Highlights
Artificial intelligence in drug discovery — what it is, where we stand and the path forward
Andreas Bender and others argue in this perspective that AI has meaningful use cases in drug discovery without _yet_ showing evidence of clinical impact, and why that might be. The paper builds on previous works of Andreas (read his blogpost about them) including the cost model that suggests optimizing phase II success rates is the most fruitful strategy (monetarily). Most of the “AI” work is in preclinical hit finding, i.e., in ligand discovery rather than drug discovery, where labelled data happens to exist. The paper also makes the data case i.e. biological data is conditional on assay parameters, cell system, dose and patient context, which makes label assignment unreliable or at best uncertain; chemical data is sparse and biased, for instance, the four common ADME datasets from Therapeutics Data Commons share only 0.1% of compounds, so applicability domains might not line up across endpoints in a multi-objective setting. The paper most interestingly argues model validation is not process validation and that’s a really neat way to think about the problem. Great read as always.
Some related links:
Read the paper here: https://rdcu.be/fyr77
So How Is AI Drug Discovery Doing, Really? by Derek Lowe
Thibault Geoui’s video on LinkedIn
Andreas’s LinkedIn post
Strategy-first synthesis planning for complex natural products
Armstrong, Nguyen et al. from Schwaller Group present SynthEx as a retrosynthesis planner built from LLM agents rather than a template library. The model writes each disconnection as a list of atom-level graph edits, and those edits are applied to the product to produce the precursors. The authors ran it on 1,098 natural products from NP-Atlas that have no published total synthesis. AiZynthFinder, finds complete routes for ~14% of them. SynthEx finds routes for ~64%. Ten synthetic chemists rated key steps blind, mixed in with steps from total syntheses published after the model’s cutoff, and could not tell them apart on feasibility / overall quality. None of it has been run in a lab, yet. The routes are public as SynthAtlas. Interesting read.
Deep Learning Foundation Models for Low-Data Regimes from Classical Molecular Descriptors
CheMeleon has a new citation now :) If you haven’t used it yet, give it a go here: https://github.com/JacksonBurns/chemeleon and https://chemprop.readthedocs.io/en/latest/chemeleon_foundation_finetuning.html . Good read!
Long List
Cheminformatics
Generative virtual screening: from search to generate and back — Great read !!
Multi-task graph neural networks for comprehensive surfactant property prediction
A decoupled alignment kernel for peptide membrane permeability predictions
Train–test splits matter more for evaluation than for performance
ALF: Open-Source Active Learning Framework for Atomistic Modeling
Explicit, Machine-Learned Two-Body Potentials for Molecular Simulations
One-Dimensional Molecular Coordinate Orders Amino Acids across
ProtLID2.0: A Residue-Specific Pharmacophore Framework for Protein–Protein Docking
An OpenMM-Based ML/MM–MMGBSA Workflow for End-point Protein–Ligand Binding Energy Ranking
Motif-Based Graph Learning for Synthetic Reaction Condition Prediction
Deep3MVPF: Multiview Deep Framework for the Prediction of Stability and m6A in mRNA 3′UTR
Odor-Sight: An Interpretable Graph Neural Network Web Platform
Uncertainty Quantification with Domain Classification Models for Acetylcholinesterase Inhibition
Novelty Assessment Method for Streaming Samples Based on Slope Entropy
Geometric Deep Learning-Based Drug Design Models for Small-Molecule Drug Discovery
Fast intermolecular interaction energy calculation with the OPLS-AA force field
Integrating AI and Molecular Modeling for Structural Prediction of a
TriGraphQA: A Triple Graph Learning Framework for Model Quality Assessment of Protein Complexes
Reliable Identification and Out-of-Library Detection in Mass Spectra
Targeting the RNA-Binding Site of SARS-CoV‑2 NSP13 by FRASE-bot
Barrel Shape and Chromophore Rigidity Predict Fluorescent-Protein
Computational Study of Allosteric Switch of NLRP3: How ATP Binding Unlocks the NACHT Domain
MedChem
PepGate: A Dual-Path Diffusion Framework for ACE Inhibitory Peptide De Novo Design
Neutral Backbone Modifications Enhance the Pharmacokinetics and
Biodistribution of Antibody−Oligonucleotide Conjugates with High
Deciphering Chemical Environments in Conformational Data Sets Using an IQA Band-Matching Protocol
Other
Harnessing Host–Guest Conformational Matching for Efficient Near-Infrared Phosphorescence
An Explicit Interaction-Prompted Diffusion Framework for High-Fidelity 3D Molecular Generation
Palate Cleanser
Best,
Manas