FIRM-DTI: Geometry-Aware Drug-Target Prediction

• The paper introduces FIRM-DTI, a novel framework that uses FiLM conditioning and triplet loss to align molecular and protein embeddings in a geometry-aware latent space.
• It leverages pretrained models to generate fixed-dimensional embeddings and modulates ligand features with protein context to achieve smooth and interpretable binding affinity predictions.
• Empirical results show competitive performance in PCC, AUPR, and AUROC while ensuring domain generalization under temporal shifts with parameter efficiency.

Searching arXiv for the specified paper to ground the article and citation. FIRM-DTI is a framework for drug–target interaction via conditioning and metric learning that addresses binding affinity prediction between drug molecules and protein targets through a geometry-aware architecture. It is introduced as a lightweight method that conditions molecular embeddings on protein embeddings through a feature-wise linear modulation (FiLM) layer, enforces metric structure with a triplet loss, and uses an RBF regression head operating on embedding distances to produce smooth, interpretable affinity predictions (Refahi et al., 25 Sep 2025). In the reported evaluation, the method is positioned against simple concatenation-based fusion schemes and against larger or more complex baselines, with particular emphasis on domain generalization under temporal shift.

1. Conceptual formulation

FIRM-DTI is defined by three key stages: generation of fixed-dimensional embeddings for ligands and protein targets using pretrained models, modulation of ligand embeddings by protein embeddings through a FiLM conditioning layer, and affinity prediction through a distance-based RBF regression head (Refahi et al., 25 Sep 2025). During training, triplet loss is used to enforce metric structure so that interacting pairs are closer than non-interacting pairs in the latent space.

The central premise is that binding affinity should be modeled not merely as a function of concatenated molecular and protein representations, but as a function of their geometry after target-aware conditioning. The method is therefore explicitly described as geometry-aware. This suggests that FIRM-DTI is organized around latent-space structure rather than solely around discriminative prediction, with metric regularization serving as a direct inductive bias for drug–target interaction modeling.

The abstract characterizes the method as a response to limitations in existing deep learning approaches for drug-target binding affinity prediction, specifically that many models fuse ligand and protein representations via simple concatenation and lack explicit geometric regularization, resulting in poor generalization across chemical space and time (Refahi et al., 25 Sep 2025). Within that framing, FIRM-DTI emphasizes conditioning and metric learning as the principal design interventions.

2. Representation pipeline and conditioning mechanism

FIRM-DTI represents molecules and proteins using fixed-dimensional embeddings. The molecule embedding, denoted $z_d$, is derived using the MolE foundation model, which represents a molecule as a graph in which atom-level features, including Daylight atomic invariants and Morgan fingerprints, plus relative bond distances as positional information, are fed into a disentangled attention transformer, yielding $z_d \in \mathbb{R}^d$ (Refahi et al., 25 Sep 2025). The protein embedding, denoted $z_t$, is computed using the ESM2 protein LLM, which encodes amino acid sequences to a per-sequence embedding by averaging residue representations, yielding $z_t \in \mathbb{R}^d$.

The conditioning mechanism is implemented through FiLM. Rather than concatenating ligand and protein embeddings, FIRM-DTI modulates the ligand embedding with target-specific transformations: $$\mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t)$$ where $\gamma$ and $\beta$ are learned linear functions of the protein embedding $z_t$, and $\odot$ denotes element-wise multiplication (Refahi et al., 25 Sep 2025).

This formulation permits the model to scale and shift every feature of the molecule embedding according to the protein context. In the source description, this is presented as enabling richer target-specific transformations than concatenation. A plausible implication is that the ligand representation is not treated as invariant across targets; instead, it is reparameterized as a protein-conditional object before affinity regression.

3. Geometry-aware regression over latent distance

After conditioning, FIRM-DTI computes a cosine distance between conditioned and normalized embeddings for drug and target, denoted $\tilde{z}_d$ and $z_d \in \mathbb{R}^d$0: $z_d \in \mathbb{R}^d$1 This distance is then mapped to a predicted affinity through a radial basis function layer with $z_d \in \mathbb{R}^d$2 centers $z_d \in \mathbb{R}^d$3: $z_d \in \mathbb{R}^d$4

$z_d \in \mathbb{R}^d$5

where $z_d \in \mathbb{R}^d$6 and $z_d \in \mathbb{R}^d$7 are learnable parameters (Refahi et al., 25 Sep 2025).

The paper summary describes this head as encouraging a smooth, interpretable mapping from embedding geometry to predicted affinity, enforcing that similar embeddings yield similar outputs. In reported out-of-domain results, the framework is further described as encoding the true geometric relation between binding affinity and distance in the latent space, with $z_d \in \mathbb{R}^d$8 correlation between cosine distance and affinity via the RBF head (Refahi et al., 25 Sep 2025).

Within the method’s internal logic, the regression head is not an independent predictor appended to arbitrary representations; it is explicitly tied to a metric structure. This suggests that the interpretability claim derives from the fact that affinity is mediated by a scalar distance and then transformed through a structured basis expansion, rather than through an unrestricted black-box multilayer predictor.

4. Metric learning and optimization

Metric structure is enforced in FIRM-DTI through triplet loss. For every anchor protein $z_d \in \mathbb{R}^d$9, a positive compound $z_t$0 interpreted as a known binder, and a negative compound $z_t$1 interpreted as a non-binder, the loss is written as

$z_t$2

where $z_t$3 is cosine distance, $z_t$4 is the composed embedding function, and $z_t$5 is a margin governing minimal separation (Refahi et al., 25 Sep 2025). The stated effect is to pull true interacting pairs closer in metric space than non-interacting pairs.

For affinity regression, the framework uses a Huber loss: $z_t$6 The total loss is the sum of the triplet loss and the Huber affinity loss: $z_t$7 (Refahi et al., 25 Sep 2025)

The ablation study assigns particular importance to the triplet component. The reported findings state that triplet loss is critical, with the largest drop in PCC when it is removed, confirming the necessity of geometric metric structure, while FiLM conditioning also adds value by contextually adapting molecular features. This positions metric learning not as an auxiliary regularizer but as one of the framework’s defining elements.

5. Benchmarks and reported empirical performance

The primary benchmark is Therapeutics Data Commons DTI-DG (Drug–Target Interaction Domain Generalization), with a temporal split that trains on 2013–2018 data and tests on 2019–2021 patent data; the metric is the Pearson correlation coefficient between predicted and actual affinities (Refahi et al., 25 Sep 2025). Additional benchmarks are DAVIS, BindingDB, and BIOSNAP.

The reported ablation study on the primary benchmark gives the following PCC values:

Model variant	PCC (↑)
Full Model	0.59
w/o FiLM conditioning	0.55
w/o triplet loss	0.32

These values are used to support two specific conclusions in the source summary: that triplet loss is critical and that FiLM conditioning contributes measurable value (Refahi et al., 25 Sep 2025).

For standard DTI benchmarks, the reported metrics are AUPR and AUROC:

Dataset	Model / comparison	Result
BIOSNAP	FIRM-DTI	AUPR 0.919, AUROC 0.910
BIOSNAP	Best other baseline	AUPR 0.897, AUROC 0.883
BindingDB	FIRM-DTI	AUPR 0.647, AUROC 0.916
BindingDB	Best other baseline	AUPR 0.652, AUROC 0.917
DAVIS	FIRM-DTI	AUPR 0.460, AUROC 0.880
DAVIS	Best other baseline	AUPR 0.511, AUROC 0.917

The accompanying interpretation states that FIRM-DTI matches or exceeds the top baselines on two out of three datasets and remains robust on the most challenging, small/imbalanced dataset, DAVIS (Refahi et al., 25 Sep 2025). Because this wording appears in the source material, it is best understood as the authors’ summary of comparative behavior rather than as an externally standardized ranking claim.

6. Domain generalization, efficiency, and comparative positioning

On the DTI-DG benchmark, FIRM-DTI is reported to achieve state-of-the-art performance with $z_t$8, outperforming large or ensemble models including PLM-SWE, TxGemma27B with 27B parameters, and the Otter-Knowledge ensemble, which requires external knowledge graphs (Refahi et al., 25 Sep 2025). The source description also highlights parameter efficiency, stating that FIRM-DTI is orders-of-magnitude smaller than LLMs and does not rely on external knowledge.

This comparative positioning is central to how the framework is presented. The distinguishing characteristics enumerated in the source are explicit conditioning via FiLM for target-specific molecule representations, alignment of embedding geometry with binding affinity using triplet loss, smooth and interpretable RBF regression over embedding distance, and competitive or superior accuracy with fewer parameters and without external knowledge graphs (Refahi et al., 25 Sep 2025).

A plausible implication is that FIRM-DTI is designed to challenge the assumption that stronger drug–target interaction performance necessarily requires either very large parameter counts or auxiliary knowledge sources. In this reading, the framework’s contribution lies less in scaling and more in the choice of inductive biases: target-conditioned modulation and explicit metric alignment.

7. Interpretation, scope, and methodological significance

FIRM-DTI is described as a lightweight, geometry-aware architecture with explicit conditional and metric learning via FiLM and triplet loss, and its conclusion states that these components fundamentally distinguish it from traditional fusion by concatenation or black-box regression pipelines (Refahi et al., 25 Sep 2025). The emphasis on smooth, interpretable affinity prediction through distance-based RBF regression further situates the method as an attempt to make latent geometry directly consequential for prediction.

A common misconception in reading this class of models is to equate any joint embedding of molecules and proteins with geometry-aware learning. The FIRM-DTI description makes a narrower claim: geometry is not merely present because vectors exist in a latent space, but is explicitly enforced through triplet loss and operationalized through a regression head that consumes embedding distance. Under this interpretation, the framework’s notion of geometry-aware modeling depends on both training objective and prediction head.

Another possible misconception is that FiLM in FIRM-DTI is simply another fusion heuristic. The provided formulation indicates a more specific role: ligand embeddings are modulated by learned linear functions of protein embeddings, producing target-specific scaling and shifting of molecular features rather than symmetric concatenation (Refahi et al., 25 Sep 2025). This suggests a directional conditioning structure in which the protein context transforms the drug representation before distance-based comparison.

In methodological terms, the reported results underscore the value of conditioning and metric learning for robust drug-target affinity prediction (Refahi et al., 25 Sep 2025). The paper’s own framing ties this value particularly to out-of-domain evaluation and temporal generalization. Given the temporal split and the contrast with models that use external knowledge graphs or much larger parameter counts, FIRM-DTI can be situated as a compact architecture whose principal claim is that structured latent geometry can improve robustness under realistic distributional shift.