This Week In Cheminformatics: Issue #017
π-conjugated chemical space exploration, shape constrained diffusion model and long list of papers
Highlights
De Novo Molecular Design via Shape-Constrained Diffusion Models
If you are tracking the evolution of 3D generative models, this paper on Diff-Shape is worth a close read because it effectively adapts a ControlNet architecture to molecular diffusion. Rather than retraining a conditional model for every new shape, Diff-Shape keeps an unconditional 3D generator frozen and injects geometric guidance via a lightweight, trainable Graph ControlNet branch. This approach is quite technically interesting. The integration of inference-time substructure inpainting is particularly useful for practical workflows, allowing the exact same backbone to handle scaffold hopping, R-group decoration, and linker generation without requiring task-specific retraining. Crucially, they backed up their benchmark metrics with actual wet-lab validation, synthesizing novel KRAS G12D and EGFR inhibitors that demonstrate nanomolar biochemical potency which is pretty cool !
Defining a Chemical Space of π-Conjugated Hydrocarbons by Unit-Based Construction
Suga et al. introduce CARBOT, a generative framework that systematically enumerates π-conjugated hydrocarbons via unit-based construction. Instead of relying on coarse predefined fragments or atom-based mutations that often break extended conjugation, CARBOT applies a minimal set of conjugation preserving transformations, namely vinyl additions, ethynyl additions, and cyclizations. What makes this tool particularly useful is its mathematical rigor: it is formally proven to be sound and complete for the defined space of neutral Kekulé structures, restricted to sp and sp2 carbons while explicitly excluding allenes and carbenes. The authors combined these structural rules with composite operations to accelerate aromatic ring formation, showing that a straightforward depth-first search guided by a substructure-matching score efficiently navigates this space. This circumvents the inefficiencies of standard fingerprint-based evolutionary searches for topologically complex targets, allowing the framework to scale to structures as large as C_200H_100 (!!) and reliably rediscover structures like Möbius annulenes and nanobelts. Good read !
Long List
Cheminformatics
A user’s guide to your first self-driving liquid handling lab
Generalization of Long-Range Machine Learning Potentials in Complex Chemical Spaces
Enhancing Predictive Modeling with Molecular Fingerprint Fusion Strategies
MolRes-DTA: a molecular-multiview fusion and residue-aware model for drug-target affinity prediction
A deep learning approach to searching property spaces of materials
ChemXDyn: Dynamics-Informed Species and Reaction Detection Methodology from Atomistic Simulations
Protein–Protein Cross-Linking by a DNA Damage-Derived Histone Modification
A Computational Modeling of ADLumin Chemiluminescence: Oxygenation and Dioxetanone Formation
Explainable Machine Learning Guided Enhanced Sampling of Protein Conformational Transition in HSP90
SAFR: Enabling Fragment-Based Drug Discovery with a Synthetic Binding Pose Data Set
Markov State Models for Tracking Reaction Dynamics on Catalytic Nanoparticles
mdBIRCH for Fast, Scalable, Online Clustering of Molecular Dynamics Trajectories
MedChem
Other
Palate Cleanser
Why do I even do this,
Manas