Finding cancer biomarkers from mass spectrometry data by decision lists

Finding accurate biomarkers is key to early diagnosis and successful treatment of many otherwise incurable diseases. In this work, we study the problem of finding biomarkers through mass spectrometry (SELDI-TOF) spectra from cancerous and normal tissues. In contrast to the common practice of using vague methods such as genetic algorithms, or uninterpretable methods such as Support Vector Machines, we look for a method that is simple, intuitive, interpretable, usable, and more accurate. We introduce decision lists to this domain. Our experiments on clinical cancer datasets demonstrate that decision lists can achieve more accurate results than other methods. More interestingly, the resulting decision lists are more interpretable for possible causal relationship between cancer and differentially expressed proteins, and directly usable in clinical biomarker design. In particular, our approach is capable of finding multiple biomarkers with high sensitivity and specificity. Such a feature will provide clues for medical experts to thoroughly investigate the roles of protein in cancer development and progression.