AI Predictors in mRNA Cancer Vaccine Neoantigen Selection

This article provides a comprehensive overview of how immuno-oncology pipelines utilize artificial intelligence (AI) to predict Major Histocompatibility Complex (MHC) binding for the creation of personalized mRNA cancer vaccines. By analyzing patient-specific tumor mutations, AI models identify which resulting mutated proteins—known as neoantigens—will successfully bind to MHC molecules and trigger a targeted immune response. The following sections break down the step-by-step computational process, from sequencing patient DNA to selecting the optimal neoantigen candidates for mRNA vaccine manufacturing.

Sequencing and Mutation Identification

The pipeline begins by obtaining two biological samples from the patient: a tumor biopsy and a healthy blood sample. Next-generation sequencing (NGS) is used to perform whole-exome sequencing (WES) on both samples. By comparing the genetic code of the tumor against the healthy tissue, bioinformaticians identify somatic mutations—genetic alterations unique to the patient's cancer. Furthermore, RNA sequencing (RNA-seq) is conducted on the tumor sample to confirm which of these mutated genes are actively being expressed and transcribed into proteins.

Generating In Silico Peptides

Once the expressed mutations are confirmed, the computational pipeline translates these genetic alterations into amino acid sequences. The software slices these sequences into short, overlapping chains of amino acids called peptides. These peptides vary in length depending on the target pathway: typically 8 to 11 amino acids for MHC Class I (which activates CD8+ killer T-cells) and 12 to 25 amino acids for MHC Class II (which activates CD4+ helper T-cells). A single tumor can generate thousands of these potential neoantigen peptides, creating a massive dataset that requires rapid filtering.

MHC Typing

Before predicting how peptides will behave, the pipeline must determine the patient's specific MHC profile, known in humans as the Human Leukocyte Antigen (HLA) system. HLA genes are the most polymorphic in the human genome, meaning every patient has a distinct set of MHC molecules. Algorithms analyze the patient's sequencing data to identify their exact HLA alleles. This step is critical because a neoantigen peptide that binds strongly to one patient’s MHC molecules may not bind at all to another's.

AI-Driven MHC Binding Prediction

With the patient's HLA type and the list of mutated peptides established, AI predictors are deployed. These predictors are typically built using deep learning architectures, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or Transformer models. These AI models have been trained on vast datasets generated from binding assays and mass spectrometry-eluted ligand data, which teach the AI the biochemical rules of peptide-MHC interactions.

The AI evaluates each mutated peptide alongside the patient's specific HLA alleles and calculates two primary metrics:

  1. Binding Affinity: The AI predicts the strength of the physical interaction between the peptide and the MHC binding groove, often expressed as an IC50 value in nanomoles (nM). A lower value indicates a stronger bond.
  2. Antigen Processing and Presentation: Advanced AI models go beyond simple binding. They predict the likelihood that a peptide will survive intracellular cleavage by the proteasome, get successfully transported into the endoplasmic reticulum via TAP proteins, and stably remain on the cell surface long enough for a T-cell to notice it.

Filtering and Ranking Candidates

After the AI scores thousands of potential interactions, the pipeline filters and ranks the neoantigens to find the most viable candidates. The ranking algorithm prioritizes peptides based on several strict criteria:

  • High MHC Binding Affinity: Only the top-scoring peptides that are predicted to bind tightly to the patient's specific MHC Class I and Class II molecules are kept.
  • High Tumor Expression: The mutated gene must be heavily expressed in the tumor, ensuring enough protein is available to be processed.
  • Dissimilarity to Self: The pipeline cross-references the neoantigen against the patient's normal proteome. If a neoantigen looks too similar to a healthy protein, it is discarded to prevent the vaccine from triggering an autoimmune reaction, or because the body’s immune system will likely ignore it due to central tolerance.
  • Clonality: The pipeline favors "truncal" mutations—those present in all cancer cells—rather than sub-clonal mutations present in only a fraction of the tumor.

Formulating the mRNA Vaccine

mRNA vaccines have a limited payload capacity, typically accommodating instructions for 20 to 34 distinct neoantigens. The immuno-oncology pipeline selects the top-ranked candidates from the AI's final output list. The amino acid sequences of these chosen neoantigens are reverse-translated into an optimized mRNA sequence. These sequences are concatenated into a single mRNA strand, separated by linker sequences, and encapsulated in lipid nanoparticles. When injected into the patient, the host cells read the mRNA, manufacture the AI-selected neoantigens, and present them on their MHC molecules, training the patient's T-cells to seek out and destroy the tumor.