Huang, Kexin, et al. "MolTrans: Molecular Interaction Transformer for drug–target interaction prediction." Bioinformatics 37.6 (2021): 830-836.

1 Introduction

Challenges

Inadequate modeling of interaction mechanism

기존의 molecular representation은 전체적인 구조를 기반으로 하지만, 실제로 interaction이 일어나는 부분은 전체에서 일부임. 또 전체 구조를 기반으로한 representation은 해석하기 어려움

Restricted to limited labeled data

Drugs, proteins 각각의 수에 비해 drug-protein pair에 대한 label은 부족함

Contributions

Knowledge inspired representation and interaction modeling for more accurate and explainable prediction

• NLP에서 Ngram과 같은 방식인 Frequent Consecutive Sub-sequence (FCS) mining을 통해 sequence를 sub-structure로 나누어서 사용

Leverage massive unlabeled biomedical data

• 많은 수의 데이터를 사용함 (당연한 얘기를..)
• transformer를 이용한 augmentation (..???????)

2 Materials and methods

2.1 Problem definition

• DTI prediction
  • Drug is represented by SMILES
  • Target protein is represented by a sequence of amino acids

2.2 The MolTrans methods

Figure 1. MolTrans workflow

2.2.1 FCS mining module

• FCS aims to generate a set of hierarchy of frequent sub-sequences for sequences.
• The significance of FCS is threefold:
  • Fingerprint와 같은 encoding 방식은 intractable하지만 FCS는 discrete하고 moderate한 사이즈의 sub-structure로 explicit하게 나누기 때문에 explainable함
  • Massive unlabeled data를 다루기에 적합한 방식
    • Small dataset에서는 주요한 sub-structure의 frequency가 낮게 나오는 경향이 있음
  • FCS can capture fundamental and meaningful biomedical semantics.
    • 서로 다른 특성을 가진 데이터셋에서도 공통적으로 발견되는 sub-structure가 있었음

2.2.2 Augmented transformer embedding module

• Contextual information을 위해 transformer encoding layer를 사용
  • 이 이상의 설명은 없는데.. multi-head 요소가 augmentation의 효과를 주는 것이라고 주장하는 것 같음

2.2.3 Interaction prediction module

Pairwise interaction

(수식 작성), I dimension 설명
To provide explainability, we favor dot product as the aggregation function because it generates a single scalar that explicitly measures the intensity of interaction between individual target-drug sub-structural pair. As dot product output is one-dimensional for every pair, I becomes a two-dimensional interaction map.

문제가 발생하는 지점

Neighborhood interaction

Interaction module에 CNN을 적용함

2.3 Implementation

특별한 내용이 없어서 생략

3 Result

3.1 Experimental setup

Dataset

Metrics

• ROC-AUC, PR-AUC, sensitivity (F1 score가 가장 높은 threshold 사용)

Evaluation strategies

• Training: Validation: Test = 7:1:2
• 5 x random split

3.2 Baselines

• LR, DNN, GNN-CPI(), DeepDTI(), DeepDTA(), DeepConv-DTI()

3.3 Q1: MolTrans achieves superior predictive performance

3.4 Q2: MolTrans has competitive performance in unseen drug and target setting

we randomly select 20% drug/target proteins and all DTI pairs associated with these drugs and targets as the test set.

3.5 Q3: MolTrans performs best with scarce data

We trained each method on 5%, 10%, 20% and 30% of dataset and predict on the rest of them (we use 10% of the test edges as validation set for early stopping)

3.6 Q4: MolTrans is robust in various protein families

we test on the predictive performance on four of the largest druggable targets: enzymes, ion channels, G-protein-coupled receptors (GPCRs) and nuclear receptors. We retrieve one test set of BIOSNAP and map the target proteins to the four protein families using GtoPdb database

3.7 Q5: MolTrans allows model understanding

Interaction map에서 threshold를 통해 0/1로 나누면 1은 interaction과 관련된 sub-structure의 pair

근데 이게 진짜 이상함. 후술하겠지만 패딩은 빼버리고 보여주는데.. 패딩을 포함하면 이상하게 나옴

3.8 Q6: Ablation study

다소 뻔한 얘기라서 생략

4 Conclusion

MolTrans 짱 좋다고 주장