Drug-target affinity prediction that runs Forman curvature flow on molecular and ESM-2 residue contact graphs, from SMILES and sequence alone.
No providers recorded yet. Browse all providers
CurvFlow-DTA predicts drug-target binding affinity by treating both the drug and the protein as graphs whose geometry evolves under a discrete curvature flow. Graph neural networks pass messages along edges, and when many distant atoms or residues must route information through a few bottleneck connections, that information gets compressed. Discrete Ricci curvature flags those bottlenecks: negatively curved edges sit where the graph pinches. Earlier curvature-aware affinity models computed a single static curvature snapshot, and only for the drug, while encoding the protein as a sequence summary. CurvFlow-DTA instead runs a weighted Forman curvature flow on both graphs, producing an ordered trajectory of edge weights, and lets each drug-target pair decide how far along that trajectory each branch should read.
The model was developed by Jicheng Ma, Yunyan Yang and Juan Zhao at Renmin University of China with corresponding author Liang Zhao at Beijing Normal University, and posted to arXiv in September 2026. The protein side is built without any bound complex: a frozen ESM-2 supplies per-residue embeddings and predicted contacts, which define the residue contact graph. At inference the model needs only a SMILES string and an amino acid sequence.
Drugs are converted from SMILES to RDKit molecular graphs with standard atom and bond attributes. Proteins are embedded with the frozen 150M-parameter ESM-2 checkpoint (esm2_t30_150M_UR50D), concatenated with an amino acid type embedding and positional encoding; the graph keeps backbone edges plus the top-ranked long-range predicted contacts (sequence separation at least 6, degree capped at 8). The flow runs for 8 steps with step size 0.1, giving a maximum horizon of 0.8, with metrics bounded between 0.25 and 4 and renormalized to unit mean. Two Flow-GINE encoders, 5 layers for the drug and 3 for the protein at hidden dimension 256, apply a per-edge exponential gate derived from the interpolated metric; pooled outputs are fused and passed to a two-layer MLP. Training uses a Huber loss, AdamW at learning rate 1e-4 and up to 300 epochs with early stopping. ESM-2 and the cached trajectories are never updated.
Separate models are trained for Davis (pKd) and KIBA (KIBA scores) in each split, with five seeds per setting. On the warm split, CurvFlow-DTA reaches MSE 0.185 and CI 0.902 on Davis and MSE 0.137 and CI 0.887 on KIBA, against 0.231 and 0.169 MSE for Ricci-GraphDTA reproduced under the same protocol. Cold-start splits hold out Bemis-Murcko scaffolds, MMseqs2 protein clusters at 40% identity, or both; Davis cold-target MSE is 0.300 versus 0.413 for Ricci-GraphDTA. Ablations show a monotonic improvement from no curvature through static curvature, fixed-horizon flow and entity-only horizons to the full pair-conditioned model.
CurvFlow-DTA is aimed at virtual screening and target prioritization settings where no bound structure exists for a candidate pair and the drug scaffold or target family may be absent from training data. Because protein graphs come from sequence alone, it applies to targets without experimental structures, and its residue attributions give medicinal chemists a way to see which part of a target drives a prediction. In two held-out kinase co-crystals, MET with an inhibitor (PDB 3U6H) and CDK2 with olomoucine (PDB 1W0X), top-attributed residues line the ATP cleft with 3.2x and 3.9x pocket enrichment over chance.
CurvFlow-DTA extends curvature-aware graph learning in affinity prediction from a static drug-only descriptor to an evolving geometry on both partners, and its ablations isolate the flow and the pair-conditioned horizon as sources of the gain rather than the protein language model features. Its evaluation is limited to two kinase-dominated benchmarks, and the authors note that literature baselines use different partitions, so cross-paper rankings are indicative only. The released repository provides training, preprocessing and split-construction code for reproducing the Davis and KIBA experiments; trained checkpoints are not distributed, so applying the model requires training it first.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.