Starling oral-bioavailability transfer model

Given two molecules (SMILES + study metadata) and molecule A's measured oral bioavailability, this model predicts whether oral-bioavailability behavior transfers from A to B — i.e. whether the two molecules behave similarly under the given study context. It is self-contained: the frozen encoders are bundled with the trained head, so it runs end-to-end on raw inputs.

Architecture

Per molecule (siamese — the same encoders + projections are applied to A and B; only the head is position-aware):

• Molecule encoder — ibm-research/MoLFormer-XL-both-10pct (MolFormer-XL), frozen: SMILES → mean-pooled token embedding → 768-d, then a 2-layer MLP (768→384→256).
• Metadata encoder — sentence-transformers/all-MiniLM-L6-v2 (MiniLM), frozen: each of the 7 metadata fields is embedded separately (mean-pooled, L2-normalized) → 384-d, then a learned per-field projection → 32-d (7×32 = 224-d total). A missing/empty field uses a learned per-field "missing" embedding instead of the text embedding, so absent metadata is handled gracefully and distinctly from any real value.
• Per molecule = [mol_mlp (256) | metadata (224)] = 480-d.

Pair head:

• Concatenate [z_A, z_B] → 960-d input.
• A pre-norm residual SwiGLU MLP (6 blocks, width 384, FFN 1024) → one logit.
• sigmoid(logit) = P(transfer). ~8M trainable params; encoders frozen.

Metadata fields (order matters)

molecule_name, species_or_population, dose, oral_exposure_mode, qualifying_conditions, comparator, extra_details

Pass a dict per molecule keyed by these names. Omit a key, or pass None/"", for a missing field — the model then uses its learned per-field "missing" embedding.

Usage

Install the runtime dependencies:

pip install torch transformers safetensors sentence-transformers

Load the model with trust_remote_code=True. The repository includes the custom model wrapper plus the frozen MolFormer and MiniLM encoders, so inference runs directly from raw SMILES and metadata.

from transformers import AutoModel

repo_id = "jiosephlee/starling-transfer-shared-eval-same-species-v2-no-source-value-small-lt10m-v1-best-knn-val1"
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True).eval()

out = model(
    smiles_a=["CC(=O)Oc1ccccc1C(=O)O"],          # molecule A (bioavailability known)
    smiles_b=["CCO"],                            # molecule B (candidate)
    metadata_a=[{"species_or_population": "human", "dose": "325 mg", "oral_exposure_mode": "tablet"}],
    metadata_b=[{"species_or_population": "human"}],   # missing fields are fine
    # no source_value argument for this variant
)
p_transfer = out.logits.sigmoid()
print(float(p_transfer[0]))

Inputs are batched lists of equal length. For each pair, pass:

• smiles_a: SMILES for molecule A.
• smiles_b: SMILES for molecule B.
• metadata_a: dicts for molecule A metadata.
• metadata_b: dicts for molecule B metadata.

This is the no-source-value variant, so do not pass source_value.

Supported metadata keys are:

[
    "molecule_name",
    "species_or_population",
    "dose",
    "oral_exposure_mode",
    "qualifying_conditions",
    "comparator",
    "extra_details",
]

You may omit unknown fields, or set them to None or "". The model has learned missing-field embeddings for those cases.

The output is a logit per pair. Use out.logits.sigmoid() to get P(transfer), where higher values mean the model predicts that oral-bioavailability behavior transfers from molecule A to molecule B under the provided study context.

Example batched inference:

out = model(
    smiles_a=["CCO", "CCN"],
    smiles_b=["CCC", "CCCl"],
    metadata_a=[
        {"species_or_population": "rat", "oral_exposure_mode": "oral gavage"},
        {"species_or_population": "human", "dose": "50 mg"},
    ],
    metadata_b=[
        {"species_or_population": "rat"},
        {"species_or_population": "human", "dose": "100 mg"},
    ],
)
probs = out.logits.sigmoid().tolist()

Training & performance

Trained on the same_species_v2 no-source-value shared-eval split. The frozen embeddings were precomputed once and the head was trained on top. The label is |value_A - value_B| thresholded. This no-source-value variant learns transfer from molecule structure and metadata only.

This upload is the checkpoint selected by record_knn_validation_1_macro_f1, not the checkpoint selected by assay-transfer validation macro-F1.

Checkpoint provenance:

• Universe: same_species_v2
• Lane: no_source_value
• Model: small_lt10m_v1
• Checkpoint directory: best_record_knn_validation_1_macro_f1
• Selection metric: record_knn_validation_1_macro_f1 = 0.704843
• Selection step: 1000

Record-KNN final evaluation on condition_key_v3_record_splits_hf/full_metadata:

• validation_1: macro-F1 0.699383, accuracy 0.733871
• validation_2: macro-F1 0.766394, accuracy 0.796000

The source run and full checkpoint comparison are included in checkpoint_metadata.json.