Phenotypic virtual screening holds great promise for discovering therapeutic compounds by directly linking molecular effects to disease-relevant phenotypic outcomes. However, existing computational methods face critical limitations in generalizing across diverse biological contexts and mitigating experimental noise inherent in high-throughput screens. We present PhenotypeCLIP, a robust framework for phenotype-aware drug retrieval that integrates multi-modal representation learning and noise-invariant alignment. The framework leverages pretrained molecular encoders enriched with chemical property and protein target interaction knowledge to model compound-phenotype mechanisms, while explicitly capturing the interplay between molecules and baseline cellular states through latent variable modeling. An information bottleneck further enforces noise-robust representations by prioritizing predictive signals over batch-specific artifacts. Evaluated on cell line-stratified LINCS datasets, PhenotypeCLIP achieves a 3.85-fold improvement in top-1 retrieval recall on unseen cell lines compared to state-of-the-art methods and demonstrates superior few-shot learning performance in mechanism-of-action prediction and drug response prediction. These advancements establish a new paradigm for generalizable and interpretable phenotypic drug discovery, bridging molecular mechanisms with systemic therapeutic effects.
Role-playing agents are becoming increasingly crucial in interactive AI systems, yet current approaches mainly rely on prompt engineering or fine-tuning to condition agents on predefined character profiles. These methods typically encounter challenges in maintaining fine-grained attribute alignment and handling low-knowledge scenarios, often leading to inconsistent role expressions across diverse contexts and when background information is limited. To overcome these limitations, we propose Attribute-Aligned Policy Optimization (AAPO), a reinforcement learning framework that integrates Direct Preference Optimization (DPO) or Proximal Policy Optimization (PPO) to explicitly strengthen the alignment between agent behaviors and role attributes, enabling faithful maintenance of role personality, language style, and background traits. Experiments demonstrate that AAPO significantly outperforms prompt-based and fine-tuning baselines in both LLM automated evaluations and human evaluations. Notably, AAPO exhibits consistent superiority in high-knowledge and low-knowledge settings, ensuring stable attribute adherence even under limited role information.
As large language models (LLMs) are increasingly deployed in real-world systems, the ability to selectively remove specific training data, known as machine unlearning, has become essential for privacy protection and compliance. Existing unlearning methods often fall short in this setting: gradient-based approaches risk destabilizing generation, preference-based methods struggle with incomplete forgetting, and parameter-space editing lacks behavioral control. To address these limitations, we propose Forgetting through Adversarial Disruption and Editing (FADE), a progressive framework that decomposes unlearning into three stages. FADE first introduces distributional noise to weaken memorization, then redirects output probabilities away from forgotten targets, at last FADE recovers performance on unaffected data. Experiments on the MUSE benchmark demonstrate that FADE achieves strong forgetting and privacy protection without compromising utility, outperforming prior methods. These results highlight the importance of staged, behavior-level interventions for reliable and controllable unlearning in LLMs.
To overcome limitations in existing compound representation methods, we propose CFC-CPI, a Cross-scale Feature Fusion approach for compound-protein interaction prediction. CFC-CPI integrates pharmacophore modeling with graph neural networks to represent compounds through atomic-scale (capturing atom types/bonds) and pharmacophore-scale (using BRICS-derived features) graph structures. This dual representation enables comprehensive characterization of compounds’ structural and functional properties. Experiments show CFC-CPI achieves state-of-the-art performance across multiple benchmark datasets, outperforming existing models on various metrics. Visualization analyses confirm the model’s ability to identify actual compound-protein binding sites, addressing interpretability concerns. In a SARS-CoV-2 virtual screening case study, CFC-CPI demonstrated exceptional performance, suggesting strong potential for drug repurposing applications.
In recent years, with the rapid development of LLMs, agent and RAG technologies, multi-round human-computer interaction applications increasingly require efficient retrieval of dialogue content to understand user intent and generate coherent responses. However, existing methods mostly rely on fixed windows or single-round dialogue embeddings, which struggle to dynamically adapt to context changes, suffering from issues such as information truncation, information noise, and no-explicit semantic associations. To address these challenges, this paper proposes a dynamic context-aware embedding method, which leverages LLMs to infer the relevant historical range associated with the current utterance, filters highly related dialogue content through semantic matching, and generates natural language instructions to explicitly model cross-round semantic chains. Experiments demonstrate that this method significantly outperforms traditional embedding approaches on the LoCoMo dataset and exhibits strong robustness across various dialogue scenarios. This work provides new ideas and practical approaches for building intelligent agent systems with semantic understanding and dynamic memory capabilities.
