Protein-aware optimization
Search is driven by a docking oracle for your exact target pocket — molecules are optimized to bind, not just to look plausible.
Protein-aware molecular optimization
Given a target, find molecules that bind — drug-like and easy to make.
ChemLlama turns protein-aware molecular optimization into one loop: propose candidates, score them against docking, drug-likeness, and synthesizability oracles, and let the search converge on molecules that satisfy all three at once.
- Optimize docking scores against your exact protein pocket.
- Hold QED drug-likeness and SA synthesizability — not affinity alone.
- Oracle-guided search within a fixed call budget, no brute-force screening.
Start with a protein
Enter a PDB code or upload a structure file.
or
No account needed to explore.
- 4 generative methods
- ChemLlaMA · GenMol · Saturn · Genetic-GFN
- 3 benchmark tasks
- Hit · Lead · Specificity
- 5 protein targets
- parp1 · fa7 · 5ht1b · braf · jak2
- 4 scoring oracles
- QED · SA · docking · similarity
From target to optimized molecules
No brute-force screening — propose, score against the oracles, and let multi-objective search converge within the call budget.
- 01
Pick a target
Choose a protein pocket — parp1, fa7, 5ht1b, braf, jak2, or your own structure via its PDB code.
- 02
Generate candidates
A generative method proposes SMILES: ChemLlaMA genetic search, discrete diffusion, Mamba RL, or Genetic-GFlowNet.
- 03
Score with oracles
Every molecule is scored for docking (QuickVina), QED drug-likeness, synthetic accessibility, and similarity.
- 04
Optimize the pool
Multi-objective search keeps a diverse pool of high-scoring molecules and iterates within the oracle-call budget.
One benchmark, many ways to search
Oracle-guided, multi-objective optimization that rewards binding, drug-likeness, and synthesizability at once — not affinity alone.
Multi-objective by design
Balance binding, QED drug-likeness, and SA synthesizability together, so a high-affinity hit is also developable.
Four generative methods
ChemLlaMA genetic search, discrete diffusion (GenMol), Mamba RL (Saturn), and Genetic-GFlowNet — swap the search strategy, keep the benchmark.
Three benchmark tasks
Hit designs de novo, Lead optimizes from known actives under a similarity constraint, and Specificity rewards selective binding over antitargets.
Scalable docking
Dock inline or offload to an HTTP QuickVina service via DOCKING_VINA_URL to parallelize scoring across many candidates.
Open & reproducible
A shared benchmark package with fixed oracles, tasks, and call budgets — so results are comparable and reproducible. Apache-2.0.
Built on the PMO-Dock benchmark
ChemLlama stands on PMO-Dock from YerevaNN — a benchmark for protein-aware molecular optimization. A shared package defines the oracles (QED, SA, docking, similarity), the tasks, and the call budgets, and four generative methods plug into it to search chemical space for molecules that bind well, look drug-like, and are easy to synthesize.
- genetic_chemalacticaChemLlaMA + genetic pool
- genmoldiscrete diffusion, fragment vocab
- saturnMamba/RNN RL with replay
- genetic_gfngenetic algorithm + GFlowNet