XtalPi

XtalPi Science runs discovery through three layers: a multi agent system that interprets scientific goals and orchestrates work, more than 200 domain AI models trained on the company's own compound and reaction data for molecular design, property prediction, synthesis routes and spectra reading, and robotic laboratories that run synthesis, crystallization and assays around the clock. The models handle prediction and design, the robots handle execution and data capture, and each round of lab results goes back into the models. Physics based methods work alongside them: free energy perturbation with a proprietary force field (XFEP) and physics driven crystal structure prediction (XtalGazer), the method the company started with. Lab automation hardware and software, sold to R&D organizations, was the larger revenue line in 2024.

XFEP, XtalGazer and the automation systems are also offered to customers in their own right.

A multi agent system sits at the center of XtalPi Science: agents interpret research goals, plan experiments and coordinate domain models, robotics, data pipelines and external models, then iterate between prediction and robotic execution. The workflow is built for scientists to supervise and steer, and human scientists invoke its models and capabilities through the hub. What it automates is laboratory work and design choices rather than decisions about patients, and partners keep their own roles: META Pharmaceuticals led target research on MP-5342, and PharmaEngine is advancing PEP08 through Phase I.

No published rule says when the agents stop and route a decision to a chemist.

The platform pairs a multi agent orchestration system with more than 200 domain AI models trained on in house compound and reaction data, covering molecular generation, structure prediction for crystals, proteins and binding modes, ADMET and selectivity, route design across more than 20 reaction types, antigen and antibody modeling, and LCMS reading. Its tools include XFEP for relative and absolute free energy predictions with a proprietary force field, XMolGen for generative chemistry from a commercial building block library, XtalGazer for crystal structure prediction, and the SureRXN reaction prediction model. Method papers cover retrosynthesis (SureRoute), reaction yield quantification from UV spectra, and selective pore modulation in layered molecular crystals.

No model versions or update practice are published.

The domain models are the company's own, built on its compound and reaction data, and the multi agent hub connects them to external foundation models when needed and to external partner agents. Laboratory hardware is proprietary too, from XmartChem synthesis workstations to the ChemPlus benchtop solid dispenser and the NeoDispenser micro powder dispensing system, and XFEP runs on a proprietary force field with cloud computing.

The external foundation models, and any cloud provider behind the computing, are not named.

PEP08, a PRMT5 inhibitor discovered on the platform for PharmaEngine, received clinical clearance in June 2025 and has begun enrollment in a Phase I solid tumor trial in Australia and the Taiwan region, and the two companies have started a second synthetic lethality program. META Pharmaceuticals' MP-5342, an LDHA inhibitor in development for inflammatory bowel disease, has completed IND enabling safety and efficacy studies, and XtalPi has submitted a US FDA IND application for KQTD-126, the first from its proprietary pipeline. Viva-Thera, a company XtalPi incubated, nominated an HIV-1 preclinical candidate after seven months of discovery and optimization. Case studies describe hit and lead work on GPX4, FGFR3, SMARCA2 and SARM1, including novel GPX4 hits in 28 days.

The PEP08 study identifier and dosing timeline are not disclosed.

Partner programs start from targets the partner chooses, as when Eli Lilly nominates target pairs for Ailux to design bispecific antibodies against, and the platform's closed loop returns experimental outcomes to the models for the next cycle. The website legal statement covers personal information that visitors volunteer, such as name, email address or phone number, and promises a privacy policy.

How partner data is separated, retained or kept out of models that serve other partners is not published.

Customers are drug, chemical and materials companies buying discovery services and laboratory systems, and the in house biology covers protein expression, more than 50 target based biochemical assays, cell based assays on more than 100 tumor cell lines, and rodent DMPK and PK/PD studies. Its website legal statement is governed by the laws of the PRC.

XtalPi takes no position on business associate status or on whether its services ever receive protected health information.

The XmartChem software gives role based user permissions so experimental methods and core data stay secure and controllable, and it collects reaction conditions and experimental data automatically. XFEP is offered as cloud based software, and discovery work runs on cloud scale computing alongside the robotic labs. The Hong Kong listing brings annual and interim results announcements and corporate governance documents, published through the investor relations site.

No security certification, penetration testing or trust center is named for the discovery environment or the software.

Molecules from the platform follow the drug pathway under whoever sponsors them. PharmaEngine is taking PEP08 through Phase I in Australia and the Taiwan region after clinical clearance in June 2025, ReviR Therapeutics has secured NMPA clinical approval for RTX-117, a rare neurological drug, and XtalPi has filed its own US FDA IND application for KQTD-126, aimed at chronic intestinal pain in irritable bowel syndrome and inflammatory bowel disease. The platform, its software and its robots are sold for research and laboratory use.

No device claim is made for any of them.

Crystal structure prediction, the company's original method, was tested in a Pfizer blind test and later supported a collaboration that helped bring Paxlovid to market, and XtalPi has published a study of the nirmatrelvir solid forms. The company has open sourced an AI model for chemical reaction yield quantification and publishes method papers, including its retrosynthesis platform, SureRoute, and a multimodal workflow running from patent analysis to potency. Data from the robotic labs feeds continuous learning back into the models.

No evaluation of where the 200 plus domain models underperform, by target class, chemical series or data density, is published.

Designs and predictions from the domain models go to the robotic laboratories for synthesis and testing before a program advances. That loop can synthesize and test 3,000 to 4,000 novel molecules within two to three months at a synthesis success rate above 80%, and the XmartChem workstation keeps 100% experiment traceability. Partner programs such as PEP08 run under a milestone sharing structure that ties the platform's evolution to clinical success.

No warranty or remedy for designs that fail is published.

The automation line includes lab management software: a digital lab notebook with visual workflow orchestration and AI model integration, scheduling and control, facility management with inventory traceability, and a digital twin cockpit. The XmartChem workstation joins solid and liquid dispensing, temperature controlled reactions and online LCMS analysis in one unattended workflow. The multi agent hub coordinates domain models, robotics and data pipelines, with model access through an API.

No connection to a customer's existing electronic lab notebook or laboratory information system is named.

Delivery takes two forms. Lab automation systems go onto customer sites with customizable deployments, such as the formulation stability testing workstation deployed at BASF, while discovery work runs in XtalPi's own robotic labs in Shanghai, Shenzhen and Boston and on cloud scale computing, with model access through an API. The company lists locations in Greater Boston, Shenzhen, Shanghai, Beijing and Runcorn in the United Kingdom, and its website is operated by Shenzhen Jingtai Technology Co., Ltd.

Where partner data, models and computation are held, and whether any of it crosses between China and other jurisdictions, is not stated.

Revenue comes from two lines: drug discovery partnerships paid through upfront payments, research funding, milestones and royalties, and lab automation systems and software sold to R&D organizations. Listed in Hong Kong, XtalPi reports segment revenue: about 266 million yuan in 2024, with robotics the larger line. It also announces deal values. The DoveTree Medicines collaboration runs at up to $5.99 billion, with a $51 million upfront payment and a $19 million second payment received; Ailux's bispecific antibody agreement with Eli Lilly runs at up to $345 million, after a $250 million small molecule partnership with Lilly in 2023; and a GPCR program carries a total potential value above $400 million. PepiX peptide work is offered through asset licensing or platform based collaboration.

No rate card or system price is published for either line.

Small molecules, peptides, siRNA, antibodies, solid forms and materials all run on the platform: hit identification to preclinical candidate for small molecules, PepiX for peptides, Kodexia for siRNA, the Ailux subsidiary for antibodies including bispecifics, XtalGazer for polymorph, salt and cocrystal work, and electrolytes, polymers and battery materials on the materials side. Partner programs span oncology (PEP08), inflammatory bowel disease (MP-5342) and HIV (Viva-Thera), and Ailux has presented preclinical data on a long acting IL-4Rα antibody. On the industrial side, Fangda Carbon uses XtalPi's predictive AI in graphite electrode manufacturing.

The peptide pipelines are described without named targets or stages.