Expansion Therapeutics, Inc.

Beyang Therapeutics has participated in the “OpenADMET - ExpansionRx Blind Challenge” jointly organized by OpenADMET and Expansion Therapeutics (Oct 27, 2025 – Jan 19, 2026). OpenADMET is an open scientific initiative focused on advancing the prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties for small-molecule drugs. The project is a collaborative effort supported by multiple organizations (e.g., ARPA-H, Gates Foundation) and administered by the Open Molecular Software Foundation (OMSF). Expansion Therapeutics is a biotechnology company focused on developing therapies for RNA‑mediated diseases such as myotonic dystrophy and ALS. The company has now open‑sourced high‑quality ADMET data from over 7,000 molecules systematically collected during its preclinical optimization programs. This dataset constitutes one of the largest publicly available datasets simulating the lead‑optimization process. The time‑split evaluation—training on historically available compounds and predicting the properties of subsequently designed molecules—closely mirrors industrial lead‑optimization workflows, in which past experimental data guides future molecular design decisions. It thereby provides a scarce industrial‑grade benchmark for assessing the forward‑prediction capability of models. The challenge was focused on nine classical properties for ADME: LogD, Kinetic Solubility, Mouse/Human Liver Microsomal Clearance, Caco-2 Efflux Ratio, Caco-2 Papp A>B, Mouse Plasma Protein Binding (MPPB), Mouse Brain Protein Binding (MBPB), Mouse Gastrocnemius Muscle Binding (MGMB) The challenge has drawn over 370 participants from both industry and academia, representing diverse backgrounds and experience levels. Among the top 10 submissions, several leading organizations have been identified: Merck (3rd place, submission moka), Deepmirror (4th place, submission campfire‑capillary), and DP Technology (6th place, submission HybridADMET). The affiliations of the remaining top‑10 participants remain undisclosed. Figure 1. The final leaderboard of OpenADMET - ExpansionRx Blind Challenge. Our company (submission name: yanyn) secured 11th place in the competition (Figure 1). In this highly industrially-relevant evaluation, our method demonstrates generalization capability comparable to leading international approaches (Figure 2). It should be noted that our model was trained solely on the official challenge training data (with the exception of LogD), without the introduction of any external or internal supplementary data. Despite operating under these purely data-constrained conditions, it achieved predictive performance of industrial reference value, reflecting the inherent robustness and practical potential of the method itself. Looking forward, based on further methodological optimization, we plan to integrate public and internal datasets to explore further potential improvements in model performance. Figure 2. The left figure presents the comprehensive analysis results of the challenge as published on the official OpenADMET website, summarizing the performance distribution of the top four teams and all submitted solutions; the right figure shows the submission results of our team. Methodological Overview Chemprop [1] is used, which is a Message passing neural network (MPNN). It is a widely used Graph neural networks method. Among the top 20 results, 18/20 teams used this algorithm. CheMeleon [2] foundation model is used, which is pre-trained on deterministic molecular descriptors from the Mordred package. Kinetic GROVER Multi-Task (KERMT) is used [3], which is a pretrained graph neural network model for molecular property prediction. This method was tested in the research paper and demonstrated promising performance in multi-task prediction scenarios [4]. Chemprop [1] is used, which is a Message passing neural network (MPNN). It is a widely used Graph neural networks method. Among the top 20 results, 18/20 teams used this algorithm. CheMeleon [2] foundation model is used, which is pre-trained on deterministic molecular descriptors from the Mordred package. Kinetic GROVER Multi-Task (KERMT) is used [3], which is a pretrained graph neural network model for molecular property prediction. This method was tested in the research paper and demonstrated promising performance in multi-task prediction scenarios [4]. Task Configuration Training set: With the exception of logD (for which external data were employed), all other tasks relied solely on the competition-provided dataset. Validation strategy: Stratified 10-fold cross-validation was performed on the training set to ensure statistical robustness of performance estimates. Hyperparameter settings: Both Chemprop and the CheMeleon foundation model were trained and evaluated using sets of distinct hyperparameter configurations. KERMT employed two pretrained backbone architectures— GROVERbase, GROVERlarge [5]. Learning paradigm: All tasks except logD were trained under a multi-task learning framework. The logD task utilized publicly available data and was trained as a single-task model, with data preprocessed using the denoising methodology proposed by Adrian et al. [6]. Molecular descriptor types: Four molecular representation schemes were evaluated for both CheMeleon and Chemprop models: Built-in default representation (default) RDKit descriptors (rdkit) Morgan binary fingerprints (morgan_binary) Morgan count-based fingerprints (morgan_count) Training set: With the exception of logD (for which external data were employed), all other tasks relied solely on the competition-provided dataset. Validation strategy: Stratified 10-fold cross-validation was performed on the training set to ensure statistical robustness of performance estimates. Hyperparameter settings: Both Chemprop and the CheMeleon foundation model were trained and evaluated using sets of distinct hyperparameter configurations. KERMT employed two pretrained backbone architectures— GROVERbase, GROVERlarge [5]. Learning paradigm: All tasks except logD were trained under a multi-task learning framework. The logD task utilized publicly available data and was trained as a single-task model, with data preprocessed using the denoising methodology proposed by Adrian et al. [6]. All tasks except logD were trained under a multi-task learning framework. The logD task utilized publicly available data and was trained as a single-task model, with data preprocessed using the denoising methodology proposed by Adrian et al. [6]. Molecular descriptor types: Four molecular representation schemes were evaluated for both CheMeleon and Chemprop models: Built-in default representation (default) RDKit descriptors (rdkit) Morgan binary fingerprints (morgan_binary) Morgan count-based fingerprints (morgan_count) Built-in default representation (default) RDKit descriptors (rdkit) Morgan binary fingerprints (morgan_binary) Morgan count-based fingerprints (morgan_count) Model Selection and Ensemble Strategy Cross-validation ranking: This is used for parameter selection. A parameter configuration is retained if it ranks within the top 10% within each individual task (based on cross-validation MAE) for at least 3 out of the 9 tasks in multi-task prediction. External validation: The selected models were evaluated on the organizer-provided independent validation set. Task-specific cutoff filtering: Prior to ensemble construction, a performance cutoff was defined independently for each task. Only models meeting the respective task-specific threshold were retained for ensemble formation. Task-level direct ensemble: For each task, all models passing the cutoff were directly ensembled to produce the final prediction. Independent ensembles were constructed per task rather than employing a single shared ensemble across all tasks. Result submission: The ensemble predictions for each task were submitted as the official final results. Cross-validation ranking: This is used for parameter selection. A parameter configuration is retained if it ranks within the top 10% within each individual task (based on cross-validation MAE) for at least 3 out of the 9 tasks in multi-task prediction. External validation: The selected models were evaluated on the organizer-provided independent validation set. Task-specific cutoff filtering: Prior to ensemble construction, a performance cutoff was defined independently for each task. Only models meeting the respective task-specific threshold were retained for ensemble formation. Task-level direct ensemble: For each task, all models passing the cutoff were directly ensembled to produce the final prediction. Independent ensembles were constructed per task rather than employing a single shared ensemble across all tasks. Result submission: The ensemble predictions for each task were submitted as the official final results. Reference 1.Chemprop v2: An Efficient, Modular Machine Learning Package for Chemical Property Prediction. Graff DE, Morgan NK, Burns JW, Doner AC, Li B, Li SC, Manu J, Menon A, Pang HW, Wu H, Zalte AS, Zheng JW, Coley CW, Green WH, Greenman KP. J Chem Inf Model. 2026 Jan 12;66(1):28-33. doi: 10.1021/acs.jcim.5c02332. Epub 2025 Dec 26. PMID: 41453060. 2. Descriptor-based Foundation Models for Molecular Property Prediction. Jackson Burns, Akshat Zalte, William Green. https://doi.org/10.48550/arXiv.2506.15792 3. Multitask finetuning and acceleration of chemical pretrained models for small molecule drug property prediction. Matthew Adrian, Yunsie Chung, Kevin Boyd, Saee Paliwal, Srimukh Prasad Veccham, Alan C. Cheng https://doi.org/10.48550/arXiv.2510.12719 4. Multitask finetuning and acceleration of chemical pretrained models for small molecule drug property prediction. Matthew Adrian, Yunsie Chung, Kevin Boyd, Saee Paliwal, Srimukh Prasad Veccham, Alan C. Cheng. https://doi.org/10.48550/arXiv.2510.12719 5. Self-Supervised Graph Transformer on Large-Scale Molecular Data. Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, Junzhou Huang. https://doi.org/10.48550/arXiv.2007.02835 6. Denoising Drug Discovery Data for Improved Absorption, Distribution,Metabolism, Excretion, and Toxicity Property Prediction. Matthew Adrian, Yunsie Chung, and Alan C. Cheng. Journal of Chemical Information and Modeling 2024 64 (16), 6324-6337. DOI: 10.1021/acs.jcim.4c00639

CAMBRIDGE, Mass.--(BUSINESS WIRE)--Rhapsogen, an immunology company redefining the treatment paradigm for autoimmune diseases, today announced the appointment of Renato Skerlj, Ph.D., as the company’s new Chief Executive Officer, effective immediately. Rhapsogen was launched by Catalio Capital Management’s venture strategy earlier this year. The company was co-founded by Prof. Jeffrey Ravetch, M.D., Ph.D., of Rockefeller University, and Prof. Eric Sundberg, Ph.D. of Emory University, both of whom are leaders in the immunology field and have made fundamental contributions to the understanding of how the immune system interacts directly with antibodies to protect against infection or cause autoimmune disease. In addition to his primary appointment, Dr. Jeffery Ravetch is a Catalio Venture Partner. Dr. Skerlj is a seasoned biotechnology executive with more than 30 years of experience in drug discovery and development. He has authored over 70 scientific publications, is an inventor on more than 60 patents, and has played key leadership roles in advancing multiple therapeutics from discovery through clinical development and regulatory approval. Most recently, Dr. Skerlj served as President and Chief Executive Officer of Expansion Therapeutics, where he led the advancement of a proprietary RNA-targeting platform and overseeing company growth and financing. Previously, he was Co-Founder and Chief Scientific Officer of X4 Pharmaceuticals and a co-inventor of mavorixafor, which received FDA approval in 2024 for the treatment of WHIM syndrome. Across his career, Dr. Skerlj has built and led high-performing R&D organizations, contributed to successful financings and strategic transactions, and helped translate scientific innovation into impactful medicines. “I am very excited by what we are building at Rhapsogen and look forward to working with our world-class team of immunology experts to deliver the solutions that autoimmune patients need and deserve,” said Dr. Skerlj. “Dr. Skerlj’s leadership experience coupled with his deep drug development expertise make him well suited to maximize the potential of Rhapsogen’s first-in-class approach in autoimmune disease. We are thrilled to continue supporting the company’s growth under Dr. Skerlj’s leadership," said Matthew Hobson, Ph.D., Principal at Catalio Capital Management, who will be also joining Rhapsogen’s Board of Directors. About Rhapsogen Rhapsogen is a Cambridge, MA-based immunology company focused on redefining the autoimmune treatment paradigm by developing next generation therapies that rapidly neutralize pathogenic autoantibodies and prevent autoimmune inflammation and tissue damage while preserving the immune system’s ability to fight infection. The company was founded in 2025 by Dr. Jeffrey Ravetch (Rockefeller University) and Dr. Eric Sundberg (Emory University). To learn more, please visit www.rhapsogen.com. About Catalio Capital Management, LP: Catalio Capital Management, LP, is a New York-based investment firm focused on the full lifecycle of innovative healthcare investing across private, public and credit markets. Catalio has exclusive partnerships with world-renowned serial scientist-entrepreneurs from around the globe who each have extensive academic and scientific achievements and who have also started multiple well-established companies based on their research. Catalio has offices in New York, London, and Washington, D.C. Learn more about Catalio Capital Management by visiting www.cataliocapital.com.

Funding to support ongoing clinical programs in seizure and psychiatric disorders and precision neuromedicine discovery platform Raymond Kelleher, M.D., Ph.D., of lead investor Cormorant Asset Management, and James Healy, M.D., Ph.D., of Sofinnova Investments, join Board of Directors BOSTON & SAN DIEGO--(BUSINESS WIRE)-- Rapport Therapeutics, Inc., a clinical-stage biotechnology company leveraging the science of receptor-associated proteins (RAPs) to advance precision neuromedicines, today announced that it has raised $150 million in Series B funding. Led by Cormorant Asset Management, the round includes investments from Fidelity Management & Research Company, Goldman Sachs Asset Management, Logos Capital, Perceptive Advisors, Sofinnova Investments, Surveyor Capital (a Citadel company), funds and accounts advised by T. Rowe Price Associates, Inc., and all existing investors, which include Third Rock Ventures, ARCH Venture Partners and Johnson & Johnson Innovation – JJDC, Inc. (JJDC). Concurrent with the round, Rapport has appointed Raymond Kelleher, M.D., Ph.D., Managing Director of Cormorant Asset Management, and James Healy, M.D., Ph.D., Managing Partner of Sofinnova Investments, to its Board of Directors. “We are appreciative for the support of such a prominent group of new investors and excited to welcome Ray and Jim to our board of directors,” said Abraham Ceesay, Chief Executive Officer of Rapport. “Rapport is extremely well positioned to advance our ongoing clinical programs in seizure and psychiatric disorders, and to expand our pipeline by harnessing our unique capability to leverage receptor-associated proteins for precision neuromedicine. We are poised for tremendous growth and look forward to pursuing a wide variety of opportunities to serve patients who desperately need more effective treatments.” Rapport is leveraging discoveries made by its scientific founders to target receptors in specific neuroanatomical regions underlying the pathophysiology of neurological disorders. Its approach, based on identifying and targeting receptor-associated proteins, contrasts with current treatments that act on targets expressed ubiquitously in the nervous system and, in some cases, other areas of the body. In June, the company appointed a scientific advisory board that includes two Nobel laureates as well as world-renowned experts in ion channel signaling, seizure disorders and the genesis of pain. “I am thrilled to join the Rapport team during such an exciting period of growth, and to work with its scientists as well as one of the most esteemed groups of scientific advisors in the biotechnology community,” said Dr. Kelleher. “Rapport is pursuing a unique approach that exploits novel molecular insights with the goal of targeting ion channel receptors with unprecedented precision. I look forward to contributing to the success of a team that possesses such a deep determination to change the lives of patients living with neurological diseases.” “Rapport is an exciting company because it has the potential to meaningfully change the lives of patients with neurological diseases,” said Dr. Healy. “I am eager to support Rapport as an investor and board member as the team continues to progress its programs in the clinic and extend what is possible in precision neuromedicine.” About the New Board Members Dr. Kelleher is an Associate Neurologist at Massachusetts General Hospital, where his laboratory studies the molecular and cellular mechanisms underlying cognitive disorders and his clinical practice focuses on neurodegenerative disorders. He currently serves on the boards of Arkuda Therapeutics, BiVACOR, Expansion Therapeutics, Garuda Therapeutics, and GC Therapeutics as Director, and Chroma Therapeutics, Graviton, Interius BioTherapeutics, and Umoja Biopharma as Observer. Dr. Kelleher received his M.D. and Ph.D. from Stanford University. Dr. Healy currently serves on the boards of ArriVent, Atsena, Bolt (BOLT), Karuna (KRTX), Natera (NTRA), and Y-mAbs (YMAB). Over the course of his twenty-three year career at Sofinnova, twenty-two of his investments have gone public and twelve have been acquired. He previously served on the boards of the Executive Committee at UC Berkeley, the National Venture Capital Association (NVCA) and the Biotechnology Innovation Organization (BIO). Dr. Healy received his M.D. and Ph.D. from Stanford University. About Rapport ( ) Rapport pursues precision neuromedicine in an effort to give patients better alternatives to current treatments for neurological diseases, which often come with reduced efficacy and side effects that lead to suboptimal dosing, noncompliance and discontinuation. By more precisely targeting the neural circuits and cell types where diseases originate, Rapport’s approach has the potential to reduce the unwanted drug-target interactions that lead to these issues. The company’s platform, built and refined over the past decade by its founding scientific team, achieves precision through the discovery and targeting of receptor-associated proteins, or RAPs, which display regional or neural circuit-specific expression. The company’s first program, designed to treat seizure disorders in patients whose condition is resistant to current treatments, is in Phase 1 clinical studies. View source version on businesswire.com: Contacts Media: Kit Rodophele Ten Bridge Communications 617-999-9620 krodophele@tenbridgecommunications.com Source: Rapport Therapeutics, Inc. View this news release online at: