BioAscent’s in silico discovery team work with our clients to apply the right computational methods at all stages of the drug discovery process and progress compounds towards the clinic. Here, Dr Angelo Pugliese, Associate Director of In Silico Discovery, shares his thoughts on the role of in silico discovery in hit discovery.

Can you describe your background in computational chemistry and your role at BioAscent?

I am a computational chemist by training. I earned my PhD at the University of Nottingham and have spent over 15 years working in drug discovery across the US and the UK. Before joining BioAscent, I led the computational chemistry team at the CRUK Beatson Institute (now CRUK Scotland Institute) in Glasgow, where I also led their fragment library projects. Throughout my career, my passion has always been finding smarter, more efficient ways to design drugs by bridging the gap between cutting-edge computer science and the wet lab.

Today, as Associate Director of In Silico Discovery at BioAscent, my daily focus is on acting as a strategic partner to our clients. I lead our computational strategy and manage our "In Silico Workbench," which is our highly integrated, enterprise-grade computational discovery hub. Our goal is to deploy the most advanced in silico tools available to find potent, selective and synthesisable ligands for our clients' challenging biological targets as efficiently as possible.

How do in silico approaches fit into the hit discovery process? How does computational hit discovery differ today compared with five or ten years ago?

In silico approaches are the central nervous system of modern hit discovery. Computational chemistry is a uniquely fast-evolving field; it adopts and iterates on new technologies far faster than almost any other scientific discipline. Five to ten years ago, we were already pushing the boundaries of what was computationally possible at the time (we had GPUs back then too), but the velocity of AI development has recently pushed that evolution into overdrive.

Today, while the core process of virtual screening remains sequential, the underlying algorithms are more powerful, accurate, and efficient. Historically, building a bespoke computational tool required months of dedicated software engineering. Today, by leveraging AI agents, we can rapidly prototype and deploy custom workflows in a matter of days. At BioAscent, we don't just screen existing databases using standard methods. Through our generative AI portals, we dynamically explore novel chemical space, actively designing proprietary, IP-generating chemical matter tailored to the client's Target Product Profile. Furthermore, the advent of AI-driven structural biology (such as AlphaFold3 and Protenix) has unlocked the ability to perform "cofolding." Instead of docking molecules into rigid, static targets, we can now simultaneously fold the protein and the ligand together. This allows us to accurately model highly complex protein-ligand interactions and reveal dynamic binding pockets for our clients in a matter of hours.

What makes virtual screening a powerful starting point for identifying novel hits, and what types of virtual screening strategies does BioAscent use to support hit discovery?

There is a common misconception in the industry that virtual screening is just about throwing massive computational power at ultra-large libraries. At BioAscent, our philosophy is different: we prioritise the quality of the starting libraries and the depth of our target expertise over brute-force scale. Virtual screening is a powerful starting point because it acts as an intelligent funnel, drastically reducing the time and cost required for clients to find novel, tractable hits by prioritising only the highest-quality scaffolds for synthesis and testing.

Over the years, our team has built deep experience running highly focused virtual screens across a wide variety of target classes, from classic kinases to challenging protein-protein interactions. Instead of screening billions of random, often med chem unfriendly, molecules, we curate highly relevant virtual libraries.

Because we work with a diverse range of biotech and pharma partners, we maintain a highly adaptable toolkit. For structure-based drug design, we utilise precise docking tailored to the specific target class. For ligand-based design, we use shape and pharmacophore matches, or 2D approaches, carefully selecting the method depending on the specific case. Crucially, we also have dedicated portals for advanced modalities. For example, if a client is targeting protein degradation, our "Proximity" portal specifically models heterobifunctional degraders and molecular glues. This breadth ensures we build the computational strategy around the client's unique biology using the most appropriate tools available.

How do in silico hit discovery approaches complement more traditional methodologies such as high-throughput screening (HTS)?

They are perfectly synergistic, especially within a CRO environment where speed and data quality are paramount. A prime example is how we leverage BioAscent’s in-house physical screening library of over 100,000 diverse compounds. While HTS is central to drug discovery, assay optimisation is often the main bottleneck.

Our AI-driven in silico methods enhance these campaigns by predicting and resolving experimental issues before they derail a client's project. For example, Pan-Assay Interference Compounds (PAINS) generate false positives via mechanisms like aggregation and redox cycling. Instead of relying on traditional, static structural filters which leads to the loss of potentially valid chemical matter, we recently developed a machine learning framework using Bayesian Optimization to engineer artifact-suppressing assay buffers. By employing a two-stage XGBoost residual stacking architecture, we successfully reduced background noise down to 17% and rescued nearly 80% of 'bad actor' compounds. By optimising the environment, we preserve vital chemical diversity and save our clients from wasting resources on false positives. Once HTS yields these high-quality hits, we can combine, for example, the data with our advanced Mechanistic PK/PD (IVIVE) models to predict in vivo behaviour, rapidly accelerating the hit-to-lead timeline.

How do you see in silico hit discovery evolving with advances in AI and data science?

The (probably not that near) future of in silico discovery is the fully closed DMTA (Design-Make-Test-Analyse) loop, where AI drives multi-parameter optimisation autonomously and simultaneously rather than sequentially.

As data science advances, we are integrating highly complex endpoints much earlier in the pipeline. It is no longer enough to just predict if a drug will bind. Soon, generative AI will propose a molecule, and before a client invests in synthesis, that molecule will autonomously pass through an array of virtual stress tests. At BioAscent, we are already building this interconnected architecture to ensure we remain at the forefront of this dynamic field. For our clients, this means access to state-of-the-art capabilities and deep drug discovery expertise that significantly de-risks their drug discovery programmes.

Learn more about In Silico Discovery and Data Analysis at BioAscent here.