UnibindMD couples machine learning with physics-based simulation to move from an arbitrary protein target to experimentally actionable peptide-binder leads. The workflow chains five distinct components: a deep-learning design engine called BindCraft, an independent structure-prediction and confidence stage (AlphaFold3-class / Boltz-2), binding-affinity estimation using PRODIGY, developability profiling through ADMET-AI, and final all-atom molecular-dynamics validation with GROMACS/CHARMM36m. The study team framed this composition as a direct response to a familiar problem: single-method confidence metrics such as ipTM and pLDDT can be self-consistent within a prediction framework but are not independent indicators of physical stability. By construction, each stage produces an orthogonal check on candidate quality.
The source did not report implementation-level hyperparameters for BindCraft, specific AlphaFold3-class/ Boltz-2 confidence thresholds used to accept or reject designs, or the exact ADMET-AI scoring cutoffs. Investigators did report the software and scoring components by name and the overall chaining logic used to triage designs before dynamics testing.
Investigators demonstrated UnibindMD across three intervention points targeting Andes virus (ANDV) biology: the Gn-Gc glycoprotein complex (viral entry), the L-protein endonuclease domain (cap-snatching; PDB 6Q99), and the host VEGFR2 kinase (vascular permeability; PDB 2XIR). The team ran nine separate BindCraft campaigns and recorded 4,475 total trajectories. From that pool, 464 binders passed the platform’s acceptance criteria, yielding per-campaign acceptance rates that ranged between 2.3% and 67.6% as reported.
The authors noted campaign-level variability in acceptance, implying that target context and design constraints materially affect yield. The preprint did not detail campaign-specific trajectory counts, nor did it provide per-target acceptance breakdowns beyond the quoted overall range. The platform’s throughput—thousands of trajectories feeding into downstream orthogonal filters—illustrates a high-volume, iterative approach to de novo peptide design.
From the accepted set, the group selected three leads for deeper characterization. PRODIGY predicted binding free energies (ΔG) for these leads in the range −9.1 to −11.1 kcal·mol−1. The source mapped those values to dissociation constants (Kd) spanning roughly 8 nM to 0.2 μM. These are the numbers reported by the authors; no additional experimental affinity measurements were provided in the preprint.
All three leads were subjected to 100 ns of unrestrained molecular-dynamics simulation in the reported GROMACS/CHARMM36m environment. During those trajectories the receptor Cα RMSD remained ≤ 2.6 Å, the peptides retained their fold, and the authors reported that ≥ 99.5% of frames preserved more than half of the initial interface contacts. Helical content for the peptides was 75–82% across the MD runs. The study team presented these MD outcomes as evidence that the candidates are physically stable on the 100-ns timescale used for validation.
The platform’s developability stage used ADMET-AI to profile candidates before MD validation. ADMET-AI flagged the expected poor oral bioavailability typical of large, polar peptides; this was the only developability liability explicitly reported. The authors wrote that no PAINS alerts were raised for the selected leads. The preprint did not list additional ADMET descriptors, nor did it provide numerical oral-bioavailability predictions or other specific pharmacokinetic parameters.
The investigators emphasized that standard single-method outputs—ipTM and pLDDT among them—can be internally consistent yet fail to predict whether a design will hold up in explicit-solvent dynamics. In response, UnibindMD enforces independent corroboration: structure-confidence metrics from AlphaFold3-class / Boltz-2 are treated separately from affinity scoring and from MD stability. The pipeline’s acceptance statistics and the MD outcomes were presented as empirical support for this multi-pronged validation strategy. The source made clear that converting single-prediction design into cross-validated, dynamics-tested workflows is the platform’s primary conceptual advance.
The preprint reported no competing interest and provided numerical outputs for campaign yields, confidence scores (ipTM and pLDDT up to 0.94), predicted ΔG/Kd ranges, and 100-ns MD stability metrics. The authors acknowledged the predictable developability issue of poor oral bioavailability for large polar peptides, as flagged by ADMET-AI. The source did not report experimental binding assays, in vivo efficacy data, or peptide synthesis and characterization results. It also did not disclose detailed design hyperparameters, campaign-level trajectory counts, or regulatory milestones.
Key unanswered questions left in the preprint include whether predicted affinities translate to biochemical potency in vitro, how MD-validated candidates perform in cell-based or animal models, and what synthesis or formulation strategies would be required to address the ADMET-AI–flagged bioavailability limitations. These items represent explicit gaps in the reported work that will require experimental follow-up before clinical or translational claims can be made.