LifeSci-FM

Infrastructure and User Services

LifeSci-FM will be developed in phases, beginning with collection of relevant life sciences data through interactions with stakeholders. This includes genomic, transcriptomic, proteomic, phenomic, and metabolomic datasets, clinical records, evolutionary data, and environmental measurements. Datasets will be used to implement and train efficient foundation models (FMs) — in terms of time, compute, and feasibility — optimized for life sciences applications. The infrastructure will be hosted on the national high-performance computing resource SURF, ensuring scalability, security, and accessibility for the research community. The LSRI is not an initiative for hardware, rather, it aims to provide access to FMs developed and tunable on existing infrastructure. Notably,  training and expansion of expertise in usage of FM will be supported by LifeSci-FM.

Many existing FMs for life sciences have restrictive licensing, limiting uptake in the clinic and in industry. Also, they are pre-trained on specific data types, e.g., genomes/proteins from many diverse species, both human and non-human, incl. both prokaryotes and eukaryotes. LifeSci-FM will provide adaptations and training of FMs with appropriate licenses, tailored to data types that are not currently covered, but that are essential for life sciences in NL. These include specific groups of organisms (e.g., bacteria, fungi, plants), communities (e.g., gut and plant root microbiomes), tissues (e.g., tumours, organs, single cells, spatial), populations (e.g., patients from specified cohorts), and protein structures (e.g., subcellular localizations, accessibility). LifeSci-FM will provide integrated/multi-modal FMs based on combinations of above specified life sciences data. A suite of tailored models will be built and made available during the construction phase of the LSRI, and additional, custom-made, models added following user-specific tailoring efforts, where possible—some adaptations might be private—and applicable. 

Examples/description of the type of research

The envisaged users are academic, industrial, clinical life science researchers that want to analyse their data effectively, using state-of-the-art AI methodologies in a safe and efficient manner. FMs can accelerate research but due to lack of infrastructure and expertise this is currently not happening. By providing access and expertise to develop life sciences FMs, LifeSci-FM aims to break down the barriers that restrict their use to specialized teams and corporations, thus broadly democratizing AI access. LifeSci-FM will provide a service-like, virtual environment that facilitates AI-driven analysis of private, local, and public datasets. 

Users will be able to apply, and tailor FMs to address key challenges, such as identifying genomic features associated with disease by integrating clinical and evolutionary data, uncovering metabolic adaptations in response to environmental conditions, or predicting gene or ecosystem function from large-scale multi-omics datasets. Analysis that is currently unfeasible at local scale—be it in healthcare, industry or in academic research groups—and thus suited to port to national and even international scale. 

LifeSci-FM will facilitate these efforts in multiple ways, importantly, by including supporting expertise (helpdesk). The platform will support secure data integration, model fine-tuning, and interactive AI-driven workflows, lowering technical barriers. LifeSci-FM will allow for user side-only analysis, facilitating informative analysis of private/protected data. By providing pre-trained models and computational tools, LifeSci-FM will democratize access to AI in the life sciences, accelerating discoveries in precision medicine, agriculture, biotechnology, and environmental research. It will foster collaboration between research institutions, healthcare professionals, and industry, ensuring that AI-driven insights contribute to both fundamental biological understanding and real-world applications.

How to access the infrastructure

This is a digital LSRI. Technically, the access mechanism will be composed of a multi-step process including user subscription/access request, submission of research question/proposals, designation to automated or tailoring flow including supportive expertise. More broadly, access will be classified in several categories/services, e.g., users that want to use a FM to annotate their own data with FM features/embeddings, a helpdesk for applications to fine-tune/adapt FM models for a specific task, and requests for compute capacity and expertise to train available FM models. A mix of open access, excellence-based access and fee-based access will be applicable.