To generate potential frameshift neoAg derived from microsatellite loci, we utilized MSIsensor2 (https://github.com/niu-lab/msisensor2) (11) to scan the human reference genome (GRCh38) for the identification of all microsatellite loci. Short nucleotide repeats exceeding five units were identified as microsatellites. For larger repeated motifs (ranging from 2 to 5 base pairs), a minimum repeat number of three was used. Only microsatellites within protein-coding regions were retained to generate potential mutant proteins. Insertions/deletions differing by a multiple of 3 will share the same reading frame, resulting in identical downstream sequences, so that we can generate two types of downstream frameshift sequences for each identified microsatellite: 3n+1 and 3n+2 shifts of nucleotide bases, where n is an integer (-3, -2, -1, 0, 1, 2, 3, and so on). For example, in the 3n+1 series, two nucleotide deletions (n = -1 and 3n+1 = -2) in the sequence will share the same reading frame with one nucleotide insertion (n = 0, 3n+1 = 1), and four nucleotide insertions (n= 1 3n+1 = 4) and so on. This applies similarly to the 3n+2 series. Therefore, to efficiently represent these frameshift mutations, we introduce two in silico variants for each microsatellite: a two-nucleotide deletion (n = -1, 3n+1 = -2) and a one-nucleotide deletion (n = -1, 3n+2 = -1). These variants encompass the spectrum of frameshift mutations within each series and generate entirely distinct downstream amino acid sequences compared to the wild-type sequence, thus making them ideal neoAg candidates.

It is important to note the significant complexity of potential mutant amino acid sequences generated at the junction region, where the wild-type and mutant sequences meet within the microsatellite region. The size and location of indels at this junction significantly impact the resulting mutant sequence. However, based on our previous experimental data (10), very few (<5%) verified neoAg epitopes originate from these junction regions. Therefore, we excluded the sequences around the junction at which the frameshifted amino acid occurred to prevent this complexity. Then, putative neoAg epitopes were predicted towards a panel of higher frequency MHC-Class I (12) and MHC-Class II (13) alleles ( Table 2 ) providing coverage for 97% and 99% of the general population, respectively. Various algorithms such as MHCflurry (14), MHCnuggets (15), NetMHC (16), PickPocket (17), SMM-align (18), NNalign (19) that are implemented in pVACtools (20) were employed to predict neoAgs. Any epitope and allele pairs with a binding affinity of IC₅₀ <50 nM in any algorithm were considered potential epitopes for subsequent processes. We predicted the binding stability for Class I epitopes using NetMHCstab (21) and assessed foreignness to the human proteome using antigen.garnish (https://github.com/andrewrech/antigen.garnish) (22). Also, other characteristics that could play an important role in epitope selection such as terminal amino acids and the Gravy score (average hydropathy) were annotated.

To gain a deeper understanding of the potential epitopes, we also annotated the corresponding mutations using the Ensembl Variant Effect Predictor (VEP) (23). Annotations include associated gene names, IDs, genomic coordinates, and among others. Another critical factor for the effectiveness of neoAgs being presented by MHC is their expression level (24–26). Higher expression levels will increase the probability of these epitopes being presented by MHC molecules. We integrated the expression levels of the corresponding genes generating the neoAgs from the Cancer Genome Atlas Program (TCGA, https://www.cancer.gov/tcga) project by using all datasets that contain pairs of normal and tumor. Additionally, for the dataset with MSI status information, we distinguished and listed the MSI-H and MSS groups separately. Differentially expressed genes (Benjamini-Hochberg adjusted p-value < 0.05) between normal and cancer tissues in each of the different cancer data set were labelled. Furthermore, we annotated the frequency of the mutation and associated phenotypes using the dbSNP (RRID:SCR_002338) (27) and ClinVar (RRID:SCR_006169) database (28). This annotation provided valuable insights into the frequency and clinical implications of the mutations from where the neoAgs are derived. Moreover, we included evidence of experimentally confirmed epitopes using IEDB (RRID:SCR_006604), which contributes to the reliability and validity of the identified epitopes.

We constructed the website using a microservices architecture with Docker. The backend database (MySQL) was employed to store and manage data. Express.js served as the middleware, responsible for translating HTTP requests into MySQL queries. Vue.js was utilized to develop the user interface. Nginx served as the HTTP server communicating between the containers hosting the frontend and backend components of the website.

We selected 100 epitopes derived from mutations in microsatellite loci among the top-ranked based on their predicted immunogenicity ( Supplementary Table S2 ) in LS patients and showed that 65 out of 100 predicted neoAg candidates were validated for their immunogenicity using in vitro ELISPOT assays (10). These peptides were used to assess the coverage across different databases. Specifically, on April 9^th, 2024, we individually searched the 65 validated immunogenic epitopes within MONET and the 10 databases listed in Table 1 . Then, we evaluated the number of epitopes recorded in each database. The services of dbPepNeo, NeoPeptide, and TANTIGEN databases were unavailable at the time of our research. Therefore, we reported results from MONET and the other 7 databases.

Public data analyzed in this study was obtained from multiple sources including: https://www.ncbi.nlm.nih.gov/genome/, https://www.ncbi.nlm.nih.gov/snp/, https://www.ncbi.nlm.nih.gov/clinvar/, https://www.iedb.org/, https://www.cancer.gov/tcga. The epitope prediction data in this study are available at https://monet.mdanderson.org.

The overall data processing is depicted in Figure 1 . We identified a total of 34,067,744 microsatellite loci in the human genome GRCh38. Among them, only 3,934,634 microsatellites were located in protein-coding regions ( Table 3 ). After eliminating duplicated sequences, we obtained 492,578 unique frameshift mutation neoAg sequences ( Table 3 ). From these sequences, we identified 4,449,128 MHC Class I and 15,589,846 MHC Class II potential epitopes. The number of epitopes binding to MHC Class I alleles for each allele ranged from approximately 6,000 to 800,000 ( Figure 2A ), while binders to Class II alleles displayed a higher number of predicted epitopes ranging from approximately 5,000 to 6 million ( Figure 2B ).

After generation of predicted potential epitopes using our pipeline, the binding stability was annotated using antigen.garish ( Figure 1 ). Then, the expression levels of the corresponding genes carrying the mutations were annotated using TCGA data. A total of 64,288 epitopes have been experimentally verified and recorded in the IEDB ( Table 3 ). A total of 542,442 mutations in MONET were recorded in the dbSNP database, which contains human variants including small indels original from germline or somatic mutations. Of these mutations, 13,289 are linked to the ClinVar database, which associates human variation with their potentially clinically relevant results. While most of the variant recorded in dbSNP and ClinVar are of germline origin, 198 somatic mutations in MONET have been recorded in ClinVar. The top ClinVar phenotypes ( Table S1 ) include malignant tumor of prostate (Rank 1, n = 31), carcinoma of colon (Rank 2, n= 24), and colorectal cancer (Rank 5, n=16). In one hand, these clinical records verified our putative neoAgs that might present in dMMR/MSI-H cancers. In the other hand, the limited coverage of the current clinically available databases suggests that our computational results could be very valuable in bridging this gap.

In MONET, the landing page can be accessed at https://monet.mdanderson.org. The key functions of MONET are: ‘Search Neoantigens’ and ‘Best Neoantigens’. Both can be accessed from the sidebar.

In the ‘Search Neoantigens’ page, users have the option to search for neoAg using various parameters, including gene symbol and related IDs, epitope sequence, mutation HVGSp/HVGSc ID, and ClinVar phenotype. Moreover, users can narrow down their search by limiting it to specific MHC class I and II types. This functionality proves particularly useful when users have a specific target gene or disease in mind and would like to search into the details of a particular epitope ( Figure 3A ).

As an example, users can utilize the search function to find potential Class I neoAg epitopes resulting from mutations in TGFBR2. The search engine generates 520 summarized results, providing information such as peptide sequence, mutation location, related gene, and associated HLA (Human Leukocyte Antigen) alleles. To further refine the results, users can utilize the filter sidebar on the right-hand side. This allows filtering based on peptide sequence, affinity cutoff, and target HLA alleles. In cases where multiple genes are returned, users can also apply a filter based on the gene symbol. Once a specific epitope of interest is identified, users can select it to further access detailed information about the epitope ( Figure 3B ).

Detailed epitope information is offered in five tabs, each providing specific details ( Figure 4 ): 1. Prediction affinity will display a table of predicted affinities generated by different algorithms for various HLA types ( Figure 4 ) and other features of the epitope such as the Gravy score; 2. The mutant gene will show essential gene information related to the epitope, including the gene’s location, related IDs, and both wild-type and mutant sequences of the protein. The mutant sequence is highlighted in red ( Supplementary Figure S1 ); 3. ClinVar and dbSNP will provide access to the gene ID and corresponding link to ClinVar and dbSNP databases. Additionally, if available, basic information about related diseases and frequency data will be presented ( Supplementary Figure S2 ); 4. If the peptide has been experimentally tested in the IEDB (Immune Epitope Database), this tab will showcase relevant information from IEDB; 5. Users can explore the expression of the target gene in solid normal tissue versus primary solid tumors across different TCGA databases. Thus, this tab displays a bar plot illustrating this information ( Supplementary Figure S3 ).

Within the best neoAg search engine function, users have the capability to search for all potential epitopes specific to one or multiple HLA alleles ( Figure 3B ). This feature allows for customized filtering options, such as an affinity and expression cutoff in a specific TCGA cancer type. This functionality proves particularly valuable when users intend to identify the top potential epitopes for a specific set of HLA alleles from either an individual patient or a group of patients.

As an example, users can select HLA-A*02:01 and HLA-B*07:02 alleles along with a median affinity cutoff of 30 nM across multiple algorithms. Furthermore, they can narrow down the search to include only the top 20% most highly expressed genes in the COAD dataset ( Figure 3B ). In total, 3,186 results for HLA-A*02:01 and 2,614 results for HLA-B*07:02 returned. These potential epitopes are generated from mutations in 1,483 genes. Users have the option to download the results directly from the page, and detailed information for each peptide can be accessed on the corresponding peptide details page.

The most distinctive feature of MONET is its focus on epitopes derived from microsatellite loci that are targets of MMRd. We compared the coverage of 65 of such epitopes, the immunogenicity of which has been validated using ELISpot assays (10), across MONET and other popular databases. MONET covers 57 out of these 65 verified epitopes ( Figure 5 , Supplementary Table S2 ). In comparison to MONET, TSNAdb, only covers 15 of the 65 epitopes, and IEDB covers just one. GNIFdb, CAD, NEPdb, CAPAD, and AntiJen do not contain any of these epitopes. Therefore, MONET demonstrates excellent coverage (>85%) of our target epitopes, which are derived from microsatellite loci due to MMRd, thus significantly outperforming other available databases in this field.