Arc Co-Founder and Executive Director Silvana Konermann discusses what motivates her to use AI to understand complex human disease in this video from The Audacious Project.
Learn more about Arc's Virtual Cell Initiative: arcinstitute.org/virtual-cell...
This work reveals a new layer of cell biology connecting mTOR signaling to ferroptosis susceptibility through control of cholesterol and antioxidant uptake, and points to a potential combination strategy pairing mTOR + GPX4 inhibition.
Full paper here: www.cell.com/molecular-ce...
HDL, acting through SCARB1, can completely block ferroptosis even when GPX4 is knocked out. LDL offered no protection.
Cells without SCARB1 lost HDL protection & showed elevated lipid peroxidation, confirming an HDL-SCARB1 ferroptosis suppression system regulated by mTORC1.
To find out, the team ran a CRISPRa gain-of-function screen. SCARB1 (the HDL receptor) was the top hit for ferroptosis resistance.
The team found that mTORC1 inhibition downregulates SCARB1 expression, likely cutting cells off from a key lipid-protective pathway.
This held across multiple cancer lines & ferroptosis-inducing compounds. Adding RM-006 sensitized cells to RSL3, ML210, Erastin, FIN56 & BSO.
The question was why mTOR-inhibited cells were so specifically vulnerable to ferroptosis.
The team performed genome-scale CRISPRi screens in glioblastoma cells treated with mTORC inhibitor RM-006 & found that GPX4 ranked #1 among only 6 sensitizing hits.
Loss of GPX4 pushed mTOR-inhibited cells from cytostasis into ferroptotic death.
mTOR inhibitors stop cancer cells from dividing, but they rarely kill them. New work in Molecular Cell from Luke Gilbert & lab reveals a mechanism connecting mTOR signaling, cholesterol biology, and ferroptosis that could turn growth arrest into cancer cell death.
Together, these findings show ENPP1 protects brown fat from immunometabolic dysfunction and that K173Q undermines this by reducing cGAMP hydrolysis.
Whether the same mechanism drives human T2D is the next question.
Read the full preprint: www.biorxiv.org/content/10.6...
But it doesn’t only act within brown adipocytes. Extracellular cGAMP also acts as a local alarm signal, recruiting macrophages to brown fat and pushing them toward a pro-inflammatory state. As a result, brown fat shifts from a glucose sink into a driver of insulin resistance.
Where does that cGAMP come from? Brown adipocytes are unusually mitochondria-rich, and under nutritional stress mitochondrial DNA triggers cGAMP production and export. Without ENPP1 to clear it, an immune response shuts down glucose uptake and fat synthesis.
Screening glucose uptake across tissues, the team found brown adipose tissue was the primary organ that failed to respond to insulin. Without ENPP1's hydrolytic activity, extracellular cGAMP had built up to far higher levels than in normal mice.
Mice engineered to lack ENPP1's cGAMP-clearing activity, essentially an exaggerated model of K173Q, are far more susceptible to insulin resistance on a high-fat diet. Mice with normal ENPP1 activity were much less affected.
ENPP1 is an enzyme that destroys a key immune signal called cGAMP.
The team found that one of its common variants, K173Q, reduces ENPP1’s cGAMP-clearing activity by about 20%. But how does that influence T2D risk?
For decades, we’ve known that the ENPP1 K173Q variant raises type 2 diabetes risk, but not why.
New work from @songnanwang.bsky.social in @lingyinli.bsky.social's lab, in collaboration with @svenssonlab.bsky.social, shows that this variant may promote metabolic dysfunction in brown adipose tissue.
As genomic foundation models scale, hierarchical architectures like this may offer the efficiency edge needed to reach eukaryotic genomes and beyond.
Explore the full preprint here: arxiv.org/pdf/2602.10603
dnaHNet also predicts which genes a bacterium can't live without. Insert a premature stop codon in silico, measure how much the model's likelihood drops, and you get a zero-shot essentiality score, beating StripedHyena 2 across every compute budget tested on 62 bacterial genomes.
When matched by parameter count on an H100, dnaHNet achieves 3-6× higher throughput and 3-4× lower peak memory than StripedHyena 2 (the architecture behind Evo 2) at sequences ≥32K nt. For example, dnaHNet (218M) hits 939K tok/s at 65K nt vs. 156K tok/s for SH2 (234M).
Nobody told the model what a codon is. It found them anyway.
Stage 1 locked onto the triplet structure of coding sequences. Stage 2 went further, flagging promoters and intergenic regions as the genome's functional seams, all from raw sequence.
The model operates in two recursive stages: Stage 1 learns local sequence structure while Stage 2 captures broader functional organization. Together, they compress raw nucleotides into efficient latent representations without any labeled biological data.
Codons, promoters, and splice sites each carry biological meaning. Fixed rules cut through them mid-element, and current models learn from those fragments rather than the biological units they came from.
dnaHNet dynamically learns where boundaries should be during training.
Most genomic AI models use fixed rules to process DNA into chunks, imposing arbitrary boundaries on a sequence with its own biological structure.
Arnav Shah, Victor Li, and team developed dnaHNet, a tokenizer-free foundation model that learns its own segmentation from scratch.
The model, now on @biorxivpreprint.bsky.social, was trained on 130K proteins using GPT-5 to generate expert-style reasoning traces.
See the preprint: www.biorxiv.org/content/10.6...
Then BioReason-Pro goes even further. It takes in protein sequence and structure alongside biological context (domains, interaction partners, organism) and generates a step-by-step explanation of its prediction.
Try it here: bioreason.net
It first uses GO-GPT, a new model the team developed to predict Gene Ontology terms, the standardized vocabulary biologists use to describe what a protein does, where it acts, and processes it is involved in.
Over 250 million protein sequences are known, but fewer than 0.1% have confirmed functions. Today, @genophoria.bsky.social, @bowang87.bsky.social & team introduce BioReason-Pro, a multimodal reasoning model that predicts protein function and explains its reasoning like an expert would.
Want to contribute to similar projects? See our jobs page for open roles (including directors in ML and computational biology, postdoctoral researchers, and scientists in genome editing and functional genomics): arcinstitute.org/jobs
It's been a busy start to 2026 at Arc. We celebrated one year of Evo 2, announced findings from the Jain Lab and Thaiss Lab, and released MULTI-evolve, a new AI framework for designing complex proteins.
Stay tuned for the recording. In the meantime, see the full program and speakers: conferences.ted.com/ted2026
Arc Co-Founder and Executive Director Silvana Konermann is speaking at TED2026 in Vancouver (April 13-17). She'll be sharing her vision for Arc's Virtual Cell Initiative, recently announced as part of The Audacious Project's 2025 cohort, on one of the world's biggest stages.
