Bio Machine Learning Prediction Explanation

Version Compatibility

Reference examples tested with: numpy 1.26+, pandas 2.2+, scikit-learn 1.4+, shap 0.44+, lime 0.2+.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

Two shap drifts: the Explanation-object plotting API arrived ~0.36 (new-style shap.plots.* take an Explanation, legacy shap.summary_plot/dependence_plot take numpy arrays -- mixing them is the most common runtime error); and TreeExplainer(..., feature_perturbation='auto') became the default in 0.47 (was interventional), so providing or omitting data= silently changes the estimand. Always set feature_perturbation explicitly. If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt rather than retrying.

Model Interpretation for Omics Classifiers

"Which genes drive my classifier?" -> Compute attributions, but treat them as a description of the model (not biology), choose the Shapley conditioning deliberately, and aggregate over correlated gene modules before ranking.

• Tree models: shap.TreeExplainer(model, data=background, feature_perturbation='interventional')
• Model-agnostic local: lime.lime_tabular.LimeTabularExplainer
• Model-reliance screen: sklearn.inspection.permutation_importance

The Single Most Important Modern Insight -- Attributions Explain the Model, Not Biology, and Under Correlation the Algorithm Chooses How to Split Credit

A feature attribution describes the function the model learned on this training distribution; it is not a measurement of biology. A high-SHAP gene can be a pure correlate of a batch, scanner, or library-prep signal the model exploited (DeGrave 2021 is the canonical proof). And because genes co-express in tight modules, the attribution algorithm has genuine freedom in how it splits credit among correlated genes -- the choice of conditioning (tree_path_dependent/conditional vs interventional/marginal) is not a cosmetic knob, it changes which genes get credit, and it is a live methodological controversy with no universally correct answer (Janzing 2020). The operational rule: SHAP/LIME rankings are a debugging and hypothesis-generation tool, never a validated biomarker-selection criterion, and within a co-expression module the ordering is not a finding.

Method Taxonomy

The Correlated-Feature / Conditional-vs-Marginal Problem (the core)

When Shapley values "drop" a feature subset, they replace it by some distribution, and two incompatible choices exist:

• Conditional / observational (tree_path_dependent): dropped features drawn from p(x_dropped | x_S). A feature the model never uses can still receive nonzero attribution purely because it is correlated with a used feature. So high SHAP does not mean the model relies on this gene.
• Marginal / interventional (interventional): dropped features drawn from the marginal p(x_dropped), i.e. do(x_dropped = background). Features the model genuinely ignores get exactly zero, even if correlated. Janzing 2020 argues this is the principled "drop" for attribution; it is why modern SHAP added and (pre-0.47) defaulted to it.

There is no free lunch: every method either extrapolates off-manifold (interventional SHAP, unrestricted permutation -- Hooker 2021's "no free variable importance") or leaks credit through correlation (conditional SHAP). Decide which pathology the question can tolerate, and aggregate attributions over co-expression modules before ranking -- "gene A ranked above gene B" within a correlated module is governed by off-manifold value-function behavior, not biology.

Decision Tree by Scenario

SHAP TreeExplainer (set the conditioning explicitly)

Goal: Attribute a tree model's predictions to features with a chosen, stated estimand.

Approach: Pass a background and feature_perturbation='interventional' for "what the model uses," or omit the background and use tree_path_dependent for the conditional/descriptive view. The default 'auto' flips between them based on whether data= is given.

import shap
import numpy as np

# Interventional ('what the model uses'): needs a background (~100-1000 rows).
background = shap.utils.sample(X_train, 200)
explainer = shap.TreeExplainer(model, data=background, feature_perturbation='interventional')
sv = explainer(X_test)                                  # modern Explanation object

# Aggregate over correlated modules BEFORE ranking (clusters = a precomputed gene->module map).
mean_abs = np.abs(sv.values).mean(axis=0)
module_importance = {}
for gene, m in zip(X_test.columns, mean_abs):
    module_importance[clusters[gene]] = module_importance.get(clusters[gene], 0) + m

Attribution Is Not Causation, Mechanism, or a Validated Biomarker

Three layers of "not": (1) not biology -- the attribution describes the model, which may have exploited a batch/confounder shortcut; (2) not causation -- predictive features conflate direct effects, confounders, mediators, and colliders, and turning attribution into a causal claim needs an explicit causal model SHAP does not contain; (3) not a validated biomarker -- taking "top-20 SHAP genes" as a panel is same-data feature selection with no FDR and no replication (winner's curse). The strongest legitimate use runs the other way: because attribution exposes what the model used, it is one of the best tools to catch shortcut/batch learning -- if top features are batch indicators or depth-tracking housekeeping genes, the model is cheating (DeGrave 2021). That is where attribution is most trustworthy.

LIME and Explanation Instability

LIME fits a sparse linear surrogate to predictions on perturbed samples around one instance. It is non-reproducible by construction: different seeds, different kernel_width, and discretize_continuous=True (the default) each flip the top features, and the per-feature perturbations land off-manifold for correlated genes. Worse, perturbation-based explainers can be deliberately fooled -- a biased model can be wrapped to look innocuous on the out-of-distribution points LIME/KernelSHAP probe (Slack 2020). Use LIME only to eyeball a single prediction's local logic with a pinned seed, never for global ranking, and never as evidence a model is unbiased.

from lime.lime_tabular import LimeTabularExplainer

explainer = LimeTabularExplainer(X_train.values, feature_names=list(X_train.columns),
                                 mode='classification', discretize_continuous=True,
                                 random_state=0)         # pin the seed; still only conditional stability
exp = explainer.explain_instance(X_test.values[0], model.predict_proba, num_features=10, num_samples=5000)

Background / Baseline Choice (the silent attribution-changer)

SHAP explains the deviation from E[f(X)] over the background dataset, so the background defines what "absence of a feature" means and changes every attribution. A tumor sample explained against a tumor-heavy vs a healthy-tissue background yields different "important genes" -- only one matches the scientific question. Use a background of real samples representative of the contrast of interest (a single global mean across a heterogeneous cohort is no real sample). check_additivity=True (default) raises when the SHAP values plus the base value do not sum to the model output (a local-accuracy violation) -- often a probability-vs-raw-margin or implementation mismatch; investigate rather than disabling it. Attributions in log-odds (model_output='raw') differ from probability space -- report the scale.