Is it time to retire druglikeness?

⚡ I was struck by a statement I read today about a Techbio company that "uses large language models to create billions of druglike molecules". I have nothing against this company and I think they do very interesting work, no need to name them. But I found myself suddenly and surprisingly critical of the expression "druglike molecules", a somewhat anachronistic concept in 2025. I mentioned Lipinski's Rule of Five (focusing on molecular weight, lipophilicity, hydrogen bond donors/acceptors) in yesterday's post (link in comments) about the drug-unlike compound that looks too much like a detergent but might be acceptable by modern criteria.

Traditional druglikeness criteria provided useful guidelines for oral bioavailability at a time of little data and less powerful model (compared to now). However, modern drug discovery has shown these rules have significant limitations:

1️⃣ Many successful drugs violate these rules, particularly biologics, peptides, and natural products.

2️⃣ The pharmaceutical landscape has expanded beyond small molecule oral drugs to include antibodies, oligonucleotides, and other modalities.

3️⃣ Target-specific considerations often override general druglikeness rules.

Today, the concept is more nuanced—druglikeness is increasingly viewed as a spectrum rather than binary. We can now use more sophisticated computational models that consider:

▶️ Specific delivery routes and target tissues

▶️ Pharmacokinetic properties

▶️ Safety profiles

▶️ Target engagement

Rather than asking "is this molecule druglike?" developers now ask "does this molecule have the right properties for its intended therapeutic use?" This shift reflects the growing complexity and sophistication of modern drug discovery, with its rapidly increasing arsenal of models, AI algorithms and datasets.