MolecularWeb: Web-Native Molecular Platforms
- MolecularWeb is a web-native ecosystem integrating real-time molecular screening with immersive XR visualization tools for chemistry and materials science.
- It leverages MolFormer embeddings, DCT compression, and HNSWlib to achieve sub-10 ms retrieval and 99% recall in large molecular databases.
- Its XR suite—including moleculARweb, PDB2AR, MolecularWebXR, and HandMol—supports interactive education, collaborative research, and advanced molecular modeling.
MolecularWeb is a designation applied to web-native molecular platforms that emphasize browser-based access, real-time interaction, and removal of installation barriers. In one usage, it denotes a cloud-based real-time molecular screening platform built around MolFormer embeddings for nearest neighbor retrieval, chemical space visualization, and property prediction (Belgodere et al., 2022). In another, it denotes an ecosystem at molecularweb.org comprising moleculARweb, MolecularWebXR, PDB2AR, and HandMol, a set of web-based tools for immersive visualization, modeling, and simulations in chemistry, structural biology, and materials sciences (Abriata, 4 Sep 2025). The term therefore spans two technically distinct but conceptually related strands: cheminformatics-oriented molecular screening and XR-oriented molecular graphics and modeling.
1. Scope and principal usages
The screening-oriented MolecularWeb is described as a cloud-based, real-time molecular screening platform that allows users to virtually screen molecules of interest by leveraging molecular embeddings inferred from MolFormer, a large masked chemical LLM (Belgodere et al., 2022). Its supported tasks are nearest neighbor retrieval, chemical space visualization, and property prediction, and its stated application areas include chemistry, chemical engineering, drug discovery, and materials design.
The XR-oriented MolecularWeb is described as a web-based ecosystem for immersive visualization, modeling, and simulations, already used in education and science communication and expanding toward research applications (Abriata, 4 Sep 2025). Its main components are moleculARweb, MolecularWebXR, PDB2AR, and HandMol. MolecularWebXR is specifically presented as a multiuser WebXR platform for education, science communication, and scientific peer discussion in chemistry and biology, while HandMol extends the ecosystem toward concurrent multiuser immersive visualization and modeling with bare hands, real-time molecular mechanics, and natural language input (Rodriguez et al., 2023).
A plausible implication is that “MolecularWeb” functions less as the name of a single monolithic program than as a label for browser-delivered molecular computation and interaction environments whose common denominator is direct web access.
2. Molecular screening with MolFormer embeddings
In the screening platform, molecules from large libraries such as PubChem or FlavorDB are first preprocessed off-line, converted into canonical SMILES strings using RDKit, and embedded with MolFormer into a 768-dimensional vector (Belgodere et al., 2022). To accelerate search, embeddings are compressed from 768 to 128 dimensions using a Discrete Cosine Transform (DCT), and the resulting vectors are indexed with HNSWlib. The reported operating point is <10 ms retrieval per query with recall of 0.99.
The platform is deployed using OpenShift on IBM Cloud, and interaction occurs through a web-based GUI. Input is a line-separated list of SMILES strings or a .txt file; output includes lists of nearest neighbors, calculated properties such as logP, QED, and molecular weight, and interactive or downloadable t-SNE visualizations. The platform can also interact with external databases such as the PubChem API for cross-comparison.
Its three supported tasks can be summarized as follows:
| Function | Core method | Reported examples |
|---|---|---|
| Nearest neighbor retrieval | MolFormer embeddings + DCT + HNSWlib | Real-time search in 100M+ compound databases |
| Chemical space visualization | MolFormer embeddings projected with t-SNE | 25,595 FlavorDB molecules |
| Property prediction | Nearest-neighbor transfer or k-NN models | RO5 screening; HBA/HBD regression |
For FlavorDB, the platform applies t-SNE to MolFormer embeddings of 25,595 molecules labeled as bitter, sweet, sour, or savory, and a k-NN classifier on the embeddings is reported to yield strong predictive performance for molecular flavor (Belgodere et al., 2022). For Rule of Five (RO5) screening, the paper reports 1.8 million CheMBL molecules, with a k-NN classifier (FAISS, ) trained on 1.44M and evaluated on 360k held-out molecules, achieving 90% accuracy in RO5 classification. For continuous properties such as hydrogen-bond acceptors and donors, -NN regression is reported to yield .
These details place MolecularWeb within the recent shift toward chemical LLMs as general-purpose molecular representation engines. The platform’s design treats embedding space as a substrate for similarity search, visualization, and lightweight prediction, rather than as a purely internal model representation.
3. The immersive MolecularWeb ecosystem
The XR-centered MolecularWeb ecosystem comprises moleculARweb, PDB2AR, MolecularWebXR, and HandMol (Abriata, 4 Sep 2025). All are described as web-native, requiring no installation and supporting a range of devices from laptops and smartphones to AR/VR headsets.
| Component | Primary role | Access model |
|---|---|---|
| moleculARweb | Marker-based and markerless AR educational activities | Phones, tablets, computers |
| PDB2AR | Generation of custom AR/VR molecular content | Web interface, no programming required |
| MolecularWebXR | Multiuser immersive rooms for discussion and exploration | Headsets, smartphones, tablets, computers |
| HandMol | Concurrent multiuser immersive visualization and modeling | High-end headsets and consumer devices |
moleculARweb is presented as accessible, marker-based AR for chemistry, materials science, and structural biology education on commodity devices. It uses printed fiducial markers that, when shown to a webcam, cause 3D molecular models to appear and follow the marker’s position and orientation (Abriata, 4 Sep 2025). Its modules include activities for atomic and molecular orbitals, VSEPR geometries, hydrogen bonding and acid/base dynamics, protein and nucleic acid structures, and large biological assemblies. The “Virtual Modeling Kits” include VMK 2.0, where two molecules are loaded onto separate cube markers to explore interactions such as hydrogen bonds or atom-atom clashes, and VMK 3.0, where multiple molecules are loaded into a simulation box with rudimentary molecular mechanics via Cannon.js.
PDB2AR is described as a tool that democratizes creation of custom AR and VR molecular content without programming. It accepts PDB files, including structures fetched from RCSB or AlphaFold DB, or Wavefront OBJ files exported from VMD, and generates web links for AR on cube markers, markerless AR, stand-alone VR, and GLB content for use in MolecularWebXR (Abriata, 4 Sep 2025).
MolecularWebXR provides multiuser immersive rooms for collaborative exploration and discussion of molecular models. Users join by session codes, can communicate with built-in audio, and can participate either through high-end headsets with hand tracking or through smartphones, tablets, and computers (Rodriguez et al., 2023). Preset rooms cover topics ranging from the periodic table, atomic and molecular orbitals, symmetry, and isomerism to complex biological assemblies and courses on protein structure.
HandMol is described in two generations. The prototype allows real-time hand tracking via WebXR, direct manipulation of atoms and bonds, natural language and voice control via embedded LLMs, on-the-fly molecular mechanics with Cannon.js, and external minimization APIs using ANI-2x or AMBER14 via OpenMM (Abriata, 4 Sep 2025). The upcoming integrated version within MolecularWebXR adds server-side synchronization, role-based multiuser sessions, exportable coordinates, adjustable simulation parameters such as temperature, and simulation engine choices including Cannon.js, ANI-2x, OpenMM/AMBER14, and Calvados.
4. Technical architecture and interaction design
The XR ecosystem is described as being built with HTML5, CSS, JavaScript, together with Three.js, AR.js, and WebXR/WebGL/WebRTC APIs, plus open-source physics and AI libraries (Abriata, 4 Sep 2025). In MolecularWebXR specifically, the backend is based on Node.js, which manages rooms, users, and synchronization of scene state (Rodriguez et al., 2023). Rather than transmitting video streams, the system synchronizes shared environments by sending only essential state such as object positions, orientations, scales, and avatar states; orientations are represented using quaternions.
Device inclusivity is central to the design. In headset mode, users receive full immersive operation with hand tracking or controllers; on smartphones and tablets, they navigate with virtual joysticks, touch, or gyroscope; on laptops and desktops, they use keyboard, mouse, or touch gestures (Rodriguez et al., 2023). MolecularWebXR also distinguishes between Admin, VR-Active Users, and Passive Users, allowing the same room model to support small-group collaboration, lectures, and outreach.
The deployment evidence reported for MolecularWebXR includes use by participants aged 12–80, support for up to 8 simultaneous VR-active users plus additional passive users and laptop presenters on local Wi-Fi, and the observation that in 150+ participants during 15 min sessions, no one quit due to discomfort (Rodriguez et al., 2023). More than 80% of participants are reported to have had no prior VR experience but were still able to manipulate objects.
The screening-oriented MolecularWeb uses a different architecture but a related web-first logic. It combines off-line database preprocessing, online embedding of user queries with MolFormer, compressed vector search with HNSWlib, and cloud deployment on IBM Cloud through OpenShift (Belgodere et al., 2022). Here, responsiveness is measured in sub-10 ms retrieval rather than immersive frame-level interaction.
5. Scientific, pedagogical, and collaborative uses
The screening platform’s applications are framed in terms of automating chemistry and chemical engineering research, assisting drug discovery, and supporting material design (Belgodere et al., 2022). Nearest-neighbor retrieval supports analog identification and chemical modification analysis; chemical space visualization supports cluster discovery and class separation; property prediction supports rapid estimation of drug-likeness and physicochemical descriptors.
The XR ecosystem’s uses are broader in modality. moleculARweb is described as suitable for education in chemistry, materials science, and structural biology, with activities spanning orbitals, molecular geometry, intermolecular interactions, and biomolecular structure (Abriata, 4 Sep 2025). MolecularWebXR extends this to science communication and scientific peer discussion, with support for synchronous, geographically distributed exploration and built-in audio (Rodriguez et al., 2023). HandMol moves toward research tasks such as collaborative manual docking, conformational sampling, preparing systems for further computation, and studying allosteric changes (Abriata, 4 Sep 2025).
The papers emphasize that accessibility is achieved without sacrificing heterogeneity of use. A school or outreach participant may enter through marker-based AR or a passive role in MolecularWebXR, while a researcher may employ the same ecosystem for collaborative structure discussion or immersive molecular mechanics. This suggests that MolecularWeb’s significance lies not only in specific algorithms or interfaces but also in the consolidation of browser-based workflows across education, communication, and research.
6. Position within the broader web-based molecular software landscape
MolecularWeb belongs to a broader class of browser-delivered molecular systems that emerged across quantum chemistry, molecular interaction maps, molecular dynamics, generative design, and spectroscopy. The Molecule Calculator (MolCalc) presents a web server whose entry page is a JSmol molecular editor and whose backend uses GAMESS at RHF/STO-3G for orbitals and PM3 for other properties such as heats of formation, thermodynamic properties, and vibrational data (Jensen et al., 2013). NaviCell provides web-based exploration, curation, and maintenance of large molecular interaction maps using the Google Maps engine, semantic zooming, and an integrated WordPress blog (Kuperstein et al., 2013).
For molecular dynamics, HTMoL is a plug-in-free, hardware-accelerated web application for remote access, visualization, and analysis of raw MD trajectories, using Node.js, WebSockets, and WebGL (Carrillo-Tripp et al., 2017). MDsrv likewise streams MD trajectories into browsers, now using Mol*, and supports session creation, remote dataset access, and analyses such as alignment, distance measurements, and RMSD (Kampfrath et al., 2022). Earlier collaborative and immersive precedents include a browser app for concurrent JSmol-based molecular handling over the Internet (Abriata, 2017) and a multiuser virtual-reality framework for interactive molecular dynamics that quantitatively demonstrated faster completion of intrinsically 3D molecular tasks in VR than with conventional interfaces (Connor et al., 2018).
Other adjacent web systems extend the same design philosophy into different domains. Molecule Generation Experience is a public GUI-based web application for AI-driven molecular generation with built-in datasets and a Kubernetes-based cloud architecture (Takeda et al., 2021). OpenMX Viewer offers browser-based visualization and analysis of crystalline and molecular structures, cube data, and band structures through drag-and-drop workflows (Lee et al., 2019). Molecular Vibration Explorer provides an online database and toolbox for IR, Raman, and vibrational frequency-conversion screening over thousands of thiolated molecules (Koczor-Benda et al., 2021).
Within this landscape, MolecularWeb is notable for uniting two otherwise separate trajectories of web molecular software: high-throughput molecular informatics on one side and immersive, multiuser, XR-native molecular graphics and modeling on the other.
