Project 4: Reasoning in Biological Foundation Models

As foundation models increasingly transform biological data analysis, the integration of reasoning capabilities into these models represents a frontier for advancing scientific discovery. This Master’s thesis explores the development and evaluation of reasoning mechanisms in biological foundation models, drawing inspiration from techniques pioneered in large language models (LLMs), such as chain-of-thought prompting and reinforcement learning (RL). The project investigates how structured reasoning—through multi-step inference, intermediate representations, and feedback loops—can be adapted to biological contexts, including protein function prediction, gene regulatory inference, and multi-modal cellular understanding. By implementing chain-of-thought-style reasoning within biological models and training with RL to optimize stepwise interpretability and correctness, the thesis aims to bridge the gap between black-box prediction and transparent, causally grounded inference. Through a combination of synthetic benchmarks and real-world biological tasks, this work provides insights into the potential of reasoning-augmented foundation models to support hypothesis generation, mechanistic insight, and personalized biomedical applications.