S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search

Summary

The paper introduces S-MolSearch, a framework for ligand-based virtual screening in drug discovery that addresses the challenges of limited and noisy binding affinity data. By utilizing molecular 3D information and semi-supervised contrastive learning, S-MolSearch processes both labeled and unlabeled data to train molecular structural encoders and generate soft labels for the unlabeled data, drawing on inverse optimal transport principles. The framework outperforms existing structure-based and ligand-based virtual screening methods, as evidenced by its superior performance on the LIT-PCBA and DUD-E benchmark datasets.

Strengths

- Well-written - Well-organized experimental settings and comparison methods

Weaknesses

- There is a lack of discussion on the reasons behind the performance differences and improvements, with only numerical comparisons of the experimental results. - There is insufficient experimentation and consideration regarding the time required for virtual screening. - There are no results for experimental metrics such as AUROC or BEDROC, which were used in previous studies.

Questions

- Both S-MolSearch and existing methods experience a decline in performance as the % of EF increases. Additional discussion on the reasons for this phenomenon is needed. - Why do soft labels based on inverse optimal transport seem to have a significant impact on the DUD-E dataset but a lesser impact on the LIT-PCBA dataset? - What aspects of the semi-supervised approach in Table 4 do the authors think primarily contributed to the performance improvement compared to fine-tuning? - Is it possible to extend this method from a zero-shot setting to a few-shot setting? If so, how do the authors think its performance would compare to existing methods in that case? - In virtual screening, not only performance but also processing time is important. How does the screening time compare to that in existing studies? - How does the performance compare to existing models when using measurements like AUROC or BEDROC instead of EF #%?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper addressed limitations and potential social impacts in Appendix.

Summary

The paper introduces "S-MolSearch," a semi-supervised contrastive learning framework designed for ligand-based virtual screening in drug discovery. This framework uniquely leverages labeled binding affinity information to produce soft labels for unlabeled molecules, integrating 3D molecular structures and binding affinity data. The paper also proposes a novel semi-supervised learning paradigm that combines contrastive learning with Inverse Optimal Transport (IOT).

Strengths

1. The supervision idea is novel and useful, and the target application is very impactful with broad implications. 2. The paper is well-written and the experiments are comprehensive.

Weaknesses

1. Memory Consumption Concerns: The model employs a parallel architecture with two f_\theta encoders and one g_\phi encoder, based on the Uni-Mol framework. Although utilizing pretrained models has shown significant performance benefits, the paper should address potential memory management strategies, especially for future applications involving molecules with a greater number of atoms. 2. Utilization of 3D Structures: The paper promotes a novel semi-supervised contrastive learning paradigm, yet the core contribution does not seem to revolve around the innovative use of 3D structures, as this capability primarily stems from the Uni-Mol architecture. It would be beneficial if the authors could clarify any specific enhancements made to ensure the effective preservation and utilization of geometric information within the model. Absent such enhancements, clearer distinctions should be made regarding the role of 3D structures to prevent misconceptions about the paper presenting a new geometric deep learning technique. 3. Clarity in Section 3.4: The explanation of how $\Gamma$, which approximates the distribution of $C$ under constraints from $U(p,q)$, relates to the continuous optimal transport problem is not clear. Moreover, the motivation and necessity of soft labels, beyond experimental justifications, needs further elaboration. The section would benefit from additional visual aids or high-level descriptions, akin to the clarifications provided in sections 3.3 and 3.5, to aid in comprehension. 4. Component Efficacy in Table 3: There appears to be a discrepancy in the impact of model components across different benchmarks—soft labels are pivotal for DUD-E, whereas pretraining is more crucial for LIT-PCBA, with soft labels showing minimal importance. Insights into this inconsistency would be valuable. Furthermore, an evaluation of how the Uni-Mol encoder alone performs on these tasks would provide additional context on the effectiveness of the proposed enhancements Minor points and typos: L153-154 is not clear. L162: It would be beneficial to include illustrations of $M_{sup}$ in the figures for clarity. Formula 2 and L 168: It is better to give intuitive explanations of $1_N$. L184: Inconsistent notation. $g(\psi)$ or $g_\psi$? L281: Misplaced comma.

Questions

The same as weakness.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The focus of the paper is predominantly on molecule binding affinity, heavily relying on a pretrained encoder. This reliance could limit the model's applicability across a broader spectrum of bioinformatics data. A more detailed discussion on the dependency on pretrained models and potential strategies to mitigate this limitation would enhance the paper's breadth and applicability.

Summary

This paper proposed a Ligand-based Virtual Screening method S-MolSearch. which can leverages molecular 3D information and affinity information in semi-supervised contrastive learning.

Strengths

1. The method is able to leverage both labeled and unlabeled data simultaneously and achieves excellent performance on DUDE and Lit-PCBA benchmarks. 2. The approach of using the principles of inverse optimal transport for semi-supervised learning is quite innovative and worth adopting. 3. The ablation experiments are sufficient, and the experimental section is quite robust.

Weaknesses

1. In the method section, it is unclear to me whether during inference only encoder$g_{\psi}$ is used, or both $\psi$ and $f_{\theta}$ are used simultaneously? 2. If the application scenario involves a newly provided protein without reference molecules, how should ligand-based virtual screening methods handle this situation?

Questions

Refer to weakness.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

S-MolSearch predominantly focuses on the molecular affinity data, omitting broader biochemical interactions, which suggests a potential area for improvement.

Summary

The paper introduces a new method for ligand-based virtual screening based on contrastive learning and inverse optimal transport. Two molecule encoders are trained. The first encoder is trained using a contrastive loss function on the ChEMBL data by pairing compounds that are active toward the same protein, and compounds active toward different targets are treated as negative pairs. Next, the second encoder is trained by using the pseudo-labels produced by the first model. The proposed model is tested on two benchmark datasets, DUD-E and LIT-PCBA. Additionally, an ablation study is conducted, and the impact of the labeled data scale is visualized.

Strengths

Originality: - The approach seems to be novel. I have not found any similar papers that use optimal transport for the ligand-based virtual screening task. Quality: - The theory described in the paper is formally proven in the Appendix. - The proposed method obtains excellent results in both tested benchmarks. - The quality of the learned representation is demonstrated in Figure 2. Clarity: - The paper is written clearly and is easy to follow. - Figure 1 shows the idea of the model very clearly. Significance: - The presented method is an interesting and effective way to utilize all the available public data to build a strong model for ligand-based virtual screening.

Weaknesses

Quality: - It would be interesting to see some qualitative examples of molecules that were found to be similar to the active compounds in the virtual screening process. Do the trained similarities correlate with the Tanimoto similarity? Clarity: - Does the “sup” subscript in Section 3.4 correspond to the “label” subscript in Proposition 1? What is the difference between these two sets? Minor comments: - A typo in line 151, “we employs InfoNCE.” - In line 183, something is missing before “1”.

Questions

1. How do you solve the cases in contrastive learning where one molecule binds to multiple targets? In this example, you need to be careful not to create a negative pair, where one molecule is the one binding to both targets and the other molecule binds to only one of them. 2. How do you avoid treating two molecules as a negative pair if both can bind to the same target? For example, they are inhibitors of two similar proteins, which increases the chance of them binding to both at the same time.

Rating

6

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

The limitations have been described.