Fungal cellulase training corpus and anchor-guided generated candidates

Training datasets and model-generated candidates for a conditional variational autoencoder pipeline (AGCLS) that designs fungal cellulase variants by sampling latent-space neighbourhoods around experimentally validated anchor enzymes. Contents Dataset Description Phase 1 corpus 1,212 fungal cellulase sequences collected from UniProt and NCBI. Phase 2 corpus 3,845 sequences remaining after filtering by instability index (18.26–35.00) and isoelectric point (4.05–6.00). Anchor sequences 2 experimentally validated enzymes: PF2 (465 aa) and PF3 (411 aa). These sequences were excluded from Phase 2 training and used as latent-space anchors. Generated candidates 25 model-generated variants: 10 generated around PF2, 10 around PF3, and 5 generated from random latent-space sampling. Predicted instability indices range from 21.80–32.24. Raw assay data Raw DNS reducing-sugar assay measurements (triplicate) for six Phase 1 candidates, together with buffer controls, substrate blanks, and commercial positive controls. Experimental validation Anchor selection is supported by the assay data included in this repository. PF3 exhibited significantly higher cellulase activity than the commercial positive control (Welch's t-test, p = 0.0003, n = 3). PF2 showed higher mean activity than the commercial positive control but did not reach statistical significance with three replicates (p = 0.067). The 20 Phase 2 candidate variants generated around PF2 and PF3 are computationally prioritised predictions and have not been experimentally expressed or assayed. Code https://github.com/computational-genomics-lab/Protein_stability_prediction