Three acts, zooming in. First the tissue, where whether a drug can reach its target at all is decided by architecture rather than chemistry. Then the field at cell scale — who is present, how far apart, what is abundant and what is scarce. Then the close-up, where nine conditions in series decide whether anything happens. Both drawn to scale, running in compressed time.
Everything in acts I and II assumes the antibody can reach the antigen. This act is where that is decided, and it is decided by tissue architecture rather than by chemistry or by expression level.
It is also the cleanest evidence for the quantity this whole program governs on. We argue the variable is accessible local density, not how much the cell makes. Here is a clinical program where how much the cell makes runs the wrong way by a measured 60% — and the drug works.
Not who owns which lever — what moving it would teach us. Grouped that way, the list says where to spend an experiment.
| lever | what moving it tells us |
|---|---|
| 1 · discriminates between the two accounts — worth an experiment on its own | |
| exposure time | The criterion's only non-collinear variable. An equilibrium account contains no clock at all, so changing the timer is the one move the two accounts answer differently. This is arm A, and it is still unrun. |
| anchoring arm k_off promoted this session | The retention advantage goes as 1/k_off,A, so a tenfold slower anchoring arm predicts a tenfold larger advantage with the gating arm untouched. A clean, falsifiable number that no published series has tested. |
| 2 · changes the outcome but not our understanding | |
| engager concentration | The dose axis. Every EC value is a point on it; moving it confirms the curve rather than the mechanism. |
| gating arm k_off | Collinear with KD at r = 0.997, so it cannot be moved independently — and the retention result says it cancels from the advantage anyway. It moves the comparator, which raises a selectivity figure without improving retention. |
| valence and format | Sets the realized avidity exponent — 1.45 to 1.74, not 2. A level effect already inside our sweeps. |
| epitope height | Decides whether an arm is an engager at all, which is decisive but coarse: two clones, no ladder. Informative only if someone builds the ladder. |
| effector:target ratio | Scales contact frequency in vitro. Not a clinical variable, so it cannot carry a design conclusion. |
| 3 · not ours to set — but two are editable with someone else's tool | |
| glycocalyx thickness | Editable. Mucinase-armed effectors raised killing more than threefold. The barrier is a target in its own right, and nobody has combined clearing it with a two-antigen gate. |
| repair capacity | Editable. ESCRT inhibition sensitizes; the 56.7 min damage half-life is its functional measure. Lowering it raises the value of every hit the engager delivers. |
| target antigen density | Our own governing variable, and the tumor sets it. This is the constraint the whole gate exists to work within. |
| CD3 copy number | Sets the occupancy ceiling the engager works against, since the bridge does not turn over. |
| contact duration | Decides 95% of outcomes and appears nowhere in our work. The largest single gap between what we model and what determines the result. |
Group 1 had one entry this morning. The retention result added the second, which is the argument for grouping this way: a lever moves into group 1 the moment someone works out what its answer would decide.
The drug is not the scarce resource — at the concentration that kills there are thirty engager molecules per antigen. What is scarce is contact: two cells touching, in the right geometry, for long enough.
Every player, what it actually is, and how to hold it in your head. No prior biology needed — the numbers are all real and the analogies are chosen to be accurate rather than cute.
Blow a single cell up to the size of a two-story house, 8 m tall. That is a magnification of 800,000×. At that scale, everything else lands here:
Any molecule on the outside of a cell that something else can recognize and stick to. Here they are proteins whose stalks poke out through the membrane.
Think of it as a door handle. A cell's surface is a crowded lawn of tens of thousands of differently shaped handles. An antigen is simply whichever handle your key happens to fit — so "antigen" is a role, not a kind of molecule.
And here is the problem the whole field is built around. Almost no antigen is truly cancer-only. Tumors mostly carry the same handles healthy cells do — just more of some of them. So selectivity is a difference in amount, not in kind, and a drug that targets one antigen will also find the healthy tissue that carries it. That is why trials get halted, and it is the entire reason for requiring two antigens.
On the cell in this simulation: 52,000 copies of CD70 and 46,000 of CD4, spread over 314 µm² — about 166 and 146 per µm².
A lab-made protein with three grabbing hands. Two grab the tumor cell; one grabs a T cell. It has no payload, no toxin, and no activity of its own.
Think of it as a handcuff, not a bullet. It does not kill anything. All it does is hold two cells together long enough that one of them does the killing. The medicine is the T cell — the engager is just the thing that introduces them.
Why two tumor-side hands? Because one is deliberately made too weak to hold on by itself. A cell carrying only one of the two antigens gets grabbed and dropped. A cell carrying both gets held. That is the gate, and it is built out of weakness on purpose.
One more thing it quietly does: a T cell normally only kills what its own receptor recognizes, which tumors escape by hiding their identity markers. An engager grabs the T cell's trigger directly, so that escape route does not work against it.
A white blood cell whose job is to kill cells it has been told are wrong. It carries the weapons; it does not carry the targeting used here.
Think of it as a trained guard who will act on a signal but cannot read the address. The engager supplies the address.
It migrates about 10 µm per minute — one cell spacing every 3 minutes — and over a 6 hour assay it can cross the field roughly 119 times. Finding targets is not its problem.
Part of the T cell's own firing mechanism. Cluster enough of it together and the T cell commits to killing whatever it is touching.
Think of it as a doorbell wired straight to the alarm. It does not check who pressed it.
That is simultaneously the trick and the danger. It is why an engager can redirect a T cell at all — and why an engager that grabs the wrong cell gets a real kill on a healthy one.
An enzyme standing on the T cell surface that continuously erases the "fire" signal.
Think of it as a very tall man in a crowd. When two cells press together to 14 nm, anything 28 nm tall is physically squeezed out to the edges.
This is the most counter-intuitive thing on the page, and it is worth sitting with. A T cell is not switched on by contact. It is held off by default, and signaling happens only in the narrow gap where the off-switch cannot fit. The trigger is not a chemical event — it is a geometric one. Which is why, in this system, distances in nanometres decide outcomes.
A dense fur of sugar chains coating the outside of every cell. Tumors frequently grow theirs thicker.
Think of it as shag carpet glued to the outside wall of the house — about a fist deep — when the thing that has to reach through it is 12 mm long.
It is passive and it may be decisive. It wants nothing and does nothing; it simply occupies the space the 14 nm gap needs. Across a panel of cells engineered to differ in thickness, thickness tracked how well they were killed at r = −0.94. Clearing it with an enzyme raised killing more than threefold.
Our framework has no term for it at all, and it acts before every term we do have.
One honest caveat. Those thicknesses were measured on mucin-expressing epithelial tumor lines. The cells in Act I are lymphocytes, which carry a thinner coat — so this gate bites hardest in the solid-tumor setting rather than in the dataset the rest of these numbers come from. Carrying it across systems is a transfer, and it is flagged as one.
Perforin punches a hole in the target's membrane. Granzyme B goes through the hole and switches on the target cell's own self-destruct program.
Think of it as picking a lock and posting an instruction through the door. The T cell does not demolish the tumor cell — it tells it to take itself apart, using machinery the cell already has for exactly that purpose.
The cell's own membrane patch kit. It detects a wound and reseals it, and it starts immediately.
Think of it as a repair crew that arrives before the burglar has finished. This is why one hit is almost never lethal.
Damage from a single perforin event fades with a half-life of 56.7 minutes. Death needs about three hits landing inside that window, which turns killing into a race between arrival rate and repair rate — and the arrival rate is the thing our model does not contain.
A pair of proteins, one on each cell, that when they meet tell the T cell to stand down.
Think of it as a handshake the tumor offers at the door that means "nothing to see here." The T cell accepts it.
It works after everything else succeeded — the grip, the gap, the trigger. That is why releasing this brake is its own separate class of cancer drug, and why binding perfectly can still produce no killing.
A protein waiting inside the tumor cell that neutralizes granzyme B once it arrives.
Think of it as shredding the instruction after it came through the door.
It is the only failure in this whole sequence where every single step upstream worked. The reach, the grip, the gap, the trigger, the pore, the delivery — all of it succeeded, and nothing happened.
Three ideas carry nearly all the numbers in this program. They are not hard, and the third one is the whole paper.
kon is how quickly two molecules find each other and stick. koff is how quickly they let go once stuck.
The useful trick: 1/koff is the average time they stay together. The anchoring arm here has koff = 4.9 × 10⁻³ per second, so 1/0.0049 = 204 seconds. That is all there is to it.
Affinity is koff divided by kon. It is the single number most of this field reports, and it is reported alone.
Why that is a problem: a ratio cannot tell you how long anything lasted. Two molecules can have identical affinity and completely different stickiness-over-time — one that grabs fast and lets go fast, one that grabs slowly and holds. An equilibrium model cannot tell those apart. A timed one can.
That distinction is what this entire program is about, and it is why the paper's title says selectivity is a statement about time rather than about affinity.
Two cells held together by engager molecules come apart only when every link is simultaneously unbound. That is a countable question, and it has an exact answer.
Think of it as a boat tied to a dock with several ropes, each of which keeps coming untied and being retied. With one rope the boat drifts off quickly. With two it takes much longer, because both must be loose at the same instant. Each rope you add multiplies the time — so the time to drift away grows exponentially in the number of ropes.
And that is why it saturates. We solved this exactly from the published rate constants: at four links, the expected time for all of them to be loose at once crosses the six-hour assay. Below four the conjugate comes apart during the experiment; at four it does not. The fifth engager buys nothing — the cells were already going to stay together for longer than the assay lasts.
So four is not a fitted number or a threshold we chose. It is where an exponential crosses a clock.
Why it matters, and this is the live question. Four links somewhere is a count; four links per µm² needs an area, and nobody has measured the area. Spread over a mature synapse, four links is 0.2 per µm² — while killing is inferred to need 40.16 per µm², two hundred times more. That gap says the kill threshold is measuring something downstream of merely holding the cells together. But squeeze the same contact down to a microcluster and 40.16 per µm² is four engagers — and the gap vanishes entirely. Both readings are alive, they differ by a factor of two hundred, and one image of a real contact decides it.
To fall off, both hands have to let go — and crucially, if one hand lets go, the other is still holding you close enough to grab again almost immediately.
Think of it as climbing with two hands on nearby holds. Losing one grip is not falling; it is a moment of reaching. That is why a two-handed hold lasts not twice as long as a one-handed hold but about 900 times longer.
And this is the result we found while building this page. Work out that ratio and the gating arm's grip strength cancels out of it completely. What is left depends only on the anchoring arm — the hand nobody was adjusting. Both published studies spent their effort tuning the gating arm, so neither of them changed the retention advantage at all. They were improving the comparison, not the grip.
It yields a prediction nobody has tested: halve the anchoring arm's release rate and the advantage should double, with the gating arm untouched.
Our framework models gates 2, 3 and 4 — reach, the second key, and the clock. The other six are where most contacts actually die, and four of the players in them were not on our field at all.
Act I is drawn to scale at cell dimensions; Act II at 2 px/nm both axes with no exaggeration. Cell counts, spacing, engager-per-antigen and diffusion times are computed from the source's stated assay conditions plus geometry. Receptor and engager dimensions from structural data; glycocalyx thicknesses measured on MUC1-induced and SKOV3 cells; contact, hit and killing statistics from intravital and in vitro cytotoxicity work. The death condition simulated is the published rule applied directly — three hits with each interval under 50 minutes. Nothing here was measured by us.