This study enrolled patients with glioma undergoing treatment at Henan Provincial People’s Hospital in China between July 2016 and December 2021. These patients received intraoperative implantation of a reservoir into the surgical cavity as the reservation for postoperative diagnosis and treatment. In principle, patients received surgical resection, chemoradiotherapy for high-grade gliomas, and adjuvant temozolomide (TMZ) chemotherapy until relapse; after recurrence, patients would undergo a second surgery or receive bevacizumab (BEV), anti-programmed cell death-1 (anti-PD-1), or BEV plus anti-PD-1 combined therapy. Serial TISF samples were collected for cf-tDNA sequencing during the postoperative standard of care and were subsequently retrospectively analyzed. Blood samples for germline DNA controls from each patient were also obtained. This study was approved by the local institutional review board, and all patients or their next of kin provided written informed consent for any procedure. Samples from 116 patients were evaluated for analysis, wherein 107 patients had paired initial tumor tissue and cf-tDNA samples sequenced with a targeted panel.

EDTA tubes containing whole blood samples or in situ TISF were centrifuged for 10 min at 1,900 g, and the supernatants from these samples were further centrifuged for 10 min at 16,000 g. Samples were then collected and stored at − 80°C until extraction. Genomic DNA was extracted from fresh tumor tissue using the QIAamp DNA Tissue & Blood Kit (Qiagen, Germantown, MD, USA). Cell-free DNA was extracted from plasma and TISF using the MagMAX™ CellFree DNA Isolation Kit (Thermo Fisher Scientific, Waltham, MA, USA). Finally, all isolated DNAs were quantified with the Qubit 2.0 Fluorometer using the Qubit dsDNA HS Assay kit (Life Technologies, Carlsbad, CA, USA) according to the recommended protocol.

Genomic DNA was sheared into 150–200- bp fragments using the Covaris M220 Focused-ultrasonicator™ Instrument (Covaris, Woburn, MA, USA). Fragmented DNA and cf-tDNA libraries were constructed using the KAPA HTP Library Preparation Kit (Illumina platforms; KAPA Biosystems, Wilmington, MA, USA) following the manufacturer’s instructions. DNA libraries were captured using a designed brain tumor panel of 68 genes (Genetron Health, Beijing, China) that included major brain tumor-related genes. The captured samples were subjected to Illumina HiSeq X-Ten for paired-end sequencing.

Sequencing reads from the HiSeq X-Ten platform were demultiplexed, allowing zero mismatches in barcodes, and the read quality statistics were calculated by FastQC.

Sequence adapters and low-quality regions were removed with Trimmomatic (v0.36) and then mapped to the hg19 reference genome with BWA (v0.7.10). PCR duplicates were marked using Picard. Local realignment was run using GATK. Pileup files that were converted from bam files were generated for the genomic regions targeted by exome enrichment. With the use of the pileup file as input, single-nucleotide variant (SNV) or indel was identified using SAMtools (v0.1.1722) and Pindel; structural variation was detected using Crest. The criteria we adopted for retaining a mutation in TISF and plasma cf-tDNA were that it had an allele fraction of ≥ 0.1% and a total of ≥ 4 reads. Additionally, known recurrent loci were manually reviewed using Integrative Genomics Viewer (IGV v2.3.34) in the target tissue as compared to the normal blood DNA. All mutations were annotated for genes and function as well as repeated genomic regions using ANNOVAR, Oncotator, and Vep. The dbNSFP and the Exome Aggregation Consortium (ExAC) database were used to filter out either the benign mutations with pp2_hdiv score < 0.452 or polymorphic non-synonymous mutation sites. Finally, all mutations were annotated for genes using ANNOVAR, Oncotator, and Vep.

Three terms, tissue dominant (TD), tissue complementary (TC), and basically consistent (BC), were defined to describe the degree of consistency between a tumor tissue and the paired cf-tDNA sample. Taking the initial tumor tissue as the baseline comparison, a paired cf-tDNA sample was described TD if it was negative, harbored no dominant subclonal mutation, or only had partial tissue-shared mutations; TC if it had at least one dominant subclonal mutation that was not shared with tissue; or BC if all mutations in the tissue were detected as the whole dominant subclonal mutations in the cf-tDNA sample. If a cf-tDNA mutation had variant allele fraction (VAF) > 0.6 before postoperative antitumor treatment, VAF > 0.8 during adjuvant treatment, or VAF ≥ 1.0 at recurrence, it was considered a dominant mutation in the cf-tDNA sample.

All statistical calculations were performed using Prism 9.4.1 for MacOS. Wilcoxon matched-pairs signed rank test was used for the comparison of maximal VAF or number of mutations (NOM) in paired tissue and cf-tDNA samples. The Mann–Whitney test was used for the comparison of VAF in non-paired tissue and cf-tDNA samples. Chi-square or Fisher’s exact test was used for the comparison of contingency distribution or correlation between groups. Progression-free survival (PFS) analysis used a log-rank test. p < 0.05 was considered significant. All p- values were two-sided.

Serial cf-tDNA analysis was performed on TISF of 107 evaluable patients with glioma that had paired tumor tissues, including the three major subtypes: IDH mutant without co-deletion of chromosome 1p/19q (IDH-mutant-noncodel; n = 25), IDH mutant and chromosome 1p/19q co-deleted (IDH-mutant-codel; n = 17), and IDH wild type (glioblastoma (GBM); n = 65). The median age at the time of the first TISF cf-tDNA sample was 51 years, and a total of 205 cf-tDNA samples were sequenced (median of 2 and range 1–5 cf-tDNA samples/patient). At least one tumor-derived genomic alteration was found in 179/205 (87.3%) of samples (median of 5 cf-tDNA variants/sample; range 0–336 variants/sample). The 26 (12.7%) negative cf-tDNA samples were all collected after the initiation of postoperative radio- or chemotherapy. A total of 2,857 genomic mutations were detected in the cf-tDNA samples, and the most frequently mutated genes were FAT1 (151/2,857, 5.3%), NF1 (148/2,857, 5.2%), and TP53 (137/2,857, 4.8%) ( Table 1 ; Figures 1A–C ).

A total of 43 patients had evaluable cf-tDNA samples before the initiation of antitumor treatment after the initial surgery. All patients had tumor-derived mutations found in their liquid samples. First, we compared diagnostic genetic alterations (16) between tumor tissues and temporally paired cf-tDNA samples (within 45 days). We found that 14.0% (6/43) of patients’ cf-tDNA samples and paired tissues shared the same diagnostic alterations, 18.6% (8/43) cf-tDNA samples contained none or a part of tissue-shared diagnostic variants, and the remaining 67.4% (29/43) cf-tDNA samples were detected to have at least one diagnostic mutation that was not found in paired tissues ( Figure 1D ). The most frequent TISF-private diagnostic mutations were TSC1/2 (in 19/43 patients), NF1 (in 12/43 patients), and FGFR1/3 (in 11/43 patients) ( Figure 1D ). The three genetic alterations were commonly related to the diagnosis of low-grade or circumscribed astrocytic gliomas (16). Considering general alterations, cf-tDNA samples contained far more mutations than paired tissues (median 9 vs 3 mutations/sample, p < 0.0001, Wilcoxon matched-pairs signed rank test) ( Figure 1E ), and 86.3% of them (460/533) were private; the median percentage of tissue-shared mutations was only 20% (0%–100%) ( Figure 1F ).

Volume decompression caused by surgical resection resulted in a very low proportion of cf-tDNA to cell-free DNA in TISF. Therefore, the VAFs of cf-tDNA mutations were expectedly low (median VAF 0.2%, range 0.1%–49.9%); among them, the higher VAFs were mainly concentrated on tissue-shared mutations (73 shared vs 460 cf-tDNA-private, median VAF 2.0% vs 0.2%, p < 0.0001, Mann–Whitney test) ( Figure 1G ). However, some cf-tDNA-private mutations also had higher VAFs, which might indicate that the residual disease had a genomic structure that was not completely covered by tissue sequencing. For example, patient 31 is a 40-year-old man with GBM, and we detected three mutations— NF1:p.D1485lfs*8, PIK3R1:N/A, and PTEN:p.L23F— in his tumor tissue. The paired cf-tDNA sample was detected to have 21 mutations, including the three shared mutations with VAFs of 0.8%, 1.0%, and 0.7% and 18 private mutations, wherein 17 had VAF of less than 0.5%, except for PTCH1:p.G17del, which had a higher VAF at 1.1%. The eccentrically high mutational abundance of the private PTCH1:p.G17del may represent the supplementary role of TISF cf-tDNA for tumor tissues ( Figure 1H ).

We failed to find the NOM or the maximal VAF (mVAF) in the cf-tDNA samples correlated with MRI-measured residual disease, glioma subtypes, or other features of gliomas ( Supplementary Table 3 ), suggesting that shedding of tumor DNA into TISF may be determined by other non-volume characteristics of glioma, or TISF cf-tDNA is supplementary information not just a substitute for the tumor itself.

Collectively, the abundant mutations in the TISF cf-tDNA reflected the molecular residual disease and complemented the genomic heterogeneity of tumor tissues.

We analyzed the dynamic consistency of tumor genomes between initial tissues and the paired cf-tDNA samples of three representative timepoints (before antitumor therapy, during chemotherapy, and at recurrence) after surgery. Before antitumor therapy, among 43 evaluable patients, the percentages of TD and TC cf-tDNA samples were similar (41.9% vs 44.2%), and the remaining six (13.9%) were BC samples. During the TMZ chemotherapy, the proportion of TD cf-tDNA samples was twice that of TC samples (63.4% vs 32.4%) in 71 evaluable patients, and only three (4.2%) were BC. When it came to recurrence, of 29 evaluable cf-tDNA samples, 21 (72.4%) were TC and seven (24.1%) were BC, and only one (3.5%) sample was TD ( Figure 2A ). The longitudinal analysis of 15 patients’ cf-tDNA samples showed a similar developing trend. That is, most patients’ cf-tDNA samples evolved from TD during the chemotherapy into TC at recurrence. One example was with patient 67, a 64-year-old man with IDH-mutant-noncodel glioma; after the second period of adjuvant TMZ treatment, five low-VAF (<0.8%) variants were detected in his first cf-tDNA sample without imaging findings, while 5 months later, 24 mutations with elevated VAFs (the maximal 49.3%) emerged in the relapsed cf-tDNA sample, wherein some were not found in the initial tumor tissue ( Figure 2B ).

In addition, four patients with relapsed glioma had evaluable recurrent tumor tissues. We found that the paired closest TISF cf-tDNA sample at recurrence contained all the mutations detected in the recurrent tissue, while the matched initial and recurrent tissues only shared a portion of mutations ( Figures 2B–E ). More interestingly, two other patients were diagnosed with pseudoprogression by a second surgery, and their closest preoperative cf-tDNA samples were mutation-negative or remained stable from previous tests ( Figures 2F, G ).

Together, the dynamic consistency between tumor tissue and cf-tDNA samples indicated that the representativeness of initial tumors for the glioma decreased with tumor evolution, and the real-time genomic status could be complemented by the TISF cf-tDNA.

Next-generation sequencing studies have captured a variety of evolutionary patterns of cancer shaping and development, mainly including linear evolution (LE), branching evolution (BE), and neutral evolution (NE) (1, 2, 17–19). Combining serial cf-tDNA samples with tumor tissues could identify dynamic evolutionary patterns of the glioma ( Figure 3C ). For example, the TC cf-tDNA samples in the majority at recurrence suggested that new adaptive mutations emerged in the offspring tumor cells, which was consistent with a BE pattern. One illustrative example was again with patient 67, a 64-year-old man with IDH-mutant-noncodel glioma. Both his relapsed cf-tDNA sample and recurrent tumor tissue were detected to have newly emerging CDKN2A:p.L16Pfs*9 and PDGFRA:p.V536E mutations compared to the initial tissue ( Figure 2B ), indicating that the relapsed tumor had evolved more malignant branches featured with CDKN2A loss (20). In addition, during radio- and chemotherapy, many instant low-VAF passenger mutations appeared and disappeared in serial cf-tDNA samples, which might reflect a temporary period of neutral evolution of the glioma genome under therapeutic pressure. For example, patient 60 was a 56-year-old man with IDH-mutant-codel glioma. His first TISF cf-tDNA sample after surgery was detected to have a tissue-shared IDH1:p.R132H mutation, while in the subsequent three cf-tDNA samples spanning 7.0 months during the TMZ chemotherapy, one, eight, and eight low-VAF mutations were found, with none of them shared, and they remained stable meantime as detected by MRI ( Figure 3A ). Relapsed gliomas having BC cf-tDNA samples might experience an LE pattern, where the most significant selection pressure shaping the tumor genome happened before the diagnosis of glioma, while the following treatment only exerted a subtle effect on the relapsed genomic structure. In addition, analysis of the sequence of the emergence of the driver mutations in the serial cf-tDNA can trace the subclonal evolutionary trajectories of glioma. For example, serial cf-tDNA analysis revealed the patient 30’s GBM was linearly driven and progressed by successive appearance of subclones featured with BRAF:p.V600E and HIST1H3B:p.K37M mutations ( Figure 3B ).

Mismatch repair (MMR) deficiency and hypermutation in the glioma genome were reported as related to the TMZ-resistant mechanisms (21). In 107 patients’ initial tumor tissues, we only detected MMR mutations (MSH6, MLH1, MSH2, or PMS2) in two IDH-mutant initial tumors (2 of 107, 1.8%) ( Figure 1B ), while MMR mutations were present in 7/43 (16.3%) patients’ temporally paired cf-tDNA samples; none of their tissues had MMR mutations identified ( Figure 1D ). Of 107 patients, 25 (23.4%) had at least one cf-tDNA sample identified as MMR mutations ( Figures 1A, B ). A total of 35 cf-tDNA samples with MMR mutations had far more mutations than those 170 without (median 20 vs 3 mutations/sample, p < 0.0001, Mann–Whitney test) ( Figure 3D ). Of note, not all MMR mutations would persist after its emergence and cause hypermutation. For example, patient 72 with IDH-mutant-noncodel glioma had a cf-tDNA sample with two transient MMR mutations (VAF 0.1%) during the TMZ treatment, but the two mutations disappeared in the following cf-tDNA samples and did not cause many mutations. Low-VAF MMR mutations were found in early cf-tDNA samples of patients 34 and 56 and then vanished with treatment ( Figure 3E ). However, relapsed cf-tDNA samples with high-VAF MMR mutations were prone to a high number of mutations and to tumor dissemination or distant recurrence. For example, multiple MMR mutations (the highest VAFs were 26.1% and 87.9%) were detected in the 1-month-apart two relapsed cf-tDNA samples of patient 4 with GBM, with 189 and 267 mutations identified in the two samples; meanwhile, the tumor spread rapidly to other regions ( Figure 3F ). Patient 52 with IDH-mutant-codel glioma received BEV treatment after tumor recurrence, the first post-treatment cf-tDNA sample at 2 months showed molecular remission, while the second cf-tDNA sample at 4 months was detected to have MMR mutations (highest VAF 9.3%) with 184 somatic mutations, and the tumor disseminated toward other regions as well ( Figure 3G ).

These phenomena supported the hypothesis that MMR-deficient cells are pre-existing in the tumor and that TMZ treatment exerts certain selection pressure biased toward MMR-deficient cells (21). However, the direct survival advantage of MMR-deficient cells may derive from that treatment-induced numerous mutations are more likely to activate a variety of carcinogenic signaling pathways, making the subclones more resistant and invasive ( Figure 3H ).

The NOM or the mVAF in the baseline cf-tDNA samples (before antitumor treatment) was not prognostic for progression in evaluable GBM patients ( Figures 4A, B ). However, a larger NOM or a higher mVAF in the during-treatment cf-tDNA sample was associated with shorter PFS following TISF collection in both IDH-wild-type (NOM: hazard ratio 2.478, log-rank p = 0.0107; mVAF: hazard ratio 2.566, log-rank p = 0.0076) and IDH-mutant gliomas (NOM: hazard ratio 3.846, log-rank p = 0.0004; mVAF: hazard ratio 3.821, log-rank p = 0.0005) ( Figures 4C–F ). Most tumors whose mVAF has continuously increased progressed shortly ( Figure 4I ). The mVAFs of relapsed cf-tDNA were significantly higher than those before recurrence (median mVAF 25.2% vs 2.6%, p = 0.0001, Wilcoxon matched-pairs signed rank test) ( Figure 4G ), while the NOM did not increase at recurrence ( Figure 4H ).

The high sensitivity of TISF cf-tDNA in glioma patients may allow continuous molecular-level monitoring of the tumor evolution from a minuscular burden, making it possible to identify responsible genomic alterations that drive tumor progression and recurrence at the earliest possible time. For example, patient 9 was a 27-year-old man with GBM. During the TMZ treatment, a BRAF:p.V600E mutations were detected in three serial cf-tDNA samples spanning 3.4 months, and the VAF of the variant increased from 0.3% to 3.4%, while the contemporaneous MRI findings were negative until slight marginal progression appeared on the MRI 4 months later ( Figure 4J ). Similarly, patient 58, a 56-year-old woman with IDH-mutant-codel tumor, had a TP53:p.R248W mutation with gradually increasing VAF (from 0.3% to 0.7%) detected in three sequential cf-tDNA samples over 10.5 months, and the MRI changes were observed until 6.8 months later ( Figure 4K ).

In the two examples, before imaging changes, serial cf-tDNA identified BRAF:p.V600E and TP53:p.R248W mutations that were responsible for driving the two tumors’ progression. We called that cf-tDNA-level relapse because the serial detection and increase indicated that the tumor had developed stable resistant subclones that began to dominate the fast growth of the relapse tumor, which could be detected by imaging modalities after a subsequent period of volume expansion ( Figure 4L ).

The dynamic heterogeneity of glioma genomes cannot be fully characterized by sequencing of resected tumor tissues, thereby hindering the temporal precision medicine during the tumor evolution for the affected patients. To evaluate the potential therapeutic utility of serial TISF cf-tDNA analysis, we assessed the number of cf-tDNA variants that could potentially be clinically targeted by available drugs. We estimated 223 different cf-tDNA variants are potentially clinically targetable, of which 174 (78.0%) were not found in any patient’s tumor tissues ( Supplementary Table 4 ). These cf-tDNA-unique variants are targetable primarily with small molecules that inhibit PI3K/AKT/mTOR, RTK/RAS, or P53 signaling pathways ( Supplementary Figure 1 ), all of which have been approved in other cancers and are being tested clinically in gliomas.