Patent Drafting Analysis of KUANO LTD.’s Machine Learning Drug-Like Molecule Analysis | US 12,211,592 B2

IP Drafting Analysis · US 12,211,592 B2

Patent Drafting Analysis of KUANO LTD.'s Quantum Graph Machine Learning for Drug-Like Molecules | US 12,211,592 B2

Q: What does US 12,211,592 B2 protect?

US 12,211,592 B2 protects a machine learning based method and system for modelling a thermodynamic ensemble or representation of a drug-like molecule. The patent solves the problem of inaccurate molecular representations in ML-based drug discovery by synthetically generating samples of a drug-like molecule's thermodynamic ensemble where the ensemble is represented as a quantum graph — a molecular graph in which each node corresponds to a molecular orbital and each edge corresponds to a type of quantum correlation between those orbitals — and feeding that quantum graph into a machine learning system to predict properties such as binding affinity, lipophilicity, solubility, melting point, or pKa.

Q: Who owns patent US 12,211,592 B2?

The patent is owned by KUANO LTD., a company based in London, United Kingdom. The inventors are Hagen Triendl (London, GB), Matthias Bal (London, GB), Jarvist Moore Frost (London, GB), Lawrence Phillips (London, GB), Agisilaos Chantzis (London, GB), Graham Simpson (London, GB), Vic Stojevic (London, GB), Noor Shaker (London, GB), Michael Craig (London, GB), Usman Bashir (London, GB), and Mariana Assmann (London, GB).

Q: What are the independent claims of US 12,211,592 B2?

Claim 1 is a method claim covering the machine learning based method of modelling a thermodynamic ensemble of a drug-like molecule by synthetically generating a quantum graph sample and inputting it into a machine learning system. Claim 26 is a system claim covering a machine learning based system configured to receive and process a synthetically generated quantum graph sample of the thermodynamic ensemble or molecular orbital representation of a drug-like molecule. Claim 27 is a product-by-process claim covering a molecule or class of drug-like molecules identified using the method of Claim 1, requiring the same quantum graph orbital-correlation structure.

Q: What technology does this patent cover?

This patent covers a new way of representing drug molecules for machine learning. Traditional drug discovery software represents molecules as simple chemical graphs where atoms are nodes and chemical bonds are edges. KUANO's approach goes deeper: it uses quantum mechanical calculations to build a 'quantum graph' where each node represents a molecular orbital and each edge encodes the quantum correlation between those orbitals, capturing information about how electrons are actually distributed in the molecule. This richer representation is then fed into a neural network to predict drug-relevant properties like how strongly a molecule binds to a protein target, how soluble it is, or how it might be metabolised.

Q: What IPC classifications does this patent use?

G16C 20/70 (2019.01) — Cheminformatics, i.e. methods or systems for solving problems in the field of chemistry, including drug design and molecular simulations. G06N 3/02 (2006.01) — Computer systems based on biological models, i.e. neural networks. G06N 10/20 (2022.01) — Quantum computing. G16C 20/50 (2019.01) — Molecular design or target identification.

A structural and strategic analysis of US 12,211,592 B2, examining claim architecture, dependent claim fallback quality, §101 eligibility risk, and prosecution positioning for KUANO's quantum-inspired ML drug discovery platform.

US 12,211,592 B2Filed: Jul 17, 2019Granted: Jan 28, 2025G16C 20/70G06N 3/02G06N 10/20G16C 20/50

Draft a Patent Like This Explore Related Patents

Spec Words

38,500

Across 9 sections plus appendices

Draft now ↗

Total Claims

3 independent · 24 dependent

Draft now ↗

Figure Sheets

System diagrams, quantum equations, ML benchmarks

Draft now ↗

Published by PatSnap Insights Team · June 4, 2026 · 13 min read Verified by PatSnap Eureka Data

Overview

Structural Overview

The detailed description dominates at approximately 77% of total words, spanning 9 thematic sections and multiple appendices covering quantum chemistry, graph neural networks, conformational sampling, and generative models — indicating an exceptionally broad technical disclosure relative to the 27 claims actually granted. The claim set is narrow in architectural scope: 3 independent claims all cover method, system, and molecule-identification variants of a single core concept, with 24 dependent claims adding specific property prediction and ML architecture limitations. The 88 figure sheets are heavily weighted toward mathematical equations, performance benchmarks, and Q-Graph visualisations, with strong support for the core quantum graph representation concept but sparse dedicated structural diagrams for some dependent claim limitations.

Section Word Distribution

↗ Click bars to explore

Figure Inventory — 88 Sheets

Figure	Description	Role
FIG. 1	High-level diagram of the GTN platform showing data curation, featurization, generative model, predictive model, and MPO pipeline components.Search in Eureka ↗	System architecture
FIG. 2	Illustration of a small molecule conformation in a protein pocket, showing the progression from chemical graph to conformation to Q-Graph and tensor network representations.Search in Eureka ↗	Key embodiment
FIG. 3A	Schematic diagram comparing ligand-based and structure-based typical workflows, showing conformational sampling, featurization, Q-Graph representation, and machine learning steps.Search in Eureka ↗	Flow diagram
FIG. 3B	Diagram illustrating quantum-inspired machine learning in the context of classical, quantum-inspired, and quantum ML technology landscape across hardware types.Search in Eureka ↗	Other
FIG. 3C	Diagram illustrating GTN's technology capabilities mapped against drug discovery challenges at different biological scales and entanglement levels.Search in Eureka ↗	Other
FIG. 4	Schematic of a graph pooling step showing dark (kept) and white (dropped) nodes, edges being coarse-grained, and new effective edges being created between kept nodes.Search in Eureka ↗	Claim support
FIG. 5	Table with literature results for MoleculeNet benchmarks comparing RMSE and ROC-AUC results for ESOL, Lipophilicity, BBBP, and HIV datasets across multiple models.Search in Eureka ↗	Other
FIG. 6	Table showing speed-up of pooling runs of the HIV dataset using SimplePooling at different pooling keep ratios.Search in Eureka ↗	Other
FIG. 7	Table summarising graph-convolutional model hyperparameters (node channels, edge channels) used across ESOL, Lipophilicity, BBBP, and HIV datasets.Search in Eureka ↗	Other
FIG. 8	Table showing multi-task and single-task test set evaluation R2 results for PCE, GAP, and HOMO properties at different pooling ratios.Search in Eureka ↗	Other
FIG. 9	Visualisation of original molecular graphs before pooling for a batch of molecules from the CEP-2017 dataset.Search in Eureka ↗	Claim support
FIG. 10	Visualisation of coarse-grained graphs after the first pooling layer, showing the model grouping rings and identifying molecular backbones.Search in Eureka ↗	Claim support
FIG. 11	Visualisation of coarse-grained graphs after the second pooling layer, showing molecular graphs reduced to basic components connected by chains.Search in Eureka ↗	Claim support
FIG. 12	Section 4 equations (1)–(4) showing the molecular Hamiltonian and electronic Schrödinger equation formulations used for quantum featurisation.Search in Eureka ↗	Other
FIG. 13	Equations (5) and (6) showing the Hartree-Fock wave function as a Slater determinant of one-electron spin-orbital wave functions.Search in Eureka ↗	Other
FIG. 14	Representation of the Hartree-Fock orbital occupation (equation 7) showing occupied and virtual orbital spaces.Search in Eureka ↗	Other
FIG. 15	Equations (8)–(10) showing CI and coupled cluster wave function expansions including CI expansion and exponential coupled cluster ansatz.Search in Eureka ↗	Other
FIG. 16	Sketch of configurations and weights in a wave function expansion, distinguishing static (box 161) and dynamic (box 162) correlation contributions.Search in Eureka ↗	Other
FIG. 17	Representation of a multi-configurational self-consistent field (MC-SCF) approach showing occupied and virtual orbital occupation patterns.Search in Eureka ↗	Other
FIG. 18	Equations (11)–(16) showing the N-particle density matrix, one- and two-particle reduced density matrices used in quantum featurisation.Search in Eureka ↗	Other
FIG. 19	Equations (17)–(19) and (21) showing one-particle reduced density matrix formulations in natural orbital and natural spin-orbital bases.Search in Eureka ↗	Other
FIG. 20	Morphine molecule with relief map of electron density in the aromatic ring plane and chemical structural formula showing hydrogen bonding interactions.Search in Eureka ↗	Key embodiment
FIG. 21	Representation of bond and ring critical points for 5-formyl-2-(3-fluoro-4-bromobutadienyl)thiazole with gradient field of electron density and atomic basin partitioning.Search in Eureka ↗	Key embodiment
FIG. 22	Equations (22)–(26) showing electron density, pair density, and exchange-correlation density formulations for QTAIM analysis.Search in Eureka ↗	Other
FIG. 23	Equations (27)–(33) showing QTAIM localization/delocalization index formulations including weight functions and atomic overlap matrix.Search in Eureka ↗	Other
FIG. 24	Equations (34)–(37) showing delocalization index (DLI) and localization index (LI) formulations defining the Q-Graph edge and node features.Search in Eureka ↗	Claim support
FIG. 25	Equations (38)–(42) showing total DLI and LI formulations, relating electron counts in atomic basins to delocalization between atoms.Search in Eureka ↗	Other
FIG. 26	Workflow diagram showing the Python/bash pipeline from RDKit SDF/PDB input through DFT calculation (ORCA/Psi4), Multiwfn wavefunction analysis, to GraphConv featurisation.Search in Eureka ↗	Flow diagram
FIG. 27	Example molecule (equation 43) showing an acrylonitrile-type structure with carbon, hydrogen, oxygen, and nitrogen atoms used for Q-Graph calculation demonstration.Search in Eureka ↗	Key embodiment
FIG. 28	Comparison plot of calculated delocalization indices versus adjacency matrix for the molecule of FIG. 27, showing both linear and log-scale colour maps.Search in Eureka ↗	Claim support
FIG. 29	Delocalization indices versus adjacency matrix for the molecule of FIG. 27 shown in logarithmic scale revealing weaker long-range correlations.Search in Eureka ↗	Claim support
FIG. 30	Comparison of delocalization indices and adjacency matrix for ligand 404 bound to protein 3pj8, showing atom-index matrix representations.Search in Eureka ↗	Claim support
FIG. 31	Graph representation of (de)localization indices and chemical bonds for ligand 404, with edge weights based on DI magnitude and node sizes scaled to LI values.Search in Eureka ↗	Key embodiment
FIG. 32	Linear and log plots of histograms of delocalization index values for ligands in the lipophilicity dataset, showing distribution characteristics.Search in Eureka ↗	Other
FIG. 33	Linear and log plots of histograms of delocalization index values for ligands in the PDBbind17 subset.Search in Eureka ↗	Other
FIG. 34	Equation (46) showing the reduced density gradient (RDG) formula S(r) used for non-covalent interaction identification in the Q-Graph featurisation.Search in Eureka ↗	Other
FIG. 35	Table with density gradient, density, and RDG values for different molecular regions (nuclei, bonds, weak interactions, boundary of molecule) from the Multiwfn manual.Search in Eureka ↗	Other
FIG. 36	Schematic showing RDG versus sign(λ2)ρ(r) interaction regions identifying H-bonds, van der Waals interactions, and steric effects from electron density analysis.Search in Eureka ↗	Other
FIG. 37	NCI scatterplot showing reduced density gradient versus sign(λ2)ρ for a molecule, with annotated H-bond, vdW, and steric interaction regions.Search in Eureka ↗	Key embodiment
FIG. 38	Isosurfaces of the RDG map coloured by sign(λ2)ρ showing H-bond (blue), vdW (green), and steric (red) interaction regions in 3D molecular space.Search in Eureka ↗	Key embodiment
FIG. 39	Equations (50) and (51) showing promolecular density and electrostatic interaction potential formulas used for approximate non-covalent interaction calculation.Search in Eureka ↗	Other
FIG. 40	Bar charts showing DTNN results comparing combined scores and dipole moment only scores across different graph featurisation types (Graph, Di, Dist+Di+Li, DTNN, Dist+Li).Search in Eureka ↗	Other
FIG. 41	Plots showing absolute DFT energies and energy differences for 2694 random QM9 molecules comparing GTN calculations versus QM9 dataset values.Search in Eureka ↗	Other
FIG. 42	Plots showing dipole moments and dipole moment differences for QM9 molecules comparing GTN DFT calculations versus QM9 reference values.Search in Eureka ↗	Other
FIG. 43	Equation (52) reproducing the CAS-CI wave function as a sum over occupation number vectors, foundational to the QC-DMRG tensor network approach.Search in Eureka ↗	Other
FIG. 44	Representation of FCI and MPS (matrix product state) ansatz in terms of local tensors, illustrating DMRG tensor network decomposition.Search in Eureka ↗	Other
FIG. 45	Equation (53) showing the MPS truncation error as the norm difference between exact and approximate wave functions, parameterised by singular values.Search in Eureka ↗	Other
FIG. 46	Diagram showing DMRG as part of a wider quantum chemistry ecosystem for treating quantum active sites within a classically sampled protein environment.Search in Eureka ↗	System architecture
FIG. 47	Equations (54)–(56) showing the one-orbital and two-orbital von Neumann entropy and mutual information formulas used for orbital entanglement measures in QC-DMRG.Search in Eureka ↗	Other
FIG. 48	Free energy diagram as a function of binding reaction coordinate showing ΔGon, ΔGoff, and ΔGD for ligand-protein binding thermodynamics.Search in Eureka ↗	Key embodiment
FIG. 49	Schematic diagram of a predictive model trained on a conformational ensemble, showing representative conformations from MD/docking feeding into a predictive output.Search in Eureka ↗	Claim support
FIG. 50	Schematic depicting two approaches for designing predictive models: path A (predict quantum observables then calibrate to experiment) and path B (direct ML prediction).Search in Eureka ↗	Claim support
FIG. 51	t-SNE plots, CDK9 binding affinity distribution, and Tanimoto similarity distribution showing novel IP space exploration for CDK-9 molecule generation.Search in Eureka ↗	Key embodiment
FIG. 52	Scatter plot of CDK1 pIC50 versus CDK9 pIC50 for generated molecules, showing selectivity quadrants with a novel CDK9-selective compound highlighted.Search in Eureka ↗	Key embodiment
FIG. 53	Chemical structures of three generated CDK9-selective compounds (compounds 1, 2, 3) not seen in training data, demonstrating platform generative capability.Search in Eureka ↗	Key embodiment
FIG. 54	Comparison of standard chemical graph adjacency matrix and Q-Graph matrix representation for a QM9 molecule, highlighting additional long-range electron correlation edges.Search in Eureka ↗	Claim support
FIG. 55	Bar charts comparing predictive model R² performance on dipole moment and HOMO tasks using Graph vs Q-Graph featurisations in single and multi-task settings.Search in Eureka ↗	Other
FIG. 56	Bar charts comparing LUMO prediction performance (R²) using different featurisations including Graph vs Q-Graph in multi-task setting at various training dataset sizes.Search in Eureka ↗	Other
FIG. 57	Bar charts comparing dipole moment prediction performance (R²) using Graph vs Q-Graph featurisations in multi-task setting across training dataset sizes.Search in Eureka ↗	Other
FIG. 58	Bar charts comparing prediction performance averaged across all QM9 tasks using Graph vs Q-Graph featurisations in multi-task setting.Search in Eureka ↗	Other
FIG. 59	Bar charts comparing HOMO prediction performance using Graph 3D, Q-Graph 3D (local), and Graph 3D featurisations in multi-task setting.Search in Eureka ↗	Other
FIG. 60	Bar charts comparing LUMO prediction performance using Graph 3D, Q-Graph 3D (local), and Graph 3D featurisations in multi-task setting.Search in Eureka ↗	Other
FIG. 61	Bar charts comparing dipole moment prediction performance (multitask) using Graph 3D, Q-Graph 3D (local), and Graph 3D featurisations.Search in Eureka ↗	Other
FIG. 62	Bar charts comparing dipole moment prediction performance (single task) using Graph 3D, Q-Graph 3D (local), and Graph 3D featurisations.Search in Eureka ↗	Other
FIG. 63	Bar charts comparing prediction performance on all tasks averaged (multitask) using Graph 3D, Q-Graph 3D (local), and Graph 3D featurisations.Search in Eureka ↗	Other
FIG. 64	Representation of P450 enzyme catalytic cycle showing the heme-based Redox mechanism with compound I and compound 0 intermediates for site of metabolism prediction.Search in Eureka ↗	Other
FIG. 65	Results table for binding affinity dataset size experiment showing R² Pearson correlation results for different model architectures at training set sizes of 20–160.Search in Eureka ↗	Other
FIG. 66	Scatter plots showing edge complex R² results as a function of training set size for binding affinity prediction experiments.Search in Eureka ↗	Other
FIG. 67	Scatter plots showing edge ligand based R² results as a function of training set size.Search in Eureka ↗	Other
FIG. 68	Scatter plots showing edge ligand only scatter max R² results as a function of training set size.Search in Eureka ↗	Other
FIG. 69	Scatter plots showing pocket attention pair message R² results as a function of training set size for binding affinity prediction.Search in Eureka ↗	Other
FIG. 70	Scatter plots showing pocket gating pair message R² results as a function of training set size for binding affinity prediction.Search in Eureka ↗	Other
FIG. 71	Scatter plots showing edge ensemble ligand only (10 poses per eg) R² results as a function of training set size.Search in Eureka ↗	Other
FIG. 72	Scatter plots showing edge ensemble ligand only (2 poses per eg) R² results as a function of training set size.Search in Eureka ↗	Other
FIG. 73	Results table for binding affinity experiment based on client split showing R² Pearson correlation for different model architectures at training sizes 20–160.Search in Eureka ↗	Other
FIG. 74	Scatter plots showing Edge ligand based results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 75	Scatter plots showing Edge ligand only results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 76	Scatter plots showing pair message complex results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 77	Scatter plots showing pocket attention pair message results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 78	Scatter plots showing pocket gating pair message results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 79	Scatter plots showing edge ensemble ligand only (10 poses) results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 80	Scatter plots showing edge ensemble ligand only (2 poses) results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 81	Scatter plots showing edge ensemble pocket (2 poses) results for binding affinity (client split) as a function of training set size.Search in Eureka ↗	Other
FIG. 82	Summary table of results for edge ligand only and edge ensemble ligand only experiments at training sizes 20–160 samples.Search in Eureka ↗	Other
FIG. 83	Histogram diagram showing log P dataset targets (avdeef, martel, chembl) distribution against measured logP values.Search in Eureka ↗	Other
FIG. 84	Python code listing showing the RDKit baseline script for computing log P descriptors from SMILES and evaluating prediction metrics.Search in Eureka ↗	Other
FIG. 85	Python code listing showing the filter_other_dataset_results.json script for aggregating RMSE results across datasets and models.Search in Eureka ↗	Other
FIG. 86	Table summarising pharma applications of quantum and conformational features across developability properties (melting point, solubility, lipophilicity, pKa, permeability, mutagenicity, NMR, metabolic).Search in Eureka ↗	Other

Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent

Claims

Claim Architecture Analysis

The claim set contains 3 independent claims: Claim 1 (method), Claim 26 (system), and Claim 27 (molecule-identification), all drawn to a single inventive concept — the machine learning-based modelling of a drug-like molecule's thermodynamic ensemble via a quantum graph representation. The dependent:independent ratio of 8:1 is below the typical 10–15:1 norm seen in deep-tech ML patents in the G16C/G06N IPC space, suggesting missed fallback opportunities. The tripartite structure of Claims 1, 26, and 27 provides method, system, and product-by-process coverage, but the absence of any CRM (computer-readable medium) claim type leaves a significant enforcement gap in software distribution scenarios.

Core inventive concept: Claims 1, 26, and 27 solve the problem of inaccurate molecular representations in drug discovery ML by synthetically generating a sample of a thermodynamic ensemble or representation of a drug-like molecule — specifically one that is "a molecular orbital representation or quantum graph representation" in which "each node corresponds to a molecular orbital and edges correspond to a type of quantum correlation between molecular orbitals" — and inputting that quantum graph into a machine learning system, replacing standard chemical graph or SMILES-based representations with quantum-mechanically derived graph structures.

Independent Claim Dissection

Claim	Preamble	Transition	Key Body Elements
Claim 1	A machine learning based method of modelling a thermodynamic ensemble or representation of a drug-like molecule	comprising	synthetically generating a sample of a thermodynamic ensemble or representation, said ensemble being a molecular orbital representation or quantum graph representation; inputting said sample into a machine learning system; quantum graph representation being a molecular graph representation in which each node corresponds to a molecular orbital and edges correspond to a type of quantum correlation between molecular orbitalsSearch prior art ↗
Claim 26	A machine learning based system configured to model a thermodynamic ensemble or representation of a drug-like molecule	comprising	a machine learning based system configured to receive and process a synthetically generated sample of the thermodynamic ensemble or representation being a molecular orbital representation or quantum graph representation; quantum graph representation in which each node corresponds to a molecular orbital and edges correspond to a type of quantum correlation between molecular orbitalsSearch prior art ↗
Claim 27	A molecule or class of drug-like molecules identified using a machine learning based method of modelling a thermodynamic ensemble or representation of a drug-like molecule	comprising	the method in which a sample of the thermodynamic ensemble is synthetically generated and inputted into a machine learning system, the ensemble being a molecular orbital representation or quantum graph representation; quantum graph representation in which each node corresponds to a molecular orbital and edges correspond to quantum correlation between molecular orbitalsSearch prior art ↗

Claim Dependency Tree

1 Method of modelling thermodynamic ensemble of drug-like molecule using ML — synthetic generation of quantum graph sample inputted to ML systemSearch Claim 1 prior art ↗

2 Adds: every element of the thermodynamic ensemble can be represented as Q-graph, tensor network, or molecular orbital representationSearch in Eureka ↗

3 Adds: quantum graph representation is the molecular graph representation obtained from quantum mechanical calculationsSearch in Eureka ↗

4 Adds: quantum graph representation depends on a conformational state of the drug-like moleculeSearch in Eureka ↗

5 Adds: molecular orbital representation is a tensor network representation of molecular quantum statesSearch in Eureka ↗

6 Adds: synthetic generation of samples is based on a thermodynamic quantitySearch in Eureka ↗

7 Adds: ML system configured to output thermodynamic quantity based on approximate expectation value over the entire or representative set of the modelled thermodynamic ensembleSearch in Eureka ↗

8 Adds: ML system configured to learn distribution of Boltzmann weights of the entire or representative set of the thermodynamic ensembleSearch in Eureka ↗

9 Adds: determining the cost function or backpropagation is based on thermodynamic quantity to be outputtedSearch in Eureka ↗

10 Adds: size of synthetically generated sample of thermodynamic ensemble is tuned depending on downstream applicationSearch in Eureka ↗

11 Adds: ML system is a graph convolutional neural networkSearch in Eureka ↗

12 Adds: synthetically generated sample of the thermodynamic ensemble is inputted as a molecular graphSearch in Eureka ↗

13 Adds: ML system configured to output any quantity that is a function of the thermodynamic ensemble or representationSearch in Eureka ↗

14 Adds: ML system used to predict ligand protein binding affinity with synthetically generated ensemble samples of ligand in solution, protein in solution, and ligand-protein complexSearch in Eureka ↗

15 Adds: ML system used to predict ligand protein inhibition concentration with ensemble samples of ligand in solution and ligand-protein complexSearch in Eureka ↗

16 Adds: ML system used to predict lipophilicity with ensemble samples of unionized and/or ionized state in octanol and waterSearch in Eureka ↗

17 Adds: ML system used to predict thermodynamic solubility with ensemble samples of solid state and dissolved stateSearch in Eureka ↗

18 Adds: ML system used to predict kinetic solubility with ensemble samples of amorphous solid state and dissolved stateSearch in Eureka ↗

19 Adds: ML system used to predict melting point with ensemble samples of solid state and molecule in octanolSearch in Eureka ↗

20 Adds: ML system used to predict acidity (pKa) with ensemble samples in appropriate environment (water or protein pocket)Search in Eureka ↗

21 Adds: ML system uses GAN, VAE, or GCPN style modelsSearch in Eureka ↗

22 Adds: ML system uses generative models to learn new thermodynamic ensemble for which data is not availableSearch in Eureka ↗

23 Adds: ML system implements weight sharing when multiple generated ensemble samples are inputtedSearch in Eureka ↗

24 Adds: docking is used to generate the sample of the thermodynamic ensembleSearch in Eureka ↗

25 Adds: docking enhanced by molecular dynamics is used to generate the sampleSearch in Eureka ↗

26 System configured to model thermodynamic ensemble of drug-like molecule — quantum graph representation nodes correspond to molecular orbitalsSearch Claim 26 prior art ↗

27 Molecule or class of drug-like molecules identified using the method of Claim 1 with quantum graph orbital-correlation node/edge structureSearch Claim 27 prior art ↗

Metric	This Application	AI/ML Drug Discovery Norm
Total claims	27	20 – 35
Independent claim count	3	2 – 5
Dependent : Independent ratio	8.00 : 1	6 – 12 : 1
Method claims present?	Yes — Claim 1	Always
System / apparatus claims?	Yes — Claim 26	Common

Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent

Drafting Quality

Drafting Quality Signals

The patent demonstrates strong technical disclosure depth — the specification devotes approximately 9 thematic sections and multiple appendices to supporting the Q-graph concept, and the core claim language in Claims 1, 26, and 27 is precisely crafted around the quantum orbital/edge correlation structure. However, the claim set carries meaningful §101 vulnerability because the patent's core advance is a data representation and mathematical method, and the hardware tie-in relies solely on an implied 'machine learning system' without reciting specific processor or hardware limitations.

✅

Antecedent Basis

The claim language is largely clean with respect to antecedent basis. The key term 'the thermodynamic ensemble' in Claims 2–25 correctly finds its antecedent in the 'a thermodynamic ensemble or representation' introduced in Claim 1's preamble. The term 'the quantum graph representation' in Claims 2–5 and the dependent claims of 26 finds its antecedent in the body of Claims 1 and 26 respectively. No orphaned 'the [element]' references were identified across the 27 claims.

Explore in Eureka ↗

✅

Spec–Claim Consistency

The core limitation that the quantum graph representation has 'each node corresponding to a molecular orbital and edges corresponding to a type of quantum correlation between molecular orbitals' is directly supported by Section 2 (Quantum Graph), FIG. 2 (Q-Graph visualisation), FIG. 31 (graph representation of DLI/LI indices), and FIG. 54 (comparison of standard graph vs Q-Graph adjacency matrices). The thermodynamic ensemble sampling limitation of Claim 1 maps to Section 7 (Machine Learning on Thermodynamic Ensembles) and FIG. 49 (ensemble conformations schematic). Coverage is strong for the independent claim core.

Explore in Eureka ↗

✅

Transition Word Usage

All three independent claims use 'comprising' as the transition word, which is the broadest available choice and is strategically appropriate here — it permits the ML system to incorporate additional steps (e.g., post-processing, output formatting) without defeating infringement. The choice is consistent across Claims 1, 26, and 27, avoiding any inconsistency that would invite narrow construction. No 'consisting of' or 'consisting essentially of' transitions are used in independent claims, which is correct given the complex, multi-component nature of the method.

Explore in Eureka ↗

⚠️

§112(f) Means-Plus-Function Risk

Claim 1 recites 'a machine learning system' without further structural definition in the independent claim body — this functional label, while not using 'means for' language, could be construed under §112(f) if a court finds it lacks structural meaning in the art. The dependent claims partially mitigate this by specifying 'graph convolutional neural network' (Claim 11), GAN/VAE/GCPN (Claim 21), and generative model (Claim 22), but these are only in dependent claims. A stronger filing would have recited at least one structural neural network architecture in the independent claim or provided an explicit definition of 'machine learning system' in the specification to preclude §112(f) treatment.

Explore in Eureka ↗

⚠️

§101 Eligibility Risk

Claims 1, 26, and 27 carry meaningful Alice/Mayo exposure because the independent claims recite a mathematical/computational method (quantum graph construction and ML processing) applied to data representing molecules — a paradigmatic abstract idea under step 2A prong 1. The §101 defense rests primarily on the 'synthetically generating a sample' limitation and the quantum graph orbital structure, which could be characterized as a novel technical implementation rather than mere data manipulation; however, the examiner cited Stojevic 2021/0081804 suggesting §101 scrutiny was applied. The granted claims survived examination, providing some prosecution history support, but litigation validity under Alice remains a credible risk without hardware recitation.

Explore in Eureka ↗

✅

Dependent Claim Fallback Quality

Claims 14–20 add genuinely distinct property-specific fallbacks (binding affinity, IC50, lipophilicity, thermodynamic solubility, kinetic solubility, melting point, pKa) that each define specific thermodynamic ensemble configurations, providing real enforcement value in pharma applications. Claims 21–23 add meaningful ML architecture fallbacks (GAN/VAE/GCPN, generative models, weight sharing). Weaker fallbacks include Claims 6, 7, and 9 which restate general thermodynamic quantity relationships already implied by the independent claim framework without adding concrete limitations that would survive a dependent claim invalidity challenge.

Explore in Eureka ↗

⚠️

Abstract Quality

The abstract states: 'There is provided a method for a machine learning based method of analysing drug-like molecules by representing the molecular quantum states of each drug-like molecule as a quantum graph, and then feeding that quantum graph as an input to a machine learning system.' This accurately identifies the core mechanism but does not mention the thermodynamic ensemble sampling step, the orbital-correlation edge structure, or any specific predicted property — the three elements that most distinguish the claim from prior art message-passing graph networks. An examiner reading only the abstract may associate this with standard GNN drug discovery work (e.g., Gilmer 2017) rather than the quantum-graph thermodynamic ensemble combination that was actually allowed.

Explore in Eureka ↗

✅

Figure Support Quality

The key structural limitation — the quantum graph representation with nodes as molecular orbitals and edges as quantum correlations — is directly supported by FIG. 2 (Q-Graph derivation from chemical graph), FIG. 31 (DLI/LI graph), FIG. 54 (Q-Graph vs standard graph adjacency matrix comparison). The thermodynamic ensemble limitation maps to FIG. 49 (ensemble diagram) and FIG. 50 (two-path predictive model). The graph pooling limitation (Claim 11 fallback) is supported by FIGS. 4, 9, 10, 11. One gap: Claim 26's 'system configured to receive and process' a synthetically generated sample lacks a dedicated hardware/system architecture diagram beyond the high-level FIG. 1.

Explore in Eureka ↗

Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent

Scorecard

Strategic Intent Scorecard

Multi-dimensional assessment of this application's patent strategy quality, based on claim structure, specification depth, and prosecution positioning.

Claim Breadth

3.5

Prosecution Defensibility

3.2

Spec–Claim Consistency

4.2

Dependent Claim Coverage

3.6

Claim Type Diversity

2.8

Figure Support Quality

Key observation: The highest-scoring dimension is Spec–Claim Consistency (4.2/5.0) — the specification's nine sections and 88 figures provide detailed quantum chemistry, graph pooling, and thermodynamic ensemble coverage that maps precisely to the independent claim core, making written description challenges difficult. The lowest-scoring dimension is Claim Type Diversity (2.8/5.0) — the absence of a computer-readable medium (CRM) claim leaves a critical enforcement gap in the increasingly common software-as-a-service drug discovery deployment model, and practitioners should consider a continuation filing specifically targeting this format. Prosecution teams assessing FTO against this portfolio should note that the three independent claims, while architecturally sound, can all be designed around by characterising molecule representations as 'standard graph with quantum-derived node/edge features' rather than true quantum graphs with orbital-node correspondence.

See how your own draft compares — Open Eureka IP Drafting →

Critical Gaps

3 Critical Gaps in This Claim Set

A senior-attorney lens on the three highest-priority structural weaknesses — what each exposes in prosecution and litigation, and what a stronger filing would have done differently.

GAP 01 · HIGHEST IMPACT

No CRM or Software Claim Type Filed

The claim set covers method (Claim 1), system (Claim 26), and identified molecule (Claim 27) but contains no computer-readable medium (CRM) or storage medium claim. This structural omission means that competitors distributing the quantum graph featurisation software or ML pipeline as a cloud API, SaaS platform, or downloadable library — the primary commercial distribution model for ML drug discovery tools — cannot be directly reached by Claims 1 or 26 without additional contributory or inducement analysis. A stronger filing would have included a CRM claim reading: 'A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of Claim 1,' which is straightforward to support from the specification and the Python/bash workflow code documented in FIGS. 84–85.

GAP 02 · HIGH IMPACT

Independent Claims Lack Hardware Anchor for §101

Claims 1, 26, and 27 recite 'a machine learning system' without any structural hardware anchor (e.g., 'one or more processors,' 'GPU array,' 'quantum chipset') in the independent claim body — the hardware specificity is only implied or found in the spec's discussion of CPU/GPU and quantum chipsets (FIG. 3B). Under Alice Step 2B, the absence of a concrete hardware tie-in weakens the 'something more' argument for §101, creating litigation vulnerability where a challenger could characterise the claims as directed to an abstract mathematical method of representing molecular data. A stronger filing would have recited at least 'one or more processors configured to execute the machine learning system' in each independent claim, grounding the claims in a specific machine under the machine-or-transformation test.

GAP 03 · HIGH IMPACT

Quantum Graph Construction Steps Not Claimed

Unlock to read the full analysis.

🔒

3 Critical Gaps in This Claim Set

See the full attorney-level analysis of what this application leaves unprotected — and how to draft it more defensively for your own filings.

No CRM claim for software distribution Missing hardware anchor in independent claims Q-graph construction pipeline not claimed

Unlock Full Analysis — Free

Frequently asked questions

US 12,211,592 B2 — key questions answered

What does US 12,211,592 B2 protect?+

Who owns patent US 12,211,592 B2?+

What are the independent claims of US 12,211,592 B2?+

What technology does this patent cover?+

What IPC classifications does this patent use?+

Still have questions? PatSnap Eureka can answer them from patent data instantly. Search in Eureka

Disclaimer: This analysis is generated by PatSnap Eureka AI based on publicly available patent data from the USPTO. It does not constitute legal advice and should not be relied upon as such. Patent data may be subject to change as prosecution progresses. Scores and assessments reflect automated analysis and may not capture all relevant legal or technical nuances. Always consult a qualified patent attorney for formal legal opinions on patentability, freedom to operate, or infringement.

Ask anything about this patent.
PatSnap Eureka searches patents and data to answer instantly.

Link copied to clipboard

AI AGENTS

AI APPLICATIONS

OTHERS

INDUSTRIES

USE CASES

EXPLORE

ENGAGE

SUPPORT & SERVICES

Your Agentic AI Partner
for Smarter Innovation

Great, Please verify your email.

Patent Drafting Analysis of KUANO LTD.’s Machine Learning Drug-Like Molecule Analysis | US 12,211,592 B2

Patent Drafting Analysis of KUANO LTD.'s Quantum Graph Machine Learning for Drug-Like Molecules | US 12,211,592 B2