{
  "id": 613304,
  "title": "Update in LB - TBM + Protenix + Boltz",
  "url": "/competitions/stanford-rna-3d-folding/discussion/613304",
  "author_name": "",
  "post_date": "2025-10-25T17:11:32.606884800Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>🧬  <strong>Template-Based Matching with Protenix + Boltz-2 + Agentic Tree Search</strong><br>\n<strong>Public Score: 0.675 | Private Score: 0.635</strong><br>\nNotebook: <a href=\"https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236\" target=\"_blank\">https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236</a></p>\n<h2>Overview</h2>\n<p>This solution combines Template-Based Matching (TBM) with Protenix and Boltz-2 models, guided by an Agentic Tree Search (ATS) framework that dynamically explores and refines candidate conformations. The integration of template priors, multi-model sampling, and structured search significantly improved performance to 0.634 (private) and 0.674 (public).</p>\n<h2>Pipeline Summary</h2>\n<ol>\n<li><p>Template-Based Matching (TBM):</p>\n<ul>\n<li>John Notebook</li></ul></li>\n<li><p>Protenix Diffusion Inference:</p>\n<ul>\n<li>D4T4 Team </li></ul></li>\n<li><p>Boltz-2 Diffusion Refinement:</p>\n<ul>\n<li>D4T4 team</li></ul></li>\n<li><p>Agentic Tree Search (ATS):<br>\nThe Weighted ATS ensemble autonomously selects optimal conformations across TBM, Protenix, and Boltz-2 models using a diversity- and reliability-aware heuristic.</p>\n<ul>\n<li>Weighted Multi-Model Selection: Combines 15 conformations (5 from each model) with tuned reliability weights — TBM (0.15), Protenix (0.75), Boltz-2 (0.10).</li>\n<li>Iterative Exploration: Expands from strong priors (Protenix conf1 + TBM conf0), scoring candidates via<br>\nscore = model_weight * (w_div * diversity - w_dist * distance_to_priors).</li>\n<li>Consensus Optimization: Selects 5 diverse yet stable conformations, maximizing geometric diversity while minimizing RMSD overlap.<br>\nImplemented in the methods agent_tree_search() and build_submission().</li></ul></li>\n</ol>\n<h2>Results</h2>\n<table>\n<thead>\n<tr>\n<th>Configuration</th>\n<th>Private</th>\n<th>Public</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Protenix + Boltz-2 (diffusion ensemble)</td>\n<td>0.46388</td>\n<td>0.48194</td>\n</tr>\n<tr>\n<td>TBM + Protenix + Boltz-2 + Agentic Tree Search</td>\n<td>0.63568</td>\n<td>0.67522</td>\n</tr>\n</tbody>\n</table>\n<h2>Insights</h2>\n<ul>\n<li>Template priors provide consistent backbone alignment.</li>\n<li>Weighted Agentic Tree Search balances exploration (diversity) and exploitation (confidence).</li>\n<li>Integrating Protenix’s geometric reasoning with Boltz-2’s diffusion refinement improves accuracy and structural realism.</li>\n</ul>\n<h2>Notebook</h2>\n<p>Full implementation: <a href=\"https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523\" target=\"_blank\">https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523</a></p>",
  "messages": [
    {
      "id": "3306946",
      "postDate": "10/25/2025 17:11:32",
      "content": "<p>🧬  <strong>Template-Based Matching with Protenix + Boltz-2 + Agentic Tree Search</strong><br>\n<strong>Public Score: 0.675 | Private Score: 0.635</strong><br>\nNotebook: <a href=\"https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236\" target=\"_blank\">https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236</a></p>\n<h2>Overview</h2>\n<p>This solution combines Template-Based Matching (TBM) with Protenix and Boltz-2 models, guided by an Agentic Tree Search (ATS) framework that dynamically explores and refines candidate conformations. The integration of template priors, multi-model sampling, and structured search significantly improved performance to 0.634 (private) and 0.674 (public).</p>\n<h2>Pipeline Summary</h2>\n<ol>\n<li><p>Template-Based Matching (TBM):</p>\n<ul>\n<li>John Notebook</li></ul></li>\n<li><p>Protenix Diffusion Inference:</p>\n<ul>\n<li>D4T4 Team </li></ul></li>\n<li><p>Boltz-2 Diffusion Refinement:</p>\n<ul>\n<li>D4T4 team</li></ul></li>\n<li><p>Agentic Tree Search (ATS):<br>\nThe Weighted ATS ensemble autonomously selects optimal conformations across TBM, Protenix, and Boltz-2 models using a diversity- and reliability-aware heuristic.</p>\n<ul>\n<li>Weighted Multi-Model Selection: Combines 15 conformations (5 from each model) with tuned reliability weights — TBM (0.15), Protenix (0.75), Boltz-2 (0.10).</li>\n<li>Iterative Exploration: Expands from strong priors (Protenix conf1 + TBM conf0), scoring candidates via<br>\nscore = model_weight * (w_div * diversity - w_dist * distance_to_priors).</li>\n<li>Consensus Optimization: Selects 5 diverse yet stable conformations, maximizing geometric diversity while minimizing RMSD overlap.<br>\nImplemented in the methods agent_tree_search() and build_submission().</li></ul></li>\n</ol>\n<h2>Results</h2>\n<table>\n<thead>\n<tr>\n<th>Configuration</th>\n<th>Private</th>\n<th>Public</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Protenix + Boltz-2 (diffusion ensemble)</td>\n<td>0.46388</td>\n<td>0.48194</td>\n</tr>\n<tr>\n<td>TBM + Protenix + Boltz-2 + Agentic Tree Search</td>\n<td>0.63568</td>\n<td>0.67522</td>\n</tr>\n</tbody>\n</table>\n<h2>Insights</h2>\n<ul>\n<li>Template priors provide consistent backbone alignment.</li>\n<li>Weighted Agentic Tree Search balances exploration (diversity) and exploitation (confidence).</li>\n<li>Integrating Protenix’s geometric reasoning with Boltz-2’s diffusion refinement improves accuracy and structural realism.</li>\n</ul>\n<h2>Notebook</h2>\n<p>Full implementation: <a href=\"https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523\" target=\"_blank\">https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523</a></p>",
      "rawMarkdown": "🧬  **Template-Based Matching with Protenix + Boltz-2 + Agentic Tree Search**\n**Public Score: 0.675 | Private Score: 0.635**\nNotebook: https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236\n\nOverview\n---------\nThis solution combines Template-Based Matching (TBM) with Protenix and Boltz-2 models, guided by an Agentic Tree Search (ATS) framework that dynamically explores and refines candidate conformations. The integration of template priors, multi-model sampling, and structured search significantly improved performance to 0.634 (private) and 0.674 (public).\n\nPipeline Summary\n----------------\n1. Template-Based Matching (TBM):\n   - John Notebook\n\n2. Protenix Diffusion Inference:\n   - D4T4 Team \n\n3. Boltz-2 Diffusion Refinement:\n   - D4T4 team\n\n4. Agentic Tree Search (ATS):\n   The Weighted ATS ensemble autonomously selects optimal conformations across TBM, Protenix, and Boltz-2 models using a diversity- and reliability-aware heuristic.\n   - Weighted Multi-Model Selection: Combines 15 conformations (5 from each model) with tuned reliability weights — TBM (0.15), Protenix (0.75), Boltz-2 (0.10).\n   - Iterative Exploration: Expands from strong priors (Protenix conf1 + TBM conf0), scoring candidates via\n     score = model_weight * (w_div * diversity - w_dist * distance_to_priors).\n   - Consensus Optimization: Selects 5 diverse yet stable conformations, maximizing geometric diversity while minimizing RMSD overlap.\n   Implemented in the methods agent_tree_search() and build_submission().\n\nResults\n-------\nConfiguration                              | Private | Public\n-------------------------------------------|----------|---------\nProtenix + Boltz-2 (diffusion ensemble)     | 0.46388    |  0.48194\nTBM + Protenix + Boltz-2 + Agentic Tree Search | 0.63568 | 0.67522\n\nInsights\n--------\n- Template priors provide consistent backbone alignment.\n- Weighted Agentic Tree Search balances exploration (diversity) and exploitation (confidence).\n- Integrating Protenix’s geometric reasoning with Boltz-2’s diffusion refinement improves accuracy and structural realism.\n\nNotebook\n--------\nFull implementation: https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523",
      "votes": null
    },
    {
      "id": "3388496",
      "postDate": "01/09/2026 01:49:05",
      "content": "<p>Hi I want money </p>",
      "rawMarkdown": "Hi I want money",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3388496,
      "author_name": "phumlani93tshabalala",
      "author_url": "",
      "post_date": "01/09/2026 01:49:05",
      "content": "<p>Hi I want money </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3306946": "🧬  **Template-Based Matching with Protenix + Boltz-2 + Agentic Tree Search**\n**Public Score: 0.675 | Private Score: 0.635**\nNotebook: https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=266875236\n\nOverview\n---------\nThis solution combines Template-Based Matching (TBM) with Protenix and Boltz-2 models, guided by an Agentic Tree Search (ATS) framework that dynamically explores and refines candidate conformations. The integration of template priors, multi-model sampling, and structured search significantly improved performance to 0.634 (private) and 0.674 (public).\n\nPipeline Summary\n----------------\n1. Template-Based Matching (TBM):\n   - John Notebook\n\n2. Protenix Diffusion Inference:\n   - D4T4 Team \n\n3. Boltz-2 Diffusion Refinement:\n   - D4T4 team\n\n4. Agentic Tree Search (ATS):\n   The Weighted ATS ensemble autonomously selects optimal conformations across TBM, Protenix, and Boltz-2 models using a diversity- and reliability-aware heuristic.\n   - Weighted Multi-Model Selection: Combines 15 conformations (5 from each model) with tuned reliability weights — TBM (0.15), Protenix (0.75), Boltz-2 (0.10).\n   - Iterative Exploration: Expands from strong priors (Protenix conf1 + TBM conf0), scoring candidates via\n     score = model_weight * (w_div * diversity - w_dist * distance_to_priors).\n   - Consensus Optimization: Selects 5 diverse yet stable conformations, maximizing geometric diversity while minimizing RMSD overlap.\n   Implemented in the methods agent_tree_search() and build_submission().\n\nResults\n-------\nConfiguration                              | Private | Public\n-------------------------------------------|----------|---------\nProtenix + Boltz-2 (diffusion ensemble)     | 0.46388    |  0.48194\nTBM + Protenix + Boltz-2 + Agentic Tree Search | 0.63568 | 0.67522\n\nInsights\n--------\n- Template priors provide consistent backbone alignment.\n- Weighted Agentic Tree Search balances exploration (diversity) and exploitation (confidence).\n- Integrating Protenix’s geometric reasoning with Boltz-2’s diffusion refinement improves accuracy and structural realism.\n\nNotebook\n--------\nFull implementation: https://www.kaggle.com/code/arunodhayan/tbm-protenix-boltz?scriptVersionId=26687523",
    "3388496": "Hi I want money"
  },
  "source": "meta"
}