Discovering signals of selection in cancer mutations with Hidden Markov Models

什么样的突变会导致癌症？这个重要问题的答案在于对癌细胞DNA数据的正确解读，这些数据现在已经大量存在。导致癌症发生的真正的“驱动突变”被大量随机的“乘客突变”所掩盖，这一事实使这项任务变得复杂。有时，当单个驱动突变系统地出现在许多独立的肿瘤样本中时，可以识别出它们。但这种情况很少发生，因为癌症进化的机制似乎是复杂的，并非没有选择。该项目的目的是设计统计和计算方法，以可靠地识别对癌症进化重要的DNA区域。这可以通过寻找选择信号来实现：当癌症需要基因处于与健康细胞不同的状态时，它们将以错义突变的形式表现出更高的基因重配置率。此外，为了显著改变基因，这些突变更有可能出现在通常不能容忍太多多样性的位置。结合这两个互补的方面可能是量化癌症突变在生物学上有意义和统计上强有力的方式的功能影响的关键。实际上，选择信号的检测不仅需要对观察到的癌症突变进行广泛的统计分析，而且需要对潜在的突变目标——人类基因组——本身进行广泛的统计分析。只有将所见与所能见进行比较，才能评估发现的重要性。隐马尔可夫模型的概率方法非常适合在大型数据集上有效地执行此任务。目标是建立一个计算框架，以实现对癌症测序数据的进化知情分析，目的是确定可以作为癌症进展驱动因素的基因组区域。

What kind of mutation causes cancer? The answer to this important question lies in the correct interpretation of cancer-cell DNA data, which now exists in abundance. This task is complicated by the fact that the actual “driver mutations“ that causally contribute to the development of cancer are disguised by a large pool of random “passenger mutations“. Sometimes, individual driver mutations can be identified when they systematically appear in many independent tumor samples. But such cases are rare, for it seems that the mechanisms of cancer evolution are intricate and not without alternative. It is the aim of this project to devise statistical and computational methods to robustly identify DNA regions that are important to cancer evolution. This can be done by finding signals of selection: when genes are required by the cancer to be in a state that is different from their configuration in healthy cells, they will exhibit a higher rate of genetic reconfiguration in the form of missense mutations. Moreover, in order to alter the gene considerably, these mutations are more likely to appear at locations that usually do not tolerate too much diversity. Combining these two complementary aspects could be the key to quantify the functional effects of cancer mutations in a biologically meaningful and statistically powerful manner. Practically, the detection of signals of selection requires extensive statistical analysis not only of the observed cancer mutations but also of the potential mutation target - the human genome - itself. Only by comparing what was seen to what could have been seen can one assess the significance of findings. The probabilistic method of Hidden Markov Models is ideally suited to perform this task efficiently on large data sets. The goal is to establish a computational framework to implement an evolutionarily informed analysis of cancer sequencing data with the objective to identify genomic regions that can act as drivers for cancer progression.