compared to the mean accuracy of model using the top 25 descriptors. (B) heatmap displaying the SCC between the top 25 descriptors selected based on highest logistic regression coefficient in model with all descriptors. The dendrogram shows hierarchical clustering of descriptors based on their SCC. (C) Bar plot comparing the 10-fold CV accuracy of different models in predicting antibody non-specificity across various validation schemes: k-fold CV and Leave-one-family-out validation. Models compared include VH-based sequences embedded by ESM 1 v, all descriptors, PCA with 3, 5 and 10 components, and top 2, 3, 4, and 5 descriptors. Sensitivity and specificity bar plots can be found in Figure S12.

To further narrow the redundancy, the 25 descriptors were tested in all possible combinations of 2, 3, 4, and 5 descriptors for training of new LogisticReg models. The results of the top descriptors from this analysis are shown in Table 1. The results indicate that, among the top 5 descriptors, the theoretical pI appears to be the most important driver for non-specificity. This conclusion is also supported by the frequency of this particular descriptor among the top models (Figure S10). Additionally, as a parallel check, we performed Principal Component Analysis (PCA) for dimensionality reduction and feature selection. The primary objective of this study was again to address multicollinearity among the descriptors and to identify the most significant features contributing to the variance in the dataset. We thus evaluated the performance of the LogisticReg models trained on the 3, 5, and 10 principal components identified by PCA. In agreement with our previous findings regarding the theoretical pI, the presence of this descriptor among the selected features significantly influenced the model performance, particularly when it was included in PCA 5 and 10 components (Figure 2C). The importance of theoretical pI was captured at ≥5 components, as seen by the magnitudes of Eigenvalues of theoretical pI for PCA 5 as compared to PCA 1–4 (Figure S11). The pI is known to influence pharmacokinetic and pharmacodynamic behavior/clearance of antibodies.^57,58 This finding is consistent with the experimental work of Gupta et al.⁵⁹ that showed that IgG1 antibodies with weakly basic pI 8–8.5 display a combination of repulsive self-interactions in low ionic strength formulations at pH 5–6 and weak non-specific interactions under physiological conditions. While Gupta et al.⁵⁹ characterized this effect experimentally on 42 IgG1 variants, we computed 68 sequence-derived descriptors and performed exhaustive combinatorial testing (a total of 68 380 combinations of top 2, 3, 4, and 5 descriptors from 25 selected descriptors, see Figure S10), identifying pI as the single most important predictor among 68 candidates (65.2% single-descriptor accuracy, see Table S2), providing a systematic and comprehensive computational corroboration of their experimental observation.

Table 1.

Top 2, 3, 4, and 5 combined VH-based sequence descriptors.

Top	Descriptors
2	Theoretical pI, Bulkiness
3	Theoretical pI, Bulkiness, Aggrescan_Nr_hotspots
4	Theoretical pI, Average_Flexibility_BP, Beta_Turn_Chou_Fasman, Hplc_Hfba_Retention
5	Theoretical pI, Bulkiness, Disorder_Propensity_DisProt, Percentage_Acessible_Res, Aggrescan_av4

Altogether, a comparison between the PLM-based and descriptor-based ML models in terms of accuracy of across different validation schemes is shown in Figure 2C. The results indicate that both the ESM 1 v and all descriptors VH‑based Logistic Regression models achieve comparable accuracy across the different validation schemes. Notably, the PCA-based models show comparable performance, demonstrating the effectiveness of dimensionality reduction in maintaining model accuracy. Interestingly, the VH-based Logistic Regression models using the top descriptors, 2, 3, 4, and 5 combinations (Table 1) exhibit robust performance. This finding suggests that a smaller subset of key descriptors can achieve similar predictive power as using the full set of descriptors, highlighting the potential for model simplification without compromising accuracy.

To explore whether combining the PLM‑ and descriptor‑based models would yield improved performance, combined models were generated in which the individual VH or VH/VL‑based embeddings from both approaches were concatenated and used to train Logistic Regression classifiers. As shown in Figure S23, combined PLM-descriptors representations achieve performance comparable to the single‑model classifiers, while providing more robust and stable prediction probabilities across datasets. These results indicate that integrating complementary sequence‑derived features does not markedly increase overall accuracy but can enhance the reliability of predicted non‑specificity scores.

VH-based LogisticReg classification model is applicable to clinical-stage therapeutic antibodies

To show applicability of the non-specificity classification model on therapeutic antibodies, the ESM 1 v mean-mode VH-based LogisticReg model was tested on the Jain dataset. As in the Boughter dataset, the Jain dataset was parsed into two groups, specific (0 flags) and non-specific (>3 flags), leaving out the mildly non-specific antibodies (1–3 flags). An accuracy of 69% was obtained for the parsed Jain dataset (see confusion matrix in Figure S14A). This value is comparable to the mean accuracy of 71% obtained for the same classifier across 3, 5, and 10-Fold CV for the parsed Boughter dataset. Moreover, as in the case of the Boughter dataset, a similar distribution of prediction probability of non-specificity was obtained for the full Jain dataset, and it appears to mimic regression-like behavior when compared to the non-specificity ELISA flags, although to a slightly weaker extent (Figure 3A and S13). The same trend can be observed for the top 5 descriptors model (Figure 3B). It is important to note that the Jain dataset comprises clinical‑stage IgG1 antibodies that have already undergone developability selection, and therefore likely exhibit milder non‑specificity compared with discovery‑stage datasets such as Boughter and Shehata datasets. Nevertheless, the observed monotonic relationship between predicted probabilities and ELISA flag counts (Figure S13) indicates that the model captures gradations of non‑specificity risk even within a well‑selected clinical population. Overall, these results demonstrate that the classifier is applicable to therapeutic antibodies.

Figure 3.

Logistic regression models predicting antibody non-specificity across different datasets. (A-F) distributions of predicted probabilities of antibody non-specificity for three different datasets using two logistic regression models: predictions for the Jain dataset (A,B), for the Shehata dataset (C,D), and for the Harvey dataset (E,F). (A, C, and E) depict results from the ESM 1 v VH-based logistic regression model, while (B, D, and F) depict results from the top 5 descriptors VH-based logistic regression model. For each dataset, antibodies are classified into specific, mildly non-specific (only in the Jain dataset), and non-specific categories, represented by different colours.

Antibodies characterized by the PSR assay appear to be on a different non-specificity spectrum than that from the non-specificity ELISA assay

Alternative assays to ELISA have recently been developed to meet the demand of high-throughput screening during drug discovery, and such one is the PSR assay,³⁵ where antibodies displayed on the surface of yeast cells are counter-selected when non-specifically bound to soluble membrane protein in a flow cytometry-setup. Several studies have been reported using this assay for assessing antibody non-specificity.^12,14,53 Recently, Harvey and coauthors³⁷ developed a one-hot LogisticReg model based on >140 000 clones assessed by the PSR assay from a naïve Nb library. They found a significant correlation of the PSR with the gold-standard ELISA based on six Nbs. To find out whether our ESM 1 v mean-mode VH-based LogisticReg model can extend its applicability further to the non-specificity scored by the PSR assay, the Shehata dataset and the VH-based Nb dataset by Harvey and coauthors,³⁷ here referred to as the Harvey dataset, were tested. The classifier did not appear to separate the PSR-scored specific and non-specific antibodies well. All the specific PSR-scored antibodies of the Shehata dataset were distributed along the entire prediction probability scale, while the few non-specific ones were on the probability end toward higher non-specificity (Figure 3C,D). A similar forecast was observed for the Harvey dataset; all the specific PSR-scored Nbs resulted in a broad probability distribution, while the non-specific PSR-scored ones resulted in a narrower probability distribution toward higher non-specificity (Figure 3E,F). Thus, the classifier appears to be better at predicting non-specific PSR-scored antibodies, than specific PSR-scored antibodies. This result suggests that the spectrum of non-specificity from the PSR assay is different than the one from the non-specificity ELISA assay, of which the classifier is trained on. Thus, a specific antibody classified by the PSR assay may necessarily not translate into a specific antibody classified by the non-specificity ELISA assay. The specific PSR-scored Nbs could partly consist of mildly non-specific clones in addition to specific ones, thus resulting in this broad probability distribution.

An interesting remark can be made about the distributions of predicted probabilities obtained from the two different LogisticReg models tested on the Harvey dataset in Figure 3E,F. The ESM 1 v VH-based LogisticReg model produces a more uniform distribution of predicted probabilities across the dataset, while the top 5 descriptors VH-based LogisticReg model exhibits a clear biphasic distribution. It is no surprise that this bimodal pattern closely resembles the distribution of pI (Figures S15-S18), as this descriptor is the main driver of non-specificity in the top 2, 3, 4, and 5 descriptors LogisticReg models (Table 1 and Figure S10). Nonetheless, the distribution of non-specific antibodies in the Harvey dataset appears to exclusively be of high pI (>8) according to Figure S18A. The distributions of the other descriptors do not appear to differ significantly between specific and non-specific antibodies (see Figures S15-S18). The distinct separation suggests that the pI plays a crucial role in differentiating between specific and non-specific antibodies.

VH-based LogisticReg models performs on par or better than existing predictors

The ESM 1 v VH-based LogisticReg model can be compared to two existing predictors in the literature, the predictors reported in the Boughter et al. study,³⁶ and the Harvey et al. study.³⁷ Boughter and colleagues found noticeable differences between the gene usage frequency of poly-reactive and non-poly-reactive antibodies but concluded that the overlap in usage between the two populations suggests that gene usage alone is not sufficient to distinguish the groups. They showed to be successful in developing a binary classifier based on a position-sensitive biophysical matrix with accuracy up to 75%. Their reported performance is on par with the achieved performance of the PLM-based classifier (ESM 1 v VH-based LogisticReg), which was trained on the same data (parsed Boughter dataset).

Furthermore, Harvey and colleagues developed a one-hot LogisticReg model based on >140 000 clones assessed by the PSR assay from a naïve Nb library with an accuracy >80%. Using their published web-based predictor,⁶⁰ we tested its performance on the Boughter, Jain, and Shehata datasets. The results show that the Harvey predictor does not separate well the different antibody groups in the Boughter dataset (Figure S19A, B), with overlapping distributions of prediction scores. Similarly, the Jain and the Shehata datasets demonstrate significant overlap between specific and non-specific antibodies (Figure S19C-F), indicating some limitations for the Harvey method in predicting non-specificity of antibodies as compared to Nbs.

Prediction of non-specificity in antibody drug development programs

Consequences of non-specificity in the clinic: Non-specific binding of therapeutic antibodies can lead to significant adverse effects in the clinic. Such antibodies can bind to structurally unrelated off-targets and thereby potentially result in unwanted toxicity⁶¹ or reduced efficacy.²¹ They can also interact with tissues like subcutis and thereby result in faster clearance via pinocytosis independently of FcRn.^57,58,62 Ultimately, non-specificity can compromise the safety and efficacy of therapeutic antibodies, potentially resulting in clinical trial failures and increased development costs. Thus, early-stage prediction of non-specificity, such as during selection and optimization, is essential to reduce the risk of failing at later stages, including during clinical trials. Otherwise, the further into the development program, the harder it becomes to allow additional protein engineering to mitigate biophysical liabilities, as new in vitro and in vivo data must be reproduced.

A powerful strategy to address this problem is to combine in silico prediction with in vitro developability assessment. To identify and flag non-specific antibodies, we propose a combined strategy that integrates in silico prediction models with traditional in vitro developability assessments during the lead optimization stage in the drug development process. This hybrid approach leverages the strengths of computational predictions and experimental validations, ensuring the selection of lead candidates with high developability potential.

Practical recommendations for discovery: To support integration of our models for early screening and candidate selection, we offer the following practical recommendations for model choice, thresholding, and downstream actions.

• Model choice and usage: For rapid, high-throughput screening, we recommend the VH‑only PLM model (ESM‑1 v embeddings on VH sequence with logistic regression). This model requires only the VH amino‑acid sequence (often available earliest) and achieved the best overall performance in our cross‑validation experiments, making it suitable for large candidate screening. Candidates flagged as high‑risk by the VH‑PLM model should be evaluated with the descriptor‑based model. Descriptor outputs can highlight specific biophysical liabilities (e.g., extreme pI, exposed hydrophobic patches, or charge clusters) that can guide sequence optimization or experimental follow‑up.

• Probability thresholds: We recommend treating the model output as a probabilistic risk score, i.e., the predicted non‑specificity (p), rather than a binary decision. As a starting point (based on our cross‑validation results), the following thresholds can be used:

• p ≥ 0.70: high risk → de-prioritize or require strong justification (e.g., exceptional binding potency) before progressing. Prioritize experimental non‑specificity assays for these candidates.

• 0.30 ≤ p < 0.70: intermediate risk → proceed to in vitro non‑specificity assays (e.g., non-specificity ELISA, PSR assay, and AC‑SINS) and consider descriptor‑based models to identify modifiable liabilities.

• p < 0.30: low risk → continue standard developability testing pipeline.

It should be noted that these thresholds are intended as initial guidance. Teams should calibrate thresholds to their project priorities and operating characteristics, e.g., by benchmarking on historical data or choosing thresholds that give a desired false positive/negative trade‑off.

• How to integrate into workflows: Apply the VH‑PLM model immediately after sequence selection to remove the highest‑risk candidates before committing to expression and assays. For intermediate-to-low risk candidates, combine model predictions with descriptor diagnostics and run orthogonal in‑vitro assays. Use descriptor outputs to suggest targeted sequence edits, e.g., mutate to tune local charge, and re‑score candidate variants with the model.

• Reporting and interpretability: Report predicted probabilities and the model used, i.e., VH‑PLM or descriptor-based, rather than only binary calls. We recommend a two‑tier approach: automated computational triage followed by an experimental triage panel for candidates in the intermediate or high‑risk zones.

• Reproducibility: The analyses described here are reproducible. We have made available a GitHub repository (see link in Data Availability section below) with Jupyter notebooks that demonstrates how to compute embed sequences, run the VH‑PLM and descriptor models, and compute predicted.

Key limitations and future directions: Our study has several important limitations. First, the training datasets (primarily mouse IgA, HIV‑1 broadly neutralizing, and influenza‑reactive antibodies) do not fully represent the range of therapeutic formats such as human IgG1, bispecifics, and single‑domain antibodies (e.g., Nbs). In particular, Nbs have distinct structural and sequence features, such as absence of a light chain, longer and often more variable CDR3s, and different framework substitutions, that can reduce transfer accuracy relative to conventional antibodies. Thus, model generalizability to those formats may be limited. Second, part of the Boughter dataset is mouse‑derived; although VH sequence homology allows partial transfer of learned sequence-non‑specificity associations to human antibodies, species‑specific biases may degrade performance. In short, a model trained specifically on human IgG therapeutic datasets would likely yield more reliable clinical predictions.

Third, our outcome is a binary label (specific versus non‑specific), which inherently limits the ability to fully distinguish between the two classes. As shown in Figure 1E,F, the predicted probabilities exhibit overlap across the different degrees of non‑specificity. Nevertheless, a robustness analysis using alternative class definitions (Figures S21–S22) confirmed that the overall conclusions remain unchanged. Sensitivity and specificity shift as the boundary is adjusted, consistent with the graded and continuous nature of non‑specific binding. Fourth, while the best models achieve useful performance (around 71% accuracy in 10-Fold CV), this still implies non‑negligible false positive and false negative rates, and therefore necessitates experimental follow‑up for candidates flagged by the models. Fifth, non‑specificity was assessed using different assays across datasets, e.g., ELISA panel versus PSR, which complicates direct cross‑dataset comparisons and may introduce systematic differences in labels. Finally, although PLM embeddings provide strong predictive signal, they are inherently less interpretable than explicit biophysical descriptors, which can limit mechanistic insight.

These limitations point to several concrete next steps. We plan to expand training to larger and more diverse datasets (including human IgG‑format antibodies) as they become available, and to explore regression formulations that predict a continuous measure of non‑specificity rather than a binary class. Fine‑tuning PLMs on curated non‑specificity data may further improve discrimination, and integrating structure‑informed features, e.g., AlphaFold/IgFold‑derived solvent exposure or paratope geometry, alongside sequence embeddings may capture additional determinants of non‑specific binding.

Conclusions

In this study, we developed ML models to predict the non-specificity of antibodies, utilizing both PLMs and biophysical properties to embed antibody sequences. In agreement with previous reports on different datasets,⁵² our results indicate the VH domain, particularly the H-CDR loops, as the main contributor to non-specificity, and that the biophysical parameter pI is a key biophysical driver of non-specificity. The resulting computational models enable the prediction of non-specific interactions of antibodies with accuracy of 71% in 10-fold CV, thus providing a valuable tool to guide the design and selection of mAbs with improved specificity and efficacy. These findings have important implications for the development of safe and efficacious lead candidates with high developability potential.

Methods

Data sources

All antibody datasets used in this study were retrieved from public sources, as well as upon request from Prof. Debora Marks and Prof. Andrew Kruse (Harvey dataset), and Dr. Tushar Jain and Dr. Dane Wittrup (Jain dataset). A list of the datasets and their corresponding sources are reported Table 2.

Table 2.

List of public antibody datasets with their corresponding size, non-specificity assay, and reference.

Dataset	Size	Poly-reactive assay	Reference
Boughter dataset	>1000 antibodies (HIV-1 broadly neutralizing, influenza-reactive, IgA mouse) of varying degree of non-specificity	ELISA with a panel of 7 ligands (DNA, insulin, LPS, albumin, cardiolipin, flagellin, KLH)	36
Jain dataset	137 clinical stage IgG1-formatted antibodies	ELISA with a panel of 6 ligands (ssDNA, dsDNA, insulin, LPS, cardiolipin, KLH)	12
Shehata dataset	398 antibodies from naïve, IgG memory, IgM memory and long-lived plasma cells	Poly-specific reagent (PSR) assay	53
Harvey dataset	>140 000 naïve nanobodies	Poly-specific reagent (PSR) assay	37

Python programming

All coding was performed in Python using Spyder IDE and Jupyter Notebook (Anaconda software distribution),⁶³ and a list of used Python modules is reported in Table 3. Biophysical descriptors were computed using: (1) Schrödinger Suite (2023–3) for the majority of the descriptors (documentation available at: https://support.schrodinger.com/s/article/827119), and (2) Biopython ProteinAnalysis module for the descriptors marked with (*) in Table S1, specifically: Theoretical pI, Charge at pH 6, and Charge at pH 7.4. Germline sequence analysis was performed using RIOT.⁷⁶

Table 3.

List of software and python modules.

Module	Additional information	Reference
NumPy	https://numpy.org	64
SciPy	https://www.scipy.org	65
Statsmodels	https://www.statsmodels.org	66
Pandas	https://pandas.pydata.org	67
Matplotlib	https://matplotlib.org	68
Seaborn	https://seaborn.pydata.org	69
Scikit-Learn	https://scikit-learn.org	70
Json	https://docs.python.org/3/library/json	71
Collections	https://docs.python.org/3/library/collections	72
Itertools	https://docs.python.org/3/library/itertools	73
ANARCI	https://github.com/oxpig/ANARCI	74
Biopython	https://biopython.org/docs/1.76/api/Bio.SeqUtils.ProtParam.html	75
RIOT	https://github.com/NaturalAntibody/riot_na	76

Training and validation of binary classification models

First, the Boughter dataset was parsed into three groups as previously done in ³⁶: specific group (0 flags), mildly poly-reactive group (1–3 flags), and poly-reactive group (>3 flags). The primary sequences were annotated in the CDRs using ANARCI following the IMGT numbering scheme. Following this, embeddings were computed on the sequence fragments used as model inputs (fragment‑level embedding) by three state-of-the-art PLMs, ESM 1 v,⁵⁴ Protbert bfd,⁵⁵ and AbLang2,⁴² for representation of the physico-chemical properties and secondary/tertiary structure (Table 4). For example, VH embeddings were generated by passing only the VH amino‑acid sequence to the PLM; H‑CDR3 embeddings were generated by passing only the H‑CDR3 sequence. We did not compute embeddings from full‑length concatenated sequences and then slice the result. In total, 16 different antibody fragment sequences were assembled and embedded. Moreover, for the embeddings from the PLMs, mean (average of all token vectors) was used. The vectorized embeddings were served as features for training of binary classification models (e.g., LogisticReg, RandomForest, GaussianProcess, GradientBoosting, and SVM algorithms) for non-specificity (class 0: specific group, and class 1: poly-reactive group). The mildly poly-reactive group was left out from the training of the models. For model evaluation, we used stratified 10‑fold cross‑validation with a fixed random seed for reproducibility (see Figures S2-S8). All model fittings were performed within training folds to avoid information leakage. To obtain a more stringent assessment of generalizability across biologically distinct groups, we additionally ran a leave‑one‑family‑out evaluation of top models in which each antibody family from the Boughter dataset was withheld in turn as the test set, while models were trained on the remaining families (see Figure 2C).

Table 4.

Type and description of sequence input for the binary classification models.

Type of input	Description
VL, VH, L-CDR 1, L-CDR 2, L-CDR 3, H-CDR 1, H-CDR 2, H-CDR 3	A vectorised embedding of VL, VH or an individual CDR sequence
VH/VL joined, L-CDRs joined, H-CDRs joined, H/L-CDRs joined	A vectorised embedding of a joined sequence
VH/VL concat, L-CDRs concat, H-CDRs concat, H/L-CDRs concat	A concatenated embedding of embeddings of individual domains

The trained classification models were validated by: (1) 3, 5, and 10-Fold cross-validation (CV), (2) Leave-One-Family-Out validation, e.g., training on HIV and influenza-reactive antibodies, while testing on mouse IgA antibodies, (3) comparing probability of predicted poly-reactive class to true class, and (4) testing on the Jain dataset. The evaluation metrics included accuracy, sensitivity, and specificity (Eqs. 1-3).

$$\mathit{Accuracy}=\frac{\mathit{TP}+\mathit{TN}}{\mathit{TP}+\mathit{TN}+\mathit{FP}+\mathit{FN}}$$ (1)

$$\mathit{Sensitivity}=\frac{\mathit{TP}}{P}$$ (2)

$$\mathit{Specificity}=\frac{\mathit{TN}}{N}$$ (3)

Where TP is true positive (true poly-reactive), TN is true negative (true specific), FP is false positive (false poly-reactive), FN is false negative (false specific), P is all positives (all poly-reactive), and N is all negatives (all specific).

Funding Statement

Financial support from Novo Nordisk A/S is acknowledged. P.S. is a Royal Society University Research Fellow [grant no. URF\R\251013] and acknowledges funding from UK Research and Innovation (UKRI), and Engineering and Physical Sciences Research Council [grant no. EP/X024733/1]. The authors declare no competing financial interest.

Abbreviations

Ab

Antibody

Ab-Ag

Antibody – antigen

AC-SINS

Affinity-capture self-interaction nanoparticle spectroscopy

AI

Artificial intelligence

BVP

Baculovirus particle

CDR

Complementarity-determining region

CV

Cross-validation

dsDNA

Double-stranded DNA

ELISA

Enzyme-linked immunosorbent assay

ESM

Evolutionary Scale Modeling

FcRn

Neonatal Fc receptor

H-CDR1

Heavy-chain complementarity-determining region 1

H-CDR2

Heavy-chain complementarity-determining region 2

H-CDR3

Heavy-chain complementarity-determining region 3

H/L-CDRs

Heavy- and light-chain complementarity-determining regions

HIV

Human immunodeficiency virus

IgA

Immunoglobulin A

IgG

Immunoglobulin G

IgG1

Immunoglobulin G1

IgM

Immunoglobulin M

KLH

Keyhole limpet hemocyanin

L-CDR1

Light-chain complementarity-determining region 1

L-CDR2

Light-chain complementarity-determining region 2

L-CDR3

Light-chain complementarity-determining region 3

LPS

Lipopolysaccharide

LogisticReg

Logistic regression

LOFO

Leave-one-feature-out

mAb

Monoclonal antibody

ML

Machine learning

Nb

Nanobody

PCA

Principal component analysis

pI

Isoelectric point

PLM

Protein language model

PSR

Poly-specific reagent

RF

Random forest

SCC

Spearman’s correlation coefficient

scFv

Single-chain variable fragment

ssDNA

Single-stranded DNA

TN

True negative

TP

True positive

FN

False negative

FP

False positive

VH

Variable heavy chain/domain

VL

Variable light chain/domain

VHH

Variable domain of heavy-chain-only antibody

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data and models used in this publication are available at our GitHub repository: https://github.com/NovoNordisk-OpenSource/ML-predictions-of-Antibody-Non-Specificity.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2026.2678000