What happened
Identification of transcription factor target gene interactions and construction of the gene regulatory networks (GRNs) are essential for understanding the molecular mechanisms underlying transcriptional gene regulation. Large scale single cell transcriptomics across different tissues offers unprecedented resolution of cellular diversity and regulatory dynamics by capturing gene expression heterogeneity. However, existing methods often lack effective multimodal integration and fail to fully exploit the hierarchical structure in Gene Ontology (GO) and gene sequence level representations, which limits their ability for predictive performance and biological interpretability. We present scMGFGRN, a multi-model deep learning framework that integrates single-cell transcriptomic profiles with GO hierarchical relationships, gene sequences by leveraging denoising auto encoders, graph attention feature extraction and pertained DNA language model to capture multi-source dependencies within multi-model biological knowledge, while its gated multi head attention module effectively identifies informative regulatory signatures and integrate complementary features from different sources to predict accurate gene regulatory networks. Benchmarking on the seven datasets of human and mouse demonstrates that scMGFGRN outperforms state of the art methods in identifying GRNs. Further analyses reveal that scMGFGRN effectively identifies novel TF gene interactions (TGIs) and reconstructs cell type specific GRNs. Interpretability analysis reveals the contribution patterns of heterogeneous biological sources, demonstrating the ability of scMGFGRN to integrate transcriptomic profiles with multi model structure information. Update (2026-09-20): Chat-based Large Language Models replicate the mechanisms of a psychic's con. Article URL: https://softwarecrisis.dev/letters/llmentalist/ Comments URL: https://news.ycombinator.com/item?id=49775104 Points: 36 # Comments: 19 Update (2026-09-20): Prediction of pH-sensing histidine residues in human cells . Changes in pH play an important role in human physiology and cell behaviours, evidenced by studies of tissues and gene expression, and the biophysics of individual systems. Bridging between these scales to establish pH-sensing networks has only been partially explored. Here, amino acid conservation and structural context, and benchmarking with known pH-sensors, were used to predict histidine residues within the human proteome that could modulate pH-dependence. Proteins containing these histidines are enriched in many categories, most notably nucleic acid-binding and transcriptional regulation. DNA phosphate backbone interactions are proposed to play a major role in modulating some pH-sensors. Transcription factors with potential pH-sensors and the largest number of documented target genes are highlighted, including those involved in circadian and developmental processes. Certain CDK-cyclin pairs also feature, along with components of SWI/SNF and cohesin complexes. Focussing on histidines, including those that are solvent accessible in AlphaFold protomer models, this study generates benchmarked p Update (2026-09-22): DeepSCENIC: transfer learning from sequence-to-function models enables causal gene regulatory network inference . Sequence-to-function (S2F) deep learning models have become an important aid to decipher the genomic cis-regulatory code. However, current S2F models do not take the cellular trans-environment of transcription factors (TF) into account. Conversely, methods for gene regulatory network (GRN) inference often rely on heuristics or simple position weight matrices (PWMs), without exploiting the combinatorial grammar of genomic enhancers. Here, we present DeepSCENIC, a deep learning framework that enables causal GRN inference by performing transfer learning from S2F models to single-cell multiome atlases. We first test and validate DeepSCENIC on ENCODE cell lines, demonstrating that the framework accurately predicts single-cell gene expression and chromatin accessibility by leveraging pretrained S2F models like Enformer and Borzoi. We show that DeepSCENIC recovers TF-region (TF-RE) interactions with high fidelity, and captures de novo TF binding motifs without prior PWM knowledge. The model improves enhancer-gene associations (RE-TG) over correlation Update (2026-09-24): MoTRUST: adaptive semantic protection for mosaic single-cell integration and molecular recovery . Single-cell multi-omics reveals complementary aspects of cellular identity, but heterogeneous assays produce incomplete modality combinations. Integrating these mosaic measurements while preserving biological structure and recovering missing molecular information remains challenging. We present MoTRUST, combining neural representations, adaptive semantic protection, molecular prediction and conditional residual diffusion. MoTRUST ranked first in overall integration across 30 tasks spanning six scenarios and surpassed the strongest evaluated controls on six of eleven development recovery endpoints. Conditional residual diffusion reduced aggregate marginal prediction scores by 4.4-5.1% across three ATAC-to-RNA data sources while preserving point predictions. Together, these capabilities enable accurate mosaic integration and molecular recovery for multimodal characterization of cellular heterogeneity. Update (2026-09-24): ARSA: an autonomous research scientist for target nomination in Alzheimer's disease and related dementias . Expanding therapeutic options for Alzheimer's disease and related dementias (ADRD) requires biologically grounded targets, yet nomination demands labor-intensive analysis and multidisciplinary evidence synthesis. To address this challenge, we present ARSA, an autonomous research scientist that transforms natural-language research interests and molecular data into prioritized, evidence-grounded target shortlists. ARSA formulates and audits hypotheses, adapts molecular analyses to observed results, and prioritizes targets using cross-cohort evidence and disease-specific knowledge, preserving the evidence and decisions underlying each nomination. Across three complementary evaluations, we show that ARSA generates hypotheses corresponding to subsequent research and identifies credible candidates within and beyond community nomination records. In structured assessment by 14 experts spanning all four technical cores of the Indiana University School of Medicine-Purdue University TREAT-AD Center, every expert assigned higher mean credibility to ARSA-retaine Update (2026-09-24): NeoToxPred: a fine-tuned protein language model with orthologous and length-stratified negative controls for robust toxicity classification . Although protein toxins represent valuable pharmacological templates, predicting toxicity directly from primary sequences is inherently challenging because of their evolutionary dynamics. Active toxins and their benign homologues frequently share identical structural scaffolds and differ by only a few key residue substitutions, making global sequence similarity an unreliable indicator of function. Conventional predictors compound this challenge through flawed negative-set designs that sample non-toxic controls from arbitrary background proteins and reduce redundancy through identity-based clustering under random partitioning. This practice introduces global confounding shortcuts that inflate benchmark performance while failing to capture true functional boundaries. To address these limitations, we developed NeoToxPred, a sequence-only binary classifier that fine-tunes the pretrained ESM-C 600M protein language model end-to-end with a six-layer feed-forward classification head. Although the neural architecture is int