A Recurrent Trajectory View for Multi-Task Drug Prediction Under Label Scarcity

Deep models for drug property prediction are trained on a few hundred labeled molecules where a single learned representation must serve many endpoints. Today's strongest scarce-label methods squeeze extra signal from two sources: richer inputs (fusing a molecule's text string, its graph, and sometimes a 3D shape) and richer objectives (training several related prediction tasks jointly, occasionally with physical laws as extra loss terms). But every one of these methods treats a molecule as a static object. Real drugs are not static: they flex, bind, and unbind over time, and the timescales of those motions — not the equilibrium shape — determine how long a drug stays on its target. This proposal argues that the missing ingredient is a dynamic molecular view: a short, cheaply simulated motion trajectory of each molecule, read by a recurrent neural network, and fused with the standard static views inside a multi-task model trained under physical constraints.

The core claim is specific: adding a recurrent encoder over on-demand simulated trajectories as a third view, and training it jointly across pharmacokinetic and binding-kinetic endpoints with known physical relations, will beat the strongest existing scarce-label baselines — because the dynamic view, unlike experimental 3D structures, is available at inference time for any molecule, and because the recurrent encoder is differentiable and linear-time, letting it plug into joint training where exact sequence-alignment methods cannot.

Core Claim

Three documented facts motivate the design. First, MolMVC shows that a 3D geometric view improves molecular representations, but is forced to drop the 3D encoder entirely at fine-tuning time because experimental structures do not exist for most candidate molecules. A simulated trajectory does not have this problem: we can generate a short molecular-dynamics run for any molecule we want to score, so the dynamic view survives from pre-training through inference. Second, PEMAL shows that enforcing pharmacokinetic balance laws across multi-task heads lets a model trained on 20% of the data beat a single-task model trained on all of it — evidence that joint, physics-constrained training is the right response to label scarcity, but it does so from a single 2D graph view. Third, Ray et al. show that unbinding trajectories carry the information that determines residence time, but extract it with a non-parametric, quadratic-time alignment method that cannot scale or sit inside a differentiable pipeline; the authors themselves call for a learned approximation. The proposal takes each paper's mechanism and supplies the missing piece the others lack.

What Exists and What Is Missing

Two lines of work fuse multiple molecular views to fight label scarcity. MolMVC contrasts a 1D substructure sequence (Transformer), a 2D graph (a Graph Isomorphism Network, which passes messages along chemical bonds), and a static 3D conformer; PMMR fuses a SMILES string and a 2D graph with attention for drug–target binding under cold-start splits. Both are strictly static, and both train downstream tasks one at a time. Two further lines attack scarcity through multi-task learning: PEMAL couples four pharmacokinetic endpoints with physical equations, and ADME-DL orders 21 ADMET tasks along the body's absorption-to-excretion sequence using gradient surgery. Neither uses more than one input view, and neither contains any recurrent state. Finally, Ray et al. model molecular dynamics directly with dependent Dynamic Time Warping, achieving impressive kinetic accuracy on a single protein–ligand system, but with no neural representation, no transfer across systems, and no multi-task coupling. No existing work combines a dynamic view with multi-view fusion and multi-task training; that triple intersection is the gap this proposal occupies. Hepa-RAFT is closest in spirit to "generate the missing data" — it synthesizes virtual transcriptomic profiles as an extra channel — but its channels are static snapshots and its transfer is retrieval-based rather than gradient-based, which is exactly the design choice we test.

Research Program

We build a model with three parallel encoders. The first two are standard: a pre-trained SMILES encoder and a 2D graph network. The third is new: for each molecule we run a short, cheap simulation (solvent molecular dynamics for pharmacokinetic tasks; protein–ligand contact dynamics for binding-kinetic tasks, following the coordination-number descriptors of Ray et al.), reduce each frame to a fixed-length descriptor vector, and encode the resulting time series with a Gated Recurrent Unit — a network that carries a hidden state forward through the frames, learning which motions matter. The three embeddings are fused with the query-key-value attention used by PMMR, and the fused representation feeds task heads for clearance, half-life, volume of distribution, AUC, and (where labels exist) binding off-rate and residence time $t_R = 1/k_{off}$. Training proceeds in two stages: self-supervised pre-training of the recurrent encoder on unlabeled simulated trajectories (masked-frame reconstruction), then joint fine-tuning on the scarce labeled endpoints with the pharmacokinetic balance laws from PEMAL as auxiliary losses,

$$\mathcal{L} = \sum_{\text{tasks}} \mathcal{L}_{\text{task}} \;+\; \lambda_1\,|\widehat{\text{AUC}}\cdot\widehat{\text{CL}} - K_1| \;+\; \lambda_2\,|\widehat{\text{CL}}\cdot\widehat{T_{1/2}} - \widehat{\text{Vdss}}\cdot K_2|,$$

where hatted terms are model predictions and $K_1, K_2$ are estimated constants. A key design question we answer empirically is how long a trajectory is needed: if nanosecond-scale runs suffice, the dynamic view costs pennies per molecule; if not, the idea degrades to "3D structures in disguise" and we will say so.

Baselines

The naive reference point is a per-task predictor of the training-set median on the pharmacokinetic benchmark, which sets the floor any learned model must clear, alongside RDKit 2D descriptors fed to gradient-boosted trees — a baseline the literature takes seriously, because Huang et al. report that expert descriptors sometimes match or beat graph neural networks across TDC tasks. All baselines run on identical scaffold and temporal splits, with the same hyperparameter search budget and comparable compute (we cap our model's tuning budget to the baselines', counting simulation generation as our model's fixed preprocessing cost). The strongest existing competitor on the pharmacokinetic question is PEMAL, which on the 215-molecule PK-Mol benchmark reports reducing mean absolute error by 15.2% for AUC and 19.6% for clearance relative to a GIN baseline; no absolute MAE value is reported on Lacuna for PK-Mol, so we compare in relative terms. Beating PEMAL means lower MAE on at least three of the four endpoints at full data, and holding that margin on the 20%-data sample-efficiency curve where PEMAL's headline result (20% of data beating a GIN on 100%) was established. On the trajectory side, the competitor is the DTW protocol of Ray et al., which classified benzene–lysozyme unbinding pathways at 86–96% accuracy against manual labels and predicted a fastest-path residence time of 1.27 ms against an experimental 1.1 ms; matching that accuracy and staying within roughly 20% of the experimental timescale, at a fraction of the pairwise-comparison wall-clock, counts as success. Finally, the decisive ablation is our own architecture with the recurrent trajectory encoder removed — leaving a static 1D+2D attention-fused model in the mold of PMMR — which isolates whether the dynamic view, not merely the extra parameters or the multi-task loss, drives any gain.

Why not just…?

Why not just use handcrafted descriptors with gradient boosting? This is the most serious threat: the TDC evidence (Huang et al.; Green et al.) says classical descriptors regularly match deep models when data is scarce and splits are realistic. The counter-evidence is that in the extreme scarcity regime the deep literature wins: Hepa-RAFT beat descriptor-based forests on 165 labeled drugs, and PEMAL cut MAE by 15–20% and survived 80% label corruption where standard models failed. We settle this by running the descriptor baseline under exactly our splits and tuning budget; if it wins, the proposal's claim about the dynamic view fails, and the honest conclusion would be that the dynamic view only pays off in the multi-task physics-constrained setting — a finding worth reporting either way.

Why not just train separate single-view models and average their predictions? Late fusion is simpler and avoids negative transfer, and it demonstrably helps when views are redundant, as in multi-camera driving datasets (100-Driver). But averaging happens after each view has already collapsed its representation into a scalar, so the views never learn each other's correlations; PMMR found its isolated single-view variants clearly weaker than the attention-fused model, especially under cold-start shift. We include a late-fusion ablation (separate models per view, averaged outputs) so this is measured rather than asserted.

Why not just fine-tune one pre-trained foundation model on the target task? Single-task fine-tuning of ChemBERTa-2 or a graph model avoids gradient interference between tasks, a real failure mode of multi-task learning. The counter-evidence is direct: PEMAL showed single-task baselines fail on small subsets while its physics-constrained multi-task model trained on 20% of data surpassed them at 100%, and ADME-DL improved foundation-model representations by up to 18.2% over single-task training. We run single-task fine-tuning of each encoder as an ablation, with and without the dynamic view, to separate the benefit of the view from the benefit of joint training.

Why not just keep Dynamic Time Warping? On the single benzene–lysozyme system, DTW may well remain the better tool: it is deterministic, needs no training, and already achieves 86–96% accuracy and a 1.27 ms residence-time estimate (Ray et al.). We concede this locally and claim globally: pairwise DTW scales quadratically in trajectory count and length, which Ray et al. themselves flag as the barrier to high-throughput screening, and it is non-differentiable, so it cannot share a loss with pharmacokinetic heads. Our experiments include a direct wall-clock comparison and a transfer test (train the recurrent encoder on one ligand's trajectories, test on another) that DTW cannot attempt without re-running all pairwise alignments.

Proposed Experiments

Three studies, in order of decisiveness. First, the trajectory-encoder sanity check on the benzene–T4 lysozyme benchmark from Ray et al.: embed trajectories with the recurrent encoder, cluster them, and measure pathway-classification accuracy against their 86–96% DTW result, reproduce a fastest-path residence-time estimate near their 1.27 ms (experimental 1.1 ms), and record wall-clock against pairwise DTW at increasing trajectory counts to demonstrate the scaling gap. Second, the scarce-label test on PK-Mol (215 molecules, four endpoints), evaluated under scaffold splits with MAE, comparing against PEMAL's reported 15.2% (AUC) and 19.6% (clearance) reductions over GIN, plus a data-efficiency curve at 20%, 40%, 60%, 80% of training data. Third, generalization: the 21 TDC ADMET endpoints used by ADME-DL under scaffold and temporal splits, where the literature warns performance collapses under realistic splits (Pearson correlation falling from roughly 0.7 to 0.4 in temporal drug–target splits per TDC), so all headline numbers are reported on out-of-distribution splits, never random ones. Ablations throughout remove, one at a time, the recurrent encoder (static fusion only), the physics constraints, the attention fusion (replaced by averaging), and the multi-task heads (single-task training). The decisive check for the whole proposal is a two-way comparison: our full model versus the same model with the trajectory view deleted, on the 20%-data PK-Mol setting. If the dynamic view adds nothing there, the claim is falsified; if it holds the PEMAL-style sample-efficiency advantage while the static variant does not, the recurrent trajectory view earns its place.