Using Structure Prediction to Triage a Target List

This document outlines a generalized pattern synthesized from publicly discussed deployments across early-stage discovery teams, rather than a single verified case study. Real-world implementation details, resource requirements, and pipeline outcomes vary based on target class and organizational infrastructure.

The underlying capability — proteome-scale predicted structures — comes from the work described in AlphaFold predicts structures for 200M+ proteins. This piece covers what a discovery team does with it.

The problem

Early-stage target discovery often generates broad lists of candidate proteins derived from differential expression analysis, functional genomics, or phenotypical screens. While these hit lists provide valuable biological links to disease, moving every candidate into experimental validation is costly and slow. Wet-lab target triage requires expressing, purifying, and characterizing dozens or hundreds of proteins, many of which ultimately prove non-druggable due to lack of defined binding pockets or shallow surface topology.

Historically, teams without experimental atomic structures were forced to rely on sequence homology or invest in expensive crystallization campaigns just to evaluate basic target tractability. This created a severe bottleneck at the entry point of structure-based pipelines. Discovery groups frequently spent months attempting to resolve proteins that were intrinsically disordered or lacked small-molecule binding cavities. To streamline downstream investment, teams require a rapid, scalable computational filter to prioritize targets with tractable binding pockets before committing wet-lab resources to biophysical assays and protein production.

How teams approach it

To address this bottleneck, computational chemistry and discovery teams deploy automated 3D structure prediction pipelines across large target candidate lists [1][2][3]. Rather than treating machine-learning predictions as definitive physical structures, mature teams use predicted models primarily for hypothesis generation and initial structural triaging.

By predicting the unliganded tertiary structures of target candidates, researchers can systematically screen for pocket volume, depth, enclosure, and surface charge distribution. This triage step allows teams to categorize targets into high-priority candidates with distinct, druggable cavities, lower-priority targets requiring secondary functional analysis, and non-viable candidates dominated by structural disorder. Crucially, successful teams integrate these computational outputs with biological rationale, using structure prediction not as a single decision-maker, but as a filtering step that guides construct design and wet-lab resource allocation.

What the workflow looks like

The triage workflow typically follows a structured four-stage progression that moves from raw sequence inputs to prioritized experimental targets.

First, teams ingest sequence lists for target candidates and run automated batch prediction runs to generate three-dimensional coordinate files. Second, the pipeline extracts local and global confidence metrics from the structural output. Per-residue confidence scores (pLDDT) and predicted aligned error metrics are evaluated to map high-confidence domain structures while identifying flexible loops, disordered regions, and uncertain inter-domain orientations [4].

Third, automated pocket-finding algorithms evaluate the high-confidence domains. These tools calculate pocket volume, hydrophobicity, enclosure metrics, and druggability scores across candidate surfaces. Fourth, the combined confidence and pocket data are compiled into a central dashboard where structural biologists and computational chemists review high-ranking candidates. Targets meeting defined structural criteria are then selected for construct design, expression, and experimental biophysical screening.

How it works in practice

• What data goes in. Amino-acid sequences for the target list, any available experimental structures and homologues, multiple-sequence alignments, and the team's own annotations about tractability, expression history and prior programme outcomes.
• What the model predicts or ranks. A predicted three-dimensional structure with a per-residue confidence score, and — downstream of that — a rough ordering of which targets look structurally tractable enough to justify experimental effort.
• Technique or model family. Deep-learning structure prediction of the AlphaFold or ESMFold family, followed by ordinary pocket-detection and druggability heuristics run over the predicted coordinates.
• How it is trained or validated. The prediction models are pre-trained and used as-is; the local validation question is whether the confidence scores track reality on targets the team already has structures for. Low-confidence regions, especially loops and disordered segments, are treated as unresolved rather than as geometry.
• What the team does with the output. High-confidence, pocket-bearing targets go forward to structural biology and screening; low-confidence ones either get an experimental structure or wait. The ranking allocates crystallography and cryo-EM time, it does not replace it.
• A representative example. Input: a 480-residue kinase-adjacent sequence with no solved structure. Output: a predicted fold with mean per-residue confidence of 88 across the core and 42 across a 30-residue insertion. The team treats the core pocket as usable for early docking and flags the insertion as needing experimental work before anyone designs against it.

Where it breaks

Despite their utility, structure prediction models have well-documented failure modes that can derail discovery pipelines if interpreted uncritically. A primary vulnerability is the misinterpretation of confidence metrics. High local per-residue confidence scores frequently mask incorrect global domain arrangements or misaligned quaternary structures, leading teams to assume a rigid inter-domain relationship that does not exist in solution [4].

Furthermore, predicted structures generally reflect an unliganded, ground-state conformation. They frequently fail to capture induced-fit dynamics, cryptic pockets that open only upon ligand binding, or alternative active conformations required for biological function. Multimeric complexes, targets reliant on post-translational modifications, and integral membrane proteins also present significant challenges where automated predictions can yield misleading structural topologies.

Finally, predicted models are rarely sufficient on their own for hit expansion or fine-grained structure-based drug design. Attempting virtual screening or molecular docking against predicted pockets without experimental confirmation frequently generates high rates of false positives, as subtle side-chain errors inside binding pockets severely disrupt physics-based docking scoring functions.

What good looks like

High-performing teams treat structure prediction as a guide for wet-lab strategy rather than a replacement for experimental structural biology. They use predicted models to inform expression construct boundaries, select stable domain constructs, and design biophysical screening strategies.

These teams establish strict protocol rules regarding structural confidence metrics. Regions with low confidence scores are filtered out prior to pocket analysis to prevent false-positive cavity detection. When a target passes initial computational triage and shows encouraging hit-finding activity, mature pipelines immediately initiate experimental structure determination through X-ray crystallography or cryo-electron microscopy to obtain accurate ligand-bound structures.

Questions to ask before you start

• Does the target rely on oligomerization, cofactor binding, or post-translational modifications that current prediction models fail to capture accurately?
• What explicit threshold of local and global confidence metric scores will trigger secondary manual review or exclusion from the pipeline?
• How will predicted structural boundaries be used to inform protein expression construct design and biophysical validation assays?
• Is the team prepared to transition to experimental structural methods as soon as a target candidate moves into active chemistry and hit validation?
• Are computational pocket metrics integrated with target biology and disease relevance, rather than evaluated in isolation?

Sources

1. Jumper et al., Highly accurate protein structure prediction with AlphaFold, Nature, 2021.
2. EMBL-EBI, AlphaFold DB expands to over 200 million structures, 2022.
3. Lin et al., Evolutionary-scale prediction of atomic-level protein structure with a language model (ESMFold), Science, 2023.
4. EMBL-EBI, AlphaFold Protein Structure Database FAQ — interpreting pLDDT and PAE.