Posts by Tags

Evaluating the Boltz-2 Co-folding Model for Challenging Modalities: A Two-Case Study with Covalent Inhibitor and Macrocyclic Molecular Glue

21 minute read

Published: August 30, 2025

Introduction

In the rapidly evolving landscape of drug discovery, Artificial Intelligence (AI) co-folding models have gained significant attention, frequently highlighted in public reports for their potential to revolutionise structure-based drug design (SBDD). These computational tools promise to accelerate the identification and optimisation of novel drug candidates by predicting the intricate interactions and corresponding conformations between receptors and ligands, under induced fit effects. However, despite the widespread enthusiasm, there remains a notable gap in detailed performance investigations, particularly concerning challenging modalities such as covalent inhibitors and the emerging class of molecular glues. For the pharmaceutical industry, where real-world applications demand robust and reliable predictive capabilities, evaluating these models against those complex cases is crucial. This helps us to understand not only what co-folding models can achieve but also their current limitations that need to be improved.

Evaluating AI-based Molecular Modelling with Physical Simulation

40 minute read

Published: December 03, 2025

Introduction

In my last blog post, I explored Boltz-2 co-folding model applied to novel chemical modalities, including covalent macrocyclic molecular glue Elironrasib and its induced protein complexes. While the post received positive feedback, a detailed investigation was not performed. As both a computational chemist and a growing expert in cheminformatics, I believe it is crucial to analyse current AI models through a physics-based lens, leveraging corresponding data to inspire practical applications within the drug discovery industry.

This new article will serve as a supplement, specifically designed to address two key questions:

1. Which are those 'good' AI models that most closely approach physics as the ground truth, while also maintaining low computational cost?

2. Could physics effectively guide the identification, refinement, or even rescue of 'imperfect' AI models so far for molecular modelling?

Evaluating AI-based Molecular Modelling with Physical Simulation

40 minute read

Published: December 03, 2025

Introduction

In my last blog post, I explored Boltz-2 co-folding model applied to novel chemical modalities, including covalent macrocyclic molecular glue Elironrasib and its induced protein complexes. While the post received positive feedback, a detailed investigation was not performed. As both a computational chemist and a growing expert in cheminformatics, I believe it is crucial to analyse current AI models through a physics-based lens, leveraging corresponding data to inspire practical applications within the drug discovery industry.

This new article will serve as a supplement, specifically designed to address two key questions:

1. Which are those 'good' AI models that most closely approach physics as the ground truth, while also maintaining low computational cost?

2. Could physics effectively guide the identification, refinement, or even rescue of 'imperfect' AI models so far for molecular modelling?

Some ‘Learning’ in Cheminformatics, QSAR and Generative AI

55 minute read

Published: March 16, 2026

Introduction

In 2025, I published two posts exploring molecular modelling through AI co-folding and physics-based simulations for Structure-Based Drug Discovery (SBDD), drawing on my background in physical organic chemistry and biophysics. However, some pillars of modern computational drug discovery - Cheminformatics and Machine Learning (ML) - have yet to be discussed here. During my time in the industry, these are the fields where I have experienced the most significant professional growth.

In this post, I will share insights from my journey transitioning between the roles of a cheminformatician and an ML engineer in the biotech. To demonstrate the real-world application of data science in drug discovery, I will present a virtual screening (VS) workflow for covalent drug discovery in the Cereblon (CRBN) chemical space (Figure 1). We will walk through a rigorous pipeline: from chemical database mining and library enumeration to molecular docking, QSAR modelling, and AI-driven generation - all guided by the rational constraints of organic and medicinal chemistry.

Some ‘Learning’ in Cheminformatics, QSAR and Generative AI

55 minute read

Published: March 16, 2026

Introduction

In 2025, I published two posts exploring molecular modelling through AI co-folding and physics-based simulations for Structure-Based Drug Discovery (SBDD), drawing on my background in physical organic chemistry and biophysics. However, some pillars of modern computational drug discovery - Cheminformatics and Machine Learning (ML) - have yet to be discussed here. During my time in the industry, these are the fields where I have experienced the most significant professional growth.

In this post, I will share insights from my journey transitioning between the roles of a cheminformatician and an ML engineer in the biotech. To demonstrate the real-world application of data science in drug discovery, I will present a virtual screening (VS) workflow for covalent drug discovery in the Cereblon (CRBN) chemical space (Figure 1). We will walk through a rigorous pipeline: from chemical database mining and library enumeration to molecular docking, QSAR modelling, and AI-driven generation - all guided by the rational constraints of organic and medicinal chemistry.

Some ‘Learning’ in Cheminformatics, QSAR and Generative AI

55 minute read

Published: March 16, 2026

Introduction

In 2025, I published two posts exploring molecular modelling through AI co-folding and physics-based simulations for Structure-Based Drug Discovery (SBDD), drawing on my background in physical organic chemistry and biophysics. However, some pillars of modern computational drug discovery - Cheminformatics and Machine Learning (ML) - have yet to be discussed here. During my time in the industry, these are the fields where I have experienced the most significant professional growth.

In this post, I will share insights from my journey transitioning between the roles of a cheminformatician and an ML engineer in the biotech. To demonstrate the real-world application of data science in drug discovery, I will present a virtual screening (VS) workflow for covalent drug discovery in the Cereblon (CRBN) chemical space (Figure 1). We will walk through a rigorous pipeline: from chemical database mining and library enumeration to molecular docking, QSAR modelling, and AI-driven generation - all guided by the rational constraints of organic and medicinal chemistry.

Evaluating AI-based Molecular Modelling with Physical Simulation

40 minute read

Published: December 03, 2025

Introduction

In my last blog post, I explored Boltz-2 co-folding model applied to novel chemical modalities, including covalent macrocyclic molecular glue Elironrasib and its induced protein complexes. While the post received positive feedback, a detailed investigation was not performed. As both a computational chemist and a growing expert in cheminformatics, I believe it is crucial to analyse current AI models through a physics-based lens, leveraging corresponding data to inspire practical applications within the drug discovery industry.

This new article will serve as a supplement, specifically designed to address two key questions:

1. Which are those 'good' AI models that most closely approach physics as the ground truth, while also maintaining low computational cost?

2. Could physics effectively guide the identification, refinement, or even rescue of 'imperfect' AI models so far for molecular modelling?

Some ‘Learning’ in Cheminformatics, QSAR and Generative AI

55 minute read

Published: March 16, 2026

Introduction

In 2025, I published two posts exploring molecular modelling through AI co-folding and physics-based simulations for Structure-Based Drug Discovery (SBDD), drawing on my background in physical organic chemistry and biophysics. However, some pillars of modern computational drug discovery - Cheminformatics and Machine Learning (ML) - have yet to be discussed here. During my time in the industry, these are the fields where I have experienced the most significant professional growth.

In this post, I will share insights from my journey transitioning between the roles of a cheminformatician and an ML engineer in the biotech. To demonstrate the real-world application of data science in drug discovery, I will present a virtual screening (VS) workflow for covalent drug discovery in the Cereblon (CRBN) chemical space (Figure 1). We will walk through a rigorous pipeline: from chemical database mining and library enumeration to molecular docking, QSAR modelling, and AI-driven generation - all guided by the rational constraints of organic and medicinal chemistry.

Evaluating the Boltz-2 Co-folding Model for Challenging Modalities: A Two-Case Study with Covalent Inhibitor and Macrocyclic Molecular Glue

21 minute read

Published: August 30, 2025

Introduction

In the rapidly evolving landscape of drug discovery, Artificial Intelligence (AI) co-folding models have gained significant attention, frequently highlighted in public reports for their potential to revolutionise structure-based drug design (SBDD). These computational tools promise to accelerate the identification and optimisation of novel drug candidates by predicting the intricate interactions and corresponding conformations between receptors and ligands, under induced fit effects. However, despite the widespread enthusiasm, there remains a notable gap in detailed performance investigations, particularly concerning challenging modalities such as covalent inhibitors and the emerging class of molecular glues. For the pharmaceutical industry, where real-world applications demand robust and reliable predictive capabilities, evaluating these models against those complex cases is crucial. This helps us to understand not only what co-folding models can achieve but also their current limitations that need to be improved.

Evaluating the Boltz-2 Co-folding Model for Challenging Modalities: A Two-Case Study with Covalent Inhibitor and Macrocyclic Molecular Glue

21 minute read

Published: August 30, 2025

Introduction

In the rapidly evolving landscape of drug discovery, Artificial Intelligence (AI) co-folding models have gained significant attention, frequently highlighted in public reports for their potential to revolutionise structure-based drug design (SBDD). These computational tools promise to accelerate the identification and optimisation of novel drug candidates by predicting the intricate interactions and corresponding conformations between receptors and ligands, under induced fit effects. However, despite the widespread enthusiasm, there remains a notable gap in detailed performance investigations, particularly concerning challenging modalities such as covalent inhibitors and the emerging class of molecular glues. For the pharmaceutical industry, where real-world applications demand robust and reliable predictive capabilities, evaluating these models against those complex cases is crucial. This helps us to understand not only what co-folding models can achieve but also their current limitations that need to be improved.