Daily Summary for 2026-11-04

daily

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: It's the fourth of November, twenty twenty-six, and this is the day's research.

Marcus: 7 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 Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement.

The day's research: Ines: Today's focus is on tracking antimicrobial resistance determinants because genome and susceptibility testing link is crucial for clinical relevance.

Marcus: AMRScan 2 rewrites the toolkit to record exact origin of every data point, including software versions and database file hashes.

Yuki: This system archives susceptibility data from NCBI Pathogen Detection and builds tables offline for byte-for-byte comparison.

Ines: This detailed tracking keeps calls and phenotypes as separate layers joined only by the assembly accession.

Marcus: That is a key distinction from other methods like AMRScan 2: provenance-recorded detection of antimicrobial resistance determinants in bacterial assemblies, with archived susceptibility data from NCBI Pathogen Detection Studies that track bacterial resistance genes and their testing results back to their original sources.

Yuki: Furthermore, this workflow runs NCBI AMRFinderPlus on genome assemblies and uses a caller wrapper that processed ten thousand five hundred twenty-one public Escherichia coli assemblies without failure.

Ines: Another area explores how complex systems manage memory and navigation in the work on Neuromorphic Navigation in Static and Dynamic Environments with Dual-Timescale Neuron-Astrocyte Memory.

Marcus: This study introduces a spiking neuron-astrocyte network that uses spike timing to reinforce successful actions while astrocytic calcium transients suppress moves into recently visited states.

Yuki: This dual-timescale memory drastically improves goal completion in grid worlds, achieving near perfect success and shortening routes significantly compared to classical strategies.

Ines: In artificial intelligence, researchers look at how large language models handle memory effects as detailed in Analyzing Memory Effects in Large Language Models through the lens of Cognitive Psychology.

Marcus: This study found that LLMs mirror human associative interference patterns, such as the fan effect and list sensitivity, but they differ because they are completely position-invariant regarding list order.

Yuki: Finally, there is work exploring how brain architecture informs large models via The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models.

Ines: This proposal suggests that omni models converge on brain-like representations and that human cortical networks can serve as a usable architectural prior to improve model accuracy across various benchmarks.

Marcus: AMRScan 2: provenance-recorded detection of antimicrobial resistance determinants in bacterial assemblies, with archived susceptibility data from NCBI Pathogen Detection Studies that track bacterial resistance genes and their testing results back to their original sources.

Yuki: Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement This work introduces a framework that uses supervised learning to predict how cells would behave if their surrounding spatial environment were changed.

Ines: Neuromorphic Navigation in Static and Dynamic Environments with Dual-Timescale Neuron-Astrocyte Memory This paper describes a neuromorphic system that uses two different timescales of memory, one for long-term action sequences and one for short-term suppression, to help an agent navigate environments efficiently.

Marcus: Behavioral patterns and mean-field games in epidemiological models This research develops new epidemiological models that incorporate different behavioral patterns of subpopulations to better understand how diseases spread.

Yuki: The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models This paper proposes that the structure of the human brain can serve as a blueprint for creating powerful, multimodal artificial intelligence models.

Ines: Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability This work presents a minimal neural network model that uses neuromodulatory gating to overcome limits on how much information it can remember.

Marcus: Analyzing Memory Effects in Large Language Models through the lens of Cognitive Psychology This study compares how large language models handle various types of human memory effects, finding similarities but also key differences in their processing.

Lucky paper: 2606.08493: Ines: So we're looking at Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement today, which asks how a cell’s expression would change under altered spatial neighbor contexts.

Marcus: Cellina is the framework they introduce, and it uses supervised disentanglement to separate a cell’s intrinsic state from its spatial context.

Yuki: They formalize tissue graph counterfactuals as spatial interventions that either rewire connections between cells or modify the expression of their neighbors.

Ines: Cellina is a variational autoencoder that separates gene expression into two latent components, one encoding cell identity and the other encoding microenvironment effects.

Marcus: The authors use supervised disentanglement to anchor the cell identity representation to its type and adversarially remove spatial domain information from that representation, routing environmental variation to the other component.

Yuki: They show Cellina outperforms spatially-informed and non-spatial competitors across benchmarks spanning over two point five million spatially-resolved cells in colorectal cancer and mouse brain data.

Ines: Cellina's best variant leads the strongest baseline on colorectal cancer data by a point zero one seven on both Pearson r and Signed Precision.

Marcus: They also reveal biologically distinct cancer subdomains in an unsupervised manner by clustering the latent spatial representation over CRC cells, which they then score against PROGENy pathway gene sets to assign biological identities.

Yuki: The model is fast to train because it uses an efficient pseudobulk based phi of v computed a priori, which omits the expensive message passing from training over graphs.

Ines: They report that in convergence analyses, all four metrics improve with k for both logfold change variants and saturate by k approximately one hundred to two hundred.

Marcus: At k equal to two hundred in-distribution, the cell-type specific variant consistently outperforms the global one, reaching a Pearson r of point zero eight nine and Signed Precision of point zero five six.

Yuki: They also show that counterfactual predictions are better when conditioned on neighborhoods sampled from the matched subdomain rather than from the global CRC domain.

Ines: Cellina's supervised disentanglement objectives rely on cell-type and spatial-domain annotations, which they state are simplifications of continuous biology.

Marcus: The paper does say that future methods that integrate raw imaging signal alongside segmented counts could denoise both the expression signal and the counterfactuals derived from it.

Yuki: This latent decomposition is compatible with that interpretation because it doesn't yet impose the structural assumptions required for identifiability.

Lucky paper: 2609.10947: Ines: Okay, we're looking at The Platonic brain bridge hypothesis, which suggests human brain networks are a universal architectural prior for multimodal large language models.

Marcus: This paper proposes that the structure and organization of the human brain can guide model design because it shows omni models, those that handle video, audio, and text together, converge onto representations mirroring cortical function.

Yuki: The correspondence they show is stable across different participants and actually increases as more input channels like video or audio are added to the model architecture.

Ines: But there's a limitation here; the paper finds this correspondence localizes to sparse features, meaning removing those specific features weakens the model's ability to predict brain responses.

Marcus: That leads into their Brain-MoE mechanism, which fixes the expert partition of a frozen base model to align with the seven functional networks of the human cortex.

Yuki: When they implemented that structural alignment via Brain-MoE, held-out accuracy on all fifteen model benchmark pairs improved by an average of six point four two percentage points.

Ines: That improvement is significant, especially since in fourteen out of those fifteen cases, the resulting structure surpassed capacity matched random experts.

Marcus: The brain-to-model correspondence part uses Brain-Scope, which is a set of sparse autoencoders fitted to fMRI data to compute features for brain encoding across all three modalities.

Yuki: This mechanism allows the functional network organization, which they call the macroscopic structure of the human brain, to become a usable architectural prior built into the model.

Ines: They also decomposed performance into content pathway and routing pathway effects, and they found those two pathways sum exactly to the total effect.

Marcus: So it suggests that this decomposition provides a clear structural breakdown of the knowledge learned by the model, which is pretty useful for debugging what's happening inside an omni model.

Yuki: The main finding they validate is that "the human brain organization is therefore a usable architectural prior for multimodal large language models."

Ines: It moves beyond just observing correspondence to suggesting that this functional network organization can be engineered as an architectural prior within the model itself.

Marcus: That engineering is done through Brain-MoE, which injects the organization of those seven functional networks as a structural preference into the model's gating mechanism.

Yuki: The specific finding about pathway decomposition really shows how they map learned knowledge onto known brain structures, rather than just learning patterns from data.

Lucky paper: 2512.13859: Ines: So we're looking at Neuromodulation-inspired gated associative memory networks with extended memory retrieval and emergent multistability.

Marcus: This paper introduces a minimal, biophysically motivated network where neuropeptide-like signals are modeled by a self-adaptive, activity dependent gating mechanism that fundamentally reorganizes the attractor structure to bypass classical capacity limits.

Yuki: The gated network can bypass the sharp loss of retrieval at the classical critical capacity alpha sub c.

Ines: That means instead of just a transient retrieval, the multiplicative gating makes that transient retrieval permanent when the gates act as fully open or fully closed switches.

Marcus: The dynamics show that introducing this self adaptive gating fundamentally alters this behavior by extending the high overlap plateau significantly as the steepness gamma increases.

Yuki: It also leads to a novel memory retrieval phase of continuous multistability, where the system supports a continuous manifold of stable fixed points.

Ines: When you look at the binary gating limit with gamma equal to infinity, the system reaches a continuum of distinct fixed points depending on the specific initial overlap.

Marcus: This suggests that in these gated networks, the steady state overlap can be continuously controlled by choosing a different initial neural condition rather than being quantized to a stored pattern.

Yuki: The dynamical mean field theory confirms that these observations persist in the thermodynamic limit and in the long time limit.

Ines: They also found a characteristic spectral signature of frozen stabilization when analyzing the instantaneous Jacobian, which is reproduced during retrieval.

Marcus: This stabilization involves a subpopulation ceasing to evolve and supplying a static background field to the rest, which can stabilize active units away from the high gain region.

Yuki: The study demonstrates that hard gate retrieval remains accessible even when modulatory variables evolve substantially more slowly than the neural variables, depending strongly on the initial gate configuration.

Ines: Furthermore, they showed that gated retrieval is largely robust to rank and density of the modulatory coupling W in the hard gate limit, requiring W to be neither dense nor full rank for finite gate steepness.

Marcus: The emergent multistability suggests that the retrieved state is set continuously by the initial overlap, a feature that can be exploited for input driven computation with persistent activity.

Yuki: The utility of gating may extend from biological systems to machine learning more broadly, such as in gated attention in large language models.

Ines: The paper suggests that neuromodulatory gating may provide richer computational capabilities in biological and neuromorphic architectures.

Lucky paper: 2509.17138: Ines: Okay, so we're looking at Analyzing Memory Effects in Large Language Models through the lens of Cognitive Psychology today.

Marcus: This paper looks at seven specific memory phenomena, using human research as a benchmark for these large language models.

Yuki: The study is trying to see how LLM performance lines up with documented human cognitive behaviors across different conditions.

Ines: They are looking at things like list length effect, list strength effect, and the fan effect.

Marcus: They found that LLMs show human-like patterns of associative interference through the fan effect, where accuracy drops as the number of associations increases.

Yuki: But they also compared how LLMs handle false memories, looking at something called DRM-style false memories.

Ines: The paper suggests LLMs show parallel susceptibility to those false memories, but the study noted that LLMs actually have greater accuracy and less interference in recognition tasks than human subjects.

Marcus: They also looked at list sensitivity, finding that recall accuracy declines predictably as list length increases, which mirrors human performance trends.

Yuki: That brings us to a big difference they found: positional bias or the position effect.

Ines: LLMs show no sensitivity to the order in which information is presented; unlike humans, they aren't showing classic primacy or recency effects on recall accuracy.

Marcus: And they also showed a stronger resistance to nonsense material, meaning semantically meaningless content causes less error compared to human participants.

Yuki: That robustness against nonsense material is interesting because it points toward a different kind of processing than what we see in people.

Ines: The paper concludes that while LLMs mirror the pattern of error generation, there are key differences in how they handle order and noise.

Marcus: We have to remember that this is just one piece of the puzzle when you look at how these models actually operate at scale.

More episodes

← Home