Review of: "Protein Structure Prediction in the 3D HP Model Using Deep Reinforcement Learning"

Abstract

What the work claimsThe authors treat folding in the 3D Hydrophobic-Polar lattice model as a sequential decision problem: the chain is built as a self-avoiding walk on the cubic lattice, one residue per step, and a Deep Q-Network with experience replay and a target network learns where to place the next residue, with the reward tied to the number of non-consecutive H-H contacts ( §2.1, §2.2).Two Q-network designs are proposed.For short chains (N ≤ 36) there is FFNN-R, a xed sparse random reservoir followed by fully connected layers ( §2.3.1).For longer chains there is LSTM-A, an LSTM whose outputs pass through multi-head attention before the Q-value head ( §2.3.2).Two simpler variants, a plain FFNN and an LSTM that uses only the last hidden state (LSTM-OLH), serve as internal baselines ( §3).The abstract claims three things: that FFNN-R reaches optimal conformations "with 25% fewer training episodes"; that LSTM-A matches best-known energy values (BKVs) on longer sequences; and that both architectures show "consistent achievement of optimal conformations." The evidence is Table 3, which compares the nal energies against BKVs from reference [18], a genetic algorithm (GA), BILS and TPPSO, plus learning curves (Figures 5 and 6), attention maps (Figure 7), and a short ef ciency section ( §3.1). Whether the conclusion holds on the dataI want to start with what is genuinely good here.The problem formulation is clean and honest about its simpli cations.The benchmark is a standard one, taken with its reference values from a published source, so readers can place the results.The authors report both successes and failures in Table 3 instead of hiding the hard sequences, they give trainable parameter counts (264,445 for FFNN-R against 6,324,741 for LSTM-A, §3), and they state the hardware and framework versions (one NVIDIA H100, PyTorch 2.5.0,Qeios qeios.