In May, Lhasa Limited announced the upcoming launch of the Federated Learning Hackathon, which proceeded to launch in June this year.
There are seven contributing pharmaceutical industry members of the ongoing Lhasa Federated Learning Hackathon. Alongside Lhasa, this group are aiming to:
- Successfully share knowledge across a diverse panel of nine secondary pharmacology endpoints, whilst preserving all confidential aspects of the data.
- Create predictive models using the Lhasa Federated Learning methodology, which outperform models trained with proprietary data only.
The Lhasa Federated Learning Hackathon recently reached the halfway milestone! At this point, each contributing member has shared knowledge, generated from their proprietary data, with Lhasa for the nine priority endpoints. The approach is facilitated by Effiris which through knowledge distillation, generates shareable data while preserving all confidential aspects of the data used.
The Molecular Informatics and AI team at Lhasa have consolidated the shared data from the participants and compared this data to the shared datasets from each individual party, to determine the levels of knowledge present. Already, some promising results have been observed, most notably:
- Knowledge has been extracted and shared across all endpoints and all contributing members.
- Models built on the consolidated federated data outperform, and have an increased applicability domain, compared to models built on individual contributor datasets.
The next phase of the Federated Learning Hackathon involves the contributing members evaluating the performance and added value of Effiris models and federated data, in comparison to existing in-house models, built solely on proprietary data.
To find out more about this forward-thinking scientific initiative, please get in touch.