Unique true predicted neoantigens (TPNAs) correlates with anti-tumor immune control in HCC patients

Nine HCV chronic infected HCC patients undergoing liver resection were enrolled for the present study (Additional file 1: Table S1). Two paired liver biopsies from each patient with histologically-confirmed HCC and adjacent non-tumor liver tissue were obtained at the time of surgery and stored in RNA-stabilizing agent (RNAlater, Qiagen) for RNA sequencing. Peripheral blood was obtained by venipuncture from the HCC patient HLA-008. All human specimens were obtained and processed at the National Cancer Institute in Naples under informed consent. Fresh human peripheral blood mononuclear cells (PBMCs) were isolated by Ficoll-Hypaque density gradient centrifugation and cultured in appropriate medium consisting of RPMI 1640 (Capricorn Scientific GmbH) containing 2 mM l-Glut, supplemented with 10% heat inactivated human serum (Capricorn Scientific GmbH), 25 mM HEPES buffer solution (Capricorn Scientific GmbH), 50 IU/ml penicillin and 50 μg/ml streptomycin (Gibco Life Technologies), 20 μg/ml gentamicin (Capricorn Scientific GmbH). PBMCs were maintained at 37 °C in a humidified incubator with 5% CO₂.

Human TAP-deficient T2 cell line (174× CEM.T2; ATCC CRL-1992™) was purchased from American Type Culture Collection (ATCC; https://​www.​atcc.​org/​). T2 cell line was maintained in Iscove’s modified Dulbecco’s medium (IMDM; Gibco Life Technologies) containing 25 mM HEPES and 2 mM l-Glut, supplemented with 20% fetal bovine serum (FBS; Capricorn Scientific GmbH), 100 IU/ml penicillin and 100 μg/ml streptomycin (Gibco Life Technologies). Cells were maintained at 37 °C in a humidified incubator with 5% CO₂.

Liver samples were homogenized in TRIzol reagent using the Tissue Lyser Disruption system (Qiagen), flash frozen on dry ice, and total RNA was purified according to manufacturer’s protocol (Invitrogen). RNA samples were quantified by NanoDrop1000 spectrophotometer (Thermo Fisher Scientific) and displayed a 260/280-absorbance ratio about 1.8–2. RNA quality was assessed by digital electrophoresis on Experion System by RNA StdSens Kit and RNA chips (Bio-Rad). Preparation of sequencing cDNA libraries was carried out on 4 μg of total RNA per sample using the TruSeq RNA stranded Sample Preparation Kit (Illumina). Paired-end cDNA libraries were quantified by Qubit Fluorometer (Q32866; LifeTechnologies) and Qubit dsDNA High Sensitivity Assay Kit. The overall quality of the libraries was evaluated on Experion System by DNA 1 K Analysis Kit and DNA chips (Bio-Rad). Paired-end libraries (100× 2 bp) were sequenced at high coverage on the Illumina HiSeq 2000 NGS [27].

Quality control check was performed on the total number of raw reads using FastQC tool (http://​www.​bioinformatics.​babraham.​ac.​uk/​projects/​fastqc/​). High-quality reads were mapped to human reference transcriptome and to human reference genome using TopHat2 v2.0.10 [28]. Only uniquely mapped reads were used to quantify gene expression in each sample and to compute differential expression between non-tumour and tumour samples [29]. Principal component analysis (PCA), MA and density plots for an overall control of the experiment were generated using RNASeqGUI. Differential expression between tumour and adjacent non-tumour samples was evaluated using the GLM implemented in EdgeR. FDR < 0.01 was used as threshold to determine differentially expressed genes (DEGs).

The analysis of pathways and gene ontology for the DEGs was performed using DAVID, both implemented in the GUI. Multiplicity correction (false discovery rate, FDR) was used for enrichment of DEGs, with a threshold of 0.05.

For the SNP calling, PCR duplicates were removed using Picard tools v1.117 and SNP calling was carried out using GATK v3.3 workflow optimized for RNA-seq reads. Nucleotide variants were then annotated using ANNOVAR. Common variants reported in dbSNP (v138), in the 1000 genome database, as well as those identified in adjacent non-tumor liver tissue were filtered out from our dataset. SNVs in super-duplicated regions were also removed.

Epitope prediction was performed for non-synonymous somatic SNVs using prediction tools available at http://​www.​cbs.​dtu.​dk/​services/​.

The NetTepi version 1.0 as well as NetMHCstabpan version 1.0 were used to predict MHC class I HLA-A*02:01 allele restricted epitopes. A peptide sequence of 30 amino acid, centered on the mutated amino acid, was used to predict 9-mer neoantigens scanning the entire sequence by overlapping peptides. In both servers, epitopes were predicted based on the combination of different parameters and ranked on prediction values (%rank). Predicted epitopes were selected based on %rank ≤ 2. The Immune Epitope Database (IEDB; http://​www.​iedb.​org/​) was used for analysis of sequence homology to experimentally validated human and pathogen-derived antigens. Known antigens with homology > 70% to mutated antigens were identified, but only those with matching aa residues at the TCR binding positions (positions 1, 4, 5 and 8) were selected for subsequent analyses. Homologous validated antigens were subsequently confirmed by the NetMHCstabpan version 1.0.

Peptides were synthesized at a purity > 95%. Lyophilized powder was dissolved in dimethylsulfoxide (DMSO; Sigma-Aldrich), diluted in phosphate-buffered saline (1× PBS; Gibco Life Technologies) and stored at − 80 °C until use.

Peptide binding affinity to HLA-A*02:01 molecule and BFA decay assays were performed for each candidate peptide. Briefly, T2 were seeded at 3.5 × 10⁵ cells per well in 24-well plates and incubated overnight with peptides (final concentrations: 10 μM, 20 μM, 50 μM, 100 μM) in IMDM serum-free medium containing 3 μg/ml human β2-microglobulin (Sigma-Aldrich) in a humidified incubator at 37 °C with 5% CO₂. Following incubation, cells were harvested and centrifuged at 200×g for 5 min. Subsequently, cells were washed twice with phosphate-buffered saline (1× PBS; Gibco Life Technologies) and stained with R-PE conjugated anti-human HLA-A2 monoclonal antibody (cat. 343306; BioLegend), for 30 min at 4 °C, and analyzed with the Attune™ N×T flow cytometer (Thermo Fisher Scientific). OVA peptide was used as negative control and T2 cells without any added peptide were used as a background control. A fluorescence index (FI) was calculated using the following formula: FI = [mean fluorescence intensity (MFI) sample − MFI background]/MFI background, where MFI background represents the value without peptide. An FI > 0.5 was set as threshold to indicate peptides with affinity for the HLA-A*02:01 molecule. For the brefeldin A decay assay, T2 were seeded at 5 × 10⁵ cells per well in 24-well plates and cultured overnight with either the candidate peptides or the control peptide (CAP-1 was used as control) at a final concentration of 50 μM, at 37 °C in a humidified incubator with 5% CO₂, in serum-free IMDM medium containing 3 μg/ml β2-microglobulin (Sigma-Aldrich). Following incubation, cells were washed and incubated with 1× BFA (brefeldin A solution; cat. 420601; BioLegend) in IMDM serum-free medium, for 1 h at 37 °C. Cells were harvested every 2 h (T0, T2, T4, T6, T8), washed with phosphate-buffered saline (1× PBS; Gibco Life Technologies), stained with anti-HLA-A*0201 fluorescent monoclonal antibody (cat. 343306; BioLegend) and analyzed by flow cytometry. The stability of each peptide bound to HLA-A2 was measured as the DC₅₀ value, which was defined as an estimate of the time required for a 50% reduction of the MFI value recorded at time 0. The DC₅₀ value was calculated according to the formula: MFI at indicated time points/MFI at time 0 × 100. All the experiments were performed in triplicate.

C57BL/6 (H-2b MHC) female mice, 8 week old, were purchased from Harlan (Udine, Italy). All animals were housed at the Animal Facility of the Istituto Nazionale Tumori “Pascale” (Naples,Italy). Mice were housed in number of 2–3 per cage and maintained in a conventional facility on a 12 h light:12 h dark cycle (lights on at 7:00 a.m.) in a temperature-controlled room (22 ± 2 °C) and with food and water ad libitum at all times. The experimental protocols were in compliance with the European Communities Council directive (86/609/EEC) and were approved by the Italian Ministry of Health (approval number 835/2016).

Animals (6 for each group) were injected by sub-cutaneous route with 100 µg of each peptide, emulsified with 50 μg of Polyinosinic:polycytidylic acid [poly(I:C); InvivoGen] adjuvant formulated in PBS (200 μl total volume). Immunization was performed twice a week for a total of six administrations. At the time of sacrifice, spleens were resected and processed into single cell suspensions using a gentle MACS Dissociator (Miltenyi Biotec) according to the manufacturer’s instructions for immunological evaluation.

IFN-γ ELISpot (BD™ human IFN-γ ELISPOT Set) assay was performed on splenocytes from the immunized mice as well as PBMCs from the HLA-008 HCC patient.

For the splenocytes from immunized mice, 2 × 10⁵ splenocytes pooled from animals in each group were counted and plated in each well. Cells were stimulated with 10 μg/ml of single and pool peptide used for the immunization and incubated for 24–26 h. As negative and positive control, peptide diluents PBS and 10 μg/ml of PHA (PHA-K; Capricorn Scientific GmbH) were used respectively.

For the PBMCs from the HCC patient and healthy donors, 4 × 10⁶ PBMCs/ml/well were stimulated with peptides at a final concentration of 10 μg/ml. On day three, 10 U/ml IL-2 was added to each well. On day five, half of the volume of medium was replaced with fresh medium containing IL-2 at a final concentration of 10 U/ml. On day seven, PBMCs were re-stimulated with each peptide. On day 10, cells were harvested for IFN-γ ELISpot assay. Each peptide was added at a final concentration of 10 μg/ml to 2 × 10⁵ PBMCs per well in 100 μl RPMI 1640 medium (Capricorn Scientific GmbH). PBMCs were cultured at 37 °C in a humidified incubator with 5% CO₂ for 20 h. Stimulation with 10 μg/ml PHA (PHA-K; Capricorn Scientific GmbH) was used as positive control, PBMCs without added peptides were used as the negative control, RPMI 1640 medium (Capricorn Scientific GmbH) was used as background control.

The plates were read with an AID EliSpot Reader Systems (AID GmbH, Strassberg, Germany). Determinations from triplicate tests were averaged. Data were analyzed by subtracting the mean number of spots in the wells with cells and medium-only from the mean counts of spots in wells with cells and antigen. Spot forming units (SFU) were calculated as the frequency per 10⁶ PBMCs.

Comparison between individual data points were performed with the unpaired two-sided Student’s t-test and ANOVA, as appropriate. Normally distributed data were represented as mean ± SEM. All p values were two-tailed and considered significant if less than 0.05.

Nine HCC patients were enrolled for the present study. They were all HCV chronically infected, with an average age of 72.8 years old. They were all positive for the HLA-A*02:01 haplotype and only two (HLA-008 and HLA-026) were positive also for the HLA-A*24:02 haplotype. Three patients (HLA-008, HLA-009 and HLA-012) are still alive at the moment of submission, but only HLA-008 is still living in the area (Additional file 1: Table S1).

Samples from paired primary HCV-related HCC and non-tumour adjacent liver tissues were subjected to RNA-seq analysis, and a clear segregation of gene expression values in tumor and the adjacent liver tissues was observed (Additional file 2: Fig. S1A, B).

The expression of 2101 protein-coding genes was modulated in the HCC samples vs. the corresponding non tumour adjacent tissues, with a degree of differential expression ranging from − 8.35 to 8.98 log2FC (FDR < 0.01; Fig. 1a). Most of differentially expressed genes (DEGs) (about 78%, i.e. 1639 genes) were significantly down-modulated in HCC samples, suggesting a strong gene transcription silencing in HCC lesions (Fig. 1b). A significant alteration of 15 biological processes and 13 pathways, such as cell adhesion molecules, viral infection and drug metabolism pathways (Bonferroni adjusted p values ≤ 0.05), was observed (Fig. 1c). Notably, PI3K, Wnt, p53 and Ras signaling pathways were significantly altered in HCV-related HCC vs non-tumor adjacent tissues (Additional file 2: Fig. S2A–D).

In order to identify potential neoantigens specific for HCV-related HCC, reads generated by RNA-seq were used for the SNP calling. More than forty-thousand Single Nucleotide Variations (SNVs) were found in the exons of our samples, of which more than 50% were non-synonymous SNVs (nsSNVs). 1516 potentially somatic mutations in 1357 genes were selected, based on criteria described in Materials and methods (Additional file 2: Fig. S3A). The number of mutations identified in each sample are reported in Additional file 2: Fig. S3B. As expected, the largest fraction was constituted by nsSNVs and about 25–30% of them were predicted to be deleterious to the protein functionality (Additional file 2: Fig. S3C).

We then focused only on the 1100 nsSNVs found in a set of 249 genes highly expressed (4th quartile) both in HCC and matched non-tumor liver tissues. The vast majority of genes bearing the selected mutations were not shared among the HCC samples. Only in few cases, the same gene was mutated in different HCC samples but the observed nsSNVs were always distinct (Additional file 1: Table S2).

All the 1100 mutations nsSNVs were evaluated for MHC class I HLA-A*02:01 allele-restricted neoantigen prediction according to the NetTepi and NetMHCstabpan parameters described in “Methods” section.

The prediction analysis was performed with the two servers running a side-by-side comparison between paired wild-type and mutated peptide sequences. According to such analysis, in each sample most of the mutations were not predicted to be neoantigens (NP) (Fig. 2a). According to the default threshold of total %rank ≤ 2, the two servers identified in each samples predicted neoantigens (PNAs) which overlapped in the majority of cases (Additional file 2: Fig. S4). Overall, the NetTepi server predicted 118 neoantigens and the NetMHCstabpan predicted 104 neoantigens, of which 84 were common (Fig. 2b). Considering only the neoantigens predicted by both servers, the average number of PNAs in the HCC samples was 9.33, ranging from 19 (HLA-029) to 2 (HLA-012) (Additional file 2: Fig. S5). However, 46 of the 84 PNAs (71.5%) should be considered as “false” predicted neoantigens (FPNAs) given that their corresponding wt peptides are characterized by a similar predicted antigenicity (e.g. both with a %rank ≤ 2) (Additional file 1: Table S3). Therefore, only 38 PNAs are “true” predicted neoantigens (TPNAs) in the whole set of HCC samples, given that their corresponding wt peptides are not predicted to be antigenic (Fig. 2b; Additional file 2: Fig. S6). The average number of TPNAs in the HCC samples was 4.22, ranging from 7 (HLA-008 and 022) to 1 (HLA-012) (Additional file 2: Fig. S5).

In order to closely analyze the “false” PNAs, the sequences of FPNAs were aligned to the corresponding wild type peptides. According to previous findings [30], we compared the aminoacid residues at positions p1, p4, p5 and p8, which are exposed to the T cell receptor (TCR) when epitopes are allocated in the HLA groove. The analysis confirmed that 24 of the 46 (52.2%) FPNAs show all 4 aa residues identical to the corresponding wild type epitope; the remaining FPNAs (47.8%) show homology in 3/4 aa residues (Additional file 2: Fig. S7).

In order to experimentally confirm the binding and stability of TPNAs to HLA-A*0201 molecule, the HLA-A*02-positive T2 cell line was loaded with 10 TPNAs with high (< 0.8 in NetTepi and < 1.3 in StabPan) and low (0.8 < %Rank < 2 in NetTepi and 1.3 < %Rank < 2 in StabPan) prediction %rank. The analysis showed that only the peptides with high prediction %rank (MYO18A_mut and WDR7_mut) induced a significant increase in the HLA surface expression on T2 cells over background as well as OVA negative control level (Fig. 3a).

The peptide—MHC dissociation kinetics showed that, independent of the %rank, all peptide-MHC complexes showed a 50% dissociation value between T4 and T6, although with different MFI values for the different peptides (Fig. 3b).

Consequently, selecting only the TPNAs with the high prediction %rank in both servers, the overall number of high affinity and stability TPNAs dropped down to nine, with some of tumour samples (e.g. HLA-012, 022, 026) showing no predicted ones (Additional file 1: Table S4). As predictable, similar results have been obtained using FPNAs with different prediction %rank, confirming that only epitopes with high prediction values show binding to HLA molecules. On the contrary, none of the not predicted neoantigens (NP) showed binding to HLA molecules (data not shown). To further confirm the antigenicity of such predicted neoantigens, they were used to stimulate ex vivo also PBMCs from HLA-A*02:01 healthy donors. Results confirmed that only the peptides with high prediction %rank (MYO18A_mut and WDR7_mut) were able to induce release of IFNγ in PBMCs (Fig. 3d).

In order to identify possible consensus patterns in the TPNAs aminoacid sequences, the TPNA sequences were aligned all together or divided in high (< 0.8 in NetTepi and < 1.3 in StabPan) and low (0.8 < %Rank < 2 in NetTepi and 1.3 < %Rank < 2 in StabPan) (9 and 44 sequences, respectively). As predictable, amino acid residues at the HLA binding positions p2 and p9 showed the highest conservation across sequences, and only minor differences were observed in the three different alignments (Additional file 2: Fig. S8). On the contrary, the TCR binding positions p1, p4, p5, p8 showed a significant variability with no prevalent aminoacid residue. However, the 9 TPNAs with highest %rank showed a profile at these positions significantly different from the one observed in the other two alignments (Additional file 2: Fig. S8A–C). In particular, at positions p1, p4 and p8, it was observed a prevalence of glycine (G) instead of leucine (L) (Additional file 2: Fig. S8D–F).

In order to definitely confirm that the TPNAs are indeed neoantigens and do not share homology with any known antigen reported in the literature, a blast search was run against epitopes in the Immune Epitope Database (IEDB; http://​www.​iedb.​org/​).

According to such analysis, 31 TPNAs showed a sequence homology to known human and/or pathogen-derived epitopes (Additional file 1: Tables S5, S6). However, a prediction analysis with NetTepi confirmed that only 4 human self epitopes and 9 pathogen-derived epitopes were associated to the HLA-A02*01 haplotype. Of these, considering the homology at TCR binding positions, only 3 predicted human self epitopes and 3 pathogen-derived epitopes could really be considered homologous to the corresponding TPNAs (Additional file 2: Fig. S9). Interestingly, the sample HLA-008 was the only one showing more than one TPNA homologous to known antigens. In particular, it showed 3/7 TPNAs homologous to pathogen-derived epitopes (e.g. Lassavirus, Shigella and Vaccinia virus) without homology to any cellular self epitope.

In order to prove that the TPNAs and their homologous known epitopes were able to elicit a cross-reactive immune response, an in vivo immunization experiment was performed in a C57BL/6 mouse model. Animals were independently immunized with the HLA-008 TPNA CHD3_mut or its homologous Shigella epitope, and with the HLA-017 TPNA NOP2_mut or its homologous human self epitope DDIT3. All such epitopes, indeed, bind to human HLA-A*0201 as well as to mouse H-2-Db/Kb haplotypes. However, while the CHD3_mut and the homologous Shigella epitopes are predicted to bind both H-2-Db and H-2-Kb haplotypes, the NOP2_mut and the homologous DDIT3 epitopes are predicted to bind only the H-2-Kb haplotype (Additional file 1: Table S7).

The results confirmed that the CHD3_mut and the homologous Shigella epitopes induced a much stronger immune response than the NOP2_mut and the homologous DDIT3 epitopes. Moreover, as predicted by the bioinformatics algorithm, PBMCs from animals immunized with the TPNA CHD3_mut and the homologous Shigella epitope, respectively, cross reacted against both epitopes. The reactivity against the homologous epitope was about 50% of the one observed against the peptide used for the immunization (Fig. 4a). Such a result can be due to the mismatch in the alignment of the residues at the TCR binding positions (Additional file 2: Fig. S9). As predictable, PBMCs from animals immunized with the TPNA CHD3_mut showed stronger reactivity to CHD3_wt than to Shigella due to the 100% homology in the aa residues at the TCR binding positions. Similar results, but at much lower intensity, were obtained in mice immunized with the NOP2_mut and the homologous DDIT3 epitopes (Fig. 4a).

In order to verify whether the cross reactive immunity between pathogen-derived epitopes and TPNAs correlates with improved survival, PBMCs from the only HCC patient still surviving to date (HLA-008) were obtained. The results clearly showed that the patient had variable levels of circulating T cells reacting to all three TPNAs as well as to the pathogen-derived epitopes. However, the two most reactive peptides were the TPNA PDCD7_mut and its homologous Vaccinia virus epitope, strongly suggesting that the pre-existing memory T cell response induced by the vaccinia virus vaccination may favor a stronger reactivity against the tumor-specific neoantigen (Fig. 4b). Interestingly, the T cell reactivity against the CHD3_mut peptide was of the same magnitude as for the PDCD7_mut peptide, although the levels of pre-existing memory T cell response against the homologous Shighella epitope was significantly lower (Fig. 4b). On the contrary, the low reactivity to the MYO18_mut did not reflect the predicted high antigenicity and stability and was not boosted by the low levels of pre-existing memory T cell response specific for its homologous epitope derived from Lassavirus. Finally, as predictable by the lack of matching in the TCR binding positions, no cross-reactivity was observed between the CWF19L1_mut and hepatitis C virus epitope (data not shown).

The intratumor immune landscape for each of the HCC samples was assessed evaluating the transcription levels of immune-related genes. In particular, infiltration of effector and regulatory cell types as well as expression of immune checkpoint and antigen presenting molecules were evaluated by such analysis. The immunophenotype (IP) for each HCC patient was generated using the online tool available at The Cancer Immunome Atlas (https://​tcia.​at/​). A more pronounced immune effective microenvironment, characterized by higher gene expression of effector cell types and antigen presenting molecules, gives a higher IP score [31].

The results showed that, with the exception of the HLA-012, HLA-017 and HLA-029 with an IP < 6, all other HCC samples had an IP ≥ 6 (Fig. 5), which has been reported to be associated with clinical benefit and/or long-term survival. Interestingly, the long-term survivors HLA-008 and HLA-009 showed an IP ≥ 6, supporting the correlation between the immune determinants included in the immunophenogram and the clinical outcome.