SLxGO

Synthetic lethality (SL) — where disrupting two genes together kills a cell that tolerates loss of either one alone — is one of the most actionable ideas in precision oncology, the principle behind PARP inhibitors in BRCA-mutant cancers. But the combinatorial space of human gene pairs exceeds 200 million, so exhaustive experimental screening is infeasible and computational prioritization has become essential. The dominant machine-learning approaches lean heavily on curated protein-protein interaction (PPI) networks, which cover only about 7,500 proteins and bias predictions toward well-studied genes, effectively excluding roughly 13,000 human protein-coding genes.

SLxGO, developed at the Indian Institute of Science Education and Research Thiruvananthapuram, removes that dependency. Rather than reading SL signal off an interaction graph, it derives functional relationships between genes directly from the semantic content of their Gene Ontology (GO) annotations, encoded with a frozen BioBERT language model. This lets it score any human gene pair — including pairs where one or both genes never appear in a PPI database — and generalize to genes unseen during training.

The framework is described in a 2026 bioRxiv preprint and, following the Feng et al. benchmarking protocol, outperforms eight state-of-the-art SL predictors across multiple cross-validation and negative-sampling schemes, holding up in particular under the cold-start conditions where network-based methods degrade.

#Key Features

• Network-independent prediction: SL is inferred from GO-annotation semantics rather than a curated PPI graph, extending coverage from the ~7,500 proteins in STRING to the full protein-coding genome.
• BioBERT-encoded gene representations: Six annotation fields per gene — Wikigenes symbol, PANTHER protein class and family, molecular function, biological process, and cellular localization — are embedded with a domain-specific BERT and concatenated into a fixed 4,608-dimensional vector.
• Learned network-feature approximation: A neural network maps the GO embeddings onto the 49 topological features the SLant algorithm computes from the PPI network, so genes absent from that network still receive network-like descriptors.
• Cold-start generalization: Under the double-gene cross-validation split, where neither gene of a test pair appears in training, SLxGO maintains ranking performance — the regime that matters for poorly characterized genes.
• Context-specific predictions: Integrating cell-line-specific transcriptional profiles extends the framework to context-dependent SL across six cell lines.

#Technical Details

For each gene, the six GO annotation fields are passed through BioBERT and concatenated into a 4,608-dimensional vector; a gene pair yields a 9,216-dimensional representation. A separate artificial neural network — two hidden layers of 2,048 and 1,024 ReLU units, trained for 10 epochs under a mean-squared-error objective — predicts the 49-dimensional SLant network-feature vector for any pair, including those outside the 4,301-gene PPI network. The GO embeddings are compressed to 50 dimensions per gene by PCA, concatenated in both gene orders with the predicted network features into 149-dimensional pair embeddings, and classified by a histogram-based gradient-boosting model whose forward- and reverse-order scores are averaged into a confidence value. The release model was trained on a balanced set of 26,220 experimentally validated SL pairs from SynLethDB against 26,220 negatives estimated from DepMap transcriptomics. Benchmarking against GRSMF, SL2MF, GCATSL, DDGCN, SLMGAE, NSF4SL, KG4SL, and PTGNN used AUROC, AUPR, and F1 alongside the ranking metrics NDCG@10, Precision@10, and Recall@10 across three cross-validation splits.

#Applications

SLxGO is built for nominating context-specific therapeutic targets in cancer, where an SL partner of a tumor-associated mutation is a candidate drug target. Because it scores gene pairs regardless of prior interaction evidence, it is especially useful for poorly characterized genes that network-based tools cannot reach. The authors experimentally confirmed a context-specific EFNA1-SLC29A1 SL interaction in HeLa cells and pharmacologically validated a synergistic ACVR1-SLC29A1 vulnerability, and they released all genome-scale predictions through SLiGO, an open-access database spanning 30 million human gene pairs that researchers can query to prioritize experimental follow-up.

#Impact

By decoupling SL prediction from PPI-network coverage, SLxGO opens roughly 13,000 previously inaccessible protein-coding genes to systematic vulnerability mapping and reports the first computational prediction plus experimental validation of the EFNA1-SLC29A1 and ACVR1-SLC29A1 pairs. Its practical standing carries caveats: the work is a preprint that has not completed peer review; the pipeline is distributed as GPL-3.0 Jupyter and R notebooks — with the trained gradient-boosting and neural-network components reproducible from them — rather than a packaged, downloadable checkpoint; and the primary access point for predictions is the SLiGO web server rather than local inference.