Trained checkpoints: graph neural network Hamiltonian and direct models for DNA charge transport

Abstract

Trained PyTorch checkpoints for every run reported in the accompanying paper: graph neural network (GAT) models that predict DNA transmission and density-of-states spectra, either through a learned reduced-order Hamiltonian passed through an NEGF layer (Hamiltonian model) or directly from pooled graph features (direct model). Part of ICLR 2027 conference submission number 50206. Files:- campaign_v3.tar (691 MB): the full-factorial campaign, 120 runs. Hamiltonian model: supervision (DOS+T, LDOS+DOS+T, LDOS+T, T-only) x orbitals per base (1, 2) x GNN layers (2, 4) x geometry channel (off, on) x 3 initialization seeds = 96 runs. Direct model: supervision (DOS+T, T-only) x GNN layers (2, 4) x geometry channel (off, on) x 3 seeds = 24 runs.- capacity_probe_v5.tar (108 MB): the capacity probe, 12 runs. Both reported models retrained with a deeper readout or a wider GNN (hidden width 512), at 3 seeds each.- MANIFEST.tsv (46 KB): every archive member with its SHA-256 checksum, size, cell name and seed.- README.md (3.3 KB): extraction instructions and how to identify a run.- SHA256SUMS (249 B): checksums for the archives and the manifest. Each run directory, _s , holds the weights at the epoch that minimized that run's own validation objective (not the final epoch) and its fully resolved training configuration (resolved_config.json). Initialization seeds are 1179027592, 2129768291 and 3731635825; all runs share one 80/20 sequence-level train/validation split (split seed 42). Usage:Extract both archives at the root of the accompanying code repository (https://anonymous.4open.science/r/G3NAT/); the analysis scripts in its analysis/ directory reproduce every reported number from these checkpoints and the companion dataset (https://doi.org/10.5281/zenodo.22964053). The two models shown in the paper's figures (seed 3731635825) are also included directly in that repository under trained_models/.