This clinical study (NCT04105582) received approval from the ethics committee of Universidad del Rosario and Hospital Mayor de Méderi (Act. No. 409, February 27, 2020) in Bogotá, Colombia. Patients with a clinical diagnosis of triple-negative breast carcinoma, who provided informed consent and signed the consent form, were enrolled in the study. A blood sample was collected via apheresis for immunogenicity analysis and for generating autologous DCs (see methodology below). Tumor tissue was obtained through Tru-Cut biopsy before starting neoadjuvant chemotherapy to identify tumor neoantigens. Blood samples were also collected during treatment, at 6 months, and one year after the last dose of autologous DCs pulsed with the neoantigens (Figure 1A).

The FMR patient who participated in the clinical study is a 53-year-old female with a diagnosis of moderately differentiated nonspecific ductal NOS infiltrating carcinoma, grade II/II, triple-negative, Ki67 30%, T2 N1 M0, stage IIB. Mammography in 2019 revealed a mass in the external upper quadrant of the left breast measuring 30x22x29 mm, with an associated lymph node of 22x13 mm. The patient underwent a hysterectomy at age 39. Her family history includes her father as the only relative with gastric cancer. She received an AC-T regimen (4 doses), followed by a left quadrantectomy with lymph node dissection (one of 13 nodes was positive). After surgery, she completed 20 sessions of radiotherapy. Currently, she is being treated with tamoxifen, with no signs of local or regional recurrence.

An integrated bioinformatics workflow was developed for processing exome and transcriptome sequencing data from paired normal and tumor samples. Initially, reads were aligned to the GRCh38 reference genome using HISAT2 with parameters optimized for assembly compatibility. The resulting SAM files were converted to BAM format, sorted, and filtered to remove unmapped reads and duplicates using SAMtools and Picard. Somatic variants (SNVs and indels) were identified with GATK MuTect2 by comparing the tumor exome to its corresponding normal control (peripheral blood exome). A filtering process was then applied with GATK FilterMutectCalls to ensure high-quality variant calls. Variants were normalized (bcftools norm -m-both), decomposed (vt decompose), and compressed (bgzip), with index files generated using tabix. Functional annotation was performed using VEP (with plugins: Frameshift, Wildtype; options: –cache, –hgvs, –vcf), incorporating gene expression data from Cufflinks (FPKM) and StringTie. Read counts from bam-readcount and transcriptional/gene expression data from vcf-expression-annotator were annotated in a stratified manner for exome and transcriptome, using a specific threshold (-b 20 for base quality).

HLA typing data were obtained using OptiType (28) and seq2HLA (29) from the patient’s normal exome and transcriptome, respectively, while neoantigen prediction was performed with the pVAC-seq suite (30). This suite predicts mutated high-affinity peptides derived from identified somatic mutations. Neoepitope prediction was conducted for the tumor sample using the patient’s HLA genotype, which included both class I and class II alleles. Multiple predictive algorithms (NetMHCpan, NetMHC, and NetMHCIIpan) were used, considering peptides of 8–11 amino acids for HLA class I and 12–25 amino acids for HLA class II. Additional protein processing (NetChop) and stability (NetMHCstab) methods were employed to evaluate proteasomal cleavage characteristics and the binding stability (half-life) of the peptide/MHC complexes (NetMHCstab predicts only with HLA class I molecules). Neoantigen prioritization was based on binding affinity (IC50 < 500 nM across multiple algorithms), mutated transcript expression (FPKM > 1), and NetMHCstab half-life stability time (greater than 1 hour). Five class I neoantigens and one class II neoantigen were identified. Due to budget limitations in peptide synthesis capabilities, the class II peptide was selected, which also contained one of the identified class I epitopes. Detailed features of all five class I and the single class II neoantigens are provided in Supplementary Table 1.

The generation of autologous DCs was based on a previous study (31). Briefly, enriched monocytes were obtained from an apheresis sample of the patient. Isolated peripheral blood mononuclear cells (PBMCs) were stored in liquid nitrogen until use for DCs generation for each dose. During neoadjuvant chemotherapy, tumor and blood samples were used to extract DNA and RNA for neoantigen identification, as described above for synthesis (one peptide per patient). A week before each dose, frozen PBMCs were thawed, and enriched monocytes were then induced into immature DCs with the addition of IL-4 and GM-CSF and finally matured with two cocktails of pro-inflammatory cytokines. During maturation, the cells were stimulated with the corresponding neoantigen (20μg/mL), defined from each patient’s specific tumor. After several washes with saline solution, a total of 1.2x10^8 autologous DCs stimulated with the neoantigen were divided in six doses after neoadjuvant chemotherapy, intradermally in the draining region of the tumor (Figure 1A). Patients were monitored for 4 hours after vaccination to detect any adverse effects, then clinical follow-up was carried out every 24 hours until the next dose. Finally, follow-up continued monthly after the last dose until month 12.

The pipeline for germline variant detection was developed following GATK Best Practices (34) to ensure a robust and reproducible process. First, read quality is evaluated with FastQC (35), and reads are aligned to the reference genome (hg38) using BWA-MEM (36). The resulting files (SAM/BAM) are converted, sorted, and indexed with Samtools. Next, variant calling with GATK is optimized, including duplicate marking (MarkDuplicatesSpark), alignment sorting (SortSam), and base quality recalibration (BaseRecalibrator and ApplyBQSR). Germline variants are called with HaplotypeCaller (37). Finally, functional annotation is performed using Ensembl VEP and the Franklin Genoox platform, integrating information on allele frequency, molecular effects, and clinical annotations. The validity of the pipeline was confirmed by achieving 100% agreement in variant detection in a set of independent samples previously analyzed in a clinical setting (Euformatics Genomics Hub).

To prioritize germline variants potentially affecting cancer predisposition, they were selected from a panel of driver genes compiled from the literature (38–41) and gene panels used in clinical diagnosis (Illumina TruSight Oncology, Pan-cancer panel CD Genomics). Within the IntOGen database, genes present in at least 1% of cancer patients were chosen. From the OncoKB database, genes confirmed as cancer drivers in at least three sources were selected, citing the gene as a cancer driver.

For the identification and annotation of somatic variants in tumor exomes, a stepwise bioinformatics pipeline was designed. The raw reads were first evaluated with FastQC (35). They were then aligned to the hg38 genome using BWA-MEM (36). The generated files were processed with GATK to remove duplicates (MarkDuplicatesSpark) and recalibrate base quality scores (BaseRecalibrator and ApplyBQSR) (42). Subsequently, somatic variant calls were made in parallel using MuTect2 (43), LoFreq (44), MuSE2 (45), and Strelka2 (46). The results from the different somatic callers were integrated with SomaticSeq (47), which employs machine learning to prioritize the most reliable variants. Finally, the variants were annotated with Ensembl VEP (48) and converted to MAF format using vcf2maf to incorporate them into Memorial Sloan Kettering’s OnkoKB oncology database (41), enabling the identification of mutations with clinical significance based on each patient’s tumor genomic profile (49). The entire pipeline was executed within Docker containers to ensure reproducibility of the analyses.

Since the frequency of circulating cells specific to a tumor neoantigen is very low, an expansion protocol was developed in which the patient’s PBMCs were stimulated with cytokines to induce the differentiation and maturation of DCs in the presence of the neoantigen. We used one approach that links antigen processing to the presentation of the neoantigen to specific T cells by DCs in situ, published by Martinuzzi et al. (32), with minor modifications. The PBMCs were pulsed on day 0 and re-stimulated on day 9 with the neoantigen to evaluate cytokine secretion and the expression of activation markers on T cells, as assessed by flow cytometry, which serve as indicators of neoantigen recognition. This assay was conducted with PBMCs at multiple time points: before, post 3, post 6, and one-year post-vaccination (Post 1 yr).

As shown in Figure 1, re-stimulated cells exhibited increased expression of CD154 and CD69 activation markers in both CD4+ and CD8+ T cells compared with the non-re-stimulated control. CD154 expression was sustained across time points, peaking after the sixth dose (Post 6) and declining one year after immunization (Figures 1B, D). In contrast, CD69 expression was most pronounced at the third dose (Post 3) and decreased at subsequent time points (Figures 1C, E). Cytokine secretion reflected these activation marker profiles: both IFNγ and TNF-α increased progressively over time, reaching a peak after the sixth dose (Post 6), and then declined by the one-year follow-up (Post 1yr) (Figures 1F, G). Altogether, these results demonstrate that vaccination induced a measurable neoantigen-specific T-cell response characterized by activation marker expression and cytokine secretion, which peaked at Post 6 but waned one year after vaccination.

To determine if vaccination altered the memory phenotype, the patient’s PBMCs were stimulated using the same culture system described earlier, with CD45RA and CCR7 markers used to evaluate the memory phenotype. Different behaviors were observed in CD4+ and CD8+ T cells. For CD4+ T cells, data over time showed a decrease in naive T cells and an increase in central memory cells when stimulated with the neoantigen (Figure 1H). For CD8+ T cells, the effector phenotype is dominant before vaccination; however, as the vaccination schedule continues, the percentage of central memory cells increases. After the sixth dose (Post 6), the naive phenotype becomes predominant, but at one year, the naive population decreases while central memory and effector memory cells increase (Figure 1I). These findings indicate a shift in the memory phenotype of both CD4+ and CD8+ T cells, more noticeably in CD4+ T cells, which show an increase in the central memory phenotype. In contrast, no clear pattern was observed for CD8+ T cells.

To evaluate the immune response related to changes in the T cell repertoire at different vaccination times, we measured the frequency and sequence of the TCR-CDR3 region of T cells in PBMCs collected at various time points: before chemotherapy and vaccination (Before), after three cycles of chemotherapy (NAC), and Post 3 and Post 6 doses of the vaccine. Additionally, a comparison between unstimulated and neoantigen-stimulated T cells was performed at the Post 6 dose time point. To identify significant differences in CDR3 frequency between samples, we plotted two samples on a scatter diagram to detect rearrangements (unique sequences) that significantly increased or decreased in frequency (Figure 2). Most sequences were excluded due to low frequency (Figure 2A); however, some rearrangements were significantly more common in the pre-treatment (Before) group (orange dots, n = 357), while others increased after chemotherapy (blue dots, n = 116) (Before qx Vs. NAC in Figure 2A). When comparing the rearrangements of NAC samples versus Post 3 (Figure 2B, NAC Vs. Post 3), and Post 3 versus unstimulated Post 6 (Figure 2C, Post 3 Vs. Post 6) we observed similar patterns, as many rearrangements are shared, with a few unique to each treatment. We used a Venn diagram to assess the number of overlapping sequences in each sample (Figure 2E). There are no common sequences among the four time points (Before, NAC, Post 3, and Post 6 unstimulated); however, many sequences are shared between the last two treatment time points (Figure 2E). These results suggest that, in this patient, the combined treatment of NAC and vaccination may promote the expansion of circulating T cells with identical TCR-CDR3 sequences.

Unlike the overall analysis of the TCR repertoire described above, the comparison between unstimulated and neoantigen-stimulated Post 6 cells showed that, although many sequences were shared, 70 rearrangements were more common in the stimulated condition (Post 6 + peptide vs. Post 6 unstimulated, orange dots in Figure 2D). Paired analyses of TCR-CDR3 sequences across different time points indicated that most rearrangements expanded after chemotherapy and the Post 3 compared to pre-chemotherapy (Before) samples, while some only increased when Post 6 cells were stimulated with the neoantigen (Figure 2G). The observed polyclonal expansion is attributed to the neoantigen, based on TCR sequencing results from stimulated versus unstimulated PBMCs, suggesting expanded clonotypes that likely recognize the neoantigen. However, this interpretation depends on antigen-specific stimulation. Overall, these findings demonstrate the usefulness of the expanded cultures in showing the vaccine’s immunogenicity and identifying candidate CDR3 clonotypes that may be specific to the neoantigen used in vaccination.

CDR3 sequences were identified from RNA-seq performed on pre-treatment and post-chemotherapy/vaccination tumor biopsies, enabling comparison between tumor-derived clonotypes and those detected in peripheral blood. A higher number of shared CDR3 sequences was observed between the post-chemotherapy tumor and peripheral blood samples across all evaluated time points (Before, NAC, Post 3, and Post 6), suggesting an increased overlap of circulating and tumor-infiltrating T cells after therapy. No CDR3 sequences were shared between the pre-treatment peripheral blood sample (before) and either of the tumor samples. Interestingly, we identified one CDR3 sequence (CASSEEGLHNEQFF) present in both pre- and post-treatment tumors as well as in PBMCs from all post-treatment time points, but absent in pre-treatment PBMCs, indicating the emergence of a tumor-reactive clonotype in the circulation following chemotherapy. Additionally, analysis of PBMCs collected after the sixth vaccination dose (Post 6) and stimulated in vitro with the peptide revealed a clonotype shared with tumor samples; however, distinct clonotypes were observed when comparing pre-treatment (CSARKVASYNEQFF) and post-treatment (CASSPQGEEQFF) tumors, consistent with the induction of a vaccination-associated neoantigen-specific clonotype. The detailed results of these comparisons are provided in Supplementary Table 2.

To evaluate the immunogenicity of four identified neoantigens not included in the vaccine, we investigated whether vaccination combined with NAC promoted responses to these epitopes, a phenomenon referred to as “epitope spreading.” To this end, PBMCs collected after the sixth vaccination dose (Post 6) were subjected to an expansion protocol and stimulated with a pool of the four neoantigens, each derived from single amino acid mutations and exhibiting distinct HLA-I restriction (Table 1).

To assess the immune response to both the neoantigen used for vaccination and the pool of four neoantigens not included in the vaccine, an IFNγ ELISPOT assay was performed to quantify cytokine secretion as an indicator of T-cell activation. As previously reported, a response to the vaccine-derived peptide was confirmed. Additionally, a robust response to the neoantigen pool was observed, prompting further in vitro analyses to identify the specific peptide(s) driving IFNγ production (Figures 3A, B). To achieve this, the pool was deconvoluted by re-stimulating the cells from the stimulated pool with individual peptides. Cytokine expression was then evaluated using IFNγ ELISPOT and by measuring intracellular IFNγ and TNF-α through flow cytometry. Both assays demonstrated a clear cytokine secretion in response to peptide 2 (Pep-2) (Figures 3C, D), with 7.74% of cells co-expressing IFNγ and TNF-α (Figure 3E).

Once Pep-2 was identified as the main driver of the immune response, the response elicited by this neoantigen was further characterized at different levels. First, the response to Pep-2 was compared with that of its wild-type (WT) counterpart in PBMCs collected one year after vaccination. As shown in Figure 4, the mutated peptide elicited a stronger response compared to the WT version, evidenced by higher IFNγ secretion and increased activation of CD137 and PD-1 (Figures 4A, B). Based on these results, both in the post 1yr vaccination PBMCs and at the pool level after the sixth dose (Post 6), the immune response to the pool, Pep-2, and the WT version was also assessed at an earlier time point, using PBMCs collected after the third vaccine dose (Post 3). The results confirmed the secretion of IFNγ and TNF-α, both for the pool and Pep-2 at this early time point and one-year post-vaccination (Figure 4C). Moreover, a higher percentage of CD8+ TNF-α+ T cells was observed in PBMCs from Post 3 sample (Figure 4D). Together, these data point to an immune response to neoantigens that is distinct from those included in the vaccine, raising the possibility of epitope spreading—a phenomenon in which the immune system broadens its reactivity beyond the original vaccine targets, thereby enhancing the overall immune response. Nonetheless, additional assays using pre-vaccination samples would be necessary to confirm this.

To evaluate the response to this neoantigen in pre-vaccination cells, only PBMCs collected after NAC were available for analysis. Simultaneously, immunogenicity assays for Pep-2 were conducted on post-chemotherapy and one-year post-vaccination cells (Post 1yr). The results showed a CD4+ T-cell response, characterized by CD154 expression and elevated IFNγ levels in the culture supernatant (Figures 4E, F). However, after vaccination, the neoantigen-specific response appeared to shift toward a CD8-mediated response, as indicated by an increased frequency of CD8+CD137+ T cells and higher IFNγ levels in the supernatant following Pep-2 stimulation (Figures 4G, H). Notably, a response to the wild-type peptide was also observed, although at a lower magnitude. Consistently, unsupervised analysis using Phenograph identified population 20 (pink, Figures 4 I, J), characterized by the expression of CD8, CD137, CD154, and PD-1. This population expanded after in-vitro stimulation with Pep-2, primarily in PBMCs collected one-year post-vaccination, corroborating the findings from the manual analysis (Figures 4I, J, Supplementary Table 2).

These findings indicate that a response to Pep-2 was already present before vaccination; however, with the available experimental units, it is not possible to determine whether this response was caused by NAC or if it pre-existed both interventions. Since responses were observed from the time of NAC, they may instead be linked to immunogenic cell death induced by the neoadjuvant chemotherapy with doxorubicin given to this patient, a process previously reported in the literature (50, 51). In this context, vaccination appears to have enhanced the response to Pep-2, primarily mediated by CD8+ T cells that persisted for up to one year after vaccination. Nonetheless, it cannot be confirmed that epitope spreading is truly occurring, as only immunogenicity tests performed with pre-treatment cells—before both NAC and vaccination—would confirm this phenomenon.

Given that information from the whole exome sequencing (WES) obtained from patients’ PBMCs, together with WES information from the tumor, is required for identifying the universe of tumor mutations (tumor variants herein named neoantigens), and following the best practices of the Broad Institute, we set out to develop a pipeline for identifying germline variants that was validated with results from a set of samples previously characterized by a pipeline used in clinical diagnostic practice. The WES from the patient’s PBMCs was uploaded to the pipeline and analyzed for all germline variants found in a panel comprising 686 cancer driver genes (Supplementary Table 3). A total of 721 variants were identified in 316 driver genes, of which 520 corresponded to single nucleotide variants (SNPs) and 201 to insertions/deletions (InDels) (Figures 5A, B). The analysis of the distribution of variants based on their functional effects revealed that the highest proportion was concentrated in promoter regions (58), followed by synonymous variants (49) and missense variants (47). This suggests that these variants may affect both transcriptional regulation and protein structure. Variants in promoter flanking regions (43) and CTCF binding sites (24) indicate a potential impact on regulation at the epigenetic level. Additionally, variants in enhancers (12) and open chromatin regions (7) were identified, reflecting their possible influence on chromatin accessibility. The non-frameshift (4) and frameshift (2) variants could impact the integrity of the protein reading frame, while the start loss (1) implies the loss of the start codon. The remaining variants are concentrated in deep intronic regions, making the functional effect unclear. These results underscore the importance of investigating both coding variants and those in regulatory regions of the genome.

Among the variants identified, 557 are benign or probably benign, 163 have uncertain clinical significance, and one pathogenic variant in the ATM gene affects exon 14 at c.2250G>A (NM_000051.4). Although this is a synonymous variant (p. Lys750=), it occurs at the last nucleotide of exon 14, which encodes lysine. This nucleotide position is highly conserved, and in silico prediction tools support its potential pathogenicity: SpliceAI indicates a strong splice-altering effect (0.94), while dbscSNV Ada and RF classify it as deleterious. The RF algorithm predicts that SNVs within the consensus splice regions (−3 to +8 at the 5’ splice site and −12 to +2 at the 3’ splice site) may disrupt splicing, assigning scores from 0 to 1—higher scores suggest a greater risk. Functional studies have shown that splice site substitutions can cause single exon skipping; specifically, the 2250G>A variant has been documented to result in exon skipping and loss of ATM function (52). Additionally, analysis of the patient’s tumor RNA-Seq data confirms heterozygous expression of this variant at the mRNA level, supporting its potential role as an oncogenic driver in this case.

ATM is a key kinase in the DNA damage response, particularly in repairing double-strand breaks and playing a crucial role in cell cycle checkpoint signaling (53). ATM mutations are associated with an increased risk of breast, pancreatic, and lymphoma cancers due to loss of function, which leads to the accumulation of mutations (54). Additionally, ATM serves as a significant target for combination therapies, such as radiation therapy, because of its role in maintaining genomic stability (55).

We use SomaticSeq’s machine learning algorithm, which applies three confidence filters: PASS (P ≥ 0.7) for high-confidence mutations, LowQual (0.1 < P < 0.7) for ambiguous variants, and REJECT (P ≤ 0.1) for potential false positives. After filtering out and discarding the rejected variants, which are more likely to be false positives, we identified 373 affected genes and a total of 385 variants. These genes were then filtered based on cancer driver genes, and we proceeded to annotate them using the OncoKB database (https://www.oncokb.org/). The analysis found mutations in 18 cancer driver genes, of which 3 met clear oncogenicity criteria in the TP53, NF1, and PIK3CA genes with OncoKB annotation (Supplementary Table 4).

In our patient, high-confidence somatic mutations were identified in TP53 (p.Gly199Ter), NF1 (p.Phe1593SerfsTer31), and PIK3CA (p.Glu542Lys). The nonsense mutation in TP53 indicates a loss of function of the tumor suppressor gene, which may contribute to uncontrolled cell proliferation. The frameshift mutation in NF1 suggests inactivation of neurofibromin, leading to sustained activation of the RAS/MAPK pathway. The activating mutation in PIK3CA results in the continuous activation of the PI3K/AKT/mTOR pathway, promoting cell survival and growth (56, 57).

According to OncoKB, evidence levels are specific to indication, biomarker, and line of therapy and are presented here as contextual rather than prescriptive. In our case, the NF1 annotation (Level 1 for selumetinib; Level 4 for trametinib/cobimetinib) relates to NF1-associated plexiform neurofibromas and is not directly applicable to the tumor type presented. For PIK3CA p.E542K, OncoKB assigns Level 1 evidence for alpelisib plus fulvestrant and for capivasertib plus fulvestrant in HR+/HER2− disease; additional regimens such as inavolisib with palbociclib and fulvestrant (Level 3A) and RLY-2608 with or without fulvestrant (Level 4) are investigational. We include these annotations to provide context for potential therapeutic options if the clinical criteria are met; no treatment decision was made in this patient based solely on these findings (41, 58, 59).