TL;DR

Applied zero-shot to FMC63, the murine antibody whose single-chain variable fragment (scFv) provides the CD19-binding domain in four approved CAR-T products, ConvergeAB™ produced a humanized, de-immunized lead in a single design pass: framework humanness rose from 77.8% to 93.3% (IMGT), both predicted CD4 T-cell epitopes were removed (2 → 0), thermal stability improved by ~19°C, and monomer content by ~13 points. Affinity was purposefully held at the parent's sub-nanomolar level, since in this case higher affinity increases toxicity. That is the point of the platform: ConvergeAB optimizes an antibody to a defined, multi-parameter target profile, raising, lowering, or holding each property as a program requires, with FMC63 as the demonstration.

Background

CD19 is a B-lineage-restricted antigen and one of the most effectively targeted antigens in oncology, most prominently through chimeric antigen receptor (CAR) T-cell therapy. FMC63 is the binder behind much of that success. Its scFv is the CD19-recognition domain in four approved CD19 CAR-T therapies: tisagenlecleucel (Kymriah), axicabtagene ciloleucel (Yescarta), brexucabtagene autoleucel (Tecartus), and lisocabtagene maraleucel (Breyanzi). All four use a murine scFv, so immunogenicity associated with the non-human sequence is a property of the class rather than of any single product.

The murine origin is a recognized contributor to limited CAR-T persistence. Because the scFv is mouse-derived, the patient's immune system can recognize it as foreign: CD4 T-cells that see the murine sequence help drive antibody and T-cell responses against the CAR-T cells themselves. This immune clearance shortens how long the CAR-T cells persist and stay active, which limits the durability of the response. It is especially significant in relapse: a second infusion is attacked and cleared even faster by the immunity raised against the first, so re-treatment often fails¹⁻³. Humanizing the framework and removing the murine T-cell epitopes, while preserving CD19 binding, reduces this immunogenicity and is the modification with a clear rationale.

Figure 1A. CD19 CAR-T mechanism of action. The FMC63 scFv (its VH and VL domains) is the CD19-binding domain of the chimeric antigen receptor. Recognition of CD19 on a malignant B cell forms the CAR-immune synapse, which triggers CD3ζ and co-stimulatory signaling, perforin and granzyme release, and tumor-cell lysis.

Figure 1B. Why the murine scFv limits CD19 CAR-T therapy, and how humanization resolves it. Shown across three stages. With the murine scFv (left column): host CD4 T-cells recognize the foreign sequence and drive anti-CAR antibody production (immunogenicity); these responses clear the CAR-T cells and shorten persistence (persistence); and at relapse a pre-existing memory anti-CAR response rapidly clears a second infusion, so re-treatment fails (relapse). Humanizing the framework removes the foreign epitopes, lowering the anti-CAR response, which extends CAR-T persistence and can allow re-treatment to succeed.

Objective

To humanize FMC63, remove its murine T-cell epitopes, and improve its developability in a single, zero-shot design pass while preserving its CD19 binding profile, in the scFv format used in a CAR. Because higher affinity is not desirable for this target, affinity was to be maintained rather than increased.

Methods

Variants were generated zero-shot with ConvergeAB™ from the FMC63 VH and VL sequences and the human CD19 target sequence, with no target-specific fine-tuning. ConvergeAB introduces sequence diversity around the parent antibody using a protein language model, generates a large target-aware candidate pool, and ranks it using orthogonal predictors of binding, structural compatibility, humanness, T-cell epitope content, thermal stability, and solubility. The top-ranked candidates were expressed and characterized against the FMC63 baseline, and the lead was profiled in full.

The lead and the FMC63 baseline were characterized in the scFv format, the configuration FMC63 occupies within a CAR. Binding kinetics were measured by surface plasmon resonance (SPR), thermal stability by nano-differential scanning fluorimetry (nano-DSF), and monomer content by analytical size-exclusion chromatography (SEC). Framework humanness was scored as IMGT germline identity, and framework and CDR boundaries follow the IMGT numbering scheme (ANARCI). CD4 T-cell epitope content was assessed computationally and is reported as a prediction pending experimental validation.

Results

The lead variant improved on the parent across humanness, thermal stability, and monomer content, and held CD19 affinity at the parent level (Table 1).

Framework humanness increased from 77.8% to 93.3% overall (VH 72.5% to 89.0%; VL 83.0% to 97.7%), moving the framework from murine toward human germline. The design also targeted immunogenicity directly: both predicted CD4 T-cell epitope cores in the variable domains were removed (2 to 0). These epitopes contribute to the immunogenicity that limits CAR-T persistence and re-treatment; presentation of such epitopes on MHC class II to CD4 T-helper cells is the step that drives an anti-drug antibody response, so removing them is the basis of de-immunization. The epitopes were removed through framework-level edits, so de-immunization is framework-driven.

Thermal onset increased by approximately 19.6°C and the first melting transition by approximately 18.9°C. Monomer content improved by about 13 percentage points. These properties bear on manufacturability, formulation, and storage stability.

CD19 affinity was 0.184 nM for the lead and 0.197 nM for the parent, a difference within assay variability and in both cases sub-nanomolar. Affinity was held rather than increased. Within the same scFv panel, candidates spanned roughly 0.05 to 1.2 nM, an approximately 24-fold range, with the tightest at 0.048 nM, about fourfold better than the parent. That candidate was not advanced, because a tighter binder is not the objective for this target.

Table 1. Comparison of FMC63 and the ConvergeAB lead in the scFv format. Humanness is IMGT germline identity; the T-cell epitope values are computational predictions. A framework-only variant, carrying the 42 framework edits without the 3 CDR edits, reaches the same humanness and epitope removal but binds CD19 at 3.33 nM; the CDR edits restore the parent-level 0.184 nM.

Sequence basis

The lead carries 45 substitutions relative to FMC63: 42 in the framework, at positions where it diverges from human germline, and just 3 in the CDRs. The framework edits alone deliver the humanness and remove both predicted T-cell epitopes, so humanness and de-immunization are framework-driven and achieved in one pass.

Framework humanization on this scale is not free. A framework-only variant of the lead, carrying the same 42 framework edits but none of the CDR edits, reaches the same 93% humanness and clears both epitopes, yet its CD19 affinity falls roughly 17-fold, from 0.197 nM to 3.33 nM. Extensive framework change does perturb the paratope, as expected from the framework and Vernier-zone residues that scaffold the CDRs. The lead's 3 CDR edits compensate, restoring affinity to the parent level (0.184 nM) and adding stability. This is the platform managing the trade-off between framework and CDRs directly: it humanizes and de-immunizes through the framework, then makes the minimal CDR changes needed to recover the binding those framework changes cost. The same control can be pointed the other way, as in a separate campaign on tafasitamab where framework-only edits deliberately repositioned CDR-H3 to tighten binding. Framework space is a control surface the platform uses deliberately, to hold the binding site still or to move it.

Figure 2. Per-residue framework humanization and de-immunization of FMC63. Each cell is one residue by IMGT position (ANARCI numbering), colored by match to the nearest human germline: green, matches; orange, differs. Framework (FW1–FW4) and CDR (CDR1–CDR3) regions are annotated above each panel; CDRs (grey) are shown for position only, not scored. Within each panel the upper row is the murine FMC63 parent and the lower row the humanized lead. Red boxes mark the two predicted CD4 T-cell epitope 9-mer cores, present in the parent and absent in the lead.

A note on reading Figure 2. IMGT germline identity scores each framework residue against the single closest human germline gene, so it is a strict reference rather than a pass-or-fail test. Human antibodies are not identical to their germline genes: they acquire somatic mutations, and the human repertoire varies naturally at many framework positions, so a residue that differs from the nearest germline is often still a common, fully human residue found in other human genes. Bona fide human antibodies do not score 100% on this metric, and 100% is not the target. The aim is to bring the framework into the range seen in human antibodies, which the lead does (VH 89.0%, VL 97.7%). The orange positions in the humanized lead (Figure 2, lower row of each panel) are therefore largely sites of natural human variation, not murine residues left un-humanized.

Why affinity was held rather than increased

This project set out to optimize the existing FMC63 framework, not to design a new binder; for a CD19 CAR-T binder, optimization does not mean higher affinity. Binding affinity is determined by both the association and dissociation rates; thus, increasing affinity could reflect a slower off-rate, which keeps the CAR engaged with individual target cells longer and promotes sustained signaling. Against a high-density antigen such as CD19 on B-lineage blasts, this prolonged signaling might be unnecessary for efficient target-cell recognition; rather, it could be associated with T-cell exhaustion and with the cytokine release syndrome and neurotoxicity that limit CAR-T therapy4. The most recent FDA-approved CD19 CAR-T product reflects the same reasoning: obecabtagene autoleucel (2024) uses a lower-affinity, faster-off-rate CD19 binder and reported high response rates with comparatively low rates of severe toxicity and durable persistence5. The lead therefore preserves the clinically validated binding of the approved FMC63 products rather than increasing it.

A configurable objective

The affinity result illustrates a general property of the platform: the optimization objective is set by the user. ConvergeAB can be directed to increase, decrease, or preserve affinity, to prioritize thermal stability or solubility, and to constrain defined regions of the sequence, for example by holding the CDRs fixed to preserve an epitope. In this campaign the objective was to humanize FMC63 and improve its developability while preserving CD19 binding. A different program can specify a different target profile, including the lower-affinity, faster-off-rate profile now used in next-generation CD19 CAR-T. The properties are weighted according to the needs of the program rather than toward a fixed goal, and a customer can adjust those priorities directly.

Figure 3. Affinity was chosen, not maximized. CD19 affinity (KD by SPR, scFv format) across the design panel. In a single design pass, ConvergeAB accessed a roughly 24-fold affinity range (0.048–1.16 nM). The lead was selected at 0.184 nM, holding the parent's sub-nanomolar affinity (FMC63, 0.197 nM), rather than the tightest candidate (0.048 nM, about 4× tighter), which was not advanced. Lower KD indicates tighter binding; the lead sitting near the parent reflects a deliberate selection, not a ceiling.

Context and significance

FMC63 is not a marginal molecule. Its scFv is the CD19-binding domain in four approved CAR-T products, Kymriah, Yescarta, Tecartus, and Breyanzi, which together generated about $3.5 billion in 2025, and it remains the most widely used anti-CD19 binder in the field. That an antibody of this clinical and commercial importance is still murine, with the associated immunogenicity liability carried into every product built on it, indicates that humanizing it without degrading its function has not been straightforward. The constraint is a simultaneous one: the framework must be made human, the CD19 binding and epitope must be preserved, and developability must not regress. The lead variant satisfies these constraints in a single design pass, a concrete demonstration of what the platform does on a hard, real-world, already-optimized molecule.

One further point is relevant for a program evaluation: because binding and the CD19 epitope are preserved, the lead is a candidate humanized version of FMC63 that keeps its specificity, a route to a lower-immunogenicity generation of FMC63-based constructs without changing the binding biology.

Conclusion

Applied zero-shot to the murine FMC63 scFv used in approved CD19 CAR-T, ConvergeAB™ produced a humanized, de-immunized lead in a single design pass: framework humanness raised from 77.8% to 93.3%, both predicted variable-domain T-cell epitopes removed, thermal stability improved by approximately 19°C, and higher monomer content, with CD19 affinity held at the parent's clinically validated level rather than increased. FMC63 is the demonstration; the platform behind it is the result of interest. ConvergeAB optimizes an antibody to a defined, multi-parameter target profile, raising, lowering, or holding each property as the program requires. For a partner, that means starting from a molecule they already have and getting back the version their program needs, more human, better behaved, and binding intact, without restarting discovery.

References

1. Turtle CJ, Hanafi LA, Berger C, et al. CD19 CAR-T cells of defined CD4+:CD8+ composition in adult B cell ALL patients. J Clin Invest. 2016;126(6):2123–2138. https://doi.org/10.1172/JCI85309

2. Wagner DL, Fritsche E, Pulsipher MA, et al. Immunogenicity of CAR T cells in cancer therapy. Nat Rev Clin Oncol. 2021;18:379–393. https://doi.org/10.1038/s41571-021-00476-2

3. Shah NN, Fry TJ. Mechanisms of resistance to CAR T cell therapy. Nat Rev Clin Oncol. 2019;16(6):372–385. https://doi.org/10.1038/s41571-019-0184-6

4. Ghorashian S, Kramer AM, Onuoha S, et al. Enhanced CAR T cell expansion and prolonged persistence in pediatric patients with ALL treated with a low-affinity CD19 CAR. Nat Med. 2019;25(9):1408–1414. https://doi.org/10.1038/s41591-019-0549-5

5. Roddie C, Sandhu KS, Tholouli E, et al. Obecabtagene autoleucel in adults with B-cell acute lymphoblastic leukemia. N Engl J Med. 2024;391(23):2219–2230. https://doi.org/10.1056/NEJMoa2406526 

Appendix