A method for identifying at least one tumor-specific antigen candidate and uses thereof

The k-mer match algorithm efficiently identifies tumor-specific antigens by comparing transcript variants against a comprehensive database, addressing inefficiencies in existing methods and enhancing the precision of cancer immunotherapy for tumors with low mutational burden.

WO2026238918A1PCT designated stage Publication Date: 2026-11-19TRON TRANSLATIONALE ONKOLOGIE AN DER UNIVERSITAETSMEDIZIN DER JOHANNES GUTENBERG UNIV MAINZ GEMEINNUETZIGE GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/064140
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-13
Filing Date
2026-05-13
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing methods for identifying tumor-specific antigens in tumors with low mutational burden are inefficient and computationally expensive, as they require extensive comparison with healthy tissue datasets, which is impractical and limits sensitivity and precision.

Method used

A k-mer match algorithm is used to identify tumor-specific antigen candidates by comparing transcript variant sequences from a biological sample against a reference k-mer database of healthy and cancerous RNA-sequencing datasets, utilizing k-mer indexing tools to efficiently determine tumor-specificity.

Benefits of technology

This method allows for rapid and accurate identification of tumor-specific antigens, enabling personalized cancer immunotherapies by reducing computational burden and improving sensitivity, particularly for tumors with low mutational burden.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure IMGF000052_0001
    Figure IMGF000052_0001
  • Figure IMGF000052_0002
    Figure IMGF000052_0002
  • Figure IMGF000052_0003
    Figure IMGF000052_0003
Patent Text Reader

Abstract

The present invention relates to a method for identifying at least one tumor-specific antigen candidate in a biological sample obtained from a subject. The method uses an exact k-mer match algorithm of spliced junctions in transcript variants for the identification of at least one tumor-specific antigen candidate, whereby the transcript variants of the sample can be analyzed faster and more accurately without bias. Further subject matter of the present invention is a tumor-specific antigen candidate identified according to the method of the invention, a pharmaceutical composition comprising the tumor-specific antigen candidate identified according to the method of the invention as well as their uses in the diagnosis, prevention, and / or treatment of a proliferative disease. Further the invention relates to a computer-implemented method, system and / or computer product for identifying at least one tumor-specific antigen candidate in a biological sample obtained from a subject.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] A method for identifying at least one tumor-specific antigen candidate and uses thereof

[0002] Description

[0003] The present invention relates to a method for identifying at least one tumor-specific antigen candidate in a biological sample obtained from a subject. The method uses an exact k-mer match algorithm of spliced junctions in transcript variants for the identification of at least one tumor-specific antigen candidate, whereby the transcript variants of the sample can be analyzed faster and more accurately without bias.

[0004] Further subject matter of the present invention is a tumor-specific antigen candidate identified according to the method of the invention, a pharmaceutical composition comprising the tumor-specific antigen candidate identified according to the method of the invention as well as their uses in the diagnosis, prevention, and / or treatment of a proliferative disease.

[0005] Further the invention relates to a computer-implemented method, system and / or computer product for identifying at least one tumor-specific antigen candidate in a biological sample obtained from a subject.

[0006] Background

[0007] Cancer immunotherapies rely on tumor-specific peptides that are presented as antigens by MHC molecules and recognized by T-cells. Neoantigens are encoded by nonsynonymous somatic mutation, which are identified by comparing DNA sequencing data from tumors and matched normal samples (Lang F et al., Identification of neoantigens for individualized therapeutic cancer vaccines, Nat Rev Drug Discov. 2O22;i-22 and Sahin U et al., Personalized RNA mutanome vaccines mobilize poly-specific therapeutic immunity against cancer, Nature. 2017, 547:222-6). However, many tumors have low mutational burden (TMB) and, therefore, are insufficient targets for immunotherapies. For these patients it is crucial to expand the antigen target repertoire. Additional tumor-specific antigens are encoded by non-canonical RNA transcripts, such as gene fusions or alternatively spliced mRNA transcripts which can be detected by RNA-seq from tumor samples (Weber D et al., Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat Biotechnol.2022;40:1276-84). For example RNA-seq identified between 500 and 24,000 novel non-canonical splice junctions (SJ) per tumor sample. However, for those tumor RNA-seq-derived antigen targets, tumor-specificity cannot be unambiguouslyensured because the expression variability between tissues makes it insufficient to choose a single tissue (e.g., blood or adjacent normal) as suitable germline control. Tumor specificity is, however, necessary to ensure the efficacy of the predicted neoantigen candidates with regard to immunogenicity and preventing off-target effects in other organs.

[0008] To ensure tumor-specificity of transcript variants detected in tumor RNA-seq, different strategies have been proposed in literature as described in the section „Ap plying de novo detection pipelines to healthy samples “.

[0009] Since obtaining a comprehensive set of healthy tissue samples from the same patient is not feasible, the problem of demonstrating tumor-specificity can be approximated using diverse unmatched RNA-seq datasets from healthy donors (Frankiw L et al., Alternative mRNA splicing in cancer immunotherapy. Nat Rev Immunol.2019; 19, 675-87 and Oreper Det al., The peptide woods are lovely, dark and deep: Hunting for novel cancer antigens. Semin Immunol. 2023, 67:101758), assuming that transcript variants that are consistently absent from diverse healthy tissue samples are tumor-specific. For example, the GTEx (Lonsdale J et al., The Genotype-Tissue Expression (GTEx) project. Nat Genet. 2013, 45:580-5) resource contains RNA-seq from more than 17k samples from 54 different healthy tissues, and the public sequencing read archive (SRA) contains over 750.000 human samples and continues to grow exponentially. Applying de novo detection pipelines to healthy samples to build exclusion lists is computationally expensive, in terms of required disk space and compute resources, and has limited sensitivity because such pipelines are optimized for precision (e.g., often junctions with < 10 supporting reads are ignored). Furthermore, these expensive runs must be repeated when the detection pipeline, the reference annotation or genome is updated.

[0010] For instance, the recounts project (Wilks C et al., recounts: Summaries and queries for large-scale RNA-seq expression and splicing. Genome Biol. 2021^2:323) compiled gene, exon and splice junctions counts from public databases totaling to 316,443 human samples, allowing to use this rich source of healthy and cancerous annotation-dependent expression data for filtering. However, other transcript variants such as fusions or circular RNAs are not included in this resource, requiring extensions in the future. While cloud computing has become a viable option, it can be considered a huge effort to expand this dataset. Compiling the current recounts dataset required 29,000 node-hours of computing resources and 150 Tb of disk space for the processed data, demonstrating the limitations of the exclusion list approach.Another approach map reads from all healthy samples against a given novel target sequence (Weber D et al., Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat Biotechnol. 2022;40:1276-84). While such requantification is more sensitive due to the direct alignment, it requires repeated mapping runs with all reads from thousands of healthy samples for each individual tumor patient and is computationally too expansive for individualized neoantigen detection pipelines (e.g. for every patient 17k alignments when using normal samples from GTEx).

[0011] Therefore, there is a need for methods that can determine for a given novel transcript sequence whether it is tumor-specific, by efficiently comparing against thousands of healthy tissue RNA-sequencing datasets.

[0012] Recently, novel algorithms and data structures were developed to compress sequences into indices of short (usually 19 to 31 nucleotides) sub-sequences of length k (k-mers). While there was a strong focus on k-mer indexing algorithms and data structures in theoretical and algorithmic bioinformatics research during the past ten years, the application of such methods was, so far, limited to basic research questions, mainly in metagenomics.

[0013] Therefore, there is an unmet need to develop individualized cancer immunotherapies for tumors with low mutational burden.

[0014] The present invention addresses this unmet need by providing a k-mer match algorithm and by taking advantage of its usability to identify tumor-specificity of (neo)antigen candidates.

[0015] In a first aspect, the invention pertains to a method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from at least one subject, more in particular from at least two subjects, the method comprising:

[0016] (a) Providing RNA sequencing data from said biological sample,

[0017] (b) Identifying transcript variant sequences in the RNA sequencing data from step (a), (c) Dividing the transcript variant sequences from step (b) into short overlapping subsequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample,

[0018] (d) Providing reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples,(e) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (c) to the reference k-mer database of reference biological samples from step (d),

[0019] (f) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (e).

[0020] The method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from at least one subject, more in particular from at least two subjects, according to the present invention, the method comprises a step of (a) providing RNA sequencing data from said biological sample. This step (a) of providing RNA sequencing data from said biological sample may comprise (a-i) extracting full RNA of a subject in a biological sample comprising non-canonical RNA transcripts, such as gene fusions or alternatively spliced mRNA transcripts, (a-2) obtaining transcript variant sequence information by sequencing the entire RNA of the biological sample (sequence information analysis). The extracting of full RNA from a biological sample maybe carried out using any RNA extraction method known in the art.

[0021] In a preferred embodiment of the method of the present invention, step (b) further comprises identifying transcript variant sequences by filtering of the provided RNA sequencing data. In a further preferred embodiment of the method of the present invention, step (b) further comprises providing structural variant information to identify transcript variant sequences, in particular fusion transcripts.

[0022] In another preferred embodiment of the method of the present invention, in step (e) the identified transcript variant sequences are verified using an orthogonal verification method, in particular comprising RT-PCR, qRT-PCR and / or Sanger sequencing.

[0023] According to a preferred embodiment of the method of the present invention, step (f) further comprises predicting binding of one or more epitope peptides, comprised within the tumorspecific antigen candidates derived from the identified transcript variants, to one or more HLA molecules.

[0024] In another preferred embodiment of the method of the present invention, step (f) further comprises selecting tumor-specific antigen candidates comprising epitope peptidespredicted to bind to HLA molecules of more than one subject, in particular thereby identifying shared variants.

[0025] According to a further preferred embodiment of the method of the present invention, step (f) further comprises filtering the identified tumor-specific transcript variant sequences.

[0026] The term “RNA extraction” means herein the purification of RNA from biological samples.

[0027] RNA extraction methods are known by a person skilled in the art. The extraction method is challenging because of the ubiquitous presence of ribonuclease enzymes in cells and tissues, which can rapidly degrade RNA. Several methods are known by a person skilled in the art, the most common of these methods is guanidinium thiocyanate-phenol-chloroform extraction. Usually, the phenol-chloroform solution used for RNA extraction has lower pH, this aids in separating DNA from RNA and leads to a more pure RNA preparation. The filter paper based lysis and elution method features high throughput capacity. RNA extraction in liquid nitrogen, commonly using a mortar and pestle (or specialized steel devices known as tissue pulverizers) are useful in preventing ribonuclease activity.

[0028] The step (b) of identifying transcript variant sequences in the RNA sequencing data from step (a) comprises using the obtained transcript variant sequence information by sequencing the RNA from the entire genome of the biological sample by e.g. high throughput RNA-seq to obtain short sequencing reads (typically 50-i50bp in length) and computational analysis (e.g. read mapping to reference genome or transcriptome and downstream computational analysis and annotation) to detect transcript variants (sequence information analysis).

[0029] The term “transcript variant sequences” as used in this invention means herein the various mRNA sequences expressed from most eukaryotic genes. This complex phenomenon arises from various mechanisms, such as DNA mutations in the encoding gene, using alternative transcription start sites and alternative post-transcriptional processing events (e.g. RNA-editing or splicing), or transcription of foreign nucleotides, e.g. from viral or microbial infections. The resulting transcript variants can lead to synthesizing proteins that possess distinct functional domains or may even generate noncoding RNAs, each with unique roles in cellular processes. The generation of these transcript variants is not merely a random occurrence; it is cell-type specific and varies with developmental stages, aging processes, or pathogenesis of diseases. This highlights the biological significance of transcript variants in regulating gene expression, their potential impact on cellular functionality and their distinct occurrence in specific tumors.The sub-step (a-i) may be carried out separately and the providing RNA sequencing data from said biological sample may start with the sub-step (a-2) of obtaining transcript variant sequence information by sequencing the entire RNA of the biological sample (sequence information analysis).

[0030] Thus, the (a-i) RNA extraction step may not be included in the method for identifying at least one tumor-specific antigen candidate according to the present prevention. In the step of providing RNA sequencing data from said biological sample of the present invention, the sub-step of extracting full RNA from the biological sample is not particularly limited and may be performed in any manner known in the art for RNA extraction.

[0031] The step (c) of dividing the transcript variant sequences from step (b) into short overlapping sub-sequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample,

[0032] The sample k-mer dataset for said transcript variant sequences from said biological sample can be created by fragmenting the transcript variant sequences into k-mer-long letter strings the fragmenting site on each of the sequencing reads shifting by one base for each fragment, using a computer device.

[0033] The creation of the sample k-mer dataset is performed using k-mer indexing tool. The exemplary k-mer indexing tools are Raptor (Mehringer S et al. Hierarchical Interleaved Bloom Filter: enabling ultrafast, approximate sequence queries. Genome Biol.

[0034] 2O23;24:131) and kmindex (Lemane T et al. kmindex and ORA: Indexing and real-time user-friendly queries in terabytes-sized complex genomic datasets. bioRxiv, 2023) with k-mer size 21. The method according to this invention is not particularly limited to these k-mer indexing tools and may be performed with any k-mer indexing tool known in the art for k-mer indexing.

[0035] The term k-mer as used herein means a transcript variant composed of as the number of nucleotides. The k-mer or k-mer fragment of the transcript variant according to the present invention refers to a variant-containing sequence (termed context sequence) which is fragmented to a set of nucleotide sequences of length "k". The variant containing context sequence is defined as the transcript sub-sequence starting k-i nucleotide before the actual variant start position (e.g. splice junction position or mutation position) and ending k-i positions after the variant end position (e.g. splice junction position or end of variantsequence). When a sequence is given, the term also refers to a collection of all possible subsequences, each being a k-mer long.

[0036] In the method described herein, at least one k-mer fragment sequence is created from a transcript variant sequence of RNA sequence from a biological sample.

[0037] For example, the "AGCTCT" sequence can be divided into the 3-nt subsequences "AGC", "GCT", "CTC", and "TCT". These subsequences are each k-mer wherein k is 3. In general, K-mers may or may not be overlapped.

[0038] The method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from at least one subject, more in particular from at least two subjects, according to the present invention, the method comprises a step of (d) providing reference k-mer database comprising set of k-mers of RNA sequencing datasets of reference biological samples.

[0039] The transcript variant sequences dataset assigned reference k-mer database of reference biological samples maybe built by (d-i) extracting full RNA of a subject in a biological sample comprising non-canonical RNA transcripts, such as gene fusions or alternatively spliced mRNA transcripts, (d-2) obtaining raw RAN-seq read data by sequencing the RNA of the biological sample, (d-3) dividing the raw RNA-seq reads into k-mers and storing the entire set of unique k-mer that appeared at least twice per sample in an efficient data structure (k-mer index).

[0040] The term “Raw sequencing reads” as used herein means raw sequencing reads are the initial data produced by various high-throughput sequencing platforms, representing short DNA or RNA fragments generated during a sequencing run. These reads are typically stored in formats such as FASTQ, containing both the nucleotide sequences and associated quality scores for each base call. The raw data can originate from a variety of sequencing technologies, such as Illumina sequencing, which relies on short-read paired-end technology, or Oxford Nanopore sequencing, which produces long reads. Other methods, such as Pacific Biosciences' Single Molecule Real-Time (SMRT) sequencing, also generate long reads with different error profiles. Regardless of the platform, raw reads are essentially unprocessed and can contain sequencing artifacts, base-calling errors, or contaminants. These reads must undergo preprocessing steps like quality filtering, adapter trimming, and alignment to areference genome or transcriptome before being suitable for downstream analyses such as genome assembly, variant detection, or transcriptome analysis.

[0041] The reference k-mer database contains 18,961 samples from 52 healthy tissues and 10,328 samples from 30 tumor entities from public resources. When a new biological sample with sequencing data is provided, the reference k-mer database may be rebuilt or extended therewith.

[0042] In the reference k-mer database, sample information is assigned to individual reference k-mer sequences. Additional metadata may given further information about some known characteristics including a sample source such as a specific tumor, a phenotype associated with possibility of a specific cancer type.

[0043] In a preferred embodiment of the method according to the present invention the subject is a human suffering from a proliferative disease, in particular from cancer, and the biological sample is a proliferative disease cell sample, in particular a cancer tumor cell sample selected from a group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors, neuroendocrine tumors, in particular selected from the group consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

[0044] Further preferred is the method according to the present invention, wherein the cancer tumor cells are selected from colorectal cancer cells, bowel cancer cells, rectal cancer cells, breast cancer cells, cervical cancer cells, endocervical cancer cells, colon cancer cells, esophageal cancer cells, brain cancer cells, head and neck cancer cells, renal cancer cells, meningeal cancer cells, glioma, glioblastoma, lung cancer cells, mesothelioma, ovarian cancer cells, pancreatic cancer cells, neuroendocrine cancer cells, prostatic cancer cells, skin cancer cells, stomach cancer cells, thyroid cancer cells, uterine cancer cells, and testicular cancer cells.

[0045] In a further preferred embodiment of the method according to the present invention the reference k-mer database of reference biological samples from step (d) is a collection of RNA sequence data from biological samples of healthy subjects and / or subjects suffering from cancer, and wherein each k-mer RNA sequence dataset of biological samples was obtained by applying k-mer indexing to the plurality of raw RNA-sequencing reads.A combination of transcript variant sequences and k-mer fragments from biological samples from healthy subjects and subjects suffering from cancer enables on the one hand the identification of a tumor-specific antigen candidate, which is present in the biological sample obtained from a subject, but is not present or only present in a small subset of samples (e.g. below a given tolerance prevalence of 1%) in the sequences of biological samples of the healthy subject in the reference k-mer database and is therefore a promising candidate for antigen for immunotherapy.

[0046] On the other hand, the combination comprising the samples of subjects suffering from cancer enables the detection of recurrence in tumor samples. Given a transcript variant sequence of interest in the biological sample, the method permits to check within seconds if this transcript variant sequence is present in thousands of reference tumor samples. This requires to build a k-mer index of the reference tumor RNA-seq data next to indexing healthy tissues for tumor-specificity prediction. The reference tumor RNA-sequencing data is not directly useful to detect novel antigens that are individual to a single patient, but the reference tumor RNA-sequencing data can help to detect, select, and prioritize a lead candidate antigen that is recurrent in many tumor patients (from one or multiple tumor types). This fact is beneficial for e.g. T-cell based therapies, or antibody therapies.

[0047] The method according to this invention can be carried out with a reference k-mer database based alone on reference biological samples from healthy subjects or on reference biological samples from subjects suffering from cancer or on a reference k-mer database based on the combination of reference biological samples from healthy subjects and biological samples from subjects suffering from cancer.

[0048] In a preferred embodiment, the k-mer used for transcript variant sequences in the method of the present invention may have a size or length of to to 50 nucleotides (nt), 12 to 40 nt, 14 to 30 nt, 16 to 25 nt, or 18 to 21 nt. A too short k-mer sequence does not allow the provision of a sufficient number of k-mer sequences necessary for discriminating transcript variant sequences. On the other hand, when too long k-mers are used, an increased number of sequence combinations is obtained to make it possible to accurately match for specific transcript variant sequences, but the longer k-mers become more sensitive to sequence analysis errors. Thus, the sensitivity is lowered and the database increases in size because more combinations are added to the database. In addition, long sequences are calculated for consensus, requiring greater storage capacity and computer power.In light of the characteristics of k-mer analysis, in a preferred embodiment of the method according to the present invention the lengths of the k-mers are shorter than the length of the identified transcript variant sequences in the RNA sequencing data of the biological sample or RNA-sequencing raw reads of reference samples, preferably the length of k-mer is 18 - 32 nt, more preferably 18 - 24 nt, even more preferably the length of k-mer is 19-23 nt, even more preferably the length of k-mer is 20 - 22 nt, even more preferably the length of k-mer is 21 nt.

[0049] In light of the characteristics of k-mer analysis, in a preferred embodiment of the method according to the present invention the fraction of k-mer sequences of the entire set of possible k-mers in the context sequence of the transcript variant sequence is an additional important parameter that needs to be optimized, also referred to as “k-mer rate”. For optimal accuracy of detecting a given variant context sequence in a k-mer index, the fraction of exactly matching k-mers should be not too low to allow to be specific (and avoid false positive hits) and not too large to be tolerant against minor sequence variations, such as germline single nucleotide polymorphisms (SNP) or sequencing errors. Preferably the fraction of matching k-mers, i.e. the k-mer rate, is higher than o, in particular in the range 0.1 - 1.0, more preferably the fraction of matching k-mers is in the range 0.3-0.8 or 0.4 - 0.9, even more preferably the fraction of matching k-mers is in the range 0.5 - 0.7, even more preferably the fraction of matching k-mers is in the range 0.6 - 0.8, and even more preferably the fraction of matching k-mers is 0.7.

[0050] The lower limit allows the number of combinations that enables a high number of transcript variant sequences to be discriminated while the upper limit allows for the maintenance of sensitivity in consideration of maximal storage capacity and computer power efficiency.

[0051] In a further preferred embodiment of the present invention, the length of the k-mers in the obtained k-mer RNA sub-sequences dataset for the biological sample and the reference k-mer RNA sub-sequences dataset of reference biological samples are the same.

[0052] In a further preferred embodiment of the present invention, the sample k-mer dataset comprises k-mer obtained by dividing the transcript variant sequences into short overlapping sub-sequences of length k (k-mers),

[0053] wherein the transcript variant sequences were identified on basis of aberrant transcript features, in particular on the basis of spliced junctions in the RNA sequencesThe term “spliced junctions” as used herein means junctions produced by splicing. Splicing is an essential step of RNA processing for multi-exon genes, in which introns are removed from a precursor RNA, thereby producing mature RNAs containing splice junctions. There are three types of RNA splice junctions (linear, back-splice, and fusion junctions) that are derived from RNA-sequencing data of non-cancerous and cancerous tissues. The method according to the invention integrates and characterizes all RNA splice junctions of both healthy or pathological human cells and tissues. The characterization of the spliced junctions enables new splicing-associated targets and provides insights into the identification and assessment of potential neoepitopes for cancer treatment.

[0054] The term “gene fusion” as used herein refers to fusion transcripts on the RNA level.

[0055] The term “neoORF” refers to a transcript variant containing an open reading frame generated by a genetic or transcriptomic alteration that encodes a novel protein sequence.

[0056] In the method according to the invention suitable and optimal parameters for tumorspecificity analysis were designed for a given query sequence. Short sequences (+ / -(k- i), here + / -20 nucleotides) around splice junctions, can be detected accurately in the compressed indexes, even when expressed only in a small subset of samples.

[0057] In a preferred embodiment of the method according to the invention the reference k-mer database of reference RNA sequences is built by:

[0058] Obtaining sequence information of reference RNA sequences from raw RNA-sequencing reads and

[0059] Dividing the read sequences into k-mers to obtain one or more overlapping k-mers, and

[0060] comparing a transcript variant based on the aberrant transcript feature

[0061] k-mer data, in particular based on the spliced junction k-mer data, to biological samples of healthy subjects and subjects suffering from cancer.

[0062] In a preferred embodiment of the method according to the invention, said comparing the sample k-mer dataset with a reference k-mer database is performed by selecting a k-mer being exactly identical in transcript variant information based on the aberrant transcript feature data, in particular based on the spliced junction in the RNA sequencing data, from the biological sample to the reference k-mer database,wherein an absence of the identified transcript variants from the sample k-mer dataset in the biological samples of healthy subjects from the reference k-mer database indicates a tumorspecific transcript variant, coding for a tumor-specific antigen candidate. The aberrant transcript features in preferred embodiments of the method of the invention comprise splice junctions, fusion transcripts, intron retentions, exitrons, neo open reading frames (neoORF), viral sequences, bacterial sequences, transposable elements and chimeric transcripts. These chimeric transcripts preferably combine any of the aforementioned aberrant transcript features with the canonical human transcripts. Viral sequences can comprise viral RNA but also transcript integrated viral DNA, as well as corresponding bacteria sequences. Possible chimeric transcripts also comprise combinations of these types with reference transcripts, e.g. when a viral sequence is integrated within the human transcript in a cancer, or if an inter-genic transposon is activated and leads to a combined exon-transposon junction.

[0063] For an improved identification of tumor-specific antigen candidates based on tumor-specific transcript variants, the inventors developed the python package kqneo.

[0064] This method permits to easily perform k-mer indexing methods with optimal parameters (Raptor and kmindex, k-mer size = 21, and k-mer fraction = 0.7) to RNA sequencing data from biological sample of subjects suffering from cancer to predict tumor-specificity by annotating novel transcript variants with expression rates in healthy tissue, kqneo accepts any query transcript sequences as input and searches in the pre-built reference k-mer database composed of the 18,961 unmatched samples from 52 healthy tissues by comparing transcript variant information. The results of the k-mer search are then formatted and annotated with the relevant metadata to obtain the expression profiles (sample rate) across healthy tissue groups.

[0065] In a preferred embodiment of the method according to the invention the reference k-mer database consists of a collection of 18,961 distinct RNA-sequencing samples from 52 different tissues and six different developmental states that represents a diverse set of tissues, which in parts complements and extends GTEx as a collection of normal samples, kqneo connects powerful k-mer indices with a large and carefully curated healthy tissue dataset, allowing the integration of rapid screening of antigen candidates into personalized target detection pipelines.

[0066] In a preferred embodiment of the method according to the invention the method is a computer-implemented method.In a second aspect, the invention pertains to a tumor-specific antigen candidate identified according to the method of to the invention, wherein the tumor-specific antigen candidate is recognized by an antigen recognizing construct, wherein the antigen recognizing construct is preferably selected from an antibody, or fragment thereof, or a T cell receptor (TCR), or fragment thereof.

[0067] The term “tumor-specific antigen candidate” as used herein means a protein or peptide that a tumor cell posesses, which may be recognized by a tumor-specific T-lymphocyte and / or may induce a cytotoxic T-lymphocyte. The tumor-specific antigen candidate can be a tumor associated antigen (TAA), a mutated gene (neoantigen), or any other form of non-canonical, novel, or aberrant transcript variant that might encode a tumor-specific protein. In addition, a tumor antigen peptide means a peptide generated when the tumor antigen is degraded inside a tumor cell. This peptide may be recognized by tumor-specific T-lymphocytes and / or induce cytotoxic T-lymphocytes by binding to an HLA molecule and being presented on the surface of the cell. The tumor antigen that is generated in a tumor cell is degraded in the cell into a peptide that consists of 8 to 11 amino acids, which is a so-called tumor antigen peptide and binds to a molecule of human leukocyte antigen (hereinafter abbreviated as HLA) that is a major histocompatibility complex (hereinafter abbreviated as MHC) to be presented on the surface of the tumor cell.

[0068] The cytotoxic T-lymphocyte recognizes the HLA molecule and tumor antigen peptide complex, releases a soluble factor or the like, for example, a cytokine such as interferon-y, and then lyses the tumor cell.

[0069] Here, to "recognize" means that a recognizing entity discerns a to-be-recognized target from others and may bind to the recognized target. In particular, in the present invention, recognition of a tumor cell or a tumor antigen peptide by a cytotoxic T-lymphocyte means that the cytotoxic T-lymphocyte binds to a tumor antigen peptide presented by an HLA molecule via a T-cell antigen receptor.

[0070] To "activate" means to further enhance or bring into action an entity or a state that has a given activity or effect. In particular, in the present invention, activation of a cytotoxic T-lymphocyte means that the cytotoxic T-lymphocyte generates, for example, interferon-y (hereinafter abbreviated as IFN-y) as a result of recognizing a tumor-specific antigen presented by an HLA molecule, or that the cytotoxic T-lymphocyte demonstrates a cytotoxic activity against the recognized target cell (also called target). To "induce" means to generate a given activity or effect in an entity or a state that substantially lacks the activity or the effect.In particular, in the present invention, to induce an antigen-specific cytotoxic T-lymphocyte means to cause a cytotoxic T-lymphocyte, that specifically recognizes a given antigen, to differentiate and / or proliferate in vitro or in vivo.

[0071] The term "antibody" includes, but is not limited to, genetically engineered or otherwise modified forms of immunoglobulins, such as intrabodies, chimeric antibodies, fully human antibodies, humanized antibodies (e.g. generated by "CDR-grafting"), antibody fragments, and heteroconjugate antibodies (e.g., bispecific antibodies, diabodies, triabodies, etra-bodies, etc.). The term "antibody" includes cys-diabodies and minibodies. Thus, each and every embodiment provided herein in regard to "antibodies", or "antibody like constructs" is also envisioned as, bi-specific antibodies, iabodies, scFv fragments, chimeric antibody receptor (CAR) constructs, diabody and / or minibody embodiments, unless explicitly denoted otherwise. The term "antibody" includes a polypeptide of the immunoglobulin family or a polypeptide comprising fragments of an immunoglobulin that is capable of non-covalently, reversibly, and in a specific manner binding a corresponding tumor-specific antigen candidate, preferably the tumor-specific antigen candidate of the invention, as disclosed herein. An exemplary antibody structural unit comprises a tetramer. In some embodiments, a full-length antibody can be composed of two identical pairs of polypeptide chains, each pair having one "light" and one "heavy" chain (connected through a disulfide bond). Antibody structure and isotypes are well known to the skilled artisan (for example from Janeway’s Immunobiology, 9th edition, 2016).

[0072] In a third aspect, the invention pertains to a nucleic acid sequence encoding the tumorspecific antigen candidate according to the invention.

[0073] In the third aspect the problem is solved by an isolated nucleic acid comprising a sequence coding for the tumor-specific antigen candidate as described herein. The term “encoding” or more simply “coding” refers to the ability of a nucleotide sequence to code for one or more amino acids. The term does not require a start or stop codon. An amino acid sequence can be encoded in any one of six different reading frames provided by a polynucleotide sequence and its complement. An amino acid sequence can be encoded by desoxyribonucleic acid (DNA), ribonucleic acid (RNA), or artificially synthesized polymers similar to DNA or RNA.

[0074] In a fourth aspect, the invention pertains to a vector comprising the nucleic acid sequence according to the invention, preferably an expression vector, more preferably comprising a promoter sequence operably linked to the nucleic acid sequence according to the invention.A “vector” maybe any agent that is able to deliver or maintain a nucleic acid in a host cell and includes, for example, but is not limited to, plasmids (e.g., DNA plasmids), naked nucleic acids, viral vectors, viruses, nucleic acids complexed with one or more polypeptide or other molecules, as well as nucleic acids immobilized onto solid phase particles. Vectors are described in detail below. A vector can be useful as an agent for delivering or maintaining an exogenous gene and / or protein in a host cell. A vector may be capable of transducing, transfecting, or transforming a cell, thereby causing the cell to replicate or express nucleic acids and / or proteins other than those native to the cell or in a manner not native to the cell. The target cell may be a cell maintained under cell culture conditions or in other in vivo embodiments, being part of a living organism. A vector may include materials to aid in achieving entry of a nucleic acid into the cell, such as a viral particle, liposome, protein coating, or the like. Any method of transferring a nucleic acid into the cell may be used; unless otherwise indicated, the term vector does not imply any particular method of delivering a nucleic acid into a cell or imply that any particular cell type is the subject of transduction. The present invention is not limited to any specific vector for delivery or maintenance of any nucleic acid of the invention, including, e.g., a nucleic acid encoding the fusion protein of the invention or a fragment thereof.

[0075] Preferably the vector of the invention is an expression vector. The term “expression vector” typically refers to a nucleic acid construct or sequence, generated recombinantly or synthetically, with a series of specific nucleic acid elements that permit transcription of a particular nucleic acid in a host cell. The expression vector typically includes a nucleic acid to be transcribed - the fusion protein of the invention - operably linked to a promoter. The term “expression” includes any step involved in the production of the polypeptide including, but not limited to, transcription, post-transcriptional modification, translation, post-translational modification, and / or secretion. A preferred vector of the invention is a plantspecific, bacterial, yeast, insect, vertebrate, preferably mammalian, or a viral vector, preferably retroviral and adeno-associated viral vector. Preferred vectors of the invention are suitable for use in gene therapy, preferably gene therapy based on transformation of autologous adult stem cells.

[0076] In a fifth aspect, the invention pertains to a host cell comprising the tumor-specific antigen candidate recognizing construct, preferably wherein the host cell is a lymphocyte, more preferably a T lymphocyte or T lymphocyte progenitor, most preferably a CD4 or CD8 positive T-cell.In addition, in the present invention, a cytotoxic T-lymphocyte inducing tumor-specific antigen candidate means a candidate that demonstrates an effect such as, for example, when CD8 positive T-cells that specifically recognize a given antigen is not present, or is present only at an extremely low ratio, is changed to a state where cytotoxic T-lymphocytes that recognize this antigen are present at an extremely high ratio.

[0077] For purposes of producing a recombinant TCR, polypeptide, or protein, the host cell is preferably a mammalian cell. Most preferably, the host cell is a human cell. While the host cell can be of any cell type, can originate from any type of tissue, and can be of any developmental stage, the host cell preferably is a peripheral blood leukocyte (PBL) or a peripheral blood mononuclear cell (PBMC). More preferably, the host cell is a T cell. The T cell can be any T cell, such as a cultured T cell, e.g., a primary T cell, or a T cell from a cultured T cell line, e.g., Jurkat, SupTi, etc., or a T cell obtained from a mammal, preferably a T cell or T cell precursor from a human patient. If obtained from a mammal, the T cell can be obtained from numerous sources, including but not limited to blood, bone marrow, lymph node, the thymus, or other tissues or fluids. T cells can also be enriched for or purified. Preferably, the T cell is a human T cell. More preferably, the T cell is a T cell isolated from a human. The T cell can be any type of T cell and can be of any developmental stage, including but not limited to, CD4-positive and / or CD8- positive, CD4-positive helper T cells, e.g., Thi and Th2 cells, CD8-positive T cells (e.g., cytotoxic T cells), tumor infiltrating cells (TILs), memory T cells, naive T cells, and the like. Preferably, the T cell is a CD8-positive T cell or a CD4-positive T cell.

[0078] In a sixth aspect, the invention pertains to a pharmaceutical composition comprising the tumor-specific antigen candidate identified according to the method of the invention, and a pharmaceutical acceptable carrier, stabilizer, and / or excipient.

[0079] In a seventh aspect, the invention pertains to a tumor-specific antigen candidate identified according to the method of the invention, and pharmaceutical composition according to the invention for use in medicine, preferably for use in the diagnosis, prevention, and / or treatment of a proliferative disease, in particular cancer.

[0080] With respect to the above mentioned medical uses of the tumor-specific antigen candidate and other materials derived therefrom, pertaining thereto or encoding the same, in accordance of the present disclosure, the to be treated and / or to be diagnosed diseases can be any proliferative disorder, preferably characterized by the expression of the tumor-specific antigen candidate or tumor-specific antigen candidate epitope sequence of the invention, forexample any cancer, including any of acute lymphocytic cancer, acute myeloid leukemia, alveolar rhabdomyosarcoma, bone cancer, brain cancer, breast cancer, cancer of the anus, anal canal, or anorectum, cancer of the eye, cancer of the intrahepatic bile duct, cancer of the joints, cancer of the neck, gallbladder, or pleura, cancer of the nose, nasal cavity, or middle ear, cancer of the oral cavity, cancer of the vagina, cancer of the vulva, chronic lymphocytic leukemia, chronic myeloid cancer, colon cancer, esophageal cancer, cervical cancer, gastrointestinal carcinoid tumor, glioma, Hodgkin lymphoma, hypopharynx cancer, kidney cancer, larynx cancer, liver cancer, lung cancer, malignant mesothelioma, melanoma, multiple myeloma, nasopharynx cancer, non-Hodgkin lymphoma, cancer of the oropharynx, ovarian cancer, cancer of the penis, pancreatic cancer, peritoneum, omentum, and mesentery cancer, pharynx cancer, prostate cancer, rectal cancer, renal cancer, skin cancer, small intestine cancer, soft tissue cancer, stomach cancer, testicular cancer, thyroid cancer, cancer of the uterus, ureter cancer, and urinary bladder cancer. A preferred cancer is cancer of the uterine cervix, oropharynx, anus, anal canal, anorectum, vagina, vulva, or penis. A particularly preferred cancer is a tumor-specific antigen candidate positive cancer, including preferably ovarian carcinoma, leukemia or melanoma.

[0081] With respect to the above mentioned medical uses of the tumor-specific antigen candidate and other materials derived therefrom, pertaining thereto or encoding the same, in accordance of the present disclosure, the to be treated and / or to be diagnosed diseases are in particular selected from the group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors and neuroendocrine tumors, and even more in particular consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

[0082] In a further aspect, the invention pertains a method for providing a cancer vaccine comprising the step of identifying at least one tumor-specific antigen candidate by the method of the invention.

[0083] In a further aspect, the invention pertains a method of inducing an immune response in a cancer patient comprising the steps of:

[0084] (a) providing a vaccine using the methods according to the invention; and

[0085] (b) administering said vaccine to the patient.In a further aspect, the invention pertains a cancer vaccine produced according to the method of the invention.

[0086] For instance, medicaments comprising the tumor-specific antigen candidate of the present invention and pharmaceutical compositions containing the tumor-specific antigen candidate of the present invention can be used as so-called cancer vaccines. The term cancer vaccine as used herein means a drug that selectively damages tumor cells by inducing and / or enhancing a specific immune response against the tumor cells. The dosage thereof can be determined with appropriate modifications according to the extent of recognition of peptides by the T-cells; in general, it is between o.oi and 100 mg / day / adult human, or preferably between o.i and 10 mg / day / adult human as an active principle.

[0087] This can be administered from once eveiy few days to eveiy few months. Administration can be carried out according to well-known methods for administrating tumor-specific antigen candidate for medical use, such as subcutaneously, intravenously, or intramuscularly. In order to induce and / or enhance the immune response during administration, the peptide of the present invention can be used, in the presence

[0088] or absence of an appropriate adjuvant, with or without linking to a carrier. The carrier is not limited as long as it exerts no harmful effect by itself onto the human body and is capable of enhancing antigenicity; cellulose, polymeric amino acids, albumin, and the like can be given as examples of carriers. Adjuvants can be those used in general for peptide vaccine inoculation, and a Freund incomplete adjuvant (FIA), aluminum adjuvant (ALUM), Bordetella pertussis vaccine, mineral oil, and the like can be given as examples. In addition, the formulation can be suitably selected by applying a suitable well-known method for formulating a peptide.

[0089] Otherwise, an effective cancer vaccine effect can be obtained also by collecting a fraction of mononuclear cells from the peripheral blood of a patient, incubating them with the peptide of the present invention, and then returning the fraction of mononuclear cells in which induction of the T-cells and / or activation of T-cells was observed, into the blood of the patient. Culture conditions, such as mononuclear cell concentration, concentration of the peptide of the present invention, culture time, and the like, can be determined by simply repeating experiments. A substance having a capability to enhance the growth of lymphocytes, such as interleukin-2, maybe added during culturing.In a further aspect, the present invention provides a vaccine which is obtainable using the methods according to the invention. A vaccine provided according to the invention induces an immune response in a patient an may comprise a pharmaceutically acceptable carrier and optionally one or more adjuvants, stabilizers etc. Thus, the terms "vaccination" and immunization" describe the process of treating a subject with the purpose of inducing an immune response for therapeutic or prophylactic reasons.

[0090] The vaccine may in the form of a therapeutic or prophylactic vaccine. A prophylactic administration of an immunotherapy, for example, a prophylactic administration of a vaccine of the invention, preferably protects the recipient from the development of a disease. A therapeutic administration of an immunotherapy, for example, a therapeutic administration of a vaccine of the invention, may lead to the inhibition of the progress / growth of the disease. This comprises the deceleration of the progress / growth of the disease, in particular a disruption of the progression of the disease, which preferably leads to elimination of the disease.

[0091] The agents, composition and treatments of cancer described herein can be use alone or combined with surgical resection and / or radiation and / or traditional chemotherapy and / or bone marrow transplantation (autologous, syngeneic, allogeneic or unrelated).

[0092] According to the invention, the term "vaccine" relates to or may comprise a pharmaceutical preparation or pharmaceutical composition or product that upon administration induces an immune response, in particular a cellular immune response, which recognizes and attacks a a diseased cell such as a cancer cell. In one embodiment, the vaccine may be a “personalized cancer vaccine" or "individualized cancer vaccine" for a particular cancer patient, which means that a cancer vaccine is adapted to the needs or special circumstances of an individual cancer patient.

[0093] In one embodiment, a vaccine provided according to the invention may comprise a peptide or polypeptide comprising one or more amino acid modifications or one or more modified peptides predicted as being immunogenic by the methods of the invention or a nucleic acid, preferably RNA, encoding said peptide or polypeptide.

[0094] The cancer vaccines provided according to the invention when administered to a patent provide one or more T cell epitopes suitable for stimulating, priming and / or expanding T cells specific for the patient's tumor. The T cells are preferably directed against cells expressing antigens from which the T cell epitopes are derived, thereby inducing orpromoting a cellular response, preferably cytotoxic T cell activity, against a cancer disease. Since a vaccine provided according to the present invention will target cancer specific mutations it will be specific for the patient's tumor.

[0095] The methods of the invention may comprise a step of determining the usability of the identified sequences or modifications for cancer vaccination. Such step can involve one or more of the following: (i) assessing whether the modifications are located in known or predicted MHC presented epitopes, (ii) in vitro and / or in silico testing whether the modifications are located in MHC presented epitopes, e.g. testing whether the modifications are part of peptide sequences which are processed into and / or presented as MHC presented epitopes, and (iii) in vitro testing whether the envisaged modified epitopes are able to stimulate T cells.

[0096] The sequences identified according to the invention and provided by a vaccine of the invention are preferably present in the form of a polypeptide or a nucleic acid, in particular RNA, encoding said polypeptide. In one embodiment, the vaccine sequences are spaced by linkers, in particular neutral linkers. The term "linker" according to the invention relates to a peptide added between two peptide domains such as epitopes or vaccine sequences to connect said peptide domains. There is no particular limitation regarding the linker sequence. However, it is preferred that the linker sequence reduces steric hindrance between the two peptide domains, is well translated, and supports or allows processing of the epitopes. Furthermore, the linker should have no or only few immunogenic sequence elements. For example, Hoyt et al. (EMBOJ. 25(8), 1720-9, 2006) and Zhang et al. (J. Biol. Chern., 279(10), 8635-41, 2004) have shown that glycine-rich sequences impair proteasomal processing and thus the use of glycine rich linker sequences act to minimize the number of linker-contained peptides that can be processed by the proteasome. Schlessinger et al. (Proteins, 61(1), 115-26, 2005) had found that amino acids glycine and serine included in an amino acid sequence result in a more flexible protein that is more efficiently translated and processed by the proteasome, enabling better access to the encoded neo-epitopes. Thus, the linker is preferably enriched in glycine and / or serine amino acids. In one embodiment, the linker comprises the sequence GGSGGGGSG.

[0097] In one preferred embodiment, a cancer vaccine according to the present invention is administered to a patient in the form of a nucleic acid, preferably RNA such as in vitro transcribed or synthetic RNA, which may be expressed in cells of a patient such as antigen presenting cells to produce the polypeptide. The present invention also envisions the administration of one or more multiepitopic polypeptides", preferably in the form of a nucleicacid, preferably RNA such as in vitro transcribed or synthetic RNA, which maybe expressed in cells of a patient such as antigen presenting cells to produce one or more polypeptides.

[0098] Within certain embodiments, immunotherapy may be active immunotherapy, in which treatment relies on the in vivo stimulation of the endogenous host immune system to react against diseased cells with the administration of immune response-modifying agents, such as polypeptides and nucleic acids as provided herein.

[0099] The agents such as vaccines and compositions described herein may be administered via any conventional route, including by injection or infusion. The administration may be carried out, for example, orally, intravenously, intraperitoneally, intramuscularly, subcutaneously or transdermally. In one embodiment, administration is carried out intranodally such as by injection into a lymph node. Other forms of administration envision the in vitro transfection of antigen presenting cells such as dendritic cells with nucleic acids described herein followed by administration of the antigen presenting cells.

[0100] In an eight aspect, the invention pertains to a method of manufacturing a tumor-specific antigen candidate specific antigen recognizing construct expressing cell line, comprising

[0101] (a) Providing a suitable host cell,

[0102] (b) Providing a genetic construct comprising a coding sequence encoding an antigen recognizing construct being specific to the tumor-specific antigen candidate identified by the method according to the invention.

[0103] (d) Introducing into said suitable host cell the genetic construct and

[0104] (e) Expressing the genetic constructs by said suitable host cell,

[0105] preferably further comprising

[0106] (f) Isolating and purifying the antigen recognizing construct from the suitable host cell and, optionally,

[0107] (g) Reconstituting the antigen recognizing construct in a T-cell.

[0108] In a ninth aspect, the invention pertains to a system for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular fromleast one subject, more in particular from at least two subjects, comprising at least one processing device configured to:

[0109] (a) Receiving in the at least one processing device the identified transcript variant sequences in the RNA sequencing data from a biological sample,

[0110] (b) Dividing the transcript variant sequences from step (a) into short overlapping subsequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample using the at least one processing device.

[0111] (c) Receiving in the at least one processing device a reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples,

[0112] (d) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (b) to the reference k-mer database of reference biological samples from step (c) using the at least one processing device,

[0113] (e) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (d) using the at least one processing device.

[0114] For example, the said system is designed as a computer system, preferably comprising one or more computer hardware components, such as a computer processor or data storage device, and optionally other hardware components.

[0115] In a further aspect, the invention pertains to a computer product, such as a software, or kit comprising computer software, or to an online portal offering a software-based service for identifying at least one tumor-specific antigen candidate in a biological sample obtained from a subject according to this invention.

[0116] The invention further relates to a database and / or data structure, generated using the method, system and / or computer product according to the invention.

[0117] In a further aspect, the invention pertains to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the invention.In a further aspect, the invention pertains to a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of the invention.

[0118] Any embodiment of the invention disclosed in the context of the method, system or computer product, is considered disclosed in combination with the other related aspects of the invention. For example, the features of the method apply to the system, and vice versa.

[0119] A further embodiment of the invention therefore relates to a system for comparing the sample k-mer dataset with a reference k-mer database is performed by selecting a k-mer being exactly identical in transcript variant information based on the aberrant transcript feature data, in particular based on spliced junction in the RNA sequencing data, from the biological sample to the reference k-mer database, wherein an absence of the identified transcript variants from the sample k-mer dataset in the biological samples of healthy subjects from the reference k-mer database indicates a tumor-specific transcript variant coding for a tumor-specific antigen candidate of the sample k-mer dataset. The aberrant transcript features in preferred embodiments of the system of the invention comprise splice junctions, fusion transcripts, intron retentions, exitrons, neo open reading frames, and / or transposable elements.

[0120] In one embodiment the method as described herein is characterised in that the information stored in the reference k-mer database is an electronically-stored representation of biological data, or corresponds to, or is an abstraction of, one or more biological entities. The present invention therefore relates to a computer implemented method in which data are processed that correspond to "real-world" biological entities. The database as such is not only characterized by the means of its production, but preferably also by the particular biological context of the entities and relationships embodied in the database.

[0121] In one embodiment, sample k-mer datasets are preferably stored in an additional data structure, database, and / or library.

[0122] In one embodiment the invention is characterized in that the method and / or system is functionally connected to the reference k-mer database according to the invention, wherein the outcome of the method and / or system of the invention enables a query to be sent from an end-user of said method and / or system to said reference k-mer database for any given one or more best-matched mismatched transcript variant sequences in the RNA sequencing data corresponding to said outcome.The computer implementation of the invention enables an efficient, fast and reliable method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from at least one subject, more in particular from at least two subjects. The data processed by the software can be handled in a completely or partially automatic manner, thereby enabling in a preferred embodiment an automated computer-implemented method. Data regarding transcript variant sequences in a RNA sequencing data can be stored electronically and eventually maintained in appropriate databases. The invention therefore also relates to computer software capable of carrying out the method as described herein.

[0123] It is to be understood that application of the teachings of the present invention to a specific problem or environment, and the inclusion of variations of the present invention or additional features thereto (such as further aspects and embodiments), will be within the capabilities of one having ordinary skill in the art in light of the teachings contained herein.

[0124] Unless context dictates otherwise, the descriptions and definitions of the features set out above are not limited to any particular aspect or embodiment of the invention and apply equally to all aspects and embodiments which are described.

[0125] All references, patents, and publications cited herein are hereby incorporated by reference in their entirety. In view of the above, it will be appreciated that the present invention also relates to the following itemised embodiments:

[0126] Item i. A method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from at least one subject, more in particular from at least two subjects, the method comprising:

[0127] (g) Providing RNA sequencing data from said biological sample,

[0128] (h) Identifying transcript variant sequences in the RNA sequencing data from step (a), (i) Dividing the transcript variant sequences from step (b) into short overlapping subsequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample,

[0129] (j) Providing reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples,(k) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (c) to the reference k-mer database of reference biological samples from step (d),

[0130] (l) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (e).

[0131] Item 2. The method according to item 1, wherein

[0132] step (f) further comprises predicting binding of one or more epitope peptides, comprised within the tumor-specific antigen candidates derived from the identified transcript variants, to one or more HLA molecules.

[0133] Item 3. The method according to item 1, wherein

[0134] step (f) further comprises selecting tumor-specific antigen candidates comprising epitope peptides predicted to bind to HLA molecules of more than one subject, in particular thereby identifying shared variants.

[0135] Item 4. The method of any of the preceding items, wherein

[0136] step (b) further comprises identifying transcript variant sequences by filtering of the provided RNA sequencing data.

[0137] Item 5. The method of any of the preceding items, wherein

[0138] step (f) further comprises filtering the identified tumor-specific transcript variant sequences.

[0139] Item 6. The method of any of the preceding items, wherein

[0140] in step (e) the identified transcript variant sequences are verified using an orthogonal verification method, in particular comprising RT-PCR, qRT-PCR and / or Sanger sequencing.

[0141] Item 7. The method of any of the preceding items, wherein

[0142] step (b) further comprises providing structural variant information to identify transcript variant sequences, in particular fusion transcripts.Item 8. The method of any of the preceding items,

[0143] wherein said subject is a human suffering from a proliferative disease, in particular from cancer, and said biological sample is a proliferative disease cell sample, in particular a cancer tumor cell sample selected from a group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors, neuroendocrine tumors, in particular selected from the group consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

[0144] Item 9. The method of any of the preceding items,

[0145] wherein the reference k-mer database of reference biological samples from step (d) is a collection of RNA sequence data from biological samples of healthy subjects and / or subjects suffering from cancer, and wherein

[0146] each k-mer RNA sequences dataset of biological samples was obtained by applying k- mer indexing to the plurality of raw RNA-sequencing reads.

[0147] Item 10. The method of any of the preceding items,

[0148] wherein the lengths of the k-mers are shorter than the length of the identified transcript variant sequences in the RNA sequencing data of the biological sample or RNA-sequencing raw reads of reference samples, preferably the length of k-mer is 18 - 32 nt, more preferably 18 - 24 nt, even more preferably the length of k-mer is 19-23 nt, even more preferably the length of k-mer is 20 - 22 nt, even more preferably the length of k-mer is 21 nt.

[0149] Item 11. The method of any of the preceding items, in particular according to item 10, characterized by a

[0150] k-mer rate higher than o, preferably in the range 0.1-1, more preferably in the range 0.4 - 0.9, even more preferably in the range 0.6 - 0.8, and even more preferably the k-mer rate is 0.7.

[0151] Item 12. The method of any of the preceding items,wherein the length of the k-mers in the obtained k-mer RNA sub-sequences dataset for the biological sample and the reference k-mer RNA sub-sequences dataset of reference biological samples are the same.

[0152] Item 13. The method of any of the preceding items,

[0153] wherein the sample k-mer dataset comprises k-mer obtained by dividing the transcript variant sequences into short overlapping sub-sequences of length k (k- mers),

[0154] wherein the transcript variant sequences were identified on basis of aberrant transcript features, in particular based on spliced junctions in the RNA sequences.

[0155] Item 14. The method of any of the preceding items,

[0156] wherein the reference k-mer database of reference RNA sequences is built by:

[0157] Obtaining sequence information of reference RNA sequences from raw RNA- sequencing reads and

[0158] Dividing the read sequences into k-mers to obtain one or more overlapping k-mers, and

[0159] comparing a transcript variant based on the aberrant transcript feature

[0160] spliced junction k-mer data, in particular based on spliced junction k-mer data, to biological samples of healthy subjects and subjects suffering from cancer.

[0161] Item 15. The method of any of the preceding items, wherein said comparing the sample k-mer dataset

[0162] with a reference k-mer database is performed by selecting a k-mer being exactly identical in transcript variant information based on the aberrant transcript feature data, in particular based on the spliced junction in the RNA sequencing data, from the biological sample to the reference k-mer database,

[0163] wherein an absence of the identified transcript variants from the sample k-mer dataset in the biological samples of healthy subjects from the reference k-mer database indicates a tumor-specific transcript variant, coding for a tumor-specific antigen candidate.

[0164] Item 16. The method of any of items 13-15, whereinthe aberrant transcript features comprise splice junctions, fusion transcripts, intron retentions, exitrons, neo open reading frames, viral sequences, bacterial sequences, transposable elements and chimeric transcripts.

[0165] Item 17. The method of any of the preceding items, wherein the method is a computer-implemented method.

[0166] Item 18. A tumor-specific antigen candidate identified according to the method of any of the items 1-16,

[0167] wherein the tumor-specific antigen candidate is recognized by an antigen recognizing construct, wherein the antigen recognizing construct is preferably selected from an antibody, or fragment thereof, or a T cell receptor (TCR), or fragment thereof.

[0168] Item 19. A pharmaceutical composition comprising the tumor-specific antigen candidate identified

[0169] according to the method of any of the items 1-17, and a pharmaceutical acceptable carrier, stabilizer, and / or excipient.

[0170] Item 20. A tumor-specific antigen candidate identified according to the method of any of the items 1-17,

[0171] and pharmaceutical composition according to item 19 for use in medicine, preferably for use in the diagnosis, prevention, and / or treatment of a proliferative disease.

[0172] Item 21. The tumor-specific antigen candidate according to item 20, wherein

[0173] the proliferative disease is a cancer.

[0174] Item 22. The tumor-specific antigen candidate according to item 21, wherein

[0175] the cancer is selected from a group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors, neuroendocrine tumors.

[0176] Item 23. The tumor-specific antigen candidate according to item 21, whereinthe cancer is selected from a group consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

[0177] Item 24. A system for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from least one subject, more in particular from at least two subjects, comprising at least one processing device configured to:

[0178] (a) Receiving in the at least one processing device the identified transcript variant sequences in the RNA sequencing data from a biological sample,

[0179] (b) Dividing the transcript variant sequences from step (a) into short overlapping sub-sequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample using the at least one processing device,

[0180] (c) Receiving in the at least one processing device reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples, (d) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (b) to the reference k-mer database of reference biological samples from step (c) using the at least one processing device, (e) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (d) using the at least one processing device.

[0181] Item 25. A computer program product comprising instructions which, when the program is executed by

[0182] a computer, cause the computer to carry out the method of items 1 to 17.

[0183] Item 26. A computer-readable medium comprising instructions which, when executed by a computer,

[0184] cause the computer to carry out the method of items 1 to 17.

[0185] The invention presents an approach that uses k-mer indexing of RNA sequencing data to efficiently evaluate the tumor specificity of transcript variants. By comparing candidatesequences against a broad collection of healthy tissue data, it determines whether they are uniquely associated with tumors and validates these findings with experimental methods such as RT-qPCR and long-read sequencing. The tool accurately identifies among others tumor-specific splice-junctions, isoforms and gene fusions while maintaining a low falsepositive rate and detecting recurrent variants across different cancer types. Overall, the invention enables scalable and precise identification of tumor-specific transcript features, expanding the range of potential cancer antigen targets, which could ideally be used as a personalized or individualized cancer vaccine, and serve as well as a resource that broadens the safe and actionable antigen repertoire for cancer immunotherapy.

[0186] Figure description

[0187] Preferred embodiments of the invention are illustrated by way of example with reference to the accompanying figures, which are provided in identical sets of both in black & white and in color (“color”):

[0188] Figure 1: K-mer indexing methods detect sample-specific splice junctions with high accuracy.

[0189] (A) Overview of benchmark and method optimization. The benchmark dataset consists of 3700 canonical exon-exon junctions. For each of 148 RNA-seq samples, the junctions were labeled positive if junction read support was detected with STAR and easyquant and negative if both methods find zero read support. Splice junction sequences were split into different expression classes according to number of samples. K-mer indexing methods were compared for their detection accuracy and computational resources in building indexes in searching in them. (B) ROC curves with sensitivity and specificity for benchmarked tools by expression classes for k=2i. Each dot depicts sensitivity and i-specificity of a tested 0 ranging from 0.45 to 1.0. (C) Detection accuracy of tools for multiple expression classes of splice junctions by different parameters for k-mer size (k) and k-mer ratio.

[0190] Figure 2: Benchmarking computational efficiency of indexing and search reveals scalability to thousands of RNA-seq samples.

[0191] (A) Wallclock time in hours to construct k-mer by data set size for different tools. (B) Column chart representation of (A). (C) Peak memory usage during index construction in Gb. (D) Column chart representation of (C). (E) Final index size in Gb on disk. (F) Column chart representation of (E). (G) Performance of Raptor, Reindeer, Kmindex and COBS querymodus by searching two independent sets of 10.000 canonical splice junctions in healthy tissue (ds) index for cold and warm (consecutive executions) cashing scenarios.

[0192] Figure 3: k4neo annotates expression breadth across a diverse collection of healthy tissues to analyze tumor-specificity.

[0193] (A) Number of collected samples per tissue type. The curated collection of sequencing experiments consists of 49 different healthy tissues and 31 tumor types. (B) Cumulative healthy tissue data representation. Healthy tissue (left) and tumor types from TCGA (right).

[0194] (C) Overview of functionality implemented in k4neo python package. k4neo uses a pre-built k-mer index of RNA-seq samples. Query transcript sequences are decomposed into overlapping k-mers, which are searched with a given k-mer ratio parameter in the pre-build indexes. The matches in k-mer indices are parsed and annotated with tissue, disease and developmental state information, kqneo computes expression profiles as sample rate (faction of samples for which the sequence was detected) per annotated metadata features (tissue, developmental state, disease, and cancer type).

[0195] Figure 4: kqneo captures tissue-specific expression patterns of known cancer genes and isoforms.

[0196] (A) kqneo detection performance on an orthogonal generated qPCR splice junction benchmark dataset. (B) Comparison of expression profiles of isoforms of the CLDN18 gene, as detected with kqneo and STAR in the same set of GTEx samples. (C) Correlation of k-mer coverage along full-length transcript (k-mer fraction) and gene expression strength in TPM. Each dot depicts the TPM and k mer fraction of a CTA in an individual melanoma sample.

[0197] (D) The sample rate per tissue of 20 known full-length cancer testis antigens in TCGA tumor cohorts and healthy tissues demonstrates that kqneo recovers the expected tissue and disease specific expression profiles. (E) The sample rate by kqneo (top) and in expression databases (bottom) of 8 full-length cancer testis antigens per TCGA tumor cohorts (left) and healthy tissues (right). Kqneo recovers tissue- and disease-specific expression profiles of CTAs.

[0198] Figure 5: kqneo reliably detects tumor-specificity of known and novel neoantigen targets with genomic support for somatic origin.

[0199] (A) The fraction of germline and somatic mutations in melanoma samples that were detected by kqneo in healthy tissues (more than 1% of healthy tissue samples) or not detected. (B) ROC curve and accuracy in tumor-specificity classification with kqneo for somatic SNVs as tumor-specific and germline SNPs as not tumor-specific, based on varying k-mer fraction.(C) Percent of gene fusions detected as tumor-specific or not tumor-specific by k4neo for gene fusions associated to structural variants from whole genome sequencing (left) or gene fusion transcripts only detected from RNA-seq (right). (D) Splice junctions per melanoma sample that are induced by somatic mutations or detected only from RNA seq. (E) Condensed representation of (D). (F) Number of RNA-seq derived SJ candidates per melanoma sample that are mutation associated or predicted with kqneo by two different tumor-specificity criteria. (G) kqneo sample rate in healthy tissues of three classes of splice junctions, RNA-only (top), RNA-only tumor-specific (middle), and mutation-associated (bottom).

[0200] Figure 6: kqneo predicts tumor-specific splice junctions in individual tumor samples that encode antigen candidates and the recurrence of shared tumor-specific splice junctions can be confirmed with kqneo in TCGA tumor samples.

[0201] (A) Cumulatively applied filtering steps in the NeoRasp splice junction pipeline including kqneo tumor-specificity filtering (y-axis) and number of resulting splice-junctions per sample (x-axis) (for 19 UVM samples). (B) Detected novel tumor-specific junctions in healthy (negative control) and tumor cohorts. (C) Fraction of alternative splicing event types (left), their impact on reading frame (in-frame or frame-shift, middle), and predicted MHC class I or class II binding epitopes (right). (D) Number of tumor-specific splice junctions shared by at least 5% of patients per healthy tissue cohort (negative control) or tumor cohort.

[0202] (E) Left panel: Recurrent detection of the top 10 most recurrent junctions per tumor type (left panel, with shared junctions in y-axis and tumor-samples on x-axis). The middle panel shows junction read support as horizontal boxplot. The right panel shows that these junctions are not detected in healthy tissues, except the immune privileged tissues, whereby the point size represent the sample rate. Candidate junctions often lead to frame-shifts (fs) and long novel peptides (number at the right side of the figure indicate novel peptide length in amino acids). Right panel: Top 10 most recurrent splice junctions with annotated peptide change for AML, SCLC and UVM cohorts (y-axis). Splice junction usage across tumor samples (x-axis) is shown in left heatmap, alongside junction read support across detecting samples (middle, horizontal boxplots), kqneo healthy tissue sample rate (right heatmap), frame shift annotation, novel peptide length, splice junction event type, and log-fold change of wild- type gene expression in tumors over highest healthy tissue. (F) Left panel: Confirmation of recurrence of the top 10 most recurrent tumor-specific junctions that encode novel peptides in TCGA tumor cohorts with kqneo. The size of the point indicate the sample rate per junction (y-axis) in each tumor type (x-axis) from o (smallest dot) to 1 (largest dot). Right panel:Confirmation of recurrence of the candidate junctions from (E) as k4neo sample rate in TCGA tumor cohorts (x-axis). (G) Antigen coverage as percent of patients in tumor cohorts with RNA-seq expression of candidate splice junctions and patient-HLA-specific MHC binding prediction.

[0203] Figure 7: The shared antigen landscape from tumor-specific splice junctions in SCLC, UVM and AML.

[0204] Shared splice junctions (y-axis) are shown for each expressing sample (x-axis) and indicated as MHC class I or II binder or non-binder. For each candidate junction it is indicated whether it could be confirmed in patient- or tumor-type-matched long-read sequencing data. (A) SCLC, (B) AML and (C) SCLC.

[0205] Figure 8: K-mer indexing methods benchmark data and results.

[0206] (A) The Number of junction sequences per sample expressing the junction in the junction benchmark ground truth dataset. (B) Distribution of number of reads supporting the junctions from benchmark data by the four groups of expression breath across samples. (C) Specificity and (D) Sensitivity by method for multiple expression classes of splice junctions (rows) and different parameters for k-mer size (columns) and k-mer ratio thresholds (x-axis).

[0207] Figure 9: Computational efficiency of k-mer methods by cohort size.

[0208] (A) Time to construct index by cohort size (x-axis, indicated as nucleotide content in Tbp at bottom and number of samples at top). (B) Peak memory usage in Gb during index construction by cohort size. (C) Final total file size of index in Gb by cohort size.

[0209] Figure 10: Full-length transcript detection sensitivity.

[0210] (A) Comparison CLDN18 sample rates of kqneo (SRA index), kallisto (SRA lung & stomach samples) and GTEx STAR counts. (B) Sensitivity of k-mer tools kmindex and Raptor to detect full-length CTA transcripts at varying TPM cutoffs (1 to 5) by k-mer ratio. (C) Gene expression of CTAs in melanoma samples as transcript per millions (TPM) by the k-mer fraction in the k-mer index of the same melanoma samples.

[0211] Figure 11: Accuracy of k-mer indexing methods for splice junctions.Sensitivity and specificity for benchmarked tools by expression classes (columns) and k-mer sizes (rows). Each dot depicts sensitivity and 1-specificity of a tested k-mer ratio ranging from 0.45 to 1.0 for all tested tools (indicated shape symbol).

[0212] Figure 12: Sample rate of germline and somatic point mutations by k4neo.

[0213] (A) Sample rate in healthy tissue index by k4neo (y-axis) by population variant frequency in GNOMAD dad (x-axis) for germline variants in melanoma samples. (B) Distribution of sample rate in healthy tissue index by k4neo for germline variants (top) and somatic SNVs (bottom) for melanoma samples.

[0214] Figure 13: Detection of tumor-specific splice variants.

[0215] (A) Overview of the NeoRasp pipeline to detect splice junctions from RNA-seq and uses k4neo for tumor-specificity detection. (B) Number of tumor-specific, shared and peptide altering splicing junctions (candidates) in normal and tumor cohorts. (C) Number of tumorspecific shared splice variants affecting the coding sequence for tumor cohort and negative control healthy tissue cohorts. (D) Recurrence rate (x-axis) of tumor-specific shared splice junction candidates (y-axis) in AML, SCLC and UVM cohorts. Vertical line depicts recurrence criterion of 5%. (E) Median number of tumor-specific splice junction per sample (y-axis) by varying tumor-specificity cutoff for detection rate in none-immune-privileged healthy tissue samples (x-axis) for negative control healthy tissues (left) and tumor cohorts (right).

[0216] Figure 14: Tumor-specific shared transcript variant in AGPAT5.

[0217] (A) Transcript structure of reference transcript (canonical, bottom) and novel shared multiexon skipping variant (top). The variant is recurrently expressed in AML (36%) and confirmed in unmatched ONT long read data. The variant is higher expressed (supported by more long-reads) than the canonical transcript.

[0218] Examples

[0219] Example 1: Detection accuracy of k-mer methods and parameters

[0220] As described in recent years, k-mer indexing methods can strongly compress the raw sequencing read data by storing only non-redundant short k-mer sequences per sample. To test their suitability and find optimal parameters for tumor-specificity analysis, it was assessed whether a given query sequence, such as the short sequences (+ / -X nucleotides)around splice junctions, can be detected accurately in the compressed indexes, even when expressed only in a small subset of samples.

[0221] In addition to the detection accuracy benchmark of individual k-mer indexing methods, the detection accuracy of k4neo was experimentally validated using a targeted qRT-PCR assay and Sanger sequencing for selected splice junctions in the breast cancer cell lines MCF7 and SKBR3 (Weber, D. et al. Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat. Biotechnol. 40, 1276-1284 (2022). This experimental confirmation resulted in 49 positives and 68 negative junction sequences, on which kqneo achieved a specificity of 99%, sensitivity of 94%, and an overall accuracy of 96% (see Fig. 4A, Figure 10 B).

[0222] Methods

[0223] Benchmarking data for detection performance of splice junctions

[0224] To compare the accuracy of k-mer indexing methods to detect individual splice-junctions, a ground-truth dataset of canonical splice junctions was created. Here, paired-end RNA-seq data from 150 samples of 48 different tissues (PRJNA764684) (Frankiw L et al., Alternative mRNA splicing in cancer immunotherapy. Nat Rev Immunol. 2019;19:675-87), were used.

[0225] Expressed canonical splice junctions were identified using a combination of read mapping against the hg38 reference genome using STAR (Barrett T et al., NCBI GEO: Archive for functional genomics data sets-update. Nucleic Acids Res. 2Oi3;4i:D99i-5) and confirmation by a targeted mapping against the splice junction sequences with easyquant (Lang F et al., Prediction of tumor-specific splicing from somatic mutations as a source of neoantigen candidates. bioRxiv 2023). More specifically, the splice junctions in the SJ.out.tab output of STAR (v2.7.10b) (with parameters — runMode alignReads — alignlntronMax 500000 — alignMatesGapMax 500000 — outSAMtype None) were used and only canonical exon-exon junctions reported in GEN CODE V40 annotations were kept.

[0226] Then, the transcriptomic context sequences in a window of + / - 100 bp around each splicejunction using splice2neo (vo.6.2) were extracted. As a next step, the easyquant (vo.4.1) with options -d 10 were used to re-quantify the number of supporting junctions reads (mismatch free alignment in + / - 10 bp around the junction position) for all samples and all junction sequences. A sample-SJ combination was classified as positive (expressed) if reported by STAR and detected by easyquant with at least two reads, as negative (not expressed) if the junction was not detected with STAR and easyquant.The resulting 3,705 splice-junctions were grouped by the number of samples expressing them into sample specific (1 sample), rare (2-5 sample), medium (6-19 samples), and common (20-150 samples). To balance dataset, the negative sample-junction pairs were down-sampled to the same number of positive samples-junction pairs per group, resulting in 76, 2875, 135, and 6732 positive and negative junction-sample pairs in the respective groups (Fig. 8A). For every considered k-mer size in the benchmark, the transcript context sequences were shortened to 2*(k-i) bases around the splice-junctions.

[0227] K-mer indexing methods

[0228] To determine the feasibility and accuracy of detecting novel RNA-derived transcript candidate sequences in k-mer indices, seven k-mer indexing methods were selected and evaluated. Five methods are based on approximate membership representations: Kmtricks (1.4.0) {Lemane T et al., kmtricks: Efficient and flexible construction of Bloom filters for large sequencing data collections. Bioinforma Adv. 2O22;2:vbacO2g), HowDeSBT (2.00.02) {Harris RS et al., Improved representation of sequence bloom trees. Bioinforma OxfEngl. 2020; 36:721-7), COBS (0.2.1) {Bingmann T et al., COBS: A Compact Bit-Sliced Signature Index. ArXiv Databases. 2019), Raptor (3.0.1) {Mehringer S et al., Hierarchical Interleaved Bloom Filter: enabling ultrqfast, approximate sequence queries. Genome Biol.

[0229] 2O23;24:131 and Kmindex (0.5.2) {Lemane T et al., kmtricks: Efficient and flexible construction of Bloom filters for large sequencing data collections. Bioinforma Adv.

[0230] 2O22;2:vbacO2g and two represent exact compact De Bruijn graph-based (cDBG) methods: MetaGraph (0.3.6) {Karasikov M et al., MetaGraph: Indexing and Analysing Nucleotide Archives at Petabase-scale. bioRxiv. 2020; 2020.10.01.322164) and REINDEER (1.0.2) Marchet C et al., REINDEER: Efficient indexing of k-mer presence and abundance in sequencing datasets. Bioinforma OxfEngl. 2020; 360177-85). For Raptor, three execution modes were compared: indexing k-mers with the Interleaved Bloom Filter (IBF) data structure (Raptor_kk), indexing minimizers using either IBF (Raptor_wk), or minimizers with the Hierarchical Interleaved Bloom Filter (HIBF) data structure (Raptor_hibf_wk) (see e.g. Mehringer S et al., Hierarchical Interleaved Bloom Filter: enabling ultrqfast, approximate sequence queries. Genome Biol. 2023124:131). For all tools, indices for k-mers of size 12, 21, 25, and 31 bp were constructed. In addition, k-mers that occur only once per sample were removed, as these are likely of erroneous origin. For the Bloom filter (BF) parameters in Kmtricks, HowDeSBT and kmindex, ntcard {Mohamadi H, et al., ntCard: A streaming algorithm for cardinality estimation in genomics data. Bioinforma Oxf Engl.

[0231] 2017;33:1324-30) was employed to estimate the number of distinct k-mers in each cohort.For HowDeSBT and kmtricks, the HowDeSBT manual was followed and the size of the bloom filter set to the estimated number of unique k-mers in the union of the experiments. For Reindeer and COBS,cDBG graphs as input was used. cDBG for each sequencing sample were constructed with BCALM2 (Chikhi R et al., Compacting de Bruijn graphs from sequencing data quickly and in low memory. Bioinforma Oxf Engl. 2Oi6;32:i2Oi-8 while excluding unique k-mers.

[0232] For MetaGraph, k-mer counters were constructed using KMC3 (Kokot M et al., KMC 3: Counting and manipulating k-mer statistics. Bioinforma Oxf Engl. 2017;33:2759-61) and subsequently the succinct de Bruijn graph representation constructed. For Raptor, the window size w for indexing with minimisers was set to k+4 as suggested in the Raptor manual. For all tools, indices were built on a compute server with 96 CPU cores and 1.5 Tb of RAM. Pre-processing steps of individual tools were parallelized using GNU parallel (Tange, O. GNU Parallel - The Command- Line Power Tool. Login USENIX Mag. 36, 42-47 (2011) where applicable to reflect typical real-world applications of the implementations.

[0233] Calculating detection accuracy

[0234] The splice junction sequences from the ground-truth dataset in the indices were queried by using the default commands from the tools and by considering a junction as detected per sample if a minimal fraction of k-mers from the query sequence (k-mer ratio) was found in the index for the sample. For all methods, k-mer sizes, and k-mer detection ratios (0.45, 0.5, 0.6, 0.7, 0.8, 0.9, and 1), the detected sample-junction pairs were compared with the ground truth data and sensitivity (recall), precision (PPV), Fi-score and Accuracy calculated as detection performance.

[0235] Healthy tissue cohorts and index building

[0236] To build a healthy tissue reference dataset, studies with accessible RNA-seq data were collected from different samples of non-disease tissues (Table 1). The raw RNA-seq data were downloaded from SRA, GEO (Barrett T et al. NCBI GEO: Archive for functional genomics data sets-update. Nucleic Acids Res. 2013; 41:0991-5) and ENCODE Luo Y et al. New developments on the Encyclopedia ofDNA Elements (ENCODE) data portal. Nucleic Acids Res. 2O2O;48:D882~ . Then, fastp (vo.23.4) (Chen S. Ultrafast one-pass FASTQ data preprocessing, quality control, and deduplication using fastp. iMeta. 2023 2 (with parameters —stdout — unpairedi — unpaired2 — length_required 31) was applied to remove low quality bases and adapter sequences. For paired-end sequencing data, interleavedFASTQ files were generated per sample, that includes forward and reverse reads that pass QC filtering. Reads with length less than largest k-mer size used (31 bp) were discarded.

[0237] Table 1: Healthy tissue cohorts with publicly accessible RNA-seq used in k-mer indices.

[0238] Study ID Samples Tissues Library layout Read length

[0239] E-MTAB-2836 200 31 paired 101

[0240] E-MTAB-6814 313 7 single 101

[0241] E-MTAB-513 16 16 paired / single 50575

[0242] PRJNA764684 150 48 paired 51

[0243] ENCODE-ENTEX 91 28 paired 101

[0244] PRJEB8360 393 1 paired 101

[0245] PRJNA377745 348 1 paired 101

[0246] ENCODE-Phase4 118 23 single 101

[0247] PRJNA174224 3 1 paired 101

[0248] PRJNA526249 155 1 paired 125

[0249] PRJNA576920 90 45 paired 101

[0250] PRJNA271929 49 1 paired 101

[0251] PRJNA398141 15 1 single 100

[0252] PRJNA573087 37 3 paired 151

[0253] Computational resource benchmarks

[0254] To assess the computational efficiency indexes of subsets of varying sizes (Table 2) were built from the cohorts (Table 1) on a single cluster node using 16 CPU cores for each job and a k-mer size of 21. Then, the run time, CPU usage, peak memory usage and file size of the final index was recorded. Also, only unique k-mers were excluded and then subsequently used in follow-up benchmarks.

[0255] As a next step, two independent sets of query sequences were built by randomly selecting two times 10,000 canonical splice junctions from the GENCODE V40 reference transcriptome and annotating them with + / - 50 bp of flanking transcript sequence. Next, the required query time and peak memory usage for querying the complete healthy tissue indexes from each tool. The two sequence sets were successively queried to reflect cold and warm cashing scenarios.

[0256] Table 2: Overview of the Datasets used to assess the computational efficiency indexes of subsets of varying sizes.Dataset Dataset_composition Nucleotid Number_of_Sampl e content es

[0257] (Tbp)

[0258] di E-MTAB-2836 0,69 122

[0259] d2 PRJNA764684 1,53 148

[0260] d3 PRJNA764684+EMTAB6814 2,61 461

[0261] d4 PRJNA764684 4,21 551

[0262] +EMTAB6814

[0263] +PRJNA576920

[0264] ds Innotop_normal 13,51 1663

[0265] +EMTAB6814+PRJNA576920

[0266] + Remaining studies from table 1

[0267] Results

[0268] Detection accuracy of k-mer methods and parameters

[0269] A benchmark data set was built from RNA-seq of 148 healthy samples from 48 tissues (Weber D et al. Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat Biotechnol. 2022;40:1276-84) consisting of 3,705 unique canonical splice junctions.

[0270] Using a stringent alignment-based re-quantification approach across all samples, 271,275 sample-junction pairs were defined as expressed (positive) and 274,203 as not expressed (negative) (Figure 1A). Next, junctions were grouped by the number of samples in which they were expressed, which was correlated with expression strength (number of junction reads).

[0271] Using this benchmark data, seven k-mer indexing methods were systematically benchmarked with varying parameters for k-mer size and required rate of detected k-mers from the query sequence (k-mer ratio) for their accuracy to detect the presence or absence of expressed or not expressed splice junction sequences. First, sensitivity and specificity were compared between all methods with a fixed k-mer size of k=2i (Figure 1B). For commonly expressed junctions, the sensitivity and specificity was nearly 100% for all k-mer indexing methods and k-mer ratios. For the medium and rarely expressed junctions, slightly reduced sensitivity was observed of down to 90% only for k-mer ratios higher than 0.7 Only for themethod REINDEER, specificity dropped to 87% when using k-mer ratios lower then 0.6, while all other methods showed up to 100% specificity (Figure 8C).

[0272] Next, the more challenging junctions that were expressed in only a single sample (samplespecific) were analyzed and the detection accuracy compared for varying k-mer sizes and k-mer ratios (Figure 1C & Figure 11 & Figure 8 C). Here, all methods except for REINDEER had no false positive hits and therefore 100% specificity (Figure 11). However, high values of k-mer size and k-mer detection ratio led to reduced sensitivity as low as 50% for some methods. MetaGraph performed consistently poorer than the other methods for sample-specific junctions across all parameter settings (Figure 11), with a maximum accuracy of 0.77 for k=i9 (Figure 1C). The k-mer size did mainly affect the performance for larger k-mer ratios, where the detection performance for larger k-mer sizes (25 and 31) dropped with increasing thresholds on k-mer ratio (Figure 1C & Figure 11). However, except for MetaGraph, all methods showed for their best performing parameters a very high detection accuracy between 0.98 and 1.0 (Figure 1B).

[0273] In summary, all tested tools showed high sensitivity and specificity in identifying high and broadly expressed splice junction sequences. For specific parameters, all methods, except MetaGraph, could correctly identify the more challenging splice junctions that are expressed in only one out of 148 samples. Overall, the accuracy benchmark demonstrated sufficiently high accuracy to detect splice junction sequences in indexed raw RNA-seq data of healthy tissue samples.

[0274] Computational efficiency of indexing and query modes

[0275] After establishing the detection accuracy, the technical feasibility and computational efficiency were assessed in practice to construct k-mer indices for thousands of unmatched healthy tissue samples. Moreover, the performance of the query mode was investigated to determine the feasibility of querying thousands of neoantigen candidate sequences (e.g. RNA-derived non-canonical splice junctions) for each patient in a personalized cancer vaccine pipeline.

[0276] Frist, the computational resource consumption of the methods was systematically evaluated and their scalability to larger sets of samples assessed. More specifically, multiple batches of short-read RNA-seq experiments ranging from 122 samples with 0.7 terra bases pairs (Tbp) to 1663 samples with 13.51 Tbp were indexed (Table 2). Wallclock time was recorded to construct the index (including all pre-processing steps), peak memory usage in gigabyte (Gb)and final index size on disk in Gb as benchmark categories (Figure 2A-F). For the index construction time it was observed, that the Sequence Bloom Tree (SBT) based methods HowDeSBT and Kmtricks showed an exponential growth with increasing number of samples, already exhibiting the longest run time at batch dq with 104 h and 70 h, respectively. Conversely, for the remaining methods the required time grew linearly when indexing larger batches of data. Interestingly, Kmindex was the fastest method for batches di-dq, but increased in runtime clearly when applied to the largest batch, requiring 116 h for indexing. Raptor was the fastest method and was able to index the largest batch d5 in less than 33 h.

[0277] Next, the peak memory usage was analyzed during index construction (Figure 2C, D). With increasing data size, the Raptor_kk mode and MetaGraph exhibited an exponential increasing demand in memory, by requiring for batch dq 547 Gb and 457 Gb, respectively (Figure 9 A). For this batch, all other methods showed a RAM usage of about 200 Gb. When indexing batch ds, the Raptor_wk mode showed the highest RAM usage with 625 Gb, followed by REINDEER, COBS and Raptor_hibf_wk mode requiring 1.5-2.7X less memory. Kmindex outperformed all other methods on this criterion requiring around 20 Gb of RAM.

[0278] When analyzing the required disk space for the final index (Figure 2E, F), the indices created for small collections of samples were similar in size. However, with growing number of samples, Kmindex showed prohibitive disk usage by producing indices that were 10-25X larger than the indices of the other methods (Figure 9B, C). Besides Kmindex, the produced indices of the other methods grew at much lower rates, with COBS generating the largest index (392 Gb), slightly larger than Reindeer (380 Gb), while the Raptor modes (Raptor_wk, Raptor_hibf_wk) produce the smallest indices with 149 and 141 Gb, respectively.

[0279] Then, the performance of the query mode was investigated to determine the practical feasibility of querying thousands of antigen candidate sequences for each patient.

[0280] Here, the screening performance of Raptor, COBS, REINDEER and Kmindex was assessed. Here, two query sets of 10,000 canonical SJ sequences were created from GENCODE V40 and queried the healthy tissue index (ds).

[0281] Query time and peak memory consumption were recorded under different caching scenarios to reflect different end-user perspectives. As a cold cache scenario, queries were defined against the index on first use, while the warm cache scenario reflects successive queries with a different set of sequences. For instance, in a clinical setting where a batch of patient neoantigen candidates are queried, the cold cache scenario would be the first patient, and thewarm cache scenario would apply to subsequent patient sequences. Raptor minimiser mode (Raptor_wk and Raptor_hibf_wk) indices were the fastest methods in the cold cache scenario, requiring only 24 and 14 minutes respectively to query the 1,663 samples (Figure 2G). Kmindex and COBS exhibited in this scenario the longest query times, followed by REINDEER. With a warmed-up cache the matrix-like approaches reduced query time by 30 (16 %) and 25 (8 %) minutes, respectively. Notably, compared to the other methods, REINDEER could not benefit from a warm cache, exhibiting in both scenarios a query time of around 1.75 h.

[0282] Raptor also profited from caching by reducing query time around 16 % down to around four minutes. Compared to Raptor and REINDEER, the matrix-like approaches showed the lowest RAM profiles requiring on average 3 Gb and 65 Gb of RAM, respectively. REINDEER showed in this benchmark the highest RAM usage, as the whole index was loaded into memory, requiring over 600 Gb of RAM.

[0283] The results indicate that Raptor with parameters k =21, w =25, and a k-mer rate of 70% is sufficiently accurate and fast to query multiple antigen candidate sequences against thousands of unmatched healthy samples.

[0284] Example 2: k4neo detects expression in a diverse collection of healthy tissues

[0285] In addition, the tool “kqneo” was developed and implemented to annotate any input transcript sequence, e.g. novel splice junctions, with expression profiles in healthy tissue samples.

[0286] Methods

[0287] kqneo

[0288] kqneo is implemented as python package and relies on the k-mer indexing methods Raptor and kmindex with k-mer size 21, which showed best detection accuracy and computational efficiency in the benchmarks, kqneo comes with pre-built k-mer indices for 1,663 RNA-seq samples of 49 healthy tissues from public resources (Table 1). Curated metadata per sample, such as tissue, developmental state, disease and optionally sex and age is stored in a document-oriented TinyDB database.

[0289] In the search mode, k4neo accepts transcript sequences as input and optionally a position of interest (e.g. splice junction or fusion breakpoint). These queiy sequences are then decomposed into overlapping k-mers, which are searched with a given k-mer ratio parameterin the pre-build indexes by Raptor or kmindex. The matches in k-mer indices are parsed as pairs of input sequence and individual RNA-seq sample and annotated with tissue, disease and developmental state information. Finally, for each input sequence, kqneo computes expression profiles as sample rate (faction of samples for which the sequence was detected) per annotated metadata features (tissue, developmental state and disease).

[0290] In addition, kqneo can be used in build mode to construct k-mer indices using Raptor and kmindex for additional samples from RNA-seq raw data.

[0291] Public study cohort and datasets

[0292] 14 sequencing studies were collected from different databases comprising 1,663 unmatched healthy samples and 9.7 Tb of raw RNA-seq data (Figure 3A, B), referred to as dataset ds. Notably, ds contains 17 tissues that are currently absent from GTEx V9 and expand GTEx tissues with low sample coverage, such as 26 additional fallopian tube (GTEx n=8), 13 additional cervix (GTEx n=i9) and eight additional bladder samples (GTEx n=2i) (Figure 3A, B). Here, additional metadata from the sequence read archive (SRA) were extracted and manually curated when available including sex (n=747 samples), ethnicity (n=i6 samples) and age (n=s8y samples). Then, tissue identifiers were harmonized to match definitions from GTEx and sub-tissue definitions were added if possible, resulting in 49 tissues of which 32 matched to GTEx.

[0293] Indexing of healthy tissue samples from GTEx and tumor samples from TCGA

[0294] To index healthy tissue samples and pan-cancer tumor samples, healthy tissues samples were selected from GTEx V9 and tumor samples from TCGA. From GTEx samples cell lines and cultured fibroblasts were excluded totaling to 16,704 samples while for TCGA 10,328 samples were indexed (Figure 3A, B). Then, a Workflow Description Language (WDL) workflow was developed to generate k-mer indices from aligned RNA-seq BAM files using Raptor in TERRA bioinformatics cloud (app.terra.bio). Raptor indices were constructed using the HIBF layout and parameters ( =21, w =25).

[0295] Results

[0296] To easily use this data and the best performing k-mer indexing methods with optimal parameters (Raptor and kmindex) to predict tumor-specificity by annotating novel transcript variants with expression rates in healthy tissue, the python package kqneo was developed (Figure 3C). kqneo accepts any query transcript sequences as input and searches in pre-builtk-mer indexes composed of the 18,961 unmatched samples from 52 healthy tissues. The results of the k-mer search are then formatted and annotated with the relevant metadata to obtain the expression profiles (sample rate) across healthy tissue groups.

[0297] In summary, the dataset consists of a collection of 18,961 distinct RNA-seq samples from 52 different tissues and six different developmental states that represents a diverse set of tissues, which in parts complements and extends GTEx as a collection of normal samples, kqneo connects powerful k-mer indices with a large and carefully curated healthy tissue dataset, allowing the integration of rapid screening of antigen candidates into personalized target detection pipelines.

[0298] Example 3: Claudin 18 isoform analysis via k4neo

[0299] The detection accuracy benchmark showed the high sensitivity and specificity of k-mer methods to detect junctions from the benchmarking dataset. Consequently, it was then assessed if the k-mer approach implemented in kqneo is able to screen for known tissuespecific expression patterns of

[0300] biological and clinically relevant transcript examples. Moreover, as shared antigen candidates show a heterogeneous expression in cancer, e.g. they are not expressed in all primary or metastatic sites and are restrictively expressed in healthy germ cells and somatic tissues, the aim was to investigate whether a k-mer index composed of cancer samples together with kqneo can capture these distinct expression profiles.

[0301] First, the expression of isoforms of CLDN18, a cell surface protein associated with the formation of tight junctions and an antibody target in gastric cancer and other solid tumors, was analyzed. CLDN18 has two protein-coding isoforms with known restricted expression to lung and stomach.

[0302] Methods

[0303] To evaluate the ability of kqneo to retrieve known tissue-specific expression patterns of individual transcript variants (isoforms), the isoform 1 and 2 of the gene Claudin 18 (CLDN18) were selected, which differ in their first exons and are differentially expressed in lung and stomach samples. Each transcript variant was represented by + / - 20 bp sequence around the respective first splice junction (Table 3) and queried against the healthy tissue index using kqneo. For comparison, raw splice-junction counts were downloaded from theGenotype-Tissue Expression (GTEx) portal (bulk-gex_v8_rnaseq_GTEx_Analysis_2Oi7-o6 O5_v8_STARv2.5.3a Junctions. get).

[0304] Table 3: Overview of used Claudini8 transcripts.

[0305] Transcript Ensembl transcript id representative junction hg 38 position

[0306]

[0307] Claudini8.i ENST00000183605 0^3:138010446-138023657:+ 393 Claudini8.2 ENST00000343735 01113:137999089-138023657:+ 354

[0308] Results

[0309] The first splice junction of each isoform was selected as a representative sequence and queried these using kqneo against the healthy tissue index, kqneo found the representative junction of the CLDN18 transcripts primarily in their respective tissues (Figure 4B). CLDN18.1 was found in 96.55% of indexed lung samples and 60% of the stomach samples. For CLDN18.2, kqneo detected the representative junction primarily in stomach samples (93-33 %)• Notably, the observed expression profile was consistent with expression quantified by mapped reads in GTEx samples (Figure 4B).

[0310] Example 4: Cancer testis antigen analysis

[0311] Next, the expression of full-length genes was assessed and cancer-testis antigens (CTAs) analyzed. CTAs are a class of cancer immunotherapy target genes that are recurrently expressed in tumors but in healthy tissues restricted to immune privileged reproductive organs.

[0312] Method

[0313] To assess the ability of k4neo to detect the disease-specific expression patterns of full-length shared antigen candidates, the representative transcripts (MANE or ENSEMBL canonical) of 8 known and clinically relevant cancer testis antigens (Table 4) were queried against healthy tissue and cancer k-mer indices. Besides the healthy tissue index, melanoma patients with RNA-seq samples from Hugo W, et al. Genomic and Transcriptomic Features of Response toAnti-PD-i Therapy in Metastatic Melanoma. Cell. 2016;165:35-44 (n=27, SRA SRP070710) and Riaz N et al. Tumor and Microenvironment Evolution during Immunotherapy with Nivolumab. Cell.2Oiy;iy .934-949.ei6 (n=46, SRA SRP094781) with kqneo were analyzed as described above. For comparisons, gene expression was calculated in transcript per million (TPM) with kallisto (vo.46.0) for the melanoma samples.To assess the ability of k4neo to detect splice junction sequences in RNA-seq using parameters optimized on the in-silico ground-truth dataset, an orthogonal benchmark based on qRT-PCR validation was generated. RNA-seq data from the MCF7 and SKBR3 cell lines, generated by the applicant (Weber, D. et al. Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat. Biotechnol. 40, 1276-1284 (2022), were used for this analysis. 50 expressed canonical splice junctions (25 per cell line) were selected as positives following the same criteria used for the in-silico benchmark. The negative set comprised 25 junctions from non-expressed genes (kallisto TPM = o) and 25 randomly generated exon-exon fusion junctions without any read support. Primer design and qRT-PCR experiments were performed as described in Lang, F. et al. Prediction of tumorspecific splicing from somatic mutations as a source of neoantigen candidates. Bioinforma. Adv. 4, vbaeo8o (2024). Amplicons were evaluated using capillary gel electrophoresis (QIAGEN’s QIAxcel Advanced), with ambiguous results resolved by Sanger sequencing. Based on qRT-PCR and Sanger confirmation, junction-cell line pairs were classified as positive or negative, resulting in 49 positives and 68 negatives. These sequences were queried using kqneo against the corresponding cell line RNA-seq k-mer index, and detection performance was assessed in terms of sensitivity and specificity. Based on the above, a qRT-PCR benchmark dataset could be established.

[0314] Table 4: Overview of representative transcripts (MANE or ENSEMBL canonical) of 8 known and clinically relevant cancer testis antigens

[0315] HGNC Ensembl gene ID Ensembl Query

[0316] symbol transcript ID length

[0317] (bp)

[0318] ACTL8 ENSG00000117148 ENST00000375406 1841

[0319] IGF2BP1 ENSGOOOOOI59217 ENST00000290341 8796

[0320] MAGEAl ENSGOOOOO198681 ENSTOOOOO356661 1711

[0321] MAGI. A3 ENSGOOOOO221867 ENST00000370278 1682

[0322] MAGEA4 ENSGOOOOO147381 ENSTOOOOO276344 1724

[0323] MAGEC1 ENSGOOOOOI55495 ENSTOOOOO285879 4256

[0324] MAGEC2 ENSGOOOOOO46774 ENSTOOOOO247452 1994PLAC1 ENSG00000170965 ENSTOOOOO359237 1119

[0325] PRAME ENSGOOOOO185686 ENST00000405655 2304

[0326] Results

[0327] The fraction of k-mers along the full-length transcript sequence that were detected by k4neo was highly correlated with gene expression defined by TPM in melanoma samples (Figure 10C), allowing to accurately discriminate between expressed and not expressed genes. For instance, using a k-mer fraction of 70%, k4neo could detect expressed genes (defined by >=2 TPM) with a sensitivity of 97.4% (Figure 10B).

[0328] Next, the sample-rate of the cancer testis antigens was compared between TCGA cancer cohorts and healthy samples with k4neo (Figure 4D, E). The expected expression profiles of CTAs with prominent targets were recovered, such as PRAME and MAGEA3, and expressed in most samples of the skin melanoma (SKCM) cohort of TCGA, but in the healthy sample set restricted to immune-privileged tissues, such as testis or ovary.

[0329] As demonstrated in Fig. 4E it was assessed whether k4neo captures the distinct expression profiles of CTAs across all tumor types and healthy tissues in the k4neo index. Compared to more standard mapping-derived transcript expression profiles of the antigens, k4neo reproduced the expected expression profiles (Pearson correlation (R2= 0.87, p < 3.2 x w286), characterized by broad expression in specific tumor types, high expression in testis tissue, and rare expression in healthy tissues.

[0330] Overall, the results demonstrate that k4neo detects disease-specific expression profile of specific isoforms and clinically relevant full-length genes with high accuracy.

[0331] Example 5: Germline and somatic mutations in melanoma

[0332] Then, detection rates of k4neo in the healthy tissue indices for somatic and germline mutations were assessed.

[0333] Method

[0334] To compare the detection rates of k4neo in the healthy tissue indices for somatic and germline mutations, whole-exome sequencing (WES) and RNA-seq data of n=73 melanoma samples (see Hugo W etal., 2016 and RiazN et al., 2017) were analyzed. For somatic variants WES of tumor and normal was aligned against hgi9 reference genome using bwa (vo.7.17)Li H et al., Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinforma Oxf Engl. 2009;25:1754-60) and subsequently somatic variants called using Mutect2 in the tronflow pipelines (Riesgo-Ferreiro P et al., TRON- Bioinformatics / tronflow-mutect2: Release vi.8.0, Zenodo; 2022). Germline SNPs were detected by applying the tronflow-haplotype-caller pipeline (vi.0.1) (see Riesgo-Ferreiro P. TRON-Bioinformatics / tronflow-haplotype-caller: vi.0.1, Zenodo; 2022) which is based on GATK (V4.2.6.1) on aligned normal sequencing data, restricting detection to exonic target regions and applying filtering with variant quality score recalibration using dbSNP (Sherry ST. dbSNP: the NCBI database of genetic variation. Nucleic Acids Res. 2001;29:308-11), the 1000 Genomes Project (The 1000 Genomes Project Consortium, Corresponding authors, Auton A, Abecasis GR, Steering committee, Altshuler DM, et al. A global reference for human genetic variation. Nature. 2015;526:68-74) and HapMap (The International HapMap Consortium. The International HapMap Project. Nature. 2003;426:789-96) sets of high quality known variant sites. Germline variants were then annotated with gnomAD (Exome Aggregation Consortium, Lek M, Karczewski KJ, Minikel EV, Samocha KE, Banks E, et al. Analysis of protein-coding genetic variation in 60,706 humans. Nature.

[0335] 2016;536:285-91) population allele-frequency and filtered for a variant allele frequency (VAF) > 0.05. Next, germline and somatic variant calls were inserted into coding transcript sequences from GENCODE (v451ift37) annotation and shortened to + / -20 bp around the mutation. If multiple mutated transcripts generated the same context sequence, said transcripts were collapsed into one search query. Subsequently, context sequences were annotated with expression profiles in the healthy tissue index using kqneo.

[0336] To compare the tumor-specificity of gene fusions associated with structural variants (SVs) versus RNA-only detected events, CCLE data from 327 cell lines (Ghandi, M. et al. Nextgeneration characterization of the Cancer Cell Line Encyclopedia. Nature 569, 503-508 (2019) were reanalyzed. EasyFuse V2.0.4 was used to predict gene fusions events from RNA-seq data (model prediction score >= 0.5). Somatic SVs identified from whole genome sequencing (WGS) via SvABA were obtained from Dep Map Portal, were lifted from GRCI137 to GRCI138 with rtracklayer (Lawrence, M., Gentleman, R. & Carey, V. rtracklayer: an R package for interfacing with genome browsers. Bioinformatics 25, 1841-1842 (2009) and annotated with overlapping genes using Ensembl V89 reference annotation. RNA-seq fusion transcripts were matched to SVs based on gene pairs and breakpoint proximity (<100 kb). Fusions for which both gene partners were overlapped by at least one SV were classified as “with SV”, those with only one gene partner affected (“partial SV”) were excluded, and fusions with no overlapping SVs were classified as “without SV”. The fusion transcript sequences(context sequence around breakpoint) were searched with kqneo against the GTEx / SRA index. Based on the above, gene fusion transcript associated to structural variants was analyzed.

[0337] Hence, it was assessed whether kqneo can predict the tumor specificity of gene fusion transcripts to support their use as neoantigens. Gene fusions are often identified from tumor RNA-seq data, and without matched whole-genome sequencing (WGS) data on structural variants (SVs), their tumor specificity remains uncertain. Using 327 tumor cell line samples with matched RNA-seq and WGS-based SV data, kqneo-derived tumor specificity between fusion transcripts associated with somatic SV breakpoints and those supported only by RNA-seq were compared. Fusion transcripts identified solely at the RNA level were frequently detected in healthy tissues (50.5%), indicating lower tumor specificity. In contrast, among SV-supported fusions, only 4.5% were detected in healthy tissues and 95.5% were classified as tumor-specific (see Fig. 5C).

[0338] Results

[0339] As a next step, non-synonymous point mutations were analyzed with k neo and a mutated transcript sequence defined to be tumor specific if it is detected by k neo in <1% of healthy tissue samples.

[0340] Then, germline single nucleotide polymorphisms (SNPs) with somatic single nucleotide variants (SNVs) from whole-exome sequencing data of 73 melanoma patients were compared. While germline SNPs with a population allele frequency > 5% are expected to be contained in the healthy tissue data, the somatic SNVs should be strongly tumor-specific and therefore largely absent from the unmatched healthy tissues.

[0341] As expected, k neo detected most of the germline SNPs (87.2%) in the healthy tissues RNA-seq data and classified only 12.8% of SNPs as tumors-specific. From the somatic SNVs, only 7.2% were detected and with 92.8% were classified as tumor specific, indicating a very strong differentiation between germline and tumor-tumor specific variants using k neo with unmatched healthy tissue data (Figure 5B). Additionally, it was observed for the expression of germline SNPs, that the sample rate in the k-mer index was positively correlated with the population allele frequency (R2 = 0.56, p< 2.2 x io16)(see Figure 4C, Figure 10B, Figure 12 A, B).

[0342] Example 6: Re-analysis of mutation-induced splice junctions in melanomaAfter analyzing non-mutated shared antigen targets, the next aim was to test if k4neo can demonstrate the tumor-specificity of neoantigen candidates encoded by splice junction which are derived from somatic mutations.

[0343] Somatic mutation can disrupt canonical splicing sequence motifs or create novel splice sites, leading to highly tumor-specific splice junctions which can encode neoantigens. Lang et al., 2017 previously reported on average 1.7 mutation-associated splice junction per melanoma sample. In addition, between 641 and 5,606 novel splice junctions without any mutational support were identified in the tumor RNA per patient.

[0344] Method

[0345] Splice junction data from RNA-seq of melanoma samples were used. In previous work, tumor-specific splicing in these samples were predicted with the tool splice2neo by associating the effect of somatic SNVs in cis-regulatory regions (e.g. gain or loss of splice motifs) with splice junctions identified in tumor RNA. In short, the effect of a mutation on splicing was predicted using SpliceAI (vi.3.1) (Jag ana than K et al. Predicting splicing from primary sequence with deep learning. Cell. 2019; 176:535-48) and MMSplice (v2.i.i) (Cheng J et al. MMSplice: modular modeling improves the predictions of genetic variant effects on splicing. Genome Biol.2019;20:1-15) using default parameters. Mutation-derived splice junctions were filtered with absolute prediction scores > 0.35 returned by the Al tools. Expressed splice junctions were predicted using Spladder (V3.0.2) (Kahles A etal., SplAdder: identification, quantification and testing of alternative splicing events from RNA-Seq data. Bioinformatics. 2016;32:1840-7) and Leafcutter (vo.2.9) (Li YI et al. Annotation-free quantification of RNA splicing using LeqfCutter. Nat Genet. 2018;50:151-8) based on the spliced alignment of STAR (v2.7.oa) against the hgi9 reference genome. The R-package splice2neo was then used to combine the RNA-derived and mutation-associated splice junctions and subsequently annotated them with the corresponding transcript sequence. This analysis resulted in 59 mutation-associated junctions. Here, the purely RNA-derived junctions that were not associated with a mutation were additionally analyzed. Moreover, non-canonical junctions were considered that were absent from an exclusion list based on GEN CODE reference annotation and healthy samples from GTEx, resulting in 33,552 RNA-derived junctions. All junction sequences were queried with kqneo in the healthy tissue index and the number of junctions per tumor sample and junction class that were detected in zero compared, at least one percent or more healthy tissue samples.

[0346] ResultsSimilar to the variant analysis, tumor-specificity was analyzed with k4neo for splice junctions with and without somatic mutation association. For 88% of junctions (100399 out of 13694) of the RNA-seq derived junctions, were detected as expressed in healthy with sample rate >1% in any tissue by ignoring immuno-priviliged tissues. Among the junctions predicted to be caused by a mutation, 91.53% (54 / 58) were tumor-specific (< 1% sample rate in all tissues ignoring immuno-priviliged tissues) (Figure 5D, E). Next, the RNA-seq derived splice junction candidates were further investigated, which were already filtered for novelty using a GENCODE and GTEx based exclusion lists. It was assessed how many novel candidates per sample can be obtained by using kqneo with different tumor-specificity criteria. Screening the index for expression with a strict tumor-specificity criterion, (< 1% sample rate in all tissues ignoring immuno-priviliged tissues) resulted in a median of 310 and median of 101 additional candidates per sample in the Hugo and Riaz cohorts, respectively (Figure 5F). Allowing the expression in up to one percent of the total healthy samples increased the median number of candidates to 527 and 166.5 for the Hugo and Riaz cohorts, respectively.

[0347] It was demonstrated that kqneo and the healthy tissue index differentiates RNA-seq derived splice junctions into commonly in healthy and tumor-specific expressed candidates (Figure 5G). Moreover, it was shown with kqneo that mutations-associated splice junction candidates are strongly tumor-specific and that that the individual candidate repertoire can be additionally expanded with uniquely and recurrently expressed RNA-seq derived junctions.

[0348] As a next step, it was aimed to demonstrate the use of kqneo to expand the tumor antigen repertoire with individual splice junctions that are highly tumor-specific.

[0349] Example 7: Individual and shared tumor antigen candidates from RNA-seq derived splice junctions

[0350] To identify tumor antigen candidates from splice junction that are shared among multiple tumor patients, splice junctions recurrently detected in tumor RNA-seq were predicted using the previously described pipeline. Studies with accessible RNA-seq data from tumor-entities with low mutational burden were collected and the raw FASTQ data from SRA was downloaded (see Table 5).

[0351] Table 5: Overview of accessible RNA-seq data from tumor- entities with low mutational burden.Cancer Cohort Samples Authors PMID entity

[0352] GBM PR.JNA212047 100 Bao et al. 25135958

[0353]

[0354] GBM PR.JNA310727 32 Gabrusiewicz etal. 26973881

[0355]

[0356] UVM PR.JNA851027 19 Yao etal. 37756564

[0357]

[0358] SKCM PRJNA312948 38 Hugo et al. 26997480

[0359] Method

[0360] Furthermore, a computational pipeline, called NeoRasp, was developed and implemented to identify tumor-specific splice junctions from RNA-seq data from individual tumor samples (Figure 13 A). NeoRasp incorporates state-of-the-art tools for splicing prediction tools and is implemented as reproducible SnakeMake (V8.24.1) workflow with software dependencies managed with Open Container Initiative (OCI) containers and Apptainer. Given paired-end FASTQ files as input, reads are aligned against the reference genome (here hg38) using STAR (v2-7.il) using ENCODE3 parameters. Splice junctions from the high confidence “SJ.out.tab” are filtered based on a user-defined detection cut-off (here >= 5 uniquely mapped reads). Canonical splice junctions from the reference gene annotation (here GENDCODE V46) are removed from the detected candidates. Next, the R-package FRASER (v2.o) is used to calculate the splice junction usage (Intron Jaccard Score) of each candidate junction using the genomic alignment. Splice junctions are annotated with counts per million (CPM) normalized read count values and reference gene and transcript expression values (TPM) estimated with Salmon (vi.10) on the STAR alignment in transcript coordinates (Aligned.toTranscriptome.out.bam). Subsequently, splice2neo (vo.6.13) is used to annotate junction candidates with corresponding transcript and peptide sequence.

[0361] The read support of candidates splice junctions is then re-quantified in a targeted manner with easyquant (vo.6.o) using the transcript context sequences around the junction with stringent bowtie2 alignment parameters (-no-mixed — dpad o — gbar 99999999 — mp 1,1 — np 1 — score-min L, 0,-0.01).

[0362] In the final step, all annotations and scores of the junctions are combined into a unified table. For splice junctions leading to change in the peptide sequence, a separate output file is generated that is compatible to NeoFox to annotate MHC-I and MHC-II binding epitopes and antigen features.Results of individual samples were gathered into a unified table and annotated for tumorspecificity. Splice junctions with the following per sample criteria were defined as reliable events: (i) 5 uniquely mapped reads; (2) Junction read support CPM > 0.1; (3) Splice usage (Intron Jaccard) > 5%. Splice junction were defined as shared events in a tumor cohort when it was reliable detected in 5 % of the samples of a tumor cohort and at least in 5 samples. As negative control analysis, the same methods were applied to a set of independent healthy cohorts (Blood, Brain, Heart). It was possible to filter out splice junctions that did not alter the peptide sequence of any transcript and splicing events that involved deep intronic splicesites (> 5obp away from next canonical donor / acceptor) for which the reference-based peptide annotation was not unambiguously possible.

[0363] Results

[0364] First, an end-to-end pipeline, called NeoRasp, was implemented to derive novel splicejunctions from RNA-seq raw data and applied it to cohorts of glioblastoma (GBM, n=i32), skin melanoma (SCKM, n = 38), and uveal melanoma (UVM, n = 19) samples, resulting in about 250.000 splice junctions per sample. After filtering for reliable junction read support, exclusion from reference transcript in GEN CODE, and exclusion from a long-read-based healthy tissue exclusion list candidate splice junctions (median 881 per UVM sample) (Figure 6A) were derived. Next, the junctions were filtered with kqneo for tumor-specificity, resulting in 4 to 80 tumor-specific splice junctions per sample (Figure 6B). As control, the same method was applied to 3 additional healthy tissue cohorts from blood, brain, and heart samples, which were not used in the kqneo index. Notably, a median of zero tumor-specific junctions in the healthy tissue (Figure 6B) was observed, indicating a low false-positive rate.

[0365] Next, it was tested whether the detected tumor-specific splice junctions are shared among at least 5% of patients from the same tumor type. While no junctions were recurrently detected in all healthy tissue control cohorts, 13, 27, and 38 shared and tumor-specific splice junctions were detected in the tumor cohorts GBM, SCKM, and UVM, respectively (Figure 6D). The top 10 most recurrent splice junctions per tumor cohort were analyzed in more detail. Most of these splice junctions have high junction read count support, lead to a reading frame shift and result in long (up to 160 AA) novel peptide sequences (Figure 6E).

[0366] Next, the recurrence of the shared tumor-specific junctions in independent data with the kqneo approach was assessed. Here, RNA-seq data from 31 tumor types from TCGA was used to build a pan-cancer kqneo index. By querying the shared junctions, the recurrence of nearly all shared splice variants in their respective tumor types (Figure 6F) was confirmed. Notably,candidates identified in GBM are also detectable in low-grade gliomas (LGG), and many candidates from SKCM are also detectable in UVM.

[0367] It was sought to expand the antigen repertoire by identifying highly tumor-specific splice junctions in multiple tumor types. The splice-junction detection pipeline, NeoRasp, was applied to nine cancer cohorts comprising 1,244 samples: glioblastoma (GBM, n=i26) (Gabrusiewicz, K. et al. Glioblastoma-infiltrated innate immune cells resemble Mo macrophage phenotype. JCI Insight 1, 085841, 85841 (2016); Bao, Z.-S. et al. RNA-seq of 272 gliomas revealed a novel, recurrent PTPRZ1-MET fusion transcript in secondary glioblastomas. Genome Res. 24, 1765-1773 (2014)), skin melanoma (SKCM, n=38) (Hugo, W. et al. Genomic and Transcriptomic Features of Response to Anti-PD-1 Therapy in Metastatic Melanoma. Cell 165, 35-44 (2016)), uveal melanoma (UVM, n=i9) (Yao, T. etal. Long-Read Sequencing Reveals Alternative Splicing-Driven, Shared Immunogenic Neoepitopes Regardless of SF3B1 Status in Uveal Melanoma. Cancer Immunol. Res. 11, 1671-1687 (2023)), acute myeloid leukemia (AML, n=4O9) (Lavallee, V.-P. et al. RNA-sequencing analysis of core binding factor AML identifies recurrent ZBTB7A mutations and defines RUNX1-CBFA2T3 fusion signature. Blood 127, 2498-2501 (2016)), non-small cell lung cancer (NSCLC, n=i99) (Goldmann, T. et al. PD-L1 amplification is associated with an immune cell rich phenotype in squamous cell cancer of the lung. Cancer Immunol. Immunother. CII 70, 2577-2587 (2021)), small cell lung cancer (SCLC, n=57) (Heeke, S. et al. Tumor- and circulating-free DNA methylation identifies clinically relevant small cell lung cancer subtypes. Cancer Cell 42, 225-237.es (2024)), triple negative breast cancers (BRCA / TNBC, n=57) (Weber, D. et al. Accurate detection of tumor-specific gene fusions reveals strongly immunogenic personal neo-antigens. Nat. Biotechnol. 40, 1276-1284 (2022)), colorectal cancers (CRC, n=88) (Sun, Q. et al. Long-read sequencing reveals the landscape of aberrant alternative splicing and novel therapeutic target in colorectal cancer. Genome Med. 15, 76 (2023)) and sarcomas (SARC, 0=147) (Lesluyes, T. et al. Genomic and transcriptomic comparison of post-radiation versus sporadic sarcomas. Mod. Pathol. Off. J. U. S. Can. Acad. Pathol. Inc 32, 1786-1794 (2019)). After requiring sufficient read support, excluding junctions matching annotated GENCODE transcripts, and removing junctions present in a long-read-based healthy-tissue exclusion set (Glinos, D. A. etal. Transcriptome variation in human tissues revealed by long-read sequencing. Nature 608, 353-359 (2022)); Reese, F. et al. The ENCODE4 long-read RNA-seq collection reveals distinct classes of transcript structure diversity. bioRxiv https: / / d0i.0rg / 10.1101 / 2023.05.15.540865 (2023) doi: 10.1101 / 2023.05.15.540865), a median of 884 novel candidate junctions per UVM sample were identified (Fig. 6A). Filtering these candidates with k4neo for tumor specificityreduced this number by more than tenfold, to a median of 79 tumor-specific junctions per UVM sample. Across tumor types, a median of 4-80 tumor-specific junctions per sample were detected (Fig. 6B). Applying the same pipeline to independent healthy-tissue cohorts from blood (Aguirre-Gamboa, R. et al. Deconvolution of bulk blood eQTL effects into immune cell subpopulations. BMC Bioinformatics 21, 243 (2020)), brain (Ramaker, R. C. et al. Post-mortem molecular profiling of three psychiatric disorders. Genome Med. 9, 72 (2017)), heart (Johnson, E. K., Matkovich, S. J. & Nerbonne, J. M. Regional Differences in mRNA and IncRNA Expression Profiles in Non-Failing Human Atria and Ventricles. Sei. Rep.

[0368] 8, 13919 (2018)) and liver (Gupta, R. et al. Comparing in vitro human liver models to in vivo human liver using RNA-Seq. Arch. Toxicol. 95, 573-589 (2021)) (n=248), which were not included in the healthy-tissue index, yielded a median of zero tumor-specific junctions, consistent with a low false-positive rate of k4neo (Fig. 6B).

[0369] As can be derived from Fig. 13B, splice junction-derived antigen candidates identified in individual patients were examined next. After restricting to junctions for which an altered protein sequence could be annotated, 4,772 candidates were identified. These comprised exon skipping (53.2%), alternative 3' splice-site usage (31.5%), alternative 5' splice-site usage (9.4%) and exitrons (5.9%). Overall, 60.5% of candidates generated frameshift peptides, whereas 39.5% preserved the reading frame. Using patient-specific HLA alleles for MHC binding prediction, it was found that 59.7% of peptides contained at least one epitope predicted to be presented by MHC class I and / or class II.

[0370] It was next assessed whether these highly tumor-specific splice junctions were recurrently detected in patients of the same tumor type and potentially qualify as shared antigen candidates. While no recurrent junctions were detected in the healthy control cohorts, between 4 and 835 shared tumor-specific junctions were identified per tumor type (see Fig.

[0371] 6D). After restricting to peptide altering events, few or none shared candidates for GBM (n=i), TNBC (n=2), CRC (n=4), NSCLC (n=o), and SARC (n=o) were obtained, but between 33 and 109 shared candidates for UVM (n=33), SCLC (n=73), and AML (n=iO9), which were detected in up to 80% of patients (Fig. 13B, C). Subsequently, the focus was on the 10 most recurrent junctions in AML, UVM and SCLC (see Fig. 6F). These junctions clustered strongly by tumor type, while some were expressed across entities. Exon-skipping events predominated in AML and SCLC, whereas alternative splice-site usage prevailed in UVM. Most junctions introduced frameshifts and encoded novel peptide sequences up to 80 amino acids in length (see Fig. 6F). Next, recurrence in the larger TCGA cohorts was assessed. These junction sequences with kqneo against the TCGA index were queried and confirmedrecurrent detection of the AML and UVM junctions at high rates in the corresponding TCGA cohorts (see Fig. 6G).

[0372] Using patient-specific HLA types for epitope binding prediction, recurrent junctions that generated at least one predicted strong MHC binder in more than 25% of patients were identified next (AML n=7, SCLC n=8, UVM n=9; Fig. 6G). It was checked how many shared variants would need to be combined to provide broader patient coverage. In SCLC, combining four exon-skipping events (KNL1, PLEKHG4B, ADGRVi and SCN3A) covered 98.3% of patients with at least one predicted strong-binding epitope (Fig.7A). Adding splice variants in SDK2 and SLC6A3 increased coverage to 94.9% of patients with at least two predicted binders. In AML, the top-ranked candidates covered 83.9% of patients with at least one epitope, 63.3% with at least two epitopes and 45.7% with three or more epitopes (Fig.

[0373] 7B). In UVM, all top recurrent junctions were predicted to encode strong-binding epitopes in expressing samples (Fig. 7C).

[0374] Then, the 10 most recurrent tumor-specific splice junctions were confirmed using sample-matched or indication-matched long-read RNA sequencing. All top recurrent UVM candidates were confirmed in sample-matched long-read data (Fig. 7C), and six SCLC and three AML candidates were confirmed in indication-matched long-read datasets (Fig. 7 A,B), indicating a high rate of independent full-length transcript validation. Notably, a multi-exon-skipping transcript variant of AGAPT5 (AGAPT5_vi) was expressed at higher levels than the canonical reference transcript in the long-read data (Fig. 7C). AGAPT5 VI introduces a frameshift and encodes 19 novel amino acids; it was recurrently detected in 36% of AML samples, of which more than 80% were predicted to present MHC epitopes derived from this variant.

[0375] It was found, that k4neo reveals highly tumor-specific splice junctions as individual and shared tumor antigen targets.

[0376] Example 8: II LA- typing and neoantigen feature annotation

[0377] HLA-types of individual tumor samples were determined with seq2HLA (V2.3) (Boegel, S. et al. HLA typing from RNA-Seq sequence reads. Genome Med. 4, 102 (2012) based on RNA-seq data using default parameters. HLA-types of individual samples were collected into a unified table following the NeoFox (Lang, F., Riesgo-Ferreiro, P., Lower, M., Sahin, U. & Schrdrs, B. NeoFox: annotating neoantigen candidates with neoantigen features. Bioinformatics 37, 4246-4247 (2021) patient-data format (see NeoFox manual). Next,NeoFox (vi.2.2) (Lang, F., Riesgo-Ferreiro, P., Lower, M., Sahin, U. & Schrdrs, B. NeoFox: annotating neoantigen candidates with neoantigen features. Bioinformatics 37, 4246-4247 (2021) was used to annotate splicing derived peptide sequences with neoantigen features using default parameters. MHC binding predictions were classified into strong and weak binders using the following cut-offs: MHC-I strong binder: NetMHCpan_bestRank_rank < = 0.5

[0378] MHC-I weak binder: NetMHCpan_bestRank_rank <= 2

[0379] MHC-II strong binder: NetMHCIIpan_bestRank_rank <= 1

[0380] MHC-II weak binder: NetMHCIIpan_bestRank_rank <= 5

[0381] To confirm the existence of the short read predicted splice junctions matched and unmatched Oxford Nanopore (ONT) sequencing data for 19 UVM samples (Yao, T. et al. Long-Read Sequencing Reveals Alternative Splicing-Driven, Shared Immunogenic Neoepitopes Regardless of SF3B1 Status in Uveal Melanoma. Cancer Immunol. Res. 11, 1671-1687 (2023), five SCLC cell lines (H526, H211, SHP77, H146, H69) (You, Y. et al. Benchmarking long-read RNA-sequencing technologies with LongBench: a cross-platform reference dataset profiling cancer cell lines with bulk and single-cell approaches. Preprint at https: / / d0i.0rg / 10.1101 / 2025.09.11.675724 (2025) and one AML sample (Liu, Q. et al. LongGF: computational algorithm and software tool for fast and accurate detection of gene fusions by long-read transcriptome sequencing. BMC Genomics 21, 793 (2020) were used. Raw reads were aligned with minimap2 (Li, H. New strategies to improve minimap2 alignment accuracy. Bioinformatics 37, 4572-4574 (2021)) against HG38 using GENCODEV46 (Frankish, A. et al. GENCODE 2021. Nucleic Acids Res. 49, D916-D923 (2021) as reference (with options: -ax splice — secondary=no — junc-bed gencode_46.bed). Splice junctions were extracted and annotated from generated alignment using regtools (Cotto, K. C. et al. Integrated analysis of genomic and transcriptomic data for the discovery of splice-associated variants in cancer. Nat. Commun. 14, 1589 (2023) (with options: regtools junctions extract -a 12 -s XS -t ts <in.bam> | regtools junctions annotate -0 <out.junc> - <hg38_primary_assembly.fa> <gencode46.gtf>.). Junctions were searched within generated BED 12 files and junctions supported by two or more long-reads were considered confirmed.Together, the results demonstrate the detection of highly tumor-specific splice junctions that are shared among many tumor samples form the same tumor type, which highlights a rich source of novel tumor-specific antigen candidates.

[0382] Although modifications and changes maybe suggested by those skilled in the art, it is the intention of the applicant to embody within the patent warranted hereon all changes and modifications as reasonably and probably come within the scope of this contribution to the art. The features of the present invention which are believed to be novel are set forth in detail in the appended claims. The features disclosed in the description and in the claims could be essential alone or in every combination for the realization of the invention in its different embodiments.

Claims

Claims1. A method for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples,obtained from a subject, in particular from at least one subject, more in particular from at least two subjects, the method comprising:(a) Providing RNA sequencing data from said biological sample,(b) Identifying transcript variant sequences in the RNA sequencing data from step (a),(c) Dividing the transcript variant sequences from step (b) into short overlapping sub-sequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample, (d) Providing reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples,(e) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (c) to the reference k-mer database of reference biological samples from step (d),(f) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (e).

2. The method according to claim 1, whereinstep (f) further comprises predicting binding of one or more epitope peptides, comprised within the tumor-specific antigen candidates derived from the identified transcript variants, to one or more HLA molecules.

3. The method according to claim 1, whereinstep (f) further comprises selecting tumor-specific antigen candidates comprising epitope peptides predicted to bind to HLA molecules of more than one subject, in particular thereby identifying shared variants.

4. The method of any of the preceding claims, whereinstep (b) further comprises identifying transcript variant sequences by filtering of the provided RNA sequencing data.

5. The method of any of the preceding claims, whereinstep (f) further comprises filtering the identified tumor-specific transcript variant sequences.

6. The method of any of the preceding claims, whereinin step (e) the identified transcript variant sequences are verified using an orthogonal verification method, in particular comprising RT-PCR, qRT-PCR and / or Sanger sequencing.

7. The method of any of the preceding claims, whereinstep (b) further comprises providing structural variant information to identify transcript variant sequences, in particular fusion transcripts.

8. The method of any of the preceding claims, whereinsaid subject is a human suffering from a proliferative disease, in particular from cancer and said biological sample is a proliferative disease cell sample, in particular a cancer tumor cell sample selected from a group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors, neuroendocrine tumors, in particular selected from the group consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

9. The method of any of the preceding claims, whereinthe reference k-mer database of reference biological samples from step (d) is a collection of RNA sequence data from biological samples of healthy subjects and / or subjects suffering from cancer, and wherein each k-mer RNA sequences dataset of biological samples was obtained by applying k-mer indexing to the plurality of raw RNA-sequencing reads.

10. The method of any of the preceding claims, whereinthe lengths of the k-mers are shorter than the length of the identified transcript variant sequences in the RNA sequencing data of the biological sample or RNA- sequencing raw reads of reference samples, preferably the length of k-mer is 18-32nt, more preferably the length of k-mer is 18 - 24 nt, even more preferably the length ofk-mer is 19-23 nt, even more preferably the length of k-mer is 20 - 22 nt, even more preferably the length of k-mer is 21 nt.

11. The method of any of the preceding claims, in particular according to claim 10, characterized by ak-mer rate higher than o, preferably in the range 0.1-1, more preferably in the range 0.4 - 0.9, even more preferably in the range 0.6 - 0.8, and even more preferably the k-mer rate is 0.7.

12. The method of any of the preceding claims, whereinthe length of the k-mers in the obtained k-mer RNA sub-sequences dataset for the biological sample and the reference k-mer RNA sub-sequences dataset of reference biological samples are the same.

13. The method of any of the preceding claims, whereinthe sample k-mer dataset comprises k-mer obtained by dividing the transcript variant sequences into short overlapping sub-sequences of length k (k-mers), wherein the transcript variant sequences were identified on basis of aberrant transcript features.

14. The method of any of the preceding claims, whereinthe reference k-mer database of reference RNA sequences is built by:Obtaining sequence information of reference RNA sequences from raw RNA- sequencing readsandDividing the read sequences into k-mers to obtain one or more overlapping k-mers, andcomparing a transcript variant based on the aberrant transcript feature k-mer data to biological samples of healthy subjects and subjects suffering from cancer.

15. The method of any of the preceding claims, whereinsaid comparing the sample k-mer dataset with a reference k-mer database is performed by selecting a k-mer being exactly identical in transcript variantinformation based on the aberrant transcript feature data from the biological sample to the reference k-mer database,wherein an absence of the identified transcript variants from the sample k-mer dataset in the biological samples of healthy subjects from the reference k-mer database indicates a tumor-specific transcript variant, coding for a tumor-specific antigen candidate.

16. The method of any of claims 13-15, whereinthe aberrant transcript features comprise splice junctions, fusion transcripts, intron retentions, exitrons, neo open reading frames, viral sequences, bacterial sequences, transposable elements, and chimeric transcripts.

17. The method of any of the preceding claims, whereinthe method is a computer-implemented method.

18. A tumor-specific antigen candidate identified according to the method of any of the claims 1-17, whereinthe tumor-specific antigen candidate is recognized by an antigen recognizing construct,wherein the antigen recognizing construct is preferably selected from an antibody, or fragment thereof, or a T cell receptor (TCR), or fragment thereof.

19. A pharmaceutical composition comprising the tumor-specific antigen candidate identified according to the method of any of the claims 1-17, and a pharmaceutical acceptable carrier, stabilizer, and / or excipient.

20. A tumor-specific antigen candidate identified according to the method of any of the claims 1-17, and pharmaceutical composition according to claim 18 for use in medicine, preferably for use in the diagnosis, prevention, and / or treatment of a proliferative disease.

21. The tumor-specific antigen candidate according to claim 20, whereinthe proliferative disease is a cancer.

22. The tumor-specific antigen candidate according to claim 21, whereinthe cancer is selected from a group consisting of carcinoma, sarcoma, lymphoma, germ cell tumor, blastoma and neuroectodermal-derived tumors, neuroendocrine tumors.

23. The tumor-specific antigen candidate according to claim 21, whereinthe cancer is selected from a group consisting of glioblastoma, skin melanoma, uveal melanoma, acute myeloid leukemia, non-small cell lung cancer, small cell lung cancer, triple negative breast cancers, colorectal cancers and sarcomas.

24. A system for identifying at least one tumor-specific antigen candidate in a biological sample, in particular in at least one biological sample, more in particular in at least two biological samples, obtained from a subject, in particular from least one subject, more in particular from at least two subjects, comprising at least one processing device configured to:(a) Receiving in the at least one processing device the identified transcript variant sequences in the RNA sequencing data from a biological sample, (b) Dividing the transcript variant sequences from step (a) into short overlapping sub-sequences of length k (k-mers) to obtain a sample k-mer dataset for said transcript variant sequences from said biological sample using the at least one processing device,(c) Receiving in the at least one processing device reference k-mer database comprising k-mer RNA sequences dataset, in particular comprising k-mer transcript variant RNA sequences dataset, of reference biological samples, (d) Identifying tumor-specific transcript variant sequences in RNA sequencing data from step (a) by comparing the obtained sample k-mer dataset from said biological sample from step (b) to the reference k-mer database of reference biological samples from step (c) using the at least one processing device,(e) Identifying the at least one tumor-specific antigen candidate based on the identified tumor-specific transcript variant sequences in step (d) using the at least one processing device.

25. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of claims 1 to 17-26. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of claims i to 17.