Genomics and RNA

DeepSEA Variant Effects

HardRaw PyTorchAbout 30 min2 public examples plus private tests

Free account with email, Google or GitHub. Your code runs in an isolated PyTorch sandbox.

Your task

Combine paired genomic predictions, conservation features, and supplied null effects into learned variant scores and empirical significance measures.

Function signature

deepsea_variant_effects(
    activations: dict,
    parameters: dict,
    masks: dict,
    eps: float = 1e-06,
) -> dict[str, torch.Tensor]

Worked example

One channel and four null effects

The prediction changes from $0.25$ to $0.5$ : probability difference $0.25$ , log-odds difference $\log 3$ . Features are $[0.25,\log 3,2]$ ; the supplied classifier gives logit $1$ and probability about $0.7311$ . Effect magnitude is $0.25\log3\approx0.2747$ . Two of four null effects strictly exceed it, giving e-value $0.5$ and functional significance $0.5\times0.2=0.1$ .

import torch

activations = {
    'reference': torch.tensor([[[0.25]]], dtype=torch.float64, requires_grad=True),
    'alternate': torch.tensor([[[0.5]]], dtype=torch.float64, requires_grad=True),
    'conservation': torch.tensor([[2.0]], dtype=torch.float64, requires_grad=True),
    'conservation_impute': torch.tensor([0.0], dtype=torch.float64, requires_grad=True),
    'feature_mean': torch.tensor([0.0, 0.0, 0.0], dtype=torch.float64, requires_grad=True),
    'null_effects': torch.tensor([[0.0, 0.1, 0.5, 1.0]], dtype=torch.float64),
    'conservation_evalues': torch.tensor([[0.2]], dtype=torch.float64, requires_grad=True),
}
parameters = {
    'feature_scale': torch.tensor([1.0, 1.0, 1.0], dtype=torch.float64, requires_grad=True),
    'classifier_weight': torch.tensor([[4.0], [0.0], [0.0]], dtype=torch.float64, requires_grad=True),
    'classifier_bias': torch.tensor([0.0], dtype=torch.float64, requires_grad=True),
}
masks = {
    'conservation_valid': torch.tensor([[True]], dtype=torch.bool),
}
eps = 1e-06

result = deepsea_variant_effects(activations=activations, parameters=parameters, masks=masks, eps=eps)

Expected return

{
    'scores': torch.tensor([[0.7310585786300049]], dtype=torch.float64),
    'difference': torch.tensor([[[0.25]]], dtype=torch.float64),
    'log_odds': torch.tensor([[[1.0986122886681096]]], dtype=torch.float64),
    'features': torch.tensor([[0.25, 1.0986122886681096, 2.0]], dtype=torch.float64),
    'standardized': torch.tensor([[0.25, 1.0986122886681096, 2.0]], dtype=torch.float64),
    'evalues': torch.tensor([[0.5]], dtype=torch.float64),
    'functional_significance': torch.tensor([0.1], dtype=torch.float64),
}

Parameters

Parameter	Shape and meaning
`activations`dict	Supplied intermediate tensors.
`parameters`dict	Supplied model parameters, keyed as shown below.
`masks`dict	Supplied masks. Each field below defines what True includes or permits.
`eps`floatDefault `1e-06`	Clamp each reference and alternate probability to [eps, 1 - eps] before probability differences and natural-log odds; require 0 < eps < 0.5.

Return

dict[str, torch.Tensor]

Return exactly the named fields applicable to this operation.

Required behavior

reference and alternate have shape $[B,S,T]$ , with S aligned prediction replicates. Clamp both to $[\varepsilon,1-\varepsilon]$ , where eps defaults to $10^{-6}$ and lies in $(0,0.5)$ . Compute signed probability differences and signed alternate-minus-reference log odds.
Feature order is: mean absolute probability difference across S for every channel; mean absolute log-odds difference across S; absolute conservation values. Replace invalid conservation entries with the supplied per-feature imputation value before taking absolute values.
Standardize with supplied feature_mean and positive feature_scale. Apply the supplied classifier_weight and classifier_bias, followed by sigmoid, to produce scores.
For each prediction channel, effect magnitude is mean(abs(difference) * abs(log_odds)) across replicates. evalues are the fraction of null_effects strictly GREATER than this magnitude; ties do not count.
functional_significance is geometric_mean(evalues over prediction channels) multiplied by geometric_mean(conservation_evalues over conservation channels). Do not add a pseudocount or take a single combined geometric mean.
For probability $p$ , log odds are $\log p-\log(1-p)$ . The classifier uses $sigmoid(z)=1/(1+e^{-z})$ .
The geometric mean of $n$ nonnegative factors is $(\prod_i a_i)^{1/n}$ ; a zero factor makes the group mean zero. For positive factors, averaging logarithms is numerically stable.

Why this matters

A splice score locates the strongest predicted gain or loss along a sequence. DeepSEA instead combines effect magnitudes across outputs with supplied conservation and classifier resources. Both summarize model predictions; neither supplies clinical calibration.

Starter code

You write the body of this function in the BioTorch workspace. Your code runs in an isolated PyTorch sandbox; Run checks the public examples and Submit grades the full test suite, including gradients.

import torch


def deepsea_variant_effects(
    activations: dict,
    parameters: dict,
    masks: dict,
    eps: float = 1e-06,
) -> dict[str, torch.Tensor]:
    """Return dict[str, torch.Tensor]:
    
    scores: (B,K).
    difference: (B, S, T).
    log_odds: (B, S, T).
    features: (B, 2T+Q).
    standardized: (B, 2T+Q).
    evalues: (B, T).
    functional_significance: (B,)."""
    # Implement the operation. Inputs and test setup are supplied by BioTorch.
    raise NotImplementedError("Implement deepsea_variant_effects")

Ready to implement it?

3 progressive hints, a solution walkthrough, and grading that checks values, shapes and gradients are waiting in the workspace.

Write and run your solution Sign in

Free account with email, Google or GitHub. Your code runs in an isolated PyTorch sandbox.

Related exercises

Sources and further reading

A primitive exercise, not an end-to-end reproduction of the named models. No affiliation or endorsement is implied.