Deep-learning for biology: promises, challenges and pitfalls/ Méthodes d’apprentissage profond pour la biologie: promesses, défis et écueils
Séminaire organisé par Hervé Turlier (Collège de France, CNRS) et Virginie Uhlmann (Université de Zürich, Suisse), du 10 au 15 juin 2024
Participants
Ramya DESHPANDE (Harvard Medical School – Cambridge, États-Unis), Claire HE (Columbia University – New York, États-Unis), Hervé ISAMBERT (Institut Curie – Paris, France), Smita KRISHNASWAMY (Yale University – New Haven, États-Unis), Jean-Baptiste MASSON (Institut Pasteur – Paris, France), William PREW (Cancer Research UK Cambridge Institute – Cambridge, Royaume-Uni), Loïc ROYER (Chan Zuckerberg Biohub San Francisco – San Francisco, États-Unis), Ivo SBALZARINI (Center for Systems Biology Dresden – Dresden, Allemagne), Hervé TURLIER (Collège de France – Paris, France), Virginie UHLMANN (Universität Zürich – Zürich, Suisse), Vincenzo VITELLI (University of Chicago – Chicago, États-Unis), Martin WEIGERT (EPFL – Lausanne, Suisse), Kevin YAMAUCHI (ETH Zurich – Zürich, Suisse)
Résumé

This image was created by a generative AI model called DALL-E from OpenAI (Ramesh et al. ICML 2021), version 3, to closely illustrate the environment and the theme of the seminar using a text prompt of approximately 25 lines refined by H. Turlier. It generically illustrates the impressive and promising capabilities of the latest AI models but also their common challenges and pitfalls, such as hallucination phenomena (see the flaws at the bottom of the left window for example), or the specific difficulty here to generate correct text.
L’intelligence artificielle (IA), et plus spécifiquement l’apprentissage profond (ou deep learning, DL, en anglais,), est en passe de révolutionner de nombreux domaines des sciences et de l’industrie. En biologie, l’apprentissage profond se développe dans quelques domaines bien identifiés où les données ont une structure qui s’adapte le mieux aux algorithmes actuels, tels que la microscopie et l’analyse d’image, la détermination de structures de protéines et l’analyse de données de séquences dites “-omiques” (génomique, transcriptomique, etc.). Ce séminaire visait à explorer l’interface entre la biologie et l’IA avec des experts provenant de nombreux domaines, des mathématiques appliquées en passant par la physique, l’informatique, et la biologie, afin de faire un bilan de cette thématique scientifique émergente à la croisée de plusieurs disciplines. Avec des participants représentant un large éventail de niveaux de carrière, allant des jeunes chercheurs (4) aux chercheurs en milieu de carrière (4) et aux scientifiques confirmés (5), l’objectif était de comprendre comment l’avènement de l’IA redéfinit déjà les limites de ce qui est possible dans la recherche en sciences de la vie et d’explorer comment elle est susceptible d’impacter les méthodes des chercheurs ainsi que la nature de l’approche scientifique elle-même. Il s’agissait également de réfléchir aux perspectives futures que l’IA permet d’entrevoir, et de mettre le doigt sur les obstacles à surmonter, qu’ils soient d’ordre scientifique, structurel, ou même humain, afin de pouvoir pleinement saisir les opportunités offertes par cette nouvelle technologie. Enfin, ce séminaire visait également à débattre de l’organisation de la communauté scientifique concernée par ce sujet, dont l’hétérogénéité et l’interdisciplinarité pose des défis inédits en termes de communication, de collaboration, de structuration et de financement au-delà des frontières des domaines scientifiques classiques.
Summary
Artificial intelligence (AI), and more specifically deep learning (DL), is poised to revolutionize many fields in science and industry. In biology, deep learning is developing in several well-identified areas where the data structure best suits current algorithms, such as microscopy and image analysis, protein structure determination, and the analysis of so-called “-omics” sequence data (genomics, transcriptomics, etc.). This seminar aimed to explore the interface between biology and AI with experts from various fields, including applied mathematics, physics, computer science, and biology, to take stock of this emerging scientific theme at the crossroads of multiple disciplines. Together with participants representing a wide range of career stages, from junior researchers (4) to mid-career (4) and established scientists (5), the goal was to understand how the advent of AI is already redefining the boundaries of what is possible in life sciences research and to explore how it will likely impact researchers’ methods and the nature of the scientific approach itself. It also sought to reflect on the future perspectives that AI allows us to foresee and to highlight the obstacles to be overcome, whether scientific, structural, or even human, to fully grasp the opportunities offered by this new technology. Finally, this seminar also aimed to discuss the organization of the scientific community concerned with this subject, whose unprecedented heterogeneity and interdisciplinarity pose novel challenges in terms of communication, collaboration, structuring, and funding beyond the boundaries of traditional scientific fields.
Report
The overarching aim of the seminar was to collectively reflect on the impact and influence of AI in biology, from four different angles: what it has enabled so far, what it is poised to enable in the future, what stands in the way of its development, and how it impacts the (biology) scientific community. We deliberately organised the seminar to include a highly diverse group of participants, both from in terms of primary scientific field and career stage. The common denominator among them all was expertise in developing or applying DL methods in the context of biology research. To facilitate exchanges in the many discussion sessions we had scheduled during the week, we chose to begin the seminar with a day of scientific presentations that aimed to illustrate the expertise and interests of each participant. We loosely grouped speakers in sessions focusing on specific types of data and questions. The first session concentrating on DL in bioimage analysis featured talks from Loïc Royer (CZI Biohub, USA), Martin Weigert (EPFL, Switzerland), and Virginie Uhlmann (Universität Zürich, Switzerland). The session that followed was focused on the application of DL in omics data analysis, and included presentations from Claire Ming-Yi He (Columbia University, USA), Kevin Yamauchi (ETHZ, Switzerland), and Smita Krishnaswamy (Yale University, USA). The topic shifted to DL applied to the study of medical data in the third session, where Hervé Isambert (Institut Curie, France) and William Prew(CRUK Cambridge Institute, UK) gave an overview of their respective work. The afternoon continued with a session on the use of DL in biophysics, with talks from Hervé Turlier (Collège de France, France), Vincenzo Vitelli (University of Chicago, USA), and Ivo Sbalzarini (Max Planck Institute of Molecular Cell Biology and Genetics, Germany). Finally, the last presentation session highlighted the application of DL methods to study behaviour through the research of Ramya Deshpandes (Harvard University, USA) and Jean-Baptiste Masson (Institut Pasteur, France).
The rest of the week was devoted to discussions around our main topic, namely AI applied to biology. Discussions were organised along four themes, which first aimed at collectively reviewing the present state of AI in biology and then gradually moved towards future perspectives.
Theme 1 meant to help outline how AI transformed the way research is done in biology and what it enabled. Two discussion groups were formed to think about this question from a “data-first” (i.e., AI used in the context of data mining) and “model-first” (i.e., AI used in the context of predictive modelling) angle, before merging the thoughts of both groups into a joint discussion. The consensus that emerged highlighted that, while AI enables automation and exploratory analysis to unprecedented levels, true scientific understanding requires integrating AI-derived insights into established scientific frameworks. While AI has undoubtedly captured imaginations, its current contribution to scientific discovery may currently be more focused on enhancing established workflows than revolutionising them. AI has a strong potential to automate the tedious aspects of existing data analysis workflows and, by making feasible what used to be impossible, may therefore empower scientists to change the way they think about which questions can possibly be answered in biology. Integrating current biological and physical knowledge into this new AI-powered scientific discovery process is as challenging as it is promising to lead us towards a true AI revolution in biology.
In Theme 2, discussions moved towards identifying the future opportunities of a wider use of AI across biology research. Participants unanimously recognized the potential of integrating AI into models of biological phenomena that translate across scales and data modalities, as this would provide a holistic view of living systems that would go well beyond what individual models – scale or modality specific – can achieve. Strong enthusiasm was also expressed for AI-enabled experimental design in which AI is used as a novel way to create experiments as well as synthetic systems, making the bridge between in silico and in vivo modelling. Finally, there is excitement about the potential of AI to gain a fundamental understanding of highly complex data, an area where foundation models that are currently gaining a lot of attraction in the research space may make a significant difference in our ability to explore and mine information-rich datasets.
Having reviewed the opportunities lying ahead, Theme 3 focused on identifying what hinders the broader democratisation of AI in biology. In terms of threats, seeing extensive usage of AI tools without a good understanding of their limits of applicability is perceived as a major risk that can be mitigated by thorough training and education. The extensive computational resources required by AI systems also create a clear energy bottleneck: without the optimisation of AI algorithms, there will be a physical limit in what can be achieved with available energy – and financial – resources. This threat can be mitigated at our level through the development of energy-efficient data and algorithms, but could benefit from new, low-energy hardware and even biological computing resources. Finally, a significant challenge was identified in the fact that industry is a key player in AI research, but generally has a secretive culture and can create bias in research by investing heavily in a very focused area. New ways must be invented to work when science is open but major contributors are profit-driven. There, open-source mandates were identified as a critically important mechanism to ensure the transparent use of AI in biology. It was however recognized that implementing them is costly and that, since open data are immensely valuable from a commercial perspective, the economic value of open academic research must be appreciated and appraised accordingly.
Finally, acknowledging that the obstacles identified can only be overcome through a collective effort of all parties involved, Theme 4 focused on the question of community. There, discussions revolved around how AI affects the scientific landscape in biology and whether these changes open up opportunities for positive shifts in the research culture. AI provides an entirely new avenue to put the scientist back at the centre to ensure that analysis methods and their results remain interpretable. This forces the research community to collectively re-think the scientific method: the paradigm from hypothesis formulation to model design and finally data analysis no longer holds as AI intertwines these three steps. Additionally, different scientists with different expertise intervene in each part of the process, meaning that those who can identify the next big biology question to study are not necessarily those who can solve the next big technology problem standing in the way of this discovery. For these different individuals to collaborate and to create synergies, new ways must be found to nurture mutual respect across disciplines and increase opportunities to mingle and exchange across the boundaries of classical scientific fields.
The depth and quality of the discussions, to which each participant actively contributed, led to the production of an extensive collection of notes that we ambition to curate and edit into a perspective article to be submitted to a peer-reviewed journal. In addition to being highly productive scientifically, the seminar saw the inception of a budding community of extremely diverse DL researchers evolving in biology. This success is to a large extent due to the privileged environment offered by the Fondation des Treilles, and to the quality of the hosting provided by the team at the Domaine des Treilles. The participants collectively agreed that more such events should be repeated in the future to nurture exchanges between the many different kinds of scientists involved with DL in the context of biology research.
Lien vers la vidéo Youtube du séminaire: https://youtu.be/G_ddZ6haflg
OpenEdition vous propose de citer ce billet de la manière suivante :
ldiebold (24 juillet 2024). Deep-learning for biology: promises, challenges and pitfalls/ Méthodes d’apprentissage profond pour la biologie: promesses, défis et écueils. Les carnets de la Fondation des Treilles. Consulté le 18 février 2025 à l’adresse https://doi.org/10.58079/1235o