1 of 18

SafeDrug: Dual Molecular Graph Encoders for Recommending Effective and Safe Drug Combinations

IJCAI 2021

Chaoqi Yang1, Cao Xiao2, Fenglong Ma3, Lucas Glass2, Jimeng Sun1

1UIUC, 2IQVIA, 3Penn State University

1

2 of 18

Outline

  • Background
  • Problem Formulation
  • SafeDrug Model
  • Experiment
  • Conclusion

2

3 of 18

Problem: Drug Recommendation

3

diagnosis

procedure

drugs

diagnosis

procedure

drugs

Visit 1

Visit t-1

……

History information

diagnosis

procedure

Visit t

diagnosis

procedure

drugs

Visit 2

???

Can we predict the drug combinations at visit t?

  • Previous models: RNN + predictive model
  • Two challenges:
    • How to utilize the drug molecule information
    • How to control drug-drug interaction (DDI)

4 of 18

Outline

  • Background
  • Problem Formulation
  • SafeDrug Model
  • Experiment
  • Conclusion

4

5 of 18

Formulation

  •  

5

6 of 18

Outline

  • Background
  • Problem Formulation
  • SafeDrug Model
  • Experiment
  • Conclusion

6

7 of 18

Overall Framework of SafeDrug

7

8 of 18

Patient Representation

8

Embedding table

Embedding vector

Historical information encoding

Patient representation

9 of 18

Molecule Graph Encoders

9

Global MPNN encoder

Local bipartite encoder

Initial atom embedding

A bipartite mask from BRICS

breaking retro-synthetically interesting chemical substructures (BRICS)

10 of 18

Medication Representation

10

Medication representation

Loss design

PID controllable loss function

11 of 18

Outline

  • Background
  • Problem Formulation
  • SafeDrug Model
  • Experiment
  • Conclusion

11

12 of 18

Data set and Baselines

  • MIMIC-III dataset

​

​

​

​

  • Baselines
    • LR, ECC, RETAIN, LEAP, DMNC, GAMENet
  • Model variants

​

​

​

​

12

13 of 18

Performance Comparison

​

​

​

​

​

​

  • Our SafeDrug model outputs effective drug recommendation results with low DDI rate.

​

​

​

13

14 of 18

Model complexity comparison

​

  • Our SafeDrug model contains fewer parameters, and the training and inference is also fast.

​

​

​

​

​

​

​

​

14

15 of 18

Pre-control the DDI rate

  •  

15

16 of 18

Outline

  • Background
  • Problem Formulation
  • SafeDrug Model
  • Experiment
  • Conclusion

16

17 of 18

Conclusion

  • This paper exploits molecule global connectivity by MPNN encoder and local substructure functionality by the bipartite encoder.
  • We combine the molecule structure information with patient health status to enhance the drug recommendations.
  • We consider multiple objectives and formulate an adaptive loss function by PID controller, which enables the model to focus on reducing DDI among the recommended drugs.
  • We conduct extensive experiments on MIMIC-III dataset and show that our model can outperform the considered strong baselines.

17

18 of 18

SafeDrug: Dual Molecular Graph Encoders for Recommending Effective and Safe Drug Combinations

Thanks for Listening!

Chaoqi Yang1, Cao Xiao2, Fenglong Ma3, Lucas Glass2, Jimeng Sun1

1UIUC, 2IQVIA, 3Penn State University

18