The new ML technique quickly calculates the binding affinities between drug candidates and their targets.

Artificial intelligence and machine learning techniques are already proving effective in pharmaceutical procedures. Drug discovery is one of the crucial procedures to find new candidate medications in the field of medicine, biotechnology and pharmacology. According to the U.S. FDA, there are five steps for the development of a new drug. These include discovery and development, preclinical research, clinical research, FDA review, and FDA post-market safety monitoring. Since drug discovery requires huge amounts of data and research, many pharmaceutical companies are embracing AI and machine learning to accelerate the pace of drug discovery.

AI and ML techniques can also lower the costs of drug development. Drug discovery is a data-driven process. It involves a voluminous amount of data such as high-resolution medical images, genomic profiles, metabolites, molecular structures, and biological information. Machine learning and deep learning-fuelled artificial intelligence can correlate, integrate, and connect existing data more rapidly to help discover patterns in the data pools.

As drugs can only work based on their stickiness to their target proteins in the body, analyzing that stickiness is a key hurdle in the drug discovery and screening process. New research combining chemistry and machine learning could lower that hurdle. The new technique, called DeepBAR, can quickly calculate the binding affinities between drug candidates and their targets. DeepBAR combines traditional chemistry calculations with recent advances in machine learning. It computes binding free energy exactly, but it requires just a fraction of the calculations demanded by previous methods.

The “BAR” in DeepBAR stands for “Bennett acceptance ratio”. It is a decades-old algorithm used in exact calculations of binding free energy. According to the researchers, DeepBAR could one day quicken the pace of drug discovery and protein engineering.

The research has appeared in the Journal of Physical Chemistry Letters and led by Xinqiang Ding, a postdoc in MIT’s Department of Chemistry.

As per the study, using the Bennet acceptance ratio typically requires knowledge of two “endpoint” states. A drug molecule bound to a protein and a drug molecule completely dissociated from a protein, plus knowledge of many intermediate states, e.g., varying levels of partial binding, all of which bog down calculation speed.

The new machine learning technique slashes those in-between states by implementing the Bennett acceptance ratio in machine learning frameworks called deep generative models. These models create a reference state for each endpoint, the bound state and the unbound state, according to Bin Zhang, the Pfizer-Laubach Career Development Professor in Chemistry at MIT, and a co-author of a new paper describing the technique.

In using deep generative models, the researchers were borrowing from the field of computer vision. Though adapting a computer vision approach to chemistry was DeepBAR’s key innovation, the crossover also raised some challenges. “These models were originally developed for 2D images,” says Xinqiang Ding. “But here we have proteins and molecules—it’s really a 3D structure. So, adapting those methods in our case was the biggest technical challenge we had to overcome.”

In tests using small protein-like molecules, DeepBAR calculated binding free energy nearly 50 times faster than previous methods. The researchers then start thinking about using this to do drug screening, particularly in the context of COVID. “DeepBAR has the exact same accuracy as the gold standard, but it’s much faster,” says Zhang. They also believe that in addition to drug screening, DeepBAR could aid protein design and engineering, since the method could be used to model interactions between multiple proteins. They also plan to improve the ability of the new machine learning technique in the future to run calculations for large proteins, a task made feasible by recent advances in computer science.

The post HOW CAN MACHINE LEARNING ACCELERATE THE PACE OF DRUG DISCOVERY? appeared first on Artificial Intelligence.

Machine-learning techniques, best known for teaching self-driving cars to stop at red lights, may soon help researchers around the world improve their control over the most complicated reaction known to science: nuclear fusion.

Fusion reactions are typically hydrogen atoms heated to form a gaseous cloud called a plasma that releases energy as the particles bang into each other and fuse. Getting these reactions under better control could create huge amounts of environmentally clean energy from nuclear reactors in fusion power plants of the future.

“The connection between machine learning and fusion energy is not obvious,” said Sandia National Laboratories researcher Aidan Thompson, principal investigator for a three-year Department of Energy Office of Science award of $2.2 million annually to make that very connection. “Simply put, we have pioneered machine-learning’s use to improve simulations of the reactor’s wall material as it interacts with plasma. This has been beyond the scope of atomic-scale simulations of the past.”

The expected result should suggest procedural or structural modifications to improve nuclear energy output, he said.

Power of machine learning in modeling nuclear fusion

Machine learning is powerful because it uses mathematical and statistical means to figure out a situation, rather than analyze every piece of data in the desired category. For example, only a small number of dog photos are needed to teach a recognition system the concept of “dogginess”— in other words, “this is a dog”—rather than scanning every dog photo in existence.

Sandia’s machine-learning approach to nuclear fusion is the same, but more complicated.

“It is not a trivial problem to physically observe what is going on within a reactor’s walls as these structures are internally bombarded with hydrogen, helium, deuterium and tritium as parts of a super-heated plasma,” said Thompson.

He described components of the circling plasma striking and altering the composition of the retaining walls and heavy atoms dislodging from the struck walls and altering the plasma. Reactions take place in nanoseconds at temperatures as hot as the sun. Trying to modify components using trial and error to improve outcomes is extraordinarily laborious.

Machine-learning algorithms, on the other hand, use computer-generated data without direct measurements from experiments and can yield information that eventually could be used to make plasma interactions with containment wall material less damaging and thus improve the overall energy output of fusion reactors.

“There is no other way of getting this information,” said Thompson.

Small number of atoms predict the energy of many

Thompson’s team expects that by using large datasets of quantum-mechanics calculations under extreme conditions as training data, they can build a machine-learning model that predicts the energy of any configuration of atoms.

This model, called a machine-learning interatomic potential, or MLIAP, can be inserted into huge classical molecular dynamics codes such as Sandia’s award-winning LAMMPS, or Large-scale Atomic/Molecular Massively Parallel Simulator, software. In this way, by interrogating only a relatively small number of atoms, they can extend the accuracy of quantum mechanics to the scale of millions of atoms needed to simulate the behavior of fusion energy materials.

“So why is what we are doing machine learning and not just bookkeeping lots of data?” asks Thompson rhetorically. “The short answer is, we generate equations from an infinite set of possible variables to build models that are grounded in physics but contain hundreds or thousands of parameters that keep us within range of our target.”

One catch is that the accuracy of the MLIAP model depends on the overlap between the training data and the actual atomic environments encountered by the application, said Thompson.

These environments may be various, requiring new training data and alteration of the machine-learning model. Recognizing and adjusting for overlaps is part of the work of the next few years.

“Our model at first will be used to interpret small experiments,” Thompson said. “Conversely, that experimental data will be used to validate our model, which can then be used to make predictions about what is happening in a full-scale fusion reactor.”

The target for giving fusion researchers access to the Sandia machine-learning models to build better fusion reactors is approximately three years, said Thompson.

The post Machine-learning technique could improve fusion energy outputs appeared first on Artificial Intelligence.

In reinforcement learning, the goal generally is to spur an AI-driven agent to complete tasks via systems of rewards. This is achieved either by learning a mapping (a policy) from states to actions that maximize an expected return (policy gradients), or by inferring such a mapping by calculating the expected return for a given state-action pair.

Model-based reinforcement learning (MBRL) aims to improve this by learning a model of the dynamics from an agent’s interactions with the environment that can be leveraged across many different tasks (aka transferability) and used for planning. To this end, researchers at Google, the University of Oxford, and UC Berkeley developed an approach — Ready Policy One (a not-so-subtle nod to Ernest Cline’s hit novel Ready Player One) — to acquiring data for training world models through exploration that jointly optimizes policies for both reward and model uncertainty reduction. The end result is that the policies leveraged for data collection also perform well in the true environment and can be tapped for evaluation.

Ready Policy One takes an active learning approach rather than focusing on optimization. In other words, it seeks to directly learn the best model rather than learning the best policy. A tailored framework allows Ready Policy One to adapt the level of exploration to improve the model in the fewest number of samples, and a mechanism stops gathering new samples in any given collection phase when the incoming data resembles what’s already been acquired.

In a series of experiments, the researchers evaluated whether their active learning approach for MBRL was more sample-efficient than existing approaches. In particular, they tested it on a range of continuous control tasks from research firm OpenAI’s Gym environment, and they found that Ready Policy One could lead to “state-of-the-art” efficiency when combined with the latest model architectures.

“We are particularly excited by the many future directions from this work,” wrote the study’s coauthors. “Most obviously, since our method is orthogonal to other recent advances in MBRL, [Ready Policy One] could be combined with state of the art probabilistic architectures … In addition, we could take a hierarchical approach by ensuring our exploration policies maintain core behaviors but maximize entropy in some distant unexplored region. This would require behavioral representations, and some notion of distance in behavioral space, and may lead to increased sample efficiency as we could better target specific state-action pairs.”

The post Researchers develop technique to increase sample efficiency in reinforcement learning appeared first on Artificial Intelligence.

A team of researchers from DeepMind, University College and Harvard University has found that lessons learned in applying learning techniques to AI systems may help explain how reward pathways work in the brain. In their paper published in the journal Nature, the group describes comparing distributional reinforcement learning in a computer with dopamine processing in the mouse brain, and what they learned from it.

Prior research has shown that dopamine produced in the brain is involved in reward processing—it is produced when something good happens, and its expression results in feelings of pleasure. Some studies have also suggested that the neurons in the brain that respond to the presence of dopamine all respond in the same ways—an event causes a person or a mouse to feel either good or bad. Other studies have suggested that neuronal response is more of a gradient. In this new effort, the researchers have found evidence supporting the latter theory.

Distributional reinforcement learning is a type of machine learning based on reinforcement. It is often used when designing games such as Starcraft II or Go. It keeps track of good moves versus bad moves and learns to reduce the number of bad moves, improving its performance the more it plays. But such systems do not treat all good and bad moves the same—each move is weighted as it is recorded and the weights are part of the calculations used when making future move choices.

Researchers have noted that humans appear to use a similar strategy to improve their level of play, as well. The researchers in London suspected that the similarities between the AI systems and the way the brain carries out reward processing were likely similar, as well. To find out if they were correct, they carried out experiments with mice. They inserted devices into their brains that were capable of recording responses from individual dopamine neurons. The mice were then trained to carry out a task in which they received rewards for responding in a desired way.

The mouse neuron responses revealed that they did not all respond the same way, as prior theory had predicted. Instead, they responded in reliably different ways—an indication that the levels of pleasure the mice were experiencing were more of a gradient, as the team had predicted.

The post AI learning technique may illustrate function of reward pathways in the brain appeared first on Artificial Intelligence.

Deep learning is capturing the attention of all of us as it is accomplishing outcomes that were not previously possible.  Deep learning is a machine learning technique that teaches computers to learn by example just as we learned as a child. We see this technology in autonomous vehicles. It enables the vehicle to distinguish between different objects on the road and enables the vehicle to stop when it sees a red light. An autonomous vehicle can determine when it is safe to move forward or to remain stationary.

In deep learning, a computer becomes proficient at performing tasks from images, text, or sound, and can realize state-of-the-art accuracy, many times exceeding human implementation.

We often hear the terms: AI (artificial intelligence), machine learning and deep learning. So, what are the differences?  All machine learning is AI, but not all AI is machine learning. AI is a general term for any computer program that does something smart. Deep learning is a subset of machine learning, and machine learning is a subset of AI.

Artificial intelligence is an area of computer science that stresses the creation of intelligent machines that work and respond like humans. The basic procedure of machine learning is to provide training data to a learning algorithm, which in turn generates a new set of rules, based on inferences from the data. By using different training data, the same learning algorithm could be used to produce diverse models. Deducing new instructions from data is the strong suit of machine learning. The more data that is available to train the algorithm, the more it learns.

When the term deep learning is used, it usually refers to deep artificial neural networks. Deep artificial neural networks are a set of algorithms that have set new bests in accuracy for critical problems, such as image recognition, sound perception, and language processing. Deep learning accomplishes perception accuracy at higher levels than ever before in areas such as consumer electronics, and it is vital for safety-critical applications such as autonomous vehicles. Current developments in deep learning have improved to the point where deep learning does better than humans in performing many tasks.

Inspired by the neurons that make up the human brain, neural networks comprise layers that are connected in adjacent layers to each other. The more layers there are, the deeper the network. A single neuron in the brain, receives as many as 100,000 signals from other neurons. When those other neurons fire, they apply either an exciting or inhibiting effect on the neurons to which they are connected. If the first neuron’s inputs add up to a certain base voltage, it will fire as well.

In an artificial neural network—just like the brain—signals travel between neurons. But instead of firing an electrical signal, a neural network allocates emphases to a variety of neurons. A neuron biased a great deal more than another neuron will wield more of an effect on the next layer of neurons. The final layer patches these weighted inputs together to come up with an answer.

These neural networks are made of layers of weighted neurons. Only they are not modelled on the workings of the brain. They are inspired by the visual system.

Every layer within a neural network utilizes a filter across the image to pick up explicit shapes or characteristics. The first few layers distinguish larger features, such as diagonal lines, while the following layers pick up finer details and organizes them into complex features.

Similar to an ordinary neural network, the final output layer is fully connected, which means that all of the neurons in that layer are connected to all neurons in the previous layer. The layers of neurons that are sandwiched between the first layer of neurons (input layer) and the last layer of neurons (output layer), are called hidden layers. This is where the neural network endeavors to solve problems. Reviewing the activities of the hidden layers can tell a lot about the information the network has learned to extract from the data.

Traditional neural networks only contain 2-3 hidden layers, while deep networks can have as many as 150. Large sets of labeled data are used to train deep learning models, using neural network architectures that learn features directly from the data without the need for manual feature extraction.

Deep learning machines don’t require a human programmer. This is possible because of the extraordinary amount of data we collect and consume. Data is the power for deep-learning models. Because of this, deep learning machines are already being used for practical purposes.

As deep learning continues to mature, we can expect many businesses to use machine learning to enhance customer experience. There are already deep-learning models being used for chatbots and online self-service solutions.

Machine translation isn’t new, but deep learning is facilitating enhance automatic translation of text by using stacked networks of neural networks and allowing translations from images.

In the past, black and white movie images had to be hand colored, which was very time consuming and costly.  Now this process can be automatically done with deep-learning models, which can automatically colorize grayscale images based on Convolutional Neural Networks, which features a fusion layer that allows an artist to merge local information dependent on small image areas with largescale prior images.

Advanced natural language processing and deep learning can help to filter news subjects in which you are interested. News aggregators using this new technology can filter news based on sentiment analysis, so you can create news streams that only cover the happening news containing stories of interest.

Another remarkable ability of deep learning is to identify an image and create an intelligible caption with proper sentence structure for that image just as though a human was writing the caption.

A deep learning machine can even generate text by learning the punctuation, grammar and style of a piece of text. It can use the model it created to automatically generate completely new text with the proper spelling, grammar and style of the example text. James Patterson—watch out.

The evolution of deep-learning machines is projected to pick up the pace and create even more innovative uses in the next few years. Deep-learning applications can educate a robot just by observing the actions of a human carrying out a task or use the connection from several other AIs in order to perform an action. A human brain processes input from past experiences. A deep learning robot will execute tasks based on the input of many different AI opinions.

The post What is Deep Learning? appeared first on Artificial Intelligence.