FROM TENSORS TO BIOLOGY

Understand biological AI.Build it in PyTorch.

Implement the math inside AlphaFold, ESM and scVI in raw PyTorch, with grading that checks your values, shapes and gradients. Follow a model from its inputs to its predictions, one short exercise at a time.

Start your first exercise Explore learning paths

Free. Sign up with email, Google or GitHub and land in a 10-minute guided exercise. Or read an exercise first.

The AlphaFold 2 guide in BioTorch: numbered steps grouped into chapters, with feedback and readout branches beside the main computation. — Every model is an interactive map of its computation.Open AlphaFold 2

212 gradient-checked exercises
17 landmark model guides

THE PRACTICE LOOP

Feedback that teaches, not just pass or fail.

Every exercise is graded on its outputs and its gradients. When something breaks, you see which tensor, which case and why.

solution.py

def weighted_masked_mean(x, mask, weights, dim=-1):
    w = weights.expand(x.shape) * mask
    s = (w * x.masked_fill(~mask, 0)).sum(dim)
    d = w.sum(dim)
    return torch.where(d > 0, s / d, 0.0)

Run · Wrong answer 1 / 2 public cases

Output values passed, but gradients failed.

weights: non-finite gradient in two_rows_one_empty. An all-masked row divides by zero inside torch.where, so its gradient is NaN even though the selected value is 0.

After a safe denominator · Accepted 7 / 7 checks

Values, shapes, masks, dtypes and vector-Jacobian products are all checked, the same subtleties that break real model code.

Explore the building blocks

Attention
Protein geometry
Diffusion
Sequence models
Genomics

START WITH A QUESTION YOU CARE ABOUT

What do you want to understand?

Each path combines an interactive architecture map, worked examples, and sessions of up to three exercises. Start wherever your curiosity leads.

PROTEIN STRUCTURE

How does a sequence become a structure?

Follow AlphaFold 2 from evolutionary evidence through attention and residue geometry.

Explore AlphaFold 2 →

PROTEIN LANGUAGE

What can a model learn from sequences?

Build the embeddings, attention, and prediction operations behind ESM-2.

Explore ESM-2 →

GENE REGULATION

How can DNA predict cellular activity?

Connect Enformer’s sequence processing and long-range attention to functional readouts.

Explore Enformer →

Browse all 17 learning paths → Guides and exercises are open to everyone. Create a free account to run code and save your progress.

LEARN BY IMPLEMENTING

From an equation to something you understand.

A roadmap with a purpose

Choose a model and follow its chapters. Each component explains the problem, inputs, transformation, outputs, and the operation you will implement.

Code, run, and submit

Write Python in the workspace. Your code runs in an isolated environment, with checks for numerical values, shapes, and gradients. Results help you see what needs another look.

Practice that follows you

Keep your submission history and progress together. Your first implementation starts a review schedule, and BioTorch reminds you to revisit each idea right before you would forget it.

You implement the components of each model, not a full reproduction.

FREE ACCOUNT

What you get when you sign up

Run and submit real code in an isolated PyTorch sandbox, with value, shape and gradient checks.
Progress that follows you: completed exercises, model paths and history in one place.
Reviews at the right time, so the ideas stay with you after the first solve.
Hints and walkthroughs for every exercise when you get stuck.

Sign in with email, Google or GitHub. Your code and history are private to you, and BioTorch never stores your password. Read the privacy policy.

Start your first exercise