Fine-tuning OpenFold3 on a small set of structures: The PDE10A case study

The goal of this experiment was to assess whether fine-tuning the public OpenFold3 model on a very small set of protein–ligand structures can meaningfully improve prediction quality for a specific drug discovery context, without degrading general performance. We focused on human phosphodiesterase 10A (PDE10A), a target class where pose accuracy and SAR alignment are critical for design decisions.This short blog documents a concrete fine-tuning experiment on human phosphodiesterase 10A (PDE10A).The goal is to:

1. show that small numbers of liganded structures can materially increase the capabilities of co-folding models
2. provide enough detail that a structural modelling team could reproduce or extend the experiment in their own stack

1. Question and setup

The practical question was:

Can we correct systematic pose errors of OpenFold3 for a specific target and chemotype using only a handful of protein–ligand complexes?

We chose to evaluate OF3 finetuning on a subset of 27 structures from the PDE10A dataset published by Roche in 2022, as it met the two key requirements:

• Held-out to the OF3 training set (since published after the training cutoff in 2021)
• The base OF3 model was performing poorly on these structures

We treated this as a realistic “low-n” fine-tuning scenario:

• Target: human PDE10A
• Model: OpenFold3 (public weights as of late 2025)
• Hardware: single NVIDIA H100 GPU
• Total wall-clock for fine-tuning: ~20 hours

2. Data: training and evaluation splits

Training set (10 complexes, used for fine-tuning) PDB IDs:

5SDY, 5SIQ, 5SI7, 5SIG, 5SI5, 5SI8, 5SIY, 5SG5, 5SGL, 5SIH

Evaluation set (17 held-out complexes, no gradient updates) PDB IDs:

5SH0, 5SE0, 5SHR, 5SJL, 5SH8, 5SF4, 5SFG, 5SE5, 5SHK, 5SEE, 5SFL, 5SJU, 5SKE, 5SKU, 5SKO, 5SEA, 5SKR

3. Fine-tuning protocol

We followed a minimal, single-target fine-tuning setup:

• Starting weights: public OpenFold3
• Objective: standard OpenFold3 structure loss on protein–ligand complexes
• Steps: ~350 gradient steps over the 10 training complexes
• Optimiser / schedule: Decreased EMA decay factor to 0.99 from default 0.999 to allow meaningful model change in a small number of gradient steps; decreased learning rate warmup steps from 1000 to 50; decreased learning rate from 0.0018 to 0.0003
• MSAs / templates: as in baseline OpenFold3, but without use of templates

All training was run via ApherisFold in a private environment, with the PDE10A complexes staying inside the organisation’s own infrastructure.

4. Evaluation protocol

For each complex in the 17-structure evaluation set we:

1. Ran 5 independent co-folding samples with the baseline OpenFold3 model.
2. Ran 5 independent co-folding samples with the fine-tuned model.
3. For each model, selected the highest-confidence sample according to the model’s internal confidence (plDDT).
4. Computed the following metrics vs. the experimental structure:global GDTintra-protein lDDTintra-ligand lDDTprotein–ligand interface lDDTDockQ (interface-focused composite score)

5. Results

The bar plot (baseline vs fine-tuned across these metrics, with error bars across the 17 structures) summarises the results.

‍

Results PDE10A case study

Qualitatively:

• For 5SH8, the baseline model places the ligand in a wrong pose relative to the pocket; key ring systems are mis-registered and several interactions are missing.
• The fine-tuned model produces a pose that overlays well with the crystal structure and recovers the expected interaction pattern.
• Inspection shows that the fine-tuned model appears to combine elements of two training complexes (notably 5SDY and 5SI7) to achieve the correct binding mode.

The complex 5SH8 (see purple structure) was used as a central qualitative example:

• Pink: baseline OpenFold3 prediction (incorrect pose)
• Yellow: the significantly more accurate prediction of the finetuned model
• Purple: the reference structure of complex 5SH8
• Green and grey: two most similar structures to 5SH8 (based on spatial overlap) used for fine-tuning (5SDY and 5SI7) to the purple reference structure of the held-out structure. These two training structures are particularly close analogues and illustrate how the fine-tuned model “interpolates” between known binding modes when predicting 5SH8.

‍

Output structures of the PDE10A case study

Quantitatively (across all 17 held-out complexes):

• All structural metrics showed a clear shift in favour of the fine-tuned model.
• Improvements were most pronounced at the protein–ligand interface (interface lDDT and DockQ), which is exactly where medicinal chemists care about accuracy.
• Global protein metrics (GDT, intra-protein lDDT) improved slightly but were already high; the main gain was correcting ligand orientation and local interactions.

Overall, fine-tuning on 10 carefully chosen complexes was enough to move predictions for this chemotype from qualitatively unreliable to experimentally plausible for decision support.

For teams who want to build a co-folding capability into their discovery engine, ApherisFold provides an enterprise-ready way to run models like OpenFold3 and Boltz-2 locally, benchmark them on proprietary structures, and operate fine-tuning workflows across programs as a standard part of R&D.

ApherisFold