Nine T-cell engagers have been tried in people. They range from durable responses to a treatment-related death. The outcomes are public, the stated reasons are public, and every one of them closes at a specific gate in the cascade.
evidence grade carried on every row · no outcome here is modeled or estimated by us · all three unvetted rows now closed
the finding, and it is not what I expected
Location is not the dominant variable. Antigen restriction is. The reasonable guess is that tissue environment decides — stroma in pancreas, pressure, perfusion, the blood–brain barrier. Those are all real. But the field's own review attributes the spread in outcomes mainly to which antigen happens to be tumor-restricted, and secondarily to half-life and immunogenicity. Stroma and immune suppression are named qualitatively and left unquantified.
That reorders the question. "Why does this work in prostate and not in colon?" is answered less by where the tumor sits than by whether anything on its surface is only on the tumor — which is the problem a two-antigen gate exists to solve, and the reason this program is about selectivity rather than potency.
Every number is a published clinical result. The gate line is ours — it maps each stated reason onto the cascade.
16% partial response, 29% stable disease. The most encouraging solid-tumor engager result in the set.
Why: DLL3 is highly expressed in small cell lung cancer and minimally in normal lung. That is as close to a tumor-restricted single antigen as this field gets.
No gate closes. The antigen does the work a two-antigen gate would otherwise have to do.
Pasotuxizumab: 2 long-term responses of 16 patients, at 14 and 19 months. Acapatamab: PSA reductions in 15 patients (63%), six exceeding 50%.
Why: PSMA is overexpressed in most prostate cancers. Again an antigen story, not a location story.
No gate closes for the responders. What limits it is not stated quantitatively.
Limited data, and the source is internally inconsistent on the denominator — its table says 7 patients, its text says 14. We report that discrepancy rather than picking one.
Why limited: standard engagers do not cross the blood–brain barrier, which is why half-life-extended formats are called unsuitable for glioma.
Closes before gate 1. The drug does not arrive. Nothing downstream is tested — reach, grip and trigger are all irrelevant if the molecule is on the wrong side of a barrier.
What opens it: a transport solution, not a binding one. Brain-shuttle architectures exist for exactly this and are a separate engineering problem.
28% stable disease and no responses at all — the least responsive target in the set.
Why: short half-life, and anti-drug antibodies in 49% of patients. Half the cohort's immune systems neutralized the drug.
Closes before gate 1, for a different reason. Not a barrier but a clock: the molecule is cleared or neutralized before enough of it is present long enough to matter. Our own criterion is about exposure windows, and this is that same variable failing at the whole-body scale.
What opens it: format and immunogenicity engineering. Note this is not a selectivity failure, so a two-antigen gate would not have helped.
2% partial response, 31% stable, 52% progressive. And 15 of 65 patients (23%) had dose-limiting toxicities, with one treatment-related death from diarrhea.
Why: EpCAM is on normal bile duct and gastrointestinal epithelium. The drug worked. It worked on the wrong cells.
Every gate opened — on healthy tissue. This is not a failure of reach, grip, trigger or payload. All of it succeeded, in the liver and gut.
What opens it: this is the case for requiring two antigens, and it is the clinical evidence behind this program's premise. It is also the cleanest demonstration of why “antigen” is a role rather than a kind of molecule: EpCAM is not a cancer molecule, it is an epithelial molecule that tumors also carry. Selectivity here is a difference in amount, and a single-antigen drug cannot read an amount.
The reference success, and worth stating why it is the easy case rather than the clever one.
Why: the target is in blood. No stroma to cross, no interstitial pressure, no glycocalyx barrier of consequence, uniform antigen, and the effector cells are already in the same compartment as the target.
The transport gates pass trivially. Every barrier that closes the solid-tumor cases is simply absent. CD19 is also lineage-restricted, and losing the whole B-cell lineage is survivable in a way that losing bile duct epithelium is not.
This is the useful part: an environment does not invent new physics. It changes which gate is limiting — and therefore which intervention could open it.
| barrier | gate it closes | what opens it | evidence |
|---|---|---|---|
| Blood–brain barrier | before gate 1 · arrival | transport architecture; binding changes cannot help | clinical |
| Short half-life, anti-drug antibodies | before gate 1 · persistence | format and immunogenicity engineering | clinical, 49% |
| Stromal matrix excluding immune cells | before gate 1 · effector arrival | stromal or fibroblast targeting. Distinct from drug transport above: that barrier is quantified, this one is about getting cells in and is still only named | named, not quantified |
| Thick mucin glycocalyx | gate 1 · reach | enzymatic clearing — raised killing >3× in vitro | in vitro, r = −0.94 |
| Antigen not tumor-restricted | gate 3 · the second key | a second antigen requirement — this program | clinical, incl. 1 death |
| Antigen density too low | gate 3 · capacity, not kinetics | Each bridge consumes one antigen of each kind, so bound engager cannot exceed the scarcer density: σ > N₂₀ or no dose reaches threshold. On these cells the margin is 3.6×, and the required dose diverges as density falls toward it — a 73% drop makes the cell unkillable at any dose. Partner-capture by diffusion is excluded as the limiter: those margins are 681–61,609×. And it has been measured: a 110-line screen puts the threshold at ~10,000 binding sites/cell, with no killing of primary epithelial cells below 2,000 — a cliff at fixed dose, which is the signature of capacity rather than potency. | measured: ~10,000 sites/cell across 110 lines |
| MHC class I loss | no gate — engagers bypass it | nothing needed; this is an engager advantage | mechanistic |
| PD-L1 / exhaustion | gate 6 · trigger reversed | checkpoint blockade in combination | drug class exists |
| ESCRT membrane repair | gate 9 · damage decays | repair inhibition sensitizes | in vitro |
| Serpin B9 | gate 8 · payload neutralized | unknown to us | mechanistic |
| Penetration and the binding-site barrier | before gate 1 · arrival depth | Biot number ~0.006: tumor surface concentration sits 100–1000× below plasma before any tissue is entered. Depth goes as √dose — 4× the dose only doubles it. Needs Γ < 1 and φ² < 1 together. Uptake 20–40 %ID/g for IgG against 1–5 for fragments. Not a diffusion problem — unbound IgG crosses the 100 µm containing >95% of viable cells in ~3 min. The barrier is the boundary condition and the binding itself. | quantified |
| Soluble shed antigen as a decoy sink | before gate 1 · drug consumed | Solved by epitope choice. A clinical-stage CEA engager targets a membrane-proximal epitope, so it binds membrane-anchored antigen preferentially over shed. Binding to target cells is unaffected up to 5 µg/mL of soluble antigen. Note the same variable — epitope height — also decides whether an arm can nucleate a synapse at all: one design decision buys both. | clinical-stage construct |
The matrix generates combinations rather than guesses, because a closed gate names its own opener.
Gate 1 and gate 3 are independent failures with independent fixes. Mucin-high tumors — ovarian and pancreatic are the standard examples — close gate 1 sterically, and that is exactly where a gate-3 solution does nothing at all.
Why it is worth stating: enzymatic clearing raised killing more than threefold on its own, and a two-antigen gate raises selectivity independently. Neither addresses the other's failure, which is the signature of a combination that multiplies rather than overlaps.
Honest status: a reading-level observation, not a result. We have not searched exhaustively for prior art on the combination, and that search is the next step before the idea is worth anything.
The transport literature's own design conclusion is that lower-affinity antibodies dissociate before internalization and penetrate more uniformly — the only way past the binding-site barrier. Our retention result says the opposite about one specific arm: the advantage goes as 1/k_off,A, so a slower anchoring arm gives a larger advantage.
These are arguments about the same rate constant, and neither literature cites the other.
Why a two-antigen gate may be unusually well placed: a gate has two target-side rate constants where a conventional antibody has one — and the retention advantage is set by the anchoring arm alone, because the gating arm cancels out of it. So in principle the gating arm is free to be weak, which is what transport wants, while the anchoring arm is strong, which is what retention wants. A single-arm molecule cannot occupy both positions at once.
Honest status: a structural argument, not a result. It needs the transport arithmetic done on a two-armed molecule with the arms treated separately — which neither literature has done and we have not either. What makes it worth recording is that it is checkable, and that it falls out of the cancellation result rather than being imported.
The reason a matrix like this is worth building is that it refuses the vague answer. "The tumor microenvironment is hostile" explains nothing. "The engager's arms are 15 nm and the barrier is 100 nm" explains exactly one gate, and names its own fix.
Clinical outcomes, target antigens and the stated reasons for success or failure are from a published review of bispecific T-cell engagers in solid tumors; glycocalyx and repair figures from the in vitro work cited elsewhere in this program. The gate assignments are ours. Where a row says we have not vetted the numbers, we have not: it is listed so the gap is visible rather than filled. Nothing here was measured by us, and no outcome is modeled or estimated.