CheMeleon

Deep learning has had a hard time displacing classical machine learning in molecular property prediction. On the datasets that actually govern discovery programs — a few hundred to a few thousand measured compounds, with real assay noise — a Random Forest over expert-crafted descriptors routinely matches or beats a graph neural network trained from scratch. CheMeleon attacks that gap from the pretraining side. Instead of learning from scarce experimental labels or from quantum-mechanical calculations that carry their own systematic biases, it learns to reproduce classical molecular descriptors, which are deterministic algorithms over the molecular graph and therefore essentially noise-free.

The model was developed by researchers in the Department of Chemical Engineering at MIT, led by Jackson Burns in William Green's group, with collaborators at BASF, and posted as a preprint in June 2025. It pretrains a directed message-passing neural network (D-MPNN) — the encoder used by the Chemprop package — on one million molecules randomly sampled from PubChem, regressing onto the 1613 descriptors computed for each molecule by Mordred.

That choice puts CheMeleon in a different family from the SMILES language models such as MoLFormer-XL and from graph models pretrained with contrastive or motif-prediction objectives. Where those learn from proxy tasks that may not align with physicochemical relevance, CheMeleon's targets are the same descriptors chemists have used as features for decades — but distilled into a differentiable encoder that can be fine-tuned end to end.

#Key Features

• Descriptors as the pretraining signal: Computed topological and physicochemical descriptors are exactly reproducible, so the pretraining objective carries no measurement noise and no simulation bias.
• Dynamic target masking: 85% of the descriptor targets are randomly masked in the loss at each step, forcing the encoder to build a representation that does not lean on any single descriptor.
• Single-flag fine-tuning: The weights ship inside Chemprop 2.2.0 and later, so a downstream model is trained with --from-foundation CheMeleon on a table of SMILES and labels.
• Robustness to activity cliffs: On MoleculeACE assays the model's error is statistically indistinguishable between cliff and non-cliff compounds far more often than the baselines, meaning accuracy survives small structural edits that flip potency.
• Fingerprints without fine-tuning: A CheMeleonFingerprint helper emits fixed embeddings for SMILES strings or RDKit molecules, usable as features for any classical model.

#Technical Details

The message-passing portion of the network has a hidden dimension of 2048 and a depth of 6 iterations with mean aggregation, totalling 8.7 million parameters; the two-layer feed-forward head that maps the embedding to descriptors brings the pretrained model to 12.9 million parameters. Training minimizes mean-squared error on the unmasked descriptors, each rescaled to zero mean and unit variance and Winsorized at six standard deviations, reaching a test RMSE of 0.14 averaged across all descriptors. For downstream work the pretrained encoder is retained, the pretraining head is discarded, a fresh randomly initialized task head is attached, and the whole network is fine-tuned by stochastic gradient descent. Across 58 benchmark tasks — 28 curated from Polaris and the Therapeutic Data Commons, plus 29 ChEMBL assays from MoleculeACE — and with no per-benchmark hyperparameter optimization, CheMeleon wins 75% of Polaris tasks against minimol (71%), Random Forest (68%), MoLFormer (39%), fastprop (36%), Chemprop (32%), and MolCLR (14%), and 97% of MoleculeACE assays. k-nearest-neighbor probing on 20 ToxCast endpoints confirms that the learned embedding space groups compounds by toxicity in the read-across sense, not merely by structural similarity.

#Applications

CheMeleon is aimed at the everyday regression and classification problems of small molecule discovery: solubility and permeability, ADME endpoints, target bioactivity, and toxicity screening, all trained from SMILES strings and measured labels. Its practical appeal is that it requires no new tooling — a group already running Chemprop adds one flag and gets a pretrained encoder, and the fingerprint helper serves teams whose pipelines are built around scikit-learn rather than PyTorch.

#Impact

The result reframes what a chemistry foundation model needs to be pretrained on: a million cheap, exact descriptor calculations proved a stronger transfer signal than far larger corpora of noisy experimental data. Released under the MIT license with weights on Zenodo and distribution through Chemprop, the model has been picked up as the backbone for OpenADMET's public endpoint baselines and is offered by fine-tuning platforms as a turnkey SMILES-to-label starting point. Caveats are worth stating plainly: the paper remains a preprint, all evaluations are single-task regression or classification with no hyperparameter search, and the descriptors themselves can still be computed exactly by Mordred — the model's value lies entirely in the transferable representation, not in predicting descriptors.