ApexGPT: generated sequences, predictor outputs and analysis code for "Foundation model-guided latent spaces enable functional protein design"

Data and analysis code supporting "Foundation model-guided latent spaces enable functional protein design" (Nature Chemical Biology, NCHEMB-A260601465). ApexProt is a training framework that uses a pretrained protein language model as a latent-space teacher for functional protein generation. ApexGPT couples an autoregressive transformer initialized from ProGen2 with ESM2, optimized by policy gradient. This deposit contains the material needed to reproduce every quantitative claim in the paper. Contents: 01_generated_sequences — every sequence set behind a figure or table row, named for the row it produces, including the ten matched-checkpoint arms of the epoch sweep with the generation script and run log for each, and the 1,020,000-sequence post-hoc control pool.02_reference_sets — the AMP and MDH training sets and the four CD-HIT clusters.03_alignments — the MAFFT alignments behind the MSA entropy columns.04_derived_values — per-sequence APEX and Macrel predictions, entropy and sequence-similarity recomputations, the ESM2 rankings of the control pool, and the un-fine-tuned ProGen2 baseline.05_code — the analysis scripts, each described in INVENTORY.md.06_figure_source_data — the source data for each figure and supplementary table. MANIFEST.tsv lists every file with a SHA-256 checksum, the path it came from and what it produces. PROVENANCE.md records the metric definitions and join rules needed to reproduce the reported values, including two that are easy to get wrong: Avg_apex in Supplementary Table 1 is the mean of each sequence's median MIC across the eleven pathogen strains, not the median; and Macrel output must be joined by record order rather than sequence string, because Macrel strips the initial methionine. REPRODUCING.md walks from a fresh clone of the code repository to every metric reported in Figure 1d and Supplementary Tables 1 and 2.