The study, published Jan. 30 in the Journal of Antimicrobial Chemotherapy, adds to growing evidence that lenacapavir may be a powerful new tool in the global anti-HIV drug arsenal. Approximately, 1.5 million people are living with HIV in Uganda.

“Our data shows that only 1.6% of the individuals studied are living with HIV strains that have any known lenacapavir-associated resistance mutations,” said senior author Dr. Guinevere Lee, assistant professor of virology in medicine at Weill Cornell Medicine. “That’s important because it shows lenacapavir is likely to be effective against strains of HIV circulating in East Africa.”

Since the 1990s, HIV drug combinations targeting different steps in the virus’ life cycle have been able to reduce virus load in patients to nearly undetectable levels. But drug resistance is a growing concern as the virus has evolved ways to thwart existing therapies. Lenacapavir, however, is the first drug to disrupt the protective capsid layer surrounding HIV’s genetic material (RNA), blocking the virus’s ability to reproduce and be transmitted from person to person.

Treatment twice a year with lenacapavir has been effective in patients who have never been treated and those with HIV strains that are resistant to other drugs. Last year, clinical trials showed that lenacapavir injections were 100% effective in preventing HIV infection among women in sub-Saharan Africa, who were HIV-negative.

However, little information was available about pre-existing resistance to lenacapavir in less well-studied HIV-1 strains like subtype A1 and D, which are more common in Eastern and Southern Africa. HIV-1 subtype B strains, which predominantly affect Europe and the United States, rarely have pre-existing mutations that would cause lenacapavir drug resistance.

Dr. Lee and her colleagues at Mbarara University of Science and Technology in Uganda and Massachusetts General Hospital in Boston helped fill that gap. They sequenced the capsid proteins from HIV-1 subtypes A1 and D from 546 Ugandan patients, who had never used antiretroviral therapy before. This approach allowed the investigators to examine naturally circulating viral variants.

They found that none of the patients had genetic mutations that would lead to major lenacapavir resistance. Only nine participants had minor lenacapavir resistance mutations that could partially reduce the effectiveness, but not enough to cause full resistance to the drug.

“Our study supports lenacapavir’s potential efficacy in this region. As lenacapavir is rolled out in East Africa, further studies will need to monitor for the emergence of drug-resistant strains,” Dr. Lee said. “It is important that we ensure HIV research reaches understudied communities where unique viral strains circulate.”

MIT researchers probed the inner workings of LLMs to better understand how they process such assorted data, and found evidence that they share some similarities with the human brain.

Neuroscientists believe the human brain has a “semantic hub” in the anterior temporal lobe that integrates semantic information from various modalities, like visual data and tactile inputs. This semantic hub is connected to modality-specific “spokes” that route information to the hub. The MIT researchers found that LLMs use a similar mechanism by abstractly processing data from diverse modalities in a central, generalized way. For instance, a model that has English as its dominant language would rely on English as a central medium to process inputs in Japanese or reason about arithmetic, computer code, etc. Furthermore, the researchers demonstrate that they can intervene in a model’s semantic hub by using text in the model’s dominant language to change its outputs, even when the model is processing data in other languages.

These findings could help scientists train future LLMs that are better able to handle diverse data.

“LLMs are big black boxes. They have achieved very impressive performance, but we have very little knowledge about their internal working mechanisms. I hope this can be an early step to better understand how they work so we can improve upon them and better control them when needed,” says Zhaofeng Wu, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this research.

His co-authors include Xinyan Velocity Yu, a graduate student at the University of Southern California (USC); Dani Yogatama, an associate professor at USC; Jiasen Lu, a research scientist at Apple; and senior author Yoon Kim, an assistant professor of EECS at MIT and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the International Conference on Learning Representations.

Integrating diverse data

The researchers based the new study upon prior work which hinted that English-centric LLMs use English to perform reasoning processes on various languages.

Wu and his collaborators expanded this idea, launching an in-depth study into the mechanisms LLMs use to process diverse data.

An LLM, which is composed of many interconnected layers, splits input text into words or sub-words called tokens. The model assigns a representation to each token, which enables it to explore the relationships between tokens and generate the next word in a sequence. In the case of images or audio, these tokens correspond to particular regions of an image or sections of an audio clip.

The researchers found that the model’s initial layers process data in its specific language or modality, like the modality-specific spokes in the human brain. Then, the LLM converts tokens into modality-agnostic representations as it reasons about them throughout its internal layers, akin to how the brain’s semantic hub integrates diverse information.

The model assigns similar representations to inputs with similar meanings, despite their data type, including images, audio, computer code, and arithmetic problems. Even though an image and its text caption are distinct data types, because they share the same meaning, the LLM would assign them similar representations.

For instance, an English-dominant LLM “thinks” about a Chinese-text input in English before generating an output in Chinese. The model has a similar reasoning tendency for non-text inputs like computer code, math problems, or even multimodal data.

To test this hypothesis, the researchers passed a pair of sentences with the same meaning but written in two different languages through the model. They measured how similar the model’s representations were for each sentence.

Then they conducted a second set of experiments where they fed an English-dominant model text in a different language, like Chinese, and measured how similar its internal representation was to English versus Chinese. The researchers conducted similar experiments for other data types.

They consistently found that the model’s representations were similar for sentences with similar meanings. In addition, across many data types, the tokens the model processed in its internal layers were more like English-centric tokens than the input data type.

“A lot of these input data types seem extremely different from language, so we were very surprised that we can probe out English-tokens when the model processes, for example, mathematic or coding expressions,” Wu says.

Leveraging the semantic hub

The researchers think LLMs may learn this semantic hub strategy during training because it is an economical way to process varied data.

“There are thousands of languages out there, but a lot of the knowledge is shared, like commonsense knowledge or factual knowledge. The model doesn’t need to duplicate that knowledge across languages,” Wu says.

The researchers also tried intervening in the model’s internal layers using English text when it was processing other languages. They found that they could predictably change the model outputs, even though those outputs were in other languages.

Scientists could leverage this phenomenon to encourage the model to share as much information as possible across diverse data types, potentially boosting efficiency.

But on the other hand, there could be concepts or knowledge that are not translatable across languages or data types, like culturally specific knowledge. Scientists might want LLMs to have some language-specific processing mechanisms in those cases.

“How do you maximally share whenever possible but also allow languages to have some language-specific processing mechanisms? That could be explored in future work on model architectures,” Wu says.

In addition, researchers could use these insights to improve multilingual models. Often, an English-dominant model that learns to speak another language will lose some of its accuracy in English. A better understanding of an LLM’s semantic hub could help researchers prevent this language interference, he says.

This research is funded, in part, by the MIT-IBM Watson AI Lab.

“Breast cancer incidence is increasing in U.S. women under 40, but until now, it was unknownif incidence trends varied by U.S. geographic region,” said Rebecca Kehm, PhD, assistant professor of Epidemiology at Columbia Mailman School, and first author. “Our findings can more accurately inform whether exposures that vary in prevalence across the U.S. also contributes to breast cancer risk in younger women.”

Using the U.S. Cancer Statistics database, the researchers analyzed age-adjusted breast cancer-incidence rates from 2001 to 2020 in women aged 25-39. They calculated the average annual percent change using statistical regression formulas and performed age-distribution analyses.

“Two-thirds of all cancers identified both in the U.S. and globally are diagnosed in women,” said Mary Beth Terry, PhD, professor of Epidemiology at Columbia Mailman School of Public Health, and senior author of the study.

From 2001 to 2020, breast cancer incidence in women under 40 increased by more than 0.50 percent per year in 21 states, while remaining stable or decreasing in the other states. Incidence was 32 percent higher in the five states with the highest rates compared to the five states with the lowest rates. The Western region had the highest rate of increase from 2001 to 2020; the Northeast had the highest absolute rate among women under 40 and experienced a significant increase over time The South was the only region where breast cancer under 40 did not increase from 2001 to 2020.

The overall incidence of early-onset breast cancer ranged from 28.6 per 100,000 in Wyoming to 41 cases per 100,000 people in Connecticut. The five states with the highest early-onset incidence from 2001 to 2020 were Maryland, New York, New Jersey, Hawaii, and Connecticut. Hispanic women had the lowest early-onset frequency rates in all regions, ranging from 26 per 100,000 in the Midwest to 32.6 per 100,000 in the Northeast.

Non-Hispanic White women were the only group to experience a statistically significant increase in early-onset breast cancer incidence across all four regions of the U.S. Non-Hispanic Black women had the highest incidence of early-onset breast cancer. This was true across the regions of the country.

The authors note the importance of investigating other risk factors including alcohol consumption, an established risk factor for breast cancer and which is known to vary across states and also be influenced by state alcohol policies.

“The increase in incidence we are seeing is alarming and cannot be explained by genetic factors, alone which evolve over much longer periods nor by changes in screening practices given that women under 40 years are below the recommended age for routine mammography screening,” noted Kehm.

“While the causes behind the rising incidence of early onset breast cancer are not yet fully understood, studying how trends vary across different population subgroups can offer valuable insights and help generate hypotheses for future research,” said Professor Terry. “We also are able to gain an understanding into the increase in breast cancer incidence among women who are not currently recommended for routine screening.”

Co-authors are Josephine Daaboul, Columbia Mailman School of Public Health and Fielding School of Public Health, University of California Los Angeles; and Parisa Tehranifar, Columbia Mailman School of Public Health.

The study was supported by the National Cancer Institute (R00CA263024).