Tumor-specific T cells play a central role in tumor control, yet the attainment of their full tumor rejection potential is hindered by regulatory mechanisms, including expression of inhibitory immune checkpoints. Although immunotherapy by immune checkpoint blockade (ICB) has transformed cancer care, only a proportion of patients experience clinical benefit. Ancillary studies to clinical trials evaluating ICB have revealed an association between high tumor mutational burden (TMB) and treatment efficacy, suggesting that unleashing T cells targeting neoepitopes encoded by somatic mutations could underlie ICB-mediated tumor control or regression. Nevertheless, TMB or tumor neoantigenic burden (TNB) are not consistently related to clinical response, potentially because most studies have addressed the influence of TMB and TNB by in silico prediction without systematic analysis of actual neoantigen-specific T cells. We hypothesized that the presence of actual neoantigen-specific T-cell responses would predict response to ICB, rather than the mere presence of a high TMB or TNB. Using whole exome and transcriptome sequencing in a cohort of 29 non-small cell lung cancer patients receiving anti-PD-1/PD-L1, we assessed expressed non-synonymous exonic mutations in each patient's tumor and predicted the encoded neoepitopes restricted by each patient's HLA-I and -II alleles. In our cohort, HLA-I TNB was better correlated to clinical response and progression free survival (PFS) than TMB and HLA-II TNB. We developed a workflow allowing for the assessment of circulating CD8 T-cell responses to more than 400 prioritized neoepitopes, and to 50 long peptides harboring each several predicted CD4 neoepitopes containing the same mutated amino acid. We showed, in 28 patients from whom PBMCs were available, that circulating CD8 T-cell responses to predicted neoantigens were of higher magnitude and breadth in clinical responder compared to non-responder patients. The presence of a CD8 T-cell response to neoantigens was associated with better PFS. No correlation was observed between expressed TMB or TNB (HLA-I or -II) and the magnitude of CD8 or CD4 T-cell responses. Longitudinal ex vivo enumeration of neoantigen-specific CD8 T-cells using HLA/peptide tetramers showed that they had a memory phenotype and that they experienced significant proliferation under therapy. Our results ascertain the contribution of neoantigen-specific CD8 T-cells in therapeutic tumor rejection. They also provide an opportunity, through the analysis of tumor molecular signatures, to unveil the mechanisms supporting or restraining the priming of spontaneous T-cell responses targetable by ICB. This research was a collaborative effort by sites within the imCORE network, made possible through support from F.Hoffmann-La Roche. Célia Ramade, Noémie Thébault, Clara-Maria Scarlata, Françoise Lauzéral-Vizcaino, Caroline Fournier, Victor Sarradin, Anna Salvioni, Marie Michelas, Daniel Oreper, Suchit Jhunjhunwala, William Thrift, Adric Quade Broadwell, Nicolas Lounsbury, Kai Liu, Meng Xiao He, Martine Darwish, Catherine Ross, Hang Xu, Milena Hornburg, Amy Heidersbach, Fanny Bouquet, Catia Fonseca, Laetitia Tudeau, Nicolas Congy-Jolivet, Carlos Gomez-Roca, Thomas Filleron, Lélia Delamarre, Julien Mazières, Jean-Pierre Delord, Maha Ayyoub. Circulating CD8 T-cell responses to neoantigens predict responsiveness to PD-(L)1 axis blockade in non-small cell lung cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6037.

Antigen presentation on MHC class II (pMHCII presentation) plays an essential role in the adaptive immune response to extracellular pathogens and cancerous cells. But it can also reduce the efficacy of large-molecule drugs by triggering an anti-drug response. Significant progress has been made in pMHCII presentation modeling due to the collection of large-scale pMHC mass spectrometry datasets (ligandomes) and advances in machine learning. Here, we develop graph-pMHC, a graph neural network approach to predict pMHCII presentation. We derive adjacency matrices for pMHCII using Alphafold2-multimer and address the peptide-MHC binding groove alignment problem with a simple graph enumeration strategy. We demonstrate that graph-pMHC dramatically outperforms methods with suboptimal inductive biases, such as the multilayer-perceptron-based NetMHCIIpan-4.0 (+20.17% absolute average precision). Finally, we create an antibody drug immunogenicity dataset from clinical trial data and develop a method for measuring anti-antibody immunogenicity risk using pMHCII presentation models. Our model increases receiver operating characteristic curve (ROC)-area under the ROC curve (AUC) by 2.57% compared to just filtering peptides by hits in OASis alone for predicting antibody drug immunogenicity.

Based on the success of cancer immunotherapy, personalized cancer vaccines have emerged as a leading oncology treatment. Antigen presentation on MHC class I (MHC-I) is crucial for the adaptive immune response to cancer cells, necessitating highly predictive computational methods to model this phenomenon. Here, we introduce HLApollo, a transformer-based model for peptide-MHC-I (pMHC-I) presentation prediction, leveraging the language of peptides, MHC, and source proteins. HLApollo provides end-to-end treatment of MHC-I sequences and deconvolution of multi-allelic data, using a negative-set switching strategy to mitigate misassigned negatives in unlabelled ligandome data. HLApollo shows a 12.65% increase in average precision (AP) on ligandome data and a 4.1% AP increase on immunogenicity test data compared to next-best models. Incorporating protein features from protein language models yields further gains and reduces the need for gene expression measurements. Guided by clinical use, we demonstrate pan-allelic generalization which effectively captures rare alleles in underrepresented ancestries.

The utility of deep neural nets has been demonstrated for mapping hematoxylin-and-eosin (H&E) stained image features to expression of individual genes. However, these models have not been employed to discover clinically relevant spatial biomarkers. Here we develop MOSBY ( M ulti- Omic translation of whole slide images for S patial B iomarker discover Y ) that leverages contrastive self-supervised pretraining to extract improved H&E whole slide images features, learns a mapping between image and bulk omic profiles (RNA, DNA, and protein), and utilizes tile-level information to discover spatial biomarkers. We validate MOSBY gene and gene set predictions with spatial transcriptomic and serially-sectioned CD8 IHC image data. We demonstrate that MOSBY-inferred colocalization features have survival-predictive power orthogonal to gene expression, and enable concordance indices highly competitive with survival-trained multimodal networks. We identify and validate 1) an ER stress-associated colocalization feature as a chemotherapy-specific risk factor in lung adenocarcinoma, and 2) the colocalization of T effector cell vs cysteine signatures as a negative prognostic factor in multiple cancer indications. The discovery of clinically relevant biologically interpretable spatial biomarkers showcases the utility of the model in unraveling novel insights in cancer biology as well as informing clinical decision-making.

The study of cells and their responses to genetic or chemical perturbations promises to accelerate the discovery of therapeutics targets. However, designing adequate and insightful models for such data is difficult because the response of a cell to perturbations essentially depends on contextual covariates (e.g., genetic background or type of the cell). There is therefore a need for models that can identify interactions between drugs and contextual covariates. This is crucial for discovering therapeutics targets, as such interactions may reveal drugs that affect certain cell types but not others. We tackle this problem with a novel Factorized Causal Representation (FCR) learning method, an identifiable deep generative model that reveals causal structure in single-cell perturbation data from several cell lines. FCR learns multiple cellular representations that are disentangled, comprised of covariate-specific (Z_x), treatment-specific (Z_t) and interaction-specific (Z_tx) representations. Based on recent advances of non-linear ICA theory, we prove the component-wise identifiability of Z_tx and block-wise identifiability of Z_t and Z_x. Then, we present our implementation of FCR, and empirically demonstrate that FCR outperforms state-of-the-art baselines in various tasks across four single-cell datasets.

1.AbstractData obtained from clinical trials for a given disease often capture reliable empirical features of the highest quality which are limited to few studies/experiments. In contrast, knowledge data extracted from biomedical literature captures a wide range of clinical information relevant to a given disease that may not be as reliable as the experimental data. Therefore, we propose a novel method of training that co-optimizes two AI algorithms on experimental data and knowledge-based information from literature respectively to supplement the learning of one algorithm with that of the other and apply this method to prioritize/rank causal genes for Alzheimer’s Disease (AD). One algorithm generates unsupervised embeddings for gene nodes in a protein-protein interaction network associated with experimental data. The other algorithm generates embeddings for the nodes/entities in a knowledge graph constructed from biomedical literature. Both these algorithms are co-optimized to leverage information from each other’s domain. Therefore; a downstream inferencing task to rank causal genes for AD ensures the consideration of experimental and literature data available to implicate any given gene in the geneset. Rank-based evaluation metrics computed to validate the gene rankings prioritized by our algorithm showed that the top ranked positions were highly enriched with genes from a ground truth set that were experimentally verified to be causal for the progression of AD.

Generating knowledge graph embeddings (KGEs) to represent entities (nodes) and relations (edges) in large scale knowledge graph datasets has been a challenging problem in representation learning. This is primarily because the embeddings / vector representations that are required to encode the full scope of data in a large heterogeneous graph needs to have a high dimensionality. The orientation of a large number of vectors requires a lot of space which is achieved by projecting the embeddings to higher dimensions. This is not a scalable solution especially when we expect the knowledge graph to grow in size in order to incorporate more data. Any efforts to constrain the embeddings to lower number of dimensions could be problematic as insufficient space to spatially orient the large number of embeddings / vector representations within limited number of dimensions could lead to poor inferencing on downstream tasks such as link prediction which leverage these embeddings to predict the likelihood of existence of a link between two or more entities in a knowledge graph. This is especially the case with large biomedical knowledge graphs which relate several diverse entities such as genes, diseases, signaling pathways, biological functions etc. that are clinically relevant for the application of KGs to drug discovery. The size of the biomedical knowledge graphs are therefore much larger compared to typical benchmark knowledge graph datasets. This poses a huge challenge in generating embeddings / vector representations of good quality to represent the latent semantic structure of the graph. Attempts to circumvent this challenge by increasing the dimensionality of the embeddings often render hardware limitations as generating high dimensional embeddings is computationally expensive and often times infeasible. To practically deal with representing the latent structure of such large scale knowledge graphs (KGs), our work proposes an ensemble learning model in which the full knowledge graph is sampled into several smaller subgraphs and KGE models generate embeddings for each individual subgraph. The results of link prediction from the KGE models trained on each subgraph are then aggregated to generate a consolidated set of link predictions across the full knowledge graph. The experimental results demonstrated significant improvement in rank-based evaluation metrics on task specific link predictions as well as general link predictions on four open-sourced biomedical knowledge graph datasets.

Spatial transcriptomics allows precise RNA abundance measurement at high spatial resolution, linking cellular morphology with gene expression. We present a novel deep learning algorithm predicting local gene expression from histopathology images. Our approach employs a graph isomorphism neural network capturing cell-to-cell interactions in the tumor microenvironment and a Vision Transformer (CTransPath) for obtaining the tumor morphological features. Using a dataset of 30,612 spatially resolved gene expression profiles matched with histopathology images from 23 breast cancer patients, we identify 250 genes, including established breast cancer biomarkers, at a 100 µm resolution. Additionally, we co-train our algorithm on spatial spot-level transcriptomics from 10x Visium breast cancer data along with another variant of our algorithm on TCGA-BRCA bulk RNA Seq. data, yielding mutual benefits and enhancing predictive accuracy on both these datasets. This work enables image-based screening for molecular biomarkers with spatial variation, promising breakthroughs in cancer research and diagnostics.

The integration of multi-modal data, such as pathological images and genomic data, is essential for understanding cancer heterogeneity and complexity for personalized treatments, as well as for enhancing survival predictions. Despite the progress made in integrating pathology and genomic data, most existing methods cannot mine the complex inter-modality relations thoroughly. Additionally, identifying explainable features from these models that govern preclinical discovery and clinical prediction is crucial for cancer diagnosis, prognosis, and therapeutic response studies. We propose PONET- a novel biological pathway-informed pathology-genomic deep model that integrates pathological images and genomic data not only to improve survival prediction but also to identify genes and pathways that cause different survival rates in patients. Empirical results on six of The Cancer Genome Atlas (TCGA) datasets show that our proposed method achieves superior predictive performance and reveals meaningful biological interpretations. The proposed method establishes insight into how to train biologically informed deep networks on multimodal biomedical data which will have general applicability for understanding diseases and predicting response and resistance to treatment.

The field of cancer research has greatly benefited from the wealth of new knowledge provided by research articles and preprints on platforms like Biorxiv. This study investigates the role of scientific figures and their accompanying captions in enhancing our comprehension of cancer. Leveraging the capabilities of Multimodal Large Language Models (MLLMs), we conduct a comprehensive analysis of both visual and linguistic data in biomedical literature. Our work introduces VLIB, a substantial scientific figure-caption dataset generated from cancer biology papers on Biorxiv. After thorough preprocessing, which includes figure-caption pair extraction, sub-figure identification, and text normalization, VLIB comprises over 500,000 figures from more than 70,000 papers, each accompanied by relevant captions. We fine-tune baseline MLLMs using our VLIB dataset for downstream vision-language tasks, such as image captioning and visual question answering (VQA), to assess their performance. Our experimental results underscore the vital role played by scientific figures, including molecular structures, histological images, and data visualizations, in conjunction with their captions, in facilitating knowledge translation through MLLMs. Specifically, we achieved a ROUGE score of 0.66 for VQA and 0.68 for image captioning, as well as a BLEU score of 0.72 for VQA and 0.70 for image captioning. Furthermore, our investigation highlights the potential of MLLMs to bridge the gap between artificial intelligence and domain experts in the field of cancer biology.

Antigen presentation on MHC class I (MHC-I) is key to the adaptive immune response to cancerous cells. Computational prediction of peptide presentation by MHC-I has enabled individualized cancer immunotherapies. Here, we introduce HLApollo, a transformer-based approach with end-to-end modeling of MHC-I sequence, deconvolution, and flanking sequences. To achieve this, we develop a novel training strategy, negative set switching, which greatly reduces overfitting to falsely presumed negatives that are necessarily found in presentation datasets. HLApollo shows a meaningful improvement compared to recent MHC-I models on peptide presentation (20.19% average precision (AP)) and immunogenicity (4.1% AP). As expected, adding gene expression boosts the performance of HLApollo. More interestingly, we show that introduction of features from a protein language model, ESM 1b, remarkably recoups much of the benefits of gene expression in absence of true expression measurements. Finally, we demonstrate excellent pan-allelic generalization, and introduce a framework for estimating the expected accuracy of HLApollo for untrained alleles. This guides the use of HLApollo in a clinical setting, where rare alleles may be observed in some subjects, particularly for underrepresented minorities.

Flexible intramedullary nailing is gaining popularity as an effective method of treating long-bone fractures in children.We retrospectively reviewed the records and radiographs of 56 unstable fractures of the tibia in 54 children treated between March 1997 and May 2005. All were followed up for at least two months after the removal of the nails.Of the 56 tibial fractures, 13 were open. There were no nonunions. The mean time to clinical and radiological union was ten weeks. Complications included residual angulation of the tibia, leg-length discrepancy, deep infection and failures of fixation. All achieved an excellent functional outcome.We conclude that flexible intramedullary fixation is an easy and effective method of management of both open and closed unstable fractures of the tibia in children.