“The aim of drug discovery is to decode the molecular puzzles of disease and deliver targeted cures; the ideal of medicine is to bring those therapeutics to patients in weeks, not decades.” — Adapted from Gertrude B. Elion, Nobel Laureate in Physiology/Medicine
Project Lumina is an agent-assisted computer-aided drug design (CADD) application designed to streamline the virtual screening pipeline for drug scientists. By connecting conversational AI agents directly to established bioinformatics databases and local QSAR heuristic scoring engines, Lumina automates the tedious, manual pipeline of target alignment, structural clean-up, ligand screening, and feasibility reporting into a single, unified execution run.
As pharmacy students, my colleagues and I took classes on drug discovery and development and noticed a sobering reality: bringing a new therapeutic to market takes several years (often 10–12 years) and astronomical financial investments (averaging $2.6 billion).
As we saw the incredible progress in modern artificial intelligence, we wondered: how can we leverage the technology of today to help support the scientists fighting on the frontlines of human health?
We realized that a massive bottleneck in early-stage drug discovery isn't just the wet-lab testing, but the dry-lab setup. Drug scientists spend weeks performing manual, repetitive tasks: searching UniProt, downloading PDB coordinates, stripping out crystallographic noise, configuring local command-line docking scripts, and referencing chemical similarity tables. These are brilliant researchers whose time should be spent interpreting data and designing therapeutics, not writing glue-code.
We built Lumina to solve this. Lumina acts as an intelligent agentic partner that stands beside the drug scientist, connecting natural language instructions directly to established bioinformatics tools to run virtual screening campaigns in seconds.
Lumina replaces a disconnected, multi-week computational pipeline with an autonomous agentic virtual screen. Instead of writing custom scripts and visiting five different portals, a drug scientist can simply describe their goal in plain English (e.g., "Find me 20 ligands that I can test against 6lu7").
Behind the scenes, Lumina automates the entire process:
- Target Identification: Instantly queries UniProt to retrieve sequence data, gene definitions, and host organism metadata.
- PDB Homology Alignment: Runs sequence alignments across the Protein Data Bank (RCSB PDB) to discover homologous structural templates that contain co-crystallized small molecules.
- Crystallographic Noise Filtration: Automatically parses active sites and strips away crystallization buffers, solvents, and salts (such as water, glycerol, DMSO, PEG, and sodium) that would otherwise skew binding predictions.
- Bioactive Database Expansion: If the template structures contain fewer binders than the scientist's target count, Lumina performs a live similarity database search against ChEMBL and PubChem to pull down structurally related, tested analogs .
-
QSAR Affinity Scoring: Scores target binding affinity (
$pK_d$ ) using a hybrid sequence-weighted QSAR model based on Lipinski rule compliance, molecular weight, lipophilicity (LogP), and structural features (such as pi-stacking aromatic rings and halogens). - Feasibility Dossiers: Generates a formal CADD report detailing residue-level interactions, safety warnings, and assay verification guidelines, along with an exportable CSV report listing candidates ranked by binding index.
Lumina is engineered using a modern, type-safe fullstack architecture:
- Frontend & Routing: Next.js 16 (App Router) combined with React 19 and TypeScript 5.
- Bioinformatics APIs: Direct integrations with the RCSB PDB APIs (Search, Entry, Polymer, Non-Polymer, and ChemComp) and the UniProt REST API, with robust local database caching.
-
Large Language Model (LLM) Agent: Google Gemini API (utilizing
gemini-3.1-pro-previewmodels) via a Vertex AI. Gemini handles goal parsing, virtual screening planning, feasibility analysis, and similarity search queries. -
Scientific Math:
-
KaTeX: Renders LaTeX mathematical formulas (thermodynamic values like
$K_d$ ,$-\log K_d$ ,$\Delta G$ , and binding coefficients) within the AI interaction reports.
-
KaTeX: Renders LaTeX mathematical formulas (thermodynamic values like
-
UI Components & Styling:
- Shadcn UI & Base UI React: Reusable component primitives and structures.
- Tailwind CSS 4: Utility classes for premium styling, animations, and dark mode variables.
- Lucide React: Icon sets.
Initially, our team set out to build a machine learning neural network model from scratch to predict binding affinities. However, we quickly encountered a massive hurdle: none of us had deep machine learning model-training expertise, and training a complex structural graph neural network (GNN) from scratch requires massive compute resources, curated datasets, and weeks of training.
We paused, did further research, and realized a breakthrough: Google's pre-trained foundation models (like Gemini) already possess a deep, latent understanding of biochemical structures, sequence homologies, Lipinski rules, and QSAR principles.
Instead of building a poor ML model from scratch, we decided to pivot. We focused our engineering on building a robust agentic workflow around Google’s foundation models. The challenge shifted to:
- Parsing free-form text inputs into structured, reproducible execution plans.
- Handling real-time stream data connections (Server-Sent Events) from biological databases to feed the agent.
- Synchronizing the deterministic calculations of molecular descriptors (Lipinski scores, LogP) with the generative reasoning of the LLM.
This pivot allowed us to build a tool that is highly reliable, explainable, and infinitely more flexible than a simple black-box neural network.
We learned that in modern software engineering, orchestration is often more powerful than training from scratch. By leveraging the pre-existing intelligence of foundation models and wrapping them in strict, domain-specific software contracts (like our bioinformatics parser and scoring engine), we could build a clinical-grade CADD workflow in a fraction of the time. We also learned how to implement real-time data streaming to keep the user interface responsive during heavy database alignments.
Looking ahead, we want to give drug scientists the ability to further customize the agent to suit their unique laboratory workflows:
- Custom Tool Integration: Allow scientists to connect their own computational tools (such as local AutoDock Vina instances or molecular dynamics engines) that the agent can spin up and call.
- Private Library Mining: Enable scientists to upload their own in-house chemical databases or vendor-specific catalogs (e.g., ZINC subset files) for the agent to screen against.
- Multi-Agent Collaboration: Let the main screening agent spin up specialized sub-agents (e.g., a "Toxicity Agent" or "Synthesis Planner Agent") to run safety audits on the top-ranking hits.
Our ultimate vision is to empower drug scientists to build more world-changing drugs in less time and at fewer costs, using agentic AI to bridge the gap between biological databases and molecular breakthroughs.