iiinova — where your digital self keeps evolving

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.

100

In Pseudomonas, the TwoAYGGAY RNA family extends well past its Rfam annotation, with three further hairpins and a 3’ pseudoknot on an elongated basal stem. A covariance model rebuilt on the Minerva predictions found 16,086 matches across 1,148 of 1,324 strains.

210

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.

100

For any portion of a genome, Minerva returns a map of which positions pair with which, including pseudoknots and overlapping hairpins that are hard to capture otherwise. Nothing needs to be aligned first, and one pass per region is fast enough to scan a whole genome.

100

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.

12312

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.

100

The team asked whether antioxidants could substitute for PRDX1. Catalase, SOD1, trolox, glutathione, and NAC offered no protection. Relocalizing PRDX3 or PRDX4 from the mitochondria or ER to the cytosol restored resistance, identifying the damaging ROS as nucleocytoplasmic.

100

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.

100

More than 1,200 genes made cells more sensitive to a single inhibitor when repressed, and 300+ made them resistant. The olaparib combinations added 300 more sensitizers. All 32 screens are searchable at ddri-screens.gilbertlab.arcinstitute.org

100

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.

100

Arc Institute @arcinstitute.org · 03/09/2026

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.

141

In people, BCHAs serve primarily as a biomarker of reductive stress. They rise with resistance exercise, COVID-19, and heavy drinking. But, in mice, another isoform of lactate dehydrogenase, LDHC, has evolved to support a highly specific function: hypermotility in sperm.

100

Arc Institute @arcinstitute.org · 03/09/2026

Instead of a new source, BCHAs are generated by lactate dehydrogenase A (LDHA), the same enzyme that drives canonical lactate fermentation. In hypoxia, mitochondria stop oxidizing branched-chain ketoacids, which pile up in the cytosol and trigger LDHA to reduce them instead.

100

Arc Institute @arcinstitute.org · 03/09/2026

It started with a Jain lab dataset on hypoxia-enriched metabolites in mice (jain-lab-ucsf.github.io/hypoxia-meta...). Across all organs and timepoints, three molecules kept popping up: all branched-chain hydroxyacids (BCHAs), a product of fermentation in unicellular organisms, but not in mammals.

100

Arc Institute @arcinstitute.org · 26/08/2026

Nearly 300 teams have already submitted to the Virtual Cell Challenge leaderboard to see how their initial models rank on this year’s six metrics. Here’s the current standing:

121

Arc Institute @arcinstitute.org · 20/08/2026

Competitors will receive a validation dataset with 3 cell lines, each containing the non-targeting control profiles plus the CRISPRi knockdown gene IDs to predict for 300 perturbations per cell line. A test set with 3 distinct cell lines will be held out until the final phase (in Oct).

120

Arc Institute @arcinstitute.org · 20/08/2026

We also redesigned the scoring. We're using six complementary metrics, each scaled between the cell context mean and data from a real replicate experiment. The overall score is a single unweighted average over the six metrics and cell contexts.

120

Five patients received B12 supplementation with or without immunosuppression, and four out of five improved. These treatment effects remain anecdotal, but motivate the team to eventually run a controlled trial.

They found that anti-CD320 was associated with a low B12 in the spinal fluid but a normal B12 in the blood, pointing to a potential new disease etiology in these patients.

210

In every patient who carried it, the antibody bound the same stretch of the CD320 extracellular domain. The team created an epitope-specific immunoassay to screen multiple cohorts from all over the world and track prevalence, metabolic consequences, and comorbidities.

100

The team ran proteome-wide phage display on CSF from patients with idiopathic myelopathy and disease controls. Of 20 enriched autoantibodies, they found one (anti-CD320) that targeted a cell-surface receptor on the blood brain barrier and was present in ~50% of patients.

100

Four years ago, a mystery case led Core Investigator @jvpluv to discover a new diagnosis (ABCD). Today his team published one of the largest clinical surveys of unexplained spinal cord disease to date in @JAMANeuro, finding ABCD in roughly half of patients.

163

The team also combined mutations into more complex variants. Every nine-substitution variant increased thermal stability while maintaining strong antibody binding, showing how the model can move beyond individual mutations toward designing proteins with multiple beneficial changes.

100

ProteinDPO identified 6 of 8 known stabilizing mutations reported in earlier H3 and H1 studies, with no HA data in training. Of its top 30 H5 candidates tested, 19 were more thermostable than wild type.

100

The team then applied ProteinDPO to H5N1 hemagglutinin, where improving protein stability is important for vaccine development and is typically tackled through rational design or large-scale experimental screens.

100

Trained only on stability changes in small protein monomers, ProteinDPO generalized to larger proteins with unseen folds and the stability of multi-chain antibodies. It also improved predictions of binding affinity in protein-protein complexes.

100

The resulting ProteinDPO model outperformed both the original ESM-IF1 and supervised fine-tuning at predicting changes in protein stability. It also matched supervised models built for the task alone.

100

Direct Preference Optimization (DPO) aligns LLMs by showing them a preferred and a rejected response to the same prompt. The team applied this to ESM-IF1, using a protein backbone as the prompt, sequences as the response, and higher measured stability as the preference.

100

The same method that teaches an LLM which answers people prefer can teach a protein model which sequences are more stable. This is the basis for ProteinDPO, developed by Innovation Investigator @brianhie.bsky.social, Talal Widatalla, and team.

182

On-demand excision opens applications like removing disease-causing repeat expansions cleanly, without needing a DNA donor at all. With this paper, the authors hope to offer a practical guide for optimizing bRNA scaffolds to favor insertion or excision.

100

The team also found that Handshake Guides (HSGs), small RNA elements that set the reaction's direction, are dials driving the reaction. Wild-type HSGs are tuned almost entirely for insertion. Rewriting them boosted excision efficiency 13,000-fold in E. coli.

100

There's a second brake: in excision, the guide loops that recognize DNA can no longer cleanly tell target from donor DNA, since the substrates are now chimeric. That confusion can stall the complex entirely, possibly a safeguard against a transposon accidentally excising itself.

Using cryo-EM, the team found the imbalance is driven by the geometry of the DNA within the complex. Insertion bends DNA into a strained U-shape, energetically driving strand exchange forward. Excision leaves DNA in a flat, relaxed X-shape with less energy to favor recombination.

100

Structurally, insertion and excision complexes are identical without bound DNA, forming the same tetramer to drive recombination with a Holliday junction-like intermediate. So, why is natural excision so sluggish compared to insertion?

100

In 2024, we introduced universal DNA rearrangement with bridge RNA. Now, in @Nature, @pdhsu @hnisimasu & team solve the structural mechanism of bridge excision—how the complex assembles for intramolecular recombination—yielding new engineering handles for programmable biology.

132

Low oxygen rescues the growth defect in THAP12-deficient cells, consistent with earlier work showing that hypoxia can relieve Complex I dysfunction. This nominates low-oxygen therapy as a possible treatment for THAP12-related disease.

100

THAP12 was recently linked to a severe recessive neurodevelopmental disorder with epilepsy. Fibroblasts from these patients show the same Complex I deficiency, suggesting the disorder is rooted in the loss of Complex I regulation. (Clinical preprint: doi.org/10.64898/202...)

100

Without THAP12, Complex I subunits across all three modules fell, mitochondrial respiration dropped, and cells produced less ATP. Restoring THAP12 only in the nucleus, not the cytoplasm, rescued the complex.

THAP12 controls Complex I from outside mitochondria. In the nucleus, it binds and activates the genes that build the complex, including NDUFAF3, NDUFAF4, and the iron-sulfur cluster factor BOLA3.

100

The team ran genome-wide CRISPR screens in human cells and stained for each of Complex I's three modules to see which genes its assembly depends on. Loss of THAP12 ranked in the top 1% for destabilizing the complex, despite not being a mitochondrial protein.

Complex I is the largest complex of the electron transport chain and the entry point for electrons that power ATP production. New work from Core Investigator Isha Jain and team identifies THAP12 as a dedicated transcription factor controlling how much Complex I a cell builds.

141

Thanks to all of Arc's Core Investigators, Technology Center Leaders, Science Fellows, Innovation Investigators, and Ignite Awardees for an exciting few days of science. We're already looking forward to next year!

000

Our 3rd Annual Investigator Retreat brought together talks on genome engineering, generative protein models, and computational biology, alongside deep dives into neurodegeneration, immunology, and metabolism.

130

HTRA2 also partners with CLPB, another intermembrane-space protein that causes the same disease when mutated. HTRA2 degrades protein aggregates, while CLPB pulls them apart again. Losing either destabilizes Complex I, suggesting that they act as a single quality-control system.

100

With Complex I impaired, mitochondria burn less oxygen, and it accumulates in the tissue, damaging oxygen-sensitive proteins. The team used those proteins as a readout and found that hypoxia both cleared excess oxygen and restored protein levels.

100

HTRA2 loss destabilizes Complex I, the entry point of the respiratory chain. Low oxygen restored Complex I protein levels but not its enzymatic activity, suggesting the rescue worked through a mechanism other than respiration.

100