Artificial intelligence is still finding its way into a lot of our daily activities. What about biology, the scientific study of living things? AI is capable of sorting through tens of thousands of genomic data points to find possible new targets for treatment. Though scientists are unsure of how today’s AI models arrive at their judgments in the first place, these biological insights seem to be useful. Now, a brand-new system called SQUID shows up with the mission of cracking AI’s opaque core logic, which is like a black box.
Scientists at Cold Spring Harbor Laboratory (CSHL) developed a computational tool called SQUID, which stands for Surrogate Quantitative Interpretability for Deepnets. It’s designed to help interpret how AI models analyze the genome.
SQUID is more reliable than other analytic methods, lowers noise in the data, and can produce more precise predictions on the consequences of genetic changes.
How is it so much more effective? Assistant Professor Peter Koo of CSHL claims that SQUID’s specialized training holds the key.
“Most of the methods used to try to grasp these models come from other domains, such as natural language processing or computer vision. Although they have their uses, they are not the best for genomics. To further understand what these deep neural networks are learning, we used SQUID to draw on decades of knowledge in quantitative genetics, says Koo.
To begin, SQUID creates a library of more than 100,000 different DNA sequences.
Afterwards, a program known as MAVE-NN (Multiplex Assays of Variant Effects Neural Network) is used to examine the library of mutations and their consequences. With this instrument, scientists may run thousands of virtual tests at once.
They may, in essence, “fish out” the methods underlying the most precise predictions made by a particular AI. Their “catch” in computation might pave the way for more realistic tests.
“Actual laboratory trials cannot be replaced by in silico [virtual] investigations. Still, they can provide a wealth of information. As co-author of the study and associate professor at CSHL Justin Kinney says, “They can help scientists form hypotheses for how a particular region of the genome works or how a mutation might have a clinically relevant effect.” The ocean is filled with AI models. Every day, more come into the sea.
SQUID, according to Koo, Kinney, and associates, should assist scientists in identifying the ones that most closely match their specific requirements.
The human genome has been mapped, but it is still a very difficult terrain. SQUID might make it easier for scientists to explore the field and get them closer to the actual medicinal implications of their findings.
Reference – Prying open the AI black box