Protein Ligand Binding Affinity Prediction (Structure based AI)
In silico structure-based drug design combined with accurate binding affinity prediction can significantly reduce both time and costs needed for research projects. Via the Apheris Platform for federated and privacy preserving data science, universal scoring functions can be run. Using these functions, multiple companies can jointly train generalizable and interpretable protein-ligand interaction models in an IP-preserving way.
Situation
Two pharma companies both invest heavily in their own in silico drug discovery pipelines. Both parties primarily rely on their own in-house data to train AI models. Accurate models to predict protein-ligand binding affinity are essential for accelerated drug discovery.
Problem
Neither pharma company owns sufficiently large and unbiased datasets. Existing AI models within the in silico drug discovery pipeline have been trained on small datasets resulting in low generalizability, high rates of overfitting, and poor information output. Though the two pharma companies possess complementary data, sharing is not possible due to data-inherent intellectual property of the competing companies.
Apheris solution
Apheris offers a federated and privacy preserving system for training protein-ligand interaction (PLI) models. Companies can jointly train a generalizable and interpretable protein-ligand interaction model in an IP-preserving manner via the Apheris Platform.
Advantages of using Apheris
Highly generalizable models
Using the Apheris Platform, the two pharma companies can jointly train best-in-class PLI-models on their complementary data, resulting in highly generalizable models with strong interpretability
Novel drug candidates
A higher discovery-rate of superior drug candidates by our big data approach leads to accelerated and more successful drug development
New revenue streams
A modular platform architecture allows the pharma companies to offer PLI-prediction as a service without disclosing any of the parties’ data or intellectual property
Federated Drug Discovery: Quantitative Structure Activity Relationship (QSAR)
The Apheris Platform for federated and privacy preserving data science enables multiple pharmaceutical partners to jointly train AI models and discover novel molecular structure-property insights (Quantitative Structure-Activity Relationship/QSAR models).
Situation
Four drug research institutions have relatively small datasets at their disposal. Their individual data is strongly biased to the chemical and bioactivity space the organization operates in. As a result, models do not generalize well and tend to perform weakly on data from external sources. While the datasets of the individual institutions are highly biased, they are complementary to each other. Synergetic combinations of those data sources will lead to improved performance and generalizability of models.
Problem
The raw data needed for training the models is highly confidential as it resembles the institutions’ key competitive advantage. Thus, open sharing and combining the data between the competing companies is not possible.
Apheris solution
The Apheris Platform enables a privacy-preserving data ecosystem in which each institution can train models on complementary data of the other parties. Training is decentralized and data is always under the full control of the data owners. Only encrypted model parameters are shared in a privacy preserving manner. Our Platform allows each institution to access and compute on complementary datasets, streamlining their drug discovery pipeline.
Advantages of using Apheris
Best-in-class models
Using the Apheris Platform, the four pharma companies can jointly train best-in-class QSAR-models providing superior predictive value
Broad applicability domain
Models trained on large and diverse data of multiple parties lead to highly generalizable models with broader applicability
Multi-target modelling
The complementary data of multiple parties opens up the possibility to model multi-target behavior
Novel scientific insights
The consensus models can add significant value in gaining novel molecular structure-property insights that drive drug design decisions
Deep Multi-Task Learning on Heterogenous Drug Discovery Data
Automated in silico drug discovery pipelines guide and speed up the drug discovery and development process of pharma and biotech companies. Commercial biological and medical datasets are typically sparse, and machine learning models that are trained to predict bioactivity do not generalize well. Via the Apheris Platform, multiple companies jointly train multi-task models on heterogenous data of different modalities. The diverse data leads to a better representation of drug compounds and their biological properties. Apheris’ federated machine learning framework enables model training without revealing key secrets of each pharma company and enables faster and more accurate drug discoveries.
Situation
One pharma company and one biotech company possess drug discovery data of different modalities that complement each other. The biotech company owns drug target interaction data while the pharma company owns in vitro drug response data. Both companies wish to advance their discovery efforts via robust AI methods that allow a deeper understanding of drug specific response mechanisms.
Problem
Both companies face difficulties in developing robust and generalizable models, as their datasets are highly biased due to limited data variability. The current data science initiatives fail to give helpful guidance for the researchers on which molecules to focus on. A combination of the companies’ data would significantly enhance the model’s informative value, but openly sharing data is not possible due to strong IP concerns on both sides.
Apheris solution
The Apheris Platform for federated and privacy preserving data science allows training AI models on the heterogenous datasets of the pharma and the biotech company while preserving data confidentiality. A sophisticated model architecture enables joint learning of a global model across different tasks. The neural networks consist of two components with base layers that are trained via the Apheris Platform, and top layers for each task are tailored to the specific datasets with different modalities.
Advantages of using Apheris
Highly generalizable models
Highly generalizable multi-task methods that can make predictions for unseen drugs based on their molecular structures
Increased model performance
Significant increases in model performance and usability by leveraging heterogenous input datasets
New scientific insights
Discovery of novel molecular structure-biological effect relations due to unbiased big-data approach that fuels drug discoveries
New partnerships
Potential follow-up initiatives for joint research between the partnering companies to further improve and accelerate drug discoveries