ChemDeck · 29 cards · Dec 2018 to Oct 2026

Papers worth your time, in one card.

Swipeable TL;DRs of recent work in computational chemistry and drug discovery: the problem, the method and the result, with one visual.

No. 01 / 29 Chem. Sci. Oct 2026 Paper

Syndirella: synthesis-directed fragment elaboration enables extensive and cost-effective binding site exploration beyond catalog compounds

Fieseler, Winokan, Morrone, … Thompson

01 Problem: Fragment-to-hit design usually elaborates one or two fragments, risking early convergence on a dead series, because buying a broad, diverse set of derivatives from catalogs is prohibitively expensive.
02 Method: Syndirella takes fragment-derived scaffolds, picks a forward synthesis route from a library of 109 SMIRKS reactions chosen for high-throughput chemistry, runs superstructure searches for catalog reactants, enumerates the Cartesian product, and places every product against the inspiration fragments with Fragmenstein.
03 Result: Across three viral targets it reached chemical space absent from catalogs with the best breadth-to-depth ratio (7.4 × 10⁻⁴ vs 5.1 × 10⁻⁴ for Arthor and 1.8 × 10⁻⁴ for Manifold). On NCS-1, only 2 of 5,335 designed elaborations were purchasable, while buying 212 reactants gave 454 attempts and 252 synthesised compounds at €42.10 each, 71% below catalog price.

Figure 1.Cost per compound on the NCS-1 campaign (EUR)Numbers from the paper (Chem. Sci., Oct 2026). Accent-hatched bars are the ones the card highlights.

No. 02 / 29 J. Chem. Inf. Model. Oct 2026 Paper

Subpocket-Aware Fragment-Based Docking Pipeline for Kinase Ligand Design: Methodology and a PKA Case Study

Buchthal, Kramer, Hubach, … Volkamer

01 Problem: KinFragLib holds 7,486 fragments cut from kinase–ligand complexes, but recombining them exhaustively is computationally infeasible, and generic fragment-based design throws away the kinase-specific structure that defines where each fragment belongs.
02 Method: An automated pipeline that grows fragments subpocket by subpocket instead of enumerating the combinatorial space: a beam search over KinFragLib fragments with pharmacophore-guided template docking in FlexX, constrained by the subpocket each fragment was originally found in.
03 Result: On PKA the pipeline produced 33,758 candidates. Of 15 synthesised, seven cut kinase activity by more than half at 100 μM and five still worked at 10 μM, with IC50 values in the double- to triple-digit nanomolar range. A crystal structure of the most potent compound confirmed the predicted binding mode.

5 of 15 synthesised compounds with nanomolar IC50, the best one confirmed by crystallography

No. 03 / 29 J. Chem. Inf. Model. Oct 2026 Paper

Building a Physically Interpretable Compressed Chemical Space via Grid-Based Optimization

Yoon & Dickson

01 Problem: Generative models encode molecules into latent spaces that are compact but physically meaningless, so nobody can say how much structural information survives the compression, or how far a molecule can be compressed before its behaviour in an environment stops being reproducible.
02 Method: COMPRESS represents a molecule as a fixed number K of spatial sites, each with three coordinates plus a charge and Lennard-Jones parameters, optimised to reproduce the all-atom density, electrostatic and van der Waals fields on 3D grids. Evaluated on benzene, 20 scaffolds, benzene dimers, protein–ligand complexes and 10,000 GEOM-Drugs molecules, with a flow-matching decoder back to all-atom structures.
03 Result: Field agreement decays slowly at high resolution and then collapses abruptly below a threshold, giving a measurable limit to compression rather than a vague one. K = 25 sites kept about 97% of GEOM-Drugs molecules under a 0.05 field loss, and editing a site's charge or size, or adding and deleting sites, decodes to realistic new compounds.

K = 25 sites retain ~97% of GEOM-Drugs molecules under a 0.05 field loss

No. 05 / 29 Bioinformatics Oct 2026 Paper

Information Leakage in Enzyme Substrate Prediction

Atabaigi Elmi, Joeres & Kalinina

01 Problem: Deep learning models for enzyme–substrate prediction report remarkable accuracy on a popular benchmark, but if training and test sets share near-identical sequences and ligands, those numbers measure recall rather than generalisation.
02 Method: An audit of one of the field's most-used datasets and four models trained on it, re-splitting separately by protein sequence similarity, by substrates shared between train and test, and by ligand structural similarity.
03 Result: With the leakage removed the models are no better than statistical baselines. Controlling sequence similarity alone still leaves high performance, removing shared substrates hurts substantially, and separating structurally similar ligands drops predictions to near random: these models generalise across enzyme sequence space far better than across ligand space.

≈ random once structurally similar ligands are separated between splits

No. 06 / 29 J. Chem. Inf. Model. Oct 2026 Paper

Validation Strategy Dominates Model Choice in Chronological QSAR Evaluation

Almendra & Falcao

01 Problem: QSAR models are usually judged by random cross-validation, which may flatter them when test compounds sit close to the training data in structure or time. Is the optimism about the model, or the split?
02 Method: Fourteen curated ChEMBL classification sets, five validation schemes (random, chronological by publication year, a thinned control, and structurally filtered versions of each), and four model classes (RF, SVM, XGBoost, neural net) trained under one fixed workflow.
03 Result: Random validation looked great: MCC 0.75 in CV, 0.72 on validation. Structural filtering cut validation MCC to 0.48, and chronological validation collapsed it to 0.19 while CV still read 0.73. Model classes overlapped heavily: the split matters far more than the algorithm.

Figure 6.Validation MCC by split (CV MCC ≈ 0.73 throughout)Numbers from the paper (J. Chem. Inf. Model., Oct 2026). Accent-hatched bars are the ones the card highlights.

No. 07 / 29 J. Chem. Inf. Model. Oct 2026 Paper

DESPOT: Direction-Enhanced Scoring POTentials for Scoring of Protein–Ligand Interactions

Poelmans, Van Eynde, … Voet

01 Problem: Knowledge-based scoring functions are reliable and interpretable, but treat the space around each protein atom as isotropic and cannot represent regions that ligands should leave empty.
02 Method: An all-atom anisotropic potential: atoms are classed by symmetry from hybridisation, space around them is discretised accordingly, and a 'void' ligand atom type learns where space is preferentially left empty. Trained on 110,943 curated complexes and combined with a residue-level score.
03 Result: On CASF-2016, scoring power r = 0.61 and best-in-class docking power, 89.5% top-1 success for DESPOT-combo. All anisotropic variants beat isotropic potentials and established empirical scoring functions in virtual screening.

89.5% top-1 docking power on CASF-2016

No. 08 / 29 J. Chem. Inf. Model. Oct 2026 Paper

AcepKa: Thermodynamics-Informed pKa Prediction and Protonation-State Generation in PlayMolecule AI

Pesce, Farr & De Fabritiis

01 Problem: pKa and protonation states drive solubility, permeability and binding, yet most predictors treat pKa as a single regression number and ignore how coupled ionisation sites interact.
02 Method: An implementation of the Uni-pKa framework on a retrained Uni-Mol backbone that models the full protonation ensemble with thermodynamic consistency, plus a GPU conformer generator and a 3D mode that protonates bound ligand poses directly.
03 Result: Matches state-of-the-art accuracy on public benchmarks. The conformer generator is about 7× faster than other GPU tools and over an order of magnitude faster than multithreaded RDKit, enabling library-scale protonation before docking or MD.

~7× faster GPU conformer generation than other GPU tools

No. 09 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Fréchet ChemNet Distance Is Blind to Double-Bond Stereochemistry

Yim

01 Problem: Fréchet ChemNet Distance is a default metric for judging generative models, but nobody had checked whether it sees E/Z stereochemistry.
02 Method: Traced the ChemNet tokenizer, inverted every double bond in MolT5-large's ChEBI-20 output, and compared models whose E/Z recall differs more than threefold.
03 Result: The tokenizer maps / and \ to the same unknown token, so inverting 384 molecules shifted FCD by just 0.001 resampling SDs while per-bond accuracy fell from 98.97% to 1.03%. Report stereochemistry directly as coverage, precision and recall.

Figure 9.E/Z per-bond accuracy before and after inverting every double bondNumbers from the paper (J. Chem. Inf. Model., Sep 2026). Accent-hatched bars are the ones the card highlights.

No. 10 / 29 J. Chem. Theory Comput. Sep 2026 Paper

Improving ABFE Calculations for GPCRs by Addressing Errors in Apo-State Representation

Anand, Panescu, … Biggin

01 Problem: Absolute binding free energies can rank novel compounds without experimental anchors, but on membrane targets like GPCRs they systematically overestimate affinity.
02 Method: ABFE on A2A, D3 and mGlu5 receptors, targeting three error sources: poor pocket resolvation in apo-like states, metal-ion handling, and receptor relaxation. Fixes: a biased grand-canonical-like water transport scheme and a REUS correction for apo relaxation.
03 Result: Together the corrections reduce the systematic overestimation of binding affinities and improve agreement with experiment while keeping ranking performance.

3 GPCRs A2A, D3 and mGlu5, with apo-state errors corrected

No. 11 / 29 Nature Sep 2026 Paper

Chemist-aligned retrosynthesis by ensembling diverse inductive bias models

Maziarz, Liu, Pultar, … Segler

01 Problem: AI retrosynthesis still struggles with rare but strategically critical reactions and produces hallucinated routes that chemists don't trust.
02 Method: RetroChimera: two new models with complementary inductive biases, combined through a learned ensembling strategy, tested across several orders of magnitude of data scale.
03 Result: Outperforms leading baselines, stays robust outside its training data and learns from very few examples per reaction class. Organic chemists preferred its predictions over published reference reactions and other AI models, and it transferred zero-shot to internal data from two major pharma companies.

Chemists' pick preferred over published reference reactions and other AI models

No. 12 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid Models

Xu, Lam, Nakano & Katritch

01 Problem: Co-folding models look excellent on standard docking benchmarks, but how well do they generalise to genuinely new targets and chemotypes?
02 Method: A systematic benchmark of physics-based docking, hybrid rescoring models and co-folding approaches, stratified by how similar each test complex is to previously known ones.
03 Result: Co-folding wins on targets resembling its training data but drops to 20–40% success on novel dissimilar systems, consistent with memorisation. Physics-based and hybrid methods keep above 60% success even with minimal similarity to known complexes.

>60% vs 20–40% physics/hybrid vs co-folding success on novel, dissimilar complexes

No. 13 / 29 J. Cheminformatics Sep 2026 Paper

DockM8: an all-in-one open-source platform for consensus virtual screening in drug design

Lacour, Agha, Hirsch & Volkamer

01 Problem: Virtual screening results vary a lot with the docking program and scoring function chosen, and assembling a consensus pipeline by hand is tedious and hard to reproduce.
02 Method: One open-source workflow covering pocket detection, ligand preparation, docking, rescoring and consensus ranking: five docking engines, 17 pose-selection methods, 17 scoring functions and five consensus schemes, through a GUI or Python API.
03 Result: Median relative enrichments of 100% on DEKOIS 2.0, 82.6% on DUD-E and 8.9% on LIT-PCBA, often matching or beating state-of-the-art approaches. No single pose-selection or consensus strategy won everywhere, so workflows should be tuned per target.

Figure 13.Median relative enrichment by benchmarkNumbers from the paper (J. Cheminformatics, Sep 2026). Accent-hatched bars are the ones the card highlights.

No. 14 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Assessing the Generalizability of Machine Learning and Physics-Based Methods with DNA-Encoded Libraries

Dolorfino, Santos Perez, … Sztain

01 Problem: ML models trained on DNA-encoded library data look excellent in-distribution, but a NeurIPS competition showed they fail on out-of-distribution chemical space. Can structural modelling close that gap?
02 Method: A systematic comparison of ML, docking (Glide, Rosetta GALigandDock) and co-folding (Boltz-2) across three biologically diverse targets screened against libraries with multiple DEL synthesis formats.
03 Result: ML excels in-distribution, but out-of-distribution hit discrimination depends on both target and ligand context, with no single method consistently dominating. Benchmark performance alone does not predict OOD performance. Released as the open-source package DEL-iver.

No winner no single method dominated out-of-distribution across targets

No. 15 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge

Ebner, Bartmann, Topf, … Luukkonen

01 Problem: Tox21 was deep learning's 'ImageNet moment' for drug discovery, but when it was folded into MoleculeNet and OGB its labels were altered and imputed, so a decade of results can't be compared.
02 Method: A reproducible Hugging Face leaderboard using the original Tox21 Challenge data set, with baseline and representative methods evaluated under the same conditions.
03 Result: The original 2015 winner, the DeepTox ensemble, and descriptor-based self-normalising networks from 2017 still rank among the top methods, so measured progress over the last decade is far smaller than benchmark tables suggest.

2015 the original Tox21 winner is still among the top methods

No. 16 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Early-Enrichment Hit Discovery via Reversible-Work c(t) Estimation in Metadynamics (CTMD)

Adury, Tiwary, Gu & Shekhar

01 Problem: Docking hit lists are full of false positives, free energy calculations are too expensive to triage them, and co-folding shortcuts can be confounded by memorisation.
02 Method: CTMD ranks binding stability using the c(t) reversible-work estimator computed from a few short, independent well-tempered metadynamics runs, without needing converged binding free energies.
03 Result: Robust early enrichment across diverse targets and chemotypes with minimal parameter tuning. By contrast, Boltz-2's enrichment tracked similarity to its training set, a memorisation artefact the physics-based protocol avoids.

Physics > memory metadynamics triage resists the training-set bias seen in co-folding

No. 17 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Systematic Evaluation of Graph Neural Networks for Ligand-Based Virtual Screening on ChEMBL Datasets

Liu, Lin, Fan, … Hou

01 Problem: Ligand-based GNN performance depends heavily on data size, class imbalance and evaluation choices, yet most benchmarks hide this behind a handful of tidy data sets.
02 Method: GCN, GAT, GIN and GraphSAGE evaluated across 5,368 ChEMBL targets and 1.59 million activity records, comparing random and similarity-aware splits, single- and multi-task learning.
03 Result: GIN is best on targets with more than 100 samples (mean ROC–AUC up to 0.94), while simpler architectures hold up better when data are scarce. Similarity-aware splits drop AUC by more than 0.3 versus random splits. Multi-task learning stabilises small targets.

>0.3 AUC lost when random splits are replaced by similarity-aware ones

No. 18 / 29 J. Chem. Inf. Model. Sep 2026 Paper

Benchmarking Generative AI and Physics-Based Molecular Simulation for Sampling Conformational Heterogeneity in T4 Lysozyme

Bhakat

01 Problem: Generative structure models are increasingly sold as a shortcut to conformational ensembles. Do they actually recover the states that MD and experiment see?
02 Method: T4 lysozyme as a test case with a four-state model built on physically meaningful collective variables, comparing seven generative AI methods (AF-cluster, AlphaFlow, BioEmu and others), AI-accelerated simulation and physics-based enhanced MD.
03 Result: The generative methods largely sampled only the exposed/open state. AI-accelerated simulation, which feeds generative ensembles into physics-based MD, recovered all four states with populations matching enhanced MD and smFRET.

1 of 4 states generative AI mostly sampled; simulation-based methods recovered all four

No. 19 / 29 Nat. Rev. Drug Discov. Aug 2026 Paper

Artificial intelligence in drug discovery — what it is, where we stand and the path forward

Bender, Thomas, Scannell, … Cortes-Ciriano

01 Problem: A decade of interest has produced a wide variety of AI methods for drug discovery, developed, applied and benchmarked. The question that matters is whether any of it delivers safer and more efficacious medicines to patients faster.
02 Method: A Perspective by 16 authors from academia and industry that takes stock of where AI has and has not moved drug discovery: what drives the cost of a successful launch, how little the available proxy data predict, what makes chemical data hard to model, and why validating a model is not the same as improving the process it serves.
03 Result: Evidence of clinically relevant impact is, so far, disappointingly limited. The likely reasons: too little focus on clinical translation during model development, conditional life-science data, problems defined too loosely for real-world use, 'technology push' over 'science pull', and the time it takes to turn a capability into a scaled, usable system. The main recommendation: benchmarks should move on from model validation to whether a tool improves decision making.

Validation ≠ impact benchmarks should test whether a tool improves decisions, not only how well it validates

No. 20 / 29 arXiv Jul 2026 Paper

Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2

Sturmfels, Hiranuma, … Anishchenko

01 Problem: Co-folding networks built on co-evolutionary statistics do not transfer to peptide therapeutics, a modality full of non-canonical residues, macrocycles and unusual topologies.
02 Method: An all-atom diffusion transformer that drops residue-level tokens, extending Vilya-1 from single molecules to their complexes with protein targets, evaluated on Riptides, a new benchmark of complex protein–peptide interfaces.
03 Result: On Riptides, 59.1% of bound peptide backbones recovered below 2 Å RMSD at 1000 samples, against 40.9% for Boltz-2 and 22.0% for Schrödinger docking. On PoseX cross-docking, 71.1% success, falling to 66.5% when poses must also pass physical-validity checks.

Figure 20.Peptide backbone recovered below 2 Å RMSD, RiptidesNumbers from the paper (arXiv, Jul 2026). Accent-hatched bars are the ones the card highlights.

No. 21 / 29 bioRxiv Jul 2026 Paper

Kinase inhibitors can change protonation or tautomeric state upon binding

Ranepura, Chowdhury, Rosenzweig, … Gunner

01 Problem: Affinity depends on the protonation and tautomer states of both protein and ligand, yet docking and free energy pipelines usually commit to one state per molecule because few tools can treat the whole charge and tautomer ensemble at scale.
02 Method: Multi-Conformation Continuum Electrostatics (MCCE) applied over the Boltzmann ensemble of charge and tautomer states for nine kinase domains bound to 18 FDA-approved inhibitors.
03 Result: Protein net charge and proton distribution stay largely stable even when the inhibitor is charged, but individual inhibitor charges shift up or down depending on the target. Binding reshuffles the low-energy states (ΔG < 2.5 kcal/mol) without recruiting higher-energy ones, and tautomers of one inhibitor can differ by 3–6 kcal/mol in binding free energy, so a minority species in solution can dominate the bound state.

3–6 kcal/mol binding free energy gap between tautomers of the same inhibitor

No. 22 / 29 J. Cheminformatics Jun 2026 Paper

Count your bits: fingerprint benchmarking to assess broad chemical space representation

Huber & Pollmann

01 Problem: Tanimoto similarity on 2D fingerprints underpins virtual screening, nearest-neighbour search and model evaluation, yet its behaviour depends heavily on fingerprint type, binary versus count, and whether you fold to a fixed length. Most people use the defaults.
02 Method: A multi-criteria benchmark of dictionary-based, Morgan/FCFP, RDKit path, Atom Pair, MAP4, torsion, LINGO and Avalon fingerprints with several representation variants, probing specificity, score distributions, compound-size dependence, top-k ranking agreement and a graph-based reference.
03 Result: Count and log-count variants generally improve specificity and structural alignment, while folding-induced bit collisions badly distort similarities for high-occupancy fingerprints: unfolded variants matter for RDKit and are often necessary for MAP4. Released as the open-source library chemap.

Unfold it folding defaults distort high-occupancy fingerprints

No. 23 / 29 bioRxiv Jun 2026 Paper

Identifying and Addressing Systematic Data Leakage in Protein-Ligand Affinity Benchmarks

Mattsson & Walters

01 Problem: Co-folding models such as Boltz-2 are reported to match FEP accuracy, but if the benchmarks leak training data, those claims measure memorisation rather than biophysics.
02 Method: A meta-analysis of ChEMBL 36 for 'target mirroring', where homologous proteins with low sequence identity still show correlated binding profiles, plus a structure-blind ligand-only baseline run against the FEP+ and OpenFE benchmarks.
03 Result: Over 6,000 leaking assay pairs found, with leakage persisting down to 0.2 sequence identity. A ligand-only model with no protein structure reached r = 0.66 on FEP+. Their Novelty-Tiered Affinity Benchmark drops that baseline to r = 0.14 below 0.35 Tanimoto similarity.

Figure 23.Ligand-only baseline (no protein structure), Pearson rNumbers from the paper (bioRxiv, Jun 2026). Accent-hatched bars are the ones the card highlights.

No. 24 / 29 J. Chem. Inf. Model. Apr 2026 Paper

Struct2Query: Structure-Guided Virtual Screening via Composite-Molecule ROCS Queries Derived from Protein Pocket Similarity

Shmilovich, Suriana & Sresht

01 Problem: ROCS screens billions of molecules fast but needs a known active as the query, which biases hits toward close analogs. Docking avoids that but is too slow to deploy at the same scale.
02 Method: Protein pockets are turned into composite-molecule ROCS queries: SiteHopper searches over 78,000 crystallographic complexes for analogous pockets, transplants their ligands, and keeps all shape and colour features so dense regions emerge as hotspots.
03 Result: On DEKOIS 2.0 (81 targets) and DUDE-Z (43 targets), early enrichment matched or exceeded Glide and HYBRID docking at ligand-based throughput, with better chemotype coverage than single-ligand ROCS for three of four purchasable libraries.

78,000+ crystallographic pockets mined to build each query

No. 25 / 29 J. Chem. Inf. Model. Dec 2025 Paper

Toward Automated Physics-Based Absolute Drug Residence Time Predictions

Smith, Branduardi, … Krilov

01 Problem: Residence time matters as much as affinity for a drug's profile, but predicting it usually means hand-tuning enhanced sampling for every system.
02 Method: A two-stage scheme: Random Acceleration MD harvests plausible egress pathways, then Infrequent Metadynamics estimates the residence time, aimed at the accuracy–throughput–usability compromise real programmes face.
03 Result: Benchmarked on five diverse targets with congeneric series, 29 drug–target complexes in total: RMSE 1.22 and R² 0.80 in log₁₀(τ), without manually tuning the enhanced sampling parameters.

1.22 RMSE in log₁₀(τ) across 29 complexes, no manual tuning

No. 26 / 29 Nature Chemistry Nov 2023 Paper

Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning

Nippa, Atz, … Schneider

01 Problem: Late-stage functionalisation is the cheap way to tune a drug candidate, but the chemical complexity of real drug molecules makes site-selective C–H diversification hard to predict.
02 Method: A platform pairing high-throughput borylation screening with geometric deep learning, trained on an informer library of diverse drug molecules, plus a simple user-friendly reaction format for sharing the data.
03 Result: Yields predicted to 4–5% mean absolute error. Reactivity of novel reactions classified at 92% balanced accuracy for known substrates and 67% for unknown. Regioselectivity of major products captured at F-score 67%. Applied to 23 commercial drugs, it found many diversification opportunities.

Figure 26.Borylation model: reactivity accuracy, regioselectivity F-scoreNumbers from the paper (Nature Chemistry, Nov 2023). Accent-hatched bars are the ones the card highlights.

No. 27 / 29 J. Phys. Chem. Lett. Feb 2023 Paper

Geometric Interaction Graph Neural Network for Predicting Protein–Ligand Binding Affinities from 3D Structures (GIGN)

Yang, Zhong, Lv, … Chen

01 Problem: Most ML affinity predictors ignore the 3D structure of the complex and the physical interactions between protein and ligand, the very things that explain binding.
02 Method: A heterogeneous interaction layer that unifies covalent and noncovalent interactions in the message-passing phase, built to be invariant to translation and rotation of the complex so no data augmentation is needed.
03 Result: State-of-the-art performance on three external test sets, and visualising the learned representations of complexes shows the predictions are biologically meaningful.

No augmentation rotation and translation invariance built into the layer

No. 28 / 29 J. Chem. Inf. Model. Jan 2022 Paper

Interpretation of Structure–Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial Intelligence

Harren, Matter, Hessler, … Grebner

01 Problem: Deep neural networks predict activity well, but their black-box character hides which structural features drive it, and that is exactly what a medicinal chemist needs to guide the next round.
02 Method: Multiple explainable-AI methods applied and compared on lead optimisation data sets with well-established SARs and available X-ray structures, with a visualisation scheme based on atom-level heatmaps.
03 Result: Combining DNN models with strong interpretation methods gave comprehensive, understandable explanations, with SHAP-based methods the most promising. The authors stress that every interpretation is only meaningful in the context of its model and training data.

SHAP the most promising family for reading SAR out of a neural network

No. 29 / 29 J. Chem. Inf. Model. Dec 2018 Paper

A Distance-Based Boolean Applicability Domain for Classification of High Throughput Screening Data

Berenger & Yamanishi

01 Problem: A QSAR model needs an applicability domain as well as an activity model, but existing methods are complex, hard to interpret, or make the user pick a threshold by hand.
02 Method: A fully automatic, trivial-to-interpret Distance-Based Boolean Applicability Domain for category QSAR, with no user-selected threshold and no embedded empirical constant.
03 Result: In retrospective experiments on high-throughput screening data sets, it improved classification performance and early retrieval for both SVM and random forest classifiers, while improving scaffold diversity among the top-ranked actives.

0 knobs no user-chosen threshold, no embedded constant

All 29 cards

Newest first. Each title opens its card; Paper ↗ on the card opens the paper.

01Syndirella: synthesis-directed fragment elaboration enables extensive and cost-effective binding site exploration beyond catalog compoundsChem. Sci. · Oct 2026 · chart
02Subpocket-Aware Fragment-Based Docking Pipeline for Kinase Ligand Design: Methodology and a PKA Case StudyJ. Chem. Inf. Model. · Oct 2026
03Building a Physically Interpretable Compressed Chemical Space via Grid-Based OptimizationJ. Chem. Inf. Model. · Oct 2026
04Evaluation of Physics-Based Pose Prediction and the Role of Receptor Flexibility in a Community Antiviral Blind ChallengeJ. Chem. Inf. Model. · Oct 2026
05Information Leakage in Enzyme Substrate PredictionBioinformatics · Oct 2026
06Validation Strategy Dominates Model Choice in Chronological QSAR EvaluationJ. Chem. Inf. Model. · Oct 2026 · chart
07DESPOT: Direction-Enhanced Scoring POTentials for Scoring of Protein–Ligand InteractionsJ. Chem. Inf. Model. · Oct 2026
08AcepKa: Thermodynamics-Informed pKa Prediction and Protonation-State Generation in PlayMolecule AIJ. Chem. Inf. Model. · Oct 2026
09Fréchet ChemNet Distance Is Blind to Double-Bond StereochemistryJ. Chem. Inf. Model. · Sep 2026 · chart
10Improving ABFE Calculations for GPCRs by Addressing Errors in Apo-State RepresentationJ. Chem. Theory Comput. · Sep 2026
11Chemist-aligned retrosynthesis by ensembling diverse inductive bias modelsNature · Sep 2026
12Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid ModelsJ. Chem. Inf. Model. · Sep 2026
13DockM8: an all-in-one open-source platform for consensus virtual screening in drug designJ. Cheminformatics · Sep 2026 · chart
14Assessing the Generalizability of Machine Learning and Physics-Based Methods with DNA-Encoded LibrariesJ. Chem. Inf. Model. · Sep 2026
15Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 ChallengeJ. Chem. Inf. Model. · Sep 2026
16Early-Enrichment Hit Discovery via Reversible-Work c(t) Estimation in Metadynamics (CTMD)J. Chem. Inf. Model. · Sep 2026
17Systematic Evaluation of Graph Neural Networks for Ligand-Based Virtual Screening on ChEMBL DatasetsJ. Chem. Inf. Model. · Sep 2026
18Benchmarking Generative AI and Physics-Based Molecular Simulation for Sampling Conformational Heterogeneity in T4 LysozymeJ. Chem. Inf. Model. · Sep 2026
19Artificial intelligence in drug discovery — what it is, where we stand and the path forwardNat. Rev. Drug Discov. · Aug 2026
20Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2arXiv · Jul 2026 · chart
21Kinase inhibitors can change protonation or tautomeric state upon bindingbioRxiv · Jul 2026
22Count your bits: fingerprint benchmarking to assess broad chemical space representationJ. Cheminformatics · Jun 2026
23Identifying and Addressing Systematic Data Leakage in Protein-Ligand Affinity BenchmarksbioRxiv · Jun 2026 · chart
24Struct2Query: Structure-Guided Virtual Screening via Composite-Molecule ROCS Queries Derived from Protein Pocket SimilarityJ. Chem. Inf. Model. · Apr 2026
25Toward Automated Physics-Based Absolute Drug Residence Time PredictionsJ. Chem. Inf. Model. · Dec 2025
26Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learningNature Chemistry · Nov 2023 · chart
27Geometric Interaction Graph Neural Network for Predicting Protein–Ligand Binding Affinities from 3D Structures (GIGN)J. Phys. Chem. Lett. · Feb 2023
28Interpretation of Structure–Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial IntelligenceJ. Chem. Inf. Model. · Jan 2022
29A Distance-Based Boolean Applicability Domain for Classification of High Throughput Screening DataJ. Chem. Inf. Model. · Dec 2018