ADhunter

Transcriptional activation domains (ADs) are short, intrinsically disordered stretches of protein that recruit the transcriptional machinery and switch genes on. They are unusually hard to find computationally: they have no conserved fold and little sequence conservation, so comparative genomics largely fails on them. The high-throughput reporter assays that can measure them have been run almost entirely on two model organisms — budding yeast and Arabidopsis thaliana — leaving predictors that recognize the well-studied acidic AD class but miss much of the rest.

ADhunter is a regression model that scores AD strength directly from protein sequence, built by Lucas Waldburger, Hunter Nisonoff and colleagues in the Staller and Shih labs at UC Berkeley, with Lawrence Berkeley National Laboratory and collaborators, and posted to bioRxiv in September 2025. Two choices separate it from earlier AD predictors such as ADpred, PADDLE, and TADA. It is trained against a continuous regression objective rather than a binary classifier, matching the fact that activation is graded rather than on/off. And it is a 20-member deep ensemble, so the spread across members supplies an epistemic uncertainty estimate for every prediction — something no prior AD predictor offered.

That uncertainty drives the study around the model. The authors scored 7.8 million proteins across 2,400 fungal genomes, used uncertainty to pick which sequences were worth measuring, ran a pooled yeast reporter assay on the selection, and folded the results back into training. The retrained model generalizes across fungal evolutionary space rather than around S. cerevisiae.

#Key Features

• Quantitative activity scores: Predicts a continuous AD activity value for each 53-residue tile rather than a binary activator call, resolving domain boundaries and peak strength within a protein.
• Per-prediction uncertainty: A 20-model deep ensemble reports variance alongside its mean prediction, flagging sequences far from the training distribution instead of failing silently on them.
• Uncertainty-guided evolutionary sampling: Quantile-balanced uncertainty sampling selected 8,935 maximally informative tiles spanning 1,050 fungal genomes, versus the 21 genomes represented in the starting dataset.
• Robustness outside the model organism: After active learning, held-out performance rose from Pearson r = 0.541 to r = 0.824, with the gain concentrated in non-Saccharomycotina sequences.
• Residue-level interpretability: SHAP analysis recovers the acidic exposure model — Asp, Phe, and Trp raise predicted activity while Lys and Arg lower it — and exposes an underrepresented, leucine- and phenylalanine-enriched non-acidic AD class.

#Technical Details

ADhunter is a dilated residual CNN head on frozen per-residue embeddings from ESM-2 (esm2_t33_650M_UR50D): a convolutional layer, a stack of residual blocks with batch normalization and ReLU, pooling, and a fully connected output. It minimizes mean squared error with Adam and early stopping. The 20 released checkpoints deliberately vary in hyperparameters — hidden width 32 to 256, kernel size 3 to 10, and up to 10 residual blocks — so ensemble members disagree in useful ways. Training began with 17,609 53-residue tiles from fungal and plant proteins measured by a GFP:mCherry sort-seq assay, then added 7,681 newly characterized tiles from non-model fungi plus harmonization and yeast control tiles.

Encoding comparisons, including against ESM-1b, found ESM-2 embeddings only marginally ahead of one-hot in distribution (r = 0.744 versus 0.742) but clearly ahead outside it: on a held-out spectral cluster, one-hot ADhunter reached r = 0.416 and the ESM version r = 0.476, with the full ensemble at r = 0.512 against TADA's 0.381. On the harmonized dataset ADhunter reaches r = 0.818 (RMSE 0.556) on the regression task versus TADA's 0.621, and 0.905 accuracy versus 0.861 when its outputs are binarized. Genome-wide inference covered 72 million tiles from 7,842,516 representative proteins in the MycoCosm collection.

#Applications

The immediate use is annotating fungal genomes: the work supplies the first functional annotation for 3,416 proteins from 670 non-model fungi, including zinc cluster transcription factors that control secondary metabolism and pathogenesis. For fungal engineering, quantitative AD scores support designing transcription factors with tuned output for inducible promoters, and finding cryptic activators that could wake silent biosynthetic gene clusters. More broadly, the assay-plus-active-learning framework is a template for any sequence-to-function property whose labeled data sits in a handful of model organisms.

#Impact

ADhunter reframes AD prediction as a regression problem with explicit uncertainty, and shows that the binding constraint on generalization is training-set breadth rather than model capacity — a point that carries well beyond transcriptional activators. The characterized set of 9,836 ADs from 1,071 fungal genomes is itself a 15.5-fold expansion in genome representation over prior datasets, with raw sequencing deposited under BioProject PRJNA1183837. The work remains a preprint awaiting peer review. Both code repositories are public — a pip-installable package bundling all 20 ensemble checkpoints, plus the earlier one-hot base model — but neither carries a license file, so reuse terms rest solely on the Berkeley Lab copyright notice in the README.