Daily Positive Signals: a schizophrenia first, an open-weight AI workhorse, and a second chance for virtual cells
On this page
An FDA-cleared app treats schizophrenia’s hardest symptoms, a startup open-sources a giant efficient AI model, and a new study says the yardstick for AI cell models was broken all along.
In this edition
- A first treatment for schizophrenia’s negative symptoms · Medicine
- Reflection’s Beam, an open-weight AI workhorse · Artificial intelligence
- A second chance for AI virtual cells · Biotechnology
Signal 1 of 3 · Medicine
A first treatment for schizophrenia’s negative symptoms
62% relative improvement in negative symptom severity (CAINS-MAP) at 16 weeks versus a digital control app
0 previously authorized treatments indicated specifically for these symptoms in the United States
The U.S. Food and Drug Administration granted 510(k) clearance on October 5 to Motivista, a prescription smartphone app from Click Therapeutics, for the negative symptoms of schizophrenia, including low motivation, reduced ability to feel pleasure, and social withdrawal (Click Therapeutics, 2026). In the Phase III CONVOKE trial, patients using Motivista improved 6.8 points on the CAINS-MAP scale at 16 weeks versus 4.2 points for a digital control app, a 62% relative improvement with a Cohen’s d of −0.36 and p=0.0003 (Click Therapeutics, 2026). The trial reported no serious treatment-related adverse events, though two episodes of “psychotic disorder” in a single participant were judged possibly related (Click Therapeutics, 2026).
Why it matters: Existing antipsychotics target positive symptoms like hallucinations and delusions, while the disabling negative symptoms affect roughly 60% of the estimated 24 million people living with schizophrenia and had no authorized treatment (Click Therapeutics, 2026; Fierce Biotech, 2026).
Evidence status: Phase III randomized controlled trial plus FDA 510(k) clearance, announced by the company and corroborated by independent trade coverage. The therapy is prescription-only, an adjunct to clinician-managed antipsychotic therapy for adults 18 and older with stable positive symptoms, not a standalone or acute treatment. Commercial availability and pricing are pending.
Signal 2 of 3 · Artificial intelligence
Reflection’s Beam, an open-weight AI workhorse
501B total parameters, with only 23 billion active per token in this mixture-of-experts model
1M token context window; the model was pretrained on 23.8 trillion tokens
Reflection AI, a two-year-old startup founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, unveiled its first frontier model on October 5: Beam, a text-only mixture-of-experts model aimed at reasoning, coding, and agentic tasks (TechCrunch, 2026; Reuters, 2026). The company says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute, and that it will release Beam’s weights under an Apache 2.0 license this month, with full technical details to follow (TechCrunch, 2026).
Why it matters: An efficient open-weight model that enterprises and public-sector bodies can run and customize on their own systems would widen access to capable AI beyond closed models from the biggest labs (TechCrunch, 2026).
Evidence status: Company announcement corroborated by two independent newsrooms (TechCrunch, Reuters). The performance claims are company-reported and have not been independently verified; weights are not yet released (an early version is on a waitlist while the model completes red-teaming). Treat the efficiency numbers as provisional.
Signal 3 of 3 · Biotechnology
A second chance for AI virtual cells
18 benchmarking metrics tested; weighted metrics consistently distinguished real biological signal from uninformative baselines
9 deep learning models re-evaluated, and under calibrated metrics even earlier models beat the mean baseline
A study published in Nature Biotechnology shows that recent benchmarks concluding AI “virtual cell” models cannot beat simple baselines were measuring with a broken ruler (Shift Bioscience, 2026; Deep learning perturbation models can outperform baselines on calibrated metrics, 2026). Widely used metrics like mean squared error and control-referenced correlation suffer from control bias and signal dilution, letting uninformative predictors score deceptively high; the authors’ new “dynamic range fraction” (DRF) quantifies how well a metric distinguishes real from fake signal, and weighted metrics like WMSE consistently score higher calibration across datasets (Shift Bioscience, 2026). Under well-calibrated metrics, nine deep learning models, from scGPT and GEARS to newer systems like PRESAGE and scLambda, mostly outperformed uninformative baselines on unseen genetic perturbations (Deep learning perturbation models can outperform baselines on calibrated metrics, 2026).
Why it matters: AI virtual cells predict how cells respond to genetic interventions, the basis for in-silico drug target screening. If the failure was in the measurement rather than the models, this approach is more viable than the last round of benchmarks suggested (Shift Bioscience, 2026; Boston Life Sciences Times, 2026).
Evidence status: Peer-reviewed study in Nature Biotechnology, announced by Shift Bioscience and corroborated by independent coverage. Shift plans to use the calibrated framework in large-scale lab and computational screens for targets relevant to cell rejuvenation and fibrosis, following its first target SB-101 (Shift Bioscience, 2026). Clinical applications remain downstream of substantial further work.
Editor’s note
Today’s three signals share a quiet theme: the measurement was the breakthrough. Schizophrenia’s negative symptoms finally have an authorized treatment because a trial measured them with a dedicated scale instead of borrowing from positive-symptom psychiatry. An AI startup’s pitch rests on measured efficiency, reasoning per unit of compute, even as the field waits for independent verification. And a celebrated AI-for-biology setback turns out to be a scoring artifact. Progress often hides inside how we keep score, and fixing the yardstick can unlock more than the measurement itself.
That’s today’s three: a treatment first for schizophrenia’s negative symptoms, an open-weight AI workhorse, and a recalibrated view of virtual cells. Read past editions at /progress/. Tomorrow’s signals land at 07:15.
The Progress Brief, by email
Useful progress, without the hype
The daily Positive Signals brief and the weekly deep dive, by email. Evidence-checked, concise, no doom.
Free. One daily email, one weekly email. Unsubscribe anytime.
References
Click Therapeutics, Inc. (2026, October 5). FDA clears Motivista (CT-155), the first authorized treatment for negative schizophrenia symptoms. GlobeNewswire. https://www.manilatimes.net/2026/10/06/tmt-newswire/globenewswire/fda-clears-motivista-ct-155-the-first-authorized-treatment-for-negative-schizophrenia-symptoms/2439402
Fierce Biotech. (2026, October 5). Meet Motivista: Click Therapeutics debuts first-of-its-kind schizophrenia digital therapeutic. https://www.fiercebiotech.com/medtech/meet-motivista-click-therapeutics-first-its-kind-schizophrenia-digital-therapeutic
TechCrunch. (2026, October 5). Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost. https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/
Reuters. (2026, October 5). Nvidia-backed Reflection unveils first AI model to take on Chinese open models. https://srnnews.com/nvidia-backed-reflection-unveils-first-ai-model-to-take-on-chinese-open-models/
Shift Bioscience. (2026, October 5). Shift Bioscience publication increases confidence in AI virtual cells for novel target discovery. Business Wire. https://www.businesswire.com/news/home/20261005043215/en/Shift-Bioscience-Publication-Increases-Confidence-in-AI-Virtual-Cells-for-Novel-Target-Discovery
Boston Life Sciences Times. (2026, October 5). Shift Bioscience study supports AI virtual cells for drug target discovery. https://bostonlifesciencestimes.com/2026/10/shift-bioscience-study-supports-ai-virtual-cells-for-drug-target-discovery/
Deep learning perturbation models can outperform baselines on calibrated metrics. (2026). Nature Biotechnology. https://doi.org/10.1038/s41587-026-03307-w
