ApexAmphion: RL-Driven Peptide Design

• The paper introduces ApexAmphion, a platform that couples a 6.4B parameter protein language model with multi‐objective reinforcement learning to design potent antimicrobial peptides.
• It integrates supervised fine-tuning on 27,148 curated AMP sequences with proximal policy optimization, enabling generation under compound physicochemical and MIC objectives.
• Experimental validation demonstrated a 100% hit rate where 100 synthesized amphionins showed low MIC values and effective membrane depolarization against multidrug-resistant pathogens.

Searching arXiv for the specified paper and closely related entries to ground the article. arXiv search query: (Cao et al., 16 Sep 2025) ApexAmphion deep reinforcement learning platform for antibiotic discovery ApexAmphion is an integrated, deep‑reinforcement‑learning platform for de novo design of peptide antibiotics. It couples a large protein LLM, ProGen2‑xlarge with 6.4 billion parameters, to a multi‑objective reinforcement learning layer that generates, scores, and optimizes antimicrobial peptides in a single pipeline. The system is organized around supervised specialization on curated AMP sequences, proximal policy optimization against a composite reward combining MIC prediction and differentiable physicochemical objectives, large‑scale virtual library generation, and experimental validation of 100 synthesized peptides termed amphionins (Cao et al., 16 Sep 2025).

1. Research context and problem setting

ApexAmphion is situated in the antibiotic discovery problem defined by antimicrobial resistance, which is projected to cause up to 10 million deaths annually by 2050. The platform is motivated by several obstacles identified for computational AMP discovery: sparse and noisy data, fragmented pipelines that separate scoring from generation, poor performance of large protein LLMs used without task‑specific adaptation, and limited multi‑objective control over developability variables such as charge, hydrophobicity, and length (Cao et al., 16 Sep 2025).

The system is intended to “close the loop” across generation, scoring, and optimization. Its workflow is explicitly described as: using a foundation protein LLM as a prior over protein sequence space; specializing it by supervised fine‑tuning on curated AMP data; optimizing it with reinforcement learning using Proximal Policy Optimization; generating large virtual libraries; and experimentally validating a selected subset. Compared with traditional QSAR, standard generative models, or heuristic optimization, ApexAmphion uses a much larger LLM than typical AMP‑specific models, applies LoRA fine‑tuning to retain broad protein knowledge, and turns “screening after generation” into condition‑based generation through reinforcement learning (Cao et al., 16 Sep 2025).

This positioning matters methodologically. Rather than treating AMP discovery as a post hoc ranking problem over candidates emitted by a separate generator, ApexAmphion treats peptide design as a sequential decision process in which the generation policy is itself altered by potency and property signals. A plausible implication is that the platform is less dependent on narrow candidate libraries than pipelines in which the generator is fixed.

2. Core architecture and AMP specialization

The base model is ProGen2‑xlarge, a GPT‑like autoregressive Transformer with 6.4 billion parameters pretrained on billions of natural protein sequences across diverse families and organisms. ApexAmphion adapts this foundation model with Low‑Rank Adaptation rather than full fine‑tuning. LoRA introduces small trainable low‑rank matrices in each transformer layer, leaving only approximately 2.7 million parameters trainable for AMP specialization while preserving the broader protein prior (Cao et al., 16 Sep 2025).

The supervised fine‑tuning dataset aggregates AMP sequences from DRAMP, DADP, LAMP2, dbAMP v2, and DBAASP. It includes AMPs with antimicrobial, antibacterial, or antifungal activity and or reported MIC, is filtered to sequences shorter than 50 amino acids, deduplicated, and split using MMseqs2 clustering. The final dataset contains 27,148 unique sequences. Fine‑tuning is monitored with validation perplexity, and the resulting model is termed Amphion‑SFT (Cao et al., 16 Sep 2025).

The learned sequence regularities are reported explicitly. Amphion‑SFT enriches hydrophobic and aliphatic residues $A, F, I, L, M, V$, enriches Lys over Arg, depletes acidic and polar residues $D, E, N, Q, S, T$, strongly depletes Gly and Pro, and shows very low Cys. The resulting peptides occupy lengths from 8 to 50 amino acids and exhibit physicochemical profiles associated with membrane‑active peptides, including net charge typically $+2$ to $+9$, $pI > 10$, Eisenberg hydrophobicity approximately $0.3$ to $0.6$, and high hydrophobic moment. The paper characterizes this as a generative prior biased toward amphipathic, K‑rich, helix‑forming sequences (Cao et al., 16 Sep 2025).

Component	Function	Reported detail
ProGen2‑xlarge	Base protein LM	6.4B parameters
LoRA adapters	AMP specialization	$\sim 2.7$M trainable parameters
Amphion‑SFT	SFT model	Trained on 27,148 unique AMP sequences
ApexMIC	MIC predictor	ESM2‑8M embeddings + MLP
Amphion‑RL	RL‑tuned generator	PPO optimization of sequence policy

The two‑stage training schedule is also quantified: supervised fine‑tuning takes approximately 12 hours on the reported hardware configuration, and RL fine‑tuning takes approximately 8 hours. This supports the claim that the platform can be steered within hours (Cao et al., 16 Sep 2025).

3. Reinforcement learning formulation and composite reward

In the RL stage, ApexAmphion treats peptide generation as a token‑level sequential decision problem. A state $s_t$ is a partial peptide sequence, and an action is selection of the next amino acid token from the standard residue vocabulary, with generation terminating at a stop token or at a maximum length of 8–50 amino acids. The policy network is the LoRA‑augmented ProGen2‑xlarge initialized from Amphion‑SFT, and only LoRA parameters remain trainable during RL. A separate ProGen2‑based value function with LoRA predicts expected return (Cao et al., 16 Sep 2025).

Optimization uses standard Proximal Policy Optimization with clipping, entropy regularization, and value loss. Reward is assigned at the sequence level rather than per step and is normalized by scaling to $[0,1]$ followed by whitening across the batch before advantage estimation and value training. The design intent stated in the paper is to prevent collapse through LoRA‑only updates, entropy regularization, and a multi‑objective reward rather than a single scalar MIC objective (Cao et al., 16 Sep 2025).

The total reward is defined as

$D, E, N, Q, S, T$0

The MIC component is provided by ApexMIC, which uses ESM2‑8M to embed peptides and an MLP to output a probability that a sequence is “low MIC active.” ApexMIC is trained with focal loss on a curated dataset of 38,623 sequences, split into 30,914 train, 3,853 validation, and 3,856 test examples. The positive class consists of experimentally validated AMPs with MIC $D, E, N, Q, S, T$1 and lengths 12–50 amino acids. Negatives include the Veltri negative dataset, additional negative or uncertain sequences from UniRef, and AMPs classified as high‑MIC by the HydrAMP MIC classifier with threshold 0.0121 (Cao et al., 16 Sep 2025).

The property component includes five descriptors: global hydrophobicity, hydrophobic moment, net charge, isoelectric point, and length. The reward clamps these descriptors to desired intervals: hydrophobicity to $D, E, N, Q, S, T$2, hydrophobic moment to $D, E, N, Q, S, T$3, net charge to $D, E, N, Q, S, T$4, and isoelectric point to $D, E, N, Q, S, T$5, with length managed by a constant term and hard constraints. The stated objective is to encourage moderate hydrophobicity, high hydrophobic moment, net positive charge without extreme cationicity, high $D, E, N, Q, S, T$6, and reasonable lengths. The paper summarizes the consequence as a bias toward membrane‑active, selective AMP‑like peptides (Cao et al., 16 Sep 2025).

4. Library generation, annotation, and candidate curation

ApexAmphion produces two large virtual libraries. Amphorium contains 2.1 million non‑redundant peptide sequences generated from Amphion‑SFT. Amphorium‑RL contains 180,000 sequences generated from Amphion‑RL and enriched for low predicted MIC. Both libraries are annotated with ApexMIC scores, APEX 1.1 pathogen‑specific MIC predictions, HydrAMP, AMPScanner v2, and physicochemical descriptors (Cao et al., 16 Sep 2025).

Selection of synthesis candidates proceeds through a multi‑stage filter. First, sequences are retained only if their ApexMIC score is at least 0.4, and sequences predicted to be extremely insoluble or hard to synthesize are rejected. Second, structural and physicochemical filters are applied: length must remain at or below 50 amino acids; AlphaFold2 or ESMFold is used to identify peptides likely to be linear or simple helices; complex tertiary structures and disulfide‑rich peptides are deprioritized; high homology to human or other known proteins is removed using MMseqs2 against UniRef50, with discards at greater than 70% identity over greater than 70% of sequence length; and sequences with high aggregation risk or problematic motifs are excluded. Third, final prioritization emphasizes high ApexMIC scores, good HydrAMP and AMPScanner predictions, favorable property ranges, low predicted hemolysis, novelty, and diversity (Cao et al., 16 Sep 2025).

The final set of 100 synthesized peptides is named the amphionins. This selection protocol is central to the platform’s experimental interface: the generative model is not used in isolation, but as the front end of a structured generation–annotation–filtering workflow. A plausible implication is that ApexAmphion’s reported hit rate depends on both policy optimization and stringent downstream triage rather than on unconstrained sampling alone.

5. Experimental performance and mechanism of action

The amphionins were evaluated by broth microdilution in 96‑well plates against a bacterial panel of 16 or more strains spanning Gram‑negative and Gram‑positive pathogens, including multidrug‑resistant Acinetobacter baumannii, polymyxin‑resistant Escherichia coli, carbapenem‑resistant Klebsiella pneumoniae, Pseudomonas aeruginosa, MRSA, and vancomycin‑resistant Enterococcus strains. Peptide concentrations ranged from 0.78 to $D, E, N, Q, S, T$7 after inoculation in two‑fold dilutions, inocula were approximately $D, E, N, Q, S, T$8 CFU/mL, incubation was at $D, E, N, Q, S, T$9 for 18–20 hours, MIC was the lowest concentration with no visible growth, and assays were performed in triplicate on separate days (Cao et al., 16 Sep 2025).

The reported result is a 100% hit rate: all 100 amphionins had MIC $+2$0 against at least one tested pathogen. Moreover, 99 of 100 were active against two or more pathogens. The abstract additionally states that low MIC values reached the nanomolar range in some cases. Representative potency summaries include median MIC values of 2–4 $+2$1 against susceptible and multidrug‑resistant A. baumannii, 2–8 $+2$2 against multiple E. coli strains including polymyxin‑resistant isolates, approximately 8 $+2$3 against P. aeruginosa PAO1, and 8–16 $+2$4 against MRSA and vancomycin‑resistant E. faecium (Cao et al., 16 Sep 2025).

The most potent amphionins share the sequence features emphasized during training and reward shaping: net charge $+2$5 to $+2$6, Lys‑dominant cationicity, enrichment of hydrophobic residues $+2$7, strong depletion of acidic and polar uncharged residues, low Gly and Pro, and near absence of cysteine. The paper presents these determinants as aligned with both RL reward design and AMP biophysics (Cao et al., 16 Sep 2025).

Mechanistic studies focus on membrane activity. Outer membrane permeabilization was measured by NPN fluorescence, and cytoplasmic membrane depolarization by DiSC$+2$8-5. Many more amphionins showed strong cytoplasmic membrane depolarization than outer membrane permeabilization, and NPN versus DiSC$+2$9-5 responses were largely uncorrelated. Based on MaxRel and AUC, the peptides were grouped into potent, transient, gradual, and weak depolarizers for the inner membrane, and into analogous classes for outer membrane permeabilization. Strong depolarizers were slightly longer, typically 16–20 amino acids, with net charge at least $+9$0, K‑rich composition, and often aromatic residues $+9$1. The paper concludes that the designed amphionins primarily kill via cytoplasmic membrane depolarization rather than necessarily massive outer membrane disruption, and describes two emergent mechanistic classes: “rapid inserters” and “steady disruptors” (Cao et al., 16 Sep 2025).

6. Diversity, steering behavior, and comparative properties

ApexAmphion’s outputs are analyzed for both fidelity to AMP statistics and novelty relative to known peptides. On amino acid composition, Amphion and Amphion‑Screen are reported to have the lowest Jensen–Shannon distance and highest Pearson correlation relative to natural AMP composition among the tested models; the paper gives an example of overall amino acid composition $+9$2 and $+9$3 for Amphorium versus natural AMPs. In ESM2 latent space, UMAP visualizations place ApexAmphion samples close to natural AMP and low‑MIC centroids while still broadly covering the manifold, whereas several baselines are described as either mode‑collapsed or wandering outside the AMP manifold (Cao et al., 16 Sep 2025).

Sequence novelty is quantified with MMseqs2. The vast majority of Amphorium peptides share less than 70% identity with any known AMP, clustering suggests more than $+9$4 unique sequence families, and only a small number of amphionins show partial similarity to existing UniRef proteins. Property distributions for Amphion and Amphorium remain close to those of natural AMPs, while Amphorium‑RL and Amphion‑Screen display the intentional RL biases of increased Lys, Gly, and Leu, decreased Glu and other polar residues, increased charge and hydrophobic moment, and moderate hydrophobicity. ESMFold predictions reportedly yield higher pLDDT scores for ApexAmphion sequences, especially in filtered subsets, suggesting more native‑like folding tendencies in linear or helical forms (Cao et al., 16 Sep 2025).

Comparative analyses are presented against PepCVAE, HydrAMP, diff‑AMP, DeepAMP, and AMP‑Designer. Across amino acid composition, latent distance, and perplexity, the paper states that ApexAmphion significantly outperforms these smaller AMP‑specific generative models. The RL contribution is also evaluated indirectly: Amphorium‑RL shows markedly lower predicted MIC distributions across 11 pathogens in APEX 1.1 analyses, higher pass rates for predicted MIC $+9$5, and more than 50% of sequences passing the HydrAMP low‑MIC filter versus 22% for Amphorium and approximately 14% for AMPSphere (Cao et al., 16 Sep 2025).

These comparisons are used to argue for three design choices: a billion‑scale foundation protein LM plus LoRA, PPO‑based integration of generation and scoring, and a property‑aware objective rather than MIC‑only optimization. The paper does not provide a classical ablation table, but it treats these indirect comparisons as evidence that RL and multi‑objective control improve both predicted potency and property balance.

7. Nomenclature, limitations, and future directions

A potential source of confusion is the name. “APEX” separately denotes the Alloy Property Explorer, a cloud‑native platform for atomistic property calculations in AI‑for‑materials workflows (Li et al., 2024). “Amphion” separately denotes an open‑source toolkit for audio, music, and speech generation, including later v0.2 extensions (Zhang et al., 2023, Li et al., 26 Jan 2025). The Amphion v0.2 overview states explicitly that it does not mention “ApexAmphion” and that any relation is necessarily conceptual or inferred rather than explicit; similarly, the APEX paper describes how such a system could be embedded in something like “ApexAmphion” only as a hypothetical integration (Li et al., 26 Jan 2025, Li et al., 2024). ApexAmphion in the antibiotic context is therefore a distinct platform centered on peptide design (Cao et al., 16 Sep 2025).

The stated limitations of ApexAmphion are specific. ApexMIC and the RL reward use a binary MIC criterion of $+9$6 versus $+9$7 rather than a continuous MIC target, largely because source data are heterogeneous. The current design is broad‑spectrum rather than species‑specific. MIC data originate from diverse assay conditions and media, introducing label noise that may bias the predictor and therefore the RL signal. Experimental validation, while large relative to many de novo AMP studies, covers only 100 peptides out of a 2.1 million sequence virtual library. Hemolysis and toxicity are addressed only indirectly through physicochemical constraints and predicted hemolysis, with no detailed mammalian toxicity, immunogenicity, or ADME modeling integrated into the reward (Cao et al., 16 Sep 2025).

The future directions follow directly from these constraints. The paper proposes pathogen‑specific and condition‑specific design through richer labeled datasets and species‑specific MIC predictors; expanded property objectives for stability, protease resistance, immunogenic motifs, toxicity, hemolysis, and host selectivity; more explicit structure‑aware design using structural constraints or predictions; and extension beyond peptides to non‑peptidic or multimodal antibiotics. More broadly, ApexAmphion is presented as an instance of foundation models plus reinforcement learning converting antibiotic discovery from “generate then screen” to generation under learned constraints. This suggests a generalizable methodology for iterative steering, but the paper’s own evidence is confined to peptide antibiotics and the validation regime it reports (Cao et al., 16 Sep 2025).