Humanization STEM™ Developability pH-dependent release
We rebuilt one of the world’s best-selling antibodies — and published what went wrong on the way
Infliximab is clinically validated, commercially enormous, and biophysically awkward: a mouse–human chimera with documented Fab–Fab self-association. We humanized it, matured it with STEM™, and found that maturation alone did not fix its self-association. This is what we did next, with every number.
All data from our preprint: doi:10.64898/2026.08.18.745573 (bioRxiv, 19 August 2026). Preprint — not certified by peer review.
Why we picked a molecule we could not flatter
Most CRO case studies use a target the CRO chose because it was winnable. We wanted the opposite: a molecule whose weaknesses are already in the literature, whose benchmark is a marketed drug, and where anyone can check our claim against a competitor they already know.
Infliximab qualifies on all three counts. It is a mouse–human chimeric anti-TNFα IgG1 with murine variable domains, a known driver of human anti-chimeric antibody responses. It undergoes reversible self-association through Fab–Fab interactions, which is documented in structural and solution studies and is exactly the property that makes a molecule hard to formulate at high concentration. And it requires prolonged intravenous infusion.
In other words: a clinically proven binding site attached to a body that a modern developability panel would flag. That is the situation many of our clients are actually in.
What we did, in three moves
1. Humanization
The infliximab CDRs were grafted onto human germline acceptor frameworks. We selected IGKV6-21 and IGHV3-72 by IgBLAST on the basis of sequence identity and preservation of canonical loop structure. The resulting humanized antibody, hInfliximab, retained TNFα binding but lost potency — IC50 166.1 ng/mL against 96.8 ng/mL for the chimeric parent.
That drop is normal and it is the reason humanization and maturation are usually scoped together rather than sequentially.
2. STEM™ affinity maturation
We applied STage-Enhanced Maturation to hInfliximab. Focused libraries were designed at preselected CDR positions using the amino acid usage and positional frequencies observed in human antibodies, biasing the sequence toward a human-like profile rather than away from one. Library complexity was held within the electroporation capacity of roughly 1 × 1010 cfu.
Every selection round was heat-challenged and subtracted against a panel of polyspecificity reagents — baculovirus particles, dsDNA, ovalbumin, HSP90-Fc and KLH — before panning on TNFα. The matured clone hInBG4 recovered and exceeded parental potency at IC50 42.0 ng/mL.
3. Targeted CDR substitution — the step we would rather not have needed
Developability profiling then told us something we did not want to hear. hInBG4 had inherited the parental self-association: an AC-SINS Δλmax of 9.00 nm, against 19.33 nm for infliximab itself. The affinity work had succeeded; the biophysics had not.
So we treated hInBG4 as a new scaffold, built a panel of targeted single and combined CDR substitutions, and profiled every variant on four orthogonal readouts — AC-SINS, BVP ELISA, SE-HPLC and transient expression. Three cleared all four.
The result
| Clone | TNFα neutralization IC50 (ng/mL) | Self-association AC-SINS Δλmax (nm) | Non-specificity BVP ELISA ratio |
|---|---|---|---|
| LW2Y | 34.5 | −1.67 | 1.5 |
| LW2YHR1K | 33.2 | 0.33 | 1.3 |
| LW2YR2S | 45.4 | −0.33 | 1.5 |
| hInBG4 — STEM™ output | 42.0 | 9.00 | 2.3 |
| hInfliximab — humanized parent | 166.1 | −4.00 | 1.1 |
| Infliximab — chimeric parent | 96.8 | 19.33 | 1.3 |
| Adalimumab — benchmark | 167.0 | −1.33 | 1.6 |
LW2Y neutralizes TNFα 2.8× more potently than chimeric infliximab and 4.8× more potently than adalimumab, while self-associating less than either.
SE-HPLC showed sharp, symmetrical monomer peaks for the engineered variants where infliximab and hInBG4 both tailed, with NISTmAb RM 8671 run as a reference standard. Transient expression in HEK293 rose from roughly 50 µg/mL for the humanized parent to the high 70s and 80s for the engineered clones.
What each assay is actually predicting
| Assay | What a bad number would cost you later |
|---|---|
| AC-SINS | Weak self-interaction correlates with aggregation, high viscosity at concentration, and poor colloidal stability — the difference between a subcutaneous product and an infusion. |
| BVP ELISA | Polyspecificity against a heterogeneous membrane surface; high ratios track with fast clearance and poor PK. |
| SE-HPLC | High-molecular-weight species and peak tailing — the aggregate burden a regulator will ask about. |
| Transient expression | A first, cheap read on whether the molecule will express at manufacturing scale. |
pH-dependent antigen release, engineered rather than traded for
An antibody that holds antigen tightly in plasma but lets go in the acidified endosome can be recycled by FcRn and bind again. Where a programme is limited by target-mediated drug disposition, that behaviour can reduce antigen accumulation and, in some cases, dose and dosing frequency.
The published attempts at pH-dependent anti-TNFα antibodies have generally used histidine scanning, and in some cases the pH switch came at the cost of neutral-pH binding. The point of the readout below is that LW2Y did not pay that price.
| Clone | KD (nM) pH 7.4 | ka (M−1s−1) | kdis (s−1) pH 7.4 | kdis (s−1) pH 5.8 | Fold |
|---|---|---|---|---|---|
| LW2Y | <0.001 | 1.06 × 105 | <1.0 × 10−7 | 1.80 × 10−4 | >1800 |
| LW2YR2S | 0.056 | 1.74 × 105 | 9.82 × 10−6 | 1.80 × 10−4 | 18.3 |
| LW2YHR1K | 0.539 | 1.76 × 105 | 9.49 × 10−5 | 1.77 × 10−4 | 1.9 |
| Adalimumab | 0.012 | 1.72 × 105 | 2.01 × 10−6 | 4.35 × 10−4 | 216 |
| hInfliximab | 0.752 | 2.12 × 105 | 1.60 × 10−4 | 2.82 × 10−4 | 1.8 |
| Infliximab | 1.009 | 2.17 × 105 | 2.19 × 10−4 | 1.51 × 10−4 | 0.69 |
Measured in-house by biolayer interferometry on the Octet platform. TNFα was captured on streptavidin biosensors, association was performed at pH 7.4 in every case, and only the dissociation buffer was changed. LW2Y’s off-rate at pH 7.4 fell below the assay’s reliable detection limit, so its fold-change is reported as a lower bound.
Immunogenicity: what we can and cannot claim
We analysed the VH and VL sequences of infliximab, hInfliximab and every engineered variant with the IEDB Deimmunization tool in peptide mutant prediction mode, at a median percentile rank threshold of 20.
One peptide in the light chain of chimeric infliximab (DILLTQSPAILSVSP, median percentile rank 16.5) met the threshold as a predicted MHC class II–restricted epitope. No peptide met the threshold in the light chains of hInfliximab, hInBG4, LW2Y, LW2YR2S or LW2YHR1K, nor in any heavy-chain sequence.
Within the limits of an in silico method: humanization removed one predicted T-cell epitope and the CDR-level engineering introduced no new high-risk ones. That is a sequence-liability result, not a clinical immunogenicity result, and we will not present it as one.
What this demonstrates — and what it does not
This programme was run to test our own platform end to end on a molecule that would not flatter it. It shows three things we can do on your antibody:
- Humanize onto human germline frameworks without stranding the paratope — and recover the potency the graft costs.
- Mature affinity while screening developability inside the campaign, not after it, so polyreactive and thermally fragile clones never reach the shortlist.
- Engineer a specific biophysical property on demand — here pH-dependent release — and verify it kinetically rather than infer it.
It also shows the honest shape of the work. STEM™ did not fix everything in one pass. Self-association survived maturation and had to be engineered out deliberately in a second round guided by measurement. Any platform that claims otherwise on every molecule is describing a marketing process, not a laboratory one.
LW2Y is presented here as evidence of capability. It is a demonstration lead from an internal programme, not a product offered for licensing on this page.
Related
Bring us the antibody you cannot fix
If you have a lead that binds well but fails developability screening, a chimeric or murine antibody that needs to become human, or a programme limited by target-mediated clearance, the scoping call is with the PhD scientist who would run the work — not a salesperson.
Talk to a scientist about your lead →
Free 30-minute call with the PhD scientist who would run the work · Initial feasibility assessment within two business days · NDA available before you disclose your target