The research, led by Shaon Basu and Markus Ralser at Charité Universitätsmedizin Berlin and the Francis Crick Institute, together with a team at AstraZeneca, focused on degraders of the androgen receptor, a transcription factor that drives treatment-resistant prostate cancer. First-line therapies for prostate cancer typically block androgen receptor signaling with antagonist drugs, but tumors inevitably acquire mutations at the androgen receptor locus, rendering the drugs ineffective and the disease lethal. Degraders promise a way around this resistance by directly destroying the receptor inside tumor cells. Despite this promise, no androgen receptor degrader has yet demonstrated improved efficacy over standard treatment with an acceptable safety profile, echoing the fate of first-generation bromodomain degraders that failed in cancer patients because of dose-limiting toxicity.

The team’s strategy began with a technological foundation: a fast, scalable proteomic workflow capable of monitoring how thousands of proteins respond to drug treatment in a 96-well plate format. HepG2 cells, a well-established model for screening liver toxicity, were cultivated and treated with liquid-handling robots, and their proteins were digested into peptides and analyzed by data-independent acquisition mass spectrometry on two platforms, using scanning SWATH acquisition on a Sciex TripleTOF 6600 and dia-PASEF acquisition on a Bruker timsTOF HT. Neural network-based processing with the DIA-NN software, benchmarked against a hepatocarcinoma-specific spectral library, quantified thousands of proteins per sample with minimal batch effects and low technical variance, making the approach suitable for industrial-scale screening.

To validate the platform, the researchers exposed HepG2 cells to 78 FDA-approved drugs whose targets span human, bacterial and fungal proteins. Forty-seven of these compounds, particularly those with favorable safety profiles, triggered no significant proteomic response. The remaining 31 induced measurable changes, and the proteomes faithfully reflected each drug’s mechanism of action: methotrexate altered less than one percent of the proteome while its primary target, dihydrofolate reductase, ranked among the most upregulated proteins, whereas the chemotherapy doxorubicin reshaped seven percent of the proteome in a pattern consistent with DNA damage. Intriguingly, the dominant proteomic signatures often corresponded to off-target rather than on-target interactions, including disruption of ribosomal translation, endoplasmic reticulum stress and mitochondrial dysfunction. The antifungal clotrimazole, for example, induced broad rearrangements driven by xenobiotic response proteins and mitochondrial impairment despite the absence of its fungal target in human cells. A principal component analysis cleanly separated compounds with narrow therapeutic windows, such as chemotherapeutics and immunosuppressants, from safer drugs.

With the platform validated, the team turned to the central question: what off-target effects do experimental androgen receptor degraders produce? Because HepG2 cells do not express the androgen receptor, any proteomic response to these compounds must reflect off-target engagement. The researchers profiled a library of 204 structurally unique androgen receptor degraders, including clinically advanced candidates such as bavdegalutamide (ARV-110) and an AstraZeneca compound referred to as compound 1, alongside control degraders targeting BRD4, CDK4 and CDK9. The screening campaign generated 2,113 proteomic samples reaching depths of roughly 4,000 proteins each. The results were striking: despite the absence of their intended target, several androgen receptor degraders perturbed up to 57 percent of the proteome, with an average of 19 percent of proteins differentially expressed, a far stronger response than that elicited by the FDA-approved drug library. Notably, degraders targeting proteins actually expressed in HepG2 cells, such as BRD4, produced responses that partly reflected the biology of their targets, confirming that the platform can distinguish on-target from off-target signatures.

The off-target responses clustered according to chemical structure. Degraders recruiting the cereblon E3 ligase, particularly those using phthalimide derivatives and achiral DHU warheads, produced the broadest perturbations, while those recruiting the von Hippel-Lindau ligase produced the weakest. Crucially, the toxicity was not inherent to the recruiter alone: lenalidomide by itself elicited no proteomic response, while other cereblon ligands induced moderate or strong changes. Functional enrichment analysis pointed decisively at the mitochondrial compartment, with pathways involving mitochondrial protein-containing complexes responding in a dose-dependent manner. When the researchers tested whether the compounds impaired respiration by culturing cells in galactose medium, which forces cells to rely on mitochondrial ATP production, several cereblon-recruiting degraders reduced viability in a dose-dependent fashion. Compound 1 showed a half-maximal inhibitory concentration of 0.73 micromolar in galactose medium, while its constituent building blocks, lenalidomide and an androgen receptor-binding analog of enzalutamide, were inactive above 20 micromolar, indicating a gain-of-function toxicity arising from the assembled degrader molecule.

To pinpoint the molecular mechanism, the team applied machine learning to the proteomic data. Univariate linear regression linking protein expression changes to galactose toxicity revealed that nine of the ten best-fitting proteins localized to mitochondria, and four of these were subunits of electron transport chain complex I, the NADH dehydrogenase that initiates the proton gradient driving ATP synthesis. The complex I protein NDUFA5 showed the strongest relationship, with a coefficient of determination of 0.57. A second-pass gradient-boosted decision tree model, refined through Shapley additive explanations analysis and retrained on top features, achieved a cross-validation area under the precision-recall curve of 80 percent, with complex I proteins dominating the feature importance rankings. Biochemical validation followed: compound 1 reduced oxygen consumption in HepG2 cells, and in assays using bovine heart mitochondrial membranes it potently inhibited NADH oxidation through the complex I-III-IV axis with an IC50 of 0.2 micromolar, roughly 100-fold more potent than lenalidomide and 45-fold more potent than enzalutamide, while leaving complex II activity untouched. The degrader, in other words, was a selective complex I inhibitor.

The most consequential step came next. Because the toxicity was not inherent to the recruiter chemistry, the researchers reasoned that proteomics could identify degraders lacking complex I engagement. Their trained models assigned compound 1 a predicted toxicity score of 91 out of 100, comparable to known hepatotoxins such as troglitazone, but attributed far lower scores of 47 and 35 to two close structural analogs, compounds 2 and 3, which differed only in the linker region and carried a 7-methoxy modification on the lenalidomide moiety. Experimentally, the analogs induced no complex I proteomic signature, did not impair mitochondrial respiration, and showed a complex I IC50 of 6.5 micromolar versus 0.2 micromolar for compound 1. In primary human hepatocytes, compound 1 caused acute cell death at 2.7 micromolar, whereas compounds 2 and 3 were tolerated above 20 micromolar. The analogs also lost the unwanted degradation of the cereblon neosubstrates IKZF3 and SALL4, proteins linked to bone marrow and developmental toxicity, likely owing to the methoxy modification.

Safety gains did not come at the expense of efficacy. All three compounds inhibited proliferation of androgen receptor-positive prostate cancer cell lines, but only compound 1 suppressed androgen receptor-negative lines, suggesting improved selectivity for the refined analogs. Compound 3 degraded the androgen receptor as effectively as compound 1 in LNCaP cells and dose-dependently suppressed prostate-specific antigen expression. It also displayed favorable pharmacokinetics, with low clearance in human liver microsomes and rat hepatocytes and 50 percent oral bioavailability in mouse. RNA sequencing confirmed that compound 3 modulated androgen response signaling similarly to enzalutamide while reducing MYC oncogenic signaling and G2/M checkpoint pathways, consistent with cell-cycle arrest, and without perturbing oxidative phosphorylation at any time point.

The decisive test came in vivo. In a C4-2 xenograft model of castration-resistant prostate cancer in immunocompromised mice, three days of oral dosing with compound 3 at 30 milligrams per kilogram reduced androgen receptor protein levels in tumors by more than 70 percent. Over 28 days of treatment, compound 3 achieved 56 percent tumor growth inhibition, compared with just 15 percent for enzalutamide at 50 milligrams per kilogram, and both treatments were well tolerated with no significant body weight loss. The study establishes that systematic, high-throughput proteomics paired with interpretable machine learning can trace unwanted proteomic signatures back to specific chemical features, here a linker region, and guide the rational design of safer, more selective degraders. As polypharmacological modalities such as degraders and antibody-drug conjugates continue to enter the clinic, the authors argue that this proteome-guided framework offers a versatile route to de-risk drug development before candidates ever reach patients.

Subject of Research: Proteome-guided discovery and mitigation of toxicity mechanisms in androgen receptor-targeted protein degraders

Article Title: Proteome-guided drug discovery maps and mitigates therapeutic degrader toxicity

Article References: Basu, S., Yu, J. S. L., Bosak, J., Bagal, S. K., Crafter, C., Fallan, C., Lemke, O., Freiwald, A., White, M. E. H., Vernardis, S. I., Messner, C. B., Michaloglou, C., Pryde, K., Ramos-Montoya, A., Regan, S., Rodrigo-Brenni, M., Roth, S., Ruffilli, C., Mülleder, M., … Ralser, M. (2026). Proteome-guided drug discovery maps and mitigates therapeutic degrader toxicity. Nature Chemical Biology. https://doi.org/10.1038/s41589-026-02347-2

Image Credits: AI Generated

DOI: 10.1038/s41589-026-02347-2

Keywords: proteomics, targeted protein degradation, PROTAC, androgen receptor, prostate cancer, mitochondrial toxicity, complex I, machine learning, drug discovery, hepatotoxicity, cereblon, mass spectrometry