Variant Effect Prediction with Deep Mutational Scanning Data

Introduction

Deep mutational scanning (DMS) is a high-throughput experimental technique that systematically measures the functional consequences of all possible single amino acid substitutions in a protein of interest [1, 2, 3]. By coupling saturation mutagenesis with a functional selection and deep sequencing, DMS generates comprehensive fitness landscapes that quantify how each mutation affects protein activity, stability, binding affinity, or immune evasion [4, 5]. These empirical landscapes have become foundational for training and benchmarking computational variant effect predictors, particularly in the context of rapidly evolving viral surface proteins such as influenza hemagglutinin and coronavirus spike proteins [6, 7, 8]. In veterinary virology, DMS data enable the prospective identification of mutations that alter host range, antigenicity, or virulence, thereby informing surveillance and vaccine strain selection [9, 10, 1]. This article reviews the methodological principles of DMS, the computational frameworks that leverage these data for variant effect prediction, and their specific applications to animal pathogens.

Deep Mutational Scanning Methodology

A typical DMS experiment begins with the construction of a library of variants, often covering all single amino acid substitutions in a target domain [1, 2]. The library is expressed in a suitable system (e.g., yeast display, lentiviral particles, or cell surface expression) and subjected to a functional selection that mimics a biologically relevant phenotype, such as receptor binding, antibody neutralization, or viral entry [11, 12]. After selection, the relative abundance of each variant is determined by deep sequencing, and an enrichment score (often termed a fitness score) is calculated as the log ratio of the variant frequency in the selected versus unselected population [4]. These scores are normalized to wild-type and typically range from highly deleterious (negative) to beneficial (positive) [5].

Experimental noise and systematic biases (e.g., sequencing errors, library construction bottlenecks) must be accounted for. The DiMSum pipeline provides a rigorous error model that identifies common pathologies such as jackpot mutations and plate effects [4]. Imputation methods such as VEFill address missing scores due to incomplete coverage or dropout, using a deep learning model trained on protein domain embeddings to predict scores for unmeasured variants [13, 14]. Joint estimation frameworks that combine multiple DMS assays further improve the accuracy and completeness of functional effect maps [15].

Computational Prediction Approaches Using DMS Data

DMS data serve as ground truth for training and validating a wide range of computational variant effect predictors. Supervised learning models directly map sequence features to functional scores, with recent approaches using large-scale mutagenesis datasets from multiple proteins to improve generalization [16, 17]. Protein language models (pLMs) trained on evolutionary sequence data can be fine-tuned on DMS scores to predict variant effects with high accuracy, even for proteins not included in the training set [6, 18, 19]. For example, pLM embeddings have been shown to capture conservation and functional impact, outperforming traditional sequence-based predictors [19].

Substitution matrices derived from DMS data, such as DeMaSk, provide a quantitative measure of amino acid exchangeability that improves codon substitution models and variant impact prediction [20]. The DEX measure similarly uses a consensus of multiple DMS datasets to define exchangeability weights for evolutionary modeling [21]. Unsupervised inference methods, including Potts models and variational autoencoders, learn the fitness landscape directly from DMS data without requiring labeled functional outcomes [5]. Cross-protein transfer learning demonstrates that models trained on DMS data from one protein can substantially improve variant effect prediction for another, even across distantly related families [22].

Benchmarking studies have systematically compared computational predictors against DMS-derived functional scores. Livesey and Marsh showed that many widely used predictors (e.g., SIFT, PolyPhen-2) capture only a fraction of the information present in DMS data, and that correlation with DMS scores is reflective of clinical classification performance [23, 24, 25]. Reeb et al. similarly found that variant effect predictions correlate with DMS measurements but often miss context-dependent effects [26]. These comparisons underscore the value of DMS as a gold standard for evaluating and improving prediction algorithms.

Integration with Machine Learning for Viral Evolution

The combination of DMS data with machine learning has proven particularly powerful for forecasting viral evolution. For SARS-CoV-2, DMS of the spike receptor-binding domain (RBD) and full spike protein has been used to predict the emergence of new variants with enhanced immune escape or receptor affinity [6, 7, 8, 1, 2]. Machine learning models trained on DMS fitness landscapes can simulate viral evolutionary trajectories under selective pressures from antibodies or host receptors [7, 27]. These models have successfully forecasted the rise of Omicron sublineages by integrating DMS data with genomic epidemiology and socio-demographic factors [9, 10].

Protein language models informed by DMS data can resolve evolutionary dynamics with spatiotemporal resolution, predicting which mutations are likely to become fixed in circulating populations [6]. AlphaFold 3-assisted topological deep learning further refines these predictions by incorporating structural information from predicted protein complexes [28]. Quantitative atomistic modeling of spike-antibody complexes, combined with DMS escape data, identifies hotspots for immune evasion and guides the design of broadly neutralizing antibodies [12]. Similarly, DMS of influenza hemagglutinin has been used to predict the evolutionary success of human H3N2 clades, demonstrating the generalizability of this approach across viral families [3].

Applications in Veterinary Virology

DMS-based variant effect prediction has direct relevance to veterinary medicine, particularly for zoonotic viruses that circulate in animal reservoirs. Deep mutational scanning of avian influenza hemagglutinin has been employed to map mutations that alter receptor binding specificity from avian-type (alpha-2,3 sialic acid) to mammalian-type (alpha-2,6 sialic acid), a key determinant of pandemic potential [3]. These data inform computational models that predict zoonotic spillover risk, as discussed in the article Deep Mutational Scanning and Computational Modeling of Avian Influenza Hemagglutinin for Zoonotic Risk Prediction.

For coronaviruses, DMS of the spike RBD from SARS-CoV-2 and related bat coronaviruses has identified mutations that enhance binding to ACE2 orthologs across species, enabling host range prediction [11, 1]. These findings are integrated into machine learning frameworks that assess the zoonotic potential of emerging coronaviruses, as covered in Deep Mutational Scanning and Machine Learning Predictions of SARS-CoV-2 Spike Protein Receptor Binding Domain Escape Mutants. DMS data also guide the selection of vaccine antigens for veterinary use by predicting which spike variants are most likely to evade immunity induced by current vaccines [8, 29].

Beyond influenza and coronaviruses, DMS can be applied to other veterinary pathogens such as canine parvovirus, feline coronavirus, and avian infectious bronchitis virus. For example, DMS of the canine parvovirus capsid could map mutations that alter host receptor binding or antibody escape, aiding in the prediction of emerging variants. The computational pipelines developed for human viruses are directly transferable, provided that appropriate functional assays (e.g., virus entry, receptor binding) are established for the animal host system.

Challenges and Limitations

Despite its power, DMS has several limitations that affect variant effect prediction. Experimental noise and incomplete mutagenesis coverage can leave gaps in the fitness landscape, necessitating imputation methods that may introduce errors [13, 14, 15]. The functional selection used in DMS may not fully recapitulate the complex in vivo environment, particularly for phenotypes such as transmissibility or pathogenicity that involve multiple host factors [1, 2]. Context dependence of mutational effects (e.g., epistasis) is often not captured by single-substitution DMS libraries, although combinatorial DMS designs are beginning to address this [5].

Computational predictors trained on DMS data may overfit to the specific assay conditions and fail to generalize to other protein backgrounds or host species [16, 17]. Cross-protein transfer learning partially mitigates this, but performance drops for proteins with very different structures or functions [22]. Furthermore, the correlation between DMS scores and clinical outcomes (e.g., disease severity, vaccine breakthrough) is not always straightforward, and predictors must be validated against real-world surveillance data [23, 24].

Frequently Asked Questions

What is deep mutational scanning?

Deep mutational scanning is a high-throughput experimental method that measures the functional impact of all possible single amino acid substitutions in a protein by combining saturation mutagenesis, a functional selection, and deep sequencing [1, 2, 4].

How does DMS data improve variant effect prediction?

DMS data provide empirical fitness landscapes that serve as ground truth for training supervised machine learning models, benchmarking existing predictors, and deriving substitution matrices that capture amino acid exchangeability [16, 23, 20, 25].

What are the main computational methods used with DMS data?

Key methods include supervised learning from large mutagenesis datasets [16, 17], protein language model fine-tuning [6, 18, 19], unsupervised inference of fitness landscapes [5], cross-protein transfer learning [22], and imputation models such as VEFill [13, 14].

How is DMS used to predict viral escape from antibodies?

DMS of viral glycoproteins (e.g., spike, hemagglutinin) in the presence of neutralizing antibodies identifies escape mutations that reduce antibody binding, and these data are used to train machine learning models that forecast antigenic drift [8, 29, 12, 27].

What are the limitations of DMS-based predictions?

Limitations include experimental noise, incomplete coverage, context dependence of mutational effects, and the challenge of extrapolating from in vitro assays to in vivo phenotypes such as transmissibility or pathogenicity [13, 1, 15, 5].

How can DMS be applied in veterinary medicine?

DMS can map mutations that alter host receptor binding, immune escape, or virulence in animal viruses such as avian influenza, canine parvovirus, and feline coronavirus, informing surveillance, vaccine strain selection, and zoonotic risk assessment [9, 10, 3].

Workflow Overview

The following Mermaid diagram summarizes the integrated workflow from DMS experiment to variant effect prediction and veterinary application.

flowchart TD
    A[Design mutagenesis library], > B[Perform functional selection]
    B, > C[Deep sequencing and enrichment scoring]
    C, > D[Quality control and error modeling (DiMSum)]
    D, > E[Imputation of missing scores (VEFill)]
    E, > F[Train computational predictors (supervised, pLM, unsupervised)]
    F, > G[Validate against independent DMS or clinical data]
    G, > H[Predict variant effects for novel sequences]
    H, > I[Forecast viral evolution and immune escape]
    I, > J[Inform veterinary surveillance and vaccine design]

References

[1] Dadonaite B, Brown J, McMahon TE et al. Spike deep mutational scanning helps predict success of SARS-CoV-2 clades. Nature. 2024. URL: https://pubmed.ncbi.nlm.nih.gov/38961298/

[2] Dadonaite B, Brown J, McMahon TE et al. Full-spike deep mutational scanning helps predict the evolutionary success of SARS-CoV-2 clades. bioRxiv. 2023. URL: https://pubmed.ncbi.nlm.nih.gov/38014024/

[3] Lee JM, Huddleston J, Doud MB et al. Deep mutational scanning of hemagglutinin helps predict evolutionary fates of human H3N2 influenza variants. Proc Natl Acad Sci U S A. 2018. URL: https://pubmed.ncbi.nlm.nih.gov/30104379/ *** Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.

[4] Faure AJ, Schmiedel JM, Baeza-Centurion P et al. DiMSum: an error model and pipeline for analyzing deep mutational scanning data and diagnosing common experimental pathologies. Genome Biol. 2020. URL: https://pubmed.ncbi.nlm.nih.gov/32799905/

[5] Fernandez-de-Cossio-Diaz J, Uguzzoni G, Pagnani A. Unsupervised Inference of Protein Fitness Landscape from Deep Mutational Scan. Mol Biol Evol. 2021. URL: https://pubmed.ncbi.nlm.nih.gov/32770229/

[6] Yang S, Luo X, Luo J et al. A deep mutational scanning-informed protein language model predicts SARS-CoV-2 evolution dynamics with spatiotemporal resolution. Nat Microbiol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42204343/

[7] Durumeric AEP, McCarty S, Smith J et al. Machine Learning-Driven Simulations of the SARS-CoV-2 Fitness Landscape from Deep Mutational Scanning Experiments. J Chem Inf Model. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42089465/

[8] Nasir A, Lee D, Avena LE et al. Predictive modeling of immune escape and antigenic grouping of SARS-CoV-2 variants. J Virol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42037411/

[9] Ding Z, Yuan HY. Viral traits from deep mutational scanning and socio-demographic context predict SARS-CoV-2 lineage fitness across diverse countries. Int J Infect Dis. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41319790/

[10] Lei ZY, Zhang XM, Han JL et al. Integrating genomic epidemiology and deep mutational scanning data for prevalence forecasting of SARS-CoV-2 Omicron lineages. PLoS One. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/41183040/

[11] Xia H, Wei D, Guo Z et al. Machine Learning on the Impacts of Mutations in the SARS-CoV-2 Spike RBD on Binding Affinity to Human ACE2 Based on Deep Mutational Scanning Data. Biochemistry. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40811092/

[12] Alshahrani M, Parikh V, Foley B et al. Quantitative Characterization and Prediction of the Binding Determinants and Immune Escape Hotspots for Groups of Broadly Neutralizing Antibodies Against Omicron Variants: Atomistic Modeling of the SARS-CoV-2 Spike Complexes with Antibodies. Biomolecules. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40001552/

[13] Polunina PV, Maier W, Rubin AF. VEFill: accurate and generalizable deep mutational scanning score imputation across protein domains. Mol Syst Biol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41862706/

[14] Polunina PV, Maier W, Rubin AF. VEFill: a model for accurate and generalizable deep mutational scanning score imputation across protein domains. bioRxiv. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40463051/

[15] Yu T, Fife JD, Adzhubey I et al. Joint estimation and imputation of variant functional effects using high throughput assay data. medRxiv. 2023. URL: https://pubmed.ncbi.nlm.nih.gov/36711907/

[16] Schulze TK, Blaabjerg LM, Cagiada M et al. Supervised learning of protein variant effects across large-scale mutagenesis datasets. Protein Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41795207/

[17] Chen L, Zhang Z, Li Z et al. Learning protein fitness landscapes with deep mutational scanning data from multiple sources. Cell Syst. 2023. URL: https://pubmed.ncbi.nlm.nih.gov/37591206/

[18] Lamb KD, Hughes J, Lytras S et al. From single-sequences to evolutionary trajectories: protein language models capture the evolutionary potential of SARS-CoV-2. Nat Commun. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41714330/

[19] Marquet C, Heinzinger M, Olenyi T et al. Embeddings from protein language models predict conservation and variant effects. Hum Genet. 2022. URL: https://pubmed.ncbi.nlm.nih.gov/34967936/

[20] Munro D, Singh M. DeMaSk: a deep mutational scanning substitution matrix and its use for variant impact prediction. Bioinformatics. 2021. URL: https://pubmed.ncbi.nlm.nih.gov/33325500/

[21] Douglas GM, Bobay LM. DEX: a consensus-based amino acid exchangeability measure for improved codon substitution modelling. bioRxiv. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41959113/

[22] Jagota M, Ye C, Albors C et al. Cross-protein transfer learning substantially improves disease variant prediction. Genome Biol. 2023. URL: https://pubmed.ncbi.nlm.nih.gov/37550700/

[23] Livesey BJ, Marsh JA. Variant effect predictor correlation with functional assays is reflective of clinical classification performance. Genome Biol. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40264194/

[24] Livesey BJ, Marsh JA. Interpreting protein variant effects with computational predictors and deep mutational scanning. Dis Model Mech. 2022. URL: https://pubmed.ncbi.nlm.nih.gov/35736673/

[25] Livesey BJ, Marsh JA. Using deep mutational scanning to benchmark variant effect predictors and identify disease mutations. Mol Syst Biol. 2020. URL: https://pubmed.ncbi.nlm.nih.gov/32627955/

[26] Reeb J, Wirth T, Rost B. Variant effect predictions capture some aspects of deep mutational scanning experiments. BMC Bioinformatics. 2020. URL: https://pubmed.ncbi.nlm.nih.gov/32183714/

[27] Thadani NN, Gurev S, Notin P et al. Learning from prepandemic data to forecast viral escape. Nature. 2023. URL: https://pubmed.ncbi.nlm.nih.gov/37821700/

[28] Wee J, Wei GW. Rapid response to fast viral evolution using AlphaFold 3-assisted topological deep learning. Virus Evol. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40352163/

[29] Teruel NFB, Crown M, Rajsbaum R et al. Comprehensive analysis of SARS-CoV-2 Spike evolution: epitope classification and immune escape prediction. Virus Evol. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40510757/

[30] Fu Y, Bedő J, Papenfuss AT et al. Integrating deep mutational scanning and low-throughput mutagenesis data to predict the impact of amino acid variants. Gigascience. 2022. URL: https://pubmed.ncbi.nlm.nih.gov/37721410/

[31] Sruthi CK, Balaram H, Prakash MK. Toward Developing Intuitive Rules for Protein Variant Effect Prediction Using Deep Mutational Scanning Data. ACS Omega. 2020. URL: https://pubmed.ncbi.nlm.nih.gov/33251402/