arXiv subjects

Yingze Wang

Publications and source records attributed to Yingze Wang.

12 recordsLinked to original sources

arXiv2026

A Distributed Multi-UGV Exploration Framework With Loop-Aware Planning and Descriptor-Aided Localization in Resource-Limited Environments

Robust and efficient cooperative exploration with multiple unmanned ground vehicles (UGVs) in unknown, GPSdenied, and bandwidth-limited environments without prior maps remains challenging, as localization drift degrades map consistency and induces redundant coverage. This paper presents a fully distributed exploration framework that couples descriptoraided inter-UGV loop closure with loop-aware hierarchical planning while enabling autonomous localization and exploration. We develop a lightweight LiDAR global descriptor with range-image prealignment to enable robust cross-UGV place recognition under large yaw and lateral variations, and use verified loop closures to maintain globally consistent trajectories and a sparse topological representation. We further introduce an uncertainty-aware crossUGV loop-closure selection module that scores candidate loop closures under pose uncertainty and retains high-utility loop closures as planning anchors for global task allocation and local route refinement. Simulations and real-UGV experiments show that the loop-closure module achieves AR@1/AR@1% of 89.9%/95.5%, distributed optimization reduces absolute trajectory error, the system substantially reduces two-way communication volume, and the overall framework reduces exploration time and travel distance by 15% and 14%, respectively, compared with an mTSP baseline.

cs.RO↗

arXiv2026

SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration

We show that large language model (LLMs) can be transformed via supervised fine-tuning (SFT) of engineered prompts into SmileyLlama for exploring the chemical space of drug molecules. We benchmark SmileyLlama against pre-trained LLMs and chemical language models (CLM) trained from scratch for generating valid and novel drug-like molecules, and use direct preference optimization (DPO) to both improve SmileyLlama's adherence to a prompt and as part of the iMiner reinforcement learning framework to predict molecules with optimized 3D conformations and high binding affinity to drug targets. By training an LLM to speak directly as a CLM, while retaining most of its natural language capabilities, we show that we can reliably generate molecules with user-specified properties rather than acting only as a chatbot with knowledge of chemistry or as a virtual assistant. While SmileyLlama is geared toward drug discovery, the SFT/DPO/LLM framework can be extended to other chemical, biological, and materials applications.

physics.chem-ph↗

arXiv2026

Accelerating Scientific Discovery with Autonomous Goal-evolving Agents

There has been unprecedented interest in developing agents that expand the boundary of scientific discovery, primarily by optimizing quantitative objective functions specified by scientists. However, for grand challenges in science, these objectives may only be imperfect proxies. We argue that automating objective function design is a central, yet unmet need for scientific discovery agents. In this work, we introduce the Scientific Autonomous Goal-evolving Agent (SAGA) to address this challenge. SAGA employs a bi-level architecture in which an outer loop of LLM agents analyzes optimization outcomes, proposes new objectives, and converts them into computable scoring functions, while an inner loop performs solution optimization under the current objectives. This bi-level design enables systematic exploration of the space of objectives and their trade-offs, rather than treating them as fixed inputs. We demonstrate the framework through a wide range of design applications, including antibiotics, nanobodies, functional DNA sequences, inorganic materials, and chemical processes. Notably, our experimental validation identifies a structurally novel hit with promising potency and safety profiles for E. coli in the antibiotic design task, and three de novo PD-L1 binders in the nanobody design task. These results suggest that automating objective formulation can substantially improve the effectiveness of scientific discovery agents.

cs.AI↗

arXiv2026

UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers

Despite advances, video diffusion transformers still struggle to generalize beyond their training length, a challenge we term video length extrapolation. We identify two failure modes: model-specific periodic content repetition and a universal quality degradation. Prior works attempt to solve repetition via positional encodings, overlooking quality degradation and achieving only limited extrapolation. In this paper, we revisit this challenge from a more fundamental view: attention maps, which directly govern how context influences outputs. We identify that both failure modes arise from a unified cause: attention dispersion, where tokens beyond the training window dilute learned attention patterns. This leads to quality degradation and repetition emerges as a special case when this dispersion becomes structured into periodic attention patterns, induced by harmonic properties of positional encodings. Building on this insight, we propose UltraViCo, a training-free, plug-and-play method that suppresses attention for tokens beyond the training window via a constant decay factor. By jointly addressing both failure modes, we outperform a broad set of baselines largely across models and extrapolation ratios, pushing the extrapolation limit from 2x to 4x. Remarkably, it improves Dynamic Degree and Imaging Quality by 233% and 40.5% over the previous best method at 4x extrapolation. Furthermore, our method generalizes seamlessly to downstream tasks such as controllable video synthesis and editing.

cs.CV↗

arXiv2026

Leak Proof PDBBind: A Reorganized Dataset of Protein-Ligand Complexes for More Generalizable Binding Affinity Prediction

The majority of machine learning scoring functions used in drug discovery for predicting protein-ligand binding poses and affinities have been trained on the PDBBind dataset. However, it is unclear whether these new scoring functions are actually an improvement over traditional models since often the training and test sets are cross-contaminated with proteins and ligands with high similarity, and hence they may not perform comparably well in binding prediction of unrelated protein-ligand complexes. In this work we have carefully prepared a new split of the PDBBind data set to control for data leakage, defined as proteins and ligands with high sequence and structural similarity. The resulting leak-proof (LP)-PDBBind data is used to retrain four popular SFs: AutoDock Vina, Random Forest (RF)-Score, InteractionGraphNet (IGN), and DeepDTA, to better test their capabilities when applied to new protein-ligand complexes. In particular we have formulated a new independent data set, BDB2020+, by matching high quality binding free energies from BindingDB with co-crystalized ligand-protein complexes from the PDB that have been deposited since 2020. Based on all the benchmark results, the retrained models using LP-PDBBind consistently perform better, with IGN especially being recommended for scoring and ranking applications for new protein-ligand systems.

physics.bio-ph↗

arXiv2025

SynLlama: Generating Synthesizable Molecules and Their Analogs with Large Language Models

Generative machine learning models for exploring chemical space have shown immense promise, but many molecules they generate are too difficult to synthesize, making them impractical for further investigation or development. In this work, we present a novel approach by fine-tuning Meta's Llama3 Large Language Models (LLMs) to create SynLlama, which generates full synthetic pathways made of commonly accessible building blocks and robust organic reaction templates. SynLlama explores a large synthesizable space using significantly less data, and offers strong performance in both forward and bottom-up synthesis planning compared to other state-of-the-art methods. We find that SynLlama, even without training on external building blocks, can effectively generalize to unseen yet purchasable building blocks, meaning that its reconstruction capabilities extend to a broader synthesizable chemical space than the training data. We also demonstrate the use of SynLlama in a pharmaceutical context for synthesis planning of analog molecules and hit expansion leads for proposed inhibitors of target proteins, offering medicinal chemists a valuable tool for discovery.

cs.LG↗

arXiv2025

Improved Treatment of 1-4 interactions in Force Fields for Molecular Dynamics Simulations

Traditional force fields commonly use a combination of bonded torsional terms and empirically scaled non-bonded interactions to capture 1-4 energies and forces of atoms separated by three bonds in a molecule. While this approach can yield accurate torsional energy barriers, it often leads to inaccurate forces and erroneous geometries, and creates an interdependence between dihedral terms and non-bonded interactions, complicating parameterization and reducing transferability. In this paper, we demonstrate that 1-4 interactions can be accurately modeled using only bonded coupling terms, eliminating the need for arbitrarily scaled non-bonded interactions altogether. Furthermore by leveraging the automated parameterization capabilities of the Q-Force toolkit, we efficiently determine the necessary coupling terms without the need for manual adjustment. Our approach is first validated on a range of small molecule systems, encompassing both flexible and rigid structures, and shows a significant improvement in force field accuracy, obtaining sub-kcal/mol mean absolute error for every molecule tested. We further extend the bonded-only model for 1-4 interactions to Amber ff14sb, CHARMM36, and OPLS-AA force fields to reproduce ab initio gas and implicit solvent $ϕ,ψ$ surfaces of alanine dipeptide.

physics.chem-ph↗

arXiv2025

A Workflow to Create a High-Quality Protein-Ligand Binding Dataset for Training, Validation, and Prediction Tasks

Development of scoring functions (SFs) used to predict protein-ligand binding energies requires high-quality 3D structures and binding assay data for training and testing their parameters. In this work, we show that one of the widely-used datasets, PDBbind, suffers from several common structural artifacts of both proteins and ligands, which may compromise the accuracy, reliability, and generalizability of the resulting SFs. Therefore, we have developed a series of algorithms organized in a semi-automated workflow, HiQBind-WF, that curates non-covalent protein-ligand datasets to fix these problems. We also used this workflow to create an independent data set, HiQBind, by matching binding free energies from various sources including BioLiP, Binding MOAD and BindingDB with co-crystalized ligand-protein complexes from the PDB. The resulting HiQBind workflow and dataset are designed to ensure reproducibility and to minimize human intervention, while also being open-source to foster transparency in the improvements made to this important resource for the biology and drug discovery communities.

physics.bio-ph↗

arXiv2024

Mining for Potent Inhibitors through Artificial Intelligence and Physics: A Unified Methodology for Ligand Based and Structure Based Drug Design

The viability of a new drug molecule is a time and resource intensive task that makes computer-aided assessments a vital approach to rapid drug discovery. Here we develop a machine learning algorithm, iMiner, that generates novel inhibitor molecules for target proteins by combining deep reinforcement learning with real-time 3D molecular docking using AutoDock Vina, thereby simultaneously creating chemical novelty while constraining molecules for shape and molecular compatibility with target active sites. Moreover, through the use of various types of reward functions, we can generate new molecules that are chemically similar to a target ligand, which can be grown from known protein bound fragments, as well as to create molecules that enforce interactions with target residues in the protein active site. The iMiner algorithm is embedded in a composite workflow that filters out Pan-assay interference compounds, Lipinski rule violations, and poor synthetic accessibility, with options for cross-validation against other docking scoring functions and automation of a molecular dynamics simulation to measure pose stability. Because our approach only relies on the structure of the target protein, iMiner can be easily adapted for future development of other inhibitors or small molecule therapeutics of any target protein.

q-bio.BM↗

arXiv2023

UAV-Aided Lifelong Learning for AoI and Energy Optimization in Non-Stationary IoT Networks

In this paper, a novel joint energy and age of information (AoI) optimization framework for IoT devices in a non-stationary environment is presented. In particular, IoT devices that are distributed in the real-world are required to efficiently utilize their computing resources so as to balance the freshness of their data and their energy consumption. To optimize the performance of IoT devices in such a dynamic setting, a novel lifelong reinforcement learning (RL) solution that enables IoT devices to continuously adapt their policies to each newly encountered environment is proposed. Given that IoT devices have limited energy and computing resources, an unmanned aerial vehicle (UAV) is leveraged to visit the IoT devices and update the policy of each device sequentially. As such, the UAV is exploited as a mobile learning agent that can learn a shared knowledge base with a feature base in its training phase, and feature sets of a zero-shot learning method in its testing phase, to generalize between the environments. To optimize the trajectory and flying velocity of the UAV, an actor-critic network is leveraged so as to minimize the UAV energy consumption. Simulation results show that the proposed lifelong RL solution can outperform the state-of-art benchmarks by enhancing the balanced cost of IoT devices by $8.3\%$ when incorporating warm-start policies for unseen environments. In addition, our solution achieves up to $49.38\%$ reduction in terms of energy consumption by the UAV in comparison to the random flying strategy.

cs.NI↗

arXiv2023

DeePMD-kit v2: A software package for Deep Potential models

DeePMD-kit is a powerful open-source software package that facilitates molecular dynamics simulations using machine learning potentials (MLP) known as Deep Potential (DP) models. This package, which was released in 2017, has been widely used in the fields of physics, chemistry, biology, and material science for studying atomistic systems. The current version of DeePMD-kit offers numerous advanced features such as DeepPot-SE, attention-based and hybrid descriptors, the ability to fit tensile properties, type embedding, model deviation, Deep Potential - Range Correction (DPRc), Deep Potential Long Range (DPLR), GPU support for customized operators, model compression, non-von Neumann molecular dynamics (NVNMD), and improved usability, including documentation, compiled binary packages, graphical user interfaces (GUI), and application programming interfaces (API). This article presents an overview of the current major version of the DeePMD-kit package, highlighting its features and technical details. Additionally, the article benchmarks the accuracy and efficiency of different models and discusses ongoing developments.

physics.chem-ph↗

arXiv2022

Improving Molecular Pretraining with Complementary Featurizations

Molecular pretraining, which learns molecular representations over massive unlabeled data, has become a prominent paradigm to solve a variety of tasks in computational chemistry and drug discovery. Recently, prosperous progress has been made in molecular pretraining with different molecular featurizations, including 1D SMILES strings, 2D graphs, and 3D geometries. However, the role of molecular featurizations with their corresponding neural architectures in molecular pretraining remains largely unexamined. In this paper, through two case studies -- chirality classification and aromatic ring counting -- we first demonstrate that different featurization techniques convey chemical information differently. In light of this observation, we propose a simple and effective MOlecular pretraining framework with COmplementary featurizations (MOCO). MOCO comprehensively leverages multiple featurizations that complement each other and outperforms existing state-of-the-art models that solely relies on one or two featurizations on a wide range of molecular property prediction tasks.

cs.LG↗