transfer/code/prompt_selection/prompt_selector.py

"""Two-stage embedding-based prompt selection.

Stage 1: For each query batch, find top-K1 most similar control prompts
         based on cosine similarity in embedding space.
Stage 2: Map those K1 control prompts to their predicted perturbation
         embeddings, average, and find top-n_base_cells real perturbed
         prompts closest to that average.
"""
from __future__ import annotations

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity


def select_prompt_indices(
    query_embeddings: np.ndarray,
    batch_global_indices: np.ndarray,
    prompt_ctrl_embeddings: np.ndarray,
    predicted_pert_embeddings: np.ndarray,
    prompt_pert_embeddings: np.ndarray,
    n_base_cells: int,
    top_k1: int = 512,
) -> np.ndarray:
    """Select best prompt indices for a single query batch.

    Parameters
    ----------
    query_embeddings : (N_query, D)
        Precomputed embeddings for all query cells.
    batch_global_indices : (batch_len,)
        Indices into query_embeddings for the current batch.
    prompt_ctrl_embeddings : (N_ctrl, D)
        Embeddings of control prompt cells.
    predicted_pert_embeddings : (N_ctrl, D)
        Embeddings of predicted perturbation cells (1-to-1 with ctrl).
    prompt_pert_embeddings : (N_pert, D)
        Embeddings of real perturbed prompt cells.
    n_base_cells : int
        Number of prompt cells to select.
    top_k1 : int
        Number of control prompts to shortlist in stage 1.

    Returns
    -------
    selected_idx : (n_base_cells,)
        Indices into the prompt_pert AnnData for the selected prompts.
    """
    # Mean embedding of current query batch
    mean_query_emb = query_embeddings[batch_global_indices].mean(axis=0, keepdims=True)

    # Stage 1: cosine similarity against control prompts
    sim_ctrl = cosine_similarity(mean_query_emb, prompt_ctrl_embeddings)[0]
    k1 = min(top_k1, len(sim_ctrl))
    top_k1_idx = np.argpartition(sim_ctrl, -k1)[-k1:]

    # Stage 2: average predicted perturbation embeddings of stage-1 hits
    pred_emb_mean = predicted_pert_embeddings[top_k1_idx].mean(axis=0, keepdims=True)

    # Cosine similarity against real perturbed prompts
    sim_pert = cosine_similarity(pred_emb_mean, prompt_pert_embeddings)[0]
    n_select = min(n_base_cells, len(sim_pert))
    selected_idx = np.argpartition(sim_pert, -n_select)[-n_select:]

    return selected_idx