AI

The experiment demonstrates that natural language processing, although it was developed to read and write language text, can learn at least some of the underlying principles of biology. Salesforce Research developed the AI program, called ProGen, which uses next-token prediction to assemble amino acid sequences into artificial proteins.

Scientists said the new technology could become more powerful than directed evolution, the Nobel-prize winning protein design technology, and it will energize the 50-year-old field of protein engineering by speeding the development of new proteins that can be used for almost anything from therapeutics to degrading plastic.

“The artificial designs perform much better than designs that were inspired by the evolutionary process,” said James Fraser, PhD, professor of bioengineering and therapeutic sciences at the UCSF School of Pharmacy, and an author of the work, which was published Jan. 26, in Nature Biotechnology.

“The language model is learning aspects of evolution, but it’s different than the normal evolutionary process,” Fraser said. “We now have the ability to tune the generation of these properties for specific effects. For example, an enzyme that’s incredibly thermostable or likes acidic environments or won’t interact with other proteins.”

To create the model, scientists simply fed the amino acid sequences of 280 million different proteins of all kinds into the machine learning model and let it digest the information for a couple of weeks. Then, they fine-tuned the model by priming it with 56,000 sequences from five lysozyme families, along with some contextual information about these proteins.

The model quickly generated a million sequences, and the research team selected 100 to test, based on how closely they resembled the sequences of natural proteins, as well how naturalistic the AI proteins’ underlying amino acid “grammar” and “semantics” were.

Out of this first batch of a 100 proteins, which were screened in vitro by Tierra Biosciences, the team made five artificial proteins to test in cells and compared their activity to an enzyme found in the whites of chicken eggs, known as hen egg white lysozyme (HEWL). Similar lysozymes are found in human tears, saliva and milk, where they defend against bacteria and fungi.

Two of the artificial enzymes were able to break down the cell walls of bacteria with activity comparable to HEWL, yet their sequences were only about 18% identical to one another. The two sequences were about 90% and 70% identical to any known protein.

Just one mutation in a natural protein can make it stop working, but in a different round of screening, the team found that the AI-generated enzymes showed activity even when as little as 31.4% of their sequence resembled any known natural protein.

The AI was even able to learn how the enzymes should be shaped, simply from studying the raw sequence data. Measured with X-ray crystallography, the atomic structures of the artificial proteins looked just as they should, although the sequences were like nothing seen before.

Salesforce Research developed ProGen in 2020, based on a kind of natural language programming their researchers originally developed to generate English language text.

They knew from their previous work that the AI system could teach itself grammar and the meaning of words, along with other underlying rules that make writing well-composed.

“When you train sequence-based models with lots of data, they are really powerful in learning structure and rules,” said Nikhil Naik, PhD, Director of AI Research at Salesforce Research, and the senior author of the paper. “They learn what words can co-occur, and also compositionality.”

With proteins, the design choices were almost limitless. Lysozymes are small as proteins go, with up to about 300 amino acids. But with 20 possible amino acids, there are an enormous number (20³⁰⁰) of possible combinations. That’s greater than taking all the humans who lived throughout time, multiplied by the number of grains of sand on Earth, multiplied by the number of atoms in the universe.

Given the limitless possibilities, it’s remarkable that the model can so easily generate working enzymes.

“The capability to generate functional proteins from scratch out-of-the-box demonstrates we are entering into a new era of protein design,” said Ali Madani, PhD, founder of Profluent Bio, former research scientist at Salesforce Research, and the paper’s first author. “This is a versatile new tool available to protein engineers, and we’re looking forward to seeing the therapeutic applications.”

Further information: https://github.com/salesforce/progen

Professor Hwangbo’s research team developed a technology to model the force received by a walking robot on the ground made of granular materials such as sand and simulate it via a quadrupedal robot. Also, the team worked on an artificial neural network structure which is suitable in making real-time decisions needed in adapting to various types of ground without prior information while walking at the same time and applied it on to reinforcement learning. The trained neural network controller is expected to expand the scope of application of quadrupedal walking robots by proving its robustness in changing terrain, such as the ability to move in high-speed even on a sandy beach and walk and turn on soft grounds like an air mattress without losing balance.

This research, with Ph.D. Student Soo-Young Choi of KAIST Department of Mechanical Engineering as the first author, was published in January in the Science Robotics. (Paper title: Learning quadrupedal locomotion on deformable terrain).

Reinforcement learning is an AI learning method used to create a machine that collects data on the results of various actions in an arbitrary situation and utilizes that set of data to perform a task. Because the amount of data required for reinforcement learning is so vast, a method of collecting data through simulations that approximates physical phenomena in the real environment is widely used.

In particular, learning-based controllers in the field of walking robots have been applied to real environments after learning through data collected in simulations to successfully perform walking controls in various terrains.

However, since the performance of the learning-based controller rapidly decreases when the actual environment has any discrepancy from the learned simulation environment, it is important to implement an environment similar to the real one in the data collection stage. Therefore, in order to create a learning-based controller that can maintain balance in a deforming terrain, the simulator must provide a similar contact experience.

The research team defined a contact model that predicted the force generated upon contact from the motion dynamics of a walking body based on a ground reaction force model that considered the additional mass effect of granular media defined in previous studies.

Furthermore, by calculating the force generated from one or several contacts at each time step, the deforming terrain was efficiently simulated.

The research team also introduced an artificial neural network structure that implicitly predicts ground characteristics by using a recurrent neural network that analyzes time-series data from the robot’s sensors.

The learned controller was mounted on the robot ‘RaiBo’, which was built hands-on by the research team to show high-speed walking of up to 3.03 m/s on a sandy beach where the robot’s feet were completely submerged in the sand. Even when applied to harder grounds, such as grassy fields, and a running track, it was able to run stably by adapting to the characteristics of the ground without any additional programming or revision to the controlling algorithm.

In addition, it rotated with stability at 1.54 rad/s (approximately 90° per second) on an air mattress and demonstrated its quick adaptability even in the situation in which the terrain suddenly turned soft.

The research team demonstrated the importance of providing a suitable contact experience during the learning process by comparison with a controller that assumed the ground to be rigid, and proved that the proposed recurrent neural network modifies the controller’s walking method according to the ground properties.

The simulation and learning methodology developed by the research team is expected to contribute to robots performing practical tasks as it expands the range of terrains that various walking robots can operate on.

The first author, Suyoung Choi, said, “It has been shown that providing a learning-based controller with a close contact experience with real deforming ground is essential for application to deforming terrain.” He went on to add that “The proposed controller can be used without prior information on the terrain, so it can be applied to various robot walking studies.”

This research was carried out with the support of the Samsung Research Funding & Incubation Center of Samsung Electronics.

Where traditional robots are hard-bodied and stiff, “soft” robots have the opposite problem; they are flexible but weak, and their movements are difficult to control. “Giving robots the ability to switch between liquid and solid states endows them with more functionality,” says Chengfeng Pan, an engineer at The Chinese University of Hong Kong who led the study.

The team created the new phase-shifting material — dubbed a “magnetoactive solid-liquid phase transitional machine” — by embedding magnetic particles in gallium, a metal with a very low melting point (29.8 °C).

“The magnetic particles here have two roles,” says senior author and mechanical engineer Carmel Majidi of Carnegie Mellon University. “One is that they make the material responsive to an alternating magnetic field, so you can, through induction, heat up the material and cause the phase change. But the magnetic particles also give the robots mobility and the ability to move in response to the magnetic field.”

This is in contrast to existing phase-shifting materials that rely on heat guns, electrical currents, or other external heat sources to induce solid-to-liquid transformation. The new material also boasts an extremely fluid liquid phase compared to other phase-changing materials, whose “liquid” phases are considerably more viscous.

Before exploring potential applications, the team tested the material’s mobility and strength in a variety of contexts. With the aid of a magnetic field, the robots jumped over moats, climbed walls, and even split in half to cooperatively move other objects around before coalescing back together. In one video, a robot shaped like a person liquifies to ooze through a grid after which it is extracted and remolded back into its original shape.

“Now, we’re pushing this material system in more practical ways to solve some very specific medical and engineering problems,” says Pan.

On the biomedical side, the team used the robots to remove a foreign object from a model stomach and to deliver drugs on-demand into the same stomach. They also demonstrate how the material could work as smart soldering robots for wireless circuit assembly and repair (by oozing into hard-to-reach circuits and acting as both solder and conductor) and as a universal mechanical “screw” for assembling parts in hard-to-reach spaces (by melting into the threaded screw socket and then solidifying; no actual screwing required.)

“Future work should further explore how these robots could be used within a biomedical context,” says Majidi. “What we’re showing are just one-off demonstrations, proofs of concept, but much more study will be required to delve into how this could actually be used for drug delivery or for removing foreign objects.”