Immunotherapy has enormous potential to provide cancer patients with a treatment which is more personalised, more precise, and more effective than current therapies, but evidently its promises will only come to fruition with the assistance of improved predictive algorithms and bioinformatics tools. Recent explosions in publicly available cancer genomic data, coupled with advancements in machine learning methods is ensuring that the marriage of computation and biology will help to address challenges facing immunotherapy in the coming decade.
Quite deservingly, there has been great excitement in recent years surrounding the use of artificial intelligence and machine learning algorithms for immunotherapeutic discovery, as well as a wide range of other biological and medical applications. One area which has profited greatly from these computational tools is the discovery of neoantigens for cancer vaccine development.
Cancer Neoantigens and Vaccine Development
Underlying all immune responses is the ability of the immune system to distinguish self from non-self – to discern whether any protein it encounters is part of its own human body, or has come from some invading entity. In the context of cancer, tumour cells are recognised as ‘mutated self’ – a signal for the immune destruction. Based on this principle, there has recently been an enormous interest in developing immunotherapies capable of selectively eliminating cancer cells by amplifying the patient’s own immune response against his/her tumour. Cancer neoepitopes or neoantigens are protein products of the mutated cancer genome which serve to trigger an antitumour immune response upon presentation to immune cells called CD8+ T cells.
This places neoepitopes at the forefront of cancer immunotherapy. The challenge now for cancer biologists is to identify which neoepitopes are iliciting potent immune responses against tumours, and which ones might hold promise in developing anti-cancer vaccines. Similarly to vaccines used to immunise against viral infections, these immunotherapies would introduce some tumour-specific antigen into the body so that the immune system mounts a response against it, selectively killing tumour cells while leaving healthy cells unharmed.
To achieve tumour-specific killing, the vaccine should target MHC-1-restricted epitopes which then activate CD8+ T-cells to specifically kill cancer cells, binding to MHC-peptide complexes via their T-cell receptor. Neoantigens are presented to T-cells by Antigen Presenting Cells – usually Dendritic Cells or tumour cells themselves. Antigen presentation is preceded by a train of intracellular events including proteasomal processing and transport of the resulting peptide to the cell surface.
Designing Neoepitope Prediction Algorithms with Artificial Intelligence
As cancer cells progressively acquire somatic mutations, so too does the array of potential neoepitopes expand. The task at hand then is to predict which of these neoepitopes are likely to be presented on the tumour cell surface, to bind MHC-1 and be recognised by the T-Cell Receptor in vivo to trigger an immune response. Predicting clinically relevant neoantigens is a challenging task. Immunogenic mutations are highly specific to each individual patient, being restricted by the unique type of MHC molecules expressed on their cells. Moreover, a cancer’s mutational fingerprint is extremely heterogeneous between patients, as are cancer cells within one single tumour. The ideal neoantigen would need to be expressed homogeneously and in sufficient concentrations such that an immune response designed against it would destroy the entire tumour.
Because so few mutations are immunogenic, high throughput computational methods to predict these neoepitopes are being employed to identify promising candidates for cancer vaccine development. Artificial Neural Networks (ANNs) – a branch of artificial intelligence – have been designed to determine how likely is a given 3D structure or 2D amino acid sequence to become an epitope. Such algorithms are largely focused on predicting peptide interactions with MHC-1. As the majority of peptides presented by MHC-1 are 9 amino acids in length, the algorithms usually look at peptides of length 8-11aas. One such algorithm is NetMHC, developed by scientists at the Technical University of Denmark. which has been trained using data from a large number of different MHC alleles. The server produces as output a predicted IC50 value representing the binding affinity of a peptide with MHC-1.
A further consideration which is critical here is predicting cross-reactivity of the neoantigen with non-tumour antigens in the body. If they are targeted therapeutically, shared epitopes could result in off-target toxicity, threatening patient safety. This can also be addressed using computational methods. German researchers have developed a tool called Expitope to assess crossreactivity of immunotherapeutic antigens against naturally expressed proteins in human tissues .
What’s Next On the Horizon?
There remains a long road to travel towards the ideal predictive algorithm for cancer vaccine development. It is important to realise that the accuracy of the algorithm depends on the training data used as input. These predictive tools are typically constructed based on IC50 values obtained from binding affinity experiments in vitro. This means that the quality of in silico predictions is influenced by data gained previously from in vitro affinity studies. Basing predictions on binding affinity alone is problematic as it fails to fully reflect the complex peptide loading pathway which takes place in vivo. Only a small percentage of predicted high affinity binders are in reality recognised by the patient’s T-Cells, making them potential candidates for immunotherapy. As a result, current strategies have scientists wading through a sea of false positives in the search for neoantigens.
There remains a major need to develop reliable methods of accurately verifying that neoepitopes found in silico can actually elicit a tumor-specific immune response and ultimately achieve tumor regression. In in vitro experiments, a binding affinity threshold of < 500nM is used to determine a peptide’s potential to be a neoepitope, but this simplification may cause putative neoepitopes with lower affinity to be overlooked. It must be emphasised that many factors other than binding affinity can also play a fundamental role in presentation of neoepitopes and immune activation. Rather than solely relying on binding affinity, improved computational methods will need to be developed which consider all aspects of antigen processing and presentation in order to successfully predict neoepitopes. Indeed similar sequence-based ANN algorithms are also in development which predict other informative steps in peptide processing such as proteasome cleavage and Transporter Associated with Antigen Binding (TAP) protein binding. An even more daunting challenge which scientists hope to soon tackle is to predict interaction of the antigen-MHC complex with the T-Cell Receptor, whose binding mechanism is considerably more complex
Yarchoan, M., Johnson III, B. A., Lutz, E. R., Laheru, D. A., & Jaffee, E. M. (2017). Targeting neoantigens to augment antitumour immunity. Nature Reviews Cancer, 17(4), 209.
Brennick, C. A., George, M. M., Corwin, W. L., Srivastava, P. K., & Ebrahimi-Nik, H. (2017). Neoepitopes as cancer immunotherapy targets: key challenges and opportunities. Immunotherapy, 9(4), 361-371.
Jiang, T., Shi, T., Zhang, H., Hu, J., Song, Y., Wei, J., … & Zhou, C. (2019). Tumor neoantigens: from basic research to clinical applications. Journal of hematology & oncology, 12(1), 93.