SecondSite models
SecondSite classifiers prioritize candidate allosteric protein–ligand interactions using fixed protein and molecular embeddings. This release contains 3,360 checkpoints covering deployment, the main benchmark, the 2015 time-split study, and the scaffold and protein-representation sensitivity experiments.
Project: https://secondsite.snu.ac.kr
Analysis code: https://github.com/memgoon/SecondSite
Which weights should I use?
- Ranking new pairs: use
deployment/. It contains four training cohorts, eight architectures and three seeds (96 checkpoints). These are the models used for the four-cohort ChEMBL screening, not the fold-specific benchmark models. - Reproducing held-out evaluation: use
evaluation/and the corresponding split assignments. A deployment model trained on the full cohort is not a substitute for the checkpoint that excluded a particular test fold. - Reproducing the time-split study: use
evaluation/time_split_2015/. These models were refitted on the eligible records dated through 2015 after epoch selection within the earlier data. They are distinct from deployment models.
Scores are prioritization scores, not calibrated probabilities or confirmation of binding, an allosteric site, or functional modulation.
Checkpoint inventory
| Experiment | Checkpoints | Role |
|---|---|---|
| Deployment | 96 | Four cohorts × eight models × three seeds |
| Main five-fold benchmark | 1,080 | Every-pair, Both-class proteins, and Both-class proteins & ligands; random, Pfam and ligand held-out |
| Explicit double-held-out | 240 | Every-pair and Both-class proteins; five folds |
| Both-class ligands benchmark | 480 | Four evaluation settings; five folds |
| Both-class proteins & ligands, double-held-out leave-one-family-out | 1,128 | 47 test families × eight models × three seeds |
| Pfam plus Murcko exclusion | 120 | Every-pair scaffold sensitivity |
| Protein representation comparison | 120 | Full protein sequence versus rechunked structure-matched sequence; protein-only and C1–C3 |
| Publication-year split at 2015 | 96 | Four cohorts × eight models × three seeds |
| Total | 3,360 |
The Both-class proteins & ligands cohort uses five folds for Pfam held-out evaluation and 47 leave-one-family-out folds for double held-out evaluation.
The deployment cohort sizes are 6,854 (Every-pair), 4,637 (Both-class proteins),
1,365 (Both-class ligands), and 395 (Both-class proteins & ligands).
Both-class cohorts retain proteins, ligands, or both represented by both
allosteric and orthosteric class labels. Folder names retain both_classed_*
for compatibility. The time-split cohorts use records dated through 2015.
Files and loading
model_index.tsv: checkpoint paths, experiment, cohort, setting, model, seed, fold, architecture file, config, size and SHA-256.config.jsonbeside each checkpoint: model construction and original metadata.FIT_REPORT.jsonbeside evaluation checkpoints: fit results and training/evaluation details.architectures/: exact model-definition source snapshots, identified by hash.splits/: frozen row-ID partition assignments, without raw database content.VALIDATION.json: validation counts and scope.SHA256SUMS: checksums for all listed release files.
Install requirements.txt, download the repository with huggingface_hub, and
load a selected checkpoint using its config:
from huggingface_hub import snapshot_download
folder = snapshot_download("memgoon/SecondSite-models")
From the downloaded directory:
from load_model import load_model
model = load_model("deployment/every_pair/c2/seed_20260817/deploy.pt")
# Or choose a particular evaluation checkpoint from model_index.tsv.
load_model loads with weights_only=True, uses the recorded architecture, and
requires an exact state-dictionary match. The checkpoint formats differ between
experiments; the loader handles both model_state_dict and state_dict.
The eight heads are ligand, protein, c1, c2, c3, d1, d2, and d3.
C1/D1 concatenate pooled representations; C2/D2 use one-way cross-attention;
C3/D3 use bidirectional cross-attention. C models use protein sequence inputs,
whereas D models use selected pocket residues. Except for the explicitly named
full-sequence sensitivity models, the sequence inputs represent the selected
structure-matched sequence, not necessarily the full canonical UniProt sequence.
The forward input is a dictionary of padded floating-point tensors and boolean masks (True for real residues/atoms, False for padding):
| Key | Shape |
|---|---|
| ligand | batch × ligand heavy atoms × 512 |
| ligand_mask | batch × ligand heavy atoms |
| protein | batch × protein residues × 1536 |
| protein_mask | batch × protein residues |
| batch × selected pocket residues × 1536 | |
| pocket_mask | batch × selected pocket residues |
Supply the fields used by the selected head. model(batch) returns logits;
torch.sigmoid(logits.float()) converts them to scores. Use the study's original
preprocessing and model-specific inference batching for exact reproduction.
Arbitrary sequences or SMILES cannot be passed directly to these classifiers.
Data and input preparation
This repository contains SecondSite weights, model definitions, configurations,
and split assignments. Join the row IDs in splits/ to the study pair tables
when reproducing an evaluation. Protein and ligand inputs are prepared using
ESM3 and Uni-Mol, respectively; see the analysis code for preprocessing.
License
SecondSite downstream weights and original code are released under MIT. This
does not relicense upstream datasets. See THIRD_PARTY_NOTICES.md for the ESM3
and Uni-Mol embedding-model provenance. Original foundation-model weights are
not bundled here.