Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta charset="UTF-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> | |
| <title>MolForge: The Scientific Method as a Workflow</title> | |
| <script src="https://cdn.tailwindcss.com"></script> | |
| <link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&family=JetBrains+Mono&display=swap" rel="stylesheet"> | |
| <style> | |
| body { | |
| font-family: 'Inter', -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif; | |
| -webkit-font-smoothing: antialiased; | |
| background-color: #ffffff; | |
| color: #0f172a; | |
| } | |
| .prose-custom { | |
| max-width: 70ch; | |
| margin: 0 auto; | |
| } | |
| .shadcn-card { | |
| border: 1px solid #e2e8f0; | |
| background: #ffffff; | |
| border-radius: 0.75rem; | |
| box-shadow: 0 1px 3px 0 rgb(0 0 0 / 0.1), 0 1px 2px -1px rgb(0 0 0 / 0.1); | |
| } | |
| .shadcn-badge { | |
| background: #f1f5f9; | |
| color: #475569; | |
| font-size: 0.75rem; | |
| font-weight: 600; | |
| padding: 0.125rem 0.625rem; | |
| border-radius: 9999px; | |
| display: inline-flex; | |
| align-items: center; | |
| } | |
| .mono { | |
| font-family: 'JetBrains Mono', monospace; | |
| } | |
| .video-container { | |
| position: relative; | |
| padding-bottom: 56.25%; | |
| height: 0; | |
| overflow: hidden; | |
| border-radius: 0.75rem; | |
| border: 1px solid #e2e8f0; | |
| } | |
| .video-container iframe { | |
| position: absolute; | |
| top: 0; | |
| left: 0; | |
| width: 100%; | |
| height: 100%; | |
| } | |
| </style> | |
| </head> | |
| <body class="bg-white"> | |
| <!-- Navigation --> | |
| <nav class="border-b sticky top-0 bg-white/80 backdrop-blur-md z-50"> | |
| <div class="max-w-5xl mx-auto px-6 h-16 flex items-center justify-between"> | |
| <span class="font-bold tracking-tight text-lg">MolForge</span> | |
| <div class="flex gap-6 text-sm font-medium text-slate-600"> | |
| <a href="https://github.com/Adhitya-Vardhan/molt_lab" target="_blank" rel="noopener noreferrer" class="hover:text-black transition-colors">GitHub</a> | |
| <a href="https://huggingface.co/spaces/Adhitya122/molforge" class="hover:text-black transition-colors">Space</a> | |
| <a href="https://colab.research.google.com/drive/1c6npGkGNbbbd8XFNeS6zInBpopLnJ4W4?usp=sharing" target="_blank" rel="noopener noreferrer" class="hover:text-black transition-colors font-bold text-indigo-600">Try Training</a> | |
| </div> | |
| </div> | |
| </nav> | |
| <main class="max-w-5xl mx-auto px-6 py-20"> | |
| <!-- Header --> | |
| <div class="mb-16 text-center"> | |
| <div class="flex justify-center gap-2 mb-6"> | |
| <span class="shadcn-badge bg-indigo-50 text-indigo-700 border border-indigo-100">Hackathon Submission</span> | |
| <span class="shadcn-badge">Medical Oncology</span> | |
| </div> | |
| <h1 class="text-4xl md:text-6xl font-extrabold tracking-tight mb-6 leading-tight max-w-3xl mx-auto"> | |
| MolForge: The Scientific Method as a Workflow | |
| </h1> | |
| <p class="text-xl text-slate-500 mb-10 leading-relaxed max-w-2xl mx-auto"> | |
| How we trained an LLM to navigate a resource-constrained laboratory, optimize oncology drug candidates, and survive scientific "sunk-cost" traps. | |
| </p> | |
| <div class="flex items-center justify-center gap-4 text-sm text-slate-400 mb-12"> | |
| <div class="w-10 h-10 rounded-full bg-indigo-100 flex items-center justify-center text-indigo-600 font-bold">AV</div> | |
| <div class="text-left"> | |
| <p class="font-semibold text-slate-900 leading-none mb-1">Adhitya Vardhan</p> | |
| <p class="leading-none">OpenEnv Hackathon 2026 • 12 min read</p> | |
| </div> | |
| </div> | |
| <!-- Video Section --> | |
| <div class="max-w-3xl mx-auto mb-20"> | |
| <div class="video-container shadow-2xl"> | |
| <iframe src="https://www.youtube.com/embed/q8YoA0YhIn8" title="MolForge Explainer Video" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe> | |
| </div> | |
| <p class="mt-4 text-sm text-slate-400 italic">Watch the MolForge technical explainer (3:24)</p> | |
| </div> | |
| </div> | |
| <!-- Introduction --> | |
| <div class="prose prose-slate prose-lg max-w-none mb-24 leading-relaxed text-slate-700 border-b pb-20"> | |
| <h2 class="text-3xl font-bold text-slate-900 mb-6">Introduction</h2> | |
| <p> | |
| In traditional drug discovery tasks, LLMs are often asked to "generate a molecule" in a single shot. But science doesn't happen in a vacuum. It happens in the loop—through trial, error, and verification. | |
| </p> | |
| <p class="mt-4"> | |
| <strong>MolForge</strong> is a reinforcement learning environment that simulates a medical oncology discovery lab. It forces the model to navigate real-world constraints: limited budget, molecular toxicity, and synthesis complexity. | |
| </p> | |
| <div class="mt-12 p-8 bg-slate-900 text-slate-200 rounded-2xl shadow-lg"> | |
| <p class="text-indigo-400 font-bold mb-4 uppercase tracking-widest text-xs">Core Philosophy</p> | |
| <p class="text-2xl font-medium leading-tight"> | |
| "The model is a trainable research agent inside a controlled scientific environment, not an oracle. It is judged by chemistry and biomedical verifiers, corrected by specialist feedback, and scored by a reward system that explains exactly where the path to a discovery failed." | |
| </p> | |
| </div> | |
| </div> | |
| <!-- The Scientific Verifier Stack --> | |
| <section class="mb-32"> | |
| <div class="flex items-center gap-3 mb-8"> | |
| <div class="w-10 h-10 rounded-lg bg-emerald-600 flex items-center justify-center text-white font-bold">🧪</div> | |
| <h2 class="text-3xl font-bold tracking-tight">The Scientific Verifier Stack</h2> | |
| </div> | |
| <p class="text-slate-600 mb-10 text-lg leading-relaxed"> | |
| MolForge doesn't just predict outcomes; it utilizes multiple simulation layers to ground the model's decisions in chemical and biological reality. | |
| </p> | |
| <div class="grid md:grid-cols-3 gap-6 mb-12"> | |
| <div class="shadcn-card p-8 bg-white hover:border-emerald-500 transition-all cursor-default group"> | |
| <div class="w-12 h-12 bg-emerald-50 text-emerald-600 rounded-xl flex items-center justify-center mb-6 text-xl">🧬</div> | |
| <h4 class="font-bold text-lg mb-3">RDKit</h4> | |
| <p class="text-sm text-slate-500 leading-relaxed italic">"Keeping molecules physically possible"</p> | |
| <p class="text-sm text-slate-600 mt-4 leading-relaxed mb-6"> | |
| RDKit acts as the fundamental chemistry ruleset. It checks for molecular valency, ensures every edit is chemically plausible, and calculates core descriptors like Lipophilicity and TPSA. | |
| </p> | |
| <a href="https://www.rdkit.org" target="_blank" class="text-xs font-bold text-emerald-600 uppercase tracking-widest hover:underline">Visit RDKit.org →</a> | |
| </div> | |
| <div class="shadcn-card p-8 bg-white hover:border-blue-500 transition-all cursor-default group"> | |
| <div class="w-12 h-12 bg-blue-50 text-blue-600 rounded-xl flex items-center justify-center mb-6 text-xl">💊</div> | |
| <h4 class="font-bold text-lg mb-3">TDC Oracles</h4> | |
| <p class="text-sm text-slate-500 leading-relaxed italic">"Predicting biomedical fate"</p> | |
| <p class="text-sm text-slate-600 mt-4 leading-relaxed mb-6"> | |
| Utilizing the Therapeutics Data Commons, MolForge predicts real-world ADMET properties, toxicity risks, and synthesizability scores (SA_Score) for every candidate. | |
| </p> | |
| <a href="https://tdcommons.ai" target="_blank" class="text-xs font-bold text-blue-600 uppercase tracking-widest hover:underline">Explore TDCommons.ai →</a> | |
| </div> | |
| <div class="shadcn-card p-8 bg-white hover:border-indigo-500 transition-all cursor-default"> | |
| <div class="w-12 h-12 bg-indigo-50 text-indigo-600 rounded-xl flex items-center justify-center mb-6 text-xl">🎯</div> | |
| <h4 class="font-bold text-lg mb-3">Heuristic Docking</h4> | |
| <p class="text-sm text-slate-500 leading-relaxed italic">"Simulating receptor-drug fit"</p> | |
| <p class="text-sm text-slate-600 mt-4 leading-relaxed"> | |
| A fast, physics-inspired simulation that updates potency in milliseconds based on structural pocket matching and receptor complementarity. | |
| </p> | |
| </div> | |
| </div> | |
| <div class="p-8 bg-slate-50 border border-slate-200 rounded-2xl"> | |
| <h5 class="font-bold text-slate-900 mb-6 flex items-center gap-2"> | |
| <span class="w-2 h-2 bg-indigo-500 rounded-full"></span> | |
| The 3 Rules of Potency Simulation | |
| </h5> | |
| <div class="grid md:grid-cols-3 gap-8"> | |
| <div class="space-y-2"> | |
| <p class="font-bold text-sm text-slate-800">1. Pocket Matching</p> | |
| <p class="text-xs text-slate-500 leading-relaxed">Structural fit of the fragment (e.g., azaindole) into the KRAS G12C target pocket.</p> | |
| </div> | |
| <div class="space-y-2"> | |
| <p class="font-bold text-sm text-slate-800">2. Lipophilic Match</p> | |
| <p class="text-xs text-slate-500 leading-relaxed">Targeting the ideal LogP of <strong>3.0</strong> for optimal binding without repulsive clashes.</p> | |
| </div> | |
| <div class="space-y-2"> | |
| <p class="font-bold text-sm text-slate-800">3. Polarity Match</p> | |
| <p class="text-xs text-slate-500 leading-relaxed">Optimizing TPSA toward the ideal <strong>85.0</strong> to avoid polar clashes in hydrophobic pockets.</p> | |
| </div> | |
| </div> | |
| </div> | |
| </section> | |
| <!-- The POMDP Architecture --> | |
| <section class="mb-32"> | |
| <div class="flex items-center gap-3 mb-8"> | |
| <div class="w-10 h-10 rounded-lg bg-indigo-600 flex items-center justify-center text-white font-bold">1</div> | |
| <h2 class="text-3xl font-bold tracking-tight">The POMDP Architecture</h2> | |
| </div> | |
| <p class="text-slate-600 mb-10 text-lg leading-relaxed"> | |
| MolForge is built as a <strong>Partially Observable Markov Decision Process (POMDP)</strong>. This means the agent never sees the "hidden truth" of the receptor. It only sees what its budget allows it to assay. | |
| </p> | |
| <div class="shadcn-card p-4 bg-slate-50 mb-12 border-dashed"> | |
| <img src="assets/molforge_architecture.png" alt="Architecture" class="rounded-lg w-full"> | |
| <p class="mt-4 text-center text-xs text-slate-400 font-medium tracking-wide">THE SCIENTIFIC FEEDBACK LOOP: VERIFIER-FIRST DESIGN</p> | |
| </div> | |
| <div class="grid md:grid-cols-2 gap-8"> | |
| <div class="space-y-4"> | |
| <h4 class="font-bold text-slate-900 border-b pb-2">The Hidden State</h4> | |
| <ul class="space-y-3 text-slate-600 text-sm"> | |
| <li class="flex gap-2"><span>•</span> <strong>Ground Truth Potency:</strong> The exact hidden binding energy of the KRAS G12C pocket.</li> | |
| <li class="flex gap-2"><span>•</span> <strong>Sunk-Cost Traps:</strong> Starting scaffolds that look promising but have hidden liabilities.</li> | |
| <li class="flex gap-2"><span>•</span> <strong>Target Mutation:</strong> Late-stage shifts in the pocket (Level 2) that punish blind optimization.</li> | |
| </ul> | |
| </div> | |
| <div class="space-y-4"> | |
| <h4 class="font-bold text-indigo-600 border-b pb-2">The Visible Evidence</h4> | |
| <ul class="space-y-3 text-slate-600 text-sm"> | |
| <li class="flex gap-2"><span>•</span> <strong>RDKit & TDC Signals:</strong> Noisy, verifier-backed readings of Lipophilicity (LogP) and TPSA.</li> | |
| <li class="flex gap-2"><span>•</span> <strong>Heuristic Docking:</strong> Fast simulations of pocket matching and receptor fit.</li> | |
| <li class="flex gap-2"><span>•</span> <strong>Governance Vetoes:</strong> Objections from the Safety Specialist or Process Chemist.</li> | |
| </ul> | |
| </div> | |
| </div> | |
| </section> | |
| <!-- Search Space & Scenarios --> | |
| <section class="mb-32"> | |
| <div class="flex items-center gap-3 mb-8"> | |
| <div class="w-10 h-10 rounded-lg bg-indigo-600 flex items-center justify-center text-white font-bold">2</div> | |
| <h2 class="text-3xl font-bold tracking-tight">The Molecular Search Space</h2> | |
| </div> | |
| <p class="text-slate-600 mb-10 text-lg leading-relaxed"> | |
| We don't ask the model to memorize molecules. We ask it to navigate a <strong>combinatorial space of 256 fragments</strong> across three starting scenarios. | |
| </p> | |
| <div class="grid grid-cols-2 md:grid-cols-4 gap-4 mb-12"> | |
| <div class="shadcn-card p-4 text-center"> | |
| <p class="text-xs font-bold text-slate-400 uppercase mb-2">Warhead</p> | |
| <p class="text-sm font-semibold">4 Options</p> | |
| </div> | |
| <div class="shadcn-card p-4 text-center"> | |
| <p class="text-xs font-bold text-slate-400 uppercase mb-2">Hinge</p> | |
| <p class="text-sm font-semibold">4 Options</p> | |
| </div> | |
| <div class="shadcn-card p-4 text-center"> | |
| <p class="text-xs font-bold text-slate-400 uppercase mb-2">Solvent Tail</p> | |
| <p class="text-sm font-semibold">4 Options</p> | |
| </div> | |
| <div class="shadcn-card p-4 text-center"> | |
| <p class="text-xs font-bold text-slate-400 uppercase mb-2">Back Pocket</p> | |
| <p class="text-sm font-semibold">4 Options</p> | |
| </div> | |
| </div> | |
| <h3 class="text-xl font-bold mb-6">Benchmark Scenarios</h3> | |
| <div class="shadcn-card overflow-hidden mb-12"> | |
| <table class="w-full text-left text-sm"> | |
| <thead class="bg-slate-50 border-b"> | |
| <tr> | |
| <th class="px-6 py-4 font-semibold text-slate-900">Scenario</th> | |
| <th class="px-6 py-4 font-semibold text-slate-900">Story</th> | |
| <th class="px-6 py-4 font-semibold text-slate-900">Budget</th> | |
| <th class="px-6 py-4 font-semibold text-slate-900">Difficulty</th> | |
| </tr> | |
| </thead> | |
| <tbody class="divide-y"> | |
| <tr> | |
| <td class="px-6 py-4 font-bold text-indigo-600">Level 0: Easy</td> | |
| <td class="px-6 py-4 text-slate-500">Near-viable scaffold needs safety repair and evidence.</td> | |
| <td class="px-6 py-4">3600</td> | |
| <td class="px-6 py-4"><span class="shadcn-badge bg-emerald-50 text-emerald-700">Low</span></td> | |
| </tr> | |
| <tr> | |
| <td class="px-6 py-4 font-bold text-indigo-600">Level 1: Medium</td> | |
| <td class="px-6 py-4 text-slate-500">Potency, toxicity, and synthesis must be balanced.</td> | |
| <td class="px-6 py-4">4300</td> | |
| <td class="px-6 py-4"><span class="shadcn-badge bg-orange-50 text-orange-700">Moderate</span></td> | |
| </tr> | |
| <tr> | |
| <td class="px-6 py-4 font-bold text-indigo-600">Level 2: Hard</td> | |
| <td class="px-6 py-4 text-slate-500">Sunk-cost trap: starting series has hidden liability.</td> | |
| <td class="px-6 py-4">5200</td> | |
| <td class="px-6 py-4"><span class="shadcn-badge bg-red-50 text-red-700">Critical</span></td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| </div> | |
| </section> | |
| <!-- Reward Design --> | |
| <section class="mb-32"> | |
| <div class="flex items-center gap-3 mb-8"> | |
| <div class="w-10 h-10 rounded-lg bg-indigo-600 flex items-center justify-center text-white font-bold">3</div> | |
| <h2 class="text-3xl font-bold tracking-tight">Reward Design: Beyond Scalar Scores</h2> | |
| </div> | |
| <p class="text-slate-600 mb-10 text-lg leading-relaxed"> | |
| Training for scientific rigor requires more than a "Good/Bad" signal. We use a <strong>decomposed reward system</strong> that mixes coarse shaping with sparse terminal bonuses. | |
| </p> | |
| <div class="grid md:grid-cols-3 gap-6 mb-12"> | |
| <div class="shadcn-card p-6 bg-slate-50 border-t-4 border-t-indigo-500"> | |
| <h4 class="font-bold mb-2 text-sm uppercase tracking-wider text-slate-500">Coarse Shaping</h4> | |
| <p class="text-xs text-slate-500">Edit feedback avoids exact hidden deltas, forcing the model to rely on empirical assays.</p> | |
| </div> | |
| <div class="shadcn-card p-6 bg-slate-50 border-t-4 border-t-emerald-500"> | |
| <h4 class="font-bold mb-2 text-sm uppercase tracking-wider text-slate-500">Evidence Multipliers</h4> | |
| <p class="text-xs text-slate-500">Submissions without current potency, toxicity, and synthesis support receive massive penalties.</p> | |
| </div> | |
| <div class="shadcn-card p-6 bg-slate-50 border-t-4 border-t-orange-500"> | |
| <h4 class="font-bold mb-2 text-sm uppercase tracking-wider text-slate-500">Budget Efficiency</h4> | |
| <p class="text-xs text-slate-500">Small credits for valid evidence-backed submissions that use less than the allocated budget.</p> | |
| </div> | |
| </div> | |
| <div class="p-6 bg-indigo-50 border border-indigo-100 rounded-xl text-sm"> | |
| <p class="font-bold text-indigo-700 mb-2 italic">"Curriculum mode is the RL warm-up engine—providing the breadcrumbs needed for the model to discover the submission bonus."</p> | |
| </div> | |
| </section> | |
| <!-- Results --> | |
| <section class="mb-32"> | |
| <div class="flex items-center gap-3 mb-8"> | |
| <div class="w-10 h-10 rounded-lg bg-indigo-600 flex items-center justify-center text-white font-bold">4</div> | |
| <h2 class="text-3xl font-bold tracking-tight">Training Results</h2> | |
| </div> | |
| <div class="shadcn-card overflow-hidden mb-12"> | |
| <table class="w-full text-left text-sm"> | |
| <thead class="bg-slate-900 text-white"> | |
| <tr> | |
| <th class="px-6 py-4 font-semibold uppercase tracking-wider text-[10px]">Difficulty</th> | |
| <th class="px-6 py-4 font-semibold uppercase tracking-wider text-[10px]">Before (SFT)</th> | |
| <th class="px-6 py-4 font-semibold uppercase tracking-wider text-[10px]">After (RL)</th> | |
| <th class="px-6 py-4 font-semibold uppercase tracking-wider text-[10px]">Improvement</th> | |
| </tr> | |
| </thead> | |
| <tbody class="divide-y"> | |
| <tr> | |
| <td class="px-6 py-5 font-bold">Level 0: Easy</td> | |
| <td class="px-6 py-5 text-slate-400">0.1167</td> | |
| <td class="px-6 py-5 font-extrabold text-slate-900">0.1295</td> | |
| <td class="px-6 py-5 text-emerald-600 font-black">+10.9%</td> | |
| </tr> | |
| <tr> | |
| <td class="px-6 py-5 font-bold">Level 1: Medium</td> | |
| <td class="px-6 py-5 text-slate-400">0.1167</td> | |
| <td class="px-6 py-5 font-extrabold text-slate-900">0.1278</td> | |
| <td class="px-6 py-5 text-emerald-600 font-black">+9.5%</td> | |
| </tr> | |
| <tr> | |
| <td class="px-6 py-5 font-bold">Level 2: Hard</td> | |
| <td class="px-6 py-5 text-slate-400">0.0800</td> | |
| <td class="px-6 py-5 font-extrabold text-slate-900">0.0866</td> | |
| <td class="px-6 py-5 text-emerald-600 font-black">+8.3%</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| </div> | |
| <div class="grid md:grid-cols-2 gap-8"> | |
| <div class="shadcn-card p-6"> | |
| <img src="assets/reward_curve.png" alt="Reward Curve" class="rounded border mb-4"> | |
| <p class="text-xs font-bold text-slate-400 uppercase tracking-widest text-center">RL Training Progression</p> | |
| </div> | |
| <div class="shadcn-card p-6"> | |
| <img src="assets/Logs.png" alt="Logs" class="rounded border mb-4"> | |
| <p class="text-xs font-bold text-slate-400 uppercase tracking-widest text-center">Governance Action History</p> | |
| </div> | |
| </div> | |
| </section> | |
| <!-- Final Takeaway --> | |
| <section class="mb-32 pt-20 border-t text-center"> | |
| <h2 class="text-4xl font-black mb-6 tracking-tight">Final Takeaway</h2> | |
| <p class="text-slate-500 max-w-2xl mx-auto mb-12 text-lg leading-relaxed text-justify"> | |
| MolForge proves that scientific AI should not be built as a single-shot generator. By grounding the LLM in a <strong>closed-loop scientific environment</strong>, we can train models that respect budget, coordinate with specialists, and base their discoveries on verifiable evidence. | |
| </p> | |
| <div class="flex flex-wrap justify-center gap-4"> | |
| <a href="https://github.com/Adhitya-Vardhan/molt_lab" target="_blank" rel="noopener noreferrer" class="px-10 py-4 bg-slate-900 text-white font-bold rounded-xl hover:bg-slate-800 transition-all shadow-xl hover:-translate-y-1">Explore Code</a> | |
| <a href="https://colab.research.google.com/drive/1c6npGkGNbbbd8XFNeS6zInBpopLnJ4W4?usp=sharing" target="_blank" rel="noopener noreferrer" class="px-10 py-4 bg-white border-2 border-slate-900 text-slate-900 font-bold rounded-xl hover:bg-slate-50 transition-all hover:-translate-y-1">Run Notebook</a> | |
| </div> | |
| <div class="mt-16 flex justify-center gap-8 text-sm font-bold text-slate-400"> | |
| <a href="https://huggingface.co/spaces/Adhitya122/molforge" class="hover:text-indigo-600 transition-colors uppercase tracking-widest">Space Deployment</a> | |
| <a href="https://huggingface.co/Adhitya122/molforge-grpo-oncology" target="_blank" rel="noopener noreferrer" class="hover:text-indigo-600 transition-colors uppercase tracking-widest">Model Card</a> | |
| </div> | |
| </section> | |
| <!-- Footer --> | |
| <footer class="py-12 border-t text-xs text-slate-400 text-center"> | |
| <p>© 2026 MolForge Project • Built for the OpenEnv Hackathon</p> | |
| </footer> | |
| </main> | |
| </body> | |
| </html> | |