BioNTech hosted its annual AI Day at the Science Museum in London, the second event in their Innovation Series. The day brought together investors, analysts, media representatives and guests to explore how artificial intelligence could transform science and medicine.
As part of BioNTech’s AI ecosystem, InstaDeep showcased the progress made across multiple fronts, from advanced compute infrastructure to pioneering biological applications, highlighting how an integrated, full-stack AI approach can advance innovation.
Powering AI: Compute, Orchestration, and Software
Breakthroughs in AI depend on infrastructure that can handle compute at scale, supporting both large-model training and rapid experimentation. To strengthen our position at the forefront in AI innovation, our integrated ecosystem enables researchers to scale ideas efficiently, maximise performance across hardware and software, and accelerate cutting-edge discoveries.
At the foundation lies Kyber, InstaDeep’s supercomputer, launched last year and now among the largest in the field. With ~500 PetaFLOPS of Nvidia H100 GPUs and more than 80,000 CPU cores housed in custom in-house-engineered racks, Kyber is fully powered by renewable energy and tightly optimised for hardware–software integration.
On top of this compute layer sits AIchor, our orchestration platform that makes Kyber accessible to scientists and engineers through a streamlined GitOps workflow. In 2025, AIchor facilitated an average of 15,000 experiments per month across our research teams, underscoring its effectiveness at scale. Initially developed for internal use, AIchor is now available for external collaboration, bringing scalable orchestration to organisations tackling complex AI workloads.
Finally, our JAX-based ML software ecosystem ensures researchers can extract maximum value from every GPU. For example, our Machine Learning Interatomic Potentials (MLIP) library enables molecular simulations with near quantum-level accuracy at speeds orders of magnitude faster than reference quantum chemistry methods such as density functional theory. Thanks to in-depth engineering, MLIP can simulate tens of thousands of atoms on a single GPU, delivering outstanding accuracy in the analysis of large-scale molecular systems.
AI for Biology: Innovation and Application
Beyond infrastructure, AI Day was an opportunity to showcase how we are applying AI to five key areas of biology, each with the potential to unlock wide-reaching advances in science and medicine:
Genomics
We unveiled early results of our next generation model belonging to the Nucleotide Transformer (NT) family, which continues to lead the field with more than one million downloads and over 500 citations. NTv3, pretrained on 15 trillion tokens spanning over 150,000 species, learns from both genomic and functional data and is capable of both predictive and generative tasks. It scales efficiently to sequences of up to one million nucleotides, and has already demonstrated state-of-the-art in vitro performance in de novo promoter-specific enhancer design, surpassing validated methods for activity-specific sequence generation.
Proteins and antibodies
AbBFN2, the first genuinely multimodal Bayesian Flow Network (BFN) for antibody design, now achieves state-of-the-art performance across 23 sequence-labelling tasks, identifies germline families and can be used to stabilise heavy/light chain pairings. With AbBFN2, humanisation, an expensive process that can traditionally take months in the lab, with no guarantees of success, can be completed in under 20 minutes while preserving binding affinity, and simultaneously optimising biophysics alongside humanness. Achieving a 90% success rate with tractable starting candidates, AbBFN2 demonstrates how a single integrated model can streamline pipelines that once required multiple tools that are costly and inefficient.
🧑🔬Try it for yourself: design and optimise therapeutic antibodies via our free demo on DeepChain.
Data acquisition & refinement
InstaDeep showcased progress in advancing biological discovery through new algorithms and tools designed to maximise information extraction. Specifically, we presented our next-generation, library-free de novo peptide sequencing algorithm. Trained on 63 million labelled spectra, our latest InstaNovo model delivers a 10–15% increase in accuracy, up to twice as many peptide identifications, and 50-fold faster inference. We demonstrated how InstaNovo is being applied at BioNTech to uncover novel targets and biomarkers within the Dark Proteome.
We also highlighted our progress in developing tools for efficient histology dataset labelling, alongside the application of foundation models for clustering haematoxylin and eosin (H&E) slides by morphological patterns. These approaches streamline the exploration and analysis of histology data at scale. When paired with an external conversational agent, this technology can significantly reduce the workload for pathologists while allowing practitioners to query and validate results in real time.
De Novo Nanoparticles Design
The immune system is primed to respond to repetitive, symmetrical structures, a feature common to viruses and bacteria. Nanoparticles have an exceptionally high surface-area-to-volume ratio, meaning their surface area is large relative to their very small size, and they often exhibit highly symmetrical structures. These characteristics underscore their importance in therapeutics, particularly in targeted drug delivery.
This motivation drives our development of AI-driven de novo protein design tools, capable of assembling trimers, which are macromolecular complexes formed by three non-covalently bound proteins, into larger symmetrical nanoparticles. At AI Day, we demonstrated how, using DeepChain’s Folding Studio, we can screen up to 10,000 designs per day, accelerating the path from structural concepts to functionalised vaccine candidates. By designing nanoparticles that mimic natural viral patterns, we aim to develop next-generation mRNA vaccines that elicit more robust and durable immune responses, supported by a diverse toolkit of AI-designed nanoparticles.
AI-guided T cell receptor (TCR) optimisation
Our AI-guided pipeline overcomes the limitations of traditional display-based affinity enhancement. By rationally exploring the vast TCR sequence space, our AI models and workflows achieve average binding affinity improvements of up to 50,000-fold in just three optimisation rounds, while also surpassing existing state-of-the-art methods in TCR–pMHC structure prediction. These enhanced TCRs have shown promising, durable tumour control in preclinical models and hold the potential to unlock new opportunities for therapies targeting antigens beyond the reach of natural TCRs.
Looking Ahead
AI Day 2025 underscored the accelerating pace of progress at the intersection of AI and biology, where each year’s advances build on the last to create entirely new possibilities. Together with BioNTech, we remain committed to tackling some of the hardest challenges in science and translating fundamental research into meaningful applications.
As InstaDeep CEO Karim Beguir highlighted: “AI is not a single exponential, but a triple exponential comprising data, compute, and models. Together BioNTech and InstaDeep, are building across the full stack of AI, accelerating discoveries and innovation in support of BioNTech’s mission of building a global immunotherapy powerhouse.”
We are proud to be pushing the boundaries of what is possible in AI and biology.
👉 Interested in learning more? Access the full webcast here
Disclaimer: All claims made are supported by our presentation and webinar unless explicitly cited otherwise.