Can you model biology mechanistically?
That’s the title of blog post by @jesse johnson, Scaling Biotech. I believe, with characteristic humility, that the question is ill-posed, even though the framework the author proposes can be useful in distinguishing between different types of modeling approaches, which can sometimes be combined (from the post):
- Mechanistic: These are models that try to directly model what’s happening with the biology. They may use experimental data to define or tune the model, but typically don’t use experimental data collected for the immediate question.
- Black box: These are models that look for correlations in data, often from one or a small number of experiments designed to address the immediate question. They typically don’t model the biology or use external data in a significant way.
- Knowledge: These are models that directly use curated information extracted from structured resources like ChEMBL, or from unstructured sources like research papers. They may model biological mechanisms to some extent, but it’s usually very limited. And they rarely use experimental data.
The reason the question is ill-posed is that, at a high level, building a “model” is driven by a purpose and the data available, which constrain the level of description. Therefore, depending on these two drivers, certain biological phenomena are definitely amenable to mechanistic modeling while others are not.
In particular, Large Language “Models” (LLMs) are indeed models whose purpose, to simplify, is to predict masked words or sentences, and the data, well, anything written or recorded language data. “Foundation Models” send us down a slippery slope because they claim to (and sometimes do) perform far outside of the purpose they were trained for. When models are applied outside of the domain for which they are built, appropriate caution is warranted -for example, is the structure of knowledge in one domain similar to another domain’s?
So, to once again quote the famous Box (George, not the black box): “All models are wrong, but some are useful”, and usefulness is defined with respect to a purpose, a task, an objective, and does not exist in a vacuum. I would suggest, therefore that a more constructive question for biology (or any modeling endeavor) is what kinds of model are appropriate for what purposes.