A repeatable method for optimizing an antibody and placing it in open IP space, with FMC63 as the worked example.

FMC63 is the most widely used anti-CD19 binder in cell therapy: a murine antibody that supplies the binding domain of four of the five approved CD19 CAR-T products (Kymriah, Yescarta, Tecartus, and Breyanzi), which together reported about $3.6 billion in 2025 sales. The fifth, Aucatzyl, uses a different binder. Our generative models can take a molecule this established and make it more human and more stable while holding its CD19 binding. But for a molecule this heavily patented, making the result usable means taking the IP landscape into account from the start, not just optimizing the biology. The patent position often decides whether a candidate can go forward, and it can be established before any sequence is designed.

Our agents assess the molecule from four angles: the patent landscape around it, what the field has already tried, its structure, and its sequence liabilities, and combine them into one brief before we start designing.

Figure 1. Agent to model to experiment. AI agents read the molecule, led by the IP landscape, and distill the reads into a design brief; the generative models design a matched-pair panel; the lab builds it and verifies it against the seed.

The starting molecule, the seed, is the murine FMC63 scFv (9PKSMW007). It binds CD19 with high affinity, but its murine framework carries immunogenicity and developability liabilities. The goal is a humanized version that keeps the affinity, reduces the liabilities, and sits in open IP space.

The preplanning: reading the molecule before design

The agents’ reads fall into three groups: the IP landscape around the molecule, the prior knowledge about it (what the field has tried, and what the structure shows), and the molecule’s own risks (how human it is, and what liabilities it carries). For this molecule, the IP landscape is the largest of the three reads, because the patents are its main design constraint.

The IP landscape around the molecule

FMC63’s sequence and CDRs date to the 1980s, which makes them prior art. No patent claims the binder itself, and anyone is free to build on it. The patents instead claim the frameworks others humanized it onto and the constructs built around it (the CAR or cell therapy it becomes part of). Because the framework and the construct are design choices, the design can route around the claims they carry.

To read the landscape at this depth, we use Converge’s Patent Extractor, a model that parses granted patents into a structured dataset of antibody sequences, their CDRs, and any reported measurements. The structured data lets us screen tens of thousands of CD19 patents and read the handful that matter at the claim level.

FMC63 is widely disclosed but seldom claimed. Its heavy-chain CDR3 appears in 1,005 patent families, most of them near-identical copies of the original. We read the 39 patents whose claims could reach it. Six of them, holding 48 claims, reach the molecule, and every one wraps the binder in a construct or method rather than claiming it on its own.

Figure 2. No patent claims FMC63 on its own. Of tens of thousands of CD19 patents, 39 were read in full; six of them, holding 48 claims, reach the binder, and all wrap it in a construct or method that a program can design around.

A different framework escapes the sequence-based claims. Five patents define a humanized FMC63 by its sequence, most of them held by Promab and ICT. Each specifies a particular human framework or sets a sequence-identity cutoff (80% or 90%) that only the murine original meets. Neither kind of claim reaches a binder that has been humanized onto a different framework.

The field crowded onto one set of frameworks and left another open. Others have disclosed 187 engineered variants of FMC63, 114 of them humanized. Almost all of the humanized variants use VH4 germlines (100 of the 114), and most of those fall in one cluster of near-identical genes, the IGHV4-4 cluster. The likely reason is convenience: VH4 is the closest human framework to the mouse sequence and takes the fewest changes to graft onto. The most convenient framework is therefore also the most patented. IGHV3-11, a VH3 germline further from the murine sequence, was used by none of the 114 and stayed open.

Figure 3. Where the field humanized FMC63. Of 114 humanized patent variants, 85 fall in the IGHV4-4 cluster and none on IGHV3-11. Our designs used that open lane.

The design constraints. The landscape turned into two rules for the sequence design.

• Stay off the patented frameworks. The sequence-based patents claim humanized FMC63 on specific frameworks, all in the crowded VH4 lanes. Building on a framework none of them used clears those claims.

• Stay under the identity threshold. Two patents claim any sequence above 80% identity to the murine original. The designs we produced sit between 71.7% and 77.5% identity to the seed, under that gate.

Both rules point to one framework: IGHV3-11, the open VH3 lane no patent used and far enough from the murine seed to clear the threshold.

Not all of the exposure is in the sequence. The seed also falls under claims tied to specific constructs and methods, such as a spacer, a cell format, or a drug combination. These claims do not depend on the sequence, so the partner’s program clears them through its own build choices.

The prior knowledge: the field’s edits and the structure

The field’s prior edits. The patent record is also a record of experiments. Hundreds of groups have already humanized FMC63. The positions they changed, and the positions they left unchanged, show where the sequence can be edited without losing CD19 binding. Across 104 disclosed variants from 56 patents (those with position-level substitution data), 179 substitutions fall at 32 heavy-chain positions, with none on the light chain. The changes cluster at the edges of the binding loops (K106, V28, P35, L30) and avoid the loop cores. They also concentrate in the framework: most variants fall under 95% whole-VH identity to the seed. One disclosure went further and measured the effect on binding. ICT’s humanization kept 17% of CD19 binding on a plain VH4 graft, 46% on a reworked VH4 version, and 82% once moved to a VH3 framework. In that one case, how much binding was retained depended on the framework, and the VH3 graft retained the most. The lane the IP read had found open, IGHV3-11, is also a VH3 germline.

Figure 4. The field’s changes to FMC63, mapped onto the CDRs. The 179 substitutions across 104 disclosed variants all fall on the heavy chain, at the edges of the binding loops; the loop cores are unchanged.

Figure 5. The one disclosure with measured binding: a VH4 graft kept 17% of CD19 binding, a reworked VH4 version 46%, and a VH3 version 82%, relative to chimeric FMC63.

The structure read. Converge works from the antibody’s bound structure, not its sequence alone. We map each proposed edit onto the antibody-antigen complex to predict its effect on binding before the lab. An edit on a residue that contacts CD19 can abolish binding; an edit away from the interface is lower-risk, though framework changes can still perturb binding through the residues that scaffold the loops. For FMC63 the complex is already solved. The seed’s scFv is deposited bound to human CD19 (PDB 7URV, cryo-EM at 3.05 Å; the seed’s VH and VL match the deposited antibody chain), so every CD19 contact can be read off the model directly. Seventeen residues sit within 4 Å of CD19, and CDR-H3 carries about 47% of the buried interface. A Tyr109-Tyr110 pair makes about a third of the contact, and no disclosed variant has ever changed it. The binding determinant is the CDR-H3 core, which has to be preserved; the loop edges and the framework are where the sequence can change. The field’s most-edited framework positions, K106 and V28, sit 9 to 10 Å from CD19, well beyond the 4 Å contact distance, so the framework can be humanized without touching the interface.

The molecule’s own risks

The last read scores the molecule’s own liabilities so the design can target them: how human the sequence is, how immunogenic it is likely to be, and the chemical liabilities that affect developability. Each is measured on the seed first, then tracked through the design. Humanness is measured the same way throughout, as IMGT per-framework germline identity: 72.5% (VH) and 83.0% (VL) for the seed. ConvergeAB’s predictor finds two CD4 T-cell epitopes in the seed. The seed also carries nine sequence liabilities, three in the CDRs and six in the framework. Most sit in the framework and are removed by the same humanizing edits. The few in the binding-loop cores remain, because those cores are preserved to keep CD19 binding.

The design: from brief to lead

The decision

A scientist set the direction from the brief: humanize the framework into the open IGHV3-11 lane, leave the CDR-H3 core unchanged, and hold CD19 affinity near the parent rather than maximize it (for a CAR-T binder a tighter KD can raise toxicity). The IP read and the binding data both pointed to IGHV3-11, so a single set of framework edits did three jobs at once: humanize the molecule, move it into open IP space, and land on the acceptor family the binding data favor. The decision carried one risk: IGHV3-11 lacks the manufacturing track record of the popular frameworks, so nothing guaranteed it would express well or stay stable. Expression yield and melting temperature were measured on every candidate to confirm it would.

The optimization

ConvergeAB designed the panel zero-shot from the FMC63 sequences and the human CD19 target, optimizing binding, structure, humanness, epitope content, thermostability, and solubility together. Each design was made twice, with and without the CDR-edge edits, so the pair shows what the CDR-edge edits add. The lead (9PKSMW008) carries 45 edits, 42 in the framework and 3 at the CDR edges, leaving the CDR-H3 core unchanged. Its edits stay clear of the CD19 interface: two fall outside it, and the third makes a single atom of contact.

Three edits carry the affinity. The lead’s framework-only twin, humanized the same way, binds CD19 about 17-fold weaker than the parent (3.33 nM); the three CDR-edge edits restore it to 0.184 nM, the parent’s level. A candidate about fourfold tighter was available and not advanced, because the brief called for holding affinity rather than raising it.

Figure 6. Three edits carry the affinity. The lead and its framework-only twin differ only by three CDR-edge edits; the lead holds the parent’s affinity (0.184 versus 0.197 nM) while the twin is about 17-fold weaker (3.33 nM). Apparent KD by SPR.

Verification

Verification checks the designs against the seed, which is run back through as a control. On IP, humanizing onto the unused IGHV3-11 framework clears every claim tied to the binder’s sequence: the lead’s exposure falls from 48 claims to 21, and to 9 in a generic CAR format. The claims that remain are tied to specific constructs rather than to the molecule, and a program clears them by its own build choices.

Figure 7. Humanizing onto an unused framework clears the sequence claims. Of the 48 that reach the seed, 21 remain for the lead, all tied to specific constructs that a program clears through its own build choices.

On immunogenicity, ConvergeAB’s predictor projects both CD4 T-cell epitopes removed, 2 to 0. Humanness and developability rose together: framework humanness 72.5% to 89.0% (VH) and 83.0% to 97.7% (VL), first melt from 58.96 °C to 77.9 °C, SEC monomer from 68.6% to 79.6%. The IGHV3-11 stability risk was real: its candidates had lower first-melt temperatures than the single IGHV3-23 comparator, but the lead still reached 77.9 °C through the CDR-edge edits planned from the start.

The optimized lead

The take-home

The lead is better on every measured property, with CD19 affinity held at the parent’s level, and it falls under far fewer patent claims than the seed. Two candidates can perform the same on the bench and still differ on IP, because the framework a binder is built on determines which claims it falls under.

Both results came from one design pass. Conventional, step-by-step optimization reaches them only over several rounds. Because the patent landscape was read before design and used to choose the framework, the edits that humanized the molecule also placed it in less crowded sequence space. The hardest constraint to change later, the set of claims a sequence falls under, was handled at the start of the project instead of the end. CD19 is among the most heavily patented antibody targets, which makes FMC63 a hard case. A binder’s IP exposure does not show until the landscape is read; the read matters whether a target is crowded or not.

The repeatable process

FMC63 is one worked example; the same process runs on any target and any starting antibody. The models design zero-shot, so a new target needs no training run before design begins. The partner sets the target profile, owns every sequence the process returns, and gets the evidence behind each one.

The IP analysis is a landscape proxy over incomplete patent coverage, not a freedom-to-operate clearance or a legal opinion.