This Week In Cheminformatics: Issue #006

Binding Site Vectors to Compare Active Sites, ArtiDock, Benchmarking Boltz-2, and a long list of papers from last week of Jan

Highlights

---

Assessing Boltz-2 Performance for the Binding Classification of Docking Hits

This is yet another paper on the “Clever Hans” paradox. While the Boltz-2 model achieves impressive hit enrichment on the difficult ULVSH dataset, outperforming physics-based methods that typically fail there, it appears to do so for the different/wrong reasons. The authors demonstrate that affinity predictions are completely decoupled from pose quality (surprise, surprise!), with the model successfully classifying binders even when the predicted structure is incorrect. On adversarial tests, Boltz-2 often continues to predict high affinity even when the binding site is destroyed by mutations or, in some cases, when the target sequence is swapped entirely. This implies Boltz-2 is acting less like a “true” cofolding tool (whatever that means) and more like a sophisticated ligand-based model, likely memorizing 3D pharmacophoric shapes from its training data while largely ignoring the physical reality of the protein pocket. Good read!

Binding Site Vectors Enable Mapping of Cytochrome P450 Functional Landscapes

Kuvek et al. introduced a clever solution to compare active sites that share function but vary in sequence. Instead of traditional grids or surface probes, the authors utilize “binding site vectors” that radiate from a central anchor (the heme iron) to the pocket surface, encoding both topology and local electrostatics into a multidimensional numeric format. This allows complex 3D binding environments to be compared rapidly using simple RMSD calculations. This method effectively works on dynamic conformational ensembles. While standard backbone alignments grouped enzymes merely by lineage, the vector analysis of Molecular Dynamics trajectories successfully clustered diverse enzymes such as CYP3A4 and CYP1A2 based on their actual substrate overlap. This work also proposed “ligand vectors” for binary docking. By inverting the method to map a ligand’s surface, one can instantaneously check if a ligand “fits” a specific protein conformation without computationally expensive energy evaluations, offering a novel approach for high-throughput pharmacophore modeling.

ArtiDock: Accurate Machine Learning Approach to Protein–Ligand Docking Optimized for High-Throughput Virtual Screening

Voitsitsky et al. benchmarked their new docking model on the PLINDER dataset with cluster-based splitting, supposedly to minimize the data leakage that is known to inflate performance in many docking papers. The result is a median RMSD of 2.0 Å compared to Glide’s 2.8Å, which is pretty impressive! For high-throughput workflows, it’s quite efficient. The reported cost of roughly $5 to screen one million molecules is significantly lower than both Glide and GPU-accelerated AutoDock. ArtiDock outperforms other models most significantly in pockets containing ions and explicit water molecules. The authors also tried coupling ArtiDock with a fast UFF minimization, whereby they fix steric clashes and bond distortions without sacrificing accuracy, offering a hybrid workflow that feels robust enough for actual production use.

Long List

---

Cheminformatics

• Multi-Modal Fusion Frameworks of Subgraph-Optimized Graph Autoencoder for Molecular Property Prediction

• MOFReasoner: think like a scientist—a reasoning large language model via knowledge distillation

• ToPolyAgent: AI Agents for Coarse-Grained Bead-Spring Topological Polymer Simulations

• A multi-task learning approach for prediction of missing bioactivity values of compounds for the SLC transporter superfamily

• Accelerating Catalytic Advancements Through the Precision of High-Throughput Experiments & Calculations

• Advancing chemical safety prediction: an integrated GNN framework with DFT-augmented cyclic compound solution

• Computational prioritization of multi-target inhibitors: explainable QSAR and docking-based discovery of dual AChE/BACE1 chemotypes

• Fragment-Guided New Therapeutic Molecule Discovery and Mapping of Clinically Relevant Interactomes

• Selector: A General Python Library for Diverse Subset Selection

• Attention-Guided Multiview Deep Learning Framework Uncovers miRNA-Drug Associations for Therapeutic Discovery

• PharmGEO: A Curated Atlas of Drug-Response Transcriptomes Enabling Cross-Study Comparisons

• STRIKER: a spectral metadata repairing tool for expanding the comprehensiveness of spectral libraries

• Cut-SOAP: A Machine Learning Descriptor for Rapid Screening of Molecular Adsorption Energetics

• Bayesian Optimization of the Coarse-Grained Force Field for Ions and Ionic Surfactants Based on a Polarizable Water Model

• Machine-Learning Framework for Excitation Energies of Chromophores in Polarizable Environments

• The discovery of monoamine oxidase inhibitors: virtual screening and in vitro inhibition potencies

• XRepDDA: An Interpretable Drug–Disease Association Prediction Framework Leveraging Pretrained Chemical Language Models

• NavDB: A Comprehensive Database for Voltage-Gated Sodium Channels Modulators and Targets

• Traj2Relax: A Trajectory-Supervised Method for Robust Structure Relaxation

• Chemical genomics language model toward reliable and explainable compound-protein interaction exploration

• PLMCA: A General Multimodal Protein–Ligand Cross-Attention Framework for Pocket Identification and Binding Affinity Prediction

• Random Functions as Data Compressors for Machine Learning of Molecular Processes

• Benchmarking deep learning models for predicting anticancer drug potency (IC50) with insights for medicinal chemists

• Graph Neural Networks for Polymer Characterization and Property Prediction: Opportunities and Challenges

• AI-Guided Conformational Dynamics of p53 L1 Loop Reveal an Allosteric Switch Regulating DNA Binding and Cancer Hotspots

• NeuroTDPi: Interpretable Deep Learning Models with Multimodal Fusion for Identifying Neurotoxic Compounds

• Improving Robustness and Training Efficiency of Machine-Learned Potentials by Incorporating Short-Range Empirical Potentials

• Uncertainty-Aware Prediction of 195Pt Chemical Shifts from Limited Data

• Estimating the Hydrogen Bond Strength by Machine Learning Approaches

MedChem

• Idler Compounds: A Simple Protocol for Openly Sharing Fridge Contents for Cross-Screening

• Discovery of AZD9750, an Orally Bioavailable Androgen Receptor Degrader for the Treatment of Prostate Cancer

• Forecasting drug resistant HIV protease evolution

• Enthalpy–Entropy Trade-Off Underlies Geometric Isomer Selectivity in Histamine H1 Receptor–Doxepin Interaction

• General Binding Affinity Guidance for Diffusion Models in Structure-Based Drug Design

• Shaping Antimalarials: A Geometry-First Approach to PfCLK3 Covalent Inhibitors

• Integrating Medicinal Chemist Expertise with Deep Learning for Automated Molecular Optimization

PhotoChem

• Temperature-, Concentration-, and Solvent-Dependent M/P Helicity Switching of Double-Helical Monometallofoldamers with Inversion of Circularly Polarized Luminescence

• Interlocked Rotaxane Enables TADF with Distinct Excited-State Structural Relaxation

Other

• Molecular dynamics simulations accelerated on FPGA with high-bandwidth memory

• A data-driven approach to control stimulus responsivity of functional polymer materials: predicting thermoresponsive color-changing properties of polydiacetylene

• PCR bias impacts microbiome ecological analyses

• Development of Multiple Local Computational Models in Retrosynthetic Analysis: Total Synthesis of (−)-Deoxylimonin

• T-Cell “Rejuvenation” Nanovaccine: Enhancing Immunological Memory and Antitumor Responses through Telomere Extension

• Outpacing Emerging Drug Threats: Validation of ToxBox Kits That Automate LC-MS/MS Analyses

Palate Cleanser

---

LaurieWired@lauriewired

today’s one-sentence horror: sudo has been largely maintained by a single person for ~30+ years

6:25 PM · Jan 31, 2026 · 933K Views

---

203 Replies · 1.04K Reposts · 20.4K Likes

amrit@amritwt

wrote code by hand. zero reviews. clean merge. i still got it.

7:41 PM · Jan 28, 2026 · 112K Views

---

46 Replies · 207 Reposts · 9.2K Likes

Nostos@nostosart

Happy Franz Kafka's "Nothing, merely tired." day

4:42 AM · Feb 1, 2026 · 497K Views

---

17 Replies · 9.16K Reposts · 33.3K Likes

sudox@kmcnam1

5:20 AM · Jan 25, 2026 · 867K Views

---

253 Replies · 5.65K Reposts · 82.1K Likes

---

---

Stay alive,
Manas