2nd Place Winner – Tox24 Challenge
Challenge: Prediction of compounds activity against Transthyretin (TTR)
​
Organisers: Marie Sklodowska-Curie Innovative Training Network European Industrial Doctorate grant agreement No. 956832 “Advanced machine learning for Innovative Drug Discovery” (AIDD), Horizon Europe Marie SkÅ‚odowska-Curie Actions Doctoral Network grant agreement No. 101120466 “Explainable AI for Molecules” (AiChemist) as well as by Chemical Research in Toxicology and ICANN2024.
​
Our Approach:
-
Advanced Graph Neural Networks – Proprietary architectures tailored for molecular data
-
Multitask Learning – Leveraging correlated endpoints for improved accuracy
-
Bioactivity-Informed Descriptors – Enriching models with experimental context
-
Consensus Modeling – Combining strengths of multiple predictors for robust results
​
Results: Shared 2nd Place among 78 groups and companies worldwide
​
References:
-
Consensus Modeling Strategies for Predicting Transthyretin Binding Affinity from Tox24 Challenge Data. Cirino et al., DOI: 10.1021/acs.chemrestox.5c00018
-
Which modern AI methods provide accurate predictions of toxicological endpoints? Analysis of Tox24 challenge results. Eytcheson et al., DOI: 10.1021/acs.chemrestox.5c00273
Acquire regulatory-compliant predictions
Perform predictions
Submit results to relevant regulatory bodies
Get QMRFs
QMRF (QSAR Model Reporting Format) is a harmonized template for summarizing and reporting key information on QSAR models, including information on model validity.
Prepared by the model developer and assessed by regulators.
Prepare QPRFs
QPRF (QSAR Prediction Reporting Format) is an extensively updated reporting format that reflects the newly established OECD QSAR Prediction Principles
Prepared by the user and assessed by regulators.
Use CAF Checklist
The QAF (QSAR Assessment Framework) Prediction Checklist verifies whether a (Q)SAR model and its predictions comply with the principles outlined in the OECD guidance on model validation (OECD, 2007).
Prepared by the user or regulators
Models
Ensemble models composed of diverse conventional and cutting-edge AI/ML algorithms, including Boosting Machines, Graph Neural Networks, and Large Language Models
Reports
A detailed report on the molecule activity, as well as rule-based and read-across data are available in PDF or CSV formats.
-2.png)
RISK-SCORE: screen and rank compounds
Perform predictions for an unlimited number of compounds and expedite decision-making without the need to delve into numerous individual endpoint prediction values
The Risk-Score is estimated based on ADMET properties along with physicochemical and medicinal chemistry properties.
This efficient approach facilitates the swift ranking of a substantial pool of compounds, delivering a list of the most and least promising candidates for acceptance as pharmaceuticals.
OUR TECHNOLOGY
CHECK OUT OUR LATEST PAPERS
-
Improving VAE based molecular representations for compound property prediction. DOI: 10.1186/s13321-022-00648-x
-
BARTSmiles: Generative Masked Language Models for Molecular Representations. DOI: 10.48550/arXiv.2211.16349
-
Datasets Construction and Development of QSAR Models for Predicting Micronucleus In Vitro and In Vivo Assay Outcomes. DOI: 10.3390/toxics11090785
OUR DATABASES
-
In addition to compiling a comprehensive set of open datasets, we employ a specially trained Large Language Model to discern and filter relevant data for the relevant endpoint from the vast pool of 35 million scientific papers available on PubMed.
-
Thousands of selected scientific papers are manually reviewed and information is extracted to construct a dataset.
-
The final datasets are manually reviewed and normalized by experts, adhering to the Klimisch criteria and falling under the 'reliable without restriction' category.
CRITICAL BENEFITS/
KEY ADVANTAGES OF THE PLATFORM
All-in-one pipeline
An exceptional opportunity to consolidate your predictions onto a single platform, avoiding the use of different software for predicting specific endpoints.
Model adaptation
We adopt a bespoke approach to cater to your unique requirements. This entails the capacity to refine and retrain our models based on the specifics of your provided compounds, thereby optimizing predictions to align precisely with the characteristics of your distinct set of compounds.
Ranking system
The unique risk-score-based ranking system enables to perform predictions for an unlimited number of compounds in a very short period of time, facilitating decision-making without the need to delve into numerous individual endpoint prediction values.
Online platform
The platform's adaptive nature, distinct from traditional software, operates in real-time adjustments based on emerging data or refined models. This ensures that decision-making is not only rapid but also remains responsive to the latest scientific insights and advancements throughout the subscription period.