🤖 machine learning

GeneSpeak-FP: Target and Compound Retrieval from Observed Cell-Level Perturbation Signatures

The paper introduces GeneSpeak-FP, a Transformer-based retrieval model that successfully identifies specific molecular targets and compounds from observed single-cell perturbation signatures within a closed library setting, achieving significant performance gains over baseline controls while highlighting the need for further validation on unseen contexts.

Original authors: Kseniia Vaniushkina, Jeongmin Lim, Jinyong Park

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kseniia Vaniushkina, Jeongmin Lim, Jinyong Park

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective walking into a room where a chemical "mystery" has just happened. A cell in a petri dish has been exposed to a drug, and its internal machinery has scrambled its instructions, changing which genes are turned on or off. This change is called a "perturbation signature." For decades, scientists have tried to predict what will happen if they give a cell a specific drug (the forward question). But what if you only have the messy aftermath—the scrambled gene list—and you need to figure out which drug caused it? This is the "inverse" question. It's like smelling a strange perfume and trying to guess the exact brand and ingredients without seeing the bottle. This paper tackles that challenge using a massive library of single-cell data, where researchers have watched thousands of individual cells react to different chemicals, creating a giant map of cause and effect.

The team behind this study, working with a massive dataset called Tahoe-100M, built a new AI detective named GeneSpeak-FP. Think of this AI as a super-smart librarian who has memorized the "fingerprints" of thousands of drugs. When you hand the librarian a single cell's reaction (its gene signature), the AI doesn't just guess; it searches its memory to find the most likely suspects. Specifically, it tries to do two things at once: first, it guesses which specific "targets" (genes) the drug was supposed to hit, and second, it tries to identify the exact drug molecule from a fixed list of 379 known compounds.

Here is how the magic works. The AI looks at a treated cell and compares it to a "control" cell (one that got a harmless drop of DMSO instead of a drug) to see what changed. It then feeds this list of changes into a special type of neural network called a Transformer. This network acts like a translator, turning the messy list of gene changes into a neat, compact "vector" (a mathematical fingerprint). This fingerprint is then run through two different doors. One door leads to a list of 278 possible gene targets, and the other leads to the library of 379 drugs. The AI ranks them, saying, "I'm 40% sure this drug hit these genes, and I'm 34% sure this specific drug is the one in the top 10."

The results are promising, but with important caveats. In their tests, the AI managed to find the correct drug in its top 10 guesses about 34% of the time (Hit@10) and correctly identified the target genes in its top 10 guesses about 41% of the time (Recall@10). This is a huge jump compared to just guessing randomly, which would only get it right about 2-3% of the time. The authors show that this system is much better than simpler methods that just count genes without understanding how they interact.

However, it is crucial to understand what this AI doesn't do. The experiment was set up as a "closed-library" test. This means the AI was only asked to pick from drugs and targets it had already seen during its training. It didn't have to guess a brand-new, unseen drug or a cell type it had never met. The authors are very clear: this is a tool for organizing known data and narrowing down a list of suspects, not a crystal ball for discovering entirely new medicines or making clinical decisions for patients. The "targets" it identifies are based on existing metadata labels, not proven biological facts. So, while GeneSpeak-FP is a powerful new way to sort through the chaos of cell reactions and find patterns in a known library, it is not yet a magic wand that can solve every biological mystery or replace the need for real-world lab experiments. It's a very sharp magnifying glass for a specific, well-lit room, but the rest of the house is still dark.

Technical Summary: GeneSpeak-FP

Problem Definition
The paper addresses an inverse retrieval problem in single-cell perturbation transcriptomics. While most existing machine learning approaches focus on the "forward" question—predicting a cell's transcriptional response to a specified perturbation—this work tackles the complementary "inverse" task: given an observed cell-level transcriptional signature, can a model retrieve the most consistent annotated targets and the specific compound from a fixed, closed library that generated that response? The authors note that this task is inherently ambiguous because distinct compounds may share targets, structures, or downstream effects, and the goal is to narrow a candidate space rather than infer a unique causal mechanism.

Methodology
The authors propose GeneSpeak-FP, a multi-task Transformer model designed for a closed-library setting using the Tahoe-100M dataset.

Input Formulation: The input is a sparse, cell-level perturbation signature ( $\Delta x$ ) calculated by contrasting the log-transformed expression profile of a single treated cell against the mean log-transformed DMSO reference profile of its specific cell line. The input universe consists of 4,184 genes (high-variance DMSO genes plus mapped targets). For each sample, the 256 genes with the largest absolute changes are selected, sorted by token ID, and padded to a fixed length.
Architecture: The model employs a Gene-Token Transformer encoder. Unlike standard Transformers, it does not use positional embeddings; instead, it treats the selected genes as an unordered set where self-attention models interactions. The input sequence includes special [CLS] and [ORGAN] tokens (the latter derived from cell-line metadata). The encoder consists of four layers with 256 hidden dimensions and eight attention heads.
Dual Retrieval Heads: The shared encoder representation feeds two distinct retrieval heads:
1. Target Retrieval Head: A 256-dimensional head that ranks a closed vocabulary of 278 Tahoe-annotated target genes. It utilizes a trainable embedding for each target gene.
2. Compound Retrieval Head: A multi-layer head (256 $\to$ 1024 $\to$ 768) that ranks a fixed bank of 379 compounds. This head aligns the transcriptome-derived representation with precomputed, fixed 768-dimensional molecular embeddings.
Training Objective: The model is trained jointly using a composite loss function:
- Target Loss ( $L_T$ ): Combines cosine similarity loss, Binary Cross-Entropy (BCE) for positive targets, and a multi-positive ranking loss (cross-entropy with temperature scaling).
- Molecular Alignment Loss ( $L_{CLIP}$ ): A symmetric contrastive loss (similar to CLIP) aligning the predicted molecular vectors with the fixed embedding bank, supplemented by a cosine alignment term to stabilize training.
- Total Loss: $L = L_T + 0.05 (L_{CLIP} + 0.5 L_{align})$ .

Experimental Setup
The evaluation utilizes a subset of the Tahoe-100M dataset containing 11,673 drug–cell-line condition pairs involving 379 compounds and 278 target genes.

Data Split: A within-compound stratified 90/10 split (10,505 training, 1,168 validation) was used. Crucially, the same compounds and cell lines appear in both partitions; the experiment measures retrieval for held-out condition pairs rather than zero-shot generalization to unseen compounds or cell lines.
Evaluation Protocol: The main evaluation is a Monte Carlo estimate over 38,400 sampled validation cells. Metrics include Recall@K and Precision@K for targets, and Hit@K, Mean Reciprocal Rank (MRR), and rank statistics for compounds.
Controls: The study compares GeneSpeak-FP against a Gaussian random-vector control and three "bag-of-genes" (BoG) diagnostic controls (prototype matcher, linear molecular head, and k-NN) to assess the value of the learned contextual representations.

Key Results
In the diagnostic comparison, GeneSpeak-FP significantly outperformed all control methods:

Target Retrieval: Achieved a Recall@10 of 0.409 (compared to a maximum of 0.0312 for controls).
Compound Retrieval: Achieved a Hit@10 of 0.342 (compared to a maximum of 0.0366 for controls) and a Hit@1 of 0.129.
Extended Evaluation: A separate pass yielded consistent results (Recall@10 $\approx$ 0.408, Hit@10 $\approx$ 0.343, MRR $\approx$ 0.205).
The authors emphasize that these metrics represent the recovery of metadata labels and recorded compound identities within the observed library, not classification accuracy over unseen molecules or the full transcriptome.

Significance and Claims
The paper claims that GeneSpeak-FP demonstrates the feasibility of a single multi-task model to simultaneously recover mapped target annotations and recorded compound identities from observed cell-level responses in a closed-library setting.

Primary Contribution: The unified formulation and evaluation of simultaneous target and compound retrieval from individual cell-level signatures within the Tahoe-100M resource.
Limitations and Scope: The authors are explicit that the results are specific to the evaluated closed-library protocol. The model does not generalize to unseen compounds or cellular contexts (cold-start scenarios) within this study. The target retrieval is framed as metadata retrieval rather than mechanistic inference, acknowledging that target annotations are compound-level labels and not exhaustive causal ground truth.
Application: The demonstrated use case is strictly for organizing existing perturbation screens and narrowing a predefined candidate space for follow-up analysis. The authors explicitly state that the model is not intended for clinical decision-making, causal target discovery, or therapeutic recommendation without independent biological and prospective validation.

Technical Summary: GeneSpeak-FP

More like this