The identification of cell-type-specific 3D chromatin interactions between regulatory elements can help to decipher gene regulation and to interpret the function of disease-associated non-coding variants. However, current chromosome conformation capture (3C) technologies are unable to resolve interactions at this resolution when only small numbers of cells are available as input. We therefore present ChromaFold, a deep learning model that predicts 3D contact maps and regulatory interactions from single-cell ATAC sequencing (scATAC-seq) data alone. ChromaFold uses pseudobulk chromatin accessibility, co-accessibility profiles across metacells, and predicted CTCF motif tracks as input features and employs a lightweight architecture to enable training on standard GPUs. Once trained on paired scATAC-seq and Hi-C data in human cell lines and tissues, ChromaFold can accurately predict both the 3D contact map and peak-level interactions across diverse human and mouse test cell types. In benchmarking against a recent deep learning method that uses bulk ATAC-seq, DNA sequence, and CTCF ChIP-seq to make cell-type-specific predictions, ChromaFold yields superior prediction performance when including CTCF ChIP-seq data as an input and comparable performance without. Finally, fine-tuning ChromaFold on paired scATAC-seq and Hi-C in a complex tissue enables deconvolution of chromatin interactions across cell subpopulations. ChromaFold thus achieves state-of-the-art prediction of 3D contact maps and regulatory interactions using scATAC-seq alone as input data, enabling accurate inference of cell-type-specific interactions in settings where 3C-based assays are infeasible.

A major problem with Active Learning (AL) is high training costs since models are typically retrained from scratch after every query round. We start by demonstrating that standard AL on neural networks with warm starting fails, both to accelerate training and to avoid catastrophic forgetting when using fine-tuning over AL query rounds. We then develop a new class of techniques, circumventing this problem, by biasing further training towards previously labeled sets. We accomplish this by employing existing, and developing novel, replay-based Continual Learning (CL) algorithms that are effective at quickly learning the new without forgetting the old, especially when data comes from an evolving distribution. We call this paradigm Continual Active Learning (CAL). We show CAL achieves significant speedups using a plethora of replay schemes that use model distillation and that select diverse, uncertain points from the history. We conduct experiments across many data domains, including natural language, vision, medical imaging, and computational biology, each with different neural architectures and dataset sizes. CAL consistently provides a 3x reduction in training time, while retaining performance.

Lineage plasticity, the ability of cells to transdifferentiate between committed developmental pathways, has been proposed as a source of intratumoral heterogeneity and as a mechanism for tumor adaptation to stringent environmental conditions. Lineage plasticity is increasingly recognized as a contributor to both drug resistance and metastasis, as recently highlighted by our team (Quintanal-Villalonga et al., Nat Rev Clin Oncol 2020), thus representing a biological phenomenon with high clinical relevance. Leveraging bulk and single-cell multi-omic approaches in clinical specimens, we aimed to perform a comprehensive molecular characterization of lung cancer cellular plasticity in the settings of (1) histological transdifferentiation as a mechanism of resistance to targeted therapy, and (2) disease progression. In lung adenocarcinomas (LUADs), lineage plasticity drives small cell (SCLC) and squamous cell (LUSC) transdifferentiation in the context of acquired resistance to targeted inhibitors. In the EGFR -mutant cancers, transformation to SCLC and to LUSC as a mechanism of acquired TKI resistance has been reported in 14% and 9% of cases, respectively. Transdifferentiated tumors portend poorer prognosis than non-transformed tumors. To date, no preventative therapies for transformation are available, although tumor subsets at high risk of transformation (concomitant TP53/RB1 inactivation in the case of SCLC transdifferentiation) have been identified. Defining molecular mechanisms of histological transformation in lung cancer has been challenging due in part to a paucity of well-annotated pre- and post-transformation clinical samples. We hypothesized that mixed histology tumors (LUAD/SCLC and LUAD/LUSC), containing different histological components and sharing common driver mutations, may represent an intermediate step of transdifferentiation, and an ideal substrate to study lineage plasticity, with both histological components sharing location, time, and treatment influence. By selecting mixed histology specimens amenable for clean microdissection of each histological component (10 LUAD/LUSC and 11 LUAD/SCLC), as well as pre- and post-transdifferentiated samples (N=17), we performed comprehensive genomic, epigenomic, transcriptomic, and protein characterization. Our data supports that histological transdifferentiation from LUAD to LUSC or SCLC tumors is driven by epigenetic remodeling rather than by mutational events, and indicates that transdifferentiated tumors retain epigenomic features of their previous LUAD state. Integrative epigenomic, transcriptomic, and protein analysis revealed divergent biological pathways dysregulated for each histological outcome, such as upregulation of genes involved in Hedgehog and Notch signaling and MYC targets in LUSC-transdifferentiated tumors. Most interestingly, these analyses identified commonly dysregulated pathways in both SCLC- and LUSC-transdifferentiating tumors, including remarkable downregulation of a variety of immune-related pathways and upregulation of genes involved in the PRC2 epigenetic remodeling complex and AKT signaling. To validate drivers of transdifferentiation from our multi-omic data, preclinical in vivo experiments indicated that concurrent activation of AKT and MYC overexpression induced a LUSC phenotype with increased P40 and CK5/6 expression in EGFR -mutant LUAD patient-derived models. LUSC features in these models were further accentuated by the EGFR inhibitor osimertinib, which enriched for transdifferentiated LUSC cells. With the aim to validate potential therapeutic targets to prevent transdifferentiation from our multi-omic data, we tested the efficacy of EZH2 (the catalytic subunit of PRC2) or AKT pharmacological inhibitors in combination with osimertinib in an EGFR -mutant patient-derived xenograft model of LUAD-to-LUSC transdifferentiation and observed that inhibition of either pathway dramatically delayed relapse and prevented emergence of LUSC phenotypic markers. Additionally, pharmacological targeting of AKT and EGFR delayed relapse in a mixed LUAD/SCLC patient-derived xenograft model representing an intermediate step of LUAD-to-SCLC transdifferentiation. Interestingly, AKT inhibition selectively targeted the SCLC compartment of the tumor and prevented full SCLC transformation. This work was published in two manuscripts, of which I am the first and co-corresponding author (Quintanal-Villalonga et al., Cancer Discov 2021; Quintanal-Villalonga et al., J Hematol Oncol 2021) and defines a novel landscape of potential drivers and therapeutic vulnerabilities of histological transdifferentiation in lung cancer. Both transformed and de novo SCLCs are aggressive and rapidly metastatic lung tumors. Limited treatment options and transient responses translate to poor prognosis for patients with SCLC; 5-year survival rates are <1% for extensive disease, and SCLC accounts for >200,000 annual deaths worldwide. Metastasis is the main cause of mortality among patients with SCLC. To characterize SCLC metastasis, we combined single-cell RNA sequencing (scRNA-seq) and multiplexed ion beam imaging (MIBI) technologies to study intratumoral heterogeneity and the surrounding tumor microenvironment (TME). Efforts to apply these technologies to human SCLC tumors have been limited, as surgical resections of primary tumors are performed in <5% of patients with SCLC, and scRNA-seq processing of biopsied samples is extremely challenging. Additionally, since resection is only clinically indicated for exceptionally early stage de novo disease, these samples fail to capture the spectrum of disease progression. Through the optimization of protocols allowing single cell profiling of difficult samples such as small tissue biopsies, pleural effusions, and fine needle aspirations, along with larger volume resections, we constructed a single-cell atlas of SCLC patient tumors, with 155,098 transcriptomes, including 54,523 transcriptomes from 21 SCLC clinical specimens. Despite substantial heterogeneity among SCLC tumors, we detected a minor cell subpopulation that was shared among tumors across subtypes, treatments, and tissue locations, pointing to a potentially universal characteristic of this malignancy. This subpopulation demonstrated a pro-metastatic, highly plastic (stem-like) phenotype and exhibited profound PLCG2 overexpression. Direct genetic manipulation validated that PLCG2 expression promotes metastatic features and induced plasticity in vitro and in vivo. Consistently, we found that higher representation of this subpopulation in clinical samples, as well as PLCG2 expression itself, are strong predictors of shorter overall survival in patients with SCLC. Additionally, we found that SCLC is enriched for a profibrotic, immunosuppressive monocyte/macrophage population associated with the recurrent pro-metastatic PLCG2-high subpopulation, with potential implications in the metastatic process. I am the co-first author of this work published in Cancer Cell (Chan JM*, Quintanal-Villalonga* et al., Cancer Cell 2021) defining a novel mechanism of plasticity-mediated metastasis in SCLC. These works highlight the key role of plasticity in disease progression and therapy resistance in lung cancer, and describe molecular events occurring during fate reprogramming, thus nominating potential drivers and therapeutic vulnerabilities in plasticity-driven clinically relevant biological processes. Citation Format: Alvaro Quintanal-Villalonga, Joseph M. Chan, Vianne R. Gao, Yubin Xie, Dana Pe’er, Charles M. Rudin. Multi-omic approaches to study the role of plasticity in therapy resistance and metastasis in lung cancer. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr NG05.

The response to tumor-initiating inflammatory and genetic insults can vary among morphologically indistinguishable cells, suggesting as yet uncharacterized roles for epigenetic plasticity during early neoplasia. To investigate the origins and impact of such plasticity, we performed single-cell analyses on normal, inflamed, premalignant, and malignant tissues in autochthonous models of pancreatic cancer. We reproducibly identified heterogeneous cell states that are primed for diverse, late-emerging neoplastic fates and linked these to chromatin remodeling at cell-cell communication loci. Using an inference approach, we revealed signaling gene modules and tissue-level cross-talk, including a neoplasia-driving feedback loop between discrete epithelial and immune cell populations that was functionally validated in mice. Our results uncover a neoplasia-specific tissue-remodeling program that may be exploited for pancreatic cancer interception.

Recent deep learning models that predict the Hi-C contact map from DNA sequence achieve promising accuracy but cannot generalize to new cell types and or even capture differences among training cell types. We propose Epiphany, a neural network to predict cell-type-specific Hi-C contact maps from widely available epigenomic tracks. Epiphany uses bidirectional long short-term memory layers to capture long-range dependencies and optionally a generative adversarial network architecture to encourage contact map realism. Epiphany shows excellent generalization to held-out chromosomes within and across cell types, yields accurate TAD and interaction calls, and predicts structural changes caused by perturbations of epigenomic signals.

Background One challenge in the field of in-vitro fertilisation is the selection of the most viable embryos for transfer. Morphological quality assessment and morphokinetic analysis both have the disadvantage of intra-observer and interobserver variability. A third method, preimplantation genetic testing for aneuploidy (PGT-A), has limitations too, including its invasiveness and cost. We hypothesised that differences in aneuploid and euploid embryos that allow for model-based classification are reflected in morphology, morphokinetics, and associated clinical information. Methods In this retrospective study, we used machine-learning and deep-learning approaches to develop STORK-A, a non-invasive and automated method of embryo evaluation that uses artificial intelligence to predict embryo ploidy status. Our method used a dataset of 10 378 embryos that consisted of static images captured at 110 h after intracytoplasmic sperm injection, morphokinetic parameters, blastocyst morphological assessments, maternal age, and ploidy status. Independent and external datasets, Weill Cornell Medicine EmbryoScope+ (WCM-ES+; Weill Cornell Medicine Center of Reproductive Medicine, NY, USA) and IVI Valencia (IVI Valencia, Health Research Institute la Fe, Valencia, Spain) were used to test the generalisability of STORK-A and were compared measuring accuracy and area under the receiver operating characteristic curve (AUC). Findings Analysis and model development included the use of 10 378 embryos, all with PGT-A results, from 1385 patients (maternal age range 21-48 years; mean age 36.98 years [SD 4.62]). STORK-A predicted aneuploid versus euploid embryos with an accuracy of 69.3% (95% CI 66.9-71.5; AUC 0.761; positive predictive value [PPV] 76.1%; negative predictive value [NPV] 62.1%) when using images, maternal age, morphokinetics, and blastocyst score. A second classification task trained to predict complex aneuploidy versus euploidy and single aneuploidy produced an accuracy of 74.0% (95% CI 71.7-76.1; AUC 0.760; PPV 54.9%; NPV 87.6%) using an image, maternal age, morphokinetic parameters, and blastocyst grade. A third classification task trained to predict complex aneuploidy versus euploidy had an accuracy of 77.6% (95% CI 75.0-80.0; AUC 0.847; PPV 76.7%; NPV 78.0%). STORK-A reported accuracies of 63.4% (AUC 0.702) on the WCM-ES+ dataset and 65.7% (AUC 0.715) on the IVI Valencia dataset, when using an image, maternal age, and morphokinetic parameters, similar to the STORK-A test dataset accuracy of 67.8% (AUC 0.737), showing generalisability. Interpretation As a proof of concept, STORK-A shows an ability to predict embryo ploidy in a non-invasive manner and shows future potential as a standardised supplementation to traditional methods of embryo selection and prioritisation for implantation or recommendation for PGT-A.

Inflammation of non-barrier immunologically quiescent tissues is associated with a massive influx of blood-borne innate and adaptive immune cells. Cues from the latter are likely to alter and expand activated states of the resident cells. However, local communications between immigrant and resident cell types in human inflammatory disease remain poorly understood. Here, we explored drivers of fibroblast-like synoviocyte (FLS) heterogeneity in inflamed joints of patients with rheumatoid arthritis using paired single-cell RNA and ATAC sequencing, multiplexed imaging and spatial transcriptomics along with in vitro modeling of cell-extrinsic factor signaling. These analyses suggest that local exposures to myeloid and T cell-derived cytokines, TNF, IFN-γ, IL-1β or lack thereof, drive four distinct FLS states some of which closely resemble fibroblast states in other disease-affected tissues including skin and colon. Our results highlight a role for concurrent, spatially distributed cytokine signaling within the inflamed synovium.

While active learning (AL) improves the labeling efficiency of machine learning (by allowing models to query the labels of data samples), a major problem is that compute efficiency is decreased since models are typically retrained from scratch at each query round. In this work, we develop a new framework that circumvents this problem by biasing further training towards the recently labeled sets, thereby complementing existing work on AL acceleration. We employ existing and novel replay-based Continual Learning (CL) algorithms that are effective at quickly learning new samples without forgetting previously learned information, especially when data comes from a shifting or evolving distribution. We call this compute-efficient active learning paradigm $\textit{``Continual Active Learning" (CAL)}$. We demonstrate that standard AL with warm starting fails, both to accelerate training, and that naive fine-tuning suffers from catastrophic forgetting due to distribution shifts over query rounds. We then show CAL achieves significant speedups using a plethora of replay schemes that use model distillation, and that select diverse/uncertain points from the history, all while maintaining performance on par with standard AL. We conduct experiments across many data domains, including natural language, vision, medical imaging, and computational biology, each with very different neural architectures (Transformers/CNNs/MLPs). CAL consistently provides a 2-6x reduction in training time, thus showing its applicability across differing modalities.

The response to tumor-initiating inflammatory and genetic insults can vary amongst morphologically indistinguishable cells, suggesting yet uncharacterized roles for epigenetic plasticity during early neoplasia. To investigate the origins and impact of such plasticity, we perform single-cell analyses on normal, inflamed, pre-malignant and malignant tissues in autochthonous models of pancreatic cancer. We reproducibly identify heterogeneous cell-states that are primed for diverse late-emerging neoplastic fates and link these to chromatin remodeling at cell-cell communication loci. Using a new inference approach, we reveal signaling gene modules and tissue-level crosstalk, including a neoplasia-driving feedback loop between discrete epithelial and immune cell populations that we validate by genetic perturbation in mice. Our results uncover a neoplasia-specific tissue remodeling program that may be exploited for pancreas cancer interception. One-Sentence Summary Single-cell analysis reveals that enhanced epigenetic plasticity drives pro-neoplastic crosstalk in early pancreatic cancer.

We report a protocol for obtaining high-quality single-cell transcriptomics data from human lung biospecimens acquired from core needle biopsies, fine-needle aspirates, surgical resection, and pleural effusions. The protocol relies upon the brief mechanical and enzymatic disruption of tissue, enrichment of live cells by fluorescence-activated cell sorting (FACS), and droplet-based single-cell RNA sequencing (scRNA-seq). The protocol also details a procedure for analyzing the scRNA-seq data. For complete details on the use and execution of this protocol, please refer to Chan et al. (2021).

Monte Carlo (MC) dropout is a simple and efficient ensembling method that can improve the accuracy and confidence calibration of high-capacity deep neural network models. However, MC dropout is not as effective as more compute-intensive methods such as deep ensembles. This performance gap can be attributed to the relatively poor quality of individual models in the MC dropout ensemble and their lack of diversity. These issues can in turn be traced back to the coupled training and substantial parameter sharing of the dropout models. Motivated by this perspective, we propose a strategy to compute an ensemble of subnetworks, each corresponding to a non-overlapping dropout mask computed via a pruning strategy and trained independently. We show that the proposed subnetwork ensembling method can perform as well as standard deep ensembles in both accuracy and uncertainty estimates, yet with a computational efficiency similar to MC dropout. Lastly, using several computer vision datasets like CIFAR10/100, CUB200, and Tiny-Imagenet, we experimentally demonstrate that subnetwork ensembling also consistently outperforms recently proposed approaches that efficiently ensemble neural networks.

We propose Epiphany, a light-weight neural network to predict the Hi-C contact map from five commonly generated epigenomic tracks: DNase I hypersensitive sites and CTCF, H3K27ac, H3K27me3, and H3K4me3 ChIP-seq. Epiphany uses 1D convolutional layers to learn local representations from the input tracks as well as bidirectional Long Short Term Memory (Bi-LSTM) layers to capture long term dependencies along the epigenome. To improve the usability of predicted contact matrices, we perform statistically principled preprocessing of Hi-C data using HiC-DC+ (1) and train Epiphany using an adversarial loss, enhancing its ability to produce realistic Hi-C contact maps for downstream analysis. We show that Epiphany generalizes to held-out chromosomes within and across cell types, and that Epiphany’s predicted contact matrices yield accurate TAD and significant interaction calls.

Small cell lung cancer (SCLC) is an aggressive malignancy that includes subtypes defined by differential expression of ASCL1, NEUROD1, and POU2F3 (SCLC-A,-N, and-P, respectively). To define the heterogeneity of tumors and their associated microenvironments across subtypes, we sequenced 155,098 transcriptomes from 21 human biospecimens, including 54,523 SCLC transcriptomes. We observe greater tumor diversity in SCLC than lung adenocarcinoma, driven by canonical, intermediate, and admixed subtypes. We discover a PLCG2-high SCLC phenotype with stem-like, pro-metastatic features that recurs across subtypes and predicts worse overall survival. SCLC exhibits greater immune sequestration and less immune infiltration than lung adenocarcinoma, and SCLC-N shows less immune infiltrate and greater T cell dysfunction than SCLC-A. We identify a profibrotic, immunosuppressive monocyte/macrophage population in SCLC tumors that is particularly associated with the recurrent, PLCG2-high subpopulation.

ABSTRACT Small cell lung cancer (SCLC) is an aggressive malignancy that includes subtypes defined by differential expression of ASCL1 , NEUROD1 , and POU2F3 (SCLC-A, -N, and -P, respectively), which are associated with distinct therapeutic vulnerabilities. To define the heterogeneity of tumors and their associated microenvironments across subtypes, we sequenced 54,523 cellular transcriptomes from 21 human biospecimens. Our single-cell SCLC atlas reveals tumor diversity exceeding lung adenocarcinoma, driven by canonical, intermediate, and admixed subtypes. We discovered a PLCG2 -high tumor cell population with stem-like, pro-metastatic features that recurs across subtypes and predicts worse overall survival, and manipulation of PLCG2 expression in cells confirms correlation with key metastatic markers. Treatment and subtype are associated with substantial phenotypic changes in the SCLC immune microenvironment, with greater T-cell dysfunction in SCLC-N than SCLC-A. Moreover, the recurrent, PLCG2- high subclone is associated with exhausted CD8+ T-cells and a pro-fibrotic, immunosuppressive monocyte/macrophage population, suggesting possible tumor-immune coordination to promote metastasis.