Explore the vast range of titles available.

Due to randomness factors in the machine learning model construction process, reproducibility is compromised. This study investigates the impact of randomness on model stability and evaluates techniques for reducing this impact using the widely adopted shallow neural network model as a testbed. Randomness in this neural network model arises from three events: randomly initializing model parameters, randomly selecting a validation subset, and randomly sampling batches for parameter updates. Among these, batch randomness exerts a much weaker impact than the other two factors. In this study, the model training is stopped when the validation performance fails to improve or when a preset threshold for loss or epoch number is met. The final model stability is considerably better when using threshold criteria than when using validation criterion, as the former avoids the randomness associated with selecting a validation subset. Sensitivity experiments show that scaling the model’s initial parameters (i.e., weights) to 0.1 times their original values can mitigate the impact of initialization randomness, thereby markedly improving model stability while also substantially enhancing predictive skill. Furthermore, weight decay and multi-model ensembles, which are two commonly used techniques, can also markedly enhance model stability. From the perspective of this case study, the compression of model initial parameters yields better improvements in stability compared to weight decay, and unlike multi-model ensemble methods that entail substantial increases in computational cost, it serves as a preferable technique for improving model stability.

This study investigates ice crystal sedimentation calculation errors arising from three-moment bulk cloud scheme. Both offline tests and one-dimensional cloud model simulations indicate that sedimentation calculation errors are most pronounced at both the cloud bottom and cloud top. At the cloud bottom, the error stems from how the bulk method treats ice crystal sedimentation. Specifically, the method uses three weighted fall velocities (corresponding to the three moments) to represent instantaneous fluxes through a fixed altitude, which inherently assumes that falling ice crystals can only affect the adjacent model layer below. This assumption artificially constrains the falling distance of larger ice crystals. At the cloud top, the differences among these three weighted fall velocities can give rise to physical inconsistencies. This issue is handled by artificial adjustment, which leads to a spurious narrow size distribution shape of ice crystals, especially under model configurations with coarse temporal resolution (large dT) and fine vertical resolution (small dH). If only the sedimentation process is considered, the above calculation errors can be effectively minimized by lowering the dT/dH ratio.

Forecasting July precipitation using prophase winter sea surface temperature through a nonlinear machine learning model remains challenging. Given the scarcity of observed samples and more attention should be paid to anomalous precipitation events, the shallow neural network (NN) and several improving techniques are employed to establish the statistical forecasting system. To enhance the stability of predicted precipitation, the final output precipitation is an ensemble of multiple NN models with optimal initial seeds. The precipitation data from anomalous years are amplified to focus on anomalous events rather than normal events. Some artificial samples are created based on the relevant background theory to mitigate the problem of insufficient sample size for model training. Sensitivity experiments indicate that the above techniques could improve the stability and interpretability of the forecasting system. Rolling forecasts further indicate that the forecasting system is robust and half of the anomalous events can be successfully predicted. These improving techniques used in this study can be applied not only to the precipitation over the middle-lower reaches of the Yangtze River but also to other climate events.

Statistical climate forecast systems typically do not use preceding global gridded sea surface temperature (SST) data directly; instead, they extract a single predictor (e.g., the Niño3.4 index) or multiple predictors (e.g., time series of several SST spatial modes). In this study, four different SST predictor extracting methods (one single-predictor method and three multiple-predictor methods) are comparatively analyzed within the same climate forecast platform incorporating either the linear regression (LR) model or the neural network (NN) forecast model. Rolling forecast experiments with the LR model show that, compared to a single strong SST predictor, only multiple predictors with more high-quality information (high signal-to-noise ratio) could improve the forecast skill. Sensitivity experiments also show that the influence of multiple-predictor extracting methods on forecast skill from the NN model is much weaker than that from the LR model. Moreover, whether or not multiple SST predictors are orthogonal might also affect the forecast skill. The above analyses provide a reference for establishing statistical climate forecast system based on preceding SST data.

The development of CAR-T-cell immunotherapy has notably elevated the efficacy of treating multiple myeloma. Currently, a variety of targets, including BCMA, CS1, CD38, FcRH5, and GPRC5D, are being investigated. Despite these significant advancements, challenges such as antigen escape, limited persistence of CAR-T cells, and the intricate nature of the tumor microenvironment persist, leading to relapses following treatment. We report the case of a patient with recurrent and refractory multiple myeloma (RRMM) who developed a substantial extramedullary plasmacytoma in the muscles of the lower limb following multiple rounds of radiotherapy and chemotherapy. The patient underwent CAR-T-cell immunotherapy targeting BCMA and CS1; however, the tumor progressed despite treatment. Surgical resection of the extramedullary plasmacytoma was subsequently performed. Upon comparison of the tumor tissue with the adjacent tissue, increased expression of MYBL2 was noted in the tumor tissue, potentially contributing to the lack of improvement in extramedullary relapse after dual-targeted CAR-T cell therapy. In patients with recurrent and refractory multiple myeloma who underwent multiple cycles of chemotherapy and radiotherapy, dual-targeted CAR-T cell therapy aimed at BCMA and CS1 failed to effectively manage extramedullary relapse. Elevated expression of MYBL2 in multiple myeloma correlates with a poorer prognosis.

The model internal year-to-year variability (hereafter, internal variability) is a significant source of uncertainty when estimating anthropogenic aerosol effective radiative forcing (ERF). In this study, we investigate the impact of internal variability using large ensemble simulations (600 years in total) with the same climate model under prescribed anthropogenic aerosol forcings. A comparison of the magnitudes (i.e., standard deviation, Std) of these influences confirms that internal variability has negligible impacts on the instantaneous radiative forcing (RF) diagnosed by double radiation calls but has considerable impacts on estimating ERF through rapid adjustments (ADJ). Approximately half of the model grids exhibit a strong internal variability influence on ERF (Std > 5 W m−2). These strong internal variabilities lead to a 50% probability that the 30-year linear change can reach 2 W m−2 and the 10-year linear change can reach 4 W m−2. A 50-year simulation can provide a relatively stable annual mean map of ERF (ERF = ADJ + RF), but it fails for ADJ. The statistically significant areas in the annual mean maps of both ERF and ADJ from a 10-year simulation exhibit instability with evident chaotic features. The insights derived from our analysis aid in assessing the stability of modeled ERF and contribute to the design of comparative experiments.

This study investigates the possible mechanisms related to the wide range of relative humidity in cirrus clouds (RHi). Under the closed adiabatic assumption, the impacts of vertical motion and ice crystal deposition/sublimation on RHi are investigated through in situ observations and parcel model simulations. Vertical motion is an active external force that changes the RHi, and ice crystal deposition/sublimation plays a role in mitigating the change in the RHi. They are the two most important mechanisms involved in controlling the RHi fluctuation during cirrus evolution and could well explain the wide range of RHi in wave-related cirrus clouds. Furthermore, a comparison of statistical cloud characteristics from both observations and simulations shows that a very low value (e.g., 0.001) for the water vapor ice deposition coefficient is highly unlikely.

The multiple regression method is still an important tool for establishing precipitation forecast models with a lead time of one season. This study developed a flexible statistical forecast model for July precipitation over the middle-lower reaches of the Yangtze River (MLYR) based on the prophase winter sea surface temperature (SST). According to the characteristics of observed samples and related theoretical knowledge, some special treatments (i.e., more flexible and better-targeted methods) were introduced in the forecast model. These special treatments include a flexible MLYR domain definition, the extraction of indicative signals from the SST field, artificial samples, and the amplification of abnormal precipitation. Rolling forecast experiments show that the linear correlation between prediction and observation is around 0.5, more than half of the abnormal precipitation years can be successfully predicted, and there is no contradictory prediction of the abnormal years. These results indicate that the flexible statistical forecast model is valuable in real-life applications. Furthermore, sensitivity experiments show that forecast skills without these special treatments are obviously decreased. This suggests that forecast models can benefit from using statistical methods in a more flexible and better-targeted way.

Tumor infiltrating T cells (TIL) are key players in the anti-tumor immune response. However, chronic exposure to tumor-derived antigens drives the differentiation into ‘exhausted’ TILs. Whether intratumoral dendritic cells (DC) can mitigate TILs exhaustion and maintain function is unclear. Here, we develop a bispecific DC-T cell engager (BiDT), consisting of an anti-TIM3-IFN fusion protein, and demonstrate that, in preclinical mouse tumor models, this engager simultaneously targets TIM3 on exhausted TILs and activates DCs via the IFNAR receptor. Mechanistically, BiDT reactivates exhausted TIM3 + TILs by preventing apoptosis through increased Bcl-2 expression and enhances DC function to reactivate T cells via IL-2 signalling and co-stimulatory CD80/86-CD28 interactions within the tumor microenvironment. Finally, to mitigate IFNα-induced toxicity, we engineer a Pro-BiDT engager featuring a pro-IFNα and report potent antitumor activity with reduced systemic toxicity. Thus, by bridging DC-T cells together, BiDT treatment enhances the critical communication pathways and cellular circuits necessary for effective anti-tumor immunity. Tumor-infiltrating T cells (TIL) are key mediators of the anti-tumor immune responses. However, chronic exposure to tumor antigens leads to the acquisition of an exhausted phenotype. Here, the authors present a bispecific T-cell engager that combines an anti-TIM-3 antibody with an IFN-α fusion protein (BiDT) and facilitates T cell-dendritic cells (DC) interactions. In preclinical mouse models, the enhanced T cell-DC crosstalk, dependent on CD28 co-stimulation and IL-2 signaling, boosts anti-tumor immunity and the generation of memory T cell responses.

This study mainly introduces the development of the Flexible Global Ocean-Atmosphere-Land System Model： Grid-point Version 2 （FGOALS-g2） and the preliminary evaluations of its performances based on re- sults from the pre-industrial control run and four members of historical runs according to the fifth phase of the Coupled Model Intercomparison Project （CMIP5） experiment design. The results suggest that many obvi- ous improvements have been achieved by the FGOALS-g2 compared with the previous version, FGOALS-gl, including its climatological mean states, climate variability, and 20th century surface temperature evolution. For example, FGOALS-g2 better simulates the frequency of tropical land precipitation, East Asian Monsoon precipitation and its seasonal cycle, MJO and ENSO, which are closely related to the updated cumulus parameterization scheme, as well as the alleviation of uncertainties in some key parameters in shallow and deep convection schemes, cloud fraction, cloud macro/microphysical processes and the boundary layer scheme in its atmospheric model. The annual cycle of sea surface temperature along the equator in the Pacific is significantly improved in the new version. The sea ice salinity simulation is one of the unique characteristics of FGOALS-g2, although it is somehow inconsistent with empirical observations in the Antarctic.