Do Not Eat Me Is Not a Switch

Phagocytosis is a subtraction, not a switch. A macrophage engulfs when its activating signals exceed its inhibitory ones, and every awkward result in the myeloid checkpoint field falls out of that one sentence. Blocking a don't-eat-me receptor on a target cell that displays no eat-me signal removes a brake that was never opposing an accelerator — which is the most economical explanation for why anti-CD47 monotherapy underperformed and why combinations with opsonising antibodies did not.

Myeloid checkpoints — 36 clickable nodes across the target-cell surface, four inhibitory receptors, the shared ITIM machinery, the activating Fc and complement input, and the phagocyte itself. Open the interactive version to click any protein through to its ELISA kit, biosimilar or In Vivo antibody.

The arithmetic is the mechanism

A macrophage receives no instruction to engulf. It integrates activating input from Fc receptors and complement receptors against inhibitory input from ITIM-bearing receptors, and acts when the balance tips. Everything else on this pathway follows from that.

CD47 blockade alone does little to a cell that is neither opsonised nor already displaying calreticulin, because there is nothing for the released brake to reveal. The converse is the useful half: an opsonising antibody works better when the brake is released, which is why the productive anti-CD47 clinical work was combination work — with rituximab against CD20, and with daratumumab against CD38 — rather than monotherapy.

The same arithmetic accounts for the toxicity. CD47 sits on every nucleated cell and on erythrocytes, so a systemic blockade releases the brake everywhere, and the on-target anaemia that dogged early development was not an off-target effect at all. It also creates an antigen sink large enough to matter for dosing — which is one concrete reason the receptor side is often the better place to intervene.

There are at least four axes, not one

CD47 dominates the literature to a degree its biology does not justify. LILRB1 binds the conserved α3 domain of MHC class I. LILRB2 binds HLA-G among other ligands. Siglec-10 binds CD24. And CD200 signals through CD200R on myeloid cells. Each of these is sufficient to protect a cell on its own.

That redundancy is the most plausible reason single-axis blockade disappoints, and it produces a genuinely awkward corollary at the MHC class I node. Losing class I is the classic immune-escape mutation for CD8 T cells — but it simultaneously removes the LILRB1 brake on the macrophage. The same mutation makes a cell harder to kill by one mechanism and easier to eat by another, which means "MHC-low" is not a single phenotype with a single consequence.

The CD24–Siglec-10 axis was specifically proposed as dominant in ovarian and breast cancer, the settings where CD47 blockade has been least convincing. It is also, as the reagent table below shows, the axis that is hardest to interrogate — which is worth knowing before designing around it rather than after.

Everything converges on two phosphatases, and one behaves backwards

All four inhibitory receptors recruit SHP-1 and SHP-2 to phosphorylated ITIMs. It is SHP-1 that dephosphorylates non-muscle myosin IIA and the signalling intermediates that would otherwise assemble the phagocytic cup, and the motheaten mouse — SHP-1-null, and profoundly inflammatory — indicates how much traffic runs through that single enzyme.

SHP-2 is the complication. In most receptor contexts it is activating; here it is not. That reversal is worth holding in mind whenever a SHP-2 inhibitor developed for a RAS-driven tumour is applied to a model whose myeloid compartment is part of the question, because the prediction from the cancer-cell literature points the wrong way.

Downstream, the route of engulfment matters as much as its rate. MerTK-mediated efferocytosis of apoptotic cells suppresses type I interferon and is actively tolerogenic. A treatment that raises total phagocytosis without directing it can therefore produce more tolerance rather than more immunity — which is why CD8 cross-presentation, not phagocytic index, is the end point this map is built towards.

The activating side is where the functional-grade depth actually is

If inhibition is subtracted from activation, then the activating input deserves equal attention, and here the reagent situation is much better. FcγRII/III covers the mouse activating and inhibitory receptors together — a genuine limitation, since it is an Fc-blockade reagent rather than a dissection tool. FcγRIV is more specific and more useful: it is the mouse receptor with the highest affinity for IgG2a and carries most antibody-dependent phagocytosis in mouse models, so blocking it is how an ADCP-dependent result is separated from a complement-dependent one. On the human side CD16, clone 3G8, pairs with the Fc receptor map in this library to separate ADCC from ADCP.

Complement is the other opsonisation route and is easy to forget in a CD47 experiment. C5 and its receptor C5aR1 prime rather than trigger engulfment, and CD11b is both the complement receptor CR3 and the integrin that forms the phagocytic cup — the point where the antibody arm and the complement arm meet on one molecule. CD18 takes out the whole β2 integrin family at once, which is blunt but decisive when the question is simply whether adhesion is required.

What you can block, and what you cannot

18 of the 36 nodes carry a functional-grade antibody. The distribution matters far more than the count, and it runs contrary to what this field's literature would lead you to expect.

Said plainly: of the four inhibitory receptors this pathway is named after, exactly one can be blocked in vivo. That happens to be the better experiment anyway — blocking SIRPα sidesteps the antigen sink CD47's ubiquity creates — but it is availability rather than design, and presenting it as a deliberate choice would be dishonest. The practical consequence is that most of this map is interrogated from the ligand side and from the phagocyte side, not from the receptors themselves. For the CD24–Siglec-10 axis the ligand side is the only interruptible end.

What the phagocyte becomes, and why the markers mislead

Macrophage polarisation is the part of this field that has aged worst, and it matters here because the markers are so often used as the read-out of a checkpoint experiment. CD206 is induced by IL-4 in culture and does not map cleanly onto any state in vivo. CD163 is the other canonical M2 marker, and its soluble form at least has the advantage of being measurable in serum rather than only in a dish — one of the few nodes here that can be tracked in a patient.

Both are worth treating as descriptive rather than definitional. The M1/M2 scheme was built from polarised cultures and tumour-associated macrophages do not sit at either pole; reporting a CD206 shift as evidence that phagocytosis has been reprogrammed is a claim the marker cannot support on its own. Subset identity is on firmer ground: Ly-6C and CD11c separate classical from non-classical monocytes and tell you which cell arrived, and CD209b marks the marginal-zone population, which is a reminder that phagocyte identity is anatomical as much as molecular.

The therapeutic antigens sit at the same end of the map for a reason. CD33 is gemtuzumab's target and also a Siglec in its own right — both an inhibitory-family member and something a drug opsonises, which is the clearest single illustration of why Siglec biology resists a tidy summary. CD52 is alemtuzumab's target and the most complement-dependent of the depleting antibodies here, which puts it closer in argument to the C5 node than its position suggests. Neither is an intervention point on this pathway; both are what the pathway acts on.

One control worth planning around

Nearly every anti-CD47 experiment confounds two things: blockade of the CD47–SIRPα interaction, and the Fc-mediated effects of the antibody doing the blocking. An IgG bound to a target cell is itself an opsonin, so an anti-CD47 antibody supplies part of the activating signal it is supposed to be unmasking.

The recombinant SIRPα-Fc decoy separates them. It is a soluble SIRPα ectodomain that occupies CD47 without engaging any phagocyte receptor, so the difference between it and an anti-CD47 antibody is the antibody's Fc contribution. Running both is the cleanest way to find out how much of an effect is checkpoint blockade and how much is opsonisation — and given how much of this field rests on that distinction, it is worth the extra arm.

A last note on controls: CSF1R blockade depletes macrophages rather than modulating them. When a checkpoint result might simply reflect a change in macrophage number, that is the arm that settles it — and it is a different experiment from everything else on this pathway.

For research use only; not for use in diagnostic or therapeutic procedures. Explore the full library of interactive pathway diagrams.