UFL1 ablation in T cells suppresses PD-1 UFMylation to enhance anti-tumor immunity.

UFMylation is an emerging ubiquitin-like post-translational modification that regulates various biological processes. Dysregulation of the UFMylation pathway leads to human diseases, including cancers. However, the physiological role of UFMylation in T cells remains unclear. Here, we report that mice with conditional knockout (cKO) Ufl1, a UFMylation E3 ligase, in T cells exhibit effective tumor control. Single-cell RNA sequencing analysis shows that tumor-infiltrating cytotoxic CD8+ T cells are increased in Ufl1 cKO mice. Mechanistically, UFL1 promotes PD-1 UFMylation to antagonize PD-1 ubiquitination and degradation. Furthermore, AMPK phosphorylates UFL1 at Thr536, disrupting PD-1 UFMylation to trigger its degradation. Of note, UFL1 ablation in T cells reduces PD-1 UFMylation, subsequently destabilizing PD-1 and enhancing CD8+ T cell activation. Thus, Ufl1 cKO mice bearing tumors have a better response to anti-CTLA-4 immunotherapy. Collectively, our findings uncover a crucial role of UFMylation in T cells and highlight UFL1 as a potential target for cancer treatment.

Visualizing interfacial collective reaction behaviour of Li-S batteries.

Benefiting from high energy density (2,600 Wh kg-1) and low cost, lithium-sulfur (Li-S) batteries are considered promising candidates for advanced energy-storage systems1-4. Despite tremendous efforts in suppressing the long-standing shuttle effect of lithium polysulfides5-7, understanding of the interfacial reactions of lithium polysulfides at the nanoscale remains elusive. This is mainly because of the limitations of in situ characterization tools in tracing the liquid-solid conversion of unstable lithium polysulfides at high temporal-spatial resolution8-10. There is an urgent need to understand the coupled phenomena inside Li-S batteries, specifically, the dynamic distribution, aggregation, deposition and dissolution of lithium polysulfides. Here, by using in situ liquid-cell electrochemical transmission electron microscopy, we directly visualized the transformation of lithium polysulfides over electrode surfaces at the atomic scale. Notably, an unexpected gathering-induced collective charge transfer of lithium polysulfides was captured on the nanocluster active-centre-immobilized surface. It further induced an instantaneous deposition of nonequilibrium Li2S nanocrystals from the dense liquid phase of lithium polysulfides. Without mediation of active centres, the reactions followed a classical single-molecule pathway, lithium polysulfides transforming into Li2S2 and Li2S step by step. Molecular dynamics simulations indicated that the long-range electrostatic interaction between active centres and lithium polysulfides promoted the formation of a dense phase consisting of Li+ and Sn2- (2 < n ≤ 6), and the collective charge transfer in the dense phase was further verified by ab initio molecular dynamics simulations. The collective interfacial reaction pathway unveils a new transformation mechanism and deepens the fundamental understanding of Li-S batteries.

USP8-governed GPX4 homeostasis orchestrates ferroptosis and cancer immunotherapy.

Ferroptosis is an iron-dependent type of regulated cell death resulting from extensive lipid peroxidation and plays a critical role in various physiological and pathological processes. However, the regulatory mechanisms for ferroptosis sensitivity remain incompletely understood. Here, we report that homozygous deletion of Usp8 (ubiquitin-specific protease 8) in intestinal epithelial cells (IECs) leads to architectural changes in the colonic epithelium and shortens mouse lifespan accompanied by increased IEC death and signs of lipid peroxidation. However, mice with heterozygous deletion of Usp8 in IECs display normal phenotype and become resistant to azoxymethane/dextran sodium sulfate-induced colorectal tumorigenesis. Mechanistically, USP8 interacts with and deubiquitinates glutathione peroxidase 4 (GPX4), leading to GPX4 stabilization. Thus, USP8 inhibition destabilizes GPX4 and sensitizes cancer cells to ferroptosis in vitro. Notably, USP8 inhibition in combination with ferroptosis inducers retards tumor growth and enhances CD8+ T cell infiltration, which potentiates tumor response to anti-PD-1 immunotherapy in vivo. These findings uncover that USP8 counteracts ferroptosis by stabilizing GPX4 and highlight targeting USP8 as a potential therapeutic strategy to boost ferroptosis for enhancing cancer immunotherapy.

Ubiquitin-induced RNF168 condensation promotes DNA double-strand break repair.

Rapid accumulation of repair factors at DNA double-strand breaks (DSBs) is essential for DSB repair. Several factors involved in DSB repair have been found undergoing liquid-liquid phase separation (LLPS) at DSB sites to facilitate DNA repair. RNF168, a RING-type E3 ubiquitin ligase, catalyzes H2A.X ubiquitination for recruiting DNA repair factors. Yet, whether RNF168 undergoes LLPS at DSB sites remains unclear. Here, we identified K63-linked polyubiquitin-triggered RNF168 condensation which further promoted RNF168-mediated DSB repair. RNF168 formed liquid-like condensates upon irradiation in the nucleus while purified RNF168 protein also condensed in vitro. An intrinsically disordered region containing amino acids 460-550 was identified as the essential domain for RNF168 condensation. Interestingly, LLPS of RNF168 was significantly enhanced by K63-linked polyubiquitin chains, and LLPS largely enhanced the RNF168-mediated H2A.X ubiquitination, suggesting a positive feedback loop to facilitate RNF168 rapid accumulation and its catalytic activity. Functionally, LLPS deficiency of RNF168 resulted in delayed recruitment of 53BP1 and BRCA1 and subsequent impairment in DSB repair. Taken together, our finding demonstrates the pivotal effect of LLPS in RNF168-mediated DSB repair.

Kinetic pathway of HIV-1 TAR cotranscriptional folding.

The Trans-Activator Receptor (TAR) RNA, located at the 5'-end untranslated region (5' UTR) of the human immunodeficiency virus type 1 (HIV-1), is pivotal in the virus's life cycle. As the initial functional domain, it folds during the transcription of viral mRNA. Although TAR's role in recruiting the Tat protein for trans-activation is established, the detailed kinetic mechanisms at play during early transcription, especially at points of temporary transcriptional pausing, remain elusive. Moreover, the precise physical processes of transcriptional pause and subsequent escape are not fully elucidated. This study focuses on the folding kinetics of TAR and the biological implications by integrating computer simulations of RNA folding during transcription with nuclear magnetic resonance (NMR) spectroscopy data. The findings reveal insights into the folding mechanism of a non-native intermediate that triggers transcriptional pause, along with different folding pathways leading to transcriptional pause and readthrough. The profiling of the cotranscriptional folding pathway and identification of kinetic structural intermediates reveal a novel mechanism for viral transcriptional regulation, which could pave the way for new antiviral drug designs targeting kinetic cotranscriptional folding pathways in viral RNAs.

A disinhibitory microcircuit of the orbitofrontal cortex mediates cocaine preference in mice.

Both clinical and animal studies showed that the impaired functions of the orbitofrontal cortex (OFC) underlie the compulsive drug-seeking behavior of drug addiction. However, the functional changes of the microcircuit in the OFC and the underlying molecular mechanisms in drug addiction remain elusive, and little is known for whether microcircuits in the OFC contributed to drug addiction-related behaviors. Utilizing the cocaine-induced conditioned-place preference model, we found that the malfunction of the microcircuit led to disinhibition in the OFC after cocaine withdrawal. We further showed that enhanced Somatostatin-Parvalbumin (SST-PV) inhibitory synapse strength changed microcircuit function, and SST and PV inhibitory neurons showed opposite contributions to the drug addiction-related behavior of mice. Brevican of the perineuronal nets of PV neurons regulated SST-PV synapse strength, and the knockdown of Brevican alleviated cocaine preference. These results reveal a novel molecular mechanism of the regulation of microcircuit function and a novel circuit mechanism of the OFC in gating cocaine preference.

Painful physical symptoms and antidepressant treatment outcome in depression: a systematic review and meta-analysis.

BACKGROUND: Painful physical symptoms (PPS) are highly prevalent in patients with major depressive disorder (MDD). Presence of PPS in depressed patients are potentially associated with poorer antidepressant treatment outcome. We aimed to evaluate the association of baseline pain levels and antidepressant treatment outcomes. METHODS: We searched PubMed, Embase and Cochrane Library databases from inception through February 2023 based on a pre-registered protocol (PROSPERO: CRD42022381349). We included original studies that reported pretreatment pain measures in antidepressant treatment responder/remitter and non-responder/non-remitter among patients with MDD. Data extraction and quality assessment were performed following the Preferred Reporting Items for Systematic Reviews and Meta-analyses by two reviewers independently. The primary outcome was the difference of the pretreatment pain levels between antidepressant treatment responder/remitter and non-responder/non-remitter. Random-effects meta-analysis was used to calculate effect sizes (Hedge's g) and subgroup and meta-regression analyses were used to explore sources of heterogeneity. RESULTS: A total of 20 studies were included. Six studies reported significantly higher baseline pain severity levels in MDD treatment non-responders (Hedge's g = 0.32; 95% CI, 0.13-0.51; P = 0.0008). Six studies reported the presence of PPS (measured using a pain severity scale) was significantly associated with poor treatment response (OR = 1.46; 95% CI, 1.04-2.04; P = 0.028). Five studies reported significant higher baseline pain interference levels in non-responders (Hedge's g = 0.46; 95% CI, 0.32-0.61; P < 0.0001). Four studies found significantly higher baseline pain severity levels in non-remitters (Hedge's g = 0.27; 95% CI, 0.14-0.40; P < 0.0001). Eight studies reported the presence of PPS significantly associated with treatment non-remission (OR = 1.70; 95% CI, 1.24-2.32; P = 0.0009). CONCLUSIONS: This study suggests that PPS are negatively associated with the antidepressant treatment outcome in patients with MDD. It is possible that better management in pain conditions when treating depression can benefit the therapeutic effects of antidepressant medication in depressed patients.

Memory consolidation drives the enhancement of remote cocaine memory via prefrontal circuit.

Remote memory usually decreases over time, whereas remote drug-cue associated memory exhibits enhancement, increasing the risk of relapse during abstinence. Memory system consolidation is a prerequisite for remote memory formation, but neurobiological underpinnings of the role of consolidation in the enhancement of remote drug memory are unclear. Here, we found that remote cocaine-cue associated memory was enhanced in rats that underwent self-administration training, together with a progressive increase in the response of prelimbic cortex (PrL) CaMKII neurons to cues. System consolidation was required for the enhancement of remote cocaine memory through PrL CaMKII neurons during the early period post-training. Furthermore, dendritic spine maturation in the PrL relied on the basolateral amygdala (BLA) input during the early period of consolidation, contributing to remote memory enhancement. These findings indicate that memory consolidation drives the enhancement of remote cocaine memory through a time-dependent increase in activity and maturation of PrL CaMKII neurons receiving a sustained BLA input.

Association between alcohol consumption and sleep traits: observational and mendelian randomization studies in the UK biobank.

Previous studies have shown that excessive alcohol consumption is associated with poor sleep. However, the health risks of light-to-moderate alcohol consumption in relation to sleep traits (e.g., insomnia, snoring, sleep duration and chronotype) remain undefined, and their causality is still unclear in the general population. To identify the association between alcohol consumption and multiple sleep traits using an observational and Mendelian randomization (MR) design. Observational analyses and one-sample MR (linear and nonlinear) were performed using clinical and individual-level genetic data from the UK Biobank (UKB). Two-sample MR was assessed using summary data from genome-wide association studies from the UKB and other external consortia. Phenotype analyses were externally validated using data from the National Health and Nutrition Examination Survey (2017-2018). Data analysis was conducted from January 2022 to October 2022. The association between alcohol consumption and six self-reported sleep traits (short sleep duration, long sleep duration, chronotype, snoring, waking up in the morning, and insomnia) were analysed. This study included 383,357 UKB participants (mean [SD] age, 57.0 [8.0] years; 46% male) who consumed a mean (SD) of 9.0 (10.0) standard drinks (one standard drink equivalent to 14 g of alcohol) per week. In the observational analyses, alcohol consumption was significantly associated with all sleep traits. Light-moderate-heavy alcohol consumption was linearly linked to snoring and the evening chronotype but nonlinearly associated with insomnia, sleep duration, and napping. In linear MR analyses, a 1-SD (14 g) increase in genetically predicted alcohol consumption was associated with a 1.14-fold (95% CI, 1.07-1.22) higher risk of snoring (P < 0.001), a 1.28-fold (95% CI, 1.20-1.37) higher risk of evening chronotype (P < 0.001) and a 1.24-fold (95% CI, 1.13-1.36) higher risk of difficulty waking up in the morning (P < 0.001). Nonlinear MR analyses did not reveal significant results after Bonferroni adjustment. The results of the two-sample MR analyses were consistent with those of the one-sample MR analyses, but with a slightly attenuated overall estimate. Our findings suggest that even low levels of alcohol consumption may affect sleep health, particularly by increasing the risk of snoring and evening chronotypes. The negative effects of alcohol consumption on sleep should be made clear to the public in order to promote public health.

KRAS4A directly regulates hexokinase 1.

The most frequently mutated oncogene in cancer is KRAS, which uses alternative fourth exons to generate two gene products (KRAS4A and KRAS4B) that differ only in their C-terminal membrane-targeting region1. Because oncogenic mutations occur in exons 2 or 3, two constitutively active KRAS proteins-each capable of transforming cells-are encoded when KRAS is activated by mutation2. No functional distinctions among the splice variants have so far been established. Oncogenic KRAS alters the metabolism of tumour cells3 in several ways, including increased glucose uptake and glycolysis even in the presence of abundant oxygen4 (the Warburg effect). Whereas these metabolic effects of oncogenic KRAS have been explained by transcriptional upregulation of glucose transporters and glycolytic enzymes3-5, it is not known whether there is direct regulation of metabolic enzymes. Here we report a direct, GTP-dependent interaction between KRAS4A and hexokinase 1 (HK1) that alters the activity of the kinase, and thereby establish that HK1 is an effector of KRAS4A. This interaction is unique to KRAS4A because the palmitoylation-depalmitoylation cycle of this RAS isoform enables colocalization with HK1 on the outer mitochondrial membrane. The expression of KRAS4A in cancer may drive unique metabolic vulnerabilities that can be exploited therapeutically.

Using normative modeling to assess pharmacological treatment effect on brain state in patients with schizophrenia.

Quantifying individual differences in neuroimaging metrics is attracting interest in clinical studies with mental disorders. Schizophrenia is diagnosed exclusively based on symptoms, and the biological heterogeneity makes it difficult to accurately assess pharmacological treatment effects on the brain state. Using the Cambridge Centre for Ageing and Neuroscience data set, we built normative models of brain states and mapped the deviations of the brain characteristics of each patient, to test whether deviations were related to symptoms, and further investigated the pharmacological treatment effect on deviation distributions. Specifically, we found that the patients can be divided into 2 groups: the normalized group had a normalization trend and milder symptoms at baseline, and the other group showed a more severe deviation trend. The baseline severity of the depression as well as the overall symptoms could predict the deviation of the static characteristics for the dorsal and ventral attention networks after treatment. In contrast, the positive symptoms could predict the deviations of the dynamic fluctuations for the default mode and dorsal attention networks after treatment. This work evaluates the effect of pharmacological treatment on static and dynamic brain states using an individualized approach, which may assist in understanding the heterogeneity of the illness pathology as well as the treatment response.

RNAJP: enhanced RNA 3D structure predictions with non-canonical interactions and global topology sampling.

RNA 3D structures are critical for understanding their functions. However, only a limited number of RNA structures have been experimentally solved, so computational prediction methods are highly desirable. Nevertheless, accurate prediction of RNA 3D structures, especially those containing multiway junctions, remains a significant challenge, mainly due to the complicated non-canonical base pairing and stacking interactions in the junction loops and the possible long-range interactions between loop structures. Here we present RNAJP ('RNA Junction Prediction'), a nucleotide- and helix-level coarse-grained model for the prediction of RNA 3D structures, particularly junction structures, from a given 2D structure. Through global sampling of the 3D arrangements of the helices in junctions using molecular dynamics simulations and in explicit consideration of non-canonical base pairing and base stacking interactions as well as long-range loop-loop interactions, the model can provide significantly improved predictions for multibranched junction structures than existing methods. Moreover, integrated with additional restraints from experiments, such as junction topology and long-range interactions, the model may serve as a useful structure generator for various applications.

RAP80 phase separation at DNA double-strand break promotes BRCA1 recruitment.

RAP80 has been characterized as a component of the BRCA1-A complex and is responsible for the recruitment of BRCA1 to DNA double-strand breaks (DSBs). However, we and others found that the recruitment of RAP80 and BRCA1 were not absolutely temporally synchronized, indicating that other mechanisms, apart from physical interaction, might be implicated. Recently, liquid-liquid phase separation (LLPS) has been characterized as a novel mechanism for the organization of key signaling molecules to drive their particular cellular functions. Here, we characterized that RAP80 LLPS at DSB was required for RAP80-mediated BRCA1 recruitment. Both cellular and in vitro experiments showed that RAP80 phase separated at DSB, which was ascribed to a highly disordered region (IDR) at its N-terminal. Meanwhile, the Lys63-linked poly-ubiquitin chains that quickly formed after DSBs occur, strongly enhanced RAP80 phase separation and were responsible for the induction of RAP80 condensation at the DSB site. Most importantly, abolishing the condensation of RAP80 significantly suppressed the formation of BRCA1 foci, encovering a pivotal role of RAP80 condensates in BRCA1 recruitment and radiosensitivity. Together, our study disclosed a new mechanism underlying RAP80-mediated BRCA1 recruitment, which provided new insight into the role of phase separation in DSB repair.

Scaling of quantum interference from single molecules to molecular cages and their monolayers.

The discovery of quantum interference (QI) is widely considered as an important advance in molecular electronics since it provides unique opportunities for achieving single-molecule devices with unprecedented performance. Although some pioneering studies suggested the presence of spin qubit coherence and QI in collective systems such as thin films, it remains unclear whether the QI can be transferred step-by-step from single molecules to different length scales, which hinders the application of QI in fabricating active molecular devices. Here, we found that QI can be transferred from a single molecule to their assemblies. We synthesized and investigated the charge transport through the molecular cages using 1,3-dipyridylbenzene (DPB) as a ligand block with a destructive quantum interference (DQI) effect and 2,5-dipyridylfuran (DPF) as a control building block with a constructive quantum interference (CQI) effect using both single-molecule break junction and large area junction techniques. Combined experiments and calculations revealed that both DQI and CQI had been transferred from the ligand blocks to the molecular cages and the monolayer thin film of the cages. Our work introduced QI effects from a ligand to the molecular cage comprising 732 atoms and even their monolayers, suggesting that the quantum interference could be scaled up within the phase-coherent distance.

Machine learning in RNA structure prediction: Advances and challenges.

RNA molecules play a crucial role in various biological processes, with their functionality closely tied to their structures. The remarkable advancements in machine learning techniques for protein structure prediction have shown promise in the field of RNA structure prediction. In this perspective, we discuss the advances and challenges encountered in constructing machine learning-based models for RNA structure prediction. We explore topics including model building strategies, specific challenges involved in predicting RNA secondary (2D) and tertiary (3D) structures, and approaches to these challenges. In addition, we highlight the advantages and challenges of constructing RNA language models. Given the rapid advances of machine learning techniques, we anticipate that machine learning-based models will serve as important tools for predicting RNA structures, thereby enriching our understanding of RNA structures and their corresponding functions.

Role of laminin and collagen chains in human spermatogenesis - Insights from studies in rodents and scRNA-Seq transcriptome profiling.

Studies have demonstrated that biologically active fragments are generated from the basement membrane and the Sertoli cell-spermatid adhesion site known as apical ectoplasmic specialization (apical ES, a testis-specific actin-based anchoring junction) in the rat testis. These bioactive fragments or peptides are produced locally across the seminiferous epithelium through proteolytic cleavage of constituent proteins at the basement membrane and the apical ES. Studies have shown that they are being used to modulate and coordinate cellular functions across the seminiferous epithelium during different stages of the epithelial cycle of spermatogenesis. In this review, we briefly summarize recent findings based on studies using rat testes as a study model regarding the role of these bioactive peptides that serve as a local regulatory network to support spermatogenesis. We also used scRNA-Seq transcriptome datasets in the public domain for OA (obstructive azoospermia) and NAO (non-obstructive azoospermia) human testes versus testes from normal men for analysis in this review. It was shown that there are differential expression of different collagen chains and laminin chains in these testes, suggesting the possibility of a similar local regulatory network in the human testis to support spermatogenesis, and the possible disruption of such network in men is associated with OA and/or NOA.

The novel TERF2::PDGFRB fusion gene enhances tumorigenesis via PDGFRB/STAT5 signalling pathways and sensitivity to TKI in ph-like ALL.

Patients with Philadelphia chromosome-like acute lymphoblastic leukaemia (Ph-like ALL) often face a grim prognosis, with PDGFRB gene fusions being commonly detected in this subgroup. Our study has unveiled a newfound fusion gene, TERF2::PDGFRB, and we have found that patients carrying this fusion gene exhibit sensitivity to dasatinib. Ba/F3 cells harbouring the TERF2::PDGFRB fusion display IL-3-independent cell proliferation through activation of the p-PDGFRB and p-STAT5 signalling pathways. These cells exhibit reduced apoptosis and demonstrate sensitivity to imatinib in vitro. When transfused into mice, Ba/F3 cells with the TERF2::PDGFRB fusion gene induce tumorigenesis and a shortened lifespan in cell-derived graft models, but this outcome can be improved with imatinib treatment. In summary, we have identified the novel TERF2::PDGFRB fusion gene, which exhibits oncogenic potential both in vitro and in vivo, making it a potential therapeutic target for tyrosine kinase inhibitors (TKIs).

NMR Studies of the Interactions between Sialyllactoses and the Polysialytransferase Domain for Polysialylation Inhibition.

It is known that sialyllactose (SL) in mammalians is a major source of sialic acid (Sia), which can further form cytidine monophosphate sialic acid (CMP-Sia), and the final product is polysialic acid (polySia) using polysialyltransferases (polySTs) on the neural cell adhesion molecule (NCAM). This process is called NCAM polysialylation. The overexpression of polysialylation is strongly related to cancer cell migration, invasion, and metastasis. In order to inhibit the overexpression of polysialylation, in this study, SL was selected as an inhibitor to test whether polysialylation could be inhibited. Our results suggest that the interactions between the polysialyltransferase domain (PSTD) in polyST and CMP-Siaand the PSTD and polySia could be inhibited when the 3'-sialyllactose (3'-SL) or 6'-sialyllactose (6'-SL) concentration is about 0.5 mM or 6'-SL and 3 mM, respectively. The results also show that SLs (particularly for 3'-SL) are the ideal inhibitors compared with another two inhibitors, low-molecular-weight heparin (LMWH) and cytidine monophosphate (CMP), because 3'-SL can not only be used to inhibit NCAM polysialylation, but is also one of the best supplements for infant formula and the gut health system.

Salvage Surgery for Initially Unresectable HCC With PVTT Converted by Locoregional Treatment Plus Tyrosine Kinase Inhibitor and Anti-PD-1 Antibody.

BACKGROUND: This study aimed to compare the survival outcomes of patients with initially unresectable hepatocellular carcinoma (HCC) and portal vein tumor thrombus (PVTT) who underwent or did not undergo salvage surgery followed by a triple combination conversion treatment consisted of locoregional treatment (LRT), tyrosine kinase inhibitors (TKIs), and anti-PD-1 antibodies. METHODS: The data from 93 consecutive patients with initially unresectable HCC and PVTT across 4 medical centers were retrospectively reviewed. They were converted successfully by the triple combination treatment and underwent or did not undergo salvage resection. The baseline characteristics, conversion schemes, conversion treatment-related adverse events (CTRAEs), overall survival (OS), and progression-free survival (PFS) of the salvage surgery and non-surgery groups were compared. Multivariate Cox regression analysis was performed to identify independent risk factors for OS and PFS. Additionally, subgroup survival analysis was conducted by stratification of degree of tumor response and type of PVTT. RESULTS: Of the 93 patients, 44 underwent salvage surgery, and 49 did not undergo salvage surgery. The OS and PFS of the salvage surgery and non-surgery groups were not significantly different (P = .370 and .334, respectively). The incidence and severity of CTRAEs of the 2 groups were also comparable. Subgroup analyses revealed that for patients with complete response (CR) or types III-IV PVTT, there was a trend toward better survival in patients who did not undergo salvage surgery. Multivariate analysis showed that baseline α-fetoprotein and best tumor response per mRECIST criteria were independent prognostic factors for OS and PFS. CONCLUSIONS: For patients with initially unresectable HCC and PVTT who were successfully converted by the triple combination therapy, salvage liver resection may not be necessary, especially for the patients with CR or types III-IV PVTT.

Mining the interpretable prognostic features from pathological image of intrahepatic cholangiocarcinoma using multi-modal deep learning.

BACKGROUND: The advances in deep learning-based pathological image analysis have invoked tremendous insights into cancer prognostication. Still, lack of interpretability remains a significant barrier to clinical application. METHODS: We established an integrative prognostic neural network for intrahepatic cholangiocarcinoma (iCCA), towards a comprehensive evaluation of both architectural and fine-grained information from whole-slide images. Then, leveraging on multi-modal data, we conducted extensive interrogative approaches to the models, to extract and visualize the morphological features that most correlated with clinical outcome and underlying molecular alterations. RESULTS: The models were developed and optimized on 373 iCCA patients from our center and demonstrated consistent accuracy and robustness on both internal (n = 213) and external (n = 168) cohorts. The occlusion sensitivity map revealed that the distribution of tertiary lymphoid structures, the geometric traits of the invasive margin, the relative composition of tumor parenchyma and stroma, the extent of necrosis, the presence of the disseminated foci, and the tumor-adjacent micro-vessels were the determining architectural features that impacted on prognosis. Quantifiable morphological vector extracted by CellProfiler demonstrated that tumor nuclei from high-risk patients exhibited significant larger size, more distorted shape, with less prominent nuclear envelope and textural contrast. The multi-omics data (n = 187) further revealed key molecular alterations left morphological imprints that could be attended by the network, including glycolysis, hypoxia, apical junction, mTORC1 signaling, and immune infiltration. CONCLUSIONS: We proposed an interpretable deep-learning framework to gain insights into the biological behavior of iCCA. Most of the significant morphological prognosticators perceived by the network are comprehensible to human minds.