Chemistry & Drug Discovery

1. Chemistry AI spans two domains: (a) LLM reasoning about chemical concepts, (b) specialized ML for molecular property prediction and drug discovery.
2. For LLM-based chemistry reasoning: ollama pull deepseek-r1:32b (~20 GB). Prompt: "Predict the product of this reaction: CH3COOH + C2H5OH --[H2SO4]--> ? Show the mechanism step by step."
3. For molecular property prediction: pip install deepchem (DeepChem — open-source library for drug discovery ML). Pre-trained models for solubility, toxicity, binding affinity prediction.
4. For protein-ligand docking: pip install diffdock (DiffDock — diffusion-based molecular docking). Feed a protein structure (PDB file) + ligand SMILES → predicts binding pose. First result in 1-5 minutes on GPU.
5. For AlphaFold-class protein structure prediction: pip install colabfold (ColabFold, local AlphaFold2). Feed an amino acid sequence → predicted 3D structure. A 300-residue protein predicts in 30-60 minutes on 12 GB GPU.
6. For molecular generation: SMILES-based LLMs or diffusion models generate novel molecules with desired properties.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs ColabFold (protein structure prediction) for proteins up to ~500 residues in 30-90 minutes. DiffDock for molecular docking in 1-3 minutes per ligand. DeepChem molecular property models train in minutes. For small-molecule drug discovery workflows: 12 GB handles most single-protein targets. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe (protein databases are large). Total: ~$400-480. $400 gets you into computational chemistry research — the same workflows that required university clusters 5 years ago.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs ColabFold for large proteins (1000+ residues) in 60-120 minutes. DiffDock batched docking (1,000 ligands) in 2-4 hours. For virtual screening campaigns (dock 1M compounds against a protein target): RTX 3090 handles ~10K ligands/hour with DiffDock. For Alphafold3 (cloud-only currently) replacement workflows, the GPU handles the compute when open implementations arrive. Total: ~$1,800-2,200. Dual RTX 3090 for parallel docking or training custom molecular models. Chemistry ML is compute-hungry for large-scale screening.

The mistake: Running an LLM for chemical calculations ("What's the molecular weight of aspirin?") without verifying the answer, then using it in a lab notebook. Why it fails: LLMs hallucinate molecular formulas. Asked for "aspirin molecular weight," the model might correctly say C9H8O4 (180.16 g/mol) — or it might give you the formula for ibuprofen (C13H18O2) or a completely fictional compound. The model doesn't compute molecular weight from atoms — it recalls a training example, which may be wrong. The fix: Use computational chemistry tools for calculations. RDKit (pip install rdkit) computes molecular weight, logP, rotatable bonds deterministically from SMILES: from rdkit import Chem; mol = Chem.MolFromSmiles("CC(=O)OC1=CC=CC=C1C(=O)O"); print(Chem.Descriptors.MolWt(mol)). LLMs for conceptual reasoning ("explain SN2 reaction mechanism"); RDKit/DeepChem for calculations. Deterministic tools for numbers, LLMs for concepts.