Daily Summary for 2026-11-10
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the tenth of November, twenty twenty-six, and this is the day's research.
Marcus: 18 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 CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating.
The day's research: Ines: Today's work builds an agentic cellular world model to predict biological responses.
Marcus: That model unifies single-cell expression prediction with evidence-grounded updating.
Yuki: It boosted differential-expression AUPRC by sixty-three point three percent in held-out conditions.
Ines: The explanation component correctly answers all answerable questions with supporting proofs on a benchmark.
Marcus: We are looking at nonlinear dynamics governing function in excitable systems now.
Yuki: This explores coordinated heart contractions versus distributed brain processing via synchronization phase transitions.
Ines: Understanding biomolecular dynamics is key to drug discovery efforts here.
Marcus: Atomic trajectory modeling uses state space models to generate computationally efficient atom-level trajectories for protein systems.
Yuki: Contrastive promoter-protein pretraining shows multi-fold improvements over leading genome language models for regulatory annotations.
Ines: This even enables zero-shot co-regulated gene retrieval from genomes alone.
Marcus: We examine how neural circuits maintain stable behavior across development using time-varying connectomes.
Yuki: A single generative encoding can produce functional recurrent controllers that maintain performance across substantial developmental changes.
Ines: The most significant development maps genetic information from frozen models onto large language models without extensive training.
Marcus: This method uses a simple one-ridge regression plug derived from frozen Evo two features to inject gene information at an early layer.
Yuki: It allows real DNA sequences to outperform shuffled ones on multiple hosts, showing gains of plus twenty-two point two on Qwen3.5-4B.
Ines: The implications extend into understanding systems where dynamics are complex, like host-parasite interactions.
Marcus: Basic reproductive numbers can be misleading when a disease-free population has multiple stable states depending on transmission structure.
Yuki: Research into protein ensemble generators suggests architectural choices are less critical than training data constraints across structural regimes.
Ines: The tangle-based clustering method provides a mathematically rigorous way to extract finer structural details from personality test data.
Marcus: This reveals subtraits missed by simpler models when analyzing the Big Five paradigm in personality tests.
Yuki: Movement rules in spatially explicit population modeling can lead to localized population patterns based on habitat suitability and density dependence.
Lucky paper: 2611.05890: Ines: So we’re diving into CellCanon now, which is this agentic cellular world model that unifies prediction, explanation, and observation-driven updating.
Marcus: It introduces this shared world state that links predictions, acquired measurements, and source-linked claims to support the explanation and revision process.
Yuki: That state holds things like the continuous response ybq, DE probabilities pDEq, pupq, and Eq C t q containing claims and their dependencies.
Ines: The predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation.
Marcus: Biological evidence compilation involves the language model selecting source linked target evidence and qualifying its relevance and assay scope.
Yuki: This process results in a shared representation z q which is the concatenation of structured and semantic features, x q equals P tr x q, in R ninety-six.
Ines: Then a cell specific MLP maps this shared representation to the continuous gene panel y b d c equals mu c plus sy MLPRNAc u d c.
Marcus: For explanation and updating, Explain assesses source support for target action, control cell expression, regulatory paths, and source statements.
Yuki: It returns supported or contradicted or insufficient evidence judgments with citations and natural language rationales.
Ines: The update mechanism uses a newly measured drug response to revise predictions for the same drug in unmeasured cell contexts via a support trained kernel ridge model.
Marcus: The performance gains are substantial; it improved differential expression AUPRC by sixty-three point three percent and Macro F1 by seventeen point three percent over the best baseline on three hundred thirty-nine held out SciPlex conditions.
Yuki: Relative to gene wise reasoning on matched drug cell cohorts, the measured language model token use drops by ninety-nine point seven seven percent for Predict and ninety eight point four one percent for Explain.
Ines: CellCanon achieves differential expression AUPRC of zero point eight zero six seven and Macro F1 of zero point eight one six four, versus zero point four nine three eight and zero point six nine six one for the strongest baselines we checked out earlier today.
Marcus: In simulated experimental feedback, the update mechanism uses one measured source context per drug to exceed the best fine tuned baselines without changing model weights.
Yuki: At forty eight observations, CellCanon’s differential expression Macro F1 rises from zero point eight one nine nine to zero point eight three nine three when observation driven updating is used versus the no observation path which returns the parent's original outputs.
Ines: The evidence grounded reasoning component shows Explain correctly answers all one hundred twenty answerable questions with complete proof on a source grounded benchmark.
Marcus: It also identifies all three hundred sixty insufficient evidence cases in the observation state task.
Yuki: We looked at sensitivity analysis where removing biological annotations lowered differential expression Macro F1 from zero point eight zero seven seven to zero point seven one one seven and removing measured DMSO expression lowered the delta Pearson from zero point three zero seven four to zero point two eight one zero.
Ines: The raw source knowledge view gives differential precision of eighty-six point six one one and direction mAP of zero point nine zero six five, while the language model compilation gives eighty five point eight eight and zero point nine one three respectively.
Marcus: The compilation improves direction ranking by seventy percentage points while trading off differential ranking and a small amount of continuous error.
Yuki: With one source context per drug, CellCanon reaches differential AUPRC of zero point nine zero seven seven and Macro F1 of zero point eight two six one compared to zero point six two seven nine and zero point seven six three zero for the strongest fine tuned comparator after forty eight observations.
Ines: This observation driven updating demonstrates that predictions and biological judgments can be revised as evidence accumulates in a cumulative account of cellular responses, which is really important for drug discovery work.
Lucky paper: 2605.25242: Ines: So we're looking at C3P now, this paper is about contrastive promoter-protein pretraining for bacterial gene regulation.
Marcus: It uses self-supervised contrastive approaches to align promoters with their corresponding proteins, leveraging protein language models as a supervisory signal.
Yuki: They trained on eighty-eight million bacterial promoter-protein pairs and found multi-fold improvement over leading genome language models for inferring regulatory annotations.
Ines: That sounds like a big jump in performance compared to what those large language models are currently doing on their own.
Marcus: They also introduced zero-shot co-regulated gene retrieval, which is the ability to find co-regulated genes in a genome using no experimental data at all.
Yuki: Compared to a randomly initialized baseline, C3P training consistently provides significant zero-shot performance gains that you don't see with standard genome language models.
Ines: So this suggests we might be able to infer regulatory networks of entirely unstudied species without any prior experimental data for those species.
Marcus: That’s the implication they are pushing, suggesting a path toward building complete models of cells entirely from the genome.
Yuki: They also showed that C3P promoter embeddings outperformed the ESM2 protein embeddings they used for comparison, even at one hundred fifty million parameters.
Ines: What about the mechanics of how they built that training signal? How exactly do you align those representations?
Marcus: They tokenize a promoter sequence into overlapping three-mers and encode it with a randomly initialized transformer encoder, then contrast that with embeddings from a frozen pre-trained protein encoder like ESM2.
Yuki: The outputs from both encoders go through learnable linear projection layers to get them down to the same dimensionality, which is two hundred fifty six.
Ines: So it's essentially teaching the promoter side how to look like the protein side using that contrastive loss function called InfoNCE?
Marcus: Exactly, they use a symmetric InfoNCE loss where the model learns to maximize cosine similarity for correct pairs while minimizing similarity for incorrect pairings.
Yuki: The gains they demonstrated in nearest neighbour prediction of annotations on a held-out species showed multi-fold improvement over leading genome language models despite having far fewer parameters.
Ines: But what about the scaling aspect? How efficient is this C3P approach compared to the leading models they are trying to improve upon?
Marcus: The scaling analysis shows it achieved strong performance at a fraction of the training cost of those leading genome language models.
Yuki: That efficiency, combined with the zero-shot retrieval capability, really points toward inferring regulatory networks for entirely unstudied species.
Ines: It’s interesting because they aren't just doing better on known bacteria; they are showing potential for much broader inference.
Marcus: They also noted that their approach addresses the challenge of predicting gene regulation directly from DNA sequences, which has been a long-standing hurdle.
Lucky paper: 2611.06418: Ines: We're moving on to a paper that looks at how we can control large language models using feedback loops, specifically the one called Mind Control of LLMs via Closed-Loop Stimulation.
Marcus: It seems like they're treating LLM steering as a feedback control problem where the internal state dictates the perturbation applied to the system instead of just a fixed intervention.
Yuki: They developed an H infinity based steering controller to explicitly limit sensitivity to disturbances and linearization mismatches in these nonlinear transformers.
Ines: That brings up a question about how they defined robust steerability, Srob, which is defined as one divided by gamma, where gamma equals K times the absolute value of T w to z of K infinity.
Marcus: The formulation incorporates two main ingredients: robustness which relies on local linearization holding near the identification regime with a structured state dependent residual.
Yuki: And feedback where conditioning on the evolving state lowers output level mismatch and narrows that out of distribution gap, which they claim language models currently lack in their form.
Ines: They tested this robustness across four shifts: prompt language on TruthfulQA, adversarial prompting on HarmBench, input language on MGSM, and context length on L-CiteEval.
Marcus: Specifically for L-CiteEval the H infinity method achieved a small increase in Answer recall and Citation F1 compared to the original model while maintaining high fluency across context length shifts.
Yuki: They also used calibration derived Srob to complement outcome based evaluation with a pre deployment diagnostic of reliability under held out attack.
Ines: The computational complexity analysis shows that H infinity synthesis is more efficient in terms of latency than A LQR for the deployed configuration at the matched reduced state size.
Marcus: It suggests that robust steerability clusters according to behavioral objectives, which means it emerges from the interaction between the system being controlled and the objective imposed upon it.
Yuki: Can you tell me more about what happens when you test this against those shifts, like adversarial prompting on HarmBench?
Ines: When testing on HarmBench adversarial prompting, they found that calibration derived Srob improves adversarial robustness.
Marcus: I wonder how the system handles the shift to input language using MGSM; does that cause a bigger drop in performance than with context length?
Yuki: The study shows that for L CiteEval context length shifts, the H infinity method achieves a small increase in Answer recall and Citation F1 compared to the original model while maintaining high fluency across context length shifts.
Ines: So, even when shifting things like prompt language or input language, the system can still preserve underlying reasoning capability by focusing its benefit where the shift induces degradation.
Marcus: That makes sense if you think about how it focuses on the specific part of the system that is actually degrading under a particular shift.
Yuki: And what about those systems that are already stable, like L CiteEval? The H infinity method achieved a small increase in Answer recall and Citation F1 compared to the original model while maintaining high fluency across context length shifts.
Ines: That means for context length, we get a small boost without losing the general fluency of the model.
Marcus: It's interesting that they also showed that robust feedback improves adversarial robustness, which is a key area for deploying these kinds of systems safely.
Ines: So, they aren't just adding complexity; they're finding a more efficient way to achieve robustness.
Marcus: It implies that for real applications, this feedback loop approach might be more practical than relying on fixed interventions.
Lucky paper: 2611.06189: Ines: Let's look at this paper now, it’s about a minimal E/I circuit model explaining differential recovery from amblyopia under TTX and reverse occlusion.
Marcus: So the authors are looking at how fellow-eye inactivation creates a larger plasticity-permissive operating regime characterized by reduced inhibition.
Yuki: That means the network has lower inhibition and a bigger gap between excitatory activity and that BCM modification threshold.
Ines: It combines excitatory inhibitory population dynamics with BCM like metaplasticity and homeostatic regulation within a two population circuit.
Marcus: The authors show that TTX produces the lowest inhibitory activity while maintaining moderate excitation, putting the network in a disinhibited operating regime.
Yuki: They approximate that BCM modification threshold by alpha times E star squared, where E star is some characteristic value.
Ines: The results show that MD shifts the amblyopic pathway into a net weakening regime compared to TTX and reverse occlusion.
Marcus: But when they look at functional recovery, fellow-eye inactivation produces stronger recovery than reverse occlusion because TTX produces a larger increase in RA than RO does.
Yuki: This is characterized by TTX producing a larger redistribution of functional responsiveness toward the amblyopic eye compared to RO.
Ines: And they tested this robustness across three hundred randomly perturbed parameter sets, and TTX produced greater amblyopiceye response recovery in all cases.
Marcus: The median recovery advantage they found was zero point three hundred three which suggests the effect wasn't just a marginal numerical difference.
Yuki: That points to the differential recovery observed not being an isolated consequence of a single finely tuned parameter set.
Ines: The conclusion is that successful recovery depends on the dynamical state of the cortical circuit through which that input is processed, not just the sensory input itself.
Marcus: It suggests that interventions work by transiently shifting neural circuits into states permissive for functional reorganization, characterized by reduced inhibition.
Yuki: The authors intentionally abstract away many biological details like cortical layers or spike-based dynamics when they present this model.
Ines: That's a fair point, but what about the mechanism behind that shift in the BCM modification threshold?
Marcus: Well, the model explicitly uses equations one through two to describe the fast subsystem dynamics of those excitatory and inhibitory populations.
Yuki: It also includes homeostatic regulation of gE through equation three and BCM like synaptic plasticity rules in equations four.
Ines: I'm curious how this applies outside of visual processing, because that’s where the big payoff is for drug discovery.
Marcus: The paper connects this to atomic trajectory modeling with state space models to generate computationally efficient atom level trajectories for protein systems.
Yuki: So they are trying to bridge the gap between expensive molecular dynamics and generative models by capturing temporal dependencies in biomolecular dynamics.
Ines: That brings us back to C3P, which shows contrastive promoter protein pretraining yields representations capturing bacterial gene regulation.
Marcus: This is interesting because it suggests learning to align promoters to proteins provides multi-fold improvements over leading genome language models for inferring regulatory annotations.
Yuki: They even show this enables zero shot co regulated gene retrieval from genomes alone, which hints at decoding gene regulation from genomes alone.
Ines: That’s a pretty big claim, connecting pretraining to inferring function directly from sequences without prior training on that specific target.
Marcus: It is powerful because it allows real DNA sequences to outperform shuffled ones on multiple hosts, with gains ranging from plus twenty-two point two on Qwen3 point 5-4B.
Yuki: That scale of improvement across different models suggests the learned representation has a very strong generalizable signal related to gene regulation.
Ines: We also saw some work on protein ensemble generators where ablating design choices across structural regimes showed that training data is the main constraint shaping performance.
Marcus: So, it’s less about choosing the perfect architecture and more about what you train it on when dealing with ordered versus disordered proteins.
Yuki: That makes sense, as generative models often inherit much of their design from structure prediction for folded proteins.
Ines: It makes me think about the broader implications for complex systems where dynamics are tricky, like host parasite interactions.
Marcus: The paper on basic reproductive numbers shows that the basic reproductive number R zero can be misleading when a disease-free population has multiple stable states.
Yuki: The invasion threshold depends on the structure of pathogen transmission rather than just demographic dynamics alone.
Ines: That’s a key distinction; it means we have to look beyond simple demographic models when assessing persistence in these systems.
Marcus: Speaking of structure, there's this tangle based clustering method that re-assesses the mathematical foundations for the Big Five paradigm in personality tests.
Yuki: It uses recently developed mathematical theory of tangles to provide a new mathematically rigorous quantitative method for extracting finer structural details from those tests.
Ines: That suggests we might be missing subtraits in personality assessments if we only use simpler models that don't account for these tangles.
Marcus: And we also have spatial modeling where movement rules in spatially explicit population modeling can lead to localized population patterns based on habitat suitability and density dependence.
Yuki: So, when studying how things spread, the environment dictates where the population settles based on its suitability and how dense it gets.
Lucky paper: 2611.06363: Ines: So, we're moving into the paper titled Discrete Backbone Elasticity and Curvature-Dependent Torsion Statistics in Proteins now.
Marcus: This work uses coarse-grained models to describe protein conformational dynamics by using C alpha geometry to figure out the probabilities of regular helices and disordered conformations.
Yuki: The baseline local elasticity and torsional order significantly reduce negative log density loss from one point four nine zero down to one point one six zero nats, and the mean absolute chain error in negative torsion fraction drops from twelve point two six to two point one five percentage points.
Ines: That reduction in error is substantial when you look at the performance metrics compared to previous models, right?
Marcus: The two-state elastic model lowers continuous density loss from one point two three one down to zero point eight nine two nats and the mean absolute trajectory helix fraction error drops from ten point seven three to two point eight seven percentage points.
Yuki: Theoretically, they introduce a statistical mechanical framework by treating the rotational mismatch between neighboring bending generators in so three as an elastic energy, which transforms that deterministic geometric identity into a continuous six parameter conditional probability density for fluctuating protein trajectories.
Ines: That sounds like it’s trying to get past the limitations of just looking at the local geometry.
Marcus: The model defines the pair energy as E pair equals K two times x L squared plus x R two minus r K cos tau plus J s sin tau, where Js introduces a parity breaking energetic preference.
Yuki: They extended this by examining two elastic states with different bending preferences and anisotropic responses whose relative weights combine bend energies with the angular partition function, which integrates the torsional Boltzmann weight of each state.
Ines: So they are testing how matching one structural population can leave further angular information unresolved when using the two-state model.
Marcus: The two-state elastic model achieves a continuous density loss of zero point seven nine zero nats and a helix error of one point six five points, and when tested on controlled trajectories, the sequence associated two state model yields a loss of zero point eight nine three nats and a helix fraction error of two point eight six points.
Yuki: That comparison suggests that matching one structural population can leave more angular information unresolved than they thought.
Ines: And what about the analysis concerning atomic context and material orientation?
Marcus: They examine the structural information remaining at similar bends by projecting added material coefficient functions onto the six function span, and this dense matching analysis retains all zero point zero one ns frames.
Yuki: That projection measures redundancy on those sampled geometries, which is consistent with reparameterization at a fixed material angle, even though they aren't identifying a fitted state with a unique atomic structure.
Ines: So the conclusion points to distinct elastic backgrounds and directional stiffness improving prediction when their contributions are weighted by both bend energy and the angular partition function, denoted as xi uv q zv q.
Marcus: The paper concludes that transforming that discrete geometric identity into an analytical statistical potential demonstrates how exact differential geometric relations can be generalized to describe conformational thermodynamics in fluctuating biomolecules.
Yuki: It shows how matching one structural population still leaves information unresolved because different bend preferences and anisotropic elastic responses improve prediction when weighted by the angular partition function, xi uv q zv q.
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