How a radiopharmaceutical delivers radiation to a tumour cell
External-beam radiotherapy aims a beam at a place. A radiopharmaceutical does the opposite: it is injected or infused, travels in the blood, and finds its target by molecular recognition. The National Cancer Institute describes the construct as three building blocks — a radioactive molecule, a targeting molecule that latches onto cancer cells, and a linker joining the two.
The targeting molecule can be a small molecule, a peptide, an antibody or an antibody fragment; the radionuclide is usually held by a chelator, a cage that keeps the metal attached to the ligand while it circulates. Once the construct binds, and in many cases is taken inside the cell, the radionuclide's emission damages DNA at close range.
The approved product labels state this mechanism plainly. Lutetium Lu 177 dotatate "binds to somatostatin receptors with highest affinity for subtype 2 somatostatin receptors", is internalised, and "the beta-minus emission from lutetium-177 induces cellular damage by formation of free radicals in somatostatin receptor-positive cells and in neighboring cells". Lutetium Lu 177 vipivotide tetraxetan links lutetium-177 to a moiety that binds PSMA, "a transmembrane protein that is expressed in prostate cancer", and its beta-minus emission "induces DNA damage which can lead to cell death".
Two properties follow from that last clause, and they shape everything else on this page. The radiation reaches neighbouring cells as well as the bound one — a crossfire that helps in a tumour and hurts in a healthy tissue. And the dose is not set by a machine that can be switched off; it is set by where the molecule goes and how long it stays.
Theranostics: image first, then treat
Because the same ligand can carry either an imaging radionuclide or a therapeutic one, the modality comes with a built-in companion test. The Society of Nuclear Medicine and Molecular Imaging summarises the idea as "see what you treat, and treat what you see": a diagnostic scan shows whether a patient's tumours express the target at all, and only then is the therapeutic version given.
In practice the pair differs by radionuclide, not by the part that binds. A positron emitter such as gallium-68 makes the ligand visible on PET; a beta emitter such as lutetium-177 turns the same ligand into a treatment. Lutetium-177 also emits gamma photons, which is why post-treatment SPECT imaging of a therapeutic dose is possible at all.
This is the oldest idea in nuclear medicine, not a new one. Radioiodine has been used to image and treat thyroid tissue since the 1940s, and the NCI still lists iodine-131 as the systemic radiation therapy most often used for certain thyroid cancers. What changed recently is the range of targets: receptors and surface antigens that can be hit with a designed ligand rather than an element the organ happens to concentrate.
The approved medicines, as worked examples
Three approved products show the span of the modality, and each is a useful reference point because its label and its pivotal trial are public.
Lutetium Lu 177 dotatate is a radiolabelled somatostatin analogue approved for somatostatin-receptor-positive gastroenteropancreatic neuroendocrine tumours in adults and in patients 12 years and older. Its pivotal trial, NETTER-1, randomised 229 patients with progressive midgut neuroendocrine tumours; the estimated rate of progression-free survival at month 20 was 65.2 % in the treatment group and 10.8 % in the control group, a hazard ratio of 0.21 (95 % CI 0.13 to 0.33).
Lutetium Lu 177 vipivotide tetraxetan is described on its own label as "a radioligand therapeutic agent", approved for PSMA-positive metastatic castration-resistant prostate cancer after androgen-receptor-pathway-inhibitor therapy. Its pivotal trial, VISION, enrolled only patients whose disease was PSMA-positive on a gallium-68 PSMA-11 PET/CT scan, and reported longer imaging-based progression-free survival and longer overall survival when the radioligand was added to standard care; grade 3 or higher adverse events were more common with it than without (52.7 % versus 38.0 %).
Radium Ra 223 dichloride sits at the other end of the design space: it is not a ligand-plus-radionuclide construct at all but a calcium mimetic that concentrates in newly formed bone, approved for castration-resistant prostate cancer with symptomatic bone metastases and no known visceral metastatic disease. Its label's first warning is bone-marrow suppression — a reminder that in this modality the organ that limits treatment is a property of where the molecule goes.
How patients are selected with imaging
The companion scan is not optional garnish; on at least one label it is written into the dosing section. The PSMA-targeted therapy's label instructs prescribers to "select patients … using LOCAMETZ or another approved PSMA positron emission tomography (PET) product based on PSMA expression in tumors", and notes that further selection criteria were applied in the clinical studies.
Selection by imaging does three things at once. It confirms the target is present, which a biopsy of one lesion cannot do for a whole body of metastases. It shows the distribution of disease, and therefore how much of the injected activity a tumour burden will absorb. And it gives a baseline against which response can be read later, because the same tracer can be used again.
It also sets a hard limit on the modality that is worth stating plainly: a radiopharmaceutical can only treat what expresses its target. Heterogeneous expression within a patient, or between lesions, is a recognised reason a molecularly precise drug still fails to control disease.
Where computational design fits
Everything above concerns medicines that already exist. The design question is narrower and comes earlier: given a target, what ligand should be made, and which liabilities will it carry when it circulates in a whole body?
That is the question generative and predictive models are now applied to — designing candidate binders for a chosen site, ranking them by predicted affinity, and predicting where they would distribute and what else they would bind. AI in radiopharmaceutical design surveys the model classes and what published evidence exists for each; why off-target binding matters more here covers the liability side.
Epitorix builds in this space. Its own account of what its models predict, and every figure with the qualifier attached to it, is on the science page; the workspace those models run in is EPiQ™, and the plain-language interface over it is Xenoclea. Nothing computational replaces the imaging, the trial or the label: predictions are hypotheses to test, and the site says so.
