Developability engineering of scFvs enables robust CAR function under transient mRNA expression

Abstract

Lipid nanoparticle delivery of mRNA enables in vivo generation of CAR-T cells while avoiding viral vector manufacturing constraints. However, transient expression places stringent requirements on the developability of CAR components, particularly single-chain variable fragments (scFvs). Here we examine how intrinsic scFv biophysical quality governs CAR expression and function under mRNA delivery. Using a humanisation and optimisation workflow guided by computational developability metrics, we engineered scFv variants from three clinically relevant antibody families (FMC63, 14G2a and MGA271/chBRCA84D) spanning a range of baseline stability and aggregation propensity. Optimised variants showed increased humanness and, in most cases, improved thermal stability and reduced non-specific interactions while broadly preserving antigen recognition. Functional testing in primary human T cells following mRNA–LNP delivery revealed strong dependence on starting binder quality: optimisation produced minimal functional change for the stable, predominantly monomeric FMC63 scFv, whereas stabilisation of the aggregation-prone 14G2a scFv markedly increased CAR surface expression and cytotoxic activity. For MGA271-derived binders, optimisation reduced non-specific interactions and off-target killing despite reduced monovalent affinity. Together, these data establish developability engineering of scFvs as a critical enabling step for transient mRNA-encoded CAR therapies.

Introduction

Single-chain variable fragments (scFvs) represent a crucial architectural component of chimeric antigen receptor (CAR)-T cell therapeutics, consisting of immunoglobulin heavy (V_H) and light chain (V_L) variable regions connected by a flexible peptide linker that maintains antigen-binding properties in a significantly reduced format (∼25 kDa as a single polypeptide)¹. Incorporated at the extracellular terminus of the CAR, the scFv enables T cell recognition of diverse tumour-associated antigens in an MHC-independent fashion, mediating the specificity and efficacy of cellular immunotherapy². This modularity has supported rapid expansion of CAR-T technology across multiple malignancies, most notably haematological cancers such as B-cell acute lymphoblastic leukaemia and non-Hodgkin’s lymphoma, where CD19-targeted scFvs have enabled durable clinical responses^3–5. Moreover, emerging preclinical and clinical data continue to highlight the central role of scFv architecture and biophysical properties in shaping CAR-T efficacy, persistence and safety^2,6,7.

The field of CAR-T cells has recently witnessed a transformative transition toward lipid nanoparticle (LNP)-mediated mRNA delivery for in vivo CAR-T cell engineering, addressing critical limitations of traditional viral vector-based manufacturing including high costs, potential insertional mutagenesis, scalability constraints and restricted accessibility in resource-limited settings⁸. Recent clinical and preclinical advances demonstrate that targeted LNPs can efficiently deliver CAR mRNA to endogenous T cell populations in vivo, generating potent and functionally active CAR-T cells with significantly reduced manufacturing timelines and costs, demonstrating robust tumour control in xenograft, humanised mouse models and non-human primate studies^10,16.

However, LNP-mRNA mediated delivery fundamentally alters CAR expression dynamics in ways that have profound implications for scFv design. Unlike viral vector-based CAR-T cells, which achieve sustained and high-level CAR expression through genomic integration, transient mRNA transfection with LNPs produces CAR-T cells with highly restricted CAR expression windows (typically 6–72 hours post-transfection)¹¹. Therefore, mRNA delivery introduces fundamentally unique design challenges where scFv biophysical characteristics become crucial determinants of therapeutic efficacy. The transient nature of mRNA-based CAR expression means that protein misfolding, aggregation propensity, or insufficient conformational stability can directly compromise therapeutic efficacy without opportunity for compensatory mechanisms or recovery. Conversely, in viral vector systems, stable genomic integration of the CAR transgene drives continuous transcription over weeks, providing multiple opportunities for functional protein recovery and turnover at the cell surface. Some levels of protein aggregation and misfolding might be tolerated or rescued in these systems, ultimately leading to a good density of functional scFvs on the cell surface even when these have suboptimal biophysical properties. In contrast, for mRNA-delivered CAR systems to achieve – within the narrow window of active translation – a comparable surface density of functional CARs, most translated molecules must correctly fold and be displayed on cell surfaces, thus making the effect of suboptimal biophysical properties much more apparent. Accordingly, the rational engineering of scFv domains to prioritise expression levels, stability, aggregation resistance, specificity, and robust folding is paramount.

Antibody discovery and engineering have been accelerated by the integration of computational and deep learning technologies¹². Among these methods, AbNatiV is a deep-learning model trained to quantify the nativeness of antibody and nanobody sequences by comparing them respectively to natural human or camelid repertoire distributions^13,14. AbNatiV has been used successfully to humanise single-domain nanobodies (VHHs) across multiple contexts^13,15–17, but applications to scFvs remain lacking. Here, we introduce an AbNatiV-based framework to humanise scFvs by jointly optimising V_H and V_L domains. Higher nativeness has been associated with improved folding efficiency, reduced aggregation propensity, and lower immunogenic potential across multiple antibody engineering studies^13,18,19. In this study, we applied AbNatiV-guided computational design, alongside complementary sequence- and structure-based assessments, to identify humanised scFv variants with improved biophysical properties across three clinically relevant binder families (FMC63, 14G2a, and MGA271/chBRCA84D). Most engineered scFv variants displayed enhanced thermodynamic stability, reduced aggregation propensity, and thus improved overall scFv developability. This translated into increased CAR surface expression, reduced antigen-independent cytotoxicity and improved functional performance when expressed from mRNA constructs. These findings establish a framework for rational scFv optimisation tailored to transient mRNA-based CAR delivery and address an important biophysical bottleneck limiting the performance of next-generation in vivo CAR-T technologies.

Results

Computationally guided humanisation and developability optimisation of single-chain variable fragments

AbNatiV was originally trained separately on human V_H, V, V and camelid VHH sequences, using a structural-alignment-based sequence encoding and a convolution–transformer autoencoder with a discrete information bottleneck while masking residues during training to force the model to learn context-dependent residue distributions. The global nativeness score derived from reconstruction error successfully distinguished native repertoires from artificial sequences and correlated with humanness and clinical immunogenicity, while per-residue nativeness profiles highlighted non-native positions and enabled nanobody humanisation with preservation of binding and, in some cases, improved stability¹³. For the joint humanisation of the V_H and V_L domains of an scFv, we first modelled the paired Fv with ABodyBuilder2²⁰ to identify and discard residues buried within each domain but also at the V_H–V_L interface to preserve interdomain stability. Humanising mutations were then sampled with AbNatiV for each domain separately using the enhanced sampling pipeline described in Ref.¹³. In the iterative process, CDR displacement or compromised V_H/V_L orientation was monitored by modelling the resulting humanised Fv to select variants with minimal predicted structural deviations from the WT, and thus harbouring mutations less likely to disrupt binding (Fig. 1a).

Fig. 1.

Computationally guided humanisation and developability engineering of scFvs for mRNA CAR expression. (a) Humanisation pipeline for paired Fv domains. Wild-type scFvs are first modelled to compute solvent-accessible surface area and identify buried residues within each domain and at the V_H/V_L interface; AbNatiV humanness profiling is then applied to flag non-CDR, solvent-exposed framework residues with low humanness scores, from which humanising mutations are sampled independently for each domain. Humanised Fv variants are remodelled and structurally aligned to the wild-type framework, and CDR displacement is monitored. (b) AbNatiV humanness scores for WT (circles) and humanised (stars) scFvs derived from MGA271, FMC63 and 14G2a. For all binders, AbNatiV identifies humanised variants with humanness scores >0.8 in both domains, placing them within the distribution of human antibody repertoires. (c) Per-residue AbNatiV scores of FMC63 WT (orange) and humanised (purple) variants.

A diverse panel of scFv was assembled to evaluate this design paradigm, focusing on binders with clear clinical relevance or well-characterised developability liabilities. FMC63 was chosen as the prototypic CD19-directed binder used in most first-generation CD19 CAR-T products, 14G2a was selected as a GD2-targeting antibody with well-documented tendencies toward self-association and associated functional instability in CAR formats²¹, and the MGA271-derived scFv was included owing to its comparatively low affinity for human B7-H3 and suboptimal performance when transferred from the IgG to the scFv scaffold^22,23. In addition, FMC63 and 14G2a are murine monoclonal antibodies that had not been previously humanised^24,25, while chBRCA84D is a chimeric antibody in which the murine BRCA84D VH and VL domains are fused to human constant regions²⁶. In all three cases, the variable domain frameworks used as starting points for AbNatiV-guided humanisation are therefore of murine origin. These features provided a test bed for whether rational humanisation and biophysical optimisation of the scFv domains could mitigate known sequence-encoded liabilities and improve suitability for CAR development.

Humanised variants were generated using the AbNatiV enhanced sampling mode, which by design converges on a single humanised sequence per input without requiring manual candidate selection. This mode was preferred over the exhaustive sampling alternative, which returns a Pareto front of sequences across varying numbers of mutations, since it was shown in the original AbNatiV study to yield humanised variants with retained or improved binding and biophysical properties, performing on par with or better than exhaustive sampling¹³. For all binder scFvs, AbNatiV generated humanised variants with a humanness score >0.8 for both the V_H and V_L domains, a threshold above which sequences are considered within the distribution of human repertoires (Fig. 1b). The selected mutations are located on solvent-exposed framework regions and collectively restore humanness across most of the regions initially flagged as non-native (Fig. 1c).

Biophysical characterisation of humanised and engineered CAR binding domains

Following sequence humanisation, biophysical characterisation was carried out for all three binder families (FMC63, 14G2a and MGA271/chBRCA84D), each examined in both wild-type and humanness-enhanced forms. Binding-kinetic analyses confirmed that antigen recognition was retained following humanisation and optimisation across all three binder families, although the impact on affinity varied between scaffolds (Fig. 2a and Table 1). FMC63 displayed a modest reduction in apparent affinity following optimisation, while the MGA271/chBRCA84D variant showed a more substantial decrease in binding strength. Despite this reduction, the optimised MGA271-derived binder retained measurable antigen binding (Fig. 2a). Furthermore, the observed reduction in FMC63 affinity is not expected to be detrimental to therapeutic efficacy, as prior reports have demonstrated that attenuated CD19 CAR affinity can support improved T-cell persistence and proliferation by reducing tonic signalling and exhaustion without compromising antitumour efficacy²⁷.

Fig. 2.

Biophysical properties of engineered scFvs. (a) Representative SPR sensorgrams comparing antigen binding of wild-type (WT) and enhanced scFv variants derived from FMC63, 14G2a and MGA271/chBRCA84D. scFvs were injected as analytes over immobilised ligand surfaces, and kinetic parameters were obtained by global fitting to a 1:1 binding model. (b) Differential scanning fluorimetry (DSF) profiles of WT and enhanced scFvs showing intrinsic fluorescence ratio (350/330 nm) as a function of temperature. Apparent melting temperatures () were derived from replicate measurements and reflect overall conformational stability. (c) Quantification of monomeric species determined by size-exclusion chromatography following affinity purification. Bars represent the percentage of monomeric scFv for WT and enhanced variants; golimumab and adalimumab are shown as reference controls. Error bars indicate mean ± SD from replicate runs.

Table 1.

Summary of biophysical and binding properties of wild-type and enhanced humanised scFv variants. V mut. and V mut. indicate the number of framework mutations introduced relative to the wild-type sequence.

scFv variant	()	% monomer	(nM)	(Ms)	(s)	V mut.	V mut.
FMC63 WT	73.69	87%	0.95			–	–
FMC63 Enhanced	79.37	94%	4.76			15	7
MGA271 WT	72.34	92%	90.1			–	–
MGA271 Enhanced	70.07	89%	1,120			8	13
14G2a WT	61.61	55%	82.67			–	–
14G2a Enhanced	68.53	70%	66.09			16	8

Differential scanning fluorimetry demonstrated that humanisation did not compromise thermal stability; on the contrary, in two of the three humanised (enhanced) scFv variants analysed in this study, the apparent melting temperatures increased by more than 5 , indicating improved conformational robustness of the engineered domains (Fig. 2b and Table 1). Size-exclusion chromatography of the affinity-chromatography-purified scFvs further revealed that monomeric yields were maintained or increased post-humanisation, with enhanced FMC63 and 14G2a variants showing a substantially higher proportion of monomeric species than their wild-type counterparts, consistent with reduced aggregation propensity (Fig. 2c). Together, these biophysical data (Table 1) support advancing the optimised binders into functional CAR constructs, delivered to primary T cells by mRNA transfection, with the expectation that improved thermal stability and reduced aggregation propensity may diminish non-specific clustering and tonic signalling, thereby enhancing functional durability, T-cell persistence and downstream performance in cell-based assays and therapeutic settings.

CAR expression following mRNA–LNP delivery to primary human T cells

To determine whether these biophysical gains translated into functional improvements at the level of the CAR receptor, scFv sequences were cloned into expression cassettes tailored for efficient mRNA production and used to generate CAR constructs encoding either wild-type or enhanced variants (Fig. 3a). Primary human T cells were transfected with these mRNAs through LNPs and evaluated in parallel for CAR expression, signalling and cytotoxic activity against CD19-, B7-H3- or GD2-positive target cells, providing a direct link between intrinsic scFv properties and emergent CAR behaviour (Supplementary Fig. 1). Across antigens, functional analyses revealed striking variant-dependent profiles that closely mirrored the underlying biophysical attributes. For CD19 CARs, the enhanced FMC63 variant did not yield a major gain in cytotoxic potency or cytokine release relative to the parental construct (Supplementary Fig. 2), consistent with the high abundance and accessibility of CD19 on target cells and with prior observations that CD19 CARs can robustly retain therapeutic effect (Fig. 3b). In contrast, 14G2a-based GD2 CARs were highly sensitive to binder optimisation: the wild-type construct showed little or no detectable surface expression by flow cytometry. While this readout does not distinguish between impaired translation, intracellular retention or aggregation-prone CAR expression, the robust surface expression observed for the enhanced variant under identical mRNA transfection conditions suggests that sequence-encoded biophysical liabilities in the wild-type scFv strongly limit productive CAR display (Fig. 3c).

Fig. 3.

Functional impact of scFv developability under transient mRNA expression. (a) Schematic representation of mRNA CAR-T constructs used in this study. (b–d) Antigen-expressing SupT1 target or control cells were co-cultured with primary human T cells expressing wild-type or enhanced CARs at the indicated effector-to-target ratios for 72 hours. Viable target cells were quantified by flow cytometry, excluding CD2⁺ and CD3⁺ events to remove CAR-T cells from the analysis. The percentage of live target cells was determined by normalising to the non-treated (NT) control group. A representative example for each relative CAR-T construct mRNA expression is shown in the left panel. (e) Schematic representation of viral vector MGA271/chBRCA84D CAR-T constructs used in this study. (f) Cytotoxic activity of CAR-T cells expressing either the wild-type or enhanced CAR against B7-H3 or an empty vector control expressing SupT1 target cells. (g) Cross-Interaction Chromatography (CIC) quantification of the MGA271/chBRCA84D wild-type and enhanced scFv versions. Golimumab and adalimumab were used as control.

Functional consequences of scFv developability under transient expression

For B7-H3-directed CARs, structural and functional evaluation of MGA271/chBRCA84D constructs showed that rational modification of the MGA271-derived scFv scaffold mitigated non-specific interactions, by substantially reducing background cytotoxicity and off-target cell killing, while preserving on-target antitumour activity. Notably, this improved selectivity was achieved despite the substantial reduction in monovalent affinity reported above (Fig. 2a, Table 1), indicating that the developability gains outweighed the loss in affinity within the functional context tested. These data are consistent with recent comparative studies reporting that alternative B7-H3 binders, including engineered variants with reduced hydrophobic or charged patches, can deliver superior antigen-sensitive activation and lower tonic signalling than first-generation MGA271-based designs (Fig. 3d). When wild-type and enhanced MGA271/chBRCA84D CARs were subsequently introduced via retroviral transduction for stable integration into primary T cells, the enhanced constructs consistently displayed markedly lower background killing of antigen-negative cells and a cleaner functional window between on-target and off-target responses, confirming that the improvements in specificity and selectivity were robust across delivery platforms and not restricted to transient mRNA expression (Fig. 3e,f). Cross-interaction chromatography (CIC) on purified antibody fragments further supported this mechanistic link, demonstrating substantially reduced non-specific retention of the enhanced scFv on heterologous matrices, in line with a decrease in biophysical promiscuity and a corresponding increase in cellular specificity (Fig. 3g).

Discussion

Taken together, the results of this study demonstrate that the performance of CARs delivered via transient LNP–mRNA expression is strongly influenced by the intrinsic biophysical properties of the scFv binding domain. By systematically analysing scFvs spanning a range of baseline developability profiles, we show that the functional impact of rational optimisation depends primarily on the quality of the starting binder. scFvs that already exhibit high thermal stability and a predominantly monomeric profile, such as FMC63, appear robust to further optimisation and – despite improvements in some biophysical properties – display minimal gains in functional assays, whereas scFvs with pronounced developability liabilities, exemplified by 14G2a, require stabilisation and aggregation-resistance engineering to achieve productive surface expression and cytotoxic activity. For MGA271-derived binders, optimisation altered the balance between affinity, stability and non-specific interactions, resulting in reduced off-target killing despite a substantial reduction in binding affinity that nevertheless remained compatible with measurable on-target cytotoxicity and improved functional selectivity.

The number of framework substitutions introduced in each variant (Table 1) provides some insight into the relationship between the extent of sequence editing and the functional outcomes observed. The total mutational load was broadly comparable across the three binders (21-24 framework substitutions), yet the functional consequences differed markedly, indicating that the number of framework changes alone does not predict the magnitude or direction of the effect. For FMC63, a substantial framework load (15 VH, 7 VL) was accommodated with minimal functional change, consistent with its already favourable starting biophysics. For 14G2a, the largest and most VH-weighted load (16 VH, 8 VL) coincided with the greatest functional gain, in keeping with mutations acting on a binder carrying pronounced stability and aggregation liabilities. By contrast, the VL-biased load of the MGA271-derived variant (8 VH, 13 VL) accompanied an affinity-specificity trade-off, in which reduced non-specific interactions were offset by a decrease in monovalent affinity. As all substitutions were restricted by design to solvent-exposed, non-CDR framework positions, these effects are most consistent with framework-mediated changes in conformational stability and colloidal behaviour; the concomitant affinity reduction in the MGA271-derived variant nonetheless indicates that framework editing can propagate to the binding site, likely through subtle changes in VH-VL orientation. Given the small number of binders examined, these observations are qualitative rather than quantitative, but they reinforce that the consequences of humanising substitutions depend on the biophysical context of the starting binder rather than on mutation count per se.

Together, these findings support an scFv-centric model in which sequence-encoded biophysical properties, rather than antigen-specific factors alone, govern CAR expression and function under transient mRNA delivery. More broadly, they establish rational scFv developability optimisation as an enabling step for robust and selective mRNA-based CAR-T therapies, particularly when translating clinically relevant but biophysically suboptimal binders into this emerging therapeutic modality.

Materials and methods

In silico humanisation of scFvs

Each WT scFv was humanised with the AbNatiV humanisation pipeline, adapted here for paired V_H/V_L antibodies¹³. For each scFv, the solvent-accessible surface area (SASA) was computed by modelling the paired Fv with ABodyBuilder2²⁰. Buried residues (SASA <0.15) and CDR residues (AHo numbering) were excluded from further optimisation to minimise the risk of stability loss and binding disruption, respectively. AbNatiV was then run on the remaining sequence using the default enhanced sampling protocol for each V_H and V_L domain separately. The resulting humanised paired Fv fragment was then modelled, structurally aligned to the WT framework model, and CDR displacement was monitored to assess the quality of the humanisation.

Surface Plasmon Resonance (SPR) binding kinetics

Binding kinetics of wild-type and enhanced scFv variants were measured by surface plasmon resonance (SPR) using a Biacore T200 instrument (Cytiva) operated at 25 . For FMC63 and MGA271/chBRCA84D scFvs, recombinant Fc-tagged human CD19 or B7-H3 antigens (ACROBiosystems) were used as ligands and captured on a Series S CM5 sensor chip pre-functionalised with anti-human Fc antibody. Briefly, the CM5 surface was activated by standard amine coupling chemistry using a 1:1 mixture of 0.2 M EDC and 0.05 M NHS, followed by immobilisation of the anti-human Fc capture antibody in 10 mM sodium acetate buffer (pH 4.5). Remaining active esters were quenched with 1 M ethanolamine (pH 8.5). Fc-tagged antigens were then captured to a level of approximately 200–500 RU prior to each binding cycle. A reference flow cell containing the capture antibody but no antigen was included for background subtraction. For GD2 binding measurements, biotinylated GD2 oligosaccharide (Biosynth, GD2-oligosaccharide-sp-biotin) was captured on a Series S SA sensor chip according to the manufacturer’s instructions, typically to a density of ∼200–300 RU. In this configuration, the corresponding scFv (14G2a wild-type or enhanced variant) was injected as the analyte. Purified scFvs were injected at multiple concentrations prepared in running buffer consisting of PBS supplemented with 0.05% (v/v) Tween-20, at a flow rate of 30 µL min^-1. Association and dissociation phases were monitored for 180 s and 600 s, respectively. Where required, sensor surfaces were regenerated using brief injections of low-pH glycine buffer.

Thermostability measurements

The apparent melting temperature of the purified scFvs was measured in PBS at a protein concentration using a Tycho NT.6 instrument (NanoTemper). The intrinsic fluorescence was recorded at 330 nm and 350 nm with a fixed 30 /min temperature ramp from 35 to 95 . The experiment was repeated three times for each antibody variant. A two-state thermodynamic model, as described in Ramon et al.²⁸, was fitted on each individual 350 nm and 330 nm fluorescence signal. The representative apparent melting temperature for each antibody fragment was obtained by averaging the fitted melting temperatures from both individual signals over the three replicates.

Cross-interaction measurements via cross-interaction chromatography

Cross-interaction of each scFv was indicated by the retention time measurements between the variants and IgG from human serum (I4506, Sigma) via cross-interaction chromatography (CIC). The column was prepared as described in Wolf Pérez et al.²⁹ and the assay was modified based on Rosace et al.³⁰. Briefly, a 1 mL HiTrap NHS-activated HP affinity column (17071701, Cytiva) was washed with 6 mL ice-cold 1 mM HCl and 30 mg of human serum IgG was immobilised to the column using an ÄKTA pure system (Cytiva). The assay was run on an ÄKTA pure system connected to an ALIAS autosampler (Cytiva) at 4 . 25 µL sample for all variants, including two control samples adalimumab and golimumab, was injected at 0.03 mg/mL with PBS as running buffer. The flow rate was 0.1 mL/min and the runtime was 30 min. All variants were injected in duplicates. The retention time was calculated based on the position of the elution peak observed with UV absorbance measured at 280 nm.

mRNA synthesis and preparation

mRNAs encoding the different CARs used in this study were cloned into a modified pcDNA3.1 vector (Thermo Fisher Scientific) carrying a 5’ UTR, 3’ UTR and a 40A(G)40A polyA tail that can be transcribed from an upstream T7 promoter. mRNA production was made in-house for in vitro studies. These mRNAs were made using a HiScribe^® T7 mRNA Kit with CleanCap^® Reagent AG (New England BioLabs), according to the manufacturer’s instructions. In brief, plasmids were first linearised with BspQI (New England BioLabs), cleaned with a Monarch^® Spin PCR & DNA Cleanup Kit (New England BioLabs) and in vitro transcribed into mRNA using HiScribe^® T7 mRNA Kit with CleanCap^® Reagent AG with 100% substitution of UTP with N1-Methyl-Pseudo-UTP. After in vitro transcription (IVT), the template was removed by DNaseI treatment and mRNA was purified by LiCl precipitation and 70% ethanol wash. Final mRNA was dissolved in RNA storage buffer (Thermo Fisher Scientific) and stored at -80 .

mRNA-LNP

GenVoy-ILM™ T cell kits (Cytiva) for mRNA were used to formulate mRNA-loaded lipid nanoparticles (LNPs) and deliver them into activated primary human T cells per the manufacturer’s instructions. In brief, mRNA working solution, GenVoy-ILM lipid mix, and formulation buffer 2 were prepared in the specified volumes and concentrations to achieve the target N:P ratio and final mRNA concentration. The mixture was loaded onto a Spark cartridge and processed using the predefined GenVoy-ILM T cell method (runtime <5 min), after which the mRNA-LNP output was collected and immediately diluted in formulation buffer 2. mRNA encapsulation efficiency was assessed via RiboGreen assay following the manufacturer’s protocol. T cells, pre-activated for 72 h with ImmunoCult™ Human CD3/CD28 T Cell Activator (StemCell Technologies), were incubated with LNPs for 24 h to enable mRNA expression, then harvested for viability and CAR expression analysis.

Retroviral production and transduction

For amphotropic viral production, HEK-293GP cells were transiently transfected with a VSV-G envelope expression plasmid and a retroviral cDNA construct (ratio 1:4). The transfection was carried out with GeneJuice (Millipore; 70967-4) according to the manufacturer’s guidelines (ratio DNA:GeneJuice 1:2.5). After 72 h, the cell supernatant was filtered through a 0.45 µm filter and the viral supernatant was added to HEK-293RD114 cells. Transduction was performed by spinoculation (1500× g, 1 h at 32 ) with the addition of 10 µg/mL Polybrene (INSIGHT Biotechnology). Viral supernatant from HEK-293RD114 cells was then collected at different time points when cell confluency reached ∼100%, snap frozen and stored at -80  for subsequent titration and use. Isolated PBMCs were activated using 10 ng/mL IL-7 and IL-15 and 1:100 T Cell TransAct™ (Miltenyi Biotec). 48 hours post activation, PBMCs were plated on Retronectin (Takara Clonetech; T100B) on 6-well plates (Corning; 351146) with viral supernatant and spun for 1500× g, 1 h at 32 .

Flow cytometry

Flow cytometry was performed using a MACSQuant 10 flow cytometer (Miltenyi Biotec, Germany). All labelling was conducted at room temperature (RT) for 15 minutes in subdued light with antibodies diluted in PBS + 2% FBS. The antibodies used were: CD3-PECy7 (BioLegend; 300420), CD2-PE (BioLegend; 300208), G4S Linker (E7O2V) Alexa Fluor^® 647 (Cell Signaling Technologies).

Cytotoxicity assay

Transduced cells were co-cultured with target cells at various effector:target (E:T) ratios, with target cells remaining constant at . Target cells were counted via flow cytometry using CountBright Absolute Counting beads (Life Technologies; C36950) and cytotoxicity was calculated as a percentage by normalising to the number of target cells recovered from co-cultures with non-transduced (NT) T cells.

Statistical analysis

Data analyses were performed using GraphPad Prism 10 (GraphPad Software, La Jolla, USA). For comparisons involving more than two groups, statistical analyses were performed using one-way ANOVA followed by Tukey’s multiple-comparison test. For pairwise comparisons, two-tailed unpaired Student’s t-tests were used as indicated. Normality was assessed using the Shapiro–Wilk test. Statistical annotations were denoted with asterisks as follows: , , , , and not significant (ns) P>0.05.

Author contributions

Conception of the study: MM, AR, PS, SO. Experimental design: MM, AR, BL, ES, PS, SO, Computational data generation: AR, PS, Experimental data generation: MM, AR, ES, JH, BL, RK, Analysis and interpretation of data: MM, AR, PS, BL, SO, Writing: MM, AR, PS, SO.

Funding

This research was funded in part by UK Research and Innovation (Grant number 10104367).

Data availability

Data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

Maurizio Mangolini, Emily Souster, Jasvinder Hayre, Rajesh Karattil and Shimobi Onuoha are current (or former) employees of Chimeris Ltd and receive (or received) salary and shares for their work. The contribution to this study of Pietro Sormanni and Aubin Ramon was conducted in a consultancy capacity and was remunerated by Chimeris UK.

Ethics statement

Peripheral Blood Mononuclear Cells (PBMCs) were isolated from Leukocyte Reduction System (LRS) cones obtained from NHS Blood and Transplant (NHSBT), Cambridge, UK. The material was supplied as fully anonymised surplus clinical material under Human Tissue Authority (HTA) licence in accordance with the UK Human Tissue Act 2004. Donors provided consent for research use at the point of donation. All procedures were conducted in accordance with the principles of the Declaration of Helsinki.

Use of large language models

Large language model assistance (ChatGPT, OpenAI) was used for language editing and clarity improvement only. All scientific content and interpretations were reviewed and approved by the authors.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-60295-2.