Libusha Kelly, Albert Einstein College of Medicine

Viruses are a mysterious and poorly understood force in microbial ecosystems. Researchers know they can infect, kill and manipulate human and bacterial cells in nearly every environment, from the oceans to your gut. But scientists don't yet have a full picture of how viruses affect their surrounding environments in large part because of their extraordinary diversity and ability to rapidly evolve.

Communities of microbes are difficult to study in a laboratory setting. Many microbes are challenging to cultivate, and their natural environment has many more features influencing their success or failure than scientists can replicate in a lab.

So systems biologists like me often sequence all the DNA present in a sample – for example, a fecal sample from a patient – separate out the viral DNA sequences, then annotate the sections of the viral genome that code for proteins. These notes on the location, structure and other features of genes help researchers understand the functions viruses might carry out in the environment and help identify different kinds of viruses. Researchers annotate viruses by matching viral sequences in a sample to previously annotated sequences available in public databases of viral genetic sequences.

However, scientists are identifying viral sequences in DNA collected from the environment at a rate that far outpaces our ability to annotate those genes. This means researchers are publishing findings about viruses in microbial ecosystems using unacceptably small fractions of available data.

To improve researchers' ability to study viruses around the globe, my team and I have developed a novel approach to annotate viral sequences using artificial intelligence. Through protein language models akin to large language models like ChatGPT but specific to proteins, we were able to classify previously unseen viral sequences. This opens the door for researchers to not only learn more about viruses, but also to address biological questions that are difficult to answer with current techniques.

Annotating viruses with AI

Large language models use relationships between words in large datasets of text to provide potential answers to questions they are not explicitly “taught” the answer to. When you ask a chatbot “What is the capital of France?” for example, the model is not looking up the answer in a table of capital cities. Rather, it is using its training on huge datasets of documents and information to infer the answer: “The capital of France is Paris.”

Similarly, protein language models are AI algorithms that are trained to recognize relationships between billions of protein sequences from environments around the world. Through this training, they may be able to infer something about the essence of viral proteins and their functions.

We wondered whether protein language models could answer this question: “Given all annotated viral genetic sequences, what is this new sequence's function?”

In our proof of concept, we trained neural networks on previously annotated viral protein sequences in pre-trained protein language models and then used them to predict the annotation of new viral protein sequences. Our approach allows us to probe what the model is “seeing” in a particular viral sequence that leads to a particular annotation. This helps identify candidate proteins of interest either based on their specific functions or how their genome is arranged, winnowing down the search space of vast datasets.

By identifying more distantly related viral gene functions, protein language models can complement current methods to provide new insights into microbiology. For example, my team and I were able to use our model to discover a previously unrecognized integrase – a type of protein that can move genetic information in and out of cells – in the globally abundant marine picocyanobacteria Prochlorococcus and Synechococcus. Notably, this integrase may be able to move genes in and out of these populations of bacteria in the oceans and enable these microbes to better adapt to changing environments.

Our language model also identified a novel viral capsid protein that is widespread in the global oceans. We produced the first picture of how its genes are arranged, showing it can contain different sets of genes that we believe indicates this virus serves different functions in its environment.

These preliminary findings represent only two of thousands of annotations our approach has provided.

Analyzing the unknown

Most of the hundreds of thousands of newly discovered viruses remain unclassified. Many viral genetic sequences match protein families with no known function or have never been seen before. Our work shows that similar protein language models could help study the threat and promise of our planet's many uncharacterized viruses.

While our study focused on viruses in the global oceans, improved annotation of viral proteins is critical for better understanding the role viruses play in health and disease in the human body. We and other researchers have hypothesized that viral activity in the human gut microbiome might be altered when you're sick. This means that viruses may help identify stress in microbial communities.

However, our approach is also limited because it requires high-quality annotations. Researchers are developing newer protein language models that incorporate other “tasks” as part of their training, particularly predicting protein structures to detect similar proteins, to make them more powerful.

Making all AI tools available via FAIR Data Principles – data that is findable, accessible, interoperable and reusable – can help researchers at large realize the potential of these new ways of annotating protein sequences leading to discoveries that benefit human health.

Libusha Kelly, Associate Professor of Systems and Computational Biology, Microbiology and Immunology, Albert Einstein College of Medicine

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Mayank Kejriwal, University of Southern California

Many people understand the concept of bias at some intuitive level. In society, and in artificial intelligence systems, racial and gender biases are well documented.

If society could somehow remove bias, would all problems go away? The late Nobel laureate Daniel Kahneman, who was a key figure in the field of behavioral economics, argued in his last book that bias is just one side of the coin. Errors in judgments can be attributed to two sources: bias and noise.

Bias and noise both play important roles in fields such as law, medicine and financial forecasting, where human judgments are central. In our work as computer and information scientists, my colleagues and I have found that noise also plays a role in AI.

Statistical noise

Noise in this context means variation in how people make judgments of the same problem or situation. The problem of noise is more pervasive than initially meets the eye. A seminal work, dating back all the way to the Great Depression, has found that different judges gave different sentences for similar cases.

Worryingly, sentencing in court cases can depend on things such as the temperature and whether the local football team won. Such factors, at least in part, contribute to the perception that the justice system is not just biased but also arbitrary at times.

Other examples: Insurance adjusters might give different estimates for similar claims, reflecting noise in their judgments. Noise is likely present in all manner of contests, ranging from wine tastings to local beauty pageants to college admissions.

Noise in the data

On the surface, it doesn't seem likely that noise could affect the performance of AI systems. After all, machines aren't affected by weather or football teams, so why would they make judgments that vary with circumstance? On the other hand, researchers know that bias affects AI, because it is reflected in the data that the AI is trained on.

For the new spate of AI models like ChatGPT, the gold standard is human performance on general intelligence problems such as common sense. ChatGPT and its peers are measured against human-labeled commonsense datasets.

Put simply, researchers and developers can ask the machine a commonsense question and compare it with human answers: “If I place a heavy rock on a paper table, will it collapse? Yes or No.” If there is high agreement between the two – in the best case, perfect agreement – the machine is approaching human-level common sense, according to the test.

So where would noise come in? The commonsense question above seems simple, and most humans would likely agree on its answer, but there are many questions where there is more disagreement or uncertainty: “Is the following sentence plausible or implausible? My dog plays volleyball.” In other words, there is potential for noise. It is not surprising that interesting commonsense questions would have some noise.

But the issue is that most AI tests don't account for this noise in experiments. Intuitively, questions generating human answers that tend to agree with one another should be weighted higher than if the answers diverge – in other words, where there is noise. Researchers still don't know whether or how to weigh AI's answers in that situation, but a first step is acknowledging that the problem exists.

Tracking down noise in the machine

Theory aside, the question still remains whether all of the above is hypothetical or if in real tests of common sense there is noise. The best way to prove or disprove the presence of noise is to take an existing test, remove the answers and get multiple people to independently label them, meaning provide answers. By measuring disagreement among humans, researchers can know just how much noise is in the test.

The details behind measuring this disagreement are complex, involving significant statistics and math. Besides, who is to say how common sense should be defined? How do you know the human judges are motivated enough to think through the question? These issues lie at the intersection of good experimental design and statistics. Robustness is key: One result, test or set of human labelers is unlikely to convince anyone. As a pragmatic matter, human labor is expensive. Perhaps for this reason, there haven't been any studies of possible noise in AI tests.

To address this gap, my colleagues and I designed such a study and published our findings in Nature Scientific Reports, showing that even in the domain of common sense, noise is inevitable. Because the setting in which judgments are elicited can matter, we did two kinds of studies. One type of study involved paid workers from Amazon Mechanical Turk, while the other study involved a smaller-scale labeling exercise in two labs at the University of Southern California and the Rensselaer Polytechnic Institute.

You can think of the former as a more realistic online setting, mirroring how many AI tests are actually labeled before being released for training and evaluation. The latter is more of an extreme, guaranteeing high quality but at much smaller scales. The question we set out to answer was how inevitable is noise, and is it just a matter of quality control?

The results were sobering. In both settings, even on commonsense questions that might have been expected to elicit high – even universal – agreement, we found a nontrivial degree of noise. The noise was high enough that we inferred that between 4% and 10% of a system's performance could be attributed to noise.

To emphasize what this means, suppose I built an AI system that achieved 85% on a test, and you built an AI system that achieved 91%. Your system would seem to be a lot better than mine. But if there is noise in the human labels that were used to score the answers, then we're not sure anymore that the 6% improvement means much. For all we know, there may be no real improvement.

On AI leaderboards, where large language models like the one that powers ChatGPT are compared, performance differences between rival systems are far narrower, typically less than 1%. As we show in the paper, ordinary statistics do not really come to the rescue for disentangling the effects of noise from those of true performance improvements.

Noise audits

What is the way forward? Returning to Kahneman's book, he proposed the concept of a “noise audit” for quantifying and ultimately mitigating noise as much as possible. At the very least, AI researchers need to estimate what influence noise might be having.

Auditing AI systems for bias is somewhat commonplace, so we believe that the concept of a noise audit should naturally follow. We hope that this study, as well as others like it, leads to their adoption.

Mayank Kejriwal, Research Assistant Professor of Industrial & Systems Engineering, University of Southern California

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Benjamin Hurrell, University of Southern California and Omid Akbari, University of Southern California

You've likely heard that you can get iron from eating spinach and steak. You might also know that it's an essential trace element that is a major component of hemoglobin, a protein in red blood cells that carries oxygen from your lungs to all parts of the body.

A lesser known important function of iron is its involvement in generating energy for certain immune cells.

In our lab's newly published research, we found that blocking or limiting iron uptake in immune cells could potentially ease up the symptoms of an asthma attack caused by allergens.

Immune cells that need iron

During an asthma attack, harmless allergens activate immune cells in your lungs called ILC2s. This causes them to multiply and release large amounts of cytokines – messengers that immune cells use to communicate – and leads to unwanted inflammation. The result is symptoms such as coughing and wheezing that make it feel like someone is squeezing your airways.

To assess the role iron plays in how ILC2s function in the lungs, we conducted a series of experiments with ILC2s in the lab. We then confirmed our findings in mice with allergic asthma and in patients with different severities of asthma.

First, we found that ILC2s use a protein called transferrin receptor 1, or TfR1, to take up iron. When we blocked this protein as the ILC2s were undergoing activation, the cells were unable to use iron and could no longer multiply and cause inflammation as well as they did before.

We then used a chemical called an iron chelator to prevent ILC2s from using any iron at all. Iron chelators are like superpowered magnets for iron and are used in medical treatments to help manage conditions where there's too much iron in the body.

When we deprived ILC2s with an iron chelator, the cells had to change their metabolism and switch to a different way of getting energy, like trading in a sports car for a bicycle. The cells weren't as effective at causing inflammation in the lungs anymore.

Next, we limited cellular iron in mice with sensitive airways due to ILC2s. We did this in three different ways: by inhibiting TfR1, adding an iron chelator or inducing low overall iron levels using a synthetic protein called mini-hepcidin. Each of these methods helped reduce the mice's airway hyperreactivity – basically reducing the severity of their asthma symptoms.

Lastly, we looked at cells from patients with asthma. We noticed something interesting: the more TfR1 protein on their ILC2 cells, the worse their asthma symptoms. In other words, iron was playing a big role in how bad their asthma got. Blocking TfR1 and administering iron chelators both reduced ILC2 proliferation and cytokine production, suggesting that our findings in mice apply to human cells. This means we can move these findings from the lab to clinical trials as quickly as possible.

Iron therapy for asthma

Iron is like the conductor of an orchestra, instructing immune cells such as ILC2s how to behave during an asthma attack. Without enough iron, these cells can't cause as much trouble, which could mean fewer asthma symptoms.

Next, we're working on targeting a patient's immune cells during an asthma attack. If we can lower the amount of iron available to ILC2s without depleting overall iron levels in the body, this could mean a new therapy for asthma that tackles the root cause of the disease, not just the symptoms. Available treatments can control symptoms to keep patients alive, but they are not curing the disease. Iron-related therapies may offer a better solution for patients with asthma.

Our discovery applies to more than just asthma. It could be a game-changer for other diseases where ILC2s are involved, such as eczema and type 2 diabetes. Who knew iron could be such a big deal to your immune system?

Benjamin Hurrell, Assistant Professor of Research in Molecular Microbiology and Immunology, University of Southern California and Omid Akbari, Professor of Molecular Microbiology and Immunology, University of Southern California

This article is republished from The Conversation under a Creative Commons license. Read the original article.