A recent report in Science announced the publication of a new human blood protein atlas, describing the disease signatures of thousands of proteins circulating in the blood. Minimally invasive protein profiling marks a step forward in the personalisation of medicine. Some interesting statistical and machine learning techniques were employed.
Blood Protein Study
The researchers’ methods included a technique called proximity extension assay (PEA), which makes use of highly specific probes of DNA strands to detect minute concentrations of proteins in the blood plasma. Amplification with PCR (Polymerase Chain Reactions) allowed 5,416 proteins to be evaluated.
A longitudinal dataset showed dramatic changes as children passed through adolescence to adulthood. The central part of the study was a cross-sectional analysis, where age, sex and BMI were identified as important explanatory factors. The signatures of 59 clinically relevant diseases, in seven classes, can be viewed interactively in The Human Protein Atlas.
Into the secretome
Rather than the hideaway of a reclusive cockney, the secretome refers to the ensemble of secreted proteins. From a data science perspective, the challenge was how to find the signatures of a wide range of diseases, based on the differential abundance of over 5,400 proteins. This was complicated by the fact that many proteins elevated by a particular disease were also found to be elevated in other diseases.
“To investigate the distinct and shared proteomics signatures across diseases, we performed differential abundance analyses. Several groups were used as controls, including healthy samples, a disease background consisting of all other diseases, and samples from the same disease class.”
The differential abundance of proteins was evaluated using normalised protein expression units (NPX). The volcano chart above plots the p-values against the multiplicative (fold) change in NPX, both on log scales. The red values on the right were unusually high and the blue values on the left was exceptionally low.
The researchers used a logistic LASSO approach to identify the importance of proteins in providing a signature of each disease against its cohort. In the case of HIV above, CRTAM was the most significant explanatory factor, even though CD6 had the most extreme p-value.
How does logistic LASSO work?
A logistic model is trained on target values of one or zero, in this case representing the presence or absence of a disease. Least absolute shrinkage and selection operator (LASSO) is a version of linear regression that selects the most relevant explanatory variables using L1 regularisation. Adding the sum of the absolute values of the regression coefficients to the objective function forces the contribution of irrelevant variables towards zero as the hyper-parameter, λ, is increased. This property was particularly useful for the disease signature problem, where there were thousands of potential explanatory proteins.
The tricky aspect of LASSO is tuning the hyper-parameter, λ. You want it to be high enough to eliminate irrelevant variables, but not so high that it discounts the useful explanatory features. In the protein study, this was addressed using cross-validation: randomly splitting the data into 70:30 training and test sets, then rerunning the regression for a range of λ values. The quality of a model can be assessed in terms of both its accuracy and its required number of inputs, using criteria such as the Akaike information criterion or Bayesian information criterion, which favour parsimony. Repeating the randomisation 100 times, the researchers could home in on an optimal value of λ. The regression coefficients of the resulting model could then be used to rank the importance of the relevant proteins, as shown in the right hand side of the panel above.
Personalised health
The potential for a cheap, annual blood test to screen the whole population is immense. Proteomics adds to the arsenal of resources available to help people stay healthy. Early indications of diseases like cancer can be critical in initiating treatment. There is plenty of room to broaden the scope beyond the current 59 diseases, to include rarer conditions, such as Motor Neurone Disease, which has impacted some top sportsmen. It would be extremely helpful to find proteins related to the apparent epidemic of mental health issues, which are hard to define and lack objective, quantitative diagnostic criteria.
Science4Performance 