@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.065_proteinbinding.yaml

@@ -0,0 +1,234 @@
+schema_version: 1
+id: aisb.t3.065_proteinbinding
+name: StaB-ddG — Predicting Mutational Effects on Protein Binding from Folding Energy
+version: 0.1.0
+one_line: Transfer-learning benchmark for protein-protein binding ΔΔG prediction via
+  stability-to-binding parameterization using ProteinMPNN and two-stage fine-tuning.
+task_description: "This packaged benchmark covers the full StaB-ddG workflow for predicting\
+  \ mutational effects on protein-protein binding free energies (ΔΔG, kcal/mol). The\
+  \ core idea is the thermodynamic identity ΔG_bind(A:B) = ΔG_fold(A:B) − ΔG_fold(A)\
+  \ − ΔG_fold(B), which allows binding energy prediction to be reduced to folding\
+  \ energy prediction and enables transfer from copious stability data to scarce binding\
+  \ data.\nThe benchmark proceeds in two fine-tuning stages:\n\n  Stage 1 — Stability\
+  \ fine-tuning: ProteinMPNN is fine-tuned on the Megascale high-throughput\n  folding\
+  \ stability dataset (Tsuboyama 2023, ~350 k measurements, AlphaFold PDB structures,\n\
+  \  downloaded from Zenodo record 7992926).\n\n  Stage 2 — Binding fine-tuning: The\
+  \ stability-fine-tuned model is further fine-tuned on a\n  quality-filtered subset\
+  \ of the SKEMPI v2 database of protein-protein binding ΔΔG measurements,\n  using\
+  \ interface-homology cluster splits to prevent data leakage.\n\nEvaluation reproduces\
+  \ the SKEMPI benchmark results reported in the ICML 2025 paper, computing per-interface\
+  \ Pearson and Spearman correlations, overall correlations, RMSE, and ROC-AUC. Baseline\
+  \ comparisons against FoldX and Flex ddG are included in the packaged baselines/\
+  \ directory. StaB-ddG is the first deep-learning predictor reported to match FoldX\
+  \ accuracy while offering a >1,000× speed-up.\nInputs: PDB structure files (wild-type\
+  \ complex), chain specifications, and mutation strings. Outputs: per-mutant ΔΔG\
+  \ predictions (kcal/mol), per-interface correlation metrics, and a full evaluation\
+  \ CSV.\n"
+capability_tags:
+- protein_modeling
+- binding_energy_prediction
+- structural_biology
+- scientific_ml
+- transfer_learning
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+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:
+- per_interface_pearson
+- per_interface_spearman
+- overall_pearson
+- overall_spearman
+- roc_auc
+- ddg_predictions
+- binding_report
+launch_profiles:
+- id: single_prediction
+  label: Single Prediction
+  description: 'Run a single-mutation or multi-mutation ΔΔG prediction for one complex
+    using run_stabddg.py. Requires a PDB file, chain specification (e.g. ABC_DE),
+    and a mutation string (e.g. EA63Q,QD30V,KA66A). Uses the pre-trained stabddg.pt
+    checkpoint. Outputs output.csv with pred_1 column. No training data download required.
+
+    '
+- id: eval_only
+  label: Evaluation Only
+  description: 'Reproduce SKEMPI benchmark numbers using the pre-trained stabddg.pt
+    checkpoint via skempi_eval.py. Requires downloading SKEMPI2_PDBs from life.bsc.es.
+    Produces cache/eval.csv and per-interface correlation metrics. Fastest path to
+    paper numbers.
+
+    '
+- id: train_and_eval
+  label: Full Train + Eval
+  description: 'Run the complete two-stage fine-tuning and evaluation pipeline: (1)
+    stability fine-tuning on Megascale via stability_finetune.py, then (2) SKEMPI
+    binding fine-tuning via skempi_finetune.py, then (3) evaluation via skempi_eval.py.
+    Requires both Megascale and SKEMPI data downloads. Optional Wandb tracking via
+    --wandb flag.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: Megascale Folding Stability Dataset (Stage 1)
+    url: https://zenodo.org/records/7992926
+    files:
+    - Tsuboyama2023_Dataset2_Dataset3_20230416.csv
+    - AlphaFold_model_PDBs.zip
+    - Processed_K50_dG_datasets.zip
+    notes: 'Download via wget from Zenodo record 7992926. Unzip both archives before
+      running stability_finetune.py. Download takes several minutes. Required for
+      Stage 1 only.
+
+      '
+  - name: SKEMPI v2 Binding Energy Dataset (Stage 2 and Eval)
+    url: https://life.bsc.es/pid/skempi2/database/download/SKEMPI2_PDBs.tgz
+    files:
+    - SKEMPI2_PDBs.tgz
+    notes: 'Download via wget from life.bsc.es and extract with tar -xvzf. A quality-filtered
+      SKEMPI CSV and interface-homology cluster train/test splits are pre-packaged
+      in data/SKEMPI/. The raw PDB structures must be downloaded separately. Required
+      for Stage 2 fine-tuning and for eval_only profile.
+
+      '
+  - name: Benchmark Package Archive
+    url: https://deepscientist.cc/AISB/065_proteinbinding
+    archive_type: zip
+    local_dir_name: paper-65-ProteinBinding
+    notes: 'Contains pre-trained model checkpoints (model_ckpts/stabddg.pt, model_ckpts/stability_finetuned.pt,
+      model_ckpts/proteinmpnn.pt), packaged SKEMPI splits, baseline prediction CSVs,
+      and example inputs.
+
+      '
+  notes:
+  - Pre-trained checkpoints are included in the benchmark archive; full training requires
+    both external dataset downloads.
+  - SKEMPI PDB download is required even for eval_only; the filtered CSV is pre-packaged.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All data sources are publicly accessible without authentication.
+  - Optional Wandb account needed only if using the --wandb training flag.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: 'Sufficient for single-prediction and eval_only profiles using the pre-trained
+      checkpoint. Stability fine-tuning requires torch.device=''cuda'' and will OOM
+      below 16 GB VRAM.
+
+      '
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 'Recommended for full train+eval profile. Extra disk headroom accommodates
+      Megascale AlphaFold PDBs (~50 GB unzipped) and SKEMPI PDB structures alongside
+      checkpoints.
+
+      '
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.6.0
+  flash_attn: null
+  key_packages:
+  - biopython==1.85
+  - scipy
+  - tqdm
+  - wandb
+  - pandas
+  setup_method: conda
+  setup_file: environment.yaml
+  notes:
+  - Create environment with: conda env create -f environment.yaml && conda activate
+      stabddg
+  - Alternatively, install into an existing PyTorch environment with: pip install
+      tqdm scipy wandb pandas biopython
+  - Optional editable install: pip install -e .
+  - See environment.yaml for the full pinned dependency set.
+  - CUDA is required for training stages (torch.device='cuda'); inference can run
+    on CPU but is significantly slower.
+paper:
+  title: Predicting mutational effects on protein binding from folding energy
+  authors:
+  - Arthur Deng
+  - Karsten Householder
+  - Fang Wu
+  - Sebastian Thrun
+  - K. Christopher Garcia
+  - Brian Trippe
+  affiliation: Stanford University
+  venue: Proceedings of the 42nd International Conference on Machine Learning (ICML
+    2025)
+  year: 2025
+  url: https://arxiv.org/abs/2507.05502
+  code: https://github.com/LDeng0205/StaB-ddG
+  notes: 'Originally posted as an arXiv preprint (2507.05502) and accepted at ICML
+    2025 (PMLR 267). Introduces StaB-ddG as the first DL predictor to match FoldX
+    accuracy on SKEMPI while offering >1,000× speed-up over Rosetta-based Flex ddG.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.065_proteinbinding.zip
+  archive_type: zip
+  local_dir_name: paper-65-ProteinBinding
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.065_proteinbinding.zip
+  sha256: 07ea82db652e56e0bd609f637362cd91f74d76f1356d3b7cd3d9c14d82fda319
+  size_bytes: 2277469
+commercial:
+  annual_fee: null
+display:
+  palette_seed: sage-bone-protein
+  art_style: molecular-atlas
+  accent_priority: high
+image_path: ../image/065_aisb.t3.065_proteinbinding.jpg
+recommended_when: 'Use this benchmark when evaluating models on structural-biology
+  tasks involving PDB inputs, mutation lists, and ΔΔG regression or ranking metrics.
+  Ideal when the goal is to assess transfer learning from stability to binding, compare
+  against FoldX or Flex ddG baselines, or reproduce ICML 2025 SKEMPI benchmark results.
+  Also suitable for single-complex inference workflows where only the pre-trained
+  checkpoint is needed.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need a text-only or sequence-only
+  task with no structural inputs. Avoid if your environment cannot provide PDB structure
+  files for the wild-type interface or cannot satisfy the GPU memory requirement for
+  training. Not appropriate for benchmarks requiring fast iteration cycles, as the
+  full train+eval profile takes up to 24 h and requires large external dataset downloads.
+
+  '
+risk_flags:
+- external_data_download_required
+- gpu_required_for_training
+- no_execution_verified_in_packaging_pass
+risk_notes:
+- Metric values in metric_contract.json are code-backed but not yet verified by live
+  execution; trust only after a successful benchmark run produces cache/eval.csv.
+- Megascale AlphaFold PDB download (~50 GB unzipped) may be slow or interrupted on
+  low-bandwidth connections; resume support via wget -c.
+- SKEMPI PDB structures must be downloaded separately from life.bsc.es; the filtered
+  CSV and splits are pre-packaged but the raw PDBs are not.
+- Monte Carlo variance reduction uses ensemble=20 samples by default at inference;
+  reducing this speeds up inference but increases prediction variance.

package/AISB/catalog/aisb.t3.065_proteinbinding.zh.yaml

@@ -0,0 +1,167 @@
+schema_version: 1
+id: aisb.t3.065_proteinbinding
+name: StaB-ddG — 通过折叠能量预测蛋白质结合的突变效应
+version: 0.1.0
+one_line: 基于稳定性到结合参数化的蛋白质-蛋白质结合 ΔΔG 预测迁移学习基准，使用 ProteinMPNN 和两阶段微调。
+task_description: |
+  本基准包涵盖完整的 StaB-ddG 工作流程，用于预测突变对蛋白质-蛋白质结合自由能的影响（ΔΔG，kcal/mol）。其核心思想是热力学恒等式 ΔG_bind(A:B) = ΔG_fold(A:B) − ΔG_fold(A) − ΔG_fold(B)，它将结合能预测简化为折叠能预测，并实现了从大量稳定性数据到稀缺结合数据的迁移。
+
+  基准测试分为两个微调阶段：
+
+  第一阶段 — 稳定性微调：ProteinMPNN 在 Megascale 高通量折叠稳定性数据集上进行微调（Tsuboyama 2023，约 35 万个测量值，AlphaFold PDB 结构，从 Zenodo 记录 7992926 下载）。
+
+  第二阶段 — 结合微调：将在第一阶段微调后的模型在 SKEMPI v2 数据库中蛋白质-蛋白质结合 ΔΔG 测量的质量过滤子集上进一步微调，使用界面同源聚类划分来防止数据泄露。
+
+  评估复现了 ICML 2025 论文中报告的 SKEMPI 基准结果，计算每个界面的 Pearson 和 Spearman 相关系数、总体相关系数、RMSE 和 ROC-AUC。基准比较包含 FoldX 和 Flex ddG，在打包的 baselines/ 目录中提供。StaB-ddG 是首个被报道能够匹配 FoldX 精度的深度学习预测器，同时提供超过 1000 倍的加速。
+
+  输入：PDB 结构文件（野生型复合物）、链规格和突变字符串。输出：每个突变的 ΔΔG 预测值（kcal/mol）、每个界面的相关性指标以及完整的评估 CSV。
+capability_tags:
+- protein_modeling
+- binding_energy_prediction
+- structural_biology
+- scientific_ml
+- transfer_learning
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+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:
+- per_interface_pearson
+- per_interface_spearman
+- overall_pearson
+- overall_spearman
+- roc_auc
+- ddg_predictions
+- binding_report
+launch_profiles:
+- id: single_prediction
+  label: 单次预测
+  description: |
+    使用 run_stabddg.py 对单个复合物进行单点突变或多点突变 ΔΔG 预测。需要提供 PDB 文件、链规格（如 ABC_DE）和突变字符串（如 EA63Q,QD30V,KA66A）。使用预训练的 stabddg.pt 检查点。输出包含 pred_1 列的 output.csv。无需下载训练数据。
+- id: eval_only
+  label: 仅评估
+  description: |
+    通过 skempi_eval.py 使用预训练的 stabddg.pt 检查点复现 SKEMPI 基准数值。需要从 life.bsc.es 下载 SKEMPI2_PDBs。生成 cache/eval.csv 和每个界面的相关性指标。是获得论文数值的最快途径。
+- id: train_and_eval
+  label: 完整训练 + 评估
+  description: |
+    运行完整的两阶段微调和评估流程：（1）通过 stability_finetune.py 在 Megascale 上进行稳定性微调，然后（2）通过 skempi_finetune.py 进行 SKEMPI 结合微调，然后（3）通过 skempi_eval.py 进行评估。需要同时下载 Megascale 和 SKEMPI 数据。可通过 --wandb 标志启用可选的 Wandb 跟踪。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: Megascale Folding Stability Dataset (Stage 1)
+    url: https://zenodo.org/records/7992926
+    files:
+    - Tsuboyama2023_Dataset2_Dataset3_20230416.csv
+    - AlphaFold_model_PDBs.zip
+    - Processed_K50_dG_datasets.zip
+    notes: |
+      通过 wget 从 Zenodo 记录 7992926 下载。运行 stability_finetune.py 前请解压两个压缩包。下载需要几分钟。仅第一阶段需要。
+  - name: SKEMPI v2 Binding Energy Dataset (Stage 2 and Eval)
+    url: https://life.bsc.es/pid/skempi2/database/download/SKEMPI2_PDBs.tgz
+    files:
+    - SKEMPI2_PDBs.tgz
+    notes: |
+      通过 wget 从 life.bsc.es 下载并使用 tar -xvzf 解压。质量过滤后的 SKEMPI CSV 和界面同源聚类训练/测试划分已预装在 data/SKEMPI/ 中。原始 PDB 结构需要单独下载。第二阶段微调和仅评估模式需要此数据。
+  - name: Benchmark Package Archive
+    url: https://deepscientist.cc/AISB/065_proteinbinding
+    archive_type: zip
+    local_dir_name: paper-65-ProteinBinding
+    notes: |
+      包含预训练模型检查点（model_ckpts/stabddg.pt、model_ckpts/stability_finetuned.pt、model_ckpts/proteinmpnn.pt）、打包的 SKEMPI 划分、基线预测 CSV 和示例输入。
+  notes:
+  - 基准包中包含预训练检查点；完整训练需要下载两个外部数据集。
+  - 即使是仅评估模式也需要下载 SKEMPI PDB；过滤后的 CSV 已预装。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 所有数据源均可公开访问，无需认证。
+  - 仅在使用 --wandb 训练标志时才需要可选的 Wandb 账户。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: |
+      使用预训练检查点进行单次预测和仅评估模式时足够。稳定性微调需要 torch.device='cuda'，在 16 GB 以下显存会 OOM。
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: |
+      推荐用于完整训练+评估模式。额外的磁盘空间用于容纳 Megascale AlphaFold PDB（解压后约 50 GB）和 SKEMPI PDB 结构以及检查点。
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.6.0
+  flash_attn: null
+  key_packages:
+  - biopython==1.85
+  - scipy
+  - tqdm
+  - wandb
+  - pandas
+  setup_method: conda
+  setup_file: environment.yaml
+  notes: |
+    使用 conda env create -f environment.yaml && conda activate stabddg 创建环境。或者，使用 pip install tqdm scipy wandb pandas biopython 安装到现有 PyTorch 环境中。可选的 editable 安装：pip install -e . 有关完整的固定依赖集，请参阅 environment.yaml。训练阶段需要 CUDA（torch.device='cuda'）；推理可以在 CPU 上运行，但速度会显著变慢。
+paper:
+  title: Predicting mutational effects on protein binding from folding energy
+  authors:
+  - Arthur Deng
+  - Karsten Householder
+  - Fang Wu
+  - Sebastian Thrun
+  - K. Christopher Garcia
+  - Brian Trippe
+  affiliation: Stanford University
+  venue: Proceedings of the 42nd International Conference on Machine Learning (ICML
+    2025)
+  year: 2025
+  url: https://arxiv.org/abs/2507.05502
+  code: https://github.com/LDeng0205/StaB-ddG
+  notes: |
+    最初作为 arXiv 预印本（2507.05502）发布，并被 ICML 2025（PMLR 267）接收。介绍了 StaB-ddG 作为首个在 SKEMPI 上匹配 FoldX 精度的深度学习预测器，同时比基于 Rosetta 的 Flex ddG 提供超过 1000 倍的加速。
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.065_proteinbinding.zip
+  archive_type: zip
+  local_dir_name: paper-65-ProteinBinding
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.065_proteinbinding.zip
+  sha256: 07ea82db652e56e0bd609f637362cd91f74d76f1356d3b7cd3d9c14d82fda319
+  size_bytes: 2277469
+commercial:
+  annual_fee: null
+display:
+  palette_seed: sage-bone-protein
+  art_style: molecular-atlas
+  accent_priority: high
+image_path: ../image/065_aisb.t3.065_proteinbinding.jpg
+recommended_when: |
+  在评估涉及 PDB 输入、突变列表和 ΔΔG 回归或排序指标的结构生物学任务模型时使用此基准。当目标是评估从稳定性到结合的迁移学习、与 FoldX 或 Flex ddG 基线进行比较、或复现 ICML 2025 SKEMPI 基准结果时尤为理想。也适用于只需预训练检查点的单复合物推理工作流。
+not_recommended_when: |
+  如果需要纯文本或纯序列任务且没有结构输入，请勿使用此基准。如果您的环境无法提供野生型界面的 PDB 结构文件或无法满足训练的 GPU 内存要求，请避免使用。不适合需要快速迭代周期的基准测试，因为完整的训练+评估流程最多需要 24 小时，且需要下载大型外部数据集。
+risk_flags:
+- external_data_download_required
+- gpu_required_for_training
+- no_execution_verified_in_packaging_pass
+risk_notes: |
+  metric_contract.json 中的指标值有代码支持，但尚未通过实际执行验证；仅在基准测试成功运行生成 cache/eval.csv 后才可信。Megascale AlphaFold PDB 下载（解压后约 50 GB）在低带宽连接上可能较慢或中断；通过 wget -c 支持断点续传。SKEMPI PDB 结构必须从 life.bsc.es 单独下载；过滤后的 CSV 和划分已预装，但原始 PDB 除外。蒙特卡洛方差缩减在推理时默认使用 ensemble=20 个样本；减少此值会加快推理但会增加预测方差。

package/AISB/catalog/aisb.t3.066_tropicalattention.yaml

@@ -0,0 +1,267 @@
+id: aisb.t3.066_tropicalattention
+name: 'Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms'
+version: 0.1.0
+one_line: Tropical-geometry-based attention mechanism for neural algorithmic reasoning
+  on combinatorial tasks, achieving sharp OOD generalization beyond PTIME to NP-hard
+  problems.
+task_description: 'This packaged benchmark covers Tropical Attention, an attention
+  mechanism grounded in tropical geometry that lifts the attention kernel into tropical
+  projective space where reasoning is piecewise-linear and 1-Lipschitz. The benchmark
+  evaluates whether replacing or augmenting conventional softmax attention with tropical-geometry
+  structure improves algorithmic generalization on 11 combinatorial tasks spanning
+  polynomial-time algorithms (Floyd-Warshall, Quickselect, Convex Hull, SCC), NP-complete
+  problems (3SUM, Balanced Partition, Subset Sum), and NP-hard problems (0/1 Knapsack,
+  Bin Packing, Min Coin Change). Tasks are evaluated under out-of-distribution (OOD)
+  conditions including length extrapolation and value-range generalization. Primary
+  metric is quickselect_length_ood_f1 measuring length-OOD F1 score on the Quickselect
+  task.
+
+  '
+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:
+- quickselect_length_ood_f1
+- algorithmic_reasoning_report
+- trained_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged neural algorithmic reasoning evaluation route. Loads
+    a pre-trained checkpoint and evaluates on OOD test splits for all 11 combinatorial
+    tasks. Use --job_file pointing to jobs_to_do_evaluate.csv with --job_id selecting
+    the target task. Outputs length-OOD and value-OOD F1 scores per task.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval
+  description: 'Run the complete tropical-attention training and OOD reasoning benchmark
+    workflow. Trains models from scratch on 100k samples per task using schedulefree
+    optimizer, then evaluates on OOD test sets. Training logs saved to outputs/<timestamp>/.
+    Uses --job_file pointing to jobs_to_do_train.csv with --job_id selecting the task
+    row.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://deepscientist.cc/AISB/066_tropicalattention
+    type: benchmark_archive
+    archive_type: zip
+    local_dir_name: paper-66-TropicalAttention
+  - url: https://github.com/Baran-phys/Tropical-Attention
+    type: upstream_repository
+  notes:
+  - Archive contains full code, dataloaders, models, and evaluation scripts
+  - Datasets are procedurally generated at runtime via dataloaders.py
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: Capable of running single-task training; evaluation-only requires less
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 'Recommended for full multi-task training runs. Single GPU suffices due
+      to ~20% fewer parameters than Universal Transformer baselines.
+
+      '
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torch
+  - schedulefree
+  - pandas
+  - matplotlib
+  - scikit-learn
+  notes:
+  - See bundled requirements.txt for full dependency set
+  - TropicalAttention.py is a drop-in module for standard Transformer blocks
+  - CUDA required for training; CPU fallback supported for inference
+risk_flags:
+- ood_metric_variance
+- multi_seed_required
+risk_notes:
+- OOD F1 scores show variance across seeds (e.g., SubsetSum 87.50±6.45); multi-seed
+  evaluation recommended
+- Floyd-Warshall and Fractional Knapsack show lower OOD scores than classification
+  tasks
+- Training uses 100k samples per task; disk space needed for outputs/<timestamp>/
+  artifacts
+recommended_when: 'Use this benchmark when you need to evaluate neural reasoning over
+  graph and combinatorial problems with clear OOD behavior. Specifically suited for:
+  (1) evaluating attention mechanisms on length-extrapolation tasks, (2) benchmarking
+  NP-hard/complete problem solving beyond PTIME, (3) comparing tropical-geometry inductive
+  bias against softmax-based transformers, (4) studying polyhedral decision boundaries
+  in neural algorithmic reasoning.
+
+  '
+not_recommended_when: 'Do not use this if you need a plug-and-play application benchmark
+  without specialized reasoning datasets. Not suitable when only in-distribution accuracy
+  matters, or when PTIME-only algorithmic reasoning is sufficient. Avoid if the research
+  question does not involve combinatorial optimization, attention mechanism design,
+  or OOD generalization.
+
+  '
+paper:
+  title: 'Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.17190
+  authors:
+  - Hashemi, Baran (Origins Data Science Lab, Technical University of Munich)
+  - Pasque, Kurt (Naval Postgraduate School)
+  - Teska, Chris (Naval Postgraduate School)
+  - Yoshida, Ruriko (Naval Postgraduate School)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.066_tropicalattention.zip
+  archive_type: zip
+  local_dir_name: paper-66-TropicalAttention
+  upstream_url: https://github.com/Baran-phys/Tropical-Attention
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.066_tropicalattention.zip
+  sha256: 59c566c1c49e1649bad3e39f2100133d7cb97cea8df9665ae1dea27bcd4a80db
+  size_bytes: 1618773
+display:
+  palette_seed: lime-violet-tropical
+  art_style: geometric-reasoning
+  accent_priority: high
+image_path: ../image/066_aisb.t3.066_tropicalattention.jpg
+capability_tags:
+- research_code_optimization
+- neural_algorithmic_reasoning
+- combinatorial_optimization
+- transformers
+- mathematical_reasoning
+- tropical_geometry
+- out_of_distribution_generalization
+- attention_mechanisms
+- np_complete_problems
+- np_hard_problems
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+experiment_anchors:
+  train_script: experiment.py
+  train_job_csv: jobs_to_do_train.csv
+  eval_job_csv: jobs_to_do_evaluate.csv
+  attention_module: TropicalAttention.py
+  eval_script: eval.py
+  metric_contract: json/metric_contract.json
+task_specifications:
+- task_name: Quickselect
+  type: classification
+  ood_metric: length_ood_f1
+  training_length_range: (10,30)
+  evaluation_length_range: (10,1024)
+  epochs: 100
+  notes: Primary OOD benchmark; sharp attention maps preserve k-th element selection
+- task_name: SubsetSumDecision
+  type: classification
+  complexity: NP-complete
+  training_length_range: (8,8)
+  value_range: (-5,5)
+  epochs: 100
+- task_name: Knapsack
+  type: classification
+  complexity: NP-hard
+  training_length_range: (10,20)
+  epochs: 100
+- task_name: BalancedPartition
+  type: classification
+  complexity: NP-complete
+  training_length_range: (10,30)
+  epochs: 1000
+- task_name: BinPacking
+  type: classification
+  complexity: NP-hard
+  training_length_range: (10,30)
+  epochs: 1000
+- task_name: MinCoinChange
+  type: classification
+  training_length_range: (10,20)
+  epochs: 100
+- task_name: ConvexHull
+  type: classification
+  training_length_range: (10,30)
+  epochs: 100
+- task_name: FractionalKnapsack
+  type: regression
+  training_length_range: (10,20)
+  epochs: 100
+  notes: Lower OOD performance expected for regression tasks
+- task_name: FloydWarshall
+  type: regression
+  training_length_range: (10,30)
+  epochs: 100
+  notes: All-pairs shortest paths; lower OOD scores than classification tasks
+- task_name: SCC
+  type: classification
+  training_length_range: (10,30)
+  epochs: 100
+- task_name: 3SUM
+  type: classification
+  complexity: NP-complete
+  training_length_range: (10,30)
+  epochs: 100
+execution_notes:
+  training_invocation: 'python experiment.py --job_file jobs_to_do_train.csv --job_id
+    <row_index>
+
+    '
+  evaluation_invocation: 'python experiment.py --job_file jobs_to_do_evaluate.csv
+    --job_id <row_index>
+
+    '
+  job_csv_fields:
+  - task
+  - model_type
+  - dataset_name
+  - input_dim
+  - d_model
+  - n_heads
+  - num_layers
+  - num_epochs
+  - dropout
+  - lr
+  - seed
+  - n_samples
+  - batch_size
+  - length_range
+  - value_range
+  - noise_prob
+  output_directory: outputs/<timestamp>/
+  default_hyperparameters:
+    d_model: 64
+    n_heads: 2
+    num_layers: 1
+    dropout: 0
+    learning_rate: 0.0001
+    batch_size: 500
+    n_samples: 100000
+    seed: 999