Selected work

Representative Papers

TxPert: Leveraging Biochemical Relationships for Out-of-Distribution Transcriptomic Perturbation Prediction

May 20, 2025

This study addresses the challenge of improving out-of-distribution (OOD) generalization for predicting transcriptional responses to genetic perturbations—specifically, unseen single- and double-gene perturbations and novel cell lines. To overcome the limited experimental coverage that constrains existing methods, we introduce, for the first time, multi-source biological knowledge graphs to guide OOD modeling, establishing a rigorous benchmark framework that enforces strict cross-perturbation-type and cross-cell-line generalization. Our method integrates graph neural networks for knowledge graph encoding, multi-relational heterogeneous graph aggregation, OOD-aware training, and an interpretable attention mechanism. Across all three OOD settings, TxPert achieves a mean R² improvement of 12.7% over state-of-the-art baselines. We publicly release both the new benchmark and the model implementation.

1 citationsRead paper

Monte Carlo Estimation for KV Cache Eviction

Oct 05, 2026

This study addresses the evaluation inaccuracies in KV cache eviction caused by the absence of future query information. To overcome this limitation, it proposes LORE-KV, a training-free framework that formulates future-aware eviction as a distribution estimation problem for the first time. Specifically, the method predicts prompt utility via Monte Carlo sampling over short autoregressive continuations, leveraging response-side query states to substitute pseudo-responses. Furthermore, it computes deletion costs through projected leave-one-out attention combined with trajectory-weighted ensembling, enabling efficient token selection under a fixed memory budget. Experimental results on the LongBench and RULER benchmarks demonstrate that LORE-KV significantly improves accuracy while introducing minimal inference overhead, effectively optimizing memory utilization for large language models.

0 citationsRead paper

AquaGen: Scaling generative models to molecular dynamics precision on thousands of atoms

Jul 03, 2026

This work addresses the high computational cost of conventional molecular dynamics simulations and the limited accuracy of existing generative models, which often rely on coarse-grained representations or implicit solvent approximations that hinder precise free energy predictions. The authors propose the first deep generative model capable of operating with full-atom resolution, explicit solvent, and periodic boundary conditions, enabling efficient direct sampling of molecular configurations from the Boltzmann distribution. By integrating physical constraints into the generative framework, the method yields interpretable and refinable ensemble predictions while providing well-calibrated uncertainty estimates. In absolute hydration free energy prediction tasks, the model achieves accuracy comparable to GPU-accelerated molecular dynamics but with a 4–10× speedup, demonstrating strong scalability.

0 citationsRead paper

Knowledge Graphs and Reasoning LLMs for Finding Simple Yet Effective Transcriptomic Perturbation Predictors

Jun 07, 2026

Predicting the transcriptional effects of out-of-distribution gene knockout perturbations remains a significant challenge. This work proposes a novel approach that integrates biological knowledge graphs with reasoning-capable large language models. The method first constructs a k-nearest neighbors (k-NN) base predictor grounded in a knowledge graph, which substantially outperforms existing methods in out-of-distribution settings. Subsequently, reinforcement learning is employed to optimize the language model’s strategy for selecting relevant neighbors, thereby indirectly enhancing downstream differential expression prediction performance. This framework provides the first empirical validation of knowledge graph–guided k-NN for perturbation response prediction and achieves state-of-the-art results through a reinforcement learning–driven inference mechanism, markedly improving generalization to tasks not directly seen during training.

0 citationsRead paper

Effective Biological Representation Learning by Masking Gene Expression

May 29, 2026

Existing transcriptomic foundation models often underperform on RNA-seq data due to technical noise and batch effects, sometimes even falling short of simple linear baselines. To address this limitation, this work proposes TxFM, a masked autoencoding self-supervised model specifically designed for RNA-seq count data. By integrating inductive representation learning with a high-quality, diverse training corpus—DiverseRNA-1.4M, comprising only 1.4 million samples—TxFM substantially enhances gene representation capabilities. Systematic ablation studies validate the importance of key architectural choices, and extensive evaluations demonstrate that TxFM outperforms current foundation models trained on datasets over 100 times larger across multiple downstream tasks. These results underscore the efficacy and superiority of carefully crafted self-supervised learning strategies in transcriptomics.

0 citationsRead paper

Recent publications

Latest Papers

Monte Carlo Estimation for KV Cache Eviction

Oct 05, 2026

This study addresses the evaluation inaccuracies in KV cache eviction caused by the absence of future query information. To overcome this limitation, it proposes LORE-KV, a training-free framework that formulates future-aware eviction as a distribution estimation problem for the first time. Specifically, the method predicts prompt utility via Monte Carlo sampling over short autoregressive continuations, leveraging response-side query states to substitute pseudo-responses. Furthermore, it computes deletion costs through projected leave-one-out attention combined with trajectory-weighted ensembling, enabling efficient token selection under a fixed memory budget. Experimental results on the LongBench and RULER benchmarks demonstrate that LORE-KV significantly improves accuracy while introducing minimal inference overhead, effectively optimizing memory utilization for large language models.

0 citationsRead paper

AquaGen: Scaling generative models to molecular dynamics precision on thousands of atoms

Jul 03, 2026

This work addresses the high computational cost of conventional molecular dynamics simulations and the limited accuracy of existing generative models, which often rely on coarse-grained representations or implicit solvent approximations that hinder precise free energy predictions. The authors propose the first deep generative model capable of operating with full-atom resolution, explicit solvent, and periodic boundary conditions, enabling efficient direct sampling of molecular configurations from the Boltzmann distribution. By integrating physical constraints into the generative framework, the method yields interpretable and refinable ensemble predictions while providing well-calibrated uncertainty estimates. In absolute hydration free energy prediction tasks, the model achieves accuracy comparable to GPU-accelerated molecular dynamics but with a 4–10× speedup, demonstrating strong scalability.

0 citationsRead paper

Knowledge Graphs and Reasoning LLMs for Finding Simple Yet Effective Transcriptomic Perturbation Predictors

Jun 07, 2026

Predicting the transcriptional effects of out-of-distribution gene knockout perturbations remains a significant challenge. This work proposes a novel approach that integrates biological knowledge graphs with reasoning-capable large language models. The method first constructs a k-nearest neighbors (k-NN) base predictor grounded in a knowledge graph, which substantially outperforms existing methods in out-of-distribution settings. Subsequently, reinforcement learning is employed to optimize the language model’s strategy for selecting relevant neighbors, thereby indirectly enhancing downstream differential expression prediction performance. This framework provides the first empirical validation of knowledge graph–guided k-NN for perturbation response prediction and achieves state-of-the-art results through a reinforcement learning–driven inference mechanism, markedly improving generalization to tasks not directly seen during training.

0 citationsRead paper

Effective Biological Representation Learning by Masking Gene Expression

May 29, 2026

Existing transcriptomic foundation models often underperform on RNA-seq data due to technical noise and batch effects, sometimes even falling short of simple linear baselines. To address this limitation, this work proposes TxFM, a masked autoencoding self-supervised model specifically designed for RNA-seq count data. By integrating inductive representation learning with a high-quality, diverse training corpus—DiverseRNA-1.4M, comprising only 1.4 million samples—TxFM substantially enhances gene representation capabilities. Systematic ablation studies validate the importance of key architectural choices, and extensive evaluations demonstrate that TxFM outperforms current foundation models trained on datasets over 100 times larger across multiple downstream tasks. These results underscore the efficacy and superiority of carefully crafted self-supervised learning strategies in transcriptomics.

0 citationsRead paper

SeedER: Seed-and-Expand Retrieval from Knowledge Graphs

May 22, 2026

This work addresses the efficiency and scalability challenges posed by the irregular structure of knowledge graphs in multi-hop compositional query answering. The authors propose SeedER, a novel framework that uniquely integrates a seed-expansion mechanism with reinforcement learning. It first generates compact sets of core seed entities through lightweight dense and sparse retrieval, then iteratively expands these seeds via a graph-aware policy, decomposing global reasoning into reusable local decisions at low computational cost. By maintaining concise candidate sets while substantially improving recall, SeedER outperforms strong baselines and functions as an efficient single-stage retriever. The approach also offers theoretical advantages in compositional generalization and submodular optimization under graph constraints.

0 citationsRead paper