Four jobs, four very different evidence bases
"AI for radiopharmaceutical design" is not one capability. It is four, and conflating them is how a strong result in one becomes an implied result in another.
- Generation — propose new ligands for a chosen site on a chosen target. Strongest evidence: designed binders have been made and measured in many published campaigns.
- Affinity prediction and ranking — order candidates before synthesis. Good evidence for ranking within a target; weaker for absolute values across targets.
- Off-target prediction — ask what else a candidate binds, ideally across the proteome. Real methods exist; the ground truth is incomplete, because no assay covers the proteome.
- Biodistribution and dose prediction — ask where the construct goes in a whole body. The least mature of the four, and the one that decides whether a radiopharmaceutical programme survives.
Generative design of peptides, mini-proteins and antibodies
The technical base here is the protein-design literature rather than anything radiopharmaceutical-specific. Diffusion models generate backbones conditioned on a target site; inverse-folding models such as ProteinMPNN assign sequences to those backbones; structure-prediction models score the result. RFdiffusion demonstrated de novo binder design from this recipe in 2023, and macrocyclic peptide binders designed with the same family of methods were reported in 2025 with measured high-affinity binding.
What is new is the application to this modality's constraints. Published 2026 work generated fifty thousand candidate peptides per target with an all-atom binder-design model, filtered them by refolding agreement, physics-based interaction energies and membrane-embedded molecular dynamics, and selected twenty per target — with the wet-lab evaluation reported as still in progress. Other groups combine computational design with selection: an AlphaFold3-assisted phage-display campaign against B7-H3 and DLL3 reported a synthesised peptide at 8.2 × 10⁻⁷ M by surface plasmon resonance with more than three-fold selectivity over homologous proteins.
There is also at least one end-to-end example on the imaging side: a virtual screen of a peptide library produced twenty-three GRPR-targeting candidates, which were synthesised, radiolabelled and evaluated in cells and in animals, with one compound carried forward on improved tumour-to-background contrast. That is the shape of a complete claim — designed, made, measured in vivo, and compared against a reference compound.
Read those three examples side by side and the state of the art is clear enough: generation produces testable molecules, and the campaigns that have closed the loop are small, target-specific and recent.
Affinity, off-target and biodistribution prediction
Affinity prediction has a public benchmark culture and open models — Boltz-2 is the current widely used open example, and it reports its own accuracy and speed limits. The honest summary is that ranking within a target is useful today and absolute cross-target prediction is not yet reliable; what those numbers mean, and on which scale they should be quoted, is a separate explainer here.
Off-target prediction is a harder evaluation problem than a modelling problem. The experimental comparators — selectivity panels, protein arrays and cell-based membrane-proteome arrays — each cover thousands of proteins rather than all of them, so a proteome-wide prediction cannot be scored against a proteome-wide truth. Any recall figure on this task is therefore recall against a known set, and should say so.
Biodistribution is where the gap is widest, and it is the gap that matters most for a radioactive drug. The published approaches are physiologically based pharmacokinetic models and machine learning built on top of them — including PBPK-constrained deep learning for pre-therapy voxel-level dosimetry. Their own authors note that validation is largely in silico or against virtual patients, with clinical voxel-level validation still emerging. A vendor claiming to predict whole-body behaviour is claiming something the peer-reviewed literature has not yet closed.
How to read a vendor's claim
The questions below are the ones that separate a result from a slide. None of them requires disclosing proprietary data to answer.
- Prospective or retrospective? A model that recovers known liabilities from historical programmes is doing a different task from one that called an outcome before it was known. Both are informative; only one is predictive, and the word "retrospective" should appear in the claim rather than in a footnote.
- What is n, and n of what? "100 % hit rate" is uninterpretable without the number of designs taken to the bench, the number of targets, and the binding threshold that counted as a hit. The field's convention is "x of n designs across m targets".
- Which assay, and direct or functional? A Kd from surface plasmon resonance and an IC50 from a cell assay are not the same claim. A number without an assay is not a measurement.
- How was the data split? A model evaluated on ligands or targets similar to its training set will look better than it is. Similarity-aware partitioning, or a test structure that postdates the training cut-off, is what rules memorisation out.
- Does the metric match the decision? Top-k recovery matches "what should I synthesise"; correlation on pKd matches "how good is the model"; a recall figure needs the set it recalls from. A metric that flatters is usually the wrong one.
- What is not reported? Campaigns that failed, targets dropped, and the designs that did not bind. Silence on negatives is the most common defect in this category's marketing.
- Is the endpoint the one that ends programmes? Affinity is the easiest thing to predict and the least common reason a radiopharmaceutical fails. Ask about biodistribution, organ exposure and cell-type risk.
- Is there anything a third party can check? A preprint, a benchmark entry, a deposited structure, an independent evaluation. Until one exists, the number is a company statement — which is not the same as false, but is a different weight of evidence.
The same checklist, applied to this site
It would be poor faith to publish that list and exempt its publisher, so here is where Epitorix sits against it, using only what the site already states.
The models are described as Epitorix's own rather than an assembly of published ones — the science page's words are "We are not just combining published models. We have our own." — and the workspace also exposes fifteen named generator engines, published ones and in-house ones side by side, listed on the engines beat. So both claims are on the table and can be asked about separately.
The figures are fenced, and the fences say what they are. Off-target recall of over 90 % on known off-target interactions is quoted "in retrospective analyses". The up-to-93 % liability-detection figure is quoted "in retrospective analyses of radiopharmaceuticals with documented clinical outcomes". The affinity correlations — Pearson R 0.84 for peptides, 0.70 for small molecules, on predicted versus measured pKd — are quoted "in internal benchmarks". By the first question on the checklist, these are retrospective and internal, and the site says so in the same breath as the number.
Two of the checklist's questions are currently open, and the site does not hide either. The number of designs, targets and the assay threshold behind the 100 % hit-rate figure are, in the site's own words, "not yet published". And the independent-check question is answered by a promise rather than a document: "a preprint describing the evaluation is forthcoming." Until that preprint exists, every figure here is a company statement with its method named — which is exactly the weight a reader should give it.
The relevant limit is also the site's own sentence, printed next to the performance claims: Epitorix is a research and development platform, and its predictions are not intended for patient-level or clinical decision-making. The full set of figures with their qualifiers is on the science page; the report that packages them for a specific candidate is PharmPhax™.
What would move the field
Three things, none of them a bigger model.
Prospective, registered evaluations. A published set of candidates and predicted liabilities, filed before the animal study, read out afterwards. One such record is worth more than a dozen retrospective analyses, and nobody in this category has published one at scale.
Shared benchmarks for the endpoints that matter. Affinity has benchmarks; biodistribution and organ exposure do not. Until a common evaluation set exists for whole-body behaviour, every vendor's number is measured against its own yardstick.
Reported negatives. The attrition literature is unambiguous that toxicity and pharmacokinetics — not potency — are what end programmes. A field that only publishes its successful designs is not yet measuring the thing it claims to fix.
