Introduction
Imagine an AI that can predict with eerie accuracy what video you'll watch next on TikTok—now imagine that same technology predicting how a molecule will behave inside your body. That's exactly what ByteDance's drug discovery unit, Anew Labs, is doing. By repurposing the machine learning models that power TikTok's recommendation engine, they're tackling diseases that pharmaceutical companies have long labeled 'undruggable.' This guide walks you through the step-by-step process of applying similar AI techniques to drug discovery, from understanding the problem to validating AI-designed therapies.
What You Need
- Access to biological data – Protein structures (from databases like PDB), molecular interaction datasets, and clinical trial results.
- AI/ML expertise – Familiarity with deep learning, recommendation algorithms, and predictive modeling frameworks (e.g., TensorFlow, PyTorch).
- High-performance computing – GPUs or cloud clusters to train large models on molecular data.
- Domain knowledge – Understanding of pharmacology, biochemistry, and disease biology (or close collaboration with experts).
- Experimental validation resources – Lab facilities or partnerships for in vitro/in vivo testing of AI-predicted candidates.
Step 1: Define the 'Undruggable' Problem
Start by identifying a disease target that conventional drug design has failed to address. These are often proteins with challenging structures—like flat surfaces, highly flexible regions, or deep binding pockets—that don't respond to small molecules. Research conditions like cancer pathways (e.g., RAS mutations), neurodegenerative disorders, or rare genetic diseases to pinpoint a specific protein target. Document the biological mechanism and existing failed approaches.
Step 2: Gather and Curate High-Quality Data
The AI needs to learn from vast amounts of molecular interaction data. Collect protein 3D structures from the Protein Data Bank (PDB), ligand-protein binding affinities from ChEMBL or BindingDB, and chemical libraries like ZINC. Clean the data by removing duplicates, filling missing values, and standardizing units. Annotate each data point with known activity levels (e.g., IC50 values) to create a training set for the AI. For targets with limited data, use transfer learning from related proteins.
Step 3: Design the AI Architecture – Borrow from Recommendation Systems
Recommendation algorithms (like TikTok's) use collaborative filtering and neural networks to map user-item interactions. For drug discovery, adapt this architecture: treat the protein as a 'user' and potential drug molecules as 'items.' Use graph neural networks to represent molecular graphs (atoms and bonds) and protein residues. Train the model to predict binding affinity by learning the interaction patterns between the protein's binding site and the molecule's features. Implement an attention mechanism to focus on key binding regions.
Step 4: Train the Model on Known Interactions
Split your curated dataset into training, validation, and test sets (e.g., 70/15/15). Use a loss function like mean squared error or cross-entropy for regression/classification. Train the model with mini-batches and learning rate scheduling. Monitor performance metrics (R², AUC-ROC) on the validation set. Avoid overfitting by employing dropout, weight decay, and early stopping. This stage may take days to weeks on powerful hardware.
Step 5: Screen Virtual Libraries Against Your Target
Once the AI is trained, input the 3D structure of your 'undruggable' target into the model. Run a virtual screening of millions of molecules from public or proprietary libraries (e.g., Enamine, DrugBank). The AI will rank compounds by predicted binding affinity, toxicity, and ADME properties (absorption, distribution, metabolism, excretion). Output a shortlist of top 100-500 candidates for further analysis.
Step 6: Validate Predictions with Bayesian Optimization
Use active learning loops: select a subset of top candidates, synthesize or purchase them, and test experimentally (e.g., SPR, ITC, cellular assays). Feed the results back into the AI to refine predictions. This iterative process—often called Bayesian optimization—quickly narrows down the most promising leads. Each round refines the model's understanding of the protein's druggability.
Step 7: Conduct Preclinical Testing
For the final few candidates, perform detailed pharmacokinetic studies in vitro and in vivo. Measure how the compound behaves in animal models: metabolism, half-life, bioavailability, and toxicity. If results are positive, progress to formulation and potential IND (Investigational New Drug) application. Document all steps and data for regulatory submission.
Step 8: Scale Up with Automated Labs
To speed up the cycle, integrate your AI with automated synthesis and high-throughput screening robots. Companies like Anew Labs have done this by creating 'self-driving labs' where AI designs the next experiment, a robot runs it, and results feed back into the model. This reduces the time from target selection to candidate nomination from years to months.
Tips for Success
- Collaborate widely: AI alone isn't enough—work with medicinal chemists, biologists, and clinical researchers to interpret predictions and design experiments.
- Embrace failure: Most AI-predicted compounds won't work in the first round. Treat early failures as data to improve your model.
- Focus on data quality: Garbage in, garbage out. Invest heavily in clean, well-annotated biological data.
- Stay updated: The field of AI drug discovery moves fast. Follow conferences like NeurIPS, ICML, and journals like Nature Machine Intelligence.
- Leverage pre-trained models: Build on architectures like AlphaFold or MolGPT instead of starting from scratch.
- Think regulatory early: Engage with FDA or EMA representatives to understand what data they expect for AI-generated candidates.
By following these steps, you can harness the same AI technology that powers personalized content feeds to potentially unlock treatments for the most challenging diseases. It's a journey that requires patience, interdisciplinary teamwork, and a willingness to iterate—but the rewards could be life-changing.