AI-driven blood analysis may spur cancer detection, Hong Kong prenatal test pioneer says
AI helps researchers recognise patterns in the gene fragments they find in blood plasma, sparking a revolution in the field of epigenetics, or the study of non-mutation behavioural effects on genes, said Professor Dennis Lo Yuk-ming.
The AI-powered method enables researchers to decipher epigenetic signals from DNA samples without the damaging effects of the chemical treatment that destroys 90 per cent of the genetic material, he said.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
"My team members and I were asking: is it possible to use some of the AI methods developed for facial recognition [to help us] see things we cannot see?" Lo said during a January 7 interview with the Post. "As it turned out, it is."
Professor Allen Chan Kwan-chee (left) and Professor Dennis Lo Yuk-ming (right) of the Chinese University of Hong Kong on 10 August 2017. Photo: Dickson Lee alt=Professor Allen Chan Kwan-chee (left) and Professor Dennis Lo Yuk-ming (right) of the Chinese University of Hong Kong on 10 August 2017. Photo: Dickson Lee>
Lo's team at the Centre for Novostics at the Hong Kong Science and Technology Park developed an in-house AI system based on the Convolutional Neural Network, a deep learning model that was designed mainly for image recognition. With advances in generative AI tools such as ChatGPT, the company also incorporated the transformer neural network architecture into the system, he said.
Like many biomedical companies at the forefront of drug discovery, diagnostics and treatment, Novostics is diving into the use of AI. Like many companies in Hong Kong and mainland China, the company "has faced challenges" in accessing computing power because of US curbs on access to cutting-edge AI chips.
Professor Dennis Lo Yuk-ming with a chip that could sequence up to 8 million single DNA molecules for genomic research and genetic disease detection, at the Prince of Wales Hospital in Shatin on 20 October 2020. Photo: K. Y. Cheng alt=Professor Dennis Lo Yuk-ming with a chip that could sequence up to 8 million single DNA molecules for genomic research and genetic disease detection, at the Prince of Wales Hospital in Shatin on 20 October 2020. Photo: K. Y. Cheng>
"Fortunately, we still have a pretty good system that we developed over the years", said Lo, a day before he formally took over as president of the Chinese University of Hong Kong (CUHK). "With the difficulty in getting the latest generation of Nvidia chips, we just have to try other chipmakers that can still provide what we need."
Lo's team is developing tests for the early detection of cancers, using a similar method as his pioneering work on using the blood plasma of pregnant women to check for Down syndrome and other genetic disorders.
Nasopharyngeal cancer (NPC), a throat tumour that is common in southern China's Guangdong province, was the first on Lo's target list. Using the cancer's suspected link with the Epstein-Barr virus, Lo developed a test in 2013 for finding it in blood plasma.
Single-use DNA cells for genome sequencing, at the Prince of Wales Hospital in Shatin on 20 October 2020 . Photo: K. Y. Cheng alt=Single-use DNA cells for genome sequencing, at the Prince of Wales Hospital in Shatin on 20 October 2020 . Photo: K. Y. Cheng>
After collecting 20,000 samples over four years, the study concluded in July 2023 that screening for the virus was an effective detection tool for the cancer, helping to slash the mortality rate to a mere 3.7 per cent from 40 per cent, according to Lo. The test is now available in Hong Kong for HK$1,500 (US$192.5).
"Even though Hong Kong's healthcare system is pretty good, 75 per cent of nasopharyngeal cancer was detected late" before the new test was introduced, Lo said.
The use of AI widened the possibilities of finding other cancers even if a virus may not be behind every cancer, because the discovery of each mutation was "like hunting one thing out of a million", he said.
Cirina, which Lo founded in Hong Kong after creating the NPC test, merged in 2017 with Grail, a company in California that speciliased in cancer detection. The merged company developed the Galleri test to screen for more than 50 types of virus-related cancers, a product that was named as one of The Best Inventions of 2022 by Time magazine.
Grail was bought for US$8 billion in 2021 by the biotechnology giant Illumina, based in San Diego. Its multi-cancer test is only available in the US market for US$949 (HK$7,380), and is not covered by most healthcare coverage plans.
Lo's team is trying to make a cheaper version of the test for mainland China and the rest of the world. Novostics has developed a test called Fragma, which can find the DNA fragments of lung cancer cells and liver carcinoma in blood plasma and urine.
In 2023, Lo established a US$200 million venture called Insighta with Nasdaq-listed Prenetics Group to commercialise the Fragma tests. Tencent Holdings invested US$30 million last October in Insighta, valuing the start-up at US$200 million.
Insighta, based at the Hong Kong Science and Technology Park, is conducting large-scale clinical tests of the Fragma process in Hong Kong and mainland China. It aims to eventually offer the test at US$200 (HK$1,565).
Some of the earlier products, such as the test for liver cancer, will "hopefully" be available "in two or three years from now", Lo said. "If we want to do it in a more resource-challenged environment, we need the technology to be cheaper."
This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP's Facebook and Twitter pages. Copyright © 2025 South China Morning Post Publishers Ltd. All rights reserved.
Copyright (c) 2025. South China Morning Post Publishers Ltd. All rights reserved.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Miami Herald
19 minutes ago
- Miami Herald
Jim Acosta's AI interview raises deeper questions about human connection
From Facebook to FaceTime, it is now easier than ever to stay connected with friends and family members. Thanks to technology, I can FaceTime with my parents, send TikTok videos and share photos of my dog with friends with a few clicks. But what happens when we use technology to virtually resurrect the dead and allow an avatar to speak on behalf of the deceased in a video interview sharing a political viewpoint? That question came to the forefront when former CNN White House correspondent Jim Acosta 'interviewed' an Artificial Intelligence avatar of Joaquin Oliver, a teenager killed in the 2018 Parkland high school shooting. In the video, the avatar used a chatbot to generate answers in a voice that supposedly sounded like the boy. Acosta said the boy's father had approached him to do the piece as a way of keeping Joaquin's memory alive. The interview sparked backlash and raised ethical concerns over technology's potential role in tarnishing the memory of the dead or changing their viewpoint. In this case, the Joaquin avatar advocated for 'stronger gun control laws, mental health support and community engagement.' Acosta's interview also raises a larger question: Is AI helping us connect or just simulating human connection while we become more disconnected? Grief is different for everyone, and how people grieve is evolving along with technology. Four years ago, I read about Joshua Barbeau, a guy who lost his fiancée to a rare liver disease. He used Project December — a chat website that simulates a text-based conversation with anyone, including someone who is dead — to communicate via chatbot with an AI version of his deceased fiancée. Traditionally, people processed grief through therapy or with the support of trusted friends or family members. Today, programs like ChatGPT are being used as therapists, for friendships and in some cases, as romantic partners. As Derek Thompson wrote in The Atlantic in February, 'Americans are spending less time with other people than in any other period for which we have trustworthy data, going back to 1965.' Isolation isn't accidental. Many people keep their phones on silent, prefer texting to calling and spend hours doomscrolling. Now, AI can simulate people with avatars. So when grief feels heavy, there is a program to help. When my grandmother died, the grief felt unbearable. Fifteen years later, when I talk about her, the loss still tugs at my heart. There isn't a day that I don't wish she was still here. I keep her memory alive without using AI, but I don't judge how others grieve. Grief is intimate — and that's what makes Joaquin Oliver's AI interview so eerie. It raises emotional questions. Is AI helping a family grieve? Was it a family's attempt to give their son a voice in a country that failed to protect him from gun violence? Maybe it's both. Human emotions can be messy. But what makes it all bearable is human connection — holding space for others in tough times — something technology can't replace. A chatbot can't actually resurrect the dead. It's a mirror of memories, reflecting our own words and thoughts. In the end, as the San Francisco Chronicle reported, Barbeau — the man who lost his fiancee — 'felt like the chatbot had given him permission to move on with his life in small ways, simply by urging him to take care of himself.' Perhaps that's the lesson. Technology can offer us tools for processing grief and maintaining memories, and maybe even give us permission to move on. But AI can't hug you or laugh at inside jokes. It won't sit next to you in silence when the world feels heavy. The loss of my grandmother still hurts. No amount of technology can bring her back. That's part of life's beauty — to love something death can touch. We carry those we've lost not through digital simulations, but by sharing memories and stories with others. The danger isn't just that AI will replace human connection — it's that we may settle for it. Mary Anna Mancuso is a member of the Miami Herald Editorial Board. Her email: mmancuso@
Tom's Guide
19 minutes ago
- Tom's Guide
GPT-5 vs GPT-4: Here's what's different (and what's not) in ChatGPT's latest upgrade
The GPT-5 announcement has come and gone. And now, as the dust settles, it's time to work out what we're left with. Is this really the game-changing invention it's been made out to be, and more importantly, how does it compare to the now old-school GPT-4? In the lead up to the big reveal, OpenAI, as well as early reviewers, were touting this to be the one of the biggest leaps in AI we've seen. On the other hand, there was a much smaller group whispering of this only actually being a minor upgrade. The real truth seems to be somewhere in the middle. GPT-5 brings about the most advanced version of ChatGPT to date. It makes some big changes in a variety of important areas, as well as some smaller upgrades in a few spaces. So when stacked up, where are the upgrades in GPT-5 most noticeable? OpenAI, and most of its competitors too, don't just launch one version at a time. They tend to launch a variety of progressions for each model. This allows the model to spread out its workload, using the best version for the task at hand. For example, you don't need a super-powered thinking machine to ask a passing question about what clouds are. But you do need something with more oomph to compile a complicated multi-step document while searching the web. For each GPT-4 model, there is a corresponding GPT-5 version. These are: Get instant access to breaking news, the hottest reviews, great deals and helpful tips. One of the biggest changes in GPT-4 over its predecessor is the way it is structured. GPT-5 is a unified system, that uses an efficient modelling system to decide which kind of model is best required to answer your question. In other words, where GPT-4 required you to change versions each time to best fit your needs, GPT-5 will automatically filter your request into the best version. This, on paper, feels like a small change, but in practice it will make a noticeable difference. This will allow ChatGPT to always offer its best version for whatever your asking from it. OpenAI has also said that this system will get better with time, learning from prompts what kind of model is best for the use. OpenAI claims that GPT-5 outperforms all previous models on benchmarks and answers questions more efficiently and quickly. However, the OpenAI team has also made improvements to reduce the number of hallucinations. Another feature that a lot of people will be excited to hear is that, compared to GPT-4, this new version is a lot less sycophant in nature. In other words, it won't just blindly agree with you or overly praise you. While the model is smarter overall, the improvement is most notable in three areas: writing, coding, and health. Coding This was expected in the lead-up to the GPT-5 reveal. As AI models get more advanced, they are all fighting to have the smartest version for coding. GPT-5, according to OpenAI, is the strongest coding model they've made to date. It is mostly improved in front-end generation (what your apps/websites look like when generated) and debugging larger sets of code. OpenAI says that, compared to GPT-4, it can often create complete websites, apps, and games from one single prompt. Early testers stated that the styling and typography were also a big improvement over GPT-4. Writing For a lot of ChatGPT users, the ability to use the model for writing is one of its biggest benefits. GPT-5 has seen an improvement in both creativity and writing in a more human tone. OpenAI has said that, with this model, ChatGPT is even better at taking rough ideas and turning them into compelling sets of writing. It is more reliable on writing with structural ambiguity (creative poetry, for example, or free verse that flows naturally). It is also more reliable for more serious writing such as academic essays or preparing emails. Health Chatbots in the world of health has been a concern for a lot of people, but GPT-5 seems to be one of the most advanced versions we've seen yet in this space. The model scored significantly higher compared to GPT-4 on HealthBench, an evaluation based on realistic scenarios and physician-defined criteria. Compared to GPT-4, it acts more like an active thought partner, proactively noticing potential concerns and asking questions to give more helpful answers. OpenAI claims that the model is now better equipped to personalize health advice, taking into account someone's geography, knowledge level, and specific health information. One of the other big claims OpenAI made in the lead up to this launch was that GPT-5 would be much faster. That seems to be true, with OpenAI claiming that, in its evaluations, GPT-5 performed better than OpenAI o3 with 50-80% less output tokens across most logical tasks. In other words, it could produce a better result with less effort. This, in theory, will make ChatGPT a more environmentally efficient model going forward, and also more cost effective. It should also lead to faster responses. Obviously, these claims are something we'll be putting to the test as the Tom's Guide team starts to test GPT-5 in a variety of situations. Initially, it seems like GPT-5 isn't a world-shattering update, but it has made ChatGPT a better tool in all of the areas that are most necessary. Along with writing better code and creative writing, it is now faster, more reliable, and more efficient. Plus there's improvements to multimodality thanks to an updated voice mode. More importantly, as AI is given more important roles in the world, especially with the rollout of agentic AI tools, GPT-5 has fewer hallucinations, better medical advice, and a more confident attitude to its own responses. While it is too early to see how it stacks up against Grok or Gemini, we would imagine it is going to take a top spot when benchmarked. GPT-5 is clearly a major upgrade to GPT-4. As more features become available, we will be covering them all here at Tom's Guide. Follow Tom's Guide on Google News to get our up-to-date news, how-tos, and reviews in your feeds. Make sure to click the Follow button.
Bloomberg
19 minutes ago
- Bloomberg
Uber Walks Fine Line to Satisfy Shareholders on Driverless Car Investments
Welcome to Tech In Depth, our daily newsletter about the business of tech from Bloomberg's journalists around the world. Today, Natalie Lung reports on Uber's strategy to invest in the future of robotaxis. GPT-5: OpenAI rolled out its long-awaited new artificial intelligence model, GPT-5. According to Nick Turley, who runs OpenAI's ChatGPT team, this new model can find better answers, respond more quickly and fabricate responses less frequently than earlier models.
