GenAI Beyond Just LLMs

From Words to Molecules: The Expanding Frontier of GenAI

Generative AI has changed how we create, work, and even imagine. In just a few short years, tools like ChatGPT, GitHub Copilot, and DALL·E have redefined productivity across industries  —  from software development and design to education and marketing. But the innovation curve continues to steepen, and it's no longer just about generating text or images. The same technology that can draft an email or write a poem is now designing molecules and discovering new materials.

We’re entering a world where AI doesn’t just write about science  —  it helps do the science. Big tech isn’t sitting this one out. OpenAI, Meta, Google DeepMind, Microsoft, and others are in a tight race, developing large-scale AI models that understand not just language and visuals, but also chemistry, biology, and physics.

At the heart of this transformation are foundation models  —  large-scale neural networks trained on diverse, high-volume datasets. These models don’t just memorize facts; they learn the structure and relationships within data, enabling them to generalize across tasks with remarkable agility. Whether it’s the ability to reason, generate coherent results, or predict with startling accuracy, foundation models enable it all and are the true engines behind the GenAI movement.

What Are Foundation Models, Really?

Think of foundation models as the “brains” behind today’s generative AI. They are trained on huge datasets (like all of Wikipedia, scientific papers, code repositories, or molecular structures) and learn to represent complex patterns  —  be it language grammar or molecular geometry. Once trained, they can be fine-tuned or adapted for specific tasks.

Multi-Modal Foundation Models: More Than Just Text

While the spotlight has mostly been on text-based AI powered by large language models (LLMs), one of the most fascinating aspects of foundation models is their multimodality. That means they can process and reason across different kinds of data  —  not just words. We’ve seen this with models like DALL·E, which takes in text prompts and generates images. The input and output are in different forms, yet the model learns how to connect them.

This same idea is now being applied to scientific domains. Instead of just learning from books or websites, AI is learning from chemical formulas, 3D molecular structures, crystal lattices, and physics-based energy and quantum simulations. These are complex data formats that require a deep understanding of science — and foundation models are proving they’re up to the task.

Some examples of these foundation models include:

• “Text” modality  –  GPT-4o, Llama-4, Claude Sonnet 3.5
• “Image” modality  –  SimCLR2, DeepProfiler, DINO2
• “Chemistry” modality  –  ChemBERTa, MegaMolBART, MolCLR
• “Large molecule” modality  –  ProteinBERT, Geneformer

Just as GPT models learn English syntax and semantics, chemistry foundation models learn the “grammar” of molecules.

Imagine a model that can read a SMILES string  —  a string of text that is used to represent a molecule  —  and intuitively understand how it behaves. That’s what models like MolBART, ChemDFM, and ProteinBERT are doing. They’re trained on millions of chemical compounds, learning how atoms interact, how functional groups influence reactivity, and how geometry shapes behavior.

These models don’t just memorize known molecules  —  they generalize to design new ones with desired properties. For instance, MolBART adapts the BART architecture from NLP to generate molecular candidates. ChemDFM, a diffusion-based model, simulates the molecular generation process as a kind of stochastic walk through chemical space. And ProteinBERT applies transformer-based learning to amino acid sequences, capturing how small changes can affect folding and function. This enables these models to predict what the expected properties can be from experimental molecules, materials, or formulations that R&D laboratories might be aiming to synthesize, thereby accelerating the rate at which labs conduct research for molecules / materials with desired properties.

Who’s Using These Models in the Real World?

This isn’t just academic theory  —  industry is already applying chemistry foundation models to drug discovery, material science, and biotech. Here are a few key players leading the charge:

Google DeepMind’s AlphaFold 3 and GNoME

Google DeepMind and Isomorphic Labs recently launched the AlphaFold 3 model, capable of predicting the structures of proteins, DNA, ligands, and with the potential to transform our understanding of the components that make up biological life and enable drug discovery. As part of the AlphaFold project, they have also open-sourced access to their research tool AlphaFold server and their database of over 200 million protein structures.

See the AlphaFold 3 announcement.

The GNoME project is another such effort, using deep learning to predict the stability of over 2.2 million new crystal structures, opening up huge opportunities in battery tech, semiconductors, and the discovery of green energy materials.

See the GNoME announcement.

IBM’s MolFormerXL

IBM’s large-scale molecular transformer model is helping researchers discover novel compounds for pharmaceuticals and sustainable materials. Its architecture is designed to generalize across chemical spaces, identifying high-potential candidates faster than traditional methods.

Read more here.

NVIDIA BioNeMo

NVIDIA’s BioNeMo platform is like a plug-and-play toolkit for biology and chemistry. It includes pretrained models for protein folding, molecular docking, and genomic analysis, all accessible via cloud APIs. A biotech startup could use BioNeMo to design proteins for a new therapeutic and validate them computationally — cutting months from the R&D cycle.

Explore BioNeMo.

Final Thoughts: The Era of Generative Discovery

We’re no longer just generating content  —  we’re generating hypotheses, molecules, and materials. Foundation models are becoming collaborators in science, augmenting human creativity with computational intuition.

As these models evolve, they won’t replace scientists  —  but they’ll become indispensable tools in their workflows. Imagine having a model that suggests the next experiment, simulates its outcome, and even explains the underlying rationale. That’s not science fiction  —  it’s happening now.