# q-bio.QM Published articles for q-bio.QM. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## How Deep is the Signal? Time-Context up to 30 Years for Colorectal Cancer Prediction in a Primary-Care EHR Model DevFeed: [How Deep is the Signal? Time-Context up to 30 Years for Colorectal Cancer Prediction in a Primary-Care EHR Model]() Original publisher: [Read original article]() Author: Achille Raviez, Lan Wang, Brendan Delaney Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [AI Foundation Models](), [Fine-tuning](), [Training AI Models](), [model architecture]() Tags: [2](), [about](), [abstract](), [all](), [arxiv](), [base](), [buffer](), [calendar](), [cancer](), [capture](), [cases](), [context](), [context-window](), [cs-ai](), [cs-lg](), [deep](), [model](), [prediction](), [primary](), [q-bio-qm](), [signal](), [time](), [up]() ### AI overview In this primary-care EHR study, colorectal cancer prediction improved as the model's calendar lookback expanded, then leveled off: AUROC rose from 0.551 at 2 years to 0.590 at 18 years, with about 80% of the gain reached by 10 years. The 34.6M-parameter transformer was pretrained on coded events from 3.6M North West London patients and evaluated on the same cohort of 1,008 cases and 9,582 matched controls across 15 nested windows, with a 12-month pre-diagnostic buffer. An XGBoost baseline reached a similar peak AUROC of 0.589. A simulation suggests elapsed-time rotary position embeddings may alias beyond about 20 years at the default base frequency, prompting the authors to suggest retuning θ. The findings support reporting calendar lookback separately from token context; they describe one model, task, and cohort. ### Source excerpt arXiv:2611.07265v1 Announce Type: cross Abstract: EHR foundation models read medical events within a fixed context window, and long context has so far been measured only in tokens. How far back in time a model must look to capture the earliest predictive signal is unknown. We introduce time-context, the calendar lookback window before the prediction point, and provide, to our knowledge, the first systematic evaluation of decade-scale time-context for an EHR foundation model. We pretrained a decoder-only transformer (34.6M parameters) on 1.4B coded events from 3.6M primary-care patients in North West London and fine-tuned it to predict incident colorectal cancer under a 12-month pre-diagnostic buffer. A frozen cohort of 1,008 cases and 9,582 matched controls, all registered for at least 30 years, was evaluated at 15 nested lookback windows from 2 to 30 years, so that every truncation is compared within the same patients. Discrimination rose from an AUROC of 0.551 at 2 years to 0.590 at 18 years, then plateaued, with about 80% of the gain reached within 10 years; an XGBoost baseline converged to the same level (peak AUROC 0.589). A simulation further suggests that rotary position embeddings (RoPE) over elapsed time alias beyond about 20 years at the default base frequency, motivating a retuned $\theta$. Time-context is a measurable design axis for clinical prediction models, distinct from token-context, and should be reported. ## Interpretable many-body interaction laws for collective motion from few trajectories DevFeed: [Interpretable many-body interaction laws for collective motion from few trajectories]() Original publisher: [Read original article]() Author: Timon S. Gutleb Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [few-shot]() Tags: [collective](), [cs-lg](), [interaction](), [laws](), [learning](), [machine-learning](), [physics-bio-ph](), [q-bio-qm]() ### AI overview The directional atomic cluster expansion (dACE) recovers interpretable many-body interaction laws for collective motion from relatively little trajectory data. It gives an exactly equivariant acceleration expansion with uniquely separated 1-, 2-, 3-body and higher contributions, plus a fitted self-motion term designed not to absorb interactions. In simulations in 2D and 3D, it recovered known laws body order by body order; compared with dACE, SocialNet needed more data, performed worse on many-body or chiral rules, and was limited to 2D. For tracked zebrafish and golden shiners, dACE's learned laws agreed better across disjoint data halves. In shiners, the 3-body term opposed the pairwise response in dense neighborhoods. The authors also find that next-step prediction accuracy mainly reflects self-propulsion and noise, and propose combining error measures to better assess interaction learning. ### Source excerpt arXiv:2611.08314v1 Announce Type: new Abstract: Models of collective motion usually reduce social interactions to pairwise forces, despite empirical evidence for many-body effects. We introduce the directional atomic cluster expansion (dACE), which adapts the atomic cluster expansion (ACE) of interatomic potentials to self-propelled agents. dACE provides a systematically improvable, exactly equivariant expansion of an agent's acceleration that splits uniquely into 1-, 2-, 3-body and higher contributions, along with a self term for the agent's own past motion that is fitted so that it does not absorb the interactions. On simulated collectives in two and three dimensions with known rules, dACE recovers the interaction law body order by body order from little data, while the neural model SocialNet, the current closest competitor, needs more data, has much larger errors on many-body or chiral rules and is restricted to 2D. On real data sets of tracked zebrafish and golden shiners, the laws learned by dACE agree better between disjoint halves of the data than SocialNet's. The learned laws are interpretable, and in the shiners the 3-body term opposes the pairwise response as neighbourhoods get dense. Finally, we show that next-step prediction accuracy mostly measures self-propulsion and noise rather than the interactions, and suggest more robust combinations of error measures for interaction learning. ## PharmAgents2: Towards Autonomous End-to-End Drug Discovery DevFeed: [PharmAgents2: Towards Autonomous End-to-End Drug Discovery]() Original publisher: [Read original article]() Author: Haichuan Tan, Bowen Gao, Tao Huang, Jiaxin Tan, Xin Hong, Baoguo Yao, Yifei Wang, Zhichao Deng, Hanji Du, Dapeng Jiang, Yiheng Li, Weiwei Ma, Xin Ma, Yuqi Shi, Zixiang Wang, Kelin Wu, Tianxiang Wu, We Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Agentic AI](), [AI for science](), [Multi Agent Systems](), [decision-making](), [Artificial Intelligence](), [knowledge-engineering]() Tags: [agents](), [arxiv](), [autonomous](), [bench](), [cs-ai](), [discovery](), [drug-discovery](), [experiments](), [exploration](), [interaction](), [investigations](), [knowledge](), [novel](), [parallel](), [pharmaceutical](), [prototype](), [public](), [q-bio-qm](), [real-world](), [research](), [scaling](), [site]() ### AI overview PharmAgents2 combines Paper2Skill, which extracts transferable decision knowledge from public sources alongside expert-developed skills, with Self-Scaling Discovery, which uses parallel agents to explore targets, hypotheses, and strategies while sharing research context. On PharmAgents-Bench v0.1 (121 questions across nine decision scenarios and three end-to-end tasks), it outperformed evaluated baselines in most decision categories and had the highest process scores on all three end-to-end tasks. In three studies, human-run experiments supported the system's findings: a new LpxH allosteric-site inhibitor scaffold showed inhibitory activity and binding; nominated JAK1 compounds showed activity after the system prioritized covalent inhibition; and a newly identified target for a known bioactive molecule was confirmed to bind. These results are reported as a prototype toward autonomous end-to-end drug discovery. ### Source excerpt arXiv:2611.07653v1 Announce Type: new Abstract: Drug discovery requires continuous scientific decision-making across a vast research space. We introduce PharmAgents2, an autonomous system for end-to-end drug discovery that combines knowledge-enhanced decision-making with organized large-scale exploration. Paper2Skill extracts transferable scientific decision knowledge from public sources and, together with expert-developed skills, supports research decisions. Self-Scaling Discovery expands exploration through parallel agents inheriting research context and decision-making capabilities across targets, scientific hypotheses, and discovery strategies. We introduce PharmAgents-Bench v0.1, comprising 121 questions across nine decision scenarios and three end-to-end tasks assessed with expert-defined process rubrics. PharmAgents2 outperforms evaluated baselines in most decision categories and achieves the highest process scores on all three end-to-end tasks. In three real-world studies, autonomous investigations generated findings subsequently supported by human-executed experiments. For LpxH, PharmAgents2 hypothesized a previously unreported allosteric site and screened compounds against it, identifying a novel inhibitor scaffold with experimentally confirmed inhibitory activity and binding. For JAK1, it prioritized a covalent inhibition strategy and nominated compounds with experimentally demonstrated activity. Starting from a known bioactive molecule, Self-Scaling Discovery explored candidate targets across the proteome and identified a new binding target, with the interaction subsequently confirmed experimentally. PharmAgents2 represents a prototype of a virtual pharmaceutical company, marking a step toward autonomous end-to-end drug discovery. ## CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating DevFeed: [CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating]() Original publisher: [Read original article]() Author: Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao, Hao Chen Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Agentic AI](), [Large Language Model]() Tags: [agentic](), [cellular](), [condition](), [cs-ai](), [explain](), [expression](), [feedback](), [fine](), [guide](), [knowledge](), [language](), [language-model](), [measurements](), [metric](), [model](), [model-weights](), [one](), [prediction](), [proofs](), [q-bio-qm](), [updating](), [world-model]() ### AI overview CellCanon combines single-cell expression prediction, evidence-grounded explanations, and updating from new measurements through a shared state linking predictions, measurements, and source-linked claims. Its predictive core reuses condition-level biological evidence across gene-specific readouts. On 339 held-out SciPlex3 conditions across three cell lines, it reports gains over the best baseline of 63.3% in differential-expression AUPRC, 17.3% in Macro-F1, and 7.2% in direction Macro-F1. Compared with gene-wise reasoning on matched drug-cell cohorts, Predict and Explain use 99.77% and 98.41% fewer measured language-model tokens. In simulated feedback, Update uses one measured source context per drug and exceeds the best fine-tuned baselines in AUPRC and Macro-F1 without changing model weights. These are reported benchmark results; the supplied excerpt does not establish performance beyond those settings. ### Source excerpt arXiv:2611.05890v1 Announce Type: cross Abstract: Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed. We introduce CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Its predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation. A shared world state links predictions, acquired measurements, and source-linked claims, supporting explanation and revision. On 339 held-out SciPlex3 conditions across three cell lines, CellCanon improves differential-expression AUPRC and Macro-F1 over the best baseline for each metric by 63.3% and 17.3%, respectively, and direction Macro-F1 by 7.2%. Relative to gene-wise reasoning on matched drug-cell cohorts, measured language-model token use falls by 99.77% for Predict and 98.41% for Explain. On the source-grounded benchmark, Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient-evidence cases. In simulated experimental feedback, Update uses one measured source context per drug to exceed the best fine-tuned baselines in differential-expression AUPRC and Macro-F1 without changing model weights. Together, these capabilities support a cumulative account of cellular responses, with predictions and biological judgments that can be revised as evidence accumulates. ## Hindsight Relabeling is All You Need for Reach-Avoid Learning DevFeed: [Hindsight Relabeling is All You Need for Reach-Avoid Learning]() Original publisher: [Read original article]() Author: Kevin Li, Marinka Zitnik Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Reinforcement learning](), [Machine learning]() Tags: [17](), [abstract](), [agent](), [all](), [also](), [apply](), [approach](), [arxiv](), [avoid](), [behavior](), [benchmark](), [best](), [biology](), [can](), [cost](), [counts](), [cs-ai](), [cs-lg](), [data](), [data-driven](), [learning](), [need](), [q-bio-qm]() ### AI overview RADT learns offline reach-avoid behavior from suboptimal random-policy trajectories without reward or cost supervision, using avoid-region hindsight relabeling. Its decision transformer represents goals and avoid-regions as prompt tokens, so evaluation can specify any number of avoid-regions of arbitrary size without encoding them in an augmented state space or retraining. Across 17 tasks, environments, and settings, it reportedly generalizes zero-shot to avoid-region sizes and counts outside training distributions, outperforming baselines that require retraining. In one such setting, it improves normalized cost by 35.7% over the best retrained baseline while maintaining high goal-reaching success. The abstract also reports an application to cell reprogramming in biology. ### Source excerpt arXiv:2505.19337v3 Announce Type: replace-cross Abstract: Offline goal-conditioned reinforcement learning methods have shown promise for reach-avoid tasks, where an agent must reach a target state while avoiding undesirable regions of the state space. Existing approaches typically encode avoid-region information into an augmented state space and cost function, which prevents flexible, dynamic specification of novel avoid-region information at evaluation time. They also rely heavily on meticulously designed reward and cost functions, limiting transferability to novel environments and specifications. We introduce RADT, a decision transformer approach for offline, reward-free, goal-conditioned, avoid-region-conditioned RL. RADT encodes goals and avoid-regions directly as prompt tokens, allowing any number of avoid-regions of arbitrary size to be specified at evaluation time. Using only suboptimal offline trajectories from a random policy, RADT can learn reach-avoid behavior in a completely data-driven manner using our novel avoid-region hindsight relabeling approach, without the need for reward/cost supervision. We benchmark RADT against existing offline goal-conditioned RL models across 17 tasks, environments, and experimental settings. RADT generalizes in a zero-shot manner to out-of-distribution avoid-region sizes and counts, outperforming baselines that require retraining. In one such zero-shot setting, RADT achieves 35.7% improvement in normalized cost over the best retrained baseline while maintaining high goal-reaching success. We also apply RADT to cell reprogramming in biology, demonstrating its versatility. ## BEND 2.0: Benchmarking the Capabilities Needed to Design DNA DevFeed: [BEND 2.0: Benchmarking the Capabilities Needed to Design DNA]() Original publisher: [Read original article]() Author: Andreas Bjerregaard, Rachael DeVries, Dionysia Danai Brilli, Panagiotis Antoniadis, Yan Li, Anders Krogh, Wouter Boomsma, Ole Winther, Frederikke Isa Marin Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [AI Foundation Models](), [LLM evaluation / benchmarking](), [post-training]() Tags: [benchmarking](), [cs-lg](), [dna](), [foundation-models](), [framework](), [genomics](), [q-bio-gn](), [q-bio-qm]() ### AI overview BEND 2.0 evaluates DNA models across eight tasks and five experimental datasets, testing mutation-effect prediction and whether predictions can guide sequence design. Results depend on the model and representation access: linear readouts on frozen representations can match or outperform native outputs, while functional post-training improves capability. Readouts transfer across assays and genes, but average-effect accuracy can mask weaker context-specific predictions, and predicting combined mutation effects does not establish accurate interaction prediction. Model-guided design can reduce the number of candidates needed to find a desired response, yet context-specific and multi-cell objectives remain challenging. The benchmark finds signs of general utility, with performance still below saturation even on simple tasks. ### Source excerpt arXiv:2611.03806v1 Announce Type: cross Abstract: Genomic foundation models are increasingly used to propose DNA sequences and edits. But successful design requires capabilities that current benchmarks do not measure. Here we introduce BEND 2.0 (BENchmark for DNA language models), which tests how well models predict the effects of DNA changes and whether these predictions can guide sequence design. Across eight tasks built on five experimental datasets, we find that model choice and how representations are accessed both affect performance. Linear readouts fitted to frozen representations can match or exceed models' native outputs, and functional post-training substantially improves the models' capabilities. Linear readouts transfer well across experimental assays and genes, suggesting that models capture functional patterns beyond the data used to fit individual predictors. At the same time, strong prediction of average effects can conceal weaker prediction of context-specific effects, and accurate prediction of combined mutation effects does not imply accurate prediction of their interactions. In design tasks, model predictions can reduce the number of candidates needed to find a desired response, but still leave substantial room for better design choices, particularly for context-specific and multi-cell objectives. Our benchmark reveals that genomic models now display signs of general utility but have yet to saturate on performance in even simple tasks. ## A Comprehensive Benchmark of Computational Models for Antibody--Antigen Binding Affinity DevFeed: [A Comprehensive Benchmark of Computational Models for Antibody--Antigen Binding Affinity]() Original publisher: [Read original article]() Author: Xinyan Zhao, Yi-Ching Tang, Victor J Cantu, Junseok Lee, Adam E Stogsdill, Ibraheem M Hamdi, Ashwin Kumar Ramesh, Zhiqiang An, Xiaoqian Jiang, Yejin Kim Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Model Development](), [Benchmark](), [Fine-tuning]() Tags: [abstract](), [affinity](), [all](), [arxiv](), [benchmark](), [benchmarks](), [case-study](), [cdr](), [compare](), [complex](), [cs-ai](), [cs-ce](), [cs-lg](), [design](), [developing](), [evaluation](), [experimental](), [fine](), [fine-tuning](), [folding](), [framework](), [function](), [models](), [q-bio-bm](), [q-bio-qm]() ### AI overview AbBiBench evaluates antibody-antigen complexes against 216,560 experimental measurements across 13 antibodies and 9 antigens, comparing 20 computational approaches without task-specific training or fine-tuning. Global structure-conditioned inverse folding produced the strongest and most consistent affinity rankings; antibody-specific fine-tuning did not reliably improve them. The rankings were more useful for prioritizing likely affinity gains than for ruling out weak variants. In an H1N1 case study, evaluating F045-092 CDR-H3 variants against both predicted binding potential and evolutionary sequence plausibility exposed a trade-off; all 21 selected candidates showed H1-reactive ELISA signals. ### Source excerpt arXiv:2506.04235v3 Announce Type: replace-cross Abstract: Computational antibody design requires benchmarks that measure functional binding rather than similarity to natural sequences. Here we present AbBiBench, which treats the antibody--antigen complex as the evaluation unit and harmonizes 216,560 experimental measurements spanning 13 antibodies and 9 antigens. We compare 20 computational approaches without task-specific training or fine-tuning. The analyses show that global structure-conditioned inverse folding yields the strongest and most consistent affinity rankings, whereas antibody-specific fine-tuning does not consistently improve performance. Across models, rankings are more useful for prioritizing likely affinity gains than for excluding weak variants. In a subsequent case study, multi-objective evaluation of model-generated F045-092 CDR-H3 variants for H1N1 binding revealed a trade-off between predicted binding potential and evolutionary sequence plausibility; all 21 selected candidates produced H1-reactive ELISA signals. AbBiBench provides an open framework for developing and evaluating function-aware antibody design methods. ## Learning Directional Randers Geometry for Latent-Space Exploration DevFeed: [Learning Directional Randers Geometry for Latent-Space Exploration]() Original publisher: [Read original article]() Author: Th\\'eau Blanchard (HeKA | U1346, GE Healthcare), St\\'ephanie Allassonni\\\`ere (CMAP) Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Machine learning](), [Representation learning](), [Artificial Intelligence]() Tags: [cs-ai](), [cs-lg](), [datasets](), [deep-neural-networks](), [disease](), [eess-sp](), [framework](), [growth](), [improvement](), [machine-learning](), [math-dg](), [metric](), [metric-learning](), [q-bio-qm]() ### AI overview The paper proposes learning a Finsler-Randers metric for neural image latent spaces so distances can depend on travel direction. Unlike symmetric interpolation costs, the directional metric is designed to represent processes with monotonic sequential constraints, such as disease progression. The authors argue that modeling asymmetry addresses these problems where adding curvature with a symmetric Riemannian metric does not. They demonstrate the framework's potential for interpolation and extrapolation on controlled cross-sectional datasets; the excerpt gives no further evaluation details. ### Source excerpt arXiv:2611.03675v1 Announce Type: new Abstract: Deep neural networks extract powerful latent representations from images, enabling fast and efficient downstream analysis. However, these latent spaces lack explicit geometric structure, leading to possibly poor interpolation fidelity and misleading extrapolation. Latent interpolation techniques assume symmetric transition costs which are often not the case for real world data. Their symmetric nature cannot capture directional processes, such as disease progression, where certain transitions are impossible (e.g. tumor growth dynamics or cognitive improvement in Alzheimer's disease). Although Riemannian metrics extend Euclidean geometry by incorporating curvature, they remain symmetric. We show that understanding asymmetry and not curvature alone directly addresses these problems. To do so we propose a new Finsler-Randers metric learning framework. It incorporates direction-dependent distances, thus enabling the computation of trajectories respecting monotonic sequential constraints. We demonstrate its potential for interpolation and extrapolation on controlled cross-sectional datasets. ## One Mistake, Three Layers: Persistence Rules Under Depth Correlation DevFeed: [One Mistake, Three Layers: Persistence Rules Under Depth Correlation]() Original publisher: [Read original article]() Author: Maryam Rahimimovassagh, Michael Brown, Sanjeda Sara Jennifer, Niloofar Yousefi Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Representation learning](), [Decoding](), [Neural Network]() Tags: [alignment](), [blocks](), [checkpoint](), [code](), [correlation](), [cs-lg](), [error](), [estimator](), [evaluation](), [family](), [folding](), [information](), [layers](), [noise](), [novel](), [null](), [permutation](), [persistence](), [q-bio-qm](), [surviving](), [synthetic](), [transformer]() ### AI overview A three-consecutive-layer threshold rule can report false early onsets when neighboring depth measurements are correlated: in synthetic curves it placed onset 19 stages too early, and on real co-folding data it triggered on 12-17% of null trajectories instead of the nominal 5%. The proposed calibration permutes labels within each evaluation unit while keeping each permutation fixed across depth, sets per-layer thresholds from order statistics, and calibrates the persistence length. In 1,820 post-cutoff targets, pocket membership was decodable across all 48 blocks in every familiarity regime, so onset did not distinguish regimes; a pre-declared accumulation-depth measure was latest for familiar protein families with structurally novel pockets, which also showed the weakest performance after reported controls. The abstract separately reports that a public co-folding checkpoint omits 128 trunk parameters when loaded with its own code, leaving them randomly initialized and extracted representations partly noise. ### Source excerpt arXiv:2611.03968v1 Announce Type: new Abstract: Probing frozen representations layer by layer is now standard practice, but reporting when a signal first becomes decodable requires a rule for reading a depth curve, and the obvious rule is wrong. On synthetic curves with known ground truth, requiring a probe score to exceed a null-calibrated threshold for three consecutive stages reports an onset nineteen stages early under transformer-like depth correlation, with every resample returning an onset. The failure follows from correlation between neighboring layers, a hazard known in neuroimaging but not, to our knowledge, checked for depth curves. We give an estimator whose error rate is calibrated rather than assumed. Null trajectories are built by permuting labels within each evaluation unit and holding each permutation fixed across depth, preserving the correlation structure the rule must survive; thresholds are per-layer order statistics; and the persistence requirement is calibrated rather than chosen. Applied to pocket information in a frozen co-folding trunk, stratifying 1,820 post-cutoff targets by protein-family and pocket familiarity, the five-stage rule fires on 4 to 7 per cent of null trajectories against a nominal 5 per cent and the three-stage rule on 12 to 17: the synthetic failure reproduces on real data. Pocket membership is decodable above threshold at all 48 blocks in all four regimes, so the onset cannot discriminate them; the pre-declared accumulation depth does, and is latest for familiar families with structurally novel pockets in every feature set. That regime is also the weakest, by a margin surviving controls for length, pocket measure and alignment depth. Separately, a widely used public co-folding checkpoint loads against its own code with 128 trunk parameters missing; those would be randomly initialized, and every representation extracted would be partly noise while appearing normal. ## From FAIR to AI-FAIR Metadata-First Infrastructure for AI-Enabled Science DevFeed: [From FAIR to AI-FAIR Metadata-First Infrastructure for AI-Enabled Science]() Original publisher: [Read original article]() Author: Elizabeth G. Campolongo, Eric R. Sokol, Marty Downs, Gina Glenne, Gregory E. Maurer, Sydne Record, Dustin R. Rubenstein, Matthew J. Thompson, Christine Laney, Hilmar Lapp, Max Tarjan, Tanya Berger-Wol Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Databases]() Topics: [AI for science](), [AI-Ready Data](), [Machine learning](), [Artificial Intelligence]() Tags: [agentic](), [agentic-evaluation](), [ai](), [ai-readiness](), [ai-workflows](), [annotations](), [cs-ai](), [cs-db](), [cs-hc](), [cs-ir](), [data-infrastructure](), [datasets](), [evaluation](), [feedback](), [framework](), [infrastructure](), [machine](), [machine-learning](), [metadata](), [model](), [modular](), [principles](), [q-bio-qm](), [question](), [science]() ### AI overview The authors propose AI-FAIR metadata criteria to address fragmented infrastructure and one-way data flows in ecological and biodiversity AI, where derived results rarely return to data-generating communities. Framed as a bidirectional Question-Data-Model triangle, the criteria aim to support content-based analysis across collections and return AI-derived value to source scientists. An agentic workflow evaluates datasets against both AI-FAIR and traditional FAIR criteria and gives feedback on missing or incomplete metadata. Paired with a domain-specific checklist, the modular tool assesses machine-learning readiness, readiness for specified AI tasks, traceability, and ethical, CARE-aligned stewardship. The abstract presents this as a proposed, community-informed approach; it does not provide evaluation results. ### Source excerpt arXiv:2611.03327v1 Announce Type: cross Abstract: Artificial intelligence (AI) technologies and their capabilities have advanced rapidly over the past few years, leading to widespread interest and adoption in the biological and ecological sciences. However, the data infrastructure and computational resources that provide the foundation for AI workflows in ecology and biodiversity science remain fragmented. Moreover, the flow of value remains largely unidirectional: data move from collectors, through repositories and aggregators into models, and finally to research products or management actions. The derived value, such as corrected taxonomies, annotations, quality flags, and insights, rarely returns to source communities. We propose an extension of the original FAIR principles to address the needs of AI-ready infrastructure: the AI-FAIR metadata criteria, which place AI-enabled science in the context of a bidirectional Question-Data-Model triangle. This framework is designed to facilitate large-scale content-based analysis across diverse collections and ensure that the scientists generating the data benefit from the AI-enabled advances built on their data. Alongside these criteria, we introduce an agentic evaluation workflow that both rates existing datasets on the AI-FAIR and traditional FAIR criteria and provides actionable feedback where critical metadata is missing or incomplete. This modular dataset evaluation tool, paired with a domain-appropriate checklist, translates these principles into a reusable assessment of machine-learning readiness, AI-readiness for specified tasks, traceability, and ethical and CARE-aligned stewardship. Through this approach we aim to create a community-informed system to close the model data loop, so that AI also enriches data to further enable scientific discovery. ## Comparative Strategies for Handling Imbalanced QSAR Regression: A Case Study on hERG Activity Prediction DevFeed: [Comparative Strategies for Handling Imbalanced QSAR Regression: A Case Study on hERG Activity Prediction]() Original publisher: [Read original article]() Author: Ming-Yen Sung, Hsin Liu, Tung-Sheng Cho, Cheng-Wei Lee, Yufeng J. Tseng Published: 2026-11-03T05:00:00Z Content type: comparison Language: en Sources: [arXiv: Machine Learning]() Topics: [Model Development](), [scikit-learn](), [Enhancing AI model performance]() Tags: [algorithm](), [case-study](), [classification](), [compare](), [cs-lg](), [data](), [distribution](), [domain](), [evaluation](), [fingerprint](), [interest](), [machine](), [machine-learning](), [metric](), [model](), [modeling](), [models](), [performance](), [prediction](), [q-bio-qm](), [regression](), [strategies]() ### AI overview The study compares data-level and algorithm-level strategies for imbalanced hERG activity regression across GCN and fingerprint-based models. It partitions the target region into subregions using domain knowledge and data distribution, then evaluates models in the region of greatest practical interest. Selecting models by performance in that interest region was more reliable than using a total-region metric. The effects of imbalance-handling strategies also differed by model type, supporting region-aware evaluation when choosing strategies for imbalanced QSAR regression. ### Source excerpt arXiv:2611.01709v1 Announce Type: new Abstract: Data imbalance frequently occurs in QSAR modeling, yet most studies focus on classification tasks, and few address regression tasks, as it is challenging to determine which regions of the data are sparse or dense. Data-level and algorithm-level imbalance-handling strategies can have different effects on model performance. In this study, we systematically compare various strategies across GCN and fingerprint-based machine learning models. To evaluate model performance more meaningfully, we partitioned the target region into several subregions based on both domain knowledge and data distribution. We focused on the region of greatest practical interest. Our results demonstrate that selecting models based on this interest-region metric leads to more reliable performance than using the total region metric. We also observed that different imbalance-handling strategies can have divergent effects depending on the model type. These findings highlight the importance of region-aware evaluation in regression QSAR modeling and provide practical guidance for selecting appropriate strategies for imbalanced datasets. ## FACET: Feasibility-Automaton Complete Enumeration and Truncation-Free Scoring of Molecular Formulas from MS/MS DevFeed: [FACET: Feasibility-Automaton Complete Enumeration and Truncation-Free Scoring of Molecular Formulas from MS/MS]() Original publisher: [Read original article]() Author: Xujun Che, Depeng Xu, Xiuxia Du Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Neural Network](), [Automaton]() Tags: [2022](), [benchmark](), [complete](), [cs-lg](), [element](), [enumeration](), [exhaustive](), [fixed](), [model](), [neural](), [pass](), [q-bio-qm](), [recall](), [spectrum]() ### AI overview FACET separates candidate generation from ranking for molecular formula identification from MS/MS spectra. Its feasibility automaton exhaustively generates formulas matching precursor mass and ring-and-double-bond and valence constraints; a neural head ranks every candidate using one pass of a pretrained spectrum encoder. This avoids model-based candidate truncation, while exhaustive enumeration fixes recall to the chemically feasible set. On NPLIB1, reported top-1 accuracy is 0.859 with known adducts and 0.729 with unknown adducts, versus 0.766 and 0.660 for the strongest established method; FACET also reports the highest recall. On CASMI 2022, recall is 1.000, and with unknown adducts it is the most accurate method compared. ### Source excerpt arXiv:2611.01352v1 Announce Type: new Abstract: Determining the neutral molecular formula of a precursor ion from its tandem mass (MS/MS) spectrum is the first step in identifying an unknown small molecule. Existing methods trade recall, whether the correct formula is among the candidates, against the cost of examining each candidate. Methods that examine each candidate in depth must first shorten the list, by limiting element counts or truncating it, and that is where the correct formula is lost. End-to-end regressors are fast but guarantee neither validity nor recall. We present FACET, which separates deciding which formulas to consider from deciding which one to choose. A feasibility automaton builds a formula one element count at a time and admits exactly the formulas that fit the precursor mass and satisfy the ring-and-double-bond and valence constraints. Run as an exhaustive search, it enumerates the complete feasible set, so every candidate is valid and recall is fixed by the chemical constraints rather than by a trained model. A small neural scoring head then ranks every feasible formula from one pass of a pretrained spectrum encoder, so no candidate is discarded. On the NPLIB1 benchmark FACET reaches top-1 accuracies of $0.859$ with the adduct known and $0.729$ with it unknown, compared with $0.766$ and $0.660$ for the strongest established method, and its recall is the highest. On the external CASMI 2022 set its recall is $1.000$, and with the adduct unknown it is the most accurate of the methods compared. ## HistoTx: Early fusion of H&E images and spatial transcriptomics at varying spatial transcriptomics resolution with self-supervised learning DevFeed: [HistoTx: Early fusion of H&E images and spatial transcriptomics at varying spatial transcriptomics resolution with self-supervised learning]() Original publisher: [Read original article]() Author: Alice Driessen, Emil Malmsten, Steffen Wolf, Bernd Illing, Isinsu Katircioglu, Adriano Martinelli, Rolf Stahel, Solange Peters, Enriqueta Felip, Sanjay Popat, Krisztian Homicsko, Raphael Gottardo, Mar Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Machine learning](), [multimodal-ai](), [Self-supervised learning](), [Transformers]() Tags: [architecture](), [cs-ai](), [cs-lg](), [data](), [deep](), [expression](), [foundation-model](), [fusion](), [images](), [inference](), [input](), [interaction](), [learning](), [modal](), [model](), [multimodal](), [patch](), [performance](), [prediction](), [pretraining](), [q-bio-gn](), [q-bio-qm](), [self-supervised-learning](), [spatial]() ### AI overview HistoTx combines H&E image patches and spatial transcript tokens at arbitrary spatial resolutions in a shared vision transformer. It is trained by continuing self-supervised pretraining of a vision-only pathology foundation model on paired image and spatial transcriptomics data. The abstract reports that image-only inference performs on par with the baseline on most tasks and improves tasks that benefit from molecular knowledge, such as gene expression prediction. When transcript inputs are also available at inference, the joint model outperforms either modality alone. ### Source excerpt arXiv:2611.02681v1 Announce Type: new Abstract: Spatial transcriptomics maps gene expression across tissue sections and, when integrated with whole slide images, jointly captures both morphological architecture and molecular activity to provide a rich, multimodal view of tissue biology. ML approaches combining spatial transcriptomics and WSIs typically learn a mapping from morphological features to molecular profiles via contrastive alignment or supervised fine-tuning of pathology foundation models. In both cases, transcriptomics serves as a training signal rather than a true input modality, leaving inference fundamentally image-driven. We propose an early-fusion vision transformer architecture in which transcript tokens are constructed at arbitrary spatial resolution, merged with image patch tokens, and jointly processed through a shared transformer stack for deep cross-modal interaction. Starting from a pretrained vision-only pathology foundation model, we train HistoTx by continuing self-supervised pretraining on paired image and spatial transcriptomics data and demonstrate that: 1) at inference with image only, performance remains on par with the baseline on most tasks while improving on tasks that inherently benefit from molecular knowledge, such as gene expression prediction; 2) when both modalities are available at inference, jointly providing image and transcript inputs outperforms either modality alone. ## Attribution-based diagnostics for single-cell foundation models DevFeed: [Attribution-based diagnostics for single-cell foundation models]() Original publisher: [Read original article]() Author: Farnoush Rezaei Jafari (Technical University of Berlin, BIFOLD - Berlin Institute for the Foundations of Learning and Data), Tristan Lazard (Microsoft Research), Kenza Bouzid (Microsoft Research), Shr Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [AI Foundation Models](), [ai observability](), [Ground truth / benchmark quality]() Tags: [annotation](), [architecture](), [context](), [cs-lg](), [diagnostics](), [feature](), [foundation-models](), [input](), [learning](), [performance](), [pitfalls](), [prediction](), [q-bio-qm](), [quality](), [reference](), [regulatory](), [rna](), [shortcuts](), [synthetic](), [tasks]() ### AI overview The study uses post-hoc attribution to diagnose what single-cell foundation models rely on in two tasks: cell-type annotation with pretrained Cell2Sentence and scGPT, and perturbation prediction with scGPT+STATE in a controlled synthetic setting. For annotation, it examines reliance on marker genes across architecture, expression rank, and cellular context; for perturbation prediction, it compares attribution patterns with regulatory ground truth. The authors report that attribution can expose feature use, contextual interactions, input pathway contributions, learning progression, and associations with prediction quality that standard performance metrics may miss. Its conclusions depend strongly on the biological reference, baseline, and null model, especially when regulatory knowledge is limited. ### Source excerpt arXiv:2611.02335v1 Announce Type: new Abstract: Single-cell RNA foundation models (scFMs) are increasingly used for in-silico biological discovery, yet predictive performance alone does not reveal whether their decisions rely on biologically meaningful signals or model-specific shortcuts. We systematically investigate post-hoc attribution of downstream predictions across two tasks: cell-type annotation using pretrained Cell2Sentence and scGPT models, and perturbation prediction using scGPT+STATE in a controlled setting. For cell annotation, we use attribution to assess whether predictions rely on established marker genes and how their relevance varies with model architecture, expression rank, and the surrounding cellular context. For perturbation prediction, we compare attribution patterns with known regulatory ground-truth in a synthetic setting to assess what they reveal about the model's learned computations. Across these tasks, attribution reveals biological feature use, contextual interactions, contributions from different input pathways, learning progression, associations with prediction quality, and potential pitfalls that are not fully captured by standard performance metrics. Experiments with complete and limited regulatory knowledge further demonstrate that conclusions drawn from attribution depend strongly on the choice of biological reference, baseline, and null model. Together, these findings highlight the value of attribution as a flexible post-hoc diagnostic for understanding model behavior, identifying potential issues, and informing the development of scFMs. ## Stand-In Effects and Pooling Amplification: Design-Based Tests of Whether Biological World Models Use Actions DevFeed: [Stand-In Effects and Pooling Amplification: Design-Based Tests of Whether Biological World Models Use Actions]() Original publisher: [Read original article]() Author: Bowen Xu, Boyu Chen Published: 2026-11-03T05:00:00Z Content type: news Language: en Sources: [arXiv: Machine Learning]() Topics: [World models](), [Machine learning](), [AI Models]() Tags: [actions](), [cells](), [cs-lg](), [diagnostics](), [errors](), [experiments](), [files](), [pixels](), [protocol](), [q-bio-qm](), [stat-me](), [synthetic](), [tests](), [training](), [world-models]() ### AI overview Replacing or removing actions to test whether a biological world model uses them can produce false findings. With autocorrelated designs, future actions can stand in for unseen history, and models that learn this proxy better may appear more action-responsive. A conditional randomization test expressed as an e-value is valid in finite samples for any model when its resampler conditions on the full state used by the known design; a simpler resampling law can recreate the false effect. A second failure occurs when many sites share one action sequence: pooled scores can amplify small model errors and overwhelm evidence for a true effect. In a pixel-world example, the median log e-value fell from −17.3 to −6832 as frames grew from 64 to 4096 pixels, while the oracle's rose. Tempering the score using a design effect estimated on training data changes the bet without changing validity. In the reported optogenetic ERK analysis, this approach rejected the null on a previously unopened field (E~=1.1x10¹⁰), after the tempered statistic was selected on synthetic data and preregistered. ### Source excerpt arXiv:2611.01666v1 Announce Type: new Abstract: Action-conditioned world models are increasingly trained on data from randomized experiments, and whether their predictions respond to actions is usually checked by replacing or removing the actions and measuring the loss in accuracy. We show that such diagnostics report action effects where there are none: under autocorrelated designs, future actions stand in for history the model does not see, and a better model learns the stand-in better. When the design law is known, a conditional randomization test written as an e-value is valid in finite samples whatever the model, provided the resampler conditions on the full state the design used; resampling from a simpler law reproduces the false effect. We identify a second failure, specific to world models: when many sites share one action sequence (cells in a microscope field, pixels in a frame), the natural pooled score turns small model errors along the action contrast into overwhelming evidence against a true effect. In a pixel world, its median log e-value fell from $-17.3$ to $-6832$ as frames grew from 64 to 4096 pixels, while the oracle's rose. Tempering the score by a design effect estimated on training data changes the bet, not the validity. Our own preregistered analysis of optogenetic ERK experiments failed in exactly this way; a tempered statistic, selected on synthetic data and registered before the data were obtained, then rejected the null on a never-opened field ($E\approx 1.1\times10^{10}$). Every analysis is preregistered, fixed in a dated protocol before it was run, or labelled exploratory, and every reported result is generated from hashed result files. ## From Tool State to Scientific Action: Auditing Decision Provenance in Biomedical Research Agents DevFeed: [From Tool State to Scientific Action: Auditing Decision Provenance in Biomedical Research Agents]() Original publisher: [Read original article]() Author: Xingyu Yi, Xinya Zhang, Yunshen Zhou, Haoyu Bao Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [AI for science](), [Operations research and optimization]() Tags: [agents](), [auditing](), [biomedical](), [cs-ai](), [cs-lg](), [decision-provenance](), [execution](), [provenance](), [q-bio-qm](), [recovery](), [safety](), [testing](), [verification]() ### AI overview BioAgent-DF audits whether biomedical research agents use the same evidence even when their verdicts match. It restores or transplants paired values at their original computation sites and reruns agents on healthy hardware. Across eight SciFact batches, it found repeatable read or citation changes in seven model-capture-claim combinations covering two claims; full restoration recovered reference evidence sets without changing verdicts. Verdict changes seen with the baseline prompt did not recur with an explicit-example prompt. In hERG assay allocation, silent numerical faults changed testing decisions while the fixed-budget shortlist retained the same number of compounds labeled blockers. Restoring earlier execution state and supplying certified scores recovered all 39 changed decisions; replacing the prediction file recovered none, and no incorrect action was committed. Streaming arithmetic checks cut forward-execution-plus-verification time 42.4% versus the original FP64 implementation, with identical flags and recovery outputs. ### Source excerpt arXiv:2611.01527v1 Announce Type: cross Abstract: A scientific agent can give the same verdict while reading and citing different evidence. BioAgent-DF tests this separation using paired reference and candidate values saved at the same scientific computation site. It restores or transplants those values at their original coordinates, then runs the agent on healthy hardware. Across eight SciFact batches, Qwen3.5-4B shows repeatable citation changes in three capture-claim combinations, and Gemma 3 4B IT shows repeatable read and citation changes in four. These seven combinations cover two claims. Full restoration recovers the reference evidence sets while verdicts remain unchanged. Verdict changes observed under the baseline prompt did not recur under the explicit-example prompt. Complementary tool and optimizer experiments distinguish recovery of numerical state, recovery of ranked lists, and restoration of an agent's visible history. These results show that an audit should track both the evidence an agent uses and its final answer. In hERG cardiac-safety assay allocation, silent numerical faults changed testing decisions while the fixed-budget shortlist kept the same number of compounds labeled as blockers. After the agent had used faulty values, restoring its earlier execution state and supplying certified scores recovered all 39 changed decisions, whereas replacing the prediction file recovered none; no incorrect action was committed. A streaming implementation of the arithmetic checks reduced the total time for forward execution and verification by 42.4% relative to the original FP64 implementation, with identical flags and recovery outputs. ## Physiological Trajectories and Observation Patterns in ICU Mortality Prediction: Transportability Under Changing Measurement Conditions DevFeed: [Physiological Trajectories and Observation Patterns in ICU Mortality Prediction: Transportability Under Changing Measurement Conditions]() Original publisher: [Read original article]() Author: Benjamin T. Pike Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Predictive AI](), [AI Models](), [ai evaluation]() Tags: [cs-lg](), [features](), [measurement](), [models](), [patterns](), [prediction](), [q-bio-qm]() ### AI overview Across external ICU datasets, longitudinal physiological summaries improved mortality prediction more consistently than measurement frequency and timing features. In eICU, adding trajectories to latest-value snapshots reduced log loss by 0.00922; adding observation-pattern features helped locally developed models but showed no clear benefit after transfer from MetaVision. With observation counts held constant, measurements spread across the prediction window benefited the richer representation more than random retention, while removing the final 12 hours increased its loss more. Under reduced measurement availability, the local benefit of observation-pattern features became harmful. The results suggest evaluating prediction models under changing measurement conditions as well as in new patient populations. ### Source excerpt arXiv:2610.42126v1 Announce Type: new Abstract: Objective: To distinguish the transferable predictive value of physiological trajectories from that of observation patterns in ICU mortality prediction, and to assess whether measurement timing affects performance independently of observation quantity. Materials and Methods: In-hospital mortality was predicted at 48 hours using MIMIC-III, MIMIC-IV, and eICU. Nested representations progressively integrated latest-value snapshots (R1), longitudinal physiological summaries (R2), and measurement frequency and timing features (R3). Models fitted on 7,762 MetaVision patients were evaluated on 21,682 MIMIC-IV patients and 43,878 eICU patients across 204 hospitals. The effect of measurement timing was assessed while holding the number of retained observations constant. Results: Physiological trajectories improved external prediction in eICU (R2-R1 log-loss difference -0.00922, 95% CI -0.01148 to -0.00704). Observation-pattern features improved locally developed eICU models (R3-R2 -0.00257, 95% CI -0.00370 to -0.00146), but not clearly after transfer from MetaVision (-0.00012, 95% CI -0.00206 to +0.00208). R3 was favoured in 10 of 30 source bootstrap refits. At matched observation counts, retaining measurements spread throughout the window reduced log loss relative to random retention by 0.03280 for R3 versus 0.00703 for R2. Removing the final 12 hours increased log loss relative to count-matched random removal by 0.01947 for R3 versus 0.00822 for R2. Under reduced availability, the local R3 increment changed from beneficial to harmful. Discussion: Physiological trajectories transferred more consistently, while observation-pattern features were particularly sensitive to the timing and availability of measurements. Conclusion: Clinical prediction models should be evaluated not only in new patient populations but also under the measurement conditions that supply their inputs. ## A small HER2 reader for routine H&E slides: real-world deployment DevFeed: [A small HER2 reader for routine H&E slides: real-world deployment]() Original publisher: [Read original article]() Author: Thang Tran (CloudKites AI Lab, New South Wales, Australia), Lan Dang (Monash Business School, Monash University, Victoria, Australia) Published: 2026-11-02T05:00:00Z Content type: news Language: en Sources: [arXiv: Machine Learning](), [arXiv: Software Engineering]() Topics: [Advanced AI workflows](), [MONAI](), [Fine-tuning](), [data-engineering]() Tags: [attention](), [benchmark](), [best](), [call](), [cases](), [classifiers](), [cross](), [cs-cv](), [cs-lg](), [cs-se](), [data](), [deployment](), [distillation](), [external](), [first](), [gpu](), [hospitals](), [laboratory](), [point](), [q-bio-qm](), [real-world](), [routine](), [slides](), [small]() ### AI overview The 22-million-parameter HER2 reader analyzes routine H&E slides and is designed as a research-use pre-read to help pathologists prioritize ISH for equivocal IHC 2+ cases. It combines a ViT-S/16 encoder trained with self-supervision, distillation from H-optimus-0 and JEPA refinement with five cross-validated slide classifiers. Across external hospitals, reported AUCs were 0.82 on Yale-HER2, 0.75 on HEROHE and 0.63 on BCNB biopsies; for 85 HEROHE IHC 2+ cases, it ranked ISH-amplified tumors above non-amplified ones at 0.76. It performed comparably to readers using encoders 4 to 50 times larger in the stated benchmarks. Per-laboratory score normalization and local slides improved performance at a new site. CPU inference takes 5-6 seconds per slide. The excerpt presents it as decision support alongside IHC, not a replacement for pathology review or confirmatory testing. ### Source excerpt arXiv:2610.40662v1 Announce Type: cross Abstract: HER2 status decides who receives HER2-targeted therapy. Pathologists score HER2 protein by immunohistochemistry (IHC); equivocal IHC 2+ tumours, their hardest calls, wait for in situ hybridisation (ISH). We built a HER2 reader for routine H&E slides, owned end to end, as a baseline for pathologists. A 22M-parameter ViT-S/16 tile encoder was trained on public TCGA slides by self-supervision, distillation from H-optimus-0 and two rounds of JEPA refinement; five cross-validated attention-based slide classifiers average their probabilities; and an operating point is fixed on training data. Compared under one paired ten-seed protocol, readers on our encoder rank first of 7 on external validation from TCGA-BRCA to Yale-HER2 (AUC 0.812), level with readers on encoders 4 and 50 times larger. We also rebuilt PathBench's encoder benchmark on public cohorts; our replica reproduces its published order of reference encoders, and under it our encoder ranks first of 7 on TCGA-BRCA (0.689) and sits level with H-optimus-0 and Phikon on Yale-HER2 (0.897). On three external hospitals our deployed reader reaches AUC 0.82 on Yale-HER2 (about 0.81 with training slides balanced for tissue size), 0.75 on HEROHE and 0.63 on BCNB biopsies, and among HEROHE's 85 IHC 2+ cases it ranks ISH-amplified above non-amplified tumours at 0.76, best among our TCGA-trained readers. Per-laboratory score normalisation and a laboratory's own slides improve it further on a new site. Without a GPU our reader needs 5-6 s per slide. It serves as a research-use pre-read in a pathology platform, where it proposes a probability and a call beside IHC and helps prioritise ISH for equivocal cases. Weights, per-slide predictions and the data behind every figure are public at https://huggingface.co/cloudkites/her2-reader. ## A richness-sensitive Acoustic Biodiversity Index from species-level detections: classifier benchmarking, false-positive control and validation against standard soundscape indices in an African savanna and an urban zoo DevFeed: [A richness-sensitive Acoustic Biodiversity Index from species-level detections: classifier benchmarking, false-positive control and validation against standard soundscape indices in an African savanna and an urban zoo]() Original publisher: [Read original article]() Author: Muhammad Saeed, Dmitry Mikhaylov, Mohamed H. Alhosani, Yasser F. Al Wahedi, Rory Pilossof, Ahmed Arabi, Walid Shaaba Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning]() Tags: [abi](), [benchmarking](), [cs-lg](), [cs-sd](), [false-positive](), [index](), [q-bio-qm](), [validation]() ### AI overview The study proposes ABI_E, a species-detection-based biodiversity measure calculated as Shannon diversity divided by the log of the expected regional species pool; it reports temporal stability separately. In classifier benchmarking, BirdNET embeddings with logistic regression achieved the highest macro-F1 (0.750 ± 0.029), while random frame-level splits inflated accuracy compared with recording-grouped validation. Applying BirdNET v2.4 to savanna and zoo recordings, the authors found that a stricter location filter removed 517 high-confidence detections of nine implausible species, all at the zoo. After matching recordings to six hours of effort, ABI_E was 0.38 at the savanna site and 0.03-0.21 across zoo habitats. Across 149 recorded hours, ABI_E tracked richness strongly (Spearman ρ = 0.93); the best of six standard acoustic indices, ACI, reached 0.48. The separate stability measure was lower in the savanna, which the authors attribute to penalising natural daily changes in bird activity. They argue that species-level indices need false-positive control, a declared regional reference pool and effort standardisation. ### Source excerpt arXiv:2610.42042v1 Announce Type: new Abstract: Acoustic indices are widely used as rapid proxies for biodiversity, but they summarise the distribution of sound energy rather than the species producing it and often track richness poorly. We revise and test an Acoustic Biodiversity Index (ABI) built from species-level detections. Its primary form, ABI_E, is Shannon diversity of detections normalised by the regional species pool expected at the site and season (H / ln S_ref); temporal stability T is reported as a separate component. We first benchmarked four classifiers on a strongly labelled tropical dataset (1,216 calls, 24 classes) with recording-grouped cross-validation. BirdNET embeddings + logistic regression performed best (macro-F1 0.750 +/- 0.029), and random frame-level splits inflated accuracy from 0.41 to 0.70. We then applied BirdNET v2.4 with a location and season filter to recordings from a riverine savanna in the Zambezi Valley, Zimbabwe, and from three habitats at Emirates Park Zoo and Resort, Abu Dhabi, United Arab Emirates. A stricter post-hoc location filter removed 517 high-confidence detections of 9 implausible species, all at the zoo. After diel-matched rarefaction to equal effort (6 h), the savanna site held 21 species (Hill number 1D = 9.83; ABI_E = 0.38) against 1.3-13 species (ABI_E = 0.03-0.21) at the zoo sites. Across 149 recorded hours, ABI_E correlated strongly with richness (Spearman rho = 0.93), whereas six standard indices did not (best: ACI, rho = 0.48). The stability term was lower at the savanna site (0.39) than at the zoo (0.65-0.67), showing that it penalises the natural diel cycle of bird activity. Species-level indices with explicit false-positive control, a declared regional reference pool and effort standardisation provide a more interpretable biodiversity indicator than energy-based acoustic indices. ## GraphTFactor: A graph-based approach for proteins transcription factor prediction DevFeed: [GraphTFactor: A graph-based approach for proteins transcription factor prediction]() Original publisher: [Read original article]() Author: Domenico Amato, Salvatore Calderaro, Giosu\\\`e Lo Bosco, Riccardo Rizzo, Filippo Vella Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [graph theory](), [reproducibility]() Tags: [abstract](), [application](), [approach](), [cs-ai](), [cs-lg](), [factor](), [graph](), [prediction](), [proteins](), [q-bio-qm](), [transcription]() ### AI overview GraphTFactor predicts transcription factors by representing protein sequences as graphs and classifying them end to end. The authors say this handles variable-length sequences without length constraints while retaining relational information. They report better performance than state-of-the-art methods on benchmark datasets and interpretable indications of which sequence regions contribute to predictions. An open-source package and interactive web application are provided for predicting TFs from custom sequences. ### Source excerpt arXiv:2610.40867v1 Announce Type: cross Abstract: Transcription factors (TFs) play an important role in regulating gene expression, making their identification a fundamental task in computational biology. In this work, we propose GraphTFactor, a novel end-to-end framework that models protein sequences as graphs and directly classifies transcription factors. By encoding sequences into graph structures, the proposed approach naturally handles variable-length inputs while preserving relevant relational information, without imposing length constraints. Experimental results show that GraphTFactor outperforms state-of-the-art methods on benchmark datasets, demonstrating the effectiveness of graph-based representations for TF prediction while providing interpretable insights into the sequence regions contributing to the model's predictions. To support reproducibility and practical adoption, we release an open-source software package and an interactive web application that let users easily perform TF prediction on custom sequences. ## Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database DevFeed: [Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database]() Original publisher: [Read original article]() Author: Madhulatha Mandarapu, Sandeep Kunkunuru Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Databases]() Topics: [graph-database](), [etl](), [knowledge-engineering]() Tags: [22](), [agent](), [ai](), [ai-agent](), [ai-agent-access](), [biomedical](), [clinical-trials](), [compose](), [construction](), [cs-ai](), [cs-db](), [cypher](), [data](), [database](), [databases](), [graph](), [graph-database](), [graphs](), [knowledge](), [knowledge-graphs](), [open](), [q-bio-qm](), [scale]() ### AI overview The paper describes a reproducible biomedical knowledge graph ecosystem built with Samyama, combining 22 graphs and two cohorts into a store of 328.1 million nodes and 1.44 billion edges. Portable .sgsnap snapshots can be loaded individually or composed; cross-graph integration uses query-time property joins or load-time identifier deduplication. Across 24 snapshots, the reported load took 37 minutes 49 seconds, used 366.5 GB peak memory, and merged 132,511 cross-source entities across five audited identifiers. In evaluation, 195 of 200 oracle queries passed, while a frozen 81-query suite had 17 data-gap empty results and two timeouts. On a 40-question BiomedQA benchmark using the three foundational graphs, domain-specific MCP tools scored 98% accuracy, compared with 85% for schema-aware text-to-Cypher and 75% for standalone GPT-4o; the MCP tools had zero schema errors. ### Source excerpt arXiv:2603.15080v4 Announce Type: replace Abstract: Biomedical knowledge is fragmented across siloed databases---Reactome, STRING, Gene Ontology, ClinicalTrials.gov, DrugBank, DGIdb, SIDER, PubMed, ClinVar, ChEMBL, OpenTargets, and more. We present an open, reproducible biomedical knowledge graph ecosystem built on Samyama, a Rust-native graph database: three foundational graphs detailed end-to-end---Pathways KG (118,686 nodes, 834,785 edges), Clinical Trials KG (7,774,446 nodes, 26,973,997 edges), and Drug Interactions KG (32,726 nodes, 191,970 edges)---and their expansion into a 22-graph, 2-cohort merged biomedical store (328,130,739 nodes and 1,441,425,733 imported edges, including 295,000 synthetic OMOP CDM patients from Synthea). First, we describe a reproducible ETL and portable .sgsnap snapshot pattern where graphs load alone or compose into one store. Second, we analyze single-store cross-KG integration via query-time property joins and load-time identifier deduplication (--dedup-keys): across all 24 snapshots (Samyama-Graph v1.13.0), 328.1M nodes and 1.44B edges load in 2,269.1 s (37m 49s) within 366.5 GB peak RSS (254 B/edge), merging 132,511 cross-source entities across five audited keys (mondo_id,uniprot_id,go_id,drugbank_id,symbol)---87.8% of the 150,965 Bound A sound-identifier ceiling, with zero conflicting-ID refusals. On a frozen 81-query suite, 62 queries return results (17 data-gap empties, 2 timeouts); on an expanded 200-query oracle suite, 195 of 200 pass (97.5%). Third, we introduce schema-driven Model Context Protocol (MCP) server generation: on the 40-question BiomedQA benchmark over the 3-KG subset, domain-specific MCP tools achieve 98% accuracy vs. 85% for schema-aware text-to-Cypher and 75% for standalone GPT-4o, with zero schema errors. ## What to randomize matters: a controlled study of volume, axis and representation in low-data molecular generation DevFeed: [What to randomize matters: a controlled study of volume, axis and representation in low-data molecular generation]() Original publisher: [Read original article]() Author: Neerav Kaushal, Alisa Teige, Ajay MNV Penmatsa Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [data augmentation](), [machine learning overfitting]() Tags: [autoregressive](), [cs-ce](), [cs-cl](), [cs-lg](), [data](), [generation](), [machine-learning](), [q-bio-bm](), [q-bio-qm]() ### AI overview Using an autoregressive sequence model trained on 1,000 seed molecules, the study compares SAFE representations and randomized SMILES for low-data molecular generation. Randomizing SAFE increased the genuine yield of novel molecules by more than 450-fold over canonical SAFE. Atom-order and fragment-order randomization worked jointly: each axis alone achieved only a fraction of the yield when both varied. With matched training corpora and compute, randomized SAFE produced 2.8 times the genuine yield of randomized SMILES. These findings concern the study's reported low-data setup; the supplied text does not describe broader validation. ### Source excerpt arXiv:2610.38835v1 Announce Type: new Abstract: Chemical language models can struggle to generate new molecules when trained on small datasets, where a model may reproduce training examples rather than learn a useful generative distribution. Randomized SMILES augmentation varies atom traversal, whereas Sequential Attachment-based Fragment Embedding (SAFE) also allows the order of fragment blocks to vary. How these distinct sources of representational variation affect low-data molecular generation has not been clearly studied. Here, using 1,000 seed molecules to train an autoregressive sequence model, we compare canonical SAFE with randomized SAFE at increasing corpus sizes, isolate atom-order and fragment-order axes of SAFE randomization, and compare fully randomized SAFE with randomized SMILES at molecule-by-molecule matched training corpora. Three findings emerge. First, randomization of SAFE alone raises the genuine yield of novel molecules more than 450-fold, turning a model that merely reproduces its training set into a practical molecular generator. Second, the two axes of randomization act jointly rather than additively: atom order alone reaches a tenth, and fragment order alone a seventh, of the genuine yield obtained when both vary, which is four times their combined contribution. Third, at matched corpus size and compute, randomized SAFE delivers 2.8 times the genuine yield of randomized SMILES. Welch tests with Games-Howell post-hoc comparisons resolve every difference on the primary outcomes. In the low-data regime, thus, what one randomizes rather than only how much determines whether augmentation produces new molecules or merely new spellings of the training set. ## Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups DevFeed: [Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups]() Original publisher: [Read original article]() Author: Maris Basha, Yuhang Wang, Xiaoran Chen, Longbiao Cheng, Luca Yapura, Anja T. Zai, Mathieu Salzmann, Richard Hahnloser Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Benchmark](), [Ground truth / benchmark quality](), [Transformers]() Tags: [3d](), [accelerometer](), [audio](), [benchmark](), [birds](), [call](), [camera](), [candidate](), [code](), [communication](), [cs-cv](), [cs-lg](), [cs-sd](), [evaluation](), [events](), [experiment](), [groups](), [identity](), [models](), [multimodal](), [q-bio-qm](), [radio](), [spatial](), [spatial-reasoning]() ### AI overview V33DA is a benchmark for identifying which visible zebra finch produced a detected vocalization, with physically verified caller labels from an on-body accelerometer signal withheld from benchmark models. It contains 33,625 events combining five-channel audio, three camera views, calibrated 3D poses, and radio telemetry for 10 birds across three experiments. Evaluations cover held-out dates, transfer to a new cohort, and a bird excluded from training. The abstract reports that methods relying on familiar vocal identity lose accuracy on unseen individuals, while candidate-aware methods that preserve spatial reasoning transfer better; the accelerometer-based reference remains near-perfect. V33DA++ provides verified overlapping-call events and surrounding context, but is unevaluated. ### Source excerpt arXiv:2610.38109v1 Announce Type: cross Abstract: Deciding which member of an animal group produced a vocalization and where that vocalization was produced are central to studying social communication but are rarely evaluated with physically verified ground truth. Existing resources either localize sounds without identifying the caller, or recognize individuals from isolated vocalizations without reliance on the 3D candidate geometry. We close this gap with V33DA, a dataset-centered multimodal benchmark for spatial vocal attribution in social zebra finches: it comprises 33,625 verified events coupling 5-channel audio with three camera views, calibrated 3D pose for every visible candidate, and per-bird radio telemetry across 10 individuals and three experiments. Caller identity is verified from an on-body accelerometer-derived vibration signal; this channel is withheld from benchmark models and used only by an on-body reference ceiling. Given a detected vocalization and the visible candidate birds, the task is to select the caller. We evaluate it under three regimes: session-disjoint testing on familiar birds from held-out recording dates, held-out-experiment transfer to a new cohort, and leave-one-individual-out evaluation on a bird whose vocalizations are held out from training. Reference methods expose a consistent gap between in-domain accuracy and transfer to unseen individuals: methods that can exploit familiar vocal identity perform well on known callers but degrade sharply on unseen ones, whereas candidate-aware methods preserving explicit spatial reasoning transfer substantially better. The on-body reference stays near-perfect across regimes, confirming the internal consistency of the accelerometer-verified labels. We release V33DA with fixed protocols and code for the evaluated methods, and V33DA++, an unevaluated companion set of verified overlapping-call events and +/-2 s context windows. ## Structure, Not Sequence: Tracking and Validating Folding Mechanisms in ESMFold's Trunk with Sparse Autoencoders DevFeed: [Structure, Not Sequence: Tracking and Validating Folding Mechanisms in ESMFold's Trunk with Sparse Autoencoders]() Original publisher: [Read original article]() Author: Bohan Liu, Xiaohan Fan, Wenqian Ye, Aidong Zhang Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Neural Network](), [Artificial Intelligence](), [Transformer]() Tags: [3d](), [analysis](), [autoencoders](), [biology](), [blocks](), [concepts](), [cs-ai](), [cs-lg](), [folding](), [gradient](), [intermediate](), [labor](), [language](), [language-model](), [meaning](), [mechanism](), [model](), [modules](), [network](), [neural](), [q-bio-qm](), [sequence](), [structure]() ### AI overview The authors use sparse autoencoders to interpret ESMFold's folding trunk and report that its single representation carries inherited sequence semantics while its pair representation computes pairwise geometry. Their analysis locates distinct spatial computation across modules and recycles, traces how concepts compose across blocks with gradient-based attribution, and checks interpreted activations against folding outputs and known biochemistry. The learned concepts reportedly forecast ESMFold's fold quality. These findings are summarized in the supplied abstract; it does not provide detailed methods or quantitative results. ### Source excerpt arXiv:2610.36226v1 Announce Type: cross Abstract: Structure predictors like ESMFold fold proteins using neural network computation and have transformed structural biology. However, the driving mechanism from amino-acid (AA) sequence to 3D structure is insufficiently understood, and biological meaning remains locked in the dense network. Sparse autoencoders (SAEs) have shown promising results in obtaining interpretable concept features over an overcomplete basis of intermediate protein sequence-model representations. Yet folding models, which predict physical shape from AA sequence semantics, have no such account. We hypothesize that the features carrying the folding computation evolve in the 3D structural regression process. In this paper, we interpret the folding trunk in three steps. First, a pilot analysis on the representation distribution across modules and recycles locates distinct spatial computation in the trunk. Second, the dictionaries expose a division of labor: the single representation carries inherited semantics, and the pair representation computes the pairwise geometry. Third, we interpret the dictionary at scale with a reasoning language model, trace how the concepts compose across blocks by gradient-based attribution, and follow the interpreted code to the model's folding output, verifying where the computation aligns with, or departs from, known biochemistry. Together, the sparse dictionaries we train capture the relational geometry the trunk computes beyond the sequence model, and their concept activations forecast the model's own fold quality, turning the black-box folding trunk into a sparse, interpretable, and scientifically verifiable circuit. [Next page]()