Arc Institute

Much of the functional biology encoded in genomes remains hidden in noncoding DNA. In a preprint out today, Patrick Hsu and team introduce COMPASS, a genomic language model-based approach to systematically map this underexplored sequence space.

To explore the atlas at scale, the team deployed AI agents to develop analyses and investigate biological hypotheses under researcher guidance. They identified candidate divergent riboswitches, ribozymes, and previously uncharacterized PhoH2-associated RNAs that existing annotation methods missed.

Much of the functional biology encoded in genomes remains hidden in noncoding DNA. In a preprint out today, Patrick Hsu and team introduce COMPASS, a genomic language model-based approach to systematically map this underexplored sequence space.

Periodicity in Evo-PIN token entropies revealed the internal architecture of tandem repeat arrays. The pattern recovered guide arrays alongside their cognate effector proteins from sequence alone, and nominated new repeat architectures that don't belong to any known RNA-guided system.

Much of the functional biology encoded in genomes remains hidden in noncoding DNA. In a preprint out today, Patrick Hsu and team introduce COMPASS, a genomic language model-based approach to systematically map this underexplored sequence space.

The team fine-tuned Evo 2 on 750M prokaryotic intergenic sequences to develop Evo-PIN, which captures shared features of noncoding RNAs beyond sequence similarity. COMPASS used these representations to cluster 376M sequences into 5.14M clusters without sequence alignment or prior annotations.

Our cells carry the same DNA, but how this DNA is interpreted differs wildly from cell to cell. In a new preprint, Science Fellow Jingtian Zhou and team introduce Bolero, a model conditioned on cell state that predicts chromatin accessibility and expression from DNA alone.

Trained only on human data, Bolero's AP-1 model correctly predicted enhancer activity in mouse and macaque neurons it never saw in training. In effect, Bolero turns genome sequence into a queryable map of cell-state-specific regulation across cell types, activity states, and species.

Our cells carry the same DNA, but how this DNA is interpreted differs wildly from cell to cell. In a new preprint, Science Fellow Jingtian Zhou and team introduce Bolero, a model conditioned on cell state that predicts chromatin accessibility and expression from DNA alone.

The team gave Bolero a dial for AP-1 activity, a per-cell score that conditions the model alongside cell state. Turning this dial up in a virtual simulation, Bolero predicted the unique ways that specific cells changed their DNA accessibility up to 29x better than a motif scan.

Our cells carry the same DNA, but how this DNA is interpreted differs wildly from cell to cell. In a new preprint, Science Fellow Jingtian Zhou and team introduce Bolero, a model conditioned on cell state that predicts chromatin accessibility and expression from DNA alone.

That embedding comes from Bolero-10M, a new atlas of 10M+ single cells spanning the human brain and body, development and aging, and multiple species. This lets Bolero predict regulation for cell states it's never directly seen.

Our cells carry the same DNA, but how this DNA is interpreted differs wildly from cell to cell. In a new preprint, Science Fellow Jingtian Zhou and team introduce Bolero, a model conditioned on cell state that predicts chromatin accessibility and expression from DNA alone.

More than 5,000 people have registered for the Virtual Cell Challenge, and with over 1,500 teams on the leaderboard, we wanted to know who they are and why they’re competing: arcinstitute.org/news/vcc-com...

RNA-binding proteins (RBPs) recognize RNA by sequence and shape, but which matters varies by site. Introducing Pythia, an interpretable AI model trained on a new high-resolution atlas of 67 RBPs—a collab between @genophoria.bsky.social, @czbiohub.bsky.social, UCSF, Calico, and @vectorinstitute.ai.

Read a conversation between co-authors @genophoria.bsky.social and Manuel Leonetti about the work: arcinstitute.org/news/eclair-...

RNA-binding proteins (RBPs) recognize RNA by sequence and shape, but which matters varies by site. Introducing Pythia, an interpretable AI model trained on a new high-resolution atlas of 67 RBPs—a collab between @genophoria.bsky.social, @czbiohub.bsky.social, UCSF, Calico, and @vectorinstitute.ai.

Pythia learns RNA sequence and 2D folding directly, without external structural predictors, so it can pinpoint binding sites driven by structure. In tests on five RBPs, disrupting structures flagged by Pythia weakened binding, while preserving the fold rescued it.

Hundreds of human tRNA genes form a redundant repertoire for decoding the genetic code. But does redundancy imply functional equivalence? A pan-cancer in vivo CRISPRi atlas from @genophoria.bsky.social and Vijay Ramani at @Gladstone Institutes reveals a structured landscape of tRNA dependencies.

The changes were coordinated across cancer cell lines, revealing tRNA isoacceptors associated with tumor-related adaptation. Distinct codon usage preferences of proliferation- and tumor stress-associated programs provided further biological context.

Hundreds of human tRNA genes form a redundant repertoire for decoding the genetic code. But does redundancy imply functional equivalence? A pan-cancer in vivo CRISPRi atlas from @genophoria.bsky.social and Vijay Ramani at @Gladstone Institutes reveals a structured landscape of tRNA dependencies.

Are tRNA requirements fixed, or shaped by where cells grow? To find out, the team compared matched CRISPRi screens during tumor growth in vivo and cell culture in vitro.

Hundreds of human tRNA genes form a redundant repertoire for decoding the genetic code. But does redundancy imply functional equivalence? A pan-cancer in vivo CRISPRi atlas from @genophoria.bsky.social and Vijay Ramani at @Gladstone Institutes reveals a structured landscape of tRNA dependencies.

Across 16 cancer cell lines from 7 tissue types, tRNA dependencies in living tumors showed tissue-specific organization. At isodecoder resolution, dependency profiles classified tissue of origin with 80% accuracy.

Most perturbation models learn only from experiments. Our new model, PIE, also draws on decades of curated biological knowledge to predict which genes respond to a perturbation, even in biological contexts and for perturbations it has never seen.

On the Replogle-Nadig dataset, PIE was up to 3.2x better than the strongest baseline at identifying differentially expressed genes, with the biggest gain on unseen perturbations. It still made meaningful predictions when context and perturbation were both unseen.

Most perturbation models learn only from experiments. Our new model, PIE, also draws on decades of curated biological knowledge to predict which genes respond to a perturbation, even in biological contexts and for perturbations it has never seen.

Virtual cell models are only as good as the data they learn from. To predict how any cell type responds to a perturbation, models need high-quality causal data generated across contexts, so technical noise doesn't masquerade as biology.

Many important bacterial non-coding RNAs and structural elements remain undiscovered. To help, David Li, Garyk Brixi, Michael Fischbach, @brianhie.bsky.social & team introduce Minerva, which uses a genome language model to predict RNA base pairing, repeats, & other interactions from sequence alone.

In prophages, Unknown Group 27 reverse transcriptase loci hold arrays of roughly 150nt mcRNA units, diverse in sequence but sharing a predicted structure. Expressed in E. coli, all five systems reverse transcribed the central hairpin of each unit into 50-100nt ssDNA.

Many important bacterial non-coding RNAs and structural elements remain undiscovered. To help, David Li, Garyk Brixi, Michael Fischbach, @brianhie.bsky.social & team introduce Minerva, which uses a genome language model to predict RNA base pairing, repeats, & other interactions from sequence alone.

Across 150 bacterial genomes, Minerva predicted 62,083 loci carrying at least two neighboring hairpins, and 70.2% of those sit outside existing annotations. The full scan took 100 minutes on one H100.

Many important bacterial non-coding RNAs and structural elements remain undiscovered. To help, David Li, Garyk Brixi, Michael Fischbach, @brianhie.bsky.social & team introduce Minerva, which uses a genome language model to predict RNA base pairing, repeats, & other interactions from sequence alone.

Loss of function for 1 gene can make a cancer cell resist a DNA damage response (DDR) inhibitor, while another gene can make it easy to kill. Arc Core Investigator Luke Gilbert and collaborators mapped both across five DDR inhibitors and found loss of PRDX1 sensitizes cells to all five.

The team then repeated the screens in PRDX1 knockout cells. Knocking down IREB2, PAX7, or MRGBP restored growth under DNA-PK inhibition and lowered γH2AX. All three lowered labile iron, and chelating iron partly rescued the cells, indicating PRDX1 protects the genome from iron-driven oxidation.

Loss of function for 1 gene can make a cancer cell resist a DNA damage response (DDR) inhibitor, while another gene can make it easy to kill. Arc Core Investigator Luke Gilbert and collaborators mapped both across five DDR inhibitors and found loss of PRDX1 sensitizes cells to all five.

Six genes consistently sensitized cells to all inhibitors when repressed. One was PRDX1, a peroxiredoxin that reduces hydrogen peroxide. Knocking it down raised γH2AX and sensitized cells to the PARP, ATR, ATM, and DNA-PK inhibitors.

Loss of function for 1 gene can make a cancer cell resist a DNA damage response (DDR) inhibitor, while another gene can make it easy to kill. Arc Core Investigator Luke Gilbert and collaborators mapped both across five DDR inhibitors and found loss of PRDX1 sensitizes cells to all five.

The team screened a genome-scale CRISPRi library in A549 lung cancer cells against the PARP inhibitor olaparib and against inhibitors of the repair kinases including ATR, ATM, DNA-PK, and WEE1. They screened each drug on its own as well as combined with olaparib.