Uncertainty-aware synthetic lethality prediction with pretrained foundation models

This paper introduces CO_SCPLOWILANTROC_SCPLOWO_SCPCAP-C_SCPCAPO_SCPLOWSLC_SCPLOW, a two-stage, graph-free framework that leverages pretrained biological foundation models and conformal prediction to accurately predict synthetic lethality pairs with calibrated uncertainty, thereby enabling robust discovery of therapeutic targets without relying on traditional protein-protein interaction networks or Gene Ontology annotations.

The Gist

The Big Problem: Finding the "Fatal Flaw" in Cancer

Imagine cancer cells as a fortress. Usually, if you attack one wall (a gene), the fortress has a backup generator (another gene) that keeps it running. But sometimes, there is a secret rule: If you break Wall A AND Wall B at the same time, the whole fortress collapses. However, if you only break Wall A, or only Wall B, the fortress survives.

In biology, this is called Synthetic Lethality (SL). Finding these specific pairs of "walls" is the holy grail of cancer treatment because it allows doctors to kill cancer cells without hurting healthy ones (which only have one wall broken).

The Catch: There are billions of possible wall combinations. Testing them all in a lab is like trying to find a specific grain of sand on a beach by picking up every grain one by one. It's too slow, too expensive, and we often only know about the "famous" grains (well-studied genes), ignoring the unknown ones.

The Solution: CILANTRO-SL

The researchers built a new AI tool called CILANTRO-SL. Think of it as a super-smart, crystal-ball-equipped detective that can guess which wall combinations will cause a collapse, without needing to physically test every single one.

Here is how it works, broken down into three simple steps:

1. The "What-If" Simulator (Stage 1)

Most old AI models rely on a pre-drawn map of how genes talk to each other (like a phone book of friends). But that map is incomplete; we don't know who talks to whom for many genes.

CILANTRO-SL doesn't use a map. Instead, it uses a massive library of biological stories (called a "Foundation Model," specifically Geneformer).

• The Analogy: Imagine you have a library containing the "life story" of millions of cancer cells.
• The Trick: The AI reads the story of a cancer cell. Then, it uses a "digital eraser" to pretend a specific gene was deleted. It asks the library: "How does the story change if we remove this character?"
• The Result: It creates a "Delta" (a difference score). This tells the AI exactly how much the cell struggles when that specific gene is gone. It does this for thousands of genes and cell types, learning a "viability signature" for each one.

2. The "Matchmaker" (Stage 2)

Now the AI has a list of "struggle signatures" for individual genes. It needs to find pairs that are deadly together.

• The Analogy: Imagine you are a matchmaker trying to find two people who, if they met, would cause a chaotic party (the cancer dying).
• The Process: The AI takes the "struggle signature" of Gene A and Gene B and smashes them together. It asks: "Do these two struggles combine to create a total disaster for the cell?"
• The Secret Sauce: It uses a special technique called FiLM (Feature-wise Linear Modulation). Think of this as a volume knob. The AI can turn the "volume" of Gene A's signature up or down based on what it knows about Gene B's personality (using a separate gene database called Gene2vec). This helps it understand the context better than just looking at the genes in isolation.

3. The "Confidence Meter" (Uncertainty)

This is the most important part. Old AI models just say "Yes, this pair is lethal" or "No, it isn't." But what if the AI is guessing?

• The Analogy: Imagine a weather app. A bad app says "It will rain." A good app says "There is a 90% chance of rain, and here is the confidence interval."
• The Innovation: CILANTRO-SL uses a math trick called Conformal Prediction. Instead of giving a single "Yes/No," it gives a Confidence Set.
  • High Confidence: "I am 99% sure this pair will kill the cancer." (The AI is confident).
  • Low Confidence: "I'm not sure. It might kill the cancer, or it might not." (The AI admits it doesn't know).
• Why it matters: This saves scientists time. They only waste money testing the "High Confidence" guesses. They ignore the "I'm not sure" ones until they have more data.

Why is this a Big Deal?

1. It works on the "Unknowns": Most AI models fail when they see a gene they've never studied before. CILANTRO-SL is like a detective who can solve a crime even if the suspect has never been in the system before, because it understands the behavior of the crime, not just the criminal's name.
2. It's Honest: By telling scientists when it is unsure, it prevents them from chasing dead ends.
3. It's Context-Aware: It knows that a gene pair might be lethal in a lung cancer cell but harmless in a skin cancer cell. It tailors its advice to the specific "neighborhood" (cell type).

The Bottom Line

CILANTRO-SL is a new way to hunt for cancer cures. Instead of blindly testing billions of combinations, it uses a digital simulator to learn how genes behave, a matchmaker to find deadly pairs, and a confidence meter to tell scientists exactly which leads are worth chasing. It turns a needle-in-a-haystack problem into a targeted search, making the discovery of new cancer therapies faster and more reliable.

Technical Summary

1. Problem Statement

Synthetic Lethality (SL) occurs when the simultaneous loss of function of two genes leads to cell death, while the loss of either gene alone is non-lethal. This mechanism offers a powerful strategy for targeted cancer therapy, particularly for tumors with specific loss-of-function mutations. However, identifying SL pairs faces three major challenges:

• Combinatorial Explosion: The search space of human gene pairs is enormous (~200 million), making exhaustive experimental screening infeasible.
• Data Scarcity and Bias: Experimental SL labels are sparse, noisy, and heavily biased toward well-studied genes. Existing computational methods often rely on curated resources like Protein-Protein Interaction (PPI) networks and Gene Ontology (GO) annotations, which limits their ability to generalize to novel or sparsely annotated genes.
• Context Dependency: SL interactions are highly context-dependent (varying by cell line), yet many models produce global, cell-line-agnostic predictions.
• Lack of Uncertainty Quantification: Current models often provide point estimates without calibrated confidence scores, making it difficult to prioritize candidates for costly experimental validation.

2. Methodology: CILANTRO-SL

The authors propose CILANTRO-SL, a two-stage, graph-free framework that leverages pretrained biological foundation models to predict SL pairs with calibrated uncertainty. It avoids reliance on static interaction topologies (PPI/GO) in favor of learning directly from perturbation and expression data.

Stage 1: Viability Pretraining (Representation Learning)

The goal is to learn viability embeddings that capture how knocking out a specific gene affects cell survival in a specific molecular context.

• In Silico Knockouts: The framework uses Geneformer, a single-cell foundation model (scFM) pretrained on 30 million transcriptomes.
  • For each cell line, the model generates a baseline embedding ($X_c$) from bulk RNA-seq data.
  • It simulates a gene knockout ($g$) by removing the corresponding gene token from the input sequence and re-embedding to generate a perturbed embedding ($X'_{c,g}$).
  • A Delta Embedding is calculated: $\Delta X_{c,g} = X_c - X'_{c,g}$. This represents the model-defined counterfactual shift caused by the gene removal.
• Gene Identity Priors: To capture global gene behavior not present in a single cell context, the model integrates Gene2vec embeddings (learned from large-scale co-expression data) as a gene identity prior ($e_g$).
• FiLM Conditioning: The delta embeddings are modulated by the gene priors using a Feature-wise Linear Modulation (FiLM) layer. This allows the model to selectively scale and shift the perturbation signal based on global gene identity.
• Regression Objective: A lightweight Multi-Layer Perceptron (MLP) is trained on DepMap CRISPR viability data to regress a viability score. The output of the penultimate layer (32-dimensional) serves as the Viability Embedding ($V_{c,g}$), encoding the gene-specific perturbation effect in cell line $c$.

Stage 2: SL Prediction and Uncertainty Quantification

• Pairwise Classification: For a candidate gene pair $(g_1, g_2)$ in cell line $c$, the model concatenates their viability embeddings $[V_{c,g_1}; V_{c,g_2}]$. A second MLP classifies the pair as SL or non-SL.
• Conformal Prediction: To address label noise and enable reliable prioritization, the framework applies Conformal Prediction.
  • Instead of a fixed probability threshold, it generates prediction sets with finite-sample coverage guarantees.
  • It calculates non-conformity scores based on the classifier's output and a held-out calibration set.
  • The result is a set of possible labels (e.g., {SL}, {non-SL}, or {SL, non-SL}) that contains the true label with a user-specified probability ($1-\alpha$).
  • Singleton sets (size 1) indicate high-confidence predictions, while larger sets indicate ambiguity.

3. Key Contributions

1. Graph-Free Foundation Model Approach: CILANTRO-SL is the first SL prediction framework to utilize pretrained single-cell foundation models (Geneformer) and gene co-expression embeddings (Gene2vec) to learn perturbation-aware representations, eliminating the need for curated PPI or GO graphs.
2. Viability Pretraining Strategy: The authors introduce a novel pretraining objective that learns "viability embeddings" by regressing CRISPR knockout effects on in silico perturbed embeddings, effectively transferring single-gene dependency knowledge to pairwise SL prediction.
3. Uncertainty-Aware Prioritization: By integrating conformal prediction, the model provides statistically rigorous confidence intervals, allowing researchers to filter for high-confidence candidates suitable for experimental follow-up.
4. Zero-Shot Generalization: The framework demonstrates robust performance on "unseen genes" (Gene-holdout), a critical capability for discovering SL interactions in understudied genomic regions where traditional graph-based methods fail.

4. Results

The model was evaluated on the SynLethDB 2.0 dataset (35k+ pairs) and DepMap data (1,021 overlapping cell lines), using two cross-validation regimes: Pair-holdout (unseen pairs, seen genes) and Gene-holdout (unseen genes).

• Performance Superiority:
  • CILANTRO-SL outperformed state-of-the-art baselines (including graph-based models like KG4SL, DDGCN, and embedding-based models like ESM4SL).
  • In the strict Gene-holdout setting (zero-shot for genes), CILANTRO-SL achieved an F1 score of 0.7148, significantly outperforming KG4SL (0.558) and ESM4SL (0.478).
  • It showed superior stability, with a much smaller performance drop between Pair-holdout and Gene-holdout compared to baselines.
• Ablation Studies:
  • Delta vs. Perturbed: Using delta embeddings ($\Delta X$) for pretraining yielded better SL prediction transferability than using raw perturbed embeddings ($X'$), despite $X'$ fitting the regression task better.
  • Gene Priors: Integrating Gene2vec priors via FiLM significantly improved performance over using delta embeddings alone or other priors (ESM2, GenePT).
• Uncertainty Calibration:
  • Conformal prediction achieved empirical coverage matching the target coverage (e.g., 90% target $\approx$ 90% observed coverage).
  • High-confidence predictions (singleton sets) were strongly enriched for CRISPR-validated SL pairs and biologically coherent pathways (e.g., DNA repair, mitotic control).
• Biological Validation:
  • High-confidence predictions recovered known mechanisms (e.g., TP53-PARP1) and identified novel, biologically plausible candidates (e.g., AURKA-BUB1B) involved in DNA damage response and mitosis, even without prior CRISPR validation in the training set.

5. Significance

CILANTRO-SL represents a paradigm shift in synthetic lethality prediction. By moving away from static, curated interaction networks toward dynamic, context-conditioned representations derived from foundation models, it overcomes the "cold start" problem for novel genes. The integration of conformal prediction transforms the model from a simple ranking tool into a reliable decision-support system for experimentalists, providing statistically guaranteed confidence levels. This approach enables the scalable discovery of therapeutic targets in cancer contexts that were previously inaccessible due to data sparsity or lack of prior knowledge.