What to build:

A well-documented ML solution to our problems, paired with a team presentation! Problems are gonna be presented during the open ceremony, and are listed in the official Pharmahacks 2026 Discord server in the #challenges channel. There are also listed below:

🧪 Challenge 1: Predicting Acute Toxicity (LD50),  Beginner

Can you predict how toxic a chemical compound is just from its molecular structure? In this challenge, you'll build machine learning models to predict the acute toxicity (LD50) of compounds using their SMILES representations. The dataset contains ~7,400 molecules with experimentally measured LD50 values, and your goal is a regression model that predicts log-transformed LD50 as accurately as possible. Start with molecular fingerprints and a simple baseline, then push further with richer descriptors, model ensembles, and interpretability analysis. 

🧬 Challenge 2: Compound-Target Binding Affinity, Advanced (Sponsored by Molecular Forecaster Inc.)

Given a drug compound and a protein target, can you predict how strongly they'll bind? This challenge provides ~421K drug-target interactions with experimental pIC50 measurements. You'll need to decide how to encode both SMILES strings and amino acid sequences into meaningful representations, then train a single protein-agnostic model that generalizes across targets. Your model will be evaluated on four test splits: warm, cold compound, cold target, and full cold.

🧠 Challenge 3: Detecting Dementia from EEG, Advanced

Can you detect Alzheimer's disease or frontotemporal dementia from brain activity alone? You'll work with EEG recordings from 88 subjects (19 channels, 500 Hz) and tackle a binary classification task: AD vs healthy or FTD vs healthy. The recordings vary in length, so you'll need to handle feature extraction (e.g., relative band power, spectral coherence) or padding, and aggregate epoch-level predictions into a single diagnosis per subject. This one demands comfort with signal processing and careful validation on a small dataset.