Viva Biotech's MARS (Multi-Modality AI-Rooted Solutions) AI algorithm team recently published an innovative study in the Journal of Chemical Information and Modeling (JCIM), entitled “EnsembleCycPerm: Interpretable Modeling of Cyclic Peptide Permeability through Solvent-Dependent Conformational Ensembles.” The study was led by Dr. Yue Qian, Vice President and Head of the MARS Division at Viva Biotech (Shanghai), as the corresponding author. Sicheng Wen, an AI Algorithm Researcher in the MARS Business Unit, served as the first author and completed the work in collaboration with Dr. Yang Wang from Viva Biotech's CADD/AIDD platform.
（Source：JCIM Official Website）
To improve cyclic peptide permeability prediction, Viva Biotech's MARS AI algorithm team developed EnsembleCycPerm, an interpretable framework integrating sequence, physicochemical, and solvent-dependent conformational information. The model generates candidate three-dimensional conformations, optimizes them in water and chloroform, and encodes their atomic and geometric features using a graph neural network. Descriptors, including intramolecular hydrogen bonds, three-dimensional polar surface area, and radius of gyration, are used to weight conformations and capture solvent-induced structural changes. This enables the model to characterize the “molecular chameleon” behavior of cyclic peptides - specifically their transition from a hydrophilic open state to a nonpolar closed state - and predict PAMPA permeability. Residue-level attribution and matched single-residue analogue analyses further identify positions that may influence permeability, providing guidance for residue substitution, N-methylation, stereochemical modification, and intramolecular hydrogen-bond design.
From Single-Property Prediction to R&D Decision Support: EnsembleCycPerm Accelerates Cyclic Peptide Screening and Optimization
During cyclic peptide drug discovery, researchers typically need to select a small number of promising candidates from a much larger pool for synthesis and experimental testing. Target activity alone is often insufficient to determine whether a candidate has strong development potential, as permeability, solubility, metabolic stability, bioavailability, and safety may all influence its overall drug-like profile.
For AI-generated cyclic peptide libraries and large-scale virtual compound collections, the EnsembleCycPerm framework can serve as an early-stage screening tool. It can reduce the number of low-potential molecules entering downstream synthesis and testing, allowing experimental resources to be concentrated on the most promising candidates.
This screening strategy helps clients identify both the strengths and potential liabilities of candidate molecules at an earlier stage, reducing the risk of committing substantial resources to compounds with a high probability of failure during later development. When several candidates display comparable performance, EnsembleCycPerm and related platform capabilities can also support the establishment of clearer development priorities and inform subsequent synthesis, activity testing, and pharmacokinetic evaluation.
Toward Real-World Molecular Design and Lead Optimization: Precise Guidance on Structural Modification and Multidimensional Property Balancing
EnsembleCycPerm not only helps answer the question of which molecules should be prioritized, but also assists researchers in analyzing how a candidate should be structurally modified next. Starting from an existing cyclic peptide scaffold, the model can compare structurally related analogues and assess the potential permeability effects of residue substitutions, N-methylation, side-chain modifications, and localized structural changes. It can also identify positions that may offer the greatest value for further optimization.
By integrating predictive results, analogue comparisons, and key structural analyses, the model can provide clearer guidance for subsequent rounds of molecular design. For example, when a candidate displays insufficient permeability, EnsembleCycPerm can help localize structural regions that may limit its performance and highlight positions suitable for modification. When multiple analogues show similar overall profiles, the model can assist researchers in determining which design strategies should be prioritized for synthesis and experimental validation.
For candidates affected by poor solubility, insufficient metabolic stability, limited bioavailability, or potential safety concerns, EnsembleCycPerm can be used in conjunction with Viva Biotech's broader AIDD/CADD modeling capabilities and experimental platforms to evaluate trade-offs among multiple properties. This integrated approach helps prevent narrowly focused optimization in which improvement of one property compromises biological activity, safety, or overall developability. As a result, molecular design can be more closely aligned with the practical requirements of lead discovery and optimization.
From Algorithmic Components to a Closed-Loop R&D Ecosystem: Industrial Empowerment via the Pep2MARS Platform
At a broader level, EnsembleCycPerm can function as an intelligent decision-support module within Viva Biotech's cyclic peptide discovery platform. It connects molecular generation, structural analysis, property prediction, risk assessment, design recommendations, compound synthesis, and experimental validation. Through this workflow, clients can more rapidly identify high-potential candidates from a large design space, recognize developability risks at an earlier stage, and obtain more targeted recommendations for structural optimization. This can reduce unnecessary synthesis and repetitive experimentation while improving the efficiency of cyclic peptide programs from early discovery through lead optimization.
EnsembleCycPerm will continue to expand toward the prediction of additional ADMET properties that are critical to cyclic peptide development. It will also be further integrated with Viva Biotech's existing AIDD/CADD models and computational and experimental platforms, supporting comprehensive multiparameter evaluation of biological activity, permeability, solubility, metabolic stability, bioavailability, and potential safety risks.
As an advanced tool developed by the MARS team, EnsembleCycPerm is an important component of Pep2MARS, Viva Biotech's proprietary suite of algorithms specifically designed for peptide discovery and optimization. Following the earlier development of Pep2MARS, which introduced an innovative solution to the challenges associated with molecular dynamics simulations of structurally complex cyclic peptides (To read the full article, please click here >> " Viva Biotech's AIDD Team Publishes in JCIM: Automation of Complex Cyclic Peptide Molecular Dynamics Simulations in Pep2MARS"), the launch of EnsembleCycPerm represents another important advance by addressing a major bottleneck in cyclic peptide permeability prediction. As additional technologies from the MARS team are introduced, Viva Biotech aims to establish a comprehensive peptide design platform spanning the entire process from de novo design to multiparameter developability optimization.
Serving as a key technological link across this process, EnsembleCycPerm strengthens Viva Biotech's integrated service capabilities in drug design, medicinal chemistry, peptide synthesis, biological activity testing, preclinical development, and CMC/CDMO services. Together, these capabilities form a complete R&D workflow encompassing: Molecular design and generation → virtual screening → activity and ADMET evaluation → risk filtering → design recommendations → synthesis and experimental validation → candidate development. Through this integrated computational and experimental framework, Viva Biotech provides end-to-end support covering the full lifecycle of cyclic peptide discovery and optimization.
For further information about the study, please refer to:
Wen, S., Wang, Y. and Qian, Y. (2026). EnsembleCycPerm: Interpretable Modeling of Cyclic Peptide Permeability through Solvent-Dependent Conformational Ensembles. Journal of Chemical Information and Modeling. doi:10.1021/acs.jcim.6c01213.