> Markdown version of [/videos/264-geometric-deep-learning-for-drug-discovery?t=495](https://www.wearedevelopers.com/videos/264-geometric-deep-learning-for-drug-discovery?t=495). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Geometric deep learning for drug discovery

How do you search a chemical space of 10^60 to cure Alzheimer's? See how engineers leverage geometric deep learning and graph neural networks to generate protein-destroying molecules from scratch.

- **Speakers:** Noah Weber
- **Event:** WeAreDevelopers LIVE
- **Published:** October 13, 2021
- **Duration:** 42:56
- **URL:** https://www.wearedevelopers.com/videos/264-geometric-deep-learning-for-drug-discovery

## Summary

Around 80% of pathogenic proteins linked to devastating conditions like Alzheimer's and Parkinson's remain "undruggable" today, as traditional interventions like symptom inhibition or gene editing are often ineffective or applied too late. A revolutionary solution lies in targeted protein degradation (TPD), an approach that shifts the paradigm from simple inhibition to actively eliminating the root cause of the disease. TPD works by designing a "linker" molecule that forcibly bridges an E3 ligase with the pathogenic protein of interest, triggering cellular destruction of the pathogen. However, manually discovering the precise multiparty interactions required for this phenomenon within a vast chemical space of 10^60 potential molecules is mathematically and financially unfeasible for traditional wet labs.

To automate and accelerate this process, pioneering researchers leverage geometric deep learning and graph neural networks (GNNs). Because non-euclidean graph structures serve as the most native way to represent molecular geometries, atomic nodes, and intersecting bonds, tools like PyTorch Geometric allow teams to model biological constraints authentically. The machine learning pipeline utilizes AlphaFold-derived 3D protein structures and breaks immense combinatorial problems into smaller pairwise predictions. From there, active learning constraints and Bayesian optimization loops systematically parse the high-dimensional chemical space without relying on brute-force computation. Instead of merely scanning existing databases, generative deep learning models can design de novo linker molecules out of thin air, optimizing simultaneously for minimal interaction energy and strict pharmacokinetic safely requirements like non-toxicity.

Processing these high-dimensional, active-learning datasets requires a highly customized cloud infrastructure. To manage massive computational overhead efficiently, teams construct automated spot-instance fleets on AWS that actively serialize model weights to persistent EBS volumes, ensuring that long-running GPU training survives random node terminations. Yet, pushing these boundaries requires more than just clever architecture. From a company-building perspective, solving complex AI life-science problems mandates aggressively interdisciplinary talent acquisition—uniting PhD chemists, biologists, physicists, and software engineers. Because domain validation is the only way to prove computational hypotheses, building a culture of cross-discipline communication is essential, reinforcing the reality that "he who owns the data, owns the AI."

**Keywords:** targeted protein degradation, geometric deep learning, graph neural networks, active learning pipelines, bayesian optimization algorithms, de novo molecule generation, non-euclidean data modeling, pytorch geometric, molecular graph representations, computational drug discovery, alphafold 3d structures, high-dimensional chemical space, aws spot instance orchestration, interdisciplinary talent acquisition, computational cheminformatics

## Chapters

1. **Challenges of treating diseases with undruggable proteins** (00:08) — Understanding why current pathogenic protein inhibition and DNA therapies often arrive too late or lack efficacy.
1. **Utilizing targeted protein degradation for therapeutic intervention** (02:44) — How connecting E3 ligase and a protein of interest with bridging molecules leads to successful degradation.
1. **Representing molecular interactions through non-Euclidean graph data** (06:06) — Graph architectures natively capture the complexities of inter-atomic states by using nodes and multiple edge types.
1. **Applying geometric deep learning to generalized graph operations** (08:15) — Generalized convolutional frameworks like PyTorch Geometric allow localized operations across spaces without directional Euclidean constraints.
1. **Structuring an end-to-end predictive pipeline for molecule pairing** (13:21) — Incorporating biological constraint calculations into predictions of how proteins and ligands establish initial bonding pairs.
1. **Exploring multidimensional spatial poses via Bayesian optimization** (18:29) — Smart parsing through complex molecule rotations limits excessive force-computation while discovering optimal relative spatial orientations.
1. **Validating probabilistic linker suggestions with chemoinformatics simulations** (21:29) — Using density functional theory tests to guarantee high dimensional probability rankings map to authentic physics constraints.
1. **Generating novel bridging molecules with constrained property models** (25:31) — Creating de novo linkers that satisfy strict graph metrics for acceptable toxicity limits and minimized positional energy.
1. **Orchestrating an AWS cloud architecture for spot instances** (28:04) — Serializing neural network architectures on EBS volumes guarantees resilient processing when running high-computation active learning batches.
1. **Building interdisciplinarity for viable biophysics machine learning** (31:07) — Why unifying biochemists, mathematicians, and engineers ensures predictive models represent accurate reality rather than isolated computational artifacts.
1. **Addressing technical strategies and interdisciplinary computing dynamics** (34:04) — Exploring questions on cloud provider deployment logistics, interaction energy formulation, and applying deep learning toward specific disease metrics.

## Related Moments

- [Leveraging large quantitative models for intermediate quantum advantage](https://www.wearedevelopers.com/videos/100314-quantum-computing-separating-the-breakthroughs-from-the-buzz) (from "Quantum Computing: Separating the Breakthroughs from the Buzz")
- [Solving optimization and simulation problems with nature's efficiency](https://www.wearedevelopers.com/videos/100293-quantum-computing-how-does-it-work-how-will-it-affect-us-and-when) (from "Quantum Computing - How Does It Work, How Will it Affect Us and When?")
- [Designing an automated and orchestrated machine learning target process](https://www.wearedevelopers.com/videos/185-effective-machine-learning-managing-complexity-with-mlops) (from "Effective Machine Learning - Managing Complexity with MLOps")
- [Solving complex engineering challenges in artificial intelligence deployment](https://www.wearedevelopers.com/videos/392-mlops-what-s-the-deal-behind-it) (from "MLOps - What’s the deal behind it?")
- [Advancing biocatalyst research through quantum machine learning algorithms](https://www.wearedevelopers.com/videos/260-getting-started-with-machine-learning) (from "Getting Started with Machine Learning")
- [Exploring pathways into the machine learning engineering field](https://www.wearedevelopers.com/videos/368-introduction-to-azure-machine-learning) (from "Introduction to Azure Machine Learning")

## Related Articles

- [Graph and AI Trends 2026: Why Is AI Running but Not Yet Delivering?](https://www.wearedevelopers.com/magazine/680-graph-and-ai-trends-2026-why-is-ai-running-but-not-yet-delivering)
- [Stephan Gillich - Bringing AI Everywhere](https://www.wearedevelopers.com/magazine/489-stephan-gillich-bringing-ai-everywhere)
- [MLOps And AI Driven Development](https://www.wearedevelopers.com/magazine/82-mlops-and-ai-driven-development)
- [MLOps – What’s the deal behind it?](https://www.wearedevelopers.com/magazine/125-mlops-what-s-the-deal-behind-it)

## Related Jobs

- [Data Scientist](https://www.wearedevelopers.com/jobs/ext/1351648-data-scientist) at **Almedia**
- [Senior AI Agent Software Engineer (Go, Python) (m/f/x)](https://www.wearedevelopers.com/jobs/48277-senior-ai-agent-software-engineer-go-python-m-f-x) at **Dynatrace**
- [Principal Software Engineer, Enterprise AI Platform](https://www.wearedevelopers.com/jobs/ext/1467292-principal-software-engineer-enterprise-ai-platform) at **GitHub**
- [AI Software Engineer (Germany)](https://www.wearedevelopers.com/jobs/48317-ai-software-engineer-germany) at **Sunhat**
- [Lead Software Engineer - Data Engineering](https://www.wearedevelopers.com/jobs/ext/2000968-lead-software-engineer-data-engineering) at **Dynatrace**
- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/1597388-machine-learning-engineer) at **ZEISS Group**