Tyler O'Neal, Staff Editor PROCESSORS

German biologists simulate phage capsid against flu on supercomputer; perfectly fitting inhibitor prevents viral infection

A new approach brings the hope of new therapeutic options for suppressing seasonal influenza and avian flu: On the basis of an empty – and therefore non-infectious – the shell of a phage virus, researchers from Berlin have developed a chemically modified phage capsid that “stifles” influenza viruses. Perfectly fitting binding sites cause influenza viruses to be enveloped by the phage capsids in such a way that it is practically impossible for them to infect lung cells any longer. This phenomenon has been proven in pre-clinical trials, also involving human lung tissue. Researchers from the Leibniz-Forschungsinstitut für Molekulare Pharmakologie (FMP), Freie Universität Berlin (FU), Technische Universität Berlin (TU), Humboldt-Universität (HU), the Robert Koch Institute (RKI) and Charité were involved in this groundbreaking work. The results are also being used for the immediate investigation of the coronavirus. The findings have now been published in an academic journal.

Influenza viruses are still highly dangerous: The World Health Organization (WHO) estimates that flu is responsible for up to 650,000 deaths per year worldwide. Current antiviral drugs are only partially effective because they attack the influenza virus after lung cells have been infected. It would be desirable – and much more effective – to prevent infection in the first place. 

This is exactly what the new approach from Berlin promises. The phage capsid, developed by a multidisciplinary team of researchers, envelops flu viruses so perfectly that they can no longer infect cells. “Pre-clinical trials show that we are able to render harmless both seasonal influenza viruses and avian flu viruses harmless with our chemically modified phage shell,” explained Professor Dr. Christian Hackenberger, Head of the Department Chemical Biology at the Leibniz-Forschungsinstitut für Molekulare Pharmakologie (FMP) and Leibniz-Humboldt Professor for Chemical Biology at HU Berlin. “It is a major success that offers entirely new perspectives for the development of innovative antiviral drugs.”  preview f7309{module INSIDE STORY}

Multiple bonds fit like hook-and-loop tape

The new inhibitor makes use of a feature that all influenza viruses have: There are trivalent receptors on the surface of the virus, referred to as hemagglutinin protein, that attaches to sugar molecules (sialic acids) on the cell surface of lung tissue. In the case of infection, viruses hook into their victim – in this case, lung cells – like a hook-and-loop fastener. The core principle is that these interactions occur due to multiple bonds, rather than single bonds. 

It was the surface structure of flu viruses that inspired the researchers to ask the following initial question more than six years ago: Would it not be possible to develop an inhibitor that binds to trivalent receptors with a perfect fit, simulating the surface of lung tissue cells? 

We now know that this is indeed possible – with the help of a harmless intestinal inhabitant: The Q-beta phage has the ideal surface properties and is excellently suited to equip it with ligands – in this case, sugar molecules – as “bait”. An empty phage shell does the job perfectly. “Our multivalent scaffold molecule is not infectious, and comprises 180 identical proteins that are spaced out exactly as the trivalent receptors of the hemagglutinin on the surface of the virus,” explained Dr. Daniel Lauster, a former Ph.D. student in the Group of Molecular Biophysics (HU) and now a postdoc at FU Berlin. “It, therefore, has the ideal starting conditions to deceive the influenza virus – or, to be more precise, to attach to it with a perfect spatial fit. In other words, we use a phage virus to disable the influenza virus!”

To enable the Q-beta scaffold to fulfill the desired function, it must first be chemically modified. Produced from E. coli bacteria at TU Berlin, Professor Hackenberger’s research group at FMP and HU Berlin use synthetic chemistry to attach sugar molecules to the defined positions of the virus shell. 

The virus is deceived and enveloped

Several studies using animal models and cell cultures have proven that the suitably modified spherical structure possesses considerable bond strength and inhibiting potential. The study also enabled the Robert Koch Institute to examine the antiviral potential of phage capsids against many current influenza virus strains, and even against avian flu viruses. Its therapeutic potential has even been proven on human lung tissue, as fellow researchers from the Medical Department, Division of Infectiology and Pneumology, of Charité were able to show: When tissue infected with flu viruses was treated with the phage capsid, the influenza viruses were practically no longer able to reproduce. 

The results are supported by structural proof furnished by FU scientists from the Research Center of Electron Microscopy (FZEM): High-resolution cryo-electron microscopy and cryo-electron microscopy show directly and, above all, spatially, that the inhibitor completely encapsulates the virus. In addition, mathematical-physical models were used to simulate the interaction between influenza viruses and the phage capsid on the supercomputer. “Our computer-assisted calculations show that the rationally designed inhibitor does indeed attach to the hemagglutinin, and completely envelops the influenza virus,” confirmed Dr. Susanne Liese from the AG Netz of FU Berlin. “It was therefore also possible to describe and explain the high bond strength mathematically.”

Therapeutic potential requires further research

These findings must now be followed up by more preclinical studies. It is not yet known, for example, whether the phage capsid provokes an immune response in mammals. Ideally, this response could even enhance the effect of the inhibitor. However, it could also be the case that an immune response reduces the efficacy of phage capsids in the case of repeated-dose exposure, or that flu viruses develop resistances. And, of course, it has yet to be proven that the inhibitor is also effective in humans. 

Nonetheless, the alliance of Berlin researchers is certain that the approach has great potential. “Our rationally developed, three-dimensional, multivalent inhibitor points to a new direction in the development of structurally adaptable influenza virus binders. This is the first achievement of its kind in multivalency research,” emphasized Professor Hackenberger. The chemist believes that this approach, which is biodegradable, non-toxic and non-immunogenic in cell culture studies, can in principle also be applied to other viruses, and possibly also to bacteria. It is evident that the authors regard the application of their approach to the current coronavirus as one of their new challenges. The idea is to develop a drug that prevents coronaviruses from binding to host cells located in the throat and subsequent airways, thus preventing infection.

Berlin university alliance at its best

Cooperation between scientists from different disciplines played a major role in the discovery of the new influenza inhibitor. Biologists, chemists, physicists, virologists, medical scientists and imaging specialists from three Berlin universities HU, FU and TU, the Robert Koch Institute, Charité and, last but not least, FMP were all involved in the project. “In my opinion, such a complex project could only have been undertaken in Berlin, where there truly are experts for every issue,” stated Professor Dr. Andreas Herrmann, Head of Molecular Biophysics at HU Berlin. “It was the Berlin university alliance at its best,” he added, “and I hope that the follow-up studies will be equally successful.”

The project was funded within Collaborative Research Center 765 (Speaker Professor Dr. Rainer Haag, FU Berlin) of the German Research Foundation (DFG). 

Army researchers measure reliability, confidence for next-gen AI

Tyler O'Neal, Staff Editor PROCESSORS

A team of Army and industry researchers have developed a metric for neural networks--computing systems modeled loosely after the human brain--that could assess the reliability and confidence of the next generation of artificial intelligence and machine learning algorithms.

Deep neural network, or DNNs, are a form of machine learning that use training data to learn. Once trained, they can make predictions when given new information or inputs; however, they can be easily deceived if the new information is too far outside its training.

Researchers said given the diversity of information in training data and potential new inputs, coming up with a solution is challenging.

"This opens a new research opportunity to create the next generation of algorithms that are robust and resilient," said Dr. Brian Jalaian, a scientist at the U.S. Army Combat Capabilities Development Command's Army Research Laboratory. "Our approach is versatile and can be added as an additional block to many of the Army's modern algorithms using modern machine learning algorithms that are based on deep neural networks used for visual imagery." CAPTION The researchers observe a similar difference in conformance for out-of-distribution examples and the adversarial examples, which motivates the use of conformance in the attribution-neighborhood as a confidence metric.{module INSIDE STORY}

This new confidence metric will help the Army create safe and secure machine learning techniques, and will apply in command and control systems, precision fire and decision support systems, Jalaian said.

Since 2018, researchers from the Army and SRI International, through the lab's Internet of Battlefield Things Collaborative Research Alliance, have investigated methods to harden Army's machine learning algorithms to provide greater dependability and safety, and be less susceptible adversarial machine learning techniques.

The researchers published their work, "Attribution-Based Confidence Metric for Deep Neural Networks", at the 2019 Neural Information Processing Systems Conference.

"While we had some success, we did not have an approach to detect the strongest state-of-the-art attacks such as (adversarial) patches that add noise to imagery, such that they lead to incorrect predictions," Jalaian said. "In this work, we proposed a generative model, which adjusts aspects of the original input images in the underlying original deep neural network. The original deep neural network's response to these generated inputs are then assessed to measure the conformance of the model."

This differs from the existing body of research, as it does not require access to the training data, the use of ensembles or the need to train a calibration model on a validation dataset that is not the same as the training set, Jalaian said.

Within the Army, researchers continue to work with the test and evaluation community to develop containerized algorithms that measure the confidence of various algorithms across different applications.

Jalaian said they are exploring variations of generative models that could harden Army AI systems against adversarial manipulations, as well as investigating the resiliency of neural network models, both theoretically and empirically, that could be executed within small smart devices, such as those that would be part of the Internet of Battlefield Things.

The Army continues to move forward with its modernization priorities, which place a high value on next-generation cyber solutions, which will enable the Army to deliver technology capabilities to warfighters.

Japanese machine learning puts a new spin on spin models

Tyler O'Neal, Staff Editor PROCESSORS

New insights into phase transitions using artificial intelligence

Researchers from Tokyo Metropolitan University have used machine learning to study spin models, used in physics to study phase transitions. Previous work showed that image/handwriting classifying AI could be applied to distinguish states in the simplest models. The team showed the approach is applicable to more complex models and found that an AI trained on one model and applied to another could reveal key similarities between distinct phases in different systems.

Machine learning and artificial intelligence (AI) are revolutionizing how we live, work, play, and drive. The self-driving car, the algorithm that beat a go grandmaster and advances in finance are just the tip of the iceberg of a wide range of applications which are having a significant impact on society. AI is also making waves in scientific research. A key attraction of these algorithms is how they can be trained with pre-classified data (e.g. images of handwritten letters) and be applied to classify a much wider range of data. CAPTION Simulated low temperature (left) and high temperature (right) phase of a 2D Ising model, where blue points are spins pointing up, and the red points are spins pointing down. Notice that the spins in the low temperature phase are mostly in the same direction. This is called a ferromagnetic phase. On the other hand, at high temperature, the ratio of up to down spins is closer to 50:50. This is called a paramagnetic phase.{module INSIDE STORY}

In the field of condensed matter physics, recent work by Carrasquilla and Melko (Nature Physics (2017) 13, 431-434) has shown that the same kind of AI used to interpret handwriting, neural networks, could be used to distinguish different phases of matter (e.g. gas, liquid and solid) in simple physical models. They studied the Ising model, the simplest model for the emergence of magnetism in materials. A lattice of atoms with a spin (up or down) has an energy that depends on the relative alignment of adjacent spins. Depending on the conditions, these spins can line up into a ferromagnetic phase (like iron) or assume random directions in a paramagnetic phase. Usually, studies of this kind of system involve analyzing some averaged quantity (e.g. sum of all the spins). The fact that an entire microscopic configuration can be used to classify a phase presented a genuine paradigm shift. CAPTION The input (correlation configurations) is fed into a system of interconnected nodes known as a neural network, giving a series of outputs telling us which phase the configuration belongs to. During training, the algorithm is told whether the outputs are right or wrong, and the network is adjusted over and over again to get better agreement i.e. it learns.

Now, a team led by Professors Hiroyuki Mori and Yutaka Okabe of Tokyo Metropolitan University are collaborating with the Bioinformatics Institute in Singapore to take this approach to the next level. In its existing form, the method of Carrasquilla and Melko cannot be applied to more complex models than the Ising model. Take the q-state Potts model, where atoms can take one of q states instead of just "up" or "down". Though it also has a phase transition, telling the phases apart is not trivial. In fact, if we consider a 5-state model, there are 120 states which are physically equivalent. To help an AI tell the phases apart, the team gave it more microscopic information, specifically how the state of a particular atom relates to the state of another atom some distance away, or how the spins correlate over the separation. Having trained the AI with many of these correlation configurations for 3 and 5-state Potts models, they found that it was able to correctly classify phases and identify the temperature where the transition took place. They were also able to correctly account for the number of points in their lattice, the finite-size effect. CAPTION The low and high temperature phases are found in the right proportions at different temperatures relative to the transition point for different sizes of lattice. (inset) The size of the lattice may be accounted for to give a single master curve.{module INSIDE STORY}

Having demonstrated that their method works, they tried the same approach on a q-state clock model, where spins adopt one of q orientations on a circle. When q is greater than or equal to 5, there are three phases which the system can take: an ordered low-temperature phase, a high-temperature phase, and a phase in between known as the Berezinskii-Kosterlitz-Thouless (BKT) phase, the investigation of which won John M. Kosterlitz, David J. Thouless, and Duncan Haldane the 2016 Nobel Prize for Physics. They proceeded to successfully train an AI to tell the three phases apart with a 6-state clock model. When they applied it to configurations from a 4-state clock model, where there are only two phases expected, they discovered that the algorithm could classify the system as being in a BKT phase near the phase transition. This goes to show there is a deep connection between the BKT phase and the critical phase arising at the smooth 'second-order' phase transition point in the 4-state system.

The method presented by the team is generally applicable to a wide range of scientific problems. A key part of physics is universality, identifying traits in seemingly unrelated systems or phenomena which give rise to unified behavior. Machine learning is uniquely placed to tease these features out of the most complex models and systems, letting scientists take a peek at the deep connections that govern nature and our universe.