Find open-source science resources

First any-to-any generative foundation model for Earth Observation, enabling unified multimodal understanding and generation across diverse satellite sensors and geospatial tasks through a single architecture (258+ stars)

Simple and accurate de novo protein binder design pipeline using AlphaFold2 backpropagation, MPNN, and PyRosetta for automated binder discovery (bioRxiv 2024)

AlphaFold/ESMFold accessible implementation with AF3 JSON export, database updates

Benchmark evaluating AI agents for end-to-end automated research from re-discovery to new-discovery, with 40 real-science tasks across 10 disciplines, curated datasets from published papers, and expert-curated multimodal rubrics (170+ stars, MIT License)

Community-driven model zoo and deployment infrastructure for AI-powered bioimage analysis, enabling standardized sharing, validation, and cross-platform execution of deep learning models across Fiji, Ilastik, napari, and other scientific imaging tools (EPFL, EMBL, and global collaborators, actively maintained)

Phylogeny-aware genomic language model trained on whole-genome alignments across multiple evolutionary timescales, predicting functional constraints and variant effects for human, mouse, chicken, fly, worm, and Arabidopsis genomes (344+ stars, MIT License)

Frontier protein language models (300M/600M/6B) trained on billions of protein sequences, establishing a new unsupervised scaling frontier beyond ESM2 with emergent long-range structural understanding; ships with ESMFold2 structure prediction (SOTA DockQ pass-rates on Foldbench protein-protein and antibody-antigen complexes, lab-validated de novo binder/scFv design protocol) and the ESM Atlas mapping 6.8B proteins with sparse-autoencoder-interpretable world-model features (2.9K+ stars, 2025-2026)

Polymathic AI's large omnimodal foundation model for astronomical surveys, seamlessly integrating 39 distinct data modalities including imaging, spectra, photometry, and catalog entries for similarity search, property prediction, and generative modeling across legacy surveys (MIT)

Open-source deep learning toolbox for bioimage analysis providing a unified, configuration-driven framework for 2D/3D semantic segmentation, instance segmentation, classification, denoising, super-resolution, and self-supervised learning; integrates state-of-the-art architectures including U-Net, Vision Transformers, and ConvNeXt, designed for microscopy and biomedical imaging researchers without extensive coding expertise (MIT License, actively maintained)

Segment Anything Model for microscopy: interactive and automatic segmentation of light, electron, and fluorescence microscopy images in 2D and 3D, with domain-specific fine-tuning workflows for scientific imaging (1.5K+ stars)

Curated open dataset collection of 602M+ observational and perturbational single-cell profiles for accelerating virtual cell model creation, integrating Tahoe-100M and scBaseCount data with Google Cloud Marketplace distribution (Arc Institute, 2025-2026)

Multi-modal geospatial ML platform for agriculture and sustainability, fusing satellite imagery (RGB, SAR, multispectral), drone imagery, weather data, and sensor data for crop identification, carbon footprint estimation, and microclimate prediction (Microsoft Research, MIT License)

98B-parameter frontier generative model jointly reasoning over protein sequence, structure, and function, trained on 2.78 billion proteins; generated a novel fluorescent protein (esmGFP) with only 58% sequence identity to known GFPs (EvolutionaryScale, 2024)

Latent-space probabilistic denoising diffusion model for predicting coarse-grained conformational ensembles of intrinsically disordered proteins and regions from sequence, with GPU/CPU inference, trajectory export, and FAISS-based similarity search (67+ stars, LGPL-3.0)

A toolbox for machine learning in seismology, providing unified interfaces for deep learning seismic phase picking, earthquake detection, and waveform analysis across multiple benchmark datasets and pretrained models (397+ stars, actively maintained)

Molecular dynamics in JAX

Large transformer-based single-cell foundation model pretrained on 50 million cells for robust gene network inference, expression denoising, cell embedding, and zero-shot label prediction, leveraging ESM2 protein embeddings and bidirectional transformer architecture (Cantini Lab, 148+ stars, GPL-3.0)

Multimodal AI system generating virtual populations for tumor microenvironment modeling from H&E and multiplex immunofluorescence pathology images, enabling large-scale spatial analysis of cancer biology and therapeutic response prediction (Microsoft Research & Providence, 370+ stars)

15TB collection of 16 large-scale numerical simulation datasets spanning fluid dynamics, MHD, astrophysics, biological systems, and acoustic scattering, with unified PyTorch dataloaders and benchmarks for training foundation models on physical sciences (Polymathic AI, NeurIPS 2024)

Generalized Attribute Based Ratings Information Extraction Library; official OpenAI toolkit that turns messy qualitative corpora into analysis-ready datasets for social scientists and data scientists, measuring quantitative attributes in text, images, or audio using the GPT API. See the [official blog post](https://openai.com/index/scaling-social-science-research/) and [NBER working paper](http://www.nber.org/papers/w34834) (413+ stars, Apache 2.0)

Computational fluid dynamics in JAX, enabling differentiable Navier-Stokes simulations with automatic differentiation for ML-accelerated CFD research, supporting turbulence modeling, convection-diffusion, and complex boundary conditions on CPUs and GPUs (Google Research, 947+ stars)

Family of large language models for materials research via continued pretraining of LLaMA-2/3 on ~30B materials science tokens, outperforming commercial LLMs on materials science tasks while identifying "adaptation rigidity" in overtrained models; includes MatNLP benchmark and CIF crystal generation capabilities (IIT Delhi M3RG, MIT License)

DeepMind's graph neural network for materials exploration, discovering 2.2M new crystal structures (380K most stable) equivalent to 800 years of traditional research, with 520K+ materials dataset open-sourced (Nature 2023)

Meta FAIR's foundation model of vision, audition, and language for in-silico neuroscience, predicting fMRI brain responses to naturalistic multimodal stimuli (video, audio, text) through unified Transformer architecture mapped to the cortical surface (2026)

Google's open multimodal medical LLM family built on Gemma 3, including a 4B multimodal model handling medical images (radiology, pathology, dermatology) alongside text and a 27B text model for clinical reasoning; trained on de-identified medical data with checkpoints and inference code released under Apache 2.0 (1.6K+ stars)

Arc Institute's 40B-parameter genome foundation model trained on 9 trillion nucleotides from all domains of life, supporting 1M base pair context for generalist DNA/RNA/protein prediction and design (Nature 2026)

Gene expression prediction

Transformer foundation model for tandem mass spectrometry (MS/MS) self-supervised on millions of unannotated spectra from the GeMS dataset via masked peak prediction and chromatographic retention-order objectives, producing 1024-dimensional molecular representations; achieves SOTA on spectral similarity, chemical property, and molecular fingerprint prediction, and powers the DreaMS Atlas annotating 201M+ MS/MS spectra for metabolomics and natural product discovery (Pluskal Lab, IOCB Prague & MIT, 211+ stars, MIT License)

Universal time series forecasting via the UNI2TS library, training a single transformer with shared self-attention and specialized mixture-of-experts feed-forward blocks to achieve strong zero-shot generalization across heterogeneous domains including energy, weather, transportation, and health time series (1.6K+ stars, Apache 2.0, 2024-2026)

Efficient differentiable n-dimensional PDE solvers built on JAX and Equinox, shipping 46+ built-in equations with Fourier spectral methods, exponential time differencing, and full auto-differentiation for physics-based deep learning workflows (MIT, 200+ stars, 2024)

First architecture deeply integrating a DNA foundation model with an LLM for multimodal biological reasoning, achieving 98% accuracy on KEGG disease pathway prediction and 15%+ average gains on variant effect prediction with interpretable step-by-step reasoning traces (bowang-lab, 390+ stars)

Multimodal LLM-based AI agent enabling deep research in spatial transcriptomics, automating analysis and interpretation of spatial gene expression data (Harvard LiuLab, bioRxiv 2025)

Multimodal AI bridging transcriptomics data and natural language, enabling intuitive chat-based exploration and analysis of single-cell RNA-seq datasets through conversational interaction without coding; fine-tuned Mistral 7B LLaVA model emulating biologist-bioinformatician discussions (207+ stars, GPL-3.0)

Single-cell analysis with transformers

Arc Institute's single-cell foundation model enabling in-context learning at inference time via a novel tabular attention architecture, trained on 150M uniformly-preprocessed cells for generalizing biological effects and generating unseen cell profiles in novel contexts (2025)

Generative AI system for antibiotic discovery that searches billions of synthesizable molecules by combining molecular building blocks through real chemical reactions, experimentally validating novel compounds active against drug-resistant bacteria

Dataset and benchmarking framework integrating histology and spatial transcriptomics, enabling multimodal analysis of whole-slide images with matched spatial gene expression for advancing computational pathology and tissue microenvironment research (Mahmood Lab, Harvard Medical School, 411+ stars)

Google Colab-based no-code toolbox democratizing deep learning in microscopy for biologists without programming experience, enabling AI-powered image segmentation, denoising, super-resolution, and object tracking across diverse imaging modalities (Henriques Lab, 640+ stars)

Target-aware peptide design framework that treats receptor sequence and structure as context via multimodal adapter tuning of protein language models (ESMC + ProteinMPNN features), with reinforcement-learning-based 3D dynamic feedback (ESMFold structure evaluation) to suppress unrealistic peptide conformations; ships with a systematic assessment pipeline covering peptide-target affinity, structure quality, physicochemical properties, diversity, and novelty (142+ stars, Apache 2.0, 2026)

Bilingual protein language model translating between protein sequence and structure, finetuned from ProtT5-XL on 17M AlphaFoldDB structures using Foldseek's 3Di structural alphabet, enabling sequence-to-structure prediction, structure-to-sequence inverse folding, and unified protein representation learning (RostLab, 310+ stars)

Foundation models for genomics and transcriptomics pretrained on 3,000+ human genomes and 850+ diverse species, enabling chromatin accessibility prediction, splice site detection, and promoter classification across multiple model scales (InstaDeep, NVIDIA & TUM, Nature Methods 2023)

AI-human collaborative research platform where a human researcher works with a team of LLM agents via team and individual meetings to perform scientific research; demonstrated by designing new SARS-CoV-2 nanobodies with wet-lab validation

Deep equivariant generative model predicting ligand-specific protein-ligand complex structures with dynamic receptor conformational flexibility, enabling accurate docking for flexible protein targets

First fully autonomous open-ended scientific discovery system with official implementation: hypothesis→experiment→writing→review simulation (13.8K+ stars, 2024)

Generative AI framework for inverse design of 3D RNA structure and function using geometric deep learning, learning design rules from 3D structures to capture complex tertiary interactions (pseudoknots, non-canonical base pairs) with expert-level accuracy for designing functional RNAs including aptamers and ribozymes (bioRxiv 2025)

Foundation model jointly trained on single-cell and spatial transcriptomics data, enabling unified representation learning across cellular and tissue spatial contexts for cell type prediction, spatial domain inference, and cross-modal integration (theislab, bioRxiv 2024, 164+ stars)

100M-parameter foundation model pretrained on 50M+ human single-cell transcriptomes covering ~20,000 genes, achieving SOTA on gene expression enhancement, drug response and perturbation prediction (Nature Methods 2024)

Family of generative single-cell foundation models (TF-Metazoa, TF-Exemplar, TF-Sapiens) jointly modeling genes and their expression levels via expression-aware autoregressive transformers, trained on up to 112M cells across 12 species spanning 1.53 billion years of evolution; achieves robust zero-shot cell type classification across species, disease state identification in human cells, and prediction of cell-type-specific transcription factors and gene-gene regulatory relationships, pip-installable with pretrained weights (CZI, 166+ stars, MIT License)

Teaching Large Language Models the Language of Biology through single-cell transcriptomics (ICML 2024)

End-to-end deep learning approach for RNA tertiary structure prediction with a flexible nucleobase center representation, achieving ~7 Å C1' RMSD across test RNAs and predicting ~545,000 structures covering 2,200+ RNA families (Kihara Lab, Purdue University, 50+ stars)