AI and Drug Discovery — From Protein Folding to Clinical Candidates

Cos'è: An analysis of AI's role in accelerating new drug discovery: from protein folding models to automated molecule generation, through to the first compound entirely designed by AI to enter clinical trials.

The Unsustainable Cost of Traditional Pharmaceutical Research

Finding a new FDA-approved drug requires an average of 12 years of work and around $2.6 billion in investment, according to estimates from the Tufts Center for the Study of Drug Development. Of this, the discovery phase — biological target identification, molecule design, compound optimization — absorbs years of high-throughput screening and hundreds of millions of dollars in laboratory infrastructure. The failure rate is brutal: fewer than 10% of candidates entering phase 1 clinical trials receive final approval. Most failures occur due to unforeseen toxicity or insufficient efficacy, flaws that could in theory be intercepted much earlier with adequate predictive models.

AlphaFold and AlphaFold 3: Beyond Protein Structure

DeepMind's launch of AlphaFold 2 in 2020 represented a genuine breakthrough: the protein folding problem, unsolved for fifty years, was resolved with accuracy comparable to X-ray crystallography. DeepMind made the database of over 200 million protein structures freely available, radically transforming the starting point of target research. But AlphaFold 2 predicts static structures of isolated proteins. AlphaFold 3, presented in May 2024 in Nature, extends capabilities to modeling interactions between proteins and ligands — the molecules that are candidates to become drugs. This is the critical leap for drug design: it's not enough to know the structure of the target protein, you need to predict how a molecular candidate binds to it, with what affinity and selectivity. AlphaFold 3 introduces a diffusion-based architecture that enables modeling of protein-DNA, protein-RNA and protein-ligand complexes with significantly greater precision than previous methods, opening the way to AI-guided molecular optimization cycles.

Insilico Medicine and ISM001-055: The First AI-Designed Drug in Clinical Trials

Insilico Medicine is the company that has taken the concept of AI-driven drug discovery from theory to concrete clinical practice. In 2023 their candidate ISM001-055, a TNIK (Traf2 and NCK-interacting kinase) inhibitor developed entirely with a generative AI pipeline, completed phase 1 and received authorization to proceed to phase 2 for the treatment of idiopathic pulmonary fibrosis (IPF). The entire process from target identification to candidate selection required approximately 18 months, compared to the typical 5-6 years for this phase with conventional methods. Insilico's Pharma.AI platform combines three modules: PandaOmics for target identification through multi-omic analysis, Chemistry42 for molecule generation and optimization via reinforcement learning, and InClinico for predicting clinical trial outcomes. This is not AI as a support tool, but AI as the primary engine of the decision-making pipeline.

Recursion Pharmaceuticals and the Phenomics Approach

A radically different approach is that of Recursion Pharmaceuticals, listed on the NYSE in 2023. Instead of starting from the molecular target, Recursion builds ultra-high-resolution phenotypic maps: cells are treated with thousands of compounds and photographed with highly automated microscopy, generating enormous visual datasets that are then analyzed by computer vision models. The idea is to identify phenotypic patterns — how cells change shape, distribution, expression — that correlate with desired biological activity, regardless of the precise molecular mechanism. This so-called phenomics-first approach is particularly useful for rare diseases where target biology is not fully understood. Recursion has active partnerships with Roche and Bayer, and its pipeline includes candidates for central nervous system diseases and fibrosis. The risk is that phenomic correlation does not guarantee pharmacological selectivity: a compound may produce the desired phenotype for off-target reasons with unpredictable toxicological consequences.

What AI Accelerates and What Remains Unchanged: Biological Validation as the Bottleneck

Intellectual honesty requires distinguishing between phases where AI brings measurable advantage and those where the process remains structurally unchanged. AI is effective in: identification and validation of biological targets through genomic and proteomic analysis at scale; generation and optimization of candidate molecules with prediction of ADMET properties (absorption, distribution, metabolism, excretion, toxicity); reduction of hit identification time from months to days. AI does not replace and is unlikely to replace in the short term: phase 1, 2 and 3 clinical trials, which require years for safety and statistical reasons; in vivo biological validation, where predictive models still fail systematically on complex phenomena such as immunogenicity or idiosyncratic hepatotoxicity; the FDA/EMA regulatory process, which by definition evaluates empirical data rather than computational predictions. The bottleneck has shifted: it is no longer the generation of molecular candidates, but their progressive biological validation. The ability to generate thousands of plausible candidates in silico creates a new problem of selection and prioritization that physical laboratories cannot absorb at the same speed.

---