Challenging negative data helps AI models better identify effective antibodies

Imagine you're developing antibodies. Drugs that target accurately, such as viral proteins and Onco-Marker. I tested a series of antibodies and found that some antibodies work, while others don't.

I'd like to continue changing them and see if I can make them even better. However, you certainly don't want to waste time testing something that doesn't work. To test only those antibodies that may work, antibodies that do not bind to the target must be isolated before proceeding to a costly and time-consuming experiment.

One way to do this is to train a computational model that can support you in the process. Today, machine learning models are already helping experimental scientists narrow their searches.

Furthermore, machine learning models are once shown, allowing you to learn what antibodies bind. The binder stands out from something that is not. Without such a model, this is not entirely clear, as it goes beyond human perception and intuition,” says Aygul Minnegalieva. Candidate for University of Oslo.

She is investigating how to optimally train AI models with Grillabo. Minnegalieva and colleagues recently published a study on this Nature Machine Intelligence.

“But not all machine learning models do that right. They can only be used if the models are trained with the right data to gain an understanding of biological determinants, for example, those that make antibodies a binder,” she explains.

“One approach to achieving this is to present the model with examples of both accurate and false responses about what we want to recognize,” explains the Ph.D. Candidate.

Such false examples or errors are called negative data, and correct examples are classified as positive data.

Errors must pose challenges for the model to recognize. In the latest research, Minnegalieva and her colleagues found that the negative data that the models are exposed to must be challenging enough.

“You need to display in your model with incorrect examples that are very similar to the correct example. This way, data models learn more effectively,” Minnegalieva points out.

Specifically, researchers have presented the model as negative data with antibodies targeting proteins, for example with viruses, but they do, but are not optimal.

“In this way, the model has improved its ability to accurately communicate antibodies that are effective in combating pathogens,” she explains.

Most importantly, this method allows the model to capture the underlying sequencing factors of antibodies that help bind to pathogen proteins.

“These determinants have created more biological meaning,” says Minegalieva. “Essentially, the model has gotten better with reasoning.”

Accelerate the development of antibodies and drugs by AI

Machine learning is increasingly adopted in the development of new drugs, allowing researchers to reduce the number of experimental tests they need.

“We can reduce the number of errors when developing new antibodies or drug candidates to target pathogens and cancers,” says Minnegalieva. “The models we use must be accurate and reliable. We really need to understand what is important from a biological standpoint. Only then can we make healthy predictions and save time.”

The new study outlines how models can be trained to better meet these requirements.

Although this study focused on antibodies, the results can be broadly generalized in a variety of areas where machine learning is applied.

“Fields such as language modeling, protein design, and prediction of molecular properties also depend on sampling negative data. All of these areas face the risk of models taking shortcuts if the negative examples are too simple,” concludes Minnegalieva.

Professor Victor Grafe, director of Grief Lavo, also highlighted the relevance and potential impact of the research. “Our work shows that curation of data is not a preprocessing step. It's a scientific choice that codes assumptions and determines what machine learning can discover.

“In immunology, drug discovery, and beyond, careful dataset design may be key to building machine learning models that generalize and uncover true biological principles,” Graif says.