Why is “Systems Biology” more relevant than you may think?

How to actually use “Systems Biology”?

If we keep dreaming, certain ideas will eventually come true—that is, unless we wake up while falling from a building before hitting the ground. But, despite the concurring “falling into the void” moments in systems biology research, there is one emerging tool that might lead to a bright future for the field. This is Machine Learning (ML).

As you've already heard, Machine Learning employs enormous datasets to model, analyse, and understand complicated data, and what's more complex than biological data? Both machine learning and systems biology appear to be a natural fit, but the truth is a little more challenging.

To begin, Systems Biology is dealing with overfitting data size while training ML algorithms. Some researchers have noticed that many ML applications contravene the rule-of-thumb that a dataset should include at least ten times as many data points as the number of independent variables [8] . Let’s consider a theoretical example of predicting drug response in cancer patients.  Suppose we want to use ML to predict how patients with a specific type of cancer will respond to a new drug. We have the following data:

1. Gene expression data for 100 genes

2. Protein levels for 50 proteins

3. 10 clinical parameters (age, tumour size, etc.)

In total, we have 160 independent variables (features). But, according to the rule-of-thumb mentioned, we would need approximately 1,600 patients (10 times the number of features) to build a reliable ML model. However, in reality, we might only have data from 100 patients due to the rarity of the cancer type and the newness of the drug.

Another example works the other way around: suppose we wish to utilise machine learning to forecast the risk of inflammatory bowel disease (IBD) based on gut microbiota composition. In this example, we have significantly more data:

1. Microbiome composition (relative abundance of 1000 bacterial species)

2. Metabolomic data (levels of 500 metabolites)

3. Host genetic information (100 relevant Single Nucleotide Polymorphism (SNPs))

4. Environmental factors (10 variables like diet, stress levels)

The sum of this example would be 1,610 independent variables, therefore, we would need about 16,100 samples. But, in reality, we might have only data from 500 individuals due to the cost and recruitment challenges.

In this regard, in addition to sample sizes and independent variables, other factors such as data heterogeneity, generalisability, and biological complexity must be considered while employing ML. To address these challenges, researchers could use strategic feature selection using the most relevant independent variables, combine multiple models, transfer knowledge across related studies, integrate diverse biological data types, and carry out long-term observations to capture system dynamics over time. It will always rely on the conditions of the research settings and the requirements of the ML algorithm.

Nevertheless, there is a solution that caught my attention more than others. Maybe being an avid gamer helps to come to this association but, Voit et al. proposed using the gaming industry's AI and ML efforts as an example. For starters, a videogame always requires what is called a game engine to run the different elements integrated, and these game engines are usually built as a user-friendly environment that serves as a toolbox for game creation. In other words, a game engine integrates rendering, physics, scripting capabilities, sound integration, AI, streaming, networking, memory management, localization support, etc. Put differently, it is a complex system interconnecting multiple functionalities and modules. In this regard, Voit et al. suggests creating a “bioengine” that would function in Systems Biology similar to a game engine for a video game. This solution could have endless applications facilitating interactive visualisations, real-time simulations of biological phenomena, ecosystem modelling, or even creating a virtual lab environment for running simulated experiments. Now, despite what some may believe, this is not merely wishful thinking about futuristic uses; it is a reality in the AlphaFold (2023) project. Google DeepMind created an AI system that computationally predicts the 3D structure of 230 million proteins. This technology functions as a "bioengine" that can aid in understanding, among other things, how to discover the structure of SARS-CoV-2 viral proteins for innovative therapies, or how to characterise the structure of haemoglobin to create treatments for hereditary haemoglobin disorders. These breakthroughs have helped to reduce costs and time-consuming processes to figure out the exact structure of a protein as these can be generated from their 3D structure [9].