Why a radioactive payload changes the question
Off-target pharmacology is usually framed as a binding problem: the drug engages a protein it was not designed for, and something happens because that protein does something. Selectivity panels are built on that logic, and so is most of the industry's screening infrastructure.
Radiation breaks the logic in two places. First, the effect does not require a functional interaction — a beta emitter deposits energy in whatever is within roughly a millimetre of it, and an alpha emitter within a few cell diameters, regardless of what the ligand is doing there. Second, the relevant quantity is time-integrated: a weak, transient interaction that concentrates a percentage of injected activity in one tissue for a day can matter more than a tight interaction that clears in minutes.
So three separate failure modes have to be kept apart. Off-target binding is engagement of a protein other than the intended one. On-target, off-tumour exposure is engagement of the correct target in a healthy tissue that also expresses it. And non-specific retention is accumulation with no molecular target at all — renal reabsorption of a filtered peptide is the textbook case. All three end at the same place: absorbed dose in a tissue nobody wanted to treat.
Which healthy tissues the approved labels actually watch
This is not speculative. The labels of the approved lutetium-177 ligands, and the professional-society guidance written around them, name the tissues.
The PSMA-targeted therapy's own radiation-dosimetry section, built from a 29-patient sub-study of the VISION trial, states that "the organs with the highest radiation absorbed doses are lacrimal glands, salivary glands, large intestine (left and right colon), kidneys, and urinary bladder wall". Its adverse-reaction table lists dry mouth among the most common reactions at any grade, and its warnings section carries myelosuppression and renal toxicity.
The somatostatin-receptor ligand's label reports kidney absorbed dose as the highest of the organs tabulated from NETTER-1, at 0.654 Gy/GBq (SD 0.295). Its warnings include myelosuppression, renal toxicity, hepatotoxicity, and secondary myelodysplastic syndrome and leukaemia, with median times to onset of 29 months and 55 months respectively in the ERASMUS cohort.
The EANM dosimetry committee's recommendations for lutetium-177-labelled somatostatin-receptor and PSMA ligands summarise dosimetry data and radiation-induced effects for exactly this set: kidneys, bone marrow, salivary glands, lacrimal glands, pituitary glands, tumours, and the skin in the case of extravasation. The same document notes that the liver "is generally not considered an organ at risk" for these two ligand classes — a useful corrective, because a liver-first intuition carried over from small-molecule toxicology is simply wrong here.
- Kidneys — filtered peptides are reabsorbed in the proximal tubule; approved labels include a renal-protection measure for the somatostatin ligand for precisely this reason.
- Salivary and lacrimal glands — the dose-leading tissues for PSMA ligands, and the source of the dry-mouth reaction on that label.
- Bone marrow — low absorbed dose per unit activity, but the site of the haematological toxicity that most often limits treatment.
- Bowel and bladder wall — exposed by the excretion route rather than by any target.
On-target, off-tumour: when the target itself is the problem
The kidney case for PSMA ligands is instructive because it is not an off-target effect in the classical sense. PSMA is expressed in the renal proximal tubule as well as on prostate cancer cells; the dosimetry sub-study of VISION discusses renal uptake with that expression in mind. A perfectly selective ligand would still go there.
That is why target choice, not only ligand quality, is a safety decision. A target expressed on a rare but essential healthy cell population will carry its liability into every candidate built against it, no matter how the binder is optimised. The failure shows up late, in animals or in patients, and it is attributed to the molecule when it belonged to the target.
Whole-tissue expression data hides this. A transcript averaged across a biopsy of kidney cortex tells you little about a proximal tubular cell population that is a small fraction of the tissue and a large fraction of the risk.
What "cell-type resolution" means
Cell-type resolution means asking the expression question per cell type rather than per organ: which specific populations express the target, at what level, and are any of them tissues where radiation is poorly tolerated.
The method is borrowed from cell-therapy target triage, where the same problem is older and better documented. Single-cell expression atlases have been used to map the potential landscape of on-target, off-tumour toxicity for CAR-T antigens, and to predict which organs a given engineered therapy would put at risk. The underlying logic transfers directly to a radioligand: the ligand does not care whether the cell it binds is a tumour cell.
What it buys, in practice, is a different shape of answer. Instead of "kidney: some uptake expected", the output is a named cell population with a level attached — which is what a programme needs in order to decide whether to change the target, change the ligand's clearance, or stop.
Epitorix's own account of this is on the science page, where the organ and cell-type map is published together with the note that it shows a predicted profile for one candidate rather than a general ranking for the modality.
What predicting it earlier actually changes
Roughly nine in ten candidates that enter clinical development fail, and in the most-cited recent analysis unmanageable toxicity accounts for about 30 % of those failures, with poor pharmacokinetic or drug-like properties a further 10 to 15 %. The same authors argue that optimisation over-weights potency and selectivity and under-weights tissue exposure — which is the gap this modality feels most sharply.
Moving that question earlier does not remove the liability; it moves the decision. A target with an unavoidable cell-type problem can be dropped before a chemistry campaign. A ligand with a predicted retention profile can be reworked while it is still a sequence. And a candidate that survives the screen enters animal studies with a specific hypothesis to test, rather than with a hope.
In retrospective analyses, Epitorix's proteome-wide screen reports over 90 % recall on known off-target interactions, and its models detected previously reported development and safety liabilities with up to 93 % accuracy in retrospective analyses of radiopharmaceuticals with documented clinical outcomes. Both figures travel with those qualifiers everywhere on this site, and the report built on them is PharmPhax™. They are retrospective, and the site states the limit itself: predictions are hypotheses, not results.
