@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.092_kmeanslocalopt.zh.yaml

@@ -0,0 +1,110 @@
+id: aisb.t3.092_kmeanslocalopt
+name: 具有局部最优保证的改进 K 均值算法
+version: 0.1.0
+one_line: 在保证局部最优性的情况下优化 k 均值，同时保留经典算法的简单性和相同的计算复杂性。
+task_description: 该打包基准测试涵盖了修改后的 K 均值算法 (LO-K-means)，可保证收敛到局部最优解决方案，解决了该领域的基本误解。尽管自 20 世纪 80 年代以来人们普遍相信，最初的 K 均值算法（劳埃德算法）并不总是收敛到局部最优解。该基准测试实现了两种变体：C-LO_K_means（连续局部最优）和 D-LO_K_means（离散局部最优），以及平衡提高精度和计算效率的 Min_D_LO_K_means。所有变体都保持与原始 K 均值算法相同的每次迭代时间复杂度。该实现支持使用一般 Bregman 散度的加权 K 均值聚类，超出平方欧几里得距离。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mean_clustering_loss
+- minimum_clustering_loss
+- average_computation_time_s
+- average_number_of_iterations
+launch_profiles:
+- id: python_eval
+  label: Python 评估
+  description: 运行打包的 Python 实现，比较标准 K-means、C-LO_K_means、D-LO_K_means 和 Min_D_LO_K_means 变体。使用 Iris 数据集评估聚类损失和迭代计数。主要代码锚点是 py/run.py，算法实现在 py/LO_K_means.py 中。
+- id: cpp_eval
+  label: C++ 评估
+  description: 使用主要的 C++ 实现来实现全面的局部最优实验系列。包括 test_multiple_runs.cpp、test_real_world.cpp 和 test_news20.cpp 变体。 cpp/LO_K_means.hpp 中的标头实现，以及 cpp/Setting.hpp 中的设置。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - name: Iris dataset
+    url: embedded
+    path: data/Iris.txt
+    format: custom
+    description: Tab-separated text format with N, D header followed by data points
+  notes:
+  - Iris.txt 与 N 行和 D 维点捆绑在一起
+  - C++ 实现包括额外的真实世界和综合测试场景
+  - 由 make_unique() 函数处理数据集加载以进行加权点处理
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.9'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  notes:
+  - 仅 CPU 基准测试；不需要 GPU 加速，也没有好处
+  - 纯数值计算；手电筒/闪光灯注意不适用
+  - Python路线需要numpy进行矩阵运算和随机生成
+  - C++ 路线使用标准 g++/clang++ 工具链进行编译
+  - 有关完整的依赖项集，请参阅捆绑的 README/requirements.txt
+risk_flags:
+- no_runtime_trust
+- static_audit_only
+risk_notes:
+- 静态代码审核确认所有阶段指标的可执行锚点
+- 在此打包过程中未执行基准测试
+- 指标值标记为临时值，直到运行时验证完成
+- 主要指标 mean_clustering_loss 通过 py/LO_K_means.py 代码支持
+recommended_when: 当您需要具有强大算法清晰度、适度硬件要求和仅 CPU 执行的经典集群任务时，请使用此基准测试。适用于评估局部最优性保证、比较 K 均值变体或研究低维表格数据的收敛行为。非常适合以算法为中心的研究，并且基础设施开销最小。
+not_recommended_when: 如果您需要 GPU 加速、深度学习工作流程、高维嵌入聚类或大规模分布式训练，请勿使用此基准测试。不适合需要神经网络主干或视觉语言组件的基准测试。
+paper:
+  title: Modified K-means Algorithm with Local Optimality Guarantees
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/li25bt.html
+  github_url: https://github.com/lmingyi/LO-K-means
+  authors:
+  - Li, Mingyi (University of Tokyo)
+  - Metel, Michael R. (Huawei Noah's Ark Lab)
+  - Takeda, Akiko (University of Tokyo / RIKEN AIP)
+display:
+  palette_seed: sand-slate-centroid
+  art_style: algorithmic-minimal
+  accent_priority: medium
+  tags:
+  - clustering
+  - optimization
+  - local-optimality
+  - bregman-divergence
+  - cpu-only
+  - icml2025
+image_path: ../image/092_aisb.t3.092_kmeanslocalopt.jpg
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.092_kmeanslocalopt.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.092_kmeanslocalopt.zip
+  archive_type: zip
+  sha256: 86ab62322776504678fb2616e825a32d89f9598e1f0a5900bfeb1231ba8a4ed6
+  size_bytes: 133834

package/AISB/catalog/aisb.t3.093_fedwmsam.yaml

@@ -0,0 +1,134 @@
+id: aisb.t3.093_fedwmsam
+name: FedWMSAM
+version: 0.1.0
+one_line: Federated optimization under non-IID data using weighted momentum and sharpness-aware
+  minimization for stable training and improved generalization.
+task_description: 'This packaged benchmark implements FedWMSAM (Federated Weighted
+  Momentum with Sharpness-Aware Minimization), a federated optimization algorithm
+  built upon the FL-simulator framework. The method addresses the non-IID (non-independently
+  and identically distributed) data challenge inherent to federated learning by combining
+  two complementary mechanisms: weighted momentum to stabilize training under long-tail
+  data distributions, and momentum-driven Sharpness-Aware Minimization (SAM) to improve
+  model generalization on heterogeneous data. The benchmark supports CIFAR-10 and
+  CIFAR-100 with configurable data split rules (Dirichlet and Pathological non-IID
+  partitions) and model architectures (ResNet-18, LeNet). Training proceeds in federated
+  rounds where a central server coordinates client-side local training and aggregates
+  model updates.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- top1_test_accuracy
+- federated_checkpoint
+- training_logs
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run one packaged federated optimization route on CIFAR-10 data with
+    default IID split for rapid validation.
+- id: full_federated_run
+  label: Full Federated Run
+  description: Execute the full FedWMSAM training workflow across all supported split
+    rules (IID, Dirichlet 0.1, Pathological 2) and datasets (CIFAR-10, CIFAR-100).
+dataset_download:
+  primary_method: bundled
+  sources: []
+  notes:
+  - CIFAR-10 and CIFAR-100 are downloaded automatically via PyTorch torchvision datasets.
+  - Non-IID data partitions are generated at runtime using the specified split rule.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    gpu_arch: NVIDIA GPU with CUDA support recommended
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_arch: NVIDIA A100 or equivalent for mixed-precision training
+environment:
+  python: '3.8'
+  cuda: '11.1'
+  pytorch: 1.9.0
+  flash_attn: null
+  key_packages:
+  - torchsampler==0.1.0
+  - scipy
+  notes:
+  - Create environment with: conda create -n fedwmsam python=3.8 && conda activate
+      fedwmsam
+  - Install dependencies via bundled install.sh script
+  - See bundled README/requirements.txt for full dependency set
+risk_flags:
+- non_iid_convergence_risk
+- client_dropout_simulation
+risk_notes:
+- Non-IID data distributions may cause slower convergence compared to IID settings
+- Full federated run with multiple split rules and datasets increases total execution
+  time
+recommended_when: 'Use this benchmark when evaluating federated optimization methods
+  under heterogeneous (non-IID) data distributions, studying the impact of weighted
+  momentum and sharpness-aware minimization on generalization, or comparing against
+  baseline federated optimizers (FedAvg, FedProx, FedSAM) on image classification
+  tasks.
+
+  '
+not_recommended_when: 'Do not use this benchmark for centralized training evaluation,
+  natural language processing tasks, or when single-GPU memory constraints cannot
+  be met (minimum 12GB VRAM recommended for ResNet-18 training).
+
+  '
+paper:
+  title: 'FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware
+    Minimization'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=75JiIa0fU1
+  citation: null
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.093_fedwmsam.zip
+  archive_type: zip
+  local_dir_name: paper-93-FedWMSAM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.093_fedwmsam.zip
+  sha256: 82f848ec812f83c028ed40090155cda3ea04ca716f1e804c11c8d1272169fbf8
+  size_bytes: 91274
+display:
+  palette_seed: terracotta-blue-federated
+  art_style: optimization-dashboard
+  accent_priority: high
+image_path: ../image/093_aisb.t3.093_fedwmsam.jpg
+capability_tags:
+- research_code_optimization
+- federated_learning
+- sharpness_aware_minimization
+- image_classification
+- optimization
+- non_iid_learning
+- weighted_momentum
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.093_fedwmsam.zh.yaml

@@ -0,0 +1,106 @@
+id: aisb.t3.093_fedwmsam
+name: 美联储WMSAM
+version: 0.1.0
+one_line: 使用加权动量和锐度感知最小化在非独立同分布数据下进行联合优化，以实现稳定的训练和改进的泛化。
+task_description: 该打包基准测试实现了 FedWMSAM（具有锐度感知最小化的联合加权动量），这是一种基于 FL 模拟器框架构建的联合优化算法。该方法通过结合两种互补机制来解决联邦学习固有的非独立同分布（non-IID）（非独立同分布）数据挑战：加权动量以稳定长尾数据分布下的训练，动量驱动的锐度感知最小化（SAM）以改进异构数据上的模型泛化。该基准测试支持 CIFAR-10 和 CIFAR-100，具有可配置的数据分割规则（Dirichlet 和 Pathological 非 IID 分区）和模型架构（ResNet-18、LeNet）。训练以联合轮次进行，其中中央服务器协调客户端本地训练并聚合模型更新。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- top1_test_accuracy
+- federated_checkpoint
+- training_logs
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 使用默认 IID 分割对 CIFAR-10 数据运行一种打包的联合优化路线，以进行快速验证。
+- id: full_federated_run
+  label: 完全联合运行
+  description: 跨所有支持的分割规则（IID、Dirichlet 0.1、Pathological 2）和数据集（CIFAR-10、CIFAR-100）执行完整的 FedWMSAM 训练工作流程。
+dataset_download:
+  primary_method: bundled
+  sources: []
+  notes:
+  - CIFAR-10 和 CIFAR-100 通过 PyTorch torchvision 数据集自动下载。
+  - 非 IID 数据分区是在运行时使用指定的分割规则生成的。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    gpu_arch: NVIDIA GPU with CUDA support recommended
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_arch: NVIDIA A100 or equivalent for mixed-precision training
+environment:
+  python: '3.8'
+  cuda: '11.1'
+  pytorch: 1.9.0
+  flash_attn: null
+  key_packages:
+  - torchsampler==0.1.0
+  - scipy
+  notes:
+  - Create environment with: conda create -n fedwmsam python=3.8 && conda 激活 fedwmsam
+  - 通过捆绑的 install.sh 脚本安装依赖项
+  - 有关完整的依赖项集，请参阅捆绑的 README/requirements.txt
+risk_flags:
+- non_iid_convergence_risk
+- client_dropout_simulation
+risk_notes:
+- 与 IID 设置相比，非 IID 数据分布可能会导致收敛速度变慢
+- 具有多个分割规则和数据集的完全联合运行增加了总执行时间
+recommended_when: 在评估异构（非 IID）数据分布下的联邦优化方法、研究加权动量和锐度感知最小化对泛化的影响或与图像分类任务上的基线联邦优化器（FedAvg、FedProx、FedSAM）进行比较时，请使用此基准。
+not_recommended_when: 请勿将此基准用于集中式训练评估、自然语言处理任务或无法满足单 GPU 内存限制时（建议 ResNet-18 训练至少使用 12GB VRAM）。
+paper:
+  title: 'FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware Minimization'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=75JiIa0fU1
+  citation: null
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.093_fedwmsam.zip
+  archive_type: zip
+  local_dir_name: paper-93-FedWMSAM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.093_fedwmsam.zip
+  sha256: 82f848ec812f83c028ed40090155cda3ea04ca716f1e804c11c8d1272169fbf8
+  size_bytes: 91274
+display:
+  palette_seed: terracotta-blue-federated
+  art_style: optimization-dashboard
+  accent_priority: high
+image_path: ../image/093_aisb.t3.093_fedwmsam.jpg
+capability_tags:
+- research_code_optimization
+- federated_learning
+- sharpness_aware_minimization
+- image_classification
+- optimization
+- non_iid_learning
+- weighted_momentum
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.094_boundre.yaml

@@ -0,0 +1,147 @@
+schema_version: 1
+id: aisb.t3.094_boundre
+name: 'BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like
+  One-Class Boundary Optimization'
+version: 0.1.0
+one_line: Drug-likeness boundary prediction via knowledge-graph alignment and EM-like
+  one-class optimization on molecular fingerprints.
+task_description: 'This packaged benchmark covers two-stage drug-likeness boundary
+  prediction. Step 1 performs multi-modal alignment between molecular fingerprints
+  (Morgan fingerprints) and biomedical knowledge-graph embeddings into a unified space.
+  Step 2 trains a deep one-class classifier using EM-like optimization to iteratively
+  refine a hyperspherical boundary around approved drugs, pushing out-of-boundary
+  non-drug compounds while contracting drug-like compounds toward the center. The
+  model outputs a normalized distance score where values greater than 1 indicate compounds
+  outside the drug-like boundary.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- f1
+- auroc
+- avgpr
+- drug_boundary_checkpoint
+launch_profiles:
+- id: single_score
+  label: Single Score
+  description: 'Run drug-likeness scoring on a SMILES input file using pre-trained
+    alignment and boundary models. Outputs normalized distance scores where inference
+    values > 1 indicate compounds outside the drug boundary. Usage: python score.py
+    --input_smiles test.smi --project_dir [project_dir] --device cuda:0
+
+    '
+- id: train_and_eval
+  label: Train + Eval
+  description: 'Run the full two-stage pipeline: (1) train multimodal alignment encoder
+    on knowledge-graph and fingerprint embeddings, then (2) train EM-like boundary
+    optimizer and evaluate F1, AUROC, and AVPR metrics.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://deepscientist.cc/AISB/094_boundre
+    type: archive
+  - type: internal_benchmark
+    path: benchmark/
+    description: Pre-processed drug and compound SMILES files with train/valid/test
+      splits
+  notes:
+  - Benchmark data and splits provided in benchmark/ folder
+  - Demo files available in demo/ folder with test.smi for quick validation
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '10.1'
+  pytorch: 1.8.1
+  flash_attn: null
+  key_packages:
+  - rdkit-pypi==2022.9.5
+  - numpy==1.24.1
+  - pandas==2.0.1
+  notes:
+  - rdkit is required for molecular fingerprint computation and SMILES parsing
+  - See bundled requirements.txt for full dependency set
+risk_flags:
+- em_iteration_convergence
+- knowledge_graph_alignment_quality
+risk_notes:
+- EM-like optimization may require careful tuning of initialization and iteration
+  count for stable boundary convergence
+- Knowledge-graph alignment quality depends on the coherence between ATC similarity
+  matrix and chemical structure
+recommended_when: 'Use this benchmark when you need a biomedical one-class boundary-learning
+  task with chemistry inputs, knowledge-graph features, and iterative boundary refinement.
+  Suitable for evaluating drug-likeness scoring pipelines, multi-modal embedding alignment,
+  and open-world compound classification.
+
+  '
+not_recommended_when: 'Do not use this if you need a lightweight non-biomedical benchmark,
+  binary classification with hard negatives, or a pre-trained model without training
+  customization.
+
+  '
+paper:
+  title: 'BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and
+    EM-like One-Class Boundary Optimization'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/bang25a.html
+  authors:
+  - Dongmin Bang
+  - Inyoung Sung
+  - Yinhua Piao
+  - Sangseon Lee
+  - Sun Kim
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.094_boundre.zip
+  archive_type: zip
+  local_dir_name: paper-94-BounDrE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.094_boundre.zip
+  sha256: 2753a34922d6d0c3d0b021300ab177f667a76ff28dfd357b39fce26571e9bad7
+  size_bytes: 14967140
+display:
+  palette_seed: sage-crimson-boundary
+  art_style: biomedical-geometry
+  accent_priority: high
+image_path: ../image/094_aisb.t3.094_boundre.jpg
+capability_tags:
+- research_code_optimization
+- drug_discovery
+- one_class_classification
+- multimodal_learning
+- biomedical_ai
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.094_boundre.zh.yaml

@@ -0,0 +1,114 @@
+schema_version: 1
+id: aisb.t3.094_boundre
+name: BounDr.E：通过生物医学知识对齐和类 EM 一类边界优化来预测药物相似性
+version: 0.1.0
+one_line: 通过知识图谱对齐和分子指纹的类 EM 一类优化进行药物相似性边界预测。
+task_description: 该打包基准涵盖两阶段药物相似性边界预测。步骤 1 将分子指纹（摩根指纹）和生物医学知识图嵌入到统一空间中进行多模态比对。第 2 步使用类似 EM 的优化来训练深度一类分类器，以迭代地细化已批准药物周围的超球形边界，推动超出边界的非药物化合物，同时将类药物化合物收缩到中心。该模型输出归一化距离分数，其中大于 1 的值表示化合物在类药物边界之外。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- f1
+- auroc
+- avgpr
+- drug_boundary_checkpoint
+launch_profiles:
+- id: single_score
+  label: 单分
+  description: 使用预先训练的对齐和边界模型对 SMILES 输入文件运行药物相似度评分。输出归一化距离分数，其中推断值 > 1 表示化合物在药物边界之外。用法： python score.py --input_smiles test.smi --project_dir [project_dir] --device cuda:0
+- id: train_and_eval
+  label: 火车+评估
+  description: 运行完整的两阶段管道：(1) 在知识图和指纹嵌入上训练多模态对齐编码器，然后 (2) 训练类似 EM 的边界优化器并评估 F1、AUROC 和 AVPR 指标。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://deepscientist.cc/AISB/094_boundre
+    type: archive
+  - type: internal_benchmark
+    path: benchmark/
+    description: Pre-processed drug and compound SMILES files with train/valid/test splits
+  notes:
+  - benchmark/ 文件夹中提供的基准数据和分割
+  - 演示文件可在 demo/ 文件夹中使用 test.smi 进行快速验证
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '10.1'
+  pytorch: 1.8.1
+  flash_attn: null
+  key_packages:
+  - rdkit-pypi==2022.9.5
+  - numpy==1.24.1
+  - pandas==2.0.1
+  notes:
+  - 分子指纹计算和SMILES解析需要rdkit
+  - 请参阅捆绑的 requirements.txt 以获取完整的依赖项集
+risk_flags:
+- em_iteration_convergence
+- knowledge_graph_alignment_quality
+risk_notes:
+- 类似 EM 的优化可能需要仔细调整初始化和迭代计数以实现稳定的边界收敛
+- 知识图比对质量取决于 ATC 相似矩阵和化学结构之间的一致性
+recommended_when: 当您需要具有化学输入、知识图特征和迭代边界细化的生物医学一类边界学习任务时，请使用此基准。适用于评估药物相似性评分管道、多模式嵌入对齐和开放世界化合物分类。
+not_recommended_when: 如果您需要轻量级非生物医学基准、具有硬负例的二元分类或无需训练定制的预训练模型，请勿使用此选项。
+paper:
+  title: 'BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/bang25a.html
+  authors:
+  - Dongmin Bang
+  - Inyoung Sung
+  - Yinhua Piao
+  - Sangseon Lee
+  - Sun Kim
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.094_boundre.zip
+  archive_type: zip
+  local_dir_name: paper-94-BounDrE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.094_boundre.zip
+  sha256: 2753a34922d6d0c3d0b021300ab177f667a76ff28dfd357b39fce26571e9bad7
+  size_bytes: 14967140
+display:
+  palette_seed: sage-crimson-boundary
+  art_style: biomedical-geometry
+  accent_priority: high
+image_path: ../image/094_aisb.t3.094_boundre.jpg
+capability_tags:
+- research_code_optimization
+- drug_discovery
+- one_class_classification
+- multimodal_learning
+- biomedical_ai
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null