CovalentAgent

The problem

Most AI-driven drug discovery tooling targets non-covalent binding. Covalent drugs — which form permanent bonds with their target proteins — represent a growing therapeutic class that includes blockbusters like osimertinib (Tagrisso) and sotorasib (Lumakras). Yet the open-source tooling for AI-powered covalent drug design is nearly nonexistent.

CovalentAgent is the first open-source attempt to fill that gap.

The approach

A LangGraph supervisor coordinates six specialist agents that mirror the covalent drug discovery workflow:

Agent	Responsibility	Models / tools
TargetAnalyst	Identify reactive residues, score ligandability	ESM-2 protein language model
WarheadSelector	Choose optimal warhead class for residue + pocket	Reactivity models + selection rules
MoleculeDesigner	Scaffold generation with warhead attachment	RDKit + fragment-based design
PropertyPredictor	ADMET, selectivity, drug-likeness	Chemprop message-passing NN
LiteratureRAG	Citation-backed rationale	ChromaDB over CovPDB + PubMed
Reporter	Ranked candidate report with provenance	Structured Pydantic output

Every design decision is grounded in literature. Every property prediction is reproducible. Every output ships with the SMILES, scores, and rationale needed for a chemist to act on it.

Data sources

CovPDB — Covalent protein-ligand complex database
CysDB — Cysteine chemoproteomics annotations
ChEMBL — Bioactivity data for covalent inhibitors
PDB — Protein structures
PubMed — Covalent drug design literature

Why this matters

Covalent drug design is a good test case for foundation biology models. A protein language model like ESM-2 can point at a reactive cysteine, but reactivity alone doesn’t tell you if it’s druggable. The interesting engineering is composing multiple signals (sequence, chemistry, structure, literature) into a ranking that a medicinal chemist can actually act on. CovalentAgent is my attempt at that composition.

Try it live

Try CovalentAgent

Pick a UniProt ID and get a ranked list of reactive cysteines. Needs a Python GPU backend — coming once the Modal sandbox lands.

live

UniProt ID (e.g. P01116 for KRAS)

Try:

Run it locally

Every project is open source. Clone the repo, install dependencies with uv sync, drop your API keys into .env, and run the CLI. The README walks you through each step.

git clone https://github.com/deepmind11/CovalentAgent.git
cd CovalentAgent
uv sync
cp .env.example .env  # add your keys
uv run python -m covalentagent --help