| Field | Response |
| :------------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------- |
| Intended Task/Domain: | Molecular property prediction for drug discovery — specifically pretrained molecular representation learning intended for downstream multi-task ADMET (absorption, distribution, metabolism, excretion, toxicity) property prediction. |
| Model Type: | Transformer-based graph neural network (graph transformer with local atom/bond message passing and global self-attention), pretrained with a joint probabilistic objective combining SMILES reconstruction, contrastive discrimination, and chemistry-specific self-supervision. |
| Intended Users: | Computational chemistry and machine-learning researchers in drug discovery, particularly those working on ADMET / DMPK property prediction. |
| Output: | Molecular embeddings (float tensors representing learned molecular representations); when downstream task-specific heads are present, continuous-valued scalar predictions per ADMET endpoint. |
| Describe how the model works: | A SMILES string is parsed into a 2D atom-and-bond graph via RDKit and encoded by a graph-transformer (local message passing combined with global self-attention) into atom-level and bond-level representations. A learned projection maps a pooled encoder readout to a Gaussian latent distribution. During pretraining, four log-probability factors are jointly maximized: (i) SMILES reconstruction from a sampled latent via a transformer decoder with rotary positional encoding; (ii) symmetric latent-density regularization (the A-MIM term); (iii) a binary auxiliary variable that classifies each input as matched (with its own latent embedding) vs mismatched, with the negative expectation estimated implicitly from in-batch samples; and (iv) chemistry-specific vocabulary-prediction targets (atom-context, bond-context, functional group). At inference time only the encoder is required; the decoder and pretraining-only heads are discarded. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
| Technical Limitations & Mitigation: | (1) Predictions are statistical estimates and should not be used as substitutes for experimental measurement in safety-critical drug development decisions. (2) The pretrained encoder's chemical coverage is biased toward drug-like commercially-available compounds and bioactive molecules; predictions on molecules far outside this distribution may be unreliable. Mitigation: monitor latent-space neighborhood-distance signals (e.g., nearest-neighbor distance to training molecules) when applying to new chemical space. (3) Downstream ADMET-prediction performance is benchmarked on Biogen, ExpansionRX, and ChEMBL-MT; generalization to other ADMET assay protocols or to non-ADMET endpoints has not been characterized. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Mean Absolute Error (MAE), Pearson correlation coefficient (r), and Spearman correlation coefficient (ρ), reported per ADMET endpoint and averaged across multiple random seeds. |
| Potential Known Risks: | The model may provide inaccurate predictions of molecules for drug design candidates. Predictions made by this model are statistical estimates and should not be used as substitutes for experimental measurement in safety-critical drug development decisions.<br><br>Users are responsible for ensuring the physical properties of model-generated molecules and the predicted ADMET properties are appropriately evaluated and that any downstream decisions comply with applicable safety regulations and ethical standards. |
| Licensing: | Apache License, Version 2.0. See `LICENSE` in the source repository at https://github.com/NVIDIA-BioNeMo/KERMT. |
|
|
