Loading BioAtlas...
Curated milestones, foundation model releases, clinical readouts, and breakthroughs across biological and pharmaceutical AI over the past 6 months.
Google DeepMind and Isomorphic Labs have released the open-source model weights and inference code for AlphaFold 3 on GitHub for non-commercial academic research. The release allows researchers worldwide to predict the joint 3D structures of proteins, DNA, RNA, ligands, and chemical modifications on local compute infrastructure.
EvolutionaryScale announced the general availability of its 98B-parameter generative biology model, ESM3, on Amazon SageMaker JumpStart and NVIDIA BioNeMo NIM microservices. The model reasons over sequence, structure, and function simultaneously, enabling de novo generation of complex proteins including novel fluorescent variants.
Chai Discovery released Chai-1, a multi-modal foundation model that predicts molecular structures across proteins, nucleic acids, small-molecule ligands, and post-translational modifications. Benchmark evaluations demonstrate a 77% success rate on protein-ligand complexes without requiring multiple sequence alignments (MSAs).
Valence Labs, in partnership with Recursion Pharmaceuticals and MIT, released Boltz-1 under a permissive MIT license. Boltz-1 matches commercial co-folding performance for multi-chain assemblies, nucleic acid complexes, and covalent protein-ligand interactions while running efficiently on consumer and cloud GPUs.
Insilico Medicine presented topline Phase IIa clinical trial data for Rentosertib (ISM001-055), a first-in-class small molecule designed from target discovery to molecular structure using its generative AI platform Pharma.AI. The drug demonstrated significant forced vital capacity (FVC) improvements and favorable tolerability in patients with idiopathic pulmonary fibrosis (IPF).
Profluent Bio published the design and experimental characterization of OpenCRISPR-1, a completely synthetic CRISPR-Cas9 endonuclease generated by a large language model trained on millions of CRISPR operons. The AI-designed enzyme demonstrates on-target human genome editing efficiency comparable to wild-type SpCas9 with reduced off-target cleavage.
Recursion Pharmaceuticals and Exscientia completed their strategic business combination, uniting Recursion's high-content phenomics and automated wet labs with Exscientia's AI precision chemistry and active learning engine. The combined pipeline includes multiple clinical-stage oncology and immunology programs.
David Baker's laboratory at the University of Washington Institute for Protein Design (IPD) released RFdiffusion All-Atom (RFdiffusionAA). The generative diffusion architecture can co-design protein backbones and side chains around arbitrary non-protein ligands, opening new avenues for de novo enzyme engineering and biosensor development.
NVIDIA expanded its BioNeMo platform with cloud-native NIM microservices for AlphaFold2, ESM-2, DiffDock, and MolMIM. Optimized for NVIDIA Blackwell and Hopper architectures, the microservices reduce structural inference and docking latency by up to 10-fold for biopharma enterprise deployments.
BioMap announced the successful completion of initial computational design milestones in its collaboration with Sanofi. Utilizing the xTrimo multi-scale foundation model suite, BioMap generated multi-specific antibody candidates with optimized affinity and developability profiles across high-priority immunology targets.
OpenAI announced a cooperative research partnership with Los Alamos National Laboratory (LANL) to benchmark multimodal frontier models on complex biological laboratory protocols, synthetic biology evaluation, and biosafety verification frameworks.
10x Genomics began commercial shipments of the Xenium Prime 5K Pan-Tissue assay, capable of measuring 5,000 RNA targets simultaneously in intact FFPE and fresh-frozen tissue sections at subcellular resolution, integrated with AI-driven cell segmentation algorithms.
Schrödinger released its Suite 2026 platform update, featuring seamless integration between machine-learning QSAR models and free energy perturbation (FEP+). The hybrid approach enables screening of billions of virtual compounds while maintaining chemical-accuracy free energy calculations for lead optimization.
Google DeepMind published AlphaProteo, an AI system trained on structural and sequence databases that designs novel protein binders for target proteins with nanomolar and picomolar affinity on the first experimental round, outperforming traditional directed evolution methods.
Ginkgo Bioworks and Google Cloud reported training milestones on massive biological sequence models leveraging Ginkgo's proprietary metagenomic repository of over 2 billion enzymatic sequences. The partnership also launched Ginkgo Datapoints for standardized experimental assay licensing.
Owkin and the EU OPTIMA consortium deployed federated learning algorithms across European clinical institutions to analyze multimodal electronic health records and digital pathology slides for personalized prostate, breast, and lung cancer treatment selection without moving patient data.
Deepcell introduced software and hardware updates for the REM-I benchtop platform, combining high-resolution brightfield single-cell imaging with self-supervised Vision Transformers for label-free real-time cellular phenotyping and sorting.
Illumina deployed DRAGEN 4.3 on NovaSeq X Plus systems, adding transformer-based secondary analysis for liquid biopsy and low-frequency somatic variant detection with 99.8% precision at 50x sequencing depth.
Cradle announced an expanded funding round to accelerate its generative protein engineering software. The platform uses predictive machine learning to design enzymes, vaccines, and engineered proteins with dramatically fewer experimental screening rounds.
Basecamp Research unveiled BaseFold, an AI protein structure prediction model trained on uncultured biodiversity from extreme global environments. BaseFold achieves superior accuracy on non-canonical enzyme active sites and extremophile proteins compared to standard AlphaFold2 baselines.
Benchling introduced Bioprocess AI, embedding predictive models directly into its Electronic Lab Notebook (ELN) to forecast bioreactor yields, optimize chromatography purification parameters, and auto-populate regulatory filing dossiers.
Generate:Biomedicines shared Phase I clinical trial data demonstrating safety and sustained pharmacokinetic profiles for GB-0669, a computationally designed monoclonal antibody targeting conserved regions of viral spike glycoproteins generated by the Chroma architecture.
Iambic Therapeutics published NeuralPLexer-2, an end-to-end deep generative model that predicts protein-ligand complex 3D coordinates while explicitly capturing induced-fit structural rearrangements across target binding pockets.
Moderna integrated deep learning models to predict mRNA secondary structure, ribosome elongation speed, and uridine depletion, resulting in therapeutic mRNA candidates with doubled protein expression half-lives in vivo.
Terray Therapeutics announced its tNova platform has generated over 5 billion physical ligand-target binding interaction data points, feeding closed-loop generative chemistry algorithms for small-molecule autoimmune targets.
Charm Therapeutics shared experimental validation of DragonFold, a deep learning platform that co-folds small molecules and flexible oncogenic targets simultaneously, identifying novel cryptic allosteric pockets on oncogenic transcription factors.
Relation Therapeutics published an atlas of human bone and joint disease mechanisms, combining single-cell spatial transcriptomics, patient tissue biopsies, and geometric graph neural networks to pinpoint target candidates for osteoarthritis.
Mila (Quebec AI Institute) and Valence Discovery updated the open-source TorchDrug framework, adding equivariant graph neural network backbones, automated retrosynthesis planning, and 3D molecular generation modules.
Hugging Face established a dedicated biological machine learning initiative, introducing native support for FASTA, FASTQ, SMILES, and mmCIF formats along with standardized model evaluation pipelines for protein language models.
Element Biosciences commenced commercial distribution of AVITI24, a benchtop system combining avidity sequencing with in situ multiomic cell profiling to measure RNA, protein, and cell morphology on a single flow cell.
Oxford Nanopore Technologies released Dorado 0.8, featuring neural network basecallers that detect m6A, pseudouridine, and 5mC modifications in direct RNA and DNA sequencing reads with single-molecule resolution.
PacBio expanded installations of its Onso sequencing system, utilizing sequencing-by-binding (SBB) chemistry to achieve 90% of bases at Q40 (99.99% accuracy) for ultra-sensitive minimal residual disease (MRD) monitoring.
Tempus AI received FDA clearance for an intelligent clinical algorithm that analyzes comprehensive genomic profiling data alongside electronic medical records to identify targeted therapy options for solid tumors.
Paige AI and Microsoft Research detailed Paige-Virchow 2, an optical foundation model trained on 3.1 million whole-slide images across 40 tissue types, achieving state-of-the-art diagnostic performance in rare tumor identification.
Dyno Therapeutics reported preclinical milestone completion in its partnership with Roche, designing engineered AAV vectors with enhanced central nervous system (CNS) tropism and reduced hepatic sequestration using generative sequence models.
Mammoth Biosciences unveiled ultra-compact CRISPR nucleases under 500 amino acids in size discovered through neural sequence search across uncharacterized microbial metagenomes, facilitating single-AAV delivery for in vivo editing.
Enveda Biosciences profiled over 1 million plant metabolites using high-resolution tandem mass spectrometry and deep learning, identifying novel natural product chemotypes with potent anti-inflammatory activity.
PostEra announced performance milestones in its partnership with Pfizer, using its Manifold synthesis platform to evaluate synthetic accessibility and design multi-step robotic synthesis routes for medicinal chemistry campaigns.
Relay Therapeutics shared Phase I/II dose expansion data for RLY-2608, a mutant-selective allosteric PI3Kα inhibitor designed using molecular dynamics and machine learning to avoid hyperinsulinemia associated with orthosteric inhibitors.
Atomwise announced the completion of 200 academic target screening projects using AtomNet, discovering novel chemical starting points with sub-micromolar affinity against historically challenging protein interfaces.
The Chan Zuckerberg Initiative (CZI) announced the activation of a high-performance compute cluster equipped with 1,000 GPUs to train predictive multimodal AI models simulating the behavior of healthy and diseased human cells.
The Broad Institute and industry partners in the Cell Painting Gallery made a petabyte-scale collection of high-content microscopy images publicly available, providing benchmark data for representation learning in chemical biology.
The OpenFold consortium released OpenFold 2.0, providing fully trainable PyTorch code and weights matching AlphaFold2 performance while introducing modular components for training on custom cryo-EM and proprietary structural datasets.
MIT researchers released DiffDock-L, an updated diffusion-based docking framework that scales to large multi-domain proteins while reducing runtime by 75% compared to original DiffDock implementations.
RosettaCommons announced the release of PyRosetta 2026, incorporating CUDA-accelerated pairwise residue potential evaluations and seamless interoperability with PyTorch and JAX tensor representations.
PathAI expanded enterprise deployments of AISight, its cloud-based digital pathology workflow platform, incorporating AI algorithms for breast biomarker quantification and automated tissue quality assessment.
Microsoft Research published MatterGen, a diffusion-based generative model for designing inorganic crystalline structures and metal-organic frameworks with user-specified mechanical, optical, and chemical properties.
Microsoft Research introduced BioGPT-2, a domain-specialized transformer model trained on PubMed abstracts and full-text biomedical literature, achieving superior accuracy on clinical question answering and gene-disease association extraction.
Anthropic rolled out domain-specific enhancements to Claude 3.5 Sonnet, adding native support for interpreting 2D chemical structure diagrams, validating SMILES strings, and assisting in biochemistry protocol troubleshooting.
The NIH announced public availability of initial multimodal datasets generated under the Bridge2AI program, spanning functional genomics, voice biomarkers, and critical care physiological data designed specifically for machine learning.
Recursion open-sourced Phenom-Beta, a vision transformer trained on millions of fluorescent cellular microscopy images from the RxRx3 dataset, enabling researchers to extract high-dimensional biological representations of chemical perturbations.
The AlphaFold Protein Structure Database, maintained by EMBL-EBI and Google DeepMind, reached a milestone of 250+ million 3D model predictions, serving over 1.8 million researchers across 190 countries.
NVIDIA and Amazon Web Services announced an expanded joint initiative providing TechBio startups with dedicated Blackwell GPU clusters and pre-packaged BioNeMo development environments for large-scale biomolecular training.
Rosalind introduced new cohort analysis modules within its cloud genomics platform, automating cross-dataset differential expression, pathway enrichment, and immune cell deconvolution for translational oncology researchers.
The Bio-IT World Conference & Expo 2026 in Boston convened leaders across computational biology, spotlighting end-to-end closed-loop discovery where generative algorithms design molecules directly synthesized by robotic cloud laboratories.
