Protein artificial intelligence is, without question, the most mature and publicly celebrated discipline within the “omics” family. When we discuss AI in biology, the conversation inevitably drifts toward the 2024 Nobel Prize in Chemistry—awarded jointly to David Baker for computational protein design, and to Demis Hassabis and John Jumper for protein structure prediction via AlphaFold.

However, structure prediction was merely the opening act. Today, the frontier has rapidly shifted from static structure prediction to protein design (creating entirely new proteins), function prediction, and complex interaction modeling. In this seventh installment of the Agentic Omics series, we will dissect the current state of AI in proteomics, evaluate the monumental shifts from AlphaFold 2 to AlphaFold 3 and ESM-3, explore generative models like ProGen and RFdiffusion, and critically assess their real-world clinical impact in drug discovery.

The AlphaFold Lineage: From Monomers to the Molecules of Life

AlphaFold 2 (AF2) shattered the 50-year-old grand challenge of protein folding during the CASP14 competition, eventually yielding a database of over 200 million predicted protein structures—nearly every known protein cataloged by science. It demonstrated that deep learning, specifically a highly modified Evoformer architecture relying on multiple sequence alignments (MSAs), could infer 3D coordinates from 1D amino acid sequences with atomic accuracy.

In 2024, DeepMind and Isomorphic Labs published AlphaFold 3 (AF3) in Nature (Abramson et al., 2024). The conceptual leap from AF2 to AF3 was profound. While AF2 was a protein folding model, AF3 is a biomolecular interaction model.

What Makes AlphaFold 3 Different?

  1. Beyond Proteins: AF3 predicts joint structures of complexes including proteins, DNA, RNA, small molecule ligands, and post-translational modifications.
  2. Architecture: It replaced the Evoformer with a simpler Pairformer module and transitioned the structural generation phase to a Diffusion Module, similar to the architecture underlying modern AI image generators.
  3. Reduced MSA Dependency: AF3 demonstrated that the lack of cross-entity evolutionary information (co-evolution) is not a substantial blocker for predicting biomolecular interactions. It learns the physics and chemistry of interactions more directly.

The Ligand Docking Debate: Upon release, AF3 showed a massive improvement over traditional physics-based docking tools like Vina. However, the exact accuracy of its protein-ligand docking remains a subject of active debate in the computational chemistry community. Independent baselines (e.g., from Inductive Bio and others) have shown that while AF3 performs exceptionally well on natural ligands and common co-factors (ATP, nucleosides), its performance on entirely novel, out-of-distribution synthetic drug-like molecules can sometimes degrade compared to heavily optimized traditional or hybrid docking pipelines. Furthermore, the initial restricted access to AF3’s code sparked intense debate about open science, though DeepMind eventually committed to broader code releases.

Evolutionary-Scale Language Models: ESM-3

While AlphaFold relies on evolutionary context via MSAs, the opposing philosophical approach is the pure “protein language model” (pLM). Driven largely by Meta’s FAIR and now EvolutionaryScale, models like ESM-2 and the recent ESM-3 (15B parameters) treat protein sequences purely as text.

ESM-3, released in 2024, is a multimodal generative language model that reasons over the sequence, structure, and function of proteins simultaneously. By discretizing 3D structures and functional annotations into tokens, ESM-3 can be prompted to generate proteins that satisfy specific constraints.

For instance, EvolutionaryScale demonstrated ESM-3’s capabilities by prompting it to design a novel Green Fluorescent Protein (GFP) that shares only 58% sequence identity with known natural fluorescent proteins—a jump across the fitness landscape that would take millions of years of natural evolution.

Generative Protein Design: Building What Nature Didn’t

If AlphaFold reads the book of life, generative AI writes new chapters. This is the domain of computational protein design, pioneered by the Baker Lab (University of Washington) and others.

RFdiffusion

Built upon the RoseTTAFold architecture, RFdiffusion (Watson et al., 2023) applies denoising diffusion probabilistic models to protein backbones. It allows researchers to specify a target shape, a binding site, or a symmetric architecture, and the model “diffuses” from random noise into a highly stable, completely novel protein structure that fulfills the design criteria. This is particularly transformative for designing de novo binders for therapeutic targets (e.g., designing a protein that binds to a cancer cell receptor to block a signaling pathway).

ProGen and ProGen2

Developed initially by Salesforce Research and expanded upon by Profluent, ProGen operates like a ChatGPT for proteins. It is an autoregressive transformer trained on millions of protein sequences conditioned on taxonomic and functional metadata. You can “prompt” ProGen with a desired function (e.g., “lysozyme enzyme from a thermophilic bacteria”), and it will generate sequences that fold into active enzymes, even if those exact sequences have never existed in nature.

ProteinMPNN

Once you have a 3D backbone designed by RFdiffusion, you need the actual amino acid sequence that will fold into that shape. ProteinMPNN is an incredibly robust graph neural network that performs “inverse folding.” It takes a 3D structure and generates the 1D sequence required to form it. The combination of RFdiffusion (for structure generation) and ProteinMPNN (for sequence generation) has become the gold-standard pipeline for de novo protein design.

Clinical Impact: From Bytes to Bedsides

The ultimate validation of these models is clinical utility. Are AI-designed drugs entering the clinic?

Isomorphic Labs (an Alphabet company) is utilizing AlphaFold 3 alongside proprietary generative AI models to build a commercial drug discovery pipeline, partnering with pharmaceutical giants like Novartis and Eli Lilly. The goal is to slash the time required in the “hit-to-lead” optimization phase by accurately predicting how small molecules will interact with novel protein targets.

Other companies are already putting AI-designed molecules into human trials:

  • Insilico Medicine: Their AI-designed drug for Idiopathic Pulmonary Fibrosis (ISM001-055), developed using generative AI for both target identification and molecule generation, successfully progressed into Phase II clinical trials.
  • Recursion Pharmaceuticals: Leveraging phenomics and structural AI to match molecules to cellular phenotypes at an industrial scale.

Honest Assessment: Despite the immense hype, the timeline from in silico discovery to FDA approval is still 7-10 years. While AI drastically accelerates the pre-clinical phase (target ID, virtual screening, lead optimization), it does not accelerate Phase I/II/III human clinical trials. Biology is infinitely more complex than a static 3D coordinate map; issues like toxicity, bioavailability, and off-target effects remain significant hurdles that purely structural AI cannot fully predict yet. Conformational dynamics—how proteins move and breathe in a living cell—remain a major challenge for models like AlphaFold, which mostly output static, low-energy state snapshots.

Conclusion

AI for proteomics has crossed the threshold from academic curiosity to an industrial necessity. The field has evolved from predicting how natural proteins fold (AlphaFold 2) to predicting how biological systems interact (AlphaFold 3) and generating entirely new biological machines (ESM-3, RFdiffusion).

As we will see later in this series, the true power of “Agentic Omics” emerges when these structural models are orchestrated by Large Language Models. Imagine an autonomous agent that reads a paper about a new disease pathway, identifies the target protein, uses AlphaFold 3 to predict its structure, deploys RFdiffusion to design a bespoke inhibitor, and runs molecular dynamics simulations to verify binding—all before the human scientist has finished their morning coffee. That future is closer than it appears.

Glossary

  • MSA (Multiple Sequence Alignment): A technique to align three or more biological sequences to identify conserved regions, evolutionary relationships, and co-evolving residues.
  • De novo Protein Design: The computational generation of entirely new proteins with structures and functions not found in nature.
  • Inverse Folding: The process of predicting a 1D amino acid sequence that will fold into a given, desired 3D structural backbone.
  • Diffusion Model: A generative AI architecture that learns to construct data (like images or protein backbones) by reversing a process that gradually adds noise to the data.
  • Ligand: A substance (usually a small molecule) that forms a complex with a biomolecule (like a protein) to serve a biological purpose.
  • Conformational Dynamics: The various 3D shapes (conformations) a protein can adopt and the transitions between them, critical for protein function and enzyme catalysis.

References

  1. Abramson, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630(8016), 493-500.
  2. EvolutionaryScale. (2024). ESM-3: Simulating 500 million years of evolution with a language model. (Preprint/Technical Report).
  3. Watson, J. L., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620(7976), 1089-1100.
  4. Nijkamp, E., et al. (2023). ProGen2: exploring the boundaries of protein language models. Cell Systems, 14(11), 968-978.
  5. Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589.
  6. The Nobel Prize in Chemistry 2024. NobelPrize.org. Nobel Prize Outreach AB 2024. Awarded to David Baker, Demis Hassabis, and John Jumper.