Daily Summary for 2026-11-05
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the fifth of November, twenty twenty-six, and this is the day's research.
Marcus: 10 new papers came out today.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We'll take the day in one pass, then pull out the papers we're staying with.
Marcus: Start us off with CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness.
The day's research: Ines: Today we move beyond narrow AI by building systems with adaptive intelligence which is crucial because current AI often lacks human-like generality.
Marcus: CTM-AI proposes a blueprint for a multi-agent system inspired by the Conscious Turing Machine. It uses specialized processors and general learners to solve tasks.
Yuki: That approach allows AI to evolve its internal connections through feedback mechanisms, contrasting with studies on molecular notation shortcuts in evaluation.
Ines: Those studies showed input representation influences model behavior regardless of high predictive scores in those evaluations.
Marcus: In biological modeling, differential dynamic causal nets construct modular networks directly from EEG data during epileptic seizures.
Yuki: That contrasts with analyses of static and dynamic functional brain networks focusing on Alzheimer's disease changes.
Ines: Those analyses highlighted unique temporal variability in the amygdala and hippocampus compared to healthy controls.
Marcus: Deeper dives suggest a shift from synapse-centric to neuron-centric models for better neuroplasticity capture.
Yuki: A neuron-centric model gives each neuron its own internal memory for processing signals, unlike reaction-diffusion models for tumor growth.
Ines: That contrast is important when looking at phenotypic heterogeneity driven by oxygen gradients in brain tissue.
Marcus: The most significant development is MIST addressing a core limitation in computational pathology with molecularly informed staining transform.
Yuki: MIST introduces virtual molecular stains using paired spatial transcriptomics data during training to create prototypes.
Ines: These prototypes anchor gene expression profiles into the foundation model's feature space for adaptation without needing transcriptomics at inference.
Marcus: MIST clusters gene expression profiles into prototypes, then uses them to generate low-rank residual transforms reorganizing H&E patch features along molecularly guided axes.
Yuki: This means MIST injects spatial transcriptomic priors during training, allowing molecularly informed adaptation without needing transcriptomics at inference.
Ines: MIST improved 240 out of 256 configurations across various tasks, showing improvements like five point two percent in survival prediction.
Marcus: It also showed three point three percent improvement in tissue subtyping, primarily driven by gene-derived prototypes capturing spatially coherent molecular programs.
Yuki: MIST operates at the projection layer to introduce molecular axes the encoder never observed, separating patches based on distinct molecular signatures beyond morphology.
Ines: The immediate next steps involve controlled ablation studies to pinpoint components contributing most to performance gains, bounded by benchmark data availability.
Marcus: This suggests a need for sustained community investment in creating new problems suitable for benchmarking MIST's capabilities.
Lucky paper: 2605.04097: Ines: So we're talking about CTM-AI today, and it’s this early blueprint for a general AI system by combining the Conscious Turing Machine with foundation models.
Marcus: It shows state of the art accuracy on MUStARD with a score of seventy-two point two eight and UR-FUNNY with seventy-two point one three, outperforming multimodal frameworks.
Yuki: The paper operationalizes this architecture using multiple specialized processors and general learners that competitively select information to solve complex problems iteratively.
Ines: That iterative loop is the core inference process, involving four stages: chunk inference, up-tree competition into the workspace, down-tree broadcast, and link formation.
Marcus: Step one involves all K LTM processors running in parallel to produce chunks with a gist h it and a follow up query q it for potential cross processor consultation.
Yuki: That chunk score s i t is decomposed into weight, intensity, and mood, which is then used in step two to select the final chunk that goes into the limited capacity workspace.
Ines: Step three is the down-tree broadcast where the winning chunk is distributed to all LTM processors causing them to pay conscious attention.
Marcus: Step four enables unconscious communication where LTM processors form links and exchange information, leading to link fusion which enriches long term memories.
Yuki: The paper also shows that the iterative loop emerges as the most critical mechanism with a negative F1 score of minus six point seven.
Ines: That means even just scaling the iterations from one to four yields continuous performance gains, like MUStARD improving from sixty-five point zero to seventy-three point eight.
Marcus: When we look at tool use on StableToolBench, CTM-AI yielded substantial gains, improving over the strongest single model baseline by up to sixteen point eight points on multi tool scenarios.
Yuki: The paper suggests that reasoning models fall short of CTM-AI when the backbone model is fixed, and pairing Qwen3 eight billion think with it yields an additional plus sixteen point eight to plus twenty-seven point three point improvement on StableToolBench.
Ines: That confirms that reasoning deepens deliberation within individual processors while CTM-AI broadens coordination across them.
Marcus: Regarding implementation, the backbone model used for most processors is Gemini two point five flash lite because it supports audio, vision, and text inputs while being relatively small scale.
Yuki: The system uses several models including GPT four o open ai twenty twenty four and Gemini two point five flash lite from Comanici et al. twenty twenty five.
Ines: For multimodal perception tasks like MUStARD and URFUNNY, CTM-AI is instantiated with three modality specific experts for each.
Marcus: For StableToolBench, each available tool acts as an independent LTM processor powered by a lightweight LLM constrained to utilize only its assigned api.
Yuki: The WebArenaLite processors include an html processor and a screenshot processor that perceive the current page state through distinct modalities.
Ines: They define operations like execute, read, and write for a memory state M t updated over time within the LTM Processor implementation.
Marcus: The chunk inference function Ctmcollect generates chunks represented as a tuple containing a gist h it, an additional internal question q it, and a composite score s i t.
Yuki: Link formation is triggered by the down-tree broadcast in step three, relying solely on the weight sub score where a weight greater than zero point eight establishes a directed link.
Ines: The evaluation parsing prompts are tailored for each task to ensure accurate final generation, like the specific prompt for UR-FUNNY which requires identifying specific humor techniques with confidence greater than or equal to zero point six.
Marcus: One case study showed that the failure in URFunny was attributed to a vision only misleading effect where the video processor failed to provide necessary facial expression information.
Yuki: The success in StableToolBench came from a processor failing to invoke its designated api call for qr code generation, which is what the paper flagged as a failure case.
Ines: The system's efficiency analysis indicates that most tasks resolve within one to three iterations, and the performance boost from additional iterations is proportional to an increase in the average number of inter processor links.
Marcus: The optimal balance for the link formation threshold eta was found between zero point four and zero point seven for both datasets.
Yuki: In conclusion, CTM-AI offers a robust foundation for cognitively inspired AI by translating consciousness theory into a practical, effective system to enhance decision making and tool use.
Lucky paper: 2601.21478: Ines: Let's get into this one, we have Differential Dynamic Causal Nets: Model Construction, Identification and Group Comparisons.
Marcus: This paper tackles how to build differential causal networks directly from electroencephalogram data by modeling interacting neuronal populations as conditionally coupled dynamic causal models.
Yuki: So each node in the network represents a parameterized local neural system, and the directed edges are node-wise connections with transmission parameters, which are varying across subjects following a mixed-effects model.
Ines: That sounds complex because you're dealing with many parameters simultaneously.
Marcus: They use an evolutionary optimization algorithm for parameter inference that minimizes a loss function derived from Chen-Fliess expansions of stochastic differential equations.
Yuki: That loss function is constructed by the summation of the squared prediction errors over the set of increasing time instances Tq, which is how they estimate the parameters.
Ines: So they're using these neural mass differential equations to drive this optimization process.
Marcus: Exactly, and for identification between cases and controls, they use post-multivariate analysis of variance based on a package of permutational multivariate analysis of variance.
Yuki: The significance of the case-control factor implies that case-control memberships contribute to the group differences while the significance of the channel factor shows that group difference should be accounted for by multiple channels.
Ines: That suggests they aren't just looking at one connection, but a whole pattern across different parts of the brain.
Marcus: The analysis reveals various differential dynamic causal nets associated with different parameters, like A sub i j k over B sub i l j k and v sub i j k zero n q.
Yuki: They found that for the parameter A sub i j k over B sub i l j k, there were six channel pairs to be claimed to differ cases from controls significantly.
Ines: And what was the most pronounced difference they saw in terms of location?
Marcus: The most pronounced differences between the case and control groups were in channel pairs with output channels in the frontal regions.
Yuki: They also identified an altered excitation inhibition balance, often involving decreased inhibition, as a robust biomarker in various neurodegenerative diseases like Alzheimer’s disease and frontotemporal dementia.
Ines: So this NccDCM tool is being proposed for detecting disease-predisposed effective connectivity using EEG data.
Marcus: They also noted that the existing imaging studies often focused on single subject studies or models for homogeneous brain regions, which they are relaxing by proposing a heterogeneous composite network model, the NccDCM.
Yuki: In this NccDCM, each node represents a local Jansen-Rit neural mass model for describing synaptic interactions in excitatory inhibition and pyramidal neuron populations.
Ines: That heterogeneity should help capture subject variability better than models where all nodes are identical.
Marcus: To account for individual subject variability, they introduced a mixed-effects model for the network parameters, which is a key part of their methodology.
Yuki: The estimation process involves solving the Jansen-Rit neural mass differential equations first, then constructing the forward pairwise loss functions, and finally minimizing those loss functions by some evolutionary optimization algorithms.
Ines: That whole sequence from solving equations to minimizing loss functions sounds like a lot of computational heavy lifting.
Marcus: It is heavy, but it allows for the inference of a large number of node parameters even when nodes are heterogeneous.
Yuki: The gains come from post-analysis of variance on estimated parameters in cases and controls, specifically showing significant differences in frontal region channel pairs for the excitation inhibition ratio A sub i j k over B sub i l j k.
Ines: So the numbers point toward frontal regions being a key area for these differential patterns.
Marcus: They also found that altered excitation inhibition balance involving decreased inhibition is a robust biomarker, and they demonstrated bifurcation behavior in simulation studies when parameters like A over B, C, K, and v zero take certain ranges.
Yuki: The simulation study on three-nodes NccDCMs showed good performance when nodes are dynamically coupled each other.
Ines: What about the limitations you mentioned earlier regarding the method's application?
Marcus: The present model has limitations including the algorithms for estimating the NccDCM and omitting subjects’ demographic information like age and gender.
Yuki: They use a marginal screening approach to estimate channel-wise connection parameters to avoid overfitting, but that comes at the cost of estimation efficiency.
Ines: So while it detects disease-predisposed connectivity, it's not perfectly efficient for every subject profile yet.
Marcus: Further work is needed to take age and gender as additional predictors in each ccDCM, and they can extend this to source-level EEG data in the deep brain.
Yuki: The paper is published on arXiv as 2601 point 21478v1, and the research of Kang You and Jian Zhang is supported by an EPSRC grant EP/Xthirty-eight thousand two hundred ninety-seven/one.
Lucky paper: 2505.03458: Ines: Today we're looking at Static and Dynamic Functional Brain Networks Reveal Distinct Connectivity Alterations Across the Alzheimer's Disease Spectrum, a paper by Rubido et al.
Marcus: It looks like they analyzed age- and sex-matched data from three hundred fifteen individuals with AD, mild cognitive impairment, and healthy controls.
Yuki: While healthy controls and mild cognitive impairment show similar patterns at the node level, clearer differences emerge specifically in Alzheimer's disease patients.
Ines: They identified stable stationary differences in functional connectivity between white matter regions and parietal and somatosensory cortices.
Marcus: But they also found that temporally varying differences were consistently observed in connections involving the amygdala and hippocampal formation.
Yuki: The study highlights shared and unique functional connectivity patterns across both static and dynamic functional networks, emphasizing the importance of incorporating dynamic information into brain network analyses of the Alzheimer's spectrum.
Ines: They constructed these networks using a Juelich brain atlas, resulting in one hundred twenty-one regions for the nodes.
Marcus: The signal for each node is calculated by averaging the resting-state fMRI bold time series across all voxels within that corresponding region.
Yuki: Functional links were defined by the Pearson correlation coefficient calculated pairwise between all those nodes.
Ines: To construct static functional network, they used the formula where you calculate correlation coefficient applied to the full time series of one hundred ninety-seven points.
Marcus: For dynamic functional networks, they used a half-overlapping sliding window approach applied to the entire time series of each node with a step size of delta t over half.
Yuki: They characterized node importance using node strength, which quantifies the strength of functional connections rather than just their count.
Ines: Eigenvector centrality is another measure where you assume connections to more influential nodes are more important than connections to less relevant ones.
Marcus: They used a random forest classifier trained on regional bold time series informed by both dynamic and static network metrics to distinguish between the groups.
Yuki: The highest accuracy, around seventy-nine percent, was obtained when classifying healthy controls versus Alzheimer's disease using the first set of nodes, which are the amygdala and hippocampal formation.
Ines: They also looked at link level analysis where they quantified how often each connection shows significant differences across sliding windows in between groups.
Marcus: The most consistent differences were driven by connections between white matter regions and somatosensory regions, specifically the somatosensory cortex.
Yuki: They found that while Alzheimer's versus healthy controls show similar static patterns, the dynamic functional network deviations from healthy controls show less consistent changes compared to static networks.
Ines: The paper provides evidence that differences in eigenvector centrality were driven by the amygdala and hippocampal formation connectivity, which is an early biomarker of the interplay between early Alzheimer’s disease pathology and cognitive decline.
Lucky paper: 2608.16951: Ines: So we're looking at The Little Scientist today, which is about an LLM agent following the scientific method to discover new algorithms using hypothesis-driven search and a Kuhn agent for paradigm shifts.
Marcus: It consumed seven hundred four million tokens on a single virtual machine with no GPUs, so that’s a serious computational cost we need to keep in mind.
Yuki: The framework is built around an LLM agent behaving like a scientist, working inside an evaluation environment that benchmarks its code against real data and feeds structured diagnostics back.
Ines: That three component interaction loop—the inner search loop, the outer loop with paradigm-shift triggers—is what makes the architecture unique.
Marcus: The hypothesis-driven search trades search breadth for depth because failed hypotheses constrain the design space for subsequent proposals, which is different from evolutionary approaches being parallelizable.
Yuki: But it’s sequential because each iteration builds on the diagnostic feedback of the previous one, and that structured per-instance diagnostics surface patterns that let the scientist reason about causal structures.
Ines: Take this example where motif discovery scored zero point six five AUROC because it found an 8bp motif when the reference was 19bp, which points directly at a design lever like width selection.
Marcus: That shift from just presenting data to generating insight is key, but I wonder how robust that structured feedback is when the problem moves away from sequence motifs.
Yuki: The framework demonstrated a capability to produce genuinely novel algorithms, like Delta V, which is an ensemble calibration strategy requiring no inference-time LLM calls.
Ines: Delta V ranks first on the ProteinGym DMS Substitutions Zero-Shot leaderboard across all five official evaluation metrics, and it’s a static strategy that doesn't need inference calls.
Marcus: The infrastructure evolution part is interesting because they transformed a stalled loop by replacing synthetic benchmark data with real ChIP-seq TFs, which contrasts with prior agent systems.
Yuki: That contrast in evaluation cost is striking; their evaluations take seconds on a single core compared to AlphaEvolve’s hundred compute hour evaluations.
Ines: The implication for agentic science is that the research productivity metric should measure the acceptance rate, which they found was twenty-six point six percent, not just end-state accuracy.
Marcus: But I have to ask, what is the structural limitation here? The paper says it requires a well-defined benchmark with clear metrics and ground truth evaluation data.
Yuki: That’s a real constraint because the results are computational benchmarks only, and wet-lab validation of strategies like Delta V would substantially strengthen that part of the work.
Ines: So, controlled ablation studies are recommended as future work to quantify the contribution of individual framework components, which is a necessary step since benchmark availability is currently bounded.
Marcus: The applicability is clearly limited by benchmark availability, which means sustained community investment in creating new problems as benchmarks or oracles will be needed for this method to scale.
Lucky paper: 2605.16392: Ines: So we're looking at MIST, the paper is titled Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation.
Marcus: It’s a plug-in replacement for the MIL projection layer that uses paired spatial transcriptomics only during training to build virtual molecular stains.
Yuki: The authors call it addressing a fundamental mismatch where frozen pathology foundation models are pretrained mainly from H andE morphology, but clinical outcomes are governed by molecular state.
Ines: They identify the projection layer as this "largely morphology-only bottleneck" that limits endpoints governed by molecular state.
Marcus: MIST converts paired spatial transcriptomics into these virtual molecular stains, which then reorganize H andE patch features along gene-defined axes.
Yuki: How does this work in practice? The paper says they derive these prototypes by clustering gene expression profiles and anchoring them in the pathology foundation model feature space.
Ines: So you cluster the gene expressions to get prototypes, and those prototypes live inside the existing model’s feature space so they can condition the H andE embeddings directly.
Marcus: Each patch gets associated with multiple prototypes through affinities, which reflects that local tissue regions often express overlapping molecular programs.
Yuki: And then prototype-conditioned low-rank residual transforms generate one stained view per molecular concept, which are combined according to the affinity weights and added back to the original patch embedding.
Ines: The empirical evaluation showed MIST improved two hundred forty out of two hundred fifty-six configurations over the standard projection layer.
Marcus: That improvement translates to an average gain of plus3 point 5 percent across twenty-three downstream tasks and eight MIL aggregators.
Yuki: Specifically, the paper shows improvements like plus5 point two percent in survival prediction and plus3 point three percent in tissue subtyping.
Ines: The ablation studies confirm that gene-derived prototypes are the primary source of those gains, while replacing them with random or morphology-derived alternatives removes most of the improvement.
Marcus: That means MIST is actually injecting molecular priors during training without requiring transcriptomics at inference, which is a big deal for deployment.
Yuki: But they also note a limitation on the scale of their data, because MIST relies on paired spatial transcriptomics data which remains limited in scale and tissue diversity.
Ines: That’s the trade-off: molecularly informed adaptation without needing transcriptomics at inference, but you still need that paired training data.
Marcus: The paper also mentions computational efficiency, stating that with K equals eight prototypes and rank r equals sixteen the MIST module contains only zero point eight seven million parameters.
Yuki: That parameter count is quite low for a complex mechanism, and they show it reduces the parameter count of the ILRA by nine point zero percent through shared feature compression.
Ines: So MIST is designed to be a reusable adaptation module that doesn't require modifying the frozen encoder, which saves you from retraining the whole foundation model.
Marcus: The paper also shows that MIST can increase separation between morphologically similar but molecularly distinct patches beyond what morphology alone permits.
Yuki: They decompose the output difference into a term capturing morphology-only transformation and another term capturing molecular routing induced by differences in prototype affinity.
Ines: That second term is key because it shows the model is routing based on those learned molecular concepts, not just shape.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models