Machine Learning Models Human Protein: Virus Interaction Network
By Ruth Seeley
We know that the body’s protein molecules carry out essential cell function and interact with each other in varying degrees of complexity. Viruses to which we’re exposed disrupt proteins’ interactions and manipulate them so they can replicate.
It was adversaries like viruses that inspired researchers at USC to study how exactly they interact with proteins in the human body. “We tried to reproduce this problem using a mathematical model,” said Paul Bogdan, associate professor in the Ming Hsieh Department of Electrical and Computer Engineering at USC, lead author of “Reconstructing missing complex networks against adversarial interventions,” published in Nature Communications this April.
The “protein interaction network” models each protein as a node. If two proteins interact, an edge connects them. Virus attacks essentially remove network nodes and links, making the original network no longer observable.
“Some networks are highly dynamic. The speed at which they change may be extremely fast or slow,” Bogdan said. “We may not have sensors to get accurate measurements. Part of the network cannot be observed and hence becomes invisible.”
To trace the effect of a viral attack, scientists needed to reconstruct the original network by finding a reliable estimate of the invisible part. Relying on a statistical machine learning framework allows researchers to trace all possibilities and find the most probable estimate.
The lab actively incorporates the influence and causality of the attack, or “adversarial intervention,” into their learning algorithm rather than treat it as a random sampling process. Bogdan explained, “Its real power lies in its generality—it can work with any type of attack and network model.”
As a result, the research has far-reaching applications to any network reconstruction problem involving adversarial attack, in fields as diverse as ecology, social science, neuroscience, and network security. The USC paper can also determine the influence of trolls and bots on social media users.
Bogdan plans to extend their work by experimenting with a range of attack models, more complex and varied datasets, and larger network sizes to understand their effect on the reconstructed network.
