Optimization of genetic sequences

A computer-implemented method for generating an optimized codon sequence encoding a target polypeptide for expression in a host organism, the method executed by one or more processors coupled to a non-transitory computer-readable memory, the method comprising:

• (a) receiving an input amino-acid sequence defining the target polypeptide;
• (b) generating, using an autoregressive transformer-neural-network-based codon-generation model stored in the memory, a candidate codon sequence corresponding to the input amino-acid sequence, wherein the autoregressive transformer-neural-network-based model generates each successive codon by:
• (i) computing a probability distribution over a codon vocabulary conditioned on the input amino-acid sequence and all previously generated codons, and
• (ii) applying a synonym-mask constraint at the logit level prior to sampling, the synonym-mask constraint assigning zero probability to codons not synonymous with the amino acid at the current position;
• (c) computing, by the one or more processors, a composite biological-fitness reward for the candidate codon sequence, the composite reward comprising:
• (i) a codon adaptation index (CAI) component computed using a host-organism-specific codon usage table and shaped using a sigmoid function according to:
• CAI_score=1/(1+exp(−k*(CAI_raw−CAI_target))
• wherein CAI_raw is an unnormalized codon adaptation index, CAI_target is a configurable target threshold, and k is a steepness parameter;
• (ii) a GC-content component shaped using a Gaussian penalty function according to:
• GC_penalty=exp(−((GC_actual−GC_target){circumflex over ( )}2/(2*sigma{circumflex over ( )}2)))
• wherein GC_actual is an actual GC percentage of the candidate codon sequence, GC_target is a target GC percentage, and sigma is a configurable width parameter;
• and
• (iii) at least one additional component selected from:
• predicted RNA folding free energy normalized per nucleotide, a codon repetition penalty based on adjacent identical codons, and a codon sequence diversity score based on positional Shannon entropy;
• (d) computing a total scalar reward as a weighted sum of the components in step (c), wherein weights are stored in the memory;
• (e) updating parameters of the autoregressive transformer-neural-network-based model using a REINFORCE policy-gradient algorithm, wherein the REINFORCE algorithm computes a parameter update according to:
• gradient=(reward−baseline)*gradient of log P(candidate codon sequence)
• wherein baseline is an exponential moving average of previous reward values;
• (f) iteratively performing steps (b) through (e) until a convergence criterion is satisfied;
• (g) selecting an optimized codon sequence from generated candidate codon sequences based on cumulative reward;
• and
• (h) outputting the optimized codon sequence in a format selected from FASTA, GenBank, CSV, or JSON, together with associated reward component values.

A computer system for generating optimized codon sequences encoding target polypeptides, the system comprising:

• (a) one or more processors;
• (b) a non-transitory computer-readable memory operably coupled to the one or more processors and storing:
• (i) an autoregressive transformer-neural-network-based codon-generation model comprising an embedding layer, one or more transformer decoder layers, and an output layer configured to produce probability distributions over a codon vocabulary,
• (ii) a synonym-mask constraint table indexed by amino acid, the table listing valid synonymous codons for each amino acid according to a genetic code, and
• (iii) one or more host-organism codon usage tables for computation of codon adaptation index;
• (c) an input interface configured to receive an input amino-acid sequence defining a target polypeptide;
• (d) a codon-generation module configured to execute on the one or more processors and to generate candidate codon sequences using the autoregressive transformer-neural-network-based model, wherein the codon-generation module applies the synonym-mask constraint at each amino-acid position to restrict codon selection to synonymous codons;
• (e) a reward-computation module configured to execute on the one or more processors and to compute a multi-objective biological-fitness reward for each candidate codon sequence, the reward comprising a weighted combination of at least two metrics selected from codon adaptation index, GC content, predicted RNA folding free energy, codon repetition penalty, and codon sequence diversity;
• (f) a reinforcement-learning module configured to update parameters of the autoregressive transformer-neural-network-based model using a policy-gradient algorithm based on the computed reward;
• and
• (g) an output interface configured to export an optimized codon sequence in a machine-readable nucleic-acid sequence format configured for integration with gene synthesis pipelines or bioinformatics tools.

A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to:

• (a) receive an input amino-acid sequence defining a target polypeptide;
• (b) generate, using an autoregressive transformer-neural-network-based codon-generation model, a candidate codon sequence corresponding to the input amino-acid sequence, wherein the autoregressive transformer-neural-network-based model generates each successive codon conditioned on the input amino-acid sequence and previously generated codons;
• (c) apply, at each amino-acid position during generation, a synonym-mask constraint that restricts codon selection to codons synonymous with the amino acid at that position by modifying a probability distribution produced by the model;
• (d) compute a multi-objective biological-fitness reward for the candidate codon sequence, the reward comprising a weighted combination of at least two metrics selected from codon adaptation index, GC content, predicted RNA folding free energy, codon repetition penalty, and codon sequence diversity;
• (e) update parameters of the autoregressive transformer-neural-network-based model using a policy-gradient reinforcement-learning algorithm based on the computed reward;
• (f) iteratively generate candidate codon sequences using the updated model until a convergence criterion is satisfied;
• (g) select an optimized codon sequence based on the computed reward;
• and
• (h) output the optimized codon sequence in a machine-readable format together with associated reward component values.