---
title: 'DrugFormDTA: Sequence-Only DTA Predictor'
url: https://www.emergentmind.com/topics/drugformdta
type: topic
---

# DrugFormDTA: Sequence-Only DTA Predictor

DrugFormDTA is a sequence-only machine learning model for drug–target binding affinity prediction that represents ligands as SMILES strings and protein targets as full amino acid sequences, encodes them with pretrained Chemformer and ESM-2 backbones, learns a joint compound–target representation, and predicts affinity through separate regression heads for Ki/Kd and IC50 [2605.04265]. In the antiviral setting, it has been studied as a fast, general-purpose affinity regressor whose inference uses log-scaled affinity values such as $ \mathrm{p}K_d = -\log_{10}(K_d) $ and $ \mathrm{p}K_i = -\log_{10}(K_i) $, and whose practical performance depends strongly on curation choices, especially viral polyprotein splitting, endpoint normalization, and task-specific fine-tuning [2605.04265].

## 1. Core problem formulation

DrugFormDTA addresses drug–target affinity (DTA) prediction as a continuous regression problem. The model takes a ligand/drug as a SMILES sequence and a protein/target as a full amino acid sequence, then predicts an affinity value rather than a binary interaction label. In the antiviral benchmark, inference is reported in log-scaled units, with $ \mathrm{p}K_d = -\log_{10}(K_d) $ and $ \mathrm{p}K_i = -\log_{10}(K_i) $; affinities were consistently transformed to pKi/pKd for stability and interpretability [2605.04265].

A defining feature is its multi-endpoint design. DrugFormDTA predicts affinity with two separate regression heads: one head for Ki/Kd and one head for IC50. During original training, if one endpoint is missing for a pair, the model leverages the other head’s prediction rather than collapsing IC50 into Ki/Kd via a crude approximation. This is a materially different design choice from pipelines that enforce a single homogeneous label space at preprocessing time, and it is particularly relevant in virology, where Ki, Kd, and IC50 measurements often coexist with uneven coverage across targets and assays [2605.04265].

The sequence-only formulation places DrugFormDTA within a broader DTA tradition that prioritizes standardized, widely available inputs. ResDTA, for example, also formulates DTA as direct regression from 1D representations of drugs and proteins and argues that SMILES and amino acid sequences are readily available and standardized, enabling end-to-end learning without additional chemical domain knowledge [2303.11434]. This suggests that DrugFormDTA belongs to the same family of models that trade explicit structural input for coverage, speed, and deployment simplicity.

## 2. Architecture and training design

In the antiviral benchmarking study, DrugFormDTA is described as a sequence-only model with a pretrained Chemformer encoder for molecules and the ESM-2 150M-parameter transformer for proteins. Chemformer is described there as a bidirectional autoregressive transformer trained on more than 100M molecules, while ESM-2 150M is described as trained on approximately 27M protein sequences and as implicitly learning structural features such as contact and conservation [2605.04265]. The model learns a joint representation of compound and target and then applies the two regression heads noted above.

The fine-tuning experiments reported for antivirals used two regimes. The first used mean squared error (MSE) loss with two staged learning rates. The second used Huber loss with a reduce-on-plateau learning-rate schedule. DrugFormDTA trains four sub-models and ensembles them at inference; the antiviral study replicated that ensembling and also compared it with a single fine-tuned model [2605.04265]. Batch sizes, optimizer type, and epoch counts are not explicitly reported for those fine-tuning runs, and no augmentation or negative sampling was used.

The model’s operational profile is notable. On the 141-pair small antiviral test, inference throughput was reported as approximately $0.8 \,\mathrm{s/mol}$ on an RTX 3060 and approximately $0.02 \,\mathrm{s/mol}$ on an RTX 5090. On the 853-compound benchmark, inference was approximately $0.03 \,\mathrm{s/mol}$ on GPU. In the same study, docking-based tools were orders of magnitude slower, with GNINA at $e=32$ reported at approximately $120 \,\mathrm{s/mol}$ [2605.04265]. This places DrugFormDTA in a regime where large libraries can be screened quickly without requiring explicit 3D docking for every candidate.

The architectural minimalism of this setup should not be confused with methodological triviality. Contemporary DTA literature shows several routes by which such a backbone can be modified. FIRM-DTI introduces protein-conditioned FiLM modulation of drug embeddings, cosine-distance geometry, triplet loss with margin $ \alpha = 0.9 $, and an RBF regression head; LaPro-DTA introduces latent dual-view drug representations and salient protein top-$k$ pooling; and $k$NN-DTA adds non-parametric retrieval at inference time [2509.20693] [2603.14792] [2407.15202]. These are not intrinsic components of DrugFormDTA as benchmarked for antivirals, but they define the methodological neighborhood in which DrugFormDTA has been discussed and adapted.

## 3. Antiviral adaptation and data curation

The most consequential adaptation reported for DrugFormDTA is virology-aware dataset curation. A dedicated antiviral dataset of 43,005 high-confidence viral protein–ligand binding measurements was assembled from BindingDB, SMACC, and Heli-SMACC, spanning more than 10 human virus species across multiple families, including coronaviruses, flaviviruses, HCV, HIV, and influenza [2605.04265]. Endpoints included Kd, Ki, and IC50, with a small number of Kd/Ki mixtures.

A central issue was viral polyproteins. The study reports that approximately 31.1% of viral targets in BindingDB are listed as polyproteins and that for SARS-CoV-2 the figure is 99.3%. DrugFormDTA’s original training pipeline, as described there, filtered BindingDB to mammalian and human-virus proteins, weighted human higher by $\sqrt{10}$, and removed protein targets longer than 2,500 amino acids, thereby dropping 91% of viral polyprotein types and 31% of virus rows. In the antiviral adaptation, these polyproteins were instead split into mature proteins and mapped to their true enzymatic entities, for example mapping “Replicase polyprotein 1ab [3264–3569]” to SARS-CoV-2 main protease [2605.04265].

The curation pipeline proceeded as follows: 65,950 viral BindingDB rows were reduced to 56,881 after removing threshold-only values such as “>1000 nM”; to 51,127 after removing polyproteins that could not be confidently mapped; to 50,413 after removing proteins missing a reference sequence; and to 42,379 after aggregating duplicates at the protein–ligand level. Heli-SMACC contributed 409 entries and SMACC contributed 238 entries, giving 43,026, after which removal of invalid SMILES yielded the final 43,005-entry training set [2605.04265].

This curation was not merely clerical. The study explicitly attributes major gains to correcting target identity, normalizing affinities to pKi/pKd, and removing threshold-only values rather than shifting them toward the reported threshold. A plausible implication is that, for DrugFormDTA, data semantics and target definition can matter as much as architectural change when the application domain contains systematic annotation artifacts, as virology does with long translated polyproteins.

## 4. Evaluation and empirical performance

DrugFormDTA was evaluated in two antiviral regimes. The first was a small, strict Ki/Kd-only test set of 141 pairs spanning six viral proteins: Dengue, HCV NS3/NS4A, MERS PLpro, SARS-CoV-2 Mpro, West Nile NS2B/NS3, and Zika NS2B/NS3. The second was a broader benchmark comprising 853 compounds across 16 crystal structures, 8 target types, and 10 virus species, including coronavirus Mpro and PLpro targets, HCV NS3/NS4A and NS5B, HIV protease and reverse transcriptase, flaviviral NS2B/NS3 proteases, and influenza A neuraminidase [2605.04265].

On the 141-pair strict set, the base model achieved $r = 0.332$, $\rho = 0.313$, $\mathrm{RMSE} = 1.229$, $\mathrm{MAE} = 1.031$, and $Q^2 = 0.080$. Fine-tuning substantially improved performance: the ensemble fine-tune under run 2 achieved $r = 0.660$, $\rho = 0.692$, $\mathrm{RMSE} = 1.072$, $\mathrm{MAE} = 0.889$, and $Q^2 = 0.290$, while a single-model fine-tune achieved $r = 0.677$, $\mathrm{RMSE} = 0.980$, $\mathrm{MAE} = 0.810$, and $Q^2 = 0.407$ [2605.04265].

On the expanded 853-compound benchmark, the fine-tuned model ranked highest among the ML-based methods reported there, with $r = 0.701$, $Q^2 = 0.492$, $\rho = 0.698$, $\mathrm{RMSE} = 1.19$, $\mathrm{AUROC} = 0.845$, $\mathrm{BEDROC} = 0.962$, fail\% $= 0.0$, and approximately $0.03 \,\mathrm{s/mol}$ inference time. The base model on the same benchmark had $r = 0.498$, $Q^2 = 0.248$, $\rho = 0.472$, $\mathrm{RMSE} = 1.45$, $\mathrm{AUROC} = 0.725$, and $\mathrm{BEDROC} = 0.899$ [2605.04265].

| Setting | Base DrugFormDTA | Fine-tuned DrugFormDTA |
|---|---|---|
| 141-pair strict test | $r=0.332$; RMSE $=1.229$; $Q^2=0.080$ | Ensemble: $r=0.660$; RMSE $=1.072$; $Q^2=0.290$ |
| 853-compound benchmark | $r=0.498$; RMSE $=1.45$; AUROC $=0.725$ | $r=0.701$; RMSE $=1.19$; AUROC $=0.845$ |

The same benchmark found that Boltz-2 and DrugFormDTA ranked highest overall among ML-based approaches, with GNINA strongest among docking approaches. On the 141-pair strict set, GNINA at $e=32$ had $r = 0.279$ and $\mathrm{RMSE} = 1.377$, while Boltz-2 had $r = 0.504$ and $\mathrm{RMSE} = 1.416$. On the expanded benchmark, Boltz-2 had $r = 0.316$, $\mathrm{RMSE} = 1.59$, and $\mathrm{AUROC} = 0.634$; GNINA (CNN) had $r = 0.302$, $\mathrm{RMSE} = 1.60$, and $\mathrm{AUROC} = 0.652$; FlowDock had $r = 0.295$, $\mathrm{RMSE} = 1.60$, and $\mathrm{AUROC} = 0.640$ [2605.04265].

Performance was nevertheless target-dependent. Boltz-2 was reported as strong on HIV reverse transcriptase with $r \approx 0.68$ but weak on SARS-CoV-2 Mpro with $r \approx 0.15$, while docking performance also varied substantially by target. Fine-tuned DrugFormDTA was described as the most consistent across targets, though it still exhibited protein-dependent variance in $r$ [2605.04265].

## 5. Relationship to adjacent DTA architectures

DrugFormDTA sits within a heterogeneous DTA ecosystem spanning sequence CNNs, geometry-aware metric-learning systems, cold-start-oriented latent-variable architectures, and retrieval-augmented inference. ResDTA exemplifies the sequence-CNN branch: it encodes SMILES and protein sequences with two residual 1D CNN streams plus a third combined stream, concatenates 256-dimensional drug and protein vectors with a 512-dimensional combined vector, and regresses KIBA scores through a five-layer fully connected head with dropout $p=0.1$ [2303.11434]. On KIBA, ResDTA reported $ \mathrm{CI} = 0.885 \pm 0.001 $ with residual connections, compared with $0.877 \pm 0.001$ without them. This establishes a sequence-only baseline tradition in which improvements come from better feature propagation and cross-modality fusion rather than explicit geometry.

A distinct line is represented by FIRM-DTI, which provides a geometry-aware blueprint for incorporation into a DrugFormDTA-like system. FIRM-DTI conditions the ligand embedding on the protein embedding through FiLM, normalizes both embeddings, uses cosine distance, and structures the latent space with a triplet loss; affinity is then predicted by an RBF regression head with $ \sigma = 0.2 $ and centers evenly spaced on $[0,2]$ [2509.20693]. On the Therapeutics Data Commons DTI-DG temporal benchmark, the full model achieved $ \mathrm{PCC} = 0.59 $; removing FiLM reduced PCC to $0.55$, whereas removing triplet loss reduced it to $0.32$. The learned distance-to-affinity curve at $ \sigma = 0.2 $ had correlation $r = 0.91$ between distance and predicted affinity. In relation to DrugFormDTA, this suggests a route by which a sequence-only backbone could acquire explicit geometric regularization and more interpretable distance-based regression.

LaPro-DTA addresses a different failure mode: cold-start degradation for unseen drugs, targets, and pairs. Its latent dual-view drug representation combines an instance-level stochastic encoding strategy with a distribution-level semantic-remapping path, while the protein encoder uses pattern-aware top-$k$ pooling with $K=4$ to isolate salient residues before cross-view multi-head attention [2603.14792]. On cold-start Davis, it reported an 8.3% MSE reduction in the unseen-drug setting, a 6.3% reduction in unseen-target, and an 11.3% reduction plus a 3.9% CI improvement in unseen-pair relative to the strongest baselines described there. As a point of comparison for DrugFormDTA, LaPro-DTA emphasizes that sequence-only models can still be redesigned for cold-start robustness if they explicitly regularize drug representations and avoid diluting binding-pocket signals with global protein pooling.

A fourth methodological axis is inference-time retrieval. $k$NN-DTA and Ada-$k$NN-DTA operate on top of a frozen pretrained DTA backbone by building pair-wise and point-wise FAISS datastores from training embeddings, retrieving nearest neighbors at test time, and aggregating labels or representations with a fusion parameter $ \alpha $ [2407.15202]. On BindingDB IC$_{50}$, $k$NN-DTA improved RMSE from $0.717$ to $0.684$, and Ada-$k$NN-DTA further improved it to $0.675$; on BindingDB Ki, the corresponding sequence was $0.785 \rightarrow 0.750 \rightarrow 0.735$. Because this procedure is explicitly described as model-agnostic and plug-and-play, it is directly relevant to DrugFormDTA as a post hoc route to accuracy gains without full architectural retraining.

## 6. Limitations, misconceptions, and prospective directions

DrugFormDTA’s antiviral performance depends on a set of assumptions that are easy to understate. First, it is sequence-only. The antiviral benchmark explicitly notes that structure-aware tools can be competitive on some targets, and related DTA papers repeatedly identify the lack of explicit 3D information as a limitation. FIRM-DTI uses sequence-only protein features and no 3D structural information, while LaPro-DTA likewise notes that future work could integrate geometric graphs or 3D conformers to improve spatial reasoning [2509.20693] [2603.14792]. For DrugFormDTA, the absence of explicit structural input is therefore better understood as a design trade-off than as a universal advantage.

Second, assay heterogeneity constrains ceiling performance. The antiviral benchmarking study notes that assay noise and heterogeneity limit performance, and that mixing IC50 with Ki/Kd introduces additional variance. It recommends normalizing to pKi/pKd, avoiding crude IC50 $\approx$ Ki assumptions, aggregating duplicates only after unit harmonization and log transformation, and considering assay metadata such as cell type, temperature, and substrate concentrations [2605.04265]. A common misconception is that a sufficiently strong encoder can erase such measurement inconsistencies; the curation results argue the opposite.

Third, out-of-distribution generalization remains unresolved. The same antiviral study states that OOD generalization remains a challenge for ML models, while FIRM-DTI argues that naive concatenation lacks metric structure and therefore generalizes poorly across chemical space and time [2605.04265] [2509.20693]. LaPro-DTA reaches a similar conclusion from the perspective of cold-start settings, attributing degradation to overfitting and protein-sequence noise [2603.14792]. These accounts converge on the view that improved generalization requires explicit inductive bias—metric learning, dual-view regularization, protein saliency selection, or retrieval—rather than only larger pretrained encoders.

The future directions suggested across this literature are comparatively concrete. For DrugFormDTA in antiviral discovery, proposed directions include better assay metadata, more Ki/Kd measurements for viral enzymes, broader coverage of underrepresented targets, multi-task learning, structure-aware hybrids where reliable holo structures exist, deeper integration with docking, and target-aware calibration or tool selection [2605.04265]. Related work adds more specific modular possibilities: FiLM-conditioned geometry with triplet loss and RBF distance regression from FIRM-DTI, latent dual-view drug encoding and top-$k$ protein pooling from LaPro-DTA, and inference-time retrieval layers from $k$NN-DTA [2509.20693] [2603.14792] [2407.15202]. A plausible implication is that DrugFormDTA’s long-term significance may lie less in a fixed architecture than in its role as a high-throughput sequence backbone that can absorb increasingly specialized regularization, fusion, and retrieval mechanisms while preserving the speed advantages demonstrated in antiviral benchmarking.

Source: https://www.emergentmind.com/topics/drugformdta