Fifteen Drugs by Blocking. None by Switching On.

What the TNF receptor superfamily does: live-or-die signals

The tumour necrosis factor receptor superfamily is about two dozen related proteins sitting on the surface of cells. They do not carry information about what a cell should attack; that is the job of the antigen receptor. What they carry is a verdict on the cell itself: live or die.

Concretely, the family splits by what the receptor has attached to its inner end. One group carries a stretch of protein called a death domain. Trigger those, and the receptor recruits machinery that activates caspases, the enzymes that dismantle a cell from within. TNFR1, Fas and the TRAIL receptors work this way. Their message is: destroy yourself.

The other group has no death domain. Instead it binds adaptor proteins called TRAFs, which switch on the transcription factor NF-κB, which in turn tells the cell to make survival proteins that block the suicide machinery. OX40, CD40, 4-1BB, CD27 and GITR are in this group, and they sit almost exclusively on immune cells. Their message is the opposite: keep going, divide, stay.

This is what makes them such tempting drug targets. A T-cell responding to a tumour either fades out within days or becomes a lasting population, and this family is what decides which. The antigen receptor decides what a T-cell recognises. This family decides whether it is still there next month.

Three ways to drug a TNFRSF receptor — and the two that work

Block the ligand. Rather than touching the receptor at all, you remove the ligand that would have activated it, so the signal never arrives. Adalimumab and infliximab soak up TNF, which is how they treat rheumatoid arthritis, Crohn’s disease, ulcerative colitis and psoriasis. Denosumab soaks up RANKL and is used for osteoporosis. Belimumab soaks up BAFF, for lupus and lupus nephritis. Etanercept does the same job with a decoy receptor, a piece of the real receptor floating free to absorb TNF before it reaches a cell. In every case the receptor is left untouched and simply never fires.

Use the receptor as an address. Here the signalling is irrelevant. The receptor is a marker that happens to sit on a cell you want destroyed, and the antibody is a delivery label: attach a toxin to it, as brentuximab vedotin does for CD30 in Hodgkin lymphoma, or use it to drag a T-cell into contact, as teclistamab and cilta-cel do against BCMA in multiple myeloma. The receptor is a marker, not a switch.

Switch the receptor on. Bind it and deliberately activate it. On a costimulatory receptor that tells an immune cell to survive and multiply; on a death receptor it tells a tumour cell to destroy itself. This is the approach with the most upside, because it drives an immune attack instead of merely lifting a brake.

The first two have produced fifteen medicines. The third has produced nothing, in twenty years of trying.

Pfizer, Roche, GSK, AbbVie, Bristol Myers Squibb and AstraZeneca have between them taken nineteen agonist programmes into trials against six of these receptors.

The split holds receptor by receptor, as the table below shows: everything in the top half reached the market, and nothing in the bottom half did.

Agonist trials build through the 2010s to twenty-one a year, then fall away. The approvals sit at either end of the chart, and none of them are agonists.

There is a long gap in the middle. Between belimumab and brentuximab vedotin in 2011 and belantamab mafodotin in 2020, this entire protein family produced nothing for eight years. The agonist wave sits inside that gap almost exactly.

The obvious reading is that the field had run out of easy ligands. TNF, RANKL and BAFF were taken, agonism was the next idea, and a decade went into it without result. What ended the drought was not agonism finally working but a third approach becoming possible: from 2020 the approvals are almost all BCMA drugs that use the receptor as an address, which needed CAR-T and T-cell-engager technology that did not exist in useful form when the agonist programmes began.

The trials came in waves. TRAIL-receptor agonists ran hardest from 2008 to 2013 and then thinned to almost nothing. The CD40 and OX40 agonist programmes ran from roughly 2014 to 2021 and then shrank to a handful a year. Each time, several companies moved at once, spent five or six years, and left. Then in 2023 and 2024 the field came back to OX40 from the opposite direction, blocking it instead of pushing it. That wave is not in these bars, which count agonists only.

Why TNFRSF agonist antibodies fail: the clustering problem

The problem is that an agonist antibody cannot finish its own job. These receptors do not signal when a single molecule is engaged. They signal when several are clustered tightly together on the cell surface. A conventional antibody has two arms, so at best it can bring two receptors into proximity, which is not enough. The clustering has to come from somewhere else.

In practice it comes from a second receptor on neighbouring cells, FcγRIIB, which grabs the tail of the agonist antibody and holds many of them together, so that the receptors underneath are pushed into a working cluster. FcγRIIB contributes nothing but scaffolding: experiments using only its external portion, with its own signalling machinery removed, still produce agonism.

That creates the problem. If the clustering is supplied by FcγRIIB, then how well an agonist antibody works depends on how many FcγRIIB-bearing cells happen to be present where the drug ends up. Cellular distribution matters as much as binding affinity. That density differs between tissues, between patients, and between mice and people.

So the drug's potency is partly a property of the tissue rather than of the molecule. Animal models mislead, the therapeutic window shifts depending on where the drug lands, and no amount of affinity engineering fixes it. A ligand-blocking antibody has no such dependency: it binds its target and the job is done.

Agonism works when the receptor is easier to switch on

Before blaming agonism in general, it is worth checking whether agonist biologics work anywhere. One does, and the difference explains why.

Romiplostim has been approved since 2008 and does exactly what the failed drugs were trying to do: it binds a receptor and switches it on, in this case to make patients produce platelets. Two things about it matter. It is not an antibody but a peptibody, two antibody Fc domains each carrying receptor-binding peptides. And the thrombopoietin receptor it targets is activated by dimerisation: bring two copies together and the signal fires.

A molecule with two arms can bring two receptors together, which is all this receptor asks for.

TNF-family receptors ask for something else. They need higher-order clusters, and recent work suggests the requirement is more specific than simply "more receptors": experiments comparing antibodies that differ only in their backbone dispel the idea that bigger clusters produce stronger agonism, pointing instead to receptor density and the resulting super-structure; the same work finds the IgG2 backbone clusters receptors more tightly than IgG1 and is correspondingly more agonistic. A structural study of GITR showed the antibody working by clustering the receptor rather than by mimicking its ligand. The most active antibody in that study produced a distinctive rod-shaped assembly. Antibody fragments with a single arm do not work at all; only bivalent antibodies do, and even then only with help.

The task is not to bind the receptor tightly. It is to arrange several copies of it into a particular pattern. The cluster experiments make the point directly: antibodies identical except for their backbone produce different arrangements and correspondingly different agonism, and it is receptor density and super-structure that predict activity rather than how tightly the antibody binds. That is why better affinity does not rescue these programmes, and why the same molecule can look potent in a mouse and inert in a person: the arrangement depends on the surrounding cells, not only on the drug.

What it costs: two antibodies against 4-1BB

What that looks like in patients is best shown by a pair of antibodies against 4-1BB, developed nearly twenty years ago.

Urelumab was a potent 4-1BB agonist and it worked, in the sense that it activated T-cells vigorously. It also caused severe, dose-dependent, on-target liver injury. Two patients died, at 1 mg/kg and 5 mg/kg, and all trials were halted in 2008. Development later resumed at doses low enough to be safe, at which point the anti-tumour results were disappointing; the largest of those restarted studies, NCT02253992, was itself terminated.

Pfizer's utomilumab was the considered response: a different binding site and a gentler design, built so it would not do that. It was safer, exactly as intended, and caused no significant toxicity. It also produced very limited efficacy, even in combination with pembrolizumab, and was eventually discontinued. Its Phase 3 in combination with avelumab, NCT02951156, was terminated, as was the Phase 1/2 NCT02554812.

Urelumab was potent enough to work and potent enough to kill two patients. Utomilumab was safe and did very little. Nobody found a dose or a design in between.

The one place TNFRSF agonism worked: 4-1BB inside a CAR-T

There is one place where this family's activating signal has been used successfully.

4-1BB is a member of the superfamily. Its signalling domain is built into approved CAR-T therapies including tisagenlecleucel and lisocabtagene maraleucel, where it is one of the two standard costimulatory domains that keep engineered T-cells alive and working.

But those products never press the receptor from outside. They take its internal signalling tail and wire it directly into an engineered receptor built inside the cell, so the clustering problem never arises. They use the signal without ever touching the receptor, and it is the only version of this that has worked.

OX40: the current test case for the whole pattern

The clearest current illustration is OX40, a costimulatory member of this family that has been drugged from both ends. Ten companies tried to switch it on for cancer between 2012 and 2025, and none produced a drug, exactly as this pattern predicts. Then the field inverted the idea and blocked it instead, for eczema, which got two programmes all the way to Phase 3. The switch was fast: seventeen OX40-blocking trials started in 2024, against three agonist trials that year.

Both were discontinued in 2026, four months apart, and both had met their Phase 3 endpoints. Rocatinlimab was stopped in March over a malignancy safety signal; amlitelimab was stopped in July because its benefit over the current standard of care was judged insufficient. One did too much, the other too little. The shape is the same as urelumab and utomilumab fifteen years earlier, but the cause is not: those two were fighting the clustering problem, while rocatinlimab and amlitelimab were making a different trade, between destroying an activated T-cell and merely quietening it.

Next-generation TNFRSF agonists: building the clustering in

The field has drawn the conclusion. If the weakness of an agonist antibody is that it depends on the patient’s own FcγRIIB to cluster the receptor, then stop depending on it: build the clustering into the molecule, or make it happen only where a tumour marker holds the drug in place.

That generation is in the clinic now. Eighteen trials of multispecific TNFRSF agonists have started, most of them pairing 4-1BB or CD40 with a tumour marker such as PD-L1, HER2 or FAP so the clustering happens at the tumour and nowhere else. Ten are still recruiting or active. INBRX-106, which carries six binding sites instead of two and so brings its own clustering, started its most recent trial in September 2024.

The early attrition is not encouraging: several of the PD-L1 and 4-1BB bispecifics have already been terminated. But these molecules bring their own clustering instead of borrowing the patient's.

Six receptors, no approved agonist antibody. What has changed is the diagnosis: the binding was never the hard part, the clustering was. The molecules in trials today are the first built to supply that clustering themselves.