Artificial intelligence is already transforming disciplines like physics, economics, and linguistics. So why does applying AI to biology seem more difficult, slower, and at times more frustrating?
In this post, I explore five key reasons that explain this challenge — and why, paradoxically, AI may be exactly what biology needs most.
1. Biology Has No “Master Equations”
In physics, fundamental laws like Newton’s F = ma allow us to model systems with precision. These equations offer clear frameworks for prediction.
Biology, however, doesn’t operate under such universal rules. Biological systems are context-sensitive, adaptive, and shaped by millions of years of evolution. There is no single equation that captures how a liver regenerates, how protein folds, or how neurons encode meaning.
2. Data in Biology Is Messy and Sparse
Unlike physics experiments, biological measurements often vary across labs, individuals, and contexts. Omics data (like transcriptomics or proteomics) can have thousands of features but few consistent labels. Labeling protein function or cell type is expensive and error-prone.
3. The Ground Truth Is Uncertain
Often, biology lacks clear ground truths. For example, even experts may disagree on the classification of certain diseases or protein interactions. This makes supervised learning tricky and complicates model evaluation.
4. Interpretability Is Critical
In many fields, a black-box model is acceptable if it gets good performance. In biology and medicine, interpretability is essential — scientists and clinicians want to understand why a model makes a decision, not just the decision itself.
5. Biological Systems Are Dynamical
Cells adapt. Proteins interact in networks. Feedback loops abound. Unlike a static object under gravity, biological systems change in response to inputs. Capturing these dynamics requires moving beyond simple static datasets.
Despite these challenges, machine learning — especially deep learning and generative models — is offering new ways to discover patterns in biological data. But it also invites us to rethink how we know in science.
That’s why I believe that…
Epistemology meets AI in biology.
Thinking about how we know what we know — with code, models, and molecules.
Reference
This post was inspired by the talk:
Jean-Philippe Vert (2024). Foundation models for biology, across scales. View on YouTube