INBIOSIS · UKM
A daily digest of the three most important new papers on AI and LLM applications in bioinformatics, computational biology, genomics, proteomics and drug discovery. Curated every morning.
30 issues · 90 papers · updated 2026-08-09
Cutting‑edge AI tools are streamlining data‑intensive workflows across bioinformatics, from molecular networking to gene‑set interpretation.
An open‑source toolkit (MN‑Suite) built with large language model assistance integrates multiple similarity algorithms for mass‑spec molecular networking, offering flexible, server‑free analysis of natural product datasets.
Why it matters. It demonstrates how LLM‑guided software engineering can accelerate deployment of specialized bio‑informatics pipelines.
The GeneInsight framework extracts functional annotations from STRING‑DB, clusters semantically related terms using sentence embeddings, and generates concise thematic summaries via LLM prompting, streamlining gene‑set interpretation.
Why it matters. It showcases LLMs as powerful assistants for turning sprawling gene‑set outputs into clear biological insights.
Using LLM‑based text mining, quantitative context‑dependent weights are assigned to literature‑extracted gene regulations, producing GRNs that better reflect disease‑specific biology and enhance drug‑target prediction and ODE model construction.
Why it matters. It highlights the value of LLMs for building more accurate, condition‑specific regulatory models.
Cutting‑edge AI tools are reshaping how we predict structures, generate single‑cell data, and assess biosecurity risks.
Large Language Models accelerate biological research but also enable the design of harmful toxin‑like proteins, posing a biosecurity threat. The authors introduce SPIKE‑Bench, a benchmark that evaluates LLMs on toxin‑design prompts and propose a classifier to mitigate risks.
Why it matters. Highlights the need for safety testing of AI systems in biotech.
Traditional protein‑structure prediction yields a single dominant conformation. This work reframes the problem as state‑space inference, reviewing ensemble generators, physics‑based simulations, and experimental constraints to predict multiple functional conformations.
Why it matters. Moves protein modelling towards capturing the dynamic ensembles essential for function and drug design.
Introduces a causal transformer paired with a quantized VAE tokenizer to generate realistic single‑cell gene‑expression vectors. The study characterises biological fidelity and scaling laws, enabling downstream perturbation‑response modelling.
Why it matters. Provides a scalable foundation model for synthetic single‑cell data, accelerating benchmarking and method development.
Three papers showing how LLMs and structured biological priors are reshaping perturbation prediction, single-cell representation learning, and spatial transcriptomics from histology images.
Rather than training a complex predictor from scratch, this work uses an LLM to rank biologically related drugs whose transcriptomic profiles have already been measured, then averages their gene-expression signatures to approximate the response of an untested compound. Tested on the Tahoe-100M perturbation atlas, the method beats chemistry-based and mean-aggregation baselines, especially when generalizing to unseen cell lines.
Why it matters. Shows that retrieval quality — not just model complexity — drives zero-shot perturbation prediction, positioning LLMs as practical biological priors in drug discovery pipelines.
Most single-cell foundation models pretrain by reconstructing masked gene expression, which captures gene-gene dependencies but not whole-cell structure. This paper introduces a contrastive pretraining scheme that splits each cell into two co-expression-guided gene-partition views and trains the model to produce matching cell embeddings, with hard negatives built by shuffling expression values. The resulting representations rank among the best for cell-type annotation and gene regulatory network inference across six networks.
Why it matters. Demonstrates a principled move beyond masked reconstruction toward learning truly cell-level representations, a key step toward more transferable single-cell foundation models.
Predicting spatial transcriptomics from H&E histology images typically treats genes as a flat output vector, ignoring biological relationships. This work plugs the Gene Ontology hierarchy into the decoder, refining predictions progressively from broad functional domains down to individual genes via residual corrections. Evaluated across nine HEST-1k datasets, GO-guided decoding consistently outperforms flat and random-hierarchy baselines — and drops in as a plug-in that requires zero changes to the image-side backbone.
Why it matters. Proves that incorporating curated ontological structure as an inductive bias measurably boosts transcriptomic prediction from histology, making large-scale, cost-free spatial profiling more reliable.
Diffusion-based protein structure prediction, faster backbone generation on Lie groups, and a new framework for auditing AI bioinformatics agents.
Vilya-2 is a diffusion transformer that extends all-atom representation from individual molecules to protein-ligand interfaces. It achieves 59.1% of peptide interfaces at sub-2 Å backbone RMSD — far outperforming co-folding models — and sets state-of-the-art on small-molecule docking. The model generalizes to macrocycles and disulfide-stapled miniproteins several-fold larger than any in training, and can be fine-tuned for hit-to-lead campaigns.
Why it matters. Bridges the gap AlphaFold left open for peptide therapeutics, enabling reliable structure prediction for molecules with non-canonical residues and complex topologies that dominate the next generation of drug candidates.
As LLM agents increasingly plan and execute biological analyses, this review introduces the Function–Evidence–Validation (FEV) framework for assessing their scientific accountability. The authors map 109 agentic systems and 28 benchmarks across genomics, single-cell omics, protein science, and drug discovery, finding that planning and tool use outpace reproducibility and external validation. They advocate evaluating workflows — not just final answers.
Why it matters. Provides the first systematic rubric for auditing whether an AI bioinformatics agent's output is scientifically credible rather than merely fluent.
Current protein backbone generators require hundreds of network evaluations with expensive Lie-group operations at each step. SE(3)-MeanFlow derives closed-form average-velocity identities in the Lie algebra, eliminating the Jacobian-vector product from the rotation branch. The result: a few-step generative model that matches or exceeds flow-matching baselines using several times fewer sampling steps, with an advantage that widens at every matched computational budget via rectification.
Why it matters. Dramatically speeds up de novo protein design inference — making high-throughput backbone generation feasible without sacrificing quality.
LLMs are moving from generic text generators to purpose‑built assistants for life‑science data, enabling faster literature mining, enzyme annotation, and even cognitive‑style reasoning about biology.
The paper introduces a “knowledge layer” that sits on top of Europe PMC, letting AI agents ask natural‑language questions and receive concise, evidence‑backed answers instead of having to craft complex search queries and read whole papers. A single LLM orchestrates sub‑queries, fetches articles, and extracts the needed statements.
Why it matters. By turning literature‑search into a plug‑and‑play service, it lowers the barrier for bio‑agents to stay up‑to‑date, accelerating hypothesis generation and data‑driven discovery.
This benchmark isolates four “levers” that affect an LLM’s ability to predict enzyme EC numbers—output format, external knowledge, reasoning structure, and robustness. It shows that open‑book (retrieval‑augmented) access dramatically lifts performance, while chain‑of‑thought reasoning helps only when the model already knows the answer.
Why it matters. It gives developers a clear diagnostic tool to spot why LLMs fail on enzyme annotation, guiding the design of more reliable bio‑AI pipelines.
The authors compare LLMs to human cognition across five dimensions (inferential organization, architecture, representations, prediction‑driven learning, and reinforcement‑like mechanisms). They argue that despite substrate differences, LLMs independently arrive at principles long identified in cognitive science.
Why it matters. Recognizing these convergences helps bio‑informaticians borrow well‑tested cognitive frameworks for interpreting LLM behavior in biological reasoning tasks.
RelAgent decomposes molecular relationship grounding into three stages — entity extraction, substructure localization, and ontology-guided reasoning — with verifier agents ranking plausible candidates. On LLaMA3.1-8B, it lifts relationship F1 from 0.1% to 54.6%, far outperforming vanilla Gemini-3.1-Pro baselines.
Why it matters. Demonstrates that agentic, structure-aware reasoning dramatically outperforms off-the-shelf LLMs for grounding natural-language descriptions of molecular substructures — a critical step for AI-assisted patent analysis and controllable molecular design.
A rigorous nine-step evaluation pipeline was applied to frozen scGPT and Geneformer — models never trained on age — across five cohorts totaling ~5M cells. Despite matching PCA on raw predictive accuracy, the foundation models encode a recoverable aging signal concentrated in NF-κB and IFN-γ inflammation programs, validated by directional intervention tests and cross-cohort transfer.
Why it matters. Provides a concrete framework for distinguishing real biology from sampling artifacts in single-cell foundation model embeddings — a problem that affects every downstream application of these models in precision medicine and biomarker discovery.
BERT-HemoPep60 uses domain-adaptive pretraining with a prefix-prompt architecture to quantitatively predict peptide hemolysis across six mammalian species. It achieves PCC scores of 0.74–0.81 for HC₅/HC₁₀/HC₅₀ predictions, outperforming conventional ML and DL models built on handcrafted sequence encodings, covering peptides up to 60 amino acids.
Why it matters. Hemolysis is a major safety bottleneck in peptide drug development. A fast, accurate, transformer-based predictor can triage thousands of peptide candidates before wet-lab screening, accelerating the discovery pipeline.
AI meets quantum chemistry, molecular glue design, and biomedical causal reasoning — three breakthroughs bridging generative models, quantum hardware, and clinical evidence.
ADAPT-GQE learns to synthesize quantum circuits for molecular ground-state preparation via a generative AI pipeline combining supervised training and reinforcement learning. Generated circuits are executed on Quantinuum Helios-1 real quantum hardware, achieving order-of-magnitude speedups over traditional ADAPT-VQE while maintaining comparable accuracy — demonstrated on imipramine, a drug-relevant molecule.
Why it matters. First AI-generated quantum chemistry circuits deployed on utility-scale quantum hardware, accelerating drug stability protocols.
TriGlue formulates molecular glue design as a ternary complex generation problem. An SE(3)-equivariant interface estimation module predicts the unknown protein-protein interface from unbound structures, then a flow-matching network jointly generates the glue molecule and the rigid-body assembly transformation, producing chemically valid molecules and plausible ternary complexes.
Why it matters. Molecular glue degraders are a growing drug class with virtually no computational design tools — TriGlue fills a critical gap.
DAGForge automates causal DAG construction for biomedical studies. An LLM generates pairwise causal judgments grounded in verbatim evidence excerpts from a reproducible literature snapshot, assembles them into a constraint-checked graph with confidence estimates, and provides an interactive browser for evidence review, adjustment-set computation, and export.
Why it matters. Causal DAGs are essential for study design but remain a manual bottleneck — DAGForge makes them auditable, evidence-linked, and reproducible.
Trust and rigor in biomedical AI — from auditing foundation-model benchmark claims to measuring confidence in multi-omics fusion.
Single-cell foundation models (Geneformer, scGPT, UCE) are trained on public repositories that overlap with popular benchmarks. The authors introduce scContam, an audit framework showing that ~80% of cells in widely used benchmarks like PBMC 3k and CELLxGENE come from the model's pretraining data — meaning benchmark scores may reflect data exposure, not genuine generalization.
Why it matters. Raises a red flag for the entire single-cell AI field: many published benchmark results may be inflated by hidden data leakage.
Most molecular models use a single 3D conformation, even though real molecules exist as ensembles of structures. EnsembleEGNN encodes entire conformational ensembles using equivariant graph neural networks and a set-attention pooler, then fuses with a BERT sequence encoder. The hybrid reaches R²=0.538 for cyclic peptide property prediction — significantly outperforming sequence-only baselines.
Why it matters. Provides a physically grounded way to model molecular flexibility, moving beyond the common "one structure per molecule" simplification in drug design pipelines.
Multi-omics fusion models often struggle when data streams are noisy or uninformative. The ACE framework dynamically reweighs each modality by reliability before fusion and produces a calibrated global trust score for every prediction, validated across BRCA, KIPAN, LGG, and ROSMAP cancer datasets.
Why it matters. Makes multimodal omics models more deployable in the clinic, where knowing how much to trust a prediction is as important as the prediction itself.
AI agents hit their limits in genomic surveillance, a GAN learns to read cancer histology and predict gene expression, and reinforcement learning cracks amyloid docking for neurodegenerative disease drug design.
A generative adversarial network that takes histopathology images and clinical metadata as input to produce realistic, biologically meaningful gene expression profiles. Its attention-based design is interpretable — it highlights exactly which image regions drove each prediction.
Why it matters. Lowers the barrier to multi-omics cancer research by generating affordable, clinically grounded transcriptomic data from widely available H&E slides.
A 100-evaluation benchmark testing whether frontier AI agents can analyze raw pathogen sequencing data under real-world surveillance conditions. Even the strongest configuration (Opus 4.8) achieved only ~50% success, with failures concentrated in pipeline choices rather than core reasoning.
Why it matters. Establishes the first standardized way to measure whether AI can be trusted with real-time outbreak surveillance — and reveals we're not there yet.
A reinforcement learning framework for docking small molecules to amyloid fibrils — the protein aggregates driving Alzheimer's and Parkinson's. Unlike standard docking, CORAL's reward captures the unique cooperative stacking binding geometry of fibril targets, significantly improving pose quality over existing methods.
Why it matters. Opens a new computational pathway for designing amyloid-targeting therapeutics, where traditional docking has long struggled due to scarce structural data.
From graph-based spatial transcriptomics to probing the hidden reasoning of chemistry LLMs and fusing genomic language models with clinical imaging — a week of foundations that push interpretability and integration forward.
A dual-graph GNN architecture that predicts spatially resolved gene expression from routine H&E stained histology slides. It explicitly models tissue heterogeneity via domain-aware edge weighting on a spot-level graph and fuses STRING protein-protein interaction priors with co-expression on a gene-level graph, all wrapped in evidential uncertainty for calibrated confidence intervals.
Why it matters. If robust, this brings transcriptome-wide spatial profiling from specialist instruments to widely available H&E slides — a major step toward clinic-ready spatial omics.
Systematic analysis across four chemistry LLM families and twelve tasks reveals that chain-of-thought traces in molecular reasoning are riddled with structural hallucinations — yet still serve a causally important "scratchpad" function. Perturbing these fabricated SMILES drafts degrades output, showing CoT is neither faithful explanation nor mere post-hoc rationalization.
Why it matters. Cautions the field against treating chemical CoT as truthful reasoning and points toward process-level supervision over answer-only evaluation for trustworthy molecular AI.
Pairs the Evo 2 genomic language model with radiomic features extracted from clinical tumor segmentations across 340 TCGA patients in three cancer types. This hypothesis-free sweep recovers known drivers and identifies 46 additional genes reaching FDR significance in kidney cancer — several previously unassociated with cancer but linked to Mendelian ciliopathies and cytoskeletal disease.
Why it matters. Demonstrates that combining zero-shot genomic foundation models with routine imaging can surface gene-phenotype associations invisible to conventional mutation-frequency approaches.
Three papers on how AI can do more than predict — it can model antibody–antigen pairing, explain its own genomic reasoning, and diagnose failure modes in perturbation biology.
A new foundation model (AAMFM) jointly represents antibody sequences and 3D structures conditioned on antigen context. Using a cross-modal adapter for geometric interfaces and epitope annotations, plus preference optimization guided by a structural prior, it achieves state-of-the-art results on functional antibody design benchmarks.
Why it matters. Most protein models treat the antibody in isolation — this approach treats the antibody–antigen pair as a unit, which is how the immune system actually works.
Sparse autoencoders trained on hidden activations of two genomic foundation models recover thousands of monosemantic features mapping to transcription-factor motifs. By ablating individual features during a forward pass and measuring the shift in the model's predictions, the authors establish which features the model causally uses — not just correlates with — TF binding.
Why it matters. Moves beyond correlation to prove that genomic AI models internally encode real biological concepts they actually rely on, setting a new standard for interpretability.
A new benchmark that tests whether models can generate mechanistically sound explanations for gene or chemical perturbations across different cell states — not just predict correct outcomes. The accompanying LLM, PertReasonLM, is trained to align outcome predictions with context-specific pathway reasoning.
Why it matters. Reveals systematic gaps between predictive accuracy and reasoning in current single-cell models, flagging failure modes invisible to standard benchmarks.
Three papers published the same day show how on-policy training, structure-predictor guidance, and RLVR are converging to make AI-generated biomolecules more structurally accurate and task-aligned.
Antibody CDR loops are notoriously hard to design generatively because backbone errors accumulate during the denoising process. This work introduces on-policy distillation — using the native structure as a privileged signal to guide only the intermediate states the model actually visits — reducing CDR-H3 RMSD by 0.42 Å over standard fine-tuning. It closes the gap between denoising training and autoregressive inference distributions for flexible loop regions.
Why it matters. Distilling on the model's own generation path instead of offline samples could become a new standard for any protein generative model, improving realism without extra data.
Rather than training dedicated scoring functions, DBMol directly uses structure prediction models (Boltz-2, AlphaFold-3) as differentiable optimizers for binding affinity. An alternating cycle of gradient-based pocket interaction optimization and flow-matching projection produces discrete, chemically valid molecules. Despite no reference-ligand supervision, DBMol substantially improves pocket coverage and maintains diversity, with competitive performance under held-out evaluation.
Why it matters. Repurposing structure predictors as optimization oracles sidesteps the need for costly affinity benchmarks and could generalize to any target with a predicted complex.
Molecular design via LLMs has been limited by supervised fine-tuning's inability to handle complex multi-objective optimization. LLMol applies RLVR — specifically GRPO — to directly reward molecules by their properties (logP, QED, structural constraints). A two-stage paradigm first teaches chemical syntax, then uses verifiable reward signals to steer generation, outperforming baselines across diverse benchmarks.
Why it matters. It adapts the RLVR paradigm that's driving frontier LLM reasoning gains to molecular design, offering a principled path to optimize any quantifiable drug property without labeled datasets.
Foundation models, LLMs, and reliability-aware AI methods are reshaping how we approach molecular design, single-cell analysis, and drug discovery — with three new papers pushing the boundaries of what these tools can do in spatial domains and trustworthy prediction.
Six prominent foundation models (Nicheformer, CellPLM, scGPT-spatial, GenePT, scELMo, Novae) were benchmarked across scRNA-seq, spatial transcriptomics, and Perturb-seq using a unified framework. No single model dominated across all tasks — rankings shifted dramatically depending on modality, preprocessing, tokenisation, and biological domain.
Why it matters. Provides the community with practical, evidence-based guidance for choosing the right foundation model for each specific biological question, rather than relying on scale or leaderboard scores alone.
This study systematically tests whether general-purpose LLMs can reason in 3D for structure-based drug design, introducing a new benchmark called 3D-Fit that evaluates pocket-conditioned molecule generation with spatial constraints like anchor fragments and mandatory interactions. LLMs still trail diffusion models but can simultaneously handle multiple spatial conditions.
Why it matters. Opens a promising path for using widely available LLMs as flexible, multi-condition drug design tools, complementing specialized diffusion approaches.
RELIABLE-BA treats each docking engine as an evidential expert, scaling uncertainty through learned reliability from molecular context. The framework fuses experts via closed-form aggregation, delivering substantially better uncertainty calibration and up to 25% prediction-error reduction when filtering to high-confidence pairs.
Why it matters. Moves binding affinity prediction beyond simple consensus scoring by providing principled, interpretable confidence measures — a step toward truly trustworthy AI-guided drug discovery.
This week, AI methods leap beyond static structure prediction — modeling protein motion across time, fusing multi-modal spatial omics, and fine-tuning LLMs for molecular geometry.
A generative emulator that simulates protein dynamics by jointly enforcing geometric symmetry (via invariant point attention), structural consistency, and temporal coherence across 100-ns to microsecond trajectories — reproducing MD-level flexibility and free-energy landscapes at a fraction of the computational cost.
Why it matters. Moves the field closer to time-resolved protein modeling, bridging the gap between static structure prediction and understanding how proteins actually move during ligand binding, allostery, and catalysis.
A graph-based framework that unifies five spatial omics modalities (Visium RNA, scMultiome RNA, scMultiome ATAC, spatial ATAC, and spatial CUT&Tag) through a TransformerConv encoder trained with masked reconstruction and cross-modal alignment — boosting clustering concordance and spatial contiguity on a melanoma cohort of 54,912 spots.
Why it matters. Offers a practical blueprint for integrating the flood of multi-assay spatial data now becoming routine in cancer and developmental biology studies.
Fine-tuning frontier LLMs on molecular coordinates (Z-matrices or Cartesian) yields surprisingly accurate equilibrium structures and diverse conformers of drug-like molecules — outperforming specialized deep learning models. Z-matrices, which encode relational geometry, prove to be a superior "grammar" for LLM adaptation to molecular structure.
Why it matters. Demonstrates that general-purpose LLMs can be repurposed for computational chemistry tasks with minimal training data, potentially democratizing access to conformer generation and molecular geometry prediction.
Instead of treating each cell as a gene-token sequence, this work renders transcriptomes as images — using optimal transport to place genes at fixed spatial positions so co-expressed programs appear as local visual texture. A vision transformer pretrained by masked image modeling learns rich representations that outperform language-model baselines on cell-type classification.
Why it matters. Replaces language-centric single-cell foundations with a vision paradigm that preserves gene-gene relationships and expression magnitudes.
Binding kinetics are critical for drug efficacy but notoriously hard to compute. The authors merge unbiased path sampling with a descriptor-free equivariant graph neural network that models the committor probability directly — eliminating hand-crafted collective variables while converging escape rates orders-of-magnitude faster.
Why it matters. Brings kinetics calculations — long a bottleneck in rational drug design — firmly into the AI-native, feature-free regime.
Generative protein models excel at structure prediction, but do they truly "understand" proteins? This systematic benchmark reveals that many top-tier generative models learn suboptimal representations when tested on downstream understanding tasks like function annotation — echoing the generation-understanding gap known from computer vision.
Why it matters. Highlights a blind spot in protein AI: strong generation ≠ strong representation learning, urging model developers to evaluate understanding alongside generation.
Cross-domain AI for molecular systems and single-cell vision models are reshaping how we represent, generate, and understand biological data at atomic and cellular scales.
Most single-cell foundation models treat cells as token sequences, discarding gene-gene relationships and expression magnitudes. scVision instead renders each cell's transcriptome as a continuous image by placing genes at fixed spatial positions via optimal transport — co-expressed genes become spatial neighbors, so gene programs appear as local texture. Pretrained via masked image modeling on 72 million human cells, it outperforms existing foundation models on zero-shot cell-type annotation and multi-study integration without ever seeing a batch label.
Why it matters. Reframes single-cell representation learning as a vision problem, unlocking decades of computer-vision advances for transcriptomics.
Molecules, crystals, and proteins each have their own specialized generative models with graph, equivariant, or frame-based operators — fragmenting the field. SinAE uses a vanilla Transformer encoder-decoder with no domain-specific architectures, shifting the reconstruction burden to an iterative flow-matching decoder. It achieves near-lossless reconstruction across all three domains and demonstrates that joint molecule-crystal training strictly improves both, proving cross-domain transfer through a shared atomic latent.
Why it matters. A truly unified generative architecture could end the fragmentation of molecular AI and enable data-scarce domains to leverage cross-domain signal.
Structure-based drug design models often optimize only for binding affinity while ignoring ADMET properties critical for real-world drug development. conDitar-dev combines a multi-scale pocket representation module, a pocket-conditioned diffusion model, and a generation-time property optimizer — delivering molecules with both strong binding and favorable developability. Generated candidates for PD-L1 and CSF1R were experimentally synthesized and biologically tested, with PD-L1 hits showing Kd values of 3.49–3.75 μM and CSF1R hits reaching 200 nM IC50.
Why it matters. One of the first diffusion-based SBDD frameworks to bridge the gap between computational design and experimental validation with developable molecules.
Three exciting papers on AI-driven modeling in biomedical research — from causal gene networks and adaptive digital twins for tumor scheduling to parameter-efficient quantum-classical architectures for molecular prediction.
Instead of relying on standard differential expression analysis, this work applies causal discovery algorithms to jointly model radiation perturbations and RNA-seq gene expression data, learning directed gene networks that go far beyond correlations. The inferred graphs reveal structured biology — high in-degree housekeeping genes, high out-degree transcription factors, and a clear hierarchical separation between stress response and cell death pathways.
Why it matters. Demonstrates that causal inference can uncover interpretable regulatory targets from transcriptomic data, offering a scalable alternative to pathway-dependent annotation for personalized treatment understanding.
A hybrid framework that pairs an interpretable reaction-diffusion model with a 3D residual deep learning module to simulate brain tumor growth. The system continuously updates a patient-specific digital twin during predictive rollouts, then feeds it into a model-predictive-control optimizer to schedule chemotherapy and radiotherapy. On 387 synthetic trajectories, the hybrid cutting errors by 84% and the DT controller reduced final tumor burden by 22% over fixed schedules.
Why it matters. Brings us closer to truly personalized, adaptive cancer treatment planning where AI-driven digital twins learn from each patient's real-time response data to keep optimizing therapy.
Proposes a topology-aligned inductive bias where the model architecture directly mirrors the molecular bond graph — atoms map to computational units, bonds determine parameter sharing. Tested on both a variational quantum circuit and a classical message-passing network with a matched baseline of just 64 trainable parameters. Achieves AUC ~0.91 on HOMO-LUMO gap classification (QM9) and reaches 90% of asymptotic performance within only 250 training molecules.
Why it matters. Shows that architectural symmetry matching chemistry's graph structure is the real efficiency driver — critical insight as the field races toward parameter-efficient models for low-data drug discovery and fair quantum-vs-classical benchmarking.
LLMs tackle reaction mechanisms and two new approaches cut costs in generative molecular design.
This paper introduces a generative model that conditions drug candidate design on gene-expression profiles from diseased cells — so the generated molecules are tailored not just to a protein target but to the actual dysfunctional pathway activity of a patient's tissue. The authors show improved in-silico hit rates across multiple cancer types compared to target-only baselines.
Why it matters. Moves beyond static protein-target docking toward truly patient-specific molecular generation, a key step for precision medicine.
Reaction mechanisms — the step-by-step electron-flow explanations of how chemicals transform — are notoriously hard for LLMs. The team trained models on curated mechanistic corpora with chain-of-thought prompts and a novel mechanism-reasoning benchmark. Their approach significantly improves the ability of base LLMs to predict elementary reaction steps correctly.
Why it matters. Chemical intuition via mechanistic reasoning is a prerequisite for reliable AI-assisted synthesis planning and retrosynthesis.
Real drug discovery budgets allow only hundreds of experimental evaluations. This work proposes a generative-optimization framework that uses Bayesian acquisition functions fused with a property-conditioned diffusion model, reducing the number of required oracle evaluations by up to 20× while maintaining competitive molecular quality scores.
Why it matters. Shrinks the gap between computationally cheap screening and expensive wet-lab validation — critical for making AI drug pipelines financially sustainable.
This week's papers expose cracks in standard evaluation pipelines for drug-discovery AI, show how uncertainty-aware models can self-correct predictions at test time without retraining, and demonstrate that clinical language bridges the gap between massive pathology foundation models and deployable clinic-ready systems.
The authors combine evidential neural networks (which decompose uncertainty into aleatoric and epistemic components) with a "neighbor fusion" strategy that corrects each molecular-property prediction by Bayesian-updating from the nearest labeled training molecules. Their key advance, PG-EVIKAL, learns a property-aware distance metric to re-rank neighbors by property relevance rather than raw structural similarity. Across 16 datasets, it cuts RMSE by a median of 19.4% versus the baseline evidential model — no retraining required.
Why it matters. Uncertainty isn't just a calibration metric here; it becomes an actionable inference engine that refines predictions as new assay data arrive in real time.
While scaffold-based validation is the gold standard for evaluating molecular-property models, this work shows it systematically hides failures on truly novel chemistries. The authors define a "structural-frontier" split that reserves the sparsest, most physicochemically remote scaffold groups, revealing up to 130% higher error on frontier molecules compared to conventional scaffold splits. Even tested robust-penalty methods fail to close the gap.
Why it matters. Current benchmarking may overstate the readiness of ADMET models for real-world drug discovery, where compounds often occupy structural regions far from training data.
Whole-slide image foundation models are computationally heavy and hard to deploy. LaGuadia distills knowledge from multiple large pathology foundation models into an 87M-parameter student using clinical pathology reports as semantic glue — a vision-language meta-teacher (MedSigLIP) aligns visual features with extracted clinical keywords, then adaptively weights each teacher's contribution depending on how well it matches the clinical narrative. The compact student matches or exceeds foundation-scale models on captioning, VQA, and classification tasks.
Why it matters. Clinical language serves as a principled distillation signal, making powerful pathology models practical for routine hospital use.
Foundation models and diffusion-based generative AI are reshaping how we train, design, and predict across the molecular and protein sciences.
A 52.6B-token training corpus that consolidates fragmented biological databases — small molecules, proteins, genomic annotations, single-cell atlases, and pathways — into a unified, training-ready format. Paired with TheBioCollection-Eval, it more than doubles BioLM performance across every domain while preserving general language ability.
Why it matters. Solves the bottleneck of scattered multi-omics data for training truly cross-domain biological LLMs.
Introduces GPFlow, a generative framework that learns the rate function of an inhomogeneous generalized Poisson process, enabling protein design without pre-specifying chain length. It achieves top distributional fitness on sequence design and ranks first on 10/16 motif scaffolding tasks, perfectly recovering the target length distribution.
Why it matters. Removes the fixed-length constraint that has limited all prior diffusion- and flow-based protein design models, opening the true design space.
Surprisingly, tabular foundation models (TabPFN, TabICL) — pretrained on synthetic causal-graph tables — achieve or exceed state-of-the-art protein-fitness regression when paired with ESM-Cambrian representations, and remain competitive on small-molecule ADMET/benchmark sets. Performance hinges on coupling with expressive pretrained encoders.
Why it matters. Shows that generic in-context tabular learners transfer to biomolecular prediction, offering a drop-in, data-efficient few-shot predictor for wet-lab teams without fine-tuning.
From biologically-informed neural networks that discover hidden mechanistic operators to LLM-powered biomedical QA and patient-specific precision oncology — this week showcases how AI is closing the gap between black-box prediction and interpretable biological insight.
This paper delivers a large-scale empirical study on biologically-informed neural networks (BINNs) — models that embed mechanistic differential equations directly into training so that interpretable operators can be recovered from noisy sparse data. Across canonical PDE benchmarks, the authors show that moderate expressivity, balanced data/PDE losses, and intermediate learning rates outperform over-parameterized designs. Practical diagnostics for detecting overfitting and failure modes are provided.
Why it matters. Provides evidence-based design guidelines for PINN-style models, accelerating their adoption as credible tools for biological model discovery rather than just curve-fitting.
PREDIKTOR aligns a patient-specific gene regulatory network (from DysRegNet + DrugBank) with a frozen gene-gene attention model pretrained on LINCS L1000 data, using a CLIP-style contrastive objective. The joint embeddings predict clinical drug response from pre-treatment transcriptomes alone — outperforming baselines on TCGA splits and transferring zero-shot to the I-SPY2 trial with a 5.6% AUROC gain, while recovering known mechanisms as stable pathway attributions.
Why it matters. Bridges the interpretability-stability trade-off in precision oncology by merging knowledge-graph grounding with large-scale perturbation transfer learning in a single framework.
Rather than treating all biomedical questions the same way, this framework routes yes/no, factoid, and list questions to tailored LLM inference pipelines — snippet shuffling and self-reflection for yes/no, chain-of-thought in-context learning for factoids, and a four-agent collaboration for list answers. It achieved first place in the BioASQ 14b factoid subtask, proving that question-type-aware orchestration outperforms a one-size-fits-all LLM prompting strategy.
Why it matters. Demonstrates a practical blueprint for deploying LLMs in biomedical QA systems where reliability and evidence grounding matter more than raw fluency.
The week saw major strides in agentic AI for drug discovery, autonomous computational pharmacology, and scalable epigenomic deconvolution.
Most AI drug generators condition only on target proteins or molecular properties — DrugGen 2 adds disease ontology as a conditioning signal. Built on fine-tuned GPT-2 with GRPO reinforcement learning, it generates molecules aware of how the disease context shapes target behavior, outperforming prior models across five diabetic nephropathy targets in novelty, structural similarity to approved drugs, and predicted binding affinity.
Why it matters. Integrating disease context into molecule generation better mirrors the clinical reality that the same target can behave differently across disease states, enabling smarter de novo design and repurposing.
LLM agents struggle with long-horizon scientific workflows because their context windows fill up. This work introduces a three-layer hierarchical memory that keeps injected context bounded — median 301 tokens, max 4,050 across 104 runs — by capping each state category and evicting completed work. Five specialist agents operate autonomously under domain-expert "PI agents" that enforce physics-based checklists and structured-domain knowledge, achieving robust pharmacokinetic model selection and improved parameter recovery versus single-agent baselines.
Why it matters. Solves the context degradation bottleneck that has held back multi-session autonomous AI in computational biology, a key step toward reliable AI co-researchers for long-running drug development projects.
Cell-type deconvolution from DNA methylation data is bottlenecked by the many-to-many mapping between methylation patterns and cell types, which breaks traditional hard-label classifiers at scale. The authors introduce data-driven soft labels that estimate a conditional cell-type probability distribution for each DNA read rather than forcing a single class assignment. Integrated into a new framework called Syto, it reduces MSE by 2.56× over the state of the art on a 39-cell-type whole-body atlas and generalizes to out-of-distribution 16-tissue data.
Why it matters. Soft-labeling turns a fundamental statistical obstacle — conflicting read-to-cell-type mappings — into a tractable probabilistic formulation, paving the way for large-scale epigenomic deconvolution in precision medicine.
Foundation models and LLM agents are pushing deeper into structural biology and single-cell analysis.
A multimodal scientific foundation model that unifies proteins, small molecules, and inorganic crystals into a single structure-aware token vocabulary, then reasons over them. It lifts low-homology Gene Ontology prediction F₁ from 0.42 to 0.55, retrosynthesis accuracy from 0.63 to 0.72, and hits SOTA on 67 of 86 benchmarks.
Why it matters. One model that both predicts and explains structural reasoning across biology and chemistry — interpretable AI for biomolecular design.
An autonomous multi-agent LLM system that converts natural-language requests into full spatial/single-cell trajectory-inference pipelines, with adaptive tool selection and a self-evolution module that refines outputs. Achieves >40% efficiency gains with expert-aligned results across six datasets.
Why it matters. Democratizes complex spatiotemporal transcriptomics by letting biologists skip the manual pipeline wrangling.
A lightweight graph-rewiring framework that updates biomedical concept co-occurrence edges as scientific language evolves, avoiding costly retraining. On BIOMRC, AUROC jumps from 0.633 to 0.699 with no precision loss.
Why it matters. Keeps biomedical retrieval and knowledge-discovery models current as new terms and findings explode — a practical fix for fast-evolving life-science text.
Agentic AI moves from toy demos to full research workflows — multi-agent systems that ground claims, run real code, and rank antibodies by in-context learning.
A multi-agent framework that generates complete bioinformatics manuscripts with every claim grounded in 60–100 retrieved papers, experiments executed by an autonomous coding agent (no fabricated results), and an 8-dimensional quality scorer with hallucination penalties. Validated on five case studies with zero out-of-range citations and ~$0.31 per paper.
Why it matters. Tackles the three biggest blockers for AI-generated science — ungrounded claims, fake results, and no quality bar — in one reproducible pipeline.
Reframes therapeutic antibody ranking as in-context learning: a pretrained structural encoder plus context ranking head uses a few labeled comparisons per antigen to adapt at test time without gradient updates, outperforming baselines on the AbRank benchmark — especially under distribution shift.
Why it matters. Brings LLM-style ICL to protein engineering, letting a single model handle diverse antigens instead of retraining per target.
A diffusion model that generates temporally coherent protein dynamics by denoising in the Fourier (spectral) domain, separating slow conformational modes from fast atomic jitter and conditioning on temperature. Trained on mdCATH, it reaches 0.844 RMSF Pearson r on held-out proteins.
Why it matters. Cheaper, scalable alternative to molecular dynamics that captures flexible vs. rigid motion explicitly via spectral inductive bias.
Open biomolecular foundation models, interpretable multi-omics, and grounding LLMs in real molecular structure are reshaping how AI meets biology.
Introduces OpenDDE, an open-source all-atom biomolecular foundation model that treats co-folding as a shared reasoning layer for sequence–structure–function modeling across complexes — enabling structure prediction today and a foundation for downstream therapeutic design.
Why it matters. A transparent alternative to closed protein-folding systems that researchers can fine-tune for their own drug-discovery pipelines.
Pathway Activity Autoencoders embed prior knowledge as architectural constraints on multi-omics data, balancing interpretability with non-linear capacity. The framework is benchmarked in breast cancer for survival analysis and risk stratification.
Why it matters. A rare deep model that keeps biology visible — pathway-level explanations without sacrificing representational power for clinical prediction.
MolBasic is a structure-first pretraining framework that teaches molecular LLMs to translate SMILES into explicit graphs, fixing their poor baseline grasp of structure. The approach improves downstream molecular understanding and generation tasks.
Why it matters. It exposes a weak spot in today's molecular LLMs and offers a lightweight fix — chemistry-grounded pretraining that respects "structure determines function."
LLMs are being wired into every layer of the biology stack — from reading decades of gene literature to seeing chemical structures to mapping single-cell gene regulation — turning siloed tasks into integrated, scalable pipelines.
An LLM pipeline reads the primary literature once per gene and writes a reusable mechanistic annotation. Applied to all 19,293 human protein-coding genes, it fills in thousands of UniProt stubs and beats the curated reference on 99.1% of head-to-head comparisons in a blind LLM-judge evaluation. All annotations are released openly as a continuously-updated census of gene function.
Why it matters. It turns the unfilled gaps in reference databases into a single, reproducible LLM-curated resource — finally scaling literature-grounded gene annotation genome-wide.
A vision-language model that injects chemical-bond adjacency into vision tokens and aligns visual features with SMILES semantics, giving molecular VLMs genuine topological awareness. It outperforms prior molecular LLMs, general VLMs, and task-specific tools across chemical image reasoning benchmarks.
Why it matters. Drug discovery still lives in scanned papers and chemistry textbooks — MolSight makes "read the molecule in the figure" a reliable, top-tier capability rather than a demo.
A multi-task learning framework that jointly infers cell-type-specific gene regulatory networks and models their dynamics from paired scRNA-seq and scATAC-seq data. Applied to a cellular reprogramming dataset, it identifies key regulators driving cell-fate transitions along the trajectory.
Why it matters. It packages trajectory-aware GRN inference into a single tool — letting biologists ask "which gene flips the switch?" directly from multi-omic snapshots, with no bespoke pipeline.
LLM-driven gene curation, chemical vision-language models, and transformer-based mass-spec prediction push practical bio-AI forward.
A two-pass LLM pipeline reads primary literature once per gene and writes structured mechanistic annotations, covering 19,293 human protein-coding genes and beating UniProt on 99.1% of head-to-head genes as judged by a cross-family LLM judge.
Why it matters. First literature-grounded, genome-scale gene-annotation resource that any lab can query, audit, and extend.
Reframes MS/MS prediction as single-stage object detection on molecular graphs, hitting 70.0% Top-1 retrieval on MassSpecGym (up from 64.0% SOTA) and 52.5% on NIST'20, with ~8× faster inference than the previous two-stage model.
Why it matters. Sharper, faster spectrum prediction directly accelerates metabolite identification in clinical metabolomics and systems biology.
A vision-language model that injects chemical-bond adjacency into visual tokens and aligns them with symbolic chemistry, beating existing molecular LLMs and VLMs across multiple chemical image reasoning tasks.
Why it matters. Bridges image-based chemistry (patents, papers, whiteboards) with structured molecular data, a long-standing gap in cheminformatics tooling.
Foundation-model methods keep maturing across genomics, proteomics, and clinical AI — with new benchmarks, fresh LM assessments, and a clever reformulation of an old analytical-chemistry problem.
An LLM-driven pipeline that reads primary papers at scale to assign mechanistic Gene Ontology annotations across the human proteome, closing the gap where curated databases lag years behind new evidence.
Why it matters. Could replace thousands of hours of manual curation and give every gene a continuously updated, evidence-linked functional description.
A systematic benchmark comparing transformer-based (DNABERT2) and convolutional (ConvNova) genomic LMs across downstream fine-tuning tasks, isolating what pre-training choices actually transfer.
Why it matters. Gives practitioners an evidence-based recipe for picking and tuning DNA LMs instead of relying on leaderboard hype.
Reformulates MS/MS spectrum prediction from molecular graphs as an object-detection task on peaks and fragments, outperforming prior ML models on metabolomics benchmarks.
Why it matters. A faster, more accurate spectra predictor directly accelerates metabolite identification in clinical and systems-biology pipelines.
From LLM-curated gene functions to chemistry-vision models and self-driving labs, AI agents now span the entire biology stack.
An LLM-driven pipeline that reads primary papers and assigns mechanistic functions to every human protein-coding gene, closing the multi-year lag of curated databases by combining retrieval, reading, and evidence-grounded reasoning into auditable, citation-backed annotations.
Why it matters. A long-standing functional genomics bottleneck is solved at scale — researchers get a continuously refreshed, traceable map of what every gene actually does.
MolSight fuses molecular graphs with a vision-language backbone so it can interpret chemical structure images, reaction diagrams, and tabular data in one model, outperforming text-only molecular LLMs on image-grounded tasks central to drug design.
Why it matters. Chemistry has always been a vision problem as much as a language one; this model finally lets molecular AI "see" structures the way chemists do.
A multi-agent LLM system that drafts wet-lab protocols, translates them to executable code, runs them on lab hardware, and refines itself from experimental feedback, keeping biological intent, device constraints, and quantitative procedures aligned end-to-end.
Why it matters. Closes the loop between natural-language biology and physical experimentation, paving the way for genuinely self-driving labs.
The authors propose Active-GRPO, a reinforcement-learning recipe that lets an LLM decide per-instance whether to imitate a reference molecule or reinforce its own discoveries, while continuously upgrading the reference. On TOMG-Bench it lifts average SR×Sim from 0.10 (GRPO) and 0.17 (RePO) to 0.18 with significant gains on LogP, MR, and QED.
Why it matters. It makes LLM-based molecule design more data- and compute-efficient, a practical step toward AI agents that autonomously discover better drug-like compounds.
A controlled benchmark comparing transformer-based DNA LMs (e.g. DNABERT2) against convolutional baselines across fine-tuning tasks, isolating how much of the gain comes from pretraining versus architecture, and probing the role of Byte-Pair Encoding tokenization for genomic data.
Why it matters. It gives the community a sober, reproducible read on when expensive DNA pretraining actually pays off — and when simpler models suffice.
The authors train a lightweight alignment module on top of frozen histopathology and RNA-seq foundation models, enabling open-vocabulary gene-set queries on H&E slides without sequencing; on a 1,720-sample multi-cancer cohort they achieve a 25× retrieval boost and validate predictions on the POSEIDON NSCLC trial.
Why it matters. It turns routine histology slides into a low-cost, slide-native molecular readout, opening a path to scalable, sequencing-free biomarker discovery.