Breast cancer is the most common cancer with the highest mortality rate. Swift detection and diagnosis diminish the impact of the disease. However, classifying breast cancer using histopathology images — tissues and cells examined under a microscope — is a challenging task because of bias in the data and the unavailability of annotated data in large quantities. Automatic detection of breast cancer using convolutional neural network (CNN), a machine learning technique, has shown promise — but it is associated with a high risk of false positives and false negatives.

Without any measure of confidence, such false predictions of CNN could lead to catastrophic outcomes. But a new machine learning model developed by Michigan Technological University researchers can evaluate the uncertainty in its predictions as it classifies benign and malignant tumors, helping reduce this risk. 

In their study, mechanical engineering graduate students Ponkrshnan Thiagarajan and Pushkar Kharinar and Susanta Ghosh, assistant professor of mechanical engineering and machine learning expert, outline their novel probabilistic machine learning model, which outperforms similar models.

“Any machine learning algorithm that has been developed so far will have some uncertainty in its prediction,” Thiagarajan said. “There is little way to quantify those uncertainties. Even if an algorithm tells us a person has cancer, we do not know the level of confidence in that prediction.”

From Experience Comes Confidence

In the medical context, not knowing how confident an algorithm is has made it difficult to rely on computer-generated predictions. The present model is an extension of the Bayesian neural network — a machine learning model that can evaluate an image and produce an output. The parameters for this model are treated as random variables that facilitate uncertainty quantification. 

The Michigan Tech model differentiates between negative and positive classes by analyzing the images, which at their most basic level are collections of pixels. In addition to this classification, the model can measure the uncertainty in its predictions.

In a medical laboratory, such a model promises time savings by classifying images faster than a lab tech. And, because the model can evaluate its level of certainty, it can refer the images to a human expert when it is less confident.

But why is a mechanical engineer creating algorithms for the medical community? Thiagarajan’s idea kindled when he started using machine learning to reduce the computational time needed for mechanical engineering problems. Whether a computation evaluates the deformation of building materials or determines whether someone has breast cancer, it’s important to know the uncertainty of that computation — the key ideas remain the same.

“Breast cancer is one of the cancers that have the highest mortality and highest incidence,” Thiagarajan said. “We believe that this is an exciting problem wherein better algorithms can make an impact on people’s lives directly.”

Next Steps

Now that their study has been published, the researchers will extend the model for multiclass classification of breast cancer. They will aim to detect cancer subtypes in addition to classifying benign and malignant tissues. And the model, though developed using breast cancer histopathology images, can also be extended for other medical diagnoses.

“Despite the promise of machine learning-based classification models, their predictions suffer from uncertainties due to the inherent randomness and the bias in the data and the scarcity of large datasets,” Ghosh said. “Our work attempts to address these issues and quantifies, uses, and explains the uncertainty.”

Ultimately, Thiagarajan, Khairnar, and Ghosh’s model itself — which can evaluate whether images have high or low measures of uncertainty and identify when images need the eyes of a medical expert — represents the next steps in the endeavor of machine learning.

Researchers show that a neural network trained exclusively to predict protein shapes can also generate new ones

Just as convincing images of cats can be created using artificial intelligence, new proteins can now be made using similar tools. In a study, a team including researchers at the University of Washington in Seattle, Washington, Rensselaer Polytechnic Institute, and Harvard University describe the development of a neural network that “hallucinates” proteins with new, stable structures. 

“The potential to hallucinate brand-new proteins that bind particular biomolecules or form desired enzymatic active sites is very exciting,” said Gaetano Montelione, a professor of chemistry and chemical biology at Rensselaer, where synthesized versions of “hallucinated” proteins invented by a neural network were analyzed. 

Proteins are string-like molecules found in every cell that spontaneously fold into intricate three-dimensional shapes. These folded shapes are key to nearly every process in biology, including cellular development, DNA repair, and metabolism. But the complexity of protein shapes makes them difficult to study. Biochemists often use supercomputers to predict how protein strings, or sequences, might fold. In recent years, artificial intelligence techniques like neural networks and deep learning have revolutionized the accuracy of this work.

“For this project, we made up completely random protein sequences and introduced mutations into them until our neural network predicted that they would fold into stable structures,” said co-lead author Ivan Anishchenko, a postdoctoral scholar in the Baker lab in the Institute for Protein Design at the University of Washington School of Medicine. “At no point did we guide the software toward a particular outcome — these new proteins are just what a computer dreams up.”

In the future, the team believes it should be possible to steer artificial intelligence so that it generates new proteins with useful features. “We’d like to use deep learning to design proteins with function, including protein-based drugs, enzymes, you name it,” said co-lead author Sam Pellock, a postdoctoral scholar in the Baker lab.

The research team generated 2,000 new protein sequences that were predicted to fold. Over 100 of these were produced in the laboratory and studied. Detailed analysis on three such proteins confirmed that the shapes predicted by the supercomputer were indeed realized in the lab.

“Our solution NMR studies, along with X-ray crystal structures determined by the University of Washington team, demonstrate the remarkable accuracy of protein designs created by the hallucination approach,” said co-author Theresa Ramelot, a senior research scientist in the Montelione lab within the Rensselaer Center for Biotechnology and Interdisciplinary Studies.  

Montelione notes, “The hallucination approach builds on earlier observations we made together with the Baker lab revealing that protein structure prediction with deep learning can be quite accurate even for a single protein sequence, without recourse to contact predictions usually obtained by analysis of many evolutionary-related protein sequences.”  

“This approach greatly simplifies protein design,” said senior author David Baker, recipient of the 2021 Breakthrough Prize in Life Sciences. “Before, to create a new protein with a particular shape, people first carefully studied related structures in nature to come up with a set of rules that were then applied in the design process. New sets of rules were needed for each new type of fold. Here, by using a deep-learning network that already captures general principles of protein structure, we eliminate the need for fold-specific rules and open up the possibility of focusing on just the functional parts of a protein directly.” 

“Exploring how to best use this strategy for specific applications is now an active area of research, and this is where I expect the next breakthroughs,” said Baker.