AI-Assisted Drug Discovery

Part of the worlds-biggest-software-project initiative.

An open-source, AI-native platform that accelerates small-molecule drug discovery by unifying generative molecular design, ADMET prediction, virtual screening, and literature mining in a single accessible workflow.

Drug discovery costs an average of $2.6 billion per approved drug and takes 10-15 years, with only ~10% of candidates surviving from Phase I to approval. Most failures stem from poor target selection and late-stage toxicity surprises that could have been caught computationally. This project builds an integrated, open-source platform that applies generative models, graph neural networks, and large-scale literature mining to compress early discovery timelines and make AI-driven drug design accessible beyond the handful of organisations that can afford $500K-$5M annual platform licences.

Why AI-Assisted Drug Discovery?

Prohibitive cost of incumbents. Commercial platforms like Schrodinger ($500K-$5M enterprise licences), Insilico Medicine, and Recursion OS are accessible only through expensive contracts or formal partnerships, shutting out academic labs, small biotechs, and researchers in developing countries.
No open-source full-stack alternative exists. Open tools (RDKit, DeepChem, REINVENT4, OpenFold3) each solve one piece of the puzzle, but assembling them into an end-to-end workflow requires significant engineering effort that most research teams cannot afford.
Predictions without uncertainty are dangerous. Most platforms output point estimates for molecular properties. Multi-task ADMET models drop from 85-90% accuracy on benchmark scaffolds to 70-75% on novel ones, yet few tools surface calibrated confidence intervals to flag unreliable predictions.
Fragmented workflows lose signal. Literature mining, generative design, ADMET prediction, and synthesis planning typically live in separate tools with no data flow between them, forcing researchers into manual copy-paste cycles and increasing the risk of missed insights.
Regulatory documentation is an afterthought. FDA and EMA are developing AI credibility frameworks, but almost no existing platform produces the model provenance, uncertainty documentation, and applicability-domain summaries needed for IND/CTA submissions.

Key Features

Virtual Screening and Molecular Design

Structure-based docking against protein binding sites using AutoDock Vina or DiffDock via NVIDIA NIM microservices
De novo molecule generation via REINVENT4 with four specialised generators: de novo (RNN), R-group replacement, linker design, and molecular optimisation
Diversity filtering and drug-likeness scoring powered by RDKit
AlphaFold3/OpenFold3 integration for protein structure input when no experimental structure is available

ADMET Prediction and Property Analysis

Multi-task ML models predicting solubility, membrane permeability, metabolic stability, hERG toxicity, and CYP inhibition from molecular structure
Calibrated uncertainty quantification (conformal prediction or Bayesian approaches) to flag predictions outside the applicability domain
Molecular property calculations: molecular weight, logP, hydrogen bond donors/acceptors, TPSA, Lipinski's Rule of Five

Compound Management and Collaboration

Compound registration, deduplication, annotation, and export in SMILES/SDF format
Multi-user project workspaces with compound versioning, run history, and annotation
SAR (structure-activity relationship) analysis and visualisation

Literature Mining and Knowledge Graph

Continuous ingestion of PubMed, patent databases, and preprint servers
Entity relationship extraction: drug-target, target-disease, drug-adverse-effect
Automatic surfacing of relevant findings for a given research programme

Synthesis Planning and Lab Integration

Retrosynthetic feasibility scoring via open-source tools (ASKCOS, IBM RXN)
Synthesis routes ranked by step count, reagent availability, and estimated yield
API connectors to electronic lab notebooks (Benchling) to ingest assay results and close the active-learning loop

Regulatory and Governance

Auto-generated model provenance summaries aligned with emerging FDA AI credibility frameworks
Full reproducibility tracking: every model version, training dataset, hyperparameter, and prediction result logged
Federated model training allowing multiple organisations to contribute data without sharing raw structures

AI-Native Advantage

This platform is not a traditional cheminformatics toolkit with an AI layer bolted on. Generative molecular design sits at the core: reinforcement learning agents optimise simultaneously for target binding affinity, synthetic accessibility, and ADMET properties, producing ranked candidate lists with diversity filters. Graph neural networks and equivariant 3D models (SE(3)-Transformer, DiffDock) provide physically grounded scoring that adapts to novel scaffolds. The active-learning loop feeds wet-lab assay results back into models for iterative refinement, compressing the design-make-test-learn cycle from months to days. Literature mining with NLP continuously surfaces new evidence, ensuring that computational hypotheses stay current with the latest published research.

Tech Stack and Deployment

The platform is designed for self-hosted, cloud, or hybrid deployment. GPU-accelerated virtual screening runs on Kubernetes-based clusters with automatic spot-instance interruption handling (AWS, GCP). Core cheminformatics uses RDKit (BSD-3-Clause), generative chemistry uses REINVENT4 (Apache 2.0), deep learning property prediction uses DeepChem (MIT), and protein structure prediction uses OpenFold3 (Apache 2.0). NVIDIA BioNeMo NIM microservices provide containerised, GPU-optimised inference endpoints deployable via Docker or Kubernetes. Standard molecular formats (SMILES, SDF, MOL, PDB, FASTA) ensure interoperability with existing tools and databases (ChEMBL, PubChem, ZINC, UniProt).

Market Context

The AI in pharmaceuticals market was valued at $1.8 billion in 2023 and is projected to reach $13.1 billion by 2030 (World Economic Forum, January 2026). As of early 2026, more than 29 AI-designed compounds have entered human clinical trials, with AI-assisted Phase I trials reporting success rates of 80-90% versus the historical average of 40-65%. Commercial platforms command $500K-$5M annual licences, placing them out of reach for the vast majority of the estimated 30,000+ academic drug discovery labs, small biotechs, and researchers in low- and middle-income countries who stand to benefit most from AI-accelerated discovery.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.

Note: The platform's core open-source dependencies are permissively licensed (RDKit BSD-3-Clause, DeepChem MIT, REINVENT4 Apache 2.0, OpenFold3 Apache 2.0, NVIDIA BioNeMo Framework Apache 2.0). AlphaFold3 model weights are CC BY-NC 4.0 and cannot be used in commercial discovery pipelines without an Isomorphic Labs agreement. AI-generated molecule IP ownership remains a live legal question in most jurisdictions.