Biology+AI Daily

[image] Deep-learning Predictions of Biomolecular Structures: Persistent Limitations and New Horizons Extended by Explicit Ion Addition 1. This perspective argues that explicit ion inputs in AlphaFold3 can be “hijacked” as a practical knob to explore alternative conformations—sometimes

[image] Task- and dataset-specific information in protein language models 1. The study systematically shows that “use the last layer embedding” is often suboptimal for protein language models (PLMs): across 13 PLMs and 15 downstream tasks, the deepest layer is best only ~17.9% of the

[image] Accurate ∆Tm Prediction Without Protein Structure Inputs for Biomolecular Stability 1 They show that state-of-the-art ∆Tm (mutation-induced melting temperature change) prediction can be achieved using sequence only: a carefully trained ESM2-650M model reaches the best error

[image] De novo design of a protein fold for small-molecule binding through aromatic π stacking 1 They computationally designed compact de novo proteins that bind the anticancer drug doxorubicin by anchoring the binding mode on a minimal aromatic π-stacking “Trp sandwich” motif, rather

[image] TFBindFormer: A Cross-Attention Transformer for Transcription Factor–DNA Binding Prediction 1. TFBindFormer predicts genome-wide TF–DNA binding by explicitly conditioning DNA representations on TF-specific protein features, addressing a core limitation of many prior “DNA-only”

[image] AlphaConformers: Structure-guided sampling enables prediction of multiple protein conformations 1. The paper presents AlphaConformers, a structure-guided pipeline that steers AlphaFold2 (AF2) toward alternative protein conformations (apo/holo), addressing AF2’s tendency to

[image] Safety first: Input screening for protein design tools 1 Biological AI models can now design protein binders with novel sequences/structures, so classic sequence-similarity screening (e.g., BLAST) is no longer sufficient for biosecurity. This work proposes a practical

[image] Multi-peptide Prompting Enables In-context Learning in Protein Language Models 1. The paper shows that off-the-shelf protein language models (PLMs), trained on single sequences, can still do in-context learning for peptide tasks: no gradient updates, no finetuning, and no

[image] Nesso-1: Accelerating Open-Source Binding Affinity Predictions @valence_ai 1. Nesso-1 introduces an open-source, coarse-grained cofolding framework that predicts protein–ligand binding affinity in approximately one second on a single GPU. It is more than 10× faster than Boltz-2

[image] Transplanting Enzyme Active Site Geometry into Antibody CDRs for Catalytic Antibody Design 1. This preprint proposes a computational framework to design catalytic antibodies by transplanting only the local, experimentally supported enzyme active-site geometry (residues, ligand

[image] Benchmarking Deep Learning Predictions of Mutation-Induced Fold Switching 1. The study introduces a quantitative, NMR-characterized benchmark for mutation-induced fold switching using the GA/GB metamorphic model system, enabling residue-level evaluation of whether structure

[image] A Medicinal Chemistry-Centered Evaluation of AlphaFold 3 and Boltz-2 Across Diverse Binding Modalities 1. The study benchmarks AlphaFold 3 (AF3) and Boltz-2 for medicinal-chemistry use cases (lead optimization, SAR interpretation, virtual screening), emphasizing that “pose

[image] Elucidating enzyme–substrate specificity through co-folding foundation model 1 Boltz2ESI is presented as an end-to-end enzyme–substrate specificity predictor that replaces “pocket definition + rigid docking” with native enzyme–ligand co-folding, aiming to capture ligand-induced

[image] Expanding the Scope of Protein Language Modeling to Protein-Protein Interactions With MSA Pairformer @CellCellPress 1. MSA Pairformer extends protein language modeling from individual chains to protein-protein interactions. Despite being trained exclusively on single-chain MSAs,

[image] A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability @NatureBiotech 1. The AIntibody challenge delivers a rare prospective, blinded, experimentally validated test of computational antibody discovery. Across 511

[image] CatESO: Differentiable Enzyme Sequence Optimization Guided by Substrate-Aware kcat Prediction 1 CatESO connects enzyme design directly to a kinetic objective: instead of “generate sequences then score kcat”, it makes substrate-conditioned kcat prediction differentiable and

[image] A map of human protein-protein interaction embeddings for functional discovery 1 MAPPIE represents each protein-protein interaction (PPI) as its own embedding, rather than collapsing everything onto protein-centric annotations. This makes partner-specific function explicit: the

[image] Humanized Anti-PD-1 Antibodies Generated Using the Conditional Kernel-Elastic Autoencoder 1. The work proposes an AI-guided “interpolation” strategy to generate new humanized anti–PD-1 antibodies between two FDA-approved endpoints—pembrolizumab and nivolumab—that bind the same

[image] Evaluating Protein Language Model Embeddings for Structural Similarity in the Protein-Sequence Twilight Zone 1. The study tests a simple, alignment-free idea: take mean-pooled protein language model (PLM) embeddings (one vector per protein) and ask whether embedding similarity

[image] Automated synthetic cell-based screening for designed proteins with emergent functions @NatureComms 1. The paper introduces PUREdrop, an automated microfluidic screening platform that expresses protein libraries inside thousands of picoliter “synthetic cells” (water-in-oil

[image] De novo design of small-molecule–induced conformational change 1. Chang & Polizzi present a general, de novo strategy to turn a static small-molecule binder into a ligand-induced conformational switch by adding a mobile “lid” domain that closes behind the ligand, inspired by

[image] Probing and steering biology across Boltz-1’s trunk–diffusion boundary 1. The paper asks what biological information survives the architectural boundary in AF3-class predictors: a representational “trunk” (Pairformer) that processes sequence/context, followed by a diffusion

[image] TriGlue: A Biology-Inspired Generative Model for Generating Molecular Glue-Induced Ternary Complex 1. TriGlue frames molecular glue discovery as a single generative task: given a bound E3 ligase (receptor) and an unbound target protein, it generates both a novel glue molecule

[image] A ligand-property-guided computational framework for prioritizing de novo protein binders for small molecules 1. The preprint presents a ligand-property-guided, multi-stage computational workflow to prioritize de novo small-molecule protein binders, explicitly showing that

[image] Evolution-inspired Multi-objective Bayesian Optimization for Protein Engineering 1. The preprint introduces EvoMOBO, an active-learning framework for protein engineering that explicitly handles multiple objectives while operating under tight evaluation budgets, aiming to

[image] Inter-residue geometry attention for antibody-specific epitope prediction 1. The paper reframes “positional encoding” for proteins: instead of 1D sequence offsets, it uses folded 3D inter-residue displacement as the positional mechanism inside attention, targeting

[image] Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design 1 AAMFM is an antigen-specific antibody multimodal foundation model built on ESM3 that jointly models antibody sequence and structure while being explicitly conditioned on antigen context,

[image] NACraft: Programmatic nucleic-acid aptamer design via all-atom structure-model feedback 1 NACraft introduces a training-free, programmatic framework to design RNA/DNA aptamers by directly optimizing nucleotide sequences through all-atom structure-model feedback, rather than

[image] AFilter: Improved Antibody Epitope Prediction by Machine Learning-Optimized Interface Energy Filtering of AlphaFold3-Predicted Complex 1. AlphaFold3 can generate near-experimental protein-complex structures, but for antibody epitope prediction it still shows a high failure rate;

[image] ProtJEPA: A Multimodal Joint-Embedding Predictive Architecture for Protein Biological World Modeling with Multi-Teacher Modality-Attentive Fusion 1 ProtJEPA trains a sequence-only student model to predict a joint “biological context” embedding distilled from 10 frozen teacher

[image] Unlocking Multimodal Protein Language Models at Inference Time 1. The paper argues that multimodal protein language models (pLMs) are often judged by training alone, but their real-world performance can hinge on inference-time sampling. It shows default decoding protocols can be

[image] Evolutionary Profiles for Protein Fitness Prediction 1 They introduce EvoIF, a lightweight protein fitness predictor that explicitly fuses two complementary evolutionary signals: (i) within-family evolutionary profiles from retrieved homologs (sequence- or Foldseek

[image] Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design 1. The paper targets a practical bottleneck in de novo protein binder design: upstream pipelines can generate thousands of candidates, but wet-lab capacity is small, so the key problem becomes how

[image] Assessment of generative de novo peptide design methods for G protein-coupled receptors 1. Junker & Schoeder benchmark deep learning pipelines for designing GPCR-binding peptides, asking a practical question: are failures mainly due to inadequate sampling (generation) or

[image] CodonMamba: A Foundation Model for Programmable mRNA Coding Sequence Design 1. CodonMamba reframes codon optimization as an inference-time steerable generation problem: the pretrained model provides a general CDS distribution, while a user-specified codon-usage prior can be

[image] On the Generalization and Usability of Cofolding Models for GPCR Drug Discovery 1. The study benchmarks Boltz (Boltz-1x), a diffusion-based protein–ligand “co-folding” model, on a curated set of 253 ligand-bound human GPCR structures from 74 GPCRdb families that were unseen

[image] PpigFinder: An integrated desktop application for bacterial genome annotation and AlphaFold 3-based protein–protein interaction screening 1. ppigFinder is a cross-platform desktop GUI that turns a raw bacterial genome (nucleotide sequence) into a structured workflow for

[image] CLDN18.2 Antibody Design with Protein Language Models: A Deep Learning Optimization Framework 1. Qu et al. present cdrGPT, a GPT-2–style antibody language-model framework that designs new heavy-chain CDRH3 loops for CLDN18.2 antibodies while simultaneously optimizing predicted

[image] AlphaFolding: 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance 1. AlphaFolding introduces a 4D (space + time) score-based diffusion model that predicts an entire protein trajectory at once: up to 32 time steps simultaneously for proteins

[image] Accurate and efficient prediction of protein conformations with ProtMonomer 1. ProtMonomer is built around a simple training insight: protein structure predictors trained under different MSA-depth distributions (i.e., different levels of evolutionary information) generalize

[image] HInt: Interaction-based homology discovery through accelerated genome-scale AlphaFold screening 1 Rouger et al. propose “interaction-based similarity” as a third axis for homology detection, complementary to sequence (e.g., BLAST) and structure (e.g., Foldseek), targeting cases

[image] Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling 1. This review reframes protein structure prediction as a sequence of methodological shifts rather than a simple timeline of “better models”, organizing the field into four phases and three

[image] RegimeFormer: A Large Protein Model of Global Perturbation Regimes 1. The paper frames mutation effects as a global, protein-level “perturbation regime” coordinate—intended to explain why the same substitution class can be tolerated in one protein but disruptive in another—then

[image] SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign 1 SimpleDesign proposes a minimalist alternative to structure-tokenizer pipelines: a single Transformer jointly models discrete amino-acid sequences and continuous 3D Cα coordinates directly in data space,

[image] Boltz-Perturb: Improving Diversity and Accuracy in Protein-Ligand Co-Folding through Training-Free Conditioning Perturbation 1. The paper argues that many protein–ligand co-folding failures are not purely “model can’t represent the right pose”, but “sampling doesn’t reach it”: a