New methods and algorithms combined with access to high quality data and computational power have given AI a new boost. Life sciences is no exception.
Therefore, the goal of the session was to explore the forefront of research and try to explain what we can expect in the field of AI specifically for life sciences in the years to come.
Professor Stephan Oepen, head of the Department of Informatics at the University of Oslo (UiO), chaired the session that included six presentations on AI applications in life sciences.
AI's Role in Patient Centric Cancer Research and Treatment: Prospects, Hopes, and Future Dilemmas
First up was Ingrid Stenstadvold Ross, CEO of the Norwegian Cancer Society, who explored AI's role in patient-centric cancer research and treatment.
– If you want to shape our own future, avoid the negative consequences, but reap the positive ones. We must start by understanding the technology and then putting us humans in the driver's seat.
Stenstadvold Ross discussed AI's growing influence in daily life and healthcare, highlighting its potential for hope and prosperity alongside risks like discrimination. She showcased AI's role in improving cancer diagnosis, treatment, and research, stressing the need for rapid implementation to enhance patient care. Ross also addressed the challenges of balancing data privacy with healthcare advancements, regulatory structures for safe AI integration, and building public trust. She emphasised the critical role of human leadership in navigating these challenges to fully harness AI's benefits in healthcare??.
From Data to Decisions: The Power of AI in Life Sciences and Healthcare
Ishita Barua, Chief Medical Officer, and Co-Founder of Livv Health, discussed AI's power in leveraging data for healthcare decision-making.
– We are not dependent on luck or talent. We don't need Alexander Fleming to by chance discover penicillin, as was the case in 1928. We can actually go and look for these antibiotics with the help from AI.
Ishita Barua's presentation focused on leveraging data for impactful actions in healthcare. She highlighted how advanced data analysis had predicted COVID-19 earlier than traditional methods and discussed the potential for AI to prevent drug shortages and improve antimicrobial resistance efforts. Barua showcased AI's role in discovering new antibiotics and emphasised the significance of AlphaFold's contributions to understanding protein structures for drug development. AlphaFold is an AI system developed by DeepMind that predicts a protein’s 3D structure from its amino acid sequence.
She further advocated for clinical validation to ensure AI's promises translated into real-world outcomes, underscoring the importance of moving from data to decisive action in healthcare advancements.
Deep learning approaches for medical imaging datasets
Anne Solberg, Professor at the University of Oslo, showcased deep learning's impact on medical imaging analysis.
Anne Solberg's presentation focused on the transformative impact of AI on medical imaging. She recounted her transition from scepticism a decade ago about AI's ability to interpret image contents to the present, where AI, particularly through deep neural networks and substantial computational resources, can identify complex patterns, including partial images.
Solberg highlighted the unique challenges associated with medical imaging compared to processing natural images, such as the need for expert labelling, privacy considerations, and the potential for bias in training data, which complicates the effective training of AI models.
She emphasised AI's crucial role in enhancing the efficiency of medical examinations by automating analyses and reducing radiologists' and sonographers' workload. Despite the progress, she acknowledged the limitations of AI, including prediction inaccuracies and the challenges in developing models that confidently manage uncertainty.
Solberg also discussed the potential future of AI in healthcare, envisaging scenarios where AI could provide comprehensive analyses by combining various data types (e.g., text, images, voice) to deliver precise diagnostic predictions. She underscored the importance of secure data access, powerful computing resources, and domain expertise to realise these advancements, suggesting a future where AI significantly enhances patient care and medical research.
Deciphering immune recognition through machine learning
Geir Kjetil Sandve, Professor at the University of Oslo, explored machine learning's role in understanding immune recognition.
– To me, deep learning is this transparent, open approach where I can understand what's happening and have a role in it.
Sandve's presentation delved into the captivating challenge of how our immune system recognises threats, such as differentiating foreign bacteria from the body's own molecules. He explored this question from an evolutionary standpoint, highlighting the immune system's ancient necessity to prepare for future pathogens, like COVID-19, with minimal genetic information. Sandve recounted the evolution from his early experiences with machine learning, which he characterised as "fiddling," to today's more scientific approach facilitated by streamlined tools and deep learning.
He stressed the interdisciplinary nature of contemporary machine learning, necessitating a deep understanding of the domain, the problem at hand, and the models used. His project, unachievable without the collaborative environment provided by the UIO Life Science convergence, aimed to re-engineer the immune system's solution to pathogen recognition through machine learning.
Sandve discussed the challenges in acquiring quality biological data and the significance of synthetic data for developing methodologies. His team explored the applicability of large language models, like ChatGPT, to proteins, questioning the need for a "protein linguistics" to better comprehend the rules governing protein data.
The presentation covered the development of models that integrate different layers of analysis, from spatial binding with convolutional neural networks (CNNs) to the significance of grammar in protein sequences, highlighting the necessity for novel methodologies to tackle complex issues.
Sandve concluded by reflecting on the increasingly enjoyable, tailored, and scientific nature of machine learning, emphasising the importance of interdisciplinary collaboration from data collection through to methodology and dissemination.
AI-assisted surgery
Ingerid Reinertsen, Research Manager at SINTEF, discussed AI-assisted surgery's advancements.
– With AI, we can train models that can do accurate and quantitative measurements of these tumours. These measurements can be done in just a few minutes, while the radiologist will spend much more time. We'll have no variability, the same image in, the same answer out.
Ingerid Reinertsen's presentation centred on the application of AI in enhancing outcomes for brain tumour surgery, where precision and efficiency are critical for patient survival. She detailed the significant challenges in diagnosing and treating the most aggressive brain cancers, underlining the necessity for advancements beyond the current therapies of surgery, chemotherapy, and radiation.
Reinertsen highlighted the role of AI in improving the accuracy and speed of tumour measurements from MR images, which traditionally are time-consuming and subject to variability among radiologists. By establishing a comprehensive database with contributions from numerous hospitals across Norway, Scandinavia, Europe, and the US, her team developed AI models capable of providing consistent, 24/7 analysis for better surgical planning and outcomes.
The presentation covered the use of AI in pre-surgical planning to accurately characterise and measure tumours, ensuring surgeons have essential information before and during operations. Reinertsen also discussed the innovative application of ultrasound in conjunction with AI during surgery to monitor the progress of tumour resection, offering real-time updates and enhanced interpretation of images.
Furthermore, she presented findings from a study comparing the performance of AI models with human doctors in detecting and measuring residual tumours, demonstrating that AI can perform on par with experienced radiologists and neurosurgeons. This supports the potential of AI as a valuable tool for improving prognostic assessments of brain tumour surgeries.
Concluding her presentation, Reinertsen advocated for the integration of AI tools into clinical practice, citing their ability to provide new insights, reduce diagnostic variability, and ultimately contribute to improved patient care. She stressed the importance of prospective testing and validation to fully realise AI's potential in enhancing the quality of medical research and healthcare delivery for brain tumour patients.
Norwegian Large Language Models: current status and the road ahead
Lastly, Lilja ?vrelid, Professor at the UiO and co-director at Integreat, provided insights into Norwegian Large Language Models.
– The real game changer for generative language models has been the stage of instruction fine-tuning. This process involves taking models pre-trained on vast amounts of text and refining them with specific instructions in natural language.
?vrelid's presentation discussed the significant advancements in computational linguistics research propelled by large language models (LLMs). She began by explaining the basics of language modelling, highlighting their ability to predict words in context and their training on vast amounts of unlabelled textual data using neural machine learning methods. ?vrelid noted the dual use of these models: for generating text and providing numeric representations of text for further machine learning processing.
She touched on the staggering development of LLMs in recent years, including the increase in model sizes, with some models reaching up to 45 billion parameters. A key focus of her talk was the game-changing potential of instruction fine-tuning, where pre-trained models are further refined with specific instructions formulated in natural language, enabling them to generalise to unseen tasks.
?vrelid also discussed the importance of human involvement in developing models like ChatGPT, which allows interaction through natural language without requiring technical expertise. She highlighted the role of benchmarking in understanding the capabilities, limitations, and biases of these models, using the medical domain as an example where LLMs have shown promising results, sometimes even preferred over physicians' answers.
The presentation shifted towards the Norwegian perspective, emphasising the need for LLMs that align with Norwegian language, culture, and privacy standards. ?vrelid introduced the Mímir project, aiming to assess the value of copyrighted material in training LLMs and the necessity for benchmarks relevant to the Norwegian context.
Concluding, she outlined the requirements for a Norwegian medical language model, stressing the importance of access to Norwegian medical texts and instructions, and the challenge of benchmarking to ensure the model's efficacy in the Norwegian medical domain. ?vrelid's presentation underscored the transformative impact of LLMs in research and their potential to adapt to specific linguistic and cultural contexts.
Thank you!
At the session's conclusion, Stephan Oepen expressed gratitude to the presenters for their valuable insights and thanked the attendees for their engagement with the critical subject of artificial intelligence in life sciences.
To stay informed about these topics, please follow dScience on LinkedIn or subscribe to our newsletter.
About the Norway Life Science 2024 conference
Norway Life Science is the leading conference for life sciences in Norway. It connects leading researchers with industrial innovators, visionary leaders and strategic minds.
The aim is to showcase excellent research and inspire to the development of new ideas. With ground-breaking research and collaboration, the conference aims to stimulate to the emergence of a new Norwegian health industry, and support Nordic solutions to global challenges.
In 2023, the University of Oslo relaunched the conference as an Oslo Science City Arena event together with partners.