Biology subjects

Sahin, U.

Publications and source records attributed to Sahin, U..

At least 19 recordsLinked to original sources

bioRxiv2026

Durable suppression of viremia by lipid nanoparticle-formulated RNA encoding for a highly potent HIV-1 neutralizing antibody

Recombinant broadly neutralizing antibodies (bNAbs) are promising tools to treat and prevent HIV-1 infection but are associated with manufacturing challenges. To overcome this limitation, we encoded the potent bNAb 1-18-LS on RNA (1-18-LS RibobNAb) for delivery via lipid nanoparticles (LNPs). Indicating high in vivo antibody expression, a 30 ug RNA-LNP intravenous injection resulted in an average peak 1-18-LS RibobNAb serum concentration of 1,061 ug/mL in human neonatal Fc receptor-transgenic mice. Notably, weekly 30 ug RNA-LNP doses maintained higher trough bNAb levels than 500 ug protein injections. Highlighting potent antiviral activity, RNA-LNP-mediated 1-18-LS RibobNAb monotherapy of viremic HIV-1YU2-infected humanized mice resulted in durable HIV-1 suppression without emerging viral escape. Importantly, 1-18-LS RNA-LNP treatment also fully controlled infection after interruption of antiretroviral therapy in humanized mice infected with different patient-derived polyclonal HIV-1 isolates. Our findings provide a proof-of-principle for effective RNA-mediated bNAb immunotherapy of HIV-1 infection.

immunology↗

bioRxiv2026

Scalable spatial DNA sequencing from archival tissue maps copy number subclones

Spatially resolved DNA sequencing holds promise due to its potential utility in understanding cancer intra-tumour heterogeneity and tumour evolution in relation to tissue architecture. However, it has so far been used to a limited extent due to technical challenges and high cost of existing methods. Hence we aimed to develop a high throughput spatial genomic assay to obtain copy number alteration (CNA) information at user-defined spatial resolution. We derived CNA profiles from ultra-low coverage whole genome sequencing at sub-millimetre resolution from archival samples using a novel method called Adaptive Resolution Multiscale Spatial DNA sequencing (ARMS DNAseq). We used it to profile CNAs from more than 766 regions (tiles) from 3 patients, covering a total area of over 300 mm2, with 1.2-2.6 million mapped reads per tile and tile sizes of 0.1-0.99mm2. Using ARMS DNAseq, we delineate tumour evolution in a spatial context, and identify more tumour subclones that were obscured or incompletely represented in bulk multi-region whole genome sequencing. Next, we show associations between tumour subclones and morphology, and prediction of subclone identity from deep learning-derived image representations. Finally, we demonstrate multi-omic integration by alignment with spatial transcriptomic data, showing subclone-specific immune cell co-occurrence as well as transcriptional programmes cutting across subclone boundaries. ARMS DNAseq converts low-throughput, region-by-region profiling into a scalable and adaptable workflow for direct spatial copy number profiling from archival tissue sections.

genomics↗

bioRxiv2026

Computing tumor specificity of cancer antigen targets by k-mer indexing of healthy tissue transcriptomes

Individualized cancer immunotherapies rely on tumor-specific T-cell antigens, often predicted from somatic mutations as neoantigens. For tumors with low mutational burden, mRNA transcript variants, including gene fusions and novel splice junctions, can serve as important alternative targets. A main challenge in their identification from tumor RNA-seq is to confirm that their expression is tumor-restricted. Although large public collections of healthy-tissue RNA-seq exist, verifying tumor-specific expression requires computationally expensive re-analysis of these data for every novel candidate. To address this, we benchmarked nine k-mer indexing algorithms and developed k4neo, which leverages k-mer indexing of raw RNA-seq reads to compute the tumor specificity of any transcript variant. This mapping-free and transcript-class agnostic approach screens any candidate sequence against 18,960 samples across 51 healthy tissue types. We confirmed k4neos detection accuracy with qRT-PCR and showed that k4neo accurately classifies somatic and germline variants, gene fusions, and isoforms by tumor specificity. Applied to nine tumor cohorts, it nominated a median of 4-80 tumor-specific splice junctions per patient, including recurrent, long-read-confirmed novel antigen candidates. Together, k4neo enables efficient access to large-scale sequencing cohorts and accurately computes tumor specificity for any input transcript sequence, thereby expanding the repertoire of individual and shared cancer antigen targets.

bioinformatics↗

bioRxiv2026

Mapping the immune landscape in small cell lung cancer unveils a distinct tumor-reactive CD8+ T cell molecular signature

Small cell lung cancer (SCLC) is a highly aggressive malignancy with limited therapeutic advances. Unlike many other cancers, its immune landscape, particularly immune competence and T cell recognition, remains poorly characterized. Here, we generate a single-cell transcriptome atlas of the SCLC immune microenvironment with paired T cell receptor (TCR) sequencing. By linking T cell states with clonality and a multilayered functional screening, we identify 6 tumor-reactive TCRs that recognize and eradicate autologous SCLC cell lines. We delineate a novel SCLC-reactive CD8+ T cell signature (SCLC_TR), enabling the identification of 47 further SCLC-reactive TCRs. The SCLC_TR signature performs extremely well in pancreatic ductal adenocarcinoma (PDAC), another immune-cold tumor indication, and, most strikingly, patients with elevated SCLC_TR signature scores exhibited significantly improved survival, underlining its prognostic potential. Comparative cell-cell interaction analyses implicate several immunosuppressive mechanisms, with myeloid cells and CD4 regulatory T cells possibly acting as counterbalances to effector T cell activity in SCLC. In summary, our study challenges the prevailing notion of SCLC as an immune-cold tumor type by providing direct evidence of tumor-reactive T cell responses and introduces the SCLC_TR signature as a tool to identify tumor-specific T cells and their microenvironmental restraints and escape mechanisms, ultimately shaping next-generation immunotherapeutic strategies. O_FIG O_LINKSMALLFIG WIDTH=195 HEIGHT=200 SRC="FIGDIR/small/735200v1_ufig1.gif" ALT="Figure 1"> View larger version (69K): org.highwire.dtl.DTLVardef@1c25f2eorg.highwire.dtl.DTLVardef@1f6f519org.highwire.dtl.DTLVardef@553e57org.highwire.dtl.DTLVardef@6fe7e1_HPS_FORMAT_FIGEXP M_FIG C_FIG

immunology↗

bioRxiv2026

RNA-Encoded PGT121-LS Anti-HIV Antibody: Comprehensive Preclinical Characterization and Translational Pharmacokinetics

Human Immunodeficiency Virus (HIV)-1 broadly neutralizing antibodies (bNAbs) have demonstrated clinical efficacy, but face manufacturing challenges associated with recombinant protein production and purification. Here, we present a ribonucleic acid (RNA)-encoded bNAb (RibobNAb) platform that enables in vivo antibody production of the clinically validated bNAb PGT121 via lipid nanoparticle (LNP) delivery, supporting rapid evaluation of Fc variants (LS, del294, LS-del294) in vitro and in vivo. We confirmed expression, sub-nanomolar HIV-1 Env binding, and potent neutralization across all RibobNAb variants in vitro. In mice, single RNA-LNP administrations yielded in vivo expression of all RibobNAb variants, with PGT121-LS exhibiting a prolonged half-life compared with PGT121. In non-human primates (NHPs), a single intravenous administration of PGT121-LS RNA-LNP was well tolerated without anti-drug antibody (ADA) formation over 180 days and resulted in PGT121-LS half-lives comparable to the reference protein. Single intramuscular administration showed RibobNAb expression but resulted in ADA development from Day 14 onwards and lower bioavailability. In vivo-expressed PGT121-LS RibobNAb retained identical antiviral functionality to PGT121-LS reference protein. An NHP pharmacokinetics model integrating RNA transfection and translation dynamics enabled allometric scaling and first-in-human dose prediction. We highlight RibobNAbs as an alternative to conventional purified protein antibodies for rapid development of bNAb-based therapeutic strategies. O_FIG O_LINKSMALLFIG WIDTH=193 HEIGHT=200 SRC="FIGDIR/small/734219v1_ufig1.gif" ALT="Figure 1"> View larger version (44K): org.highwire.dtl.DTLVardef@141319forg.highwire.dtl.DTLVardef@120caadorg.highwire.dtl.DTLVardef@1da2782org.highwire.dtl.DTLVardef@158064d_HPS_FORMAT_FIGEXP M_FIG Graphical abstract C_FIG

immunology↗

bioRxiv2026

Lung-targeted cytokine-coding RNA-lipoplexes induce T and NK cell-mediated anti-tumor immune response

Lung is a major site of metastases for many primary cancers associated with poor outcomes. A central challenge in cancer immunotherapy is overcoming tumor immune evasion, which limits effective antitumor responses. Here, we investigated whether combinatorial mRNA-encoded cytokine therapy can overcome tumor immune evasion by coordinately engaging innate and adaptive immunity, using murine models of pulmonary metastases. We employed intravenously administered cationic nucleoside-modified mRNA-lipoplexes (RNA-LPX) for targeted delivery of mRNA-encoded cytokines to the lung. The cytokine mix containing interferon-, half-life extended interleukin (IL)-7, and a half-life extended IL-2 variant with reduced CD25-binding modulated the tumor immune microenvironment resulting in a potent and broad anti-tumor response and prolonged survival with good tolerability at the conditions tested. Using cell depletion experiments, we demonstrated that both T and natural killer (NK) cells are crucial mediators of the observed anti-tumor efficacy of the cytokine RNA mix, which induced activation and effector function of NK and T cells, coupled with reduced regulatory T cells (Treg) numbers and Treg activation in the lung. Importantly, antitumor efficacy was maintained in models of impaired antigen presentation, including loss of an immunodominant tumor antigen and MHC class I deficiency, where NK cells served as the primary effectors. The cytokine RNA mix induced immune cell activation in the primary human lung tumor culture, suggesting potential for translational application. Collectively, these findings demonstrate that combinatorial cytokine therapy can drive both antigen-dependent and antigen-independent tumor control for the treatment of lung metastases.

immunology↗

bioRxiv2025

Preclinical immunogenicity of the LP.8.1-adapted BNT162b2 COVID-19 vaccine

SARS-CoV-2 evolution toward antigenically distinct lineages drives escape from host immunity. JN.1 lineage derivatives have recently dominated the global epidemiologic landscape. In preclinical models, an LP.8.1-adapted BNT162b2 vaccine elicited higher serum neutralizing antibody responses against contemporary, circulating JN.1 sublineages, including the currently dominant XFG, as compared to JN.1-, KP.2- and XEC-adapted vaccines. These findings supported the selection of an LP.8.1-adapted vaccine for the composition of the 2025-26 COVID-19 vaccine formula.

immunology↗

bioRxiv2025

Recurrent RNA-lipoplex vaccination is required to sustain functional tumor-infiltrating neoantigen-specific CD8 T cells and therapeutic efficacy

Cancer vaccines induce durable, polyepitopic T cell responses, and show promising clinical benefit in adjuvant settings, yet they are largely ineffective in advanced disease. Using a clinically relevant RNA-lipoplex vaccine, we investigated the efficacy constraints in a preclinical model. Vaccination remodeled the tumor microenvironment (TME), increasing T cell infiltration and promoting a proinflammatory myeloid compartment. This was associated with complete regression of smaller, immature tumors, but only delayed growth of larger, established tumors. While vaccine-induced T cells were long-lived and functional in peripheral tissues, intratumoral T cells declined rapidly in abundance, diversity, and function, reverting to a prevaccine-like state. scRNA-seq suggested that this was driven by a pro-apoptotic program, with surviving T cells showing signatures of cellular stress and impaired activation. Importantly, recurrent vaccination replenished functional T cells in the TME and enhanced efficacy. These findings highlight the importance of optimizing vaccine schedules and tailoring therapeutic strategies to tumor stage.

immunology↗

bioRxiv2025

Cross-competition shapes CD8+ T cell hierarchies and differentiation after RNA vaccination

Short SummaryImmunodominance is a universal feature of adaptive immunity that constrains T cell expansion, clonal diversity and breadth resulting in a narrowly focused T cell response. While observed across diverse priming settings and vaccine platforms, the influence of immunodominance on T cell phenotype remains unclear. Using an mRNA lipoplex vaccine encoding multiple antigens to study how immunodominance influences CD8+ T cell fate, we found that dominant CD8+ T cell responses alter the magnitude and phenotype of subdominant responses through peptide-MHC-I stability-mediated T cell cross-competition. Dominant CD8+ T cell responses preferentially acquired markers associated with terminal differentiation and cytotoxic function, while sub-dominant responses adopted memory-precursor and stem-like features. Removal of dominant responses allowed increased expansion of sub-dominant T cell responses and adoption of terminally differentiated effector phenotypes. These findings reveal that immunodominance dynamically shapes the magnitude, breadth and differentiation of CD8+ T cell responses and highlights opportunities to fine-tune T cell responses for therapeutic vaccination.

immunology↗

bioRxiv2025

mRNA-based tuberculosis vaccines BNT164a1 and BNT164b1 are immunogenic, well-tolerated and efficacious in rodent models

We designed and preclinically tested two mRNA-LNP-based vaccine candidates to protect against tuberculosis (TB). BNT164a1 and BNT164b1 encode the same eight Mycobacterium tuberculosis (Mtb) antigens expressed across different infection stages: Ag85A, Hrp1, ESAT-6, RpfD, RpfA, HbhA, M72, and VapB47. BNT164a1 utilizes nucleoside-unmodified mRNA, while BNT164b1 utilizes N1-methyl pseudouridine-modified mRNA. Prime-boost immunization with BNT164 candidates elicited antibody and/or T-cell responses against all antigens in three mouse strains (C57BL/6, BALB/c, and HLA-A2.1/DR1 humanized mice). The candidates demonstrated favorable safety profiles in a rat toxicity study and significantly reduced bacterial burdens of two Mtb strains in murine aerosol challenge models. BNT164 protection correlated with granuloma infiltration by CD8+ T cells with memory precursor phenotypes. In conclusion, BNT164a1 and BNT164b1 were immunogenic, well tolerated and efficacious in preclinical models and are the first mRNA-based TB vaccines to enter phase I/II clinical trials (NCT05537038, NCT05547464).

immunology↗

bioRxiv2025

A single MHCII neoepitope mRNA vaccine elicits CD4 T- and B- cell responses promoting endogenous CD8 anti-tumor immunity

Recent progress in therapeutic cancer vaccines has shown promising clinical activity, especially when targeting MHC class I (MHCI) neoantigen-specific CD8+ T cell responses in post-surgical patients. To explore the role of CD4+ T cells in vaccine-dependent tumor rejection, we constructed an mRNA lipoplex vaccine encoding a single MHCII-restricted neoantigen. The vaccine elicited Tfh and Th1 cell responses while decreasing Tregs, leading to rejection of established tumors in mice. IL-21 and IFN-{gamma}, crucial for Tfh and Th1 function respectively, contributed to anti-tumor activity. B cells and neoantigen-specific antibodies were also shown to participate in vaccine efficacy. Additionally, conventional type 1 dendritic cells (cDC1s) were essential for eliciting vaccine-induced CD4+ T cells, and both cDC1s and CD4+ T cells were required to enhance endogenous CD8+ responses, which were crucial for tumor control. Our results suggest that immunizing against MHCII neoantigens alone is sufficient to orchestrate a potent and cooperative immune response against cancer.

immunology↗

bioRxiv2024

ALS driven by mutant NEK1 aggregation is accelerated by Pml loss, but clinically reversed through pharmacologic induction of Pml-mediated degradation

Germinal mono-allelic loss-of-function mutations of NEK1 drive Amyotrophic Lateral Sclerosis (ALS) at variable penetrance, presumably through haploinsufficiency. Modeling the ALS-associated Arg812Ter mutation in mice revealed that the resulting truncated Nek1 (Nek1t) is aggregation-prone, particularly in alpha-motoneurons (MNs), and drives canonical ALS symptoms when bi-allelically expressed (Nek1t/t). Promyelocytic leukemia (Pml) ablation allows for ALS symptoms to occur even in heterozygote Nek1wt/t animals, mimicking the human situation. Pml precludes disease occurrence by promoting SUMO-facilitated degradation of Nek1t proteins through PML nuclear bodies (NBs). Conversely, Pml induction, achieved by activating the interferon pathway via poly(I:C) treatment, clears Nek1t aggregates in MNs, dramatically reducing ALS-associated symptoms and extending survival by 5 months. Our studies highlight the role of NEK1 aggregates in ALS pathogenesis and identifies activation of interferon pathways as a candidate therapeutic strategy that not only promotes Pml-triggered SUMOylation/degradation of toxic misfolded proteins in vivo, but also facilitates the clearance of protein aggregates, yielding dramatic clinical improvement. These observations validate PML as a relevant therapeutic target in neurodegenerative conditions associated with protein aggregation.

pathology↗

bioRxiv2024

Immunologic and Biophysical Features of the BNT162b2 JN.1- and KP.2-Adapted COVID-19 Vaccines

Vaccines remain a vital public health tool to reduce the burden of COVID-19. COVID-19 vaccines that are more closely matched to circulating SARS-CoV-2 lineages elicit more potent and relevant immune responses that translate to improved real-world vaccine effectiveness. The rise in prevalence of the Omicron JN.1 lineage, and subsequent derivative sublineages such as KP.2 and KP.3, coincided with reduced neutralizing activity and effectiveness of Omicron XBB.1.5-adapted vaccines. Here, we characterized the biophysical and immunologic attributes of BNT162b2 JN.1- and KP.2-adapted mRNA vaccine-encoded spike (S) protein immunogens. Biophysical interrogations of S revealed the structural consequences of hallmark amino acid substitutions and a potential molecular mechanism of immune escape employed by JN.1 and KP.2. The vaccine candidates were evaluated for their immunogenicity when administered as fourth or fifth doses in BNT162b2-experienced mice or as a primary series in naive mice. In both vaccine-experienced and naive settings, JN.1- and KP.2-adapted vaccines conferred improved neutralizing responses over the BNT162b2 XBB.1.5 vaccine against a broad panel of emerging JN.1 sublineages, including the predominant KP.3.1.1 and emerging XEC lineages. Antigenic mapping of neutralizing responses indicated greater antigenic overlap of JN.1- and KP.2-adapted vaccine responses with currently circulating sublineages compared to an XBB.1.5-adapted vaccine. CD4+ and CD8+ T cell responses were generally conserved across all three vaccines. Together, the data support the selection of JN.1- or KP.2-adapted vaccines for the 2024-25 COVID-19 vaccine formula. ONE-SENTENCE SUMMARYThe Omicron JN.1- and KP.2-adapted BNT162b2 mRNA vaccines encoding prefusion S proteins elicit similar preclinical neutralizing antibody responses against circulating JN.1 sublineage pseudoviruses that are more potent than those elicited by past iterations of BNT162b2 licensed vaccines, thus demonstrating the importance of annual strain changes to the COVID-19 vaccine.

immunology↗

bioRxiv2024

Co-occurring Mutations in Different Genes Can Fuel Oncogenic Signaling and Serve as Metastatic Tumor Markers

Interrogation of big genomic data and integration with large-scale protein-protein interaction networks and pathways, can provide deep patterns that are rare- yet can prompt dramatic phenotypic alterations and serve as clinical signatures. Mapping cancer-specific co-occurring mutation-pair signatures, in primary and metastatic tumors, is indispensable in precision oncology. The additivity of co-occurring driver mutations in different genes (in trans) can lead to powerful proliferation signals. Co-occurring rare in trans combinations can serve as metastasis markers; excluded combinations may indicate candidates for oncogene-induced senescence (OIS), a tumor-suppressive mechanism. Our statistical framework of the pan-cancer mutation profiles of [~]60,000 tumor sequences from the TCGA and AACR GENIE databases, identified 3424 statistically significant different double mutations in non-redundant pathways, that is, have different downstream targets that may promote specific cancers through single or multiple pathways. Our analysis indicates that they are mostly in primary tumors. We list actionable in trans mutations for 2385 metastatic tumors and provide co-occurrence trees of metastatic breast- cancer markers. This innovative work clarifies the mechanistic conceptual basis and establishes the first of its kind tool for identifying and predicting metastasis. Crucially, when coupled with their proliferative functions and pathways, and linked with drugs, it could provide an invaluable metastasis-targeting resource.

systems biology↗

bioRxiv2024

ChatNT: A Multimodal Conversational Agent for DNA, RNA and Protein Tasks

Language models are thriving, powering conversational agents that assist and empower humans to solve a number of tasks. Recently, these models were extended to support additional modalities including vision, audio and video, demonstrating impressive capabilities across multiple domains including healthcare. Still, conversational agents remain limited in biology as they cannot yet fully comprehend biological sequences. On the other hand, high-performance foundation models for biological sequences have been built through self-supervision over sequencing data, but these need to be fine-tuned for each specific application, preventing transfer and generalization between tasks. In addition, these models are not conversational which limits their utility to users with coding capabilities. In this paper, we propose to bridge the gap between biology foundation models and conversational agents by introducing ChatNT, the first multimodal conversational agent with an advanced understanding of biological sequences. ChatNT achieves new state-of-the-art results on the Nucleotide Transformer benchmark while being able to solve all tasks at once, in English, and to generalize to unseen questions. In addition, we have curated a new set of more biologically relevant instructions tasks from DNA, RNA and proteins, spanning multiple species, tissues and biological processes. ChatNT reaches performance on par with state-of-the-art specialized methods on those tasks. We also present a novel perplexity-based technique to help calibrate the confidence of our model predictions. Our framework for genomics instruction-tuning can be easily extended to more tasks and biological data modalities (e.g. structure, imaging), making it a widely applicable tool for biology. ChatNT is the first model of its kind and constitutes an initial step towards building generally capable agents that understand biology from first principles while being accessible to users with no coding background.

genomics↗

bioRxiv2024

SegmentNT: annotating the genome at single-nucleotide resolution with DNA foundation models

Genome annotation models that directly analyze DNA sequences are indispensable for modern biological research, enabling rapid and accurate identification of genes and other functional elements. This capability is paramount as the volume of sequenced genomes rapidly expands, making the need for efficient and accurate annotation methods increasingly critical, particularly in the context of genetic variant prediction and in-silico sequence design. Current annotation tools are typically developed for specific element classes and trained from scratch using supervised learning on datasets that are often limited in size. This approach constrains their performance and ability to generalize to new genomes. Here, we frame the genome annotation problem as instance segmentation and introduce a novel methodology for fine-tuning pre-trained DNA foundation models to segment 14 different genic and regulatory elements at single-nucleotide resolution. We leverage the self-supervised pre-trained model Nucleotide Transformer (NT) to develop a general segmentation model, SegmentNT, capable of processing DNA sequences up to 50kb long. By utilizing pre-trained weights from NT, SegmentNT surpasses the performance of several ablation models and baselines, including convolutional networks with one-hot encoded nucleotide sequences and large models trained from scratch. We demonstrate state-of-the-art performance on gene annotation, splice site and regulatory elements detection throughout the genome. We also leveraged our framework to accommodate two extra DNA foundation models, Enformer and Borzoi, extending the sequence context up to 500kb and enhancing performance on regulatory elements. Finally, we show that a SegmentNT model trained on human genomic elements generalizes to elements of different species, and a multi-species SegmentNT model achieves strong generalization across unseen species. Our approach is readily extensible to additional genomic elements and species. We have made our SegmentNT human and multi-species models, as well as the SegmentEnformer and SegmentBorzoi models, available on our github repository in Jax and HuggingFace space in Pytorch.

genomics↗

bioRxiv2023

LightMHC: A Light Model for pMHC Structure Prediction with Graph Neural Networks

The peptide-major histocompatibility complex (pMHC) is a crucial protein in cell-mediated immune recognition and response. Accurate structure prediction is potentially beneficial for protein interaction prediction and therefore helps immunotherapy design. However, predicting these structures is challenging due to the sequential and structural variability. In addition, existing pre-trained models such as AlphaFold 2 require expensive computation thus inhibiting high throughput in silico peptide screening. In this study, we propose LightMHC: a lightweight model (2.2M parameters) equipped with attention mechanisms, graph neural networks, and convolutional neural networks. LightMHC predicts full-atom pMHC structures from amino-acid sequences alone, without template structures. The model achieved comparable or superior performance to AlphaFold 2 and ESMFold (93M and 15B parameters respectively), with five-fold acceleration (6.65 seconds/sample for LightMHC versus 36.82 seconds/sample for AlphaFold 2), potentially offering a valuable tool for immune protein structure prediction and immunotherapy design.

bioinformatics↗

bioRxiv2023

FrameDiPT: SE(3) Diffusion Model for Protein Structure Inpainting

Protein structure prediction field has been revolutionised by deep learning with protein folding models such as AlphaFold 2 and ESMFold. These models enable rapid in silico prediction and have been integrated into de novo protein design and protein-protein interaction (PPI) prediction. However, biologically relevant features dependent on conformational distributions cannot be estimated with these models. Diffusion models, a novel class of generative models, have been developed to learn conformational distributions and applied to de novo protein design. Limited work has been done on protein structure inpainting, where a masked section is recovered by simultaneously conditioning on its sequence and the rest of the structure. In this work, we propose FrameDiff inPainTing (FrameDiPT), a generalised model for protein inpainting. This is important for T-cells given the hyper-variability of the complementarity determining region (CDR) loops. We evaluated the model on CDR loop design for T-cell receptors and achieved comparable prediction accuracy to ProteinGenerator and RFdiffusion with limited training data and learnable parameters. Different from deterministic structure prediction models, FrameDiPT captures the conformational distribution at different regions and binding states, highlighting a key advantage of generative models. The model and inference code have been released1.

immunology↗