acceptodds
Under review as a conference paper at ICLR 2027

Simple Siamese Networks on Foundation Model Embeddings Outperform Complex Models on Unseen Synthetic Lethality Prediction

Abstract

Synthetic lethality (SL) is a biological phenomenon whereby co-deletion of two genes results in cell death, whereas deletion of either has minimal effect, and is widely considered a promising route for cancer therapeutic target discovery Over the past decade, various machine learning methods have been proposed to predict synthetic lethal interactions from gene-level features. However, recent advances in AI for biology have produced a rich set of pretrained embeddings, each capturing distinct multi-modal representations defining genes and proteins, which have as-yet not been integrated into SL prediction models. Here, we propose a lightweight deep learning framework that uses multiple pretrained embeddings for the prediction of SL. On the largest database of SL interactions, we show such lightweight models achieve comparable performance to existing methods despite having orders of magnitude fewer trainable parameters, while the advantage of graph-based methods on previously seen genes disappears under strict gene holdout, likely due to existing SL methods leveraging the SL graph itself. We conduct a comprehensive ablation analysis, finding embeddings of curated knowledge generalize best, and several individual embeddings outperform naive multi-modal fusion. We further find integrating cellular context improves generalization capacity. Finally, we evaluate SL prediction on small focussed screens unseen during training, to our knowledge first evaluation to test whether such models could help actively guide future SL screens. We find poor zero-shot performance across all models, and evaluate fine-tuning strategies possible with lightweight models to improve predictive performance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biology, Genetic Interactions, Foundation Model Embeddings, Siamese Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8oBNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SLeuth, a lightweight Siamese network that uses a shared encoder to predict synthetic lethality via an inner product in a latent space over six categories of off-the-shelf multimodal pretrained representations. Under three settings of progressively harder generalization, the authors compare it against three complex baselines and find broadly comparable performance despite having orders of magnitude fewer trainable parameters.

Critical strengths

(1) It applies multimodal pretrained representations to synthetic lethality prediction and performs fusion with a minimal Siamese network, achieving performance competitive with complex graph models while using orders of magnitude fewer trainable parameters. This is a valuable direction with clear research significance.

(2) It systematically compares the performance differences of different modalities and different pretrained models on synthetic lethality prediction, providing empirical reference for subsequent representation selection.

Critical weaknesses

(1) The integration strategy of the method is relatively simple, and the architectural innovation is limited. Its core lies in integrating the pre-trained embeddings from six sources. The main integration methods include concatenation, standardization, and PCA dimensionality reduction, followed by learning through a lightweight encoder. There is no explicit modeling of the differentiated contributions of each data source to the synthetic lethal prediction, nor is there any innovation regarding the fusion mechanism itself. Table 2 shows that single-modal models such as go2vec and text_embed outperform multi-modal models in each Tier. This indicates that the necessity of multi-source fusion has not been fully demonstrated, and it also suggests that the gains may mainly come from the pre-trained representations themselves.

(2) There is a lack of analysis of potential issues that may exist in the integration of multi-source pre-trained embeddings. The embeddings from different models are not consistent in the feature space, semantic scale, and data distribution. Even with direct concatenation supplemented by matrix normalization and PCA, there may still be redundant information, semantic conflicts, alignment difficulties, and negative transfer. The paper merely mentions "PCA reduces redundancy" in a passing manner, without discussing semantic scale alignment and the risk of negative transfer, nor providing evidence to show that this fusion strategy has indeed alleviated these problems; and the multi-modal models perform worse than the single optimal modal in multiple Tiers, which may precisely be the manifestation of these problems.

Questions for authors

Table 5 reports only the mean and standard deviation of Spearman correlations across five fold models, without confidence intervals or significance tests, and most correlations are close to 0 or even negative. I suggest adding confidence intervals for the correlations and testing whether the differences between models are statistically significant, to establish the reliability of these results.

2 comments
AuthorsOpenReview ↗

Revision Changes

W1. Fusion is not justified, and the novelty is limited. Without PCA, the multimodal model is significantly better than every single embedding at Tier I and on par with the best ones at Tiers II and III. No single embedding is significantly better at any tier. On novelty, we see the paper as an application and benchmark of pretrained embeddings for SL prediction, and we kept the architecture simple on purpose, also due to the data size limitations as the effective sample size (i.e. the number of positive SL pairs) is not that large.

W2. Scale, redundancy and negative transfer. Our PCA standardized every feature, which discards the relative scale of the features. Without PCA, AUPRG is higher in most model–tier combinations, but significantly so only at Tier I for SLeuth and SLeuth-all. Since PCA also changes the model size, this does not isolate the effect of magnitude. Using all 16 embeddings gives results close to the six selected ones, lower only for SLeuth at Tier III, and removing any one category has no significant effect.

Q. Table 5 has no confidence intervals or tests. The table now shows, under each model, the Holm-adjusted bootstrap p-value against SLeuth, and marks correlations whose 95% CI excludes zero. The full CIs are in Appendix Tables. Because screen pairs share genes, this bootstrap is optimistic. A gene-level bootstrap gives larger standard errors, and some of SLeuth's correlations on unseen genes then no longer exclude zero (Appendix). We agree that most correlations are near zero or negative. Outside Adamson's growth phenotype, all are below 0.11 in absolute value. SLeuth is negative on that phenotype, which mostly reflects single-gene fitness, and positive on Adamson's GI score.

Thank you for the clarification from the author. If the author positions this paper as an application and benchmark test of pre-trained embeddings in synthetic lethal prediction, and intends to keep the architecture simple, then the current title, abstract, and contribution statement do not fully align with this positioning:

The current writing makes it more likely for readers to think that the main contribution of this paper is the proposal of a new method or architecture. Considering the limited novelty at the method level of this paper and the main value lies in the problem setting, the systematic evaluation of existing pre-trained embeddings/foundation models representations, and the examination of the gains brought by multimodal information under various settings through simple fusion experiments.

I suggest the author adjust the presentation framework of the paper: Weaken the emphasis on methodological innovation, and strengthen the contributions in the aspects of benchmark testing and problem definition. Specifically, in the title, abstract, introduction, and contribution list, it should be clearly stated that this paper is an application/benchmark study, explaining the scope of evaluation, evaluation protocol, and main findings, and presenting the simple fusion architecture as an empirical component rather than a core method innovation. This will enable the paper's contributions to be consistent with the writing style and present the value of this paper more accurately.

In addition, the author needs to consider redesigning Figure 1 and Figure 2: The current two figures contain insufficient information and fail to fully support the positioning of benchmark testing/application research; if the paper positioning is adjusted as per the above suggestions, the content and narrative of these two figures will no longer be applicable. In conclusion, I will not modify my rating.

Reviewer_22DYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SLeuth predicts synthetic lethality in a cell line by combining PCA-reduced frozen gene features with learned cell line embeddings in a Siamese encoder and symmetric bilinear head. On SynLethDB, the authors compare SLeuth to three baselines across three gene holdout tiers and benchmark 16 embeddings from six categories, then test transfer on three external screens. SLeuth matches the baselines with few trainable weights, though single embeddings often beat six-category fusion and external transfer stays weak.

Critical strengths

SLeuth is a lightweight baseline that matches SLMGAE on held-out genes with about 29,000 trainable parameters in Table 1, and it notes that frozen encoders are excluded. Its benchmark of 16 representations is the best part of the paper, showing in Table 2 that text features reach 0.841±0.0390.841\pm0.039 Tier III AUPRG and beat six-category fusion at 0.732±0.1360.732\pm0.136.

The evaluation also avoids common synthetic lethality pitfalls. It uses experimentally screened negatives and masks missing context labels while fixing splits across baselines, and reporting the weak external transfer in Tables 5 and 6 shows how badly random pair splits overstate generalization to unseen genes.

Critical weaknesses
  1. The baseline comparisons did not convince me that SLeuth beats the closest prior work, which along with the split leakage leaves me leaning toward a weak reject. The paper misses ELISL and Cilantro-sl. ELISL already pairs frozen SeqVec features with cell line omics, and Cilantro-sl tests biological foundation models with lightweight heads on gene holdouts, contradicting Section 1's claim that pretrained embeddings are untested here. The abstract also overstates claiming the first external screen evaluation, as Feng et al. already trained on SynLethDB and tested on an external Horlbeck K562 screen.
  2. Two baselines are weakened by input changes, and gains over the third are tiny. Section 4 and Appendix A replace the SynLethKG inputs of SLGNN and NSF4SL with ComplEx features from a two-relation STRING and GO graph, likely causing NSF4SL's 0.495±0.0400.495\pm0.040 Tier I AUROC. In Table 1, six-category SLeuth beats SLMGAE by just 0.0070.007 AUPRG in Tier II and 0.0190.019 in Tier III while losing on Tier III AUROC (0.6770.677 vs. 0.7050.705). Without paired tests or a simple bilinear baseline on the same features, I couldn't tell if the SLeuth head helps.
  3. My other main worry is that modality selection in Section 3.5 leaks information across splits. The authors pick the best modalities once using validation scores averaged over all five Tier I folds and reuse that choice across all folds and tiers, so each Tier I test fold contributed to feature selection and genes held out in Tier III shaped the chosen modalities.
  4. Section 5.3 and Table 6 omit external screen phenotype columns and sign conventions. In Corn, synthetic lethality means negative GEMINI scores near −1-1, so correlating positive SL probabilities with raw scores rewards inverted predictions. Table 6 also lacks a scratch baseline, and moving from random pairs to unseen gene pairs drops full-model correlation from 0.9740.974 to 0.4270.427 on Adamson, 0.4420.442 to −0.026-0.026 on Corn, and to 0.1050.105 on Gilbert.
Questions for authors
  1. Was Tier I modality selection disjoint from all test folds, and do results hold when selecting features inside each outer training split?
  2. How does SLeuth compare to ELISL and Cilantro-sl, along with a simple bilinear probe and unmodified SLGNN and NSF4SL?
  3. Which phenotype columns and sign conventions were used on external screens, what was the loss, and were SynLethDB overlaps removed?
  4. Does SynLethDB pretraining beat scratch training in Table 6, and how does retraining without cell line features compare to zeroing them in Table 4?
1 comment
AuthorsOpenReview ↗

Revision Changes

W1 / Q(b). We added ELISL, run on its context-free sources (SeqVec and PPI embeddings). It is below SLeuth at every tier, and Appendix lists what we changed to fit our shared folds. We added Cilantro-SL to related work, but we could not run it because we did not find its accessible data or model weights from its release. The abstract and introduction now say that pretrained embeddings have only recently begun to be used for SL prediction, and the "first evaluation" claim is gone. Feng et al. already used an external K562 screen (Gilbert). Since SynLethDB includes that study, we treat it only as an in-distribution check.

W2 / Q(b). With the cited paper authors' original inputs (SynLethKG for SLGNN and its TransE embeddings for NSF4SL), SLGNN has seen a lot of improvements. SLeuth is still ahead of both at every tier. We select the SLMGAE, SLGNN and NSF4SL checkpoints on validation AUPRG, as we do for SLeuth, rather than by each method's own rule (see the appendix). A bilinear probe on SLeuth's inputs is below SLeuth on every split and tier, although three of these gaps are not significant (appendix). SLeuth also beats SLMGAE on every split; p-values are reported in the paper now.

W3 / Q(a). The previous submitted selection averaged Tier I validation scores over the five splits, and each split's validation pairs are test pairs of the other splits. On top of that, about 72% of each Tier II and III test split lies in the Tier I training split with the same index (Appendix). Selection now uses only held-out training pairs from each split and tier, and using all 16 embeddings without any selection gives similar results.

W4 / Q(c). Appendix Table 8 lists the column we use from each screen. All targets are minus the raw phenotype, so a positive correlation is the correct direction. For Adamson we added the GI score, because its growth phenotype mostly reflects single-gene fitness. Fine-tuning uses an MSE loss. Pairs in SynLethDB are removed from all test sets.

Q(d). SynLethDB pretraining adds little with this much screen data. When both genes are unseen, full fine-tuning loses 66–91% of its correlation. Zeroing the cell-line embedding doesn't change performance much. Retraining without cell lines raises AUROC at every tier, significantly at Tiers I and II, but it also changes the head and the label, so it does not isolate the cell-line input. We have withdrawn the claim that cell-line context helps generalization.

Reviewer_84QLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • In this article, the authors present synthetic lethality (SL) prediction, emphasizing generalization to previously unseen genes. The authors propose SLeuth, a lightweight Siamese neural network that leverages frozen pretrained biological embeddings from multiple modalities, including protein sequences, gene expression, biomedical text, protein interaction networks, Gene Ontology, and knowledge graphs.

  • The framework incorporates cell-line-specific representations and a positive-definite similarity function to predict SL interactions using approximately 29,000 trainable parameters. Experiments on SynLethDB 3.0 compare SLeuth with established graph-based methods across three evaluation tiers with increasingly strict gene holdouts.

  • The paper also presents extensive modality and cell-line ablations and evaluates zero-shot transfer and fine-tuning on three external CRISPR screening datasets. The experimental results demonstrate that lightweight embedding-based models can achieve competitive performance against substantially larger graph-based approaches under unseen-gene evaluations, although multimodal fusion does not consistently outperform individual embeddings and zero-shot generalization to independent screening datasets remains challenging.

Critical strengths
  • The paper addresses an important and practically relevant challenge in synthetic lethality prediction by emphasizing inductive generalization to previously unseen genes.
  • The proposed SLeuth architecture provide a computationally efficient alternative to complex graph-based methods, using substantially fewer trainable parameters while achieving competitive performance under challenging gene-holdout settings.
  • The evaluation on three independent CRISPR screening datasets, together with experiments on zero-shot prediction and subsequent fine-tuning, provides useful insights into the limitations and adaptability of pretrained biological representations for synthetic lethality prediction.
Critical weaknesses

There are some weaknesses that I want more clarification on from the authors:

(1) The paper concluded that lightweight Siamese networks outperform more complex models on unseen synthetic lethality prediction, but this is not consistently supported by the reported results.

  • For example, in Table 1, under the Tier III setting, SLeuth achieves a lower AUROC than SLMGAE (0.677 ± 0.031 vs. 0.705 ± 0.038), although it obtains slightly higher AUPR (0.419 vs. 0.371) and AUPRG (0.732 vs. 0.713).
  • The relatively large variability across the five evaluation splits makes it difficult to establish a robust performance advantage. It would be great if the authors could provide additional paired statistical comparisons to clarify whether the observed differences are meaningful.

(2) The proposed multimodal framework does not consistently outperform simpler single-modality alternatives, particularly under strict unseen-gene evaluation. In Table 2, text-only embeddings achieve a Tier III AUPRG of 0.841 ± 0.039, substantially exceeding the full six-modality model's 0.732 ± 0.136. In experiments, it remains unclear whether the degradation primarily arises from PCA dimensionality reduction, inappropriate modality selection, or interactions among heterogeneous biological representations.

(3) Finally, I believe that the tier-specific feature selection using properly separated training and validation partitions would provide a more reliable assessment of the proposed approach under different generalization settings.

Questions for authors

There are several questions for the authors:

  • Could the authors investigate why text-only embeddings substantially outperform the full multimodal model under Tier III evaluation, particularly by separating the effects of PCA dimensionality reduction and modality selection?
  • Given the poor zero-shot transfer results on external CRISPR screens, could the authors provide the main sources of distribution shift, and how much target-screen supervision is necessary to achieve reliable performance under strict unseen-gene evaluation?
1 comment
AuthorsOpenReview ↗

Revision Changes

W1. Advantage over SLMGAE and paired comparisons. The revised SLeuth is now ahead of all six other methods on every metric, tier and split. Holm-corrected paired t-tests are reported in the revision.

W2 / Q(a). Single embeddings vs fusion, and PCA vs selection. We now test the two separately: the same embeddings with and without PCA, and the selected embeddings against all 16. Without PCA, no single embedding is significantly better than the multimodal model at any tier. At Tier III, some single modality models score higher, but by less than the standard deviation across splits. With PCA, 9 of the 16 embeddings beat the multimodal model at Tier III, same phenomenon present in our previous submission. Removing PCA closes most of that gap, though not significantly. Without PCA, SLeuth-all is still below those 9 embeddings, while SLeuth-all-CL beats every single embedding at Tier III.

W3. Tier-specific selection on separate validation data. Selection now runs separately for each split and tier on held-out training pairs. At Tiers II and III, though, these are held-out pairs rather than held-out genes, which ranks embeddings poorly for unseen genes. The all-16-modalities models avoid selection altogether.

Q(b). Sources of distribution shift and how much screen data is needed. For Adamson and Corn, the screen phenotypes barely separate SynLethDB's SL pairs from its non-SL pairs. Agreement is better for SL pairs reported in the screen's own cell line, although for the Adamson GI score the difference is not significant under a gene-level bootstrap. We see cell-line context and the phenotype being measured as the main shifts. Training from scratch on the screens gets close to fine-tuning.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

  • We no longer apply PCA to the multimodal input.
  • Embeddings are now selected separately for each split and tier, on a random 10% of the training pairs held out from training. SLeuth-all and SLeuth-all-CL skip selection and use all 16 embeddings.
  • We now report two variants: SLeuth, the pooled model, predicts whether a pair is SL in any cell line, and SLeuth-CL keeps the cell-line heads.
  • SLGNN and NSF4SL were rerun with the authors' own inputs, and we added ELISL, a bilinear probe and a degree prior.
  • On SynLethDB we use paired t-tests across splits, and on the screens a bootstrap, both with Holm correction. Results from the tables mark significant differences, and the appendix gives p-value.
  • The model changed after submission. It is wider (32 instead of 16), has no layer normalization, learns τ directly, and adds a per-head offset, a per-gene term and a pair term. With PCA also removed, SLeuth has more parameters but still orders of magnitude fewer than SLMGAE.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.