| # Choose and compare BERTose checkpoints |
|
|
| Use the [shared code](https://huggingface.co/supanthadey1/bertose-affinose-training-code) and [workbench notebook](https://huggingface.co/supanthadey1/bertose-affinose-training-code/blob/main/releases/2026-10-02/notebooks/BERTose_workbench.ipynb). The code version is `2026-10-02`; model assets remain under `releases/2026-09-28/`. |
|
|
| ## Sequence embeddings |
|
|
| | Checkpoint | Purpose | |
| |---|---| |
| | **Variant 3 — default** | BERTose reference used in the manuscript comparison; start here for glycan embeddings. | |
| | **Variants 1, 2, 4 and 5** | Compare the effect of different pretraining objective weights on the same inputs. | |
|
|
| All five variants used pretraining seed 42 and their own selected epoch. They are alternative weighting configurations, not independent seeds or models for different tasks. The reported comparison does not establish universally best weights. |
|
|
| Run from the code directory after following the [quickstart](https://huggingface.co/supanthadey1/bertose-affinose-training-code/blob/main/releases/2026-10-02/docs/QUICKSTART.md). Keep the same WURCS strings, sample order and pooling for each variant: |
|
|
| ```python |
| from bertose_workbench import Encoder, cosine_similarity |
| |
| wurcs = [ |
| "WURCS=2.0/1,1,0/[a2122h-1a_1-5]/1/", |
| "WURCS=2.0/1,2,1/[a2122h-1a_1-5]/1-1/a4-b1", |
| ] |
| reference_variant, comparison_variant = "3", "1" # comparison: "1", "2", "4" or "5" |
| vectors, hashes = {}, {} |
| for variant in (reference_variant, comparison_variant): |
| encoder = Encoder.from_pretrained(variant=variant, device="cpu") |
| vectors[variant] = encoder.encode(wurcs, batch_size=8, pooling="first") |
| hashes[variant] = encoder.metadata["sha256"] |
| |
| reference = cosine_similarity(vectors[reference_variant], vectors[reference_variant]) |
| comparison = cosine_similarity(vectors[comparison_variant], vectors[comparison_variant]) |
| print(reference, comparison, hashes) |
| ``` |
|
|
| Each matrix compares glycans within one variant. Compare those matrices or neighbor rankings; do not mix vectors from different variants in one search index or use cross-variant cosine scores without validating their alignment. Similarity describes representation geometry, not binding performance. |
|
|
| `first` pooling uses the first sequence token. `mean` averages attended tokens, including START and END. Keep pooling fixed when comparing vectors. Inputs must be WURCS 2.0 within the trained 256-token limit; use `encode_with_diagnostics` to inspect tokenizer warnings. |
|
|
| The loader checks the checkpoint against its recorded SHA-256. Save that hash and pooling with your exports. `from_pretrained` also accepts `revision="main"` (or a Hub commit) and `local_checkpoint="/path/to/sequence_encoder.pt"` for a local file. |
|
|
| ## Binding: choose a regime and seed |
|
|
| The nine fits use the same prediction interface. Regime names describe held-out evaluation splits: glycans, proteins, or both. Choose the intended regime and record the exact `cell_id`: |
|
|
| | Regime | Seed 42 | Seed 43 | Seed 44 | |
| |---|---|---|---| |
| | Glycan-cold | `mo_A_glycan_s42` | `moc_A_glycan_s43` | `moc_A_glycan_s44` | |
| | Protein-cold | `mo_B_protein_s42` | `mof_B_protein_s43` | `mof_B_protein_s44` | |
| | Doubly-cold | `mo_C_doubly_s42` | `mof_C_doubly_s43` | `mof_C_doubly_s44` | |
|
|
| ```bash |
| python download_assets.py --task binding --cell-id mo_A_glycan_s42 --output-dir assets |
| ``` |
|
|
| Use the paths in `assets/paths_binding.json` with the [binding instructions](https://huggingface.co/supanthadey1/bertose-affinose-training-code/blob/main/releases/2026-10-02/docs/QUICKSTART.md#binding-from-cached-inputs). Prediction requires pairs represented in the supplied glycan/protein cache. The adapter does not generate features for new glycans or protein FASTA. Outputs are regression scores, not calibrated probabilities or affinity constants. |
|
|
| For a backend, keep the selected model loaded and reuse a `CachedInputs` context across batches. The existing `load_model`, `read_pairs` and `predict_batch` functions provide the same preprocessing and prediction used by the command. |
|
|
| ## IAR and website reuse |
|
|
| IAR needs the **full reference checkpoint** and fitted linkage readout. Sequence encoder exports have no trained masked-token output head and cannot substitute for that checkpoint. `python download_assets.py --task iar --output-dir assets` retrieves it. The supplied command runs only the two recorded single-linkage cases; general uncertain-glycan input still needs an adapter and validation. |
|
|
| The collaborator can reuse `Encoder.encode_with_diagnostics`, `cosine_similarity`, `nearest_neighbors` and `pca` for embedding, similarity and visualization tools. The notebook shows embedding exports. Use the existing binding API and IAR command for their supported inputs; no replacement inference engine is needed. Follow [WEBSITE_INTEGRATION.md](https://huggingface.co/supanthadey1/bertose-affinose-training-code/blob/main/releases/2026-10-02/docs/WEBSITE_INTEGRATION.md) for backend reuse, output labels and deployed checkpoint identity. |
|
|
