Lhasa Limited shared knowledge shared progress

Effiris

A secondary pharmacology model suite leveraging value from federated learning

  • Effiris is a qualitative model suite for predicting the secondary pharmacology effects of your compounds.
  • Effiris contains models which have used privacy-preserving knowledge transfer to learn from multiple sources of proprietary data. 
  • Effiris will enable more efficient, informed triage and prioritisation of compounds.
Effiris
Effiris Publications

Insight

Why do you need Effiris?

Evaluation

Why Choose Lhasa?

Get Involved

References

 

Features and Benefits

The Teacher - Student Approach (for Effiris Members only)

 

Insight

Effiris is a suite of secondary pharmacology models which facilitate more efficient triage of your compounds, as well as more informed compound design.

Effiris will leverage artificial intelligence (AI) to allow the transfer of knowledge from multiple pharmaceutical partners into the federated models. Member data will always be kept confidential, thanks to privacy-preserving knowledge transfer methods, but all users of Effiris will benefit from the shared data.

Therefore, Effiris will aid secondary pharmacology screening through this privacy-preserving knowledge transfer.

Lhasa is already in the process of building Effiris models using privacy-preserving knowledge transfer and has delivered its first round of target models to the Effiris consortium members.

Aims of the project:

Lhasa is always looking for new, innovative ways to add value for its members and Effiris is another step in this direction. Effiris will facilitate more efficient, informed triage and prioritisation of compounds. It will also assist medicinal chemists who are working in compound design, and help to focus downstream testing efforts.

History of the project:

The privacy-preserving element of the Effiris project was a concept that was proven through a collaborative research effort during 2018. This effort was initially named the Cronos Proof of Concept Project (publication pending) and demonstrated positive results. The project developed hERG SOHN (self-organising hypothesis network) models which were trained on private data1. These models were further improved once they had the ability to learn from multiple sets of proprietary data through semi-supervised Knowledge Transfer2. These models had a very strong performance, see below.

The intention of the proposed Effiris collaborative consortium is to evolve this research; applying the techniques to other secondary pharmacology targets and moving towards building a suite of Teacher - Student models over the course of the three-year project and beyond. This will result in a suite of secondary pharmacology models, as part of Effiris. 

 

Why do you need Effiris?

  • Effiris will be an essential tool for teams within organisations who work in discovery toxicology and who are looking to streamline their secondary pharmacology screening.
  • Effiris will help you to design the most low-risk compounds which can be prioritised for further development.
  • For those in pre-clinical safety, Effiris will help you to better understand unexpected off-target effects observed in animal studies. 

Back to top

 

Evaluation

A validation was carried out for the original hERG Student model. This validation used the dataset as described by Preissner et al.3 to compare the Teachers and the Student model.

The Teacher models created from this initial proof of concept project had performances ranging from 0.270 to 0.539* (Figure 1). Not only did the Student model outperform the Teacher average, it had a higher performance than any individual Teacher model. The Student model also had a very good balanced accuracy of 81% (Figure 2).

Figure 1: A comparison of the performance of the Student hERG model with its Teacher models*. 

*MCC – Matthews correlation coefficient. Matthews, B. W. (1975)4

 

 

A

F

D

H

E

C

G

B

<Teacher>

Student

MCC

0.539

0.325

0.411

0.270

0.364

0.468

0.321

0.472

0.396

0.546

Kappa

0.529

0.309

0.389

0.256

0.339

0.468

0.255

0.457

0.375

0.499

BAcc

0.793

0.634

0.731

0.610

0.643

0.737

0.683

0.704

0.692

0.813

Recall +

0.772

0.353

0.841

0.310

0.350

0.615

0.854

0.473

0.571

0.911

Recall -

0.814

0.916

0.722

0.910

0.937

0.858

0.511

0.934

0.825

0.714

Precis +

0.582

0.584

0.472

0.535

0.650

0.593

0.370

0.707

0.562

0.517

Precis -

0.914

0.808

0.893

0.797

0.811

0.869

0.913

0.841

0.856

0.960

Coverage

0.587

0.392

0.454

0.328

0.200

0.442

0.307

0.706

0.427

0.774

 

Figure 2: Evaluation metrics for the Student model.  

  • The Student could improve precision (52%)
  • Very good balanced accuracy (81%)
  • Excellent recall (>91%)
  • Excellent negative precision (96%). 

Back to top

 

Why Choose Lhasa?

Lhasa Limited - an honest broker and trusted holder of data

Lhasa is a not-for-profit organisation and believes that shared knowledge can lead to shared progress. Recognised as the original "Honest Broker", Lhasa Limited has repeatedly been trusted with proprietary data and this can be seen with our involvement in other collaborative data sharing projects.

Get Involved

A consortium has been initiated in order to apply the knowledge transfer methodology to create Effiris. A second round of collaborative research is already in progress.

Lhasa is open to new members and is actively encouraging participation. Please contact the Global Alliances team for more information.

 

References

  1. T. Hanser et al. (2019) ‘Avoiding hERG-liability in drug design via synergetic combinations of different (Q)SAR methodologies and data sources: a case study in an industrial setting’, Journal of Cheminformatics, vol. 11, no. 9.
    https://www.lhasalimited.org/publications/avoiding-herg-liability-in-drug-design-via-synergetic-combinations-of-different-qsar-methodologies-and-data-sources-a-case-study-in-an-industrial-setting/5233 
  2. N. Papernot et al. (2017) ‘Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data’, ICRL Conference 2017.
    https://arxiv.org/abs/1610.05755
  3. R. Preissner et al. (2018) ‘The Catch-22 of Predicting hERG Blockade Using Publicly Accessible Bioactivity Data’, Journal of Chemical Information and Modeling, vol. 58, no. 6, pp.1224-1233.
    https://doi.org/10.1021/acs.jcim.8b00150
  4. B. W. Matthews. (1975) ‘Comparison of the predicted and observed secondary structure of T4 phage lysozyme’, Biochimica et Biophysica Acta (BBA) - Protein Structure, vol. 405, no. 2, pp.442–451.
    https://doi.org/10.1016/0005-2795(75)90109-9

Back to top

© 2019 Lhasa Limited | Registered office: Granary Wharf House, 2 Canal Wharf, Leeds, LS11 5PS, UK Tel: +44 (0)113 394 6020
VAT number 396 8737 77 | Lhasa Limited is registered as a charity (290866)| Company Registration Number 01765239 (England and Wales).

Thanks to QuestionPro's generosity, we now have survey software that powers our data intelligence.