Overview
| Name | KERMT · aliases: KERMT, Contrastive KERMT, NV-KERMT-70M-v2 |
|---|---|
| Full name | Kinetic GROVER Multi-Task |
| Ambiguous names (not auto-linked) | None |
| Family | v1 and the contrastive v2.0 are two papers with separate weights; results must not be carried across
|
| Organization | Merck & Co. (modeling and informatics), NVIDIA (BioNeMo team) |
| Organization type | Mixed |
| Country | United States |
| Model class | Property prediction (graph-transformer pretraining + multitask fine-tuning) |
| Domain | Small molecules |
| First release | † partly verified · paper revision Basis: First preprint (arXiv 2510.12719) Note: The arXiv id and date are carried over from the site record and were not re-fetched this round source |
| Other milestones | |
| Latest update | · new version Basis: GitHub release v2.0.0 (contrastive version) Contrastive v2.0: code release v2.0.0, v2 weights uploaded 2026-06-09 with a matching preprint (2026-06-09); commits through 2026-09-15 are documentation and tooling maintenance source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · New version, new weights and a new preprint within 12 months |
| Openness | Code: Open · Apache-2.0 Weights: Research use only · NVIDIA Open Model License (v2.0 weights) Code is open; v2.0 weights are offered on a model hosting platform under an open model licence, while the pharma company's own page states no licence |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | None recorded |
| Unverified fields | † first_release, release_milestones.paper_v1, release_milestones.first_public_evidence |
| As of | 2026-11-16 |
What it does
Adapts a third-party graph-transformer chemical pretraining architecture - atom and bond message passing plus a transformer - pretrained on tens of millions of molecules from two public compound libraries, with the emphasis on multitask fine-tuning and hardware-side acceleration and memory optimisation. It is positioned as a production ADMET model inside a pharma company, compared against the multitask variant of Chemprop and sitting on the same pretrained-ADMET track as MiniMol and MolE.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Merck internal ADMET (30 endpoints); Public literature ADMET (25 endpoints) and Biogen set
Other evaluations (outside the 44)
- In-house pharma ADMET endpoint set (time split) (Own task set) Self-reported: A time-split evaluation over 30 endpoints and about 800k compounds against a multitask graph network and another pretrained model; the data are not public and cannot be reproduced externally
- Own public-literature ADMET set and a third-party public set (Own task set) Self-reported: 25 public endpoints plus a six-endpoint public set; in the same paper these two reverse the ordering seen on the in-house set, so the ranking depends on the setting
- Similarity-binned generalisation analysis (Comparison against baselines only) Self-reported: Comparison by chemical-similarity bin, with the authors reporting an advantage over the multitask graph-network baseline in every bin; all self-reported
Developers report: multitask fine-tuning clearly beats non-pretrained graph-network baselines, with the advantage growing at larger data sizes, plus an accelerated implementation.
Independent evidence: No independent evidence found. No site-benchmark pairing and no third-party evaluation found; the key results rest on non-public in-house data
Training data
Group in Figure 7: Private or licensed data.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 25 public ADMET endpoints with 114,112 compounds plus another public set of 6 endpoints and 3,521 compounds; pretraining on 11 million compounds from two public libraries source |
| b. Private or licensed data | Yes | In-house pharma ADMET data: 30 endpoints, 800,733 compounds measured through 2024, not public source |
| c. Own wet-lab data used for training | No | The in-house data are the company's historical measurements, not generated for this model source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: A modelling and informatics department at a large US pharma company co-authoring with the biological-models team of a GPU company; compute and funding come from both companies' internal R&D budgets, and the model is also listed in the GPU company's open-model catalogue.
- Funding: Not disclosed
- Products and partnerships: The architecture comes from an open graph-transformer pretraining model published by a Chinese internet company's lab; the code is open under the GPU company's digital-biology GitHub organisation and the pharma company published a corporate story page (2026-03-19).
- Related models: Chemprop v2, MiniMol, MolE, GenMol
Reading notes
- The most informative results come from the in-house dataset, which is not public and cannot be reproduced or checked externally.
- There are no site-benchmark results, so it cannot be compared directly with ADMET-AI or MiniMol.
- The advantage over graph-network baselines holds only in the multitask setting on the in-house assays; the same paper's public and third-party sets reverse the order, and all of it is self-reported.
- The TDC pairing does not hold: that benchmark does not appear in the paper.
Caveats
- Zero pairings: no results on any site benchmark.
- The key comparison depends on non-public in-house pharma data.
- v1 and the contrastive v2.0 are different weights; results cannot be carried across.
- The first-release date is carried over from the site record and was not re-fetched this round.
Sources
- First preprint (preprint)
- First preprint full text (preprint)
- Contrastive-version preprint (preprint)
- Code release list (repository)
- v2 weights page (package registry)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| benchmarks_reported | Candidate pairing TDC | Not a pairing: the benchmark does not appear in the paper; the actual benchmarks are an internal set and two public sets | source | 2026-11-16 |
| latest_update | 2025-10 preprint | 2026-06-10 contrastive-learning version v2.0 (code, weights and new preprint) | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.