ALCHEMY

The Problem: Traditional drug discovery takes over a decade and costs $2B+, with a devastating 90% clinical failure rate. The core bottleneck: disconnected data silos and the inability to rapidly predict drug-target binding while simultaneously auditing for human toxicity. The Solution — ALCHEMY: We built ALCHEMY, an agentic operating system that mimics a real-world pharmaceutical lab. Enter a disease (e.g., "Alzheimer's disease"), and ALCHEMY autonomously orchestrates the entire discovery pipeline: Target Agent: Queries UniProt to identify the exact protein sequences responsible for the disease. Repurposing Agent: Scans a FAISS vector database of 1,000+ FDA-approved drugs in milliseconds using custom 1280-dimensional protein embeddings. PharmaDuel (The Swarm): Two autonomous AI agents — a BioChemist (proposer) and a Toxicologist (critic) — debate drug safety and efficacy in real-time, aggressively filtering out toxic candidates through adversarial reasoning. Hardware & AMD Integration: This project pushes the limits of AMD Instinct MI300X GPUs and ROCm 6.1: Native Fine-Tuning (Track 2): We leveraged the 192GB VRAM of the MI300X to fine-tune a 650M parameter ESM2 protein language model. To overcome NaN gradient collapses common in protein transformers, we bypassed mixed precision and trained entirely in FP32 — achieving 77.6% validation accuracy with stable exponential loss decay. Agentic Workflows (Track 1): The PharmaDuel orchestration relies on Qwen2.5-72B running locally on AMD hardware for blazing-fast reasoning and debate generation. Vector Embeddings: The entire drug database was re-embedded into a 1280-dimension vector space using our fine-tuned model directly on the MI300X. ALCHEMY proves that with AMD's cloud infrastructure, a small team can compress the full computational pipeline of a pharmaceutical company into an interactive, real-time dashboard — accelerating drug discovery from years to minutes.