qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning

qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning View ORCID Profile Haoyu Wang , View ORCID Profile Ziyan Zhang , Haipeng Gong doi: https://doi.org/10.1101/2025.09.07.674462 Haoyu Wang 1 MOE Key Laboratory of Bioinformatics, School of Life Sciences, Tsinghua University , Beijing, 100084, China 2 Beijing Frontier Research Center for Biological Structure, Tsinghua University , Beijing, 100084, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Haoyu Wang Ziyan Zhang 1 MOE Key Laboratory of Bioinformatics, School of Life Sciences, Tsinghua University , Beijing, 100084, China 2 Beijing Frontier Research Center for Biological Structure, Tsinghua University , Beijing, 100084, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Ziyan Zhang Haipeng Gong 1 MOE Key Laboratory of Bioinformatics, School of Life Sciences, Tsinghua University , Beijing, 100084, China 2 Beijing Frontier Research Center for Biological Structure, Tsinghua University , Beijing, 100084, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: hgong{at}tsinghua.edu.cn Abstract Full Text Info/History Metrics Preview PDF A bstract Recent advancements in deep learning have greatly prompted the de novo design of drugs and materials. Previous studies have shown that a well-designed molecular representation is critical for improving the accuracy of deep-learning-based molecular property prediction methods. However, the lack of large-scale data enriched with detailed physicochemical information hinders effective learning of an informative molecular representation. To fill this data gap, we introduce qcMol, a dataset consisting of 1.2 million molecules from 95 datasets with high-quality quantum chemical annotations, to facilitate molecular representation learning as well as downstream molecular property prediction. Chemicals in this dataset include drug-like compounds, metabolites and molecules with matched experimental data, covering 247,448 kinds of scaffolds and a broad spectrum of molecular sizes. Each compound in qcMol is annotated with detailed quantum chemical information, obtained through reliable quantum chemical calculations based on B3LYP-D3/def2-SV(P)//GFN2-xTB as well as the follow-up wave function post-analysis. These features are organized into multiple formats, allowing for flexible integration into diversified molecular representation learning frameworks. The broad data distribution, comprehensive quantum chemical annotations and flexible data formats jointly enable qcMol to serve as the pre-training resource as well as the benchmark test set for deep learning models, benefiting the practical in silico drug discovery. qcMol is freely accessible from https://structpred.life.tsinghua.edu.cn/qcmol/ . Introduction The discovery of new drugs and materials is time-consuming and financially demanding. On average, this process takes 12 years from start to commercialization, costing 9 billion dollars 1 – 3 . Hence, previous developments in this field have been largely supported by the collaboration between physics-based calculations and laboratory experiments 4 – 6 . Benefiting from advances in artificial intelligence, various deep-learning-based methods have been developed, aiming to accelerate this process 7 – 10 . Compared with physics-based methods 11 , 12 , deep learning methods are advantageous not only in the improved performance but also in the markedly reduced inference time 13 . However, as a data-driven approach, the power of deep learning models is highly dependent on the quality of input data. Particularly, an informative and comprehensive molecular representation can improve the performance of deep learning models, effectively accelerating molecular discovery 14 , 15 . This principle has motivated the extensive investigation on molecular representation learning 10 , 16 . Conventional molecular descriptors can be used as input features for representation learning models in multiple formats, including strings 17 , 18 , images 19 and graphs 9 . Generated using cheminformatics tools like RDKit 20 and Open Babel 21 , these descriptors are often too simplistic to capture the complex electronic structures of molecules, and conversely, the lack of physical accuracy in the representation of learning targets tends to misguide model training, potentially disrupting prediction power and generalizability 22 , 23 . Density functional theory (DFT) 24 , a reliable and computationally efficient quantum chemical method, provides access to electronic structures and relevant properties like electron density, molecular orbitals and energy levels. Wave functions obtained from such quantum chemical calculations can be localized using methods like Natural Bond Orbital (NBO) 25 analysis to extract chemically interpretable electronic features, enabling the generation of structured representations that are inherently compatible with graph neural networks (GNN) and related frameworks without losing the quantum chemical fidelity. Moreover, incorporation of quantum chemical information, whether as explicit input features or through integrated training strategies, has been reported to benefit molecular property prediction tasks by the consideration of molecular electronic structures 26 . Hence, a well-maintained dataset composed of versatile molecular features derived through reliable quantum chemical calculations as well as the follow-up wave function post-analysis is essential for the training and benchmarking of deep-learning-based molecular representation learning and property prediction models. In the past years, several quantum chemistry databases and platforms have been developed to support machine learning applications in molecular science 27 – 33 . Among them, QM9 32 is one of the most widely used, containing computed electronic, energetic, geometric and thermodynamic properties for 134,000 stable small organic molecules that are composed of C, H, O, N and F elements and are chosen from GDB-17 34 , a dataset constructed with the purpose of exploring the chemical space. However, this dataset restricts the molecular size to be no more than nine heavy atoms, which substantially impairs the chemical diversity, and consequently is mainly used as a benchmark set for the evaluation of molecular property prediction models. Another relevant dataset, the PubChemQC Project 29 , provides ground-state electronic structures for over 3 million molecules computed by DFT at the B3LYP/6-31G * level, as well as low-lying excited states of more than 2 million molecules calculated by time-dependent DFT (TD-DFT) with B3LYP/6-31+G * . Unfortunately, the involved molecules are randomly selected from the PubChem database, without prioritizing the relevance to drug discovery or material science applications. More importantly, this dataset primarily provides global molecular properties, such as the HOMO–LUMO gap and dipole moment, while lacking fine-grained electronic features at the levels of atoms or bonds. QuanDB 31 , a recently released dataset, incorporates 154,610 compounds composed of nine elements (H, C, O, N, P, S, F, Cl and Br) from public databases and literature. In addition to global molecular properties, it provides five local quantum chemical descriptors and the most stable 3D conformation for each molecule. However, the dataset covers a limited range of molecular sizes and features, lacking molecules of over 40 atoms and essential features. In this work, to address the limitations of existing data resources, we constructed a new dataset named qcMol, which contains high-quality and abundant molecular features with distributions closely matching the real-world compounds, aiming to facilitate representation learning on chemical molecules. Unlike contemporary datasets that focus on small-molecule compounds, qcMol additionally includes drug-like molecules and common metabolites, many of which are associated with paired experimental measurements. Moreover, it provides detailed quantum chemical descriptors in a suitable format to simplify feature processing by graph-based deep learning frameworks. Incorporation of experimental results also allows qcMol to serve as a benchmark set for the training and refining of deep learning models developed for practical molecular design and drug discovery tasks. Methods The qcMol computational workflow integrates several advanced computational chemistry tools, including RDKit 20 (2022.09.5 version), Open Babel 3.1.1 21 , GFN2-xTB 6.5.1 35 , ORCA 5.0.3 36 – 41 , NBO 7.0 25 , and Multiwfn 3.8 42 . The complete dataset construction pipeline is described below and illustrated in Figure 1 . Download figure Open in new tab Figure 1. Task scheduling system and quantum chemical calculation pipeline of qcMol. Tasks are automatically distributed to computing nodes to enable parallel quantum chemical calculations for large-scale molecules. Dataset collection and curation To fully utilize computational resources, we carefully selected the molecules for calculation and applied quality control procedures to the input data. Data collection qcMol integrates small-molecule data from various sources, including published literature and established biochemical databases. HMDB 43 , MoleculeNet 44 , MoleculeACE 45 , PDBbind 46 , and several other publicly available datasets 47 are among the commonly used sources incorporated into qcMol. We collect each molecule’s identifier and associated experimental data while retaining as much relevant metadata as possible for downstream applications. Data curation Due to the limited computational resources and potential inconsistencies among the raw datasets, quality control procedures are necessary to ensure data reliability and usability. As molecular structures were recorded using SMILES, a data cleaning process was first applied to address inconsistencies in encoding formats and the lack of standardization ( Figure 1 ). Specifically, we applied the following data cleaning criteria: 1) entries with inconsistent results assigned to identical SMILES strings were removed; 2) multi-component SMILES ( e . g ., representing mixtures or salts) were flagged for further processing; 3) molecules containing elements beyond the first four periods of the Periodic Table were filtered out; 4) entries containing metal ions were excluded; 5) all SMILES strings were converted into the RDKit-canonicalized SMILES, InChI, and InChIKey formats; 6) each molecule was assigned a unique internal identifier ( i . e . qcMol ID) for reference within the database; 7) a deduplication step was conducted against existing entries, preventing redundant quantum chemical calculations. In qcMol, different conformers of the same molecule are treated as the same molecular entity, whereas molecules with different charge states are considered as distinct entries. Finally, to ensure the completeness of the database, we retrieved additional molecular metadata from the PubChem 48 database for each compound. Workflow for quantum chemical calculations Following standard notation in quantum chemistry, our computational scheme is denoted as B3LYP-D3/def2-SV(P)//GFN2-xTB, which indicates that single-point energy calculations using the B3LYP-D3 49 , 50 functional with the def2-SV(P) 41 basis set were performed on geometries optimized at the GFN2-xTB 35 level ( Figure 1 ). Initial Structure Generation Initial molecular structures were generated using RDKit 20 , an open-source cheminformatics toolkit. Specifically, RDKit was used to construct molecular geometries and perform preliminary optimization using the MMFF94 force field 51 . The resulting structures were saved in the XYZ format. If RDKit failed to generate a valid structure, the molecule was processed using Open Babel 21 as a fallback. The generated structure served as the initial input for subsequent geometry optimization and quantum chemical calculations. Preliminary Structure Optimization In consideration of the trade-off between computational accuracy and efficiency, the initial structures were first optimized using GFN2-xTB before being passed to ORCA for higher-level quantum chemical calculations. GFN2-xTB is an extended tight-binding method parameterized from DFT, allowing efficient geometry optimization for large molecular systems composed of main-group elements. Specifically, the ANCopt algorithm was employed for geometry optimization, which iteratively minimized energy through structural refinement 35 . Notably, we adopted the “extreme” optimization level in xTB to maximize geometric precision while maintaining computational efficiency, yielding high-quality initial structures for subsequent quantum chemical calculations. DFT-Based Quantum Chemical Calculations Based on the optimized structures, advanced ab initio calculations were performed using ORCA 36 – 41 at the B3LYP-D3 level with the def2-SV(P) basis set. B3LYP-D3 is a hybrid functional that combines Becke’s three-parameter exchange functional with the Lee-Yang-Parr correlation functional, and includes semi-empirical D3 dispersion corrections with atom-pair-specific parameters to account for long-range interactions accurately. The def2-SV(P) basis set is a double- ζ quality basis set that includes polarization functions. Wave function Post-analysis NBO 7.0 25 and Multiwfn 42 were employed to analyze the wave function obtained from ORCA comprehensively. The NBO method decomposes wave functions into localized bonding and non-bonding orbitals, providing a chemically meaningful picture of molecular structure and electron population. In addition, Multiwfn was used to compute a range of electronic properties, including electron density, molecular orbitals, partial atomic charges and bond-level descriptors. The hierarchical strategy described above combines the computational efficiency of xTB-based geometry optimization with the high electronic accuracy of advanced DFT methods. Compared to traditional protocols such as B3LYP/6-31G * , our method achieves significantly higher efficiency and better scalability to large molecules by employing fast geometric optimization with GFN2-xTB. Furthermore, by incorporating D3 dispersion correction, the method captures more accurately the long-range non-covalent interactions, such as π - π stacking and hydrogen bonding, which are often underestimated by standard DFT approaches. Additionally, compared to 6-31G * , the def2-SV(P) basis set provides greater accuracy, better numerical stability, and broader element coverage, while maintaining similar computational efficiency. Finally, the wave function was analyzed to extract molecular properties at multiple levels, including global, atomic, and bond-level descriptors. Computational scheduling and time consumption In our scheduling system, batches of molecules are randomly assigned to different computing nodes to enable efficient large-scale computation ( Figure 1 ). To avoid performance degradation from parallel interference, all molecular calculations are constrained to single-core execution. On each calculation node, tasks are evenly distributed across available CPU cores and executed sequentially through structure optimization, quantum property computation, wave function analysis, and feature extraction. A dynamic recycling strategy ensures full utilization of computational resources. When the overall CPU usage falls below a pre-defined threshold, completed tasks are collected and remaining molecules are redistributed for processing. This scheduling mechanism ensures near-full utilization of all CPUs across the computing cluster, thereby maximizing computational throughput. The calculations used dual Intel Skylake 6132 (14-core, 2.6 GHz) CPUs with 96 GB RAM per node, costing over 6 million CPU core hours in total. Results Overview qcMol is a well-annotated and chemically diverse dataset designed to support drug and material discovery. This section provides a statistical overview of its key features, characterizing the composition and scope of the dataset. DFT calculations were performed at the B3LYP-D3 level with the def2-SV(P) basis set to derive the quantum chemical information. To balance computational cost, accuracy and chemical diversity, we limited the composing elements to the first four periods of the Periodic Table. The qcMol dataset comprises a total of 1,200,434 curated molecules, covering a broad spectrum of molecular sizes and weights. Each molecule is annotated with 31 quantum chemical properties, including 15 global and 16 local descriptors. The overall molecular features involve these detailed quantum chemical features as well as traditional properties computed using RDKit. Notably, 63.71% of molecules in the dataset are associated with systematic experimental results obtained through wet-lab studies, providing a valuable resource for data-driven drug discovery. Data characteristics Data Diversity for Robust Learning qcMol features a broad distribution in terms of elemental diversity, molecular size, and the variation of molecular scaffolds. In consideration of the salient relativistic effects in heavy elements, the database primarily focuses on non-metal elements from the first four periods of the Periodic Table ( Figure 2a ). The top elements in qcMol molecules include hydrogen (H), carbon (C), nitrogen (N), oxygen (O), sulfur (S) and phosphorus (P) ( Figure 2b ), consistent with the preference of life systems. The dataset also contains a notable number of molecules with halogen atoms, including fluorine (F), chlorine (Cl), bromine (Br) and iodine (I). Download figure Open in new tab Figure 2. Elemental composition and molecular size distribution in qcMol. a) Distribution of elements involved in all existing datasets, schematically shown within the Periodic Table. Each element is colored with the degree of darkness proportional to the logarithm of its appearing percentage. The number below an element reflects the occurring count. The occurring ratio is shown instead for the top elements. b) Ranking of the top elements (excluding hydrogen) in qcMol. c) The distribution of molecular size in qcMol. d) The distribution of molecular weight in qcMol. e) Boxplot of the molecular size across different datasets, where the box limits correspond to upper and lower quartiles and the whiskers extend to 1.5 inter-quartile range. Compared to QM9 and QuanDB, qcMol offers a wider distribution of molecular sizes ( Figure 2c ) and molecular weights ( Figure 2d ). For instance, the number of atoms per molecule in the qcMol dataset ranges from 1 to 604, covering the molecular sizes of most existing chemical datasets ( Figure 2e ) and thus achieving an enhanced reflection of the real-world chemical space, which is expected to be conducive to the training of robust, well-performed deep learning models. Molecular scaffolds are widely used to split datasets for the out-of-distribution (OOD) evaluation, since molecules sharing similar scaffolds often exhibit similar properties 52 . We analyzed the scaffold diversity of qcMol and identified 247,448 unique scaffold types, with an average of 4.85 molecules per scaffold. Figure 3 provides a visualization of the most frequently occurring scaffolds, along with a quantitative analysis of the top 10 scaffold types. Notably, monocyclic scaffolds represent the dominant structural motif in the qcMol dataset, among which the most frequent ones are benzene, pyridine and cyclopropane rings. Benzene is a simple and widely occurring aromatic ring with a fully conjugated π -electron system, pyridine is a six-membered heteroaromatic ring commonly found in drug molecules, whereas cyclopropane is a strained three-membered ring used to restrict molecular flexibility in medicinal chemistry. These diverse monocyclic scaffolds illustrate the structural richness of the dataset and its alignment with the chemical space of drug-like compounds. Overall, the broad coverage of molecular scaffolds in qcMol makes it a potential benchmark for assessing the generalizability of molecular representation learning models. Download figure Open in new tab Figure 3. The scaffold composition of qcMol. a) Statistics of the top 10 scaffolds. b) Schematic view of the top 50 scaffolds after excluding benzene. Quantum Chemical Annotation of Widely Used Molecular Datasets Datasets like QM7, QM8 and QM9 are composed of molecules enumerated from the chemical space, many of which do not correspond to known or synthesizable compounds, weakening their applicability in real-world scenarios 32 , 34 . Other studies, such as the PubChemQC Project 29 , focus on calculating a large number of relatively simple molecules chosen from the PubChem database, prioritizing small-scale compounds feasible for quantum chemical calculations under the limited computational resources, which leads to the overlook of compounds with higher levels of chemical relevance or practical meaning. Unlike the above mentioned works, qcMol emphasizes the significance of the molecules being computed ( Figure 4 ). Download figure Open in new tab Figure 4. Versatile data in qcMol. Molecules in qcMol cover a broad range of compound types and are collected from a number of different datasets. Moreover, considering the rapid data accumulation of chemical and biochemical experiments, qcMol integrates compounds from multiple datasets that capture fundamental physicochemical properties of small molecules ( Supplementary Table 1 ). Quantum chemical annotations are especially meaningful for these molecules rather than those generated purely through computational enumeration, because molecules investigated in past experiments are often closely associated with disease processes and human physiological environment. A key focus of qcMol is to facilitate the understanding of interactions between small molecules and biological entities, the basis on which small molecules perform specified biological functions and modulate cellular states. Accordingly, qcMol includes data covering protein-ligand interactions, RNA-ligand interactions, drug-drug interactions, and cellular responses to drugs ( Supplementary Table 1 ). Another important molecular category covered by qcMol falls into the class of metabolites, which are intermediate or end products of biochemical reactions within organisms and play crucial roles in maintaining cellular function, signal transduction, energy metabolism, and other physiological processes. Moreover, metabolites reflect the states of gene expression and protein activity, since they are actively involved in the regulation of internal and external cellular environments. Hence, monitoring metabolite changes allows direct detection of alterations in disease-related pathways, facilitating diagnosis and target discovery. In this consideration, qcMol incorporates metabolites in the Human Metabolism Database (HMDB) and Pseudomonas aeruginosa Metabolism Database (PAMDB) ( Figure 4 and Supplementary Table 1 ). In addition, this dataset also includes a set of drug-like molecules from PubChem ( Supplementary Table 1 ). Noticeably, we have accomplished quantum chemical annotations for all molecules involved in qcMol. In summary, qcMol currently includes five major categories of data from 95 datasets: molecular properties, intermolecular interactions, molecule-cell interactions, drug-like molecules, and metabolites. The molecules are collected from multiple different datasets, as detailed in Supplementary Table 1 . By combining quantum chemical information with experimental molecular data, qcMol provides a comprehensive representation of molecules and predictive modeling for drug-like compounds. Integrated Multi-Scale Features for Molecular Representation Molecular representation learning requires that the input data be converted into formats compatible with machine learning models. Commonly used formats include strings, images and graphs, which are well-suited for natural language models 18 , computer vision models 19 and graphbased neural networks 9 , respectively. With the growing complexity of molecular tasks, there is an increasing interest in incorporating quantum chemical information into molecular representations, enabling models to capture more detailed electronic and structural properties. Prior studies have mainly focused on global quantum chemical descriptors, such as HOMO, LUMO and the HOMO-LUMO gap, while neglecting local information at the atomic or bonding level. Although the wave function contains complete quantum information, its raw form is incompatible with most machine learning models. To address this limitation, we employed the wave function localization methods to extract structured, interpretable features at both atomic and bonding levels, enabling their effective use in downstream learning tasks. Figure 5a summarizes the main molecular features curated by qcMol. Download figure Open in new tab Figure 5. Diverse molecular features in qcMol. a) Overview of molecular features in qcMol. b) Distribution of the HOMO-LUMO gap (in atomic unit) for molecules in qcMol. c) Distribution of the dipole moment for molecules in qcMol. d) The most stable conformation of an exemplar molecule. e) LI representation for atoms in the exemplar molecule in panel ( d ). f) DI representation for bonds in the exemplar molecule in panel ( d ). g) ELF representation for bonds in the exemplar molecule in panel ( d ). Basic information ( Supplementary Table 3 ): This aspect includes fundamental molecular identifiers, such as qcMol ID, IUPAC name, SMILES, InChI, InChIKey, chemical formula, etc., which facilitate indexing, deduplication, and cross-database referencing. Global features ( Supplementary Table 4 ): The qcMol dataset provides 15 global features and 16 local features for the quantum chemical annotation of each molecule. The global feature includes the HOMO-LUMO gap ( Figure 5b ), dipole moment ( Figure 5c ), and other relevant properties, which primarily characterize general molecular characteristics. qcMol also includes a set of shape-related features, such as isosurface area, isosurface volume and sphericity parameters, to comprehensively capture the geometric aspects of molecular structure. Local features ( Supplementary Tables 5 & 6 ): The 16 local features consist of node-level (atom) and edge-level (bond) descriptors, organized into a molecular graph format. Exemplar node-level features are the natural atomic orbital (NAO) descriptors and local ionization (LI) values ( Figure 5e ), while the representative edge-level features include the delocalization index (DI) matrix ( Figure 5f ), electron localization function (ELF) values ( Figure 5g ), and Laplacian bond order (LBO). Structural features ( Supplementary Table 5 ): Previous studies have demonstrated that the three-dimensional (3D) geometric structure of a molecule plays a crucial role in determining its physicochemical properties and intermolecular interactions. In qcMol, we provide reliable geometry-optimized structures for all molecules, representing energetically stable conformations that serve as meaningful structural representations for down-stream applications ( Figure 5d ). Database Usage and Data Download Web Interface and Functionality The qcMol database is accessible through a user-friendly web interface via https://structpred.life.tsinghua.edu.cn/qcmol/ . The homepage presents the core highlights of qcMol, including a clear navigation of major functionalities ( Figure 6a ). Download figure Open in new tab Figure 6. Example of the web interface for qcMol. a) The main interface of the web server. b) The browse interface of qcMol. c) The download interface to obtain the subset annotation. d) The upload interface for new molecule calculation. e) Exemplar page showing molecular details at the Atom Information level. By clicking “Browse”, users can explore all molecules in qcMol and apply various filters to identify compounds that meet specific criteria ( Figure 6b ). The “Search” interface allows users to locate individual molecules by entering versatile identifiers including the IUPAC name, SMILES, chemical formula, InchI and other synonyms. The detailed information of each identified molecule could be visualized through further clicking ( Figure 6e ). The “Download” interface enables users to obtain dataset files in structured formats ( Figure 6c ). All raw data are publicly available for download from the qcMol website. Moreover, the dataset is further divided into multiple subsets, allowing download by specified data types. Users can also upload molecules through the “Upload” interface to request annotation of special molecules via the qcMol calculation pipeline ( Figure 6d ). Furthermore, the “Related Work” page will be continuously updated with recent advances and studies building on qcMol. For inquiries, please reach out via the “Help” page for more discussion. To ensure long-term usability and expansion, qcMol will be updated every six months with newly calculated and processed molecules. Future releases will also include quantum chemical calculations for more complex molecular structures. All data are available under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Details of the Data Organization We provide an example to illustrate the structure of the data after downloading. Each compound is assigned an internal qcMol ID ( e . g ., qcMol-Dataset-000000000). As shown in Supplementary Table 2 , the molecule is stored in a dedicated folder with the following files: (1) the.info file for recording basic information; (2) the.global file to record the general properties of the molecule; and (3) the.local file to record details of atoms, bonds and structure. The following file is also available upon request for further analysis: (4) the.mol file to record the structure; (5) the.inp and.out files of the DFT calculation; (6) the.gbw and.47 files for wave function post analysis. Conclusion and Discussion Conclusion Here, we summarize the key contributions of this work. To bridge the gap between existing quantum chemical databases and downstream applications, qcMol focuses on molecules with high chemical or biological relevance. Considering the trade-off between the limited computational resources and the vastness of chemical space, qcMol selected relevant molecules from various public datasets and performed quantum chemical calculations at the B3LYP-D3/def2-SV(P)//GFN2-xTB level. Subsequently, qcMol performed wave function post-analysis to extract structured, localized descriptors at the atomic and bonding levels. Currently, the qcMol dataset offers quantum chemical annotations for five major categories of molecular data, including molecular properties, intermolecular interactions, molecule-cell interactions, drug-like compounds, and metabolites. This also ensures that qcMol aligns more closely with realistic chemical space by capturing diversity in molecular size, elemental composition, and scaffold structure. As a result, it promotes more robust and transferable molecular representations. By combining quantum chemical calculations with wave function analysis, qcMol provides a rich set of multi-level features, including basic information, global properties, local descriptors and geometric structures, to support diverse molecular representation learning tasks. The hierarchical organization and adaptability of these features make them especially suitable for multi-modal or hybrid modeling frameworks, which aim to comprehensively capture the structural, electronic and topological aspects of molecular systems. Combining large-scale quantum chemical computation with feature engineering, qcMol substantially extends the scale of existing datasets and provides multi-level formats for diverse downstream tasks. This contribution is especially valuable for deepening our understanding of molecules, enhancing representation learning, and accelerating the development of novel drugs and functional materials. Discussion Although qcMol offers advantages in data distribution, dataset scale and feature formats, which make it well-suited for molecular representation learning, several aspects still merit further discussion and improvement. First, although the dataset covers a broad range of drug-like molecules, metabolites, and experimentally relevant compounds, the involved molecules only occupy a few limited local regions of the vast chemical space. The diversity in molecular size, elemental composition and scaffold structure helps improve model robustness, but rare or complex molecules remain underrepresented. Future work would focus on expanding coverage to include macrocycles, organometallics, as well as larger-sized molecules. Second, the quantum chemical calculations were performed at the B3LYP-D3/def2-SV(P) level, which balances accuracy and efficiency well. However, more challenging tasks, like accurate reactivity prediction and excited-state property modeling, require more advanced methods such as CCSD(T) or TD-DFT. Third, each molecule in qcMol is currently represented by a single optimized geometry. However, many molecules exhibit notable conformational flexibility, and a single structure may not adequately capture their full spatial and energetic diversity. Incorporating conformational ensembles in future dataset versions could enhance the robustness and reliability of molecular modeling, particularly for tasks sensitive to the 3D structural variability. Lastly, how to best utilize the curated data in specific model architectures remains an open question. Systematic benchmarks and ablation studies are needed to guide model design and feature utilization strategies. Further improvements in APIs, data loaders, and integration with frameworks like PyTorch Geometric 53 and DeepChem 54 would enhance usability and community adoption. Author contribution H. W. and H. G. proposed the concept and idea. H. W. designed the computational pipeline, collected data and performed all quantum chemical computations. Z. Z. constructed the database and the web interface. H. G. acquired funding and supervised the whole process. All authors are involved in the manuscript writing and agreed with the final manuscript. Ethics declarations Competing interests The authors declare no competing interests. Supplementary Information Supplementary Tables View this table: View inline View popup Supplementary Table 1. Statistics of molecular datasets View this table: View inline View popup Download powerpoint Supplementary Table 2. Download file structure View this table: View inline View popup Download powerpoint Supplementary Table 3. Basic information for molecules View this table: View inline View popup Download powerpoint Supplementary Table 4. Global features of molecules View this table: View inline View popup Download powerpoint Supplementary Table 5. Local features of molecules (atom information) View this table: View inline View popup Download powerpoint Supplementary Table 6. Local features of molecules (relationships between atoms) Acknowledgements This work has been supported by the National Natural Science Foundation of China (#32171243), by the Ministry of Science and Technology of China (#2023YFF1204400), and by the Beijing Frontier Research Center for Biological Structure. We thank Hui Zeng for discussions and valuable suggestions. Funder Information Declared National Natural Science Foundation of Chin , #32171243 Ministry of Science and Technology of China , #2023YFF1204400 Beijing Frontier Research Center for Biological Structure References 1. ↵ Morgan , S. , Grootendorst , P. , Lexchin , J. , et al. The cost of drug development: A systematic review . Health Policy 100 , 4 – 17 . ISSN: 0168-8510 . https://www.sciencedirect.com/science/article/pii/S0168851010003659 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 2. Rennane , S. , Baker , L. & Mulcahy , A. Estimating the cost of industry investment in drug research and development: a review of methods and results . INQUIRY: The Journal of Health Care Organization, Provision, and Financing 58 , 00469580211059731 . https://journals.sagepub.com/doi/full/10.1177/00469580211059731 ( 2021 ). OpenUrl 3. ↵ Schlander , M. , Hernandez-Villafuerte , K. , Cheng , C.-Y. , et al. How Much Does It Cost to Research and Develop a New Drug? A Systematic Review and Assessment . PharmacoEconomics 39 , 1243 – 1269 . ISSN: 1179-2027 . doi: 10.1007/s40273-021-01065-y ( 2021 ). OpenUrl CrossRef PubMed 4. ↵ Sinha , S. K. , Prasad , S. K. , Islam , M. A. , et al. Identification of bioactive compounds from Glycyrrhiza glabra as possible inhibitor of SARS-CoV-2 spike glycoprotein and non-structural protein-15: a pharmacoinformatics study . Journal of Biomolecular Structure and Dynamics 39 , 4686 – 4700 . doi: 10.1080/07391102.2020.1779132 ( 2021 ). OpenUrl CrossRef 5. Sadybekov , A. A. , Sadybekov , A. V. , Liu , Y. , et al. Synthon-based ligand discovery in virtual libraries of over 11 billion compounds . Nature 601 , 452 – 459 . ISSN: 1476-4687 . doi: 10.1038/s41586-021-04220-9 ( 2022 ). OpenUrl CrossRef PubMed 6. ↵ Manglik , A. , Lin , H. , Aryal , D. K. , et al. Structure-based discovery of opioid analgesics with reduced side effects . Nature 537 , 185 – 190 . ISSN: 1476-4687 . doi: 10.1038/nature19112 ( 2016 ). OpenUrl CrossRef PubMed 7. ↵ Lu , W. , Wu , Q. , Zhang , J. , et al. TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction in Advances in Neural Information Processing Systems 35 ( Curran Associates, Inc ., 2022 ), 7236 – 7249 . https://proceedings.neurips.cc/paper_files/paper/2022/file/2f89a23a19d1617e7fb16d4f7a049ce2-Paper-Conference.pdf . OpenUrl 8. Li , Y. , Hsieh , C.-Y. , Lu , R. , et al. An adaptive graph learning method for automated molecular interactions and properties predictions . Nature Machine Intelligence 4 , 645 – 651 . ISSN: 2522-5839 . doi: 10.1038/s42256-022-00501-8 ( 2022 ). OpenUrl CrossRef 9. ↵ Zhou , G. , Gao , Z. , Ding , Q. , et al. Uni-Mol: A Universal 3D Molecular Representation Learning Framework in The Eleventh International Conference on Learning Representations ( 2023 ). https://openreview.net/forum?id=6K2RM6wVqKu . 10. ↵ Feng , S. , Yang , L. , Huang , Y. , et al. Unimap: universal smiles-graph representation learning 2023 . https://arxiv.org/abs/2310.14216 . 11. ↵ Friesner , R. A. , Banks , J. L. , Murphy , R. B. , et al. Glide: A New Approach for Rapid, Accurate Docking and Scoring. 1. Method and Assessment of Docking Accuracy . Journal of Medicinal Chemistry 47 , 1739 – 1749 . ISSN: 0022-2623 . doi: 10.1021/jm0306430 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 12. ↵ Halgren , T. A. , Murphy , R. B. , Friesner , R. A. , et al. Glide: A New Approach for Rapid, Accurate Docking and Scoring. 2. Enrichment Factors in Database Screening . Journal of Medicinal Chemistry 47 , 1750 – 1759 . ISSN: 0022-2623 . doi: 10.1021/jm030644s ( 2004 ). OpenUrl CrossRef PubMed Web of Science 13. ↵ Ji , M. , Gong , X. , Li , X. , et al. Advanced research on the antioxidant activity and mechanism of polyphenols from Hippophae species—A review . Molecules 25 , 917 . ISSN: 1420-3049 . https://www.mdpi.com/1420-3049/25/4/917 ( 2020 ). OpenUrl PubMed 14. ↵ Bengio , Y. , Courville , A. & Vincent , P. Representation Learning: A Review and New Perspectives . IEEE Transactions on Pattern Analysis and Machine Intelligence 35 , 1798 – 1828 ( 2013 ). OpenUrl CrossRef 15. ↵ David , L. , Thakkar , A. , Mercado , R. , et al. Molecular representations in AI-driven drug discovery: a review and practical guide . Journal of Cheminformatics 12 , 56 . ISSN: 1758-2946 . doi: 10.1186/s13321-020-00460-5 ( 2020 ). OpenUrl CrossRef PubMed 16. ↵ Gao , B. , Qiang , B. , Tan , H. , et al. Drugclip: Contrastive protein-molecule representation learning for virtual screening in . 36 ( 2023 ), 44595 – 44614 . https://proceedings.neurips.cc/paper_files/paper/2023/file/8bd31288ad8e9a31d519fdeede7ee47d-Paper-Conference.pdf . OpenUrl 17. ↵ Weininger , D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules . Journal of Chemical Information and Computer Sciences 28 , 31 – 36 . ISSN: 0095-2338 . doi: 10.1021/ci00057a005 ( 1988 ). OpenUrl CrossRef 18. ↵ Ross , J. , Belgodere , B. , Chenthamarakshan , V. , et al. Large-scale chemical language representations capture molecular structure and properties . Nature Machine Intelligence 4 , 1256 – 1264 . ISSN: 2522-5839 . doi: 10.1038/s42256-022-00580-7 ( 2022 ). OpenUrl CrossRef 19. ↵ Zeng , X. , Xiang , H. , Yu , L. , et al. Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework . Nature Machine Intelligence 4 , 1004 – 1016 . ISSN: 2522-5839 . doi: 10.1038/s42256-022-00557-6 ( 2022 ). OpenUrl CrossRef 20. ↵ Landrum , G. , Tosco , P. , Kelley , B. , et al. Rdkit/Rdkit: 2025_03_2 (Q1 2025) Release 2025 . http://rdkit.org/docs/Cookbook.html . 21. ↵ O’Boyle , N. M. , Banck , M. , James , C. A. , et al. Open Babel: An open chemical toolbox . Journal of Cheminformatics 3 , 33 . ISSN: 1758-2946 . doi: 10.1186/1758-2946-3-33 ( 2011 ). OpenUrl CrossRef PubMed 22. ↵ Shurrab , S. & Duwairi , R. Self-supervised learning methods and applications in medical imaging analysis: A survey . PeerJ Computer Science 8 , e1045 . ISSN: 2167-9843 . https://europepmc.org/articles/PMC9455147 ( 2022 ). OpenUrl CrossRef 23. ↵ Krishnan , R. , Rajpurkar , P. & Topol , E. J. Self-supervised learning in medicine and healthcare . Nature Biomedical Engineering 6 , 1346 – 1352 . ISSN: 2157-846X . doi: 10.1038/s41551-022-00914-1 ( 2022 ). OpenUrl CrossRef 24. ↵ Kohn , W. & Sham , L. J. Self-Consistent Equations Including Exchange and Correlation Effects . Physical Review 140 , A1133 – A1138 . https://link.aps.org/doi/10.1103/PhysRev.140.A1133 ( 1965 ). OpenUrl CrossRef 25. ↵ Glendening , E. D. , Landis , C. R. & Weinhold , F. NBO 7.0: New vistas in localized and delocalized chemical bonding theory . Journal of Computational Chemistry 40 , 2234 – 2241 . https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.25873 ( 2019 ). OpenUrl CrossRef PubMed 26. ↵ Liu , S. , Wang , H. , Liu , W. , et al. Pre-training Molecular Graph Representation with 3D Geometry in International Conference on Learning Representations ( 2022 ). https://openreview.net/forum?id=xQUe1pOKPam . 27. ↵ Nakata , M. , Shimazaki , T. , Hashimoto , M. , et al. PubChemQC PM6: Data Sets of 221 Million Molecules with Optimized Molecular Geometries and Electronic Properties . Journal of Chemical Information and Modeling 60 , 5891 – 5899 . ISSN: 1549-9596 . doi: 10.1021/acs.jcim.0c00740 ( 2020 ). OpenUrl CrossRef PubMed 28. Nakata , M. & Maeda , T. PubChemQC B3LYP/6-31G * //PM6 Data Set: The Electronic Structures of 86 Million Molecules Using B3LYP/6-31G * Calculations . Journal of Chemical Information and Modeling 63 , 5734 – 5754 . ISSN: 1549-9596 . doi: 10.1021/acs.jcim.3c00899 ( 2023 ). OpenUrl CrossRef PubMed 29. ↵ Nakata , M. & Shimazaki , T. PubChemQC Project: A Large-Scale First-Principles Electronic Structure Database for Data-Driven Chemistry . Journal of Chemical Information and Modeling 57 , 1300 – 1308 . ISSN: 1549-9596 . doi: 10.1021/acs.jcim.7b00083 ( 2017 ). OpenUrl CrossRef PubMed 30. Isert , C. , Atz , K. , Jiménez-Luna , J. , et al. QMugs, quantum mechanical properties of drug-like molecules . Scientific Data 9 , 273 . ISSN: 2052-4463 . doi: 10.1038/s41597-022-01390-7 ( 2022 ). OpenUrl CrossRef 31. ↵ Yang , Z. , Huang , T. , Pan , L. , et al. QuanDB: a quantum chemical property database towards enhancing 3D molecular representation learning . Journal of Cheminformatics 16 , 48 . ISSN: 1758-2946 . doi: 10.1186/s13321-024-00843-y ( 2024 ). OpenUrl CrossRef PubMed 32. ↵ Ramakrishnan , R. , Dral , P. O. , Rupp , M. , et al. Quantum chemistry structures and properties of 134 kilo molecules . Scientific Data 1 , 140022 . ISSN: 2052-4463 . doi: 10.1038/sdata.2014.22 ( 2014 ). OpenUrl CrossRef 33. ↵ Smith , D. G. A. , Altarawy , D. , Burns , L. A. , et al. The MolSSI QCArchive project: An open-source platform to compute, organize, and share quantum chemistry data . WIREs Computational Molecular Science 11 , e1491 . https://wires.onlinelibrary.wiley.com/doi/abs/10.1002/wcms.1491 ( 2021 ). OpenUrl 34. ↵ Blum , L. C. & Reymond , J.-L. 970 Million Druglike Small Molecules for Virtual Screening in the Chemical Universe Database GDB-13 . Journal of the American Chemical Society 131 , 8732 – 8733 . ISSN: 0002-7863 . doi: 10.1021/ja902302h ( 2009 ). OpenUrl CrossRef PubMed Web of Science 35. ↵ Bannwarth , C. , Ehlert , S. & Grimme , S. GFN2-xTB—An Accurate and Broadly Parametrized Self-Consistent Tight-Binding Quantum Chemical Method with Multipole Electrostatics and Density-Dependent Dispersion Contributions . Journal of Chemical Theory and Computation 15 , 1652 – 1671 . ISSN: 1549-9618 . doi: 10.1021/acs.jctc.8b01176 ( 2019 ). OpenUrl CrossRef PubMed 36. ↵ Neese , F. Software update: The ORCA program system—Version 5.0 . WIREs Computational Molecular Science 12 , e1606 . https://wires.onlinelibrary.wiley.com/doi/abs/10.1002/wcms.1606 ( 2022 ). OpenUrl CrossRef 37. Neese , F. , Wennmohs , F. , Hansen , A. , et al. Efficient, approximate and parallel Hartree–Fock and hybrid DFT calculations. A ‘chain-of-spheres’ algorithm for the Hartree–Fock exchange . Chemical Physics 356 , 98 – 109 . ISSN: 0301-0104 . https://www.sciencedirect.com/science/article/pii/S0301010408005089 ( 2009 ). OpenUrl CrossRef Web of Science 38. Helmich-Paris , B. , de Souza , B. , Neese , F. , et al. An improved chain of spheres for exchange algorithm . The Journal of Chemical Physics 155 , 104109 . ISSN: 0021-9606 . doi: 10.1063/5.0058766 ( 2021 ). OpenUrl CrossRef PubMed 39. Izsák , R. , Neese , F. & Klopper , W. Robust fitting techniques in the chain of spheres approximation to the Fock exchange: The role of the complementary space . The Journal of Chemical Physics 139 , 094111 . ISSN: 0021-9606 . doi: 10.1063/1.4819264 ( 2013 ). OpenUrl CrossRef PubMed 40. Izsák , R. & Neese , F. An overlap fitted chain of spheres exchange method . The Journal of Chemical Physics 135 , 144105 . ISSN: 0021-9606 . doi: 10.1063/1.3646921 ( 2011 ). OpenUrl CrossRef PubMed 41. ↵ Weigend , F. & Ahlrichs , R. Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for H to Rn: Design and assessment of accuracy . Physical Chemistry Chemical Physics 7 , 3297 – 3305 . ISSN: 1463-9076 . doi: 10.1039/B508541A ( 2005 ). OpenUrl CrossRef PubMed Web of Science 42. ↵ Lu , T. & Chen , F. Multiwfn: A multifunctional wavefunction analyzer . Journal of Computational Chemistry 33 , 580 – 592 . eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/jcc.22885 . https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.22885 ( 2012 ). OpenUrl CrossRef PubMed 43. ↵ Wishart , D. S. , Guo , A. , Oler , E. , et al. HMDB 5.0: the Human Metabolome Database for 2022 . Nucleic Acids Research 50 , D622 – D631 . ISSN: 0305-1048 . doi: 10.1093/nar/gkab1062 ( 2021 ). OpenUrl CrossRef 44. ↵ Wu , Z. , Ramsundar , B. , Feinberg , E. N. , et al. MoleculeNet: a benchmark for molecular machine learning . Chemical Science 9 , 513 – 530 . ISSN: 2041-6520 . doi: 10.1039/C7SC02664A ( 2018 ). OpenUrl CrossRef PubMed 45. ↵ Van Tilborg , D. , Alenicheva , A. & Grisoni , F. Exposing the Limitations of Molecular Machine Learning with Activity Cliffs . Journal of Chemical Information and Modeling 62 , 5938 – 5951 . ISSN: 1549-9596 . doi: 10.1021/acs.jcim.2c01073 ( 2022 ). OpenUrl CrossRef PubMed 46. ↵ Liu , Z. , Li , Y. , Han , L. , et al. PDB-wide collection of binding data: current status of the PDBbind database . Bioinformatics 31 , 405 – 412 . ISSN: 1367-4803 . doi: 10.1093/bioinformatics/btu626 ( 2014 ). OpenUrl CrossRef PubMed 47. ↵ Sterling , T. & Irwin , J. J. ZINC 15 – Ligand Discovery for Everyone . Journal of Chemical Information and Modeling 55 , 2324 – 2337 . ISSN: 1549-9596 . doi: 10.1021/acs.jcim.5b00559 ( 2015 ). OpenUrl CrossRef PubMed 48. ↵ Wang , Y. , Xiao , J. , Suzek , T. O. , et al. PubChem: a public information system for analyzing bioactivities of small molecules . Nucleic Acids Research 37 , W623 – W633 . ISSN: 0305-1048 . doi: 10.1093/nar/gkp456 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 49. ↵ Becke , A. D. Density-functional Thermochemistry. III. The Role of Exact Exchange . The Journal of Chemical Physics 98 , 5648 – 5652 . ISSN: 0021-9606 . https://pubs.aip.org/aip/jcp/article-abstract/98/7/5648/842114/Density-functional-thermochemistry-III-The-role-of?redirectedFrom=fulltext ( 1993 ). OpenUrl CrossRef Web of Science 50. ↵ Lee , C. , Yang , W. & Parr , R. G. Development of the Colle-Salvetti Correlation-Energy Formula into a Functional of the Electron Density . Physical Review B 37 , 785 – 789 . https://journals.aps.org/prb/abstract/10.1103/PhysRevB.37.785 ( 1988 ). OpenUrl CrossRef PubMed Web of Science 51. ↵ Halgren , T. A. Merck molecular force field. I. Basis, form, scope, parameterization, and performance of MMFF94 . Journal of Computational Chemistry 17 , 490 – 519 . ISSN: 0192-8651 . https://onlinelibrary.wiley.com/doi/abs/10.1002/%28SICI%291096-987X%28199604%2917%3A5/6%3C490%3A%3AAID-JCC1%3E3.0.CO%3B2-P ( 1996 ). OpenUrl CrossRef Web of Science 52. ↵ Bemis , G. W. & Murcko , M. A. The properties of known drugs. 1. Molecular frameworks . Journal of Medicinal Chemistry 39 , 2887 – 2893 . ISSN: 0022-2623 . doi: 10.1021/jm9602928 ( 1996 ). OpenUrl CrossRef PubMed Web of Science 53. ↵ Fey , M. & Lenssen , J. E. Fast Graph Representation Learning with PyTorch Geometric in ICLR Workshop on Representation Learning on Graphs and Manifolds ( 2019 ). https://pytorch-geometric.readthedocs.io/en/latest/index.html . 54. ↵ Ramsundar , B. , Eastman , P. , Walters , P. , et al. Deep Learning for the Life Sciences https://www.amazon.com/Deep-Learning-Life-Sciences-Microscopy/dp/1492039837 ( 2019 ). Supplementary References 1. Jumper , J. , Evans , R. , Pritzel , A. , et al. Highly accurate protein structure prediction with AlphaFold . Nature 596 , 583 – 589 . doi: 10.1038/s41586-021-03819-2 ( 2021 ). OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted September 12, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning Haoyu Wang , Ziyan Zhang , Haipeng Gong bioRxiv 2025.09.07.674462; doi: https://doi.org/10.1101/2025.09.07.674462 Share This Article: Copy Citation Tools qcMol: a large-scale dataset of 1.2 million molecules with high-quality quantum chemical annotations for molecular representation learning Haoyu Wang , Ziyan Zhang , Haipeng Gong bioRxiv 2025.09.07.674462; doi: https://doi.org/10.1101/2025.09.07.674462 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Biophysics Subject Areas All Articles Animal Behavior and Cognition (7640) Biochemistry (17706) Bioengineering (13902) Bioinformatics (41978) Biophysics (21465) Cancer Biology (18611) Cell Biology (25528) Clinical Trials (138) Developmental Biology (13387) Ecology (19920) Epidemiology (2067) Evolutionary Biology (24332) Genetics (15615) Genomics (22519) Immunology (17747) Microbiology (40424) Molecular Biology (17194) Neuroscience (88662) Paleontology (667) Pathology (2839) Pharmacology and Toxicology (4827) Physiology (7650) Plant Biology (15160) Scientific Communication and Education (2046) Synthetic Biology (4302) Systems Biology (9826) Zoology (2271)