Academic Jobs - Home of Higher Ed Logo

AI-driven drug discovery and protein design

Submit News
a sign with a question mark and a question mark drawn on it
Photo by Nahrizul Kadri on Unsplash

AI-driven drug discovery and protein design

For most of pharmaceutical history, finding a new drug has been an exercise in expensive luck. A single approved medicine takes ten to fifteen years and roughly $2.6 billion to develop, with most of that cost buried in the thousands of candidate molecules that fail along the way. Researchers screened vast chemical libraries, hoping a promising compound would survive the gauntlet of optimization, toxicity testing, and clinical trials. Today, a wave of artificial intelligence is trying to replace that lottery with something closer to engineering — and the early results suggest the shift is real.

From prediction to design

The turning point came when DeepMind's AlphaFold2 and the Baker Lab's RoseTTAFold cracked a problem that had stumped biologists for fifty years: predicting a protein's three-dimensional shape from its amino acid sequence alone. Because a protein's structure determines its function, knowing that shape is the foundation for everything from understanding disease to building a drug that fits a target like a key in a lock. The public AlphaFold database has since ballooned to more than 214 million predicted structures, covering nearly every protein catalogued in science.

Prediction was only the beginning. The more radical leap is generative design — using AI not just to model proteins that already exist, but to invent new ones from scratch. Diffusion-based tools such as RFdiffusion and FrameDiff, conceptually similar to the models that generate images from text, can now sketch novel protein backbones tailored to a specific job. The work earned David Baker a share of a Nobel Prize, and the phrase that captures the field's ambition has stuck: generative biology is moving drug discovery, as one industry CEO put it, from "a process of chance to one of design."

Designs that reach the lab — and the clinic

What separates 2026 from the hype of earlier years is that AI-designed molecules are no longer just topping computational leaderboards; they are being physically tested. Researchers have reported a designed antimicrobial peptide that cleared infection in mice and engineered enzymes that outperform both nature and traditional directed evolution. NVIDIA's Proteina-Complexa model, part of its BioNeMo platform, is being used by Novo Nordisk and others to design protein binders that companies then validate in the lab rather than only on a screen.

The clinical evidence is starting to arrive too. Generate:Biomedicines presented encouraging data from asthma patients treated with an AI-assisted antibody, where a twice-yearly injection lowered disease-triggering proteins without notable side effects. Isomorphic Labs, the DeepMind spinoff led by Demis Hassabis, made the biggest technical splash with IsoDDE, a model that roughly doubles AlphaFold 3's accuracy on novel systems and dramatically improves binding-affinity prediction. The company runs 17 drug programs across cancer, immunology, and cardiovascular disease, and expects its first AI-designed candidates to enter human trials by the end of 2026.

The money has followed the science. Eli Lilly and NVIDIA committed a billion dollars to a joint AI drug-discovery lab; Isomorphic struck partnerships worth nearly $3 billion in potential milestones with Lilly and Novartis; and frontier AI labs are now buying their way directly into biology. The overall market is growing at well over 25 percent a year and is projected to reach $8–10 billion by 2030.

A dose of realism

For all the momentum, sober voices warn against declaring victory. Early discovery — the part AI excels at — is only a small slice of the time and cost of bringing a drug to market. The real test happens in clinical trials, which still take years and remain stubbornly resistant to acceleration. A revealing detail: the gap between announced deal values and actual upfront payments often runs fifty to one, a sign of how much remains speculative.

Data is the other bottleneck. Unlike the internet text that trains language models, pharmaceutical data is fragmented, inconsistently formatted, and scarce, which is why infrastructure and curated datasets increasingly matter more than algorithmic cleverness.

The honest verdict is that 2026 is a proving year. Several AI-designed drugs are approaching pivotal trials, and their outcomes will offer the first large-scale answer to the question the field has been promising for a decade: not whether AI can design a plausible molecule, but whether those molecules actually make people better. The first approvals, if they come, are most likely in 2027 or 2028 — and they will mark the moment design finally meets the patient.

the word discovery spelled with scrabble letters on a wooden surface

Photo by Ling App on Unsplash

Portrait of Chongjie Chen
About the author

Chongjie ChenView author

Distinguished Visiting Scholar

Discussion

Sort by:

Be the first to comment on this article!

You

Please keep comments respectful and on-topic.

New0 comments

Join the conversation!

Add your comments now!

Have your say

Engagement level

Browse by Faculty

Browse by Subject

Frequently Asked Questions

How long and costly has traditional drug development been?

A single approved medicine takes ten to fifteen years and roughly $2.6 billion to develop, with most of that cost buried in the thousands of candidate molecules that fail along the way.

🧬What breakthrough solved protein structure prediction?

DeepMind's AlphaFold2 and the Baker Lab's RoseTTAFold cracked the fifty-year problem of predicting a protein's three-dimensional shape from its amino acid sequence alone.

📊How large is the public AlphaFold database now?

The public AlphaFold database has ballooned to more than 214 million predicted structures, covering nearly every protein catalogued in science.

What is generative protein design?

Generative design uses AI to invent new proteins from scratch rather than only modeling existing ones, with tools such as RFdiffusion and FrameDiff sketching novel protein backbones.

🏅Which Nobel recognition is mentioned for this work?

The work earned David Baker a share of a Nobel Prize.

🔬What real-world lab results have AI-designed molecules shown?

Researchers have reported a designed antimicrobial peptide that cleared infection in mice and engineered enzymes that outperform both nature and traditional directed evolution.

💊Which companies are using NVIDIA's BioNeMo platform?

NVIDIA's Proteina-Complexa model is being used by Novo Nordisk and others to design protein binders validated in the lab.

🩺What clinical data has Generate:Biomedicines presented?

Generate:Biomedicines presented encouraging data from asthma patients treated with an AI-assisted antibody where a twice-yearly injection lowered disease-triggering proteins without notable side effects.

📅When does Isomorphic Labs expect first human trials?

Isomorphic Labs expects its first AI-designed candidates to enter human trials by the end of 2026.

📈What is the projected market size for this field by 2030?

The overall market is growing at well over 25 percent a year and is projected to reach $8–10 billion by 2030.

What remains the main bottleneck after early discovery?

Clinical trials still take years and remain stubbornly resistant to acceleration, with the gap between announced deal values and actual upfront payments often running fifty to one.

When are the first AI-designed drug approvals most likely?

The first approvals, if they come, are most likely in 2027 or 2028.