Aya Abdelsalam Ismail

Concept Bottleneck Language Models For protein design We introduce Concept Bottleneck Protein Language Models (CB-pLM), a generative masked language model with a layer where each neuron corresponds to an interpretable concept. Our architecture offers thr...

[8/n] More details in the paper: arxiv.org/abs/2411.06090 and our blog post ncfrey.substack.com/p/building-t....

You can also find the code on GitHub: github.com/prescient-de... and model weights on 🤗.

[7/n] Our architecture allows us to know what concepts the model learned and which concepts the model uses during inference by inspecting the weights of the final linear layer; this offers a way to debug and asses the model's quality.

[6/n] Interpretability: The concept bottleneck can be used to understand which concept the model uses to predict a certain amino acid. Reliably controlling model behavior: The concepts can be used as knobs to control the model's output.

[5/n] We train a mask language model with up to 3 Billion parameters with a layer that directly encodes biophysical and biochemical concepts that biologists care about. These models match the performance of unconstrained masked language model.

[4/n] Joint work with @ncfrey.bsky.social Tuomas Oikarinen @amywang1.bsky.social @juliusad.bsky.social @activelearner.bsky.social Taylor Joren @allen.bsky.social Hector Corrada Bravo @kyunghyuncho.bsky.social and Joseph Kleinhenz in @prescientdesign.bsky.social

[3/n] In our concept bottleneck protein language model paper, we show that we can train the model with billions of parameters, with interpretability constraints, without performance degradation.

[2/n] But the thing is, more often than not, we know beforehand what we want/expect our model to learn, especially in very well-studied domains like Biology. So, instead of playing the guessing game, we trained a model that explicitly learns different concepts that biologists care about.

🧵
[1/n] Does AlphaFold3 "know" biophysics and the physics of protein folding? Are protein language models (pLMs) learning coevolutionary patterns? You can try to guess the answer to these questions using mechanistic interpretability.