A multiscale model-guided screening workflow produced seven HPK1 compounds that exceeded 80% inhibition at 50 micromolar in a kinase activity assay. One, HQ2026-03, had a reported IC50 of 0.41 nanomolar, the lowest of the seven. IC50 refers to the concentration at which a compound inhibits half the measured activity.
The finding is a narrow laboratory result, not a demonstration of selectivity or cellular efficacy. The compounds were selected candidates rather than an unbiased sample of the library: about 200,000 ChemDiv compounds were screened, 286 overlapping candidates were retained after ranking, and only 20 were chosen for testing. The seven positives therefore describe that selected group, not the full library.
A model built from several views
MF-Net was designed to examine a drug-target pair through sequence-, atomic- and fragment-level representations. It fused those views with a convolutional neural network, used a multilayer perceptron to predict affinity, and included contrastive alignment during training.
For the HPK1 case study, MF-Net was trained without HPK1 data. It screened about 200,000 ChemDiv compounds, then intersected the top 2,000 docking-ranked compounds with the top 2,000 MF-Net-ranked compounds. That overlap contained 286 candidates, from which 20 were selected for kinase testing at 50 micromolar.
Where the model held up
The broader evaluation drew on separate public benchmark datasets. PDBBind v2016 supplied 10,451 training samples, 1,000 validation complexes and 250 core-test samples. Screening tests also used 92,883 DUD-E training and validation samples and 28,334 DEKOIS2.0 test samples, with LIT-PCBA added as another benchmark.
On the PDBBind core test, MF-Net had the best reported scores: RMSE 1.127 plus or minus 0.004, MAE 0.865 plus or minus 0.005, Pearson correlation 0.862 plus or minus 0.005 and SD 1.116 plus or minus 0.008. RMSE and MAE measure prediction error, while Pearson correlation measures how closely predictions track the reported values.
The advantage weakened as the test cases moved further from the training examples. In out-of-distribution tests, novel ligands and proteins were separated using similarity or sequence-identity thresholds below 20%. RMSE was 1.212 plus or minus 0.027 for novel ligands and 1.326 plus or minus 0.029 for novel proteins, while the novel-protein correlation was 0.804 plus or minus 0.009. Novel complexes were the weakest setting.
Component tests showed that sequence information was especially important within this implementation, although the comparisons do not establish a biological mechanism. Removing sequence features raised RMSE by 20.6%, MAE by 23.7% and SD by 19.4%, while lowering Pearson correlation by 8.5%. Removing the contrastive loss raised RMSE by 7.3% and MAE by 7.9%.
MF-Net also outperformed the other tested fusion designs, including direct concatenation, attention-based, gated and transformer-based versions. It had the lowest reported RMSE, MAE and SD and the highest Pearson correlation among those variants.
The edge was strongest at the top of the list
The screening results were more mixed. On generic structure-based tests, MF-Net was strongest at the earliest cutoffs, with enrichment factors of 29.05 at EF0.1% and 23.72 at EF0.5%. IGN and EHIGN performed better at higher screening thresholds, and the analysis notes lower sensitivity to weak or moderately active compounds.
On seven representative LIT-PCBA targets, MF-Net was reported as the best model at EF0.1% for every target and best overall on five. For KAT2A, its EF0.1% was 62.21, compared with 41.47 for EHIGN, a reported 50% improvement.
Seven hits, with a wide range of potency
In the follow-up assay, seven of the 20 selected compounds exceeded 80% inhibition at 50 micromolar. The study then used a ten-point, four-fold serial dilution to determine IC50 values. The reported values included 0.41 nanomolar for HQ2026-03, 5.63 nanomolar for HQ2026-01, 11.24 nanomolar for HQ2026-02 and 5697.00 nanomolar for HQ2026-07.
HQ2026-01 showed similar potency to the reference inhibitor Sunitinib, at 5.63 nanomolar versus 5.44 nanomolar. HQ2026-03 was the most potent reported compound at 0.41 nanomolar and was described as surpassing the reference inhibitor. HQ2026-03, HQ2026-04, HQ2026-05 and HQ2026-06 also exceeded 90% inhibition at 50 micromolar.
The assay numbers need careful handling. Confidence intervals and independent reproducibility estimates were not reported for the IC50 results, and the comparison with Sunitinib did not establish kinase selectivity or cellular efficacy. The generic screening analysis also found that other methods performed better at higher thresholds.
The study makes a strong case for MF-Net under the benchmark and early-ranking conditions it tested, while the experimental evidence remains limited to a kinase activity assay on 20 selected compounds. The results leave open whether the model will hold up across broader ligand and protein shifts, or whether the reported HPK1 activity will prove selective and reproducible.
Disclosure and access
The article is peer-reviewed and openly licensed. The authors declare no conflict of interest and say that the public data sources, supporting data and source code are openly available through MF-Net. The work was supported by programs from the National Administration of Traditional Chinese Medicine, the Chinese Medicine Guangdong Laboratory and the Macao Science and Technology Development Fund.
Paper data and sources
Original title: A Unified Hierarchical Multiscale Fusion Framework for Drug-Target Affinity Prediction: From Benchmark Performance to Nanomolar Inhibitor Discovery.
Authors: Shuo Liu, Xiang Zhang, Haixia Feng et al.
Journal/Repository: Advanced science (Weinheim, Baden-Wurttemberg, Germany)
Status: Peer-reviewed
First online: 2026-08-21
DOI: 10.1002/advs.77345
Original paper