Health and life sciences
Machine learning for functional protein design
Machine learning for protein design generates new backbones with diffusion models, designs amino-acid sequences that fit a backbone, and predicts whether a designed sequence folds as intended. Designs that pass these computational checks then go to laboratory assays that test the intended function.
How machine learning is used in protein design
Generating new protein backbones
Diffusion models start from noise and refine it into a protein backbone, optionally built around a motif or binding site the designer specifies. This produces shapes that do not exist in nature but are plausible to build.
Real exampleRFdiffusion (Watson et al., 2023)Fine-tuned the RoseTTAFold structure prediction network on structure denoising to generate protein backbones, and experimentally characterised hundreds of designed symmetric assemblies, metal-binding proteins and binders.
Designing a sequence for a fixed backbone
Inverse folding models take a backbone and predict amino-acid sequences likely to fold into it. Designers can fix residues that must not change, such as an active site, and let the model fill in the rest.
Real exampleProteinMPNN (official repository)The project repository provides the trained model weights and helper scripts for parsing structures, choosing which chains to design and fixing residues, with runnable examples.
Checking a design before the laboratory
Structure prediction runs on each designed sequence, and designs whose predicted structure does not match the intended shape are dropped. This filter costs computer time instead of laboratory time.
Real exampleAlphaFold (Jumper et al., 2021)Presented a neural network that regularly predicts protein structures with atomic accuracy even when no similar structure is known, validated in the CASP14 blind assessment.
Generating sequences with protein language models
Language models trained on millions of natural sequences generate new members of a protein family, steered by fine-tuning or control tags. The designs are then expressed and assayed for the intended activity.
Real exampleProGen (Madani et al., 2023)Artificial proteins fine-tuned to five lysozyme families showed catalytic efficiencies similar to natural lysozymes, with sequence identity to natural proteins as low as 31.4%.
Can you learn machine learning for protein design?
Yes, provided you learn some structural biology alongside the models. Biochemists, structural biologists and machine learning engineers all enter this field. Start with Python and structure files, add machine learning and deep learning, and run the published design tools before trying to train your own.
- Python and structure filesRead PDB and mmCIF files, extract chains and residues with Biopython, and drive command-line tools from scripts.Learn it in Python
- Protein structure basicsLearn secondary structure, folding, binding interfaces and how designs are expressed and assayed, from a biochemistry text or course.
- Machine learning foundationsUnderstand training, validation and why a model confidence score is not evidence that a design works.Learn it in Machine Learning
- Deep learning and generative modelsLearn neural networks, attention and diffusion so the published design pipelines stop being black boxes.Learn it in Deep Learning
- Run a published pipeline end to endGenerate a backbone, design sequences with ProteinMPNN and predict the structure of each design, then filter on agreement with the target shape.
Moving into machine learning for protein design
If you already work in the field
Protein scientists already know expression, purification and the assays that decide whether a design works. Add Python and the open design tools, then redesign a protein your laboratory can already express and test, so the computational result meets a real measurement.
If you come from data science or ML
Machine learning engineers should learn what a backbone is, what makes a sequence expressible and how long an experimental round takes. Reproduce a published design pipeline, then work with a laboratory that can test designs, because only an assay settles whether a design has the intended function.
Roles that do this workComputational protein designerProtein engineerMachine learning scientist, protein designStructural bioinformatician
Tools you will use
| Tool | What you use it for |
|---|---|
| ProteinMPNN | Design amino-acid sequences that are predicted to fold onto a given protein backbone. |
| RFdiffusion | Generate new protein backbones, binders and scaffolds around a functional site with a diffusion model. |
| ColabFold | Predict the structure of a designed sequence to check whether it refolds to the target backbone. |
| Biopython | Read PDB and mmCIF files, extract chains and sequences, and compare structures. |
Public data to practise on
- RCSB Protein Data Bank: Experimentally determined protein structures to use as design targets and to test sequence recovery.
- CATH: Protein domains classified by fold, useful for building splits that hold out whole structural families.
Your first project
Download 20 structures from the RCSB PDB, redesign their sequences with ProteinMPNN, and compare native sequence recovery with a baseline that samples amino acids at background frequencies.
Common questions
Do you need a wet laboratory to work on protein design?
Not for the computational half, which runs on a workstation or a cloud GPU. You do need laboratory collaborators, because a design is only validated when it is made and assayed.
Does a high structure prediction confidence mean the design works?
No. Confidence says the model expects the sequence to fold into the predicted shape. Function, stability and expression are separate questions answered by experiments.
What hardware do the design tools need?
A single modern GPU runs sequence design and structure prediction for small proteins, and free Colab notebooks exist for the main tools. Large backbone generation runs are faster with more GPU memory.
Should you start with structure generation or sequence design?
Start with sequence design on an existing backbone. It is quicker, easier to evaluate and teaches the file formats and filters you need before generating new shapes.
Courses for these skills
- Generative AI: LLMs, RAG and AgentsUnderstand generative objectives and evaluating generated candidates.Syllabus available · lessons open as produced
- Deep Learning with PyTorchBuild representation-learning foundations before specialised graph models.Syllabus available · lessons open as produced
- Practical Machine Learning in PythonMatch losses and validation to measured molecular endpoints.51 published lessons · more lessons in production
New to programming or maths? Start with Python, or follow the machine learning path in order. Published lessons are free to read; compare Free and Pro.