Trove1-70m

A single-cell foundation model trained on hundreds of millions of transcriptomes has still never seen a protein measurement. Transcript abundance is an imperfect stand-in for the molecules that actually do the work of a cell — post-transcriptional regulation, differential degradation and secretion all break the correlation — so the layer at which biological function is expressed is absent from pretraining. The authors of Tahoe-x1 named proteomics as the key missing modality in their own model. Trove1-70m, from Tesorai, is an answer that goes sideways rather than up: instead of training a bigger RNA-only model, take the published Tahoe-x1 70M checkpoint and continue pretraining it for one epoch on mass-spectrometry proteomes.

What makes that cheap is the tokenization. Each protein is mapped to the gene token of the gene encoding it, and its abundance is written into the expression-value slot, so a proteomic sample enters the model in exactly the format a transcriptomic cell does. Tahoe-x1's per-sample quantile binning absorbs the very different intensity scales of label-free, DIA and isobaric-labelling experiments. There is no separate protein encoder and no new vocabulary — which also means proteomic input is consumed through the RNA tokenizer rather than modelled natively.

The payoff is measured on the RNA side. Pretrained further on proteomes alone, the 70M model matches or exceeds RNA-only Tahoe-x1 checkpoints 14-fold and 50-fold larger on the transcriptomic benchmarks those checkpoints were published against. A control arm given a million extra bulk RNAseq samples instead of proteomes helps on some of those benchmarks but not on gene essentiality, which locates the gain in the second modality rather than in extra training.

#Key Features

• Proteins enter as gene tokens: Protein identity becomes an Ensembl gene token and abundance becomes an expression value, letting one encoder consume bulk RNAseq, single-cell RNAseq and mass-spec proteomes without architectural change.
• Agent-curated proteomics corpus: Public mass-spec deposits vary in layout, identifiers and metadata; per-deposit parsing agents, backed by human expert review of ambiguous cases, harmonise them into one training table.
• Format-aware augmentation: Isobaric-labelling data is sometimes deposited as ratios rather than absolute abundances, so each sample is presented in both raw and batch-corrected form.
• Held-out-drug perturbation benchmark: A protein-level benchmark scores whether a proteome representation predicts drug-induced abundance changes for compounds withheld from pretraining, under leave-one-pair-out cross-validation.
• Evaluated frozen: Results come from frozen embeddings fed to a cosine-similarity kNN or a probe, so nothing is refit per dataset.

#Technical Details

The training corpus comprises 48,843 samples from 440 public mass-spectrometry sources spanning cancer cell lines, tissues, body fluids, primary cells, organoids and drug-perturbation studies; sources contributing fewer than ten samples were dropped. No acquisition strategy dominates — 15,526 isobaric-labelling, 14,170 DIA and 12,541 label-free samples — making platform heterogeneity a property of the data rather than a confound. Deposits whose quantification was missing or computed with outdated search algorithms were re-analysed with Tesorai Search, and a share of the 370 agent-parsed deposits required human intervention the agent could not automate. On DepMap gene essentiality the continued-pretrained model reaches 0.796 ± 0.002 AUPRC against 0.739 for the 1B RNA-only baseline, and it improves tissue-of-origin separation, though a classical PCA embedding remains the strongest method on that task. Pathway membership on MSigDB Hallmarks is where cross-modal training helps least: it matches the 1B and 3B checkpoints within seed variability rather than surpassing them. On the protein-perturbation benchmark — 21 kinase inhibitors across four cardiomyocyte donors and 56 anticancer drugs across three cancer lines — top-100 differential-expression Pearson rises from 0.542 to 0.611 and 0.746 to 0.813 over the base checkpoint, while the 1B RNA-only model scores below the 70M base on both. Gains grow with corpus size from 300 to 48,843 samples.

#Applications

The checkpoint produces gene-level and cell-level embeddings for transcriptomic or proteomic input, supporting cross-study analysis of public proteomes, tissue and cell-line characterisation, and prioritising context-specific essential genes in a cancer background. Its most distinctive use is forecasting how a proteome shifts under a drug never seen in training, which suits target deconvolution and mechanism-of-action work where protein-level readouts are the endpoint but scarce.

#Impact

The result argues that curating a complementary modality buys more than adding parameters, with direct consequences for how single-cell foundation models are budgeted. It also sharpens an uncomfortable finding from the perturbation literature: on the protein-perturbation benchmark, highly-variable-protein selection and PCA outscore every foundation model tested, including this one. The work sits alongside ProteinTalks, which used perturbed-proteome prediction as a pretraining objective without evaluating held-out-drug transfer, and CAPTAIN, which learns from co-assayed single-cell RNA and surface-protein panels rather than bulk mass spectrometry. Reproduction is limited: the corpus pipeline depends on Tesorai's commercial search engine and expert curation, and while the preprint names a GitHub repository for the benchmark and a Hugging Face repository for the checkpoint, neither has been populated with code or weights.