The clinical logic behind the probe rests on the growing importance of PARP inhibitors in oncology. Drugs such as olaparib, rucaparib, niraparib, and talazoparib have been approved by the FDA and the European Medicines Agency for multiple cancer types, including breast, ovarian, prostate, and small-cell lung cancers. PARP-1 is the most abundantly expressed member of the PARP family, accounting for more than 90 percent of the family’s DNA repair functions, and its activity is heightened in many tumors with dysregulated DNA damage repair pathways. Because tumor PARP-1 expression may predict sensitivity to PARP-targeted therapy, the researchers argue that a noninvasive way to measure that expression across the whole body could guide patient selection before treatment and quantify tumor response during and after therapy.

Today, assessing PARP protein levels relies on pathological biopsy, which the authors describe as the clinical gold standard. But biopsies have well-known drawbacks in this context. Tumors are heterogeneous, so a sample from a single lesion may not reflect gene expression patterns at other metastatic sites, and obtaining reliable tissue from organs such as the lungs or brain requires highly invasive procedures. Radiolabeled PARP inhibitor derivatives, imaged with positron emission tomography (PET) or single-photon emission computed tomography (SPECT), offer a systemic alternative. Existing compounds such as [18F]FTT and [18F]PARPi have already shown clinical value in multiple trials, but all known PARP-1 radiotracers accumulate strongly in the abdomen, particularly the liver, spleen, kidneys, and intestines, which can obscure lesions in exactly the organs where breast, ovarian, prostate, and pancreatic cancers commonly spread.

The innovation in this study lies in how the new tracer was designed. Rather than relying purely on experience-driven, trial-and-error optimization, the team built a multi-stage AI prioritization workflow. Starting from fuzuloparib, China’s first independently developed PARP inhibitor, approved in December 2020 and highly potent against tumor cells with an IC50 of 1.46 nanomolar, they used a SMILES-based conditional chemical language model, a recurrent neural network implemented in the Lib-INVENT framework, to generate candidate analogs. Two scaffold-constrained decoration runs, one attaching substituents directly to the triazole-containing ring system and one using a methylene-extended attachment point, produced a compact set of seven analogs bearing methyl, ethyl, propyl, amino, aminomethyl, trifluoromethyl, or phenyl groups.

Computational triage then narrowed the field. A synthetic complexity model, SCScore, trained on reaction data scored the candidates between 4.6 and 4.8, with compound 005, named NH2-FZPF, receiving the lowest value of 4.6, indicating the lowest predicted synthetic difficulty. Its derivatization-ready amine and the commercial availability of a key amine-containing building block made it the practical choice. Convolutional neural network-guided docking against the human PARP-1 crystal structure (PDB ID 5DS3) showed that the fuzuloparib-derived cores of the resulting gallium-complexed ligands occupied the same binding pocket as the parent drug, sharing interactions with key residues S904 and G863. A deep neural network trained on 7,546 ChEMBL targets using combined ECFP4 and MHFP6 fingerprints ranked PARP-1 as the top predicted target for fuzuloparib, NH2-FZPF, NOTA-FZPF, and DOTA-FZPF alike, with confidence scores between 0.98 and 1.00, suggesting that adding the chelator did not compromise selectivity. Explicit-solvent molecular dynamics simulations in GROMACS, run for 100 nanoseconds, confirmed that the protein backbone converged and the complexes remained stable, even though the bulky DOTA conjugate showed greater ligand-level fluctuations than the parent compound.

The two lead candidates, [68Ga]Ga-DOTA-FZPF and [68Ga]Ga-NOTA-FZPF, were then labeled with generator-produced gallium-68 and rigorously characterized. Radiochemical purities reached 99.35 plus or minus 0.56 percent and 98.48 plus or minus 1.22 percent, respectively, with molar activities of roughly 7 to 11 GBq per micromole. Both tracers proved highly hydrophilic, with log D7.4 values of -1.15 for the DOTA version and -2.23 for the NOTA version, and retained more than 95 percent integrity after 180 minutes in saline, mouse serum, and human serum, as well as after three hours of in vivo metabolism in mice. In cell experiments, both tracers were taken up significantly more by 22Rv1 prostate cancer and A2780 ovarian cancer cells, which express high levels of PARP-1, than by AsPC-1 pancreatic cancer cells with low PARP-1 expression, and co-incubation with a 500-fold excess of fuzuloparib blocked that uptake, confirming specificity.

Micro-PET/CT imaging in tumor-bearing mice delivered the decisive preclinical result. One hour after injection, [68Ga]Ga-DOTA-FZPF accumulated at 5.03 plus or minus 0.77 percent of the injected dose per gram in 22Rv1 tumors and 5.27 plus or minus 0.40 percent in A2780 tumors, significantly higher than the 1.90 plus or minus 0.10 percent seen in low-PARP-1 AsPC-1 tumors. Tumor-to-muscle ratios reached 14.80 plus or minus 8.80 in the 22Rv1 model, and blocking experiments cut tumor uptake dramatically, from 5.03 to 1.53 percent ID/g in 22Rv1 mice and from 5.27 to 1.83 percent ID/g in A2780 mice. Critically, the DOTA tracer outperformed its NOTA sibling on every front: higher tumor uptake, and far lower accumulation in the liver (2.43 versus 7.27 percent ID/g), spleen (0.57 versus 2.64), and kidney (3.37 versus 6.90) at 60 minutes. The NOTA compound was therefore excluded, and [68Ga]Ga-DOTA-FZPF advanced to clinical studies. The tracer also functioned as a pharmacodynamic readout: in mice treated with fuzuloparib, tumor uptake of the probe fell during therapy even though PARP-1 expression by immunohistochemistry remained similar, a pattern consistent with the unlabeled drug occupying the binding sites, exactly what a target-engagement imaging agent should show.

The first-in-human phase enrolled three breast cancer patients and one ovarian cancer patient in a small exploratory cohort approved by the Ethics Committee of Fudan University Shanghai Cancer Center. No adverse events or vital sign abnormalities were observed after injection, and the tracer cleared rapidly from blood, with a distribution half-life of less than two minutes in preclinical pharmacokinetics and predominantly renal excretion. In one patient with left breast cancer, the scan showed markedly higher uptake in the affected breast than in the contralateral normal breast, with a maximum standardized uptake value of 3.7 versus 1.8, and pathology confirmed high PARP-1 expression in the lesion. In a second patient with hepatic metastases after mastectomy, uptake in the liver lesion was close to the liver background, SUVmax 1.3 versus 1.1, and immunohistochemistry showed low PARP-1 expression. The authors note that, compared with clinical data for [18F]FTT and [18F]PARPi, their trial showed significantly reduced uptake in non-specific organs such as the liver, spleen, gastrointestinal tract, and bones, although tumor uptake itself was modest, possibly because of the small cohort, the relatively low PARP-1 expression at the enrolled lesions, and the pronounced hydrophilicity of the DOTA chelator, which may limit cell membrane penetration. The modest tumor-to-liver ratio of 1.38 plus or minus 0.28 may also constrain sensitivity for detecting hepatic metastases.

Despite those limitations, the study marks a genuine milestone. According to the authors, no prior clinical investigation of a gallium-68-labeled PARP-1-targeting radiotracer has been reported, and no AI-designed radiotracer has previously undergone clinical evaluation, making this a preliminary attempt at both. The work demonstrates that established computational tools, generative chemistry, CNN-based docking, synthetic complexity scoring, reverse target screening, and molecular dynamics, can be integrated with medicinal chemistry judgment into a reproducible closed loop that delivers experimentally verifiable imaging agents. The team frames the result as a case study that could be extended to radiotracers targeting other molecules, and the broader implications are striking: if AI pipelines can reliably compress the years of iterative medicinal chemistry normally required to tune a radiotracer’s affinity, synthesizability, and pharmacokinetics, the path from molecular target to clinic-ready imaging agent could shorten considerably. For now, [68Ga]Ga-DOTA-FZPF remains an early-stage candidate, and the authors themselves call for further optimization and larger clinical studies to validate PET imaging of PARP-1, but the bench-to-bedside journey it completed offers a template for how artificial intelligence may reshape radiopharmaceutical development.

Subject of Research: AI-assisted development and first-in-human evaluation of the PARP-1-targeted PET radiotracer [68Ga]Ga-DOTA-FZPF

Article Title: An artificial intelligence-assisted PARP-1-targeted PET imaging agent [ 68 Ga]Ga-DOTA-FZPF: From bench to clinic

Article References: Ji, M., Jiang, C., Xi, Y., Liu, S., He, S., Zhang, J., Xu, X., Yang, Z., Wang, X., & Song, S. (2026). An artificial intelligence-assisted PARP-1-targeted PET imaging agent [68Ga]Ga-DOTA-FZPF: From bench to clinic. Materials Today Bio, 41, Article 103728. https://doi.org/10.1016/j.mtbio.2026.103728

Image Credits: AI Generated

DOI: Not provided

Keywords: PARP-1, PET imaging, radiotracer, artificial intelligence, gallium-68, fuzuloparib, molecular imaging, drug design, breast cancer, ovarian cancer, PARP inhibitors, radiopharmaceuticals