iiinova — where your digital self keeps evolving

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

8/ Microscale to macroscale mechanism A biophysical model simulating the prolonged inhibitory postsynaptic currents of GABAergic anaesthetics at the microscale, reproduces 461/485 (95%) of macroscale empirical features 🤯 Shortening excitation (à la ketamine) also gets >90% ✅

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

7/ Do conserved dynamics have conserved molecular underpinnings? 🧬 We integrate gene expression from human, macaque & mouse, plus new marmoset genes 🐒🧬 We find a conserved anterior-posterior cortical gradient linking dynamics to genes for excitation & inhibition, including known anesthetic targets

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

6/ Spatio-temporal isolation 🌐 Anaesthesia desynchronises regions & decouples their dynamical profiles Neural activity fails to persist over time, but it also fails to propagate in space.

132

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

5/ We measured intrinsic neural timescales directly ⏱️ In every single species, timescales get shorter under anaesthesia, and regions with the longest timescales lose the most ⚖️ DBS of the centromedian thalamus reverses this: timescales lengthen as the macaque wakes up 🐒⚡

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

4/ Do any features change in the SAME direction in EVERY contrast & drug & species? 🎯 485 survive (vs ~1 expected by chance!) We consistently find ⬇️ autocorrelation, ⬆️ forecasting error 🔮. Under anaesthesia, information does not persist in neural activity

130

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

3/ We compare every feature in every region, awake vs anaesthesia Across 7 anaesthetics, 9 orders of magnitude in scale & ~700M years of evolution 🪱🐟🐁🐒🧑, many features shift the same way in every species. And when centromedian thalamic DBS re-awakens the macaque, they flip back⚡

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 02/10/2026

1/ 🚨 Just out in @natneuro.nature.com! 🧠 We profiled >6,000 features of brain dynamics across humans, macaques, marmosets, mice, fish and worms 🧑🐒🐒🐁🐟🪱 to find a shared neural endpoint of anaesthesia: spatiotemporal isolation of local activity 🧵👇 🔗 doi.org/10.1038/s41593-026-02460-4

23922

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

10/ Working Model of Thalamic Restoration Species-specific biophysical modelling of macaque 🧠 successfully recapitulates restoration of integration (ΦR) by central but not ventral thalamus DBS: connectivity profile is key! Next: model-based virtual clinical trial in disorder of consciousness?

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

9/ Inhibition Controls Controllability Species-specific models with heterogeneous inhibition also show that PVALB is anatomically well-suited to tune 🧠 controllability Consciousness may rely on conserved interplay between the brain's wiring diagram and its molecular tuning🧬

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

8/ Biophysical Modelling To go beyond correlation, we built computational models 💻🔬 integrating species-specific anatomical wiring with 🧬 gradients for humans, macaques, & mice Only models with PVALB-driven inhibition successfully recapitulate breakdown of ΦR/integration

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

7/ PVALB Inhibition? We mapped the regional breakdown of ΦR against species-specific gene expression 🧬 in 3 species 👱🐒🐭 Spatial topography of ΦR breakdown coincides with PVALB gene expression (parvalbumin interneurons). This suggests a role for PV-mediated inhibition in governing integration

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

6/ Out of Control As the system disintegrates, the brain's dynamics become significantly harder to "control" in a network-theoretic sense 🌀 Loss of controllability scales with breakdown of integration (& is restored by CT-DBS)

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

5/ Tracking Behaviour Dominance Analysis shows that ΦR outperforms traditional Φ and other proposed information-theoretic measures of "integration" 📊⚖️ for tracking DBS-induced changes in behavioural arousal

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

4/ Reversing the Collapse via Thalamus⚡️🧠 When macaque🐒 is re-awakened by deep-brain stimulation of central thalamus (CT), we also re-ignite ΦR & the brain’s integrative capacity 🔥 Not so for DBS of a control site! CT is a local switch for the global informational architecture

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

3/ Convergent Breakdown The results were striking: across 4 mammalian species 👱🐒🐵🐭 and 6 molecularly diverse drugs, anaesthetic-induced disconnection from environment is marked by a convergent collapse of ΦR 📉 (but *not* traditional Φ or related measures) - rising again upon recovery 📈

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

2/ From Φ to ΦR 🔥 How do you measure "integration"? Information decomposition reveals that influential measures of "whole-minus-sum-of-parts" double-count the redundancy in the parts. Our revised Φ (ΦR) is a principled solution📊⚖️

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 19/03/2026

1/ The Big Picture We combined fMRI, information theory, & species-specific computational modelling in human, macaque, marmoset and mouse 👱🐒🐵🐭 to ask: does anaesthesia reduce the 🧠's capacity to be "more than the sum of its parts"?

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

8/ Computational consequences Competitive interactions improve computational capacity, when the model’s generative connectivity is used as the wiring diagram for a connectome-based reservoir network 🤖 Could competition lie behind the efficiency of mammalian👱🐒🐭 brains? 🤔

101

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

7/ Not just "more parameters"... Allowing negative connections provides greater performance gains than doubling (!) the number of model parameters. Want more? Check out our Supplementary for extensive validations & more sanity checks

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

6/ Consistent network topology of competition Competitive connections are not randomly distributed. In each species🐒🐭, negative connections tend to be longer-range than positive ones, and less modular & locally clustered

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

5/ Biological annotations Across species 🐭🐒, competitive interactions are grounded in 🧠 biology They link regions at opposite ends of the cortical hierarchy, with opposite molecular 🧬 and cytoarchitectonic profiles. Competition may help segregate functionally diverse systems

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

4/ Dynamical consequences It is not just FC that improves, either. After all, that was the model objective. But by allowing competitive interactions, we also get more realistic 🧠 dynamics : synergy, local-global broadcasting, metastability, and irreversibility. For free!

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

3/ From distant cousins to digital twins ♊️ Each 🧠 has a unique fingerprint. But traditional cooperative-only models miss its essence: your model and a random stranger's are easily confused. Hardly "personalised"... Allow competition, and it becomes clear which model was yours!

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

2/ Improved fit Competitive interactions dramatically improve model fit to human, 🐒 & 🐭 FC. The model does not *have to* include negatives – but it chooses to! In humans, competition improves model "cognitive matching" to >100 meta-analytic patterns from NeuroSynth

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 13/03/2026

1/ Introducing competition Most whole-brain models assume that regions cooperate (as A goes up, so does B) Our generative model also allows competitive (negative-sign) interactions. As A gets more active, it can suppress B We test if this improves model fit in human, 🐒 & 🐭

100

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

7/ Mapping disorder involvement ⚕️ We derive regional risk maps for 30+ brain disorders 🧠🩹. Their clustering aligns with official ICD-11 clinical classification 🩺 better than clustering maps of risk-gene expression 🧬, & recover symptom co-morbidities

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

6/ From shared function to shared co-activation Regions’ LLM-derived cognitive similarity predicts functional co-activation from fMRI, and effects of direct stimulation ⚡️ better than anatomical connectivity, molecular profile, or spatial proximity - also in 🐭&🐒

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

5/ Conserved molecular circuits for cognition 🧬🧠 We can also extend to macaque🐒 & mouse🐭 previous human studies based on NeuroSynth (Hansen 2021 Nat Hum Behav; Luppi 2024 Nat Biomed Eng). Integration w/ species-specific gene expression 🧬 reveals a cross-species molecular circuit for cognition

120

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

4/ A functional cortical hierarchy is latent in the literature Asking the LLMs about regions and functions one-by-one spontaneously reveals a unimodal-transmodal functional hierarchy 🧠 matching species-specific anatomical hierarchy

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

3/ Unlocking Cross-Species Discovery 🐒🐭 We can use the LLM-generated maps to decode brain maps from other species. We show this with 🐒 macaque working memory & the recent Neuropixel maps of 🐭 mouse neural activity during cognitive tasks, from International Brain Laboratory (doi.org/10.1038/s415...)

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

2/ Beyond traditional meta-analysis LLM maps go beyond traditional meta-analysis tools: they correlate better with 🧠 circuits derived from direct causal interventions from intracranial electrical stimulation ⚡️ and Lesion Network Mapping

110

Andrea Luppi, PhD @loopyluppi.bsky.social · 08/12/2025

1/ AI-driven cognitive mapping We use Large Language Models (GPT-4, Gemini, Llama, Mistral, DeepSeek) as neuroscience experts, to synthesise the evidence that region X is involved in function/disorder Y. We can do this for humans but also macaque 🐒 & mouse 🐭

110