Senescence, or in simple terms, biological ageing is an intricate process present in any living organism. While there has been constant research in simplifying the ageing notion and developing anti-ageing processes, these have limited practical applications in real life. One of the developments that has grabbed attention is biomarkers, which helps measure biological factors.
Biomarkers are generally chemicals found in living substances (genes, cells) that ascertain physiological features. They are usually obtained through biological samples and through biopsies. In areas such as clinical trials and experiments, biomarkers have become a standard predictor of clinical outcomes.
Now, with the advancement of deep learning (DL), biomarkers can be improved significantly. Medical researchers have been tapping the potential of neural networks, which forms the essence of DL. The neural network and biomarker combination has led to developments ranging from detecting cancer to even distinguishing mental illnesses such as Alzheimer’s disease. In this article, we review a clinical study where biomarkers give out predictors for human ageing through neural networks.
The ageing Process
Ageing in humans or any living entity is a continuous process and there are plenty of factors that affect ageing in general. Be it from a lifestyle perspective (food choices, exercise, health complications) or environmental impact (living space,surroundings), they all contribute to the ageing process equally.
Studies have shown that chemical reactions in the human body drastically impact the process of ageing. One such popular and accepted study that has been going around with biology researchers, is the free radical theory of ageing by Denham Harman in the 1950s. The theory posits that the free radicals present in the cells and tissues eventually destroy these living components as humans age.
Consequent studies holding this theory in the hindsight have led to biomarker development to assess human ageing. Certain unique biomarkers worth mentioning are Glycans and methylomes. However, these substances have partly helped indicate diseases through human age. Now, researchers are relying on machine learning as well as deep learning to develop biomarkers which predict ageing almost accurately.
Deep Neural Networks To The Rescue
In a research by Evgeny Putin and team at Insilico Medicine, US, they designed an ensemble for 21 deep neural networks (DNNs) which derives information from a simple blood test. Once the DNNs are developed, they are trained with data such as blood biochemistry and cell count tests, which are linked to age and gender. For this study, five markers in blood are considered for analysis,:
- Albumin
- Glucose
- Alkaline phosphatase
- Urea
- Erythrocytes
The researchers consider a data set of 62,419 anonymous blood biochemistry records of healthy individuals, which had details such as the individual’s age, gender and 46 standardised blood markers for prediction.
The study process involved a sequence of operations such as data preprocessing, normalisation, outlier reduction and splitting the dataset to train and test it. In the words of the researchers, the basis for age prediction is described below:
“Since we treated human age prediction as a regression problem, we used two metrics to estimate the performance of the method: standard coefficient of determination (R2) and ε-prediction (epsilon-prediction) accuracy. When using epsilon-prediction accuracy, the sample is considered correctly recognised if the predicted age is in the range of [true age -ε; true age +ε], where ε controls the level of certainty in the prediction. So if ε = 0, then it is a simple classification accuracy. In this study, we considered ε = 10. The key advantage of using epsilon-prediction accuracy is that it allows cohort analysis without fixed age ranges (e.g. 10-20, 20-30)”.
This is done to optimise prediction through DNNs. Now, a single DNN showed a standard coefficient of determination value (R2) of 0.80 which means that the accuracy is around 82 percent. Now, a total of 21 DNNs are combined with a technique called Stacking to increase this accuracy. Stacking fits other machine learning (ML)models to DNNs. For this study, it was found that ElasticNet was the best fit ML model that goes with DNNs. Also, the predictions should not vary significantly. So, DNNs are trained with different hyperparameters, tweaked with a number of neural layers and neuron count in the network, activation functions and so on.
Now these DNNs are mapped with the blood markers mentioned earlier. It is done with a feature selection technique called modified Permutation Feature Importance (PFI). With this, the markers are revealed better compared to ML models.
Conclusion
By using DNNs in biomarkers, complex interactions between molecules can be analysed to ascertain ageing almost correctly. DL techniques will definitely help in tracing rare diseases and ageing syndromes, and it will be quick and efficient.
