Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

Among nuclear receptors, the androgen receptor (AR) is unique in that its ligand-binding domain (LBD) interacts with the FXXLF motif in the N-terminal domain, resembling coactivator LXXLL motifs. We compared AR- and estrogen receptor alpha-LBD interactions of the wild-type AR FXXLF motif and coactivator transcriptional intermediary factor 2 LXXLL motifs and variants of these motifs. Random mutagenesis revealed a key role for the F residues in FXXLF motifs in high-affinity and selective AR LBD interaction. The FXXLF motif in full-length AR and transcriptional intermediary factor 2 LXXLL motifs competed for an overlapping binding site. A computer model of the AR LBD/AR FXXLF complex showed that the bulky F residues are buried in a deep coactivator-binding groove. The corresponding groove in estrogen receptor alpha LBD is considerably shallower, explaining lack of binding of any of the FXXLF motifs tested. FXXLF and LXXLL motif interaction depended on different charged amino acid residues in the AR LBD present at opposite ends of the coactivator groove. In conclusion, our data demonstrate the importance of a deep hydrophobic groove and alternative usage of charged amino acids in specifying peptide binding to the AR LBD.

Whole genome sequencing (WGS) using fresh-frozen tissue and matched blood samples from cancer patients may become the most complete genetic tumor test. With the increasing availability of small biopsies and the need to screen more number of biomarkers, the use of a single all-inclusive test is preferable over multiple consecutive assays. To meet high-quality diagnostics standards, we optimized and clinically validated WGS sample and data processing procedures, resulting in a technical success rate of 95.6% for fresh-frozen samples with sufficient (≥20%) tumor content. Independent validation of identified biomarkers against commonly used diagnostic assays showed a high sensitivity (recall; 98.5%) and precision (positive predictive value; 97.8%) for detection of somatic single-nucleotide variants and insertions and deletions (across 22 genes), and high concordance for detection of gene amplification (97.0%; EGFR and MET) as well as somatic complete loss (100%; CDKN2A/p16). Gene fusion analysis showed a concordance of 91.3% between DNA-based WGS and an orthogonal RNA-based gene fusion assay. Microsatellite (in)stability assessment showed a sensitivity of 100% with a precision of 94%, and virus detection (human papillomavirus), an accuracy of 100% compared with standard testing. In conclusion, whole genome sequencing has a >95% sensitivity and precision compared with routinely used DNA techniques in diagnostics, and all relevant mutation types can be detected reliably in a single assay.

Purpose: In a post hoc analysis of the CATNON trial (NCT00626990), we explored whether adding temozolomide to radiotherapy improves outcome in patients with IDH1/2 wildtype (wt) anaplastic astrocytomas with molecular features of glioblastoma [redesignated as glioblastoma, isocitrate dehydrogenase–wildtype (IDH-wt) in the 2021 World Health Organization (WHO) classification of central nervous system tumors]. Patients and Methods: From the randomized phase III CATNON study examining the addition of adjuvant and concurrent temozolomide to radiotherapy in anaplastic astrocytomas, we selected a subgroup of IDH1/2wt and H3F3Awt tumors with presence of TERT promoter mutations and/or EGFR amplifications and/or combined gain of chromosome 7 and loss of chromosome 10. Molecular abnormalities including MGMT promoter methylation status were determined by next-generation sequencing, DNA methylation profiling, and SNaPshot analysis. Results: Of the 751 patients entered in the CATNON study, 670 had fully molecularly characterized tumors. A total of 159 of these tumors met the WHO 2021 molecular criteria for glioblastoma, IDH-wt. Of these patients, 47 received radiotherapy only and 112 received a combination of radiotherapy and temozolomide. There was no added effect of temozolomide on either overall survival [HR, 1.19; 95% confidence interval (CI), 0.82–1.71] or progression-free survival (HR, 0.87; 95% CI, 0.61–1.24). MGMT promoter methylation was prognostic for overall survival, but was not predictive for outcome to temozolomide treatment either with respect to overall survival or progression-free survival. Conclusions: In this cohort of patients with glioblastoma, IDH-wt temozolomide treatment did not add benefit beyond that observed from radiotherapy, regardless of MGMT promoter status. These findings require a new well-powered prospective clinical study to explore the efficacy of temozolomide treatment in this patient population.

Upon hormone binding, a hydrophobic coactivator binding groove is induced in the androgen receptor (AR) ligand-binding domain (LBD). This groove serves as high affinity docking site for α-helical FXXLF motifs present in the AR N-terminal domain and in AR cofactors. Study of the amino acid requirements at position +4 of the AR FXXLF motif revealed that most amino acid substitutions strongly reduced or completely abrogated AR LBD interaction. Strong interactions were still observed following substitution of Leu+4 by Phe or Met residues. Leu+4 to Met or Phe substitutions in the FXXLF motifs of AR cofactors ARA54 and ARA70 were also compatible with strong AR LBD binding. Like the corresponding FXXLF motifs, interactions of FXXFF and FXXMF variants of AR and ARA54 motifs were AR specific, whereas variants of the less AR-selective ARA70 motif displayed increased AR specificity. A survey of currently known AR-binding proteins revealed the presence of an FXXFF motif in gelsolin and an FXXMF motif in PAK6. In vivo fluorescence resonance energy transfer and functional protein-protein interaction assays showed direct, efficient, and specific interactions of both motifs with AR LBD. Mutation of these motifs abrogated interaction of gelsolin and PAK6 proteins with AR. In conclusion, we have demonstrated strong interaction of FXXFF and FXXMF motifs to the AR coactivator binding groove, thereby mediating specific binding of a subgroup of cofactors to the AR LBD.