| --- |
| license: mit |
| --- |
| # ESMC Sparse Autoencoder Features Table |
|
|
| This dataset contains a Parquet table of the 16,384 features from the [ESMC-6B-sae-layer60-k64-codebook16384](https://huggingface.co/Biohub/ESMC-6B-sae-layer60-k64-codebook16384), that was used for analysis in the [ESMC paper](https://biohub.ai/papers/esm_protein.pdf) and to construct the [ESM Atlas](https://biohub.ai/esm/protein/atlas). This table provides descriptions of the precomputed features that can be activated through the spotlight SAE model, assisting users for downstream interpretation of the insights revealed by ESMC. |
|
|
| Download the table [here](https://biohub.ai/models/esmc#model-data). |
|
|
| The features descriptions are in the table below. |
|
|
| | Feature | Description | |
| |---|---| |
| | `feature_id` | The feature index. | |
| | `description` | Full description of the concept the feature activates on, produced by a multi-agent system based on activations of the feature in Swissprot. This description has 3 sections which are used to derive the `summary`, `activation_pattern`, and `exemplar_protein_families` columns. Note that these represent our best understanding of what each feature represents but are not necessarily comprehensive and may not capture nuances of features that have complex meaning or limited activation in Swissprot. | |
| | `summary` | Summary of the biological concept that is activated. | |
| | `activation_pattern` | The residue-level activation pattern of the feature with respect to the concepts described in the summary (ex. is the activation diffuse across an entire domain, or is it highly localized to one or more residues in specific locations). | |
| | `exemplar_protein_families` | Some of the common protein families/functions that the feature is found in. These are derived from the Swissprot data used to generate the description. These should not be considered exhaustive. | |
| | `threshold` | The activation threshold above which the feature description can be relied upon for accuracy. | |
| | `category` | Agent-generated categorization of the feature into 1 of 12 categories based on its description. Categories include things like "catalytic function", "compositional bias", and "structural motif". | |
| | `uniref90_frequency` | Frequency of activation of this feature in Uniref90 (total 208M proteins). | |
| | `uniref90_idf` | Inverse document frequency (IDF) of this feature in Uniref90 (derived from the frequency by `log(N / freq`), used for feature normalization. | |
| | `uniref90_max_activation` | Maximum activation value in Uniref90 (used alongside IDF for feature normalization). | |
| | `top_100_uniref_ids` | Top 100 proteins in Uniref90 by activation strength of this feature. | |
| | `top_swissprot_activations` | Top 100 proteins in a non-redundant subset of Swissprot (30% sequence identity clustered and filtered to high annotation score). | |
| | `decoder_nearest_neighbors` | 10 nearest neighbor features in the SAE decoder latent space, i.e. the most semantically similar features according to the model. | |
|
|
|
|
