InitialMoleculeGeneration
The AI agent's performance will be evaluated based on its ability to propose novel drug candidates that meet specific binding affinity, drug-likeness, and ADMET criteria for multi-target activity. The evaluation will focus on the integrated workflow, the quality of the final molecular proposals, and the effective use of tracking and search tools.
Evaluation overview
How the linked challenge is judged: tasks, benchmarks, and criteria count.
Task templates
Inputs and expected outputs.
InitialMoleculeGeneration
Using Llama 3.1 via Groq and Semantic Kernel, generate an initial set of 5-10 diverse small molecules that are plausible starting points for multi-target agonists (e.g., GLP-1R, GIPR, GCGR).
JSON object with target receptor names and desired activity profiles (e.g., {"targets": ["GLP-1R", "GIPR"], "affinity_target": "low nM"})
JSON array of initial SMILES strings and associated descriptors (e.g., molecular weight, cLogP).
PropertyPredictionAndScoring
For each generated molecule, predict key properties: binding affinity to specified receptors (simulated via QSAR/docking), Lipinski's rule compliance, and preliminary ADMET scores (e.g., hERG, solubility). Record these with Weights & Biases.
JSON array of SMILES strings from initial generation.
JSON array of molecules, each with SMILES, predicted binding affinities, Lipinski violations count, and ADMET scores.
OptimizationAndRefinement
Utilize Semantic Kernel with Llama 3.1 to iteratively refine the top 3-5 candidates from the previous stage, focusing on improving binding affinity, minimizing ADMET risks, and maintaining drug-likeness. Leverage Tavily for real-time information retrieval on similar compounds or protein-ligand interactions. Track all iterations using Weights & Biases.
JSON array of scored molecules from previous task.
JSON object representing the single best optimized drug candidate, including its SMILES, comprehensive predicted properties, and a detailed rationale for its selection and optimization pathway.