报告题目：RNA-targeted Drug Discovery: RNA-Small Molecule Bindin…

报 告 人：吴敏

时 间：2025年6月16日（周一）上午10:00

地 点：国交2号楼201会议室

简 介：The drug discovery pipeline is a complex, multi-stage process encompassing target identification, hit discovery, lead optimization, and clinical development. While traditional drugs mainly target proteins, RNA-targeted drug discovery focuses on developing therapeutics that directly interact with RNA molecules. In this talk, we use the hit discovery stage as an example and introduce deep learning techniques designed for RNA–small molecule binding affinity and binding site prediction. DeepRSMA is a cross-attention-based deep learning model for predicting RNA–small molecule binding affinity. It employs nucleotide- and atomic-level feature extraction modules for RNA and small molecules, respectively, and uses a transformer-based cross-fusion module to model their interactions. The final prediction integrates features from both the extraction and fusion modules. MVRBind is a multi-view, multi-scale graph convolutional network for predicting RNA–small molecule binding sites. It capture...

报告题目：基于正则化重心映射的scATAC-seq数据整合方法

报 告 人：陈盛泉

时 间：2025年6月18日（周三）上午9:30-11:00

地 点：腾讯会议号：335-427-641

简 介：单细胞染色质开放性测序（scATAC-seq）为基因调控机制研究提供了重要视角，而多源scATAC-seq数据分析面临着批次效应等关键挑战，亟需建立高效的数据整合方法。针对现有方法普遍存在的两大局限：过度校正导致生物异质性丢失、忽略原始特征空间的校正导致下游分析可靠性受影响，本报告将介绍一种基于正则化重心映射的深度学习模型Fountain。Fountain引入批次内数据的几何信息作为批次间重心映射的正则化约束，在保持生物异质性的前提下实现跨批次精准对齐，且训练后的模型无需重新训练即可整合新批次数据，实现数据的在线整合。此外，Fountain可提供批次校正后的原始维度特征，显著提升细胞异质性刻画能力，并在细胞类型特异性遗传力分析等下游任务中展现出更强的生物学解释力。

报告题目：Catalyzing Quantum Communication: From Embezzling to N…

报 告 人：肖运龙

时 间：2025年6月11日（周三）下午16:00-17:00

地 点：新文科楼四楼402会议室

简 介：We report three recent advances in quantum communication: first, we derive fundamental limits via dynamical quantum resource theory; second, we introduce a catalyst-driven framework that amplifies teleportation and channel capacities—achieving arbitrarily high fidelity with negligible disturbance to the catalyst; and third, we use transformer-based language models (BERT) to enhance superdense coding for large-text transmission over noisy channels without conventional error correction. Together, these results inaugurate a hybrid quantum–classical paradigm that unites catalytic protocols and NLP techniques to enable scalable, high-fidelity quantum networks.