exhibit991biontechaiday-

This slide presentation includes forward-looking statements This presentation contains forward-looking statements within the meaning of the Private Securities Litigation Reform Act of 1995, as amended. In some cases, forward-looking statements can be identified by terminology such as “will,” “may,” “should,” “expects,” “intends,” “plans,” “aims,” “anticipates,” “believes,” “estimates,” “predicts,” “potential,” “continue,” or the negative of these terms or other comparable terminology, although not all forward-looking statements contain these words. The forward-looking statements in this presentation are neither promises nor guarantees, and you should not place undue reliance on these forward-looking statements because they involve known and unknown risks, uncertainties, and other factors, many of which are beyond BioNTech’s control; and which could cause actual results to differ materially from those expressed or implied by these forward-looking statements. You should review the risks and uncertainties described under the heading “Risk Factors” in BioNTech's Quarterly Report on Form 6-K for the period ended June 30, 2025; and in subsequent filings made by BioNTech with the SEC, which are available on the SEC’s website at https://www.sec.gov/. Except as required by law, BioNTech disclaims any intention or responsibility for updating or revising any forward-looking statements contained in this presentation in the event of new information, future developments or otherwise. These forward-looking statements are based on BioNTech’s current expectations and speak only as of the date hereof. Furthermore, certain statements contained in this presentation relate to or are based on studies, publications, surveys and other data obtained from third-party sources and BioNTech’s own internal estimates and research. While BioNTech believes these third-party sources to be reliable as of the date of this presentation, it has not independently verified, and makes no representation as to the adequacy, fairness, accuracy or completeness of, any information obtained from third-party sources. In addition, any market data included in this presentation involves assumptions and limitations, and there can be no guarantee as to the accuracy or reliability of such assumptions. While BioNTech believes its own internal research is reliable, such research has not been verified by any independent source. This presentation contains references to our trademarks and to trademarks belong to other entities. Solely for convenience, trademarks and trade names referred to, including logos, artwork and other visual displays, may appear without the ® or TM symbols, but such references are not intended to indicate, in any way, that their respective owners will not assert, to the fullest extent under applicable law, their rights thereto. We do not intend our use or display of other companies’ trade names or trademarks to imply a relationship with, or endorsement or sponsorship of us by, any other companies. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 2

BioNTech – Building a global immunotherapy powerhouse translating science into survival 14:00 Advancing a disruptive tech-bio company Prof. U. Sahin, M.D. 14:15 Developing the future of AI at BioNTech K. Beguir InstaDeep – Delivering across the full AI stack 14:25 Compute & model scaling A. Laterre 14:35 AI innovation B. Almeida, B. Guloglu 15:00 Data acquisition & refinement N. Lopez Carranza, Y. Ben Dhieb 15:20 Applications C. Zhang, L. Walls, A. Delaunay, M. Rooney 15:40 Audience Q&A Prof. U. Sahin, M.D., K. Beguir Agenda AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 3

Margaret Keenan Dec. 8, 2020 London Science Museum

BioNTech’s AI capabilities with worldwide reach BioNTech locations BioNTech AI centers Rwanda Kigali Austria Vienna Germany (HQ + 6 sites) Singapore U.S. Cambridge Gaithersburg China Shanghai Turkey Istanbul Australia Victoria United Kingdom London South Africa Cape Town UAE DubaiTunis Tunisia France Paris AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 6

Multiplatform oncology company Infectious diseases pipeline COVID-19 vaccine global impact Leader in integrated AI capabilities In-house manufacturing 16 >20Clinical programs Ongoing Phase 2 or 3 trials 7 Clinical programs in high unmet need indications 5 Billion doses distributed 4 Platforms including individualized mRNA and bispecific antibodies BioNTech – disruptive tech-bio company with pioneering technologies developed through full AI integration AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 7

We are uniquely positioned to combine approaches to transform cancer care ADC = antibody-drug conjugate. Immunomodulators • Focus on the critical IO pathways • Targeting different complementary pathways in cancer immunity cycle may promote a durable anti-tumor effect mRNA cancer immunotherapies • Eliminate polyclonal residual disease with multi-antigen and individualized approaches • Polyspecific activity by targeting multiple antigens at once • Establish long-lasting immunological memory to prevent relapses Targeted therapies • Precise and potent modalities for fast onset tumor reduction • ADC as potential “augmenters” of immunomodulators and mRNA cancer immunotherapies • Focus on HER2, HER3, TROP2, B7H3 ADCs as combination partners mRNA cancer immuno- therapies Targeted therapies Immunomodulators Potentially curative approaches Potential for synergy Potential for synergy Potential for synergy AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 9

Potential for synergy mRNA cancer immuno- therapies Targeted therapies We are uniquely positioned to combine approaches to transform cancer care ADC = antibody-drug conjugate. Immunomodulators • Focus on the critical IO pathways • Targeting different complementary pathways in cancer immunity cycle may promote a durable anti-tumor effect • BNT327 pumitamig mRNA cancer immunotherapies • Eliminate polyclonal residual disease with multi-antigen and individualized approaches • Polyspecific activity by targeting multiple antigens at once • Establish long-lasting immunological memory to prevent relapses Targeted therapies • Precise and potent modalities for fast onset tumor reduction • ADC as potential “augmenters” of immunomodulators and mRNA cancer immunotherapies • Focus on HER2, HER3, TROP2, B7H3 ADCs as combination partners AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 10 Immunomodulators Potentially curative approaches Potential for synergy Potential for synergy

Local neutralization of angiogenic and immunosuppressive VEGF-A effects VEGF-A PD-L1 Targeting the TME and blockade of PD-1/PD-L1 signaling Anti-VEGF-A (Fab) Anti-PD-L1 (VHH) VEGF-A dimer in the TME PD-L1 ligand on tumor cells Pumitamig VEGF-A dimer in the TME Anti-VEGF-A (Fab) Anti-PD-L1 (VHH) PD-L1 ligand on tumor cells NSCLC IHC2 1. Partnered with Bristol Myers Squibb; 2. IHC data: Human Protein Atlas Pumitamig’s synergistic targeting of PD-L1 and VEGF1 Tumor microenvironment (TME) AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 11

Anti-PD-L1 Anti-VEGF >1.4 M estimated new cancer cases in the US and EU annually that cannot be addressed by current IO therapies Anti-PD-(L)1 approved Anti-PD-(L)1 not approved US and EU cancer incidence2 1. NCI SEER https://training.seer.cancer.gov/index.html. 2.US incidence source: NIH and American Cancer Society data EU incidence source: European Cancer Information System Next-generation bispecific can potentially expand the reach of IO therapy Breast (non TNBC) TNBC PD-L1 CRC (MSS) EGFRmut NSCLC Pancreatic Ovarian GBM Lung Melanoma RCC Endometrial HNSCC/nasopharyngeal TNBC PD-L1 >10% HCC Gastric CRC (MSI-H) Anti-PD-(L)1 therapy addresses ~1.5 M new cancer cases in the US and EU annually with medical need remaining high (5-year survival < 50%)1 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 12

Potential to transform standard of care and establish new IO backbone treatment option for patients with high unmet medical needs Maximizing potential of next-generation immunomodulator pumitamig1 with global co-development and co-commercialization partnership • Bispecific antibody targeting PD-L1 and VEGF-A • Over 1,200 patients treated in clinical trials across multiple tumor types • Broad development ongoing in 10+ indications, including initial registrational trials Anti-VEGF-A Anti-PD-L1 VHH 1. Partnered with Bristol Myers Squibb. Landmark strategic collaboration with BMS to advance pumitamig1 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 13

Interindividual variability & intratumoral heterogeneity Cancer evolution 5-20 years – up to 10.000 mutations Healthy cell Pre-cancer cell Mutations Mutations Mutations Individual patients DNA mutations Root cause of cancer treatment failure AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 14

Targeted therapies Immunomodulators Potential for synergy We are uniquely positioned to combine approaches to transform cancer care ADC = antibody-drug conjugate. Immunomodulators • Focus on the critical IO pathways • Targeting different complementary pathways in cancer immunity cycle may promote a durable anti-tumor effect • BNT327 pumitamig mRNA cancer immunotherapies • Eliminate polyclonal residual disease with multi-antigen and individualized approaches • Polyspecific activity by targeting multiple antigens at once • Establish long-lasting immunological memory to prevent relapses Targeted therapies • Precise and potent modalities for fast onset tumor reduction • ADC as potential “augmenters” of immunomodulators and mRNA cancer immunotherapies • Focus on HER2, HER3, TROP2, B7H3 ADCs as combination partners AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 15 Potential for synergy Potential for synergy Potentially curative approaches mRNA cancer immuno- therapies

1 6 Individual patient samples (blood and tissue) Artificial intelligence- driven neoantigen prediction On-demand tailored RNA manufacturing Individualized immuno- therapy Mapping of mutations Fixed combination of shared tumor antigens2 Multi-antigen approach tailored to each indication Neoantigens Individualized therapy Multiple shared antigens Off-the-shelf therapy iNeST1 FixVac individualized Neoantigen- Specific immunotherapy Fixed Antigen Vaccine ANTIGEN 1 ANTIGEN 2 ANTIGEN 3 ANTIGEN 4 mRNA- lipoplex platform Described in Castle et al., Cancer Res 2011, Kreiter et al., Nature 2025 Described in Sahin et al., Nature 2017, Rojas et al. Nature 2023 Leveraging our leadership in mRNA to fully exploit cancer immunotherapy target space with two approaches 1. Partnered with Genentech, a member of the Roche Group. 2 Antigens vary across programs; 3. T-cell responses analyzed by ex vivo multimer staining analysis in blood. Described in Kranz et al., Nature 2016 and In Sahin et. al., Nature 2020 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 16

Individual patient samples Mapping of mutations Neoantigen prediction On-demand tailored mRNA manufacturing Individualized immunotherapy NORMAL G G G AAACT T T T T CC TUMOR G G G AAACG T T T T CC 1 2 3 4 5 Just-in-time manufacturing Dedicated mRNA GMP production facilities Selection algorithms AI and ML optimization Driven by data Potential for continued improvement as more data are generated and analyzed iNeST is being developed in collaboration with Genentech, a member of the Roche Group. Autogene cevumeran is an investigational candidate. iNeST: autogene cevumeran driving continuous innovation with data AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 17

Neoantigen prediction: how do we identify, predict, and characterize neoantigens? Neoantigen rank Gene Mutation Length (aa) Transcript VAF MHC I score MHC II score Coverage in tumor VAF in tumor Coverage in normal tissue VAF in normal tissue 1 SNF8 V183M 27 16.05 0.1 2.16 155 0.33 119 0.00 2 SEMA7A G340S 27 1.44 0.04 8.6 113 0.44 120 0.01 3 DUS4L S305P 26 2.07 0.28 8.54 213 0.48 150 0.00 20 Types of mutation and clonality of mutations Mutated transcription expression level Characterization of neoantigen peptide Peptide-MHC binding affinity/quality Similarity/richness across tumors Lack of expression in healthy tissues Representative data. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 18

Defining the complex TCR-tumor antigen interaction is an unsolved computational problem Tumor Cell T Cells HLA Allele Diversity > 30,000 TCR Diversity > 5x108 Peptide Diversity > 100,000 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 19

Sahin & Türeci, Science 2018 Tumor Heterogeneity Mutation and neoantigen profile, epigenetics, biology and evolution Immune system HLA restriction and immune SNPs Host and environment Factors include HLA haplotype, microbiome, epigenome, age, antigen exposure, drugs, and comorbidities Tumor Environment Immunosuppression Recognition and editing Individual Patient Specificity, quantity, and functional state of immune effectors Hereditary and environmental factors So m at ic m ut at io ns Immune Diversity Y Z X Dimensions of cancer heterogeneity AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 20

Personalized immunotherapy AI-powered bio-engineering • iNeST1: Personalized immunotherapy platform levering AI to create therapies unique to each patient’s tumor • 4 ongoing trials • >450 patients treated2 • 18,000 neoantigens selected2 • Computational extension of immunotherapy target space3 • Semi-automated manufacturing capabilities for iNeST1 • Development of novel DeepChain platform combining cutting-edge AI and bio-engineering • Optimization of mRNA design & structure • Automated dry-wet lab to enhance discovery capabilities • In-house supercomputing cluster with ~500 PetaFLOPS of Nvidia H100 GPUs Our leading scientific capabilities are fueled by AI to pioneer personalized immunotherapies 1. Partnered with Genentech, a member of the Roche Group; 2. From trials BNT122-01, GO39733, GO40558 and ML41081; 3. Castle et al. 2011 Cancer Res. A I AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 21

Individualized treatment platforms to address inter-individual variability Deep genomics & immunology expertise to analyze patient data Automated in-house manufacturing to serve patients on time and globally AI-infused & digitally-integrated target & drug discovery and development ` Drug classes Inter- individual variability Off-the-shelf drugs Tailored on-demand immunotherapies Clinical samples Engineered cell therapies mRNA therapeutics T cell receptors Antibodies Antibody conjugates Small molecule immunomodulators Personalized omics Capabilities to build tomorrow’s personalized precision medicines Fully-integrated tech-bio company BioNTech is uniquely positioned with complete AI integration and personalized medicine capabilities under one roof AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 22

Moore’s Law: efficiency of hardware compute doubles every two years 1. Ray Kurzweil Q&A - The Singularity, Human-Machine Integration & AI | EP #83 - Peter Diamandis & Ray Kurzweil Singularity Q&A - (ref) - March 2024 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 25

While the compute efficiency of AI models is also doubling every 8 months 1. Situational Awareness June The Decade Ahead Leopold Aschenbrenner - (ref) - June 2024 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 26

This supercycle, the Era of Experience, is driven by techniques such as Reinforcement Learning (RL) and Optimization that focus on "learning by doing" with AI agents. Welcome to the Era of Experience David Silver, Richard S. Sutton - (ref) - Apr 2025 Environment New State New Action AI Agent AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 29

InstaDeep is active in Reinforcement Learning research Oryx: a Performant and Scalable Algorithm for Many-Agent Coordination in Offline MARL NeurIPS Conference 2025 Breaking the Performance Ceiling in Complex Reinforcement Learning requires Inference Strategies NeurIPS Conference 2025 Oral Memory-Enhanced Neural Solvers for Efficient Adaptation in Combinatorial Optimization NeurIPS Conference 2025 Spotlight AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 30

Nucleotide Transformer Building and Evaluating Robust Foundation Models for Human Genomics Nature Methods 2024 ProtBFN & AbBFN Protein Sequence Modelling with Bayesian Flow Networks Nature Communications 2025 ChatNT A Multi-Modal Conversational Agent for Genomics Nature Machine Intelligence 2025 Matchgate Classical Shadows Unified Matchgate Classical Shadows for Quantum Fermionic Systems Nature Partner Journals Quantum Information 2025 InstaNovo ML for de novo peptide sequencing for large- scale mass spectrometry proteomics Nature Machine Intelligence 2025 SegmentNT Annotating the genome at single-nucleotide resolution with DNA foundation model Nature Methods 2025 coming soon Biology and AI know-how: 6 Nature journal publications in less than 12 months AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 31

Compute & Model Scaling Applications Data Acquisition & Refinement AI Innovation InstaDeep and BioNTech are building across the full stack of AI: AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 33

Gold medal at IMO 2025 achieved1 1. DeepMind achieves gold medal-level performance on the 2025 International Mathematical Olympiad with a general-purpose reasoning LLM! (ref) – 21st of July 2025 2. OpenAI general-purpose reasoning models solved all 12 problems at the 2025 International Collegiate Programming Contest (ICPC) World Finals (ref) – 17th of September 2025 3. DeepMind released Gemini Robotics 1.5 - AI models that let robots perceive, plan, think and act across diverse physical environments to complete complex, multi-step tasks with explainable reasoning (ref ) – 25th of September 2025 1st place at International Programming Contest2 Growing capabilities for physical agents3 Scaling laws drive AI innovation AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 35

A fully integrated AI ecosystem InstaDeep's AI supercomputer Purpose-built for AI, delivering full control, visibility, and performance. AIChor orchestration platform A Kubernetes-native AI training platform for seamless scaling and fast experimentation. ML software ecosystem in JAX Software for high-performance computing and advanced model optimization. AI innovation Pioneer work in generative models, representation learning, and reasoning. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 37

~500 PetaFLOPS of Nvidia H100 GPUs 86,000 CPU Cores 1.2 Tons of Hardware per Rack ▪ Custom rack design engineered in-house ▪ Optimised for AI performance and cost efficiency ▪ Powered 100% by renewable energy ▪ Designed to scale seamlessly with next-generation hardware ▪ Tight hardware–software integration for control and efficiency Kyber, InstaDeep's AI supercomputer Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 38

+15,000 experiments / month in 2025 +75% GPU usage Simple GitOps workflow: Commit → Build → Run → Monitor Scalable Kubernetes-native provisioning and auto-scaling across clusters Flexible Modular data plane for multi-cluster and multi-cloud compute AIchor, a complete AI training platform, ready for production and built for scale. AIchor orchestration platform https://aichor.ai/ Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 39

▪ Scale→ from rapid prototyping to large-scale training and deployment ▪ Efficiency→ “better, faster, cheaper" AI workloads that maximise hardware usage ▪ Modularity → interoperable, reliable, and optimised components working together ML software ecosystem in JAX Foundation models Decision-making & reasoning Scientific computing AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 40

1 | Efficiently train 100B-parameter foundation models Hierarchical model sharding Rack #7 Switch H100 DGX 8x H100 GPUs H100 DGX 8x H100 GPUs Switch Rack #11 Switch H100 DGX 8x H100 GPUs H100 DGX 8x H100 GPUs Switch Rack #14 Switch H100 DGX 8x H100 GPUs H100 DGX 8x H100 GPUs Switch Spine switch #1 Spine switch #2Intra-node: fully sharded data parallelism (NVLink) Kyber Inter-node: data parallelism (RoCE) AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 41

1 | Efficiently train 100B-parameter foundation models Hierarchical model sharding ✓ Intra-node: fully sharded data parallelism (NVLink) ✓ Inter-node: data parallelism (RoCE) ✓ Tensor and sequence parallelism available Code optimizations ✓ CuDNN kernels (e.g. Flash Attention) ✓ Mixed precision with FP8 quantisation ✓ XLA compiler and RoCE configuration tuning ✓ NUMA binding affinity ✓ … Model FLOPs Utilization (MFU) This is the ratio of observed throughput (tokens per second) to the theoretical maximum throughput of a system running at peak FLOPs. e.g, Llama 3.1 405B achieves 38 to 41% MFU on 16,384 H100 GPUs. +66 % Model FLOPs Utilization (MFU) on 64 x H100 GPUs AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 42

Simulating molecular properties at scale is key to many industries, including drug discovery, materials, chemicals. Machine Learning Interatomic Potentials allow quantum accuracy orders of magnitude faster on molecular simulations 2 | Scaling molecule screening with Machine Learning Interatomic Potential Scientists Candidates Experiments Looking for any type of new molecule or material Success rate depends on selection quality AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 43

Scientists Candidates Experiments Looking for any type of new molecule or material Success rate depends on selection quality Classical Force Fields Quantum Chemistry MLIP S pe ed Accuracy 2 | Scaling molecule screening with Machine Learning Interatomic Potential Simulating molecular properties at scale is key to many industries, including drug discovery, materials, chemicals. Machine Learning Interatomic Potentials allow quantum accuracy orders of magnitude faster on molecular simulations Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 44

Up to 5x speed-up in simulation speed 160 atoms molecule for 1ns (runtime in min) Re-engineered OS code InstaDeep OS MLIP library InstaDeep backends & models Over 100,000 atoms on one GPU Imipenem binding to L,D-transpeptidase Better Cheaper Scalable Quantum Chemistry-level accuracy 2 | Scaling molecule screening with Machine Learning Interatomic Potential M LI P E ne rg ie s Fo rc e Fi el d s En er g ie s DFT reference energies Method Hardware Runtime Relative Compute cost DFT 64 CPU cores ~ 145 days $12,500 MLIP 1 H100 GPU < 20 min $1 0 50 100 150 200 250 +10,000 times cheaper than DFT Estimated on a 150 atoms molecule Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 45

+1 Million Downloads Across model sizes1 +500 Citations Across model types2 1. Hugging Face Statistics. 2. Google Scholar. Nucleotide Transformer: one of the most popular genomics AI models on Hugging Face AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 49

Generative AI for genomics – publications in top-tier journals Nucleotide Transformer Building and Evaluating Robust Foundation Models for Human Genomics Nature Methods, 2024 SegmentNT Segmenting the Genome at Single- Nucleotide Resolution with DNA Foundational Models Nature Methods, 2025 Isoformer Multi-modal Transfer Learning between Biological Foundation Models NeurIPS Conference, 2024 ChatNT A Multi-Modal Conversational Agent for Genomics Nature Machine Intelligence, 2025 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 50

Exploiting the data available with the aim of building a best-in-class model for genomics Learn from genomes OR Learn from functional data Nucleotide Transformer v3 NTv3 NT Evo Borzoi AlphaGenome Pre-training on genomes from >150,000 species Post-training on >17,000 functional tracks across 16 species AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 51

Introducing NTv3: a new, truly foundational, model for genomics with a million nucleotide context Multi-species more than 150,000 species genomes Multimodal genomes + functional tracks + genome annotation Multi-domains human genomics, plants genomics, metagenomics Long-range up to 1 million input nucleotides Generative capacities design of DNA sequences de novo with in-vitro validation Suite of models from 10M -> 4B parameters Designed for efficiency fastest foundation models available Genomic tracks (e.g. gene expression) Genome annotation (e.g. genes, splice sites finding) Variant prediction (e.g. eQTL, deleterious mutation) Sequence generation (e.g. enhancer design) NTv3 DNA sequence A C T A T C T A G A T A GT G G AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 52

Fine-tuning | NTv3, The Fastest Genomic Foundation Models NTv3 scales up to 1 million nucleotides an order of magnitude more efficiently than competitive models *RC equivariant AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 56

NTv3 scales up to 1 million nucleotides an order of magnitude more efficiently than competitive models *RC equivariant Fine-tuning | NTv3, The Fastest Genomic Foundation Models AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 57

Fine-tuning | NTv3 is amongst the best models for fine-tuning on downstream tasks Evaluation of different foundation models on 44 long-range downstream tasks, including gene expression, DNA accessibility and genome annotation across various human tissues. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 58

Fine-tuning | NTv3 is amongst the best models for fine-tuning on downstream tasks Average performance across quantitative tasks Evaluation of different foundation models on 44 long-range downstream tasks, including gene expression, DNA accessibility and genome annotation across various human tissues. Average performance across classification tasks NTv3 Best small foundation model (10M) Top performance with larger model size AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 59

Post-training | Learning from +17k genomic tracks and genome annotation Predictions for 1mb input sequence at single-nucleotide resolution AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 60

Post-training | NTv3 accurately predicts genomic tracks at single-nucleotide resolution Predictions for 1 million nucleotide genomic region Example of NTv3 predictions for experiments in human K562 leukemia cells AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 61

Post-training | NTv3 accurately predicts genomic tracks at single-nucleotide resolution *Linder et al., Nature Genetics 2025 NTv3 outperforms state-of-the-art model (Borzoi*) at experimental track prediction. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 62

*de Almeida et al., Nature Methods 2025 Post-training | NTv3 accurately annotates genomes at single-nucleotide resolution NTv3 outperforms state-of-the-art model (SegmentNT*) at gene finding, regulatory element detection and splicing. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 65

Exploiting the data available with the aim of building a best-in-class model for genomics Predictive OR Generative Nucleotide Transformer v3 NTv3 NT Enformer Evo Native predictions and can be finetuned De-novo and conditional sequence generation Thanks to the masked discrete diffusion framework, NTv3 both exhibits strong representation capabilities for downstream tasks and is generative! AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 66

Generation | Designing regulatory enhancer sequences with NTv3 In collaboration with Alex Stark Experiment Design promoter-specific enhancers, at different levels of activity, in Drosophila cell line. Motivation Enhancers are sequence elements that modulate the expression of genes and can be used for gene therapy. Approach Fine-tune NTv3 to become a generative model using Masked Diffusion Language Models (MDLM) Validation Experimental validation through in vitro MPRAs AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 67

Generation | NTv3 designs have in-vitro state-of-the-art performance for activity-specific design Experimental validation of enhancers with different strengths (RpS12 promoter) NTv3 successfully generated de novo enhancers matching the prompted activity levels. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 68

Generation | NTv3 designs have in-vitro state-of-the-art performance for promoter-specific design Experimental validation of enhancers with promoter-specific activities *de Almeida et al., Nature Genetics 2023 NTv3 successfully generated de novo promoter-specific enhancers, achieving fold-change specificity significantly superior to previously validated state-of-the-art in vitro methods. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 69

NTv3: a new generation foundational model for genomics applications Genomic tracks (e.g. gene expression) Genome annotation (e.g. genes, splice sites finding) Variant prediction (e.g. eQTL, deleterious mutation) Sequence generation (e.g. enhancer design) NTv3 DNA sequence A C T A T C T A G A T A GT G G AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 70

AGL… AI (BFN-X) Our goal is to model as much of the data as possible ▪ Superior performance by learning a joint distribution across multiple data types and sources. ▪ Unparalleled flexibility in the hands of scientists with task-specific inference. Our vision: one model, many tasks AI AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 73

AGL… AI (BFN-X) Laying the Groundwork 1. Atkinson, T., Barrett, T.D., Cameron, S. et al. Protein sequence modelling with Bayesian flow networks. Nat Commun 16, 3197 (2025) AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 74

Why design Fvs? Therapeutic antibodies have diversified in formats across the years. The Fv region is the common key recognition component in all modalities. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 76

AbBFN2 CDR-H1 CDR-H2 CDR-H3 VH: EVQLLESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISWNSGSIYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCARGWSQVDTAMDLDYGQGTLVTVSS D geneV gene J gene CDR-L1 CDR-L2 CDR-L3 VL: DIQMTQSPSSVSASVGDRVTITCRASQSVSSNLAWYQQKPGKAPKLLIYGASSLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQYNNWLTFGQGTRLEIK V gene J gene Antibodies have unusual properties, necessitating fine-grained control over their genetics, sequence, and overall biophysics. It is estimated that >1016 antibody sequences are possible: needle-in-a-haystack problem with multiple design objectives. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 77

AbBFN2 achieves SOTA results on 23/23 sequence labelling tasks1, demonstrating robust learning of the genetic and biophysical attributes of antibody sequences. Sequence annotation • Sequence labelling is a prerequisite for steered generation and design. • AbBFN2 is a one-stop labelling tool, • Simplifies traditional computational pipelines and improves accuracy. 1. Experiment conducted on 10,000 unseen antibodies which were labelled by competitor models and traditional tools. Categorical data assessed via balanced F1 scores, continuous data assessed via Pearson’s R and root mean squared error. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 79

Using an unstable starting candidate, AbBFN2 is able to refine the interfaces and the total antibody to increase stability1, which allows more stable pairing, better storage, and higher expression levels. Stabilisation of existing antibodies 1. Interface energies are calculated using the Rosetta Protein Modelling Suite. 5,000 samples are generated in each case and compared to a background distribution 5,000 randomly picked unseen antibodies. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 80

• AbBFN2 optimises sequences with multiple conditions1, using inference time compute scaling to generate diverse candidates for early discovery and optimisation. Multi-objective design using AbBFN2 >80%: Success rate (overall) >90%: Success rate (tractable candidates) 46.6: Number of mutations (1 objective) 56.9: Number of mutations (5 objectives) EVQLLESGGGLVQPGGSLRLSCAAS... QVQLLESGGSLVQPGGSLRLSCAAS... QVQLLESGGSLVQPGGSIRLSCARS... Inference-time compute scaling 1. 91 high-risk unseen sequences with multiple sequence liabilities were optimised. For each sequence, 4 candidates were generated with up to 15 recycling iterations. Results are reported for the best variant for each candidate. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 81

Experimental validation Traditional humanization is a trial-and-error bottleneck: it often relies on arbitrary back-mutations, is time- and cost-intensive and risks disrupting binding through extensive changes. 1. Based on Marks et al, Bioinformatics, 2021; Tennenhouse et al, Nat. Biomed. Eng, 2024 1 AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 82

Experimental validation Objective Humanise antibodies to reduce side-effects, starting from the precursor sequence1 to generate designs Results Our sequences achieved similar expression with an average fewer edits compared to their manually designed clinical-stage counterparts. 1. For each precursor antibody, designs were generated as described previously. Designs were converted into scFv format and expressed in cell free E. coli expression systems. Expression levels are compared to the expression levels of the experimentally humanised reference sequence and binding was assessed using bio-layer interferometry. Experimentally humanised control sequences are those reported in Marks et al, Bioinformatics, 2021 2. Kd values might vary between reported and literature values due to experimental setup and selected scaffolds. 3. Exp. (Experimental) refers to a single molecule (per target) that is present in Marks et al, Bioinformatics, 2021 Target # mutations Binding (Kd, nM)2 Exp3. Ours Exp3. Ours IL-6Ra 42 37 10.9 13.6 IL-5 41 34 0.304 0.577 Her2 55 63 14.1 29.3 IgE 60 42 2.82 7.92 AbBFN2 enables efficient in silico humanization, preserving antigen binding while eliminating the need for lengthy and expensive wet lab experiments. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 84

AbBFN2: Cutting-edge in silico antibody design • Training on both sequence and associated metadata of interest produces a rich syntax for “prompt/task engineering”. • The “condition anywhere, generate anywhere” paradigm of AbBFN2 admits a wide variety of tasks that can be decided at inference time. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 85

Database searchReference library JETs lncRNAs Known Dark Tumor Antigens (dTAs) TMDPKTGYQ TAYEPTTFW ATGPVMPTR MS2 spectra TMDPKTGYQ GARVEMEYR ATGPVMPTR All Possible Peptides Incl. any unanticipated targets TAYEPTTFW SWHADEQV IGEYKTSLS … InstaNovo’s library-free approach allows discovery of unanticipated dark proteome antigens AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 92

InstaNovo (auto-regressive) and InstaNovo+ (diffusion) combine to outperform SOTA methods. Has already shown potential in detecting tumour specific epitopes from undocumented ORFs and aberrant splicing. Published in Nature Machine Intelligence Covered by Science Magazine Eloff, K., Kalogeropoulos, K., Mabona, A. et al. InstaNovo enables diffusion-powered de novo peptide sequencing in large-scale proteomics experiments. Nature Machine Intelligence 7, 565–579 (2025). https://doi.org/10.1038/s42256-025-01019-5 InstaNovo SOTA de novo peptide sequencing AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 93

Introducing InstaNovo V2 Larger Dataset 63 million labelled spectra Faster Prediction Up to 50x faster inference Higher Accuracy 10–15% increase in peptide recovery More Identifications Up to 2× the number of identifications "Introducing the next generation of InstaNovo models", https://instanovo.ai/introducing-the-next-generation-of-instanovo-models/ The next generation of InstaNovo models AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 94

High valency nanoparticle vaccines yield strong antibody responses towards tough infectious disease targets Hepatitis B vaccine1 Human papilloma virus vaccine2 Malaria vaccine3 Goal: Leverage AI to build nanoparticles suited to harness the power of mRNA vaccines Nanoparticle vaccines have a crown of repeating antigen on a scaffold They yield improved immune responses compared to solitary antigens All nanoparticle vaccines in humans are protein based 1. Valenzuela et al. Nature. 1982. 2. Kirnbauer et al. Proc. Natl. Acad. Sci. 1992. 3. Collins et al. Sci Rep. 2017. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 110

111 I AI Day © 2025 InstaDeep Ltd. & BioNTech SE I October, 2025 Goal: mRNA launched nanoparticle vaccines mRNA delivery ER/Golgi Nanoparticle assembly Nanoparticle secretion Nanoparticle vaccine Generated by BioRender

Shape generation Generate thousands of de novo trimer shapes to enhance diversity of building blocks Utilizing AI protein design to build the nanoparticle components Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 114

Symmetric assembling Assembling the nanoparticle building blocks into desired shapes Assembled to thousands of symmetric shapes Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 115

Sequence design Generate hundreds of thousands sequences to match the desired shapes and assemblies N L G V T F K W S … V D E V T A T Q T H Hundreds of sequences per particle S P R H T L A L R … A T M K E S V A E Designing amino acid sequences to form the protein shape Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 116

DeepChain Folding Studio Filter and enrich to tens - hundreds of high-quality designs prior to laboratory testing Computationally rank and filter the nanoparticle models Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 117

Negative stain electron micrographs confirming nanoparticle assembly Computationally designed nanoparticle model Candidate A Wet-lab experiments Candidate B In vitro: confirming nanoparticle design and assembly Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 118

Candidate + antigenCandidate + antigen Computationally designed nanoparticle + antigen model Negative stain electron micrographs confirming nanoparticle displays antigen In vitro: showcasing nanoparticles can display vaccine antigens Wet-lab experiments Internal AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 119

TCR affinity enhancement Antoine Delaunay Senior Research Engineer InstaDeep Michael Rooney Senior Director Comp Biology BioNTech Can you put this in as my head shot? Currently the slide shows John. MY title is senior director, computational biology AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 120

T cell receptors (TCRs) can access highly tumor-specific cancer targets Extracellular targets Modest tumor specificity Intracellular targets High tumor specificity HPV P53 RAS PRAME NY-ESO-1 TROP2CEACAM c-MET MHC TCRAntibody AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 121

T cells can achieve durable responses Checkpoint blockade in NSCLC Brahmer, JCO, 2023. PRAME-directed TCR-T Wermke, Nature Medicine, 2025. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 122

Affinity enhancement is required to unlock the full potential of T cell receptors (TCRs) TCR-T solTCR 10-3 M 10-6 M 10-9 M 10-12 M “micromolar” “nanomolar” “picomolar” Weak binding TCRs Strong binding TCRs Natural TCRs Cell therapy Off-the-shelf biologic AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 123

Conventional display-based affinity enhancement is labor- intensive but explores tiny sliver of TCR sequence space CDR1α library CDR2α library CDR3α library CDR1β library CDR2β library CDR3β library Displays Measurement Re pe at Natural TCR Enhanced TCR developability high potency high specificity ~1032 mutants with <15 mutations AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 124

AI-guided exploration of TCR sequence space enables efficient discovery of optimized variants AI-guided rational exploration of sequence space Measurement Re pe at Natural TCR Enhanced TCR ~1032 mutants with <15 mutations developability high potency high specificity AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 125

Residue environment determines optimal substitutions but varies from TCR to TCR Overall TCR:pMHC docking is similar, but exact CDR loop positions are highly diverse Divergent germline CDR2β-MHC interactions in two structures that share both V-genes and MHC allele (PDB ID: 5nht, 6vm9) Twelve TCRpMHC structures superimposed by MHC (PDB ID: 1ao7, 1mi5, 2ak4, 2nx5, 2ypl, 3dxa, 3ffc, 3h9s, 3vxm, 4g8g, 4jrx, 4mji) peptide CDR3β CDR3α Learning the rules of TCR optimization is hard due to high structural diversity of TCRpMHC interactions AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 126

0 10 20 30 40 50 Boltz-1¹ Chai-1² Our model For each model, fraction of targets where at least one of 5 generated structures achieved CDR3 RMSD < 2.5Å. Test set contains only unseen targets. Performance benchmark on test targets (CDR3 RMSD < 2.5Å ) 1. Wohlwend et al., Boltz-1: Democratizing Biomolecular Interaction Modeling, bioRxiv, 2025. 2. Chai Discovery et al., Chai-1: Decoding the molecular interactions of life, bioRxiv, 2024. 20 25 30 35 40 45 50 0 200 400 Number of structures generated per target Boltz-1¹ Chai-1² Our model % ta rg et s w ith to p- 5 m od el C D R3 R M SD < 2 .5 Å Our method outperforms sampling models that quickly saturate % o f t ar ge ts w ith C D R3 R M SD < 2 .5 Å Our model outperforms state-of-the-art in TCR–pMHC structure prediction AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 127

solTCR natural TCRs TCR-T measurement re pe at natural TCR enhanced TCR round K d, M variant sampler affinity predictor structure Source: Internal Example of a TCR affinity improvement of more than a 100,000-fold in 3 rounds Our AI pipeline achieves an average 50,000-fold TCR binding enhancement increase over WT, in three rounds or less, on the four considered targets. We repeatedly reach picomolar affinity. AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 128

Affinity-enhanced TCRs lead to strong and durable in vivo tumor control in a pre-clinical model pHLA tumor target 1 pHLA tumor target 2 Source: Internal Source: Internal TCR dosing starts TCR dosing ends TCR dosing starts TCR dosing ends AI Day © 2025 BioNTech SE & InstaDeep Ltd. I October, 2025 I 129