Artificial Intelligence and Machine Learning News

Neural Networks for Drug Discovery and Design

Architectures suited for molecules streamline the identification of pharmaceutical candidates.
  1. Introduction
  2. High-Throughput Screening
  3. Architectures
  4. Mechanisms
  5. Author
connected neurons over a graph plain, illustration

Drugs play a central role in modern medicine, but bringing new ones to market is a lengthy, expensive process. Pharmaceutical companies are exploring ways to streamline all aspects of their complex pipelines with artificial intelligence (AI).

A key early step is the discovery and design of new molecules that have a desired biochemical effect that can modulate known disease-related processes. To succeed, the molecules must also be suitable for manufacture and drug formulation, and have an acceptably low number of side effects. Finding better candidates and eliminating losers at an early stage makes this process faster and cheaper.

In recent years, researchers in academia and industry have been devising machine learning tools both to screen known compounds computationally for desirable properties, and to suggest completely new ones. The new tools use some techniques that have powered the deep learning revolution of the last decade. However, efficiently analyzing chemistry and biology have favored other architectures, such as graph neural networks, whose structure naturally mimics that of molecules.

“There is really an opportunity to do a lot of interesting novel computer science, because the context of chemistry and biology is very different than language and imaging,” said Regina Barzilay, a computer scientist at the Massachusetts Institute of Technology who has worked on this topic for years. “It’s an area that’s going to grow for the field, and we will see a lot of computer science discoveries along the way.”

Back to Top

High-Throughput Screening

The revenues from a single successful drug can be enormous, especially for “blockbusters” that generate more than $1 billion in sales every year. The costs of developing a single drug, however, are also enormous, currently estimated at $2.5 billion or more. The most expensive step is late-stage clinical trials, but candidates often fail earlier in the pipeline, for example if they prove ineffective in animals or humans, or have unacceptable side effects.

By some estimates, some 10,000 candidate molecules fail for each one that is ultimately commercialized. AI that predicts a proposed drug’s properties—desired and undesired—well enough to improve these odds will be extremely lucrative.

Pharmaceutical companies have long employed automated laboratory screening of biochemical properties using proprietary libraries that may contain millions of compounds, including known drugs and naturally occurring compounds. These resources are curated to include diverse molecular features, and thus to span a large space of structures.

However, even these huge libraries represent a vanishingly small fraction of an estimated 1060 or more “small molecules,” which can be manufactured and administered more straightforwardly and cheaply than large, biologically produced molecules. “The chemical space that one has to screen against for a given protein target or a genome is exponentially huge,” said Neeraj Kumar, a computational data scientist at the U.S. Department of Energy’s Pacific Northwest National Laboratory.

“The chemical space that one has to screen against for a given protein target or a genome is exponentially huge.”

Computationally predicting the properties of compounds with AI can accelerate drug discovery and, ideally, commercialization. Even more significant is finding completely new kinds of molecules in the enormous universe of possible molecules. “You can never be sure that there isn’t something very interesting out there that’s really important, but you don’t know about it because it’s not part of your existing library,” Barzilay said.

Nonetheless, exploring the many possible structures is daunting. One common strategy to reduce the search burden starts with the backbones of known drugs and other molecules to define “scaffolds.” Modest chemical modifications can lead to even better candidates.

A related technique is known as fragment-based sampling. In the laboratory, for example, researchers can include small pieces of molecules in their library and search for their interactions with a target protein. Using AI, Barzilay and others also have adapted this strategy to find building blocks, which can be used as seeds for building new compounds.

Back to Top


Over the last decade, neural networks have shown amazing versatility in image and language tasks when given massive amounts of labeled data, even with little guidance about relevant features on which to focus. Nonetheless, network architectures whose representations match the task can make a big difference. For images, for example, convolutional neutral networks can be constructed to help identify objects irrespective of their position, size, or orientation.

For molecules, graph neural networks (GNNs) are showing increasing success. These systems represent data in terms of nodes (which correspond to atoms) and the edges that connect them (which correspond to the bonds between atoms in a molecule).

Other promising deep learning architectures include recurrent neural networks and graph convolutional neural networks.

“There are lots of pharmaceutical companies that are using the tools quite successfully,” Barzilay said, based on her experience in MIT’s Machine Learning for Pharmaceutical Discovery and Synthesis Consortium, a collaboration between the pharmaceutical and biotechnology industries and MIT’s departments of Chemical Engineering, Chemistry, and Computer Science.

Assessing whether a molecule binds to a target protein is “a classification problem in machine learning,” Kumar said, resulting in a binary binding/non-binding decision. Importantly, however, his group and others formulate their task as a regression problem, quantifying the interaction strength, based on structures where the binding affinity is experimentally known.

Many other properties also are important; for example, toxicity and other side effects in non-targeted tissues, and the ease of synthesizing a novel molecule. In addition, such factors as solubility, degradation lifetime on the shelf and in the body, and other properties affect the deliverability of a drug. “When you design drug candidates, you have multiple properties to optimize,” Kumar said. Various properties can be combined into a single reward function that is used in training.

How to best design the algorithms and formulate the objectives are still open questions, however, Barzilay said. “There are many different ways how creatively you can think about it.” Moreover, although the first generation of tools produced reasonable results, she said, “Right now, people increasingly understand that you need to design to a new algorithm foundation.”

Beyond deciding which drugs are worth bringing into trials, Barzilay said there is longer-term potential for “personalization” of drugs to determine what is best of for an individual with a particular profile. “I think machine learning can really excel” at this, she said.

Back to Top


One biological mechanism that powers many successful drugs is targeting receptor proteins that reside on cell surfaces. Drug molecules in the bloodstream can gum them up and inhibit the signals they otherwise would pass to the cellular machinery inside. The critical binding between a target protein and a potential small-molecule drug can be measured in the lab. Alternatively, if the folded configuration of a protein is known, whether from experiment or theory, computational modeling can assess whether possible drugs can match the shape and atomic attractions in the nooks and crannies of the protein that would otherwise detect signaling molecules.

“When you design drug candidates, you have multiple properties to optimize.”

This strategy has gotten a huge boost from recent progress in AI prediction of protein structure from its known genetic sequence. In 2020, Alphabet-owned DeepMind trounced other participants in the long-running Critical Assessment of protein Structure Prediction project. In 2022, the project used its updated attention-based transformer AlphaFold 2 to predict structures for all known proteins, which it made public.

Like the much more labor-intensive experimental structures for proteins, these predictions provide a solid starting point for predicting how small molecules will interact with them. Indeed, DeepMind CEO Demis Hassabis has founded a new company, Isomorphic Laboratories, specifically targeting drug discovery.

“AlphaFold 2 is definitely changing the way we have been thinking,” said Kumar, adding the protein-structure database will be a powerful resource. “The only caveat with that database is we don’t have measured properties that we need as part of the training.”

Advanced drugs also include large proteins, such as antibodies, that are made biologically. The interactions between proteins also are essential in other biological processes, such as the formation of multiprotein complexes.

Rapid progress in protein-structure prediction is transforming this field as well. David Baker’s group at the University of Washington, for example, has adapted some of the new methods to their long-running project to design completely unknown proteins, as well as smaller peptides.

Unlike some of the big-data-driven successes of deep learning, drug development also relies on deep scientific understanding. “You cannot go to Mechanical Turk or something like this. You really need to do the experiments,” Barzilay said. “So typically incorporating some kind of biases into the model, some knowledge of chemistry, really helps,” as does knowledge of biology and medicine.

Researchers continue to extend AI techniques to specialized fields that already have powerful specialized frameworks. “It is already science, it is invented,” Barzilay said. “The question is, ‘What is the best way to inject this information into the model?’ This, I think, is not a solved problem by far.”

*  Further Reading

U.S. National Institute of General Medical Sciences Protein Structure Prediction Center;

S.B.A. Turzo, E.R. Hantz, and S. Lindert. Applications of machine learning in computer-aided drug discovery; QRB Discovery 3: e14, 1–16 (2022).

SCORR Marketing, How Artificial Intelligence & Blockchain Are Changing Drug Discovery (2021),

U.S. Government Accountability Office, Artificial Intelligence in Health Care: Benefits and Challenges of Machine Learning in Drug Development,” (2020);

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More