The open-source ecosystem for computational drug discovery has matured dramatically. Researchers today have access to AI models and cheminformatics tools that rival — and in some cases surpass — commercial alternatives. This guide covers the most impactful open-source tools available in 2025, organized by drug discovery pipeline stage.

A newer edition of this survey is available: Best Open-Source Computational Drug Discovery Tools in 2026, with licences and releases re-checked tool by tool.

Protein Structure Prediction#

ESMFold (Meta AI)#

What it does: Predicts protein 3D structure from a single amino acid sequence using a protein language model. No multiple sequence alignment (MSA) required.

Speed: 5-15 seconds per structure (60x faster than AlphaFold2)
Accuracy: Median GDT-TS 87.0 on CASP14
Best for: High-throughput screening, orphan proteins, rapid triage
License: MIT
Paper: Lin et al., Science 379, 1123–1130 (2023)

AlphaFold2 (DeepMind)#

What it does: Predicts protein structure using evolutionary information from multiple sequence alignments. The gold standard for accuracy.

Speed: 5-60 minutes per structure (MSA generation dominates)
Accuracy: Median GDT-TS 92.4 on CASP14
Best for: Maximum accuracy, multi-domain proteins, publication-grade structures
License: Apache 2.0
Paper: Jumper et al., Nature 596, 583–589 (2021)

ESM3 (Meta AI)#

What it does: Multimodal protein language model that generates and reasons about protein sequences, structures, and functions simultaneously.

Best for: Protein design, functional annotation, structure-function exploration
License: Open (research), ESM3 Open for non-commercial use
Paper: Hayes et al., bioRxiv (2024)

Molecular Docking#

DiffDock (MIT)#

What it does: Diffusion-based molecular docking that predicts ligand binding poses without requiring a predefined binding site (blind docking).

Success rate: 38% top-1 (RMSD < 2 Angstroms) on PDBBind, versus 23% for the best traditional docking method and 20% for the best deep learning baseline reported in the paper
Best for: Blind docking when binding site is unknown, novel protein targets
License: MIT
Paper: Corso et al., ICLR (2023)

AutoDock Vina#

What it does: Classical grid-based docking with exhaustive search. Fast and well-validated across thousands of publications.

Best for: Known binding sites, large-scale virtual screening, well-characterized targets
License: Apache 2.0
Validation: >15,000 citations, extensive experimental validation

GNINA#

What it does: CNN-based molecular docking that uses convolutional neural networks for scoring and pose refinement.

Best for: Rescoring docking poses, when classical scoring functions underperform
License: dual-licensed under GPL and Apache (the GPL is required by its OpenBabel dependency; Apache-only use requires removing OpenBabel references from the source)
Paper: McNutt et al., Journal of Cheminformatics 13, 43 (2021)

ADMET Prediction#

ADMET-AI#

What it does: Predicts 58 pharmacokinetic and toxicity endpoints using graph neural networks with calibrated confidence scores.

Best for: Comprehensive ADMET profiling with uncertainty quantification
License: MIT
Key feature: Applicability domain warnings flag unreliable predictions

ProTox3#

What it does: Acute toxicity prediction (LD50, GHS classification) plus 33 toxicity endpoints including organ toxicity and endocrine disruption.

Best for: Detailed toxicity profiling, safety assessment
License: Academic free use

pkCSM#

What it does: Graph-based pharmacokinetic prediction covering 30 ADMET parameters.

Best for: Quick orthogonal ADMET assessment
License: Free for academic use

Molecular Generation and Design#

REINVENT4 (AstraZeneca)#

What it does: De novo molecular design using reinforcement learning with customizable scoring functions. Generates novel molecules optimized for multiple objectives simultaneously.

Best for: Multi-objective lead optimization, scaffold hopping
License: Apache 2.0
Paper: Blaschke et al., Journal of Cheminformatics 16, 20 (2024)

GenMol#

What it does: Generative molecular design using SAFE (Sequential Attachment-based Fragment Embedding) notation for fragment-aware generation.

Best for: Fragment-based generative design, library enumeration
License: MIT

Retrosynthesis#

AiZynthFinder (AstraZeneca)#

What it does: AI-powered retrosynthetic route planning. Given a target molecule, it generates multi-step synthesis trees using commercially available starting materials.

Best for: Planning practical synthetic routes, checking synthesizability
License: MIT
Paper: Genheden et al., Journal of Cheminformatics 12, 70 (2020)

Cheminformatics#

RDKit#

What it does: The foundational open-source cheminformatics toolkit. Handles molecular I/O (SMILES, SDF, MOL2), descriptor calculation, fingerprinting, substructure search, conformer generation, and chemical reactions.

Best for: Everything cheminformatics — it's the industry standard
License: BSD
Note: Used internally by most other tools on this list

Binding Site Detection#

P2Rank#

What it does: Machine learning-based prediction of ligand binding sites on protein surfaces with pocket ranking and druggability scores.

Best for: Identifying binding sites on predicted structures, druggability assessment
License: MIT/Apache 2.0
Paper: Krivák & Hoksza, Journal of Cheminformatics 10, 39 (2018)

Single-Cell and Genomics#

Scanpy (Theis Lab)#

What it does: Complete single-cell RNA-seq analysis pipeline — quality control, normalization, dimensionality reduction, clustering, differential expression, and trajectory inference.

Best for: scRNA-seq analysis, cell type annotation, pathway analysis
License: BSD 3-Clause
Paper: Wolf et al., Genome Biology 19, 15 (2018)

scVI-tools#

What it does: Probabilistic models for single-cell omics data — variational autoencoders for batch correction, imputation, differential expression, and multi-modal integration.

Best for: Batch effect removal, integrating multiple datasets, probabilistic analysis
License: BSD 3-Clause
Paper: Gayoso et al., Nature Biotechnology 40, 163–166 (2022)

Biomedical NLP#

BioBERT#

What it does: Pre-trained biomedical language model for named entity recognition (NER), relation extraction, and question answering on scientific text.

Best for: Extracting genes, proteins, drugs, and diseases from literature
License: Apache 2.0
Paper: Lee et al., Bioinformatics 36, 1234–1240 (2020)

Putting It All Together#

The challenge with open-source tools isn't availability — it's integration. Each tool has different input formats, dependencies, GPU requirements, and output schemas. A typical drug discovery workflow might chain 5-10 tools across protein structure, docking, ADMET, and design.

Ivy Biosciences addresses this by orchestrating 80+ open-source tools into a unified platform where researchers can chain analyses through a chat interface or CLI, without managing infrastructure or format conversions.

Key Takeaways#

Open-source drug discovery tools have reached production quality — ESMFold, DiffDock, and ADMET-AI compete with commercial offerings
The biggest bottleneck is no longer tool availability but integration and accessibility
A practical drug discovery workflow chains multiple tools: structure prediction, docking, ADMET screening, molecular generation, and retrosynthesis
All tools listed here are freely available and actively maintained as of 2025
Platforms like Ivy Biosciences, Galaxy, and Knime provide integration layers to connect these tools into workflows

References#

Lin, Z. et al. "Evolutionary-scale prediction of atomic-level protein structure with a language model." Science 379, 1123–1130 (2023). doi:10.1126/science.ade2574
Jumper, J. et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2
Corso, G. et al. "DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking." ICLR (2023). arXiv:2210.01776
McNutt, A.T. et al. "GNINA 1.0: molecular docking with deep learning." Journal of Cheminformatics 13, 43 (2021). doi:10.1186/s13321-021-00522-2
Swanson, K. et al. "ADMET-AI: A machine learning ADMET platform." Bioinformatics 40, btae416 (2024). doi:10.1093/bioinformatics/btae416
Genheden, S. et al. "AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning." Journal of Cheminformatics 12, 70 (2020). doi:10.1186/s13321-020-00472-1
Blaschke, T. et al. "REINVENT 4: Modern AI-driven generative molecule design." Journal of Cheminformatics 16, 20 (2024). doi:10.1186/s13321-024-00812-5
Wolf, F. et al. "SCANPY: large-scale single-cell gene expression data analysis." Genome Biology 19, 15 (2018). doi:10.1186/s13059-017-1382-0
Lee, J. et al. "BioBERT: a pre-trained biomedical language representation model." Bioinformatics 36, 1234–1240 (2020). doi:10.1093/bioinformatics/btz682