Life Science: From AlphaFold to Protein Design

Bottom line: Structure prediction ("reading" proteins) has been solved by the AlphaFold family to near-experimental accuracy, while protein design ("writing" proteins) is being redefined by diffusion models. For AI-Infra engineers, the real challenge is not model innovation — it is fitting three compute patterns (MSA depth, equivariant graph networks, and diffusion sampling) into one GPU engineering stack.

1. From Sequence to Structure: Why MSA Is "First Principles"

AlphaFold 2's most counter-intuitive design: it does not learn a direct "sequence → structure" map; it learns "evolutionary co-variation → structure".

• The input is not a single sequence but a multiple sequence alignment (MSA) — homologous sequences aligned into a matrix;
• The core signal is co-evolution: two columns that mutate together are likely in 3D contact;
• MSA depth directly determines quality: for orphan proteins with <100 homologs, even AlphaFold 2 cannot recover.

Engineering implication: MSA search (JackHMMER/MMseqs2) is itself the bottleneck. A single protein's MSA search can take minutes on CPU; prediction runs on GPU. The pipeline is classic heterogeneous orchestration — "CPU search → GPU predict" — and batch pipelines need queues to keep both compute stages fed.

2. AlphaFold 2: Evoformer and the Structure Module

AlphaFold 2's skeleton splits into two blocks:

Module	Role	Compute Profile
Evoformer	Exchanges information across MSA rows/columns, distilling evolutionary signal	2D attention; memory grows quadratically with sequence length
Structure module	Folds abstract representations into 3D coordinates with equivariant attention	Per-residue iterative refinement, 28 recycling passes

Key engineering points:

• Memory is dominated by Evoformer MSA attention: long sequences (>1000 residues) need chunking or downsampling to fit a single GPU;
• Recycling dominates latency: 28 serial passes cannot be parallelized; inference latency scales roughly linearly with residue count;
• Confidence outputs (pLDDT/PAE) are not decoration: regions below threshold must be treated as untrustworthy; downstream consumers (e.g. docking) should filter by PAE first.

3. AlphaFold 3 and the Diffusion Era

AlphaFold 3 extends scope from proteins to protein–ligand–nucleic-acid–ion complexes, and swaps structure generation for a diffusion model:

• Pairformer partially replaces Evoformer logic, injecting covalent bonds and chemical constraints;
• Diffusion sampling: start from random coordinates, iteratively denoise to a plausible conformation — prediction becomes a "conditional generation" problem;
• The payoff: complexes and modifications (glycosylation, metal ions) can be predicted; the cost: sampling is far more expensive than a single forward pass — many denoising steps per prediction.

Engineering view: AlphaFold 3 resembles a "structure diffusion model" — multi-step sampling × a large model, structurally the same compute profile as text-to-image. For batch docking (drug-screening pipelines over hundreds of thousands of complexes), sampling steps are the #1 cost lever — 10-step accelerated sampling often suffices for ranking.

4. Equivariant GNNs and RoseTTAFold

• RoseTTAFold: a three-track architecture (sequence/structure/MSA streams feeding each other), lighter than AlphaFold 2, runs on a single GPU, roughly an order of magnitude faster;
• Equivariant GNNs (e.g. SE(3)-Transformers): make the network invariant to rotations/translations — "rotate the coordinates, predictions rotate along." This is the fundamental constraint distinguishing structural modeling from CV;
• In practice RoseTTAFold serves as a fast screen: filter candidates first, refine with AlphaFold second — a two-stage pipeline spends compute where it counts.

5. Protein Design: Diffusion's "Writing" Power

The shift from "reading" to "writing":

• RFdiffusion: diffuses over residue coordinates to generate entirely new backbones, then backfills sequences;
• Generative design pipeline: objective functions (binding affinity, stability) as conditions → diffusion-sampled backbones → sequence design (ProteinMPNN-class models) → structural validation (AlphaFold re-scoring);
• The key loop: designs must be "re-read" by structure-prediction models to verify — generation–validation–iteration is the only firewall before the experimental loop.

6. Engineering Checklist

• Heterogeneous pipeline: MSA search (CPU, multicore) → prediction (GPU) → confidence filtering, decoupled by queues;
• Batch inference: deduplicate and cluster MSAs first — homolog redundancy can waste 50%+ of compute;
• Sampling-vs-quality tiers: define "ranking-grade" and "publication-grade" accuracy, configure diffusion steps per tier;
• Memory planning: estimate attention memory as L² for long complexes; chunk sequences when needed;
• Data versioning: MSA database versions must be stored with predictions — otherwise results are not reproducible.

Success criterion: closure. A protein that goes from sequence → MSA → prediction → confidence filter → design → re-scoring, end-to-end reproducible and auditable, is what it means to be inside the engineering system of AI life science.