TXpredict

Most sequenced microbes have never had their transcriptome measured. Many cannot be cultured at all, and even when they can, RNA-seq demands organism-specific lysis and rRNA-depletion protocols that make cross-species surveys impractical. The result is a widening gap: genome sequences accumulate far faster than any functional readout of which genes those genomes actually express.

TXpredict, from the Advanced Research Institute of Multidisciplinary Sciences at the Beijing Institute of Technology with collaborators at Harvard University and the Shenzhen Institute of Synthetic Biology (SIAT), closes part of that gap by predicting a microbe's transcriptome from its annotated genome and nothing else. Rather than modelling promoters or untranslated regions — the usual route, and one that generalises poorly across species — it embeds every annotated protein-coding gene with ESM-2 and learns to map those embeddings to expression level. The premise is that evolutionary constraints on protein sequence carry information about how much of that protein a cell needs.

Trained on public RNA-seq for 22 bacterial, 9 archaeal, and 9 fungal species and applied zero-shot to genomes it has never seen, TXpredict was used to build TXpredictDB, predicted expression profiles covering 9.8 million genes across 2,685 genomes and 1,744 genera — 82% of which have no RNA-seq data in GEO at all.

#Key Features

• Genome-only input: The model needs a genome assembly and its annotation, no expression measurements, no close relative, and no organism-specific protocol development.
• Domain-specific checkpoints: Separate bacterial, archaeal, and fungal models are released, and cross-domain testing shows they encode partly shared principles of expression regulation while performing best within their own taxon.
• Beats homology transfer: Against a BLASTP nearest-neighbour baseline that copies a homolog's expression value, TXpredict wins on 21 of 22 bacterial genomes, with mean Spearman 0.53 vs. 0.38 for bacteria, 0.44 vs. 0.37 for archaea, and 0.62 vs. 0.39 for fungi.
• A condition-dependent variant: A second model adds a one-hot 5'UTR branch (-100 to +100 bp around the transcription start site) and predicts how genes respond across experimental contexts, reaching an average correlation of 0.52 over 13 bacterial strains and 4,619 conditions.
• Interpretable sequence attribution: Gradient-times-input analysis on the UTR branch recovers real regulatory features, including the -10 box of the sigmaS-dependent yadV promoter in E. coli and a Zur binding motif in P. putida.

#Technical Details

The transcriptome model is deliberately small: a single transformer encoder layer with 4 attention heads over a 128-dimensional projection with sinusoidal positional encoding, followed by 512- and 128-dimensional feedforward layers and a scalar output. Input is 1,301-dimensional — an ESM2-650M (esm2_t33_650M_UR50D) embedding plus normalised protein length and the 20-dimensional amino acid composition — with proteins over 1,500 residues excluded from training. Training used Adam at a learning rate of 2e-4 for 5 epochs on a single 24 GB GPU, averaged over five random seeds.

Ground truth came from NCBI GEO: 24.8 million expression measurements across 2,649 bacterial, 169 archaeal, and 2,078 fungal RNA-seq samples, normalised to RPKM and averaged per species. Leave-one-genome-out cross-validation gives mean Spearman correlations of 0.53 (bacteria), 0.44 (archaea), and 0.62 (fungi), reaching 0.70 for the fungus O. oligospora. Prospective RNA-seq on Pseudomonas sp. WBC-3 gave 0.40 across all proteins and 0.50 for GO-annotated ones. On ProteinGym yeast deep mutational scanning sets, predicted expression shifts correlate with measured fitness less well than ESM-IF1 or ProteinMPNN but better than ESM-2 itself. Codon-aware embeddings from CodonTransformer underperformed the protein-based inputs.

#Applications

TXpredict serves microbial ecology, comparative genomics, and synthetic biology groups working on organisms that resist culture. Predicted profiles give a prior for which genes matter in an uncharacterised genus — conserved highly expressed genes in a pathogen are candidate drug targets, and predicted expression of biosynthetic gene clusters (11,400 identified by antiSMASH in TXpredictDB) helps prioritise natural-product discovery. A whole 4,600-gene genome takes about 22 minutes in the provided Colab notebook, so the tool is usable without local GPU infrastructure.

#Impact

TXpredict is a compact demonstration that a protein language model trained without any expression supervision still encodes signal about expression level, transferable to genomes and taxa outside the training distribution. Its accuracy is moderate by design and by data: correlations in the 0.4-0.6 range are useful for ranking and screening but not for quantitative modelling of a single gene. The authors are explicit that the model ignores cis-regulatory elements, that archaeal training data is thin, that TXpredictDB gives one reference profile per genome and therefore no dynamics, and that the condition-dependent model cannot extrapolate to experimental conditions absent from training. Code and the three pretrained checkpoints are released on GitHub under the MIT licence, with weights and TXpredictDB mirrored on HuggingFace; the work remains a preprint.