Find open-source science resources

Diffusion-based molecular docking achieving SOTA blind docking performance, treating ligand pose prediction as generative diffusion over SE(3), with DiffDock-L update for improved generalization (MIT CSAIL, ICLR 2023)

This is https://huggingface.co/kingabzpro/Qwen-3-32B-Medical-Reasoning applied to https://huggingface.co/Qwen/Qwen3-32B Original model card created by @kingabzpro

Detects Gene Ontology and/or other user defined categories which are over/under represented in RNA-seq data.

Comprehensive collection of papers on unifying LLMs and knowledge graphs

A spline based scRNA-seq method for identifying differentially variable (DV) genes across two experimental conditions. Spline-DV constructs a 3D spline from 3 key gene statistics: mean expression, coefficient of variance, and dropout rate. This is done for both conditions. The 3D spline provides the “expected” behavior of genes in each condition. The distance of the observed mean, CV and dropout rate of each gene from the expected 3D spline is used to measure variability. As the final step, the spline-DV method compares the variabilities of each condition to identify differentially variable (DV) genes.

GP-MoLFormer is a class of models pretrained on SMILES string representations of 0.65-1.1B molecules from ZINC and PubChem. This repository is for the model pretrained on all the unique molecules from both datasets.

Model Repo: xformai/qwen-0.6b-mentalhealth-support Base Model: Qwen/Qwen-0.5B Task: Empathetic Conversational AI for mental health & emotional support Fine-Tuned By: XformAI

Unified benchmarking framework for protein representation learning, providing standardized interfaces for pre-training and diverse downstream tasks including structure prediction, fitness prediction, and property prediction across multiple protein datasets and model architectures (ICLR 2024, 273+ stars, MIT License)

### Note: This model is copied version of DNABERT-2 which removes the FlashAttention integration with Trition. This allows the model to be installed off HuggingFace without having to uninstall Triton. Running the below example code yields identical output compared to the original verison.

## Model Description EYE-Llama_gqa is a large language model specifically designed for ophthalmic question-answering (QA). It is built upon the Llama 2 architecture and fine-tuned on a the EYE-lit and EYE-QA+ dataset.

## Overview This project focuses on curating and modeling bioactivity data of small molecules targeting immune receptors. Using datasets from ImmtorLig_DB, we applied machine learning techniques to predict interactions between small molecules and immune receptors or cytokines, aiding drug discovery…

SEraster is a rasterization preprocessing framework that aggregates cellular information into spatial pixels to reduce resource requirements for spatial omics data analysis. SEraster reduces the number of spatial points in spatial omics datasets for downstream analysis through a process of rasterization where single cells’ gene expression or cell-type labels are aggregated into equally sized pixels based on a user-defined resolution. SEraster is built on an R/Bioconductor S4 class called SpatialExperiment. SEraster can be incorporated with other packages to conduct downstream analyses for spatial omics datasets, such as detecting spatially variable genes.

Long-range genomic foundation model using subquadratic Hyena operators instead of Transformer attention, enabling context lengths up to 1 million nucleotides for chromosome-scale DNA sequence modeling and downstream genomics tasks (Stanford Hazy Research, NeurIPS 2023, 784+ stars, Apache 2.0)

This package provides functions for differential chromatin interaction analysis between two single-cell Hi-C data groups. It includes tools for imputation, normalization, and differential analysis of chromatin interactions. The package implements pooling techniques for imputation and offers methods to normalize and test for differential interactions across single-cell Hi-C datasets.

本模型基于 trueto/medbert-base-chinese 预训练模型，结合 BiLSTM 和 CRF 构建而成，用于中文医疗命名实体识别（NER）任务。

XAItest is an R Package that identifies features using eXplainable AI (XAI) methods such as SHAP or LIME. This package allows users to compare these methods with traditional statistical tests like t-tests, empirical Bayes, and Fisher's test. Additionally, it includes simThresh, a system that enables the comparison of feature importance with p-values by incorporating calibrated simulated data.

Implements bindings for SQL tables that are compatible with Bioconductor S4 data structures, namely the DataFrame and DelayedArray. This allows SQL-derived data to be easily used inside other Bioconductor objects (e.g., SummarizedExperiments) while keeping everything on disk.

Package designed to aid in classifying cells from single-cell RNA sequencing data using external reference data (e.g., bulk RNA-seq, scRNA-seq, microarray, gene lists). A variety of correlation based methods and gene list enrichment methods are provided to assist cell type assignment.

Starting from one SBML file, it extracts information from each listOfCompartments, listOfSpecies and listOfReactions element by saving them into data frames. Each table provides one row for each entity (i.e. either compartment, species, reaction or speciesReference) and one set of columns for the attributes, one column for the content of the 'notes' subelement and one set of columns for the content of the 'annotation' subelement.

SpatialDE is a method to find spatially variable genes (SVG) from spatial transcriptomics data. This package provides wrappers to use the Python SpatialDE library in R, using reticulate and basilisk.

An R package that offers a workflow to predict condition-specific enhancers from ChIP-seq data. The prediction of regulatory units is done in four main steps: Step 1 - the normalization of the ChIP-seq counts. Step 2 - the prediction of active enhancers binwise on the whole genome. Step 3 - the condition-specific clustering of the putative active enhancers. Step 4 - the detection of possible target genes of the condition-specific clusters using RNA-seq counts.

An ontology encoding the core metadata based on the Intelligent Information Request and Delivery (iiRDS) standard

Mass cytometry enables the simultaneous measurement of dozens of protein markers at the single-cell level, producing high dimensional datasets that provide deep insights into cellular heterogeneity and function. However, these datasets often contain unwanted covariance introduced by technical variations, such as differences in cell size, staining efficiency, and instrument-specific artifacts, which can obscure biological signals and complicate downstream analysis. This package addresses this challenge by implementing a robust framework of linear models designed to identify and remove these sources of unwanted covariance. By systematically modeling and correcting for technical noise, the package enhances the quality and interpretability of mass cytometry data, enabling researchers to focus on biologically relevant signals.

This model card describes the ClinicalBERT model, which was trained on a large multicenter dataset with a large corpus of 1.2B words of diverse diseases we constructed. We then utilized a large-scale corpus of EHRs from over 3 million patient records to fine tune the base language model.

CARD is a reference-based deconvolution method that estimates cell type composition in spatial transcriptomics based on cell type specific expression information obtained from a reference scRNA-seq data. A key feature of CARD is its ability to accommodate spatial correlation in the cell type composition across tissue locations, enabling accurate and spatially informed cell type deconvolution as well as refined spatial map construction. CARD relies on an efficient optimization algorithm for constrained maximum likelihood estimation and is scalable to spatial transcriptomics with tens of thousands of spatial locations and tens of thousands of genes.

This is the full precision (f16) GGUF version of a model trained for medical chatbot and dental implant assistant tasks. It combines general doctor–patient dialogue understanding with domain-specific Q&A derived from Straumann® dental implant system manuals.

signifinder is an R package for computing and exploring a compendium of tumor signatures. It allows to compute a variety of signatures coming from public literature, based on gene expression values, and return single-sample (-cell/-spot) scores. Currently, signifinder collects more than 70 distinct signatures, relating to multiple tumors and multiple cancer processes.

Efficient foundation models for PDEs with pretrained transformer-based neural operators and downstream task fine-tuning pipelines, HuggingFace integration for models and datasets (ETH Zurich CAMLab, arXiv 2024)

This project fine-tunes the meta-llama/Llama-4-Scout-17B-16E-Instruct model using a medical reasoning dataset (FreedomIntelligence/medical-o1-reasoning-SFT) with 4-bit quantization for memory-efficient training.

Covering 12 Major Languages including English, Chinese, French, Hindi, Spanish, Arabic, Russian, Japanese, Korean, German, Italian, Portuguese and 38 Minor Languages So far.

Strand specific peak-pair calling in ChIP-exo replicates. The cumulative Skellam distribution function is used to detect significant normalised count differences of opposed sign at each DNA strand (peak-pairs). Then, irreproducible discovery rate for overlapping peak-pairs across biological replicates is computed.

Computational evaluation of variability across DNA or RNA sequencing datasets is a crucial step in genomics, as it allows both to evaluate reproducibility of replicates, and to compare different datasets to identify potential correlations. fCCAC applies functional Canonical Correlation Analysis to allow the assessment of: (i) reproducibility of biological or technical replicates, analyzing their shared covariance in higher order components; and (ii) the associations between different datasets. fCCAC represents a more sophisticated approach that complements Pearson correlation of genomic coverage.

Supplies AnnotationHub with MassBank metabolite/compound annotations bundled in CompDb SQLite databases. CompDb SQLite databases contain general compound annotation as well as fragment spectra representing fragmentation patterns of compounds' ions. MassBank data is retrieved from https://massbank.eu/MassBank and processed using helper functions from the CompoundDb Bioconductor package into redistributable SQLite databases.

!zxdfdsxf.png

Protein solubility is a critical factor in both pharmaceutical research and production processes, as it can significantly impact the quality and function of a protein. This is an example for finetuning ibm/biomed.omics.bl.sm-ted-458m for protein solubility prediction (binary classification) based…

PyTorch implementation of neural ODEs

!11.png

High performance functions for row and column operations on sparse matrices. For example: col / rowMeans2, col / rowMedians, col / rowVars etc. Currently, the optimizations are limited to data in the column sparse format. This package is inspired by the matrixStats package by Henrik Bengtsson.

LLM-based molecular optimization tool

Pathview is a tool set for pathway based data integration and visualization. It maps and renders a wide variety of biological data on relevant pathway graphs. All users need is to supply their data and specify the target pathway. Pathview automatically downloads the pathway graph data, parses the data file, maps user data to the pathway, and render pathway graph with the mapped data. In addition, Pathview also seamlessly integrates with pathway and gene set (enrichment) analysis tools for large-scale and fully automated analysis.

Functions to reconstruct case and control AFs from summary statistics. One function uses OR, NCase, NControl, and SE(log(OR)). The second function uses OR, NCase, NControl, and AF for the whole sample.

!fffffff.png

General-purpose pathology foundation model pretrained on 100K+ diagnostic whole-slide images across 20 major tissue types, achieving state-of-the-art transfer learning across 30+ clinical tasks and serving as a universal feature extractor for digital pathology (Mahmood Lab, 722+ stars)

Vision-language pathology foundation model using contrastive learning on histopathology image-text pairs, enabling zero-shot classification, slide-level retrieval, and multimodal reasoning across diverse cancer types (Mahmood Lab, 494+ stars)

Biofactoid is a web-based system that empowers authors to capture and share machine-readable summaries of molecular-level interactions described in their publications.