Small-molecule foundation model pairing a graph encoder with a Transformer SMILES decoder so compounds can be optimized directly in encoding space.
No providers recorded yet. Browse all providers
Ouroboros is a chemical foundation model for small molecules that treats representation learning and molecular generation as two directions through the same latent space. A graph neural network compresses a molecule into a single 2,048-dimensional encoding vector, and an autoregressive Transformer decompresses that vector back into a SMILES string. Because both directions share one encoding space, a compound can be edited by optimizing its vector against a differentiable objective and then decoded at each step, turning iterative medicinal-chemistry optimization into gradient descent on a continuous representation.
That framing addresses a persistent split in computational chemistry. Representation models learn embeddings that transfer well to property prediction but cannot emit new structures, while generative models produce structures from latents that are poorly organized for downstream prediction. Chemical language models such as ChemBERTa sit on the representation side of that split, and in the paper's screening benchmark show no substantial advantage over molecular fingerprints at retrieving molecules with comparable biological activity. Ouroboros decouples the two modules — the encoder is pretrained without any generation constraint, and the decoder is trained afterwards to invert it — so the encoding space is shaped by chemistry rather than by reconstruction convenience.
The model was developed by the Center for AI and Computational Biology at the Institute of Systems Medicine, Chinese Academy of Medical Sciences, with collaborators at the National University of Singapore and DeepMed Technology. It was posted to bioRxiv in March 2025 and published in Advanced Science in 2026.
Molecules are converted to graphs with atom features (type, hybridization, formal charge, chirality, ring and aromatic flags) and bond features, passed through a Weisfeiler-Lehman message-passing network, and pooled by a global self-attention module into the 2,048-dimensional encoding. Two projection heads map encodings to conformational and fingerprint similarity scores under an MSE loss; the message-passing weights are inherited from the earlier GeminiMol model and frozen for the first 2,000 steps. The encoder was trained on similarity matrices built from 126,248 molecules drawn from GeminiMol's diverse set, the CycPeptMPDB macrocyclic peptide database, the Cell Painting Gallery, PDB ligands, and a hand-curated cofactor set. The decoder is a 4-layer, 32-head Transformer that takes the encoding as memory and emits logits over a 43-token SMILES vocabulary; it was trained on the 48.2M-compound Enamine REAL Diversity set combined with the core molecules. Three checkpoints are released: M0 reproduces the published methodology, while M1c and M1d train the decoder on broader data using SMILES and SELFIES respectively. On virtual screening, Ouroboros encodings achieve the highest BEDROC of seven methods on both DUD-E (1,463,336 molecules) and LIT-PCBA (2,808,885), and its property predictor gives the best Spearman correlation on 6 of 10 ADMET and physicochemical datasets against GeminiMol, CombineFP, and FP-GNN.
Ouroboros supports ligand-based virtual screening, QSAR and ADMET property modeling, and lead optimization from a single frozen encoder, which suits medicinal chemistry teams that would otherwise maintain separate models per task. The authors applied it to a multi-target oncology campaign spanning 10 drug targets tied to five cancer-driver genes: similarity screening of a 48.2 million-compound library followed by docking produced 18 synthesized compounds, of which 7 were active and 3 inhibited multiple targets, with one reaching nanomolar IC50 against AURKA alongside PI3Kα and PI3Kγ activity.
The wet-lab validation is what distinguishes Ouroboros from purely in-silico generative chemistry work, demonstrating that navigation through a learned encoding space can reach synthesizable, experimentally active scaffolds. Its clearest limitation is that it does not predict drug-target binding affinity, so target-directed campaigns still depend on docking to rank the molecules generated along an optimization trajectory, and none of the 18 tested compounds inhibited more than four targets at once. The code is MIT licensed and the pretrained checkpoints are distributed from the developers' web page, so the encode-optimize-decode loop runs outside the originating lab — but it cannot be rebuilt there: the repository ships the encoder and decoder architectures and the inference entry points, with no script that pretrains the encoder or trains the decoder. The preprocessed DUD-E and LIT-PCBA sets come from GeminiMol's Zenodo deposit under an Academic Free License v3.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.