Paper2Agent turns papers into AI agents that apply methods to new data

Paper2Agent takes a paper’s text, code, and data and builds an agent that can answer questions about the work and collaborate with agents from other papers. Applying a method from an article is usually locked behind technical

🔗 Read original →

Human brain circuits grown inside mice after cortex replacement

Scientists genetically modified mice so that from birth they lacked almost the entire cerebral cortex, leaving only about 2% of normal cortical volume. Into this void they implanted laboratory‑grown human brain organoids in two‑day‑old newborn mice. After three months, more than 90% of the rodents’ cortical volume consisted of engrafted, functional human neurons that had integrated into the spinal cord and nervous system.

The approach overcame past limits where human cells competed with mouse neurons for space. Each implanted organoid contained roughly 100,000 cells. With the freed niche, researchers were able to grow the rare spindle‑shaped von Economo neurons (VENs), which occur in the human brain at a frequency of about 1 per 90,000 neurons and are linked to social behavior and decision‑making.

Previously VENs could only be studied post‑mortem, as they fail to form in petri dishes. This “xenocortical” model may aid research into neurological disorders. In additional tests, mice were subjected to a five‑hour oxygen deprivation, showing that the human tissue is many times more sensitive to hypoxia than mouse tissue and can provoke cerebral‑palsy‑like symptoms.

The method will allow personalized organoids to be grown from a patient’s skin and used to test precise drugs directly on living neural networks.

🔗 Read original →

Blocking ceramide synthesis pathways shows opposite effects on Alzheimer's symptoms in mice

On 11 September, researchers published a bioRxiv preprint comparing two strategies to lower ceramide levels in the PDAPP‑J20 mouse model of Alzheimer’s disease. Eight‑month‑old female mice received either GW4869, which blocks the neutral sphingomyelinase (nSMase) pathway, or myriocin, which blocks the serine‑palmitoyltransferase (SPT) pathway, for three weeks.

Mice treated with GW4869 learned the Barnes maze at the level of healthy controls, showed a reduced amyloid‑plaque area in one hippocampal region, and exhibited fewer signs of microglial activation around plaques. In contrast, mice given myriocin retained learning deficits and had a larger plaque area in that region compared with untreated transgenic animals.

In microglial cell culture, amyloid induced NFκB nuclear translocation; GW4869 blocked this translocation and lowered intracellular fluorescent amyloid. A chemically distinct nSMase inhibitor, cabotin, decreased TNF‑α production but did not reduce intracellular amyloid in the same assay.

Thus, the effect depended on the ceramide‑synthesis route: inhibiting nSMase with GW4869 improved learning and lessened amyloid burden, whereas inhibiting SPT with myriocin worsened some outcomes.

🔗 Read original →

Clarifying the Fly Brain Hype: No Consciousness in Bitcoin Trading

Media hype surrounding the digital fly brain often mischaracterizes simple machine‑learning tricks as signs of consciousness. For example, making the model “trade” Bitcoin can be extrapolated to any other stunt — playing Bad Apple, solving a Rubik’s cube, or waving a lightsaber.

Journalistic framings replace the technical process with anthropomorphic ideas like a “digital mind” or “conscious activity.” In reality, the fly connectome is not a functioning biological system nor a computational program; it is a static topological map of spatial connections among ~160,000 neurons and tens of millions of synapses.

This graph itself lacks neurotransmitter dynamics, electrical activity, or subjectivity. Any manipulation using the digital model follows a single algorithm: input data are converted into a signal matrix that mimics visual or sensory neuron stimulation.

To activate the static graph, it is embedded in external machine‑learning architectures such as Graph Neural Networks. Trading or key presses are executed not by the connectome itself but by an optimization model that uses the brain topology as a specific structured weight matrix.

Meanwhile, the model’s dopaminergic and aversive neural circuits are forcibly stimulated: profit from a trade or gameplay advantage is supplied as a positive mathematical reinforcement signal. Understand? No fly is learning to trade or play poker.

The procedure is a standard loss‑function optimization where the biological graph serves merely as an alternative connectivity architecture. Interpreting these experiments as steps toward “transferring consciousness” or “creating digitized intelligence” is inaccurate; they represent an engineering adaptation of biological data to applied machine‑learning tasks.

🔗 Source: @solid_state_humanity

UK AI Health Commission Recommends Staged Approval and Post‑Launch Monitoring

On September 10 the National Commission for AI Regulation in Healthcare published 44 recommendations for future UK rules on medical AI. The advice is directed to the UK Medicines and Healthcare products Regulatory Agency (MHRA), the government, and healthcare organisations.

Medical AI can change after updates and perform differently in various clinics, with outcomes influenced by data, workflows, and conditions of use. In the current UK system safety and performance checks focus mainly on pre‑deployment data, while post‑market surveillance relies heavily on incident reports.

According to recommendation 14, MHRA could initially authorise a device only under defined conditions of use — after reviewing initial data, risk‑mitigation measures, and reporting requirements. Later, the scope of use could expand if the system meets pre‑set safety and performance thresholds, with the temporary status made clear to patients and the path to full authorisation set out in advance.

After launch the commission calls for mandatory monitoring plans, real‑world studies, and regular performance reports; if performance deteriorates an escalation procedure must apply. Another recommendation proposes inserting the device identifier and its version into the patient’s health record to help clinicians trace outcomes. The report also links safety to how a clinic adopts the technology, requiring manufacturers to describe safe‑use conditions — including user training and clinic readiness — and for contracts to allocate responsibility for these measures, with clinics tasked to train staff on the specific technology.

🔗 Read original →

Study identifies two MRI subtypes of Alzheimer's with distinct progression patterns

On September 11, the preprint described two MRI‑based subtypes of Alzheimer’s disease that show different patterns of worsening. The authors combined baseline MRI scans from 1,396 participants across four North American research cohorts. Using the SuStaIn model (SuStaIn algorithm, 2018), they identified two statistical trajectories of change.

One trajectory begins with atrophy of memory‑related regions (parahippocampal and temporal areas) followed later by growth of white‑matter lesion volume. The other starts with an increase in visible white‑matter lesion volume, especially in the occipital lobe, then expands ventricular spaces and shrinks memory‑related regions. Similar‑looking changes on a scan can belong to different disease stages, and standard MRI grouping often mixes trajectory type with stage.

When the model was tested on ten data splits, two subtypes best explained the data. In both groups, amyloid and tau levels were abnormal. The subtype with early white‑matter lesions showed higher blood pressure and, in one cohort, a larger volume of enlarged perivascular spaces. The subtype with early atrophy had faster decline on memory, attention, language and global cognition tests.

Among the 1,145 participants with clinical follow‑up, a model adjusted for age and sex linked early atrophy to higher odds of progressing to an Alzheimer’s diagnosis compared with early lesions. Future studies could use these two trajectories for participant selection and prognosis, as they are associated with different rates of cognitive decline.

🔗 Read original →

Consistency metric selects cases for autonomous AI diagnosis

On 15 September Nature Medicine published a study of a locally deployed diagnostic AI agent. Researchers ran the same simulated clinical case five times and compared the meaning of diagnoses. Using a consistency threshold of 0.90, they allowed 272 cases of 551 cases for autonomous processing in the main configuration.

The simulations used MIRA‑v2, a set of de‑identified patient records from MIMIC‑IV. In each case the doctor‑agent interviews a patient‑agent, requests permitted tests, and issues a diagnosis with explanation, all on local hospital infrastructure. In June MIRA turned the “diagnose from description” task into a multi‑step virtual chart encounter where the agent asks questions, orders studies, and picks a plan.

For each of the 551 cases the team made five independent runs to capture variability. They defined ConsistencyDx as the agreement among the five diagnoses; it outperformed the model’s internal word probability (AUC 0.860 vs 0.747). When ConsistencyDx was at least 0.90, 272 cases went to autonomous flow, and 269 of those matched the benchmark label (98.9%); the remaining 279 cases were sent for physician review.

A stress test that removed the patient‑history prompt dropped accuracy from 90.6% to 70.2% and lowered ConsistencyDx, while internal probabilities and linguistic confidence features stayed high. Five runs required roughly five times more tokens than a single run. On the external VivaBench set, ConsistencyDx remained the best metric; at a threshold of 0.85, 32.0% of cases were kept for autonomous processing with an accuracy of 89.9%. The optimal threshold must be recalibrated for each deployment because it depends on model, generation settings, number of repeats, and the method used to compare diagnosis meaning.

🔗 Read original →

Lund University Starts Microelectrode Brain Stimulation Trial for Parkinson's

On **September 1

🔗 Read original →

Preprint links P5 DNA site to Rep/Cap fragments in liver

🔗 Read original →

Mitchell McLennan calls pre‑birth PCSK9 editing an inheritable choice

Mitchell McLennan made the comment on **

🔗 Read original →

Guide RNA length shapes base editor activity in human globin genes

A preprint released on 13 September shows that the length of the guide RNA determines which nearby DNA letters a base editor can modify in human cells. The study focused on the beta‑ and gamma‑globin genes, where changing the guide length altered both editing efficiency and the pattern of nucleotide changes.

In the beta‑globin region targeting the HbE mutation, the desired change occurs at nucleotide A9 while the adjacent A10 can generate the Hb Aubenas variant. With a 15-nt guide, A9 was edited most frequently and the ratio of A10 to A9 edits was about ~3%.

For the gamma‑globin target, the shortest guide length that yielded detectable editing depended on how the editor was delivered. In HUDEP‑2 cells that continuously produced the editor, a 15-nt guide gave measurable editing, whereas after transient delivery in HUDEP‑2 and primary progenitors editing began only with a 17-nt guide.

Because the globin genes are similar, Cas9 can cut both sites, a situation linked in the group’s 2024 work to large deletions between them. PCR assays for the deletion products did not detect the expected fragments when using guides of 15–16 nt, and a nickase version (dABE8e) gave higher editing frequencies than the standard ABE8e with a 16‑nt guide.

🔗 Read original →

ProtScape ranks Parkinson’s drug targets far ahead of previous model

ProtScape is a model that builds separate protein‑interaction maps for 207 cell types and states; it was released as a preprint on bioRxiv, September 14.

To test target discovery, the authors hid 18 hidden Parkinson’s drug targets during training; ProtScape placed all of them within the first 511 ranks, whereas the earlier PINNACLE model required 8,186 ranks to achieve the same coverage.

Unlike a generic interaction map that blends signals from many tissues, ProtScape constructs context‑specific networks using only proteins whose genes are active in each cell type, then predicts additional links by matching a protein’s amino‑acid sequence to its neighbors in that network.

In the validation step, some links were removed from the original maps and the model reconstructed them from the remaining neighbors and sequence; each protein pair was assigned entirely to either training or testing data across all cell contexts, ensuring novelty for the test set.

For 15 diseases, target rankings were trained using drugs with completed phase‑II human trials or stronger clinical evidence as positives; the ranking depth shows how many proteins must be examined to reach all 18 Parkinson’s hits: 511 for ProtScape versus 8,186 for PINNACLE (about a 16‑fold difference). After applying a preset threshold to unlabeled proteins, the authors obtained 102 candidates for follow‑up experiments.

🔗 Read original →

AI model using routine data improves lung cancer immunotherapy outcome prediction

On 13 September, Nature Medicine published work of the I3LUNG project analyzing 2,396 patients from six clinical centers. Authors checked whether an AI prediction based on routinely collected pretreatment data could outperform standard biomarkers and change the assessment of completed cases.

Immunotherapy helps the immune system attack tumors, yet in metastatic non‑small‑cell lung cancer it is hard to know who will achieve disease control and who will live longer. PD‑L1 tumor level is one guide but captures only part of the patient’s situation. The study examined how much prognostic information already exists in ordinary medical records.

For the main model the authors selected nine pretreatment features: sex, smoking status, ability to perform daily activities, PD‑L1, metastasis location, and blood‑test results. In an independent patient cohort the model surpassed single biomarkers and the LIPI blood‑based index in distinguishing disease control and several survival outcomes.

Adding CT scans and digital tumor‑slice images improved performance when tested on data from the same centers, but this advantage was inconsistent in external groups. The model relying only on clinical data and blood work proved more stable, and the most reliable prediction came from the pretreatment data set that physicians already have.

To test usability, twenty physicians reviewed 100 case histories, first with patient data alone and then with model output and explanations of which features shifted the prediction. After the model’s suggestion, correct identification of cases where disease was kept under control rose from 0.72 to 0.87, and overall accuracy increased from 0.57 to 0.65, although false‑positive control predictions rose slightly.

The next phase of I3LUNG is already underway, testing the system on more than 2,000 patients. The published paper describes its retrospective phase—analysis of accumulated clinical data and review of finished cases together with clinicians.

🔗 Read original →

OpenAI Foundation allocates over $125 million for health‑AI data projects

OpenAI Foundation — the nonprofit parent of OpenAI — launched its second scientific initiative, Public Data for Health, and awarded the first tranche of **over $125 million

🔗 Read original →

Step‑wise AI diagnostic route improves rare disease detection

A review of seven AI diagnostic systems assembled a six‑step route for rare‑disease identification. On a set of 50 test cases the route yielded the correct first diagnosis in 11 instances, whereas a single‑query GPT‑5 model got it right only 5 times. The findings were posted as a preprint on Research Square on 14 September.

Across 19 evaluations of 33 738 cases the average share of correct first diagnoses was 51.2 %, with individual systems ranging from 22.0 % to 77.5 %. The authors screened 1 193 publications, read 111 full texts, and kept seven studies that together supplied those 19 accuracy measurements.

The six‑phase route—extracting salient features, building and revising disease candidates, checking medical references, re‑ranking options, and recording the reasoning—was tested against a direct GPT‑5 query

🔗 Read original →

Selective Vagus Nerve Cuff Enables Targeted Stimulation Without Battery

In a pilot experiment a cuff with 14 pairs of electrodes divided the vagus nerve into sectors that control the heart and larynx. On 14 September a team from University College London demonstrated a temporary, battery‑less implant that receives power and pulse settings through NFC, like contact‑less phone payment. The device sequentially activated each cuff sector in four pigs and then, during a 30‑minute human operation, mapped

🔗 Read original →

Tuberous sclerosis shows Alzheimer’s-like tau marker levels due to chronic mTOR activation

Duke University researchers on 11 September compared plasma p‑tau217 in 63 people with tuberous sclerosis, 25 with Alzheimer’s disease, and 116 cognitively healthy controls. The marker was significantly higher than in healthy peers, began rising at a younger age, and was statistically indistinguishable from the Alzheimer’s group.

Tuberous sclerosis is a rare inherited disorder affecting 7–12 per 100,000 individuals, caused by loss‑of‑function mutations in TSC1 or TSC2, which normally restrain mTOR. Without this brake, mTOR remains hyperactive throughout life, leading to benign tumors in brain, kidneys, heart and skin, and epilepsy in most patients.

Earlier Duke work found neuronal damage markers in cerebrospinal fluid and post‑mortem tau aggregates resembling Alzheimer’s pathology but lacking amyloid plaques. The link lies in chronic mTOR activation blocking autophagy, damaging synapses and allowing tau to accumulate — processes mirroring Alzheimer’s neurodegeneration.

The study shifted to blood using p‑tau217, the most accurate early tau marker available. After adjusting for age and sex, p‑tau217 remained significantly higher than healthy (p < 0.0001 vs healthy; p = 0.071 vs AD) and kidney contribution was noted as modest. Independent age‑sex matching confirmed the same result.

mTOR is the same pathway targeted by rapamycin longevity hypotheses; rapamycin extended lifespan in worms, yeast, flies and, in 2009, in normal mice. Human longevity trials have been inconclusive, but tuberous sclerosis offers a natural model of lifelong mTOR hyperactivity producing Alzheimer’s‑level tau pathology. Authors plan to test whether rising p‑tau217 predicts memory decline. Clinically, Aeovian is testing the selective mTORC1 inhibitor AV078 for epilepsy in this disease; if mTOR drives tau pathology, AV078 could also treat early Alzheimer‑like tau changes.

🔗 Read original →

Cardiac spheroids boost early survival after heart attack in mice

Danish biologists from Professor Ditte Caroline Andersen’s lab at the University of Southern Denmark transplanted mice with induced heart attacks and weakened immunity with cardiomyocytes derived from stem cells — either as loose single cells or as dense spheres of 250 cells. In the first two days, 92% of mice receiving spheres survived, compared with 36% receiving single cells and 50% receiving no cells. The results were posted as a preprint on bioRxiv on 11 September.

After a heart attack the heart can lose up to a billion cardiomyocytes and cannot replace them; the adult heart lacks resident stem cells. Regenerative medicine produces such cells from induced pluripotent stem cells (iPSC) and transplants them — worldwide 12 clinical trials have been registered. Single cells wash out of the injection site and die without neighbor contact, while engineered patches persist longer but integrate poorly with cardiac tissue. Spheres represent a compromise: aggregates of several hundred cells delivered through the same thin needle, with preserved intercellular contacts that aid survival during injection.

To eliminate immune confounding, the authors used the new mouse line NXG B2m, which lacks B and T lymphocytes, natural killer cells, and the gene that normally displays the “self” marker for immune surveillance. The procedure itself was nearly lethal: of 49 mice, only 50% survived the first two days with the vehicle

🔗 Read original →

Japan’s Centenarian Surge and Fertility Drop Point to Transhumanist Solution

Another "record" in Japan. The number of centenarians over 100 years old has exceeded 107,000 people, while the total fertility rate has fallen to a catastrophic 1.14.

The population is rapidly heading toward 87 million by 2070, of whom nearly 40% will be elderly. Prime ministers call it a "silent existential crisis," and traditional family subsidy and birth‑stimulus programs are openly and predictably fraying at the seams.

The only remedy for the demographic crisis, they say, is transhumanism. First, transhumanism shifts the focus from purely palliative care to biomedical aging therapy. Japan’s problem is not longevity itself but the period of frailty that creates a massive burden on health‑care and pension systems. Introducing "healthy longevity" technologies and treatments that repair cellular damage could turn 80- and 90-year-olds from consumers of social resources into active participants in economic and social life. Second, the gap between biological limits and modern lifestyle is closed by augments and cybernetics. Exoskeletons, implants, and neural interfaces already let older people retain physical autonomy, and when these technologies aim at enhancement rather than treatment, functionally old and young people become indistinguishable. Large‑scale development of biomedical interfaces and robotics solves the labor‑shortage problem without destabilizing society through sudden migration spikes.

For policymakers, promoting a transhumanist vision is the only realistic technological survival strategy. Instead of futile attempts to make youth have more children,

🔗 Read original →

MoleculeMind Announces QuantaMind Neural Force Field Study of PETase

On September 14, MoleculeMind announced that QuantaMind had modeled the PETase enzyme reaction in a system of 17,792 atoms. The corresponding article appeared in Science Advances on September 11.

The simulation included the PETase protein, a PET substrate fragment, and 4,898 water molecules. Molecular dynamics moves atoms step by step according to interatomic forces, producing a trajectory where bonds can form or break and protons can transfer to neighboring atoms.

In the hybrid approach, the reacting site is treated quantum‑mechanically while the surroundings are described with a simplified force field, with the boundary chosen in advance. QuantaMind is a neural‑network force field: given atomic coordinates it predicts the energy and forces for the next step of the trajectory.

The training set comprised quantum‑chemical calculations and configurations near transition states — rare atomic arrangements where one bond breaks while another forms. PETase, a bacterial enzyme that degrades PET plastic, was selected because its structures are known but the proton transfers in its mechanism remained debated.

In the calculated trajectory

🔗 Read original →