The T cells of the immune system have a key role in the identification and elimination of cells that pose a threat to the body, such as infected cells and cancer cells. Two papers by Balachandran et al.1 (page 512) and Łuksza et al.2 (page 517), which have many authors in common, propose a framework to assess how effectively tumours can be detected by T cells — a tumour property known as immunogenicity. The authors demonstrate that their models for assigning tumour-immunogenicity scores can be used to predict clinical responses to a type of cancer immunotherapy called checkpoint blockade.
Most cells in the body present peptide fragments known as antigens on their cell surface, which are generated from intracellular proteins. Each peptide is bound in a complex with a specialized receptor called an MHC class I protein (HLA class I in humans). T cells known as cytotoxic T cells police the body in search of cells displaying specific antigens, especially antigens from infectious organisms, or in the case of cancer, antigens known as neoantigens that have arisen as a result of a mutation (Fig. 1). If the T-cell receptor (TCR) of a cytotoxic T cell recognizes and binds an antigen that is not normally present, the T cell will often unleash an attack that kills the cell displaying that antigen. TCRs are highly variable and have slightly different antigen-binding regions, enabling the immune system to recognize millions of antigens3. Antigen binding to MHC proteins and TCR recognition of antigen–MHC complexes are key determinants of an immune response.
Tumour cells often fight back against this immune-system surveillance by hijacking the natural mechanisms that dampen immune responses, which are normally intended to block autoimmmune attacks against healthy tissue. Checkpoint-blockade therapies can block these immuno-inhibitory signals, such as those generated by the ‘checkpoint’ PD-L1 protein4. However, only a subset of tumours treated with such therapies regress. Therefore, approaches are needed to identify the tumours that are most likely to respond to immunotherapy.
Current ways of predicting the effectiveness of checkpoint-blockade therapy rely on measuring the level of PD-L1 protein expressed by tumour cells, counting the number of T cells in a tumour, and estimating the number of different neoantigens that a tumour contains5. The work by Łuksza and Balachandran and their respective colleagues offers a new type of integrated model to predict whether a tumour will be attacked by T cells, a characteristic that they refer to as tumour fitness (low fitness being associated with a strong immune response against the tumour).
The authors calculate tumour fitness by assessing the immunogenicity of the neoantigens that a tumour contains. To estimate the immunogenicity of each neoantigen, the authors first considered how tightly each patient’s MHC protein binds to each given neoantigen compared with the non-mutant version of the antigen, and second, they assigned a score for the likelihood of a neoantigen–MHC complex being recognized by a TCR. Although MHC shape can vary depending on the version of MHC protein present, computational algorithms can accurately predict the affinity of a given antigen for any version of MHC protein encoded in a patient’s genome.
However, predicting which antigens are more likely than others to be recognized by TCRs remains a challenge. To tackle this problem, both groups of authors make the simplifying assumption that neoantigens are more likely to be immunogenic if they resemble infectious-disease-associated antigens that are known to stimulate T cells and therefore might have a higher probability of being recognized as ‘non-self’.
In both studies, the authors calculated the fitness of each individual tumour on the basis of the combined fitnesses of the subpopulations of tumour cells, known as clones, that contain different mutations. Each clone in the tumour was represented by a score for its neoantigen that was most likely to bind MHC and be recognized by a TCR. Łuksza and colleagues validated their model in an analysis of three groups of individuals diagnosed with cancer (two groups with melanoma and one group with non-small-cell lung cancer) and being treated with checkpoint-blockade therapy. The tumours predicted by their model to have lower fitness were indeed associated with longer patient-survival times. Furthermore, they demonstrated that the predictive power of their full model is superior to partial versions that used only some of the scoring criteria.
Balachandran and colleagues demonstrated that this type of modelling approach could distinguish between long-term and short-term survivors after a diagnosis of pancreatic cancer. An alternative model that they tested, which assessed the number of different neoantigens, could not. Furthermore, when blood samples from patients were analysed, Balachandran et al. found that, compared with short-term survivors, the long-term survivors had a higher frequency of neoantigens generated from mutations in the gene MUC16 and had T-cell responses against mutant MUC16. The authors suggest that MUC16 mutations generate neoantigens that are key antitumour targets of the immune system.
The tumour-fitness model is a mathematical model in which several parameters need to be set. These parameters relate to the timescale of response to therapy and the assessment of the probability that a neoantigen will be recognized by a TCR. To determine the numerical values to use for these parameters, both groups varied these parameter values so that the model would best match the observed patient-survival data.
Łuksza and colleagues employed an approach in which the parameters determined using data from one group of patients were used to predict survival of an independent group of patients. This is a commonly accepted way of avoiding a problem known as ‘over-fitting’, in which a model works only for a particular data set. Balachandran and colleagues determined the parameter values and predicted outcomes using the same group of patients, but argued that similar values were obtained when subsamples of the data were used for this parameter-setting purpose. The need to set the model parameters for each group of patients raises the question of how predictive the model will be for any given patient, considering the unique characteristics of their individual tumour type and the particular type of immunotherapy used.
Many other factors — in addition to affinity of antigen for MHC proteins and TCR antigen recognition — can influence the potency of neoantigens or correlate with checkpoint-blockade effectiveness. Thus, it would be desirable to be able to extend the model easily. Indeed, Łuksza and colleagues demonstrate the incorporation of additional parameters, such as the expression of cytotoxic genes in the tumour microenvironment, which is associated with immune-cell targeting of the tumour, and show that this addition improves the model’s predictive power. In the same spirit, a superior model could be achieved by taking yet more factors into account, such as biases in the protein-degradation process that generates antigens, or the level of neoantigen expression.
It remains possible that the similarity of a neoantigen to non-microbial antigens, such as antigens associated with autoimmunity, could also be used to predict immunogenicity and patient survival. Thus, more research is needed to pinpoint the basis of TCR recognition of antigens. Although many factors must be considered in the prediction of effective antitumour immune responses in a given patient (including non-mutant antigens, which can sometimes trigger an anticancer response), neoantigens are emerging as crucial targets that T cells can use to detect and destroy cancer cells, and represent important targets for immunotherapy6–8.
Nature 551, 444-446 (2017)