AI for Scientific Discovery

AI for Scientific Discovery refers to the application of machine learning models — including graph neural networks, foundation models, and differentiable simulators — to accelerate hypothesis generation, experimental design, and knowledge extraction across scientific domains such as biology, chemistry, physics, and materials science. It matters because traditional computational science scales poorly with system complexity, and AI methods can compress decades of experimental iteration into days of compute. The counterintuitive fact: AlphaFold2 did not merely predict protein structure faster — it solved a 50-year grand challenge so thoroughly that the entire structural biology experimental pipeline is now being redesigned around it.

The core insight

Scientific discovery is fundamentally a search problem over an astronomically large hypothesis space, and the "aha moment" is realizing that pretrained AI models encode latent scientific knowledge that can be steered with far less data than training from scratch. Rather than learning physics from millions of random experiments, a model pretrained on existing literature or simulations already understands the geometry of the solution space — fine-tuning then becomes precision targeting instead of brute-force exploration. This collapses the traditional hypothesis → experiment → analysis loop into a tightly coupled differentiable pipeline where the model, the simulator, and the optimizer all speak the same gradient language.

How It Works

Energy Functions and Force Fields

At the heart of molecular AI is learning a potential energy surface (PES). Given atomic coordinates R∈R3N\mathbf{R} \in \mathbb{R}^{3N} for NN atoms, the model learns:

1
E=fθ(R,Z)\quad E = f_\theta(\mathbf{R}, \mathbf{Z})

Explanation: The neural network fθf_\theta maps atomic positions R\mathbf{R} and atomic numbers Z\mathbf{Z} to a scalar energy EE, replacing expensive quantum-mechanical calculations. Forces are obtained by automatic differentiation — the same backpropagation used in model training:

2
Fi=−∂E∂Ri\quad \mathbf{F}_i = -\frac{\partial E}{\partial \mathbf{R}_i}

Explanation: The force on atom ii is the negative gradient of energy with respect to its position, making the entire simulation pipeline end-to-end differentiable.

Message-Passing on Molecular Graphs

Molecular graphs represent atoms as nodes and bonds as edges. The message-passing neural network (MPNN) update at layer ℓ\ell is:

3
hi(ℓ+1)=ϕ ⁣(hi(ℓ),  ⨁j∈N(i)ψ ⁣(hi(ℓ),hj(ℓ),eij))\quad \mathbf{h}_i^{(\ell+1)} = \phi\!\left(\mathbf{h}_i^{(\ell)},\; \bigoplus_{j \in \mathcal{N}(i)} \psi\!\left(\mathbf{h}_i^{(\ell)}, \mathbf{h}_j^{(\ell)}, \mathbf{e}_{ij}\right)\right)

Explanation: Each atom ii updates its hidden state hi\mathbf{h}_i by aggregating messages from its neighbors N(i)\mathcal{N}(i) using edge features eij\mathbf{e}_{ij}, where ϕ\phi is an update function and ψ\psi is a message function, both parameterized by neural networks. Here ⨁\bigoplus is a permutation-invariant aggregation (sum, mean, or attention-weighted), ensuring the model's output is invariant to atom ordering.

Property Prediction via Equivariant Networks

For 3D molecular systems, we need the network output to respect physical symmetries. For a rotation R∈SO(3)\mathbf{R} \in \text{SO}(3):

4
fθ(R x)=ρ(R) fθ(x)\quad f_\theta(\mathbf{R}\,\mathbf{x}) = \rho(\mathbf{R})\, f_\theta(\mathbf{x})

Explanation: Applying a rotation to the input coordinates must produce a predictably transformed output under the representation ρ\rho, so scalars (energy) are invariant and vectors (forces) rotate correctly. Networks satisfying this include SchNet, DimeNet, PaiNN, and MACE.

Bayesian Optimization for Experimental Design

Active learning couples the surrogate model with an acquisition function. The Expected Improvement (EI) acquisition is:

5
αEI(x)=E ⁣[max⁡ ⁣(f(x)−f∗,0)]=σ(x) ⁣[z Φ(z)+ϕ(z)],z=μ(x)−f∗σ(x)\quad \alpha_{\text{EI}}(\mathbf{x}) = \mathbb{E}\!\left[\max\!\left(f(\mathbf{x}) - f^*, 0\right)\right] = \sigma(\mathbf{x})\!\left[z\,\Phi(z) + \phi(z)\right], \quad z = \frac{\mu(\mathbf{x}) - f^*}{\sigma(\mathbf{x})}

Explanation: The acquisition function balances exploitation (high predicted mean μ\mu) and exploration (high uncertainty σ\sigma) by computing the expected gain over the current best observation f∗f^*, where Φ\Phi and ϕ\phi are the standard normal CDF and PDF. ```python

# Minimal Bayesian Optimization loop (BoTorch) import torch from botorch.models import SingleTaskGP from botorch.acquisition import ExpectedImprovement from botorch.optim import optimize_acqf

# Surrogate model on observed (X_obs, Y_obs) model = SingleTaskGP(X_obs, Y_obs).eval()

# Acquisition: maximize Expected Improvement EI = ExpectedImprovement(model, best_f=Y_obs.max()) candidate, acq_value = optimize_acqf( EI, bounds=bounds, q=1, num_restarts=10, raw_samples=256 )

# candidate is the next experiment to run ```

researcher · Demis Hassabis | b. 1976 | AlphaFold and AI-driven structural biology

Hassabis co-founded DeepMind after a career spanning neuroscience research at UCL and game AI development, bringing an explicitly neuroscience-inspired philosophy to building general AI systems. His most consequential contribution to scientific discovery is leading the AlphaFold project, which in 2020–2021 predicted the 3D structure of virtually every known protein to near-experimental accuracy, depositing over 200 million structures in a freely accessible database. A surprising personal fact: before AlphaFold, Hassabis was better known in AI circles for creating the Bullfrog Productions game Theme Park at age 17 — long before his neuroscience PhD.

researcher · Max Welling | b. 1968 | Geometric deep learning and molecular property prediction

Welling is a professor at the University of Amsterdam and a Research Director at Microsoft Research, known for foundational work on variational autoencoders and equivariant neural networks. His group developed SE(3)-equivariant networks that respect the rotational and translational symmetries of 3D molecular systems, directly enabling accurate force-field learning. Welling has argued publicly that the scientific community vastly underestimates how quickly AI will replace ab-initio simulation — a position that has sparked significant debate at NeurIPS workshops on ML for physical sciences.

Figure
AI Scientific Discovery Pipeline
AI for Scientific Discovery — End-to-End Pipeline Raw Data sequences, spectra, crystal structures Representation graph, SMILES, voxel, sequence ML Model GNN / Transformer Equivariant Net Diffusion Model Predictions property, structure, energy + uncertainty Experiment wet lab, synchrotron, accelerator run Active Learning Feedback Loop (Bayesian Optimization / Acquisition Function) Biology / Genomics Chemistry / Drug Design Materials Science Physics / Climate
Figure
Equivariant GNN Message Passing
Equivariant GNN — Molecular Message Passing i h_i j1 h_j1 j2 h_j2 j3 h_j3 e_ij1 dist e_ij2 dist e_ij3 dist ⊕ (permutation-invariant sum) h_i^(ℓ+1) SO(3) equivariance: rotate coords → forces rotate consistently

Intuition

The Mental Model

Imagine the entire space of possible drug-like molecules as an ocean. Traditional drug discovery is like randomly casting fishing nets — you might catch something useful after years of effort. Classical computational chemistry is like carefully sailing to specific coordinates with sonar — rigorous but glacially slow. AI for scientific discovery is like having a map drawn by someone who has already explored most of the ocean: the model has "read" millions of known molecules and, without being told explicitly, has learned which regions of chemical space contain calm waters (stable, non-toxic, bioavailable compounds) and which are treacherous shoals. The active learning loop is your navigation system: after each expedition (experiment), you update the map with new soundings and choose your next destination by asking "where does my map have the highest uncertainty AND the highest expected reward?" This is exactly what the Expected Improvement acquisition function computes — it combines the map's confidence (surrogate model uncertainty) with its predicted treasure (predicted property value) to suggest the single best next place to explore. The equivariance constraint is like insisting your map doesn't change if you rotate your compass — the laws of physics don't care which way north points, and your model shouldn't either.

Historical Development

Origins · 1 minhow this was discovered
History
1989
Yann LeCun's convolutional networks establish the principle of building symmetry (translation invariance) into neural architectures — the conceptual ancestor of equivariant molecular networks.
1995
Hohenberg-Kohn theorem is already 30 years old; density functional theory (DFT) becomes the gold standard for quantum chemistry but scales as O(N³), motivating neural network surrogates.
2007
Behler & Parrinello introduce the first generation of neural network potentials (NNPs), learning DFT-quality energies at molecular dynamics speed — the direct precursor to modern force fields.
2017
Gilmer et al. (Google Brain) publish "Neural Message Passing for Quantum Chemistry" (MPNN), unifying graph-based molecular property prediction under a single framework and achieving state-of-the-art on QM9.
2018
Olivecrona et al. demonstrate reinforcement learning for de-novo drug design using recurrent networks, generating molecules with desired bioactivity profiles from scratch.
2020
AlphaFold2 (DeepMind) achieves near-experimental accuracy on CASP14 protein structure prediction, winning decisively; the scientific community describes it as solving the protein folding problem.
2021
RFdiffusion and ProteinMPNN (Baker Lab, UW) combine diffusion models with inverse folding to design entirely novel proteins with experimentally validated function — discovery, not just prediction.
2022
GNoME (DeepMind) discovers 2.2 million stable crystal structures — 10× the previously known total — and experimentally validates hundreds, demonstrating AI-driven materials discovery at scale.
2023
AIMNet2 and MACE-MP-0 establish universal neural network potentials pre-trained on entire periodic tables, enabling zero-shot simulation of arbitrary materials without task-specific retraining.

Why It Matters

Real-World Impact
AlphaFold2 / ESMFoldPredicted 3D structures for >200 million proteins across nearly all known organisms; researchers at Wellcome Sanger Institute used the database to identify novel drug targets for neglected tropical diseases within months of release, compressing years of crystallography.
RFdiffusion (Baker Lab)Uses SE(3)-equivariant diffusion to hallucinate protein backbones conditioned on functional constraints; experimentally validated binders have been designed against cancer targets, viral glycoproteins, and cell-surface receptors, with binding affinities rivaling evolved antibodies.
GNoME (Google DeepMind)Graph Networks for Materials Exploration screened 381,000 novel stable inorganic crystal structures; 736 have already been synthesized by autonomous robotic labs, dramatically accelerating the hunt for next-generation battery electrolytes and superconductors.
NVIDIA FourCastNet / Google GraphCastMedium-range weather forecasting models trained on ERA5 reanalysis data; GraphCast achieves 10-day global forecasts in under 1 minute vs. 60 minutes for the ECMWF IFS ensemble, with comparable accuracy — a direct science-accelerating application of transformer-based spatiotemporal modeling.
Insilico Medicine's INS018_055AI-designed drug candidate for idiopathic pulmonary fibrosis generated using a generative chemistry pipeline; entered Phase II clinical trials in 2023, representing the first AI-designed small molecule to reach mid-stage clinical evaluation.
CERN Anomaly Detection (CWoLa / LHC)Weakly-supervised and self-supervised anomaly detection models now scan LHC collision events in real time, searching for beyond-Standard-Model physics signatures that traditional cut-based triggers would discard — turning particle physics into an ML-accelerated discovery pipeline.
Evo (Arc Institute)A 7B-parameter sequence foundation model trained on 2.7 million microbial genomes predicts gene essentiality, CRISPR guide RNA efficiency, and designs synthetic regulatory elements, demonstrating that genomic foundation models transfer across scales from nucleotides to whole genomes.

Prerequisites

What This Leads To

See Also

Graph Neural Networks · Protein Structure Prediction · Generative Models for Drug Discovery · Neural Ordinary Differential Equations · Active Learning and Experimental Design

Before this