In silico prediction of cancer immunogens: current state of the art

Cancer is a catch-all term for a constellation of diseases typically characterised by abnormal cell division. The term cancer can be traced to the Greek physician Hippocrates (460-370 BC), who used the terms carcinoma and carcinos to refer to ulcer-forming tumours and non-ulcer forming tumours. In Greek, these words refer to a crab. The Roman physician, Celsus (28-50 BC), translated this to cancer, the Latin for crab. Galen (130-200 AD) used the Greek word oncos, meaning swelling to describe tumours. Almost all cells and tissues can become cancerous, but fortunately most cancers are very rare. Yet cancer remains one of the prime health issues of our time [1].

In 2012, there were about 14 million new cancer cases worldwide and 8.2 million deaths. Deaths caused by cancer is very high in developed countries [1]. In 2014, the US recorded 591,700 deaths from cancer, with approximately 197,233 deaths in women and 394,466 deaths in men; about 22% of all deaths. The equivalent UK figures were 163,000 deaths, or 450 deaths per day; with approximately 86,500 cancer deaths in men and 76,900 deaths in women; about 25% of all deaths. Yet over half of the global cancer burden occurs in less well developed countries. Lung, bowel, liver, and stomach, are the commonest cancers globally, equating to 4 in 10 deaths worldwide. At about 1 in 10 cases, smoking-related lung cancer is the commonest male cancer.

A cancer can be classed as either “common” or “rare” based on relative prevalence. The precise threshold between classes remains open. The US National Cancer Institute (NCI) identifies “rare” as those cancers with a prevalence below 15 in 100,000 [2]. This means only 11 adult cancers are defined as common in the US: prostate, breast, lung, bowl, cervical, bladder, rectum, ovary, kidney, melanoma, and non-Hodgkin lymphoma [3]. Other adult cancers - about 25% of all adult cancers - are, by this definition, “rare” [3].

Driven by the financial exigencies governing drug discovery and development, effective cancer treatment is significantly skewed towards common cancers. As an example, there are over 20 Category 1 intervention - uniform consensus that intervention is appropriate and based on significant evidence - for prostate and breast cancer, the commonest cancers in men and women [4‐6]. Yet none exist for say the bone cancers, chondrosarcoma or chordoma, which affect under 1000 individuals annually in the US [7‐9].

Survival varies considerably between different cancers. It ranges from 98% for testicular cancer to about 1% for pancreatic cancer. Most common cancers have a 10-year survival above 50%. Over 80% of those with cancers which are easy to treat and/or diagnose survive for 10+ years, yet less than 1 in 5 people with hard-to-treat or hard-to-diagnose cancers survive for 10 years or more [10]. Thus cancer remains a pivotal unmet medical need, driving both technical innovations and improved clinical practice, resulting in dramatic improvement in cancer treatment. In the UK, mortality rates peaked in the 1980s, with overall cancer mortality falling by 14% since the early 1970s, with a 22% decrease in men and an 8% decrease in women. In the UK, mortality for all cancers is predicted to decrease by 15% in the period 2014-2035, reaching less than 280 deaths per 1 hundred thousand by the year 2035 [10].

According to somatic mutation theory, mutations in DNA and epi-mutations disrupt the programmed regulation of cell division, upsetting the balance between proliferation and apoptotic cell death, resulting in excessive and uncontrolled division. Many mutations lead to cancer, but most do not. The treatment of solid tumours in particular has changed dramatically in recent years due to enhanced molecular diagnostics helping to identify a burdening number of addressable oncogenic abnormalities including in-frame insertions/deletions and amplification or rearrangements and gene activating point mutations.

Historically, cancer has been treated by small molecule drugs. A number of anti-cancer drugs are classed as agents of so-called chemotherapy. These are typically characterised by significant side-effects, as many affect cells indiscriminately. The main types of chemotherapy include DNA-damaging alkylating agents, including structurally-simple reactive molecules such as Busulfan; Antimetabolites, which compete with natural nucleotides for incorporation into DNA or RNA, impairing DNA replication, such as 5-fluorouracil; Anti-tumour antibiotics, such as complex natural product Epirubicin; Topoisomerase inhibitors, which interfere with DNA unzipping prior to replication, such as Topotecan; Mitotic inhibitors, such as plant-derived natural product Paclitaxel; and Corticosteroids, such as Prednisone [11]. Other, more targeted therapies are now appearing. Precision medicine can be defined as therapy individualised to each tumour, achieving this by exploiting quantifiable genetic alterations as de fact predictive biomarkers and/or as therapeutic or prophylactic targets for the next generation of cancer treatments.

Most recently, immune based approaches have gained significant saliency. Immunotherapy directed against cancer, include a triumvirate of main approaches: monoclonal antibodies, immune checkpoint inhibitors, and vaccines. The immune response has two arms: the humoral, or antibody-mediated, arm and the cellular arm, mediated primarily by T cells. Historically, almost all vaccine prophylactic responses have been mediated by Antibodies. Each human has billions of potential antibodies capable of recognizing proteins and tagging them for elimination. The individual ‘baseline’ for addressing antigen challenge is the primary naïve antibody repertoire. The structural and sequence diversity of this baseline enables the immune system to recognize, at least weakly, a very large set of antigens. Unfortunately, only a subset of Tumour Associated Antigens (TAAs) are amendable to the antibody mediated responses necessitating the exploration of cellular immune mechanisms as a replacement or adjunct therapy.

The effectiveness of potential therapeutic cancer vaccines is often reduced by mechanisms in cancer patients that suppress T-cells and antigen presenting cells (APCs). Most cancer vaccines induce anti-tumour immune responses when formulated with strong adjuvants, due to the general lack of immunogenicity exhibited by vaccines not derived from whole pathogens. Vaccination against cancer takes several forms: DNA-based vaccines, RNA-based vaccines, and DC-based vaccines.

DNA vaccines: trials to evaluate the efficacy of Inovio Pharmaceuticals combination vaccine INO-3112 are planned against cervical, head, and neck cancers (NCT02172911, NCT02163057) “http://​ir.​inovio.​com/​news-and-media/​news/​press-release-details/​2017/​Inovio-Begins-Phase-3-Clinical-Trial-of-VGX-3100-for-the-Treatment-of-HPV-Related-Cervical-Pre-Cancer/​default.​aspx”. INO-3112 contains plasmids encoding E6 and E7 (VGX-3100) [12] combined with DNA-based IL12 delivery (INO-9012). Inovio’s preventive anti-HIV DNA vaccine, PENNVAX-G, used in a prime-boost protocol with altered pox virus vector, has a satisfactory safety and immunogenicity profile [13]. This study should foment design of anti-cancer therapeutic vaccines by exploring prime-boost regimens using DNA vaccines and viral boosts. The Vaccibody-developed DNA-based vaccine VB10.16 targets HPV16 “http://​www.​vaccibody.​com/​vb10-16/​”. A trial (NCT02529930) is set to launch; if successful it should provide an innovative and much needed non-invasive way to treat HPV-induced cervical cancers.

RNA vaccines: Sahin’s group pioneered use of lipid-based positively-charged nanoparticles delivering RNA encoding TAAs, to target DCs in vivo and thus simulate an anti-viral response [14]. This is currently undergoing a phase I trial in melanoma patients (NCT02410733). A two component RNA vaccine platforms launched by Curevac has also yielded promising results in early trials (NCT00923312) [15].

DC-based vaccines: multiple platforms are being developed to harness ex vivo activated DC vaccines for cancer immunotherapy. These platforms include the with-antigen loading vaccine DCVax-Direct “https://​www.​nwbio.​com/​dcvax-direct/​” and the without-loading vaccine DCVaxL “https://​www.​nwbio.​com/​dcvax-technology/​”. Similarly, the Individualized Vaccines Against Cancer (IVAC) platform uses autologous DCs loaded with individually sequenced neo-antigens (NCT02035956, NCT02316457). The potential of DC vaccines is only beginning to be explored.

Protein-based vaccines: As TAA are poorly immunogenic, an adjuvant able to generate effective immune response should be added in the protein-based vaccines [16, 17]. Aluminum salts (alum) are used as adjuvants promoting protective humoral immunity, while for the activation of cell-mediated immunity are used conserved moieties associated with pathogen or endogenous alarmins like head shock proteins (HSPs). HSPs are able to induce both innate and addaptive immune responses. The first autologous HPS vaccine, Oncophage, failed to demonstrate survival benefits in Stage IV melanoma patients although stage I and II patients seemed to benefit from vaccination [18]. Wang et al. [19] have developed a platform for generating of chaperone complexes between HSPs and clinically relevant TAA.

Computational prediction can give important insight into both antibody and cellular immune responses. Here we examine non-experimental approaches to the cataloguing and prediction of TAAs. We describe the classification of TAAs into separate categories, databases that curate and classify TAAs, servers that facilitate the accurate and robust prediction of TAAs, and the role of DC vaccines to fight cancer and deliver pre-loaded epitope ensemble vaccines.

The worth, value, and utility of vaccines, though clear for all to see, is not yet unchallenged; yet most reasonable people are likely to agree that they are, qualifications apart, a thing of inestimable value and utility. Existing vaccines are not perfect. One might argue that their intrinsic complexity, and the highly empirical nature of their discovery over decades, and the fraught nature of their manufacture, is a root of current mistrust. In some senses this also hampered the progress of cancer vaccines and immunotherapy. Finally, these are beginning to make some headway.

Computational prediction has a part to play, one of the strongest messages to emerge from this review is that immunogenicity is a multi-factorial property: some protein antigens are immunogenic for one reason, or set of reasons, while another protein will be immunogenic for another possibly-tangential reason or set of reasons. Each such a causal manifold seems dauntingly complex and confusing. The prediction of immunogenicity for cancer antigens is a greater problems still in multi-factorial prediction since we must factor in the high degree of antigenic similarity to other host proteins. Thus the search for new antigens is a search through a multi-factorial landscape of contingent causes. As noted above, the immunoinformatic prediction of cancer antigens potentially underpins several important therapeutic strategies, including epitope ensemble vaccines, CAR T-cell therapy, and DC vaccines.

To develop proper predictive approaches to the prediction of cancer and other immunogenic antigens we need to address several issues. We require more “positive” and carefully curated, validated data focussed on cancer. While there are databases of vaccine antigens - AntigenDB [53] is a dedicated resource directly addressing this, as well as IEDB [54] - similar yet better data resources are still required, suggesting the need to enlarge, deepen, and broaden available data collections. We also require much better and much deeper representations of the sequence data. Single descriptors characterising the whole sequence [55], and other multivariate descriptors of sequences. One could envisage a phase space of disjoint descriptor variables from which variable selection protocols could extract a compact, near-optimal choice of indicative variables. Also, better algorithms are needed. Powerful machine learning toolkits, such as Weka, are already available, and these are more than capable of delivering robust and extensible methods provided the data and the data representation are adequate. Yet, as new algorithms appear we must not be complacent but open, embracing proven innovations.

Better protocols for establishing the immunogenicity of identified potential vaccines are desperately needed. This work is that of the experimentalist. Here a fast, straightforward methodology is required which projects a more consistent, clearer, and much more accurate picture of the immunogenicity of individual proteins. As with many computational studies of real world problems, there is also general need for experiments able to validate predictions. The in silico analyses of pathogen genomes and virtual proteomes, has led to the publication of innumerable papers reporting potential but unverified vaccine candidates [56‐58]. Such papers typically use methodology largely embodied in web-servers: operating such systems is facile, and the resulting analysis straightforward. Publishing unverified papers ultimately becomes counterproductive. Science progresses through independent corroboration by verification by peers. Science progresses faster when people do not waste time on fruitless research. Many are rightly alarmed by the increasing perception that the complex results of present day science cannot be reproduced and validated. Explanations are legion, including increased levels of scrutiny and institutional pressure on research and individual researchers. Arguably, the greatest issues are the increasing complexity and instrumentality of modern experimentation, in the opaqueness of many systems being studied, and the daunting technicality of analysing and teasing out the nature of many experiments. Computational experiments may be reproducible in themselves but without robust and reproducible experimental validation mean little. Other vaccine prediction studies give credibility to their results [59, 60] by linking vaccine design to experimental validation. Even in the current atmosphere of hysteria and hyperbole over AI, prediction lacking validation exerts slight influence and convinces few.

The utility of vaccines, though clear to most of us, is not yet unchallenged. Existing vaccines are not perfect. This also hampered cancer vaccines and immunotherapy. Finally, these are beginning to make some headway. Computational prediction has a part to play, one of the strongest messages to emerge from this review is that immunogenicity is a multi-factorial property. The prediction of immunogenicity for cancer antigens is a greater problems still in multi-factorial prediction since we must factor in the high degree of antigenic similarity to other host proteins. Immunoinformatics is poised to deliver on its potential and open up a whole new era in Cancer immunotherapy.