Predicting how a mutation changes the binding strength between two proteins — the quantity $\Delta\Delta G$, measured in kcal/mol — is a workhorse of antibody engineering and protein design. The best current models are deep geometric networks that read the 3D structure around the mutated residue, and they are accurate but expensive: screening all combinations of mutations across a few sites means hundreds of thousands of forward passes, each one a stream of floating-point multiplications on a GPU. In parallel, a neuromorphic line of work has shown that replacing continuous neuron activations with binary pulses ("spikes") lets a network replace power-hungry multiply-accumulate operations with simple additions, cutting energy by one to two orders of magnitude on images, point clouds, and text. What has never been shown is whether spiking networks can do continuous scalar regression — every spiking result in the literature is classification, and discrete spike counts naturally quantize away the fine differences that separate a good mutation from a useless one.
This proposal builds Spike-DDG, a spiking graph network for $\Delta\Delta G$ regression, and tests one central claim: that the accuracy–energy frontier for binding-affinity screening can be moved by combining event-driven message passing on residue graphs with an analog membrane-potential readout — reading the continuous, sub-threshold internal state of the final spiking layer instead of counting its spikes. The application is combinatorial mutational screening at a scale (hundreds of thousands of multi-point variants) where per-inference energy, not wall-clock time, is the binding constraint, for instance on edge hardware or neuromorphic chips.
Core Claim
Spiking networks fail at regression when the output must be reconstructed from spike counts, because a layer firing over $T$ time steps can only express $T+1$ output levels per neuron. But the membrane potential — the analog accumulator that a spiking neuron integrates before deciding whether to fire — is itself a continuous variable. If the readout consumes the membrane potential rather than the spike train, the bulk of the network still runs on additions (spike inputs make each synaptic operation a pure accumulate, at $0.9$ pJ versus $4.6$ pJ per floating-point multiply-accumulate in 45nm CMOS, per SpikingBERT), while only the single final layer pays for real multiplication. Our claim is concrete: a residue-graph spiking network with this readout, distilled from a strong teacher, can match the accuracy of the best published ANN predictors on SKEMPI v2.0 while using at least an order of magnitude less arithmetic energy per inference, making million-variant screens feasible at millijoule-per-variant budgets.
This is not the same as "make $\Delta\Delta G$ prediction fast." Light-DDG already did that: it distilled Prompt-DDG into a small floating-point transformer, got a 15.45% performance gain over its teacher on SKEMPI v2.0 and an 89.7× inference speedup. Light-DDG is fast on a GPU, but every one of its operations is still a multiplication; our target is the arithmetic energy itself and compatibility with event-driven hardware, which no $\Delta\Delta G$ work has touched.
Research Program
Three components, each reusing a technique proven elsewhere and each addressing a known failure mode.
Architecture. The model is a message-passing network over the $k$-nearest-neighbor residue graph of the mutated interface, following DDAffinity, which captures spatial neighbors, sequential neighbors, and residue burial (how deeply a residue is packed, measured by nearby C$\beta$ atom counts) and enforces the physical anti-symmetry $\Delta\Delta G_{\text{mut}\to\text{wt}} = -\Delta\Delta G_{\text{wt}\to\text{mut}}$. We convert its message-passing layers to spiking form: hidden units become leaky integrate-and-fire neurons, and messages are carried by spikes, so synaptic operations are accumulations rather than multiplications. Node inputs come from the hierarchical microenvironment prompt codes of Prompt-DDG, which summarize the local structural context at three scales (residue identity, inter-residue angles, local conformation); this imports the structural prior that Prompt-DDG showed dominates mutational effects, without requiring us to run a large protein language model at inference time.
The analog readout. The final layer's neurons never spike. They integrate their inputs over the $T$ inference steps, and the prediction is a single linear map of the time-averaged membrane potential:
$$\hat{y} = \mathbf{w}^{\top} \bar{U}, \qquad \bar{U} = \tfrac{1}{T}\textstyle\sum_{t=1}^{T} U_t,$$
where $U_t$ is the membrane potential vector of the readout layer at step $t$ and $\mathbf{w}$ is a learned weight vector. Because this is one narrow layer over one graph, its multiply-accumulate cost is a negligible fraction of total operations; everything upstream runs on accumulations. This is the proposal's key ingredient, and we will ablate it directly against the standard spike-count readout.
Training. SKEMPI v2.0 has only about 7,000–7,800 measured mutations — far too few to directly train a spiking network, whose surrogate gradients are notoriously unstable in the small-data regime. We therefore train by distillation, following two precedents. First, SKEMPI-Aug-style synthetic labeling: the Prompt-DDG teacher labels millions of uncharacterized mutations on PDB structures, giving the student a large supervision corpus. Second, the "trained-less but learning-more" regime of Spiking PointNet: train at a single time step $T=1$ (avoiding the memory blowup of backpropagation through time), inject membrane-potential perturbation during training so the network stays robust when inference runs at $T>1$. We additionally use threshold-dependent batch normalization from STBP-tdBN to keep firing rates stable across depth. This training recipe is borrowed, not new; what is new is whether it transfers to molecular graphs and to a regression head.
The closest neuromorphic works are all classification systems and we build on each while departing from it. SpikingBERT distilled a language model into a spiking transformer via steady-state average spiking rate — a mechanism that presupposes a classification head and cannot express a continuous kcal/mol target; we keep its distillation philosophy but replace the rate-matching head with the analog readout. Spiking PointNet and 3DSMT process raw 3D point clouds (CAD objects) with no notion of chemical nodes, residue identity, or paired interfaces; Geometry-Aware Spiking GNN brings spiking dynamics to graphs but only flat citation and co-purchase networks, never physical coordinates or energy regression. On the binding side, MPBind uses equivariant graph networks for binding-site prediction with no efficiency mechanism at all, and Binding Oracle speeds up fine-tuning (about 2 minutes on one GPU) while inference still runs full dense forward passes. No paper in the corpus runs spikes on molecules or regressions on spikes; both gaps meet here.
Baselines
The "do nothing" reference point is the zero predictor, which always outputs $\Delta\Delta G = 0$ (no change); its RMSE equals the spread of the test set and its Pearson correlation is zero, so it calibrates how much signal any model must recover. No reported result found on Lacuna for this reference on the specific splits we use, so we compute it ourselves.
The strongest accuracy competitor is DDAffinity, the spatial-sequential message-passing network whose protocol defines the splits we adopt; on its 10-fold cross-validation over the multi-point subset M1707 it reports Pearson $r = 0.693$, Spearman $0.640$, RMSE $1.944$, MAE $1.496$, and AUROC $0.812$. We will re-run its released model on the identical S1131/M1707/M595/M1340/Ssym splits under the same evaluation code, and beat it if we match its Pearson within 0.02 on M595 (the blind mutation-level test) while showing at least 10× lower arithmetic energy per inference under a shared accounting convention. The strongest energy competitor is Light-DDG, which reports a 15.45% gain over Prompt-DDG on SKEMPI v2.0 and 89.7× inference acceleration; no absolute Pearson value for Light-DDG on SKEMPI v2.0 is reported on Lacuna, so we will reproduce it from its code under our splits. Beating Light-DDG means Pareto-dominating it: equal or better Pearson at strictly lower energy per inference when both are measured with the same operation counts and the 0.9 pJ/4.6 pJ per-operation figures from SpikingBERT. We also compare against H3-DDG (Pearson 0.75 on SKEMPI v2, the top reported number in the corpus, but achieved with 3-body hypergraph and 4-body attention that is the most energy-hungry design in the literature) as an accuracy upper bound rather than an efficiency rival.
The key ablation is the ANN twin: the exact same residue-graph architecture, prompt codes, distillation data, and tuning budget, but with standard continuous activations (ReLU) and a conventional regression head, run on a GPU. It isolates what spiking itself buys or costs in accuracy. A second ablation swaps the analog membrane readout for a spike-count readout with identical everything else, isolating the proposal's central ingredient. A third is an INT8-quantized version of the ANN twin, included because it is a fair cheaper-hardware competitor, not a strawman. All models receive the same hyperparameter search budget (the same number of random-search trials on the same validation fold) and the same teacher labels, so differences reflect the mechanisms, not tuning effort.
Why not just…?
Why not just distill into a small ANN like Light-DDG? This is the most serious objection, and we treat it as such. Light-DDG already dominates on GPU wall-clock and accuracy-per-parameter, and if the deployment target is a data-center GPU, our proposal loses. Our claim is narrower and different: the arithmetic energy floor of a floating-point network is set by its multiplier circuits (4.6 pJ per MAC versus 0.9 pJ per accumulation in 45nm CMOS, per SpikingBERT), and dynamic sparsity means silent neurons cost nearly nothing — advantages no ANN distillation can access. The BindingGYM screening experiment, which measures total energy for a fixed half-million-variant screen under one accounting convention for all models, is designed to settle exactly this: if the Light-DDG-style ANN twin matches our accuracy at equal or lower measured energy, the proposal fails and we will report it as such.
Why not just quantize the ANN to INT8 or use binary graph networks? Quantization shrinks each operation but keeps evaluation dense: every neuron still computes an activation for every input, so dynamic power is paid regardless of whether the input changed. Full-Precision Free Binary GNNs show binary-weight GNNs achieving large training speedups on link prediction, but they are non-spiking and have never been applied to molecular interfaces; no reported result found on Lacuna for any quantized $\Delta\Delta G$ predictor. Because INT8 is nevertheless a credible cheap baseline, it is included head-to-head in the experiments rather than argued away. Evidence that event-driven sparsity buys something multipliers cannot: Geometry-Aware Spiking GNN reduced graph inference energy to 0.051 mJ versus 18.444 mJ for a standard GCN on the same task, and 3DSMT consumed 4.3 mJ where comparable efficient ANNs consumed 78.7–207.0 mJ.
Why not just use fast sequence-only predictors and skip the 3D graph? Sequence-only models (MVSF-AB and its predecessors) are cheap and need no structure, but the corpus is unambiguous that geometry carries the signal on multi-point mutations: DDAffinity needs spatial neighbor graphs and residue burial to reach $r = 0.693$ on multi-point mutations, and H3-DDG shows that higher-order 3D interactions beat pairwise and flat models by large margins on multi-point and combinatorial benchmarks. Dropping 3D structure would trade away exactly the regime (epistatic multi-point effects) where screening matters most. Our design keeps the geometry and makes it cheap, instead of discarding it.
Why not accept that spikes ruin regression and multi-step latency? The latency half is answered by prior work: Spiking PointNet trains at a single time step and 3DSMT achieves 142 ms latency, faster than its non-spiking Mamba3D baseline at 256 ms. The regression half is genuinely open — no paper in the corpus demonstrates continuous physical-property regression with spiking dynamics — and we do not pretend otherwise. It is precisely why the analog membrane readout is the core claim rather than an implementation detail, and why the spike-count-versus-membrane ablation is the single most decisive experiment in the program: if the membrane readout does not substantially beat spike counting on RMSE, the proposal's mechanism is wrong and the negative result should be published.
Proposed Experiments
The primary evaluation follows the DDAffinity protocol on SKEMPI v2.0 (about 7,085–7,805 mutations across 345 complexes): the single-point set S1131, the multi-point set M1707, the blind mutation-level tests M1340 and M595, and Ssym for the anti-symmetry check, reporting Pearson, Spearman, RMSE, MAE, and AUROC alongside DDAffinity's published values ($r = 0.693$, RMSE $1.944$ on M1707). Because Binding Oracle showed that random splits leak information between near-identical complexes, we add a leakage-safe secondary evaluation on its 30%-sequence-similarity splits (B1816 train / T1889 test) for every model, including the re-run baselines.
The energy claim is tested on BindingGYM, the large-scale benchmark of over 500,000 entries dominated by multi-point mutations, where H3-DDG beat Prompt-DDG by 62.61%. We run the full screening workload through Spike-DDG, the ANN twin, the INT8 twin, and reproduced Light-DDG, and report total arithmetic energy under a single convention: per-operation energies of 0.9 pJ per accumulation and 4.6 pJ per multiply-accumulate, times measured operation counts, times measured spike density for the spiking model. We also report the scaling of energy with number of screened variants, which should be sublinear for the spiking model when shared wild-type context keeps most neurons silent across variants of the same complex.
The applied test is antibody-style screening on SARS-CoV-2 RBD data: 285 single mutations on PDB 6M0J and 494 on 7FAE, as used by DDAffinity. The target is ranking quality against the reported reference points — H3-DDG's average favorable-mutation rank of 9.67% (top 10% of candidates) and Light-DDG's best average ranking on the same case studies. A final robustness check repeats the SKEMPI evaluation with AlphaFold3-predicted structures instead of crystal structures, since H3-DDG showed only a 5.4% drop under that substitution and any energy advantage is moot if the spiking model is uniquely structure-sensitive.
The decisive checks, in order of importance, are: the readout ablation (membrane versus spike count, expected to show a large RMSE gap on S1131 and M595); the ANN-twin comparison (same architecture, same training, activations swapped, establishing the accuracy cost — if any — of spiking); and the BindingGYM energy audit (establishing the benefit). If the first two hold and the third shows the projected order-of-magnitude energy reduction at matched accuracy, the claim stands; if the ANN twin at INT8 matches both accuracy and energy, the honest conclusion is that neuromorphic dynamics are unnecessary for this task, and that too would be a finding worth publishing, since nothing in the literature currently answers it either way.
References
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer
- DDAffinity: predicting the changes in binding affinity of multiple point mutations using protein 3D structure
- Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning
- Spiking PointNet: Spiking Neural Networks for Point Clouds
- Going Deeper With Directly-Trained Larger Spiking Neural Networks
- 3DSMT: A Hybrid Spiking Mamba-Transformer for Point Cloud Analysis
- Geometry-Aware Spiking Graph Neural Network
- MPBind: a multitask protein binding site predictor using protein language models and equivariant GNNs
- Binding Oracle: Fine-Tuning From Stability to Binding Free Energy
- Accurately Predicting Protein Mutational Effects via a Hierarchical Many-Body Attention Network
- Full-Precision Free Binary Graph Neural Networks
- MVSF-AB: accurate antibody-antigen binding affinity prediction via multi-view sequence feature learning