ProMIM: Dual Frameworks for Protein and Vision Models

Updated 8 July 2026

ProMIM is a term that encompasses distinct frameworks: one for protein mutational effect prediction using hierarchical self-supervised learning, and another for masked image-guided prompt learning in vision-language models.
In protein science, ProMIM integrates three objectives—PIM, BIM, and SIM—to model sidechain, backbone, and complex-level interactions, thereby improving ΔΔG prediction accuracy.
In vision-language modeling, ProMIM refines prompt generation by conditioning on masked-image features, enhancing generalization in CLIP-style pipelines.

ProMIM is an acronym used for distinct methods in recent literature. In protein science, it denotes a self-supervised multi-level pre-training framework for protein mutational effect prediction, designed to estimate mutation-induced changes in binding free energy, $\Delta\Delta G$ , by modeling protein–protein interactions across sidechain, backbone, and complex levels (Mo et al., 2024). In vision-language modeling, the same acronym denotes Masked Image Modeling-guided Conditional Prompt Learning, a plug-and-play framework that conditions prompt generation on masked-image features rather than full-image features to improve generalization to unseen classes in CLIP-style pipelines (Bui et al., 7 Aug 2025). The term does not denote the ProM/RapidProM prototype-selection method (Sani et al., 2019), the process-mining alignment framework PIMA (Chen et al., 2017), or Projected Markov Models and their HMM realization (Noe et al., 2013).

1. Nomenclature and scope

In the protein-interaction literature, ProMIM is introduced as Multi-level Interaction Modeling for Protein Mutational Effect Prediction (Mo et al., 2024). Its central application is mutation-effect prediction for protein–protein interactions, where the target is the mutation-induced change in binding free energy, $\Delta\Delta G$ . The method is explicitly motivated by the observation that mutations can affect interactions hierarchically across three levels: sidechain conformation, backbone conformation, and binding affinity.

In a separate line of work, ProMIM is introduced as Masked Image Modeling-guided Conditional Prompt Learning for vision-LLMs such as CLIP (Bui et al., 7 Aug 2025). There, the acronym refers to a prompt-learning regularization mechanism rather than a protein model. This naming collision is substantive rather than superficial: the two methods address different tasks, use different architectures, and rely on different training objectives.

Several neighboring terms are easy to conflate with ProMIM but are explicitly distinct in the supplied literature. “Process-oriented Iterative Multiple Alignment” is abbreviated PIMA, not ProMIM (Chen et al., 2017). The ProM/RapidProM paper on prototype selection is implemented in the ProM ecosystem but “does not introduce or discuss any method, package, or concept named ProMIM” (Sani et al., 2019). Likewise, the paper on Projected Markov Models develops PMMs and HMMs for molecular kinetics rather than a method named ProMIM (Noe et al., 2013).

2. ProMIM in protein mutational effect prediction

ProMIM in its protein-science sense is a structure-based, self-supervised framework for predicting the effects of mutations on protein–protein interactions, with the downstream output being $\Delta\Delta G$ (Mo et al., 2024). The paper denotes a protein complex as

$\mathbf{c}\in\mathbb{A}^n,$

where $\mathbb{A}=\{1,\dots,20\}$ is the amino-acid alphabet. A complex contains two binders, indexed by disjoint residue sets $\mathcal{P}^r$ and $\mathcal{P}^l$ . Each residue $i$ is described by amino-acid type $a_i$ , position $\boldsymbol{p}_i\in\mathbb{R}^3$ , orientation $\Delta\Delta G$ 0, and sidechain torsions $\Delta\Delta G$ 1 with $\Delta\Delta G$ 2.

The framework is organized around a hierarchical biological hypothesis. In the paper’s causal picture,

$\Delta\Delta G$ 3

This view is used to distinguish ProMIM from earlier methods that, in the authors’ characterization, largely focus on sidechain-level modeling. The motivation is especially explicit for mutations involving glycine and proline, which are described as cases where backbone flexibility and conformation may change substantially.

The three interaction levels are formalized as follows. At the sidechain / residue level, mutations alter residue identity and rotamer distribution. At the backbone / conformation level, altered local packing may perturb interface geometry and inter-chain distances. At the protein / complex level, the mutation may change whether two proteins bind well and thus affect the overall binding relation. This suggests that mutational effect prediction should not be treated as a purely local sidechain problem.

3. Architecture and self-supervised objectives

ProMIM starts from a structure encoder derived from RDE (Mo et al., 2024). During pretraining, the two binding proteins are randomly assigned as receptor and ligand; to simulate an unbound state, the receptor remains fixed while the ligand undergoes random rotation and translation. The model extracts single features for each residue—residue type, backbone dihedral angles, and local atom coordinates—and pair features for residue pairs—relative position and residue-type pair information. These are embedded as $\Delta\Delta G$ 4 and $\Delta\Delta G$ 5 and processed by an Invariant Point Attention (IPA) module to produce residue-wise hidden representations $\Delta\Delta G$ 6. The appendix specifies a single representation size of 128, a pair representation size of 64, and 6 IPA blocks.

A key implementation detail is that two encoders with the same architecture are trained independently. One is trained on PPIRef50K for protein-level interaction modeling and backbone-level interaction modeling; the other is trained on PDB-REDO for sidechain-level interaction modeling. For downstream $\Delta\Delta G$ 7 prediction, hidden representations from the pretrained encoders are concatenated with single representations and fused with a one-layer MLP; max-pooling then yields global wild-type and mutant representations, which are passed to an MLP trained with MSE loss.

The protein-level objective, PIM, is a contrastive protein-matching loss. If $\Delta\Delta G$ 8 and $\Delta\Delta G$ 9 are global ligand and receptor embeddings from a batch of $\Delta\Delta G$ 0 complexes, positives are same-complex pairs and negatives are cross-complex pairs. The loss is

$\Delta\Delta G$ 1

with cosine similarity and temperature $\Delta\Delta G$ 2 used in the ligand-to-receptor and receptor-to-ligand matching terms. This objective is intended to capture complex-level binding compatibility.

The backbone-level objective, BIM, predicts the inter-chain $\Delta\Delta G$ 3 distance map. Starting from residue representations $\Delta\Delta G$ 4, the model constructs pairwise representations, refines them with a Transformer, predicts a distance map $\Delta\Delta G$ 5, and optimizes

$\Delta\Delta G$ 6

This objective is the component specifically designed to capture mutation-sensitive backbone and interface geometry changes.

The sidechain-level objective, SIM, is adopted from RDE and uses a conditional normalizing flow to model sidechain torsion distributions. With residue representation $\Delta\Delta G$ 7, the conditional density obeys

$\Delta\Delta G$ 8

and the training loss is

$\Delta\Delta G$ 9

In implementation terms, ProMIM is therefore not a single monolithic pretraining loss but a coordinated multi-level design in which PIM and BIM are trained jointly on PPIRef50K and SIM is trained separately on PDB-REDO.

4. Data, training protocol, and benchmark results

The pretraining corpora are structurally distinct. PPIRef50K, used for PIM and BIM, contains 45,553 PPIs, each involving two interacting protein chains (Mo et al., 2024). PDB-REDO, used for SIM, contains over 130,000 refined X-ray structures; after clustering at 50% sequence identity, it yields 38,413 chain clusters. This split reflects the differing requirements of the objectives: PIM and BIM need explicit interacting binders, whereas SIM benefits from larger structural coverage.

For PIM and BIM training, the model randomly samples 64 residues from each binder, yielding 128 residues total. Training uses Adam, runs for 200K iterations, starts at learning rate $\mathbf{c}\in\mathbb{A}^n,$ 0, decays the learning rate by multiplying by 0.8 if validation loss does not decrease in the last 5 validation steps, enforces a minimum learning rate of $\mathbf{c}\in\mathbb{A}^n,$ 1, and uses batch size 48. SIM training follows the preprocessing and hyperparameters of RDE.

The supervised benchmark is SKEMPI2, evaluated with 3-fold cross-validation split by structure (Mo et al., 2024). On all mutations, ProMIM achieves Pearson 0.672, Spearman 0.573, RMSE 1.516, MAE 1.089, AUROC 0.760, per-structure Pearson 0.464, and per-structure Spearman 0.431. Compared to the strongest pretraining baselines reported in the paper, RDE-Net attains Pearson 0.654, Spearman 0.555, AUROC 0.749, and per-structure Spearman 0.424, while DiffAffinity attains Pearson 0.661, Spearman 0.544, AUROC 0.742, and per-structure Spearman 0.392.

On single-point mutations, ProMIM reports Pearson 0.668, Spearman 0.534, RMSE 1.279, MAE 0.924, AUROC 0.738, per-structure Pearson 0.466, and per-structure Spearman 0.439 (Mo et al., 2024). On multi-point mutations, where the paper emphasizes the clearest gains, ProMIM reports Pearson 0.666, Spearman 0.614, RMSE 1.963, MAE 1.491, AUROC 0.825, per-structure Pearson 0.458, and per-structure Spearman 0.425. This is presented as evidence that backbone- and complex-level modeling become particularly important when mutations induce larger conformational changes.

Ablation studies are used to isolate the contribution of each pretraining level. ProMIM improves over the unpretrained variant ProMIM* on per-structure Spearman from 0.398 to 0.431, described in the paper as an 8.29% relative improvement (Mo et al., 2024). For all mutations, PIM alone gives Pearson 0.649 and Spearman 0.544; BIM gives 0.623 and 0.528; SIM gives 0.654 and 0.555; the full PIM + BIM + SIM model gives 0.672 and 0.573. This establishes SIM as the strongest single objective overall while supporting the paper’s claim that the combined multi-level representation is the most robust general-purpose predictor.

5. Backbone-sensitive analysis and zero-shot generalization

The paper’s most direct test of its central hypothesis uses a backbone-sensitive subset consisting of single-point mutations from or to glycine or proline (Mo et al., 2024). On this subset, SIM alone performs worst, even worse than the unpretrained model; BIM improves significantly over SIM; and PIM + BIM performs best. The authors use this result to argue that sidechain-only modeling is insufficient when mutation effects propagate through backbone rearrangement.

Zero-shot results extend the argument beyond SKEMPI2. On 285 single-point mutations across 15 significant mutation sites on SARS-CoV-2 RBD affecting ACE2 binding, ProMIM achieves Pearson 0.483, compared with 0.385 for FoldX, 0.438 for RDE-Net, and 0.466 for DiffAffinity (Mo et al., 2024). In antibody optimization against SARS-CoV-2, the task is to rank 494 possible mutations and identify 5 favorable single-point mutations across 26 sites in the heavy-chain CDR. The reported rank ratios for ProMIM are 5.33% for TH31W, 4.79% for AH53F, 19.43% for NH57L, 75.78% for RH103M, and 8.37% for LH104F. The paper highlights that ProMIM is the only method to rank 3 of 5 favorable mutations within the top 10% and 4 of 5 within the top 20%.

These results suggest a specific practical interpretation. ProMIM’s advantage is not merely that it uses structural pretraining, since several baselines also do so; rather, the paper argues that its advantage comes from aligning three self-supervised objectives with three hypothesized biological levels of mutation propagation. A plausible implication is that ProMIM is especially attractive when the relevant mutation regime is not known in advance: some mutations may be dominated by sidechain repacking, whereas others may involve substantial backbone-mediated interface change.

In vision-language modeling, ProMIM denotes a different method: a plug-and-play framework that augments conditional prompt learning by feeding masked-image features to the prompt-generation meta-network while retaining full-image features for classification (Bui et al., 7 Aug 2025). In the 11-dataset base-to-new benchmark with CLIP ViT-B/16 and 16-shot, CoCoOp reports Base 80.47, New 71.69, and harmonic mean 75.83, while CoCoOp + ProMIM reports Base 80.64, New 73.96, and harmonic mean 77.16. The paper also reports essentially unchanged learnable parameter count and negligible overhead: 69K vs. 69K parameters, 420 min vs. 423 min training time, 127.06 ms vs. 127.16 ms inference time, and 11310 MB vs. 11312 MB memory.

Several other papers are conceptually adjacent but do not define ProMIM. The prostate MRI paper on 3D Masked Modelling studies a 3D CNN masked autoencoder for lesion classification and is described as “directly relevant conceptually” to ProMIM-style masked pretraining, but it does not introduce a method called ProMIM (Fernandez-Quilez et al., 2022). ProtFIM addresses protein sequence fill-in-the-middle design and is described as a relevant baseline or precursor for protein infilling, not as ProMIM (Lee et al., 2023). In process mining, the prototype-selection paper is relevant to the ProM ecosystem but “is not about ‘ProMIM’” (Sani et al., 2019). Likewise, PIMA is the method introduced for medical trace alignment, and the string “ProMIM” does not appear in that paper (Chen et al., 2017). In molecular kinetics, the foundational framework is Projected Markov Models (PMMs) with HMM approximations rather than a method named ProMIM (Noe et al., 2013).

The acronym therefore has no single cross-domain referent. In current research usage represented here, “ProMIM” most often requires disambiguation by field: in protein interaction modeling it refers to hierarchical self-supervised mutation-effect prediction (Mo et al., 2024), whereas in vision-language learning it refers to masked-image-conditioned prompt adaptation (Bui et al., 7 Aug 2025).