The sheer volume of data in biomedical research today is staggering. From intricate radiological simulations to the deep dives into genomics and proteomics, it's a landscape that can easily overwhelm even the most seasoned specialists. But what if there was a way to not just process this data, but to find hidden patterns and unlock new insights? That's where generative AI, particularly models like LG AI Research's EXAONE, is starting to shine.
Think of generative AI as a remarkably adept student. It learns from vast amounts of existing information and then, based on specific instructions, can create entirely new data. This capability is proving to be a game-changer. Multimodal AIs, which can understand and generate various types of data – text, images, even audio – are opening up exciting avenues. We're talking about suggesting novel chemical combinations for drug discovery, or even helping diagnose and treat diseases by analyzing medical records and pathology images. As Kyunghoon Bae, president of LG AI Research, points out, the key here is the quality of the data used for training; garbage in, garbage out, as the saying goes.
This is precisely why the partnership between LG AI Research and The Jackson Laboratory (JAX) is so significant. JAX, a renowned nonprofit biomedical research organization with a rich history dating back to 1929, has long been at the forefront of using mouse models to understand human diseases. Their work at the intersection of mouse genetics and human genomics, combined with advanced computational tools, makes them an ideal collaborator. Together, they're aiming to build AI models that can swiftly diagnose cancer and predict treatment outcomes, drawing on pathology images, genomic data, and clinical information. They're also delving into JAX's extensive mouse datasets, looking at genomic, behavioral, and metabolic information across a mouse's lifespan to better understand the trajectory of Alzheimer's disease and predict responses to therapies.
One of the persistent challenges with current generative AI is something called 'hallucination' – when the AI confidently presents false information as fact. This can stem from algorithmic limitations or, as mentioned, poor training data. LG AI Research has taken a proactive approach to mitigate this. They have a unique advantage in accessing high-quality data from various LG affiliated companies in electronics, chemistry, and ICT. Furthermore, EXAONE has been trained on licensed data from over 45 million research papers and patents, specifically to bolster its expertise in chemistry.
To further combat hallucination, EXAONE incorporates a framework called retrieval-augmented generation (RAG). Essentially, RAG acts as a fact-checker, retrieving reliable information from external sources to ensure more accurate responses. Instead of solely relying on its learned patterns, the AI cross-references with curated external data. LG AI Research has been using RAG since early last year, giving them a head start in this area. Even with these safeguards, they acknowledge that hallucination can't be entirely eliminated. Therefore, EXAONE's outputs clearly reference the data sources or models used, allowing users to scrutinize the reasoning and make their own informed judgments.
Chemistry, in particular, is a highly visual field, and EXAONE is designed to handle this. Trained on 350 million licensed image-text data pairs, it can learn from both written descriptions and visual representations, including molecular structures, charts, and graphs. This visual prowess is enhanced by advanced 'transfer learning,' where AI models share knowledge to adapt to new, related tasks. For chemistry, LG AI Research developed a 'geometrically aligned transfer encoder' that allows EXAONE to predict molecular properties by understanding fundamental chemical principles. As research scientist Daewoong Jeong explains, molecules, despite their diverse properties, are governed by underlying principles – like how boiling point and surface tension are interconnected molecular characteristics. This ability to grasp these fundamental connections is what makes generative AI such a powerful new tool in the quest for scientific breakthroughs.
