Helico masked-contact model, exp22
Full-size 368.5M-parameter Helico initialized from Protenix base_default_v1.0.0, fine-tuned for 20,000 optimizer updates and 640,000 weighted crop draws on the pinned Protenix-v1 PDB split. This release contains regular and exponential-moving-average weights; optimizer state remains in the durable training archive.
The model preserves MSA input and adds three-state unknown/no-contact/contact pair conditioning plus a symmetric contact-prediction head. Contacts use minimum token heavy-atom distance <5 Å with conservative observed-atom validity. Absorbing corruption is value-independent, with endpoint/sparse-mask oversampling and 20% MSA dropout. Coordinate diffusion and distogram parameters are fine-tuned jointly. Confidence and dummy-template parameters remain frozen; real template retrieval is disabled.
Training uses 384-token crops, effective batch 32, four diffusion samples/crop, one trunk cycle and up to 512 MSA rows. No curated nonbinding or extra pMHC/TCR examples were added. Negative contact labels describe observed geometry, not experimental nonbinding.
Evaluation and intended use
Experiment 23 is complete: all measured results, official FoldBench scores and candidate predictions. The original 32-crop contact BCE was 0.019414, but the completed evaluation provides more informative structure, calibration and control measurements.
On 384 full validation complexes, EMA confidence-selected observed-atom lDDT is 0.799787 versus 0.787484 for the initialization, a difference of +0.012303 (95% interval +0.008583 to +0.016490). Interface improvement is unresolved. Contact AP with MSA is 0.9066 within proteins, 0.4847 between proteins and 0.5345 for protein–ligand pairs. Official FoldBench protein–ligand success is 44.44% versus initialization 43.91%; protein–protein is 56.99% versus 58.06%; protein monomer lDDT is 0.82783 versus 0.82891. None of these differences is resolved. Protein–DNA success decreases by 3.64 percentage points (nominal 95% interval −7.32 to −0.29). Fixed-correspondence validation metrics are distinct from official FoldBench scores; the linked release contains all nine categories with complete failure denominators.
Sparse supplied contacts do not yet reliably control generated structures. Oracle requests recover only 3.125% intraprotein, 3.125% interprotein and 6.25% protein–ligand missed contacts, with no resolved positive gain over repeated unconditioned inference. Contact-only diffusion and discrepancy search do not establish a useful matched-compute advantage. This checkpoint supports continued research in contact prediction and structure generation; it has not demonstrated a general controlled conformational ensemble model or binding specificity. No nonbinding, Flock, pMHC or TCR training additions were used.
Required inference correction
The sampler's rigid coordinate rotation previously ran under BF16 autocast, distorting internal distances. The corrected implementation preserves coordinate precision; weights are unchanged. Use source containing fix 5806c0f, preferably the complete evaluation source below. Earlier structure-generation evaluation outputs are superseded. Every released EMA tensor exactly matches the training checkpoint, but identical inference seeds do not guarantee identical floating-point predictions.
Reproduce one prediction
Use source d764b577019b and its CUDA dependencies:
hf download timodonnell/helico-masked-contact-exp22 ema.pt --revision 01a6025408d078aa8305f302d1d664a0291c45ee --local-dir /tmp/helico-model
hf download timodonnell/helico-contact-eval-v1 7qeh.pt.gz --type dataset --revision ec4de019de89f79f03da31c5e026ddf7d858e77e --local-dir /tmp/helico-input
python experiments/exp23_contact-evaluation/predict_released.py --checkpoint /tmp/helico-model/ema.pt --features /tmp/helico-input/7qeh.pt.gz --output /tmp/helico-prediction.pt
The output contains coordinates, contact probabilities and confidence, with a JSON checksum/settings receipt. Native coordinates, labels and observation masks are removed before prediction. This example uses a prepared full input with MSA and unknown contacts; arbitrary sequences/ligands require the full input-preparation pipeline. See the evaluation notebook for settings and limitations.
Provenance
- Dataset: timodonnell/helico-protenix-v1-pdb@164e41cf98d7fdf4d437192a9947138ca2464011
- Data lock SHA256: 2855c34acc97c8990eb49268f0f17a35f014f76fcd1fac9195dbdd9aaa699114
- Training source: Open-Athena/helico@46cffd6939c9ea17d2103d5b8f5d79ba9e35444d
- Full training checkpoint SHA256: 355788e40f610c8cc662e4c10f8225204668dfdcaf92ee1f04d744f8d5ea8b9a
- Training: https://github.com/Open-Athena/helico/issues/22
- Evaluation: https://github.com/Open-Athena/helico/issues/23
- W&B: https://wandb.ai/timodonnell/helico/runs/exp22-contact-scale-v11
Load the saved model_config into HelicoConfig and model_state_dict into Helico from the linked source. The files are trusted PyTorch checkpoints and include configuration/provenance dictionaries. Omitted conditioning uses the learned all-unknown embedding. Raw and EMA weights must be selected explicitly when comparing results.